{"QuestionId":388172,"AnswerCount":6,"Tags":"<python><lua><scipy><scientific-computing><torch>","CreationDate":"2008-12-23T04:25:07.537","AcceptedAnswerId":388184.0,"OwnerUserId":41718.0,"Title":"Scientific libraries for Lua?","Body":"<p>Are there any scientific packages for Lua comparable to Scipy?<\/p>\n","answers":[{"AnswerId":"9254474","CreationDate":"2012-02-13T01:20:28.603","ParentId":null,"OwnerUserId":"221509","Title":null,"Body":"<p>I'm not sure if it is comparable to Scipy, but there is <a href=\"http:\/\/www.nongnu.org\/gsl-shell\/\" rel=\"nofollow\">GSL Shell<\/a> which is based on LuaJIT and GNU Scientific Library, which offers many numerical algorithms and vector\/matrix linear algebra operations.<\/p>\n"},{"AnswerId":"9253469","CreationDate":"2012-02-12T22:52:02.930","ParentId":null,"OwnerUserId":"1205758","Title":null,"Body":"<p>There's a Numpy-like extension for Lua which runs without dependencies at<\/p>\n\n<p><a href=\"https:\/\/github.com\/jzrake\/lunum\" rel=\"nofollow\">https:\/\/github.com\/jzrake\/lunum<\/a><\/p>\n\n<p>In the future it will provide FFT's and linear algebra like Numpy+Scipy. Presently it supports numeric array manipulation like in Numpy.<\/p>\n"},{"AnswerId":"18776865","CreationDate":"2013-09-13T01:33:04.577","ParentId":null,"OwnerUserId":"18403","Title":null,"Body":"<p>You have some options:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/numlua.luaforge.net\/\" rel=\"noreferrer\">Numeric Lua<\/a> - C module for Lua 5.1\/5.2, provides matrices, FFT, complex numbers and others<\/li>\n<li><a href=\"http:\/\/www.nongnu.org\/gsl-shell\/\" rel=\"noreferrer\">GSL Shell<\/a> - Modification of Lua (supports Lua libraries) with a nice syntax. Provides almost everything that Numeric Lua does, plus ODE solvers, plotting capabilities, and other nice things. Has a great documentation.<\/li>\n<li><a href=\"http:\/\/www.scilua.org\/\" rel=\"noreferrer\">SciLua<\/a> - Pure LuaJIT module. Aims to be a complete framework for scientific computing in Lua. Provides vectors and matrices, random numbers \/ distributions, optimization, others. Still in early development.<\/li>\n<li><a href=\"https:\/\/bitbucket.org\/lucashnegri\/lna\" rel=\"noreferrer\">Lua Numerical Algorithms<\/a> - Pure LuaJIT module (uses blas\/lapack via LuaJIT FFI). Provides matrices \/ linear algebra, FFT, complex numbers, optimization algorithms, ODE solver, basic statistics (+ PCA, LDA), and others. Still in early development, but has a somewhat complete documentation and test suits.<\/li>\n<\/ul>\n"},{"AnswerId":"10631596","CreationDate":"2012-05-17T07:30:57.633","ParentId":null,"OwnerUserId":"1400386","Title":null,"Body":"<p>You should try <strong><a href=\"http:\/\/www.torch.ch\" rel=\"noreferrer\">Torch7<\/a><\/strong> (<a href=\"https:\/\/github.com\/andresy\/torch\" rel=\"noreferrer\">github<\/a>).<\/p>\n\n<p>Torch7 has a very nice and efficient vector\/matrix\/tensor numerical library\nwith a Lua front-end. It also has a bunch of functions for computer vision\nand machine learning. <\/p>\n\n<p>It's pretty recent but getting better quickly.<\/p>\n"},{"AnswerId":"388184","CreationDate":"2008-12-23T04:36:07.220","ParentId":null,"OwnerUserId":"33252","Title":null,"Body":"<p>There is the basis for one in <a href=\"http:\/\/numlua.luaforge.net\/\" rel=\"nofollow noreferrer\">Numeric Lua<\/a>.<\/p>\n"},{"AnswerId":"388587","CreationDate":"2008-12-23T10:33:39.823","ParentId":null,"OwnerUserId":"17160","Title":null,"Body":"<p>One can always use <a href=\"http:\/\/labix.org\/lunatic-python\" rel=\"noreferrer\">Lunatic Python<\/a> and access scipy inside lua.<\/p>\n\n<pre><code>&gt; require(\"python\")\n&gt; numpy = python.import(\"numpy\")\n&gt; numpy.array ... etc ..\n<\/code><\/pre>\n"}]}
{"QuestionId":11987325,"AnswerCount":4,"Tags":"<numpy><scipy><gfortran><theano>","CreationDate":"2012-08-16T12:39:30.310","AcceptedAnswerId":null,"OwnerUserId":380038.0,"Title":"Theano fails due to NumPy Fortran mixup under Ubuntu","Body":"<p>I installed <a href=\"http:\/\/deeplearning.net\/software\/theano\/\" rel=\"noreferrer\">Theano<\/a> on my machine, but the nosetests break with a Numpy\/Fortran related error message. For me it looks like Numpy was compiled with a different Fortran version than Theano. I already reinstalled Theano ( + ) and Numpy \/ Scipy (), but this did not help.<\/p>\n\n<p>What steps would you recommend?<\/p>\n\n<h3>Complete error message:<\/h3>\n\n<pre><\/pre>\n\n<h3>My research:<\/h3>\n\n<p>The <a href=\"http:\/\/www.scipy.org\/Installing_SciPy\/BuildingGeneral\" rel=\"noreferrer\">Installing SciPy \/ BuildingGeneral<\/a> page about the  error:<\/p>\n\n<p>If you see an error message<\/p>\n\n<p><\/p>\n\n<p>when building SciPy, it means that NumPy picked up the wrong Fortran compiler during build (e.g. ifort). <\/p>\n\n<p>Recompile NumPy using:<\/p>\n\n<p><\/p>\n\n<p>or whichever is appropriate (see ).<\/p>\n\n<p>But:<\/p>\n\n<pre><\/pre>\n\n<h3>Used software versions:<\/h3>\n\n<ul>\n<li>scipy 0.10.1 (scipy.test() works)<\/li>\n<li>NumPy 1.6.2 (numpy.test() works)<\/li>\n<li>theano 0.5.0 (several tests fails with )<\/li>\n<li>python 2.6.6<\/li>\n<li>Ubuntu 10.10<\/li>\n<\/ul>\n\n<h2>[UPDATE]<\/h2>\n\n<p>So I removed numpy and scipy from my system with  and using  of what was left.<\/p>\n\n<p>Than I installed numpy and scipy from the github sources with .<\/p>\n\n<p>Afterwards I entered again  and . <\/p>\n\n<p>Error persists :\/<\/p>\n\n<p>I also tried the  ... +  approach, but for my old Ubuntu (10.10) it installs too old version of numpy and scipy for theano: <\/p>\n","answers":[{"AnswerId":"18241213","CreationDate":"2013-08-14T20:03:51.183","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>A better fix is to remove atlas and install openblas. openblas is faster then atlas. Also, openblas don't request gfortran and is the one numpy was linked with. So it will work out of the box.<\/p>\n"},{"AnswerId":"18238732","CreationDate":"2013-08-14T17:48:02.127","ParentId":null,"OwnerUserId":"430281","Title":null,"Body":"<p>I had the same problem, and after reviewing the source code, user212658's answer seemed like it would work (I have not tried it). I then looked for a way to deploy user212658's hack without modifying the source code.<\/p>\n\n<p>Put these lines in your <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANORC\">theanorc<\/a> file:<\/p>\n\n<pre><code>[blas]\nldflags = -lblas -lgfortran\n<\/code><\/pre>\n\n<p>This worked for me.<\/p>\n"},{"AnswerId":"12142579","CreationDate":"2012-08-27T13:06:16.947","ParentId":null,"OwnerUserId":"1491200","Title":null,"Body":"<p>Have you tried to recompile NumPy from the sources? <\/p>\n\n<p>I'm not familiar with the Ubuntu package system, so I can't check what's in your <code>dist-packages\/numpy<\/code>. With a clean archive of the <code>NumPy<\/code> sources, you should have a <code>setup.py<\/code> at the same level as the directories <code>numpy<\/code>, <code>tools<\/code> and <code>benchmarks<\/code> (among others). I'm pretty sure that's the one you want to use for a <code>python setup.py build<\/code>.<\/p>\n\n<p><strong>[EDIT]<\/strong><\/p>\n\n<p>Now that you have recompiled <code>numpy<\/code> with the proper <code>--fcompiler<\/code> option, perhaps could you try to do the same with <code>Theano<\/code>, that is, compiling directly from sources <em>without<\/em> relying on a <code>apt-get<\/code> or even <code>pip<\/code>. You should have a better control on the build process that way, which will make debugging\/trying to find a solution easier. <\/p>\n"},{"AnswerId":"13579399","CreationDate":"2012-11-27T07:34:10.480","ParentId":null,"OwnerUserId":"212658","Title":null,"Body":"<p>I had the same problem. The solution I found is to add a hack in theano\/gof\/cmodule.py to link against gfortran whenever 'blas' is in the libs. That fixed it.<\/p>\n\n<pre><code>class GCC_compiler(object):\n   ...\n    @staticmethod\n    def compile_str(module_name, src_code, location=None,\n                    include_dirs=None, lib_dirs=None, libs=None,\n                    preargs=None):\n        ...\n        cmd.extend(['-l%s' % l for l in libs])\n        if 'blas' in libs:\n            cmd.append('-lgfortran')\n<\/code><\/pre>\n"}]}
{"QuestionId":13757826,"AnswerCount":2,"Tags":"<numpy><pypy><theano>","CreationDate":"2012-12-07T06:07:58.497","AcceptedAnswerId":13762698.0,"OwnerUserId":901263.0,"Title":"NumPyPy vs Theano?","Body":"<p>I am wondering: do these two projects basically have the same goal -- to speed up numerical work in Python?<\/p>\n\n<p>What are the similarities and differences?<\/p>\n\n<p>I know that Theano does not aim to re-implement all of NumPy like NumPyPy does, but from what I've read, Theano can already lead to some really impressive speedup results.  So why do we need NumPyPy if we can just write code for Theano that runs fast?<\/p>\n","answers":[{"AnswerId":"20026491","CreationDate":"2013-11-17T02:34:11.297","ParentId":null,"OwnerUserId":"179081","Title":null,"Body":"<p>Theano, seeks to improve NumPy,\nNumPy is a prerequisite for Theano.<\/p>\n\n<p>A large feature of Theano, is it's transparent use of CUDA GPU's, when possible.<\/p>\n"},{"AnswerId":"13762698","CreationDate":"2012-12-07T12:01:37.330","ParentId":null,"OwnerUserId":"1098041","Title":null,"Body":"<p>Well for one thing : millions of lines of code use numpy, so porting Numpy to pypy would be a great step forward for the porting of many other (scientific and other) librairies to Pypy.<\/p>\n\n<p>Re-implementing all Numpy in pypy may sound like a chore, and it is, but the alternative in just insane : re-implementing hundreds or librairies to use XXX instead.<\/p>\n\n<p>And by the way I don't know theano really well, but I know it isn't a substitute for Numpy.\nthey are different projects, with different features.<\/p>\n"}]}
{"QuestionId":14178339,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2013-01-06T01:24:08.887","AcceptedAnswerId":null,"OwnerUserId":199785.0,"Title":"theano define function which repeatedly calls another function?","Body":"<p>My training function:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Then from elsewhere:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>So what I'd like this to look like is<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I'm really not even sure how to start on this, as theano is still pretty mind-bending for me. I was able to get this far but loops are seriously challenging.<\/p>\n\n<p>I have the vague notion that if I can turn the theano.function into a theano.scan, and then put an outer theano.function around it - that might work. However, theano.scan is still magical to me (despite my best efforts). <\/p>\n\n<p>How can I make it so that the looping over minibatches is incorporated into a single function call? <\/p>\n\n<p>Update:<\/p>\n\n<p>I thought I had it! I got this:<\/p>\n\n<pre><\/pre>\n\n<p>But unfortunately it seems like since I use index to calculate the batches in the givens, I can't also update on it:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Update 2:<\/p>\n\n<pre><\/pre>\n\n<p>This actually runs, but it's output is weird:<\/p>\n\n<pre><\/pre>\n\n<p>Everytime I run it I get the same output, even though X &amp; y are initialized to random values each run.<\/p>\n","answers":[{"AnswerId":"16868431","CreationDate":"2013-06-01T02:15:08.277","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Everybody I know do the loop over minibatch in python. This can be done with scan, but all your tries here did not used scan. So it is normal that they didn't worked. You need to call the scan function somewhere to use it (or its higher level interface like map). In fact, in your case, I think you can use <code>theano.scan(fn, theano.tensor.arange(N))<\/code>.<\/p>\n\n<p>I can't answer all your questions in this post as the snippet of code is incomplete, but here is some information:<\/p>\n\n<pre><code>return theano.function(\n    inputs=[], outputs=[self.cost] * n_batches,\n<\/code><\/pre>\n\n<p>Here: <code>[self.cost] * n_batches<\/code> is pure python code. This create a list of <code>n_batches<\/code> element, where each elements are <code>self.cos<\/code>t. So if <code>n_batches<\/code> is 3, you will have <code>outputs=[self.cost, self.cost, self.cost]<\/code>. That is why you got the same value outputed multiple time.<\/p>\n\n<p>I can't tell you why you always add the same answer as I need information that wasn't provided.<\/p>\n"}]}
{"QuestionId":14311688,"AnswerCount":1,"Tags":"<python><warnings><theano>","CreationDate":"2013-01-14T02:40:27.977","AcceptedAnswerId":null,"OwnerUserId":1030101.0,"Title":"how to set the Theano flag warn.sum_div_dimshuffle_bug to False","Body":"<p>I am using the theano package to find the derivative of a sigmoid function, using the cross entropy as the cost. This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>when I run my code, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>But I dont know how to do that. I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>but it is giving me an error on warn, saying it is not recognized as a variable.<\/p>\n","answers":[{"AnswerId":"14631722","CreationDate":"2013-01-31T17:33:43.570","ParentId":null,"OwnerUserId":"359944","Title":null,"Body":"<p>Try <\/p>\n\n<pre><code>from theano import config\nconfig.warn.sum_div_dimshuffle_bug = False\n<\/code><\/pre>\n\n<p>This worked for me<\/p>\n"}]}
{"QuestionId":15542043,"AnswerCount":2,"Tags":"<python><g++><theano>","CreationDate":"2013-03-21T07:51:40.970","AcceptedAnswerId":15542158.0,"OwnerUserId":2193800.0,"Title":"Theano install warning: g++ not detected","Body":"<p>After I installed Theano I tried to run it but got following error message:<\/p>\n\n<pre><\/pre>\n\n<p>Why?<\/p>\n","answers":[{"AnswerId":"37846374","CreationDate":"2016-06-15T21:48:49.483","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>You can try installing mingw (<a href=\"http:\/\/www.mingw.org\/\" rel=\"nofollow noreferrer\">http:\/\/www.mingw.org\/<\/a>) as mentioned in the answer posted here - <a href=\"https:\/\/stackoverflow.com\/questions\/36722975\/theano-g-not-detected\/37846308#37846308\">theano g++ not detected<\/a><\/p>\n"},{"AnswerId":"15542158","CreationDate":"2013-03-21T08:00:17.403","ParentId":null,"OwnerUserId":"1060350","Title":null,"Body":"<p>Try installing the <code>g++<\/code> compiler.<\/p>\n\n<p>Python is rather slow, so for performance-critical parts you need a compiled language such as C++.<\/p>\n"}]}
{"QuestionId":15917849,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2013-04-10T05:42:09.473","AcceptedAnswerId":null,"OwnerUserId":2168406.0,"Title":"How can I assign\/update subset of tensor shared variable in Theano?","Body":"<p>When compiling a function in , a shared variable(say X) can be updated by specifying .\nNow I am trying to update only subset of a shared variable:<\/p>\n\n<pre><\/pre>\n\n<p>The codes will raise a error \"update target must be a SharedVariable\", I guess that means update targets can't be non-shared variables. So is there any way to compile a function to just udpate subset of shared variables?<\/p>\n","answers":[{"AnswerId":"19216429","CreationDate":"2013-10-07T02:56:48.000","ParentId":null,"OwnerUserId":"1319683","Title":null,"Body":"<p>Use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.set_subtensor\" rel=\"noreferrer\">set_subtensor<\/a> or <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.inc_subtensor\" rel=\"noreferrer\">inc_subtensor<\/a>:<\/p>\n\n<pre><code>from theano import tensor as T\nfrom theano import function, shared\nimport numpy\n\nX = shared(numpy.array([0,1,2,3,4]))\nY = T.vector()\nX_update = (X, T.set_subtensor(X[2:4], Y))\nf = function([Y], updates=[X_update])\nf([100,10])\nprint X.get_value() # [0 1 100 10 4]\n<\/code><\/pre>\n\n<p>There's now a page about this in the Theano FAQ: <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/faq_tutorial.html\" rel=\"noreferrer\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/faq_tutorial.html<\/a><\/p>\n"},{"AnswerId":"16739636","CreationDate":"2013-05-24T16:28:42.940","ParentId":null,"OwnerUserId":"380038","Title":null,"Body":"<p>This code should solve your problem:<\/p>\n\n<pre><code>from theano import tensor as T\nfrom theano import function, shared\nimport numpy\n\nX = shared(numpy.array([0,1,2,3,4], dtype='int'))\nY = T.lvector()\nX_update = (X, X[2:4]+Y)\nf = function(inputs=[Y], updates=[X_update])\nf([100,10])\nprint X.get_value()\n# output: [102 13]\n<\/code><\/pre>\n\n<p>And here is the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#using-shared-variables\" rel=\"nofollow\">introduction about shared variables in the official tutorial<\/a>.<\/p>\n\n<p>Please ask, if you have further questions!<\/p>\n"}]}
{"QuestionId":16426641,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2013-05-07T19:04:11.087","AcceptedAnswerId":16446170.0,"OwnerUserId":413345.0,"Title":"Disconnected Input in Gradient of Theano Scan Op","Body":"<p>I have a number of items in groups of varying size. For each of these groups, one (known) item is the \"correct\" one.  There is a function which will assign a score to each of item. This results in a flat vector of item scores, as well as vectors telling the index where each group begins and how big it is.  I wish to do a \"softmax\" operation over the scores in each group to assign the items probabilities, and then take the sum of the logs of the probabilities of the correct answers. Here is a simpler version, where we simply return the score of the correct answer without the softmax and the logarithm.<\/p>\n\n<pre><\/pre>\n\n<p>This correctly calculates the output, but when I attempt to take the gradient with respect to the parameter , I get (paths abbreviated):<\/p>\n\n<pre><\/pre>\n\n<p>Now, the  are constant, so there's no reason to need to take any gradients with respect to it.  Ordinarily you could deal with this by either suppressing s or telling Theano to treat  as a constant in your  call (see the last lines of the sample script). But there doesn't seem to be any way to pass such things down to the internal  calls in the gradient computation for the .<\/p>\n\n<p>Am I missing something? Is these a way to get the gradient computation to work through the ScanOp here?<\/p>\n","answers":[{"AnswerId":"16446170","CreationDate":"2013-05-08T16:56:35.097","ParentId":null,"OwnerUserId":"413345","Title":null,"Body":"<p>This turns out to be a Theano bug as of mid-Feb. 2013 (0.6.0rc-2). It is fixed in the development version on github as of the date of this post.<\/p>\n"}]}
{"QuestionId":16512803,"AnswerCount":1,"Tags":"<python><mingw><neural-network><theano>","CreationDate":"2013-05-12T22:44:11.393","AcceptedAnswerId":16868378.0,"OwnerUserId":420774.0,"Title":"Compilation error when importing Theano on Windows 7","Body":"<p>I'm trying to use Theano on Windows 7. I was able to install Theano and import Theano, but after seeing the warning about not having a C compiler installed I also installed mingw. Now when I try \"import theano\" I get a compilation error. The message is rather long, but the relevant parts (from what I could tell) look like this:<\/p>\n\n<pre><\/pre>\n\n<p>and later in the error message this:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what I'm doing wrong? The files referenced in the compile statement don't exist, so that may be part of the problem, but it doesn't explain why Theano thinks they should be there.<\/p>\n","answers":[{"AnswerId":"16868378","CreationDate":"2013-06-01T02:06:05.347","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Installation on Windows isn't trivial. The simplest way is to use EPD(not the free version, it miss some depenency). Are you an academic? If so, it is free. It install all Theano dependency, so you only need to install Theano after that and we link to MKL for BLAS.<\/p>\n\n<p><a href=\"https:\/\/www.enthought.com\/products\/epd\" rel=\"nofollow\">https:\/\/www.enthought.com\/products\/epd<\/a><\/p>\n\n<p>We had an Theano installer that worked with Anaconda 1.4, but the newer released Anaconda 1.5 broke it. We haven't repaired it yet.<\/p>\n"}]}
{"QuestionId":16528804,"AnswerCount":1,"Tags":"<enthought><epd-python><theano>","CreationDate":"2013-05-13T18:35:02.220","AcceptedAnswerId":null,"OwnerUserId":2378916.0,"Title":"linking to python library using EPD Canopy","Body":"<p>Summary: I'm trying to install the theano python package, and the theano install can't find \"-lpython2.7\" in my EPD Canopy installation.<\/p>\n\n<p>More details: Recently I installed the Enthought EPD Canopy python distribution (64-bit academic) in OS X 10.6.8.  Next I installed pip via \"easy_install pip\".<\/p>\n\n<p>Next I installed Theano via \"sudo pip install theano\".  The install looks OK, but then python -c \"import theano\" fails.  The full output is at <a href=\"https:\/\/gist.github.com\/anonymous\/5548936\" rel=\"nofollow\">https:\/\/gist.github.com\/anonymous\/5548936<\/a>, but it seems like the main point is:<\/p>\n\n<p>Problem occurred during compilation with the command line below:<\/p>\n\n<pre><\/pre>\n\n<h1>===============================<\/h1>\n\n<p>ld: library not found for -lpython2.7\ncollect2: ld returned 1 exit status<\/p>\n\n<p>I've had some discussions with the theano google group, and the main message I get is to look for \"libpython2.7.so\", which I can't find.  I checked \/Users\/rkeisler\/Library\/Enthought\/Canopy_64bit\/User\/lib\/.  Inside is python2.7\/os.py and python2.7\/site-packages, but no \"libpython*\" files.<\/p>\n\n<p>I also did a more thorough check for libpython* files.  The only things I could find were:<\/p>\n\n<pre><\/pre>\n\n<p>Finally, on the EPD Canopy package list, I see \"libpython\" listed.  However, when I try to install libpython using the Canopy package manager, \"libpython\" doesn't appear.  It's not an available package.  I'm not sure where to go from here.<\/p>\n","answers":[{"AnswerId":"16926207","CreationDate":"2013-06-04T19:45:34.950","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Following @RobertKern@ information, it is now fixed in the development version of Theano.<\/p>\n\n<p>To update to the development version, do:<\/p>\n\n<pre><code>pip -U --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n"}]}
{"QuestionId":16641014,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2013-05-20T00:37:34.880","AcceptedAnswerId":16641107.0,"OwnerUserId":2276326.0,"Title":"Theano import error: cannot import name stacklists","Body":"<pre><\/pre>\n\n<p>this is my program.\ni am getting following error.can anybody help<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"16641107","CreationDate":"2013-05-20T00:51:03.093","ParentId":null,"OwnerUserId":"1330293","Title":null,"Body":"<p>You probably have an old version of Theano; <code>stacklist<\/code> was <a href=\"https:\/\/github.com\/mrocklin\/Theano\/commit\/8eefe5a0094e84d470865e42cc16edcbdeda58e3\" rel=\"nofollow\">recently introduced\/renamed<\/a> (a month ago). Sou you should update to the latest\/dev version. If you want to stay in your version try importing <code>tensor_of_scalars<\/code> instead of <code>stacklist<\/code>.<\/p>\n\n<p>To update follow the instructions <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#install\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"16641059","CreationDate":"2013-05-20T00:43:30.710","ParentId":null,"OwnerUserId":"1103045","Title":null,"Body":"<p>This error can be caused by one of two things.<\/p>\n\n<p>The first one is pretty obvious: does <code>theano.tensor<\/code> define a name <code>stacklists<\/code>?  Should it be, for example, <code>stacklist<\/code>?<\/p>\n\n<p>Secondly, it can happen if something else you're importing has already imported the name in a way where doing so again would cause a circular reference.  The second would have to be fixed by looking at your source files.<\/p>\n"}]}
{"QuestionId":16683390,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2013-05-22T04:18:56.527","AcceptedAnswerId":16683692.0,"OwnerUserId":2276326.0,"Title":"Python RuntimeError: Failed to import pydot","Body":"<p>I am learning concepts of logistic regression concepts. When i implement it in python, it shows me some error mentioned below. I am beginner in python. Could anybody help to rectify this error?<\/p>\n\n<p>RuntimeError                              Traceback (most recent call last)\n in ()<\/p>\n\n<pre><\/pre>\n\n<p>C:\\Anaconda\\lib\\site-packages\\theano\\printing.pyc in pydotprint(fct, outfile, compact, format, with_ids, high_contrast, cond_highlight, colorCodes, max_label_size, scan_graphs, var_with_name_simple, print_output_file, assert_nb_all_strings)<\/p>\n\n<pre><\/pre>\n\n<p>RuntimeError: Failed to import pydot. You must install pydot for  to work.<\/p>\n","answers":[{"AnswerId":"39842355","CreationDate":"2016-10-04T00:15:59.297","ParentId":null,"OwnerUserId":"6901690","Title":null,"Body":"<p>I got the same error and I did the following sequence to make it work, in a Python 3:<\/p>\n\n<pre><code>source activate anaconda\npip install pydot\npip install pydotplus\npip install pydot-ng\n<\/code><\/pre>\n\n<p>Then you download and install Graphviz from here according to your OS type:\n<a href=\"http:\/\/www.graphviz.org\/Download..php\" rel=\"nofollow noreferrer\">http:\/\/www.graphviz.org\/Download..php<\/a><\/p>\n\n<p>If you are running Python on Anaconda, open Spyder from terminal, not from Anaconda. Go to terminal and type:<\/p>\n\n<pre><code>spyder\n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n.\n.\n.\nimport pydot\nimport graphviz\nimport pydot_ng as pydot\n<\/code><\/pre>\n\n<p>Develop your model and:<\/p>\n\n<pre><code>theano.printing.pydotprint(prediction, outfile=\"\/Volumes\/Python\/prediction.png\", var_with_name_simple=True)\n<\/code><\/pre>\n\n<p>You will have a picture like this:\n<a href=\"https:\/\/i.stack.imgur.com\/CO8GR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CO8GR.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"21513225","CreationDate":"2014-02-02T16:46:32.297","ParentId":null,"OwnerUserId":"3262258","Title":null,"Body":"<p>I also have the same issue.  I would suggest you post this in the Github Theano Issues forum:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues?direction=desc&amp;sort=updated&amp;state=open\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues?direction=desc&amp;sort=updated&amp;state=open<\/a><\/p>\n\n<p>It seems to me, that since this instance of the pydotprint() function is actually part of the printing module within the Theano library that this should not be an issue (but it is) and hence it should be brought to the attention of the developers in order to fix it.<\/p>\n\n<p>Please correct me if this is not the case.<\/p>\n"},{"AnswerId":"16683692","CreationDate":"2013-05-22T04:51:24.647","ParentId":null,"OwnerUserId":"2318122","Title":null,"Body":"<p>It mainly depends on where you put the pydot files. If you are running it straight from the Python Shell then you should have them installed in the modules folder which is most commonly the \"Lib\" folder inside the main python folder.<\/p>\n"}]}
{"QuestionId":16738937,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2013-05-24T15:49:47.633","AcceptedAnswerId":16782594.0,"OwnerUserId":380038.0,"Title":"Using Theano.scan with shared variables","Body":"<p>I want to calculate the <a href=\"http:\/\/office.microsoft.com\/en-001\/excel-help\/sumproduct-HP005209293.aspx\" rel=\"nofollow\"><\/a> of two arrays in Theano. Both arrays are declared as shared variables and are the result of prior computations. Reading the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">tutorial<\/a>, I found out how to use scan to compute what I want using 'normal' tensor arrays, but when I tried to adapt the code to shared arrays I got the error message . (See minimal running code example below)<\/p>\n\n<p>Where is the mistake in my code? Where is my misconception? I am also open to a different approach for solving my problem.<\/p>\n\n<p>Generally I would prefer a version which takes the shared variables directly, because in my understanding, converting the arrays first back to Numpy arrays and than again passing them to Theano, would be wasteful. <\/p>\n\n<p>Error message producing  code using <strong>shared<\/strong> variables:<\/p>\n\n<pre><\/pre>\n\n<p>Error message: <\/p>\n\n<pre><\/pre>\n\n<p>Working  code using <strong>non-shared<\/strong> variables:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"16782594","CreationDate":"2013-05-28T01:33:56.643","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You need to change the compilation line to this one:<\/p>\n\n<pre><code>Tsumprod = theano.function([], outputs=Tsumprod_result)\n<\/code><\/pre>\n\n<p>theano.function() always need a list of inputs. If the function take 0 input, like in this case, you need to give an empty list for the inputs.<\/p>\n"}]}
{"QuestionId":16787358,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2013-05-28T08:29:27.513","AcceptedAnswerId":16787359.0,"OwnerUserId":380038.0,"Title":"Theano TypeError: function() takes at least 1 argument (1 given)","Body":"<p>One of my Theano functions does not take any inputs and only uses shared variables to calculate the output. But this function throws a . <\/p>\n\n<p>Here a minimal example: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"16787359","CreationDate":"2013-05-28T08:29:27.513","ParentId":null,"OwnerUserId":"380038","Title":null,"Body":"<p>From: <a href=\"https:\/\/stackoverflow.com\/a\/16782594\/380038\">https:\/\/stackoverflow.com\/a\/16782594\/380038<\/a><\/p>\n\n<p>\"<code>theano.function()<\/code> always needs a list of inputs. If the function takes 0 input, like in this case, you need to give an empty list for the inputs.\"<\/p>\n\n<p><code>f2 = th.function(outputs=z)<\/code> has to be <code>f2 = th.function([], outputs=z)<\/code><\/p>\n"}]}
{"QuestionId":16810371,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2013-05-29T09:36:14.473","AcceptedAnswerId":16815974.0,"OwnerUserId":380038.0,"Title":"Using Theano.scan with multidimensional arrays","Body":"<p>To speed up my code I am converting a multidimensional sumproduct function from Python to Theano. My Theano code reaches the same result, but only calculates the result for one dimension at a time, so that I have to use a Python for-loop to get the end result. I assume that would make the code slow, because Theano cannot optimize memory usage and transfer (for the gpu) between multiple function calls. Or is this a wrong assumption?<\/p>\n\n<p>So how can I change the Theano code, so that the sumprod is calculated in one function call?<\/p>\n\n<p>The original Python function:<\/p>\n\n<pre><\/pre>\n\n<p>For the following input <\/p>\n\n<pre><\/pre>\n\n<p>the output would be:  that is 1*1 + 5*5, 2*2 + 6*6 and 4*4 + 7*7<\/p>\n\n<p>The Theano version of the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"16815974","CreationDate":"2013-05-29T14:00:42.667","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>First, there is more people that will answer your questions on theano mailing list then on stackoverflow. But I'm here:)<\/p>\n\n<p>First, your function isn't a good fit for GPU. Even if everything was well optimized, the transfer of the input to the gpu just to add and sum the result will take more time to run then the python version.<\/p>\n\n<p>Your python code is slow, here is a version that should be faster:<\/p>\n\n<pre><code>def sumprod(a1, a2):\n    \"\"\"Sum the element-wise products of the `a1` and `a2`.\"\"\"\n    a1 = numpy.asarray(a1)\n    a2 = numpy.asarray(a2)\n    result (a1 * a2).sum(axis=0)\n    return result\n<\/code><\/pre>\n\n<p>For the theano code, here is the equivalent of this faster python version(no need of scan)<\/p>\n\n<pre><code>m1 = theano.tensor.matrix()\nm2 = theano.tensor.matrix()\nf = theano.function([m1, m2], (m1 * m2).sum(axis=0))\n<\/code><\/pre>\n\n<p>The think to remember from this is that you need to \"vectorize\" your code. The \"vectorize\" is used in the NumPy context and it mean to use numpy.ndarray and use function that work on the full tensor at a time. This is always faster then doing it with loop (python loop or theano scan). Also, Theano optimize some of thoses cases by moving the computation outside the scan, but it don't always do it.<\/p>\n"}]}
{"QuestionId":16848650,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2013-05-31T01:54:35.713","AcceptedAnswerId":16860115.0,"OwnerUserId":1658908.0,"Title":"Theano: Why does indexing fail in this case?","Body":"<p>I'm trying to get the max of a vector given a boolean value.<\/p>\n\n<p>With Numpy:<\/p>\n\n<pre><\/pre>\n\n<p>But with Theano:<\/p>\n\n<pre><\/pre>\n\n<p>Why does this happen? Is this a subtle nuance that i'm missing?<\/p>\n","answers":[{"AnswerId":"16860115","CreationDate":"2013-05-31T14:51:21.860","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You are using a version of Theano that is too old. In fact, tensor_var.nonzero() isn't in any released version. You need to update to the development version.<\/p>\n\n<p>With the development version I have this:<\/p>\n\n<pre><code>&gt;&gt;&gt; that[~(that&gt;=5).nonzero()].max().eval()\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nTypeError: bad operand type for unary ~: 'tuple'\n<\/code><\/pre>\n\n<p>This is because you are missing parenthesis in your line. Here is the good line:<\/p>\n\n<pre><code>&gt;&gt;&gt; that[(~(that&gt;=5)).nonzero()].max().eval()\narray(9, dtype=int32)\n<\/code><\/pre>\n\n<p>But we still have unexpected result! The problem is that Theano do not support bool. Doing ~ on int8, is doing the bitwise invert on 8 bits, not 1 bit. It give this result:<\/p>\n\n<pre><code>&gt;&gt;&gt; (that&gt;=5).eval()\narray([0, 0, 0, 0, 0, 1, 1, 1, 1, 1], dtype=int8)\n&gt;&gt;&gt; (~(that&gt;=5)).eval()\narray([-1, -1, -1, -1, -1, -2, -2, -2, -2, -2], dtype=int8)\n<\/code><\/pre>\n\n<p>You can remove the ~ with this:<\/p>\n\n<pre><code>&gt;&gt;&gt; that[(that&lt;5).nonzero()].max().eval()\narray(4, dtype=int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":17026496,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2013-06-10T14:39:54.977","AcceptedAnswerId":17026497.0,"OwnerUserId":380038.0,"Title":"Theano shared variable update causes `ValueError: length not known`","Body":"<p>Minimal example code: <\/p>\n\n<pre><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><\/pre>\n\n<p>What is the cause of this error?<\/p>\n","answers":[{"AnswerId":"17026497","CreationDate":"2013-06-10T14:39:54.977","ParentId":null,"OwnerUserId":"380038","Title":null,"Body":"<p><code>Updates<\/code> must contain <code>a list of pairs<\/code>. See official tutorial on <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#using-shared-variables\" rel=\"nofollow\">using shared variables<\/a>.<\/p>\n\n<p>Correct code:<\/p>\n\n<pre><code>import theano as th\nimport theano.tensor as T\nimport numpy as np\n\nx = T.dscalars('x')\nz = th.shared(np.zeros(2))\nupdates = [(z, z+x)]\n\nf1 = th.function(inputs=[x], updates=updates) \nf1(3)\nprint z.get_value()\n<\/code><\/pre>\n"}]}
{"QuestionId":17039908,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2013-06-11T08:45:59.827","AcceptedAnswerId":null,"OwnerUserId":380038.0,"Title":"How to reuse Theano function with different shared variables without rebuilding graph?","Body":"<p>I have a Theano function that is called several times, each time with different shared variables. The way it is implemented now, the Theano function gets redefined every time it is run. I assume, that this make the whole program slow, because every time the Theano functions gets defined the graph is rebuild.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>For non shared (normal) variables I can define the function once and then call it with different variables without redefining. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Is this possible also for shared variables?<\/p>\n","answers":[{"AnswerId":"17296524","CreationDate":"2013-06-25T11:49:33.330","ParentId":null,"OwnerUserId":"857617","Title":null,"Body":"<p>You can use the givens keyword in theano.function for that. Basically, you do the following.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>m1 = theano.shared(name='m1', value = np.zeros((3,2)) )\nm2 = theano.shared(name='m2', value = np.zeros((3,2)) )\n\nx1 = theano.tensor.dmatrix('x1')\nx2 = theano.tensor.dmatrix('x2')\n\ny = (x1*x2).sum(axis=0)\nf = theano.function([],y,givens=[(x1,m1),(x2,m2)],on_unused_input='ignore')\n<\/code><\/pre>\n\n<p>then to loop through values you just set the value of the shared variables to the value you'd like. You have to set the on_unused_input to 'ignore' to use functions with no arguments in theano, by the way. Like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>array1 = array([[1,2,3],[4,5,6]])\narray2 = array([[2,4,6],[8,10,12]])\n\nfor i in range(10):\n    m1.set_value(i*array1)\n    m2.set_value(i*array2)\n    print f()\n<\/code><\/pre>\n\n<p>It should work, at least that's how I've been working around it.<\/p>\n"},{"AnswerId":"17045919","CreationDate":"2013-06-11T13:54:41.687","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Currently it is not easily possible to reuse a Theano function with different shared variable.<\/p>\n\n<p>But you have alternative:<\/p>\n\n<ol>\n<li>Is it really a bottleneck? In the example, it is, but I suppose it is a simplified case. The only way to know is to profile it.<\/li>\n<li>You compile 1 Theano function with the first shared variable. Then you can call the get_value\/set_value on those shared variables before calling the Theano function. This way, you won't need to recompile the Theano function.<\/li>\n<\/ol>\n"}]}
{"QuestionId":17125845,"AnswerCount":4,"Tags":"<python-2.7><enthought><win64><theano>","CreationDate":"2013-06-15T16:48:56.567","AcceptedAnswerId":null,"OwnerUserId":1961354.0,"Title":"Installing Theano on EPD (Windows x64), g++ not detected","Body":"<p>I am trying to run theano on Enthought Python Distribution (academic license) under windows 7 64.\nFollowin the topic <a href=\"https:\/\/stackoverflow.com\/questions\/10270871\/installing-theano-on-epd-windows-x64\">Installing Theano on EPD (Windows x64)<\/a> I installed bleeding edge version of theano since I got the same error. But now I have this problem:<\/p>\n\n<pre><\/pre>\n\n<p>EPD installs its own version of mingw, so I do not uderstand why the problem occurs.\nI tried to find g++ (assuming EPD installed it) through window search to put in PATH but there is nothing. <\/p>\n\n<p>I've separatly installed mingw64, but when I type in command prompt<\/p>\n\n<pre><\/pre>\n\n<p>it's hanging<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"17134799","CreationDate":"2013-06-16T15:24:24.933","ParentId":null,"OwnerUserId":"1961354","Title":null,"Body":"<p>I solved this problem by adding Visual C++ Compilers feature to my current VS2010 installtion.\nNow I can import theano and console shows that I use gpu<\/p>\n\n<pre><code>&gt;&gt;import theano\nForcing DISTUTILS_USE_SDK=1\nUsing gpu device 0: GeForce GT 630M\n<\/code><\/pre>\n\n<p>But when I am trying to run this code:<\/p>\n\n<pre><code>from theano import function, config, shared, sandbox\nimport theano.tensor as T\nimport numpy\nimport time\n\nvlen = 10 * 30 * 768  # 10 x #cores x # threads per core\niters = 1000\n\nrng = numpy.random.RandomState(22)\nx = shared(numpy.asarray(rng.rand(vlen), config.floatX))\nf = function([], T.exp(x),mode='DebugMode')\n<\/code><\/pre>\n\n<p>I get \n<code>NVCC: nvcc : fatal error : Could not set up the environment for Microsoft Visual Studio using 'c:\/Program Files (x86)\/Microsoft Visual Studio 10.0\/VC\/bin\/..\/..\/VC\/bin\/amd64\/vcvars64.bat<\/code><\/p>\n"},{"AnswerId":"37846439","CreationDate":"2016-06-15T21:53:14.907","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>Instead of using git command to install Theano try downloading the zip from the theano repository on GitHub. To install theano, use <code>python setup.py install<\/code> command. \nAlso try using Anaconda distribution to install Python3.4 or older versions. Then use <code>conda install<\/code> command to install mingw for g++ support.<\/p>\n"},{"AnswerId":"17132910","CreationDate":"2013-06-16T11:38:28.800","ParentId":null,"OwnerUserId":"1961354","Title":null,"Body":"<p>The problem was that I installed Enthought Canopy and it does not contain mingw. The issue can be solved by installing Enthought Python Distribution. Following <a href=\"https:\/\/stackoverflow.com\/questions\/2970493\/cuda-linking-error-visual-express-2008-nvcc-fatal-due-to-null-configuratio\">CUDA linking error - Visual Express 2008<\/a> I created vcvars64.bat in \nc:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\bin\\amd64\\ to avoid nvcc fatal   : Visual Studio configuration file '(null)' error.\n But now I get this exception:<\/p>\n\n<pre><code>c:\\program files (x86)\\microsoft visual studio 10.0\\vc\\include\\codeanalysis\\sourceannotations.h(29): error:  invalid redeclaration of type name \"size_t\"\n<\/code><\/pre>\n"},{"AnswerId":"17149578","CreationDate":"2013-06-17T14:06:54.983","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem isn't with Theano, but with nvcc. For Theano to use the GPU, it need a working nvcc installation. But it is not the case currently.<\/p>\n\n<p>To help you fix this problem, try to compile nvcc example. They will also fail. When you fix this problem, Theano will work. For this, check nvcc installation\/tests documentations.<\/p>\n\n<p>I suspecct that you didn't used the right microsoft compiler version. nvcc don't accept version of msvc.<\/p>\n"}]}
{"QuestionId":17323040,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2013-06-26T14:33:11.717","AcceptedAnswerId":17358567.0,"OwnerUserId":1323010.0,"Title":"What is the prupose\/meaning of passing \"input\" to a function in Theano?","Body":"<p>Example will make that clearer I hope, (This is a Logistic Regression object, the Theano Tensor library is imported as T)<\/p>\n\n<pre><\/pre>\n\n<p>Which is called down in main...<\/p>\n\n<pre><\/pre>\n\n<p>If these snippits aren't enough to get an understanding, the code is on this page under \"Putting it All Together\"-\n<a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg<\/a><\/p>\n","answers":[{"AnswerId":"17323199","CreationDate":"2013-06-26T14:39:56.270","ParentId":null,"OwnerUserId":"64250","Title":null,"Body":"<p>This is a Python feature called <a href=\"http:\/\/www.diveintopython.net\/power_of_introspection\/optional_arguments.html\" rel=\"nofollow\">named parameters<\/a>. For functions with optional parameters or many parameters it is helpful to pass the parameters by name, instead of just relying on the order on which they were passed to the function. In your specific case you can see the meaning of the <code>input<\/code> parameter <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#creating-a-logisticregression-class\" rel=\"nofollow\">here<\/a>. <\/p>\n"},{"AnswerId":"17323872","CreationDate":"2013-06-26T15:09:31.163","ParentId":null,"OwnerUserId":"1634191","Title":null,"Body":"<p>Named parameters, or default keyword arguments, like <code>input<\/code>, <code>n_in<\/code>, and <code>n_out<\/code> are useful for several reasons.<\/p>\n\n<ul>\n<li>If a function\/method have many parameters, it becomes easier to pass them by name instead of having to remember the functional order of the parameters.<\/li>\n<li>Many functions\/methods have default use cases which are used often, and specialty use cases which are used rarely. If the specialty use case requires passing additional arguments to the function, those will most likely take the form of named parameters with default values. This way, when the function is used in the default use case, the user does not have to specify any additional parameters. Only when someone wants to use the specialty case will they have to specify something extra. This keeps function and method calls readable and simple when they aren't used in complex or specialty ways.<\/li>\n<\/ul>\n"},{"AnswerId":"17358567","CreationDate":"2013-06-28T06:29:44.473","ParentId":null,"OwnerUserId":"857617","Title":null,"Body":"<p>so... Theano builds graphs for the expressions it computes before evaluating them. By passing a theano variable such as 'x' in the example to the initialization of the logistic regression object, you will create a number of expressions such as p_y_given_x in your object which are theano expressions dependent on x. This is later used for symbolic gradient calculation.<\/p>\n\n<p>To get a better feel for it you can do the following:<\/p>\n\n<p><\/p>\n\n<pre><code>import theano.pp #pp is for pretty print\nx = T.dmatrix('x') #naming your variables is a good idea, and important i think\nlr = LogisticRegression(x,n_in = 28*28, n_out= 10)\nprint pp(lr.p_y_given_x)\n<\/code><\/pre>\n\n<p>This should given you an output such as<\/p>\n\n<pre><code>softmax( W \\dot x + b)\n<\/code><\/pre>\n\n<p>And while you're at it go ahead and try out<\/p>\n\n<p><\/p>\n\n<pre><code>print pp(T.grad(lr._y_given_x,x)) #might need syntax checkng\n<\/code><\/pre>\n\n<p>which is how theano internally stores the expression. Then you can use these expressions to create functions in theano, such as<\/p>\n\n<p><\/p>\n\n<pre><code>values = theano.shared( value = mydata, name = 'values')\nf = theano.function([],lr.p_y_given_x , \n                    givens ={x:values},on_unused_input='ignore')\nprint f()\n<\/code><\/pre>\n\n<p>then calling f should give you the predicted class probabilities for the values defined in mydata. The way to do this in theano (and the way it's done in the DL tutorials) is by passing a \"dummy\" theano variable and then using the \"givens\" keyword to set it to a shared variable containing your data. That's important because storing your variables in a shared variable allows theano to use your GPU for matrix operations.<\/p>\n"}]}
{"QuestionId":17426742,"AnswerCount":2,"Tags":"<python><numpy><scipy><sparse-matrix><theano>","CreationDate":"2013-07-02T13:20:51.677","AcceptedAnswerId":null,"OwnerUserId":2542699.0,"Title":"How to calculate (1 - SparseMatrix) of a huge sparse matrix?","Body":"<p>I researched a lot on this but couldn't find a practical solution to this problem. I am using scipy to create csr sparse matrix and want to substract this matrix from an equivalent matrix of all ones. In scipy and numpy notations, if matrix is not sparse, we can do so by simply writing 1 - MatrixVariable. However, this operation is not implemented if Matrix is sparse.  I could just think of the following obvious solution:<\/p>\n\n<p>Iterate through the entire sparse matrix, set all zero elements to 1 and all non-zero elements to 0.<\/p>\n\n<p>But this would create a matrix where most elements are 1 and only a few are 0, which is no longer sparse and due its huge size could not be converted to dense.  <\/p>\n\n<p>What could be an alternative and effective way of doing this?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"17426874","CreationDate":"2013-07-02T13:26:35.310","ParentId":null,"OwnerUserId":"832621","Title":null,"Body":"<p>You can access the data from your sparse matrix as a <code>1D array<\/code> so that:<\/p>\n\n<pre><code>ss.data *= -1\nss.data += 1\n<\/code><\/pre>\n\n<p>will work like <code>1 - ss<\/code>, for all non-zero elements in your sparse matrix.<\/p>\n"},{"AnswerId":"17427551","CreationDate":"2013-07-02T13:57:01.747","ParentId":null,"OwnerUserId":"110026","Title":null,"Body":"<p>Your new matrix will not be sparse, because it will have <code>1<\/code>s everywhere, so you will need a dense array to hold it:<\/p>\n\n<pre><code>new_mat = np.ones(sps_mat.shape, sps_mat.dtype) - sps_mat.todense()\n<\/code><\/pre>\n\n<p>This requires that your matrix fits in memory. It actually requires that it fits in memory 3 times. If that is an issue, you can get it to be more efficient doing something like:<\/p>\n\n<pre><code>new_mat = sps_mat.todense()\nnew_mat *= -1\nnew_mat += 1\n<\/code><\/pre>\n"}]}
{"QuestionId":17445280,"AnswerCount":4,"Tags":"<python><debugging><theano>","CreationDate":"2013-07-03T10:13:41.920","AcceptedAnswerId":17453598.0,"OwnerUserId":869402.0,"Title":"theano - print value of TensorVariable","Body":"<p><strong>How can I print the numerical value of a theano TensorVariable?<\/strong>\nI'm new to theano, so please be patient :)<\/p>\n\n<p>I have a function where I get  as a parameter.\nNow I want to debug-print the shape of this  to the console.\nUsing<\/p>\n\n<pre><\/pre>\n\n<p>results in the console output (i was expecting numbers, i.e. ):<\/p>\n\n<pre><\/pre>\n\n<p>Or how can I print the numerical result of for example the following code (this counts how many values in  are bigger than half the maximum):<\/p>\n\n<pre><\/pre>\n\n<p> should be a single number because  sums up all the values.\nBut using<\/p>\n\n<pre><\/pre>\n\n<p>gives me (expected something like ):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38538089","CreationDate":"2016-07-23T04:12:09.713","ParentId":null,"OwnerUserId":"615130","Title":null,"Body":"<p>print Value of a Tensor Variable.<\/p>\n\n<p>Do the following:<\/p>\n\n<p><code>print tensor[dimension].eval()<\/code> # this will print the content\/value at that position in the Tensor<\/p>\n\n<p>Example, for a 1 d tensor:<\/p>\n\n<pre><code>print tensor[0].eval()\n<\/code><\/pre>\n"},{"AnswerId":"29963238","CreationDate":"2015-04-30T08:54:07.183","ParentId":null,"OwnerUserId":"828368","Title":null,"Body":"<p>For future readers: the previous answer is quite good. \nBut, I found the 'tag.test_value' mechanism more beneficial for debugging purposes (see <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#using-test-values\" rel=\"noreferrer\">theano-debug-faq<\/a>):<\/p>\n\n<pre><code>from theano import config\nfrom theano import tensor as T\nconfig.compute_test_value = 'raise'\nimport numpy as np    \n#define a variable, and use the 'tag.test_value' option:\nx = T.matrix('x')\nx.tag.test_value = np.random.randint(100,size=(5,5))\n\n#define how y is dependent on x:\ny = x*x\n\n#define how some other value (here 'errorCount') depends on y:\nerrorCount = T.sum(y)\n\n#print the tag.test_value result for debug purposes!\nerrorCount.tag.test_value\n<\/code><\/pre>\n\n<p>For me, this is much more helpful; e.g.,  checking correct dimensions etc.<\/p>\n"},{"AnswerId":"17453598","CreationDate":"2013-07-03T16:40:18.020","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>If y is a theano variable, y.shape will be a theano variable. so it is normal that <\/p>\n\n<pre><code>print y.shape\n<\/code><\/pre>\n\n<p>return:<\/p>\n\n<pre><code>Shape.0\n<\/code><\/pre>\n\n<p>If you want to evaluate the expression y.shape, you can do:<\/p>\n\n<pre><code>y.shape.eval()\n<\/code><\/pre>\n\n<p>if <code>y.shape<\/code> do not input to compute itself(it depend only on shared variable and constant). Otherwise, if <code>y<\/code> depend on the <code>x<\/code> Theano variable you can pass the inputs value like this:<\/p>\n\n<pre><code>y.shape.eval(x=numpy.random.rand(...))\n<\/code><\/pre>\n\n<p>this is the same thing for the <code>sum<\/code>. Theano graph are symbolic variable that do not do computation until you compile it with <code>theano.function<\/code> or call <code>eval()<\/code> on them.<\/p>\n\n<p><strong>EDIT:<\/strong> Per the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/gof\/graph.html#theano.gof.graph.Variable.eval\" rel=\"noreferrer\">docs<\/a>, the syntax in newer versions of theano is<\/p>\n\n<pre><code>y.shape.eval({x: numpy.random.rand(...)})\n<\/code><\/pre>\n"},{"AnswerId":"51192683","CreationDate":"2018-07-05T13:32:34.520","ParentId":null,"OwnerUserId":"1712419","Title":null,"Body":"<p>Use <code>theano.printing.Print<\/code> to add print operator to your computational graph.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import numpy\nimport theano\n\nx = theano.tensor.dvector('x')\n\nx_printed = theano.printing.Print('this is a very important value')(x)\n\nf = theano.function([x], x * 5)\nf_with_print = theano.function([x], x_printed * 5)\n\n#this runs the graph without any printing\nassert numpy.all( f([1, 2, 3]) == [5, 10, 15])\n\n#this runs the graph with the message, and value printed\nassert numpy.all( f_with_print([1, 2, 3]) == [5, 10, 15])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<p><code>this is a very important value __str__ = [ 1.  2.  3.]<\/code><\/p>\n\n<p>Source: <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#how-do-i-print-an-intermediate-value-in-a-function\" rel=\"nofollow noreferrer\">Theano 1.0 docs: \u201cHow do I Print an Intermediate Value in a Function?\u201d<\/a><\/p>\n"}]}
{"QuestionId":17641566,"AnswerCount":1,"Tags":"<python><eclipse><pydev><theano><mnist>","CreationDate":"2013-07-14T17:00:30.507","AcceptedAnswerId":17641624.0,"OwnerUserId":1908423.0,"Title":"PyDev can't find machine learning data","Body":"<p>I have a problem I hoped someone would be able to help me with regarding the tutorial at <a href=\"http:\/\/deeplearning.net\/tutorial\/gettingstarted.html#gettingstarted\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/gettingstarted.html#gettingstarted<\/a><\/p>\n\n<p>I keep receiving an error when I try to run the code to load the data set, that is this code here: <\/p>\n\n<pre><\/pre>\n\n<p>I am using Eclipse with PyDev and have numpy, Scipy and Theano working. I ran the command to clone the git repository and have downloaded the data set as per the instructions, however running the code above still returns <\/p>\n\n<pre><\/pre>\n\n<p>I am new to python in general and this really has me stumped as I am not even sure what the cause of the problem could be, nor how what to search for in order to resolve it myself.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"17641624","CreationDate":"2013-07-14T17:07:12.220","ParentId":null,"OwnerUserId":"506544","Title":null,"Body":"<p>The file mnist.pkl.gz is probably not in the same directory as the script that you are trying to run.<\/p>\n\n<p>You would be better off receiving the actual location of the file as a command line parameter to the script and then load the file using that path<\/p>\n"}]}
{"QuestionId":17678905,"AnswerCount":1,"Tags":"<c++><tensorflow><ros><ofstream>","CreationDate":"2013-07-16T14:12:00.800","AcceptedAnswerId":17679464.0,"OwnerUserId":1569693.0,"Title":"writing a tf::transform object to a file","Body":"<p>I'm attempting to do something resembling the following block of code:<\/p>\n\n<pre><\/pre>\n\n<p>Assuming that I've properly set up that  and  when I'm trying to write  to , how would I do so? I tried a number of variants of this, and all of them result in a huge wall of text to the effect that  doesn't know how to handle a tf::transform object, which isn't too surprising.<\/p>\n\n<p>Is there some way to make  take arbitrary objects? Is there some format that I could readily convert it to that's more conducive to streaming? Ideally, if I convert it, I'd like to have a way to reversibly convert it to some matrix that I can pipe straight into and out of a file.<\/p>\n","answers":[{"AnswerId":"17679464","CreationDate":"2013-07-16T14:35:20.347","ParentId":null,"OwnerUserId":"1648011","Title":null,"Body":"<p>Implement the operator<br>\nI'm not sure of the contents of the transform struct in this case, but assuming it is:<\/p>\n\n<pre><code>struct transform { float mat[16]; }\n<\/code><\/pre>\n\n<p>Then the implementation can be something like:<\/p>\n\n<pre><code>std::ostream&amp; operator&lt;&lt; (std::ostream&amp; os, const tf::transform&amp; t)\n{\n  os &lt;&lt; t.mat[0];\n  for(int i=1;i&lt;16;++i) os &lt;&lt; ',' &lt;&lt; t.mat[i];\n  return os;\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":17857737,"AnswerCount":1,"Tags":"<python><numpy><atlas><theano>","CreationDate":"2013-07-25T12:14:23.193","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"Running Python code for theano: \/usr\/bin\/ld: cannot find -latlas","Body":"<p>I am trying to run theano on ubuntu which requires .<\/p>\n\n<p>I have already installed libatlas but I can find it in <\/p>\n\n<p>I have also copied all of the files to a new folder called :<\/p>\n\n<pre><\/pre>\n\n<p>But still, when I run the python code I see:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried adding to environment variables but didn't work:<\/p>\n\n<pre><\/pre>\n\n<p>Also I tried adding the path path to ld file:<\/p>\n\n<pre><\/pre>\n\n<p>or<\/p>\n\n<pre><\/pre>\n\n<p>None of them worked and I still see the error running the Python code.<\/p>\n","answers":[{"AnswerId":"17859308","CreationDate":"2013-07-25T13:23:44.903","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>To change how Theano link to BLAS, you need to use Theano flags[1]. They can be set with an environment variable THEANO_FLAGS or with a configuration file.<\/p>\n\n<p>How did you told Theano to use atlas? If you just installed the atlas packages, it won't work. You need to install the libatlas-dev pacakge as per this Theano installation instruction for Ubuntu[2]<\/p>\n\n<p>A last point, we don't recommand atlas, especially for Ubuntu. OpenBLAS is packaged for Unbuntu and is faster. See [2] for detail on how to installed them. You will need to remove atlas before installing openblas, otherwise, there will be conflict.<\/p>\n\n<p>[1]<a href=\"http:\/\/www.deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANO_FLAGS\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANO_FLAGS<\/a>\n[2]<a href=\"http:\/\/www.deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu<\/a><\/p>\n"}]}
{"QuestionId":18006270,"AnswerCount":1,"Tags":"<python><function><theano>","CreationDate":"2013-08-01T23:22:13.783","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"Function declaration in Python","Body":"<p>I am using Theano in Python.  I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I cannot find any declaration of the  function, while I can only find a piece of code before the previous one as:<\/p>\n\n<pre><\/pre>\n\n<p>I can find the declaration of  but the two functions ( and ) do not have the same signature (input parameters).<\/p>\n\n<p>What does this mean, and is the second code segment the declaration of ?<\/p>\n","answers":[{"AnswerId":"18006380","CreationDate":"2013-08-01T23:32:20.517","ParentId":null,"OwnerUserId":"1053992","Title":null,"Body":"<p>Mostly when name starts with uppercase letter then it is class name (unless it is completely upper cased, then it mostly means constant). There is no <code>new<\/code> operator in Python so this naming pattern is pretty important to distinguish functions from classes (anyway it is general pattern). I don't know Theano, but I suppose that <code>TrainFn1Member<\/code> is a class that implements <code>__call__<\/code> method, so you can call it's instance like a function. Search for <code>__call__<\/code> in <code>TrainFn1Member<\/code> class definition.<\/p>\n\n<hr>\n\n<p>UPDATE:<\/p>\n\n<p>According to your comment, <code>TrainFn1Member<\/code> is a function (what is pretty strange according to what I said above and what is not my idea ;)). In this case it has to return some <code>callable<\/code> what means that it returns one of 3 things (I hope I have not missed anything): <\/p>\n\n<ol>\n<li>function (<code>def<\/code> or <code>lambda<\/code>)<\/li>\n<li>instance of class that implements <code>__call__<\/code><\/li>\n<li>method of some object (function bound to some class instance)<\/li>\n<\/ol>\n\n<p>As I don't know <code>Theano<\/code> at all, I can only suggest to search deeper being aware of those above.. (and welcome to Hogwarts ;))<\/p>\n"}]}
{"QuestionId":18165131,"AnswerCount":2,"Tags":"<eclipse><gpu><pydev><theano>","CreationDate":"2013-08-10T18:32:54.210","AcceptedAnswerId":null,"OwnerUserId":1908423.0,"Title":"Getting Theano to use the GPU","Body":"<p>I am having quite a bit of trouble setting up Theano to work with my graphics card - I hope you guys can give me a hand.<\/p>\n\n<p>I have used CUDA before and it is properly installed as would be necessary to run Nvidia Nsight. However, I now want to use it with PyDev and am having several problems following the 'Using the GPU' part of the tutorial at <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#gpu-linux\" rel=\"noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install.html#gpu-linux<\/a><\/p>\n\n<p>The first is quite basic, and that is how to set up the environment variables. It says I should '<strong>Define a $CUDA_ROOT environment variable<\/strong>'. Several sources have said to create a new '.pam_environment' file in my home directory. I have done this and written the following: <\/p>\n\n<pre><\/pre>\n\n<p>I am not sure if this is exactly the way it has to be written - apologies if this is a basic question. If I could get confirmation that this is indeed the correct place to have written it, too, that would be helpful.<\/p>\n\n<p>The second problem is in the following part of the tutorial. It says to '<strong>change the device option to name the GPU device in your computer<\/strong>'. Apparently this has something to do with THEANO_FLAGS and .theanorc, but nowhere am I able to find out what these are: are they files? If so where do I find them? The tutorial seems to be assuming some knowledge that I don't have!<\/p>\n\n<p>Thanks for taking the time to read this: any and all answers are greatly appreciated - I am very much completely stuck at the moment!<\/p>\n","answers":[{"AnswerId":"18189699","CreationDate":"2013-08-12T14:32:25.263","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p><code>THEANO_FLAGS<\/code> is an environment variable and .theanorc is a configuration file. You can use both mechanism to configure Theano. This is described <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I never heard of the .pam_environment file. Also, you shouldn't just override the value of <code>LD_LIBRARY_PATH<\/code>, but append\/prepend to it like this:<\/p>\n\n<pre><code>LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/usr\/local\/cuda-5.5\/lib64\/lib\n<\/code><\/pre>\n\n<p>For Theano, if you define <code>CUDA_ROOT<\/code>, you don't need to modify <code>LD_LIBRARY_PATH<\/code>, so I would just remove the last line.<\/p>\n\n<p>Normally, if your shell is bash, people define the env variable <code>CUDA_ROOT<\/code> in the .bashrc file like this:<\/p>\n\n<pre><code>export CUDA_ROOT=\/usr\/local\/cuda-5.5\/bin\n<\/code><\/pre>\n\n<p>The change to .bashrc will only be used if you log out and log it again.<\/p>\n"},{"AnswerId":"21170740","CreationDate":"2014-01-16T19:11:04.887","ParentId":null,"OwnerUserId":"851699","Title":null,"Body":"<p>On Linux\/OSX:<\/p>\n\n<p>Edit or create the file <code>~\/.theanorc<\/code>. The file should contain:<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = gpu0\n\n[nvcc]\nfastmath = True\n\n[cuda]\nroot=\/usr\/local\/cuda-5.5\/  \n# On a mac, this will probably be \/Developer\/NVIDIA\/CUDA-5.5\/\n<\/code><\/pre>\n\n<p>You need to add cuda to the $LD_LIBRARY_PATH variable. If you're running eclipse, you can go to Project properties > Interpreters > Configure and interpreter ... > Environment, and then add an LD_LIBRARY_PATH variable that points to your cuda lib folder (probably \/Developer\/NVIDIA\/CUDA-5.5\/lib64)<\/p>\n\n<p>Now when you import theano it should print a message about finding the gpu.  You can run the test code at <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html<\/a> to see if it's using the gpu.<\/p>\n"}]}
{"QuestionId":18174876,"AnswerCount":1,"Tags":"<python><scipy><theano>","CreationDate":"2013-08-11T17:49:07.123","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"how to find elements and dimentions of theano csr_matrix?","Body":"<ul>\n<li>What is the difference between theano.sparse and scipy.sparse?<\/li>\n<li>How can I find the dimensions and elements of a scipy.sparse.csr_matrix()?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"18189847","CreationDate":"2013-08-12T14:38:05.080","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>theano.sparse.csr_matrix is a symbolic variable. It don't contain any data. So if you access its .shape and .data, you also obtain a symbolic variable.<\/p>\n\n<p>Symbolic variable are used to create a Theano graph that you can compile to a function. So this mean Theano is a compiler. As all compiler, it work in 2 steps:<\/p>\n\n<ol>\n<li>Create a Theano graph and compile it<\/li>\n<li>Use the compiled function.<\/li>\n<\/ol>\n\n<p>There is a shortcut, you can do this to hide the compilation phase:<\/p>\n\n<pre><code>a_theano_symbolique_variable.eval().\n<\/code><\/pre>\n\n<p>If this need input to evalute the symbolic variable, you can pass it as parameters to eval() like this:<\/p>\n\n<pre><code>a_theano_symbolique_variable.eval(another_theano_var=the_value_of_the_theano_var)\n<\/code><\/pre>\n"}]}
{"QuestionId":19489259,"AnswerCount":2,"Tags":"<python-2.7><machine-learning><computer-vision><theano><mnist>","CreationDate":"2013-10-21T08:13:19.543","AcceptedAnswerId":null,"OwnerUserId":667127.0,"Title":"Accessing Data of a Theano Shared Variable","Body":"<p>I'v successfully loaded the MNIST dataset into Theano shared variables as follows<\/p>\n\n<pre><\/pre>\n\n<p>My question is how do I access the data in both train_set_x and train_set_y. Each image in the data set is 28 * 28 pixels. That is a vector of length 784 with all elements in the vector as floats representing values between 0.0 and 1.0 inclusive. The labels are casted into int because it represents the label associated to each vector image and is a value between 0 and 9. I want to be able to loop over the train_set_x matrix images and train_set_y labels to view the data of each image and its label separately and eventually plot the images on screen.<\/p>\n","answers":[{"AnswerId":"19515782","CreationDate":"2013-10-22T10:50:24.620","ParentId":null,"OwnerUserId":"667127","Title":null,"Body":"<p>@Nouiz has pointed out the right way to show the values of both train_set_x and train_set_y. The problem was related to the environment variable \"DYLD_FALLBACK_LIBRARY_PATH\" which was not setup.\nI have a couple of python installations on my mac machine. An installation that was there as part of XCode. Another one which I have installed from python.org and a third installation which I installed from anaconda. Internally only the anaconda' python was able to run native c code for theano. My problem was due to me using the other python installations.\nThe way I fixed this was by editing the some lines in .bash_profile in my home directory. I pointed the default version to be the one I installed with anaconda and also set the environment variable DYLD_FALLBACK_LIBRARY_PATH=\"\/Users\/Me\/anaconda\/lib\". This solved the problem and everything works like a charm.<\/p>\n"},{"AnswerId":"19495252","CreationDate":"2013-10-21T12:59:18.563","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>First, <code>train_set_x<\/code> and <code>train_set_y<\/code> <em>(before the cast)<\/em> and <code>train_set<\/code> are separate copy of the same train set. So I suppose you simplified your example too much, as you say that <code>train_set_x<\/code> is the input and <code>train_set_y<\/code> is the corresponding label and this don't make sense with the code.<\/p>\n\n<p>The answer of you question depend of the contain of <code>mnist.pkl.gz<\/code>. Where did you get it? From the Deep Learning Tutorial? For my answer, I'll suppose <code>train_set<\/code> is a 2d numpy <code>ndarray<\/code>. So that you use a different <code>mnist.pkl.gz<\/code> file then the one from DLT.<\/p>\n\n<p>With that supposition, you can call <code>train_set_x.get_value()<\/code> and this will return a copy of the <code>ndarray<\/code> in shared variable. If you don't want a copy, you can do <code>train_set_x.get_value(borrow=True)<\/code> and this will work. If the shared variable is on the GPU, this will copy the data from the GPU to the CPU, but it won't copy the data if it is already on the CPU.<\/p>\n\n<p><code>train_set_y<\/code> is a <strong>Theano graph<\/strong>, not a <strong>Theano shared variable<\/strong>. So you can't call <code>get_value()<\/code> on it. You need to compile and run the graph that give <code>train_set_y<\/code>. If you want to evaluate it only once, you can call <code>train_set_y.eval()<\/code> as a shortcut to compile and run it as it do not take any input except shared variable.<\/p>\n\n<p>So you can do this:<\/p>\n\n<pre><code>for x,y in zip(train_set_x.get_value(), train_set_y.eval()):\n   print x, y\n<\/code><\/pre>\n"}]}
{"QuestionId":19581711,"AnswerCount":1,"Tags":"<theano><pymc>","CreationDate":"2013-10-25T05:23:32.727","AcceptedAnswerId":19644730.0,"OwnerUserId":2918385.0,"Title":"Assembling univariate priors into a matrix for use in MvNormal","Body":"<p>When using pymc 3 , is it possible to assemble univariate random variables into a matrix which is then used as a prior for a multivariate distribution? If so, how can I best go about this?<\/p>\n\n<p>Here is a specific example. I would like to take three R.V.'s and create a triangle matrix, A, with them:<\/p>\n\n<pre><\/pre>\n\n<p>After some manipulation I would then use this matrix as the prior for the precision parameter in the multivariate normal distribution.<\/p>\n\n<p>I assume this probably has more to do with operations with tensor variables in theano, so I will add the theano tag as well.<\/p>\n\n<p>Thank you for your time!<\/p>\n\n<p>Edit 1: Here is a minimal example of what I am trying to do:<\/p>\n\n<pre><\/pre>\n\n<p>Edit 2: Here is a test to show that  seems to do the job outside of pymc<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"19644730","CreationDate":"2013-10-28T20:43:04.490","ParentId":null,"OwnerUserId":"359944","Title":null,"Body":"<p>Looks like <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#tensor.stacklists\" rel=\"nofollow\"><code>stacklist<\/code><\/a> might do what you want. <\/p>\n"}]}
{"QuestionId":19720332,"AnswerCount":2,"Tags":"<pymc><theano>","CreationDate":"2013-11-01T03:30:17.457","AcceptedAnswerId":null,"OwnerUserId":1502840.0,"Title":"How to build a model requiring external package in PyMC3?","Body":"<p>I'm not sure if this is a PyMC3 question or a Theano question. I've used PyMC2 for a long time to fit a cosmology to supernova data. This requires some messy integrals (see i.e. <a href=\"http:\/\/arxiv.org\/abs\/astroph\/9905116\" rel=\"nofollow noreferrer\">http:\/\/arxiv.org\/abs\/astroph\/9905116<\/a> )<\/p>\n\n<p>So I use a package in python called Cosmolopy to do the integration and for some other convenience functions. Whereas this used to work fine with PyMC2, with the reliance on theano in PyMC3, I can't figure out if there is even a way to use Cosmolopy.<\/p>\n\n<p>Here is some example code of my current understanding of how to build a model in PyMC3<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This code crashes because Cosmolopy expects a float for omega_matter but receives a theano.TensorVariable instead.<\/p>\n\n<p>So the question is two-fold:<\/p>\n\n<ol>\n<li><p>Am I just missing something syntactically with PyMC3 that would allow me to do this (possibly because I am still stuck somehow on PyMC2 model-building)?<\/p><\/li>\n<li><p>If not 1, then do I need to find a way to do the integrals in theano?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"24258828","CreationDate":"2014-06-17T08:13:07.617","ParentId":null,"OwnerUserId":"3609835","Title":null,"Body":"<p>I think one possible solution would be to write a custom Theano Op following the instructions at <a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/extending\/<\/a> <\/p>\n\n<p>I would write a pure Python op without support for gradient computation, in which you would only have to implement the make_node() and perform() methods.<\/p>\n"},{"AnswerId":"19734538","CreationDate":"2013-11-01T20:05:25.737","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I don't know well PyMC3, but I know well Theano. Theano use symbolic compiler and TensorVariable are such symbolic variable. You need to compile and execute the function to get a value out of it. I don't know where to do this in PyMC3. A fast thing to try that will work if the variable depend only on constant and shared variable is to do this call::<\/p>\n\n<pre><code>the_tensor_variable.eval()\n<\/code><\/pre>\n\n<p>This will compile the function and suppose it don't take any variable input and if it compile, it will run it and return the value.<\/p>\n"}]}
{"QuestionId":19798958,"AnswerCount":1,"Tags":"<python><pymc><theano>","CreationDate":"2013-11-05T21:06:17.270","AcceptedAnswerId":19868858.0,"OwnerUserId":2918385.0,"Title":"Supplying test values in pymc 3","Body":"<p>I am exploring the use of bounded distributions in pymc. I am trying to bound a Gamma prior distribution between two values. The model specification seems to fail due to the absence of test values. How may I pass a testval argument such that I am able to specify these sorts of models?<\/p>\n\n<p>For completeness I have included the error, as well as a minimal example below. Thank you!<\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>edit: for reference purposes, here is a simple working model utilizing a bounded gamma prior distribution:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"19868858","CreationDate":"2013-11-08T21:31:53.353","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Use that line:<\/p>\n\n<pre><code>xbound = BoundedGamma('xbound', alpha=1, beta=2, testval=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":20135786,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2013-11-22T02:18:51.443","AcceptedAnswerId":null,"OwnerUserId":550879.0,"Title":"Can I avoid using `Theano.scan`?","Body":"<p>I have 3-dimensional tensor ( -- an array of matrices), and I'd like to compute the determinant () of each matrix. Is there a way to compute each determinant without using ? When I try calling  directly on the tensor I get the error<\/p>\n\n<pre><\/pre>\n\n<p>But I read that  is slow and doesn't parallelize well, and that one should use only tensor operations if possible. Is that so? Can I avoid using scan in this case?<\/p>\n","answers":[{"AnswerId":"20149947","CreationDate":"2013-11-22T16:38:10.997","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I see 3 possibilities:<\/p>\n\n<ul>\n<li>If you know before compiling the Theano function the number of matrix in the tensor3 variable, you could use the split() op or just call det() on all matrix in the tensor3.<\/li>\n<li>If you don't know the shape, you can make your own op, that will loop over the input and call the numpy fct. See for <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/tutorial\/extending_theano.html\" rel=\"nofollow\">an example<\/a> on how to make an op.<\/li>\n<li>Use scan. It is easy to use it for this case. See <a href=\"https:\/\/github.com\/nicholas-leonard\/Theano\/commit\/74a6265f15ef332d98e3e55aef0bc3c979682269\" rel=\"nofollow\">this example<\/a>, just change the call from tensordot to det().<\/li>\n<\/ul>\n"}]}
{"QuestionId":20235070,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2013-11-27T06:04:31.567","AcceptedAnswerId":20235613.0,"OwnerUserId":179081.0,"Title":"Theano element wise maximum","Body":"<p>I'm trying to find find the value of\n element-wise on a matrix in theano.\nI don't have much experience with theano.<\/p>\n\n<p>So far I have<\/p>\n\n<pre><\/pre>\n\n<p>This works, but isn't pretty and I thought I should be able to use  for this.<\/p>\n\n<p>In matlab, to do what I want to do, I would just write \n<\/p>\n","answers":[{"AnswerId":"20235613","CreationDate":"2013-11-27T06:41:09.193","ParentId":null,"OwnerUserId":"2015778","Title":null,"Body":"<p>This works for me:<\/p>\n\n<pre><code>import theano.tensor as T\nfrom theano import function\n\nx = T.dmatrix('x')\nlinmax = function([x], T.maximum(x,0))\n<\/code><\/pre>\n\n<p>Testing:<\/p>\n\n<pre><code>linmax([[-1,-2],[3,4]])\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>array([[0.,0.],[3.,4.]])\n<\/code><\/pre>\n"},{"AnswerId":"24887725","CreationDate":"2014-07-22T12:41:56.670","ParentId":null,"OwnerUserId":"3760518","Title":null,"Body":"<p>I have seen this implemented as <\/p>\n\n<pre><code>s = x*(x&gt;0)\n<\/code><\/pre>\n\n<p>several times. Dont know if that's faster than T.maximum()<\/p>\n"}]}
{"QuestionId":20237812,"AnswerCount":1,"Tags":"<python><function><operation><theano>","CreationDate":"2013-11-27T08:59:28.013","AcceptedAnswerId":20244704.0,"OwnerUserId":179081.0,"Title":"What is the difference between a Op and a Function","Body":"<p>Theano has Ops and functions.<br>\n<strong>What is the difference?<\/strong><\/p>\n\n<p>Functions seem nice and easy to define,\neg: <\/p>\n\n<pre><\/pre>\n\n<p>Ops seem complex to define. All abstract classes and such\nbut things like  and  are defined as Ops.\nI'm not to sure on the difference.<\/p>\n\n<p>How would I write the above  function as a Op?<\/p>\n","answers":[{"AnswerId":"20244704","CreationDate":"2013-11-27T14:11:09.900","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>theano.function() return a python object that is callable. So you can use it do the the computation you described when it was called.<\/p>\n\n<p>Theano Ops are part of the symbolic graph that describe the computation that you want. Do not forget that Theano have two step as many other language as C and others. You first need to describe the computation that you want, then compile. In C, you define that computation in text file. In Theano, you describe it with a Theano symbolic graph and that graph include Ops.<\/p>\n\n<p>Then you compile, with possible gcc for C and with theano.function() in Theano.<\/p>\n\n<p>So Op is the element op the symbolic graph. It describe the computation done at one point in the graph. This page in Theano tutorial describe the graph in more detail:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/symbolic_graphs.html#theano-graphs\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/symbolic_graphs.html#theano-graphs<\/a><\/p>\n\n<p>This page describe how to make an Op in Theano:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/extending_theano.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/extending_theano.html<\/a><\/p>\n\n<p>You can skip the section for optional part. So you can skip most of that page if you don't plan to make one and just want to understand the usage.<\/p>\n"}]}
{"QuestionId":20284663,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2013-11-29T11:25:19.870","AcceptedAnswerId":20285275.0,"OwnerUserId":522000.0,"Title":"Can somebody help explain a line of code in an example of Theano tutorial?","Body":"<p>In the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">logistic regression example<\/a> provided in the Theano tutorial, there is one line of code in the  function as below:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone help explain what exactly the use of square bracket in the last line of the above code? How is  gonna be interpreted? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"20285275","CreationDate":"2013-11-29T11:56:41.720","ParentId":null,"OwnerUserId":"1398797","Title":null,"Body":"<p>You have most of the information you need in the comments of the function.<\/p>\n\n<p><code>T.log(self.p_y_give_x)<\/code> returns a numpy matrix.<\/p>\n\n<p>So the [T.arange(y.shape[0]), y] is a slice of the matrix. Here we are using numpy advanced slicing. See: <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html\" rel=\"nofollow\">http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html<\/a><\/p>\n"},{"AnswerId":"35822682","CreationDate":"2016-03-06T03:13:59.817","ParentId":null,"OwnerUserId":"5424661","Title":null,"Body":"<p>I'm also confused about the matrix slicing here. T.arange(y.shape[0]\uff09 is a 1d list. y.shape[0] depends on the size of the mini batch you set. y is a list of labels who has the same dimension as  T.arange(y.shape[0]). Therefore, this slicing, according to @William Denman's reference, means: for every row in the matrix of T.log(self.p_y_give_x), we select a column index y(in which y denotes the golden label, who is also been used as index here).<\/p>\n"}]}
{"QuestionId":20376435,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><autoencoder>","CreationDate":"2013-12-04T13:17:32.023","AcceptedAnswerId":null,"OwnerUserId":2873565.0,"Title":"Auto-encoder to reduce input data size","Body":"<p>Currently, I want to use the autoencoder for reducing the input data size in order to use the reduced data for another neural networks. My task is to take a video and then give the images of the video to the autoencoder. When I use only a few images as input, the autoencoder works well but when I want to have a sequences of images, it does not. <\/p>\n\n<p>Imagine taking video from a ball moving. We have for example 200 images. If I use autoencoder for 200 images the error is big but if I use only for 5 images, the reconstruction error is small and acceptable. It seems that autoencoder does not learn the sequence or temporal movement of the ball circulating. I also tries denoting stacked autoencoder but the results are not good. <\/p>\n\n<p>Does any one know what the problem is or it is possible to use the autoencoder for this task?  <\/p>\n","answers":[{"AnswerId":"53204813","CreationDate":"2018-11-08T09:27:55.523","ParentId":null,"OwnerUserId":"5108062","Title":null,"Body":"<p>Autoencoders\/Variational Autoencoders does not learn about sequences, it learns to \"map\" the input data to a latent space which has fewer dimensions. For example if the image is <code>64x64x3<\/code> you could map that to a <code>32 dim<\/code> tensor\/array.<\/p>\n\n<p>For learning a sequence of images, you would need to connect the output of the autoencoder encoder part to a RNN (LSTM\/GRU) which could learn about the sequence of the encoded frames (consecutive frames in latent space). After that, the output of the RNN could connect to the decoder part of the autoencoder so you could see the reconstructed frames.<\/p>\n\n<p><a href=\"https:\/\/github.com\/gaborvecsei\/Neural-Network-Dreams\" rel=\"nofollow noreferrer\">Here you can find a GitHub project which tries to encode the video frames and then predict sequences<\/a><\/p>\n"}]}
{"QuestionId":20415867,"AnswerCount":1,"Tags":"<generics><printing><types><probability><theano>","CreationDate":"2013-12-06T04:16:58.417","AcceptedAnswerId":null,"OwnerUserId":2514749.0,"Title":"Theano printing probabilities for test-set samples","Body":"<p>in DL tutorials I'm trying to print the probability of the test samples according to   <a href=\"https:\/\/stackoverflow.com\/questions\/17323040\/what-is-the-prupose-meaning-of-passing-input-to-a-function-in-theano\">What is the prupose\/meaning of passing &quot;input&quot; to a function in Theano?<\/a>\nbut I get the following Error. Do I need to add some theano_flags?   <\/p>\n\n<p>how to solve the problem?<\/p>\n\n<blockquote>\n  <p>TypeError: Cannot convert Type Generic (of Variable ) into\n  Type TensorType(float64, matrix). You can try to manually convert\n   into a TensorType(float64, matrix).<\/p>\n  \n  <p>(number of features of my data=120,classes=2,test_set batch size=1)<\/p>\n<\/blockquote>\n\n<p>part of the code is:<\/p>\n\n<p>from theano import pp<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"20481585","CreationDate":"2013-12-09T21:52:39.857","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>There is an error in your code when you create values. test_set_x.get_value is a python function. So it should be test_set_x.get_value() as you want the value, not the callable that return it. As theano.shared() receive as input a callable, it create a generic Theano varible that isn't a tensor. So when you try to replace x that is a tensor variable with a generic variable, it raise an error as it is not an allowed  replacement.<\/p>\n\n<p>But even better, you don't need to create a new shared variable, just compile the function like this:<\/p>\n\n<pre><code>            f=theano.function([],classifier.p_y_given_x, \n                              givens={x:test_set_x},on_unused_input='ignore')\n<\/code><\/pre>\n"}]}
{"QuestionId":20530874,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2013-12-11T22:03:44.663","AcceptedAnswerId":null,"OwnerUserId":648896.0,"Title":"Finding closest point with Theano","Body":"<p>among a set of weight vector how could I find the closest to a given instance vector in optimize theano way. Or do I need to use numpy?<\/p>\n","answers":[{"AnswerId":"20570978","CreationDate":"2013-12-13T16:13:28.727","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This question was answered on Theano mailing list:<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/J-l9UmpSG2Y\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/J-l9UmpSG2Y<\/a><\/p>\n"}]}
{"QuestionId":20590909,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2013-12-15T03:32:05.730","AcceptedAnswerId":20602231.0,"OwnerUserId":429726.0,"Title":"Returning the index of a value in Theano vector","Body":"<p>What is the procedure in Theano for returning the index of a particular value in a Vector? In NumPy, this would be . Theano's  is a switch statement.<\/p>\n","answers":[{"AnswerId":"20602231","CreationDate":"2013-12-16T01:46:20.903","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>There is 2 behavior of numpy.where(condition, [x ,y]). Theano always support you provide 3 parameter to where(). As said in NumPy doc[1], numpy.where(cond) is equivalent to nonzero().<\/p>\n\n<p>You can do it like this in Theano:<\/p>\n\n<pre><code>import theano\nimport numpy as np\nv = np.arange(10)\nvar = theano.tensor.vector()\nout = theano.tensor.eq(var, 2).nonzero()[0]\nprint out.eval({var: v})\n<\/code><\/pre>\n\n<p>Check line 5. NumPy nonzero() return a tuple. Theano do the same. There is one vector in that tuple per dimensions in the input of nonzero().<\/p>\n\n<p>[1] <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.where.html\" rel=\"noreferrer\">http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.where.html<\/a><\/p>\n"}]}
{"QuestionId":21112016,"AnswerCount":1,"Tags":"<c++><python-2.7><theano>","CreationDate":"2014-01-14T11:07:55.593","AcceptedAnswerId":21121992.0,"OwnerUserId":2546106.0,"Title":"Reusing compiled Theano functions","Body":"<p>Suppose I have implemented the following function in Theano:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to run it a graph of computations is constructed, the function gets optimized and compiled. <\/p>\n\n<p>How can I reuse this compiled chunk of code from within a Python script and\/or a C++ application?<\/p>\n\n<p><strong>EDIT:<\/strong>\nThe goal is to construct a deep learning network and reuse it in a final C++ app.<\/p>\n","answers":[{"AnswerId":"21121992","CreationDate":"2014-01-14T19:14:33.227","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Currently this isn't possible. There is user that modified Theano to allow pickling the Theano function, but during unpickling we already re optimize the graph.<\/p>\n\n<p>There is a Pull Request that allow Theano to generate a C++ library. The user can then compile it himself and use it as a normal C++ library. The lib links against the python lib and requires numpy to be installed. But this isn't ready for broad usage.<\/p>\n\n<p>What is your goal? To save on the compilation time? If so Theano already caches the c++ module that it compiles, so the next time it is reused, the compilation will be faster. But for a big graph, the optimization phase is always redone as told above, and this can take a significant time.<\/p>\n\n<p>So what is your goal?<\/p>\n\n<p>This is something that we are working on. Make sure to use the latest Theano release (0.6) as it compiles faster. The development version is also a little faster.<\/p>\n"}]}
{"QuestionId":21290373,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2014-01-22T17:55:41.377","AcceptedAnswerId":21323254.0,"OwnerUserId":899275.0,"Title":"Theano integration with already existing python code","Body":"<p>I have a simple question, for which I expected to already have found an easy answer online, but I did not.<\/p>\n\n<p>I have a python project that works a lot with numpy, due to matrix operations. I wanted to speed the code up, so I did some profiling to find out the right tool for the job. Turns out the overhead is not python, but the matrix operations. Hence I thought I should use Theano (especially given the case that I am implementing machine learning algorithms, and this is what is was made for).<\/p>\n\n<p>Most of the overhead of my project is in one function, and I was wondering it is somehow possible to just rewrite that function with theano and then get the numpy arrays out of it and continue the computation as usual.<\/p>\n\n<p>This is again just to test how much speed up I will obtain without committing myself to changing a lot of code.<\/p>\n\n<p>Thank you!<\/p>\n\n<p>Edit: function in case is this one<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"21320362","CreationDate":"2014-01-23T22:08:51.633","ParentId":null,"OwnerUserId":"425797","Title":null,"Body":"<p>If you have code running in Numpy, translating it to Theano is not going to make it magically faster, and it's going to take a significant coding effort.<\/p>\n\n<p>Theano is really nice, but choosing it is more of a design-time decision - you wanna have all the niceties of symbolic differentiation so you don't have to calculate your gradients by hand for backpropr, so you use that framework. <\/p>\n\n<p>If you already have the Numpy code ready, just optimize the bottleneck, which in your case must be that dot product. I would try using the dot function instead of einsum, i.e. instead of:<\/p>\n\n<pre><code>dw = np.einsum('ij,ik-&gt;jk', layerValues[layer - 1], deDz)\n<\/code><\/pre>\n\n<p>Try:<\/p>\n\n<pre><code>dw = layerValues[layer-1].T.dot(deDz)\n<\/code><\/pre>\n\n<p>Sometimes einsum is dumb so that might be faster. If not, consider either <\/p>\n\n<ol>\n<li>using the GPU for the matrix-vector multiplication, via gnumpy <em>or<\/em><\/li>\n<li>using a better algorithm that'll converge faster than plain old gradient descent (which I assume is what you're using): adagrad or a second-order optimization algorithm - lbfgs, say - <\/li>\n<\/ol>\n"},{"AnswerId":"21323254","CreationDate":"2014-01-24T02:24:03.667","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You don't need to change all your script to use Theano. You can re implement just part of your code with Theano and it will work. Theano take numpy ndarray as input and return numpy ndarray as output by default. So the integration is easy with numpy.<\/p>\n\n<p>Theano don't implement einsum. So I would recommend you to start by replacing that with call to dot as Patrick said. I have see and hear many times that einsum is slower then call to dot in some cases. If that speed up isn't good enough, Theano can help you. If you just move the dot call to Theano, Theano won't be faster if Theano and NumPy are linked to the same BLAS library. But Theano should make you code faster if you move more computation to it. But it isn't perferct and some case don't have speed up and rare cases have slowdown compared to NumPy(mostly when the input shapes aren't big enough like with scalar) <\/p>\n\n<p>About Patrick answer, you don't need to use the symbolic gradient of Theano to benefit from Theano. But if you want to use the symbolic gradient, Theano can only compute the symbolic gradient inside computation graph done in Theano. So you will need to convert that part completely to Theano. But as your code already work, it mean you have manually implement the grad. This is fine and don't cause any problem. You can move that manual implementation of the grad to Theano without using the symbolic gradient.<\/p>\n\n<p>About Patrick comment on GPU. Don't forget that the transfer CPU\/GPU of data is the most costly operation on the GPU. It can completely cancel the GPU speed up in many cases. So it isn't sure that doing only the dot on the GPU will help. In Theano, we put the weight on the GPU and without doing that I don't think you can get speed up from the GPU (which ever gnumpy\/Theano\/something else). The cases when doing only the DOT on the GPU will still give speed up is with gigantic matrices.<\/p>\n"}]}
{"QuestionId":21342931,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2014-01-24T21:36:38.557","AcceptedAnswerId":21345831.0,"OwnerUserId":217802.0,"Title":"Error importing Theano","Body":"<p>After installing python, numpy, scipy and theano to ~\/.local, I tried to import theano but it threw an error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm installing on a Red Hat box:<\/p>\n\n<pre><\/pre>\n\n<p>What should I do...?<\/p>\n","answers":[{"AnswerId":"54492982","CreationDate":"2019-02-02T12:13:35.980","ParentId":null,"OwnerUserId":"1744914","Title":null,"Body":"<p>In case you are using <a href=\"https:\/\/github.com\/pyenv\/pyenv\" rel=\"nofollow noreferrer\">pyenv<\/a> you can do that with<\/p>\n\n<pre><code>CONFIGURE_OPTS=--enable-shared pyenv install 3.6.5\n<\/code><\/pre>\n"},{"AnswerId":"21345831","CreationDate":"2014-01-25T02:25:28.387","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You didn't build correctly python. It wasn't compiled with the -fPIC parameter. Look at how to compile python with a shared library.<\/p>\n\n<p>EDIT:\nYou need to compile python like this:<\/p>\n\n<pre><code>.\/configure --enable-shared\nmake\nmake install\n<\/code><\/pre>\n"},{"AnswerId":"44847761","CreationDate":"2017-06-30T13:32:29.110","ParentId":null,"OwnerUserId":"8047453","Title":null,"Body":"<p>I used a virtualenv on Ubuntu 16.04, and before installing python with <code>pyenv install 3.6.1<\/code> I ran the following command to ensure python is built with the fPIC flag:<\/p>\n\n<pre><code>export CONFIGURE_OPTS=\"OPT=\\\"-fPIC\\\"\"\n<\/code><\/pre>\n\n<p>Then I was able to install and run theano properly.<\/p>\n"}]}
{"QuestionId":21474782,"AnswerCount":1,"Tags":"<machine-learning><neural-network><theano>","CreationDate":"2014-01-31T07:37:33.620","AcceptedAnswerId":21483126.0,"OwnerUserId":3256363.0,"Title":"How can we see the transformed vaue z, in autoencoder link: http:\/\/deeplearning.net\/tutorial\/dA.html","Body":"<p>How can we see the z value , which is the reconstruction of x (dataset )<\/p>\n\n<p>Please see the link : <a href=\"http:\/\/deeplearning.net\/tutorial\/dA.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/dA.html<\/a><\/p>\n","answers":[{"AnswerId":"21483126","CreationDate":"2014-01-31T15:01:28.823","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The function \"get_reconstructed_input\" will return a Theano variable that represent z from the hidden representation. Check the function \"get_cost_updates\" It use it to train the model.<\/p>\n\n<pre><code>   tilde_x = self.get_corrupted_input(self.x, corruption_level)\n   y = self.get_hidden_values( tilde_x)\n   z = self.get_reconstructed_input(y)\n<\/code><\/pre>\n\n<p>If you don't want to train the model, you can do this:<\/p>\n\n<pre><code>   y = self.get_hidden_values(self.x)\n   z = self.get_reconstructed_input(y)\n<\/code><\/pre>\n\n<p>To make an executable function that compute this:<\/p>\n\n<pre><code>f = theano.function([x], z)\n<\/code><\/pre>\n"}]}
{"QuestionId":21608025,"AnswerCount":5,"Tags":"<theano>","CreationDate":"2014-02-06T16:07:51.887","AcceptedAnswerId":21703319.0,"OwnerUserId":1724926.0,"Title":"How to set up theano config","Body":"<p>I'm new to Theano.\nTrying to set up a config file.<\/p>\n\n<p>First of all, I notice that I have no .theanorc file:<\/p>\n\n<ol>\n<li> - returns nothing<\/li>\n<li> - returns nothing<\/li>\n<li> - passes ok<\/li>\n<\/ol>\n\n<p>I'm guessing some default configuration was created wen i installed theano. Where is it?<\/p>\n","answers":[{"AnswerId":"21703319","CreationDate":"2014-02-11T13:32:15.160","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano does not create any configuration file by itself, but has default values for all its configuration flags. You only need such a file if you want to modify the default values.<\/p>\n\n<p>This can be done by creating a .theanorc file in your home directory. For example, if you want floatX to be always float32, you can do this:<\/p>\n\n<pre><code>echo -e \"\\n[global]\\nfloatX=float32\\n\" &gt;&gt; ~\/.theanorc\n<\/code><\/pre>\n\n<p>under Linux and Mac. Under windows, this can also be done. See this page for more details:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"noreferrer\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a><\/p>\n"},{"AnswerId":"39484318","CreationDate":"2016-09-14T07:02:45.287","ParentId":null,"OwnerUserId":"6713172","Title":null,"Body":"<p>I had a similar question and this is what helped me:<\/p>\n\n<pre><code>import theano\n\/\/...\ntheano.config.floatX = 'float32' \/\/or 'float64' whatever you want\n<\/code><\/pre>\n"},{"AnswerId":"47619039","CreationDate":"2017-12-03T13:47:55.660","ParentId":null,"OwnerUserId":"4494896","Title":null,"Body":"<p>I have been having similar problems. I have NVIDIA 1070 GPU on a desktop machine with Asus Z270E motherboard and was able to import theano after setting up the .theanorc file as below. (And rebooting afterwards)<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = gpu\n\n[cuda]\nroot = \/usr\/local\/cuda\n[lib]\ncnmem = 1   \n<\/code><\/pre>\n"},{"AnswerId":"44446622","CreationDate":"2017-06-08T22:36:05.993","ParentId":null,"OwnerUserId":"1673784","Title":null,"Body":"<p>This worked for me:<\/p>\n\n<pre><code>nano ~\/.theanorc\n<\/code><\/pre>\n\n<p>Then I entered:<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = cuda\n<\/code><\/pre>\n\n<p>Code to check if Theano is using the GPU is on the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"nofollow noreferrer\">Theano doc page<\/a>.<\/p>\n\n<p>(I am using Ubuntu 14.04, Theano 0.9.0 (conda), NVIDIA 1080 Ti GPU).<\/p>\n"},{"AnswerId":"40015751","CreationDate":"2016-10-13T08:19:01.857","ParentId":null,"OwnerUserId":"4755098","Title":null,"Body":"<p>In Linux in terminal Home directory write:<\/p>\n\n<pre><code>nano .theanorc\n<\/code><\/pre>\n\n<p>In the file copy the following lines<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = gpu0\n\n[lib]\ncnmem = 1   \n<\/code><\/pre>\n\n<p>Save it.<\/p>\n\n<p>When I import theano in python I was having cnmem memory problems. Seems that is because the monitor is connected to the gpu. To resolve it change cnmem to 0.8. This number below 1 is the percentage of gpu reserved for theano<\/p>\n"}]}
{"QuestionId":21644214,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-02-08T09:19:04.863","AcceptedAnswerId":null,"OwnerUserId":1953384.0,"Title":"IPython module import error: \/usr\/bin\/ld: cannot find -lpython2.7. collect2: ld returned 1 exit status","Body":"<p>I'm using the Python module <a href=\"https:\/\/github.com\/Theano\/\" rel=\"nofollow\">Theano<\/a> on a server. It is not pre-installed on there so I installed it in my home folder along with some other modules that weren't on the server. I get the following error when I \"import theano\" in IPython.<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix the above error?<\/p>\n\n<p>Another thing is that whenever I run a Python job on the server, I first do<\/p>\n\n<pre><\/pre>\n\n<p>before executing my Python script and the server has libpython2.6.so in its \/usr\/lib64 folder. I think this is related to the problem.<\/p>\n","answers":[{"AnswerId":"21703090","CreationDate":"2014-02-11T13:23:03.837","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Your python installation is incomplete. When you do:<\/p>\n\n<pre><code>module load libs\/python\/2.7.3\n<\/code><\/pre>\n\n<p>information is added to your environment variables to make you use python 2.7.3. But this version doesn't include the development header of python (Theano needs this). Or it doesn't put the right path in your environment.<\/p>\n\n<p>To debug this, run this before and after running \"module load libs\/python\/2.7.3\" to compare what module load does:<\/p>\n\n<pre><code>env &amp;&gt; BEFORE.txt\nmodule load libs\/python\/2.7.3\nenv &amp;&gt; AFTER.txt\ndiff BEFORE.txt AFTER.txt\n<\/code><\/pre>\n\n<p>Then check the paths added to your env variables and look in those directories. It should have modified the LD_LIBRARY_PATH variable, but it should have done the same modification to LIBRARY_PATH. If it doesn't, do that modification yourself and tell your system admin about this.<\/p>\n\n<p>This will solve your problem.<\/p>\n\n<p>Otherwise, use the python 2.6 from the OS, maybe in include the development version.<\/p>\n\n<p>Otherwise, check if other python versions are available via module.<\/p>\n\n<p>Lastly, contact your system admin to add the development and shared version of python or install it yourself:<\/p>\n\n<pre><code>wget -c http:\/\/www.python.org\/ftp\/python\/2.7.6\/Python-2.7.6.tgz\ntar -jxf Python-2.7.6.tar.bz2 \ncd Python-2.7.6\n.\/configure --prefix=~\/python2.7.6 --enable-shared\nmake\nmake install\n<\/code><\/pre>\n"}]}
{"QuestionId":21713148,"AnswerCount":1,"Tags":"<python><macos><pymc><theano><psycopg>","CreationDate":"2014-02-11T21:13:55.817","AcceptedAnswerId":21714067.0,"OwnerUserId":283296.0,"Title":"psycopg2, pymc, theano and DYLD_FALLBACK_LIBRARY_PATH","Body":"<p>I am unable to use  along with . The following simple snippet from the tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>results in the following error:<\/p>\n\n<blockquote>\n  <p>Exception: The environment variable 'DYLD_FALLBACK_LIBRARY_PATH' does\n  not contain the '\/Users\/josh\/anaconda\/envs\/py27\/lib' path in its\n  value. This will make Theano unable to compile c code. Update\n  'DYLD_FALLBACK_LIBRARY_PATH' to contain the said value, this  will fix\n  this error.<\/p>\n<\/blockquote>\n\n<p>I was able to fix this problem by adding:<\/p>\n\n<pre><\/pre>\n\n<p>to my shell init file . <strong>However<\/strong>, and this is the part I don't understand, that line breaks :<\/p>\n\n<pre><\/pre>\n\n<p>How can I have  and  (here ) live happily together?<\/p>\n\n<p>This is on OS X with a Python 2.7.6 installation with an Python environment created with <a href=\"http:\/\/docs.continuum.io\/conda\/index.html\" rel=\"nofollow\">conda<\/a>.<\/p>\n","answers":[{"AnswerId":"21714067","CreationDate":"2014-02-11T22:04:06.240","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The development version of Theano don't need changes to DYLD_FALLBACK_LIBRARY_PATH. So undo the change to it and update your Theano version. From:<\/p>\n\n<p><a href=\"http:\/\/www.deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\">http:\/\/www.deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>Run one of those 2 command depending of your need:<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git --install-option='--prefix=~\/.local'\npip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>EDIT: I removed the link to an answer elsewhere and copied the answer here. Thanks<\/p>\n"}]}
{"QuestionId":21945079,"AnswerCount":0,"Tags":"<python><machine-learning><theano>","CreationDate":"2014-02-21T21:01:12.113","AcceptedAnswerId":null,"OwnerUserId":742616.0,"Title":"Theano maximum likelihood for geometric brownian motion","Body":"<p>I'm trying to do maximum likelihood estimation (MLE) for a geometric brownian motion with Theano. I know, Theano is primarily a ML library, but it <em>should<\/em> work... (of course I'm just trying out Theano and my eventual goal is to do MLE for a slightly more complicated model....)<\/p>\n\n<p>Writing math at SO is a bitch so look up GBM e.g. <a href=\"http:\/\/en.wikipedia.org\/wiki\/Geometric_Brownian_motion\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>Here's the Theano code, which is slightly modified from the logistic regression example in the Theano tutorial:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>loghood is the log likelihood, which should be minimized w.r.t mu and sigma2<\/li>\n<li>Xdata is simulated from a GBM, dXdata[n] = Xdata[n+1] - Xdata[n]<\/li>\n<li>autodifferentiation should work OK, I also tried putting the gradients by hand<\/li>\n<li>costlist, mulist, siglist were added for debugging purposes<\/li>\n<\/ul>\n\n<p>So what I get is that the train function evaluates OK for the initial mu, sigma2, but right at step 2 where the mu, sigma2 are updated as per the gradients, the train function seems to backfire by giving either crazy numbers or nan:s...<\/p>\n\n<p>I compared the result with exact MLE and also minimization by scipy.optimize.minimize, which both work just fine. I think the problem is somewhere in the \"loghood\" above, but I just can't figure it out...<\/p>\n\n<p><strong>So can anyone please figure out where the code goes wrong?<\/strong><\/p>\n","answers":[]}
{"QuestionId":22036710,"AnswerCount":3,"Tags":"<theano><word2vec><deep-learning>","CreationDate":"2014-02-26T09:15:35.600","AcceptedAnswerId":34802772.0,"OwnerUserId":956730.0,"Title":"How to compute a language model with word2vec tool?","Body":"<p>I'm trying to build a neural network language model and it seems that word2vec tool by Mikolov et al is a good tool for this purpose. I tried that but it just produces word representations.\nDoes anybody know how i can produce a language model by that tool or any other reasonable deep learning framework?<\/p>\n","answers":[{"AnswerId":"26458679","CreationDate":"2014-10-20T04:42:00.403","ParentId":null,"OwnerUserId":"2440487","Title":null,"Body":"<p>Microsoft Research has released a toolkit for language modelling with word2vec-style vectors. You can find it <a href=\"http:\/\/research.microsoft.com\/en-us\/projects\/rnn\/\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"22037003","CreationDate":"2014-02-26T09:27:12.083","ParentId":null,"OwnerUserId":"3153644","Title":null,"Body":"<p><code>word2vec<\/code> is a tool to represent a single word (o a group of words) as a numerical vector. So it is not directly related to a language model.<\/p>\n\n<p>To generate a Language model you can use the <a href=\"https:\/\/code.google.com\/p\/mitlm\/\" rel=\"nofollow\">MITLM<\/a> to do it. For example you can create a N-gram model using the corpus <code>Lectures.txt<\/code> with this command:<\/p>\n\n<pre><code>estimate-ngram -text Lectures.txt -write-lm Lectures.lm\n<\/code><\/pre>\n\n<p>A great tutorial can be found <a href=\"http:\/\/projects.csail.mit.edu\/cgi-bin\/wiki\/view\/SLS\/MITLMTutorial\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"34802772","CreationDate":"2016-01-15T01:12:05.107","ParentId":null,"OwnerUserId":"956730","Title":null,"Body":"<p>Doc2Vec implemented in Gensim does the job. The trick is that they use the document ID as a context word, which is present in all window sizes of all the words in the document.<\/p>\n\n<p>Code is <a href=\"https:\/\/radimrehurek.com\/gensim\/models\/doc2vec.html\" rel=\"nofollow\">here in Python\/Gensim<\/a><\/p>\n"}]}
{"QuestionId":22350482,"AnswerCount":1,"Tags":"<python><python-2.7><theano>","CreationDate":"2014-03-12T11:44:59.643","AcceptedAnswerId":22352914.0,"OwnerUserId":888849.0,"Title":"Theano installation in Windows 32bit","Body":"<p>I am trying to install Theano library on a windows 32-bit machine. I've already installed python 2.7, numpy, scipy, mingw.  The next dependency is blas.  How can I install it on Windows? I've also installed canopy in order to install pip. The next steps are to install Theano with:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing any step except in blas installation?  When I try to use the canopy platform to perform the installation, I've noticed that I couldn't install from the package manager the needed dependencies  \u201cmingw 4.5.2\u201d and \u201clibpython 1.2\u201d.<\/p>\n\n<p>I also tried to follow AnacondaCE instructions. I've downloaded it and configured it using Windows installer for Theano on AnacondaCE for Windows configuration. <\/p>\n","answers":[{"AnswerId":"22352914","CreationDate":"2014-03-12T13:23:40.827","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>How did you install numpy\/scipy\/mingw? I think they are installed by default with canopy. As you tried many things, I would suggest that you remove all other python installation and only keep canopy. Canopy provide blas, so you don't need to install one. It is better to install the academic\/full version of canopy as it provide a faster BLAS. Their should be newer version of mingw and libpython in canopy since the instruction was done. Install them.<\/p>\n\n<p>Then this is done, you only need to install theano with this command:<\/p>\n\n<pre><code>pip install --upgrade --no-deps Theano\n<\/code><\/pre>\n\n<p>All the command you provides are different way to install Theano. You only need one! If you want the development version, you can use this instead:<\/p>\n\n<pre><code>git clone git:\/\/github.com\/Theano\/Theano.git\ncd Theano\npython setup.py develop\n<\/code><\/pre>\n"}]}
{"QuestionId":22527072,"AnswerCount":1,"Tags":"<python><neural-network><backpropagation><theano><deep-learning>","CreationDate":"2014-03-20T08:15:29.947","AcceptedAnswerId":24844120.0,"OwnerUserId":2489122.0,"Title":"Multilayer perceptron with target variable as array instead of a single value","Body":"<p>I am new to deep learning and I have been trying to use the theano library to train my data. <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">MLP tutorial<\/a> here has a scalar output value while my use case has an array with a 1 corresponding to the value depicted in the output.<\/p>\n\n<p>For example (assume the possible scalar values are 0,1,2,3,4,5),<\/p>\n\n<pre><\/pre>\n\n<p>I have only modified the code to read my input and output (output now is a 2 dimensional array or matrix in the parlance of theano). Other parts of the code is as is from the MLP tutorial pasted above.\n<br><\/p>\n\n<p>The error I am getting is in the following function<\/p>\n\n<pre><\/pre>\n\n<p>Error stack:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to know how to change this theano.function to accommodate the y value as a matrix. <\/p>\n","answers":[{"AnswerId":"24844120","CreationDate":"2014-07-19T19:23:07.013","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to define <code>y<\/code> as <code>T.imatrix()<\/code> instead of <code>T.lvector()<\/code>.<\/p>\n"}]}
{"QuestionId":22579246,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-03-22T15:00:41.380","AcceptedAnswerId":22612374.0,"OwnerUserId":826983.0,"Title":"Theano: Get matrix dimension and value of matrix (SharedVariable)","Body":"<p>I would like to know how to retrieve the dimension of a SharedVariable from theano.<\/p>\n\n<p>This here e.g. does not work:<\/p>\n\n<pre><\/pre>\n\n<p>and only returns<\/p>\n\n<pre><\/pre>\n\n<p>I am also interested in printing\/retriving the values of a matrix or vector ..<\/p>\n","answers":[{"AnswerId":"22612374","CreationDate":"2014-03-24T14:33:38.670","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You can get the value of a shared variable like this:<\/p>\n\n<pre><code>w.get_value()\n<\/code><\/pre>\n\n<p>Then this would work:<\/p>\n\n<pre><code>w.get_value().shape\n<\/code><\/pre>\n\n<p>But this will copy the shared variable content. To remove the copy you can use the borrow parameter like this:<\/p>\n\n<pre><code>w.get_value(borrow=True).shape\n<\/code><\/pre>\n\n<p>But if the shared variable is on the GPU, this will still copy the data from the GPU to the CPU. To don't do this:<\/p>\n\n<pre><code>w.get_value(borrow=True, return_internal_type=True).shape\n<\/code><\/pre>\n\n<p>Their is a simpler way to do this, compile a Theano function that return the shape:<\/p>\n\n<pre><code>w.shape.eval()\n<\/code><\/pre>\n\n<p><code>w.shape<\/code> return a symbolic variable. .eval() will compile a Theano function and return the value of shape.<\/p>\n\n<p>If you want to know more about how Theano handle the memory, check this web page: <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/tutorial\/aliasing.html\">http:\/\/www.deeplearning.net\/software\/theano\/tutorial\/aliasing.html<\/a><\/p>\n"}]}
{"QuestionId":22591812,"AnswerCount":1,"Tags":"<python><performance><optimization><parallel-processing><theano>","CreationDate":"2014-03-23T13:55:33.583","AcceptedAnswerId":22787775.0,"OwnerUserId":826983.0,"Title":"Parallelize iterative calculation using Theano","Body":"<p>This is a piece of python code that basically just calculates the activations of a neural network and then updates the new state for the next input value  with respect to an arbitrary leaking rate .<\/p>\n\n<pre><\/pre>\n\n<p>It is not that important to understand what this is exactly doing. My question is: Can I parallelize this using the GPU with ? You can see that the new  depends on the previous value of  so what I would want to do is to parrallelize the calculations for those vectors and matrices. If those arrays become considerably large this will result in a much better performance.<\/p>\n\n<p>Could anybody tell me how to do this?<\/p>\n","answers":[{"AnswerId":"22787775","CreationDate":"2014-04-01T13:43:32.170","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This ticket was closed as off-topick, so I answered it in the comment. So here my answer as a post.<\/p>\n\n<p>My answer is generic and apply to all system, not just Theano. As each iteration of your loop depend on the previous one, you can't paralelize your iterations completly. You could parallelize the <code>u=data[t]<\/code> as it don't depend on the previous x. You could parallelize <code>dot( Win, vstack((1,u)) )<\/code> for the same reason. But you can't parallelize <code>dot(W,x)<\/code> and what depend on it like tanh and the lines afters.<\/p>\n\n<p>If you want to optimize this, you can move outside the loop all computation that don't depend on x. This will allow to work with more data at the same time and so could be faster. So the <code>dot(win, ...)<\/code> could be speed up. But this will raise the memory usage.<\/p>\n"}]}
{"QuestionId":22613364,"AnswerCount":2,"Tags":"<sentiment-analysis><text-classification><theano><deep-learning>","CreationDate":"2014-03-24T15:15:03.183","AcceptedAnswerId":22682341.0,"OwnerUserId":956730.0,"Title":"Theano Classification Task always gives 50% validation error and test error?","Body":"<p>I am doing a text classification experiment with Theano's DBN (Deep Belief Network) and SDA (Stacked Denoising Autoencoder) examples.\nI have produced a feature\/label dataset just as Theano's MINST dataset is produced and changed the feature length and output values of those examples to adopt to my dataset (2 outputs instead of 10 outputs, and the number of features is adopted to my dataset).\nEvery time i run the experiments (both DBN and SDA) i get an exact 50% validation error and test error.\nDo you have any ideas what i'm doing wrong? because i have just produced a dataset out of Movie Review Dataset as MINST dataset format and pickled it.<\/p>\n\n<p>my code is the same code you can find in <a href=\"http:\/\/www.deeplearning.net\/tutorial\/DBN.html\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/tutorial\/DBN.html<\/a>\nand my SDA code is the same code you can find in\n<a href=\"http:\/\/www.deeplearning.net\/tutorial\/SdA.html\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/tutorial\/SdA.html<\/a><\/p>\n\n<p>The only difference is that i have made my own dataset instead of MINST digit recognition dataset. My dataset is Bag of Words features from Movie Review Dataset which of course has different number of features and output classes so i just have made tiny modifications in function parameters number of inputs and output classes.\nThe code runs beautifully but the results are always 50%.\nThis is a sample output:<\/p>\n\n<pre><\/pre>\n\n<p>The pretraining code for file DBN_MovieReview.py ran for 430.33m<\/p>\n\n<pre><\/pre>\n\n<p>The fine tuning code for file DBN_MovieReview.py ran for 5.48m<\/p>\n\n<p>I ran both SDA and DBN with two different feature sets. So i got this exact 50% accuracy on all these 4 experiments.<\/p>\n","answers":[{"AnswerId":"33166614","CreationDate":"2015-10-16T09:13:11.607","ParentId":null,"OwnerUserId":"5453108","Title":null,"Body":"<p>I had same problem. I think this problem is because of Overshooting.\nSo I decreased learning rate 0.1 to 0.013 and increase epoch.\nThen it works. \nBut I'm not sure your problem is same.<\/p>\n"},{"AnswerId":"22682341","CreationDate":"2014-03-27T08:46:24.453","ParentId":null,"OwnerUserId":"956730","Title":null,"Body":"<p>I asked the same question in Theano's user groups and they answered that feature values should be between 0 and 1.<\/p>\n\n<p>So i used a normalizer to normalize feature values and it solved the problem.<\/p>\n"}]}
{"QuestionId":22675274,"AnswerCount":2,"Tags":"<python><eclipse><macos><machine-learning><theano>","CreationDate":"2014-03-26T23:31:32.547","AcceptedAnswerId":22793080.0,"OwnerUserId":1568741.0,"Title":"Python won't find variable in module","Body":"<p>I just started playing around with Theano but have a strange problem in Eclipse. I am trying to import the config module to run some example code. The import works fine and I can see what's in the module.<\/p>\n\n<p>Here is the simple code I am trying:<\/p>\n\n<pre><\/pre>\n\n<p>This works fine and I get an output like:<\/p>\n\n<pre><\/pre>\n\n<p>and some more lines like that. Unfortunately if I use the following code, I get an \"undefined variable from import\"-error for the floatX:<\/p>\n\n<pre><\/pre>\n\n<p>This is only happening in Eclipse. In the console I get \"float32\", which is the correct output. Any idea why this is happening and how I can get to give me the value behind that variable? Thank you!<\/p>\n\n<p>System: OSX 10.9.2 \/ Python: 2.7.6 (Macports installation) \/ Theano: 0.6.0 (Macports installation) \/ Eclipse: Kepler Service Release 2<\/p>\n","answers":[{"AnswerId":"22701190","CreationDate":"2014-03-27T23:16:05.883","ParentId":null,"OwnerUserId":"2317533","Title":null,"Body":"<p>Is Eclipse using the same version of python as what you are running in the shell (console)? Does Eclipse know where to find theano- does it have a PYTHONPATH setting for it?<\/p>\n\n<p>What OS are you using?<\/p>\n"},{"AnswerId":"22793080","CreationDate":"2014-04-01T17:43:38.147","ParentId":null,"OwnerUserId":"1568741","Title":null,"Body":"<p>ok, I found the answer finally. I never really had an error. I did not find that out because I never tried to actually run the script because the editor indicated there was an error... The maker of PyDev answered the following question himself and provides a workaround:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/2112715\/how-do-i-fix-pydev-undefined-variable-from-import-errors\">How do I fix PyDev &quot;Undefined variable from import&quot; errors?<\/a><\/p>\n\n<p>For code in your project, the only way is adding a comment saying that you expected that (the static code-analysis only sees what you see, not runtime info -- if you opened that module yourself, you'd have no indication that main was expected).<\/p>\n\n<p>You can use ctrl+1 (Cmd+1 for Mac) in a line with an error and pydev will present you an option to add a comment to ignore that error.<\/p>\n"}]}
{"QuestionId":22735082,"AnswerCount":1,"Tags":"<python><serialization><memory-management><theano>","CreationDate":"2014-03-29T18:28:40.250","AcceptedAnswerId":22774654.0,"OwnerUserId":1323010.0,"Title":"Python Memory Error with Python pkl files","Body":"<p>I'm using a python library for deep learning and neural networks. The computer i'm running on has 16 gb of ram@1866 MHz. At first my input data file was too large, so I broke it smaller:<\/p>\n\n<p><\/p>\n\n<p>caused:<\/p>\n\n<p><\/p>\n\n<p>Since the file was just an numpy array of numpy arrays, I could break it into seperate files, and would recreate the larger file dynamically in the program by loading numerous pickle files.<\/p>\n\n<p><\/p>\n\n<p>And this solution worked fine. <strong>My problem is<\/strong> that now I have an enourmous file that is causing a  that I don't know how to break up further. It is a theano tensor variable representing a 30,000x30,000 matrix of floating point numbers. My questions:<\/p>\n\n<ol>\n<li>Is there a method to save something across multiple pkl files even if you are unsure of how to divide the underlying data structure?<\/li>\n<li>Will running this on our labs server (48 gb) work better? Or is this memory error independent of the architecture?<\/li>\n<li>Is the huge pkl file I have now that is too large to use worthless? I hope not, it was around 8 hours of neural network training.<\/li>\n<li>Are there any other solutions besides using a database that anyone can think of? If at all possible, I would strongly prefer to <strong>not<\/strong> use databases because I've already had to transfer the software to numerous servers, many of which I do not have root access to and is a pain to get other things installed.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"22774654","CreationDate":"2014-04-01T00:27:01.520","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>First, pkl aren't great to save binary data and aren't memory friendly. It must copy all the data in ram before writing to disk. So this double the memory usage! You can use numpy.save and numpy.load to stores ndarray without that memory doubling.<\/p>\n\n<p>For the Theano variable, I guess you are using a Theano shared variable. By default, when you get it via <code>get_value()<\/code>, it copy the data. You can use <code>get_value(borrow=True)<\/code> to don't copy this.<\/p>\n\n<p>Both of those change together could lower the memory usage by 3x. If this isn't enough or if you are sick of handling multiple files yourself, I would suggest that you use pytables: <a href=\"http:\/\/www.pytables.org\/\" rel=\"nofollow\">http:\/\/www.pytables.org\/<\/a> It allow to have one big ndarray stored in a file bigger then the ram avaiable, but it give an object similar to ndarray that you can manipulate very similarly to an ndarray.<\/p>\n"}]}
{"QuestionId":22799362,"AnswerCount":1,"Tags":"<pymc><theano><pymc3>","CreationDate":"2014-04-02T00:22:38.377","AcceptedAnswerId":22803664.0,"OwnerUserId":3297752.0,"Title":"Deterministic variables and a Fortran Scipy function in PyMC 3","Body":"<p>I am trying to build a simple PyMC 3 model in which I estimate two cut points and a correlation parameter in a latent bivariate Gaussian density, producing four predicted probabilities for a vector of (multinomial) counts. (This will, I hope, eventually be part of a larger model in which these and other parameters are estimated for a number of latent multivariate Gaussian densities.)<\/p>\n\n<p>So, I want to model the cut points cx and cy as normal random variables and the correlation parameter rho as a scaled Beta random variable (as a side note, I'd love to hear a better way to deal with rho - does PyMC 3 have truncated normal random variables, for example?). And I want to use the function mvnun to calculate the predicted probabilities for given values of cx, cy, and rho. The function mvnun is part of scipy.stats.mvn, which is a bit of compiled Fortran code that has two functions for calculating very accurate multivariate normal CDF values.<\/p>\n\n<p>If I try to stick rho in a correlation matrix S or if I put cx or cy into arrays indicating the integration limits, I get this:<\/p>\n\n<pre><\/pre>\n\n<p>If I use fixed numerical values for cx, cy, and\/or rho, mvnun works just fine (in or out of the 'with model:' block). I've been poking around, trying to figure out why the PyMC RVs cause this error, but I'm stumped. I gather that cx, cy, and rho are theano TensorVariables, but I can't figure out what, if anything, about theano TensorVariables would cause these problems.<\/p>\n\n<p>Is there a fundamental problem with trying to call a Fortran function with PyMC RVs as arguments? Or is my code flawed in some way? Both? Something else entirely?<\/p>\n\n<p>I'm new to PyMC, and I installed PyMC 3 thinking it was the current version (I swear the note about it being the alpha version wasn't there when I installed it a few weeks ago). Perhaps I should install 2.3 and figure out how to put this together with that?<\/p>\n\n<p>In any case, any advice about how to fix things would be much appreciated.<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"22803664","CreationDate":"2014-04-02T06:54:31.207","ParentId":null,"OwnerUserId":"359944","Title":null,"Body":"<p>This will be somewhat easier in PyMC 2.3, so installing 2.3 is a pretty good option.<\/p>\n\n<p>In PyMC 3 that doesn't work because cx, cy, and rho are theano TensorVariables and mvncdf expects numpy arrays. Theano variables are not a kind of numpy array. Instead they <em>represent<\/em> a computation leading to a numpy array. <\/p>\n\n<p>You'll need to make mvncdf into a Theano operator (theano analog of a numpy function). There unfortunately isn't a super simple way to do that right now, though <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/1784\" rel=\"nofollow\">it is coming<\/a>. You can also take a look at the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/extending_theano.html\" rel=\"nofollow\">guide to making a Theano Op<\/a>.<\/p>\n"}]}
{"QuestionId":22860717,"AnswerCount":1,"Tags":"<python><entropy><theano>","CreationDate":"2014-04-04T10:50:43.743","AcceptedAnswerId":31435097.0,"OwnerUserId":862629.0,"Title":"Entropy and Probability in Theano","Body":"<p>I have a written a simple python code to calculate the entropy of a set and I am trying to write the same thing in Theano.<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to write the equivalent code in Theno, but I am not sure how to do it:<\/p>\n\n<pre><\/pre>\n\n<p>However, the scan line is not correct and I get tons of errors. I am not sure if there is a simple way to do it (to compute the entropy of a vector), or if I have to put more effort on the scan function. Any ideas?<\/p>\n","answers":[{"AnswerId":"31435097","CreationDate":"2015-07-15T15:46:09.383","ParentId":null,"OwnerUserId":"5116849","Title":null,"Body":"<p>Other than the point raised by nouiz, P should not be declared as a T.vector because it will be the result of computation on your vector of values.<\/p>\n\n<p>Also, to compute something like entropy, you do not need to use Scan (Scan introduces a computation overhead so it should only be used because there's no other way of computing what you want or to reduce memory usage); you can take a approach like this :<\/p>\n\n<pre><code>values = T.vector('values')\nnb_values = values.shape[0]\n\n# For every element in 'values', obtain the total number of times\n# its value occurs in 'values'.\n# NOTE : I've done the broadcasting a bit more explicitly than\n# needed, for clarity.\nfreqs = T.eq(values[:,None], values[None, :]).sum(0).astype(\"float32\")\n\n# Compute a vector containing, for every value in 'values', the\n# probability of that value in the vector 'values'.\n# NOTE : these probabilities do *not* sum to 1 because they do not\n# correspond to the probability of every element in the vector 'values\n# but to the probability of every value in 'values'. For instance, if\n# 'values' is [1, 1, 0] then 'probs' will be [2\/3, 2\/3, 1\/3] because the\n# value 1 has probability 2\/3 and the value 0 has probability 1\/3 in\n# values'.\nprobs = freqs \/ nb_values\n\nentropy = -T.sum(T.log2(probs) \/ nb_values)\nfct = theano.function([values], entropy)\n\n# Will output 0.918296...\nprint fct([0, 1, 1])\n<\/code><\/pre>\n"}]}
{"QuestionId":22883149,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2014-04-05T15:54:55.013","AcceptedAnswerId":22883792.0,"OwnerUserId":826983.0,"Title":"Can I configure Theano's x divided by 0 behavior","Body":"<p>I have a little problem using Theano. It seems that a  results in  not as using e.g. Numpy this results in 0 (at least the inverse function do behave like that). Take a look:<\/p>\n\n<pre><\/pre>\n\n<p>Results:<\/p>\n\n<pre><\/pre>\n\n<p>This is a problem for some of my calculations I want to perform in my GPU and I don't want to get back the data from it to replace  by  to continue my calculations of course since this would slow down the process considerably.<\/p>\n","answers":[{"AnswerId":"22883792","CreationDate":"2014-04-05T16:48:54.983","ParentId":null,"OwnerUserId":"102441","Title":null,"Body":"<p><code>theano.tensor<\/code> calculates the <em>elementwise<\/em> inverse<\/p>\n\n<p><code>np.linalg.inv<\/code> calculates the inverse <em>matrix<\/em><\/p>\n\n<p>These are not the same thing mathematically<\/p>\n\n<hr>\n\n<p>You're probably looking for the <strong>experimental<\/strong> <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/linalg.html#theano.sandbox.linalg.ops.MatrixInverse\" rel=\"nofollow\"><code>theano.sandbox.linalg.ops.MatrixInverse<\/code><\/a><\/p>\n"}]}
{"QuestionId":22914786,"AnswerCount":3,"Tags":"<python><numpy><neural-network><theano>","CreationDate":"2014-04-07T14:06:03.850","AcceptedAnswerId":22930872.0,"OwnerUserId":643014.0,"Title":"Error because of Theano and NumPy variable types","Body":"<p>I am writing this code using numpy 1.9 and the latest version of Theano but I get an error which I can't fix. I doubt it could be the way I declare variable types but I can't work it around. I appreciate your suggestions. I want to product a matrix with a vector and sum it with a bias. <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for your time!<\/p>\n","answers":[{"AnswerId":"27993778","CreationDate":"2015-01-16T22:25:59.767","ParentId":null,"OwnerUserId":"1639345","Title":null,"Body":"<p>I had a similar error and was able to resolve it by adding a <code>.theanorc<\/code> file containing the following two lines:<\/p>\n\n<pre><code>[global]\n\nfloatX = float32\n<\/code><\/pre>\n\n<p>That seemed to fix everything. However, your problem shows a slightly different error. But I think it's worth trying.<\/p>\n"},{"AnswerId":"22914819","CreationDate":"2014-04-07T14:07:50.993","ParentId":null,"OwnerUserId":"1590306","Title":null,"Body":"<p>Error looks quite self-explanatory; have you tried:<\/p>\n\n<pre><code>dt = np.dtype(np.float32) \n<\/code><\/pre>\n\n<p>??<\/p>\n"},{"AnswerId":"22930872","CreationDate":"2014-04-08T07:53:35.570","ParentId":null,"OwnerUserId":"643014","Title":null,"Body":"<p>This answer comes from <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/ofTRZJz7oZQ\" rel=\"nofollow noreferrer\">Theano-users google group<\/a>.<\/p>\n\n<p>You define your <code>x<\/code> variable as:<\/p>\n\n<pre><code>x=T.vector(dtype=theano.config.floatX)\n<\/code><\/pre>\n\n<p>This is it is a vector(i.e. it only have 1 dimensions).<\/p>\n\n<pre><code>x_inp = np.matrix('2;1',dtype=dt)\n<\/code><\/pre>\n\n<p>create a matrix, not a vector.<\/p>\n\n<p>Theano graph are strongly typed, you must defined the good number of\ndimensions. Just use:<\/p>\n\n<pre><code>x_inp = np.asarray([2,1]) \n<\/code><\/pre>\n\n<hr>\n\n<p>I actually ended up defining <code>x<\/code> and <code>b<\/code> as matrices.<\/p>\n"}]}
{"QuestionId":23314351,"AnswerCount":1,"Tags":"<python><scipy><theano>","CreationDate":"2014-04-26T17:30:52.767","AcceptedAnswerId":23314906.0,"OwnerUserId":283296.0,"Title":"What exactly is a tensor in theano?","Body":"<p>What exactly is a <strong>Tensor<\/strong> in <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/index.html#\">Theano<\/a>, and what is the precise connection with <a href=\"http:\/\/en.wikipedia.org\/wiki\/Tensor#Definition\">Tensors<\/a> as they are typically understood in Physics or Math?<\/p>\n\n<p>I went through the <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/introduction.html#introduction\">Theano at Glance<\/a> and the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\">Basic Tensor functionality<\/a>, but I could not find a clear connection.<\/p>\n","answers":[{"AnswerId":"23314906","CreationDate":"2014-04-26T18:26:33.160","ParentId":null,"OwnerUserId":"773960","Title":null,"Body":"<p>There is a good breakdown of different physics\/math ways to think of tensor in Jim Belk's <a href=\"https:\/\/math.stackexchange.com\/a\/192441\">answer<\/a> to a question on math.stackexchange. After looking over the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/index.html\" rel=\"nofollow noreferrer\">documentation<\/a> on tensor and the various operations Theano provides I'd say Theano's notion of tensor corresponds to the first way of thinking of tensor. In Jim's words:<\/p>\n\n<blockquote>\n  <p>Tensors are sometimes defined as multidimensional arrays, in the same way that a matrix is a two-dimensional array. From this point of view, a matrix is certainly a special case of a tensor.<\/p>\n<\/blockquote>\n\n<p>In any case, I don't see anything myself in the docs indicating that Theano's tensor implementation knows about global properties of manifolds or tensor products in linear algebra beyond defining dot-products and the like.  This would indicate Theano is taking a local point of view in it's implementation as opposed to the global.<\/p>\n"}]}
{"QuestionId":23634582,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2014-05-13T14:48:14.470","AcceptedAnswerId":null,"OwnerUserId":3627108.0,"Title":"What's the difference between the shapes of variable in Theano","Body":"<p>Could you tell me what's the difference with the shapes 0, (?,), (1,?), (?,?) in Theano?\nWhy the array I defined as<\/p>\n\n<pre><\/pre>\n\n<p>is an array of (3,) ? How could I defined an array of (3,1) ? <\/p>\n\n<p>Besides, I write the code as below:<\/p>\n\n<pre><\/pre>\n\n<p>Why it report error:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you very much\uff01<\/p>\n","answers":[{"AnswerId":"23635590","CreationDate":"2014-05-13T15:33:08.417","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>You are trying to pass <code>a.z<\/code> to <code>a.ac()<\/code>, whereas <code>a.z<\/code> is actually the <em>result<\/em> of <code>a.ac(x)<\/code>!<\/p>\n\n<p>Instead, you probably want do this:<\/p>\n\n<pre><code>a.ac(a.inputs)\n# array([  8.61379147, -13.0183053 ,  -4.41056323])\n<\/code><\/pre>\n\n<p>The value of the symbolic variable <code>a.z<\/code> is undetermined until <code>a.x<\/code>, <code>a.W<\/code> and <code>a.b<\/code> can all be evaluated. The syntax for <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/function.html\" rel=\"nofollow\"><code>theano.function<\/code><\/a> works like this:<\/p>\n\n<pre><code>find_x = theano.function([&lt;inputs needed to compute x&gt;], &lt;output x&gt;)\n<\/code><\/pre>\n\n<p>When you actually want to call <code>find_x()<\/code>, you only need to give it the stuff in the square brackets, and the second argument to <code>theano.function<\/code> will be the return value of <code>find_x()<\/code>.<\/p>\n"}]}
{"QuestionId":23643850,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2014-05-14T00:41:46.220","AcceptedAnswerId":null,"OwnerUserId":3633343.0,"Title":"How to swich theano.tensor to numpy.array?","Body":"<p>I have simple codes as shown below:<\/p>\n\n<pre><\/pre>\n\n<p>However, I get the following error information:<\/p>\n\n<pre><\/pre>\n\n<p>Furthermore, when I use the function of theano.tensor, it seems that what it returns is called \"tensor\", and I can't simply switch it to the type numpy.array, even though what the result should shape like a matrix.<\/p>\n\n<p>So that's my question:how can I switch outxx to type numpy.array? <\/p>\n","answers":[{"AnswerId":"23660792","CreationDate":"2014-05-14T16:58:56.603","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano \"tensor\" variable are symbolic variable. What you build with them are like a programme that you write. You need to compile a Theano function to execute what this program do. There is 2 ways to compile a Theano function:<\/p>\n\n<pre><code>f = theano.function([testxx.input], [outxx])\nf_a1 = f(a)\n\n# Or the combined computation\/execution\nf_a2 = outxx.eval({testxx.input: a})\n<\/code><\/pre>\n\n<p>When you compile a Theano function, your must tell what the input are and what the output are. That is why there is 2 parameter in the call to theano.function(). eval() is a interface that will compile and execute a Theano function on a given symbolic inputs with corresponding values.<\/p>\n"},{"AnswerId":"31883372","CreationDate":"2015-08-07T17:22:07.730","ParentId":null,"OwnerUserId":"3497273","Title":null,"Body":"<p>Since <code>testxx<\/code> uses <code>sum()<\/code> from <code>theano.tensor<\/code> and not from <code>numpy<\/code>, it probably expects a <code>TensorVariable<\/code> as input, and not a numpy array.<\/p>\n\n<p>=> Replace <code>a = np.array(...)<\/code> with <code>a = T.matrix(dtype=theano.config.floatX)<\/code>.<\/p>\n\n<p>Before your last line, <code>outxx<\/code> will then be a <code>TensorVariable<\/code> that depends on <code>a<\/code>. So you can evaluate it by giving the value of <code>a<\/code>.<\/p>\n\n<p>=> Replace your last line <code>outxx = np.asarray(...)<\/code> with the following two lines.<\/p>\n\n<pre><code>f = theano.function([a], outxx)\noutxx = f(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype = np.float32))\n<\/code><\/pre>\n\n<p>The following code runs without errors.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nclass testxx(object):\n    def __init__(self, input):\n        self.input = input\n        self.output = T.sum(input)\na = T.matrix(dtype=theano.config.floatX)\nclassfier = testxx(a)\noutxx = classfier.output\nf = theano.function([a], outxx)\noutxx = f(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype = np.float32))\n<\/code><\/pre>\n\n<p>Theano documentation on <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/adding.html\" rel=\"nofollow\">adding scalars<\/a> gives other similar examples.<\/p>\n"}]}
{"QuestionId":23656777,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-05-14T13:57:07.707","AcceptedAnswerId":null,"OwnerUserId":3627108.0,"Title":"Using Theano to compute forward_propagation","Body":"<p>I tried to code the forward propagation with Theano. I define a class name hiddenLayer as below:<\/p>\n\n<p>import theano.tensor as T\nfrom theano import shared\nimport numpy as np\nfrom theano import function<\/p>\n\n<pre><\/pre>\n\n<p>I want to set up a list of hiddenLayer objects and the current's hiddenLayer is the input of the next hiddenLayer. Finally I defined a function named forward buy it raises error and the code is as follow:<\/p>\n\n<pre><\/pre>\n\n<p>The error message is:<\/p>\n\n<pre><\/pre>\n\n<p>Could you tell me why what's the problem with and how to solve it? Thanks<\/p>\n","answers":[{"AnswerId":"23661302","CreationDate":"2014-05-14T17:29:04.027","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is that you override the l.x in the second loop. You can't do that. Once self.x is used in the <strong>init<\/strong>, the result based on it are based on the current instance of self.x. So when you override it, it don't recreate the other stuff on the new x.<\/p>\n\n<p>You should pass x as in input to <strong>init<\/strong>. If None, create one. This is for the first layer. For the other, it should be the previous layer output.<\/p>\n\n<pre><code>def __init__(self, n_in, n_out, x=None):\n   if x is not None:\n      self.x = x\n   else:\n      x = T.dvector('x')\n<\/code><\/pre>\n"}]}
{"QuestionId":23709983,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-05-17T10:01:50.540","AcceptedAnswerId":null,"OwnerUserId":3627108.0,"Title":"Theano function raises ValueError with 'givens' attribute","Body":"<p>I use  function and want to use  to iterate all the input samples. The code is as below:<\/p>\n\n<pre><\/pre>\n\n<p>It eventually raises an error:<\/p>\n\n<pre><\/pre>\n\n<p>Could you tell me why, and how to solve the problem?<\/p>\n","answers":[{"AnswerId":"23996593","CreationDate":"2014-06-02T14:10:47.983","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is this: <code>train_set[index]<\/code><\/p>\n\n<p>Here train_set is a numpy ndarray and index a Theano variable. NumPy don't know how to work with Theano variables. You must convert train_set to a Theano variable like a shared variable:<\/p>\n\n<pre><code>train_set = theano.shared(train_set)\n<\/code><\/pre>\n\n<p>You also need to change your declaration of index as Theano don't support real value for index:<\/p>\n\n<pre><code>index = T.iscalar()\n<\/code><\/pre>\n"}]}
{"QuestionId":23729919,"AnswerCount":1,"Tags":"<python><theano><summarization><deep-learning>","CreationDate":"2014-05-19T04:53:35.457","AcceptedAnswerId":null,"OwnerUserId":3025272.0,"Title":"Text summarization using deep learning techniques","Body":"<p><strong>I am trying to summarize text documents that belong to legal domain.<\/strong> <\/p>\n\n<p>I am referring to the site deeplearning.net on how to implement the deep learning architectures. I have read quite a few research papers on document summarization (both single document and multidocument) but I am unable to figure to how exactly the summary is generated for each document. <\/p>\n\n<p>Once the training is done, the network stabilizes during testing phase. So even if I know the set of features (which I have figured out) that are learnt during the training phase, it would be difficult to find out the importance of each feature (because the weight vector of the network is stabilized) during the testing phase where I will be trying to generate summary for each document. <\/p>\n\n<p>I tried to figure this out for a long time but it's in vain. <\/p>\n\n<p>If anybody has worked on it or have any idea regarding the same, please give me some pointers. I really appreciate your help. Thank you.<\/p>\n","answers":[{"AnswerId":"23765727","CreationDate":"2014-05-20T16:53:55.277","ParentId":null,"OwnerUserId":"1724013","Title":null,"Body":"<p>I think you need to be a little more specific. When you say \"I am unable to figure to how exactly the summary is generated for each document\", do you mean that you don't know how to <em>interpret<\/em> the learned features, or don't you understand the algorithm? Also, \"deep learning techniques\" covers a very broad range of models - which one are you actually trying to use?<\/p>\n\n<p>In the general case, deep learning models do not learn features that are humanly intepretable (albeit, you can of course try to look for correlations between the given inputs and the corresponding activations in the model). So, if that's what you're asking, there really is no good answer. If you're having difficulties understanding the model you're using, I can probably help you :-) Let me know. <\/p>\n"}]}
{"QuestionId":23744826,"AnswerCount":1,"Tags":"<python><networking><theano><nnet>","CreationDate":"2014-05-19T18:34:16.810","AcceptedAnswerId":23819788.0,"OwnerUserId":1735729.0,"Title":"Binary output Neural Network in Python Theano","Body":"<p>As part of a personal project I'm trying to modify the example code given in Theano documentation (<a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">Multilayer Perceptron<\/a>) with my own data.<\/p>\n\n<p>Till now I managed to bring my own (text) data in the format required and I want to build a binary classifier. The thing is that when I write that the number of outputs is 1 i.e.<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The output of my training data (before casting to theano shared type) is like this:<\/p>\n\n<pre><\/pre>\n\n<p>The strange thing is that if I use as a number of output neurons ANYTHING above 1 (e.g. n_out=2), the code is running without any errors but of course now there are many output neurons that have no practical meaning.<\/p>\n\n<p>Could some please explain why the code with binary output seems to give me an error? How can I get this working?<\/p>\n\n<p>Thank you! <\/p>\n","answers":[{"AnswerId":"23819788","CreationDate":"2014-05-23T01:47:43.927","ParentId":null,"OwnerUserId":"534969","Title":null,"Body":"<p>The logistic regression class used as output layer in the MLP tutorial is not the \"standard\" logistic regression, which gives as output a single value and discriminates between just two classes, but rather a Multinomial Logistic Regression (a.k.a Softmax Regression), which gives as output one value for each class, telling the probability of the input belonging to them. So, if you have 10 classes, you'll also need 10 units and obviously the sum of all output units equals 1, since it's a probability distribution.<\/p>\n\n<p>Despite of the class name used (\"LogistRegression\"), its doctring in the <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"nofollow\">linked source code<\/a> leaves no doubts of its real intent (<code>'''Multi-class Logistic Regression Class [...]'''<\/code>).<\/p>\n\n<p>Whereas in your problem you have two classes, you'll also need 2 output units and the value for your <code>n_out<\/code> must be 2 instead of 1. Of course, with two classes the value for one output will be always 1 minus the value of the other.<\/p>\n\n<p>Also, check if you really need int64 instead of int32. Theano has much better support for the second.<\/p>\n"}]}
{"QuestionId":23783391,"AnswerCount":1,"Tags":"<theano><deep-learning>","CreationDate":"2014-05-21T12:37:08.187","AcceptedAnswerId":null,"OwnerUserId":2166433.0,"Title":"ImportError: No module named 'theano.floatX'","Body":"<p>I am following a tutorial to create <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet\" rel=\"nofollow\">convolution neural network with Theano<\/a>. Although, I got a problem in a piece of code:<\/p>\n\n<pre><\/pre>\n\n<p>I loaded floatX with: <\/p>\n\n<pre><\/pre>\n\n<p>and checked with:<\/p>\n\n<pre><\/pre>\n\n<p>But still cannot load the module , which should be in , judging from <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/floatX.html\" rel=\"nofollow\">documentation<\/a>. Does somebody know where can I find it?<\/p>\n\n<p>Thank you in advance! <\/p>\n","answers":[{"AnswerId":"23789646","CreationDate":"2014-05-21T17:05:01.477","ParentId":null,"OwnerUserId":"534969","Title":null,"Body":"<p>This sections of the convnet tutorial has a bug or is very outdated. Symbolic variables in Theano are located in theano.tensor package. This package theano.floatX even doesn't exist!<\/p>\n\n<p>The current version in the tutorial github repository works fine. They allocate the symbolic variable the right way:<\/p>\n\n<pre><code># allocate symbolic variables for the data\n    index = T.lscalar()  # index to a [mini]batch\n    x = T.matrix('x')   # the data is presented as rasterized images\n    y = T.ivector('y')  # the labels are presented as 1D vector of\n                        # [int] labels\n<\/code><\/pre>\n\n<p>Browsing the tutorial repository I found the <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/commit\/21065c8271913be6667357cbd3df5d37bba144d1\" rel=\"nofollow\">revision where this bug was corrected<\/a>. \nThey seem to have forgotten to update the tutorial text with this fix.<\/p>\n"}]}
{"QuestionId":23788179,"AnswerCount":2,"Tags":"<numpy><pycuda><theano><deep-learning>","CreationDate":"2014-05-21T15:53:16.577","AcceptedAnswerId":23811048.0,"OwnerUserId":2462245.0,"Title":"Is there a GPU accelerated numpy.max(X, axis=0) implementation in Theano?","Body":"<p>Do we have a GPU accelerated of version of  in Theano.\nI looked into the documentation and found , but it is 4-5 times slower than the numpy implementation. <\/p>\n\n<p>I can assure you, it is not slow because of some bad choice of matrix size. Same matrix under theano.tensor.exp is 40 times faster than its numpy counterpart.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"23797179","CreationDate":"2014-05-22T03:05:50.957","ParentId":null,"OwnerUserId":"2014584","Title":null,"Body":"<p>The <code>max<\/code> and <code>exp<\/code> operations are fundamentally different; <code>exp<\/code> (and other operations like addition, <code>sin<\/code>, etc.) is an elementwise operation that is embarrassingly parallelizable, while <code>max<\/code> requires a parallel-processing scan algorithm that basically builds up a tree of pairwise comparisons over an array. It's not impossible to speed up <code>max<\/code>, but it's not as easy as <code>exp<\/code>.<\/p>\n\n<p>Anyway, the <code>theano<\/code> implementation of <code>max<\/code> basically consists of the following lines (in theano\/tensor\/basic.py):<\/p>\n\n<pre><code>try:\n    out = max_and_argmax(x, axis)[0]\nexcept Exception:\n    out = CAReduce(scal.maximum, axis)(x)\n<\/code><\/pre>\n\n<p>where <code>max_and_argmax<\/code> is a bunch of custom code that, to my eye, implements a max+argmax operation using <code>numpy<\/code>, and <code>CAReduce<\/code> is a generic GPU-accelerated scan operation used as a fallback (which, according to the comments, doesn't support <code>grad<\/code> etc.). You could try using the fallback directly and see whether that is faster, maybe something like this:<\/p>\n\n<pre><code>from theano.tensor.elemwise import CAReduce\nfrom theano.scalar import maximum\n\ndef mymax(X, axis=None):\n    CAReduce(maximum, axis)(X)\n<\/code><\/pre>\n"},{"AnswerId":"23811048","CreationDate":"2014-05-22T15:24:39.330","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The previous answer is partial. The suggestion should not work, as the work around is the one used in the final compiled code. There is optimization that will do this transformation automatically.<\/p>\n\n<p>The title of the question isn't the same as the content. They differ by the axis argument. I'll answer both questions.<\/p>\n\n<p>If the axis is 0 or None we support this on the GPU for that operation for matrix. If the axis is None, we have a basic implementation that isn't well optimized as it is harder to parallelize. If the axis is 0, we have a basic implementation, but it is faster as it is easier to parallelize.<\/p>\n\n<p>Also, how did you do your timing? If you just make one function with only that operation and test it via the device=gpu flags to do your comparison, this will include the transfer time between CPU and GPU. This is a memory bound operation, so if you include the transfer in your timming, personnaly I don't expect any speed op for that case. To see only the GPU operation, use Theano profiler: run with the Theano flag profile=True.<\/p>\n"}]}
{"QuestionId":23978598,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2014-06-01T09:50:55.863","AcceptedAnswerId":23978780.0,"OwnerUserId":1828289.0,"Title":"Why does theano conv2d add empty dimension?","Body":"<p>I am playing around with some simple Theano code, and I ran into the following:<\/p>\n\n<pre><\/pre>\n\n<p>Result:\n(1, 91, 100)<\/p>\n\n<p>The result of a 2d convolution of 2d inputs is expected to be 2d, but it is actually 3d.  Why?<\/p>\n","answers":[{"AnswerId":"23978780","CreationDate":"2014-06-01T10:14:47.000","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The docstring of <code>conv2d<\/code> says <em>signal.conv.conv2d performs a basic 2D convolution of the input with the\ngiven <strong>filters<\/strong>.<\/em> (note the plural)<\/p>\n\n<p>You could pass it <strong>several<\/strong> filters and it will return the convolutions with all of those. Try e.g.<\/p>\n\n<pre><code>c = conv2d(m,np.array([w, w, w]))\nf = theano.function([m], c)\nprint f(numpy.ones([100,100], dtype=numpy.float32)).shape  # outputs (3, 91, 100)\n<\/code><\/pre>\n\n<p>So it seems that by default it will add a degenerate axis if you only pass 1 filter (probably because it adds this axis internally to your filter if you didn't pass it in that way yourself. In other words, it doesn't keep track of the input shape in order to return something that corresponds. Looks like a design choice more than anything else.)<\/p>\n"}]}
{"QuestionId":24172613,"AnswerCount":1,"Tags":"<python-2.7><lambda><theano>","CreationDate":"2014-06-11T21:08:54.840","AcceptedAnswerId":24187030.0,"OwnerUserId":3113501.0,"Title":"Theano scan function","Body":"<p>Example taken from: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/scan.html<\/a><\/p>\n\n<pre><\/pre>\n\n<p>What is prior_result? More accurately, where is prior_result defined?<\/p>\n\n<p>I have this same question for lot of the examples given on:<a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/scan.html<\/a><\/p>\n\n<p>For example, <\/p>\n\n<pre><\/pre>\n\n<p>Where is power and free_variables defined?<\/p>\n","answers":[{"AnswerId":"24187030","CreationDate":"2014-06-12T14:31:15.513","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This is using a Python feature call \"lambda\". lambda are unnamed python function of 1 line. They have this forme:<\/p>\n\n<pre><code>lambda [param...]: code\n<\/code><\/pre>\n\n<p>In your example it is:<\/p>\n\n<pre><code>lambda prior_result, A: prior_result * A\n<\/code><\/pre>\n\n<p>This is a function that take prior_result and A as input. This function, is passed to the scan() function as the fn parameter. scan() will call it with 2 variables. The first one will be the correspondance of what was provided in the output_info parameter. The other is what is provided in the non_sequence parameter.<\/p>\n"}]}
{"QuestionId":24183108,"AnswerCount":1,"Tags":"<theano><matrix-indexing>","CreationDate":"2014-06-12T11:22:30.603","AcceptedAnswerId":24186817.0,"OwnerUserId":1099534.0,"Title":"Indexing in Theano","Body":"<p>How can I index a matrix in Theano by a vector of indices?\n<br>\nMore precisely, say that:<\/p>\n\n<ul>\n<li><em>v<\/em> has type theano.tensor.vector (e.g. [0,2]) <\/li>\n<li><em>A<\/em> has type theano.tensor.matrix (e.g. [[1,0,0], [0,1,0], [0,0,1]])<\/li>\n<\/ul>\n\n<p>The desired result is [[1,0,0], [0,0,1]].\n<br>\nI mention that my goal is to convert a list of indices into a matrix of one-hot row vectors, where the indices indicate the hot position. My initial attempt was to let A = theano.tensor.eye and index it using the vector of indices.<\/p>\n","answers":[{"AnswerId":"24186817","CreationDate":"2014-06-12T14:22:19.933","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You can do:<\/p>\n\n<pre><code>A[v]\n<\/code><\/pre>\n\n<p>It will do what you want.<\/p>\n"}]}
{"QuestionId":24203451,"AnswerCount":2,"Tags":"<numpy><matrix><theano>","CreationDate":"2014-06-13T10:43:07.220","AcceptedAnswerId":24203540.0,"OwnerUserId":1099534.0,"Title":"Matrices with different row lengths in numpy","Body":"<p>Is there a way of defining a matrix (say <em>m<\/em>) in numpy with rows of different lengths, but such that <em>m<\/em> stays 2-dimensional (i.e. m.ndim = 2)?<\/p>\n\n<p>For example, if you define <em>m = numpy.array([[1,2,3], [4,5]])<\/em>, then <em>m.ndim<\/em> = 1. I understand why this happens, but I'm interested if there is any way to trick numpy into viewing <em>m<\/em> as 2D. One idea would be padding with a dummy value so that rows become equally sized, but I have lots of such matrices and it would take up too much space. The reason why I really need <em>m<\/em> to be 2D is that I am working with Theano, and the tensor which will be given the value of <em>m<\/em> expects a 2D value.<\/p>\n","answers":[{"AnswerId":"24203540","CreationDate":"2014-06-13T10:47:55.023","ParentId":null,"OwnerUserId":"166749","Title":null,"Body":"<p>No, this is not possible. NumPy arrays need to be rectangular in every pair of dimensions. This is due to the way they map onto memory buffers, as a pointer, itemsize, stride triple.<\/p>\n\n<p>As for this taking up space: <code>np.array([[1,2,3], [4,5]])<\/code> actually takes up more space than a 2\u00d73 array, because it's an array of two pointers to Python lists (and even if the elements were converted to arrays, the memory layout would still be inefficient).<\/p>\n"},{"AnswerId":"24206476","CreationDate":"2014-06-13T13:26:46.603","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I'll give here very new information about Theano. We have a new TypedList() type, that allow to have python list with all elements with the same type: like 1d ndarray. All is done, except the documentation.<\/p>\n\n<p>There is limited functionality you can do with them. But we did it to allow looping over the typed list with scan. It is not yet integrated with scan, but you can use it now like this:<\/p>\n\n<pre><code>import theano\nimport theano.typed_list\n\na = theano.typed_list.TypedListType(theano.tensor.fvector)()\ns, _ = theano.scan(fn=lambda i, tl: tl[i].sum(),\n                   non_sequences=[a],\n                   sequences=[theano.tensor.arange(2, dtype='int64')])\n\nf = theano.function([a], s)\nf([[1, 2, 3], [4, 5]])\n<\/code><\/pre>\n\n<p>One limitation is that the output of scan must be an ndarray, not a typed list.<\/p>\n"}]}
{"QuestionId":24205187,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2014-06-13T12:22:09.910","AcceptedAnswerId":null,"OwnerUserId":1099534.0,"Title":"Variable-length tensors in Theano","Body":"<p>This question refers to best practices in Theano. Here is what I am trying to do:<\/p>\n\n<p>I am building a neural network for an SMT system. In this context, I conceptually represent sentences as variable-length lists of words, and words as fixed-length lists of integers. Ideally, I would like to represent my corpus as a 3D tensor (first dimension = sentences in corpus, second dimension = words in sentence, third dimension = integer features in words). The difficulty is that sentences have variable length and, to my knowledge, tensors in Theano have the strict requirement that all lengths in one dimension must be the same.<\/p>\n\n<p>Solutions I have thought of include:<\/p>\n\n<ol>\n<li> Use padding with dummy words so that sentences become equally sized. But this means that whenever I iterate over a sentence, I need to include special code to discard the padding.\n<li> Represent the corpus as a vector of matrices. However, this makes it hard to work with certain functions. For instance, if I want to add up the representations of all the words in a sentence, I can't simply use *corpus.sum(axis=1)*. I would have to loop over sentences, do *sentence.sum(axis=0)*, and then gather the results into another tensor.\n<\/ol>\n\n<p>My question is: which of these alternatives are preferred, or is there a better one?<\/p>\n","answers":[{"AnswerId":"30507375","CreationDate":"2015-05-28T13:02:29.613","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The first option is probably the best option in most cases. It's what I do though it does mean passing around a separate vector of sentence lengths and masking certain results to eliminate the padding region when needed.<\/p>\n\n<p>In general, if you want to perform a consistent operation to all sentences then you'll usually get much better speed applying that operation to a single 3D tensor than sequentially to a series of matrices. This is especially true for operations running on a GPU.<\/p>\n\n<p>If you're using scan operations the speed differences will become even more magnified. You'll be better off scanning over a 3D tensor and operating on a per-word matrix in your step function that covers all (or a minibatch of) sentences. If needed, you may need to know which rows of that matrix are real data and which are padding. As an aside, I find that setting the first dimension of a 3D tensor to be the temporal\/sequence position dimension helps when using scan, which always scans over the first dimension.<\/p>\n\n<p>Often, using the value zero as your padding value will result in the padding have no impact on your operations.<\/p>\n\n<p>The other option, looping over the sentences, would mean mixing Theano and Python code which can make some computations difficult or impossible. For example, getting the gradient of a cost function with respect to some parameters over a all (or batch) of your sentences may not be possible if the data is stored in lots of separate matrices.<\/p>\n"}]}
{"QuestionId":24225948,"AnswerCount":1,"Tags":"<math><numpy><theano>","CreationDate":"2014-06-15T02:04:34.090","AcceptedAnswerId":null,"OwnerUserId":1641608.0,"Title":"Revisit forcing python math functions to operate on float32","Body":"<p>I am aware of howto force python operations to work on float32:\n<a href=\"https:\/\/stackoverflow.com\/questions\/20768590\/how-to-force-python-float-operation-on-float32-rather-than-float64\">How to force python float operation on float32 rather than float64<\/a><\/p>\n\n<p>But there is no Q or A about forcing the built in functions to work on float32. \nI wanted to ask how about forcing the built-in math or numpy functions such as math.sqrt or numpy.sqrt to work on float32. FYI, I could not comment on the question yet.<\/p>\n\n<p>In theano, we can easily configure the function, e.g., sqrt to work on float32 or float64 as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and the result is:<\/p>\n\n<pre><\/pre>\n\n<p>Now I tried to force math.sqrt and numpy.sqrt to do the same as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But the result is still seems in float64 (I confirmed it \nthat the result would be the same, i.e., 3.87298334621,\nif I set theano.config.floatX='float64'):<\/p>\n\n<pre><\/pre>\n\n<p>I am curious to know howto force math.sqrt, numpy.sqrt to work on float32?<\/p>\n","answers":[{"AnswerId":"24225977","CreationDate":"2014-06-15T02:10:54.737","ParentId":null,"OwnerUserId":"2357112","Title":null,"Body":"<pre><code>&gt;&gt;&gt; type(numpy.sqrt(numpy.float32(2)))\n&lt;type 'numpy.float32'&gt;\n<\/code><\/pre>\n\n<p><code>numpy.sqrt<\/code> already does what you want. On the other hand, the <code>math<\/code> functions always cast their input to <code>float<\/code> and return <code>float<\/code>, with no option to change that. Stick with NumPy operations instead of the <code>math<\/code> module for NumPy data types, and you should be fine.<\/p>\n"}]}
{"QuestionId":24229361,"AnswerCount":1,"Tags":"<python><gpgpu><theano>","CreationDate":"2014-06-15T11:58:23.743","AcceptedAnswerId":24329482.0,"OwnerUserId":826983.0,"Title":"Theano: Indexing inside a compiled function (GPU)","Body":"<p>Since Theano allows to <em>update<\/em> the memory on the graphics card DRAM by simply defining which memory area has to be updated and how this should be done, I was wondering if the following is somehoe possible (imho it should be).<\/p>\n\n<p>I have a 2x5 randomly initialized matrix which first column will be initialized with start-values. I would like to write a function that depends of the preceeding column and updates the next one based on arbitrary calulations. <\/p>\n\n<p>I think this code explains it very well:<\/p>\n\n<p><strong>Note:<\/strong> This code is <em>not<\/em> working, it's just an illustration:<\/p>\n\n<pre><\/pre>\n\n<p>My desired output would be:<\/p>\n\n<pre><\/pre>\n\n<p>but instead I get <\/p>\n\n<pre><\/pre>\n\n<p>Can anybody help me here?<\/p>\n\n<p><strong>Wow:<\/strong> This question is 9 minutes after asking already in the Top 4 of google results for \"<em>Theano indexing gpu<\/em>\" for me. O_o<\/p>\n","answers":[{"AnswerId":"24329482","CreationDate":"2014-06-20T14:22:41.983","ParentId":null,"OwnerUserId":"3760518","Title":null,"Body":"<p>Have a look at: <a href=\"https:\/\/stackoverflow.com\/questions\/15917849\/how-can-i-assign-update-subset-of-tensor-shared-variable-in-theano\">How can I assign\/update subset of tensor shared variable in Theano?<\/a><\/p>\n\n<p>For your code this translates to:<\/p>\n\n<pre><code>from theano import function, sandbox, shared\nimport theano.tensor as T\nimport numpy as np\n\nreservoirSize = 2\nsamples       = 5\n\n# To initialize _mat first column\n_vec      = np.asarray([1 for i in range(reservoirSize)], np.float32)\n\n# Random matrix, first column will be initialized correctly (_vec)\n_mat      = np.asarray(np.random.rand(reservoirSize, samples), np.float32)\n_mat[:,0] = _vec\n\nprint \"Init:\\n\",_mat\n\n_mat      = shared(_mat)\n\nidx  = T.iscalar()\ntest = function([idx], updates= {\n            # -&gt; instead of\n            #_mat[:,idx]:sandbox.cuda.basic_ops.gpu_from_host(_mat[:,idx-1] * 2)\n            # -&gt; do this:\n            _mat:T.set_subtensor(_mat[:,idx], _mat[:,idx-1]*2) \n            })\n\nfor i in range(1, samples):\n    test(i)\n\nprint \"Done:\\n\",_mat.get_value()  # use get_value() here to retrieve the data\n<\/code><\/pre>\n"}]}
{"QuestionId":24230227,"AnswerCount":1,"Tags":"<theano><machine-translation>","CreationDate":"2014-06-15T13:47:18.537","AcceptedAnswerId":24244862.0,"OwnerUserId":1099534.0,"Title":"Typed Lists in Theano","Body":"<p>Consider the following machine translation problem. Let  be a source sentence and  be a target sentence. Both sentences are conceptually represented as lists of indices, where the indices correspond to the position of the words in the associated dictionaries. Example:<\/p>\n\n<pre><\/pre>\n\n<p>Note that  and  don't necessarily have the same length. Now let  and  be sets of such instances. In other words, they are a parallel corpus. Example:<\/p>\n\n<pre><\/pre>\n\n<p>Note that not all 's in  have the same length. That is, sentences have variable numbers of words.\n<p>\nI am implementing a machine translation system in Theano, and the first design decision is what kind of data structures to use for  and . From one of the answers posted on <a href=\"https:\/\/stackoverflow.com\/questions\/24203451\/matrices-with-different-row-lengths-in-numpy\">Matrices with different row lengths in numpy<\/a> , I learnt that typed lists are a good solution for storing variable length tensors.\n<p>\nHowever, I realise that they complicate my code a lot. Let me give you one example. Say that we have two typed lists  and  and aim to calculate the negative loss likelihood. If they were regular tensors, a simple statement like this would suffice:<\/p>\n\n<pre><\/pre>\n\n<p>But  can only be applied to tensors, so in case of typed lists I have to iterate over them and apply the function separately to each element:<\/p>\n\n<pre><\/pre>\n\n<p>On top of making my code more and more messy, these problems propagate. For instance, if I want to calculate the gradient of the loss, the following doesn't work anymore:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know exactly why it doesn't work. I'm sure it has to do with the type of , but I am not interested in investigating any further how I could make it work. The mess is growing exponentially, and what I would like is to know whether I am using typed lists in the wrong way, or if it is time to give up on them because they are not well enough supported yet.<\/p>\n","answers":[{"AnswerId":"24244862","CreationDate":"2014-06-16T13:25:57.407","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Typed list isn't used by anybody yet. But the idea for having them is that you iterate on them with scan for each sentence. Then you do everything you need in 1 scan. You don't do 1 scan for each operation.<\/p>\n\n<p>So the scan is only used to do the iteration on each example in the minibatch, and the inside of scan is all what is done on one example.<\/p>\n\n<p>We haven't tested typed list with grad yet. It is possible that it is missing some implementations.<\/p>\n"}]}
{"QuestionId":24233605,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-06-15T20:01:43.980","AcceptedAnswerId":24234094.0,"OwnerUserId":826983.0,"Title":"Concatenation\/creation of symbolic expressions\/statements","Body":"<p>I want to create something like this programmatically:<\/p>\n\n<pre><\/pre>\n\n<p>We got  so a is that whole statement. This is not doing anything yet. Then there is  which is depending on  and is another statement itself.This goes on until . I know it looks ugly but it just updates the column of a vector like  for  returning a vector looking like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now Imagine I would want to do this a thousand times.. therefore I am wondering if I could generate such statements since they follow an easy pattern. <\/p>\n\n<p>These \"<em>concatenated<\/em>\" statements are not evaluated until <\/p>\n\n<pre><\/pre>\n\n<p>which is when they are getting compiled to  and<\/p>\n\n<pre><\/pre>\n\n<p>does execute everything.<\/p>\n","answers":[{"AnswerId":"24234094","CreationDate":"2014-06-15T21:02:48.470","ParentId":null,"OwnerUserId":"3741258","Title":null,"Body":"<p>You can use function nesting:<\/p>\n\n<pre><code>def nest(f, g):\n    def h(x):\n        return f(g(x), x)\n    return h\n\nexpr = lambda (a,b) : a\nexpr = nest((lambda x, (a, b): x + (a - b)), expr)\nexpr = nest((lambda x, (a, b): x + (a - b)), expr)\n\nprint expr((1,2)) # prints -1\n<\/code><\/pre>\n\n<p>Regarding the example code, you could do something like (modifying <code>nest<\/code> to use no arguments):<\/p>\n\n<pre><code>def nest(f, g):\n    def h():\n        return f(g())\n    return h\n\nexpr = lambda: (_vec, _init)[1]\nexpr = nest(lambda x: T.set_subtensor(x[1], x[0] * 2)[1], expr)\nexpr = nest(lambda x: T.set_subtensor(x[2], x[1] * 2)[1], expr)\nexpr = nest(lambda x: T.set_subtensor(x[3], x[2] * 2)[1], expr)\nexpr = nest(lambda x: T.set_subtensor(x[4], x[3] * 2)[1], expr)\n\ntest_vector = function([], outputs=expr)\nsubt = test_vector()\n<\/code><\/pre>\n"}]}
{"QuestionId":24263390,"AnswerCount":3,"Tags":"<python><matrix><theano>","CreationDate":"2014-06-17T12:07:19.943","AcceptedAnswerId":24268694.0,"OwnerUserId":1099534.0,"Title":"From list of indices to one-hot matrix","Body":"<p>What is the best (elegant and efficient) way in Theano to convert a vector of indices to a matrix of zeros and ones, in which every row is the one-of-N representation of an index?<\/p>\n\n<pre><\/pre>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40622953","CreationDate":"2016-11-16T02:11:18.507","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>There's now a built in function for this <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html#theano.tensor.extra_ops.to_one_hot\" rel=\"nofollow noreferrer\"><code>theano.tensor.extra_ops.to_one_hot<\/code><\/a>.<\/p>\n\n<pre><code>y = tensor.as_tensor([3,2,1])\nfn = theano.function([], tensor.extra_ops.to_one_hot(y, 4))\nprint fn()\n# [[ 0.  0.  0.  1.]\n#  [ 0.  0.  1.  0.]\n#  [ 0.  1.  0.  0.]]\n<\/code><\/pre>\n"},{"AnswerId":"24268694","CreationDate":"2014-06-17T16:14:20.890","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I didn't compare the different option, but you can also do it like this. It don't request extra memory.<\/p>\n\n<pre><code>import numpy as np\nimport theano\n\nn_val = 4\nv_val = np.asarray([1,0,3])\nidx = theano.tensor.lvector()\nz = theano.tensor.zeros((idx.shape[0], n_val))\none_hot = theano.tensor.set_subtensor(z[theano.tensor.arange(idx.shape[0]), idx], 1)\nf = theano.function([idx], one_hot)\nprint f(v_val)[[ 0.  1.  0.  0.]\n [ 1.  0.  0.  0.]\n [ 0.  0.  0.  1.]]\n<\/code><\/pre>\n"},{"AnswerId":"24264952","CreationDate":"2014-06-17T13:19:09.633","ParentId":null,"OwnerUserId":"1099534","Title":null,"Body":"<p>It's as simple as:<\/p>\n\n<pre><code>convert = t.eye(n,n)[v]\n<\/code><\/pre>\n\n<p>There still might be a more efficient solution that doesn't require building the whole identity matrix. This might be problematic for large n and short v's.<\/p>\n"}]}
{"QuestionId":24286238,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-06-18T12:58:30.263","AcceptedAnswerId":24314997.0,"OwnerUserId":1099534.0,"Title":"Obscure Looping Error in Theano","Body":"<p>I am trying to simulate a repeat-until loop in Theano:<\/p>\n\n<pre><\/pre>\n\n<p>I encountered the following strange behaviour. Let  be a constant. Whenever  is a constant, everything works fine. However, when  is a scalar, I get an error related to optimisation:<\/p>\n\n<pre><\/pre>\n\n<p>I'd appreciate if someone could help me understand the error. I'm assuming the  doesn't refer to  or , because I can print their  and see that they do have one. Other than that, I can't make any sense out of it.<\/p>\n\n<p>[Edit] This is not a fatal error. The code runs normally and the process finishes with exit code 0. It looks like Theano is trying to optimise the graph and fails to do so, which doesn't really impact the program.<\/p>\n","answers":[{"AnswerId":"24314997","CreationDate":"2014-06-19T19:41:47.303","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The traceback indicate that in the function equal_compuations(), we didn't cover all case, when doing some comparison.<\/p>\n\n<p>I have a PR with a fix for it here:<\/p>\n\n<pre><code>https:\/\/github.com\/Theano\/Theano\/pull\/1928\n<\/code><\/pre>\n\n<p>thanks for the report.<\/p>\n\n<p>Your [edit] section, indicate me that you cut some of the errors message. If this happear during optimization with a warning, it mean an optimization was just skipped. It is possible that the optimization just don't apply, but it could be possible that with the fix, now the optimization apply. If that is the case, there could be some speed up with the fix.<\/p>\n"}]}
{"QuestionId":24331893,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2014-06-20T16:35:15.843","AcceptedAnswerId":null,"OwnerUserId":1784599.0,"Title":"Theano SDE example","Body":"<p>I'm trying to work through an example on solving SDEs on GPU using Theano found <a href=\"http:\/\/www.nehalemlabs.net\/prototype\/blog\/2013\/10\/17\/solving-stochastic-differential-equations-with-theano\/\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>I'm stuck with a GpuDimShuffle error, but I'm not seeing how any of the dims don't match...<\/p>\n\n<pre><\/pre>\n\n<p>Yielding:<\/p>\n\n<p>TypeError: ('The following error happened while compiling the node', Rebroadcast{0(GpuDimShuffle{x,0}.0), '\\n', 'super(type, obj): obj must be an instance or subtype of type')<\/p>\n\n<p>I'm on Theano 0.6.0<\/p>\n","answers":[{"AnswerId":"24410185","CreationDate":"2014-06-25T13:33:15.273","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>If you update to the development version, it work for me:<\/p>\n\n<p><a href=\"http:\/\/www.deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>We try to keep the development version very stable and I recomment everybody to use it, since it contain many fixes since the last release.<\/p>\n\n<p>If that don't fix it, the error is specific to your OS or python version. We will need more information about it. Also always provide the full error message with the traceback. This provide much more debugging information.<\/p>\n"}]}
{"QuestionId":24347565,"AnswerCount":1,"Tags":"<theano><deep-learning><automatic-differentiation><autoencoder>","CreationDate":"2014-06-22T02:15:41.530","AcceptedAnswerId":null,"OwnerUserId":2782619.0,"Title":"Does Theano support variable split?","Body":"<p>In my Theano program, I want to split the tensor matrix into two parts, with each of them making different contributions to the error function. Can anyone tell me whether automatic differentiation support this?<\/p>\n\n<p>For example, for a tensor matrix variable M, I want to split it into M1=M[:300,] and M2=M[300:,], then the cost function is defined as 0.5* M1 * w + 0.8*M2*w. Is it still possible to get the gradient with T.grad(cost,w)? <\/p>\n\n<p>Or more specifically, I want to construct an Autoencoder with different features having different weights in contribution to the total cost. <\/p>\n\n<p>Thanks for anyone who answers my question.<\/p>\n","answers":[{"AnswerId":"24410773","CreationDate":"2014-06-25T13:59:26.653","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano support this out of the box. You have nothing particular to do. If Theano don't support something in the crash, it should raise an error. But you won't have it for this, if there isn't problem in the way you call it. But the current pseudo-code should work.<\/p>\n"}]}
{"QuestionId":24402213,"AnswerCount":1,"Tags":"<python><ubuntu><theano><theano-cuda>","CreationDate":"2014-06-25T07:03:05.193","AcceptedAnswerId":null,"OwnerUserId":3568055.0,"Title":"theano.test() : optimization failure due to constant_folding (on ubuntu)","Body":"<p>When running theano.test() on an Ubuntu operating system, some error message about an optimization failure is produced as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know a way to fix these problem, or what exactly is going on?<\/p>\n","answers":[{"AnswerId":"24409647","CreationDate":"2014-06-25T13:10:35.310","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This can be caused by many things. The error is related to the GPU. So first, make sure you can compile nvidia example and that they run fine. To be sure this isn't the problem.<\/p>\n\n<p>The problem is that Theano isn't able to import a GPU module it compiled, because he didn't found the symbol it need. This missing symbol \"_Z25CudaNdarray_CopyFromArrayP11CudaNdarrayP23tagPyArrayObject_fields\" is in a shared library that Theano already pre-compiled.<\/p>\n\n<p>What is your OS? Make sure to update to the latest development version of Theano. There was a fix recently (Monday if my memory is exact) that could solve this on some OS.<\/p>\n"}]}
{"QuestionId":24431621,"AnswerCount":1,"Tags":"<python><gradient><backpropagation><theano>","CreationDate":"2014-06-26T13:17:00.947","AcceptedAnswerId":24437169.0,"OwnerUserId":1099534.0,"Title":"Does Theano do automatic unfolding for BPTT?","Body":"<p>I am implementing an RNN in Theano and I have difficulties training it. It doesn't even come near to memorising the training corpus. My mistake is most likely caused by me not understanding exactly how Theano copes with backpropagation through time. Right now, my code is as simple as it gets:<\/p>\n\n<pre><\/pre>\n\n<p>My question is: given that my network is recurrent, does this automatically do the unfolding of the architecture into a feed-forward one? On one hand, <a href=\"https:\/\/github.com\/gwtaylor\/theano-rnn\/blob\/master\/rnn_minibatch.py\" rel=\"noreferrer\">this<\/a> example does exactly what I am doing. On the other hand, <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/RAfkjvI2CEU\" rel=\"noreferrer\">this<\/a> thread makes me think I'm wrong.<\/p>\n\n<p>In case it does do the unfolding for me, how can I truncate it? I can see that there is a way, from the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"noreferrer\">documentation<\/a> of , but I can't come up with the code to do it.<\/p>\n","answers":[{"AnswerId":"24437169","CreationDate":"2014-06-26T17:57:21.027","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>I wouldn't say it does automatic \"unfolding\" - rather, Theano has a notion of what variables are connected, and can pass updates along that chain. If this is what you mean by unfolding, then maybe we are talking about the same thing. <\/p>\n\n<p>I am stepping through this as well, but using <a href=\"https:\/\/11350770138305416713.googlegroups.com\/attach\/2628c6c56beb34bf\/rnn.py?part=0.1&amp;view=1&amp;vt=ANaJVrFlEcr9KydEXVwhub2Lhwv0bs8OZIc678NDDHXGzyPIfr39rH0EiiXXbi4Q9GHbNmAhFQzp7l3hapna0fV3lyR1ICmPe_rij_aL292mq8ioP2lccQg\" rel=\"nofollow noreferrer\">Rasvan Pascanu's rnn.py<\/a> code (from <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/OcdVuTZ19Dc\/3m1Xlb1xxmoJ\" rel=\"nofollow noreferrer\">this thread<\/a>) for reference. It seems much more straightforward for a learning example. <\/p>\n\n<p>You might gain some value from visualizing\/drawing graphs from the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/printing_drawing.html\" rel=\"nofollow noreferrer\">tutorial<\/a>. There is also set of slides online with a <a href=\"http:\/\/www.cs.bham.ac.uk\/~jxb\/INC\/l12.pdf\" rel=\"nofollow noreferrer\">simple drawing<\/a> that shows the diagram from a 1 layer \"unfolding\" of an RNN, which you discuss in your post. <\/p>\n\n<p>Specifically, look at the <code>step<\/code> function:<\/p>\n\n<pre><code>def step(u_t, h_tm1, W, W_in, W_out):\n    h_t = TT.tanh(TT.dot(u_t, W_in) + TT.dot(h_tm1, W))\n    y_t = TT.dot(h_t, W_out)\n    return h_t, y_t\n<\/code><\/pre>\n\n<p>This function represents the \"simple recurrent net\" shown in <a href=\"http:\/\/www.cs.bham.ac.uk\/~jxb\/INC\/l12.pdf\" rel=\"nofollow noreferrer\">these slides, pg 10<\/a>. When you do updates, you simply pass the gradient w.r.t. W, W_in, and W_out, respectively (remember that y is connected to those three via the <code>step<\/code> function! This is how the gradient magic works).<\/p>\n\n<p>If you had multiple W layers (or indexes into one big W, as I believe gwtaylor is doing), then that would create multiple layers of \"unfolding\". From what I understand, this network only looks 1 step backward in time. If it helps, <a href=\"https:\/\/github.com\/lmjohns3\/theano-nets\" rel=\"nofollow noreferrer\">theanonets<\/a> also has an RNN implementation in Theano.<\/p>\n\n<p>As an additional note, training RNNs with BPTT is <em>hard<\/em>. <a href=\"http:\/\/www.cs.utoronto.ca\/~ilya\/pubs\/ilya_sutskever_phd_thesis.pdf\" rel=\"nofollow noreferrer\">Ilya Sutskever's dissertation<\/a> discusses this at great length - if you can, try to tie into a <a href=\"https:\/\/github.com\/boulanni\/theano-hf\" rel=\"nofollow noreferrer\">Hessian Free optimizer, there is also a reference RNN implementation here<\/a>. Theanets also does this, and may be a good reference.<\/p>\n"}]}
{"QuestionId":24459062,"AnswerCount":1,"Tags":"<python><macos><matplotlib><theano>","CreationDate":"2014-06-27T18:57:28.570","AcceptedAnswerId":null,"OwnerUserId":3784367.0,"Title":"Compilation error ld: library not found for -lgcc_ext in MacOSX","Body":"<ol>\n<li><p>I'm trying to replicate a tutorial example in the <a href=\"http:\/\/deeplearning.net\/software\/pylearn2\/\" rel=\"nofollow\">pyLearn2<\/a> documentation. When I run , in the example, I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>Referenced from: \/System\/Library\/Frameworks\/ImageIO.framework\/Versions\/A\/ImageIO<\/p>\n\n<p>Expected in: <\/p>\n\n<p>I can import  and image in python in command line. Could someone help me understand what it is complaining about and how to fix the error?<\/p><\/li>\n<li><p>Another problem (which may or may not be related to the problem above) is a linking error<\/p>\n\n<p>Problem occurred during compilation with the command line below:<\/p>\n\n<pre><\/pre>\n\n<p>I am running  on Mac OSX Mavericks. I am not sure how to fix the error: the library seems to be in my system in several places:<\/p>\n\n<pre><\/pre>\n\n<p>I am not sure which one to link and how I should link.<\/p>\n\n<p>I have changed (added) the paths to  which didn't fix the problem.<\/p>\n\n<p>Any help would be greatly appreciated.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"29567965","CreationDate":"2015-04-10T18:03:24.547","ParentId":null,"OwnerUserId":"1318153","Title":null,"Body":"<p>It looks like <a href=\"https:\/\/lists.macosforge.org\/pipermail\/macports-users\/2010-November\/022662.html\" rel=\"nofollow\">one possible solution<\/a> is to manually create a link to the .dylib file:<\/p>\n\n<blockquote>\n  <p>sudo ln -sf\n  \/System\/Library\/Frameworks\/OpenGL.framework\/Versions\/A\/Libraries\/libGL.dylib\n  \/opt\/local\/lib\/libGL.dylib<\/p>\n  \n  <p>sudo ln -sf\n  \/System\/Library\/Frameworks\/ApplicationServices.framework\/Versions\/A\/Frameworks\/ImageIO.framework\/Versions\/A\/Resources\/libPng.dylib\n  \/opt\/local\/lib\/libpng.dylib<\/p>\n  \n  <p>sudo ln -sf\n  \/System\/Library\/Frameworks\/ApplicationServices.framework\/Versions\/A\/Frameworks\/ImageIO.framework\/Versions\/A\/Resources\/libTIFF.dylib\n  \/opt\/local\/lib\/libtiff.dylib<\/p>\n  \n  <p>sudo ln -sf\n  \/System\/Library\/Frameworks\/ApplicationServices.framework\/Versions\/A\/Frameworks\/ImageIO.framework\/Versions\/A\/Resources\/libJPEG.dylib\n  \/opt\/local\/lib\/libjpeg.dylib<\/p>\n  \n  <p>Why does this work? Because for whatever reason, when Apple requests\n  libJPEG.dyld, whatever system searches for it finds\n  \/opt\/local\/lib\/libjpeg.dyld because the search is case insensitive.\n  For a Unix operating system. Go figure. The lines above will force any\n  program looking for libjpeg.dyld to be redirected to Apple's\n  libJPEG.dyld.<\/p>\n<\/blockquote>\n\n<p>Another user suggested <a href=\"http:\/\/comments.gmane.org\/gmane.os.apple.macports.user\/22154\" rel=\"nofollow\">setting DYLD_FALLBACK_LIBRARY_PATH<\/a>:<\/p>\n\n<blockquote>\n  <p>You could also use DYLD_FALLBACK_LIBRARY_PATH. I had to use that to\n  avoid Apple's libJPEG getting in the way of apps that expected the\n  libjpeg port. See the man page for dlopen() for details.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":24468482,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2014-06-28T15:49:52.687","AcceptedAnswerId":24496769.0,"OwnerUserId":3054726.0,"Title":"Defining a gradient with respect to a subtensor in Theano","Body":"<p>I have what is conceptually a simple question about Theano but I haven't been able to find the answer (I'll confess upfront to not really understanding how shared variables work in Theano, despite many hours with the tutorials).<\/p>\n\n<p>I'm trying to implement a \"deconvolutional network\"; specifically I have a 3-tensor of inputs (each input is a 2D image) and a 4-tensor of codes; for the ith input codes[i] represents a set of codewords which together code for input i.<\/p>\n\n<p>I've been having a lot of trouble figuring out how to do gradient descent on the codewords.  Here are the relevant parts of my code:<\/p>\n\n<pre><\/pre>\n\n<p>(here codes and dicts are shared tensor variables).  Theano is unhappy with this, specifically with defining<\/p>\n\n<pre><\/pre>\n\n<p>The error message I'm getting is:  <em>theano.gradient.DisconnectedInputError: grad method was asked to compute the gradient with respect to a variable that is not part of the computational graph of the cost, or is used only by a non-differentiable operator: Subtensor{int64}.0<\/em><\/p>\n\n<p>I'm guessing that it wants a symbolic variable instead of codes[idx]; but then I'm not sure how to get everything connected to get the intended effect.  I'm guessing I'll need to change the final line to something like<\/p>\n\n<pre><\/pre>\n\n<p>Can someone give me some pointers as to how to define this function properly?  I think I'm probably missing something basic about working with Theano but I'm not sure what.  <\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>-Justin<\/p>\n\n<p>Update:  Kyle's suggestion worked very nicely.  Here's the specific code I used<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"24496769","CreationDate":"2014-06-30T18:40:12.860","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>To summarize the findings:<\/p>\n\n<p>Assigning <code>grad_var = codes[idx]<\/code>, then making a new variable such as: \n<code>subgrad = T.set_subtensor(codes[input_index], codes[input_index] - learning_rate*del_codes[input_index])<\/code><\/p>\n\n<p>Then calling \n<code>train_codes = function([input_index], loss, updates = [[codes, subgrad]])<\/code><\/p>\n\n<p>seemed to do the trick. In general, I try to make variables for as many things as possible. Sometimes tricky problems can arise from trying to do too much in a single statement, plus it is hard to debug and understand later! Also, in this case I think theano needs a shared variable, but has issues if the shared variable is <em>created<\/em> inside the function that requires it. <\/p>\n\n<p>Glad this worked for you!<\/p>\n"}]}
{"QuestionId":24505682,"AnswerCount":1,"Tags":"<python><python-2.7><spyder><theano>","CreationDate":"2014-07-01T08:28:16.847","AcceptedAnswerId":24738726.0,"OwnerUserId":3592346.0,"Title":"spyder: python: theano: How to disable warnings in spyder?","Body":"<p>I running machine learning algo's with theano.  I have been getting lot of warnings of DeprecationWarning.  coming from numpy package. I want to disable this warnings pls suggest option. \nwarning nature:fromnumeric.py:932: DeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future<\/p>\n\n<p>I tried following  run configure added command line option  -W ignore  or -W ignore::DeprecationWarning but none of this is working<\/p>\n\n<p>alternatively fixing the warning solution is fine for me. looks like its fixed in theano <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/Hf7soRrnh8w\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/Hf7soRrnh8w<\/a> but I dont know where to find this updated version of theano<\/p>\n\n<p>I am using Anaconda distribution 2.0.1 windows 8.1 - 64 bit<\/p>\n\n<p>thanks<\/p>\n","answers":[{"AnswerId":"24738726","CreationDate":"2014-07-14T14:20:27.470","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>See this link to update Theano to the development version. I should be fixed:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>In resume, run one of those 2 command:<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>or (if you want to install it for the current user only):<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git --user\n<\/code><\/pre>\n"}]}
{"QuestionId":24524621,"AnswerCount":1,"Tags":"<python><pip><theano>","CreationDate":"2014-07-02T06:51:19.163","AcceptedAnswerId":null,"OwnerUserId":1358855.0,"Title":"use package from pip local instead pip global","Body":"<p>I have Theano library installed on <\/p>\n\n<pre><\/pre>\n\n<p>but the Theano installed is the old one and I am using some library that can't import some packages.<\/p>\n\n<p>So I tried to install the new one using <\/p>\n\n<pre><\/pre>\n\n<p>but everytime I import theano, the version is the old one which come from <\/p>\n\n<pre><\/pre>\n\n<p>So I need to know how to make import theano load my  theano, not the  theano.<\/p>\n\n<p>Thank you :)<\/p>\n","answers":[{"AnswerId":"24709025","CreationDate":"2014-07-12T02:26:05.297","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is that the old version wasn't installed with pip, but probably with easy_install. This cause many type of problems.<\/p>\n\n<p>You can fix it by changing the import order after starting python. To do so, in your python script before importing theano do something like this:<\/p>\n\n<pre><code>import sys\nsys.path[0:0] = [\"THE_PYTHON_PATH_YOU_WANT_TO_ADD\"]\n<\/code><\/pre>\n\n<p>THE_PYTHON_PATH_YOU_WANT_TO_ADD is something like <code>~\/.local\/lib\/python2.7\/site-packages\/<\/code><\/p>\n"}]}
{"QuestionId":24540358,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2014-07-02T20:21:45.677","AcceptedAnswerId":null,"OwnerUserId":1460123.0,"Title":"Pylearn2 Tutorial Import Error","Body":"<p>While running python make_dataset in the quick start example for Pylearn2, I've run across an import error in a Theano  file. The heart of the issue seems to be this: . I'm running development versions of numpy, scipy, Theano, and Pylearn2.  Any ideas?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"24540875","CreationDate":"2014-07-02T20:55:42.133","ParentId":null,"OwnerUserId":"1460123","Title":null,"Body":"<p>The problem was resolved after removing ATLAS, installing OpenBLAS, reinstalling numpy + scipy + theano, and then installing Pylearn2 from source. <\/p>\n\n<p>See <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/doc\/install_ubuntu.txt#L70\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/doc\/install_ubuntu.txt#L70<\/a> of the Theano install guide for a more detailed explanation.<\/p>\n"}]}
{"QuestionId":24549256,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-07-03T08:56:45.340","AcceptedAnswerId":24626805.0,"OwnerUserId":3279225.0,"Title":"How to access theano.tensor.var.TensorVariable?","Body":"<p>Let's say I have a matrix w of size(1152, 10) like this:<\/p>\n\n<pre><\/pre>\n\n<p>and I have a input of size(1152, 1) like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I wanna calculate the dot multilication of them like this:<\/p>\n\n<pre><\/pre>\n\n<p>and it gave me:<\/p>\n\n<pre><\/pre>\n\n<p>Does theano.tensor.dot return a symbolic expression instead of a value?<\/p>\n","answers":[{"AnswerId":"24626805","CreationDate":"2014-07-08T08:15:09.180","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>In one word: yes.<\/p>\n\n<p>To see the result of the operation, use <code>result.eval()<\/code><\/p>\n\n<p>Minimal working example:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nfrom theano import tensor as T\nw_values = np.random.randn(1152, 10).astype(theano.config.floatX)\ninput_values = np.random.randn(1152, 1).astype(theano.config.floatX)\nw = theano.shared(w_values, 'w')\ninput = theano.shared(input_values, 'input')\nresult = T.dot(input.T, w)\nprint(result.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":24555984,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-07-03T14:08:37.663","AcceptedAnswerId":24609262.0,"OwnerUserId":1099534.0,"Title":"Invert Theano tensor element-wise","Body":"<p>My goal is to invert the values in a Theano tensor element-wise. For instance, I want to turn  into . If there is a zero element, I want to keep it unchanged (e.g.  should be turned into ). What is the most elegant way of doing this?<\/p>\n\n<p>My solution was to add a very small value to all the elements:<\/p>\n\n<pre><\/pre>\n\n<p>I am aware that I could use a  to iterate over the elements and check individually whether they can be inverted, but I want my code to look as mathsy as possible.<\/p>\n","answers":[{"AnswerId":"24609262","CreationDate":"2014-07-07T11:12:53.310","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>You can get this using <code>T.sgn()<\/code> since the sign of 0 is 0, though it will probably still fail for very, very small values <em>near<\/em> 0<\/p>\n\n<p>Minimal example<\/p>\n\n<pre><code>import numpy as np\nimport theano\nfrom theano import tensor as T\n\nX_values = np.arange(10).astype(theano.config.floatX)\nX = T.shared(X_values, 'X')\nY = (X ** (-1 * abs(T.sgn(X)))) * abs(T.sgn(X))\nprint(Y.eval())\n<\/code><\/pre>\n\n<p>I think what user189 was saying is that mathematically this is not really correct - the result <em>should<\/em> go to infinity as the denominator approaches 0. What is the application where you need this? You can always test for T.isinf and T.isnan to try and catch numerical problems without resorting to tricks, and raising an exception is probably a cleaner approach than altering the math.<\/p>\n"}]}
{"QuestionId":24577630,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2014-07-04T15:42:32.153","AcceptedAnswerId":24605866.0,"OwnerUserId":1099534.0,"Title":"How can a tensor be flipped in Theano?","Body":"<p>Given a tensor , how can I flip it?\nFor instance,  flipped is .<\/p>\n","answers":[{"AnswerId":"24605866","CreationDate":"2014-07-07T08:13:57.537","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>You can simply do <code>v[::-1].eval()<\/code>, or just <code>v[::-1]<\/code> in the middle of your computational graph.<\/p>\n\n<p>Minimal example:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nfrom theano import tensor as T\n\nX_values = np.arange(10).astype(theano.config.floatX)\nX = T.shared(X_values, 'X')\nprint(X.eval())\nprint(X[::-1].eval())\n<\/code><\/pre>\n\n<p>See the section on indexing <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">here<\/a> for more details.<\/p>\n"},{"AnswerId":"26808859","CreationDate":"2014-11-07T19:41:57.103","ParentId":null,"OwnerUserId":"1245262","Title":null,"Body":"<p>OK, I know I'm late to the party here, but I've just started playing with Theano, and thought I'd throw in this variation, since I don't think shared values were necessary:<\/p>\n\n<pre><code> from theano import tensor as T\n from theano import function as Tfunc\n\n z = T.vector()\n f = Tfunc([z],z[::-1])\n<\/code><\/pre>\n\n<p>This gives:<\/p>\n\n<pre><code> &gt;&gt;&gt; f([1,3,5,7,9])\n array([ 9.,  7.,  5.,  3.,  1.])\n<\/code><\/pre>\n"}]}
{"QuestionId":24606346,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><autoencoder>","CreationDate":"2014-07-07T08:40:34.263","AcceptedAnswerId":null,"OwnerUserId":3568055.0,"Title":"Theano implementation of Stacked DenoisingAutoencoders - Why same input to dA layers?","Body":"<p>In the tutorial Stacked DenoisingAutoencoders on <a href=\"http:\/\/deeplearning.net\/tutorial\/SdA.html#sda\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/SdA.html#sda<\/a>, the pretraining_functions return a list of functions which represent the train function of each dA layer. But I don't understand why it gives all the dA layers the same input (). Actually, the input of each dA layer should be the output of the layer below except the first dA layer. Can anybody tell me why these codes are correct?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"24810827","CreationDate":"2014-07-17T18:21:17.970","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Since the inputs of each hidden layer are configured as the outputs of the previous layer:<\/p>\n\n<pre><code># the input to this layer is either the activation of the hidden\n# layer below or the input of the SdA if you are on the first\n# layer\nif i == 0:\n    layer_input = self.x\nelse:\n    layer_input = self.sigmoid_layers[-1].output\n<\/code><\/pre>\n\n<p>When setting <code>self.x<\/code> to <code>train_set_x[batch_begin:batch_end]<\/code> in <code>givens<\/code> section of the pretraining function it actually makes theano propagate the inputs from one layer to the other, so when you pre-train the second layer the inputs will first propagate through the first layer and then be processed by the second. <\/p>\n\n<p>If you look closely at the end of the <a href=\"http:\/\/deeplearning.net\/tutorial\/SdA.html#tips-and-tricks\" rel=\"nofollow\">tutorial<\/a> there is a tip how to reduce the training run-time by precomputing the explicit inputs per layer.<\/p>\n"}]}
{"QuestionId":24654389,"AnswerCount":1,"Tags":"<neural-network><pooling><theano>","CreationDate":"2014-07-09T12:58:26.130","AcceptedAnswerId":null,"OwnerUserId":2166433.0,"Title":"Average pooling with Theano","Body":"<p>I am trying to implement another pooling function for neural network with Theano, expect of already existing maxpool, for example average pool.<\/p>\n\n<p>Using to <a href=\"https:\/\/github.com\/quaizarv\/deeplearning-benchmark\/blob\/master\/common\/theanoWrappers.py\" rel=\"nofollow\">this source<\/a>, where average pooling is already implemented, my code looks like:<\/p>\n\n<p>Random initialization just to test:<\/p>\n\n<pre><\/pre>\n\n<p>Definition of Theano scalars and functions:<\/p>\n\n<pre><\/pre>\n\n<p>TSN is <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/neighbours.html\" rel=\"nofollow\">theano.sandbox.neighbours<\/a><\/p>\n\n<p>And the call of the function:<\/p>\n\n<pre><\/pre>\n\n<p>And I am getting an error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't really understand this error. Would be glad to have any suggestions how to correct this error or example of other pooling techniques, programmed in Theano.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Edit: with the ignoring the border, it works perfectly<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"24865492","CreationDate":"2014-07-21T12:47:39.920","ParentId":null,"OwnerUserId":"216880","Title":null,"Body":"<p><code>invals<\/code> has shape <code>(5, 5)<\/code> in the last two dimensions, however you want to pool over <code>(2, 2)<\/code> subsets. This only works if you ignore the border (i.e. the last column and the last row of <code>invals<\/code>).<\/p>\n"}]}
{"QuestionId":24665056,"AnswerCount":5,"Tags":"<python><cuda><gpu><nvidia><theano>","CreationDate":"2014-07-09T23:06:20.643","AcceptedAnswerId":null,"OwnerUserId":3822367.0,"Title":"Using Theano with GPU on Ubuntu 14.04 on AWS g2","Body":"<p>I'm having trouble getting Theano to use the GPU on my machine.  <\/p>\n\n<p>When I run:\n\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/misc$ THEANO_FLAGS=floatX=float32,device=gpu python check_blas.py\nWARNING (theano.sandbox.cuda): CUDA is installed, but device gpu is not available  (error: Unable to get the number of gpus available: no CUDA-capable device is detected)<\/p>\n\n<p>I've also checked that the NVIDIA driver is installed with: lspci -vnn | grep -i VGA -A 12<\/p>\n\n<p>with result: Kernel driver in use: nvidia<\/p>\n\n<p>However, when I run: nvidia-smi\nresult: NVIDIA: could not open the device file \/dev\/nvidiactl (No such file or directory).\nNVIDIA-SMI has failed because it couldn't communicate with NVIDIA driver. Make sure that latest NVIDIA driver is installed and running.<\/p>\n\n<p>and \/dev\/nvidiaactl doesn't exist. What's going on?<\/p>\n\n<p>UPDATE: \/nvidia-smi works with result:<\/p>\n\n<pre><\/pre>\n\n<p>and after compiling the NVIDIA_CUDA-6.0_Samples then running deviceQuery I get result:<\/p>\n\n<p>cudaGetDeviceCount returned 35\n-> CUDA driver version is insufficient for CUDA runtime version\nResult = FAIL<\/p>\n","answers":[{"AnswerId":"29877206","CreationDate":"2015-04-26T11:47:34.713","ParentId":null,"OwnerUserId":"2930037","Title":null,"Body":"<p>Had the same problem and reinstalled Cuda and at the end it says i have to update PATH to include \/usr\/local\/cuda7.0\/bin and LD_LIBRARY_PATH to include \/usr\/local\/cuda7.0\/lib64. The PATH (add LD_LIBRARY_PATH in same file) can be found in \/etc\/environment. Then theano found gpu. Basic error on my part...<\/p>\n"},{"AnswerId":"49819761","CreationDate":"2018-04-13T14:52:26.577","ParentId":null,"OwnerUserId":"1052899","Title":null,"Body":"<p>I got <\/p>\n\n<pre><code>-&gt; CUDA driver version is insufficient for CUDA runtime version\n<\/code><\/pre>\n\n<p>and my problem is related with the selected GPU mode.\nIn other words, the problem may be related to the selected GPU mode (Performance\/Power Saving Mode), when you select (with nvidia-settings utility, in the \"PRIME Profiles\" configurations) the integrated Intel GPU and you execute the <code>deviceQuery<\/code> script... you get this error:<\/p>\n\n<p>But this error is misleading,\nby <strong>selecting<\/strong> back the <strong>NVIDIA(Performance mode)<\/strong> with <strong>nvidia-settings<\/strong> utility the  problem disappears.<\/p>\n\n<p><strong>This is not a version problem<\/strong>. <\/p>\n\n<p>Regards<\/p>\n\n<p>P.s: The selection is available when Prime-related-stuff is installed. Further details: <a href=\"https:\/\/askubuntu.com\/questions\/858030\/nvidia-prime-in-nvidia-x-server-settings-in-16-04-1\">https:\/\/askubuntu.com\/questions\/858030\/nvidia-prime-in-nvidia-x-server-settings-in-16-04-1<\/a> <\/p>\n"},{"AnswerId":"38781071","CreationDate":"2016-08-05T04:22:02.097","ParentId":null,"OwnerUserId":"1787555","Title":null,"Body":"<p>If you are using CUDA 7.5, make sure follow official instruction:\nCUDA 7.5 doesn't support the default g++ version. Install an supported version and make it the default.<\/p>\n\n<pre><code>sudo apt-get install g++-4.9\n\nsudo update-alternatives --install \/usr\/bin\/gcc gcc \/usr\/bin\/gcc-4.9 20\nsudo update-alternatives --install \/usr\/bin\/gcc gcc \/usr\/bin\/gcc-5 10\n\nsudo update-alternatives --install \/usr\/bin\/g++ g++ \/usr\/bin\/g++-4.9 20\nsudo update-alternatives --install \/usr\/bin\/g++ g++ \/usr\/bin\/g++-5 10\n\nsudo update-alternatives --install \/usr\/bin\/cc cc \/usr\/bin\/gcc 30\nsudo update-alternatives --set cc \/usr\/bin\/gcc\n\nsudo update-alternatives --install \/usr\/bin\/c++ c++ \/usr\/bin\/g++ 30\nsudo update-alternatives --set c++ \/usr\/bin\/g++\n<\/code><\/pre>\n\n<p>If theano GPU test code has error:<\/p>\n\n<blockquote>\n  <p>ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu:\n  libcublas.so.7.5: cannot open shared object file: No such file or\n  directory  WARNING (theano.sandbox.cuda): CUDA is installed, but\n  device gpu is not available (error: cuda unavilable)<\/p>\n<\/blockquote>\n\n<p>Just using <code>ldconfig<\/code> command to link the shared object of cuda 7.5:<\/p>\n\n<pre><code>sudo ldconfig \/usr\/local\/cuda-7.5\/lib64\n<\/code><\/pre>\n"},{"AnswerId":"24665354","CreationDate":"2014-07-09T23:38:24.583","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>CUDA GPUs in a linux system are not usable until certain \"device files\" have been properly established.<\/p>\n\n<p>There is a note to this effect in <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-getting-started-guide-for-linux\/index.html#runfile-verifications\" rel=\"nofollow\">the documentation<\/a>.<\/p>\n\n<p>In general there are several ways these device files can be established:<\/p>\n\n<ol>\n<li>If an X-server is running.<\/li>\n<li>If a GPU activity is initiated as root  user (such as running nvidia-smi, or any CUDA app.)<\/li>\n<li>Via startup scripts (refer to the documentation linked above for an example).<\/li>\n<\/ol>\n\n<p>If none of these steps are taken, the GPUs will not be functional for non-root users.  Note that the files do not persist through re-boots, and must be re-established on each boot cycle, through one of the 3 above methods.  If you use method 2, and reboot, the GPUs will not be available until you use method 2 again.<\/p>\n\n<p>I suggest reading the linux getting started guide entirely (linked above), if you are having trouble setting up a linux system for CUDA GPU usage.<\/p>\n"},{"AnswerId":"24913032","CreationDate":"2014-07-23T14:19:43.733","ParentId":null,"OwnerUserId":"2304604","Title":null,"Body":"<p>I've wasted a lot of hours trying to get AWS G2 to work on ubuntu but failed by getting exact error like you did. Currently I'm running Theano with gpu smoothly with <a href=\"https:\/\/aws.amazon.com\/marketplace\/pp\/B00FYCDDTE\" rel=\"nofollow\">this<\/a> redhat AMI. To install Theano on Redhat follow the process of <em>Installing Theano in CentOS<\/em> in Theano documentation.<\/p>\n"}]}
{"QuestionId":24697064,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2014-07-11T11:52:35.777","AcceptedAnswerId":null,"OwnerUserId":499439.0,"Title":"How to get predictions vector from Theano stacked autoencoder","Body":"<p>I'm trying to modify Stacked Autoencoder for classification from Theano <a href=\"http:\/\/deeplearning.net\/tutorial\/deeplearning.pdf\" rel=\"nofollow\">deep learning tutorial<\/a>, chapter 8. The code of autoencoder I'm dealing with is available <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/blob\/master\/code\/SdA.py\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>My dataset consists of 4 arrays: test_set_x, test_set_y, valid_set_x, valid_set_y. The names are self-explained.<\/p>\n\n<p>This is how the trained autoencoder is checked on validation set:<\/p>\n\n<pre><\/pre>\n\n<p>This code prints out \"0.87\" on my dataset, so it does work.<\/p>\n\n<p>Expressing it more verbose<\/p>\n\n<pre><\/pre>\n\n<p>still gives correct answer 87%.<\/p>\n\n<p>But whenever I'm trying to get directly the real class prediction vector, I get some very wrong result: all elements of result vector are equal to 4 (one on my classes).<\/p>\n\n<p>My try looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>This prints out \"[4, 4, 4, ....., 4, 4]\". Comparing this result with valid_set_y vector gives about 12% correctness, not even nearly 87%.<\/p>\n\n<p>I don't understand what I'm doing wrong.<\/p>\n\n<p>Please help me if you've ever had a deal with theano autoencoders and\/or the mentioned tutorial.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"24722883","CreationDate":"2014-07-13T12:50:41.893","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The <code>valid_score<\/code> output is the <strong>error<\/strong> rate on the validation set. A validation score of <code>87%<\/code> means that you managed to classify correctly only ~12% of your validation examples. This result seems to be consistet with an \"all 4\" prediction rule.<\/p>\n"}]}
{"QuestionId":24728255,"AnswerCount":3,"Tags":"<python><macos><theano><deep-learning>","CreationDate":"2014-07-14T00:51:42.317","AcceptedAnswerId":null,"OwnerUserId":3835247.0,"Title":"How to install theano library on OS X?","Body":"<p>In my machine learning course we are going to start using <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#install\" rel=\"nofollow\">theano<\/a>, a very well known library for deep learning architectures. I all ready installed it with the following command:<\/p>\n\n<pre><\/pre>\n\n<p>By the way, when i want to test if it installed correctly, the python interpreter can\u00b4t fin the module. I dont know if im installing it right. Reading the documentation i found <a href=\"http:\/\/continuum.io\/downloads\" rel=\"nofollow\">anaconda<\/a>, is it right to first install anaconda and then try to install again with  theano?. Is this the right way to install this library on MAC OS X?. How can i install this library correctly in order to use theano succesfully?<\/p>\n","answers":[{"AnswerId":"32152584","CreationDate":"2015-08-22T05:15:37.183","ParentId":null,"OwnerUserId":"4127806","Title":null,"Body":"<p>Consider installing Theano in a virtual enviroment as oppose to installing it globally via sudo. The significance of this step is nicely described <a href=\"http:\/\/www.dabapps.com\/blog\/introduction-to-pip-and-virtualenv-python\/\" rel=\"nofollow\">here<\/a>. \nIn your terminal window, do the following:<\/p>\n\n<pre><code>virtualenv --system-site-packages -p python2.7 theano-env\nsource theano-env\/bin\/activate\npip install -r https:\/\/raw.githubusercontent.com\/Lasagne\/Lasagne\/v0.1\/requirements.txt\n<\/code><\/pre>\n"},{"AnswerId":"44283294","CreationDate":"2017-05-31T11:14:29.553","ParentId":null,"OwnerUserId":"3232179","Title":null,"Body":"<p>Anaconda is indeed highly recommended for python libraries. In fact, Anaconda is not only for python in contrast to pip. You can read more here: <a href=\"https:\/\/stackoverflow.com\/questions\/20994716\/what-is-the-difference-between-pip-and-conda\">What is the difference between pip and conda?<\/a><\/p>\n\n<p>For installing Theano, I had already installed Anaconda. I just simply did:<\/p>\n\n<p><code>conda install theano<\/code><\/p>\n\n<p>Then, in Ipython, I successfully imported theano:<\/p>\n\n<p><code>import theano<\/code><\/p>\n"},{"AnswerId":"28209171","CreationDate":"2015-01-29T07:30:07.693","ParentId":null,"OwnerUserId":"666043","Title":null,"Body":"<p>Installing Python with Homebrew and installing Theano with pip after that worked fine for me. I just needed to install nose after that to be able to run the tests.<\/p>\n\n<pre><code>brew install python\npip install Theano\npip install nose\n<\/code><\/pre>\n\n<p>I cannot help much installing the framework with Anaconda though.<\/p>\n"}]}
{"QuestionId":24744701,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><gpu><theano>","CreationDate":"2014-07-14T19:49:53.977","AcceptedAnswerId":24765554.0,"OwnerUserId":1460123.0,"Title":"Convert CudaNdarraySharedVariable to TensorVariable","Body":"<p>I'm trying to convert a  GPU model to a CPU compatible version for prediction on a remote server -- how can I convert 's to 's to avoid an error calling cuda code on a GPU-less machine?  The experimental theano flag  seems to have left a few 's hanging around (specifically ).<\/p>\n","answers":[{"AnswerId":"24765554","CreationDate":"2014-07-15T18:29:32.457","ParentId":null,"OwnerUserId":"2616754","Title":null,"Body":"<p>For a plain CudaNdarray variable, something like this should work:<\/p>\n\n<p>'''x = CudaNdarray...    x_new=theano.tensor.TensorVariable(CudaNdarrayType([False] * tensor_dim))<br>\nf = theano.function([x_new], x_new)<\/p>\n\n<p>converted_x = f(x)\n'''<\/p>\n"}]}
{"QuestionId":24752655,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><unsupervised-learning>","CreationDate":"2014-07-15T07:47:24.470","AcceptedAnswerId":26047542.0,"OwnerUserId":1714410.0,"Title":"Unsupervised pre-training for convolutional neural network in theano","Body":"<p>I would like to design a deep net with one (or more) convolutional layers (CNN) and one or more fully connected hidden layers on top.<br>\nFor deep network with fully connected layers there are methods in theano for unsupervised pre-training, e.g., using <a href=\"http:\/\/www.deeplearning.net\/tutorial\/SdA.html\">denoising auto-encoders<\/a> or <a href=\"http:\/\/www.deeplearning.net\/tutorial\/rbm.html\">RBMs<\/a>.<\/p>\n\n<p>My question is: How can I implement (in theano) an unsupervised pre-training stage for convolutional layers?<\/p>\n\n<p>I do not expect a full implementation as an answer, but I would appreciate a link to a good tutorial or a reliable reference.<\/p>\n","answers":[{"AnswerId":"26047542","CreationDate":"2014-09-25T20:29:26.867","ParentId":null,"OwnerUserId":"1936499","Title":null,"Body":"<p><a href=\"http:\/\/people.idsia.ch\/~ciresan\/data\/icann2011.pdf\" rel=\"noreferrer\">This paper<\/a> describes an approach for building a stacked convolutional autoencoder.  Based on that paper and some Google searches I was able to implement the described network.  Basically, everything you need is described in the Theano convolutional network and denoising autoencoder tutorials with one crucial exception: how to reverse the max-pooling step in the convolutional network.  I was able to work that out using a method from <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/7t7_hxbAMdA\/V1tp7YZ50PYJ\" rel=\"noreferrer\">this discussion<\/a> - the trickiest part is figuring out the right dimensions for W_prime as these will depend on the feed forward filter sizes and the pooling ratio.  Here is my inverting function:<\/p>\n\n<pre><code>    def get_reconstructed_input(self, hidden):\n        \"\"\" Computes the reconstructed input given the values of the hidden layer \"\"\"\n        repeated_conv = conv.conv2d(input = hidden, filters = self.W_prime, border_mode='full')\n\n        multiple_conv_out = [repeated_conv.flatten()] * np.prod(self.poolsize)\n\n        stacked_conv_neibs = T.stack(*multiple_conv_out).T\n\n        stretch_unpooling_out = neibs2images(stacked_conv_neibs, self.pl, self.x.shape)\n\n        rectified_linear_activation = lambda x: T.maximum(0.0, x)\n        return rectified_linear_activation(stretch_unpooling_out + self.b_prime.dimshuffle('x', 0, 'x', 'x'))\n<\/code><\/pre>\n"}]}
{"QuestionId":24763371,"AnswerCount":1,"Tags":"<python><numpy><scipy><gpu><theano>","CreationDate":"2014-07-15T16:24:03.990","AcceptedAnswerId":null,"OwnerUserId":1460123.0,"Title":"Theano OSError on function declaration","Body":"<p>On declaration of a Theano symbolic function, I get an OSError and traceback.  Interestingly enough, the same code functions on a different machine.  One machine is configured to use the GPU, while the other (with the error) is CPU-only. Has anyone else experienced this sort of behavior and have a clue how to proceed? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"24765088","CreationDate":"2014-07-15T18:02:24.630","ParentId":null,"OwnerUserId":"1460123","Title":null,"Body":"<p>I'm not sure why it was triggered on declaration of Theano's symbolic function, but it was a simple memory issue.  Cutting down on memory usage solved the problem for me.<\/p>\n"}]}
{"QuestionId":24804516,"AnswerCount":1,"Tags":"<python><numpy><theano><deep-learning>","CreationDate":"2014-07-17T13:09:29.357","AcceptedAnswerId":null,"OwnerUserId":2688733.0,"Title":"Error while calculating dot product in Theano","Body":"<p>I have the following simple code written in Theano and I am getting error while compiling function f:<\/p>\n\n<pre><\/pre>\n\n<p>What is going wrong on my side?<\/p>\n","answers":[{"AnswerId":"24805730","CreationDate":"2014-07-17T14:02:55.163","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>From the error message you get it seems like theano cannot find the mkl library file libmkl_intel_lp64.dylib.<\/p>\n\n<p>How do you configure your theano? especially, the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.config.blas.ldflags\" rel=\"nofollow\">blas library<\/a>?<\/p>\n\n<p>Please check <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#troubleshooting-make-sure-you-have-a-blas-library\" rel=\"nofollow\">this section<\/a> of theano installation manual.<\/p>\n"}]}
{"QuestionId":24892171,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-07-22T15:55:33.640","AcceptedAnswerId":null,"OwnerUserId":1099534.0,"Title":"Printing whole matrix in Theano","Body":"<p>I'm debugging my Theano code and printing the values of my tensors as advised <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html\" rel=\"nofollow\">here<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>The issue is that, when  is a relatively large matrix, the value is truncated to the first couple of rows and the last couple of rows. However, I would like the whole matrix to be printed. Is this possible?<\/p>\n","answers":[{"AnswerId":"24893256","CreationDate":"2014-07-22T16:52:10.763","ParentId":null,"OwnerUserId":"3821154","Title":null,"Body":"<p>I believe you can print the underlying numpy, accessed as <code>a.get_value()<\/code>. Within numpy you can modify printing by <\/p>\n\n<pre><code>numpy.set_printoptions(threshold=10000000)\n<\/code><\/pre>\n\n<p>where threshold should be bigger than the number of elements expected, and then the whole array will show. See the documentation for <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.set_printoptions.html\" rel=\"nofollow\">set_printoptions<\/a>. Note that if outputted to a console, this may freeze up because of the possibly very large amount of text.<\/p>\n"}]}
{"QuestionId":24917916,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2014-07-23T18:14:29.777","AcceptedAnswerId":24919518.0,"OwnerUserId":1099534.0,"Title":"Thresholding in Theano","Body":"<p>Is there a way to threshold the values in a Theano tensor? For instance, if , I would like to create another tensor  which contains the same values as , except that the ones that exceed a certain threshold  are replaced by  itself:<\/p>\n\n<pre><\/pre>\n\n<p>More generally, what is there a standard framework to create your own operations on tensors?<\/p>\n","answers":[{"AnswerId":"24919518","CreationDate":"2014-07-23T19:42:06.163","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Here is code that do that. Use the clip function.<\/p>\n\n<pre><code>import theano\nv = theano.tensor.vector()\nf = theano.function([v], theano.tensor.clip(v, 0, 100))\nf([1, 2, 3, 100, 200, 300])\n# array([   1.,    2.,    3.,  100.,  100.,  100.])\n<\/code><\/pre>\n\n<p>If you you don't want a min you can use a switch:<\/p>\n\n<pre><code>import theano\nv = theano.tensor.vector()\nf = theano.function([v], theano.tensor.clip(v, 0, 100))\nf([1, 2, 3, 100, 200, 300])\n# array([   1.,    2.,    3.,  100.,  100.,  100.])\nf = theano.function([v], theano.tensor.switch(v&lt;100, v, 100))\nf([1, 2, 3, 100, 200, 300])\n# array([   1.,    2.,    3.,  100.,  100.,  100.])\n<\/code><\/pre>\n"}]}
{"QuestionId":25003733,"AnswerCount":2,"Tags":"<python><cuda><gpu><nvcc><theano>","CreationDate":"2014-07-28T20:26:44.753","AcceptedAnswerId":null,"OwnerUserId":562769.0,"Title":"Why does Theano print \"cc1plus: fatal error: cuda_runtime.h: No such file or directory\"?","Body":"<p>I am trying to use the GPU with Theano. I've read <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#gpu-linux\" rel=\"nofollow\">this tutorial<\/a>.<\/p>\n\n<p>However, I can't get theano to use the GPU and I don't know how to continue.<\/p>\n\n<h2>Testing machine<\/h2>\n\n<pre><\/pre>\n\n<h2>Cuda sample<\/h2>\n\n<p>Compiling and executing worked as a super user (tested with ):<\/p>\n\n<pre><\/pre>\n\n<p>When I try this as normal user, I get:<\/p>\n\n<pre><\/pre>\n\n<p>How can I get cuda to work with non-super users?<\/p>\n\n<h2>Testing code<\/h2>\n\n<p>The following code is from \"<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#testing-theano-with-gpu\" rel=\"nofollow\">Testing Theano with GPU<\/a>\"<\/p>\n\n<pre><\/pre>\n\n<h2>The error message<\/h2>\n\n<p>The complete error message is much too long to post it here. A longer version is on <a href=\"http:\/\/pastebin.com\/eT9vbk7M\" rel=\"nofollow\">http:\/\/pastebin.com\/eT9vbk7M<\/a>, but I think the relevant part is:<\/p>\n\n<pre><\/pre>\n\n<p>The standard stream gives:<\/p>\n\n<pre><\/pre>\n\n<h2>theano.rc<\/h2>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"25011528","CreationDate":"2014-07-29T08:55:00.370","ParentId":null,"OwnerUserId":"571812","Title":null,"Body":"<p>Try exporting C_INCLUDE_PATH to cuda toolkit include files on your system, something like:<\/p>\n\n<pre><code>export C_INCLUDE_PATH=${C_INCLUDE_PATH}:\/usr\/local\/cuda\/include\n<\/code><\/pre>\n"},{"AnswerId":"25049424","CreationDate":"2014-07-31T01:08:46.523","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>As some comments told, the problem is the permissio of \/dev\/nvidia*. As some told this mean during your startup, it don't get initialized correctly. Normally, this is done correctly when the GUI is started. My guess is that you didn't enable it or install it. So you probably have an headless server.<\/p>\n\n<p>To fix this, just run as root <code>nvidia-smi<\/code>. This will detect that it isn't started correctly and will fix it. root have the permission to fix things. Normal user don't have the permission to fix this. That is why it work with root (it get automatically fixed), but not as normal user.<\/p>\n\n<p>This fix need to be done each time the computer boot. To automatise this, you can create as root this file <code>\/etc\/init.d\/nvidia-gpu-config<\/code> with this content:<\/p>\n\n<pre><code>#!\/bin\/sh\n#\n# nvidia-gpu-config    Start the correct initialization of nvidia GPU driver.\n#\n# chkconfig: - 90 90\n# description:  Init gpu to wanted states\n\n# sudo \/sbin\/chkconfig --add nvidia-smi\n#\n\ncase $1 in\n'start')\nnvidia-smi\n;;\nesac\n<\/code><\/pre>\n\n<p>Then as root run this command: <code>\/sbin\/chkconfig --add nvidia-gpu-config<\/code>.<\/p>\n\n<p>UPDATE: This work for OS that use the init system SysV. If your system use the init system systemd, I don't know if it work.<\/p>\n"}]}
{"QuestionId":25045897,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><conv-neural-network>","CreationDate":"2014-07-30T20:03:26.727","AcceptedAnswerId":25062985.0,"OwnerUserId":3479456.0,"Title":"Adding additional features in Theano (CNN)","Body":"<p>I'm using Theano for classification (convolutional neural networks)<\/p>\n\n<p>Previously, I've been using the pixel values of the (flattened) image as the features of the NN.\nNow, I want to add additional features. <br>I've been told that I can concatenate that vector of additional features to the flattened image features and then use that as input to the fully-connected layer, but I'm having trouble with that.<\/p>\n\n<p>First of all, is that the right approach?<\/p>\n\n<p>Here's some code snippets and my errors:<br>\nSimilar to the provided example from their site with some modifications<br><\/p>\n\n<p>(from the class that builds the model)<\/p>\n\n<pre><\/pre>\n\n<p>Below, variables  and  are defined previously. What's important is :<\/p>\n\n<pre><\/pre>\n\n<p>(from the class that trains)<\/p>\n\n<pre><\/pre>\n\n<p>However, I get an error when the train_model is called:<\/p>\n\n<pre><\/pre>\n\n<p>Do the input shapes represent the shapes of ,  and , respectively?<\/p>\n\n<p>If so, the third seems correct (batchsize=5, 2 extra features), but why is the first a scalar and the second a matrix?<\/p>\n\n<p>More details:<\/p>\n\n<pre><\/pre>\n\n<p>Do I have the right idea or is there a better way of accomplishing this?\nAny insights into why I'm getting an error?<\/p>\n","answers":[{"AnswerId":"25062985","CreationDate":"2014-07-31T15:28:02.613","ParentId":null,"OwnerUserId":"3479456","Title":null,"Body":"<p>Issue was that I was concatenating on the wrong axis.<\/p>\n\n<pre><code>layer2_input = T.concatenate([layer2_input, self.f.flatten(2)])\n<\/code><\/pre>\n\n<p>should have been<\/p>\n\n<pre><code>layer2_input = T.concatenate([layer2_input, self.f.flatten(2)], axis=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":25046108,"AnswerCount":1,"Tags":"<python><machine-learning><gpu><theano>","CreationDate":"2014-07-30T20:16:18.570","AcceptedAnswerId":25084895.0,"OwnerUserId":1460123.0,"Title":"Theano ValueError: Some matrix has no unit stride","Body":"<p>Recently have run into issues getting predictions for a trained pylearn2 model.  The relevant bits of the traceback are provided below.  I have ensured that  matches the shape of the numpy array I'm passing in to the theano prediction function I've generated, but still receive the following error.    <\/p>\n\n<pre><\/pre>\n\n<p>Interestingly enough, behavior seems to be machine-dependent.  I have the prediction script working on my local machine, but execution on a Google Compute Engine instance produces the above error.  <\/p>\n\n<p>Any ideas where I might start debugging?  The input strides look pretty odd, but I'm unsure how to begin debugging that value.<\/p>\n","answers":[{"AnswerId":"25084895","CreationDate":"2014-08-01T16:38:42.150","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is that NumPy create an ndarray with wrong strides. This was fixed in more recent version of NumPy. So update NumPy and it should work.<\/p>\n\n<p>This is the line that show NumPy strides are bad:<\/p>\n\n<pre><code>Inputs strides: [(9223372036854775807, 4), (4, 9223372036854775807)]\n<\/code><\/pre>\n\n<p>Did you compile NumPy with some special flags to test strides for dimensions of size 1?<\/p>\n\n<p>Here I did a PR to be more tolerant to those invalid strides:<\/p>\n\n<pre><code>https:\/\/github.com\/Theano\/Theano\/pull\/2008\n<\/code><\/pre>\n"}]}
{"QuestionId":25057977,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2014-07-31T11:40:05.210","AcceptedAnswerId":25097908.0,"OwnerUserId":54564.0,"Title":"Defining a function with a loop in Theano","Body":"<p>I want to define the following function of two variables in Theano and compute its Jacobian:<\/p>\n\n<pre><\/pre>\n\n<p>How do I make a Theano function for the above expression - and eventually minimize it using its Jacobian?<\/p>\n","answers":[{"AnswerId":"25097908","CreationDate":"2014-08-02T18:00:18.447","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Since your function is scalar, the Jacobian reduces to the gradient. Assuming your two variables <code>x1, x2<\/code> are scalar (looks like it from the formula, easily generalizable to other objects), you can write<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nx1 = T.fscalar('x1')\nx2 = T.fscalar('x2')\n\nk = T.arange(1, 10)\n\nexpr = ((2 + 2 * k - T.exp(x1 * k) - T.exp(x2 * k)) ** 2).sum()\n\nfunc = theano.function([x1, x2], expr)\n<\/code><\/pre>\n\n<p>You can call <code>func<\/code> on two scalars<\/p>\n\n<pre><code>In [1]: func(0.25,0.25)\nOut[1]: array(126.5205307006836, dtype=float32)\n<\/code><\/pre>\n\n<p>The gradient (Jacobian) is then<\/p>\n\n<pre><code>grad_expr = T.grad(cost=expr, wrt=[x1, x2])\n<\/code><\/pre>\n\n<p>And you can use <code>updates<\/code> in <code>theano.function<\/code> in the standard way (see theano tutorials) to make your gradient descent, setting <code>x1, x2<\/code> as shared variables in givens, by hand on the python level, or using <code>scan<\/code> as indicated by others.<\/p>\n"}]}
{"QuestionId":25121749,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2014-08-04T15:04:21.173","AcceptedAnswerId":null,"OwnerUserId":2688733.0,"Title":"Non-unit value on shape on a broadcastable dimension","Body":"<p>I am a theano newbie and I am implementing a simple perceptron based learning rule and I am getting following error, I fail to understand why am I getting this error?<\/p>\n\n<p>Here my code:<\/p>\n\n<pre><\/pre>\n\n<p>ERROR:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"25127898","CreationDate":"2014-08-04T21:32:23.217","ParentId":null,"OwnerUserId":"2688733","Title":null,"Body":"<p>So I found the problem in my code, <\/p>\n\n<p>The \"train\" function expects a row, but the value being passed is a (3, 1) column vector. <\/p>\n\n<p>So here is the corrected version:<\/p>\n\n<pre><code>train = function([z2],\n                 updates=[(w,w-T.transpose(z2))]\n                 )\ncount = 0\nwhile np.abs(np.sum(predictedOut(D[0])-D[1])) &gt; 0 and count &lt; 1000 :\n    print 'on example ',count\n    a1 = D[0][count%N,:].reshape(1,feats+1)\n    b1 = D[1][count%N].reshape(1,1)\n    a2 = prod(a1).reshape(1,1)\n    a3 = w_up(a1,b1,a2)[0].reshape(1,feats+1)\n    train(a3)\n    count += 1\n<\/code><\/pre>\n"}]}
{"QuestionId":25143328,"AnswerCount":2,"Tags":"<python><numerical-methods><theano>","CreationDate":"2014-08-05T15:59:42.247","AcceptedAnswerId":null,"OwnerUserId":1508226.0,"Title":"How to group several output variables together in Theano?","Body":"<p>I am trying to implement a function in Theano that maps a vector to a vector, but each dimension of the output vector is specified by hand. If I create a Theano function like so:<\/p>\n\n<pre><\/pre>\n\n<p>then  gives  as the output, when I'd like it to return . What is the Theanic way of doing this?<\/p>\n","answers":[{"AnswerId":"25190520","CreationDate":"2014-08-07T19:21:38.877","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The other answer by Kyle Kastner will work, but you can have Theano do that for you (I fixed the division by 0 from your example):<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nx = T.dvector('x')\ndx = 28.0 * (x[1] - x[0])\ndy = x[0] * (10.0 - x[1]) - x[2]\ndz = x[0] * x[1] - 8.0\/3.0 * x[2]\no = T.as_tensor_variable([dx,dy,dz])\nf = theano.function([x],o)\nf([1,2,3])\n# output array([ 28.,   5.,  -6.])\n<\/code><\/pre>\n"},{"AnswerId":"25158728","CreationDate":"2014-08-06T11:04:03.477","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>The output of the function is just a list of numpy arrays - you can do <code>np.array(f([1, 2, 3]))<\/code> to convert the output list into a numpy vector.<\/p>\n"}]}
{"QuestionId":25159498,"AnswerCount":1,"Tags":"<python><numpy><neural-network><convolution><theano>","CreationDate":"2014-08-06T11:42:17.727","AcceptedAnswerId":25161521.0,"OwnerUserId":1714410.0,"Title":"Theano conv2d and max_pool_2d","Body":"<p>When implementing a <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">convolutional neural network (CNN) in theno<\/a> one comes across two variants of  operator:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/conv.html#theano.tensor.nnet.conv.conv2d\" rel=\"nofollow\"><\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/conv.html#theano.tensor.signal.conv.conv2d\" rel=\"nofollow\"><\/a><\/li>\n<\/ul>\n\n<p>And an implementation of max-pooling:  <\/p>\n\n<ul>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/downsample.html#theano.tensor.signal.downsample.max_pool_2d\" rel=\"nofollow\"><\/a><\/li>\n<\/ul>\n\n<p>My questions are:  <\/p>\n\n<ol>\n<li>What is the difference between the two implementations of ?  <\/li>\n<li><p>What is the difference between the use of  argument of  and the application of  subsampling after ?<br>\nThat is what is the difference between:<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n","answers":[{"AnswerId":"25161521","CreationDate":"2014-08-06T13:19:09.840","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>In answer to your first question, <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/conv.html\" rel=\"nofollow\">here is the section of the Theano docs that addresses it<\/a>:<\/p>\n\n<blockquote>\n  <p>Two similar implementation exists for conv2d:<\/p>\n\n<pre><code>signal.conv2d and nnet.conv2d.\n<\/code><\/pre>\n  \n  <p>The former implements a traditional 2D convolution, while the latter\n  implements the convolutional layers present in convolutional neural\n  networks (where filters are 3D and pool over several input channels).<\/p>\n<\/blockquote>\n\n<p>Under the hood they both call the same function, so the only difference is the user interface.<\/p>\n\n<p>Regarding your second question, the result is different. The equivalent call to:<\/p>\n\n<pre><code>conv2(..., subsample=(2,2))\n<\/code><\/pre>\n\n<p>would be:<\/p>\n\n<pre><code>conv2d(...,subsample=(1,1))[:,:,::2,::2]\n<\/code><\/pre>\n\n<p>In other words <code>conv2d<\/code> doesn't take the max over the whole pooling region, but rather the element at index <code>[0,0]<\/code> of the pooling region.<\/p>\n"}]}
{"QuestionId":25166657,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2014-08-06T17:36:06.393","AcceptedAnswerId":25205016.0,"OwnerUserId":1120370.0,"Title":"Index gymnastics inside a Theano function","Body":"<p>I am using <a href=\"http:\/\/deeplearning.net\/software\/theano\/\" rel=\"nofollow noreferrer\">Theano<\/a> to implement a neural n-gram language model along the lines of <a href=\"https:\/\/www.google.com\/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=1&amp;cad=rja&amp;uact=8&amp;ved=0CCIQFjAA&amp;url=http%3A%2F%2Fmachinelearning.wustl.edu%2Fmlpapers%2Fpaper_files%2FBengioDVJ03.pdf&amp;ei=lmDiU5HuLJSlyAStv4LoBA&amp;usg=AFQjCNGlSitd2Z9Z-MSvGtcoJOwpuu28Ug&amp;sig2=GHJW6buocK6-jft5XZ6nIg&amp;bvm=bv.72197243,d.aWw\" rel=\"nofollow noreferrer\">Bengio et al 2003<\/a>. This model uses a distributed representation for words, and I'm having trouble writing a symbolic expression that allows me to take the gradient with respect to the word representation vectors.<\/p>\n\n<p>Following the notation in the paper, I have a word representation matrix  of size <em>V x m<\/em>, where <em>V<\/em> is the vocabulary size and <em>m<\/em> is the dimensionality of the word embedding. Each row of  is a vector representation of a word.<\/p>\n\n<p>My training data consists of n-grams drawn from a corpus. Let's say I let <em>n = 3<\/em>. Then I am trying to estimate <em>P(w<sub>t<\/sub>|w<sub>t-1<\/sub>, w<sub>t-2<\/sub>)<\/em>. A neural network estimates this probability by using the concatenated embedding vectors for <em>w<sub>t-1<\/sub><\/em> and <em>w<sub>t-2<\/sub><\/em> to predict <em>w<sub>t<\/sub><\/em> via a non-linear function. (See the paper for details.) Each word is represented by an index into a vocabulary which also indexes its representation row in . If these indexes are <em>i<sub>1<\/sub><\/em>, <em>i<sub>2<\/sub><\/em>, and <em>i<sub>3<\/sub><\/em>, I am trying to write a Theano expression for.<\/p>\n\n<pre><\/pre>\n\n<p>where  contains a hidden layer and a non-linear function, and  is the concatenation of the arrays  and . The first thing I have to do is write a symbolic Theano expression for . Also, this function needs to take not just a single training instance, but a mini-batch of multiple training instances.<\/p>\n\n<p>I know how to do this if I'm working directly with numpy matricies instead of abstract Theano expressions. For example, if  is shared, and  is a <em>N x n - 1<\/em> minibatch of <em>N<\/em> training vectors of word indexes, I can look up the concatenated vectors like so:<\/p>\n\n<pre><\/pre>\n\n<p>(A bit of index gymnastics I learned <a href=\"https:\/\/stackoverflow.com\/questions\/25146874\/concatenate-over-dimension-in-numpy\">elsewhere<\/a> on StackOverflow.)<\/p>\n\n<p>When I try to compile this expression into a Theano function, however, I run into errors.<\/p>\n\n<pre><\/pre>\n\n<p>The preceding gives me this error<\/p>\n\n<pre><\/pre>\n\n<p>I think this means that the index trick of putting -1 as the final reshape parameter is not supported by the Theano compiler.<\/p>\n\n<p>The equivalent command gives a different error.<\/p>\n\n<pre><\/pre>\n\n<p>I need to write the symbolic expression for  so that I can take its gradient with respect to . Can anyone help me do this?<\/p>\n\n<p>Alternately, can someone point me to example Theano code that works with word embeddings. All the tutorial material I've found has been for writing neural nets over image data, but I haven't seen any examples of how to do distributed representations.<\/p>\n","answers":[{"AnswerId":"25205016","CreationDate":"2014-08-08T13:46:59.560","ParentId":null,"OwnerUserId":"1120370","Title":null,"Body":"<p>Well, I'm an idiot. Kinda. I was missing an extra pair of parentheses around my <code>reshape<\/code> argument. The following works.<\/p>\n\n<pre><code>function([X_var], C[X_var].reshape((X_var.shape[0], -1)))\n<\/code><\/pre>\n\n<p>It's confusing though because the <code>reshape<\/code> method of an <code>array<\/code> will take either two arguments like I have above or a tuple like I have in the answer, but Theano will only compile the latter.<\/p>\n"}]}
{"QuestionId":25237039,"AnswerCount":4,"Tags":"<cuda><gpu><pickle><theano><deep-learning>","CreationDate":"2014-08-11T06:28:42.483","AcceptedAnswerId":null,"OwnerUserId":1505986.0,"Title":"Converting a theano model built on GPU to CPU?","Body":"<p>I have some pickle files of deep learning models built on gpu. I'm trying to use them in production. But when i try to unpickle them on the server, i'm getting the following error. <\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<br>\n    File \"score.py\", line 30, in <br>\n      model = (cPickle.load(file))<br>\n    File \"\/usr\/local\/python2.7\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/sandbox\/cuda\/type.py\", line 485, in CudaNdarray_unpickler<br>\n      return cuda.CudaNdarray(npa)<br>\n  AttributeError: (\"'NoneType' object has no attribute 'CudaNdarray'\", , (array([[ 0.011515  ,  0.01171047,  0.10408644, ..., -0.0343636 ,<br>\n           0.04944979, -0.06583775],<br>\n         [-0.03771918,  0.080524  , -0.10609912, ...,  0.11019105,<br>\n          -0.0570752 ,  0.02100536],<br>\n         [-0.03628891, -0.07109226, -0.00932018, ...,  0.04316209,<br>\n           0.02817888,  0.05785328],<br>\n         ...,<br>\n         [ 0.0703947 , -0.00172865, -0.05942701, ..., -0.00999349,<br>\n           0.01624184,  0.09832744],<br>\n         [-0.09029484, -0.11509365, -0.07193922, ...,  0.10658887,<br>\n           0.17730837,  0.01104965],<br>\n         [ 0.06659461, -0.02492988,  0.02271739, ..., -0.0646857 ,<br>\n           0.03879852,  0.08779807]], dtype=float32),))  <\/p>\n<\/blockquote>\n\n<p>I checked for that cudaNdarray package in my local machine and it is not installed, but still i am able to unpickle them. But in the server, i am unable to. How do i make them to run on a server which doesnt have a GPU?<\/p>\n","answers":[{"AnswerId":"41057109","CreationDate":"2016-12-09T09:29:28.790","ParentId":null,"OwnerUserId":"4087317","Title":null,"Body":"<p>I solved this problem by just saving the parameters W &amp; b, but not the whole model. You can save the parameters use this:<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loading_and_saving.html?highlight=saving%20load#robust-serialization\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/loading_and_saving.html?highlight=saving%20load#robust-serialization<\/a>\nThis can save the CudaNdarray to numpy array. Then you need to read the params by numpy.load(), and finally convert the numpy array to tensorSharedVariable use theano.shared().<\/p>\n"},{"AnswerId":"25243316","CreationDate":"2014-08-11T12:28:34.000","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>There is a script in pylearn2 which may do what you need:<\/p>\n\n<p><code>pylearn2\/scripts\/gpu_pkl_to_cpu_pkl.py<\/code><\/p>\n"},{"AnswerId":"31711656","CreationDate":"2015-07-29T21:39:58.067","ParentId":null,"OwnerUserId":"309653","Title":null,"Body":"<p>This works for me. Note: this doesn't work unless the following environment variable is set: <code>export THEANO_FLAGS='device=cpu'<\/code><\/p>\n\n<pre><code>import os\nfrom pylearn2.utils import serial\nimport pylearn2.config.yaml_parse as yaml_parse\n\nif __name__==\"__main__\":\n\n_, in_path, out_path = sys.argv\nos.environ['THEANO_FLAGS']=\"device=cpu\"\n\nmodel = serial.load(in_path)\n\nmodel2 = yaml_parse.load(model.yaml_src)\nmodel2.set_param_values(model.get_param_values())\n\nserial.save(out_path, model2)\n<\/code><\/pre>\n"},{"AnswerId":"28958979","CreationDate":"2015-03-10T08:23:15.550","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>The related Theano code is <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/type.py\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>From there, it looks like there is an option <code>config.experimental.unpickle_gpu_on_cpu<\/code> which you could set and which would make <code>CudaNdarray_unpickler<\/code> return the underlying raw Numpy array.<\/p>\n"}]}
{"QuestionId":25247373,"AnswerCount":3,"Tags":"<ubuntu><numpy><nose><theano>","CreationDate":"2014-08-11T15:53:16.853","AcceptedAnswerId":39937822.0,"OwnerUserId":3930228.0,"Title":"Need nose >= 0.10.0 error while attempting to install Theano on Ubuntu","Body":"<p>Overall I've had a hell of a time getting Theano to work, I've gotten to the stage where I <em>think<\/em> everything in stalled correctly.\nRunning:<\/p>\n\n<pre><\/pre>\n\n<p>and the console tells me that I have the latest version of everything.<\/p>\n\n<p>Running:<\/p>\n\n<pre><\/pre>\n\n<p>and I'm told that all requirements are satisfied.<\/p>\n\n<p>Yet when I try to do the tests as the guide recommends<\/p>\n\n<pre><\/pre>\n\n<p>gives me <\/p>\n\n<pre><\/pre>\n\n<p>I've done a thorough search online and all the solutions I've seen seem to be based around installing nose but I've definitely got nose installed and above 0.10.0.<\/p>\n","answers":[{"AnswerId":"25324947","CreationDate":"2014-08-15T10:31:31.473","ParentId":null,"OwnerUserId":"3930228","Title":null,"Body":"<p>As suggested I installed <a href=\"https:\/\/store.continuum.io\/cshop\/anaconda\/\" rel=\"nofollow\">Anaconda<\/a>, that did the trick.<\/p>\n"},{"AnswerId":"39937822","CreationDate":"2016-10-08T22:04:31.360","ParentId":null,"OwnerUserId":"4863734","Title":null,"Body":"<p>If you use <code>Anaconda<\/code> simply try this whet your <code>conda<\/code> environment is activated:<\/p>\n\n<pre><code>conda install nose\n<\/code><\/pre>\n\n<p>I had the same issue and <code>conda install<\/code> works without any <code>pip<\/code>!<\/p>\n"},{"AnswerId":"37839321","CreationDate":"2016-06-15T15:11:07.860","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>I faced similar issue in running theano.test() under Windows 7.\nI used Anaconda to create separate environment for python3.4.\nI installed theano using conda install for the above python3.4 environment.\nBut theano.test() gave me same error message as you have.\nSo I downloaded the nose-parameterized zip from github\nand installed it using the following :<\/p>\n\n<pre><code>((theanoBasic)) C:\\pythonZipsPy34\\manualZip\\nose-parameterized-master&gt;python setup.py install --record installNotes.txt\n<\/code><\/pre>\n"}]}
{"QuestionId":25248472,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2014-08-11T16:55:25.607","AcceptedAnswerId":25249911.0,"OwnerUserId":163173.0,"Title":"How to convert torch Tensor\/ Storage to a lua table?","Body":"<p>If I have a tensor:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any way get this data as a Lua table?<\/p>\n","answers":[{"AnswerId":"25249911","CreationDate":"2014-08-11T18:22:59.900","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>There is a dedicated constructor to <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#torchtensortable\">create a tensor from a table<\/a> but so far there is no method out-of-the box to convert the other way around.<\/p>\n\n<p>Of course you can do that <em>manually<\/em>:<\/p>\n\n<pre><code>-- This assumes `t1` is a 2-dimensional tensor!\nlocal t2 = {}\nfor i=1,t1:size(1) do\n  t2[i] = {}\n  for j=1,t1:size(2) do\n    t2[i][j] = t1[i][j]\n  end\nend\n<\/code><\/pre>\n\n<p>--<\/p>\n\n<p><strong>Update<\/strong>: as of <a href=\"https:\/\/github.com\/torch\/torch7\/commit\/10f3323\">commit 10f3323<\/a> there is now a dedicated <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/ff11731\/doc\/utility.md#table-torchtotableobject\"><code>torch.totable(object)<\/code><\/a> converter.<\/p>\n"}]}
{"QuestionId":25269922,"AnswerCount":1,"Tags":"<python><ubuntu><numpy><theano>","CreationDate":"2014-08-12T16:56:27.573","AcceptedAnswerId":25282710.0,"OwnerUserId":3930228.0,"Title":"Can't get Theano to work on ubuntu 14.04","Body":"<p>I'm trying to use Theano on ubuntu 14.04, I've followed the guide for an easy install located here <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu<\/a><\/p>\n\n<p>Everything says it's installed fine, if I run:<\/p>\n\n<pre><\/pre>\n\n<p>Then I get in return <\/p>\n\n<pre><\/pre>\n\n<p>And when running<\/p>\n\n<pre><\/pre>\n\n<p>I get<\/p>\n\n<pre><\/pre>\n\n<p>But when I go to run the tests they just don't work.<\/p>\n\n<pre><\/pre>\n\n<p>gives me<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>gives me<\/p>\n\n<pre><\/pre>\n\n<p>The last test gives very similar results<\/p>\n\n<pre><\/pre>\n\n<p>I'm a complete linux newbie so I'm completely baffled by what could be the problem.<\/p>\n","answers":[{"AnswerId":"25282710","CreationDate":"2014-08-13T09:28:32.990","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>If you can, I would simply use a \"scientific\" Python - either <a href=\"https:\/\/store.continuum.io\/cshop\/anaconda\/\" rel=\"nofollow\">Anaconda<\/a> (my preference) or Enthought Python. In addition to avoiding systemwide installation of packages, it is easy to install things with pip and numpy and scipy come preinstalled. For updating numpy and scipy you can also use the built-in conda package manager - it does an excellent job of handling the nasty work behind installing a new numpy or scipy. <\/p>\n"}]}
{"QuestionId":25284195,"AnswerCount":1,"Tags":"<python><gpu><theano>","CreationDate":"2014-08-13T10:49:35.220","AcceptedAnswerId":null,"OwnerUserId":2447425.0,"Title":"How to stack vectors in Theano without using scan?","Body":"<p>I am using theano.scan for creating stacked vector of contexts like this:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that scan is so slow, that this slows down whole processing. \nIn signal processing this is fairly standard operation, so I was thinking about creating a special op, just for this. \nUnfortunately, I would need also GPU implementation and grad for this op and it looks like a long shot for me. \nCan you kick me in the right direction? I have already read Extending theano documentation, but still doesn't help a lot.<\/p>\n\n<p>Example:<\/p>\n\n<p>in case of <\/p>\n\n<pre><\/pre>\n\n<p>matrix:<\/p>\n\n<pre><\/pre>\n\n<p>would be converted to <\/p>\n\n<pre><\/pre>\n\n<p>Thank you\nJ<\/p>\n","answers":[{"AnswerId":"25293490","CreationDate":"2014-08-13T18:26:44.580","ParentId":null,"OwnerUserId":"2447425","Title":null,"Body":"<p>So the problem is solvable in this way. Works now.<\/p>\n\n<pre><code>Y_= T.concatenate([Y_[c:Y_.shape[0]+c-left_ctx-right_ctx] for c in range(left_ctx+right_ctx+1)], axis=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":25326462,"AnswerCount":1,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2014-08-15T12:33:06.230","AcceptedAnswerId":25334094.0,"OwnerUserId":1461210.0,"Title":"Initializing a symmetric Theano dmatrix from its upper triangle","Body":"<p>I'm trying to fit a Theano model that is parametrized in part by a symmetric matrix . In order to enforce the symmetry of , I want to be able to construct  by passing in just the values in the upper triangle.<\/p>\n\n<p>The equivalent numpy code might look something like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, since symbolic tensor variables don't support item assignment, I'm struggling to find a way to do this in Theano.<\/p>\n\n<p>The closest thing I could find is , which allows me to construct a symbolic matrix from its diagonal:<\/p>\n\n<pre><\/pre>\n\n<p>Whilst there is also a  function, this cannot be used to construct a matrix from the upper triangle, but rather returns a copy of an array with the lower triangular elements zeroed.<\/p>\n\n<p>Is there any way to construct a Theano symbolic matrix from its upper triangle?<\/p>\n","answers":[{"AnswerId":"25334094","CreationDate":"2014-08-15T20:55:50.580","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You could use the <code>theano.tensor.triu<\/code> and add the result to its transpose, then subtract the diagonal.<\/p>\n\n<p>Copy+Pasteable code:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\ntheano.config.floatX = 'float32'\n\nmat = T.fmatrix()\nsym1 = T.triu(mat) + T.triu(mat).T\ndiag = T.diag(T.diagonal(mat))\nsym2 = sym1 - diag\n\nf_sym1 = theano.function([mat], sym1)\nf_sym2 = theano.function([mat], sym2)\n\nm = np.arange(9).reshape(3, 3).astype(np.float32)\n\nprint m\n# [[ 0.  1.  2.]\n#  [ 3.  4.  5.]\n#  [ 6.  7.  8.]]\nprint f_sym1(m)\n# [[  0.   1.   2.]\n#  [  1.   8.   5.]\n#  [  2.   5.  16.]]\nprint f_sym2(m)\n# [[ 0.  1.  2.]\n#  [ 1.  4.  5.]\n#  [ 2.  5.  8.]]\n<\/code><\/pre>\n\n<p>Does this help? This approach would require a full matrix to be passed, but would ignore everything below the diagonal and symmetrize using the upper triangle.<\/p>\n\n<p>We can also take a look at the derivative of this function. In order not to deal with a multidimensional output, we can e.g. look at the gradient of the sum of the matrix entries<\/p>\n\n<pre><code>sum_grad = T.grad(cost=sym2.sum(), wrt=mat)\nf_sum_grad = theano.function([mat], sum_grad)\n\nprint f_sum_grad(m)\n# [[ 1.  2.  2.]\n#  [ 0.  1.  2.]\n#  [ 0.  0.  1.]]\n<\/code><\/pre>\n\n<p>This reflects the fact that the upper triangular entries figure doubly in the sum.<\/p>\n\n<hr>\n\n<p>Update: You can do normal indexing:<\/p>\n\n<pre><code>n = 4\nnum_triu_entries = n * (n + 1) \/ 2\n\ntriu_index_matrix = np.zeros([n, n], dtype=int)\ntriu_index_matrix[np.triu_indices(n)] = np.arange(num_triu_entries)\ntriu_index_matrix[np.triu_indices(n)[::-1]] = np.arange(num_triu_entries)\n\ntriu_vec = T.fvector()\ntriu_mat = triu_vec[triu_index_matrix]\n\nf_triu_mat = theano.function([triu_vec], triu_mat)\n\nprint f_triu_mat(np.arange(1, num_triu_entries + 1).astype(np.float32))\n\n# [[  1.   2.   3.   4.]\n#  [  2.   5.   6.   7.]\n#  [  3.   6.   8.   9.]\n#  [  4.   7.   9.  10.]]\n<\/code><\/pre>\n\n<hr>\n\n<p>Update: To do all of this dynamically, one way is to write a symbolic version of <code>triu_index_matrix<\/code>. This can be done with some shuffling of <code>arange<\/code>s. But probably I am overcomplicating.<\/p>\n\n<pre><code>n = T.iscalar()\nn_triu_entries = (n * (n + 1)) \/ 2\nr = T.arange(n)\n\ntmp_mat = r[np.newaxis, :] + (n_triu_entries - n - (r * (r + 1)) \/ 2)[::-1, np.newaxis]\ntriu_index_matrix = T.triu(tmp_mat) + T.triu(tmp_mat).T - T.diag(T.diagonal(tmp_mat))\n\ntriu_vec = T.fvector()\nsym_matrix = triu_vec[triu_index_matrix]\n\nf_triu_index_matrix = theano.function([n], triu_index_matrix)\nf_dynamic_sym_matrix = theano.function([triu_vec, n], sym_matrix)\n\nprint f_triu_index_matrix(5)\n# [[ 0  1  2  3  4]\n#  [ 1  5  6  7  8]\n#  [ 2  6  9 10 11]\n#  [ 3  7 10 12 13]\n# [ 4  8 11 13 14]]\nprint f_dynamic_sym_matrix(np.arange(1., 16.).astype(np.float32), 5)\n# [[  1.   2.   3.   4.   5.]\n#  [  2.   6.   7.   8.   9.]\n#  [  3.   7.  10.  11.  12.]\n#  [  4.   8.  11.  13.  14.]\n#  [  5.   9.  12.  14.  15.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":25330635,"AnswerCount":2,"Tags":"<python><enthought><theano>","CreationDate":"2014-08-15T16:49:54.653","AcceptedAnswerId":null,"OwnerUserId":2688733.0,"Title":"theano NotImplementedError","Body":"<p>I am running some theano code making use of tensor.advanced_subtensor\nI am getting the following error :<\/p>\n\n<pre><\/pre>\n\n<p>I have the latest version of theano (0.6.0.dev-60b5ccc2bcabb1010714376764daf8a50722cee9) and numpy (1.8.0). Why am I still getting this error? How can I resolve this error? How do I clear theano cache?<\/p>\n","answers":[{"AnswerId":"44953776","CreationDate":"2017-07-06T15:56:29.133","ParentId":null,"OwnerUserId":"5468983","Title":null,"Body":"<p>You need to clear Theano cache. Cache live in <strong>~\/.theano\/<\/strong> folder.\nFollow below steps to clear it manually.<\/p>\n\n<pre><code> import theano\n print (theano.config.compiledir)\n # and then delete directory returned from above.\n<\/code><\/pre>\n\n<p>If you don not want to delete manually then use below command.<\/p>\n\n<pre><code>theano-cache purge\n<\/code><\/pre>\n"},{"AnswerId":"25334422","CreationDate":"2014-08-15T21:22:35.227","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The theano cache is usually in <code>~\/.theano\/<\/code> if you are using *ix.<\/p>\n"}]}
{"QuestionId":25339439,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2014-08-16T10:53:08.580","AcceptedAnswerId":25343861.0,"OwnerUserId":596046.0,"Title":"How to print values from inside a theano function?","Body":"<p>I've recently moved from Matlab\/C++ to theano and have the following function<\/p>\n\n<pre><\/pre>\n\n<p>and I'd like to print the values between 2 layers of the net in every iteration (for debugging, better control of the function etc.)\nI've tried editing the function setting the classifier so that it prints (either using print() or theano.printing.Print\/theano.pp()) and all I get is a single print while the model is being set.<\/p>\n","answers":[{"AnswerId":"25343861","CreationDate":"2014-08-16T20:50:33.680","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>In your example <code>classifier.cost<\/code> is an expression, probably consisting of several other expressions building on the same input. You can turn any of those intermediate expressions into functions just as you are doing with <code>classifier.cost<\/code>, e.g.<\/p>\n\n<pre><code>f_first_layer = theano.function([x], first_layer)\n<\/code><\/pre>\n\n<p>You can then call and print the output of this function e.g. after every call to <code>train_model<\/code>. If you call it <em>before<\/em> <code>train_model<\/code> with the same params as you will call <code>train_model<\/code> with just after, then you will have the exact output of the layers as they will be evaluated by <code>train_model<\/code> (calling it after <code>train_model<\/code> will be different due to updating).<\/p>\n"}]}
{"QuestionId":25365440,"AnswerCount":1,"Tags":"<python><oop><numpy><pickle><theano>","CreationDate":"2014-08-18T14:13:09.920","AcceptedAnswerId":25394371.0,"OwnerUserId":1461210.0,"Title":"Why am I allowed pickle instancemethods that are Theano functions, but not normal instancemethods?","Body":"<p>In the process of using joblib to parallelize some model-fitting code involving Theano functions, I've stumbled across some behavior that seems odd to me.<\/p>\n\n<p>Consider this very simplified example:<\/p>\n\n<pre><\/pre>\n\n<p>I understand why the first case fails, since  is clearly an instancemethod of . What I don't understand is why the second case <em>works<\/em>, since ,  and are only declared within the  method of .  can't be evaluated until the  instance has been created. Surely, then, it should also be considered an instancemethod, and should therefore be unpickleable? In fact, even still works if I explicitly declare  and  to be attributes of the  instance.<\/p>\n\n<p>Can anyone explain what's going on here?<\/p>\n\n<hr>\n\n<h2>Update<\/h2>\n\n<p>Just to illustrate why I think this behaviour is particularly weird, here are a few examples of some other callable member objects that don't take  as the first argument:<\/p>\n\n<pre><\/pre>\n\n<p>None of them are pickleable with the exception of the !<\/p>\n","answers":[{"AnswerId":"25394371","CreationDate":"2014-08-19T23:08:39.283","ParentId":null,"OwnerUserId":"322806","Title":null,"Body":"<p>Theano functions aren't python functions. Instead they are python objects that override <code>__call__<\/code>.  This means that you can call them just like a function but internally they are really objects of some custom class. In consequence, you can pickle them. <\/p>\n"}]}
{"QuestionId":25366863,"AnswerCount":1,"Tags":"<python><numpy><theano><gradient-descent><deep-learning>","CreationDate":"2014-08-18T15:27:52.347","AcceptedAnswerId":25367188.0,"OwnerUserId":3670532.0,"Title":"Clarification in the Theano tutorial","Body":"<p>I am reading <a href=\"http:\/\/nbviewer.ipython.org\/github\/craffel\/theano-tutorial\/blob\/master\/Theano%20Tutorial.ipynb\" rel=\"noreferrer\">this tutorial<\/a> provided on the <a href=\"http:\/\/deeplearning.net\/software\/theano\/index.html\" rel=\"noreferrer\">home page of Theano documentation<\/a><\/p>\n\n<p>I am not sure about the code given under the gradient descent section.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Vu9t4.png\" alt=\"enter image description here\"><\/p>\n\n<p><strong>I have doubts about the for loop<\/strong>.<\/p>\n\n<p>If you initialize the '<strong>param_update<\/strong>' variable to zero.<\/p>\n\n<pre><\/pre>\n\n<p>and then you update its value in the remaining two lines.<\/p>\n\n<pre><\/pre>\n\n<p>Why do we need it?<\/p>\n\n<p>I guess I am getting something wrong here.\nCan you guys help me!<\/p>\n","answers":[{"AnswerId":"25367188","CreationDate":"2014-08-18T15:46:25.873","ParentId":null,"OwnerUserId":"3953341","Title":null,"Body":"<p>The initialization of <code>param_update<\/code> using <code>theano.shared(.)<\/code> only tells Theano to reserve a variable that will be used by Theano functions. This initialization code is only called once, and will not be used later on to reset the value of <code>param_update<\/code> to 0.<\/p>\n\n<p>The actual value of <code>param_update<\/code> will be updated according to the last line <\/p>\n\n<pre><code>updates.append((param_update, momentum*param_update + (1. - momentum)*T.grad(cost, param)))\n<\/code><\/pre>\n\n<p>when <code>train<\/code> function that was constructed by having this update dictionary as an argument ([23] in the tutorial):<\/p>\n\n<pre><code>train = theano.function([mlp_input, mlp_target], cost,\n                        updates=gradient_updates_momentum(cost, mlp.params, learning_rate, momentum))\n<\/code><\/pre>\n\n<p>Each time <code>train<\/code> is called, Theano will compute the gradient of the <code>cost<\/code> w.r.t. <code>param<\/code> and update <code>param_update<\/code> to a new update direction according to momentum rule. Then, <code>param<\/code> will be updated by following the update direction saved in <code>param_update<\/code> with an appropriate <code>learning_rate<\/code>.<\/p>\n"}]}
{"QuestionId":25390978,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2014-08-19T18:56:47.417","AcceptedAnswerId":25428302.0,"OwnerUserId":3670532.0,"Title":"What does 'no_inplace' mean in theano?","Body":"<p>Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>and the output is<\/p>\n\n<pre><\/pre>\n\n<p>add shows that the apply node has the add as the operation.<\/p>\n\n<p><strong>But what does no_inplace mean?\nand why we have a \".0\" at the end of the output?<\/strong><\/p>\n","answers":[{"AnswerId":"25428302","CreationDate":"2014-08-21T13:58:59.820","ParentId":null,"OwnerUserId":"3670532","Title":null,"Body":"<p>Inplace computations are computations that destroy their inputs as a side-effect. For example, if you iterate over a matrix and double every element, this is an inplace operation because when you are done, the original input has been overwritten. Ops representing inplace computations are destructive, and by default these can only be inserted by optimizations, not user code.<\/p>\n\n<p>So no_inplace is just the opposite.<\/p>\n\n<p>From <a href=\"http:\/\/deeplearning.net\/software\/theano\/glossary.html#glossary\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/glossary.html#glossary<\/a><\/p>\n"}]}
{"QuestionId":25422826,"AnswerCount":2,"Tags":"<python><ipython><theano><pydot>","CreationDate":"2014-08-21T09:28:59.143","AcceptedAnswerId":25426249.0,"OwnerUserId":3670532.0,"Title":"Error using python pydot","Body":"<p>I get error using the theano.printing.pydotprint() function<\/p>\n\n<p>following lines work fine without any error:<\/p>\n\n<pre><\/pre>\n\n<p>Also when I run<\/p>\n\n<pre><\/pre>\n\n<p>in the python interpreter I get output as<\/p>\n\n<pre><\/pre>\n\n<p>but the problem is when I execute the script using the function I get following error<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what is the problem?<\/p>\n\n<p>P.S: I am running the python tutorial given here:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/printing_drawing.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/printing_drawing.html<\/a>\nSo the call to the function is surely correct.<\/p>\n\n<p>Here is the traceback of the error I am getting:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"25426249","CreationDate":"2014-08-21T12:21:16.510","ParentId":null,"OwnerUserId":"3670532","Title":null,"Body":"<p>Tried reinstalling pydot as given by the <a href=\"https:\/\/stackoverflow.com\/questions\/15951748\/pydot-and-graphviz-error-couldnt-import-dot-parser-loading-of-dot-files-will\">solution to this problem<\/a>, but this was not working.<\/p>\n\n<p>That is<\/p>\n\n<pre><code>pip uninstall pyparsing\npip install -Iv https:\/\/pypi.python.org\/packages\/source\/p\/pyparsing\/pyparsing-1.5.7.tar.gz#md5=9be0fcdcc595199c646ab317c1d9a709\npip install pydot\n<\/code><\/pre>\n\n<p>there was some problem with this installation, even though every time, installed successfully message was given.<\/p>\n\n<p>But<\/p>\n\n<pre><code>sudo apt-get install python-pydot\n<\/code><\/pre>\n\n<p>this worked.<\/p>\n\n<p>\"Because the solution was not to install pydot from somewhere, but \"python-pydot\" from official ubuntu repositories.\" - <a href=\"https:\/\/stackoverflow.com\/questions\/15951748\/pydot-and-graphviz-error-couldnt-import-dot-parser-loading-of-dot-files-will\">answer by sadik<\/a> worked<\/p>\n\n<p>We must note that on successful installation of pydot, it can be checked at two places.<br\/>\n\/usr\/share\/doc\/python-pydot<br\/>\nand<br\/>\n\/usr\/share\/python-support\/python-pydot<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/nuAw5.png\" alt=\"enter image description here\"><\/p>\n"},{"AnswerId":"25424972","CreationDate":"2014-08-21T11:16:27.173","ParentId":null,"OwnerUserId":"1403430","Title":null,"Body":"<p>Abhishek : Check if you could see the \"pydot\" folder under the lib folder. Looks like you are in <em>ix<\/em> machine. Ideally you would find it is installed or not within 'lib' folder \/ or within site-packages.<\/p>\n\n<p>Meanwhile, I would suggest you to try re-installing the package(pydot) and see if it helps. <\/p>\n"}]}
{"QuestionId":25423173,"AnswerCount":2,"Tags":"<python><attributeerror><theano>","CreationDate":"2014-08-21T09:47:00.147","AcceptedAnswerId":null,"OwnerUserId":1521172.0,"Title":"Name conflicting in Theano","Body":"<p>I am trying to import theano in a module, but I am getting a traceback:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that there is some name conflict in some config. Can anybody please point me to the same.<\/p>\n","answers":[{"AnswerId":"41095855","CreationDate":"2016-12-12T07:18:19.797","ParentId":null,"OwnerUserId":"3935797","Title":null,"Body":"<p>I got similar error using when using the jupyter notebook. Restarting kernel solved the issue.<\/p>\n"},{"AnswerId":"31148162","CreationDate":"2015-06-30T20:44:06.220","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>This error happens because some module, probably <code>theano.gof<\/code>, is imported twice. Usually, this is because a first call to <code>import theano.gof<\/code> gets started, registering <code>'gcc.cxxflags'<\/code> in the configuration parser a first time, but then raises <code>ImportError<\/code>, which is catched and ignored.\nThen, <code>import theano.gof<\/code> gets called again, tries to register the option again, which raises the exception you get.<\/p>\n\n<p>Is there any traceback or error message before this one, or something that would give a hint of why the first import failed?<\/p>\n"}]}
{"QuestionId":25449271,"AnswerCount":1,"Tags":"<python><linux><machine-learning><theano><deep-learning>","CreationDate":"2014-08-22T14:23:21.350","AcceptedAnswerId":null,"OwnerUserId":3968255.0,"Title":"Why is Theano (much) slower on Windows than on Linux?","Body":"<p>I implemented a recursive autoencoder with Theano and tested it on both Linux and Windows. It tooks ~3 hours, 2.3G memory on Linux, while ~9 hours, 0.5G memory on Windows. config.allow_gc=True for both cases.<\/p>\n\n<p>It could be a Python issue, as discussed in the thread: <a href=\"https:\/\/stackoverflow.com\/questions\/10150881\/why-is-python-so-much-slower-on-windows\">Why is python so much slower on windows?<\/a><\/p>\n\n<p>Is there any specific setting in Theano that could slow things down on Windows as well?<\/p>\n\n<p>Thanks,<\/p>\n\n<p>Ya<\/p>\n","answers":[{"AnswerId":"25457043","CreationDate":"2014-08-22T23:44:54.880","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>It could be that they use different BLAS librairies. From memory, autoencoder bottleneck is the matrix product, that call BLAS. Different BLAS implementation can have up to 10x speed difference.<\/p>\n\n<p>So check if you used the same BLAS. I would recommand to install python via EPD\/Canopy or Anaconda python packages. There not free version link to a good blas and Theano reuse it. The now free version is free for academic.<\/p>\n"}]}
{"QuestionId":25500045,"AnswerCount":1,"Tags":"<python><cuda><gpu><theano>","CreationDate":"2014-08-26T07:12:44.277","AcceptedAnswerId":null,"OwnerUserId":3670532.0,"Title":"GPU with Theano giving poor result compared to CPU with theano","Body":"<p>I am doing this tutorial:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#exercise\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#exercise<\/a><\/p>\n\n<p>and the solution to the tutorial is given here:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/_downloads\/using_gpu_solution_1.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/_downloads\/using_gpu_solution_1.py<\/a><\/p>\n\n<p>But my issue is when I run the code<\/p>\n\n<p>with GPU:<\/p>\n\n<pre><\/pre>\n\n<p>I got the following output:<\/p>\n\n<pre><\/pre>\n\n<p>and with CPU:<\/p>\n\n<pre><\/pre>\n\n<p>I got the following output:<\/p>\n\n<pre><\/pre>\n\n<p>In <a href=\"http:\/\/deeplearning.net\/software\/theano\/_downloads\/using_gpu_solution_1.py\" rel=\"nofollow\">the solution<\/a> they have mentioned the speed up of almost double with GPU.\nBut I am getting more elapsed time with GPU than CPU.<\/p>\n\n<p>Is it that the code ran on multiple core on CPU and it got improvements over GPU?<\/p>\n\n<p>Can anyone tell me what I am getting wrong?\nOnly thing I can see is improve in system time using the GPU.\nIs this what they mean by speed up? as the overall elapsed time is still more with GPU.<\/p>\n","answers":[{"AnswerId":"25518636","CreationDate":"2014-08-27T03:48:28.580","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>There is one problems (1) and consideration about this(2-3):<\/p>\n\n<p>1) You don't time correctly. The way you time, you include the theano compilation time. Theano compilation is not made to be included in the profile. You should only time the time spent inside Theano function. For this, modify the script of use the Theano profiler as in the profile.<\/p>\n\n<p>2) This is a toy example. It do pure Stochastic Gradient Descent(SGD). To get good speed up from the GPU, we need to use minibatch with SGD. If we don't the GPU don't have enough data to parallellize the computation.<\/p>\n\n<p>3) As this is a toy example with only a small model, the speed up will vary highly depending of the CPU and GPU used. It is probably that you did your timming on a better CPU then the original, or that you did it with a parallel BLAS.<\/p>\n"}]}
{"QuestionId":25619458,"AnswerCount":2,"Tags":"<algorithm><convolution><theano>","CreationDate":"2014-09-02T08:52:44.560","AcceptedAnswerId":null,"OwnerUserId":3897995.0,"Title":"perform the exact same convolution as in theano's conv2d","Body":"<p>I have an existing classification model that was trained using theano's conv2d under theano.tensor.nnet. Now I have to use this model to do some sort of prediction in Java. <\/p>\n\n<p>I implement a simple convolution in Python(In the end, I will code it in Java) as per some documentation(<a href=\"https:\/\/developer.apple.com\/Library\/ios\/documentation\/Performance\/Conceptual\/vImage\/ConvolutionOperations\/ConvolutionOperations.html\" rel=\"nofollow\">https:\/\/developer.apple.com\/Library\/ios\/documentation\/Performance\/Conceptual\/vImage\/ConvolutionOperations\/ConvolutionOperations.html<\/a>).  For example, for a 2*2 kernel (k11,k12,k21,k22), one of the areas under the kernel is (a11,a12,a21,a22). The convolution is performed by a11*k11 + a12*k12 + a21*k21 + a22*k22.<\/p>\n\n<p>Unfortunately, as I test my convolution code and theano's conv code with some dummy matrix and kernels, they give different results. Only in some rare cases, they give same results.<\/p>\n\n<p>It seems to me that there are many variants of convolution algorithm and I have to implement the exact same convolution algorithm as used by theano's ConvOp. However, I can't find any material describing theano's Conv2d algorithm. <\/p>\n\n<p>Could you explain a little bit about theano's conv2d algorithm?<\/p>\n\n<p>The following is my python code for convolution:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"26234696","CreationDate":"2014-10-07T11:12:37.737","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Theano convolution does exactly the same thing as <code>scipy.signal.convolve2d<\/code>. This is exploited\/tested e.g. <a href=\"https:\/\/github.com\/sklearn-theano\/sklearn-theano\/blob\/master\/sklearn_theano\/tests\/test_base.py#L10\" rel=\"nofollow\">here<\/a>. For self-containedness, try copy+pasting:<\/p>\n\n<pre><code>import numpy as np\nfrom scipy.signal import convolve2d\nimport theano\nimport theano.tensor as T\n\nrng = np.random.RandomState(42)\nimage = rng.randn(500, 500).astype(np.float32)\nconv_filter = rng.randn(32, 32).astype(np.float32)\n\nimg = T.tensor4()\nfil = T.tensor4()\n\nfor border_mode in [\"full\", \"valid\"]:\n    scipy_convolved = convolve2d(image, conv_filter, mode=border_mode)\n    theano_convolve2d = theano.function([img, fil], T.nnet.conv2d(img, fil,\n                                        border_mode=border_mode))\n    theano_convolved = theano_convolve2d(image.reshape(1, 1, 500, 500),\n                                         conv_filter.reshape(1, 1, 32, 32))\n    l2_discrepancy = np.sqrt(((scipy_convolved - theano_convolved) ** 2).sum())\n    print \"Discrepancy %1.5e for border mode %s\" % (l2_discrepancy, border_mode)\n    print \"Norms of convolutions: %1.5e, %1.5e\" % (\n        np.linalg.norm(scipy_convolved.ravel()),\n        np.linalg.norm(theano_convolved.ravel()))\n<\/code><\/pre>\n\n<p>It outputs<\/p>\n\n<pre><code>Discrepancy 9.42469e-03 for border mode full\nNorms of convolutions: 1.65433e+04, 1.65440e+04\nDiscrepancy 9.03687e-03 for border mode valid\nNorms of convolutions: 1.55051e+04, 1.55054e+04\n<\/code><\/pre>\n\n<p>Since scipy implements a standard convolution, so does theano.<\/p>\n"},{"AnswerId":"39976915","CreationDate":"2016-10-11T12:08:48.733","ParentId":null,"OwnerUserId":"3876563","Title":null,"Body":"<p>Actually you and the (Theano, Scripy) are all right. the reason is that: your are using the different convolution2D. the Theano and Script using the convolution2D defined in Math that should rotate the kernel. but you did not (ref: <a href=\"http:\/\/www.songho.ca\/dsp\/convolution\/convolution.html#convolution_2d\" rel=\"nofollow\">http:\/\/www.songho.ca\/dsp\/convolution\/convolution.html#convolution_2d<\/a>).\nSO if your kernel is like this:<\/p>\n\n<pre><code>[1, 2, 3]\n[4, 5, 6]\n[7, 8, 9]\n<\/code><\/pre>\n\n<p>than change it to (center symmetry):<\/p>\n\n<pre><code>[9, 8, 7]\n[6, 5, 4]\n[3, 2, 1]\n<\/code><\/pre>\n\n<p>So using your method will get the same answer as Theano\/Scripy<\/p>\n"}]}
{"QuestionId":25685104,"AnswerCount":1,"Tags":"<pymc><theano><dirichlet><pymc3>","CreationDate":"2014-09-05T11:40:06.987","AcceptedAnswerId":25846440.0,"OwnerUserId":601308.0,"Title":"Dirichlet process in PyMC 3","Body":"<p>I would like to implement to implement the Dirichlet process example referenced in\n<a href=\"http:\/\/stronginference.com\/post\/implementing-dirichlet-processes-for-bayesian-semi-parametric-models\" rel=\"nofollow\">Implementing Dirichlet processes for Bayesian semi-parametric models<\/a> (source: <a href=\"https:\/\/github.com\/fonnesbeck\/pymc_radon\/blob\/master\/radon_dp.py\" rel=\"nofollow\">here<\/a>) in PyMC 3.<\/p>\n\n<p>In the example the stick-breaking probabilities are computed using the \ndecorator:<\/p>\n\n<pre><\/pre>\n\n<p>How would you implement this in PyMC 3 which is using Theano for the gradient computation?<\/p>\n\n<p>edit:\nI tried the following solution using the  method:<\/p>\n\n<pre><\/pre>\n\n<p>Which sadly is really slow and does not obtain the original parameters of the synthetic data.<\/p>\n\n<p>Is there a better solution and is this even correct?<\/p>\n","answers":[{"AnswerId":"25846440","CreationDate":"2014-09-15T10:54:31.927","ParentId":null,"OwnerUserId":"2288595","Title":null,"Body":"<p>Not sure I have a good answer but perhaps this could be sped up by instead using a theano blackbox op which allows you to write a distribution (or deterministic) in python code. E.g.: <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/examples\/disaster_model_arbitrary_deterministic.py\" rel=\"nofollow\">https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/examples\/disaster_model_arbitrary_deterministic.py<\/a><\/p>\n"}]}
{"QuestionId":25729969,"AnswerCount":7,"Tags":"<python><windows><cuda><mingw><theano>","CreationDate":"2014-09-08T17:37:48.177","AcceptedAnswerId":26073714.0,"OwnerUserId":1663093.0,"Title":"Installing theano on Windows 8 with GPU enabled","Body":"<p>I understand that the Theano support for Windows 8.1 is at experimental stage only but I wonder if anyone had any luck with resolving my issues. Depending on my config, I get three distinct types of errors. I assume that the resolution of any of my errors would solve my problem.<\/p>\n\n<p>I have installed Python using WinPython 32-bit system, using MinGW as described <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\">here<\/a>. The contents of my  file are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>When I run  the error is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I have also tested it using  which is installed on my system with the following error:<\/p>\n\n<pre><\/pre>\n\n<p>In the latter error, several pop-up windows ask me how would I like to open (.res) file before error is thrown.<\/p>\n\n<p> is present in both folders (i.e. VS 2010 and VS 2013). <\/p>\n\n<p>Finally, if I set VS 2013 in the environment path and set  contents as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>If I run  without the GPU option on, it runs without a problem. Also CUDA samples run without a problem. <\/p>\n","answers":[{"AnswerId":"34711198","CreationDate":"2016-01-10T21:51:00.907","ParentId":null,"OwnerUserId":"5771005","Title":null,"Body":"<p>I used <a href=\"https:\/\/stackoverflow.com\/a\/26073714\/1709587\">this guide<\/a>, and it was quite helpful.\nWhat many of Windows Theano guides only mention in passing (or not at all) is that you will need to compile theano from mingw shell, not from your IDE.<\/p>\n\n<p>I ran mingw-w64.bat, and from there \"python\" and \"import theano\". Only after that importing it from pycharm works.<\/p>\n\n<p>Additionally, official instructions on deeplearning.net are bad because they tell you to use CUDA 5.5, but it won't work with newer video cards.<\/p>\n\n<p>The comments are also quite helpful. If it complains about missing crtdefs.h or basetsd.h, do what Sunando's answer says. If AFTER THAT it still complains that identifier \"Iunknown\" is undefined in objbase.h, stick the following in \nC:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\\Include\\objbase.h file, on line 236:<\/p>\n\n<pre><code>#include &lt;wtypes.h&gt;\n#include &lt;unknwn.h&gt;\n<\/code><\/pre>\n\n<p>I had to do this last part to make it work with bleeding edge install (required for parts of Keras).<\/p>\n\n<p>I also wrote a list of things that worked for me, here:\n<a href=\"http:\/\/acoupleofrobots.com\/everything\/?p=2238\" rel=\"nofollow noreferrer\">http:\/\/acoupleofrobots.com\/everything\/?p=2238<\/a>\nThis is for 64 bit version.<\/p>\n"},{"AnswerId":"33128977","CreationDate":"2015-10-14T15:04:49.807","ParentId":null,"OwnerUserId":"3816062","Title":null,"Body":"<p>I could compile the cu files by adding the required dependencies in the nvcc profile located in \u201cC:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\bin\\nvcc.profile\u201d<\/p>\n\n<p>I modified the include and the lib path and it started working.<\/p>\n\n<p>INCLUDES += \u201c-I$(TOP)\/include\u201d $(<em>SPACE<\/em>) \u201c-IC:\/Program Files (x86)\/Microsoft Visual Studio 12.0\/VC\/include\u201d $(<em>SPACE<\/em>) \u201c-IC:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\\Include\u201d $(<em>SPACE<\/em>)\nLIBRARIES =+ $(<em>SPACE<\/em>) \u201c\/LIBPATH:$(TOP)\/lib\/$(_WIN_PLATFORM_)\u201d $(<em>SPACE<\/em>) \u201c\/LIBPATH:C:\/Program Files (x86)\/Microsoft Visual Studio 12.0\/VC\/lib\/amd64\u201d $(<em>SPACE<\/em>) \u201c\/LIBPATH:C:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\\Lib\\x64\u201d $(<em>SPACE<\/em>)<\/p>\n\n<p>I have made a full documentation of the install, hope it helps <a href=\"https:\/\/planetanacreon.wordpress.com\/2015\/10\/09\/install-theano-on-windows-8-1-with-visual-studio-2013-cuda-7-5\/\" rel=\"nofollow\">https:\/\/planetanacreon.wordpress.com\/2015\/10\/09\/install-theano-on-windows-8-1-with-visual-studio-2013-cuda-7-5\/<\/a><\/p>\n"},{"AnswerId":"29081593","CreationDate":"2015-03-16T16:08:33.083","ParentId":null,"OwnerUserId":"4677134","Title":null,"Body":"<p>Following the tutorial by Matt, I ran into issues with nvcc.\nI needed to add the path to VS2010 executables in nvcc.profile (you can find it in the cuda bin folder): <\/p>\n\n<p><code>\"compiler-bindir =  C:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\bin\\amd64\"<\/code><\/p>\n"},{"AnswerId":"33011235","CreationDate":"2015-10-08T09:03:23.100","ParentId":null,"OwnerUserId":"3079329","Title":null,"Body":"<p>Here are my simple steps for installing theano on a\n64-bit windows 10 machine. It's tested on the code listed <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#testing-theano-with-gpu\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>(All installation are with default installation path)<\/p>\n\n<ul>\n<li>install anaconda python 3.x distribution (it already includes numpy,\nscipy, matlibplot, etc.) <\/li>\n<li>run 'conda install mingw libpython' in command-line<\/li>\n<li>install theano by downloading it from the official website and do `python setup.py install'<\/li>\n<li>install lastest CUDA toolkit for 64-bit windows 10 (now is 7.5)<\/li>\n<li>install visual studio 2013 (free for windows 10)<\/li>\n<li>create .theanorc.txt file under %USERPROFILE% path and here are\nthe content    in the .theanorc.txt file to run theano with GPU<\/li>\n<\/ul>\n\n<hr>\n\n<p>[global]<\/p>\n\n<p>floatX = float32<\/p>\n\n<p>device = gpu<\/p>\n\n<p>[nvcc]<\/p>\n\n<p>fastmath = True<\/p>\n\n<p>compiler_bindir=C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin\\cl.exe<\/p>\n\n<p>[cuda]<\/p>\n\n<p>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5<\/p>\n\n<hr>\n"},{"AnswerId":"30539337","CreationDate":"2015-05-29T21:26:55.810","ParentId":null,"OwnerUserId":"4853662","Title":null,"Body":"<p>In case you want to upgrade to MS Visual Studio 2012 and CUDA 7 on Windows 8.1 x64, check out this tutorial here:<\/p>\n\n<p><a href=\"http:\/\/machinelearning.berlin\/?p=383\" rel=\"nofollow\">http:\/\/machinelearning.berlin\/?p=383<\/a><\/p>\n\n<p>It should work as long as you stick to it exactly.\nAll the best<\/p>\n\n<p>Christian<\/p>\n"},{"AnswerId":"26073714","CreationDate":"2014-09-27T10:30:52.703","ParentId":null,"OwnerUserId":"1663093","Title":null,"Body":"<p>Theano is a great tool for machine learning applications, yet I found that its installation on Windows is not trivial especially for beginners (like myself) in programming. In my case, I see 5-6x speedups of my scripts when run on a GPU so it was definitely worth the hassle. <\/p>\n\n<p>I wrote this guide based on my installation procedure and is meant to be verbose and hopefully complete even for people with no prior understanding of building programs under Windows environment. Most of this guide is based on these <a href=\"https:\/\/github.com\/Theano\/Theano\/wiki\/Windowsinstallation\" rel=\"nofollow noreferrer\">instructions<\/a> but I had to change some of the steps in order for it to work on my system. If there is anything that I do that may not be optimal or that doesn't work on your machine, please, let me know and I will try to modify this guide accordingly.  <\/p>\n\n<p>These are the steps (in order) I followed when installing Theano with GPU enabled on my Windows 8.1 machine:<\/p>\n\n<h3>CUDA Installation<\/h3>\n\n<p>CUDA can be downloaded from <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">here<\/a>. In my case, I chose 64-bit Notebook version for my NVIDIA Optimus laptop with Geforce 750m. <\/p>\n\n<p>Verify that your installation was successful by launching <code>deviceQuery<\/code> from command line. In my case this was located in the following folder: <code>C:\\ProgramData\\NVIDIA Corporation\\CUDA Samples\\v6.5\\bin\\win64\\Release<\/code> . If successful, you should see PASS at the end of the test. <\/p>\n\n<h3>Visual Studio 2010 Installation<\/h3>\n\n<p>I installed this via <a href=\"http:\/\/www.dreamspark.com\" rel=\"nofollow noreferrer\">dreamspark<\/a>. If you are a student you are entitled for a free version. If not, you can still install the <a href=\"http:\/\/www.visualstudio.com\/en-us\/products\/visual-studio-express-vs.aspx\" rel=\"nofollow noreferrer\">Express version<\/a> which should work just as well. After install is complete you should be able to call Visual Studio Command Prompt 2010 from the start menu. <\/p>\n\n<h3>Python Installation<\/h3>\n\n<p>At the time of writing, Theano on GPU only allows working with 32-bit floats and is primarily built for 2.7 version of Python. Theano requires most of the basic scientific Python libraries such as <code>scipy<\/code> and <code>numpy<\/code>. I found that the easiest way to install these was via <a href=\"http:\/\/sourceforge.net\/projects\/winpython\/files\/latest\/download?source=files\" rel=\"nofollow noreferrer\">WinPython<\/a>. It installs all the dependencies in a self-contained folder which allows easy reinstall if something goes wrong in the installation process and you get some useful IDE tools such as ipython notebook and Spyder installed for free as well. For ease of use you might want to add the path to your python.exe and path to your Scripts folder in the <a href=\"https:\/\/superuser.com\/questions\/284342\/what-are-path-and-other-environment-variables-and-how-can-i-set-or-use-them\">environment variables<\/a>.<\/p>\n\n<h3>Git installation<\/h3>\n\n<p>Found <a href=\"http:\/\/msysgit.github.io\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<h3>MinGW Installation<\/h3>\n\n<p>Setup file is <a href=\"http:\/\/www.mingw.org\/wiki\/Getting_Started\" rel=\"nofollow noreferrer\">here<\/a>. I checked all the base installation files during the installation process. This is required if you run into g++ error described below. <\/p>\n\n<h3>Cygwin installation<\/h3>\n\n<p>You can find it <a href=\"https:\/\/cygwin.com\/install.html\" rel=\"nofollow noreferrer\">here<\/a>. I basically used this utility only to extract PyCUDA tar file which is already provided in the base install (so the install should be straightforward). <\/p>\n\n<h3>Python distutils fix<\/h3>\n\n<p>Open <code>msvc9compiler.py<\/code> located in your <code>\/lib\/distutils\/<\/code> directory of your Python installation. Line 641 in my case reads: <code>ld_args.append ('\/IMPLIB:' + implib_file)<\/code>. Add the following after this line (same indentation): <\/p>\n\n<pre><code>ld_args.append('\/MANIFEST')\n<\/code><\/pre>\n\n<h3>PyCUDA installation<\/h3>\n\n<p>Source for PyCUDA is <a href=\"http:\/\/pypi.python.org\/packages\/source\/p\/pycuda\/pycuda-2012.1.tar.gz#md5=b67c4fce6c258834339073f2537fa84f\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p><strong>Steps:<\/strong><\/p>\n\n<p>Open cygwin and navigate to the PyCUDA folder (i.e. <code>\/cygdrive\/c\/etc\/etc<\/code>) and execute <code>tar -xzf pycuda-2012.1.tar.gz<\/code>.<\/p>\n\n<p>Open Visual Studio Command Prompt 2010 and navigate to the directory where tarball was extracted and execute <code>python configure.py<\/code><\/p>\n\n<p>Open the .\/siteconf.py and change the values so that it reads (for CUDA 6.5 for instance):<\/p>\n\n<pre><code>BOOST_INC_DIR = []\nBOOST_LIB_DIR = []\nBOOST_COMPILER = 'gcc43'\nUSE_SHIPPED_BOOST = True\nBOOST_PYTHON_LIBNAME = ['boost_python']\nBOOST_THREAD_LIBNAME = ['boost_thread']\nCUDA_TRACE = False\nCUDA_ROOT = 'C:\\\\Program Files\\\\NVIDIA GPU Computing Toolkit\\\\CUDA\\\\v6.5'\nCUDA_ENABLE_GL = False\nCUDA_ENABLE_CURAND = True\nCUDADRV_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32']\nCUDADRV_LIBNAME = ['cuda']\nCUDART_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32']\nCUDART_LIBNAME = ['cudart']\nCURAND_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32']\nCURAND_LIBNAME = ['curand']\nCXXFLAGS = ['\/EHsc']\nLDFLAGS = ['\/FORCE']\n<\/code><\/pre>\n\n<p>Execute the following commands at the VS2010 command prompt: <\/p>\n\n<pre><code>set VS90COMNTOOLS=%VS100COMNTOOLS%\npython setup.py build\npython setup.py install\n<\/code><\/pre>\n\n<p>Create this python file and verify that you get a result:<\/p>\n\n<pre><code># from: http:\/\/documen.tician.de\/pycuda\/tutorial.html\nimport pycuda.gpuarray as gpuarray\nimport pycuda.driver as cuda\nimport pycuda.autoinit\nimport numpy\na_gpu = gpuarray.to_gpu(numpy.random.randn(4,4).astype(numpy.float32))\na_doubled = (2*a_gpu).get()\nprint a_doubled\nprint a_gpu\n<\/code><\/pre>\n\n<h3>Install Theano<\/h3>\n\n<p>Open git bash shell and choose a folder in which you want to place Theano installation files and execute:<\/p>\n\n<pre><code>git clone git:\/\/github.com\/Theano\/Theano.git\npython setup.py install\n<\/code><\/pre>\n\n<p>Try opening python in VS2010 command prompt and run <code>import theano<\/code><\/p>\n\n<p>If you get a g++ related error, open MinGW msys.bat in my case installed here: <code>C:\\MinGW\\msys\\1.0<\/code> and try importing theano in MinGW shell. Then retry importing theano from VS2010 Command Prompt and it should be working now. <\/p>\n\n<p>Create a file in WordPad (NOT Notepad!), name it <code>.theanorc.txt<\/code> and put it in <code>C:\\Users\\Your_Name\\<\/code> or wherever your users folder is located: <\/p>\n\n<pre><code>#!sh\n[global]\ndevice = gpu\nfloatX = float32\n\n[nvcc]\ncompiler_bindir=C:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\bin\n# flags=-m32 # we have this hard coded for now\n\n[blas]\nldflags =\n# ldflags = -lopenblas # placeholder for openblas support\n<\/code><\/pre>\n\n<p>Create a test python script and run it:<\/p>\n\n<pre><code>from theano import function, config, shared, sandbox\nimport theano.tensor as T\nimport numpy\nimport time\n\nvlen = 10 * 30 * 768  # 10 x #cores x # threads per core\niters = 1000\n\nrng = numpy.random.RandomState(22)\nx = shared(numpy.asarray(rng.rand(vlen), config.floatX))\nf = function([], T.exp(x))\nprint f.maker.fgraph.toposort()\nt0 = time.time()\nfor i in xrange(iters):\n    r = f()\nt1 = time.time()\nprint 'Looping %d times took' % iters, t1 - t0, 'seconds'\nprint 'Result is', r\nif numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):\n    print 'Used the cpu'\nelse:\n    print 'Used the gpu'\n<\/code><\/pre>\n\n<p>Verify you got <code>Used the gpu<\/code> at the end and you're done! <\/p>\n"},{"AnswerId":"25918575","CreationDate":"2014-09-18T17:10:22.787","ParentId":null,"OwnerUserId":"1680562","Title":null,"Body":"<p>Here's a guide to installing theano with CUDA on 64-bit Windows.<\/p>\n\n<p>It seems straightforward, but I have not actually tested it to ensure that it works.<\/p>\n\n<p><a href=\"http:\/\/pavel.surmenok.com\/2014\/05\/31\/installing-theano-with-gpu-on-windows-64-bit\/\" rel=\"nofollow\">http:\/\/pavel.surmenok.com\/2014\/05\/31\/installing-theano-with-gpu-on-windows-64-bit\/<\/a><\/p>\n"}]}
{"QuestionId":25788268,"AnswerCount":1,"Tags":"<python-2.7><numpy><enthought><theano>","CreationDate":"2014-09-11T13:05:28.370","AcceptedAnswerId":25812219.0,"OwnerUserId":2688733.0,"Title":"Theano get unique values in a tensor","Body":"<p>I have a tensor which I convert into a vector by flattening, now I want to remove the duplicate values in this vector. How can I do this? What is equivalent for numpy.unique() in theano?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"25812219","CreationDate":"2014-09-12T15:54:16.097","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>EDIT: this is now available in Theano: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html#theano.tensor.extra_ops.Unique\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html#theano.tensor.extra_ops.Unique<\/a><\/p>\n\n<p>This question was also asked on theano-user mailing list. The conclusion is that this is one of the function NumPy function that isn't wrapped in Theano. As he don't need the grad, it can be rapidly wrapped. Here is an example who expect the outputs to be the same as the input.<\/p>\n\n<pre><code>from theano.compile.ops import as_op\n\n\n@as_op(itypes=[theano.tensor.imatrix],\n       otypes=[theano.tensor.imatrix])\ndef numpy_unique(a):\n    return numpy.unique(a)\n<\/code><\/pre>\n\n<p>More doc about as_op is available here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/extending_theano.html#as-op-example\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/extending_theano.html#as-op-example<\/a><\/p>\n"}]}
{"QuestionId":25881271,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-09-17T02:21:35.873","AcceptedAnswerId":null,"OwnerUserId":1392141.0,"Title":"Swapping rows of a Theano symbolic matrix","Body":"<p>I am implementing parallel tempering Gibbs sampling using Theano. I am trying to create a Theano function that takes a matrix  and swaps some of its rows. I have a symbolic binary vector named  that denotes which rows should be swapped (i.e., if , then  and  should be swapped). The order of swapping is not important for me.<\/p>\n\n<p>I was trying to write a  that goes through the  vector and performs swapping of  row-by-row. The problem is that Theano doesn't allow doing something like  with symbolic variables. Here is a simple code snippet of what I am trying to do.<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on how I can do this the right way?<\/p>\n","answers":[{"AnswerId":"25901769","CreationDate":"2014-09-17T23:09:29.807","ParentId":null,"OwnerUserId":"1392141","Title":null,"Body":"<p>Okay, here is a really simple solution I found.<\/p>\n\n<pre><code>import numpy as np\n\nimport theano\nimport theano.tensor as T\n\ndef swap(swp, pos, X):\n    return T.concatenate([X[:pos],X[[pos+swp]],X[[pos+1-swp]],X[pos+2:]])\n\nmax_length = 10\nswaps = T.ivector('swaps')\npos   = T.iscalar('pos')\nX     = T.vector('X')\n\nnew_X, _ = theano.scan(swap,\n                       sequences=[swaps, T.arange(max_length)],\n                       outputs_info=X)\n\ndo_swaps = theano.function([swaps, X], new_X[-1])\n\nX_swapped = do_swaps(np.array([1, 1, 0, 1], dtype='int32'), np.arange(5))\nprint X_swapped\n<\/code><\/pre>\n\n<p>However, I am not sure how it is optimal or not for executing on a GPU.<\/p>\n"}]}
{"QuestionId":25886374,"AnswerCount":2,"Tags":"<python><matlab><scipy><theano>","CreationDate":"2014-09-17T09:04:13.730","AcceptedAnswerId":26065509.0,"OwnerUserId":1714410.0,"Title":"pdist for theano tensor","Body":"<p>I have a theano symbolic matrix<\/p>\n\n<pre><\/pre>\n\n<p> will be later on populated by  vectors of dim  (at train time).<\/p>\n\n<p>I would like to have the theano equivalent of  (<a href=\"http:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.spatial.distance.pdist.html\" rel=\"noreferrer\"><\/a> of <a href=\"http:\/\/www.mathworks.com\/help\/stats\/pdist.html\" rel=\"noreferrer\"><\/a>), something like<\/p>\n\n<pre><\/pre>\n\n<p>How can I achieve this?  <\/p>\n\n<p>Calling  on  directly does not work as  at this stage is only symbolic...<\/p>\n\n<p><strong>Update:<\/strong> I would very much like to be able to mimic  \"compact\" behavior: that is, computing only ~1\/2 of the x entries of the distance matrix.<\/p>\n","answers":[{"AnswerId":"26065509","CreationDate":"2014-09-26T17:53:15.947","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p><code>pdist<\/code> from scipy is a collection of different functions - there doesn't exist a Theano equivalent for all of them at once. However, each specific distance, being a closed form mathematical expression, can be written down in Theano as such and then compiled.<\/p>\n\n<p>Take as a example the minkowski <code>p<\/code> norm distance (copy+pasteable):<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nX = T.fmatrix('X')\nY = T.fmatrix('Y')\nP = T.scalar('P')\ntranslation_vectors = X.reshape((X.shape[0], 1, -1)) - Y.reshape((1, Y.shape[0], -1))\nminkowski_distances = (abs(translation_vectors) ** P).sum(2) ** (1. \/ P)\nf_minkowski = theano.function([X, Y, P], minkowski_distances)\n<\/code><\/pre>\n\n<p>Note that <code>abs<\/code> calls the built-in <code>__abs__<\/code>, so <code>abs<\/code> is also a theano function. We can now compare this to <code>pdist<\/code>:<\/p>\n\n<pre><code>import numpy as np\nfrom scipy.spatial.distance import pdist\n\nrng = np.random.RandomState(42)\nd = 20 # dimension\nnX = 10\nnY = 30\nx = rng.randn(nX, d).astype(np.float32)\ny = rng.randn(nY, d).astype(np.float32)\n\nps = [1., 3., 2.]\n\nfor p in ps:\n    d_theano = f_minkowski(x, x, p)[np.triu_indices(nX, 1)]\n    d_scipy = pdist(x, p=p, metric='minkowski')\n    print \"Testing p=%1.2f, discrepancy %1.3e\" % (p, np.sqrt(((d_theano - d_scipy) ** 2).sum()))\n<\/code><\/pre>\n\n<p>This yields<\/p>\n\n<pre><code>Testing p=1.00, discrepancy 1.322e-06\nTesting p=3.00, discrepancy 4.277e-07\nTesting p=2.00, discrepancy 4.789e-07\n<\/code><\/pre>\n\n<p>As you can see, the correspondence is there, but the function <code>f_minkowski<\/code> is slightly more general, since it compares the lines of two possibly different arrays. If twice the same array is passed as input, <code>f_minkowski<\/code> returns a matrix, whereas <code>pdist<\/code> returns a list without redundancy. If this behaviour is desired, it can also be implemented fully dynamically, but I will stick to the general case here.<\/p>\n\n<p>One possibility of specialization should be noted though: In the case of <code>p=2<\/code>, the calculations become simpler through the binomial formula, and this can be used to save precious space in memory: Whereas the general Minkowski distance, as implemented above, creates a 3D array (due to avoidance of for-loops and summing cumulatively), which is prohibitive, depending on the dimension <code>d<\/code> (and <code>nX, nY<\/code>), for <code>p=2<\/code> we can write<\/p>\n\n<pre><code>squared_euclidean_distances = (X ** 2).sum(1).reshape((X.shape[0], 1)) + (Y ** 2).sum(1).reshape((1, Y.shape[0])) - 2 * X.dot(Y.T)\nf_euclidean = theano.function([X, Y], T.sqrt(squared_euclidean_distances))\n<\/code><\/pre>\n\n<p>which only uses <code>O(nX * nY)<\/code> space instead of <code>O(nX * nY * d)<\/code> We check for correspondence, this time on the general problem:<\/p>\n\n<pre><code>d_eucl = f_euclidean(x, y)\nd_minkowski2 = f_minkowski(x, y, 2.)\nprint \"Comparing f_minkowski, p=2 and f_euclidean: l2-discrepancy %1.3e\" % ((d_eucl - d_minkowski2) ** 2).sum()\n<\/code><\/pre>\n\n<p>yielding<\/p>\n\n<pre><code>Comparing f_minkowski, p=2 and f_euclidean: l2-discrepancy 1.464e-11\n<\/code><\/pre>\n"},{"AnswerId":"25960161","CreationDate":"2014-09-21T14:21:24.703","ParentId":null,"OwnerUserId":"97160","Title":null,"Body":"<p>I haven't worked with Theano before, but here is a solution based on pure Numpy functions (perhaps you convert it to the equivalent theano functions. Note that I'm using automatic <em>broadcasting<\/em> in the expression below, so you might have to rewrite that explicitly if Theano doesn't support it):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># X is an m-by-n matrix (rows are examples, columns are dimensions)\n# D is an m-by-m symmetric matrix of pairwise Euclidean distances\na = np.sum(X**2, axis=1)\nD = np.sqrt((a + a[np.newaxis].T) - 2*np.dot(X, X.T))\n<\/code><\/pre>\n\n<p>It is based on the fact that: <code>||u-v||^2 = ||u||^2 + ||v||^2 - 2*u.v<\/code>. (I showed this in <a href=\"https:\/\/stackoverflow.com\/a\/4171845\/97160\">previous<\/a> <a href=\"https:\/\/stackoverflow.com\/a\/7774323\/97160\">answers<\/a> of mine using MATLAB)<\/p>\n\n<p>Here is a comparison against Scipy existing functions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom scipy.spatial.distance import pdist, squareform\n\ndef my_pdist(X):\n    a = np.sum(X**2, axis=1)\n    D = np.sqrt((a + a[np.newaxis].T) - 2*np.dot(X, X.T))\n    return D\n\ndef scipy_pdist(X):\n    D = squareform(pdist(X, metric='euclidean'))\n    return D    \n\nX = np.random.rand(5, 3)\nD1 = my_pdist(X)\nD2 = scipy_pdist(X)\n<\/code><\/pre>\n\n<p>The difference should be negligible, close to machine epsilon (<code>np.spacing(1)<\/code>):<\/p>\n\n<pre><code>&gt;&gt;&gt; np.linalg.norm(D1-D2)\n8.5368137554718277e-16\n<\/code><\/pre>\n\n<p>HTH<\/p>\n\n<hr>\n\n<h1>EDIT:<\/h1>\n\n<p>Here is another implementation with a single loop:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def my_pdist_compact(X):\n    D = np.empty(shape=[0,0], dtype=X.dtype)\n    for i in range(X.shape[0]-1):\n        D = np.append(D, np.sqrt(np.sum((X[i,] - X[i+1:,])**2, axis=1)))\n    return D\n<\/code><\/pre>\n\n<p>Somewhat equivalent MATLAB code:<\/p>\n\n<pre class=\"lang-matlab prettyprint-override\"><code>function D = my_pdist_compact(X)\n    n = size(X,1);\n    D = cell(n-1,1);\n    for i=1:n-1\n        D{i} = sqrt(sum(bsxfun(@minus, X(i,:), X(i+1:end,:)).^2, 2));\n    end\n    D = vertcat(D{:});\nend\n<\/code><\/pre>\n\n<p>This returns the pairwise-distances in compact form (upper triangular part of the symmetric matrix). This is the same output as <code>pdist<\/code>. Use <code>squareform<\/code> to convert it to full matrix.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; d1 = my_pdist_compact(X)\n&gt;&gt;&gt; d2 = pdist(X)    # from scipy.spatial.distance\n&gt;&gt;&gt; (d1 == d2).all()\nTrue\n<\/code><\/pre>\n\n<p>I will leave it to you to see if it's possible to write the equivalent <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loop.html\" rel=\"nofollow noreferrer\">loop<\/a> using Theano (see <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow noreferrer\"><code>theano.scan<\/code><\/a>)!<\/p>\n"}]}
{"QuestionId":25899187,"AnswerCount":1,"Tags":"<pycuda><theano><deep-learning>","CreationDate":"2014-09-17T19:51:47.827","AcceptedAnswerId":25903319.0,"OwnerUserId":1754197.0,"Title":"pycuda vs theano vs pylearn2","Body":"<p>I am currently learning programming with GPU to improve the performance of machine learning algorithms. Initially I try to learn programming cuda with pure c, then I found pycuda which to me a wrapper of cuda library, and then I found theano and pylearn2 and got a little confused:<\/p>\n\n<p>I understand them in this way:<\/p>\n\n<ol>\n<li>pycuda: python wrapper for cuda library<\/li>\n<li>theano: similar to numpy but transparent to GPU and CPU<\/li>\n<li>pylearn2: deep learning package which build on theano and implemented several machine learning\/deep learning model<\/li>\n<\/ol>\n\n<p>Since I am new to GPU programming, should I start learning from C\/C++ implementation or starting from pycuda is enough, even starting from theano? E.g. I would like to implement randomForest model after learning the GPU programming.Thanks.<\/p>\n","answers":[{"AnswerId":"25903319","CreationDate":"2014-09-18T02:37:08.770","ParentId":null,"OwnerUserId":"534969","Title":null,"Body":"<p>Your understand is almost right. I would just add some remarks about Theano. It's much more than a Numpy which can run on the GPU. Theano is indeed is math expression compiler, which translates symbolic math expressions in highly optimized C\/CUDA code, targeted for both CPU and GPU. The code it generates is often much more efficient than the one most programmers would write. Theano also can make symbolic differentiation (very useful for gradient based optimization) and has also a feature to achieve better numerical stability (which probably is something useful, though I don't know for real to what extent). It's very likely Theano will be enough to implement what you need. If you still decide to learn CUDA or PyCUDA, choose the one based no the language you will use, C++ or Python. <\/p>\n"}]}
{"QuestionId":25962554,"AnswerCount":2,"Tags":"<python><yaml><theano><unsupervised-learning><autoencoder>","CreationDate":"2014-09-21T18:37:33.107","AcceptedAnswerId":null,"OwnerUserId":3670532.0,"Title":"Getting the learned representation of the data from the unsupervised learning in pylearn2","Body":"<p>We can train an autoencoder in pylearn2 using below YAML file (along with pylearn2\/scripts\/train.py)<\/p>\n\n<pre><\/pre>\n\n<p>What we get is the learned autoencoder model as \"dae_l1.pkl\".<\/p>\n\n<p>If I want to use this model for supervised training, I can use \"dae_l1.pkl\" to initialize the layer of an MLP. I can then train this model.\nI can even predict the output of the model using 'fprop' function.<\/p>\n\n<p>But what if I dun want to use this pretrained model for supervised learning and I just want to save the new learned representation of my data with the autoencoder.<\/p>\n\n<p>How can I do this?<\/p>\n\n<p>Even more detailed question is put <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/pylearn-users\/8FsuAts9PoA\" rel=\"nofollow\">here<\/a><\/p>\n","answers":[{"AnswerId":"31551779","CreationDate":"2015-07-22T00:14:25.513","ParentId":null,"OwnerUserId":"674069","Title":null,"Body":"<p>I think you can use the encode and decode functions of the autoencoder to get the hidden representation. E.g:<\/p>\n\n<pre><code>l1_path = 'dae_l1.pkl'\nl1 = serial.load(l1_path)\n\"\"\"encode\"\"\"\n#layer 1\nl1Input = l1.get_input_space().make_theano_batch()\nl1Encode = l1.encode(l1Input)\nl1Decode = l1.decode(l1Encode)\nl1EncodeFunction = theano.function([l1Input], l1Encode)\nl1DecodeFunction = theano.function([l1Encode], l1Decode)\n<\/code><\/pre>\n\n<p>Then, the representation will be:<\/p>\n\n<pre><code>l1encode = l1EncodeFunction(YourData)\n<\/code><\/pre>\n"},{"AnswerId":"25997299","CreationDate":"2014-09-23T14:09:31.240","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>The <code>reconstruct<\/code> method of the pickled model should do it - I believe usage is the same as <code>fprop<\/code>.<\/p>\n"}]}
{"QuestionId":25986855,"AnswerCount":1,"Tags":"<python><pymc><theano>","CreationDate":"2014-09-23T04:22:04.747","AcceptedAnswerId":null,"OwnerUserId":1559693.0,"Title":"PyMC + Theano on Debian Backports","Body":"<p>I'm attempting to run a model in pymc3 that takes advantage of  when performing the dot product in a multilevel model.  However, when I attempt to import theano I get:<\/p>\n\n<pre><\/pre>\n\n<p>I'm running .<\/p>\n\n<p>Conda:<\/p>\n\n<p>Current conda install:<\/p>\n\n<pre><\/pre>\n\n<p>I have also tried, after looking at various problem:<\/p>\n\n<pre><\/pre>\n\n<p>This seg faults.<\/p>\n\n<p>Also, it appears that I have the blas directories in .<\/p>\n\n<pre><\/pre>\n\n<p>Edit:<\/p>\n\n<p>The cause of this also breaks  in pymc3.<\/p>\n","answers":[{"AnswerId":"26678427","CreationDate":"2014-10-31T15:41:21.283","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I suppose you have an old version of Theano, as now if we we can't find or detect how to link again blas, we have a fallback that do not cause crash, but cause only a warning.<\/p>\n\n<p>Update Theano to the development version.<\/p>\n\n<p>If you installed it via conda, remove it first, the update:<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git --user\n<\/code><\/pre>\n\n<p>More info on how to update to the development version:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n"}]}
{"QuestionId":26009101,"AnswerCount":2,"Tags":"<python><classification><theano><deep-learning>","CreationDate":"2014-09-24T05:18:58.297","AcceptedAnswerId":26049622.0,"OwnerUserId":956730.0,"Title":"How to read values of Theano's LogisticRegression class predictions within DeepLearningTutorials package?","Body":"<p>I am using <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">Theano's LogisticRegression sample code<\/a> and i have not modified the code in the given package at all and i'm using the same data.<\/p>\n\n<p>I need to read the values of prediction which are within (self.y_pred) field in LogisticRegression class and also the values of prediction probabilities which are in self.p_y_given_x field of the same class.<\/p>\n\n<p>They are tensortype and tensorvariables and i don't know how to read\/print them. I need them to do a postprocessing but i can't get access to the values. The values should be read after training which should be here around the star charachters.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"26070601","CreationDate":"2014-09-27T02:08:27.450","ParentId":null,"OwnerUserId":"956730","Title":null,"Body":"<p>This is the code that worked for me as answered by Kyle. It returns the values of the prediction classes and i print a report out of it.<\/p>\n\n<pre><code>classifier = LogisticRegression(input=x, n_in=train_set_x.get_value(borrow=True).shape[1], n_out=25)\n\n# the cost we minimize during training is the negative log likelihood of\n# the model in symbolic format\ncost = classifier.negative_log_likelihood(y)\n\n# compiling a Theano function that computes the mistakes that are made by\n# the model on a minibatch\ntest_model = theano.function(inputs=[index],\n        outputs=classifier.errors(y),\n        givens={\n            x: test_set_x[index * batch_size: (index + 1) * batch_size],\n            y: test_set_y[index * batch_size: (index + 1) * batch_size]})\n\nvalidate_model = theano.function(inputs=[index],\n        outputs=classifier.errors(y),\n        givens={\n            x: valid_set_x[index * batch_size:(index + 1) * batch_size],\n            y: valid_set_y[index * batch_size:(index + 1) * batch_size]})\n\npredict = theano.function(inputs=[],\n        outputs=classifier.y_pred,\n        givens={\n            x: test_set_x})\n# compute the gradient of cost with respect to theta = (W,b)\ng_W = T.grad(cost=cost, wrt=classifier.W)\ng_b = T.grad(cost=cost, wrt=classifier.b)\n\n# specify how to update the parameters of the model as a list of\n# (variable, update expression) pairs.\nupdates = [(classifier.W, classifier.W - learning_rate * g_W),\n           (classifier.b, classifier.b - learning_rate * g_b)]\n\n# compiling a Theano function `train_model` that returns the cost, but in\n# the same time updates the parameter of the model based on the rules\n# defined in `updates`\ntrain_model = theano.function(inputs=[index],\n        outputs=cost,\n        updates=updates,\n        givens={\n            x: train_set_x[index * batch_size:(index + 1) * batch_size],\n            y: train_set_y[index * batch_size:(index + 1) * batch_size]})\n\n###############\n# TRAIN MODEL #\n###############\nprint '... training the model'\n# early-stopping parameters\npatience = 50000  # look as this many examples regardless\npatience_increase = 2  # wait this much longer when a new best is\n                              # found\nimprovement_threshold = 0.995  # a relative improvement of this much is\n                              # considered significant\nvalidation_frequency = min(n_train_batches, patience \/ 2)\n                              # go through this many\n                              # minibatche before checking the network\n                              # on the validation set; in this case we\n                              # check every epoch\n\nbest_params = None\nbest_validation_loss = numpy.inf\ntest_score = 0.\nstart_time = time.clock()\n\ndone_looping = False\nepoch = 0\nwhile (epoch &lt; n_epochs) and (not done_looping):\n    epoch = epoch + 1\n    #********************here i call the function and report based on returned class predictions.\n    report(predict())\n<\/code><\/pre>\n"},{"AnswerId":"26049622","CreationDate":"2014-09-25T23:14:04.597","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>You need to compile a function that gives back the prediction.<\/p>\n\n<p>This code may not work exactly, but this is the idea:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\n# Create some data with 100 samples, 10 features \nX = np.random.randn(100, 10)\nX_sym = T.fmatrix('X')\n# Create prediction function\npredict_function = theano.function(inputs=[X_sym], outputs=self.y_pred)\n\n# See the actual prediction\nprint(predict_function(X))\n<\/code><\/pre>\n"}]}
{"QuestionId":26091467,"AnswerCount":1,"Tags":"<github><theano>","CreationDate":"2014-09-29T01:45:10.693","AcceptedAnswerId":26091623.0,"OwnerUserId":419116.0,"Title":"Getting a particular version of a branch","Body":"<p>Is there a way to download a particular version of a branch?<\/p>\n\n<p>In particular I'd like to do a git clone of <a href=\"https:\/\/github.com\/Theano\/Theano\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano<\/a> now, and save a set of instructions on how to get the exact same version from github, regardless of future commits.<\/p>\n","answers":[{"AnswerId":"26091623","CreationDate":"2014-09-29T02:10:25.777","ParentId":null,"OwnerUserId":"1144203","Title":null,"Body":"<p><strong>UPDATE<\/strong><\/p>\n\n<p>There is an easier way to do this on github if no further changes are expected. In github, you can navigate to the 'tree view' of a repository from your browser via the URL<\/p>\n\n<pre><code>https:\/\/github.com\/&lt;repo_name&gt;\/tree\/&lt;commit_sha&gt;\n<\/code><\/pre>\n\n<p>Clicking on the 'Download ZIP' button on the right-hand navigation bar will download the codes of the repository up to that particular commit.<\/p>\n\n<p><strong>ORIGINAL ANSWER<\/strong><\/p>\n\n<p>I think one way to make your user's (or whoever will be reading your instructions) life easier is to clone the entire repository as-is. Then if the current <code>HEAD<\/code> commit is the version you want and you don't plan on making\/merging changes to your repository, you can just refer your user to this new repository; probably tagging it will be a good idea.<\/p>\n\n<p>Otherwise, you can create a branch (and a tag) in your new repository with the specific commit by doing:<\/p>\n\n<pre><code>$ git checkout -b new_branch commit_sha \/\/ where commit_sha points to the version you wanted\n<\/code><\/pre>\n\n<p>Then you can refer your user to this new branch (or tag) in your repository, after they have cloned your repository. <\/p>\n\n<p>I don't think there is a simple way to clone the <a href=\"https:\/\/github.com\/Theano\/Theano\" rel=\"nofollow\">original repository<\/a> from an old commit without using <code>git reset<\/code>.<\/p>\n"}]}
{"QuestionId":26107927,"AnswerCount":3,"Tags":"<python><theano><pickle><mnist><dbn>","CreationDate":"2014-09-29T20:04:09.283","AcceptedAnswerId":31861263.0,"OwnerUserId":4092486.0,"Title":"How to put my dataset in a .pkl file in the exact format and data structure used in \"mnist.pkl.gz\"?","Body":"<p>I'm trying to use the Theano library in python to do some experiments with Deep Belief Networks. I use the code in this address: <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/DBN.py\" rel=\"noreferrer\">DBN full code<\/a>. This code use the <a href=\"http:\/\/deeplearning.net\/data\/mnist\/mnist.pkl.gz\" rel=\"noreferrer\">MNIST Handwritten database<\/a>. This file is already in pickle format. \nIt is unpicked in: <\/p>\n\n<ul>\n<li>train_set<\/li>\n<li>valid_set<\/li>\n<li>test_set<\/li>\n<\/ul>\n\n<p>Which is further unpickled in:<\/p>\n\n<ul>\n<li>train_set_x, train_set_y = train_set <\/li>\n<li>valid_set_x, valid_set_y = valid_set<\/li>\n<li>test_set_x, test_set_y = test_set<\/li>\n<\/ul>\n\n<p>Please can someone give me the code that constructs this dataset in order to create my own?\nThe DBN example I use needs the data in this format and I don't know how to do it.\nif anyone has any ideas how to fix this, please tell me.<\/p>\n\n<p><strong>Here is my code:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"26228673","CreationDate":"2014-10-07T04:42:16.730","ParentId":null,"OwnerUserId":"309653","Title":null,"Body":"<p>The pickled file represents a tuple of 3 lists : the training set, the validation set and the testing set. (train, val, test)<\/p>\n\n<ul>\n<li>Each of the three lists is a pair formed from a list of images and a list of class labels for each of the images. <\/li>\n<li>An image is represented as numpy 1-dimensional array of 784 (28 x 28) float values between 0 and 1 (0 stands for black, 1 for white). <\/li>\n<li>The labels are numbers between 0 and 9 indicating which digit the image represents. <\/li>\n<\/ul>\n"},{"AnswerId":"31861263","CreationDate":"2015-08-06T16:39:06.013","ParentId":null,"OwnerUserId":"3497273","Title":null,"Body":"<p>A .pkl file is not necessary to adapt code from the Theano tutorial to your own data. You only need to mimic their data structure.<\/p>\n\n<h1>Quick fix<\/h1>\n\n<p>Look for the following lines. It's line 303 on <strong>DBN.py<\/strong>.<\/p>\n\n<pre><code>datasets = load_data(dataset)\ntrain_set_x, train_set_y = datasets[0]\n<\/code><\/pre>\n\n<p>Replace with your own <code>train_set_x<\/code> and <code>train_set_y<\/code>.<\/p>\n\n<pre><code>my_x = []\nmy_y = []\nwith open('path_to_file', 'r') as f:\n    for line in f:\n        my_list = line.split(' ') # replace with your own separator instead\n        my_x.append(my_list[1:-1]) # omitting identifier in [0] and target in [-1]\n        my_y.append(my_list[-1])\ntrain_set_x = theano.shared(numpy.array(my_x, dtype='float64'))\ntrain_set_y = theano.shared(numpy.array(my_y, dtype='float64'))\n<\/code><\/pre>\n\n<p>Adapt this to your input data and the code you're using.<\/p>\n\n<p>The same thing works for <strong>cA.py<\/strong>, <strong>dA.py<\/strong> and <strong>SdA.py<\/strong> but they only use <code>train_set_x<\/code>.<\/p>\n\n<p>Look for places such as <code>n_ins=28 * 28<\/code> where mnist image sizes are hardcoded. Replace <code>28 * 28<\/code> with your own number of columns.<\/p>\n\n<h1>Explanation<\/h1>\n\n<p>This is where you put your data in a format that Theano can work with.<\/p>\n\n<pre><code>train_set_x = theano.shared(numpy.array(my_x, dtype='float64'))\ntrain_set_y = theano.shared(numpy.array(my_y, dtype='float64'))\n<\/code><\/pre>\n\n<p><code>shared()<\/code> turns a numpy array into the Theano format designed for efficiency on GPUs.<\/p>\n\n<p><code>dtype='float64'<\/code> is expected in Theano arrays.<\/p>\n\n<p>More details on <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"noreferrer\">basic tensor functionality<\/a>.<\/p>\n\n<h1>.pkl file<\/h1>\n\n<p>The .pkl file is a way to save your data structure.<\/p>\n\n<p>You can create your own.<\/p>\n\n<pre><code>import cPickle\nf = file('my_data.pkl', 'wb')\n    cPickle.dump((train_set_x, train_set_y), f, protocol=cPickle.HIGHEST_PROTOCOL)\nf.close()\n<\/code><\/pre>\n\n<p>More details on <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loading_and_saving.html\" rel=\"noreferrer\">loading and saving<\/a>.<\/p>\n"},{"AnswerId":"27614025","CreationDate":"2014-12-23T04:19:04.490","ParentId":null,"OwnerUserId":"2399329","Title":null,"Body":"<p>This can help:<\/p>\n\n<pre><code>from PIL import Image\nfrom numpy import genfromtxt\nimport gzip, cPickle\nfrom glob import glob\nimport numpy as np\nimport pandas as pd\nData, y = dir_to_dataset(\"trainMNISTForm\\\\*.BMP\",\"trainLabels.csv\")\n# Data and labels are read \n\ntrain_set_x = Data[:2093]\nval_set_x = Data[2094:4187]\ntest_set_x = Data[4188:6281]\ntrain_set_y = y[:2093]\nval_set_y = y[2094:4187]\ntest_set_y = y[4188:6281]\n# Divided dataset into 3 parts. I had 6281 images.\n\ntrain_set = train_set_x, train_set_y\nval_set = val_set_x, val_set_y\ntest_set = test_set_x, val_set_y\n\ndataset = [train_set, val_set, test_set]\n\nf = gzip.open('file.pkl.gz','wb')\ncPickle.dump(dataset, f, protocol=2)\nf.close()\n<\/code><\/pre>\n\n<p>This is the function I used. May change according to your file details.<\/p>\n\n<pre><code>def dir_to_dataset(glob_files, loc_train_labels=\"\"):\n    print(\"Gonna process:\\n\\t %s\"%glob_files)\n    dataset = []\n    for file_count, file_name in enumerate( sorted(glob(glob_files),key=len) ):\n        image = Image.open(file_name)\n        img = Image.open(file_name).convert('LA') #tograyscale\n        pixels = [f[0] for f in list(img.getdata())]\n        dataset.append(pixels)\n        if file_count % 1000 == 0:\n            print(\"\\t %s files processed\"%file_count)\n    # outfile = glob_files+\"out\"\n    # np.save(outfile, dataset)\n    if len(loc_train_labels) &gt; 0:\n        df = pd.read_csv(loc_train_labels)\n        return np.array(dataset), np.array(df[\"Class\"])\n    else:\n        return np.array(dataset)\n<\/code><\/pre>\n"}]}
{"QuestionId":26265552,"AnswerCount":1,"Tags":"<python-2.7><machine-learning><theano>","CreationDate":"2014-10-08T20:10:39.467","AcceptedAnswerId":null,"OwnerUserId":1430829.0,"Title":"Converting theano tensor types","Body":"<p>I have a computation graph built with Theano. It goes like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now, the mapping from a vector to a vector. However, input is set as a matrix type so I can pass many vectors through the mapping simultaneously. I'm doing some machine learning and this makes the learning phase more efficient.<\/p>\n\n<p>The problem is that after the learning phase, I'd like to view the mapping as vector to vector so I can compute:<\/p>\n\n<pre><\/pre>\n\n<p> complains that input is not . Is there a way I can change the input tensor type without rebuilding the whole computation graph?<\/p>\n","answers":[{"AnswerId":"42166488","CreationDate":"2017-02-10T18:42:00.577","ParentId":null,"OwnerUserId":"3583290","Title":null,"Body":"<p>Technically, this a possible solution:<\/p>\n\n<pre><code>import theano\nfrom theano import tensor as T\nimport numpy as np\n\nW1 = theano.shared( np.random.rand(45,32).astype('float32'), 'W1')\nb1 = theano.shared( np.random.rand(32).astype('float32'), 'b1')\nW2 = theano.shared( np.random.rand(32,3).astype('float32'), 'W2')\nb2 = theano.shared( np.random.rand(3).astype('float32'), 'b2')\n\ninput  = T.vector('input') # it will be reshaped!\nhidden = T.tanh(T.dot(input.reshape((-1, 45)), W1)+b1)\noutput = T.nnet.softmax(T.dot(hidden, W2)+b2)\n\n#Here comes the trick\njac = theano.gradient.jacobian(output.reshape((-1,)), wrt=input).reshape((-1, 45, 3))\n<\/code><\/pre>\n\n<p>In this way <code>jac.eval({input: np.random.rand(10*45)}).shape<\/code> will result <code>(100, 45, 3)<\/code>!<\/p>\n\n<p>The problem is that it calculates the derivative across the batch index. So in theory the first <code>1x45<\/code> number can effect all the <code>10x3<\/code> outputs (in a batch of length 10).<\/p>\n\n<p>For that, there are several solutions.\nYou could take the diagonal across the first two axes, but unfortunately <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nlinalg.py#L193\" rel=\"nofollow noreferrer\">Theano does not implement it<\/a>, <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.diagonal.html\" rel=\"nofollow noreferrer\">numpy does<\/a>!<\/p>\n\n<p>I think it can be done with a <code>scan<\/code>, but this is an other matter.<\/p>\n"}]}
{"QuestionId":26301491,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2014-10-10T14:14:39.973","AcceptedAnswerId":null,"OwnerUserId":1137860.0,"Title":"Which is better for use Theano on Linux or Windows?","Body":"<p>Which is better for use Theano on Linux or Windows? I want to try some deep learning methods by Theano. <\/p>\n","answers":[{"AnswerId":"30217062","CreationDate":"2015-05-13T13:56:58.073","ParentId":null,"OwnerUserId":"1373401","Title":null,"Body":"<p>I use Theano on windows and once you use a package like enthought canopy or anaconda I find it easy to install and use. Although like a lot of this stuff on windows it is probably easier on Linux! :) <\/p>\n"},{"AnswerId":"26306031","CreationDate":"2014-10-10T18:45:09.220","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>It works fine on both, but I find better support for CUDA and much it is easier to install the scientific Python stack on Linux (usually Ubuntu).<\/p>\n"}]}
{"QuestionId":26304097,"AnswerCount":2,"Tags":"<python><python-2.7><numpy><theano>","CreationDate":"2014-10-10T16:37:22.723","AcceptedAnswerId":null,"OwnerUserId":596046.0,"Title":"Using a shared varibale to select from a matrix","Body":"<p>I'm trying to select multiple elements from a matrix using a changing vector, which is sent to my Theano function at each iteration. But when I try running the code I get the following error for the last line (selecting from W)<\/p>\n\n<pre><\/pre>\n\n<p>The declaration is:<\/p>\n\n<pre><\/pre>\n\n<p>and then I use my input (x) as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The intention is that at run time x will be a simple vector of indices of the form<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for the help<\/p>\n","answers":[{"AnswerId":"26304514","CreationDate":"2014-10-10T17:03:43.333","ParentId":null,"OwnerUserId":"736578","Title":null,"Body":"<p>The matrix W is of numpy ndarray type, so it does not know how to deal with Theano tensors such as x. If you want to index this numpy array, use a numpy index array instead of a Theano tensor. If you want to index the Theano tensor self.W, you will have to wait for the next release of Theano, or update to the development version: the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#indexing\" rel=\"nofollow\">theano manual<\/a> says that:<\/p>\n\n<blockquote>\n  <p>Like NumPy, Theano distinguishes between basic and advanced indexing. Theano fully supports basic indexing (see NumPy\u2019s indexing).<\/p>\n  \n  <p>Integer advanced indexing will be supported in 0.6rc4 (or the development version). We do not support boolean masks, as Theano does not have a boolean type (we use int8 for the output of logic operators).<\/p>\n<\/blockquote>\n"},{"AnswerId":"26306013","CreationDate":"2014-10-10T18:43:48.190","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>You may be able to use <code>inc_subtensor<\/code> or <code>set_subtensor<\/code> for this type of behavior, instead of slicing. See <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.set_subtensor\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.set_subtensor<\/a> for more details.<\/p>\n"}]}
{"QuestionId":26367114,"AnswerCount":0,"Tags":"<theano><pymc3><dirichlet>","CreationDate":"2014-10-14T17:50:42.463","AcceptedAnswerId":null,"OwnerUserId":2321303.0,"Title":"PyMC3 Dirichlet distribution","Body":"<p>I am implementing a linear regression model in pymc3 where the unknown vector of weights is constrained to be a probability mass function, hence modelled as a Dirichlet distribution, as in the following code:<\/p>\n\n<pre><\/pre>\n\n<p>After sampling the posterior by running:<\/p>\n\n<pre><\/pre>\n\n<p>I analysed the trace of the Dirichlet variables, and found that their values do not add to one (below is an example):<\/p>\n\n<pre><\/pre>\n\n<p>I am not familiar with theano variables, and found it difficult to explore how a Dirichlet RV is expressed in pymc3... Am I doing anything wrong, or should I just normalise the values returned in the trace so that they sum to one?<\/p>\n\n<p><strong>Quick update<\/strong>\nIt looks like the function  employes a sort of gradient descent optimisation. This does not take into account the constraint resulting from the fact that a vector representing a draw from a Dirichlet distribution is a probability mass function (its values should be positive and their sum should be one). This constraint is apparently not enforced at the sampling stage of the algorithm either, and causes convergence problems as the precision of the likelihood distribution drifts towards zero.<\/p>\n","answers":[]}
{"QuestionId":26372698,"AnswerCount":0,"Tags":"<torch><cudnn><cuda-gdb>","CreationDate":"2014-10-15T00:44:49.603","AcceptedAnswerId":null,"OwnerUserId":1386335.0,"Title":"Script works only when run in cuda-memcheck","Body":"<p>I'm writing a convnet using torch and cudnn and having some memory issues.\nI tried debugging the script with cuda-memcheck only to notice it actually runs when fed through cuda-memcheck (albeit slower than it should).<\/p>\n\n<p>Turns out if cuda-memcheck is running in the background, a separate instantiation of the script itself also runs fine.<\/p>\n\n<p>Any idea what might be happening here?<\/p>\n","answers":[]}
{"QuestionId":26387625,"AnswerCount":2,"Tags":"<linux><git><theano>","CreationDate":"2014-10-15T16:35:51.397","AcceptedAnswerId":26388618.0,"OwnerUserId":3681744.0,"Title":"Installing Theano from Git","Body":"<p>When I want to install pylearn2 (and Theano) I use the following command in Linux cluster:<\/p>\n\n<pre><\/pre>\n\n<p>However, I see the following error for Theano instllation:<\/p>\n\n<pre><\/pre>\n\n<p>I did not find anything on the net regarding this problem. I would appreciate if someone could help me.<\/p>\n","answers":[{"AnswerId":"26388618","CreationDate":"2014-10-15T17:35:35.183","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You can install manually the development version of Theano. From this link:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git --user\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>If for some reason that don't work, clone the repo and install it manually with <code>pip install .<\/code> in the repo.<\/p>\n"},{"AnswerId":"32239995","CreationDate":"2015-08-27T03:03:15.763","ParentId":null,"OwnerUserId":"3966933","Title":null,"Body":"<p>Theano has the following requirements make sure they are met: <\/p>\n\n<pre><code>1) Python &gt;= 2.6\n2) g++, python-dev\n3) NumPy &gt;= 1.6.2\n4) SciPy &gt;= 0.11\n5) A BLAS installation (with Level 3 functionality)\n<\/code><\/pre>\n\n<p>Then lets  not complicate things and just try<\/p>\n\n<blockquote>\n  <p>pip install Theano<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":26417524,"AnswerCount":2,"Tags":"<lua><luajit><torch><zerobrane>","CreationDate":"2014-10-17T03:43:14.287","AcceptedAnswerId":27289841.0,"OwnerUserId":936332.0,"Title":"How to set the environment variable of zerobrane studio","Body":"<p>I install all <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow\">torch<\/a> package into my local file torch-distro(Followed by this <a href=\"https:\/\/github.com\/soumith\/torch-distro\" rel=\"nofollow\">tutorial<\/a>). I want to use  to debug my code.  can't find my local path of torch. How Can I set my local path to the  environment variable.<br>\nI tried to add path.lua = \"${prefix}\/torch-distro\/install\/bin\/luajit\" into the user.lua. But it can't work<\/p>\n","answers":[{"AnswerId":"26418570","CreationDate":"2014-10-17T05:44:08.903","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>(These instructions are for the Windows version of Torch, but the steps should work for Linux\/OSX versions assuming the paths are modified).<\/p>\n\n<p>Let's say the Torch is installed in <code>C:\\Program Files\\Torch<\/code>, then to get it running as the external interpreter from ZeroBrane Studio (ZBS), you need to add <code>path.lua=[[C:\\Program Files\\Torch\\bin\\torch-lua]]<\/code> to <code>&lt;ZBS&gt;\\cfg\\user.lua<\/code> configuration file.<\/p>\n\n<p>Now, when you execute a Lua script from ZBS (<code>Project | Run<\/code> or <code>F6<\/code>), it will run inside the Torch environment:<\/p>\n\n<pre><code>local torch = require 'torch'\nlocal data = torch.Tensor{\n   {68, 24, 20},\n   {74, 26, 21},\n   {80, 32, 24}\n}\nprint(data)\n<\/code><\/pre>\n\n<p>However, there are few more steps required to get the debugging to work on Windows (these steps are likely not be needed on other systems, but I haven't tested debugging there). ZBS is using luasocket, which is compiled against <code>lua51.dll<\/code>, but Torch is using <code>libtorch-lua.dll<\/code>, so loading luasocket into your (Torch) process is likely to crash it. To make it work, you need to build a proxy DLL and put it into your <code>Torch\/bin<\/code> folder.<\/p>\n\n<p>To build the proxy DLL, you will need Visual Studio C++ or mingw\/gcc compiled and can follow these steps:<\/p>\n\n<ol>\n<li>Get <code>mkforwardlib.lua<\/code> (VS) or <code>mkforwardlib-gcc.lua<\/code> (mingw\/gcc) script from <a href=\"http:\/\/lua-users.org\/wiki\/LuaProxyDllThree\" rel=\"nofollow\">Lua Proxy DLL3 page<\/a>.<\/li>\n<li>Run <code>lua mkforwardlib.lua libtorch-lua lua51 X86<\/code>; if everything goes well, this will produce <code>lua51.dll<\/code> file in the current folder.<\/li>\n<li>Copy <code>lua51.dll<\/code> file to <code>Torch\\bin<\/code> folder.<\/li>\n<\/ol>\n\n<p>Now you should be able to debug Torch scripts by using <code>Project | Start Debugging<\/code>.<\/p>\n"},{"AnswerId":"27289841","CreationDate":"2014-12-04T08:57:43.730","ParentId":null,"OwnerUserId":"936332","Title":null,"Body":"<p>Following method works on linux platform:<\/p>\n\n<ol>\n<li><p>Configuring  the luajit interpreter by adding following code into the user.lua    <\/p>\n\n<p><code>path.lua = \"your_path\/luajit\"<\/code> <\/p><\/li>\n<li><p>Configuring the envrioment variable by adding following code into the \/opt\/zbsstudio\/lualibs\/mobdebug\/mobdebug.lua    <\/p>\n\n<p>package.path = package.path .. ';my_path\/?\/init.lua'\npackage.cpath = package.cpath .. ';my_path\/?.so'<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":26464341,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-10-20T11:15:42.283","AcceptedAnswerId":26483381.0,"OwnerUserId":498892.0,"Title":"Defining expression in theano","Body":"<p>I have a question about theano library.\nSuppose I have two matrices:  and  that respectively have shapes  and .<\/p>\n\n<p>I want to define an expression which gives tensor  of shape :<\/p>\n\n<pre><\/pre>\n\n<p>(pay attention that there is no sum, so this is not a dot product).\nOf course I can do this calculation in a loop, but I want theano to calculate derivatives for me.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"26483381","CreationDate":"2014-10-21T09:31:54.503","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Theano tensors broadcast exactly like numpy arrays. All you have to do is<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nimport numpy as np\nrng = np.random.RandomState(42)\n\nn, m1, m2  = 10, 20, 30\nA = T.fmatrix()\nB = T.fmatrix()\n\nC = A.reshape((n, m1, 1)) * B.reshape((n, 1, m2))\n\nfunc_C = theano.function([A, B], C)\n\na = rng.randn(n, m1).astype(np.float32)\nb = rng.randn(n, m2).astype(np.float32)\n\nr1 = func_C(a, b)\nr2 = np.einsum(\"ij, ik -&gt; ijk\", a, b)\nassert np.sqrt(((r1 - r2) ** 2).sum()) &lt; 1e-5\n<\/code><\/pre>\n\n<p>Where the definition of <code>C<\/code> is the crucial part. This can also be done with dynamic shapes, but this answer stays within your framework.<\/p>\n"}]}
{"QuestionId":26497564,"AnswerCount":5,"Tags":"<python><machine-learning><neural-network><theano>","CreationDate":"2014-10-21T22:38:18.433","AcceptedAnswerId":26498509.0,"OwnerUserId":3681744.0,"Title":"Theano HiddenLayer Activation Function","Body":"<p>Is there anyway to use Rectified Linear Unit (ReLU) as the activation function of the hidden layer instead of  or  in Theano? The implementation of the hidden layer is as follows and as far as I have searched on the internet ReLU is not implemented inside the Theano.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"26605167","CreationDate":"2014-10-28T09:47:24.223","ParentId":null,"OwnerUserId":"2166433","Title":null,"Body":"<p>I wrote it like this:<\/p>\n\n<pre><code>lambda x: T.maximum(0,x)\n<\/code><\/pre>\n\n<p>or:<\/p>\n\n<pre><code>lambda x: x * (x &gt; 0)\n<\/code><\/pre>\n"},{"AnswerId":"26535270","CreationDate":"2014-10-23T18:52:43.120","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>I think it is more precise to write it in this way:<\/p>\n\n<pre><code>x * (x &gt; 0.) + 0. * (x &lt; 0.)\n<\/code><\/pre>\n"},{"AnswerId":"43770203","CreationDate":"2017-05-03T21:20:14.160","ParentId":null,"OwnerUserId":"5754215","Title":null,"Body":"<p>The function is very simple in Python:<\/p>\n\n<pre><code>def relu(input):\n    output = max(input, 0)\n    return(output)\n<\/code><\/pre>\n"},{"AnswerId":"28776557","CreationDate":"2015-02-28T00:21:37.380","ParentId":null,"OwnerUserId":"498892","Title":null,"Body":"<p><strong>UPDATE:<\/strong> Latest version of theano has native support of ReLU:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#theano.tensor.nnet.relu\" rel=\"noreferrer\">T.nnet.relu<\/a>, which should be preferred over custom solutions.<\/p>\n\n<p>I decided to compare the speed of solutions, since it is very important for  NNs. Compared speed of function itself and it's gradient, in first case <code>switch<\/code> is preferred, the gradient is faster for x * (x>0). \nAll the computed gradients are correct.<\/p>\n\n<pre><code>def relu1(x):\n    return T.switch(x&lt;0, 0, x)\n\ndef relu2(x):\n    return T.maximum(x, 0)\n\ndef relu3(x):\n    return x * (x &gt; 0)\n\n\nz = numpy.random.normal(size=[1000, 1000])\nfor f in [relu1, relu2, relu3]:\n    x = theano.tensor.matrix()\n    fun = theano.function([x], f(x))\n    %timeit fun(z)\n    assert numpy.all(fun(z) == numpy.where(z &gt; 0, z, 0))\n\nOutput: (time to compute ReLU function)\n&gt;100 loops, best of 3: 3.09 ms per loop\n&gt;100 loops, best of 3: 8.47 ms per loop\n&gt;100 loops, best of 3: 7.87 ms per loop\n\nfor f in [relu1, relu2, relu3]:\n    x = theano.tensor.matrix()\n    fun = theano.function([x], theano.grad(T.sum(f(x)), x))\n    %timeit fun(z)\n    assert numpy.all(fun(z) == (z &gt; 0)\n\nOutput: time to compute gradient \n&gt;100 loops, best of 3: 8.3 ms per loop\n&gt;100 loops, best of 3: 7.46 ms per loop\n&gt;100 loops, best of 3: 5.74 ms per loop\n<\/code><\/pre>\n\n<p>Finally, let's compare to how gradient should be computed (the fastest way)<\/p>\n\n<pre><code>x = theano.tensor.matrix()\nfun = theano.function([x], x &gt; 0)\n%timeit fun(z)\nOutput:\n&gt;100 loops, best of 3: 2.77 ms per loop\n<\/code><\/pre>\n\n<p>So theano generates inoptimal code for gradient. IMHO, switch version today should be preferred.<\/p>\n"},{"AnswerId":"26498509","CreationDate":"2014-10-22T00:19:43.973","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>relu is easy to do in Theano:<\/p>\n\n<pre><code>switch(x&lt;0, 0, x)\n<\/code><\/pre>\n\n<p>To use it in your case make a python function that will implement relu and pass it to activation:<\/p>\n\n<pre><code>def relu(x):\n    return theano.tensor.switch(x&lt;0, 0, x)\nHiddenLayer(..., activation=relu)\n<\/code><\/pre>\n\n<p>Some people use this implementation: <code>x * (x &gt; 0)<\/code><\/p>\n\n<p>UPDATE: Newer Theano version have theano.tensor.nnet.relu(x) available.<\/p>\n"}]}
{"QuestionId":26499269,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-10-22T02:04:40.020","AcceptedAnswerId":26555850.0,"OwnerUserId":3681744.0,"Title":"AttributeError in Theano","Body":"<p>I am trying to run the installation test of the Theano with the following code:<\/p>\n\n<pre><\/pre>\n\n<p>However, I would see the following error corresponding to blas.py:<\/p>\n\n<pre><\/pre>\n\n<p>I understand that AttributeError is a famous error and there are questions asked about it, but for Theano the only solution I found on the internet was to add: <\/p>\n\n<pre><\/pre>\n\n<p>to the blas.py. However, this does not solve the problem and I am still facing the AttributeError.<\/p>\n","answers":[{"AnswerId":"26555850","CreationDate":"2014-10-24T20:33:47.003","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This is fixed in Theano development version.<\/p>\n\n<p>Use one of those 2 command to update Theano depending if the installation is just for your user or for the OS:<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git --user\npip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>Here is the doc for that update:<\/p>\n\n<pre><code>http:\/\/www.deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\n<\/code><\/pre>\n"}]}
{"QuestionId":26513839,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2014-10-22T17:49:41.763","AcceptedAnswerId":null,"OwnerUserId":2436416.0,"Title":"Theano import error: cannot import name Shape","Body":"<p>I am trying to import theano as <\/p>\n\n<pre><\/pre>\n\n<p>But its giving the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am currently on the bleeding edge version of theano using the following command:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":26574293,"AnswerCount":1,"Tags":"<numpy><matrix><theano>","CreationDate":"2014-10-26T14:53:45.757","AcceptedAnswerId":26588143.0,"OwnerUserId":419338.0,"Title":"Theano broadcasting different to numpy's","Body":"<p>Consider the following example of numpy broadcasting:<\/p>\n\n<pre><\/pre>\n\n<p>As expected, the vector  is added to each rows of the matrix  and the output is:<\/p>\n\n<pre><\/pre>\n\n<p>Trying to replicate the same behaviour in the git version of theano:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I understand  objects such as  have a  attribute, but I can't find a way to 1) set this correctly for the  object or 2) have it correctly inferred. How can I re-implement numpy's behaviour in theano?<\/p>\n","answers":[{"AnswerId":"26588143","CreationDate":"2014-10-27T12:59:50.387","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano need all broadcastable dimensions to be declared in the graph before compilation. NumPy use the run time shape information.<\/p>\n\n<p>By default, all shared variable dimsions aren't broadcastable, as their shape could change.<\/p>\n\n<p>To create the shared variable with the broadcastable dimension that you need in your example:<\/p>\n\n<pre><code>b = theano.shared(bval, broadcastable=(True,False))\n<\/code><\/pre>\n\n<p>I'll add this information  to the documentation.<\/p>\n"}]}
{"QuestionId":26598408,"AnswerCount":1,"Tags":"<numpy><matrix><euclidean-distance><theano>","CreationDate":"2014-10-27T23:11:57.870","AcceptedAnswerId":26613970.0,"OwnerUserId":956730.0,"Title":"A Pure Pythonic Pairwise Euclidean distance of rows of a numpy ndarray","Body":"<p>I have a matrix of size (n_classes, n_features) and i want to compute the pairwise euclidean distance of each pair of classes so the output would be a (n_classes, n_classes) matrix where each cell has the value of euclidean_distance(class_i, class_j).<\/p>\n\n<p>I know that there is this scipy spatial distances (<a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.14.0\/reference\/spatial.distance.html\" rel=\"nofollow\">http:\/\/docs.scipy.org\/doc\/scipy-0.14.0\/reference\/spatial.distance.html<\/a>)  and sklearn.metric.euclidean distances (<a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.pairwise.euclidean_distances.html\" rel=\"nofollow\">http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.pairwise.euclidean_distances.html<\/a>) but i want to use this in Theano software so i need a pure mathematical formula rather than functions that compute the results.<\/p>\n\n<p>for example i need a series of transformations like A = X * B, D = X.T-X, results = D.T something that contains just matrix mathematical operations not functions.<\/p>\n","answers":[{"AnswerId":"26613970","CreationDate":"2014-10-28T16:51:11.433","ParentId":null,"OwnerUserId":"2855342","Title":null,"Body":"<p>You can do this using numpy broadcasting as shown in <a href=\"https:\/\/gist.github.com\/kastnerkyle\/7a985f63c1ba92ac8bca\" rel=\"nofollow\">this gist<\/a>. It should be straightforward to convert this to Theano code, or just reference @eickenberg's comment above, since he's the one who showed me how to do this!<\/p>\n"}]}
{"QuestionId":26615835,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-10-28T18:29:38.693","AcceptedAnswerId":26640860.0,"OwnerUserId":596046.0,"Title":"Multiple networks in Theano","Body":"<p>I'd like to have 2 separate networks running in Theano at the same time, where the first network trains on the results of the second. I could embed both networks in the same structure but that would be a real mess in the entire forward pass (and probably won't even work because of the shared variables etc.)<\/p>\n\n<p>The problem is that when I define a theano function I don't specify the model it's applied on, meaning if I'm having a predict and a train function they'll both work on the first model I define.<\/p>\n\n<p>Is there a way to overcome that issue?<\/p>\n","answers":[{"AnswerId":"26640860","CreationDate":"2014-10-29T21:18:06.790","ParentId":null,"OwnerUserId":"596046","Title":null,"Body":"<p>In a rather simplified way I've managed to find a nice solution. The trick was to create one model, define its function and then create the other model and define the second function. Works like a charm<\/p>\n"}]}
{"QuestionId":26621341,"AnswerCount":1,"Tags":"<python><numpy><nan><divide-by-zero><theano>","CreationDate":"2014-10-29T01:21:28.740","AcceptedAnswerId":26626732.0,"OwnerUserId":956730.0,"Title":"How to prevent division by zero or replace infinite values in Theano?","Body":"<p>I'm using a cost function in Theano which involves a regularizer term that requires me to compute this term:<\/p>\n\n<pre><\/pre>\n\n<p>as some values of self.squared_euclidean_distances might be zero this produces Nan values. How can i work around this problem? I tried to use T.isinf but were not successful. One solution would be to remove zeros in self.squared_euclidean_distances into a small number or replace infinite numbers in T.sum(c \/ self.squared_euclidean_distances) to zero. I just don't know how to replace those values in Theano.<\/p>\n","answers":[{"AnswerId":"26626732","CreationDate":"2014-10-29T09:30:06.837","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Take a look at <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#condition\" rel=\"nofollow\"><code>T.switch<\/code><\/a>. You could do for example<\/p>\n\n<pre><code>T.switch(T.eq(self.squared_euclidean_distances, 0), 0, c \/ self.squared_euclidean_distances)\n<\/code><\/pre>\n\n<p>(Or, upstream, you make sure that you never compare a vector with itself using squared euclidean distance.)<\/p>\n"}]}
{"QuestionId":26669099,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-10-31T07:00:28.877","AcceptedAnswerId":null,"OwnerUserId":888051.0,"Title":"change a subset of elements' values in theano matrix","Body":"<p>I want to create a <strong>mask<\/strong> matrix dynamically, for example, in <\/p>\n\n<pre><\/pre>\n\n<p>Here is what I tried in ,<\/p>\n\n<pre><\/pre>\n\n<p>the above  code failed with error message: .\nAny other ways?<\/p>\n","answers":[{"AnswerId":"26672284","CreationDate":"2014-10-31T10:19:19.197","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>This works for me on <code>0.6.0<\/code>. I used your code and created a function from it to check the output. Try copying and pasting this:<\/p>\n\n<pre><code>import theano\nfrom theano import tensor\n\nmask = tensor.zeros((5,5))\nrow = tensor.ivector('row')\ncol = tensor.ivector('col')\nmask = tensor.set_subtensor(mask[row, col], 1)\n\nf = theano.function([row, col], mask)\n\nprint f(np.array([0, 1, 2]).astype(np.int32), np.array([1, 2, 3]).astype(np.int32))\n<\/code><\/pre>\n\n<p>This yields<\/p>\n\n<pre><code>array([[ 0.,  1.,  0.,  0.,  0.],\n       [ 0.,  0.,  1.,  0.,  0.],\n       [ 0.,  0.,  0.,  1.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])\n<\/code><\/pre>\n"}]}
{"QuestionId":26674433,"AnswerCount":2,"Tags":"<python><windows><cuda><canopy><theano>","CreationDate":"2014-10-31T12:12:06.653","AcceptedAnswerId":null,"OwnerUserId":2399329.0,"Title":"Error installing Theano for GPU 750m in Canopy on Windows 8.1","Body":"<p>I am using Windows 8.1 64 bit. I have installed canopy with academic license. My laptop has Nvidia GeForce GT750m<\/p>\n\n<p>I have installed Theano using  from Theano's git repository. I have Visual Studio 2013 and Cuda Toolkit 64 bit installed. Cuda samples are working fine.<\/p>\n\n<p>Theanoarc file:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error while doing <\/p>\n\n<pre><\/pre>\n\n<p>VisualStudio Command prompt gives the error:<\/p>\n\n<pre><\/pre>\n\n<p>  in the last line<\/p>\n\n<p>I have followed <a href=\"https:\/\/stackoverflow.com\/questions\/25729969\/installing-theano-on-windows-8-with-gpu-enabled\">Installing theano on Windows 8 with GPU enabled<\/a><\/p>\n\n<p>But <strong>Python distutils fix<\/strong> cannot be applied as I have used Canopy. Please help.<\/p>\n","answers":[{"AnswerId":"26684445","CreationDate":"2014-10-31T22:39:51.973","ParentId":null,"OwnerUserId":"1988991","Title":null,"Body":"<p>Windows Canopy, like all standard CPython 2.7 distributions for Windows, is compiled with Visual C++ 2008, and that is the compiler that you should use for compiling all non-trivial C extensions for it.<\/p>\n\n<p>The following article includes some links for obtaining this compiler:\n<a href=\"https:\/\/support.enthought.com\/entries\/26864394-Windows-Unable-to-find-vcvarsall-bat-cython-f2py-other-c-extensions-\" rel=\"nofollow\">https:\/\/support.enthought.com\/entries\/26864394-Windows-Unable-to-find-vcvarsall-bat-cython-f2py-other-c-extensions-<\/a><\/p>\n"},{"AnswerId":"26679798","CreationDate":"2014-10-31T16:58:01.570","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano do not support Visual Studio compiler. It is needed for cuda, but not for Theano CPU code. This need g++ compiler.<\/p>\n\n<p>Is the canopy package for the compiler correctly installed? What did you do to try to compile with microsoft compiler?<\/p>\n"}]}
{"QuestionId":26703136,"AnswerCount":1,"Tags":"<install><compiler-flags><theano>","CreationDate":"2014-11-02T19:02:04.990","AcceptedAnswerId":26723370.0,"OwnerUserId":1452257.0,"Title":"How to set Theano flags in the THEANORC file? And where?","Body":"<p>I'm trying to install Theano, but it is more complicated than I thought. I've used Enthought Canopy and the guide on <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#install\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#install<\/a>. In order to complete the installation I want to set the flags mentioned in the installation guide:<\/p>\n\n<blockquote>\n  <p>(Needed only for Theano 0.6rc3 or earlier) Set the Theano flags\n  blas.ldflags=-LC:\\Users\\\\AppData\\Local\\Enthought\\Canopy\\App\\appdata\\canopy-1.0.0.1160.win-x86_64\\Scripts\n  -lmk2_core -lmk2_intel_thread -lmk2_rt.<\/p>\n<\/blockquote>\n\n<p>According to <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a> I should do that by creating a .theanorc file on $HOME\/.theanorc:$HOME\/.theanorc.txt. However, I don't know how to translate $HOME into a normal windows path - what is $HOME if I have a default installation using Enthought Canopy?<\/p>\n","answers":[{"AnswerId":"26723370","CreationDate":"2014-11-03T21:17:46.967","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>To find your home on windows, you can click the \"start\" button, then select your username on the right column. This will open a window in the right folder.<\/p>\n\n<p>On Windows 7, this corredpond to this path: C:\\Users\\YOUR_USER_NAME<\/p>\n"}]}
{"QuestionId":26706686,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2014-11-03T01:51:39.993","AcceptedAnswerId":26716174.0,"OwnerUserId":4208837.0,"Title":"Theano logistic SGD with per-dimension learning rates","Body":"<p>Here's a very simple, beginner Theano question.<\/p>\n\n<p>I'm trying to modify the Logistic SGD code provided with the <a href=\"http:\/\/www.deeplearning.net\/tutorial\/logreg.html#logreg\" rel=\"nofollow\">Deep Learning Tutorials<\/a>, to switch from a single learning rate to a learning rate that would be dimension-specific. For instance if I have 3 input dimensions, I would like to use 3 different learning rates, one per dimension.<\/p>\n\n<p>The original relevant code is:<\/p>\n\n<pre><\/pre>\n\n<p>In numpy, it would be simply a matter of replacing the scalar learning rate with an array of learning rates and perform element-wise multiplication with the gradients g_W and g_b. In Theano doing so yields an error:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly there is something about Theano that I'm missing. Can anyone enlighten me?<\/p>\n","answers":[{"AnswerId":"26716174","CreationDate":"2014-11-03T14:25:25.087","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Indeed, you need to replace the learning rate scalar by an array. You can try e.g. the following:<\/p>\n\n<pre><code>learning_rate = theano.shared(np.array([0.1, 0.2, 0.05]))\n<\/code><\/pre>\n\n<p>It may need to be transposed depending on the shape of the gradient, but essentially you have stated the correct way to go and it should work using a shared variable.<\/p>\n"}]}
{"QuestionId":26718812,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-11-03T16:43:48.477","AcceptedAnswerId":26789849.0,"OwnerUserId":2663583.0,"Title":"Python - Theano scan() function","Body":"<p>I cannot fully understand the behaviour of theano.scan().<\/p>\n\n<p>Here's an example:<\/p>\n\n<pre><\/pre>\n\n<p>The above snippet prints the following sequence, which is perfectly reasonable:<\/p>\n\n<pre><\/pre>\n\n<p>However if I switch the tap index from -2 to -1, i.e.<\/p>\n\n<pre><\/pre>\n\n<p>The result becomes:<\/p>\n\n<pre><\/pre>\n\n<p>instead of what would seem reasonable to me (just take the last value of the vector and add 2):<\/p>\n\n<pre><\/pre>\n\n<p>Any help would be much appreciated.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"26789849","CreationDate":"2014-11-06T21:37:04.907","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>When you use taps=[-1], scan suppose that the information in the output info is used as is. That mean the addf function will be called with a vector and the non_sequence as inputs. If you convert x0 to a scalar, it will work as you expect:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\n\ndef addf(a1,a2):\n        print a1.type\n        print a2.type\n        return a1+a2\n\ni = T.iscalar('i')\nx0 = T.iscalar('x0') \nstep= T.iscalar('step')\n\nresults, updates = theano.scan(fn=addf,\n                   outputs_info=[{'initial':x0, 'taps':[-1]}],\n                   non_sequences=step,\n                   n_steps=i)\n\nf=theano.function([x0,i,step],results)\n\nprint f(1,10,2)\n<\/code><\/pre>\n\n<p>This give this output:<\/p>\n\n<pre><code>TensorType(int32, scalar)\nTensorType(int32, scalar)\n[ 3  5  7  9 11 13 15 17 19 21]\n<\/code><\/pre>\n\n<p>In your case as it do addf(vector,scalar), it broadcast the elemwise value.<\/p>\n\n<p>Explained in another way, if taps is [-1], x0 will be passed \"as is\" to the inner function. If taps contain anything else, what is passed to the inner function will have 1 dimension less then x0, as x0 must provide many initial steps value (-2 and -1).<\/p>\n"}]}
{"QuestionId":26725593,"AnswerCount":3,"Tags":"<python><oop><decorator><python-decorators><theano>","CreationDate":"2014-11-04T00:10:50.797","AcceptedAnswerId":26785871.0,"OwnerUserId":851699.0,"Title":"Using Python decorators to add a method to a method","Body":"<p>I want to be able to call a method according to some standard format:<\/p>\n\n<pre><\/pre>\n\n<p>, where outputs is a tuple of arrays, and each input is an array.<\/p>\n\n<p>However, in most instances, I only return one array, and don't want to be forced to return a tuple of length 1 just for the sake of the standard format. (My actual formatting problem is more complicated but lets stick with this explanation for now.)<\/p>\n\n<p>I want to be able to define a class like:<\/p>\n\n<pre><\/pre>\n\n<p>And then be able to call it as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Question is: how do I define the decorator to allow this behaviour?<\/p>\n\n<p>I tried:<\/p>\n\n<pre><\/pre>\n\n<p>, but it fails on the line with the second assert with:<\/p>\n\n<pre><\/pre>\n\n<p>Because the  requires \"self\" as an argument, and the the object has not even been created yet at the time the decorator modifies the function.<\/p>\n\n<p>So Stack, how can I do this?<\/p>\n\n<hr>\n\n<p>Notes:<\/p>\n\n<p>1) I know I could do this with base-classes and inheritance instead, but that becomes a problem when you have more than one method in the class that you want to decorate this way.<\/p>\n\n<p>2) The actual problem comes from using theano - the standard format is , but most functions don't return any updates, so you want to be able to define those functions in a natural way, but still have the option of calling them according to this standard interface.<\/p>\n","answers":[{"AnswerId":"26729151","CreationDate":"2014-11-04T06:41:03.870","ParentId":null,"OwnerUserId":"296974","Title":null,"Body":"<p>That's indeed a problem, because the way the \"bound\" method is retrieved from the function doesn't consider this way.<\/p>\n\n<p>I see two ways:<\/p>\n\n<ol>\n<li><p>You could just wrap the function:<\/p>\n\n<pre><code>def single_return_format(fcn):\n    # TODO Do some functools.wraps here...\n    return lambda *args, **kwargs: (fcn(*args, **kwargs), )\n<\/code><\/pre>\n\n<p>No fooling around with <code>.standard_format<\/code>, but a mere replacement of the function. So the function can define itself as returning one value, but can only be called as returning the tuple.<\/p><\/li>\n<li><p>If this is not what you want, you can define a class for decorating methods which overrides <code>__get__<\/code> and does the wrapping in a \"live fashion\". Of course, it can as well redefine <code>__call__<\/code> so that it is usable for (standalone, non-method) functions as well.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"26744995","CreationDate":"2014-11-04T21:04:27.653","ParentId":null,"OwnerUserId":"529630","Title":null,"Body":"<p>To get exactly what you want you'd have to write a non-data descriptor and a set of wrapper classes for your functions. The reason for this is that the process of getting functions from objects as methods is highly optimised and it's not possible to hijack this mechanism. Instead you have to write your own classes that simulate this mechanism -- which will slow down your code if you are making lots of small method calls. <\/p>\n\n<p>The very best way I can think to get the desired functionality is not to use any of the methods that you describe, but rather write a wrapper function that you use when needed to call a normal function in the standard format. eg.<\/p>\n\n<pre><code>def vectorise(method, *args, **kwargs):\n    return tuple(method(arg, **kwargs) for arg in args)\n\nobj = _SomeClass()\n\nresult = vectorise(obj.add_one, 1, 2, 3)\n<\/code><\/pre>\n\n<p>Indeed, this is how <code>numpy<\/code> takes functions that operate on one argument and turns them into a function that works on arrays.<\/p>\n\n<pre><code>import numpy\n\ndef add_one(x):\n    return x + 1\n\narr = numpy.vectorize(add_one)([1, 2, 3])\n<\/code><\/pre>\n\n<p>If you really, really want to use non-data descriptors then following will work. Be warned these method calls are considerably slower. On my computer a normal method call takes 188 nanoseconds versus 1.53 microseconds for a \"simple\" method call -- a ten-fold difference. And <code>vectorise<\/code> call takes half the time a <code>standard_form<\/code> call does. The vast majority of that time is the lookup of the methods. The actual method calls are quite fast.<\/p>\n\n<pre><code>class simple_form:\n    \"\"\"Allows a simple function to be called in a standard way.\"\"\"\n\n    def __init__(self, func):\n        self.func = func \n\n    def __get__(self, instance, owner):\n        if instance is None:\n            return self.func\n        return SimpleFormMethod(self.func, instance)\n\n\nclass MethodBase:\n    \"\"\"Provides support for getting the string representation of methods.\"\"\"\n\n    def __init__(self, func, instance):\n        self.func = func\n        self.instance = instance\n\n    def _format(self):\n        return \"&lt;bound {method_class} {obj_class}.{func} of {obj}&gt;\".format(\n            method_class=self.__class__.__name__,\n            obj_class=self.instance.__class__.__name__,\n            func=self.func.__name__,\n            obj=self.instance)\n\n    def __str__(self):\n        return self._format()\n\n    def __repr__(self):\n        return self._format()\n\n\nclass SimpleFormMethod(MethodBase):\n\n    def __call__(self, *args, **kwargs):\n        return self.func(self.instance, *args, **kwargs)\n\n    @property\n    def standard_form(self):\n        return StandardFormMethod(self.func, self.instance)\n\n\nclass StandardFormMethod(MethodBase):\n\n    def __call__(self, *args, **kwargs):\n        return tuple(self.func(self.instance, arg, **kwargs) for arg in args)\n\n\nclass Number(object):\n\n    def __init__(self, value):\n        self.value = value\n\n    def add_to(self, *values):\n        return tuple(val + self.value for val in values)\n\n    @simple_form\n    def divide_into(self, value):\n        return value \/ self.value\n\n\nnum = Number(2)\nprint(\"normal method access:\", num.add_to, sep=\"\\n\")\nprint(\"simple form method access:\", num.divide_into, sep=\"\\n\")\nprint(\"standard form method access:\", num.divide_into.standard_form, sep=\"\\n\")\nprint(\"access to underlying function:\", Number.divide_into, sep=\"\\n\")\nprint(\"simple example usage:\", num.divide_into(3))\nprint(\"standard example usage:\", num.divide_into.standard_form(*range(3)))\n<\/code><\/pre>\n"},{"AnswerId":"26785871","CreationDate":"2014-11-06T17:39:38.720","ParentId":null,"OwnerUserId":"851699","Title":null,"Body":"<p>Dunes gave the correct answer.  I've stripped it down to bare bones so that it solves the problem in the question.  The stripped-down code is here:<\/p>\n\n<pre><code>class single_return_format(object):\n\n    def __init__(self, func):\n        self._func = func\n\n    def __get__(self, instance, owner):\n        return SimpleFormMethod(instance, self._func)\n\n\nclass SimpleFormMethod(object):\n\n    def __init__(self, instance, func):\n        self._instance = instance\n        self._func = func\n\n    def __call__(self, *args, **kwargs):\n        return self._func(self._instance, *args, **kwargs)\n\n    @property\n    def standard_format(self):\n        return lambda *args, **kwargs: (self._func(self._instance, *args, **kwargs), )\n\n\nclass _SomeClass(object):\n\n    def __init__(self):\n        self._amount_to_add = 1\n\n    @single_return_format\n    def add_one(self, x):\n        return x+self._amount_to_add\n\n\nobj = _SomeClass()\nassert obj.add_one(3) == 4\nassert obj.add_one.standard_format(3) == (4, )\n<\/code><\/pre>\n"}]}
{"QuestionId":26799771,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2014-11-07T11:00:17.530","AcceptedAnswerId":26801592.0,"OwnerUserId":2963058.0,"Title":"How to print a matric or vector to screen returned by a function in theano?","Body":"<p>I'm new of theano, when I tried to run this code, I got the response below:<\/p>\n\n<pre><\/pre>\n\n<p>Initial model:\ndot.0<\/p>\n\n<p>how can I actually see the value of the returned value, thanks.<\/p>\n","answers":[{"AnswerId":"26801592","CreationDate":"2014-11-07T12:48:12.007","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You need to evaluate the expression, which is best done by turning it into a function. Remove line 18 from your code and then add<\/p>\n\n<pre><code>dot_product = T.dot(x, w)\nf = theano.function([x], dot_product)\n\nprint f(rng.randn(feats))\n<\/code><\/pre>\n"}]}
{"QuestionId":26822010,"AnswerCount":1,"Tags":"<macos><object><malformed><torch>","CreationDate":"2014-11-08T20:57:37.043","AcceptedAnswerId":null,"OwnerUserId":4231095.0,"Title":"install_name_tool malformed object (load command 23 cmdsize is zero) - Mac OS X Yosemite","Body":"<p>Installation of cunn for torch on Yosemite fails with malformed object error.  <\/p>\n\n<pre><\/pre>\n\n<p>Searching online shows that this is related to library corruption or an update to install_name_tool. I replaced the install_name_tool from XCode(6.1) into \/usr\/bin but I still get the same error.  Below are some diagnostics<\/p>\n\n<pre><\/pre>\n\n<p>I need this to work so that I can use CUDA with torch, I have already spent hours on it, please help.<\/p>\n","answers":[{"AnswerId":"26837009","CreationDate":"2014-11-10T04:55:59.550","ParentId":null,"OwnerUserId":"4234292","Title":null,"Body":"<p>I had the same problem and I solved it by building <code>libcunn.so<\/code> locally.\nRun the following commands:<\/p>\n\n<pre><code>git clone https:\/\/github.com\/torch\/cunn.git\nls cunn\ncmake -E make_directory build\ncd build\/\ncmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH=\"\/usr\/local\/bin\/..\" -DCMAKE_INSTALL_PREFIX=\"\/usr\/local\/lib\/luarocks\/rocks\/cunn\/scm-1\" &amp;&amp; make\n<\/code><\/pre>\n\n<p>Then you should have <\/p>\n\n<pre><code>Linking CXX shared module libcunn.so\n<\/code><\/pre>\n\n<p>Then just copy the library to the target folder:<\/p>\n\n<pre><code>cp libcunn.so \/usr\/local\/lib\/lua\/5.1\/libcunn.so\n<\/code><\/pre>\n"}]}
{"QuestionId":26828318,"AnswerCount":1,"Tags":"<python><theano><pylearn>","CreationDate":"2014-11-09T12:45:36.867","AcceptedAnswerId":null,"OwnerUserId":648896.0,"Title":"How can I transform a novel data by the trained Autoencoder with Pylearn2?","Body":"<p>I trained and saved two layer stacked CAE model by Pylearn2. I would like to load these models and transform a novel dataset. How should I do it?<\/p>\n\n<p>This is my model:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried something like this but it does not work.<\/p>\n\n<pre><\/pre>\n\n<p>This is what I do lately but not sure about its correctness:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"26949633","CreationDate":"2014-11-15T19:10:43.200","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>Check the <code>pylearn2.scripts.autoencoder<\/code> directory for a configuration file that shows how to use stacked, pretrained autoencoders.  You load the pretrained models as a transformer object on the dataset on it's way into the next stage of the model. <\/p>\n\n<p>If you don't want to use yaml files, you should be able to string the model functions together using the correct methods (untested as I'm writing off the top of my head):<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>inputs = T.vector()<\/p>\n    \n    <p>enc = l1.encode(inputs)  <\/p>\n    \n    <p>output = l2.encode(enc)<\/p>\n    \n    <p>f= theano.function([inputs], output)<\/p>\n    \n    <p>real_transformation = f(real_data)<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>And to go back, you can do the same with the <code>ae_layer.decode()<\/code> method.  <\/p>\n\n<p>If these are part of the layers of and <code>MLP<\/code>, you can call the <code>.fprop()<\/code> method on the <code>MLP<\/code> to do an upward pass through all the layers.<\/p>\n"}]}
{"QuestionId":26879157,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2014-11-12T04:13:59.933","AcceptedAnswerId":26900149.0,"OwnerUserId":1245262.0,"Title":"Purpose of 'givens' variables in Theano.function","Body":"<p>I was reading the code for the logistic function given at <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a>. I am confused about the difference between  &amp;  variables for a function. The functions that compute mistakes made by a model on a minibatch are:<\/p>\n\n<pre><\/pre>\n\n<p>Why couldn't\/wouldn't one just make x&amp; y shared input variables and let them be defined when an actual model instance is created?<\/p>\n","answers":[{"AnswerId":"26900149","CreationDate":"2014-11-13T02:04:57.000","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The <code>givens<\/code> parameter allows you to separate the description of the model and the exact definition of the inputs variable. This is a consequence of what the given parameter do: modify the graph to compile before compiling it. In other words, we substitute in the graph, the key in givens with the associated value.<\/p>\n\n<p>In the deep learning tutorial, we use a normal Theano variable to build the model. We use <code>givens<\/code> to speed up the GPU. Here, if we keep the dataset on the CPU, we will transfer a mini-batch to the GPU at each function call. As we do many iterations on the dataset, we end up transferring the dataset multiple time to the GPU. As the dataset is small enough to fit on the GPU, we put it in a shared variable to have it transferred to the GPU if one is available (or stay on the Central Processing Unit if the Graphics Processing Unit is disabled). Then when compiling the function, we swap the input with a slice corresponding to the mini-batch of the dataset to use. Then the input of the Theano function is just the index of that mini-batch we want to use.<\/p>\n"},{"AnswerId":"26890315","CreationDate":"2014-11-12T15:17:41.747","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>I don't think anything is stopping you from doing it that way (I didn't try the <code>updates=<\/code> dictionary using an input variable directly, but why not). Remark however that for pushing data to a GPU in a useful manner, you will need it to be in a shared variable (from which <code>x<\/code> and <code>y<\/code> are taken in this example).<\/p>\n"}]}
{"QuestionId":26907913,"AnswerCount":1,"Tags":"<python><matrix><sparse-matrix><theano><autoencoder>","CreationDate":"2014-11-13T11:30:39.073","AcceptedAnswerId":null,"OwnerUserId":1780628.0,"Title":"How can I speed up an autoencoder to use on text data written in python's theano package?","Body":"<p>I'm new to theano and I'm trying to adapt the autoencoder script <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/dA.py\" rel=\"nofollow\">here<\/a> to work on text data. This code uses the MNIST dataset as training data. This data is in the form of a numpy 2d array. <\/p>\n\n<p>My data is a csr sparse matrix of about 100,000 instances with about 50,000 features. The matrix is the result of using sklearn's tfidfvectorizer to fit and transform the text data. As I'm using sparse matrices I modify the code to use the theano.sparse package to represent my input.\nMy training set is the symbolic variable: <\/p>\n\n<pre><\/pre>\n\n<p>However, theano.sparse matrices cannot perform all of the operations used in the original script (there is a list of sparse operations <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/library\/sparse\/#list-of-implemented-operations\" rel=\"nofollow\">here<\/a>). The code uses dot and sum from the tensor methods on the input. I have changed the dot to sparse.dot but I can't find out what to replace the sum with so I am converting the training batches to dense matrices and using the original tensor methods as shown in this cost function:<\/p>\n\n<pre><\/pre>\n\n<p>The get_corrupted_input and get_reconstructed_input methods remain as they are in the link above. My question is is there a faster way to do this? <\/p>\n\n<p>Converting the matrices to dense is making running the training very slow. Currently it takes 20.67m to do one training epoch with a batch size of 20 training instances.<\/p>\n\n<p>Any help or tips you could give would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"26949508","CreationDate":"2014-11-15T18:57:44.553","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>In the most recent master branch of theano.sparse there is an sp_sum method listed.<\/p>\n\n<p><a href=\"http:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sparse\/basic.py\" rel=\"nofollow\">(see here)<\/a>  <\/p>\n\n<p>If you're not using the bleeding edge version I'd install that and see if calling it will work and if doing so speeds things up:  <\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>(And if it does, noting it here would be nice, it's not always clear that the sparse functionality is much faster than using dense calculations all the way through, especially on the gpu.) <\/p>\n"}]}
{"QuestionId":26980666,"AnswerCount":0,"Tags":"<python><numpy><neural-network><enthought><theano>","CreationDate":"2014-11-17T19:45:21.870","AcceptedAnswerId":null,"OwnerUserId":2688733.0,"Title":"Theano: Grad error","Body":"<p>I have a complicated architecture and I am calculating gradients wrt to diff parameters, but I am getting following strange error, what is this indicative of? I am not getting any clue why is this error popping up?<\/p>\n\n<pre><\/pre>\n\n<p>The sequence steps which lead to this error are: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":26992692,"AnswerCount":1,"Tags":"<neural-network><enthought><theano>","CreationDate":"2014-11-18T11:01:18.853","AcceptedAnswerId":27046695.0,"OwnerUserId":2688733.0,"Title":"Theano : Extract particular rows in matrix","Body":"<p>I have a matrix W and two vectors y1 and y2. I want to extract rows from W. The rows I am interested in are in the range [y1:y2]. What is the best way of doing this in Theano? <strong>Can this be done without using theano.map or tensor.switch method<\/strong>? This obtained matrix will be used somewhere in grad computation. For e.g.:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"27046695","CreationDate":"2014-11-20T18:18:04.917","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>So tell otherwise what you want is:<\/p>\n\n<pre><code>out=[]\nfor i,j in zip(y1,y2):\n   out.append(W[i:j])\nnumpy.asarray(out)\n<\/code><\/pre>\n\n<p>Is the lenght of y1 and y2 constant? If so, you can unroll the loop like this:<\/p>\n\n<pre><code>out=[]\nfor i in range(LEN):\n   out.append(W[y1[i]:y2[j]])\ntheano.stack(*out)\n<\/code><\/pre>\n\n<p>Theano support all the power of NumPy advanced indexing. If you can find how to do it with NumPy without the stack, you can do it the same way in Theano.<\/p>\n"}]}
{"QuestionId":27015150,"AnswerCount":3,"Tags":"<lua><luajit><torch>","CreationDate":"2014-11-19T10:58:59.830","AcceptedAnswerId":27019667.0,"OwnerUserId":2783487.0,"Title":"How to get past 1gb memory limit of 64 bit LuaJIT on Linux?","Body":"<p>The overview is I am prototyping code to understand my problem space, and I am running into 'PANIC: unprotected error in call to Lua API (not enough memory)' errors.  I am looking for ways to get around this limit.<\/p>\n\n<p>The environment bottom line is Torch, a scientific computing framework that runs on LuaJIT, and LuaJIT runs on Lua.  I need Torch because I eventually want to hammer on my problem with neural nets on a GPU, but to get there I need a good representation of the problem to feed to the nets.  I am (stuck) on Centos Linux, and I suspect that trying to rebuild all the pieces from source in 32bit mode (this is reported to extend the LuaJIT memory limit to 4gb) will be a nightmare if it works at all for all of the libraries.<\/p>\n\n<p>The problem space itself is probably not particularly relevant, but in overview I have datafiles of points that I calculate distances between and then bin (i.e. make histograms of) these distances to try and work out the most useful ranges.  Conveniently I can create complicated Lua tables with various sets of bins and torch.save() the mess of counts out, then pick it up later and inspect with different normalisations etc. -- so after one month of playing I am finding this to be really easy and powerful.  <\/p>\n\n<p>I can make it work looking at up to 3 distances with 15 bins each (15x15x15 plus overhead), but this only by adding explicit garbagecollection() calls and using fork()\/wait() for each datafile so that the outer loop will keep running if one datafile (of several thousand) still blows the memory limit and crashes the child.  This gets extra painful as each successful child process now has to read, modify and write the current set of bin counts -- and my largest files for this are currently 36mb.  I would like to go larger (more bins), and would really prefer to just hold the counts in the 15 gigs of RAM I can't seem to access.<\/p>\n\n<p>So, here are some paths I have thought of; please do comment if you can confirm\/deny that any of them will\/won't get me outside of the 1gb boundary, or will just improve my efficiency within it.  Please do comment if you can suggest another approach that I have not thought of.<\/p>\n\n<ul>\n<li><p>am I missing a way to fire off a Lua process that I can read an arbitrary table back in from?  No doubt I can break my problem into smaller pieces, but parsing a return table from stdio (as from a system call to another Lua script) seems error prone, and writing\/reading small intermediate files will be a lot of disk i\/o.<\/p><\/li>\n<li><p>am I missing a stash-and-access-table-in-high-memory module ?  This seems like what I really want, but not found it yet<\/p><\/li>\n<li><p>can FFI C data structures be put outside the 1gb?  Doesn't seem like that would be the case but certainly I lack a full understanding of what is causing the limit in the first place.  I suspect that this will just get me an efficiency improvement over generic Lua tables for the few pieces that have moved beyond prototyping?  (unless I do a bunch of coding for each change)<\/p><\/li>\n<li><p>Surely I can get out by writing an extension in C (Torch appears to support nets that should go outside of the limit), but my brief investigation there turns up references to 'lightuserdata' pointers -- does this mean that a more normal extension won't get outside 1gb either?  This also seems like it has the heavy development cost for what should be a prototyping exercise.<\/p><\/li>\n<\/ul>\n\n<p>I know C well so going the FFI or extension route doesn't bother me - but I know from experience that encapsulating algorithms in this way can be both really elegant and really painful with two places to hide bugs.  Working through data structures containing tables within tables on the stack doesn't seem great either.  Before I make this effort I would like to be certain that the end result really will solve my problem.<\/p>\n\n<p>Thanks for reading the long post.<\/p>\n","answers":[{"AnswerId":"41748149","CreationDate":"2017-01-19T17:33:35.457","ParentId":null,"OwnerUserId":"4748557","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/github.com\/torch\/tds\" rel=\"nofollow noreferrer\" title=\"torch tds\">torch tds<\/a> module. From the README:<\/p>\n\n<blockquote>\n  <p>Data structures which do not rely on Lua memory allocator, nor being limited by Lua garbage collector.<\/p>\n  \n  <p>Only C types can be stored: supported types are currently number, strings, the data structures themselves (see nesting: e.g. it is possible to have a Hash containing a Hash or a Vec), and torch tensors and storages. All data structures can store heterogeneous objects, and support torch serialization.<\/p>\n<\/blockquote>\n"},{"AnswerId":"27019667","CreationDate":"2014-11-19T14:44:31.987","ParentId":null,"OwnerUserId":"646619","Title":null,"Body":"<p>Only object allocated by LuaJIT itself are limited to the first 2GB of memory. This means that tables, strings, full userdata (i.e. not lightuserdata), and FFI objects allocated with <code>ffi.new<\/code> will count towards the limit, but objects allocated with <code>malloc<\/code>, <code>mmap<\/code>, etc. are not subjected to this limit (regardless if called by a C module or the FFI).<\/p>\n\n<p>An example for allocating a structure with <code>malloc<\/code>:<\/p>\n\n<pre><code>ffi.cdef[[\n    typedef struct { int bar; } foo;\n    void* malloc(size_t);\n    void free(void*);\n]]\n\nlocal foo_t = ffi.typeof(\"foo\")\nlocal foo_p = ffi.typeof(\"foo*\")\n\nfunction alloc_foo()\n    local obj = ffi.C.malloc(ffi.sizeof(foo_t))\n    return ffi.cast(foo_p, obj)\nend\n\nfunction free_foo(obj)\n    ffi.C.free(obj)\nend\n<\/code><\/pre>\n\n<p>The new GC to be implemented in LuaJIT 3.0 IIRC will not have this limit, but I haven't heard any news on it's development recently.<\/p>\n\n<p>Source: <a href=\"http:\/\/lua-users.org\/lists\/lua-l\/2012-04\/msg00729.html\">http:\/\/lua-users.org\/lists\/lua-l\/2012-04\/msg00729.html<\/a><\/p>\n"},{"AnswerId":"27060812","CreationDate":"2014-11-21T11:54:34.160","ParentId":null,"OwnerUserId":"2783487","Title":null,"Body":"<p>Here is some follow-up information for those who find this question later:<\/p>\n\n<p>The key information is as posted by Colonel Thirty Two, that C module extensions and FFI code can easily get outside of the limit. (and the referenced lua list post reminds that plain Lua tables that go outside the limit will be very slow to garbage collect)<\/p>\n\n<p>It took me some time to pull the pieces together to both access and save\/load my objects, so here it is in one place:<\/p>\n\n<p>I used lds at <a href=\"https:\/\/github.com\/neomantra\/lds\" rel=\"noreferrer\">https:\/\/github.com\/neomantra\/lds<\/a> as a starting point, in particular the 1-D Array code.<\/p>\n\n<p>This broke using torch.save(), as it doesn't know how to write the new objects.  For each object I added the code below (using Array as the example):<\/p>\n\n<pre><code>function Array:load(inp)\n   for i=1,#inp do\n      self._data[i-1] = tonumber(inp[i])\n   end\n   return self\nend\n\nfunction Array:serialize ()\n   local siz = tonumber(self._size)\n   io.write(' lds.ArrayT( ffi.typeof(\"double\"), lds.MallocAllocator )( ', siz , \"):load({\")\n   for i=0,siz-1 do\n      io.write(string.format(\"%a,\", self._data[i]))\n   end\n   io.write(\"})\")\nend\n<\/code><\/pre>\n\n<p>Note that my application specifically uses doubles and malloc(), so a better implementation would store and use these in self rather than hard coding above.<\/p>\n\n<p>Then as discussed in PiL and elsewhere, I needed a serializer that would handle the object:<\/p>\n\n<pre><code>function serialize (o)\n     if type(o) == \"number\" then\n       io.write(o)\n     elseif type(o) == \"string\" then\n       io.write(string.format(\"%q\", o))\n     elseif type(o) == \"table\" then\n       io.write(\"{\\n\")\n       for k,v in pairs(o) do\n          io.write(\"  [\"); serialize(k); io.write(\"] = \")\n         serialize(v)\n         io.write(\",\\n\")\n       end\n       io.write(\"}\\n\")\n     elseif o.serialize then\n        o:serialize()\n     else\n       error(\"cannot serialize a \" .. type(o))\n     end\nend\n<\/code><\/pre>\n\n<p>and this needs to be wrapped with:<\/p>\n\n<pre><code>io.write('do local _ = ')\nserialize( myWeirdTable )\nio.write('; return _; end')\n<\/code><\/pre>\n\n<p>and then the output from that can be loaded back in with<\/p>\n\n<pre><code>local myWeirdTableReloaded = dofile('myWeirdTableSaveFile')\n<\/code><\/pre>\n\n<p>See PiL (Programming in Lua book) for dofile()<\/p>\n\n<p>Hope that helps someone!<\/p>\n"}]}
{"QuestionId":27047857,"AnswerCount":1,"Tags":"<python><numpy><scipy><sparse-matrix><theano>","CreationDate":"2014-11-20T19:26:02.573","AcceptedAnswerId":27107458.0,"OwnerUserId":919431.0,"Title":"Theano gradient of sparse matrix multiplication","Body":"<p>I'm trying to implement an autoencoder with sparse inputs in Theano.<\/p>\n\n<p>I got the sparse autoencoder to work with a squared error cost function. But if I want to apply a cross-entropy error, which contains matrix multiplications, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I uploaded an example notebook illustrating the problem at <a href=\"http:\/\/nbviewer.ipython.org\/urls\/gist.githubusercontent.com\/peterroelants\/4946cdbf189c5e75f2b7\/raw\/2ee7d3e533a4a6ac2707a2ffa310b81a86e70afd\/gistfile1.json\" rel=\"nofollow\">http:\/\/nbviewer.ipython.org\/urls\/gist.githubusercontent.com\/peterroelants\/4946cdbf189c5e75f2b7\/raw\/2ee7d3e533a4a6ac2707a2ffa310b81a86e70afd\/gistfile1.json<\/a> .<\/p>\n\n<p>I distilled the problem down to the matrix multiplication . This works in the dense case [see cell 2], but gives an error in the sparse case [see cell 3]. Note that changing this cost function in the sparse case [cell 3] to the squared error () will result in a working result.<\/p>\n\n<p>Can anyone point me out what I'm doing wrong? And show me how to get a sparse input autoencoder with cross-entropy error to work in Theano?<\/p>\n","answers":[{"AnswerId":"27107458","CreationDate":"2014-11-24T14:45:01.640","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You can not use T.* function on sparse variable. In this case, you can use:<\/p>\n\n<pre><code>theano.sparse.sp_sum((x * T.log(z))\n<\/code><\/pre>\n\n<p>\\edit This diff fix in Theano fix this crash:<\/p>\n\n<pre><code>diff --git a\/theano\/sparse\/basic.py b\/theano\/sparse\/basic.py\nindex 4620c5a..a352b9a 100644\n--- a\/theano\/sparse\/basic.py\n+++ b\/theano\/sparse\/basic.py\n@@ -2244,7 +2244,7 @@ class MulSD(gof.op.Op):\n     def grad(self, (x, y), (gz,)):\n         assert _is_sparse_variable(x) and _is_dense_variable(y)\n         assert _is_sparse_variable(gz)\n-        return y * gz, x * gz\n+        return y * gz, dense_from_sparse(x * gz)\n\n     def infer_shape(self, node, shapes):\n         return [shapes[0]]\n<\/code><\/pre>\n\n<p>I'll try to get the fix merged in Theano this week.<\/p>\n"}]}
{"QuestionId":27064617,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2014-11-21T15:20:47.010","AcceptedAnswerId":27081373.0,"OwnerUserId":1150636.0,"Title":"Theano multiple tensors as output","Body":"<p>I am using Theano to create a neural network, but when I try to <strong>return two lists of tensors at the same time in a list<\/strong> I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>What kind of Theano structure should I use to return the two tensors together in an array so I can retrieve them like this:<\/p>\n\n<pre><\/pre>\n\n<p>I tried using some of the things I found in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">basic Tensor functionality page<\/a> but none of those work. (For example I tried the stack or stacklists)<\/p>\n\n<p>Here is the error I get using theano.tensor.stack or stacklists:<\/p>\n\n<pre><\/pre>\n\n<p>A little extra context to the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"27081373","CreationDate":"2014-11-22T19:12:42.537","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This is not supported by Theano. When you call <code>theano.function(inputs, outputs)<\/code>, outputs can be only 2 things:<\/p>\n\n<p>1) a Theano variable\n2) a list of Theano variables<\/p>\n\n<p>(2) does not allow you to have a list in the top level list, so you should flatten the lists in the outputs. This will return more then 2 outputs.<\/p>\n\n<p>A compossible solution to your problem is to have the inner list copied into 1 variable.<\/p>\n\n<pre><code>tensor_nabla_w = theano.tensor.stack(*nabla_w).\n<\/code><\/pre>\n\n<p>This asks that all elements in nabla_w is are the same shape. This will add an extra copy in the computation graph (so it could be a little slower).<\/p>\n\n<p>Update 1: fix call to stack()<\/p>\n\n<p>Update 2:<\/p>\n\n<p>As of now, we have the added constraint that all the elements will have different shapes, so stack can not be used. If they all have the same number of dimensions and dtype, you can use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/typed_list.html\" rel=\"nofollow\">typed_list<\/a>, otherwise you will need to modify Theano yourself or flatten the output's lists.<\/p>\n"}]}
{"QuestionId":27077788,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2014-11-22T13:14:11.653","AcceptedAnswerId":null,"OwnerUserId":1002394.0,"Title":"Theano float64 matrix product value error","Body":"<p>I need to do matrix multiplication with float64 precision matrices.\nThe following code works in float32 and matrix() instead of dmatrix().\nHowever, when it comes to float64, it fails.<\/p>\n\n<pre><\/pre>\n\n<p>I have the following error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix this? <\/p>\n","answers":[{"AnswerId":"27080228","CreationDate":"2014-11-22T17:21:05.303","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Check out <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#tips-for-improving-performance-on-gpu\" rel=\"nofollow\">the deep learning tutorial here<\/a> to see that right now theano calculations only benefit from GPU when <code>float32<\/code> are passed. See also <a href=\"http:\/\/comments.gmane.org\/gmane.comp.mathematics.theano.user\/1087\" rel=\"nofollow\">this thread<\/a> on the mailing list.<\/p>\n"},{"AnswerId":"27576976","CreationDate":"2014-12-20T04:40:32.630","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is not related to the GPU, but the value you used for the Theano flag blas.ldflags: 'asdfadf'<\/p>\n\n<p>This is passed to g++ when compiling again blas. You didn't put value g++ parameter.<\/p>\n"}]}
{"QuestionId":27106371,"AnswerCount":1,"Tags":"<neural-network><theano><deep-learning><lstm>","CreationDate":"2014-11-24T13:47:27.667","AcceptedAnswerId":null,"OwnerUserId":1367788.0,"Title":"LSTM implementation using theano","Body":"<p>I am using theano scan function for implementing LSTM(long short term memory), but I got error like <\/p>\n\n<pre><\/pre>\n\n<p>I used scan like this <\/p>\n\n<pre><\/pre>\n\n<p>and step_fprop is defined as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Anyone has any idea why I kept getting that kind of error.<\/p>\n","answers":[{"AnswerId":"27335619","CreationDate":"2014-12-06T19:34:32.167","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p><code>outputs_info<\/code> expects values for each of the variables passed in the <code>return<\/code> of the <code>step_fprop<\/code> function.  your code seems to only return the hidden state <code>hh<\/code> but <code>ouputs_info<\/code> expects two values whose initial state is defined by <code>p_c<\/code> and <code>p_hidden_inputs<\/code><\/p>\n\n<p>It looks like you'll need to <code>return [_whatever_previous_lstm_state, hh]<\/code> in the step_fprop function <\/p>\n"}]}
{"QuestionId":27120646,"AnswerCount":2,"Tags":"<machine-learning><theano>","CreationDate":"2014-11-25T07:07:18.567","AcceptedAnswerId":null,"OwnerUserId":3121136.0,"Title":"Theano gradient doesn't work with .sum(), only .mean()?","Body":"<p>I'm trying to learn theano and decided to implement linear regression (using their Logistic Regression from the tutorial as a template). I'm getting a wierd thing where T.grad doesn't work if my cost function uses .sum(), but does work if my cost function uses .mean(). Code snippet:<\/p>\n\n<p>(THIS DOESN'T WORK, RESULTS IN A W VECTOR FULL OF NANs):<\/p>\n\n<pre><\/pre>\n\n<p>(THIS DOES WORK, PERFECTLY):<\/p>\n\n<pre><\/pre>\n\n<p>The only difference is in the cost = single_error.sum() vs single_error.mean(). What I don't understand is that the gradient should be the exact same in both cases (one is just a scaled version of the other). So what gives?<\/p>\n","answers":[{"AnswerId":"27139964","CreationDate":"2014-11-26T01:42:00.320","ParentId":null,"OwnerUserId":"1209307","Title":null,"Body":"<p>Try dividing your gradient descent step size by the number of training examples.<\/p>\n"},{"AnswerId":"27576961","CreationDate":"2014-12-20T04:37:00.160","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The learning rate (0.1) is way to big. Using mean make it divided by the batch size, so this help. But I'm pretty sure you should make it much smaller. Not just dividing by the batch size (which is equivalent to using mean).<\/p>\n\n<p>Try a learning rate of 0.001.<\/p>\n"}]}
{"QuestionId":27191631,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2014-11-28T15:20:29.287","AcceptedAnswerId":27202899.0,"OwnerUserId":4279087.0,"Title":"I'm trying to use CNN with data that aren't images","Body":"<p>The code being used is the CNN from <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet<\/a> but i'm having problems to understand what i need to change in order to accept other types of data. The file that i used has the same format of the MNIST but much smaller, this is the data being used <a href=\"https:\/\/archive.ics.uci.edu\/ml\/datasets\/Iris\" rel=\"nofollow\">https:\/\/archive.ics.uci.edu\/ml\/datasets\/Iris<\/a><\/p>\n\n<p>These are the two parts that i changed:<\/p>\n\n<p>Batch_size, nkers, n_epocchs<\/p>\n\n<pre><\/pre>\n\n<p>All parameters of the layers<\/p>\n\n<pre><\/pre>\n\n<p>When i execute this code i get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Probably the answer is really easy but i'm around a week looking this code and nothing pops in my mind<\/p>\n","answers":[{"AnswerId":"27202899","CreationDate":"2014-11-29T13:35:55.523","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The C in CNN stands for convolution. In order to perform a convolution you need variables that taken together form some sort of spatial\/temporal\/in any way continuous extent, on which a group structure holds, such as translation in space, translation in time, rotations, or something more exotic. This is not the case for the data you are working with, so using a CNN does not make much sense. (That doesn't stop you from trying to arrange the variables in a 2D space and to see what comes out, but it doesn't seem at all useful.) If you want to do NNs, stick to fully connected ones and start by evaluating logistic regression.<\/p>\n"}]}
{"QuestionId":27276822,"AnswerCount":2,"Tags":"<lua><mingw><luarocks><torch>","CreationDate":"2014-12-03T16:30:46.630","AcceptedAnswerId":null,"OwnerUserId":2698948.0,"Title":"Installing Torch7 with Luarocks on Windows with mingw build error","Body":"<p>I followed the instructions <a href=\"http:\/\/www.thijsschreijer.nl\/blog\/?p=863\" rel=\"noreferrer\">here<\/a> and set up Lua and Luarocks from scratch, with Mingw. Everything worked fine and I was able to install rocks, including ones which require compiling like LuaSocket.<\/p>\n\n<p>I followed the instructions on the <a href=\"http:\/\/torch.ch\/\" rel=\"noreferrer\">Torch7<\/a> page to install Torch via luarocks. But it fails building. I do not understand why.<\/p>\n\n<p>Here is the console output. My best guess is that it has something to do with  when I think I want it to use Mingw.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31332589","CreationDate":"2015-07-10T04:41:36.040","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>The command looks mostly correct, but I think the cmake command needs <code>-G \"MSYS Makefiles\"<\/code> option to use mingw instead of VS. You may also need to pull the most recent torch version as it includes <a href=\"https:\/\/github.com\/torch\/torch7\/pull\/287\" rel=\"nofollow\">several changes<\/a> that fix some compilation issues with mingw.<\/p>\n\n<p>Note that I haven't tested the changes with LuaRocks and not sure how to pass that additional option to it, but you should be able to run the same command manually to get the desired result (I compiled it from the command line).<\/p>\n"},{"AnswerId":"34643788","CreationDate":"2016-01-06T22:10:23.100","ParentId":null,"OwnerUserId":"5498261","Title":null,"Body":"<p><strong>cmake<\/strong> appears to use Visual Studio 9 2008, but it \"wrongly\" uses <strong>mingw32-make.exe<\/strong> instead of maybe... <strong>nmake.exe<\/strong>.\nYou could run this command:<code>\"c:\\Program Files (x86)\\Microsoft Visual Studio 9.0\\VC\\bin\\vcvars32.bat\"<\/code> (adapt to your visual studio path) in the same console, before you execute the <code>luarocks<\/code> command. Now, <strong>cmake<\/strong> should use <strong>nmake<\/strong>.<\/p>\n"}]}
{"QuestionId":27288032,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2014-12-04T06:58:44.270","AcceptedAnswerId":27323988.0,"OwnerUserId":1926931.0,"Title":"How to create layer0 input for input images with 3 channels","Body":"<p>Hi I am following the <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/convolutional_mlp.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/convolutional_mlp.py<\/a> code to implement a conv neural net. I have input images where the channel is important and hence I want to have 3 channel feature map as layer 0 input.<\/p>\n\n<p>So I need something like this<\/p>\n\n<p><\/p>\n\n<p><em>instead of<\/em><\/p>\n\n<p><\/p>\n\n<p>which will be used here<\/p>\n\n<pre><\/pre>\n\n<p>where that x is provided to theano as  <\/p>\n\n<pre><\/pre>\n\n<p>So - my question is - how should I create (shape) that train_set_x ?<\/p>\n\n<p>With (gray scale intensity - i.e single channel) train_set_x is created as<\/p>\n\n<pre><\/pre>\n\n<p>where data_x is a flattened numpy array of length 784 (for 28*28 pixels)<\/p>\n\n<p>Thanks a lot for advice<\/p>\n","answers":[{"AnswerId":"27323988","CreationDate":"2014-12-05T20:15:49.867","ParentId":null,"OwnerUserId":"1926931","Title":null,"Body":"<p>I was able to get it working. I am pasting some code here which might help some one. Not very elegant - but works. <\/p>\n\n<p><code>\ndef shuffle_in_unison(a, b):\n     #courtsey http:\/\/stackoverflow.com\/users\/190280\/josh-bleecher-snyder\n    assert len(a) == len(b)\n    shuffled_a = np.empty(a.shape, dtype=a.dtype)\n    shuffled_b = np.empty(b.shape, dtype=b.dtype)\n    permutation = np.random.permutation(len(a))\n    for old_index, new_index in enumerate(permutation):\n        shuffled_a[new_index] = a[old_index]\n        shuffled_b[new_index] = b[old_index]\n    return shuffled_a, shuffled_b\n<\/code><\/p>\n\n<pre><code>def createDataSet(imagefolder):\n\nos.chdir(imagefolder)\n\n# total number of files\nnumber_of_files = len([item for item in os.listdir('.') if os.path.isfile(os.path.join('.', item))])\n\n# get a shuffled list : I needed this because my image names were of the format n_x_&lt;some details&gt;.jpg\n# where n was my target and x was a number from 0 to m-1 where m was the number of samples\n# of the target value n. So I needed so shuffle and iterate while putting images in train\n# test and validate arrays\nimage_index_array = range(0,number_of_files)\nrandom.seed(12)\nrandom.shuffle(image_index_array)\n# split 80\/10\/10 - train\/test\/val\ntrainsize = int(number_of_files*.8)\ntestsize = int(number_of_files*.1)\nvalsize = number_of_files - trainsize - testsize\n\n# create the random value arrays of train\/test\/val by slicing the total image index array\ntrain_index_array = image_index_array[0:trainsize]\ntest_index_array = image_index_array[trainsize:trainsize+testsize]\nvalidate_index_array = image_index_array[trainsize+testsize:]\n\n# initialize the data structures\ndataset = {'train':[[],[]],'test':[[],[]],'validate':[[],[]]}\n\ni_counter = 0\ntrain_X = []\ntrain_y = []\n\ntest_X = []\ntest_y = []\n\nval_X = []\nval_y = []\n\nfor item in os.listdir('.'):\n    if not os.path.isfile(os.path.join('.', item)):\n        continue\n\n    if item.endswith('.pkl'):\n        continue\n\n    print 'Processing item ' + item\n    item_y = item.split('_')[0]\n    item_x = cv2.imread(item)\n\n    height, width = item_x.shape[:2]\n\n    # this was my requirement - skip it if you do not need it\n    if(height != 135 or width != 240):\n        continue\n\n    # get 3 channels\n    b,g,r = cv2.split(item_x)\n\n    item_x = [b,g,r]\n    item_x = np.array(item_x)\n    item_x = item_x.reshape(3,135*240)\n\n    if i_counter in test_index_array:\n        test_X.append(item_x)\n        test_y.append(item_y)\n    elif i_counter in validate_index_array:\n        val_X.append(item_x)\n        val_y.append(item_y)\n    else:\n        train_X.append(item_x)\n        train_y.append(item_y)\n\n    i_counter = i_counter + 1\n\n# fix the dimensions. Flatten out the channel and intensity dimensions    \ntrain_X = np.array(train_X)\ntrain_X = train_X.reshape(train_X.shape[0],train_X.shape[1]*train_X.shape[2])\ntest_X = np.array(test_X)\ntest_X = test_X.reshape(test_X.shape[0],test_X.shape[1]*test_X.shape[2])\nval_X = np.array(val_X)\nval_X = val_X.reshape(val_X.shape[0],val_X.shape[1]*val_X.shape[2])\n\ntrain_y = np.array(train_y)\ntest_y = np.array(test_y)\nval_y = np.array(val_y)\n\n# shuffle the train and test arrays in unison\ntrain_X,train_y = shuffle_in_unison(train_X,train_y)\ntest_X,test_y = shuffle_in_unison(test_X,test_y)\n\n# pickle them\ndataset['train'] = [train_X,train_y]\ndataset['test'] = [test_X,test_y]\ndataset['validate'] = [val_X,val_y]\noutput = open('pcount.pkl', 'wb')\ncPickle.dump(dataset, output)\noutput.close`\n<\/code><\/pre>\n\n<p>Once you have this pickle file\nYou can use it in convolutional_mlp.py like this. <\/p>\n\n<pre><code>    layer0_input = x.reshape((batch_size, 3, 135, 240))\n\n# Construct the first convolutional pooling layer:\n# filtering reduces the image size to (135-8+1 , 240-5+1) = (128, 236)\n# maxpooling reduces this further to (128\/2, 236\/2) = (64, 118)\n# 4D output tensor is thus of shape (batch_size, nkerns[0], 64, 118)\nlayer0 = LeNetConvPoolLayer(\n    rng,\n    input=layer0_input,\n    image_shape=(batch_size, 3, 135, 240),\n    filter_shape=(nkerns[0], 3, 8, 5),\n    poolsize=(2, 2)\n)\n<\/code><\/pre>\n\n<p>The load_data function in logistic_sgd.py will need a small change as below<\/p>\n\n<pre><code>    f = open(dataset, 'rb')\ndump = cPickle.load(f)\ntrain_set = dump['train']\nvalid_set = dump['validate']\ntest_set = dump['test']\nf.close()\n<\/code><\/pre>\n\n<p>Hope this helps<\/p>\n"}]}
{"QuestionId":27288990,"AnswerCount":3,"Tags":"<python><python-3.x><windows-7-x64><theano>","CreationDate":"2014-12-04T08:04:24.370","AcceptedAnswerId":null,"OwnerUserId":1253251.0,"Title":"Theano installation for Windows, Python 3, 64bit","Body":"<p>Is there a procedure to install Theano for Python 3.4 64bit on Windows 7, manually, without using any of the bundles?<\/p>\n","answers":[{"AnswerId":"29552088","CreationDate":"2015-04-10T01:56:56.277","ParentId":null,"OwnerUserId":"3403018","Title":null,"Body":"<p>I wrote a step by step tutorial on this today, check it out at:<br>\n<a href=\"http:\/\/www.islandman93.com\/2015\/04\/tutorial-python-34-theano-and-windows-7.html\" rel=\"nofollow\">http:\/\/www.islandman93.com\/2015\/04\/tutorial-python-34-theano-and-windows-7.html<\/a><\/p>\n\n<p>Some specifics from my post:<\/p>\n\n<p>As user2805751 said, you'll need LibPython, Numpy, and Scipy.<\/p>\n\n<p>You'll also need: MinGW <a href=\"http:\/\/mingw-w64.yaxm.org\/doku.php\/download\" rel=\"nofollow\">http:\/\/mingw-w64.yaxm.org\/doku.php\/download<\/a><\/p>\n\n<p>Installing Theano with python setup.py install will automatically do the 2to3 conversion.<\/p>\n"},{"AnswerId":"27327337","CreationDate":"2014-12-06T01:36:35.733","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>There is no such instruction. There is a new PR with instruction that do not use any bundles, but it is for python2. You can try it, but change python2 for python3:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/pull\/2155\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/pull\/2155<\/a><\/p>\n"},{"AnswerId":"27328131","CreationDate":"2014-12-06T03:56:00.297","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>If the above PR doesn't work, I was able to get the following working for windows 8.1 64 bit (python 2.7.8, but you might try 3.4) <\/p>\n\n<p><a href=\"http:\/\/pavel.surmenok.com\/2014\/05\/31\/installing-theano-with-gpu-on-windows-64-bit\/\" rel=\"nofollow\">http:\/\/pavel.surmenok.com\/2014\/05\/31\/installing-theano-with-gpu-on-windows-64-bit\/<\/a><\/p>\n\n<p>I needed to also install the LibPython package from <a href=\"http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#libpython\" rel=\"nofollow\">http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#libpython<\/a> inorder to fix an ld linker problem.<\/p>\n"}]}
{"QuestionId":27343379,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2014-12-07T13:57:35.423","AcceptedAnswerId":27383389.0,"OwnerUserId":2278493.0,"Title":"RNN in generative mode with Theano (Scan op)","Body":"<p>I have a question regarding my RNN implementation.<\/p>\n\n<p>I have the following code<\/p>\n\n<pre><\/pre>\n\n<p>Now after training I could predict output of course as such:<\/p>\n\n<pre><\/pre>\n\n<p>However this is running the network in a predictive mode (i.e. take X steps of input and predict an output). I would like to run this in a generative mode, meaning that I want to start from an initial state and have the RNN run for any arbitrary amount of steps and feed its output back as input.<\/p>\n\n<p>How could I do this?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"27383389","CreationDate":"2014-12-09T16:02:15.810","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>You can run scan with an arbitrary number of steps using the <code>n_steps<\/code> parameter.  To pass <code>y_t<\/code> to the next computation -- and assuming <code>x_t.shape == y_t.shape<\/code> -- you can use <code>outputs_info=[h0, x_t]<\/code> as an argument to scan and modify the step function to <code>one_step(h_tm1, y_tm1, W_ih, W_hh, b_h, W_ho, b_o)<\/code> <\/p>\n\n<p>More complex termination criteria could be created with the <code>theano.scan_module.until()<\/code> but those are a question of design not implementation.  See <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">here<\/a> for examples.<\/p>\n"}]}
{"QuestionId":27357307,"AnswerCount":1,"Tags":"<neural-network><lmdb><caffe>","CreationDate":"2014-12-08T11:47:04.623","AcceptedAnswerId":27375961.0,"OwnerUserId":213615.0,"Title":"Caffe convert_imageset with one class\/ label","Body":"<p>I want to create an lmdb dataset from images which part of them contain the feature I want caffe to learn, and part of them don't.<br>\nMy question is - in the text input file transferred to convert_imageset - how should I label those images that don't contain the feature?<br>\nI know the format is<\/p>\n\n<pre><\/pre>\n\n<p>But which label should I assign to images <strong>without<\/strong> the feature?<br>\nFor example, img1.jpg contain the feature, img2.jpg and img3.jpg don't.\nSo should the text file look like -<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"27375961","CreationDate":"2014-12-09T09:54:19.587","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>Got an answer from <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/caffe-users\" rel=\"nofollow\">Caffe-users Google Group<\/a> - yes, creating a dummy feature is the right way for this.<br>\nSo it is:<\/p>\n\n<pre><code>img1.jpg 0\nimg2.jpg 1\nimg3.jpg 1\n<\/code><\/pre>\n"}]}
{"QuestionId":27382474,"AnswerCount":1,"Tags":"<pymc><theano><pymc3>","CreationDate":"2014-12-09T15:19:52.470","AcceptedAnswerId":31576199.0,"OwnerUserId":1509797.0,"Title":"pymc3: parallel computing with njobs>1 vs. GPU","Body":"<p>I am trying to speed-up pymc3 sampling with parallelisation and I see only modest benefit.<\/p>\n\n<p>I was able to decrease total running time from 25 minutes (njobs=1) to 13 minutes (njobs=6) on i7 MacBook Pro. Due to the fact that it takes about 4 minutes before pymc actually starts sampling, the increase is relatively small.<\/p>\n\n<p>The question is - does anyone successfully using GPU with pymc3 and how much benefit can I get for models that take 6-8 minutes to sample? (My MacBook has nvidia GT 750M 2Gb)<\/p>\n","answers":[{"AnswerId":"31576199","CreationDate":"2015-07-23T00:25:25.043","ParentId":null,"OwnerUserId":"1690016","Title":null,"Body":"<p>I'm running Linux on an Intel i7-4930.<\/p>\n\n<p>I ran a PyMC3 model that took 90 minutes on the CPU (utilizing all cores), but only took 18 minutes on my GeForce GTX 970.<\/p>\n\n<p>So a speed-up of almost 5x.<\/p>\n"}]}
{"QuestionId":27396664,"AnswerCount":2,"Tags":"<python><python-2.7><ubuntu><scikit-image><caffe>","CreationDate":"2014-12-10T08:49:26.193","AcceptedAnswerId":27396789.0,"OwnerUserId":1111479.0,"Title":"ImportError cannot import name BytesIO when import caffe on ubuntu","Body":"<p>I am trying to make <a href=\"http:\/\/caffe.berkeleyvision.org\/\">caffe<\/a> running on my machine equipped with Ubuntu 12.04LTS. \nAfter finishing all the steps on the <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\">Installation page<\/a>, I trained the LeNet model successfully and tried to use it as the tutorial from <a href=\"http:\/\/radar.oreilly.com\/2014\/07\/how-to-build-and-run-your-first-deep-learning-network.html\">here<\/a>. Then I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I set the  in  file before I did the above.\nWhat is the problem? Could anyone give some hint? I am really confused. \nAfter running the command in the very directory:<\/p>\n\n<pre><\/pre>\n\n<p>So, the problem becomes to: how to solve the name issue?\nP.S.: I also inserted an issue at <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1549\">the repository of caffe<\/a>. <\/p>\n","answers":[{"AnswerId":"34188511","CreationDate":"2015-12-09T20:37:44.233","ParentId":null,"OwnerUserId":"1695866","Title":null,"Body":"<p>I ran into this problem as well, installing caffe on an AWS ubuntu 14.04 instance following the script as outlined on the BVLC github repo here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Caffe-on-EC2-Ubuntu-14.04-Cuda-7\" rel=\"nofollow\" title=\"Caffe on EC2 Ubuntu 14.04\">\"Caffe on EC2 Ubuntu 14.04\"<\/a>.<\/p>\n\n<p>I have setup the python path as instructed. As diagnosed by @Martijn Pieters, the problem is that caffe is importing its own io library, which is then importing scikit-image's io library, which in turn is trying (but failing) to load the standard python io library (where BytesIO is located). Instead, due to the python path, when scikit-image tries to import BytesIO from the module io, it is circularly leading back to caffe's io module.<\/p>\n\n<p>I also found that even when not trying to import caffe, but due to having set my python path to include caffe, that this same problem hits me elsewhere.<\/p>\n\n<p>There are probably several ways to address this. But the essence is that the top-level import of caffe is at fault. To verify this, I altered the caffe code as follows:<\/p>\n\n<ol>\n<li><p>I renamed <code>...\/caffe\/io.py<\/code> module to <code>...\/caffe\/caffe_io.py<\/code> to be safe (although with correct namespace care, this shouldn't be necessary)<\/p><\/li>\n<li><p>I modified the import at the top of the <code>pycaffe.py<\/code> module from: <code>import caffe.io<\/code> to <code>import caffe.caffe_io<\/code><\/p><\/li>\n<li><p>I modified the import in <code>__init__.py<\/code> the same way (from <code>import caffe.io<\/code> to <code>import caffe.caffe_io<\/code>)<\/p><\/li>\n<\/ol>\n\n<p>Now, when you import io from python, it won't pick up the io library in caffe. When you import caffe, it will import its custom caffe_io library, and all should be well. You may want to do a more thorough scan through the python caffe modules to ensure I haven't overlooked other places where the import needs to change.<\/p>\n\n<p>I hope this helps. Perhaps when I have time, I'll issue a pull request with these (or similar) changes to the caffe github repo.<\/p>\n"},{"AnswerId":"27396789","CreationDate":"2014-12-10T08:56:52.597","ParentId":null,"OwnerUserId":"100297","Title":null,"Body":"<p>You appear to have a package or module named <code>io<\/code> in your Python path that is masking the standard library package. It is imported instead but doesn't have a <code>BytesIO<\/code> object to import.<\/p>\n\n<p>Try running:<\/p>\n\n<pre><code>python -c 'import io; print io.__file__'\n<\/code><\/pre>\n\n<p>in the same location you are running the tutorial and rename or move the file named by that import, presuming it is not the standard library version (ending in <code>lib\/python2.7\/io.pyc<\/code>).<\/p>\n\n<p>It could be you set your Python path to the wrong directory. You should include <code>path\/to\/caffe\/python<\/code>, not <code>path\/to\/caffe\/python\/caffe<\/code>, nor should you try and run python with the latter as your current working directory. In both cases <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py\" rel=\"noreferrer\"><code>caffe\/python\/caffe\/io.py<\/code><\/a> instead of the standard library version.<\/p>\n\n<p>The installation instructions are not at fault here; they clearly tell you to use:<\/p>\n\n\n\n<pre><code>export PYTHONPATH=\/path\/to\/caffe\/python:$PYTHONPATH\n<\/code><\/pre>\n\n<p>Note the lack of <code>\/caffe<\/code> at the end of that path.<\/p>\n"}]}
{"QuestionId":27413989,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2014-12-11T01:39:47.823","AcceptedAnswerId":27516347.0,"OwnerUserId":1191551.0,"Title":"Passing a symbolic theano.tensor into a compiled theano.function","Body":"<p>I'm trying to refactor my code to be easier to change architectures easily. Currently, I'm constructing a recurrent neural network as follows.<\/p>\n\n<pre><\/pre>\n\n<p>This works just fine. But the problem with this implementation is that I have to hard code the weights and make sure all the math is correct every time I change architectures. Inspired by the <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">Multilayer Perceptron tutorial<\/a>, I tried refactoring my code by introducing a Layer class.<\/p>\n\n<pre><\/pre>\n\n<p>This allows me to write the RNN code much cleaner, less error-prone, and a lot easier\nto change the architectures.<\/p>\n\n<pre><\/pre>\n\n<p>However, when I run my program I get an error<\/p>\n\n<pre><\/pre>\n\n<p>The problem here is the scan operation is passing a symbolic variable (a subtensor of Xs)\nto the compiled step function.<\/p>\n\n<p>The whole point of refactoring my code was so that I wouldn't have to define all of the computation within the step function. Now I am left with 4 symbolic variables (, , , ) that define a segment of the computational graph that I need to scan over with . However, I'm not sure how to do this because  cannot take in a symbolic variable. <\/p>\n\n<p>Here is a simplified example of what I am trying to do using the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">exponentiation example<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas how to get around this error?<\/p>\n","answers":[{"AnswerId":"27509494","CreationDate":"2014-12-16T16:36:20.050","ParentId":null,"OwnerUserId":"1675559","Title":null,"Body":"<p>You basically just can't use a compiled Theano function as a scan op.<\/p>\n\n<p>The way around this is to get your Layer classes to have a function that returns a function, which builds your computation tree, which you can then in turn use to compile the scan op. <\/p>\n"},{"AnswerId":"27516347","CreationDate":"2014-12-17T00:17:45.250","ParentId":null,"OwnerUserId":"1191551","Title":null,"Body":"<p>So the solution is to use <code>theano.clone<\/code> with the <code>replaces<\/code> keyword argument. For example, in the exponentiation example, you can define a step function as follows:<\/p>\n\n<pre><code>def step(p, a):\n    replaces = {prior_result: p, A: a}\n    n = theano.clone(next_result, replace=replaces)\n    return n\n<\/code><\/pre>\n"}]}
{"QuestionId":27436987,"AnswerCount":2,"Tags":"<image><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2014-12-12T04:40:59.780","AcceptedAnswerId":null,"OwnerUserId":3958348.0,"Title":"Caffe Multiple Input Images","Body":"<p>I'm looking at implementing a Caffe CNN which accepts two input images and a label (later perhaps other data) and was wondering if anyone was aware of the correct syntax in the prototxt file for doing this? Is it simply an IMAGE_DATA layer with additional tops? Or should I use separate IMAGE_DATA layers for each?<\/p>\n\n<p>Thanks,\nJames<\/p>\n","answers":[{"AnswerId":"30555159","CreationDate":"2015-05-31T08:01:14.683","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You may also consider using HDF5_DATA layer with multiple \"top\"s<\/p>\n"},{"AnswerId":"29162458","CreationDate":"2015-03-20T08:39:14.340","ParentId":null,"OwnerUserId":"2237635","Title":null,"Body":"<p>Edit: I have been using the HDF5_DATA layer lately for this and it is definitely the way to go. <\/p>\n\n<p>HDF5 is a key value store, where each key is a string, and each value is a multi-dimensional array. Thus, to use the HDF5_DATA layer, just add a new key for each top you want to use, and set the value for that key to store the image you want to use. Writing these HDF5 files from python is easy:<\/p>\n\n<pre><code>import h5py\nimport numpy as np\n\nfilelist = []\nfor i in range(100):\n    image1 = get_some_image(i)\n    image2 = get_another_image(i)\n    filename = '\/tmp\/my_hdf5%d.h5' % i\n    with hypy.File(filename, 'w') as f:\n        f['data1'] = np.transpose(image1, (2, 0, 1))\n        f['data2'] = np.transpose(image2, (2, 0, 1))\n    filelist.append(filename)\nwith open('\/tmp\/filelist.txt', 'w') as f:\n    for filename in filelist:\n        f.write(filename + '\\n')\n<\/code><\/pre>\n\n<p>Then simply set the source of the HDF5_DATA param to be '\/tmp\/filelist.txt', and set the tops to be \"data1\" and \"data2\".<\/p>\n\n<p>I'm leaving the original response below:<\/p>\n\n<p>====================================================<\/p>\n\n<p>There are two good ways of doing this. The easiest is probably to use two separate IMAGE_DATA layers, one with the first image and label, and a second with the second image. Caffe retrieves images from LMDB or LEVELDB, which are key value stores, and assuming you create your two databases with corresponding images having the same integer id key, Caffe will in fact load the images correctly, and you can proceed to construct your net with the data\/labels of both layers.<\/p>\n\n<p>The problem with this approach is that having two data layers is not really very satisfying, and it doesn't scale very well if you want to do more advanced things like having non-integer labels for things like bounding boxes, etc. If you're prepared to make a time investment in this, you can do a better job by modifying the tools\/convert_imageset.cpp file to stack images or other data across channels. For example you could create a datum with 6 channels - the first 3 for your first image's RGB, and the second 3 for your second image's RGB. After reading this in using the IMAGE_DATA layer, you can split the stream into two images using a SLICE layer with a slice_point at index 3 along the slice_dim = 1 dimension. If further down the road, you decide that you want to load even more complex assortments of data, you'll understand the encoding scheme and can write your own decoding layer based off of src\/caffe\/layers\/data_layer.cpp to gain full control of the pipeline.<\/p>\n"}]}
{"QuestionId":27438528,"AnswerCount":1,"Tags":"<integration><ros><caffe>","CreationDate":"2014-12-12T07:03:56.887","AcceptedAnswerId":null,"OwnerUserId":3958348.0,"Title":"Integrate Caffe model into ROS","Body":"<p>I'm looking at integrating a trained Caffe network into ROS and was wondering if anyone had any experience with this?<\/p>\n\n<p>Thanks,\nJames<\/p>\n","answers":[{"AnswerId":"30250973","CreationDate":"2015-05-15T03:18:31.250","ParentId":null,"OwnerUserId":"3113676","Title":null,"Body":"<p>Visit <a href=\"http:\/\/tzutalin.blogspot.tw\/2015\/06\/integrate-ros-into-caffe.html\" rel=\"nofollow\">http:\/\/tzutalin.blogspot.tw\/2015\/06\/integrate-ros-into-caffe.html<\/a>\nIntegrate Caffe into ROS indigo.<\/p>\n"}]}
{"QuestionId":27484484,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-12-15T12:48:12.363","AcceptedAnswerId":null,"OwnerUserId":2278493.0,"Title":"Error when using multiple taps using Scan in Theano","Body":"<p>I have made a simple RNN with Theano to test some stuff out. However for a small test I would like to add more taps to the hidden layer's output.<\/p>\n\n<p>However, this gives me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Below you can find the full code I am using to reproduce this error<\/p>\n\n<pre><\/pre>\n\n<p>This works fine without having the second tap added. (so only taps=[-1] and having only h_tm1 in the one_step() function.)<\/p>\n\n<p>Am I doing something wrong here or may this be a bug with Theano?<\/p>\n","answers":[{"AnswerId":"27509347","CreationDate":"2014-12-16T16:29:05.073","ParentId":null,"OwnerUserId":"1675559","Title":null,"Body":"<p>I'm not sure if this is your problem, but writing your <code>outputs_info<\/code> in this way expects only one return value of type matrix. The output sequence will then be a 3d tensor of (sequence_length,2,input_width).<\/p>\n"}]}
{"QuestionId":27497502,"AnswerCount":1,"Tags":"<python><python-2.7><theano>","CreationDate":"2014-12-16T04:25:21.690","AcceptedAnswerId":null,"OwnerUserId":4127806.0,"Title":"undefined symbol: ATL_chemv","Body":"<p>When I run this command<\/p>\n\n<p>sudo python -c \"import numpy; numpy.test()\"<\/p>\n\n<p>i get this error<\/p>\n\n<p>ImportError: \/usr\/lib\/liblapack.so.3gf: undefined symbol: ATL_chemv<\/p>\n\n<p>How can I fix it? <\/p>\n","answers":[{"AnswerId":"32422573","CreationDate":"2015-09-06T10:30:30.840","ParentId":null,"OwnerUserId":"5305715","Title":null,"Body":"<p>Actually, the answers in <a href=\"https:\/\/stackoverflow.com\/questions\/12249089\">this question<\/a> will help resolve things better, especially if you want to keep all the options. Most others will tell you to remove packages when you don't need to.<\/p>\n"}]}
{"QuestionId":27541733,"AnswerCount":1,"Tags":"<deep-learning><theano><caffe><logistic-regression><softmax>","CreationDate":"2014-12-18T08:00:45.000","AcceptedAnswerId":null,"OwnerUserId":1926931.0,"Title":"Loss function for ordinal target on SoftMax over Logistic Regression","Body":"<p>I am using Pylearn2 OR Caffe to build a deep network. My target is ordered nominal. I am trying to find a proper loss function but cannot find any in Pylearn2 or Caffe. <\/p>\n\n<p>I read a paper \"Loss Functions for Preference Levels: Regression with Discrete Ordered Labels\" . I get the general idea - but I am not sure I understand what will the thresholds be, if my final layer is a SoftMax over Logistic Regression (outputting probabilities). <\/p>\n\n<p>Can some help me by pointing to any implementation of such a loss function ?<\/p>\n\n<p>Thanks\nRegards<\/p>\n","answers":[{"AnswerId":"28131434","CreationDate":"2015-01-24T23:25:10.993","ParentId":null,"OwnerUserId":"1269942","Title":null,"Body":"<p>For both pylearn2 and caffe, your labels will need to be 0-4 instead of 1-5...it's just the way they work.  The output layer will be 5 units, each is a essentially a logistic unit...and the softmax can be thought of as an adaptor that normalizes the final outputs. But \"softmax\" is commonly used as an output type.  When training, the value of any individual unit is rarely ever exactly 0.0 or 1.0...it's always a distribution across your units - which log-loss can be calculated on. This loss is used to compare against the \"perfect\" case and the error is back-propped to update your network weights.  Note that a raw output from PL2 or Caffe is not a specific digit 0,1,2,3, or 5...it's 5 number, each associated to the likelihood of each of the 5 classes.  When classifying, one just takes the class with the highest value as the 'winner'. <\/p>\n\n<p>I'll try to give an example...\nsay I have a 3 class problem, I train a network with a 3 unit softmax.\nthe first unit represents the first class, second the second and third, third.<\/p>\n\n<p>Say I feed a test case through and get...<\/p>\n\n<p>0.25, 0.5, 0.25  ...0.5 is the highest, so a classifier would say \"2\". this is the softmax output...it makes sure the sum of the output units is one.<\/p>\n"}]}
{"QuestionId":27544698,"AnswerCount":1,"Tags":"<python><neural-network><backpropagation><theano>","CreationDate":"2014-12-18T10:45:20.757","AcceptedAnswerId":null,"OwnerUserId":4373828.0,"Title":"pure-python RNN and theano RNN computing different gradients -- code and results provided","Body":"<p>I've been banging my head against this for a while and can't figure out what I've done wrong (if anything) in implementing these RNNs. To spare you guys the forward phase, I can tell you that the two implementations compute the same outputs, so the forward phase is correct. The problem is in the backwards phase.<\/p>\n\n<p>Here is my python backwards code. It follows the style of karpathy's neuraltalk quite closely but not exactly:<\/p>\n\n<pre><\/pre>\n\n<p>And here's the theano code (mainly copied from another implementation I found online. I initialize the weights to my pure-python rnn's randomized weights so that everything is the same.):<\/p>\n\n<pre><\/pre>\n\n<p>Now here's the crazy thing. If i run the following:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>So, I'm getting the derivatives of the weight matrix between the hidden layer and output right, but not the derivatives of the weight matrix hidden -> hidden or input -> hidden. But this insane thing is that I ALWAYS get the LAST ROW of the weight matrix input -> hidden correct. This is insanity to me. I have no idea what's happening here. Note that the last row of the weight matrix input -> hidden does NOT correspond to the last timestep or anything (this would be explained, for example, by me calculating the derivatives correctly for the last timestep but not propagating back through time correctly). dcdWxh is the sum over all time steps of dcdWxh -- so how can I get one row of this correct but none of the others???<\/p>\n\n<p>Can anyone help? I'm all out of ideas here.<\/p>\n","answers":[{"AnswerId":"33660600","CreationDate":"2015-11-11T21:56:56.593","ParentId":null,"OwnerUserId":"2886336","Title":null,"Body":"<p>You should compute the sum of the <strong>pointwise absolute value<\/strong> of the difference of the two matrices. The plain sum could be close to zero due to the specific learning tasks (do you emulate the zero function? :), whichever that is.<\/p>\n\n<p>The last row presumably implements the weights from a constant-on-neuron, i.e. the bias, so you - seem to - always get the bias right (however, check the sum of absolute values) .<\/p>\n\n<p>It also looks like row-major and column-major notation of matrices are confused, like in<\/p>\n\n<pre><code>gWhx - bc['dcdWxh']\n<\/code><\/pre>\n\n<p>which reads like weight from \"hidden to x\" in opposite to \"x to hidden\".<\/p>\n\n<p>I'd rather post this as a comment, but I lack the reputation in doing so. sorry!<\/p>\n"}]}
{"QuestionId":27554484,"AnswerCount":2,"Tags":"<python-2.7><cuda><nvidia><theano><pylearn>","CreationDate":"2014-12-18T19:45:41.290","AcceptedAnswerId":null,"OwnerUserId":2423116.0,"Title":"import theano results in ImportError","Body":"<p>I'm trying to use theano but I get an error when I import it.\nI've installed cuda_6.5.14_linux_64.run, and passed all the recommended test in Chapter 6 of <a href=\"http:\/\/developer.download.nvidia.com\/compute\/cuda\/6_5\/rel\/docs\/CUDA_Getting_Started_Linux.pdf\" rel=\"nofollow\">this<\/a> NVIDIA PDF.\nUltimately I want to be able to install pylearn2, but I get the exact same error as below when I try to compile it.<\/p>\n\n<p>EDIT1: My theanorc looks like:<\/p>\n\n<pre><\/pre>\n\n<p>If I replace gpu with cpu, the command import theano succeeds.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29040199","CreationDate":"2015-03-13T19:13:09.187","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>We also saw this error.  We found that putting \/usr\/local\/cuda-6.5\/bin in $PATH seemed to fix it (even with the root = ... line in .theanorc).<\/p>\n"},{"AnswerId":"27833622","CreationDate":"2015-01-08T05:40:29.923","ParentId":null,"OwnerUserId":"2341715","Title":null,"Body":"<p>I encountered exactly the same question.<\/p>\n\n<p>My solution is to replace cuda-6.5 with cuda-5.5, and everything works fine.<\/p>\n"}]}
{"QuestionId":27600893,"AnswerCount":1,"Tags":"<python><matrix><theano>","CreationDate":"2014-12-22T10:32:18.887","AcceptedAnswerId":null,"OwnerUserId":2398548.0,"Title":"theano: summation by class label","Body":"<p>I have a matrix which represents a distances to the k-nearest neighbour of a set of points,\nand there is a matrix of class labels of the nearest neighbours. (both N-by-k matrix)<\/p>\n\n<p>What is the best way in theano to build a (N-by-#classes) matrix whose (i,j) element will be the sum of distances from i-th point to its k-NN points with the class label 'j'?<\/p>\n\n<p>Example:\n\n<\/p>\n\n<pre><\/pre>\n\n\n\n<p>this task in theano?<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n\n","answers":[{"AnswerId":"27603816","CreationDate":"2014-12-22T13:37:59.863","ParentId":null,"OwnerUserId":"764322","Title":null,"Body":"<p>You might be interesting in having a look to this repo: \n<a href=\"https:\/\/github.com\/erogol\/KLP_KMEANS\/blob\/master\/klp_kmeans.py\" rel=\"nofollow\">https:\/\/github.com\/erogol\/KLP_KMEANS\/blob\/master\/klp_kmeans.py<\/a><\/p>\n\n<p>Is a K-Means implementation using theano (func <code>kpl_kmeans<\/code>). I believe what you want is the matrix <code>W<\/code> used in the function <code>find_bmu<\/code>. <\/p>\n\n<p>Hope you find it useful.<\/p>\n"}]}
{"QuestionId":27609843,"AnswerCount":1,"Tags":"<python><theano><pylearn>","CreationDate":"2014-12-22T20:20:35.800","AcceptedAnswerId":null,"OwnerUserId":534080.0,"Title":"pylearn2 CSVDataset TypeError","Body":"<p>I'm having an issue loading a custom dataset into pylearn2. I'm trying to get a simple MLP to train using a tiny XOR dataset. I have a dataset named  in the same directory as my yaml file, which is not in the same directory as pylearn2's  script.<\/p>\n\n<p>Here's the entire contents of :<\/p>\n\n<pre><\/pre>\n\n<p>Here is the entire contents of my YAML file:<\/p>\n\n<pre><\/pre>\n\n<p>When I run pylearn2's  script, it fails before training (presumably while compiling the theano functions). Here's the entirety of the output:<\/p>\n\n<pre><\/pre>\n\n<p>What does this mean, exactly? I looked into the code for , and it loads the data in using , which should bring them in as floats. Nothing changes if I edit  to look like floats (, for example).<\/p>\n","answers":[{"AnswerId":"27621601","CreationDate":"2014-12-23T13:43:04.897","ParentId":null,"OwnerUserId":"4388766","Title":null,"Body":"<p>This is because type of y attribute of CSVDataset is set float64.<br>\nI've fixed __init__() of csv_dataset.py as follows, and it works.<br>\nI don't know this is pylearn2's problem or not.<\/p>\n\n<pre><code>    if self.task == 'regression':\n        super(CSVDataset, self).__init__(X=X, y=y)\n    else:\n        super(CSVDataset, self).__init__(X=X, y=y.astype(int),\n                                         y_labels=np.max(y) + 1)\n<\/code><\/pre>\n\n<p>BTW you should fix your yaml<\/p>\n\n<ul>\n<li>n_classes of Softmax layer should be 2<\/li>\n<li>\"channel_name: 'valid_y_misclass'\" causes error because you don't set \"valid\" attribute of monitoring_dataset.<br>\nTry to set \"valid\" of monitoring dataset, or use 'train_y_misclass' instead.<\/li>\n<\/ul>\n"}]}
{"QuestionId":27624812,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2014-12-23T17:01:42.577","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"Tiling Theano tensor","Body":"<p>I have two tensors A and B where the first one has size (500,10) and the second has size (500). I want to find A \/ B.\nI am using regular \/ operator and it Theano compiler says that they should be the same size.\nThen I tried using tensor.tile on B to make it the same size as A. It has three parameters(x, reps, ndim). I tried different values and I am limited to these constraints:\nx.ndim = len(reps) and ndim = len(reps)\nThen having these constraints how can I tile the array to a matrix?! Is this a bug in Theano?<\/p>\n","answers":[{"AnswerId":"27658318","CreationDate":"2014-12-26T14:23:27.687","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can just broadcast it, and there are several ways to do it. Take the following example<\/p>\n\n<pre><code>import numpy as np\nA = np.arange(1., 5001., 1.).reshape(500, 10)\nB = np.arange(1., 501., 1.)\n\nimport theano\nAs = theano.shared(A)\nBs = theano.shared(B)\n<\/code><\/pre>\n\n<p>The failsafe way of doing this is adding an appropriate axis<\/p>\n\n<pre><code>AoverB = A \/ B[:, np.newaxis]\nAoverBalso = A \/ B.reshape((-1, 1))\nAsoverBs = As \/ Bs.reshape((-1, 1))\n<\/code><\/pre>\n\n<p>Another way is to exploit the fact that there is implicit broadcasting to pad the first axes if they are missing<\/p>\n\n<pre><code>AoverBT = A.T \/ B.T   # no axis was added here\nAsoverBsT = As.T \/ Bs.T\n<\/code><\/pre>\n\n<p>In order to show that all these calculate the same thing, we use <code>numpy.testing<\/code><\/p>\n\n<pre><code>from numpy.testing import assert_array_equal\nassert_array_equal(AoverB, AoverBalso)\nassert_array_equal(AoverB, AsoverBs.eval())\nassert_array_equal(AoverB, AoverBT.T)\nassert_array_equal(AoverB, AsoverBsT.T.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":27629347,"AnswerCount":3,"Tags":"<python><python-3.x><anaconda><theano><conda>","CreationDate":"2014-12-23T23:10:04.367","AcceptedAnswerId":null,"OwnerUserId":4386560.0,"Title":"How do I install theano in Anaconda ver. 2.1 Windows 64 bit for Python 3.4?","Body":"<p>I've installed <a href=\"https:\/\/en.wikipedia.org\/wiki\/Anaconda_%28Python_distribution%29\" rel=\"nofollow noreferrer\">Anaconda<\/a>. Now I want to install the Theano library in Anaconda. I have tried:<\/p>\n\n<ol>\n<li><p>The Theano installer for Anaconda from <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#windows-installer-for-anacondace\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install.html#windows-installer-for-anacondace<\/a>, but it raises error \"The installer could not find a version of Anaconda installed. Please download and install Anaconda CE\". I have added ~\/anaconda3, ~\/anaconda3\/scripts to the environment variable path.<\/p><\/li>\n<li><p>I have tried to install it by building the package as mentioned on Stack&nbsp;Overflow, <em><a href=\"https:\/\/stackoverflow.com\/questions\/18640305\">How do I keep track of pip-installed packages in an Anaconda (Conda) environment?<\/a><\/em>, but this also fails during the testing of package. The error screenshot is below:<\/p><\/li>\n<\/ol>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/80vxe.jpg\" alt=\"Theano error message\"><\/p>\n\n<p>I have even installed Python ver. 3.4, installed Theano using pip install, and when I tried to import Theano it gave an error similar to the error in the screen shot. I tried the changes mentioned in this <a href=\"https:\/\/en.wikipedia.org\/wiki\/Google_Groups\" rel=\"nofollow noreferrer\">Google Groups<\/a> discussion, <strong><a href=\"http:\/\/groups.google.com\/d\/msg\/theano-users\/RcANq0NQgPE\/IQMHD3mggi0J\" rel=\"nofollow noreferrer\">Re: [theano-users] Install Theano on Windows for Python 3<\/a><\/strong>, but no luck.<\/p>\n","answers":[{"AnswerId":"42214273","CreationDate":"2017-02-13T21:56:08.553","ParentId":null,"OwnerUserId":"1887281","Title":null,"Body":"<p>Note to moderator: This is NOT a duplicate post. All of my other posts were deleted, so I'm leaving this one here and will flag the other questions as duplicate. <\/p>\n\n<p>I could never get a working installation of Theano using Anaconda with Python 3.4, and I also could never get the manual installation working with MinGW, but I was able to get it working flawlessly using WinPython 3.4. <\/p>\n\n<p><strong>Theano Installation and Configuration on Windows 10 with GPU Acceleration and Python 3.4<\/strong><\/p>\n\n<p>If you're using Windows, Theano can be tricky to install and configure. I was able to get it working by following a combination of these tutorials:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/ankivil.com\/installing-keras-theano-and-dependencies-on-windows-10\/\" rel=\"nofollow noreferrer\" title=\"Installing Keras and Theano with GPU-acceleration on Windows 10\">Installing Keras and Theano with GPU-acceleration on Windows 10<\/a><\/li>\n<li><a href=\"http:\/\/ankivil.com\/making-theano-faster-with-cudnn-and-cnmem-on-windows-10\/\" rel=\"nofollow noreferrer\" title=\"Making Theano faster on Windows 10 with CuDNN and CNMeM\">Making Theano faster on Windows 10 with CuDNN and CNMeM<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano_versions\/dev\/install_windows.html\" rel=\"nofollow noreferrer\" title=\"Official Theano installation instructions for Windows\">Official Theano installation instructions for Windows<\/a><\/li>\n<\/ul>\n\n<p><strong>Easier configuration of Theano with Python 3.4 using WinPython instead of Anaconda Python<\/strong><\/p>\n\n<p>It was much easier to get Theano working on Python 3.4 when using <a href=\"http:\/\/winpython.sourceforge.net\/\" rel=\"nofollow noreferrer\" title=\"WinPython download\">WinPython<\/a> instead of <a href=\"https:\/\/www.continuum.io\/downloads\" rel=\"nofollow noreferrer\" title=\"Anaconda Python download\">Anaconda Python<\/a>, but WinPython stores environment settings in its settings directory (e.g. <code>C:\\SciSoft\\WinPython-64bit-3.4.4.2\\settings\\.keras\\<\/code>) rather than looking in your <code>%USERPROFILE%<\/code> for the keras.json file when you're wanting it to pick up your environment settings (as explained in the setup guides). Also, if you are still having trouble, you might just need to set the <code>THEANO_FLAGS<\/code> system environment variable to something like this: <code>floatX=float32,device=gpu,nvcc.fastmath=True,lib.cnmem=0.8,blas.ldflags=-LC:\\src\\OpenBLAS -lopenblas<\/code>. (Note that this environment variable overrides the settings in any .theanorc setup file as detailed <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow noreferrer\" title=\"Official Theano configuration instructions\">here in the Theano configuration documentation<\/a> except if using WinPython, the .theanorc file would go into <code>C:\\SciSoft\\WinPython-64bit-3.4.4.2\\settings\\.theanorc<\/code> rather than <code>%USERPROFILE\\.theanorc<\/code>.)<\/p>\n\n<p>When installing Theano with WinPython, installation is much easier if you use the suggested Theano installation location (<code>C:\\SciSoft\\<\/code>). In that case, your installation directory should look like this:<\/p>\n\n<p><a href=\"http:\/\/ankivil.com\/installing-keras-theano-and-dependencies-on-windows-10\/\" rel=\"nofollow noreferrer\" title=\"Installing Keras and Theano with GPU-acceleration on Windows 10\"><img src=\"https:\/\/i.stack.imgur.com\/gnwme.png\" alt=\"Picture of SciSoft installation directory\" title=\"Image of installation directory for Theano\"><\/a><\/p>\n\n<p><strong>Fixing bugs in Theano environment batch file when using WinPython<\/strong><\/p>\n\n<p>The other issue I ran into with the Theano installation guides is that the batch script had some bugs in it that were causing the dependency paths to be incorrect. Here was my final version of the <code>env.bat<\/code> file:<\/p>\n\n<pre><code>REM configuration of paths\nset VSFORPYTHON=\"C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\"\nset SCISOFT=%~dp0\n\nREM add tdm gcc stuff\nset PATH=%SCISOFT%TDM-GCC-64\\bin;%SCISOFT%TDM-GCC-64\\x86_64-w64-mingw32\\bin;%PATH%\n\nREM add winpython stuff\nCALL %SCISOFT%WinPython-64bit-3.4.4.2\\scripts\\env.bat\n\nREM configure path for msvc compilers\nREM for a 32 bit installation change this line to\nREM CALL %VSFORPYTHON%\\vcvarsall.bat\nCALL %VSFORPYTHON%\\vcvarsall.bat amd64\n\nREM return a shell\ncmd.exe \/k\n<\/code><\/pre>\n\n<p>If using Theano, your .keras file will need to be setup like:<\/p>\n\n<pre><code>{\n    \"floatx\": \"float32\",\n    \"epsilon\": 1e-07,\n    \"image_dim_ordering\": \"th\",\n    \"backend\": \"theano\" \n}\n<\/code><\/pre>\n\n<p><strong>Issue with installing CuDNN<\/strong><\/p>\n\n<p>Another key thing was that the CuDNN DLLs need to be copied into their corresponding folders in the CUDA installation directory in order for them to be detected. Instructions are detailed here: <a href=\"https:\/\/stackoverflow.com\/questions\/36248056\/how-to-setup-cudnn-with-theano-on-windows-7-64-bit\" title=\"How to install CuDNN into CUDA on Windows\">Instructions for installing CuDNN into CUDA on Windows<\/a><\/p>\n\n<p><strong>If still having issues with Theano installation on Windows with Python 3.4:<\/strong><\/p>\n\n<p>Then please review the information here: <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/3376\" rel=\"nofollow noreferrer\" title=\"How to get Theano working with Python 3.5\">Full installation guide for Theano on Windows with Python 3.4, including all required environment variables and PATH directories<\/a><\/p>\n\n<p><strong>Another key issue with installing the C++ dependencies for Theano<\/strong><\/p>\n\n<p>Another thing I was tripped up by is that in the <a href=\"http:\/\/deeplearning.net\/software\/theano_versions\/dev\/install_windows.html\" rel=\"nofollow noreferrer\" title=\"Official Theano Windows installation instructions\">official Theano documentation<\/a>, it provides very specific instructions on installing the <a href=\"http:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=44266\" rel=\"nofollow noreferrer\" title=\"Official Microsoft Visual C++ libraries required for Python\">Microsoft Visual C++ Compiler for Python 2.7<\/a>. It seems to be that this compiler is also required to be installed <em>in exactly the way that the Theano documentation specifies to perform the installation on the command line<\/em> to get Python 3.4 to work. I will quote the official Theano documentation, which states:<\/p>\n\n<blockquote>\n  <ol>\n  <li>open an administrator\u2019s console (got to <code>start<\/code>, then type <code>cmd<\/code>,\n  right-click on the command prompt icon and select <code>run as\n  administrator<\/code>) <\/li>\n  <li><code>cd<\/code> to your downloads directory and execute <code>msiexec \/i\n  VCForPython27.msi ALLUSERS=1<\/code><\/li>\n  <\/ol>\n<\/blockquote>\n\n<p><strong>General advice about GPU-acceleration<\/strong><\/p>\n\n<p>And FYI, if you haven't tried configuring a neural network library, I highly recommend that you use GPU-acceleration. <\/p>\n"},{"AnswerId":"36744339","CreationDate":"2016-04-20T12:49:13.953","ParentId":null,"OwnerUserId":"3667840","Title":null,"Body":"<p>As we can see, you have tried to use Theano under Windows. Please, ensure that you have a <a href=\"http:\/\/en.wikipedia.org\/wiki\/MinGW\" rel=\"nofollow noreferrer\">MinGW<\/a> compiler. Further, ensure that you have MinGW and libpython packages.<\/p>\n\n<p>Generally, I recommend to use the answer <em><a href=\"https:\/\/stackoverflow.com\/questions\/34097988\/how-to-install-keras-and-theano-in-anaconda-python-2-7-in-windows\/34975902#34975902\">How do I install Keras and Theano in Anaconda Python 2.7 on Windows?<\/a><\/em>, but without the last step.<\/p>\n"},{"AnswerId":"27752192","CreationDate":"2015-01-03T05:57:36.073","ParentId":null,"OwnerUserId":"2900367","Title":null,"Body":"<p>Running Theano on Python 3.4 is complicated. So far I would recommend that you run Theano in Python 2.7. The libraries written for Theano are Python 2.6+ based. So in order to get Theano running in Python 3.4, you would be needing the 2to3 automated python 2 to 3 code translation tool. I haven't tested Theano using 2to3, so I can't comment on whether it would work or not. But, I am using Python 2.7 and Theano works smoothly. Also, you might want to use AnacondaCE with Python 2.7 installer which pretty much gives you everything you need to start developing.<\/p>\n\n<p>You would also need to reinstall Theano using<\/p>\n\n<pre><code>pip install Theano\n<\/code><\/pre>\n"}]}
{"QuestionId":27632440,"AnswerCount":3,"Tags":"<numpy><neural-network><protocol-buffers><deep-learning><caffe>","CreationDate":"2014-12-24T06:20:31.667","AcceptedAnswerId":27645934.0,"OwnerUserId":1714410.0,"Title":"InfogainLoss layer","Body":"<p>I wish to use a loss layer of type <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1InfogainLossLayer.html#details\" rel=\"noreferrer\"><\/a> in my model. But I am having difficulties defining it properly.<\/p>\n\n<ol>\n<li><p>Is there any tutorial\/example on the usage of  layer?<\/p><\/li>\n<li><p>Should the input to this layer, the class probabilities, be the output of a  layer, or is it enough to input the \"top\" of a fully connected layer?<\/p><\/li>\n<\/ol>\n\n<p> requires three inputs: class probabilities, labels and the matrix .\nThe matrix  can be provided either as a layer parameters .<br>\nSuppose I have a python script that computes  as a  of shape  with  (where  is the number of labels in my model).<\/p>\n\n<ol start=\"3\">\n<li><p>How can I convert my  into a  file that can be provided as a  to the model?<\/p><\/li>\n<li><p>Suppose I want  to be provided as the third input (bottom) to the loss layer (rather than as a model parameter). How can I do this?<br>\nDo I define a new data layer which \"top\" is ? If so, wouldn't the data of this layer be incremented every training iteration like the training data is incremented?\nHow can I define multiple unrelated input \"data\" layers, and how does caffe know to read from the training\/testing \"data\" layer batch after batch, while from the  \"data\" layer it knows to read only once for all the training process?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"27645934","CreationDate":"2014-12-25T09:27:10.477","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><em>1. Is there any tutorial\/example on the usage of <strong>InfogainLoss<\/strong> layer?<\/em>:<br>\nA nice example can be found <a href=\"https:\/\/stackoverflow.com\/questions\/30486033\/tackling-class-imbalance-scaling-contribution-to-loss-and-sgd?lq=1\">here<\/a>: using <strong>InfogainLoss<\/strong> to tackle class imbalance.<\/p>\n\n<hr>\n\n<p><em>2. Should the input to this layer, the class probabilities, be the output of a <strong>Softmax<\/strong> layer?<\/em><br>\nHistorically, the answer used to be <em>YES<\/em> according to <a href=\"https:\/\/stackoverflow.com\/a\/30037614\/1714410\">Yair's answer<\/a>. The old implementation of <code>\"InfogainLoss\"<\/code> needed to be the output of <code>\"Softmax\"<\/code> layer or any other layer that makes sure the input values are in range [0..1].  <\/p>\n\n<p>The <a href=\"https:\/\/stackoverflow.com\/users\/1714410\/shai\">OP<\/a> noticed that using <code>\"InfogainLoss\"<\/code> on top of <code>\"Softmax\"<\/code> layer can lead to numerical instability. <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3855\" rel=\"nofollow noreferrer\">His pull request<\/a>, combining these two layers into a single one (much like <code>\"SoftmaxWithLoss\"<\/code> layer), was accepted and merged into the official Caffe repositories on 14\/04\/2017. The mathematics of this combined layer are given <a href=\"https:\/\/stackoverflow.com\/a\/34917052\/1714410\">here<\/a>.<\/p>\n\n<p>The upgraded layer \"look and feel\" is exactly like the old one, apart from the fact that <strong>one no longer needs to explicitly pass the input through a <code>\"Softmax\"<\/code> layer.<\/strong><\/p>\n\n<hr>\n\n<p><em>3. How can I convert an numpy.array into a binproto file<\/em>:  <\/p>\n\n<p>In python<\/p>\n\n<pre><code>H = np.eye( L, dtype = 'f4' ) \nimport caffe\nblob = caffe.io.array_to_blobproto( H.reshape( (1,1,L,L) ) )\nwith open( 'infogainH.binaryproto', 'wb' ) as f :\n    f.write( blob.SerializeToString() )\n<\/code><\/pre>\n\n<p>Now you can add to the model prototext the <code>INFOGAIN_LOSS<\/code> layer with <code>H<\/code> as a parameter:<\/p>\n\n<pre><code>layer {\n  bottom: \"topOfPrevLayer\"\n  bottom: \"label\"\n  top: \"infoGainLoss\"\n  name: \"infoGainLoss\"\n  type: \"InfogainLoss\"\n  infogain_loss_param {\n    source: \"infogainH.binaryproto\"\n  }\n}\n<\/code><\/pre>\n\n<hr>\n\n<p><em>4. How to load <code>H<\/code> as part of a DATA layer<\/em>  <\/p>\n\n<p>Quoting <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1640\" rel=\"nofollow noreferrer\">Evan Shelhamer's post<\/a>:<\/p>\n\n<blockquote>\n  <p>There's no way at present to make data layers load input at different rates. Every forward pass all data layers will advance. However, the constant H input could be done by making an input lmdb \/ leveldb \/ hdf5 file that is only H since the data layer will loop and keep loading the same H. This obviously wastes disk IO. <\/p>\n<\/blockquote>\n"},{"AnswerId":"30037614","CreationDate":"2015-05-04T18:44:06.337","ParentId":null,"OwnerUserId":"101433","Title":null,"Body":"<p>The layer is summing up <\/p>\n\n<pre><code>-log(p_i)\n<\/code><\/pre>\n\n<p>and so the p_i's need to be in (0, 1] to make sense as a loss function (otherwise higher confidence scores will produce a higher loss). See the curve below for the values of log(p).<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/qxmZO.png\" alt=\"enter image description here\"><\/p>\n\n<p>I don't think they have to sum up to 1, but passing them through a Softmax layer will achieve both properties.<\/p>\n"},{"AnswerId":"51641525","CreationDate":"2018-08-01T20:07:50.010","ParentId":null,"OwnerUserId":"844293","Title":null,"Body":"<p>Since I had to search through many websites to puzzle the complete\ncode, I thought I share my implementation:<\/p>\n\n<p>Python layer for computing the H-matrix with weights for each class:<\/p>\n\n<pre><code>import numpy as np\nimport caffe\n\n\nclass ComputeH(caffe.Layer):\n    def __init__(self, p_object, *args, **kwargs):\n        super(ComputeH, self).__init__(p_object, *args, **kwargs)\n        self.n_classes = -1\n\n    def setup(self, bottom, top):\n        if len(bottom) != 1:\n            raise Exception(\"Need (only) one input to compute H matrix.\")\n\n        params = eval(self.param_str)\n        if 'n_classes' in params:\n            self.n_classes = int(params['n_classes'])\n        else:\n            raise Exception('The number of classes (n_classes) must be specified.')\n\n    def reshape(self, bottom, top):\n        top[0].reshape(1, 1, self.n_classes, self.n_classes)\n\n    def forward(self, bottom, top):\n        classes, cls_num = np.unique(bottom[0].data, return_counts=True)\n\n        if np.size(classes) != self.n_classes or self.n_classes == -1:\n            raise Exception(\"Invalid number of classes\")\n\n        cls_num = cls_num.astype(float)\n\n        cls_num = cls_num.max() \/ cls_num\n        weights = cls_num \/ np.sum(cls_num)\n\n        top[0].data[...] = np.diag(weights)\n\n    def backward(self, top, propagate_down, bottom):\n        pass\n<\/code><\/pre>\n\n<p>and the relevant part from the train_val.prototxt:<\/p>\n\n<pre><code>layer {\n    name: \"computeH\"\n    bottom: \"label\"\n    top: \"H\"\n    type: \"Python\"\n    python_param {\n        module: \"digits_python_layers\"\n        layer: \"ComputeH\"\n        param_str: '{\"n_classes\": 7}'\n    }\n    exclude { stage: \"deploy\" }\n}\nlayer {\n  name: \"loss\"\n  type: \"InfogainLoss\"\n  bottom: \"score\"\n  bottom: \"label\"\n  bottom: \"H\"\n  top: \"loss\"\n  infogain_loss_param {\n    axis: 1  # compute loss and probability along axis\n  }\n  loss_param {\n      normalization: 0\n  }\n  exclude {\n    stage: \"deploy\"\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":27661975,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2014-12-26T21:05:39.587","AcceptedAnswerId":27728366.0,"OwnerUserId":4396662.0,"Title":"ValueError : Length Not Known while mapping a list in Theano","Body":"<p>I am a beginner in Theano but despite a lot of research I still don't understand why the following code <\/p>\n\n<pre><\/pre>\n\n<p>returns this error message<\/p>\n\n<pre><\/pre>\n\n<p>When I type T.sqrt(A) in the console it returns \"Elemwise{sqrt,no_inplace}.0\" and I can't get the value of T.sqrt(A). \nI wonder if I can use matplotlib with theano tensors and otherwise how I can get a numpy array list mapped by a theano function back into the ndarray type.<\/p>\n","answers":[{"AnswerId":"27728366","CreationDate":"2015-01-01T04:55:36.990","ParentId":null,"OwnerUserId":"2580227","Title":null,"Body":"<p>The reason you get \"Elemwise{sqrt,no_inplace}.0\" is because Theano is symbolic, so tensors do not have a value until evaluated. To fix this:<\/p>\n\n<pre><code>from matplotlib.pyplot import plot\nfrom numpy import linspace\nimport theano.tensor as T\nA = linspace(0,1,100)\nplot(A, T.sqrt(A).eval())\n<\/code><\/pre>\n\n<p>to evaluate the tensor and get its numeric value.<\/p>\n\n<p>Also, you should import theano.tensor, not theano.tensors.<\/p>\n"}]}
{"QuestionId":27690125,"AnswerCount":2,"Tags":"<c++><caffe>","CreationDate":"2014-12-29T13:48:02.070","AcceptedAnswerId":27690893.0,"OwnerUserId":4402334.0,"Title":"caffe.cpp RegisterBrewFunction","Body":"<p>I was reading the caffe source codes. In the caffe.cpp which is the source of tools\/caffe, I encounter the following codes which puzzles me:<\/p>\n\n<pre><\/pre>\n\n<p>Based on my knowledge, this macro replace RegisterBrewFunction(func) with a anonymous class, and the only thing it has done is add &lt;#func, &amp;func> to g_brew_map.\nSo why not just doing this like this?<\/p>\n\n<pre><\/pre>\n\n<p>Hope someone help me about this.<\/p>\n","answers":[{"AnswerId":"27690893","CreationDate":"2014-12-29T14:40:09.110","ParentId":null,"OwnerUserId":"266487","Title":null,"Body":"<p>Your <code>#define<\/code> defines a macro that can be used inside a function. When the code reaches this function then the macro argument will be registered. You have to call the function for brewing the coffee.<\/p>\n\n<p>The original <code>#define<\/code> defines a macro that shall be used at file scope. It creates a more or less anonymous object. The constructor of that object's class will run before <code>main<\/code> is called and registers the function. There is no other function necessary.<\/p>\n"},{"AnswerId":"27693465","CreationDate":"2014-12-29T17:37:47.777","ParentId":null,"OwnerUserId":"560648","Title":null,"Body":"<p>Because an assignment may not live at global scope, only inside a function.<\/p>\n\n<p>This macro can be used in many more places, and indeed will normally live at the top of some source file so that you don't have to call a function to make it work.<\/p>\n"}]}
{"QuestionId":27698386,"AnswerCount":1,"Tags":"<python><testing><theano>","CreationDate":"2014-12-30T01:17:04.543","AcceptedAnswerId":27822231.0,"OwnerUserId":2644074.0,"Title":"Why do the Theano tests fail with many \"KnownFailureTest\"s?","Body":"<p>Theano is failing it's tests when I do:<\/p>\n\n<pre><\/pre>\n\n<p>If these are known failures, shouldn't it still pass? IE when I test other libraries, KnownFailures sometimes trigger, but the overall test still passes with \"OK\" (but will still note the KnownFails and Skipped tests).<\/p>\n\n<p>My guess is this is ok, and the test really is \"passing\", but since I'm doing a fresh install following the deeplearning.net tutorials, and I'm getting this error, I assume others might have this question as well, and a search on Google, and SO, isn't really helpful.<\/p>\n\n<p>Forgive the error-code-dump, I am sure no one will need to read all through this, but it's here for reference if someone else has this question. Here are the errors at the end of the tests:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"27822231","CreationDate":"2015-01-07T15:01:32.267","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>KnownFailureTest are a valid return value for nosetests. When Theano started, we where creating tests for features to implement and raised KnownFailureTest in them until we implement them. We do not do that anymore as we end up with to much questions from people about this. So this cause too much distraction. But we didn't changed the old tests that did that.<\/p>\n\n<p>I just created an issue to change that: <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2375\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2375<\/a><\/p>\n\n<p>I do not know when it will be changed.<\/p>\n"}]}
{"QuestionId":27720437,"AnswerCount":1,"Tags":"<c++><opencv><caffe>","CreationDate":"2014-12-31T11:24:36.997","AcceptedAnswerId":null,"OwnerUserId":648896.0,"Title":"How can I read opencv images into caffe format for real-time prediction?","Body":"<p>I try to create a real-time prediction program from video frames. All the code is backended by opencv. However, I trained a caffe model and I try to use it by this <a href=\"https:\/\/github.com\/niuzhiheng\/caffe\" rel=\"nofollow\">windows port<\/a> of caffe. It does not support opencv to caffe conversion of the images in memory. How can I do it externally? In the original caffe there is a solution that is <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/80c9a2c17f5469371b434fde4072bc13b49ee1eb\" rel=\"nofollow\">recently merged<\/a> but cannot apply this to this windows version.<\/p>\n","answers":[{"AnswerId":"30557410","CreationDate":"2015-05-31T12:27:11.957","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You should be able to define the input to your network using a <strong><a href=\"https:\/\/github.com\/niuzhiheng\/caffe\/blob\/windows\/src\/caffe\/layers\/memory_data_layer.cpp\" rel=\"nofollow\">MEMORY_DATA<\/a><\/strong> layer, then, using the <code>Reset<\/code> method you should be able to set the input according to the data stored in your opencv representation.<\/p>\n"}]}
{"QuestionId":27732543,"AnswerCount":1,"Tags":"<python><numpy><random><cuda><theano>","CreationDate":"2015-01-01T16:32:07.490","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"Why does creation of a Theano shared variable on GPU effect numpy's random streams?","Body":"<p>I'm just starting to play with Theano, and am wondering why the first creation of a shared variable on the gpu seems to effect numpy's random number generator. At times this initial creation seems to advance the random number generator.<\/p>\n\n<p>I've explored the following test cases in this code:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>rand_test0 - A sanity check to show I can reset a random stream using numpy.random.seed<\/li>\n<li>rand_test1 - Show creation of shared variable on cpu does nothing unexpected<\/li>\n<li>rand_test2 - Show creation of shared variable on gpu does have an unexpected effect<\/li>\n<li>rand_test3 - Show it is only the initial creation of the shared variable on the gpu with an unexpected effect<\/li>\n<li>rand_test4 - A verification of rand_test3<\/li>\n<\/ol>\n\n<p>The results I got were as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Does this make sense to anyone? Is it a Theano artifact? Is it a CUDA artifact, caused by my initial access to the GPU (i.e. the fact that I was playing with shared variables is only incidental to what I'm seeing). Or, am I misunderstanding something else?<\/p>\n","answers":[{"AnswerId":"32687028","CreationDate":"2015-09-21T03:43:40.663","ParentId":null,"OwnerUserId":"3299394","Title":null,"Body":"<p>Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/sandbox\/randomnumbers.html\" rel=\"nofollow\">documentation<\/a> talks about the difficulties of seeding random variables and why they seed each graph instance with its own random number generator. <\/p>\n\n<blockquote>\n  <p>Sharing a random number generator between different {{{RandomOp}}}\n  instances makes it difficult to producing the same stream regardless\n  of other ops in graph, and to keep {{{RandomOps}}} isolated.\n  Therefore, each {{{RandomOp}}} instance in a graph will have its very\n  own random number generator. That random number generator is an input\n  to the function. In typical usage, we will use the new features of\n  function inputs ({{{value}}}, {{{update}}}) to pass and update the rng\n  for each {{{RandomOp}}}. By passing RNGs as inputs, it is possible to\n  use the normal methods of accessing function inputs to access each\n  {{{RandomOp}}}\u2019s rng. In this approach it there is no pre-existing\n  mechanism to work with the combined random number state of an entire\n  graph. So the proposal is to provide the missing functionality (the\n  last three requirements) via auxiliary functions: {{{seed, getstate,\n  setstate}}}.<\/p>\n<\/blockquote>\n\n<p>They also provide <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html\" rel=\"nofollow\">examples<\/a> on how to seed all the random number generators. <\/p>\n\n<blockquote>\n  <p>You can also seed all of the random variables allocated by a\n  RandomStreams object by that object\u2019s seed method. This seed will be\n  used to seed a temporary random number generator, that will in turn\n  generate seeds for each of the random variables.<\/p>\n<\/blockquote>\n\n<pre><code>&gt;&gt;&gt; srng.seed(902340)  # seeds rv_u and rv_n with different seeds each\n<\/code><\/pre>\n\n<p>Try seeding the random variable's using Theano's seed functionality instead of numpy's. <\/p>\n"}]}
{"QuestionId":27756908,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-01-03T16:27:08.187","AcceptedAnswerId":null,"OwnerUserId":2278493.0,"Title":"Random numbers error in scan, theano.gof.fg.MissingInputError","Body":"<p>I have a small problem using random numbers together with scan.<\/p>\n\n<p>Please see this small example in which I tried to isolate my problem.<\/p>\n\n<pre><\/pre>\n\n<p>What this code should do is the following:\n- begin with an array, in this case [1,2,3,4,5]\n- generate random numbers sampled from the normal distribution with the average being the average of the previous output (or the initial observations)\n  In this case the average for the first step would be 3.\n- Let's say the sampled numbers are: [2,3,3.5,4,5], the new average is now 3.5\n- Repeat the above for 10 timesteps<\/p>\n\n<p>Instead I get the following error output:<\/p>\n\n<pre><\/pre>\n\n<p>I am probably missing something simple and obvious again here.<\/p>\n\n<p>Help is much appreciated, thanks!<\/p>\n","answers":[{"AnswerId":"28864024","CreationDate":"2015-03-04T20:15:16.470","ParentId":null,"OwnerUserId":"554606","Title":null,"Body":"<p>You need to pass the updates dictionary that you get form scan to your function, so:<\/p>\n\n<pre><code>f=th.function([initials], result, updates=updates)\n<\/code><\/pre>\n\n<p>Also, you can't have shared variables as input to your function.<\/p>\n\n<p>You can achieve what you are trying to do for example like this:<\/p>\n\n<pre><code>import theano as th\nimport numpy as np\nfrom theano import tensor as T\n\nstream=th.tensor.shared_randomstreams.RandomStreams()\n\navg=T.vector()\ninitials=T.fvector()\n\ndef get_output(prev_rand):\n    return stream.normal(size=prev_rand.shape, avg=prev_rand.mean())\n\nresult, updates = th.scan(get_output, outputs_info=[initials], n_steps=10)\n\nf = th.function([initials], result, updates=updates)\n\ninitial_values = np.array([1,2,3,4,5], dtype=th.config.floatX)\n\nprint f(initial_values)\n<\/code><\/pre>\n"}]}
{"QuestionId":27770975,"AnswerCount":1,"Tags":"<python><linear-algebra><logistic-regression><theano>","CreationDate":"2015-01-04T22:16:31.123","AcceptedAnswerId":null,"OwnerUserId":3998526.0,"Title":"Theano inner product 3d matrix","Body":"<p>thanks for reading this.<\/p>\n\n<p>I'm trying to implement a multi-label logistic regression using theano:<\/p>\n\n<pre><\/pre>\n\n<p>but -T.dot(x, w) product fails with this error:<\/p>\n\n<p>TypeError: ('Bad input argument to theano function with name \"train\"  at index 0(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (5, 10, 2).')<\/p>\n\n<p>x has shape (5, 2, 10) And W (1, 2, 10). I would expect the dot product to have shape (5,2).<\/p>\n\n<p>My questions are:\nIs there anyway to do this inner product?\nDo you think there is a better way to achieve a multi-label logistic regression?<\/p>\n\n<p>thanks!<\/p>\n\n<p>---- EDIT -----<\/p>\n\n<p>So here is an implementation of what I would like to do using numpy.<\/p>\n\n<pre><\/pre>\n\n<p>output:<\/p>\n\n<pre><\/pre>\n\n<p>But I don't know how to do this symbolically using Theano.<\/p>\n","answers":[{"AnswerId":"27772429","CreationDate":"2015-01-05T01:52:51.083","ParentId":null,"OwnerUserId":"3998526","Title":null,"Body":"<p>After some hours of fighting this seems to produce the right results:<\/p>\n\n<p>I had an error which was having the input as rng.randn(examples,features,labels) instead of rng.randn(examples,features). This means, that besides having more labels, the inputs should be the same size.<\/p>\n\n<p>And the way of computing the dot product the right way was using theano.scan method like:\nresults, updates = theano.scan(lambda label: T.dot(x, w[label,:]) - b[label], sequences=T.arange(labels))<\/p>\n\n<p>thanks everybody for their help!<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\nrng = np.random\n\nexamples = 5\nfeatures = 10\nlabels = 2\nD = (rng.randn(examples,features), rng.randint(size=(labels, examples), low=0, high=2))\ntraining_steps = 10000\n\n# Declare Theano symbolic variables\nx = T.matrix(\"x\")\ny = T.matrix(\"y\")\nw = theano.shared(rng.randn(labels ,features), name=\"w\")\nb = theano.shared(np.zeros(labels), name=\"b\")\nprint \"Initial model:\"\nprint w.get_value(), b.get_value()\n\nresults, updates = theano.scan(lambda label: T.dot(x, w[label,:]) - b[label], sequences=T.arange(labels))\n\n# Construct Theano expression graph\np_1 = 1 \/ (1 + T.exp(- results))   # Probability that target = 1\nprediction = p_1 &gt; .5                         # The prediction thresholded\nxent = -y * T.log(p_1) - (1-y) * T.log(1-p_1) # Cross-entropy loss function\ncost = xent.mean() + 0.01 * (w ** 2).sum()# The cost to minimize\ngw, gb = T.grad(cost, [w, b])             # Compute the gradient of the cost\n                                          # (we shall return to this in a\n                                          # following section of this tutorial)\n\n# Compile\ntrain = theano.function(\n          inputs=[x,y],\n          outputs=[prediction, xent],\n          updates=((w, w - 0.1 * gw), (b, b - 0.1 * gb)),\n          name='train')\npredict = theano.function(inputs=[x], outputs=prediction , name='predict')\n\n# Train\nfor i in range(training_steps):\n    pred, err = train(D[0], D[1])\n\nprint \"Final model:\"\nprint w.get_value(), b.get_value()\nprint \"target values for D:\", D[1]\nprint \"prediction on D:\", predict(D[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":27789597,"AnswerCount":1,"Tags":"<python><python-2.7><theano>","CreationDate":"2015-01-05T23:18:13.470","AcceptedAnswerId":27831625.0,"OwnerUserId":4416268.0,"Title":"Theano installation problems on a cluster","Body":"<p>I am trying to install Theano on a cluster node running \"Red Hat Enterprise Linux Client release 5.10 (Tikanga)\". I do not have admin permissions on the cluster. Hence, I installed Theano on my local user profile. The following are the version details of my installation:<\/p>\n\n<blockquote>\n  <ol>\n  <li>The Python version is : Python 2.7.3<\/li>\n  <li>The  version installed on the cluster is:  version 1.6.2<\/li>\n  <li><\/li>\n  <li>nose version 1.3.4<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>I installed Theano in the following manner<\/p>\n\n<blockquote>\n  <ol>\n  <li><\/li>\n  <li><\/li>\n  <li>The installed Theano version is: 0.6.0<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>I then tried to run theano.test() inside a python2.7 console. The test ran smoothly for a couple of minutes before I got the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Please help me install Theano correctly. I have come across solutions on this group suggesting re-installing the latest version of Theano. However, I have already installed Theano from the git repository. I have tried this possible solution already: <a href=\"https:\/\/stackoverflow.com\/a\/18238732\/4416268\">https:\/\/stackoverflow.com\/a\/18238732\/4416268<\/a> but still I get the same error.<\/p>\n","answers":[{"AnswerId":"27831625","CreationDate":"2015-01-08T01:33:53.683","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Use Theano flag: <code>blas.ldflags=-lblas -lgfortran<\/code><\/p>\n\n<p>If you do not know how to use Theano flag, check this page: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a><\/p>\n\n<p>Your problem is discussed on that page: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html<\/a><\/p>\n"}]}
{"QuestionId":27816260,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-01-07T09:33:03.903","AcceptedAnswerId":27823964.0,"OwnerUserId":4397268.0,"Title":"What Is equivalent to a [a < 0] = 0 in Theano?","Body":"<p>What is the equivalent to NumPy's  in Theano (tensor variable)?\nI want all my matrix elements that are smaller than a number equal to zero.<\/p>\n","answers":[{"AnswerId":"27823964","CreationDate":"2015-01-07T16:27:19.703","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This work:<\/p>\n\n<pre><code>import theano\na=theano.tensor.matrix()\nidxs=(a&lt;0).nonzero()\nnew_a=theano.tensor.set_subtensor(a[idxs], 0)\n<\/code><\/pre>\n\n<p>Do not forget, Theano is a symbolic language. So the variable a isn't changed in the user graph. It is the new variable new_a that contain the new value and a still have the old value.<\/p>\n\n<p>Theano will optimize this to work inplace if possible.<\/p>\n"},{"AnswerId":"38614428","CreationDate":"2016-07-27T13:30:40.920","ParentId":null,"OwnerUserId":"5424661","Title":null,"Body":"<p>This also works, and can also add upper bound limit<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\na = T.matrix()\nb = a.clip(0.0)\n<\/code><\/pre>\n\n<p>or if you want upper bound as well, you might like to try:<\/p>\n\n<pre><code>b = T.clip(a, 0.0, 1.0)\n<\/code><\/pre>\n\n<p>where 1.0 the place one want set upper bound.<\/p>\n\n<p>check the documents <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">here<\/a><\/p>\n"}]}
{"QuestionId":27816788,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-01-07T09:59:39.383","AcceptedAnswerId":null,"OwnerUserId":1465156.0,"Title":"Better way than T.eye in theano","Body":"<p>The problem is, given a arbitrary 1-d vector , expanded it into  basis vectors with  dimension. <\/p>\n\n<p>The rule of the expansion is: each element in  is the index of columns in the n*n identity matrix.<\/p>\n\n<p>For example:<\/p>\n\n<pre><\/pre>\n\n<p>Since , we have the  identity matrix:<\/p>\n\n<pre><\/pre>\n\n<p>Expand each element  using the rule, we have:<\/p>\n\n<pre><\/pre>\n\n<p>I want to solve this problem using theano, with very large  (>50k) and very long  (>10k), so efficiency is important.<\/p>\n\n<p>The solution using numpy is trivial, but the numpy.eye function may cost too much, we may use anther method to make it faster. Comparing the following methods:<\/p>\n\n<pre><\/pre>\n\n<p>However, the second method may not work in theano since loop in symbolic variable may not trivial. Is there a method which can avoid using T.eye?<\/p>\n\n<p> can be an arbitrary 1-d vector.<\/p>\n","answers":[{"AnswerId":"27817819","CreationDate":"2015-01-07T10:53:49.357","ParentId":null,"OwnerUserId":"764322","Title":null,"Body":"<p>You can try another approach. In my computer:<\/p>\n\n<pre><code>&gt;&gt;&gt; %timeit np.eye(n)[y_value]\n1 loops, best of 3: 544 ms per loop\n<\/code><\/pre>\n\n<p>However, you don't need to create the whole array if you know in advance the rows you want. You can do this:<\/p>\n\n<pre><code>&gt;&gt;&gt; n = 25500\n&gt;&gt;&gt; n_rows = y_value.size\n&gt;&gt;&gt; r = np.zeros((n_rows, n))\n&gt;&gt;&gt; r[range(n_rows), y_value] = 1\n<\/code><\/pre>\n\n<p>You create a way smaller array, only <code>y x n<\/code> where <code>y<\/code> is the size of your index vector, and populate it in every row. The timing in my computer is:<\/p>\n\n<pre><code>&gt;&gt;&gt; %%timeit \n..: r = np.zeros((n_rows, n))\n..: r[range(n_rows), y_value] = 1\n\n100 loops, best of 3: 3.8 ms per loop\n<\/code><\/pre>\n\n<p><code>x151<\/code> speedup in my laptop.<\/p>\n\n<p>Additionally, if you don't want an array full of zeros at the rear (x-axis), you could do:<\/p>\n\n<pre><code>&gt;&gt;&gt; %%timeit \n..: r = np.zeros((n_rows, y_value.max()+1))\n..: r[range(n_rows), y_value] = 1\n\n100000 loops, best of 3: 16 \u00b5s per loop\n<\/code><\/pre>\n\n<p>Which is even faster, but the resulting array is <code>y x ymax<\/code>, in this case <code>99 x 100<\/code>, which might not be what you want.<\/p>\n"}]}
{"QuestionId":27824979,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-01-07T17:21:12.387","AcceptedAnswerId":27831513.0,"OwnerUserId":4397268.0,"Title":"normalization in theano for any image?","Body":"<p>I write below code but that is very slow. And, of course, does not work correct!<\/p>\n\n<p><strong>I have what I do?<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"27831513","CreationDate":"2015-01-08T01:20:08.733","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>First that section of your code do nothing. Just remove this line<\/p>\n\n<pre><code>T.set_subtensor(self.pooled_out[i], self.pooled_out[i])\n<\/code><\/pre>\n\n<p>Without full code, I can't test my solution, but I think this would do what you want:<\/p>\n\n<pre><code>pmin = self.pooled_out.min(axis=[2,3], keepdims=True)\npmax = self.pooled_out.max(axis=[2,3], keepdims=True)\nnormalized_pooled_out = (self.pooled_out - pmin)\/pmax\n<\/code><\/pre>\n\n<p>Then normalized_pooled_out contain the symbolic variable that have the value I think you want.<\/p>\n"}]}
{"QuestionId":27827656,"AnswerCount":1,"Tags":"<cuda><deep-learning><caffe>","CreationDate":"2015-01-07T19:59:57.043","AcceptedAnswerId":null,"OwnerUserId":4096101.0,"Title":"Caffe Installation Issues","Body":"<p>I'm having some problem while installing Caffe. Please let me know if anyone have come across the same issue. Thanks.<\/p>\n\n<blockquote>\n  <p>make runtest<br>\n  .build_release\/test\/test_all.testbin 0 --gtest_shuffle<br>\n  Cuda number of devices: 1<br>\n  Setting to use device 0<br>\n  Current device id: 0<br>\n  Note: Randomizing tests' orders with a seed of 88789 .<br>\n  [==========] Running 838 tests from 169 test cases.<br>\n  [----------] Global test environment set-up.<br>\n  [----------] 3 tests from ImageDataLayerTest\/3, where TypeParam = caffe::DoubleGPU<br>\n  [ RUN      ] ImageDataLayerTest\/3.TestResize<br>\n  F0107 14:26:04.664185  3079 math_functions.cpp:91] Check failed: error == cudaSuccess (11 vs. 0)  invalid argument<br>\n  <strong>* Check failure stack trace: *<\/strong><br>\n      @     0x2ab3f5243daa  (unknown)<br>\n      @     0x2ab3f5243ce4  (unknown)<br>\n      @     0x2ab3f52436e6  (unknown)<br>\n      @     0x2ab3f5246687  (unknown)<br>\n      @           0x6bdc35  caffe::caffe_copy&lt;>()<br>\n      @           0x7439af  caffe::BasePrefetchingDataLayer&lt;>::Forward_gpu()<br>\n      @           0x428da2  caffe::Layer&lt;>::Forward()<br>\n      @           0x62ff53  caffe::ImageDataLayerTest_TestResize_Test&lt;>::TestBody()<br>\n      @           0x657363  testing::internal::HandleExceptionsInMethodIfSupported&lt;>()<br>\n      @           0x64de07  testing::Test::Run()<br>\n      @           0x64deae  testing::TestInfo::Run()<br>\n      @           0x64dfb5  testing::TestCase::Run()<br>\n      @           0x6512f8  testing::internal::UnitTestImpl::RunAllTests()<br>\n      @           0x651587  testing::UnitTest::Run()<br>\n      @           0x41d3a0  main<br>\n      @     0x2ab3f8396ec5  (unknown)<br>\n      @           0x4243d7  (unknown)<br>\n      @              (nil)  (unknown)<br>\n  make: *** [runtest] Aborted (core dumped)  <\/p>\n<\/blockquote>\n\n#\n\n<p>Ubuntu 14.04<\/p>\n\n<pre><\/pre>\n\n#\n","answers":[{"AnswerId":"27973046","CreationDate":"2015-01-15T21:10:46.567","ParentId":null,"OwnerUserId":"3646384","Title":null,"Body":"<p>There seems to be a problem with the GPU support. Maybe it does not support your GPU. I would try installing Caffe without GPU support. All you need to do is to uncomment <\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>CPU_ONLY := 1<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n\nin Makefile.config and then make again. <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">Here<\/a> are the instructions. <\/p>\n"}]}
{"QuestionId":27845374,"AnswerCount":1,"Tags":"<python><gpu><theano>","CreationDate":"2015-01-08T16:50:56.730","AcceptedAnswerId":27996360.0,"OwnerUserId":4397268.0,"Title":"memory error in my program when I set theano.config.device to gpu","Body":"<p>My graphics system is GT 550M<\/p>\n\n<p>When I run my program on the gpu It gives following error and I don't know how I fix this problem<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"27996360","CreationDate":"2015-01-17T05:07:58.200","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The error by cuda are returned async. So part of the error message can be unrelated. Here is the first line:<\/p>\n\n<p>MemoryError: error freeing device pointer 0x0000000500C60000 (the launch timed out and was terminated)<\/p>\n\n<p>The answer is in the second part: the launch timed out and was terminated<\/p>\n\n<p>You GPU is attached to a monitor. In that case, there is a limit of 5s for each GPU kernel call. It happen that it got busted and the driver killed that kernel. This is to prevent the screen from being not responsive.<\/p>\n\n<p>possible solution:\n1) use a different gpu for the monitor.\n2) make the kernel faster by using small input data (lower batch size for example)\n3) buy a faster GPU, not sure it will work and if it work with your current size, the problem will appear with bigger size.<\/p>\n"}]}
{"QuestionId":27873828,"AnswerCount":1,"Tags":"<ubuntu><lua><torch>","CreationDate":"2015-01-10T07:20:01.877","AcceptedAnswerId":null,"OwnerUserId":4439342.0,"Title":"undefined symbol: spotrs_ in torch7","Body":"<p>I am having problem with torch7 in Ubuntu 14.04. Error:<\/p>\n\n<p>error loading module \"libtorch\" from \"libtorch.so\": undefined symbol: spotrs_<\/p>\n\n<p>I am trying to import it in lua 5.1.5. I don't have any problem in Mac. <\/p>\n\n<p>Thanks in Advance.<\/p>\n","answers":[{"AnswerId":"27898847","CreationDate":"2015-01-12T09:32:37.670","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Have you used Torch <a href=\"https:\/\/github.com\/torch\/ezinstall\" rel=\"nofollow\">ezinstall<\/a> installer? <code>spotrs<\/code> is a LAPACK function, so you need to make sure you have e.g <a href=\"https:\/\/github.com\/torch\/ezinstall\/blob\/83d8ccb\/install-deps#L7-L22\" rel=\"nofollow\">OpenBLAS<\/a> properly installed.<\/p>\n"}]}
{"QuestionId":27878446,"AnswerCount":3,"Tags":"<android><java-native-interface><caffe>","CreationDate":"2015-01-10T16:30:44.100","AcceptedAnswerId":28625180.0,"OwnerUserId":1282137.0,"Title":"Storing file on Android for Native reading","Body":"<p>I'm writing app for android and I'm using caffe library. My problem is that on start I need to initialize caffe, which is done by passing two files (structures of network) to caffe.\nProblem is that I don't know how to store extra files on device. I've added model file to assets, but I don't know how can I read it using file path. Can you tell me where to store these file that could be access using file path? <\/p>\n\n<p>Thanks for any ideas.<\/p>\n","answers":[{"AnswerId":"27878470","CreationDate":"2015-01-10T16:34:14.953","ParentId":null,"OwnerUserId":"3933089","Title":null,"Body":"<p>Put them into your project as assets, and then when the app starts, you can read them from the assets and copy them into the app's private storage area. You can find this directory using <a href=\"http:\/\/developer.android.com\/reference\/android\/content\/Context.html#getFilesDir()\" rel=\"nofollow\"><code>Context.getFilesDir()<\/code><\/a>.<\/p>\n\n<p>From there, you'll be able to pass the files to Caffe.<\/p>\n"},{"AnswerId":"27878463","CreationDate":"2015-01-10T16:33:24.010","ParentId":null,"OwnerUserId":"4440363","Title":null,"Body":"<p>This should do it. Just copy those files to data directory from asset folder. If you already have those files there just load them.<\/p>\n\n<pre><code>String toPath = \"\/data\/data\/\" + getPackageName();  \/\/ Your application path\n\n\n\n\nprivate static boolean copyAssetFolder(AssetManager assetManager,\n            String fromAssetPath, String toPath) {\n        try {\n            String[] files = assetManager.list(fromAssetPath);\n            new File(toPath).mkdirs();\n            boolean res = true;\n            for (String file : files)\n                if (file.contains(\".\"))\n                    res &amp;= copyAsset(assetManager, \n                            fromAssetPath + \"\/\" + file,\n                            toPath + \"\/\" + file);\n                else \n                    res &amp;= copyAssetFolder(assetManager, \n                            fromAssetPath + \"\/\" + file,\n                            toPath + \"\/\" + file);\n            return res;\n        } catch (Exception e) {\n            e.printStackTrace();\n            return false;\n        }\n    }\n\nprivate static boolean copyAsset(AssetManager assetManager,\n        String fromAssetPath, String toPath) {\n    InputStream in = null;\n    OutputStream out = null;\n    try {\n      in = assetManager.open(fromAssetPath);\n      new File(toPath).createNewFile();\n      out = new FileOutputStream(toPath);\n      copyFile(in, out);\n      in.close();\n      in = null;\n      out.flush();\n      out.close();\n      out = null;\n      return true;\n    } catch(Exception e) {\n        e.printStackTrace();\n        return false;\n    }\n}\n\nprivate static void copyFile(InputStream in, OutputStream out) throws IOException {\n    byte[] buffer = new byte[1024];\n    int read;\n    while((read = in.read(buffer)) != -1){\n      out.write(buffer, 0, read);\n    }\n}\n<\/code><\/pre>\n"},{"AnswerId":"28625180","CreationDate":"2015-02-20T09:08:48.020","ParentId":null,"OwnerUserId":"1282137","Title":null,"Body":"<p>Assets are packaged and access only using special methods, so i solved problem by access file and then copy it to new location which i passed to native method.<\/p>\n"}]}
{"QuestionId":27890137,"AnswerCount":1,"Tags":"<c++><macos><opencv><osx-yosemite><caffe>","CreationDate":"2015-01-11T17:44:37.577","AcceptedAnswerId":27904334.0,"OwnerUserId":2288621.0,"Title":"Undefined symbols for architecture x86_64: for caffe build","Body":"<p>I got this error for <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"noreferrer\">caffe<\/a> build. How can I fix it?\nI'm using Mac OSX Yosemite 10.10.1.<\/p>\n\n<p>CONSOLE LOG<\/p>\n\n<pre><\/pre>\n\n<p>src\/caffe\/layers\/window_data_layer.cpp includes<\/p>\n\n<pre><\/pre>\n\n<p>It seems I have correct libraries. (Is this right thing to check??)<\/p>\n\n<pre><\/pre>\n\n<p>OpenCV libraries seem to be built with libstdc++, because <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"noreferrer\">the official install guide<\/a> suggests so.<\/p>\n\n<pre><\/pre>\n\n<p>All the build commands are correctly with -stdlib=libstdc++ option.<\/p>\n\n<pre><\/pre>\n\n<p>Thank for your help in advance!<\/p>\n\n<hr>\n\n<p>Now I find I can use ld with -v option.\nI keep investigating.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Dylibs seems to contain appropriate symbols.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Dylibs seems to have the appropriate architecture. mmm...<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Found a nice option for ld, which logs each file the ld loads.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"27904334","CreationDate":"2015-01-12T14:32:25.050","ParentId":null,"OwnerUserId":"2288621","Title":null,"Body":"<p>Solved!<\/p>\n\n<p><code>cv::imread(cv::String const&amp;, int)<\/code> is defined on <code>libopencv_imgcodecs.dylib<\/code> and\nMakefile is missing it.<\/p>\n\n<p>So, I added <code>opencv_imgcodecs<\/code> to Makefile.<\/p>\n\n<pre><code>LIBRARIES += glog gflags protobuf leveldb snappy \\\n        lmdb \\\n        boost_system \\\n        hdf5_hl hdf5 \\\n        opencv_imgcodecs opencv_highgui opencv_imgproc opencv_core pthread\n<\/code><\/pre>\n"}]}
{"QuestionId":27948324,"AnswerCount":2,"Tags":"<python><machine-learning><theano>","CreationDate":"2015-01-14T16:54:47.867","AcceptedAnswerId":null,"OwnerUserId":4178757.0,"Title":"Implementing LeCun Local Contrast Normalization with Theano","Body":"<p>I'm trying to use the code that I found to implement the LeCun Local Contrast Normalization but I get incorrect result. The code is in Python and uses the  library.<\/p>\n\n<pre><\/pre>\n\n<p>Here is the testing code:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the result: <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/UnCq9.jpg\" alt=\"Example image processing results\"><\/p>\n\n<p><em>(left to right: origin, my result, the expected result)<\/em> <\/p>\n\n<p>Could someone tell me what I did wrong with the code?<\/p>\n","answers":[{"AnswerId":"29746766","CreationDate":"2015-04-20T11:35:44.277","ParentId":null,"OwnerUserId":"4810149","Title":null,"Body":"<p>I think these two lines may have some mistakes on the matrix axes:<\/p>\n\n<pre><code>per_img_mean = denom.mean(axis=[1, 2])\ndivisor = T.largest(per_img_mean.dimshuffle(0, 'x', 'x', 1), denom)\n<\/code><\/pre>\n\n<p>and it should be rewritten as:<\/p>\n\n<pre><code>per_img_mean = denom.mean(axis=[2, 3])\ndivisor = T.largest(per_img_mean.dimshuffle(0, 1, 'x', 'x'), denom)\n<\/code><\/pre>\n"},{"AnswerId":"35259710","CreationDate":"2016-02-07T22:05:46.557","ParentId":null,"OwnerUserId":"5896264","Title":null,"Body":"<p>Here is how I implemented local contrast normalization as reported in Jarrett et al (<a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/jarrett-iccv-09.pdf\">http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/jarrett-iccv-09.pdf<\/a>). You can use it as a separate layer.<\/p>\n\n<p>I tested it on the code from the LeNet tutorial of theano in which I applied LCN to the input and to each convolutional layer which yields slightly better results.<\/p>\n\n<p>You can find the full code here:\n<a href=\"https:\/\/github.com\/jostosh\/theano_utils\/blob\/master\/lcn.py\">https:\/\/github.com\/jostosh\/theano_utils\/blob\/master\/lcn.py<\/a> <\/p>\n\n<pre><code>class LecunLCN(object):\ndef __init__(self, X, image_shape, threshold=1e-4, radius=9, use_divisor=True):\n    \"\"\"\n    Allocate an LCN.\n\n    :type X: theano.tensor.dtensor4\n    :param X: symbolic image tensor, of shape image_shape\n\n    :type image_shape: tuple or list of length 4\n    :param image_shape: (batch size, num input feature maps,\n                         image height, image width)\n    :type threshold: double\n    :param threshold: the threshold will be used to avoid division by zeros\n\n    :type radius: int\n    :param radius: determines size of Gaussian filter patch (default 9x9)\n\n    :type use_divisor: Boolean\n    :param use_divisor: whether or not to apply divisive normalization\n    \"\"\"\n\n    # Get Gaussian filter\n    filter_shape = (1, image_shape[1], radius, radius)\n\n    self.filters = theano.shared(self.gaussian_filter(filter_shape), borrow=True)\n\n    # Compute the Guassian weighted average by means of convolution\n    convout = conv.conv2d(\n        input=X,\n        filters=self.filters,\n        image_shape=image_shape,\n        filter_shape=filter_shape,\n        border_mode='full'\n    )\n\n    # Subtractive step\n    mid = int(numpy.floor(filter_shape[2] \/ 2.))\n\n    # Make filter dimension broadcastable and subtract\n    centered_X = X - T.addbroadcast(convout[:, :, mid:-mid, mid:-mid], 1)\n\n    # Boolean marks whether or not to perform divisive step\n    if use_divisor:\n        # Note that the local variances can be computed by using the centered_X\n        # tensor. If we convolve this with the mean filter, that should give us\n        # the variance at each point. We simply take the square root to get our\n        # denominator\n\n        # Compute variances\n        sum_sqr_XX = conv.conv2d(\n            input=T.sqr(centered_X),\n            filters=self.filters,\n            image_shape=image_shape,\n            filter_shape=filter_shape,\n            border_mode='full'\n        )\n\n\n        # Take square root to get local standard deviation\n        denom = T.sqrt(sum_sqr_XX[:,:,mid:-mid,mid:-mid])\n\n        per_img_mean = denom.mean(axis=[2,3])\n        divisor = T.largest(per_img_mean.dimshuffle(0, 1, 'x', 'x'), denom)\n        # Divisise step\n        new_X = centered_X \/ T.maximum(T.addbroadcast(divisor, 1), threshold)\n    else:\n        new_X = centered_X\n\n    self.output = new_X\n\n\ndef gaussian_filter(self, kernel_shape):\n    x = numpy.zeros(kernel_shape, dtype=theano.config.floatX)\n\n    def gauss(x, y, sigma=2.0):\n        Z = 2 * numpy.pi * sigma ** 2\n        return  1. \/ Z * numpy.exp(-(x ** 2 + y ** 2) \/ (2. * sigma ** 2))\n\n    mid = numpy.floor(kernel_shape[-1] \/ 2.)\n    for kernel_idx in xrange(0, kernel_shape[1]):\n        for i in xrange(0, kernel_shape[2]):\n            for j in xrange(0, kernel_shape[3]):\n                x[0, kernel_idx, i, j] = gauss(i - mid, j - mid)\n\n    return x \/ numpy.sum(x)\n<\/code><\/pre>\n"}]}
{"QuestionId":27971707,"AnswerCount":1,"Tags":"<python><opencl><theano>","CreationDate":"2015-01-15T19:43:18.510","AcceptedAnswerId":28749022.0,"OwnerUserId":1085946.0,"Title":"Using Python+Theano with OpenCL in an AMD GPU","Body":"<p>I'm trying to use Python with Theano to accelerate some code with OpenCL. I installed  and  as instructed (I think), and got no errors. The installation detected the OpenCL runtime installed. <\/p>\n\n<p>I just cannot run the Theano example for OpenCL, mainly because I don't know how to specify my GPU. My GPU is a , according to . All code in the Theano documentation uses  and the only place where OpenCL is mentioned, it says  where  is a number.<\/p>\n\n<p>I tried  and got a  error saying that the correct format is . I have since tried any number of combinations of numbers ( and such), and always an .<\/p>\n\n<p>My system is Ubuntu 14.04 x64 and my hardware is a Toshiba Satellite, 15\". I installed  with , and later installed  following the instructions on their site.<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"28749022","CreationDate":"2015-02-26T17:33:51.500","ParentId":null,"OwnerUserId":"1828289","Title":null,"Body":"<p>opencl0:0 is correct.  Could you confirm that pyopencl works?  You may have a problem with your opencl (or drivers\/cl compiler).<\/p>\n\n<p>However, I think Theano does not quite work with OpenCL at the moment.  The current state is there is partial support, enough for \"hello world\", but not enough to run any significant code.<\/p>\n\n<p>See:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2189\">https:\/\/github.com\/Theano\/Theano\/issues\/2189<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/1471\">https:\/\/github.com\/Theano\/Theano\/issues\/1471<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2190\">https:\/\/github.com\/Theano\/Theano\/issues\/2190<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/pull\/1732\">https:\/\/github.com\/Theano\/Theano\/pull\/1732<\/a><\/p>\n\n<p>To summarize, no, most stuff is not ported (including Elemwise, for example, which is a common op).  I would really like to see Theano on OpenCL though.  That would be a great thing for AMD to pitch in on.  Soon :)<\/p>\n"}]}
{"QuestionId":27972386,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2015-01-15T20:27:34.277","AcceptedAnswerId":null,"OwnerUserId":678392.0,"Title":"Caffe Deep Learning Library example","Body":"<p>Here is an example from\n<a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html<\/a><\/p>\n\n<p>I'm kind of lost. What am I supposed to infer from this example? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31407205","CreationDate":"2015-07-14T12:54:43.357","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>This snippet is to explain a feature of Caffe's Blob class shielding the user from the details of CPU&lt;->GPU memory transfer.<\/p>\n\n<p>My attempt at elaborating on the comments in the code:<\/p>\n\n<p>It assumes you've already declared a Blob object and populated with data. What the data represents is irrelevant. The actual declaration of the Blob object and its initialization are missing from this snippet.<\/p>\n\n<pre><code>\/\/ Assuming that data are on the CPU initially, and we have a blob.\nconst Dtype* foo;\nDtype* bar;\n<\/code><\/pre>\n\n<p>Since the data populating the blob resides in CPU memory, making use of it on a GPU device requires a transfer.<\/p>\n\n<pre><code>foo = blob.gpu_data(); \/\/ data copied cpu-&gt;gpu.\n<\/code><\/pre>\n\n<p>But if you're getting copying that same data to somewhere else on CPU memory, the Blob object won't have to perform the expensive copy operations required for CPU&lt;->GPU transfers.<\/p>\n\n<pre><code>foo = blob.cpu_data(); \/\/ no data copied since both have up-to-date contents.\n<\/code><\/pre>\n\n<p>Initializing the 'bar' object with data residing in GPU memory. It's already copied them once. No need to repeat the costly copy operations.<\/p>\n\n<pre><code>bar = blob.mutable_gpu_data(); \/\/ no data copied.\n\/\/ ... some operations ...\nbar = blob.mutable_gpu_data(); \/\/ no data copied when we are still on GPU.\n<\/code><\/pre>\n\n<p>The Blob class keeps track of whether the CPU and CPU copies are identical or if one was modified, requiring a renewed copy in order to keep them identical.<\/p>\n\n<pre><code>foo = blob.cpu_data(); \/\/ data copied gpu-&gt;cpu, since the gpu side has modified the data\nfoo = blob.gpu_data(); \/\/ no data copied since both have up-to-date contents\n<\/code><\/pre>\n\n<p>Now we're just going back and forth and seeing what will trigger a copy and will fall back on a cached copy of the data.<\/p>\n\n<pre><code>bar = blob.mutable_cpu_data(); \/\/ still no data copied.\nbar = blob.mutable_gpu_data(); \/\/ data copied cpu-&gt;gpu.\nbar = blob.mutable_cpu_data(); \/\/ data copied gpu-&gt;cpu.\n<\/code><\/pre>\n"}]}
{"QuestionId":27984493,"AnswerCount":1,"Tags":"<theano><deep-learning>","CreationDate":"2015-01-16T12:58:58.327","AcceptedAnswerId":27989893.0,"OwnerUserId":1926931.0,"Title":"Evaluating and modifying theano tensors when stuff is on GPU","Body":"<p>I am seriously stuck with something for ages now. I need some help.<\/p>\n\n<p>I am running a theano conv network on GPU.\nThe network has a loss function as such<\/p>\n\n<p>def mse(x, t):\n    return T.mean((x - t) ** 2)<\/p>\n\n<p>Here x is the predicted value of a rectified liner unit and t is the expected value.<\/p>\n\n<p>Now for a particular learning problem I am trying to modify the function such that I want to threshold the value of x. So essentially something simple as this<\/p>\n\n<p>x[x>ts] = ts<\/p>\n\n<p>But I am really struggling with this. I tried so many things<\/p>\n\n<p><\/p>\n\n<p>Apart from the three prints, which all print  everything else gives me error. \nSo I am at my wits end how to do this simple stuff.\nIs it because this stuff is on GPU ?<\/p>\n\n<p>I did test the code on local python prompt, by constructing a numpy array and converting it into a tensor shared variable. The different stuff above works. \nBut I am conscious that the type is theano.tensor.sharedvar.TensorSharedVariable and not theano.tensor.var.TensorVariable.  <\/p>\n\n<p>I would really appreciate if some one gives me a helping hand here.<\/p>\n\n<p>Regards<\/p>\n","answers":[{"AnswerId":"27989893","CreationDate":"2015-01-16T17:53:48.647","ParentId":null,"OwnerUserId":"1926931","Title":null,"Body":"<p>Please find the answer to this question given by pascal at\n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/cNnvw2rUHc8\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/cNnvw2rUHc8<\/a><\/p>\n\n<p>The failures are correct because the input values are not being provided at the time the function is being called, since it is symbolic.<\/p>\n\n<p>The answer is to use T.minimum(x,threshold)<\/p>\n"}]}
{"QuestionId":27986339,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-01-16T14:39:56.837","AcceptedAnswerId":28009396.0,"OwnerUserId":4462028.0,"Title":"Where can I find the label map between trained model like googleNet's output to there real class label?","Body":"<p>everyone, I am new to caffe. Currently, I try to use the trained GoogleNet which was downloaded from model zoo to classify some images. However, the network's output seem to be a vector rather than real label(like dog, cat). \nWhere can I find the label-map between trained model like googleNet's output to their real class label?\nThanks.<\/p>\n","answers":[{"AnswerId":"28009396","CreationDate":"2015-01-18T11:57:49.263","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you got <code>caffe<\/code> from git you should find in <code>data\/ilsvrc12<\/code> folder a shell script <code>get_ilsvrc_aux.sh<\/code>.<br>\nThis script should download several files used for ilsvrc (sub set of imagenet used for the large scale image recognition challenge) training.  <\/p>\n\n<p>The most interesting file (for you) that will be downloaded is <code>synset_words.txt<\/code>, this file has 1000 lines, one line per class identified by the net.<br>\nThe format of the line is<\/p>\n\n<blockquote>\n  <p>nXXXXXXXX description of class<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":27987691,"AnswerCount":2,"Tags":"<cuda><caffe>","CreationDate":"2015-01-16T15:48:13.890","AcceptedAnswerId":28422572.0,"OwnerUserId":1115169.0,"Title":"Issues with compiling Caffe with cuDNN","Body":"<p>i am trying to compile caffe on ubuntu14 with 750ti geforce gpu but i cant. i installed the cudnn library in \/usr\/local\/cuda\/lib64 and the cudnn.h header file in \/usr\/local\/cuda\/include still there seems to be a problem. i really think enabling the cudNN=1 in Makefile.config<\/p>\n\n<pre><\/pre>\n\n<p>thats where the problem is. What exactly are these errors?<\/p>\n\n<pre><\/pre>\n\n<p>I tested CUDA sample and GPU is working fine with cuda. Here is the result...<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28422572","CreationDate":"2015-02-10T01:27:00.183","ParentId":null,"OwnerUserId":"4548565","Title":null,"Body":"<p>A version of Caffe that works with cuDNNv2 is available from <a href=\"https:\/\/github.com\/slayton58\/caffe\" rel=\"nofollow\">S. Layton's github page<\/a>.<\/p>\n\n<p>His Caffe master branch works with cuDNNv2. You can download it from the github page.<\/p>\n\n<p>He made a pull request to the official Caffe github and the full discussion is available <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1739\" rel=\"nofollow\">here<\/a> if you want the details.<\/p>\n"},{"AnswerId":"27988298","CreationDate":"2015-01-16T16:21:34.000","ParentId":null,"OwnerUserId":"1115169","Title":null,"Body":"<p>I downloaded cuDNN R1 and now its working. Looks cudNN R2 latest one is not compatible with caffe.<\/p>\n"}]}
{"QuestionId":28009021,"AnswerCount":1,"Tags":"<python><logistic-regression><theano>","CreationDate":"2015-01-18T11:11:31.710","AcceptedAnswerId":28078636.0,"OwnerUserId":3656081.0,"Title":"Symbolic variables automatically update in theano","Body":"<p>Im following the theano tutorial given <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">here<\/a> for simple Stochastic Gradient Descent. Here however I am unable to understand in this block how the values of  and  are getting automatically updated according to the values of  and  since later down when we run test_logistic() we are only updating the values of  and ? \nThanks<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28078636","CreationDate":"2015-01-21T23:12:34.023","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p><code>p_y_given_x<\/code> and <code>y_pred<\/code> are symbolic variable (just python object from Theano). Those python variable that point to the Theano object do not get updated. They just represent the computation we want to do. Think like in pseudo-code.<\/p>\n\n<p>They will be used when compiling the Theano function. It is only then that the value will be computed. But this do not cause any change to the python variable that point to the object <code>p_y_given_x<\/code> and <code>y_pred<\/code>. The object are not changed.<\/p>\n\n<p>Understanding this distinction take time for some people. It is a new way of thinking. So don't hesitate to ask questions. One thing that help is to always ask yourself if you are in the symbolic world or the numerical world. the numerical world happen only with compiled Theano function.<\/p>\n"}]}
{"QuestionId":28011527,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-01-18T15:41:46.923","AcceptedAnswerId":null,"OwnerUserId":1382757.0,"Title":"How can I pool gradients in Theano?","Body":"<p>I'm using Theano for the first time to build a large statistical model. I'm performing a kind of stochastic gradient descent, but for each sample in the minibatch I need to perform a sampling procedure to compute the gradient. Is there a way in Theano to pool the gradients while I perform the sampling procedure for each datapoint in a minibatch, and only afterward perform the gradient update? <\/p>\n","answers":[{"AnswerId":"28078965","CreationDate":"2015-01-21T23:41:44.450","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>I don't understand what you mean by \"pool\".\nWhen you compute the gradient of your cost wrt some variables, the cost has to be a scalar. So, when using minibatches, you have to combine the individual costs for the examples in the minibatch. That can be done by a sum, a mean, a weighted sum... And then that cost is backpropagated.\nThe gradient of that cost wrt parameters will correspond (mathematically) to the sum\/mean\/weigted sum of the individual gradients (on each of the examples), but that is not the way it is computed.\nThe gradient of that cost wrt intermediate variables that are function of the inputs (hidden representations, etc.) will have the same format as the original minibatch, with the gradient wrt each of the minibatches in a different row.<\/p>\n\n<p>So, maybe what you want is expressing your final cost as a result of your sampling procedure, and then backpropagate the gradient of that cost.\nOr maybe you do not want to backpropagate the gradient of the true cost all the way, and backpropagate something that depends on the gradient instead.\nIn that case, you can do something like:<\/p>\n\n<pre><code># minibatch of inputs\ninputs = tt.matrix()\ninterm_result = f(input)\ncost = g(interm_result).sum()\ngrad_wrt_interm_result = th.grad(cost, interm_result)\nsampled_grad = sampling_procedure(grad_wrt_interm_result)\ngrad_wrt_params = th.grad(cost, params,\n                          known_grads={inter_result: sampled_grad})\n<\/code><\/pre>\n\n<p>That way, you can perform some of the backpropagation to interm_result, then change the gradient wrt inter_result to sampled_grad, and then finish the backpropagation towards the parameters.<\/p>\n"}]}
{"QuestionId":28011551,"AnswerCount":3,"Tags":"<python><gpu><theano>","CreationDate":"2015-01-18T15:44:05.900","AcceptedAnswerId":null,"OwnerUserId":22996.0,"Title":"How configure theano on Windows?","Body":"<p>I have Installed Theano on Windows machine and followed the configuration <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow noreferrer\">instructions<\/a>.<\/p>\n\n<p>I placed the following .theanorc.txt file in C:\\Users\\my_username folder:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to run the test, but haven't managed to run it on GPU. I guess the values from .theanorc.txt are not read, because I added the line print config.device and it outputs \"cpu\".<\/p>\n\n<p>Below is the basic test script and the output:<\/p>\n\n<pre><\/pre>\n\n<p>output:<\/p>\n\n<pre><\/pre>\n\n<p>I have installed CUDA Toolkit successfully but haven't managed to install pyCUDA. I guess Theano should work without pyCUDA installed anyway.<\/p>\n\n<p>I would be very thankful if anyone could help out solving this problem. I have followed <a href=\"https:\/\/stackoverflow.com\/questions\/25729969\/installing-theano-on-windows-8-with-gpu-enabled\">these<\/a> instructions but don't know why the configuration values in the program don't match the values in .theanorc.txt file.<\/p>\n","answers":[{"AnswerId":"44518712","CreationDate":"2017-06-13T10:19:59.247","ParentId":null,"OwnerUserId":"5859283","Title":null,"Body":"<p>Contrary to what has been said on a couple of pages, my installation (Windows 10, Python 2.7, Theano 0.10.0.dev1) would not interpret config instructions within a  <code>.theanorc.txt<\/code> file in my user profile folder, but would read a <code>.theanorc<\/code> file.<\/p>\n\n<p>If you are having trouble creating a file with that style of name, use the following commands at a terminal:<\/p>\n\n<pre><code>cd %USERPROFILE%\ntype NUL &gt; .theanorc\n<\/code><\/pre>\n\n<p>Sauce: <a href=\"http:\/\/ankivil.com\/making-theano-faster-with-cudnn-and-cnmem-on-windows-10\/\" rel=\"noreferrer\">http:\/\/ankivil.com\/making-theano-faster-with-cudnn-and-cnmem-on-windows-10\/<\/a><\/p>\n"},{"AnswerId":"42520401","CreationDate":"2017-02-28T23:04:23.077","ParentId":null,"OwnerUserId":"2531309","Title":null,"Body":"<p>Try to change the content in .theanorc.txt as indicating by Theano website ( <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a>). The path needs to be changed accordingly based on your installation.<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = gpu\n\n[nvcc]\nflags=-LC:\\Users\\cchan\\Anaconda3\\libs\ncompiler_bindir=C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin\n<\/code><\/pre>\n"},{"AnswerId":"28078667","CreationDate":"2015-01-21T23:15:26.967","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You are right that Theano does not need PyCUDA.<\/p>\n\n<p>It is strange that Theano does not read your configuration file. The exact path that gets read is this. Just run this in Python and you'll see where to put it:<\/p>\n\n<p><code>os.path.expanduser('~\/.theanorc.txt')<\/code><\/p>\n"}]}
{"QuestionId":28028253,"AnswerCount":2,"Tags":"<lua><dataframe><data-analysis><torch>","CreationDate":"2015-01-19T15:39:11.107","AcceptedAnswerId":37419889.0,"OwnerUserId":297094.0,"Title":"Lua library for data analysis (data frames)","Body":"<p>Is there any Lua implementation of data frames - structures for data analysis which? Something like Python pandas. I want to do some statistical operations using LuaJIT.<\/p>\n","answers":[{"AnswerId":"28029533","CreationDate":"2015-01-19T16:42:43.267","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>You may want to look at <a href=\"https:\/\/en.wikipedia.org\/wiki\/Torch_%28machine_learning%29\" rel=\"nofollow\">Torch7<\/a> that provides N-dimensional arrays with support for various statistical and mathematical operations and is based on LuaJIT.<\/p>\n"},{"AnswerId":"37419889","CreationDate":"2016-05-24T17:09:56.153","ParentId":null,"OwnerUserId":"409508","Title":null,"Body":"<p>Yes, there is now. Check out <a href=\"https:\/\/github.com\/AlexMili\/torch-dataframe\" rel=\"nofollow\">torch-dataframe<\/a> that I and Alex are developing. Our main priority is reliability that we check with a rich test suite. Performance comes second although we try to avoid performance hogs within the limits of Lua.<\/p>\n\n<p>The package is currently far from the pandas sophistication, but feel free to contribute with any methods you feel lacking.<\/p>\n"}]}
{"QuestionId":28031841,"AnswerCount":1,"Tags":"<python><numpy><hdf5><caffe>","CreationDate":"2015-01-19T19:08:02.153","AcceptedAnswerId":null,"OwnerUserId":1115169.0,"Title":"HDF5-DIAG: Error detected in HDF5 (1.8.11)","Body":"<p>I am trying to load hdf5 in caffe, it is not working. I checked the paths and even able to view hdf file using viewer. Everything is ok but caffe cant seem to load.<\/p>\n\n<p>I write the hdf5 using a python script like this, where X and labels are numpy arrays. <\/p>\n\n<pre><\/pre>\n\n<p>Here is the entire problem.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28032495","CreationDate":"2015-01-19T19:53:11.043","ParentId":null,"OwnerUserId":"1115169","Title":null,"Body":"<p>Solved :)<\/p>\n\n<p>i created a text file, placed the path of real .hd5 file inside it. The caffe prototxt file points to text file and it worked :)<\/p>\n\n<pre><code> hdf5_data_param {\n   source: \"train.txt\"\n   batch_size: 10\n  }\n<\/code><\/pre>\n\n<p>train.txt contains line..<\/p>\n\n<pre><code>facialkp.hd5\n<\/code><\/pre>\n"}]}
{"QuestionId":28054916,"AnswerCount":3,"Tags":"<lua><windows-8.1><luajit><torch>","CreationDate":"2015-01-20T20:52:32.163","AcceptedAnswerId":28058692.0,"OwnerUserId":2352671.0,"Title":"How to install Torch on windows 8.1?","Body":"<p><a href=\"http:\/\/torch.ch\/\" rel=\"noreferrer\">Torch<\/a> is a scientific computing framework with wide support for machine learning algorithms. It is easy to use and efficient, thanks to an easy and fast scripting language, LuaJIT, and an underlying C\/CUDA implementation.<\/p>\n\n<h2>Q:<\/h2>\n\n<p>Is there a way to install torch on MS Windows 8.1?<\/p>\n","answers":[{"AnswerId":"30012889","CreationDate":"2015-05-03T10:45:21.620","ParentId":null,"OwnerUserId":"1070480","Title":null,"Body":"<p><a href=\"http:\/\/cilvr.nyu.edu\/doku.php?id=software:torch:start#torch_tutorial_for_machine_learning\" rel=\"nofollow\">This webpage<\/a> hosted by New York University recommends installing a Linux virtual machine in order to run Torch7 on Windows through Linux. Another option would off course be to install a Linux dist in parallel with Windows 8.<\/p>\n\n<p>Otherwise, if you don't mind running an older version of Torch, there is a Windows installer for <a href=\"http:\/\/torch5.sourceforge.net\/download.html\" rel=\"nofollow\">Torch5<\/a> at SourceForge.<\/p>\n"},{"AnswerId":"31260640","CreationDate":"2015-07-07T05:44:55.643","ParentId":null,"OwnerUserId":"1970830","Title":null,"Body":"<p>I think to use a GPU from inside the virtual machine, the processor and the motherboard should not only support VT-x , but VT-d should be supported too.<\/p>\n\n<p>But the question is, if I use a CPU with VT-d supported, do you think there will be a significant loss in PCIe connections efficiency?<\/p>\n\n<p>From what I understand,\nVT-d is important if I want to give the virtual machines direct access to my hardware components (like PCI Express cards). Like directly attach graphics card to vm instead of host machine. Isn't that mean that the PCIe connections efficiency will be the same just like if it was the host?<\/p>\n"},{"AnswerId":"28058692","CreationDate":"2015-01-21T02:37:32.490","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>I got it installed and running on Windows (although not 8.1, but I don't expect the process to be different) following instructions in <a href=\"https:\/\/github.com\/torch\/DEPRECEATED-torch7-distro\" rel=\"nofollow noreferrer\">this repository<\/a>; it's now deprecated, but wasn't deprecated few months ago when I built it. The new instructions point to <a href=\"https:\/\/github.com\/torch\/torch7\" rel=\"nofollow noreferrer\">torch\/torch7<\/a> repository, but it has a different structure and I haven't been able to build it on Windows yet.<\/p>\n\n<p>There are instructions on how to install Torch7 from luarocks, but you may <a href=\"https:\/\/stackoverflow.com\/questions\/27276822\/installing-torch7-with-luarocks-on-windows-with-mingw-build-error\">run into issues on windows<\/a> as well; I haven't tried this process. It seems like there is <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/torch7\/rOxupFTyb28\" rel=\"nofollow noreferrer\">no official support for Windows yet<\/a>, but some work is being done by contributors (there is a link to a pull request in that thread).<\/p>\n\n<p><s>Based on my experience, compiling that deprecated repo may be your best option on Windows at the moment.<\/s><\/p>\n\n<p>Update (7\/9\/2015): I've recently submitted <a href=\"https:\/\/github.com\/torch\/torch7\/pull\/287\" rel=\"nofollow noreferrer\">several changes<\/a> that fix compilation issues with mingw, so you may try the most recent version of torch7 and follow the build instructions in the ticket. Note that the changes only apply to the core lib and additional libraries may need similar changes.<\/p>\n"}]}
{"QuestionId":28057957,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><caffe><kaggle>","CreationDate":"2015-01-21T01:08:04.857","AcceptedAnswerId":28369288.0,"OwnerUserId":1115169.0,"Title":"Multi label regression in Caffe","Body":"<p>i am extracting 30 facial keypoints (x,y) from an input image as per kaggle facialkeypoints competition.<\/p>\n\n<p>How do i setup caffe to run a regression and produce 30 dimensional output??. <\/p>\n\n<pre><\/pre>\n\n<p>How do i setup caffe accordingly?. I am using EUCLIDEAN_LOSS (sum of squares) to get the regressed output. Here is a simple logistic regressor model using caffe but it is not working. Looks accuracy layer cannot handle multi-label output.<\/p>\n\n<pre><\/pre>\n\n<p>Here is the layer file:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28369288","CreationDate":"2015-02-06T15:36:26.367","ParentId":null,"OwnerUserId":"1115169","Title":null,"Body":"<p>i found it :)<\/p>\n\n<p>I replaced the SOFTLAYER to EUCLIDEAN_LOSS function and changed the number of outputs. It worked.<\/p>\n\n<pre><code>layers {\n  name: \"loss\"\n  type: EUCLIDEAN_LOSS\n  bottom: \"ip1\"\n  bottom: \"label\"\n  top: \"loss\"\n}\n<\/code><\/pre>\n\n<p>HINGE_LOSS is also another option.<\/p>\n"}]}
{"QuestionId":28090797,"AnswerCount":4,"Tags":"<neural-network><reshape><deep-learning><caffe>","CreationDate":"2015-01-22T14:08:53.840","AcceptedAnswerId":null,"OwnerUserId":1926094.0,"Title":"How to reshape a blob in Caffe?","Body":"<p>How to reshape a blob of the shape  to  in Caffe?<\/p>\n\n<p>I want to make a convolution layer the weights of which are identical between channels.<\/p>\n\n<p>One way I come up with is to reshape the bottom blob of the shape  to  and place a convolution layer upon it. But I just don't know how to reshape a blob.<\/p>\n\n<p>Please help me out, thank you.<\/p>\n","answers":[{"AnswerId":"43199056","CreationDate":"2017-04-04T05:38:40.783","ParentId":null,"OwnerUserId":"7812167","Title":null,"Body":"<p>If I understand your final objective right, Caffe's convolution layer already can do multiple input-output convolution with common\/shared filters like:<\/p>\n\n<pre><code>layer {\n  name:   \"conv\"\n  type:   \"Convolution\"\n  bottom: \"in1\"\n  bottom: \"in2\"\n  bottom: \"in3\"\n  top:    \"out1\"\n  top:    \"out2\"\n  top:    \"out3\"\n  convolution_param {\n    num_output : 10  #the same 10 filters for all 3 inputs\n    kernel_size: 3        \n  }\n}\n<\/code><\/pre>\n\n<p>Assuming you have all streams split (slice layer can do that), and finally you may merge them if desired with a concat or eltwise layer.<\/p>\n\n<p>That avoid the needs of reshaping blob, convolved, and then reshaping it back, which might introduce cross-channel interference near the margins.<\/p>\n"},{"AnswerId":"28611043","CreationDate":"2015-02-19T16:04:55.367","ParentId":null,"OwnerUserId":"1580510","Title":null,"Body":"<p>Not sure if this fits your specs exactly, but Caffe does have flattening layers. The blob goes from n * c * h * w to n * (c<em>h<\/em>w) * 1 * 1.<\/p>\n\n<p>See <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n"},{"AnswerId":"35005464","CreationDate":"2016-01-26T01:14:21.947","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>Caffe now has a reshapeLayer for you.\n<a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1ReshapeLayer.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1ReshapeLayer.html<\/a><\/p>\n"},{"AnswerId":"35008024","CreationDate":"2016-01-26T06:20:11.730","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>As pointed by <a href=\"https:\/\/stackoverflow.com\/a\/35005464\/1714410\">whjxnyzh<\/a>, you can use <code>\"Reshape\"<\/code> layer. Caffe is quite flexible in the way it allows you to define the output shape.<br>\nSee <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L899\" rel=\"nofollow noreferrer\">the declaration of <code>reshap_param<\/code> in caffe.proto`<\/a>:<\/p>\n\n<blockquote>\n<pre><code>\/\/ Specify the output dimensions. If some of the dimensions are set to 0,\n\/\/ the corresponding dimension from the bottom layer is used (unchanged).\n\/\/ Exactly one dimension may be set to -1, in which case its value is\n\/\/ inferred from the count of the bottom blob and the remaining dimensions.\n<\/code><\/pre>\n<\/blockquote>\n\n<p>In your case I guess you'll have a layer like this:<\/p>\n\n<pre><code>layer {\n  name: \"my_reshape\"\n  type: \"Reshape\"\n  bottom: \"in\"\n  top: \"reshaped_in\"\n  reshape_param { shape: {dim: 0 dim: 1 dim: -1 dim: 0 } }\n}\n<\/code><\/pre>\n\n<p>See also on <a href=\"http:\/\/caffe.help\/manual\/layers\/reshape.html\" rel=\"nofollow noreferrer\">caffe.help<\/a>.<\/p>\n"}]}
{"QuestionId":28093753,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><caffe>","CreationDate":"2015-01-22T16:24:27.470","AcceptedAnswerId":28544226.0,"OwnerUserId":1115169.0,"Title":"Convolution issue in Caffe","Body":"<p>i have 96x96 pixel images in grayscale format stored in HDF5 files. i am trying to do multi output regression using caffe however convolution is not working. What exactly is the problem here? Why is convolutions not working?<\/p>\n\n<pre><\/pre>\n\n<p>My prototxt layer file is like this<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28544226","CreationDate":"2015-02-16T14:58:48.230","ParentId":null,"OwnerUserId":"4572256","Title":null,"Body":"<p>the lines<\/p>\n\n<pre>I0122 17:18:40.337906  5074 net.cpp:103] Top shape: 100 9216 1 1 (921600)\nI0122 17:18:40.337929  5074 net.cpp:103] Top shape: 100 30 1 1 (3000)<\/pre>\n\n<p>suggest that your input data is not in the correct shape. For an input of 100 batchs of 96x96 grey-scale image the shape should be: 100 1 96 96.\nTry to change this. (my guess is that for shape: N C H W, where N number of batches, c channels, h height, w weight)<\/p>\n"}]}
{"QuestionId":28120712,"AnswerCount":1,"Tags":"<python><windows><visual-studio><installation><theano>","CreationDate":"2015-01-24T00:01:48.690","AcceptedAnswerId":28199278.0,"OwnerUserId":1242679.0,"Title":"import theano (0.6) on Windows 8 with device= gpu (and visual studios 12.0)","Body":"<p>I'm having an compilation issue when I try to use Theano with the GPU device (it works fine with the CPU). I'm getting almost exactly the same problem to that already reported <a href=\"https:\/\/stackoverflow.com\/questions\/25729969\/installing-theano-on-windows-8-with-gpu-enabled\">here<\/a>, however following the solution provided does not work for me. Following the original solution, I can get as far as verifying that pycuda has been installed successfully, but importing theano still throws the same error:<\/p>\n\n<pre><\/pre>\n\n<p>I have Python 2.7.8 32bit and MinGW set up and CUDA 6.5. I'm using the following .theanorc config:<\/p>\n\n<pre><\/pre>\n\n<p>In order to get the pycuda example to work, I had to add visual studio 12.0  to my user path. It doesn't work with visual studios 10.0 for some reason, despite using the visual studio 10 command prompt to build pycuda.<\/p>\n\n<p>Incidentally, if I try , I get an error saying that version of visual studio is no good:<\/p>\n\n<pre><\/pre>\n\n<p>(I have visual studio express 2010 and 2013 installed)<\/p>\n\n<p>I understand that the Theano+GPU support for Windows is still somewhat experimental, but it seems like it does work for some people. Any suggestions as to what try next?<\/p>\n","answers":[{"AnswerId":"28199278","CreationDate":"2015-01-28T18:04:27.497","ParentId":null,"OwnerUserId":"1242679","Title":null,"Body":"<p>I found a bit of hack to fix it <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/theano-users\/warning$20C4273$3A$20$27round$27$20$3A$20inconsistent$20dll$20linkage\/theano-users\/C8ce20uE6Gs\/9Xg7UHazoWAJ\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>Essentially it entails finding the <code>cuda_ndarray.cuh<\/code> file in the <code>&lt;install-dir&gt;\/andbox\\cuda<\/code> and adding <code>#include &lt;algorithm&gt;<\/code>.<\/p>\n\n<p>It still leaves a warning <code>warning C4273: 'round' : inconsistent dll linkage<\/code>. <\/p>\n\n<p>Which according to <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2055\" rel=\"nofollow\">this reported issue<\/a> is due to a conflict between Python and CUDA and both providing an <code>round<\/code> functionality. This could perhaps be fixed by defining the macro HAVE_ROUND when linking with CUDA to tell Python not to try to redefine round.<\/p>\n\n<p>Not sure if this is a general fix that's applicable for everyone, but seemed to have worked for me - as far as enabling me to use the GPU with theano.<\/p>\n"}]}
{"QuestionId":28134363,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-01-25T07:53:38.470","AcceptedAnswerId":28147346.0,"OwnerUserId":1240896.0,"Title":"Given an image in Lua Torch, how can I find its dimensions?","Body":"<p>Let's say I want to know the height and width of .  What method should I call on it?  A link to a resource for methods on images would be great as the Torch  command does not work in this case.<\/p>\n","answers":[{"AnswerId":"28147346","CreationDate":"2015-01-26T09:18:29.913","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p><code>image.lena()<\/code> returns a Torch 3-dimensional tensor where the first dimension is the number of channels (3 for a RGB image), and the last ones are resp. the height (nb. of rows) and width (nb. of columns) of the image.<\/p>\n\n<p>So all you need to do is use the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/e8155b4\/doc\/tensor.md#torch.Tensor.size\" rel=\"nofollow\"><code>size(dim)<\/code><\/a> method as follow:<\/p>\n\n<pre><code>require 'image'\n\nlocal img = image.lena()\nprint(torch.typename(img)) -- torch.DoubleTensor\n\nlocal nchan, height, width = img:size(1), img:size(2), img:size(3)\nprint('nb. channels: ' .. nchan) -- 3\nprint('width: ' .. width .. ', height: ' .. height) -- 512, 512\n<\/code><\/pre>\n"}]}
{"QuestionId":28171577,"AnswerCount":1,"Tags":"<python><lua><caffe><torch>","CreationDate":"2015-01-27T13:22:44.347","AcceptedAnswerId":28189121.0,"OwnerUserId":2380470.0,"Title":"How to get a layer from a caffe model using torch","Body":"<p>In python when I want to get the data from a layer using caffe I have the following code <\/p>\n\n<pre><\/pre>\n\n<p>Hoever when I'm using torch I'm a bit stuck since I don't know how to perform the same action.\nCurrently I have the following code<\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated <\/p>\n","answers":[{"AnswerId":"28189121","CreationDate":"2015-01-28T09:51:23.283","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>First of all please note that <a href=\"https:\/\/github.com\/szagoruyko\/torch-caffe-binding\" rel=\"noreferrer\">torch-caffe-binding<\/a> (i.e the tool you use with <code>require 'caffe'<\/code>) is a direct wrapper around Caffe library thanks to LuaJIT FFI.<\/p>\n\n<p>This means that it allows you to conveniently do a forward or backward with a Torch tensor, <strong>but<\/strong> <a href=\"https:\/\/github.com\/szagoruyko\/torch-caffe-binding\/blob\/9b0892a\/caffe.cpp#L30-L69\" rel=\"noreferrer\">behind the scenes<\/a> these operations are made on a <code>caffe::Net<\/code> and not on a Torch <code>nn<\/code> network.<\/p>\n\n<p>So if you want to manipulate a plain <a href=\"https:\/\/github.com\/torch\/nn\" rel=\"noreferrer\">Torch network<\/a> what you should use is the <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\" rel=\"noreferrer\">loadcaffe<\/a> library which fully converts the network into a <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/containers.md#nn.Sequential\" rel=\"noreferrer\"><code>nn.Sequential<\/code><\/a>:<\/p>\n\n<pre><code>require 'loadcaffe'\n\nlocal net = loadcaffe.load('net.prototxt', 'net.caffemodel')\n<\/code><\/pre>\n\n<p>Then you can use <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/module.md#findmodulestypename\" rel=\"noreferrer\"><code>findModules<\/code><\/a>. However please note that you cannot use their initial label anymore (like <code>conv1<\/code> or <code>fc7<\/code>) as they are <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\/blob\/7013690\/loadcaffe.lua#L23-L40\" rel=\"noreferrer\">discarded after convert<\/a>.<\/p>\n\n<p>Here <code>fc7<\/code> (= <code>INNER_PRODUCT<\/code>) corresponds to the N-1 linear transformation. So you can get it as follow:<\/p>\n\n<pre><code>local nodes = net:findModules('nn.Linear')\nlocal fc7 = nodes[#nodes-1]\n<\/code><\/pre>\n\n<p>Then you can read the data (weights and biases) via <code>fc7.weight<\/code> and <code>fc7.bias<\/code> - these are regular <code>torch.Tensor<\/code>-s.<\/p>\n\n<hr>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>As of commit <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\/commit\/2516fac\" rel=\"noreferrer\">2516fac<\/a> loadcaffe now saves layer names in addition. So to retrieve the <code>'fc7'<\/code> layer you can now do something like:<\/p>\n\n<pre><code>local fc7\nfor _,m in pairs(net:listModules()) do\n  if m.name == 'fc7' then\n    fc7 = m\n    break\n  end\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":28177298,"AnswerCount":5,"Tags":"<python><caffe>","CreationDate":"2015-01-27T18:17:48.620","AcceptedAnswerId":28235061.0,"OwnerUserId":1115169.0,"Title":"Import caffe error","Body":"<p>i compiled caffe successfully in my ubuntu machine but cannot import in python.<\/p>\n\n<p>Caffe is installed \/home\/pbu\/Desktop\/caffe<\/p>\n\n<p>i tried adding the \/home\/pbu\/caffe\/python path  to sys.path.append, still not working<\/p>\n\n<p>i am trying to import caffe<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28235061","CreationDate":"2015-01-30T11:36:07.833","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>This happens when you have not run <code>make<\/code> for the python files separately.<\/p>\n\n<p>Run <code>make pycaffe<\/code> soon after running <code>make<\/code> in the Caffe directory.<\/p>\n\n<p>You may have to set the path to the python library correctly in <code>Makefile.config<\/code><\/p>\n"},{"AnswerId":"46405278","CreationDate":"2017-09-25T12:28:24.140","ParentId":null,"OwnerUserId":"5258060","Title":null,"Body":"<p>Adding to the above best answer. After you run <code>make<\/code> for python files by running <code>make pycaffe<\/code> where you ran your previous <code>make<\/code>s. Then you have to export that python path by running <code>export PYTHONPATH=&lt;path-to-caffe&gt;\/python<\/code>. You can choose to run this everytime before running a python code which utilizes caffe or add it to your <code>~\/.bashrc<\/code>. <\/p>\n"},{"AnswerId":"36261669","CreationDate":"2016-03-28T11:47:15.893","ParentId":null,"OwnerUserId":"5107003","Title":null,"Body":"<p>You should build caffe and pycaffe using the command:<\/p>\n\n<pre><code>cd $FRCN_ROOT\/caffe-fast-rcnn\nmake -j8 &amp;&amp; make pycaffe\n<\/code><\/pre>\n\n<p>and before the compilation, you should create a <code>Makefile.config<\/code> file and set the corresponding library path, such as python. <\/p>\n\n<p>More details are presented on the web: <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\" rel=\"nofollow\">bgirshick\/py-faster-rcnn<\/a>.<\/p>\n\n<p>What's more, when I run the \"Beyond the demo\" section, it seams that if I Create a symlink of the folder \"VOCdevkit\" as \"VOCdevkit2007\" which turns out to be \"can't find the dataset\". So, I change the folder name as \"VOCdevkit2007\", and it runs well.<\/p>\n"},{"AnswerId":"40734483","CreationDate":"2016-11-22T05:35:42.237","ParentId":null,"OwnerUserId":"1904943","Title":null,"Body":"<p>I posted my Caffe install notes (my architecture: Arch Linux x86_64 | Intel i7 CPU ...) in an Anaconda Python 2.7 virtual environment here:<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213\" rel=\"nofollow noreferrer\">Caffe Installation Notes<\/a><\/p>\n\n<pre><code>https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213\n<\/code><\/pre>\n\n<p>I also encountered the (downstream) \"Import caffe error,\" for which I needed to resolve my $PYTHONPATH to complete the make compilation and get Caffe finally installed, and also to be able to import it (in Python).<\/p>\n"},{"AnswerId":"35312906","CreationDate":"2016-02-10T10:40:01.440","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Well, I use the <code>cmake-gui<\/code> for <code>making<\/code> Caffe. There you need to set the Python paths to the Anaconda-python:<\/p>\n\n<pre><code>PYTHON_EXECUTABLE &lt;path_to_anaconda_home&gt;\/bin\/python2.7\nPYTHON_INCLUDE_DIRECTORY &lt;path_to_anaconda_home&gt;\/include\/PYTHON2.7\nPYTHON_LIBRARY &lt;path_to_anaconda_home&gt;\/lib\/libpython2.7.so\n<\/code><\/pre>\n"}]}
{"QuestionId":28187868,"AnswerCount":2,"Tags":"<cuda><theano><cublas>","CreationDate":"2015-01-28T08:41:35.660","AcceptedAnswerId":31155012.0,"OwnerUserId":133374.0,"Title":"Theano: cublasSgemm failed (14) an internal operation failed","Body":"<p>Sometimes, after a while of running fine, I get such an error with Theano \/ CUDA:<\/p>\n\n<pre><\/pre>\n\n<p>As my code runs fine for a while (I do Neural Network training, and it runs most of the time through, and even when this error occurred, it already ran fine for >2000 mini-batches), I wonder about the cause of this. Maybe some hardware fault?<\/p>\n\n<p>This is with CUDA 6.0 and a very recent Theano (yesterday from Git), Ubuntu 12.04, GTX 580.<\/p>\n\n<p>I also got the error with CUDA 6.5 on a K20:<\/p>\n\n<pre><\/pre>\n\n<p>(Another error I sometimes got in the past is <a href=\"https:\/\/stackoverflow.com\/questions\/28221191\/theano-cuda-error-an-illegal-memory-access-was-encountered\">this<\/a> now instead. Not sure if this is related.)<\/p>\n\n<p>Via <a href=\"https:\/\/stackoverflow.com\/users\/4580205\/markus\">Markus<\/a>, who got the same error:<\/p>\n\n<pre><\/pre>\n\n<p>With CUDA 6.5, Windows 8.1, Python 2.7, GTX 970M.<\/p>\n\n<blockquote>\n  <p>The error only occurs in my own network, if I run the LeNet example from Theano, it runs fine. Though the network is compiling and running fine on the CPU (and also on the GPU for some colleagues using Linux). Does anyone have an idea what the problem could be?<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34883787","CreationDate":"2016-01-19T17:56:26.523","ParentId":null,"OwnerUserId":"5811925","Title":null,"Body":"<p>Ran into a similar issue, and fwiw, in my case it was solved by eliminating the import of another library that used pycuda. It appears theano really does not like to share.<\/p>\n"},{"AnswerId":"31155012","CreationDate":"2015-07-01T07:24:54.943","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>Just for reference in case anyone stumbles upon this:<\/p>\n\n<p>This doesn't occur anymore for me. I'm not exactly sure what fixed it, but I think the main difference is that I avoid any multithreading and forks (without exec). This caused many similar problems, e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/28221191\/theano-cuda-error-an-illegal-memory-access-was-encountered\">Theano CUDA error: an illegal memory access was encountered (StackOverflow)<\/a>, and <a href=\"https:\/\/groups.google.com\/d\/topic\/theano-users\/Pu4YKlZKwm4\/discussion\" rel=\"nofollow noreferrer\">Theano CUDA error: an illegal memory access was encountered (Google Groups discussion)<\/a>. Esp. that discussion on Google Groups is very helpful.<\/p>\n\n<p>Theano functions are not multithreading safe. However, that is not a\nproblem for me because I'm only using it in one thread. However, I\nstill think that other threads might cause these problems. Maybe it is\nrelated to the GC of Python which frees some Cuda_Ndarray in some\nother thread while the theano.function is running.<\/p>\n\n<p>I looked a bit at the <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/cuda_ndarray.cu\" rel=\"nofollow noreferrer\">relevant Theano code<\/a> and not sure if it covers\nall such cases.<\/p>\n\n<p>Note that you might even not be aware that you have some background\nthreads. Some Python stdlib code can spawn such background threads.\nE.g. multiprocessing.Queue will do that.<\/p>\n\n<p>I cannot avoid having multiple\nthreads, and until this is fixed in Theano, I create a new subprocess\nwith a single thread where I do all the Theano work. This also has\nseveral advantages such as: More clear separation of the code, being\nfaster in some cases because it all really runs in parallel, and being\nable to use multiple GPUs.<\/p>\n\n<p>Note that just using the multiprocessing module did not work for me\nthat well because there are a few libs (Numpy and others, and maybe\nTheano itself) which might behave bad in a forked process (depending\non the versions, the OS and race conditions). Thus, I needed a real\nsubprocess (fork + exec, not just fork).<\/p>\n\n<p>My code is <a href=\"https:\/\/gist.github.com\/albertz\/4177e40d41cb7f9f7c68\" rel=\"nofollow noreferrer\">here<\/a>, in case anyone is interested in this.<\/p>\n\n<p>There is ExecingProcess which is modeled after multiprocessing.Process\nbut does a fork+exec. (Btw, on Windows, the multiprocessing module\nwill anyway do this, because there is no fork on Windows.)\nAnd there is AsyncTask which adds up a duplex pipe to this which works\nwith both ExecingProcess and the standard multiprocessing.Process.<\/p>\n\n<p>See also: <a href=\"https:\/\/github.com\/Theano\/Theano\/wiki\/Using-Multiple-GPUs\" rel=\"nofollow noreferrer\">Theano Wiki: Using multiple GPUs<\/a><\/p>\n"}]}
{"QuestionId":28204536,"AnswerCount":1,"Tags":"<neural-network><connectivity><theano>","CreationDate":"2015-01-28T23:40:37.007","AcceptedAnswerId":null,"OwnerUserId":4496286.0,"Title":"Theanets: Removing individual connections","Body":"<p>How do you remove connections in Theanets? I'd like to create custom connectivity between an input layer, a single hidden layer, and an output layer. But the only defaults are feedforward all-to-all architectures or recurrent architectures. I'd like to remove specific connections from the all-to-all connectivity and then train the network.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"28290504","CreationDate":"2015-02-03T02:46:35.317","ParentId":null,"OwnerUserId":"2014584","Title":null,"Body":"<p>(Developer of <code>theanets<\/code> here.)<\/p>\n\n<p>This is currently not directly possible with <a href=\"http:\/\/theanets.readthedocs.org\" rel=\"nofollow\">theanets<\/a>. For computational efficiency the underlying computations in feedforward networks are implemented as simple matrix operations, which are fast and can be executed on a GPU for sometimes dramatic speedups.<\/p>\n\n<p>You can, however, <em>initialize<\/em> the weights in a layer so that some (or many) of the weights are zero. To do this, just pass a dictionary in the <code>layers<\/code> list, and include a <code>sparsity<\/code> key:<\/p>\n\n<pre><code>import theanets\n\nnet = theanets.Autoencoder(\n    layers=(784, dict(size=1000, sparsity=0.9), 784))\n<\/code><\/pre>\n\n<p>This initializes the weights for the layer so that the given fraction of weights are zeros. The weights are, however, eligible for change during the training process, so this is only an initialization trick.<\/p>\n\n<p>You can, however, implement a custom <a href=\"http:\/\/theanets.readthedocs.org\/en\/latest\/generated\/theanets.layers.Layer.html\" rel=\"nofollow\"><code>Layer<\/code><\/a> subclass that does whatever you like, as long as you stay within the <a href=\"http:\/\/deeplearning.net\/software\/theano\" rel=\"nofollow\">Theano<\/a> boundaries. You could, for instance, implement a type of feedforward layer that uses a mask to ensure that some weights remain zeros during the feedforward computation.<\/p>\n\n<p>For more details you might want to ask on the <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/theanets\" rel=\"nofollow\">theanets mailing list<\/a>.<\/p>\n"}]}
{"QuestionId":28205589,"AnswerCount":1,"Tags":"<function><updates><theano>","CreationDate":"2015-01-29T01:32:58.660","AcceptedAnswerId":28312564.0,"OwnerUserId":1105571.0,"Title":"The update order of theano function's update list","Body":"<p>Theano function's parameter updates take a list of pair, in which each pair specifys a shared symbolic variable and its new expression after the calculating the function outputs. \nI wonder whether there is any order for the updating procedure. \nThe order will matters if two symbolic variable's new expression relies on each other and the updating procedure will use the updated symbolic variable for updating other symbolic variables that rely on it. \nFor examples, this list might looks like this,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have written some function to test this. The result seems to show that it always use the old value in the expression (second term in the pair) to update the symbolic variable in the first term, i.e., <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Is this a defined behavior?<\/p>\n\n<p>However I found the implementation of momentum <a href=\"http:\/\/nbviewer.ipython.org\/github\/craffel\/theano-tutorial\/blob\/master\/Theano%20Tutorial.ipynb\" rel=\"nofollow\">here<\/a>, \nHere are the codes for generating the update list and param_update symbolic variables<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Then in the first iteration, the param will not be updated, because param_updates are all zero. In my understanding, param_update should be updated first, and then use that for updating param.<\/p>\n","answers":[{"AnswerId":"28312564","CreationDate":"2015-02-04T02:33:30.420","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>For the update, it always use the previous value (the value before the Theano function call). So you found the right thing.<\/p>\n\n<p>For momentum, I think it is normal that there is a delay.<\/p>\n"}]}
{"QuestionId":28209743,"AnswerCount":1,"Tags":"<cuda><theano><cublas>","CreationDate":"2015-01-29T08:07:10.543","AcceptedAnswerId":31155045.0,"OwnerUserId":133374.0,"Title":"Crash in Theano\/CUDA exit in cuStreamDestroy","Body":"<p>I have an application which is linked against CPython and calls Theano+CUDA code from there.<\/p>\n\n<p>The application itself also uses CUDA and Cublas. But as they are creating their own handle, I think they should not get into problems.<\/p>\n\n<p>The GPU is in exclusive mode, i.e. only used by that process. I got that crash both on a Nvidia Tesla K20c and a Nvidia GeForce GTX 680. On Ubuntu 12.04. CUDA 6.0. Latest Theano from Git.<\/p>\n\n<p>Sometimes, but not always, it crashes when it does the CPython cleanup, where Theano indirectly will cleanup its Cublas handles (see  in the stacktrace). The CPython cleanup is done in an  call, maybe that is relevant.<\/p>\n\n<p>This is the stacktrace:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a known error? Where could I even start to try to debug this? How do I fix this?<\/p>\n","answers":[{"AnswerId":"31155045","CreationDate":"2015-07-01T07:26:18.017","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>I call <code>Py_Finalize<\/code> earlier now, not via <code>std::atexit<\/code>, and so far, I did not have this crash anymore. So I guess that is the solution.<\/p>\n"}]}
{"QuestionId":28216931,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2015-01-29T14:18:27.990","AcceptedAnswerId":null,"OwnerUserId":1452494.0,"Title":"Error following first Theano program example","Body":"<p>I'm totally new to theano and following this simple intro exercise to theano found here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/introduction.html#introduction\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/introduction.html#introduction<\/a><\/p>\n\n<p>The idea is to simply declare some tensor variables and wrap them in a function, it is the most simple thing you could possibly do with theano<\/p>\n\n<p>the exact code is:<\/p>\n\n<pre><\/pre>\n\n<p>However, I get this traceback:<\/p>\n\n<pre><\/pre>\n\n<p>My only thought is that it may be python3 related, but that should not be the case. Please help.<\/p>\n","answers":[{"AnswerId":"28219065","CreationDate":"2015-01-29T15:57:51.160","ParentId":null,"OwnerUserId":"1452494","Title":null,"Body":"<p>Problem is not including the BLAS in the most recent version of theano. Solved when you pull the bleeding-edge version:<\/p>\n\n<pre><code>pip3 install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n"},{"AnswerId":"28312623","CreationDate":"2015-02-04T02:39:59.137","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano code base do not work out of the box for python2 and python3. It need to get converted. This is done during the installation of Theano. When installed via pip, this is done automatically. If you cloned\/downloded the source code, you need to install it with:<\/p>\n\n<pre><code>python setup.py install\n<\/code><\/pre>\n\n<p>Here is a Theano ticket with more information:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2317\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2317<\/a><\/p>\n\n<p>Also, for python 3 support, you should use the development version line the other answer:<\/p>\n\n<pre><code>pip3 install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>But this isn't related to BLAS as it is written.<\/p>\n"}]}
{"QuestionId":28267549,"AnswerCount":0,"Tags":"<c++><boost><osx-yosemite><boost-python><caffe>","CreationDate":"2015-02-01T20:35:32.903","AcceptedAnswerId":null,"OwnerUserId":2831921.0,"Title":"boost: command not found on mac OS 10.10","Body":"<p>I have been trying to install caffe Deep learning framework on my MAC OS 10.10, but there is a command which needs to be executed in this process.<\/p>\n\n<pre><\/pre>\n\n<p>I installed boost library version 1.55.0 , still it gives this error.<\/p>\n\n<pre><\/pre>\n\n<p>Also in this command:     <\/p>\n\n<p>I get this error:<br>\n<\/p>\n\n<p>Where am I getting wrong?<\/p>\n\n<p>Are environment variable needed to be set for this command to run.<\/p>\n\n<p>Also,<\/p>\n\n<p>What is meant by changing homebrew formula?<\/p>\n\n<p>I am newbie in python. Before posting this on stack overflow, I tried a lot to figure our this error, but it is almost a week, I am not able to overcome. If possible, some step by step guide of caffe installation on mac 10.10 will be helpful.<\/p>\n\n<p>Thanks a lot in advance.<\/p>\n","answers":[]}
{"QuestionId":28272791,"AnswerCount":1,"Tags":"<lua><mnist><torch>","CreationDate":"2015-02-02T07:19:41.963","AcceptedAnswerId":28325844.0,"OwnerUserId":2138524.0,"Title":"Issue with neuralnetwork_turial.lua with data preprocessing","Body":"<p>I have installed the <a href=\"https:\/\/github.com\/nicholas-leonard\/dp\" rel=\"nofollow\">torch deep learning module<\/a> by first -ing and later using  and the installation was succussful. The  works well in the torch prompt.<\/p>\n\n<p>But when I try to execute the <a href=\"https:\/\/github.com\/nicholas-leonard\/dp\/blob\/master\/examples\/neuralnetwork_tutorial.lua\" rel=\"nofollow\">neuralnetwork_tutorial.lua<\/a>(), it throws the following errors.<\/p>\n\n<pre><\/pre>\n\n<p>I put some  statements in those scripts to understand the flow. I happen to notice that in <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/File.lua\" rel=\"nofollow\">File.lua<\/a> the first step after getting the object is to determine the type of the object; of which 8 have been declared. The types have been declared through 0 to 7, 0 being . However the code fails, as it detects a type 28(??).<\/p>\n\n<p>Kindly any help where I am going wrong? Or where to look into to find the issue?<\/p>\n\n<p>P.S.: The script downloads the data on its own, however due to certain standard corporate proxy setting issues, it could not download. Therefore, I personally downloaded the data <a href=\"https:\/\/stife076.files.wordpress.com\/2014\/08\/mnist2.zip\" rel=\"nofollow\">MNIST<\/a> and stored it in the specific data directory. If this could be a clue??<\/p>\n","answers":[{"AnswerId":"28325844","CreationDate":"2015-02-04T16:01:04.377","ParentId":null,"OwnerUserId":"49985","Title":null,"Body":"<p>Okay, so it was a bug in the code (serialized MNIST wasn't cross-platform). Fixed by serializing dataset using ascii format instead of binary.<\/p>\n"}]}
{"QuestionId":28287599,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-02-02T21:59:54.243","AcceptedAnswerId":null,"OwnerUserId":2831921.0,"Title":"make all gives linker command fail error in caffe installation","Body":"<p>After following all the steps in caffe installation, I find this error:<\/p>\n\n<pre><\/pre>\n\n<p>What can be done. I tried all possible cases, reinstalling and finding in githun issues.<\/p>\n\n<p>Thanx in advance.<\/p>\n","answers":[{"AnswerId":"28315765","CreationDate":"2015-02-04T07:27:15.450","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Seems like you have not installed OpenCV in your system.\nIf you are using the <a href=\"http:\/\/brew.sh\/\" rel=\"nofollow\">Homebrew<\/a> package manager, try executing the following commands<\/p>\n\n<pre><code>brew tap homebrew\/science\nbrew install homebrew\/science\/opencv\n<\/code><\/pre>\n\n<p>If the above two line of commands do perform without any error, try building Caffe again.<\/p>\n"}]}
{"QuestionId":28299422,"AnswerCount":2,"Tags":"<python><math><gradient><theano><automatic-differentiation>","CreationDate":"2015-02-03T12:52:36.083","AcceptedAnswerId":null,"OwnerUserId":4056875.0,"Title":"How does theano implement computing every function's gradient?","Body":"<p>I have a question about Theano's implementation.\nHow the theano get the gradient of every loss function by the following function(T.grad)? Thank you for your help.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28301771","CreationDate":"2015-02-03T14:49:35.437","ParentId":null,"OwnerUserId":"3088138","Title":null,"Body":"<p>Look up <strong>Automatic differentiation<\/strong> and there the <em>backwards mode<\/em> that is used to efficiently evaluate gradients.<\/p>\n\n<p>Theano is, as far as I can see, a hybrid between the code-rewriting and operator based approach. It uses operator overloading in python to construct the computational graph, then optimizes it and generates from that graph (optimized) sequences of operations to evaluate the required inkds of derivatives.<\/p>\n"},{"AnswerId":"39082247","CreationDate":"2016-08-22T14:33:40.580","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p><strong>Edit<\/strong>: this answer was wrong in saying that Theano uses Symbolic Differentiation. My apologies.<\/p>\n\n<p>Theano implements <strong>reverse mode autodiff<\/strong>, but confusingly they call it \"symbolic differentiation\". This is misleading because symbolic differentiation is something quite different.  Let's look at both.<\/p>\n\n<p><strong>Symbolic differentiation<\/strong>: given a graph representing a function <code>f(x)<\/code>, it uses the chain rule to compute a new graph representing the derivative of that function <code>f'(x)<\/code>. They call this \"compiling\" <code>f(x)<\/code>. One problem with symbolic differentiation is that it can output a very inefficient graph, but Theano automatically simplifies the output graph.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>\"\"\"\nf(x) = x*x + x - 2\nGraph =\n          ADD\n         \/   \\\n        MUL  SUB\n       \/  \\  \/  \\\n       x  x  x  2\n\nChain rule for ADD=&gt; (a(x)+b(x))' = a'(x) + b'(x)\nChain rule for MUL=&gt; (a(x)*b(x))' = a'(x)*b(x) + a(x)*b'(x)\nChain rule for SUB=&gt; (a(x)-b(x))' = a'(x) - b'(x)\nThe derivative of x is 1, and the derivative of a constant is 0.\n\nDerivative graph (not optimized yet) =\n          ADD\n         \/   \\\n       ADD    SUB\n      \/  |    |  \\\n   MUL  MUL   1   0\n  \/  |  |  \\\n 1   x  x   1\n\nDerivative graph (after optimization) =\n          ADD\n         \/   \\\n       MUL    1\n      \/   \\\n     2     x\n\nSo: f'(x) = 2*x + 1\n\"\"\"\n<\/code><\/pre>\n\n<p><strong>Reverse mode autodiff<\/strong>: works in two passes through the computation graph, first going forward through the graph (from the inputs to the outputs), and then backwards using the chain rule (if you are familiar with backpropagation, this is exactly how it computes gradients).<\/p>\n\n<p>See <a href=\"http:\/\/alexey.radul.name\/ideas\/2013\/introduction-to-automatic-differentiation\/\" rel=\"noreferrer\">this great post<\/a> for more details on various automatic differentiation solutions and their pros&amp;cons.<\/p>\n"}]}
{"QuestionId":28398705,"AnswerCount":1,"Tags":"<lua><libjpeg><torch>","CreationDate":"2015-02-08T20:00:40.180","AcceptedAnswerId":null,"OwnerUserId":620625.0,"Title":"How do I get the right libjpeg dylib for Lua?","Body":"<p>So, I am running torch on OSX (See error in bold below):<\/p>\n\n<pre><\/pre>\n\n<p>Specifically this line:\n<strong>error loading module 'libjpeg' from file '\/usr\/local\/Cellar\/jpeg\/8d\/lib\/libjpeg.dylib':\n        dlsym(0x7fd564000320, luaopen_libjpeg): symbol not found\n    warning:  could not be loaded (is it installed?)<\/strong><\/p>\n\n<p>Seems like I don't have the right dylib? If so, where do I get it?<\/p>\n","answers":[{"AnswerId":"28401169","CreationDate":"2015-02-09T00:39:01.163","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>this happens when there's two libjpeg's installed on your machine, and one conflicts with another.\nstart torch like this:<\/p>\n\n<blockquote>\n  <p>export DYLD_LIBRARY_PATH=\/usr\/local\/lib:$DYLD_LIBRARY_PATH<\/p>\n  \n  <p>th<\/p>\n<\/blockquote>\n\n<p>it should hopefully work with this.<\/p>\n"}]}
{"QuestionId":28402666,"AnswerCount":2,"Tags":"<python><machine-learning><theano><perceptron><deep-learning>","CreationDate":"2015-02-09T04:04:44.593","AcceptedAnswerId":null,"OwnerUserId":4544800.0,"Title":"How to access to a theano symbolic variable's value inside a class?","Body":"<p>I want to access to the value of my_classifier.y_binary. My goal is to compute my_classifier.error.<\/p>\n\n<p>I know how to get the value of my_classifier.y_hat using eval but I don't know how to use it when the input is a self parameter.<\/p>\n\n<p>Thanks<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"28612148","CreationDate":"2015-02-19T16:52:39.107","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>If you want the value of a node in the graph you'll need to compile a function to get it.  I think something like <\/p>\n\n<pre><code>y_binary = theano.function(inputs = [X,], outputs=my_classifier.y_binary, allow_input_downcast=True)\n<\/code><\/pre>\n\n<p>should give you the function <code>y_binary()<\/code> and calling <code>y_binary(features)<\/code> should forward propagate the network and yield the binarized output.<\/p>\n"},{"AnswerId":"37125559","CreationDate":"2016-05-09T21:12:46.273","ParentId":null,"OwnerUserId":"2498151","Title":null,"Body":"<p>A compiled function is a much better choice, but while you're setting stuff up a quick and dirty way is like this:<\/p>\n\n<p>like this:<\/p>\n\n<pre><code>while (epoch &lt; n_epochs):    \n        epoch = epoch + 1    \n        for minibatch_index in range(n_train_batches):\n            minibatch_avg_cost = train_model(minibatch_index)\n            iter = (epoch - 1) * n_train_batches + minibatch_index\n            print(\"**********************************\")\n            print(classifier.hiddenLayer.W.get_value()) \n<\/code><\/pre>\n\n<p>full code here: <a href=\"https:\/\/github.com\/timestocome\/MiscDeepLearning\/blob\/master\/MLP_iris2.py\" rel=\"nofollow\">https:\/\/github.com\/timestocome\/MiscDeepLearning\/blob\/master\/MLP_iris2.py<\/a><\/p>\n\n<p>I think in your example you'd use 'my_classifier.w.get_value()'<\/p>\n"}]}
{"QuestionId":28418823,"AnswerCount":1,"Tags":"<python><theano><softmax>","CreationDate":"2015-02-09T20:34:08.667","AcceptedAnswerId":34094065.0,"OwnerUserId":2040628.0,"Title":"getting error with softmax and cross entropy in theano","Body":"<p>I'm implementing a DNN with Theano. At the last layer of DNN, I'm using a softmax as a nonlinear function from <\/p>\n\n<p>As a lost function i'm using cross entropy from \nBut I get a strange error:\n\"The following error happened while compiling the node', GpuDnnSoftmaxGrad{tensor_format='bc01' ...\"<\/p>\n\n<p>I'm a newbie with theano and can't figure out what's wrong with this model. Your help is appreciated\nPS: my guess is it is somehow related to the fact that softmax takes a 2D tensor and returns a 2D tensor.<\/p>\n\n<p>PS2:I'm using the bleeding edge Theano (just cloned) my CUDA version is old it is 4.2 BUT I'm almost sure that that's not the problem since I'm working without error with other DNN tools written based on Theano.\nI'm using pylearn2 to accelerate and that's not the problem either since I already used it successfully with the current Theano and CUDA in another DNN.<\/p>\n\n<p>The error happens at this line: <\/p>\n\n<p>The full error message is:\n<\/p>\n\n<p>The Cross entropy funcion I'm using is defined as:\n\nwhere input is the output of the softmax layer and target_y is the labels.<\/p>\n","answers":[{"AnswerId":"34094065","CreationDate":"2015-12-04T17:27:48.567","ParentId":null,"OwnerUserId":"2040628","Title":null,"Body":"<p>solved. I had to use T.nnet.categorical_crossentropy since my target variable is an integer vector. <\/p>\n"}]}
{"QuestionId":28427145,"AnswerCount":1,"Tags":"<visual-studio><cuda><theano>","CreationDate":"2015-02-10T08:30:44.227","AcceptedAnswerId":28428405.0,"OwnerUserId":2771315.0,"Title":"CUDA and Theano install Windows 7: Is Visual Studio required?","Body":"<p>I'm in the process of installing Theano and combining it with CUDA on my Windows 7 PC. All the information\/tutorials I've seen require Visual Studio to be installed. Most of the Visual Studio versions are >= 5GB in size which seems like a ridiculous number especially for an IDE. Is Visual Studio required to run\/compile CUDA code, or can I install an alternative IDE? If it is required, what's the smallest (in size) Visual Studio version? <\/p>\n","answers":[{"AnswerId":"28428405","CreationDate":"2015-02-10T09:41:35.637","ParentId":null,"OwnerUserId":"2771315","Title":null,"Body":"<p>Unfortunately, Visual Studio is explicitly required to run some of the CUDA tools. As of February 2015, your best bet is to download VS 2013 Community Edition (<a href=\"http:\/\/www.visualstudio.com\/en-us\/products\/visual-studio-community-vs.aspx\" rel=\"nofollow noreferrer\">http:\/\/www.visualstudio.com\/en-us\/products\/visual-studio-community-vs.aspx<\/a>) and CUDA 6.5 (<a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-downloads<\/a>). <a href=\"https:\/\/stackoverflow.com\/a\/26073714\/2771315\">This<\/a> is a great guide for installing CUDA on Windows 7 and 8.<\/p>\n"}]}
{"QuestionId":28427416,"AnswerCount":2,"Tags":"<python><matrix><normalization><theano>","CreationDate":"2015-02-10T08:50:35.577","AcceptedAnswerId":28527317.0,"OwnerUserId":2040628.0,"Title":"normalize a matrix row-wise in theano","Body":"<p>Lets say I have a Matrix  with size  and I want to normalize it row-wise,i.e., \nthe sum of each row should be one. How can I do this in theano? <\/p>\n\n<p>Motivation: using softmax returns back error for me, so I try to kind of sidestep it by implementing my own version of softmax.<\/p>\n","answers":[{"AnswerId":"34205530","CreationDate":"2015-12-10T15:18:29.940","ParentId":null,"OwnerUserId":"3731804","Title":null,"Body":"<p>or you can also use<\/p>\n\n<pre><code>m\/m.norm(1, axis=1).reshape((m.shape[0], 1))\n<\/code><\/pre>\n"},{"AnswerId":"28527317","CreationDate":"2015-02-15T14:54:59.790","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>See if the following is useful for you:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nm = T.matrix(dtype=theano.config.floatX)\nm_normalized = m \/ m.sum(axis=1).reshape((m.shape[0], 1))\n\nf = theano.function([m], m_normalized)\n\nimport numpy as np\na = np.exp(np.random.randn(5, 10)).astype(theano.config.floatX)\n\nb = f(a)\nc = a \/ a.sum(axis=1)[:, np.newaxis]\n\nfrom numpy.testing import assert_array_equal\nassert_array_equal(b, c)\n<\/code><\/pre>\n"}]}
{"QuestionId":28462854,"AnswerCount":1,"Tags":"<python><matrix><shared><theano>","CreationDate":"2015-02-11T19:27:58.677","AcceptedAnswerId":null,"OwnerUserId":3214017.0,"Title":"Theano matrix multiplication","Body":"<p>I have a piece of code that is supposed to calculate a simple\nmatrix product, in python (using theano). The matrix that I intend to multiply with is a shared variable.<\/p>\n\n<p>The example is the smallest example that demonstrates my problem.<\/p>\n\n<p>I have made use of two helper-functions. floatX converts its input to something of type theano.config.floatX\ninit_weights generates a random matrix (in type floatX), of given dimensions.<\/p>\n\n<p>The last line causes the code to crash. In fact, this forces so much output on the commandline that I can't even scroll to the top of it anymore.<\/p>\n\n<p>So, can anyone tell me what I'm doing wrong?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28463917","CreationDate":"2015-02-11T20:31:38.327","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>This work for me. So my guess is that you have a problem with your blas installation. Make sure to use Theano development version:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>It have better default for some configuration. If that do not fix the problem, look at the error message. There is main part that is after the code dump. After the stack trace. This is what is the most useful normally.<\/p>\n\n<p>You can disable direct linking by Theano to blas with this Theano flag: blas.ldflags=<\/p>\n\n<p>This can cause slowdown. But it is a quick check to confirm the problem is blas.<\/p>\n\n<p>If you want more help, dump the error message to a text file and put it on the web and link to it from here.<\/p>\n"}]}
{"QuestionId":28488849,"AnswerCount":2,"Tags":"<python><opencv><ocr><caffe>","CreationDate":"2015-02-12T22:13:03.463","AcceptedAnswerId":null,"OwnerUserId":1214547.0,"Title":"What is the easiest way to obtain a state-of-the-art handwritten digit classifier?","Body":"<p>I am working on a project involving OCR-ing handwritten digits which uses a typical preprocessing-segmentation-recognition pipeline. I have done the first two stages manually by adjusting some standard algorithms from  for my particular task. For the third stage (recognition), I'd like to use a ready-made classifier.<\/p>\n\n<p>First I tried Tesseract, but it was <a href=\"https:\/\/stackoverflow.com\/questions\/27321553\/forcing-tesseract-to-give-some-answer\">really bad<\/a>. So I started looking into the progress on . Due to its popularity, I'd hoped it would be easy to get a nice high-quality classifier. Indeed, the top answer <a href=\"https:\/\/stackoverflow.com\/questions\/13319730\/suggestions-for-digit-recognition\">here<\/a> suggests using a  tandem, which is conveniently implemented in <a href=\"https:\/\/github.com\/Itseez\/opencv\/blob\/master\/samples\/python2\/digits.py\" rel=\"nofollow noreferrer\">this  sample<\/a>. Unfortunately, it isn't as good as I'd hoped. It keeps confusing 's for 's (where it is obvious for my eye that it is actually a ), which accounts by far for the greatest amount of mistakes my algorithm makes.<\/p>\n\n<p>Here are some examples of the errors made by :<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/9oS7Umy.png\" alt=\"\"><\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/DlZp8Qr.png\" alt=\"\"><\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/ArhPfDK.png\" alt=\"\"><\/p>\n\n<p>The top line are the original digits extracted from the image (higher-resolution images do not exist), the middle line are these digits deskewed, size-normalized and centered, and the bottom line is the output of .<\/p>\n\n<p>I tried to hot-fix this  error by applying a  classifier after  (if  outputs an  run  and return its output instead), but the results were the same.<\/p>\n\n<p>Then I tried to adapt <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/tree\/master\/pylearn2\/scripts\/papers\/maxout\" rel=\"nofollow noreferrer\">this  sample<\/a> which claims to achieve 0.45%  test error. However, after spending a week with  I couldn't make it work. It <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/pylearn-users\/WLm8pZ2Vb0M\" rel=\"nofollow noreferrer\">keeps crashing randomly all the time<\/a>, even in an environment as sterile as an   instance running <a href=\"http:\/\/thecloudmarket.com\/image\/ami-735bda04--ubuntu14-04-mkl-cuda-dl\" rel=\"nofollow noreferrer\">this image<\/a> (I don't even mention my own machine).<\/p>\n\n<p>I know about the existence of , but I haven't tried it.<\/p>\n\n<p>What would be the easiest way to set up a high-accuracy (say, MNIST test error &lt;1%) handwritten digit classifier? Preferably, one that does not require an  card to run. As far as I understand,  (since it heavily relies on ) does. A  interface and an ability to run on  would be a pleasant bonus.<\/p>\n\n<p>Note: I cannot create a new  tag since I don't have enough reputation, but it surely should be there.<\/p>\n","answers":[{"AnswerId":"28561241","CreationDate":"2015-02-17T12:15:35.197","ParentId":null,"OwnerUserId":"987599","Title":null,"Body":"<p>Go ahead and try caffe if you haven't done it already. \nIt is much easier than cuda-convnet to compile, it does not rely on cuda (although it speeds up things considerably) and it has an example for mnist with the Lenet algorithm.<\/p>\n\n<p>look here:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/dev\/examples\/mnist\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/tree\/dev\/examples\/mnist<\/a><\/p>\n"},{"AnswerId":"28488917","CreationDate":"2015-02-12T22:17:40.380","ParentId":null,"OwnerUserId":"764322","Title":null,"Body":"<p>In the webpage of the MINST database you can find at the bottom a benchmark of the state of the art methods with links to their papers:<\/p>\n\n<p><a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a><\/p>\n\n<p>The last entry of the table has the best result with a 0.23% of error (pretty impressive).<\/p>\n\n<p>Short answer: there is no <code>easy<\/code> way to achieve state of the art rate, unless you can accept some 2-5% of error (then use sklearn) or you find the code online.<\/p>\n"}]}
{"QuestionId":28508907,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-02-13T21:18:57.727","AcceptedAnswerId":null,"OwnerUserId":2358537.0,"Title":"Error in using 'theano-nose' command","Body":"<p>After installing Theano from Enthought Canopy on Windows, following the steps here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#id9\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#id9<\/a> , I tried to execute the command theano-nose from Canopy terminal. I got an error message saying \"unable to find theano-nose\". Can someone tell me what might be going wrong?<\/p>\n","answers":[{"AnswerId":"28573484","CreationDate":"2015-02-17T23:52:01.777","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>theano-nose is from the command line. But your shell need to be able to find it.<\/p>\n\n<p>It is highly possible it didn't got installed correctly on Windows or that you need to reboot or log out\/log in. I do not remember Windows particularities on that level.<\/p>\n\n<p>But a simple work around is to start python and run this:<\/p>\n\n<pre><code>import theano\ntheano.test()\n<\/code><\/pre>\n\n<p>This is equivalent for what you want to do.<\/p>\n"}]}
{"QuestionId":28512643,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-02-14T05:10:14.647","AcceptedAnswerId":28515146.0,"OwnerUserId":620625.0,"Title":"Why does torch.Tensor have a libjpeg field?","Body":"<p>I see this:<\/p>\n\n<pre><\/pre>\n\n<p>How does Tensor get the field libjpeg? I don't see it referenced in Tensor.lua.<\/p>\n\n<p>For more context, I am trying to debug this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28515146","CreationDate":"2015-02-14T11:26:54.557","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>A common practice in Torch packages is to use the tensor table as a namespace. This is a trick that is useful to do quick and dirty templated function dispatch.\nFor example, if you load the <em>nn<\/em> package, you will find the functions<\/p>\n\n<pre><code>torch.DoubleTensor.nn.L1Cost_updateOutput\ntorch.FloatTensor.nn.L1Cost_updateOutput\n<\/code><\/pre>\n\n<p>These are usually called according to the type of input tensor. For example:<\/p>\n\n<pre><code>input = torch.FloatTensor()\ninput.nn.L1Cost_updateOutput(...) \n<\/code><\/pre>\n\n<p>This is what you observe with <em>torch.Tensor.libjpeg*<\/em>\nif you use the image loading packages, then you will notice that there will be <em>torch.FloatTensor.libjpeg*<\/em> and <em>torch.DoubleTensor.libjpeg*<\/em><\/p>\n\n<p>I suspect that you might have set the default tensor type to <em>torch.CudaTensor<\/em>, which is when you would observe this error.\nBecause the image package's functions are not defined for a Cuda tensor, the functions <em>torch.CudaTensor.libjpeg*<\/em> will not exist.<\/p>\n\n<p>The solution for you is to set your default tensor type to FloatTensor or DoubleTensor, and create any Cuda tensors as you need them.<\/p>\n"}]}
{"QuestionId":28513586,"AnswerCount":2,"Tags":"<lua><png><torch>","CreationDate":"2015-02-14T07:43:05.743","AcceptedAnswerId":null,"OwnerUserId":4565952.0,"Title":"How to read 16bit png using Lua?","Body":"<p>I am using Ubuntu and torch7 library to deal with 16bit images.<\/p>\n\n<p>It would be best if Lua can read\/write 16bit png files.<\/p>\n\n<p>However, I found that if I try to read them by image.load function, it gives result of only higher 8 bit values.<\/p>\n\n<p>Currently I'm using preprocessed binary files instead, but it is quite cumbersome.<\/p>\n\n<p>Is there any way to read\/write 16 bit png file with Lua?<\/p>\n","answers":[{"AnswerId":"30022598","CreationDate":"2015-05-04T04:25:39.457","ParentId":null,"OwnerUserId":"456980","Title":null,"Body":"<p>I have some OpenCV bindings for torch7. Simply because OpenCV has better image resizing\/warping\/loading than the image library written in torch. It handles the 16bit png images fine.<\/p>\n\n<p>They work on Height x Width x Channel images instead of the torch7 Channel x Height x Width images. This is no problem in practice, because they're convertible by transpose. <\/p>\n\n<p>It's not documented properly but should be very useful to someone! (read the init.lua for a description). <\/p>\n\n<p><a href=\"https:\/\/github.com\/Saulzar\/lua---opencv\" rel=\"nofollow\">https:\/\/github.com\/Saulzar\/lua---opencv<\/a><\/p>\n"},{"AnswerId":"28513733","CreationDate":"2015-02-14T08:09:55.587","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/clementfarabet\/graphicsmagick\" rel=\"nofollow\">https:\/\/github.com\/clementfarabet\/graphicsmagick<\/a><\/p>\n\n<p>The graphicsmagick package should work for 16-bit pngs.<\/p>\n\n<p>You can install it via <\/p>\n\n<pre><code>luarocks install graphicsmagick\n<\/code><\/pre>\n"}]}
{"QuestionId":28552966,"AnswerCount":0,"Tags":"<python><theano><theano-cuda>","CreationDate":"2015-02-17T00:56:52.663","AcceptedAnswerId":null,"OwnerUserId":4480720.0,"Title":"nvcc : fatal error : Unsupported host compiler 'bi'","Body":"<p>I'm trying to use the gpu in theano. I made my .theanorc file and tried to run the following python code: <\/p>\n\n<pre><\/pre>\n\n<p>i got this code from deeplearning.net. <\/p>\n\n<p>The output:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":28581740,"AnswerCount":5,"Tags":"<python><cuda><theano><pycuda>","CreationDate":"2015-02-18T10:49:59.943","AcceptedAnswerId":null,"OwnerUserId":1537631.0,"Title":"Installing Theano with GPU on Windows 8.1 64-bit with Visual Studio 2013","Body":"<p>This Theano Installation is making me mad :(<\/p>\n\n<p>So, I've followed the instructions here on the most voted answer because it seemed like the most similar condiguration from mine and up-to-date version : <a href=\"https:\/\/stackoverflow.com\/questions\/25729969\/installing-theano-on-windows-8-with-gpu-enabled\">Installing theano on Windows 8 with GPU enabled<\/a><\/p>\n\n<p>1- I've installed Cuda v6.5, launched deviceQuery and it works fine.<\/p>\n\n<p>2- I already have Visual Studio 2013 so I haven't installed Visual Studio 2010<\/p>\n\n<p>3- > At the time of writing, Theano on GPU only allows working with 32-bit floats and is primarily built for 2.7 version of Python.<\/p>\n\n<p>So i don't know exactly what is the current state now but I have a friend with the same configuration than mine and he managed to make it work so I guess it's possible. I've installed Python through Anaconda.<\/p>\n\n<p>4- I've installed MinGW and Cygwin<\/p>\n\n<p>5- I've fixed msvc9compiler.py<\/p>\n\n<p>6- Here's the bottleneck : the PyCUDA Installation<\/p>\n\n<p>Here's what I've done:\n- I've used cygwin to extract the pycuda tar file\n- I've executed python configure.py through VS2013 x64 Native Tools Command Prompt than configured siteconfig.py as followed:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>I've executed python setup.py build --compiler=\"msvc\" through VS2013 x64 Native Tools Command Prompt<\/li>\n<li>I've executed python setup.py install through VS2013 x64 Native Tools Command Prompt<\/li>\n<li><p>When I execute the little test in python, here's what's happening:<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n\n<p>Could you please tell me why the hell this doesn't work ? <\/p>\n","answers":[{"AnswerId":"29081282","CreationDate":"2015-03-16T15:54:34.390","ParentId":null,"OwnerUserId":"4677134","Title":null,"Body":"<p>You probably need to add the path to the executables for Visual Studio in your <code>nvcc.profile<\/code> <\/p>\n\n<p>(you can find it in your CUDA bin folder. On my system: <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v6.5\\bin<\/code>).<\/p>\n\n<p>In my case, since I have Visual studio 2010, I added at the end of <code>nvcc.profile<\/code>:<\/p>\n\n<pre><code>\"compiler-bindir =  C:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\bin\\amd64\"\n<\/code><\/pre>\n"},{"AnswerId":"29330809","CreationDate":"2015-03-29T15:11:24.843","ParentId":null,"OwnerUserId":"4234099","Title":null,"Body":"<p>I have written a practical guide for the whole process:<\/p>\n\n<p><a href=\"https:\/\/my6266blog.wordpress.com\/2015\/01\/21\/installing-theano-pylearn2-and-even-gpu-on-windows\/\" rel=\"nofollow\">https:\/\/my6266blog.wordpress.com\/2015\/01\/21\/installing-theano-pylearn2-and-even-gpu-on-windows\/<\/a><\/p>\n\n<p>Good luck! It's not that complicated, just follow the steps one by one.<\/p>\n"},{"AnswerId":"31461817","CreationDate":"2015-07-16T18:25:50.893","ParentId":null,"OwnerUserId":"5124900","Title":null,"Body":"<p>I was able to get Theano installed on my ASUS K501LX Windows 8.1 laptop, with an NVIDIA GeForce 950M GPU, without any hassle whatsoever. I largely followed Maor's post above from March 29th. I was actually shocked at how easy it was! All I needed was the Community Edition of Visual Studio 2013 and the CUDA 7 Toolkit. I then installed Anaconda 3.4 (I used the latest version that's out there now). The one modification I made to Maor's post was installing mingw, via <code>conda install mingw libpython<\/code> immediately after installing Anaconda. Also, since I am using Python 3, I had to change the flags parameter in <code>.theanorc.txt<\/code> to point to <code>C:\\Anaconda3\\libs<\/code>. <\/p>\n\n<p>Upon importing theano, it returned that it was using my GeForce GTX 950M device, and running the <code>theano\\misc\\check_blas.py<\/code> check returned no errors and carried out its tests on my GPU. <\/p>\n\n<p>Happy times! <\/p>\n"},{"AnswerId":"34842046","CreationDate":"2016-01-17T18:32:31.140","ParentId":null,"OwnerUserId":"2924421","Title":null,"Body":"<p>The process was a pretty big hassle, so here is a tutorial for anyone that is interested:<\/p>\n\n<p>All of this has been tested using a clean install of Windows 8.1, with nothing else on it, though it should work fine if you don't have a clean install because this will install all the required versions of the software for you.<\/p>\n\n<p>You need 64-bit windows, 32 bit will not work. You will also need a <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">CUDA compatible graphics card<\/a>, so if you don\u2019t have one you\u2019re stuck for now, sadly. This means that you need a relatively modern NVIDIA graphics card, AMD will not work (it can run OpenCL but not CUDA because CUDA is lame and proprietary).<\/p>\n\n<p>I am installing this on Windows 8.1, but I suspect it should all still work on windows 7 as well.<\/p>\n\n<p>First download <a href=\"http:\/\/sourceforge.net\/projects\/winpython\/files\/WinPython_2.7\/2.7.10.3\/WinPython-64bit-2.7.10.3.exe\/download\" rel=\"nofollow noreferrer\">WinPython<\/a>, (make sure to get python 2.7, version 2.7.10.3, this link points to there) and install it TO A PATH THAT DOES NOT HAVE SPACES IN IT. OTHERWISE THINGS BREAK. I made an Other folder in C:\\ (C:\\Other) and then made a folder named Python27 (C:\\Other\\Python27) and tell the installer to install it in there.<\/p>\n\n<p>Once it is done installing, you will need to add it to your path. Press the windows key and type environment variables, then click \u201cEdit the System Environment Variables\u201d, click Environment Variables in the windows that pops up, scroll down to Path, and then append<\/p>\n\n<pre><code>C:\\Other\\Python27\\python-2.7.10.amd64\n<\/code><\/pre>\n\n<p>Or wherever else you installed WinPython to<\/p>\n\n<p>Then add a semicolon after it, so you get<\/p>\n\n<pre><code>C:\\Other\\Python27\\python-2.7.10.amd64;\n<\/code><\/pre>\n\n<p>This is how you add a specific path to the Path variables, in the future, I will just say to add it here and now give specific steps about how to do that. Note that if you update the system path, the current command prompt windows that are open won\u2019t get that update, and you will have to open a new command prompt window to actually have it use the new path.<\/p>\n\n<p>The purpose of a path is so your command prompt window knows where programs are, because if you call, say<\/p>\n\n<p>python<\/p>\n\n<p>in the command prompt, it will look through every folder in your path until it finds a python.exe. If it can\u2019t find any, it will get angry as it typically would if that program didn\u2019t exist.<\/p>\n\n<p>If you don\u2019t want to clutter your path variable\/if your path variable is full, I put a tutorial <a href=\"https:\/\/stackoverflow.com\/a\/34216503\/2924421\">here<\/a> about how to make it so your path is just appended to when you open a command prompt window via a text file that stores all the paths, instead of having to edit the environment variable itself, if you are interested.<\/p>\n\n<p>You then need to add<\/p>\n\n<pre><code>C:\\Other\\Python27\\python-2.7.10.amd64\\DLL;\nC:\\Other\\Python27\\python-2.7.10.amd64\\Scripts;\n<\/code><\/pre>\n\n<p>to your path as well (again, or wherever else you installed python. For later on I will just say where I installed it and if you installed it somewhere else it should be pretty easy to just tweak the commands accordingly)<\/p>\n\n<p>Next, install <a href=\"https:\/\/www.visualstudio.com\/post-download-vs?sku=community&amp;clcid=0x409\" rel=\"nofollow noreferrer\">visual studio 2015<\/a> community and <a href=\"https:\/\/www.visualstudio.com\/en-us\/downloads\/download-visual-studio-vs.aspx\" rel=\"nofollow noreferrer\">visual studio 2013<\/a>, making sure to install all the tools related to c++ development as well (using custom installation, then under Programming Languages). These don\u2019t need to be in a path without spaces, and they probably won\u2019t let you store them anywhere else anyway and that is OK.<\/p>\n\n<p>Add<\/p>\n\n<pre><code>C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin;\n\nC:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin\\amd64;\n\nC:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\lib;\n\nC:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\lib\\amd64;\n<\/code><\/pre>\n\n<p>To your system path.<\/p>\n\n<p>Install NOT NEWEST DRIVER BECAUSE THEY ARE UNSTABLE, but instead <a href=\"http:\/\/www.nvidia.com\/download\/driverResults.aspx\/89119\/en-us\" rel=\"nofollow noreferrer\">355.60<\/a> because it is known to be very reliable, and new enough. Then install the <a href=\"https:\/\/developer.nvidia.com\/cuda-toolkit-65\" rel=\"nofollow noreferrer\">CUDA toolkit<\/a> (it\u2019s also okay to store this in a path with spaces, it probably won\u2019t give you the option either, but even if it does, just let it store it in the default place it wants to store it to). Version 6.5 is needed because version 7 and above aren\u2019t supported by pycuda. If you have a GTX 9__ you will need to download CUDA from <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads-geforce-gtx9xx\" rel=\"nofollow noreferrer\">here<\/a> instead.<\/p>\n\n<p>This will probably automatically append<\/p>\n\n<pre><code>C:\\Program Files (x86)\\Windows Kits\\8.1\\Redist\\D3D\\x64;\nC:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\bin;\nC:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\libnvvp;\nC:\\Program Files (x86)\\NVIDIA Corporation\\PhysX\\Common;\n<\/code><\/pre>\n\n<p>To your path, if not you will need to do so now.<\/p>\n\n<p>Those are the only three things that can be stored to paths with spaces (these will be in like Program Files or Program Files (x86)), with everything else be very careful to store them to paths that don\u2019t have spaces.<\/p>\n\n<p>Download the <a href=\"http:\/\/sourceforge.net\/projects\/boost\/files\/boost-binaries\/1.55.0\/boost_1_55_0-msvc-12.0-64.exe\/download\" rel=\"nofollow noreferrer\">boost binaries<\/a> (1_55_0 for 64 bit which is the version this link points to), and run the installer, then select to store them to a path without spaces (I stored them to C:\\Other\\boost)<\/p>\n\n<p>Navigate to that directory in the command line, then run<\/p>\n\n<pre><code>bootstrap.bat\n<\/code><\/pre>\n\n<p>and then when it is done run<\/p>\n\n<pre><code>.\\b2\n<\/code><\/pre>\n\n<p>This will start building, and take a long time, and use a lot of space (about 6 GB).<\/p>\n\n<p>It will probably say that 8 targets failed, 8 targets were skipped, and 1075 were updated. This is what one should expect, and is not a problem.<\/p>\n\n<p>Install <a href=\"https:\/\/git-scm.com\/download\/win\" rel=\"nofollow noreferrer\">Git-2.7.0-64-bit<\/a> to some path without a space in it<\/p>\n\n<p>Choose to use Git from the Windows Command Prompt, checkout Windows-style, commit Unix-style endings, use Window\u2019s default console window, and do not enable file system caching.<\/p>\n\n<p>Add<\/p>\n\n<pre><code>C:\\Other\\Git\\bin;\n<\/code><\/pre>\n\n<p>To your system path<\/p>\n\n<p>Next, run the installer for <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=44266\" rel=\"nofollow noreferrer\">VCForPython<\/a>. <\/p>\n\n<p>Download <a href=\"https:\/\/pypi.python.org\/pypi\/pycuda\" rel=\"nofollow noreferrer\">pycuda<\/a> source (pycuda-2015.1.3)<\/p>\n\n<p>Navigate inside that directory, then run<\/p>\n\n<pre><code>python configure.py\n<\/code><\/pre>\n\n<p>this will create a file named siteconf.py.<\/p>\n\n<p>Open this file, and it should look something like<\/p>\n\n<pre><code>BOOST_INC_DIR = []\nBOOST_LIB_DIR = []\nBOOST_COMPILER = 'gcc43'\nUSE_SHIPPED_BOOST = True\nBOOST_PYTHON_LIBNAME = ['boost_python-py27']\nBOOST_THREAD_LIBNAME = ['boost_thread']\nCUDA_TRACE = False\nCUDA_ROOT = 'C:\\\\Program Files\\\\NVIDIA GPU Computing Toolkit\\\\CUDA\\\\v6.5'\nCUDA_INC_DIR = ['${CUDA_ROOT}\/include']\nCUDA_ENABLE_GL = False\nCUDA_ENABLE_CURAND = True\nCUDADRV_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCUDADRV_LIBNAME = ['cuda']\nCUDART_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCUDART_LIBNAME = ['cudart']\nCURAND_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCURAND_LIBNAME = ['curand']\nCXXFLAGS = []\nLDFLAGS = []\n<\/code><\/pre>\n\n<p>modify it so it looks like:<\/p>\n\n<pre><code>BOOST_INC_DIR = ['C:\/Other\/boost']\nBOOST_LIB_DIR = ['C:\/Other\/boost\/lib64-msvc-12.0']\nBOOST_COMPILER = 'msvc'\nUSE_SHIPPED_BOOST = True\nBOOST_PYTHON_LIBNAME = ['boost_python-vc120-mt-1_55']\nBOOST_THREAD_LIBNAME = ['boost_thread-vc110-mt-1_55']\nCUDA_TRACE = False\nCUDA_ROOT = 'C:\\\\Program Files\\\\NVIDIA GPU Computing Toolkit\\\\CUDA\\\\v6.5'\nCUDA_ENABLE_GL = False\nCUDA_ENABLE_CURAND = True\nCUDADRV_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCUDADRV_LIBNAME = ['cuda']\nCUDART_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCUDART_LIBNAME = ['cudart']\nCURAND_LIB_DIR = ['${CUDA_ROOT}\/lib\/Win32', '${CUDA_ROOT}\/lib\/x64']\nCURAND_LIBNAME = ['curand']\nCXXFLAGS = ['\/DBOOST_PYTHON_STATIC_LIB', '\/EHsc']\nLDFLAGS = ['\/LIBPATH:C:\\\\Other\\\\boost\\\\\/lib64-msvc-12.0', '\/FORCE']\n<\/code><\/pre>\n\n<p>then run<\/p>\n\n<pre><code>python setup.py build\n<\/code><\/pre>\n\n<p>followed by<\/p>\n\n<pre><code>python setup.py install\n<\/code><\/pre>\n\n<p>This should install pycuda for you =)<\/p>\n\n<p>To install Theano (with GPU enabled), download <a href=\"https:\/\/github.com\/Theano\/Theano\/releases\/tag\/rel-0.7\" rel=\"nofollow noreferrer\">release 0.7<\/a>, unzip it, navigate inside it using a command prompt, and then type<\/p>\n\n<pre><code>python setup.py install\n<\/code><\/pre>\n\n<p>then go and edit system environment variables, and create one named<\/p>\n\n<pre><code>THEANO_FLAGS\n<\/code><\/pre>\n\n<p>and set it\u2019s value to<\/p>\n\n<pre><code>device=gpu,floatX=float32\n<\/code><\/pre>\n\n<p>Then open a new command prompt, and if you have completed all of the steps above, this should work well =) You can run the code <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"nofollow noreferrer\">here<\/a> to make sure that you are actually running off the GPU.<\/p>\n"},{"AnswerId":"29196085","CreationDate":"2015-03-22T15:20:00.377","ParentId":null,"OwnerUserId":"3536271","Title":null,"Body":"<p>Actually you don't need to install pycuda to get theano working on your windows machine. I'm not an expert but I have Theano installed on Windows 8.1.<\/p>\n\n<p>This is my laptop config: 64-bit, nvcc\/cuda 6.5, Python 2.7.9, WinPython-64bit-2.7.9.3, Windows 8.1, VS2013 and two graphic units (Intel HD Graphics 4600 and NVIDIA GeForce GT 750M).<\/p>\n"}]}
{"QuestionId":28608114,"AnswerCount":1,"Tags":"<python><hdf5><deep-learning><caffe>","CreationDate":"2015-02-19T13:51:41.163","AcceptedAnswerId":28627633.0,"OwnerUserId":4448562.0,"Title":"Error using HDF5 data for training models in caffe","Body":"<p>I am working with caffe and I have been trying to train the caffenet model using HDF5 data I used the prototxt files from ~\/..\/caffe\/examples\/hdf5_classification. But I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>P.S: I am a newbie in python programming <\/p>\n\n<p>Can someone help me regarding this.....\nCheers<\/p>\n","answers":[{"AnswerId":"28627633","CreationDate":"2015-02-20T11:13:43.437","ParentId":null,"OwnerUserId":"4448562","Title":null,"Body":"<p>I didn't include the labels in the hdf file added labels data structure and the problem was solved <\/p>\n"}]}
{"QuestionId":28648345,"AnswerCount":1,"Tags":"<ipython><torch>","CreationDate":"2015-02-21T16:50:33.603","AcceptedAnswerId":28654908.0,"OwnerUserId":659156.0,"Title":"itorch creates a python console, not a torch console","Body":"<p>When I call  I don't get a torch console but an ipython console:<\/p>\n\n<pre><\/pre>\n\n<p>What do I do wrong?<\/p>\n","answers":[{"AnswerId":"28654908","CreationDate":"2015-02-22T05:36:39.517","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>iTorch supports iPython v2.3 or above. Please see the required dependencies.\n You seem to have iPython v 0.1.2, maybe that's a reason you see this behavior.<\/p>\n"}]}
{"QuestionId":28652645,"AnswerCount":0,"Tags":"<logistic-regression><theano>","CreationDate":"2015-02-21T23:28:50.717","AcceptedAnswerId":null,"OwnerUserId":1536499.0,"Title":"Theano - logistic regression example weight vector becomes NaN?","Body":"<p>I am doing a tutorial (code <a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/2_logistic_regression.py\" rel=\"nofollow\">here<\/a>) and video <a href=\"https:\/\/www.youtube.com\/watch?v=S75EdAcXHKk\" rel=\"nofollow\">here<\/a> (13:00 minutes in).  <\/p>\n\n<p>My only change is using the mnist training set from a different location (creating a one-hot encoding) but it is not working.  I literally copy-pasted all the code (except for the mnist loading) in this example.  Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>The weight vector updates once, and becomes all  on the second update.  I am very new to theano, but I am looking for tips to figure this out, especially if someone has already done this tutorial.<\/p>\n\n<p><strong>UPDATE<\/strong>.\nIt looks like the gradient is the issue.<\/p>\n\n<p>When I add this<\/p>\n\n<pre><\/pre>\n\n<p>It prints .  This appears to be the correct usage of  though.<\/p>\n\n<p><strong>UPDATE 2.<\/strong>\nWhen I change the cost function to this:<\/p>\n\n<pre><\/pre>\n\n<p>It is working now but I only have 70% accuracy which is really bad.<\/p>\n\n<p><strong>UPDATE 3.<\/strong>\nI downloaded the MNIST data used in the tutorial and it worked with 92% accuary.<\/p>\n\n<p>I am not sure why my first mnist datasource was dying with the crossentropy cost, and then performing really poor with mean squared error cost function.<\/p>\n","answers":[]}
{"QuestionId":28661843,"AnswerCount":1,"Tags":"<opencv><osx-mavericks><caffe>","CreationDate":"2015-02-22T18:52:01.570","AcceptedAnswerId":null,"OwnerUserId":4594305.0,"Title":"caffe installation error linker error lib\/libcaffe.dylib Error and src\/caffe\/CMakeFiles\/caffe.dir\/all Error","Body":"<p>I have been trying to install caffe on my mac OSX 10.9.5.\nI have been following the official caffe installation from  <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#compilation\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/installation.html#compilation<\/a>. <\/p>\n\n<p>When I am following the cmake installation's \"make all\" in build folder, I keep on getting the following linking error. I have been trying many possible suggestions that I have found on the web but to no avail. <\/p>\n\n<p>Any suggestion is appreciated. Thank you in advance.<\/p>\n\n<pre><\/pre>\n\n<p>The full error log is at <a href=\"https:\/\/github.com\/jackywang529\/myOpenCV\/blob\/master\/OpenCV\/OpenCVTutorial2\/errorLog\" rel=\"nofollow\">https:\/\/github.com\/jackywang529\/myOpenCV\/blob\/master\/OpenCV\/OpenCVTutorial2\/errorLog<\/a><\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"28691308","CreationDate":"2015-02-24T08:42:05.247","ParentId":null,"OwnerUserId":"4594305","Title":null,"Body":"<p>After upgrading from CUDA6.5 to CUDA7.0, the \"make all\" step completed successfully. I also made sure that I removed the edits that I had made to the formulas, which were necessary when I was using CUDA6.5. Such edits is described in (<a href=\"http:\/\/caffe.berkeleyvision.org\/install_osx.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/install_osx.html<\/a> under section libstdc++ installation).<\/p>\n\n<p>Good luck with all caffe users!<\/p>\n"}]}
{"QuestionId":28692209,"AnswerCount":4,"Tags":"<python><caffe>","CreationDate":"2015-02-24T09:33:33.283","AcceptedAnswerId":28979649.0,"OwnerUserId":1452257.0,"Title":"Using GPU despite setting CPU_Only, yielding unexpected keyword argument","Body":"<p>I'm installing Caffe on an Ubuntu 14.04 virtual server with CUDA installed (without driver) using <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM<\/a> as inspiration. During the installation process I edited the MakeFile to include  before building it. However, it seems that Caffe is still trying to make use of the GPU. When I try to run a test example I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix this and run entirely on CPU?<\/p>\n","answers":[{"AnswerId":"34022583","CreationDate":"2015-12-01T14:16:29.283","ParentId":null,"OwnerUserId":"5476099","Title":null,"Body":"<p>just one typo from user2696499<\/p>\n\n<pre><code>if ms != inputs[in_][1:]\n    print(self.inputs[in_])\n    in_shape = self.inputs[in_][1:]\n    m_min, m_max = mean.min(), mean.max()\n    normal_mean = (mean - m_min) \/ (m_max - m_min)\n    mean = resize_image(normal_mean.transpose((1,2,0)),\n                        in_shape[1:]).transpose((2,0,1)) * \\\n                        (m_max - m_min) + m_min\n    '''\n    raise ValueError('Mean shape incompatible with input shape.')\n    '''\n<\/code><\/pre>\n"},{"AnswerId":"28829083","CreationDate":"2015-03-03T10:12:04.610","ParentId":null,"OwnerUserId":"2927205","Title":null,"Body":"<p>There are some problems currently due to a lot of interface changes introduced by the caffe developers. The Python Wrapper is not yet updated with these changes.<\/p>\n\n<p>See this PR which fixes the problem: <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1964\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/1964<\/a><\/p>\n"},{"AnswerId":"33866398","CreationDate":"2015-11-23T08:23:39.993","ParentId":null,"OwnerUserId":"5055922","Title":null,"Body":"<p>The same error was solved in another google group:-<\/p>\n\n<p>Where all you need to do is to change this:<\/p>\n\n<pre><code> mean=np.load(mean_file)\n<\/code><\/pre>\n\n<p>to this:<\/p>\n\n<pre><code>mean=np.load(mean_file).mean(1).mean(1)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/C1J5cO54oRE\/bSOT3EViAgAJ\" rel=\"nofollow\">Google Group: Mean shape incompatible with input shape<\/a><\/p>\n"},{"AnswerId":"28979649","CreationDate":"2015-03-11T06:02:18.843","ParentId":null,"OwnerUserId":"2696499","Title":null,"Body":"<p>I'm gonna add a few words to Mailerdaimon's answer.<\/p>\n\n<p>I followed the installation guide (<a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM<\/a>) to setup Caffe in my vagrant virtual machine. FYI, virtual machines DO NOT support GPU accelerating. Back to the point, after I fix 'CPU \/ GPU switch in example scripts'(<a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2058\">https:\/\/github.com\/BVLC\/caffe\/pull\/2058<\/a>) and add '--print_results --labels_file' options(<a href=\"https:\/\/github.com\/jetpacapp\/caffe\/blob\/master\/python\/classify.py\">https:\/\/github.com\/jetpacapp\/caffe\/blob\/master\/python\/classify.py<\/a>) to 'python\/classify.py', this command '.\/python\/classify.py .\/examples\/images\/cat.jpg foo --print_results' still throws the following error:<\/p>\n\n<pre><code>  Traceback (most recent call last):\n  File \".\/python\/classify.py\", line 175, in &lt;module&gt;\n    main(sys.argv)\n  File \".\/python\/classify.py\", line 129, in main\n    channel_swap=channel_swap)\n  File \"\/home\/vagrant\/caffe\/python\/caffe\/classifier.py\", line 38, in __init__\n    self.transformer.set_mean(in_, mean)\n  File \"\/home\/vagrant\/caffe\/python\/caffe\/io.py\", line 267, in set_mean\n    raise ValueError('Mean shape incompatible with input shape.')\n  ValueError: Mean shape incompatible with input shape.\n<\/code><\/pre>\n\n<p>Then I dump the shape of 'mean'(which is 3*256*256) and 'input'(3*227*227). Obviously these two shapes are incompatible. But old versions of 'set_mean()' do NOT throw the error, so I dig into the python code and find out that old 'set_mean()' function looks like this(python\/caffe\/pycaffe.py, line 195-202, <a href=\"https:\/\/github.com\/jetpacapp\/caffe\/\">https:\/\/github.com\/jetpacapp\/caffe\/<\/a>):<\/p>\n\n<pre><code>if mode == 'elementwise':\n    if mean.shape != in_shape[1:]:\n        # Resize mean (which requires H x W x K input in range [0,1]).\n        m_min, m_max = mean.min(), mean.max()\n        normal_mean = (mean - m_min) \/ (m_max - m_min)\n        mean = caffe.io.resize_image(normal_mean.transpose((1,2,0)),\n                in_shape[2:]).transpose((2,0,1)) * (m_max - m_min) + m_min\n<\/code><\/pre>\n\n<p>But in latest Caffe, contributors encapsulate 'set_mean()' and other transformation functions into class \n'Transformer'. New 'set_mean()' function looks like this(python\/caffe\/io.py, line 253-254, <a href=\"https:\/\/github.com\/BVLC\/caffe\/\">https:\/\/github.com\/BVLC\/caffe\/<\/a>):<\/p>\n\n<pre><code>if ms != self.inputs[in_][1:]:\n    raise ValueError('Mean shape incompatible with input shape.')\n<\/code><\/pre>\n\n<p>Jesus, how could these two be the same function? So I change the new 'set_mean()', comment out the error raising sentence, and add shape re-sizing procedure as in old 'set_mean()'.<\/p>\n\n<pre><code>if ms != ins:\n    print(self.inputs[in_])\n    in_shape = self.inputs[in_][1:]\n    m_min, m_max = mean.min(), mean.max()\n    normal_mean = (mean - m_min) \/ (m_max - m_min)\n    mean = resize_image(normal_mean.transpose((1,2,0)),\n                        in_shape[1:]).transpose((2,0,1)) * \\\n                        (m_max - m_min) + m_min\n    '''\n    raise ValueError('Mean shape incompatible with input shape.')\n    '''\n<\/code><\/pre>\n\n<p>Voila, problem solved.<\/p>\n\n<pre><code>Classifying 1 inputs.\nDone in 1.17 s.\n[('tabby', '0.27933'), ('tiger cat', '0.21915'), ('Egyptian cat', '0.16064'), ('lynx', '0.12844'), ('kit fox', '0.05155')]\n<\/code><\/pre>\n"}]}
{"QuestionId":28710350,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-02-25T03:01:08.990","AcceptedAnswerId":28808297.0,"OwnerUserId":678392.0,"Title":"When does Caffe make copies of the data?","Body":"<pre><\/pre>\n\n<p>Why do the last two lines copy data? Don't the GPU and CPU have both up-to-date contents?<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\" rel=\"noreferrer\">http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html<\/a><\/p>\n","answers":[{"AnswerId":"28808297","CreationDate":"2015-03-02T11:08:33.943","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p><code>.gpu_data<\/code> and <code>.cpu_data<\/code> are used in cases were the <code>data<\/code> is used only as input and will not be modified by the algorithm. <code>.mutable_*<\/code> is used when the data itself gets updated while running the algorithm.<\/p>\n\n<p>Whenever a the data is called, it checks whether the previous statement was a <code>mutable_*<\/code> function call and that too using the same processor (gpu or cpu). If it is using the same processor, data need not be copied. If it is using the other processor, there is a chance that the data might have been updated in the previous <code>.mutable_*<\/code> call and hence a data copy is required.<\/p>\n\n<p><strong>Edit 1<\/strong>\nWhenever the previous instruction is 'mutable', data copy is to be done before the current instruction IF the current instruction is on a different processor. <\/p>\n\n<p>In no other case the data copy takes place except for a special initial condition, ie; when no data was present at all in the GPU memory and hence a copy of the data will take place before *_gpu_data() call.<\/p>\n"}]}
{"QuestionId":28716618,"AnswerCount":1,"Tags":"<neural-network><convolution><theano><conv-neural-network>","CreationDate":"2015-02-25T10:28:44.867","AcceptedAnswerId":28752057.0,"OwnerUserId":4078391.0,"Title":"Theano: Reconstructing convolutions with stride (subsampling) in an autoencoder","Body":"<p>I want to train a simple convolutional auto-encoder using Theano, which has been working great. However, I don't see how one can reverse the  command when subsampling (stride) is used. Is there an efficient way to \"invert\" the convolution command when stride is used, like in the image below?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Zevzj.jpg\" alt=\"Image shamelessly stolen and painted from http:\/cs231n.github.io\/convolutional-networks. \"><\/p>\n\n<p>For example, I want to change the following ... <\/p>\n\n<pre><\/pre>\n\n<p>... into the situation where . The first layer will work work just as expected. However, the second layer will effectively \"do a convolution with stride 1, then throw away half of the outputs\". This is clearly a different operation than what I'm looking for -  won't even have the same number of neurons as length as . What should the second  command be to \"reconstruct\" the original ?<\/p>\n","answers":[{"AnswerId":"28752057","CreationDate":"2015-02-26T20:24:33.360","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>I deduce from this that you intend to have tied weights, i.e. if the first operation were are matrix multiplication with <code>W<\/code>, then the output would be generated with <code>W.T<\/code>, the adjoint matrix. In your case you would thus be looking for the adjoint of the convolution operator followed by subsampling.<\/p>\n\n<p>(EDIT: I deduced wrongly, you can use any filter whatsoever to 'deconvolve', as long as you get the shapes right. Talking about the adjoint is still informative, though. You will be able to relax the assumption afterwards.)<\/p>\n\n<p>Since the convolution operator and subsampling operators are linear operator, lets denote them by <code>C<\/code> and <code>S<\/code> respectively and observe that convolution + subsampling an image <code>x<\/code> would be<\/p>\n\n<pre><code>S C x\n<\/code><\/pre>\n\n<p>and that the adjoint operation on <code>y<\/code> (which lives in the same space as <code>S C x<\/code>) would be <\/p>\n\n<pre><code>C.T S.T y\n<\/code><\/pre>\n\n<p>Now, S.T is nothing other than upsampling to the original image size by adding zeros around all entries of <code>y<\/code> until the right size is obtained.<\/p>\n\n<p>From your post, you seem to be aware of the adjoint of the convolution operator of stride (1, 1) - it is the convolution with reversed filters and reversed <code>border_mode<\/code>, i.e. with <code>filters.dimshuffle(1, 0, 2, 3)[:, :, ::-1, ::-1]<\/code> and switch from <code>border_mode='valid'<\/code> to <code>border_mode='full'<\/code>.<\/p>\n\n<p>Concatenate upsampling and this reverse filter convolution and you obtain the adjoint you seek.<\/p>\n\n<p>Note: There may be ways of exploiting the gradient <code>T.grad<\/code> or <code>T.jacobian<\/code> to obtain this automatically, but I am never sure how this is done exactly.<\/p>\n\n<p>EDIT: There, I wrote it down :)<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nfilters = theano.shared(np.random.randn(4, 3, 6, 5).astype('float32'))\n\ninp1 = T.tensor4(dtype='float32')\n\nsubsampled_convolution = T.nnet.conv2d(inp1, filters, border_mode='valid', subsample=(2, 2))\n\ninp2 = T.tensor4(dtype='float32')\nshp = inp2.shape\nupsample = T.zeros((shp[0], shp[1], shp[2] * 2, shp[3] * 2), dtype=inp2.dtype)\nupsample = T.set_subtensor(upsample[:, :, ::2, ::2], inp2)\nupsampled_convolution = T.nnet.conv2d(upsample,\n     filters.dimshuffle(1, 0, 2, 3)[:, :, ::-1, ::-1], border_mode='full')\n\nf1 = theano.function([inp1], subsampled_convolution)\nf2 = theano.function([inp2], upsampled_convolution)\n\nx = np.random.randn(1, 3, 10, 10).astype(np.float32)\nf1x = f1(x)\ny = np.random.randn(*f1x.shape).astype(np.float32)\nf2y = f2(y)\n\np1 = np.dot(f1x.ravel(), y.ravel())\np2 = np.dot(x.ravel(), f2y[:, :, :-1].ravel())\n\nprint p1 - p2\n<\/code><\/pre>\n\n<p><code>p1<\/code> being equal to <code>p2<\/code> corroborates that f2 is the adjoint of f1<\/p>\n"}]}
{"QuestionId":28717780,"AnswerCount":2,"Tags":"<virtualenv><theano>","CreationDate":"2015-02-25T11:24:39.870","AcceptedAnswerId":28720491.0,"OwnerUserId":326849.0,"Title":"Issue installing Theano in my virtualenv","Body":"<p>I'm trying to install Theano in a virtualenv:<\/p>\n\n<pre><\/pre>\n\n<p>but I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I would like not to depend on any system package, so I didn't use the option \"--system-site-packages\" to create my virtualenv.<\/p>\n\n<p>Anybody can help?<\/p>\n","answers":[{"AnswerId":"28720491","CreationDate":"2015-02-25T13:39:12.883","ParentId":null,"OwnerUserId":"326849","Title":null,"Body":"<p>As pointed out by user1615070, I just had to install numpy and scipy in my virtualenv before installing Theano (to not use the system versions).<\/p>\n"},{"AnswerId":"44598600","CreationDate":"2017-06-16T22:04:26.203","ParentId":null,"OwnerUserId":"3287820","Title":null,"Body":"<p>To be more specific, Theano (at least version 0.8) has the <a href=\"http:\/\/deeplearning.net\/software\/theano_versions\/0.8.X\/install_ubuntu.html#virtualenv\" rel=\"nofollow noreferrer\">specific command in its documentation<\/a>. Here it is:<\/p>\n\n<pre><code>virtualenv --system-site-packages -p python2.7 theano-env\n<\/code><\/pre>\n"}]}
{"QuestionId":28756503,"AnswerCount":1,"Tags":"<theano><pymc3>","CreationDate":"2015-02-27T02:26:10.873","AcceptedAnswerId":null,"OwnerUserId":1361822.0,"Title":"Unable to create lambda function in hierarchical pymc3 model","Body":"<p>I'm trying to create the model shown below with PyMC 3 but can't figure out how to properly map probabilities to the observed data with a lambda function.<\/p>\n\n<pre><\/pre>\n\n<p>The error I get is<\/p>\n\n<pre><\/pre>\n\n<p>In the model I'm attempting to build, the elements of  indicate which element in  gives the probability of the corresponding observed value in  (placed in RV ). In other words,<\/p>\n\n<pre><\/pre>\n\n<p>I'm guessing I need to define the probability with a Theano expression or use Theano  but I don't see how it can be done for this model.<\/p>\n","answers":[{"AnswerId":"28838717","CreationDate":"2015-03-03T18:01:40.623","ParentId":null,"OwnerUserId":"429726","Title":null,"Body":"<p>You should specify your categorical <code>p<\/code> values as <code>Deterministic<\/code> objects before passing them on to <code>w<\/code>. Otherwise, the <code>as_op<\/code> implementation would look something like this:<\/p>\n\n<pre><code>@theano.compile.ops.as_op(itypes=[t.lscalar, t.dscalar, t.dscalar],otypes=[t.dvector])\ndef p(z=z, phi=phi):\n    return [phi[z[i,j]] for i in range(D) for j in range(W)]\n<\/code><\/pre>\n"}]}
{"QuestionId":28764801,"AnswerCount":0,"Tags":"<python><spyder><theano>","CreationDate":"2015-02-27T12:09:17.563","AcceptedAnswerId":null,"OwnerUserId":4614444.0,"Title":"Installing Theano in Winpython using tutorial from deeplearning.net","Body":"<p>I'm trying to install Theano in my Windows Server 2008 OS(64-bit) and I'm following the tutorial given in <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a><\/p>\n\n<p>I followed the \"Winpython\" approach and <\/p>\n\n<p>have reached this step <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#configuring-theano\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#configuring-theano<\/a><\/p>\n\n<p>As Instructed I executed the command<\/p>\n\n<p><\/p>\n\n<p>The tutorial said \"this step will add the  directory to you  environment variable\"<\/p>\n\n<p>However after executing the code when I try to import theano, it shows that the module cannot be found.<\/p>\n\n<p>If I directly copy the  directory into the  folder of my  directory , I get a lot of jumbled messages and errors (over 100 lines) In my ipython console.<\/p>\n\n<p>Please tell me what I am doing wrong. <\/p>\n\n<p>Is there a better tutorial that I can follow ? <\/p>\n","answers":[]}
{"QuestionId":28776650,"AnswerCount":0,"Tags":"<lua><scientific-computing><gsl><torch>","CreationDate":"2015-02-28T00:32:42.440","AcceptedAnswerId":null,"OwnerUserId":297094.0,"Title":"How does GSL-shell compare to Torch","Body":"<p>It looks that <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow\">Lua Torch<\/a> supersedes <a href=\"http:\/\/www.nongnu.org\/gsl-shell\/\" rel=\"nofollow\">GSL Shell<\/a>, especially with the <a href=\"https:\/\/github.com\/facebook\/iTorch\" rel=\"nofollow\">iTorch<\/a> package. How do they compare to each other? When someone will like to use GSL Shell vs Torch? Are they compatible (can one use GSL Shell objects in Torch functions)? <\/p>\n","answers":[]}
{"QuestionId":28779133,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-02-28T07:05:03.637","AcceptedAnswerId":null,"OwnerUserId":115781.0,"Title":"How to tune a training schema for a different data set in Caffe?","Body":"<p>Currently I am following the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"nofollow\">caffe imagenet example<\/a> but apply it on my own training data set. My dataset is about 2000 classes and about 10 ~ 50 images each class. Actually I was classifying vehicle images and the images were cropped to the front, so the images within each class have the same size, the same view angle(almost). <\/p>\n\n<p>I've tried the imagenet schema but looks like it didn't work well and after about 3000 iterations the accuracy was down to 0. So I am wondering is there a practical guide on how to tune the schema?<\/p>\n","answers":[{"AnswerId":"29223230","CreationDate":"2015-03-24T01:02:59.893","ParentId":null,"OwnerUserId":"2484466","Title":null,"Body":"<p>You can delete the last layer in imagenet, add your own last layer with a different name(to fit the number of classes), specify it with a higher learning rate, and specify a lower overall learning rate. There does exist an official example here: <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html<\/a><\/p>\n\n<p>However, if the accuracy was 0 you should check the model parameters first, perhaps it's an overflow<\/p>\n"}]}
{"QuestionId":28806203,"AnswerCount":1,"Tags":"<lua><neural-network><deep-learning><torch>","CreationDate":"2015-03-02T09:20:54.793","AcceptedAnswerId":null,"OwnerUserId":2138524.0,"Title":"How to predict using model generated by Torch?","Body":"<p>I have executed the <a href=\"https:\/\/github.com\/nicholas-leonard\/dp\/blob\/master\/examples\/neuralnetwork_tutorial.lua\" rel=\"nofollow\">neuralnetwork_tutorial.lua<\/a>. Now that I have the model, I would like to test it with some of my own handwritten images. But I have tried many ways by storing the weights, and now by storing the complete model using <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/serialization.md\" rel=\"nofollow\">torch save and load methods<\/a>.<\/p>\n\n<p>However now that I try to predict my own handwritten images(converted to 28X28 DoubleTensor) using <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28814333","CreationDate":"2015-03-02T16:09:32.560","ParentId":null,"OwnerUserId":"49985","Title":null,"Body":"<p>You have two options. <\/p>\n\n<p>One. Use the encapsulated <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/module.md#module\" rel=\"nofollow\">nn.Module<\/a> to forward your <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor\" rel=\"nofollow\">torch.Tensor<\/a>:<\/p>\n\n<pre><code>mlp2 = mlp:toModule(datasource:trainSet():sub(1,2))\ninput = testImageTensor:view(1, 1, 32, 32)\noutput = mlp2:forward(input)\n<\/code><\/pre>\n\n<p>Two. Encapsulate your torch.Tensor into a <a href=\"http:\/\/dp.readthedocs.org\/en\/latest\/view\/index.html#dp.ImageView\" rel=\"nofollow\">dp.ImageView<\/a> and forward that through your <a href=\"http:\/\/dp.readthedocs.org\/en\/latest\/model\/index.html#dp.Model\" rel=\"nofollow\">dp.Model<\/a> :<\/p>\n\n<pre><code>inputView = dp.ImageView('bchw', testImageTensor:view(1, 1, 32, 32))\noutputView = mlp:forward(inputView, dp.Carry{nSample=1})\noutput = outputView:forward('b')\n<\/code><\/pre>\n"}]}
{"QuestionId":28844498,"AnswerCount":0,"Tags":"<machine-learning><anaconda><theano><intel-mkl><passage>","CreationDate":"2015-03-04T00:14:38.263","AcceptedAnswerId":null,"OwnerUserId":3718836.0,"Title":"mkl exception after installing anaconda accelerate","Body":"<p>When trying to train an RNN using Passage I get the following exception when running: model = RNN(layers=layers, cost='bce')<\/p>\n\n<p>Exception: ('The following error happened while compiling the node', Dot22(Reshape{2}.0, Reshape{2}.0), '\\n', 'Compilation failed (return status=1): ld: library not found for -lmkl. clang: error: linker command failed with exit code 1 (use -v to see invocation). ', '[Dot22(, )]')<\/p>\n\n<p>This is a new issue since installing Anaconda Accelerate.<\/p>\n\n<p>Appreciate any and all advice!<\/p>\n","answers":[]}
{"QuestionId":28848133,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-03-04T06:44:18.607","AcceptedAnswerId":28848389.0,"OwnerUserId":639973.0,"Title":"caffe: libglog.so.0 missing (error while loading shared libraries)","Body":"<p>I've installed caffe on a server a while ago, and back then it worked properly.<\/p>\n\n<p>Now I'm following the LeNet MNIST tutorial again (<a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html<\/a>), and running<\/p>\n\n<pre><\/pre>\n\n<p>returns <\/p>\n\n<pre><\/pre>\n\n<p>I've noticed that liibglog.so.0 is not in \/lib which might be the reason for it, but I'm not allowed to copy that file into \/lib directory, since I'm not a root user.<\/p>\n\n<p>Is there workaround for this?<\/p>\n","answers":[{"AnswerId":"28848389","CreationDate":"2015-03-04T07:02:10.050","ParentId":null,"OwnerUserId":"391161","Title":null,"Body":"<p>The easiest way to work around the lack of shared libraries in system directories is to use <code>LD_LIBRARY_PATH<\/code> with the directory where the shared library lives.<\/p>\n\n<p>Before running the the command that requires a library, run the following in the following the same shell.<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=~\/local\/lib\n<\/code><\/pre>\n\n<p>You can also stick this in your <code>.bashrc<\/code> for convenience.<\/p>\n\n<p>An alternate solution is to use the following command line flag while compiling, but that requires mucking with other people's build scripts.<\/p>\n\n<pre><code> -Wl,-rpath,$(DEFAULT_LIB_INSTALL_PATH)\n<\/code><\/pre>\n"}]}
{"QuestionId":28858743,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-03-04T15:47:46.263","AcceptedAnswerId":null,"OwnerUserId":1259448.0,"Title":"Theano, structured v.s. regular gradient","Body":"<p>I don't know whether SO is a good place to ask this question, but I can't find the document talking about the different between structured gradient and regular gradient in the Theano library. What are they ?<\/p>\n","answers":[{"AnswerId":"30510229","CreationDate":"2015-05-28T15:01:04.163","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Do you mean with respect to sparse operations?<\/p>\n\n<p>As mentioned <a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/other_ops.html#sparse-gradient\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<blockquote>\n  <p>There are 2 types of gradients for sparse operations: normal gradient\n  and structured gradient.<\/p>\n<\/blockquote>\n\n<p>The difference only applies to sparse matrices because with sparse matrices you may or may not care about the effect of the \"empty space\" in a sparse matrix on the gradients.<\/p>\n\n<p>More <a href=\"http:\/\/deeplearning.net\/software\/theano\/sandbox\/sparse.html#dot-vs-structureddot\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":28867791,"AnswerCount":3,"Tags":"<ubuntu><boost><include><caffe>","CreationDate":"2015-03-05T00:23:05.107","AcceptedAnswerId":null,"OwnerUserId":1054435.0,"Title":"Building Caffe on Ubuntu: make can't find Boost's include files","Body":"<p>I am following <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"noreferrer\">these instructions<\/a> to install and build Caffe along with its dependencies. I built Boost and got this at the end:<\/p>\n\n<pre><\/pre>\n\n<p>When I run  in the caffe directory, I get this:<\/p>\n\n<pre><\/pre>\n\n<p>What do I need to modify for it to find the include (and lib) files? A specific environment variable? A setting in caffe's Makefile? Something else?<\/p>\n","answers":[{"AnswerId":"43689645","CreationDate":"2017-04-28T22:21:09.837","ParentId":null,"OwnerUserId":"4539629","Title":null,"Body":"<p>This worked for me:<\/p>\n\n<pre><code>cd \/usr\/include\/boost\/thread\nsudo ln -s locks.hpp latch.hpp\n<\/code><\/pre>\n"},{"AnswerId":"28952718","CreationDate":"2015-03-09T22:12:25.020","ParentId":null,"OwnerUserId":"3646384","Title":null,"Body":"<p>Just copy your boost folder that you have built (must be named as \"boost\") to your <code>\/usr\/local\/include<\/code> or <code>\/usr\/include<\/code>. Then run <code>make all<\/code> again. <\/p>\n"},{"AnswerId":"34915365","CreationDate":"2016-01-21T04:23:09.787","ParentId":null,"OwnerUserId":"2351184","Title":null,"Body":"<p>Probably you do not have <code>boost<\/code> installed on your machine<\/p>\n\n<pre><code> sudo apt-get install --no-install-recommends libboost-all-dev\n<\/code><\/pre>\n"}]}
{"QuestionId":28877653,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning>","CreationDate":"2015-03-05T12:14:35.277","AcceptedAnswerId":null,"OwnerUserId":4193583.0,"Title":"Using theano with single examples out of batch","Body":"<p>I'm currently using the machine learning library theano to produce a project which utilises deep learning techniques, to find an internal representation of the MNIST dataset.<\/p>\n\n<p>So far my codebase is pretty much the same as this convolutional net tutorial <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html<\/a><\/p>\n\n<p>Mt problem is arising from the fact that I want to be able to select individual examples out of the test batches (or indeed my own handwritten characters for testing) but I can't seem to get theano to do this.<\/p>\n\n<p>I have tried quite a few things (been stuck for 4+ days) but my latest attempt which I think is getting somewhere looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>tData is a theano.shared variable which contains the testing images, layer3.y_pred is the output I'm interested in.<\/p>\n\n<p>But I keep getting errors related to shape, non-tensor types, and conversions. Anybody who has any experience with theano and knows what's going on under the hood I would really appreciate some input.<\/p>\n","answers":[{"AnswerId":"29837052","CreationDate":"2015-04-24T01:16:07.830","ParentId":null,"OwnerUserId":"3098048","Title":null,"Body":"<p>You can change <code>tData[index]<\/code> to <code>tData[index:index+1]<\/code>.<\/p>\n\n<p>If <code>tData<\/code> is an n by m matrix containing n samples, then theano is expecting training data with dimension (batch_size, m). In your case, <code>tData[index]<\/code> has dimension (m,) while <code>tData[index:index+1]<\/code> has dimension (1,m)<\/p>\n"}]}
{"QuestionId":28912683,"AnswerCount":1,"Tags":"<python><numpy><theano><deep-learning>","CreationDate":"2015-03-07T07:56:54.990","AcceptedAnswerId":28919818.0,"OwnerUserId":844373.0,"Title":"Pass in matrix of images of variables sizes into Theano","Body":"<p>I'm trying to use Theano to do some recognition. All my images are different sizes, and I don't want to resize them because they're paintings so they shouldn't be the same size. I was wondering how to pass in a matrix of images of variable image size lengths into the Theano function.<\/p>\n\n<p>I'm under the impression that this not possible with numpy. Is there an alternative?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"28919818","CreationDate":"2015-03-07T20:49:27.330","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>Unless I'm mistaken in my interpretation of your code, I don't think what you're trying to do makes sense.<\/p>\n\n<p>If I understand correctly, in <code>model()<\/code> you are computing a weighted sum over your image pixels using <code>dot(X, w)<\/code>, where I assume that <code>X<\/code> is an <code>(nimages, npixels)<\/code> array of image data, and <code>w<\/code> is a weight matrix with <em>fixed<\/em> dimensions <code>(784, 10)<\/code>.<\/p>\n\n<p>In order for that dot product to even be computable, <code>X.shape[1]<\/code> (the number of pixels in each of your input images) <em>must<\/em> be equal to <code>w.shape[0]<\/code>.<\/p>\n\n<p>If the sizes of your input images vary, how can you expect to learn a single weight matrix with fixed dimensions?<\/p>\n"}]}
{"QuestionId":28954294,"AnswerCount":1,"Tags":"<cuda><theano>","CreationDate":"2015-03-10T00:40:06.090","AcceptedAnswerId":null,"OwnerUserId":2004857.0,"Title":"Getting Theano to use the GPU and all CPU cores (at the same time)","Body":"<p>I managed to get Theano working with either GPU or multicore CPU on Ubuntu 14.04 by following <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow\">this tutorial<\/a>.<\/p>\n\n<p>First I got multicore working (I could verify that in System Monitor).\nThen, after adding the config below to .theanorc, I got GPU working:<\/p>\n\n<pre><\/pre>\n\n<p>I verified it by running the test from the tutorial and checking the execution times, and also by the log message when running my program: <\/p>\n\n<blockquote>\n  <p>\"Using gpu device 0: GeForce GT 525M\"<\/p>\n<\/blockquote>\n\n<p>But as soon as GPU started working I wouldn't see multicore in System Monitor anymore. It uses just one core at 100% like before.<\/p>\n\n<p>How can I use both? Is it even possible?<\/p>\n","answers":[{"AnswerId":"28976870","CreationDate":"2015-03-11T01:06:30.300","ParentId":null,"OwnerUserId":"4656401","Title":null,"Body":"<p>You can't fully utilize both multicore and GPU at the same time. \nMaybe this can be impoved in the future.<\/p>\n"}]}
{"QuestionId":28956321,"AnswerCount":1,"Tags":"<gpu><restart><convolution><caffe>","CreationDate":"2015-03-10T04:42:26.577","AcceptedAnswerId":28959690.0,"OwnerUserId":3510476.0,"Title":"In CNN with caffe, Can I set up initial caffemodel?","Body":"<p>I have operated CNN using caffe.\nhowever, system was forced termination.\nI have caffemodel so far.<\/p>\n\n<p>Can I restart learning from now on using current caffemodel?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"28959690","CreationDate":"2015-03-10T09:06:52.487","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Caffe supports resuming as explained <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html#resume-training\" rel=\"nofollow\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>We all experience times when the power goes out [...] Since we are snapshotting intermediate results during training, we will be able to resume from snapshots.<\/p>\n<\/blockquote>\n\n<p>This is available via the <code>--snapshot<\/code> option of the main caffe command-line tool, e.g:<\/p>\n\n<pre><code>.\/build\/tools\/caffe train [...] --snapshot=caffenet_train_10000.solverstate\n<\/code><\/pre>\n\n<p>As explained within the doc <code>caffenet_train_10000.solverstate<\/code> is:<\/p>\n\n<blockquote>\n  <p>the solver state snapshot that stores all necessary information to recover the exact solver state.<\/p>\n<\/blockquote>\n\n<p>In particular you can find more precisions about how to configure snapshotting from the <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/solver.html#snapshotting-and-resuming\" rel=\"nofollow\">solver documentation<\/a>.<\/p>\n"}]}
{"QuestionId":28958117,"AnswerCount":1,"Tags":"<android><opencv><image-processing><object-detection><caffe>","CreationDate":"2015-03-10T07:23:39.263","AcceptedAnswerId":null,"OwnerUserId":2631051.0,"Title":"Not able to build caffe in android","Body":"<p>I'm trying to create an android app that can identify the object in the image and gives its name as result. I know caffe-library can be used for this but getting error when i run .\/build.py .<\/p>\n\n<p>command :<\/p>\n\n<pre><\/pre>\n\n<p>Error :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30062929","CreationDate":"2015-05-05T20:44:41.370","ParentId":null,"OwnerUserId":"4476156","Title":null,"Body":"<p>Make sure that the clone is recursive to include the dependencies:<\/p>\n\n<blockquote>\n  <p>*** caffe-android-lib Dependency:<br>\n  Boost-for-Android<br>\n  protobuf<br>\n  Eigen<\/p>\n<\/blockquote>\n\n<p>E.g.:<\/p>\n\n<blockquote>\n  <p>git clone --recursive <a href=\"https:\/\/github.com\/sh1r0\/caffe-android-lib.git\" rel=\"nofollow\">https:\/\/github.com\/sh1r0\/caffe-android-lib.git<\/a><br>\n  cd caffe-android-lib<br>\n  .\/build.py $(NDK_PATH)<\/p>\n<\/blockquote>\n\n<p>Or have you tried to install the dependencies (especially protobuf) from source? <\/p>\n\n<p>Given the dependencies are correctly installed, then you will have a successful caffe-android build:<\/p>\n\n<p>E.g.:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/gist.github.com\/melvincabatuan\/6b5e37444b77326ae7b3\" rel=\"nofollow\">https:\/\/gist.github.com\/melvincabatuan\/6b5e37444b77326ae7b3<\/a><br>\n  ...updated 10980 targets...<br>\n  Done!<br>\n  ...<br>\n  [armeabi-v7a] Install        : libcaffe_jni.so => libs\/armeabi-v7a\/libcaffe_jni.so<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":28985551,"AnswerCount":1,"Tags":"<python><ubuntu><numpy><cuda><caffe>","CreationDate":"2015-03-11T11:33:24.063","AcceptedAnswerId":null,"OwnerUserId":3786150.0,"Title":"Caffe installation in ubuntu 14.04","Body":"<p>I'm following <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-ec2-instance\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-ec2-instance<\/a> link for installing caffe on my machine.\nBut when i write command  all i get these errors<\/p>\n\n<pre><\/pre>\n\n<p>I'm new to this and can't really figure out what package is missing <\/p>\n","answers":[{"AnswerId":"29021849","CreationDate":"2015-03-12T22:54:47.347","ParentId":null,"OwnerUserId":"4665062","Title":null,"Body":"<p>(This worked for an unmerged longjon caffe branch.)<\/p>\n\n<p>In\ncaffe\/include\/caffe\/util\/math_functions.hpp<\/p>\n\n<p>try changing<\/p>\n\n<pre><code>using std::signbit;\nDEFINE_CAFFE_CPU_UNARY_FUNC(sgnbit, y[i] = signbit(x[i])); \n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>\/\/ using std::signbit;\nDEFINE_CAFFE_CPU_UNARY_FUNC(sgnbit, y[i] = std::signbit(x[i]));\n<\/code><\/pre>\n"}]}
{"QuestionId":29023614,"AnswerCount":1,"Tags":"<python><glibc><libc><theano>","CreationDate":"2015-03-13T01:55:32.070","AcceptedAnswerId":29046172.0,"OwnerUserId":1245262.0,"Title":"Problems with a local installation of libc","Body":"<p>I'm trying to run a Theano implementation of alexNet on some machines at work. When I first tried to run it I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Normally, what I would do in this situation is use my package manager to update libc, but this is not an option for work\/administrative reasons. So, what I did instead was to install a local version of libc and to change LD_LIBRARY_PATH in my .bashrc file to point to it. Now, however, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Oddly enough, I get a similar 'relocation error' when I just try to 'ls'.<\/p>\n\n<p>Does anyone know how I can install a version of libc that will only be used by my Python interpreter, but not by everything else?<\/p>\n\n<p>Note: The libstdc++ I'm using is a locally installed version. I installed it for the same reason that I'm trying to install a local version of libc. <\/p>\n\n<p>...<\/p>\n\n<p>OK, I've progressed a little further, but am still stuck. <\/p>\n\n<p>If I return to my old LD_LIBRARY_PATH, in order to avoid errors with commands like 'ls', I can verify that the new (local) libc does, in fact, use the old (system) ld-linux-x86-64.so.2<\/p>\n\n<pre><\/pre>\n\n<p>So, I thought I would change LD_LIBRARY_PATH to look for the new (local) ld-linux-x86-64.so.2<\/p>\n\n<pre><\/pre>\n\n<p>So, my libc should see the new ld-linux-x86-64.so.2. I can verify that this so has the symbol _dl_starting_up:<\/p>\n\n<pre><\/pre>\n\n<p>But, when I try to run the alexNet implementation using LD_PRELOAD to guarantee the use of the right libc, I still get the same error:<\/p>\n\n<pre><\/pre>\n\n<p>Why is the new (local) ld-linux not overriding the old (system) ld-linux? Shouldn't my setting of LD_LIBRARY_PATH have taken care of that?<\/p>\n\n<p>If I try to force the issue, by preloading ld-linux, I get a segmentation fault:<\/p>\n\n<pre><\/pre>\n\n<p>So, now I'm stuck. I don't know<\/p>\n\n<ol>\n<li><p>Why my setting of LD_LIBRARY_PATH didn't force my new (local) libc to use the new (local) ld-linux.<\/p><\/li>\n<li><p>What I should do about the segmentation fault when I force libc to use the new (local) ld-linux.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"29046172","CreationDate":"2015-03-14T06:35:31.273","ParentId":null,"OwnerUserId":"50617","Title":null,"Body":"<blockquote>\n  <p>So, what I did instead was to install a local version of libc and to change LD_LIBRARY_PATH in my .bashrc file to point to it<\/p>\n<\/blockquote>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/a\/8658468\/50617\">this answer<\/a> for why that can't work.<\/p>\n\n<blockquote>\n  <p>Why my setting of LD_LIBRARY_PATH didn't force my new (local) libc to use the new (local) ld-linux.<\/p>\n<\/blockquote>\n\n<p>As above answer explains, <code>ld-linux<\/code> is <em>baked<\/em> into your executable at link time and can't be changed by <code>LD_LIBRARY_PATH<\/code>.<\/p>\n\n<blockquote>\n  <p>What I should do about the segmentation fault when I force libc to use the new (local) ld-linux.<\/p>\n<\/blockquote>\n\n<p><code>LD_PRELOAD<\/code>ing <code>ld-linux<\/code> can <em>never<\/em> work: <code>LD_PRELOAD<\/code> is interpreted <em>by<\/em> <code>ld-linux<\/code>, so you are effectively forcing two separate <code>ld-linux<\/code>es into your process, and that greatly confuses both of them.<\/p>\n\n<p>So, what can you do?<\/p>\n\n<p>It is in fact possible to install two separate versions of glibc on a single host, we do that every day. Instructions <a href=\"https:\/\/stackoverflow.com\/a\/851229\/50617\">here<\/a>.<\/p>\n"}]}
{"QuestionId":29035181,"AnswerCount":1,"Tags":"<python><list><function><matrix><theano>","CreationDate":"2015-03-13T14:44:31.567","AcceptedAnswerId":null,"OwnerUserId":1622610.0,"Title":"theano list of matrices as function argument","Body":"<p>Is it possible to define a function in theano that takes a list of matrices (scalars, vectors) as an input argument?<\/p>\n\n<p>Here a simple example:<\/p>\n\n<pre><\/pre>\n\n<p>What happens if list 'a' has not two but thousands of elements with different shapes?\nthe only thing that comes in my mind is to concatenate these elements to one large matrix and then proceed with it. \nIs not there a nicer solution?<\/p>\n","answers":[{"AnswerId":"29036741","CreationDate":"2015-03-13T15:57:48.457","ParentId":null,"OwnerUserId":"4652942","Title":null,"Body":"<p>Maybe you could try something like this:<\/p>\n\n<pre><code>def my_func(*args): \n    for list in args:\n        # stuff here\n\n# ln being your lists.\nmy_func(l1, l2, l3...)\n<\/code><\/pre>\n\n<p>What happens there is that when you use *args in a function definition it means you can pass any number of positional arguments.<\/p>\n"}]}
{"QuestionId":29044706,"AnswerCount":2,"Tags":"<neural-network><gpu><theano><conv-neural-network>","CreationDate":"2015-03-14T02:13:50.243","AcceptedAnswerId":null,"OwnerUserId":3568055.0,"Title":"How to speed up GPU mode convolutional neural network with theano?","Body":"<p>I'm using theano to implement a convolution neural network. My CPU RAM is 32G and GPU RAM is 2G, but the data is also very big -- almost 5G training data.<\/p>\n\n<p>When the program is running, the computer seems to be frozen and each operation is really slow, even didn't respond. And the CPU mode seems to be at least 2x faster than GPU mode.<\/p>\n\n<p>Is there any way to speed up the GPU convolutional neural network?<\/p>\n","answers":[{"AnswerId":"29529863","CreationDate":"2015-04-09T04:36:45.057","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Make sure to use Theano 0.7 with cudnn, this speed up convolution heavily:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/cuda\/dnn.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/cuda\/dnn.html<\/a><\/p>\n"},{"AnswerId":"37107540","CreationDate":"2016-05-09T03:58:42.723","ParentId":null,"OwnerUserId":"4326150","Title":null,"Body":"<p>In order to use GPU accelleration first thing you need to install CUDA.\nOn the level of Theano configuration(Theano flags\/TheanoRC) there are few ways you can speed-up your model with GPU:<\/p>\n\n<ol>\n<li>Specify usage of GPU \"device = gpu\"<\/li>\n<li>Enable Cuda memory allocation (CnMem) \"cnmem = 0.75\"<\/li>\n<li>Enable CUDNN optimization \"optimizer = cudnn\"<\/li>\n<\/ol>\n\n<p>You can read more about Theano config <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANO_FLAGS\" rel=\"nofollow\">here<\/a> <\/p>\n"}]}
{"QuestionId":29055715,"AnswerCount":1,"Tags":"<python><ubuntu-12.04><pydot><caffe>","CreationDate":"2015-03-14T23:59:34.450","AcceptedAnswerId":29173396.0,"OwnerUserId":1726140.0,"Title":"Drawing network in Caffe causes pydot to throw End of Line errors","Body":"<p>So I just pulled the latest revision of Caffe from the master branch, and went through all the initialization steps. As a quick test, I was trying to run the  script provided, in order to visualize the MNIST Autoencoder example network. \nOn executing the following command:<\/p>\n\n<p><\/p>\n\n<p>Pydot complained, and threw the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I see many more  messages like the ones shown above, and my error log was getting too big, so I didn't post the entire log.\n<a href=\"https:\/\/stackoverflow.com\/questions\/24442094\/pydot-not-playing-well-with-line-breaks\">This post<\/a>, seems to be seeing the same error as me, so I tried to replicate their solution, and changed all the strings in the  method in  to raw strings. But that didn't seem to work. <\/p>\n\n<p>Any suggestions on how I can sort this issue?<\/p>\n\n<p>Thanks!! :)<\/p>\n","answers":[{"AnswerId":"29173396","CreationDate":"2015-03-20T18:25:59.227","ParentId":null,"OwnerUserId":"4695219","Title":null,"Body":"<p>I think the key is in the <em>determine_node_label_by_layertype<\/em> function. This is a block of code that should look something like this (or at least it does in my current version of the repository):<\/p>\n\n<pre><code>def determine_node_label_by_layertype(layer, layertype, rankdir):\n\"\"\"Define node label based on layer type\n\"\"\"\n\n    if rankdir in ('TB', 'BT'):\n        # If graph orientation is vertical, horizontal space is free and\n        # vertical space is not; separate words with spaces\n        separator = ' '\n    else:\n        # If graph orientation is horizontal, vertical space is free and\n        # horizontal space is not; separate words with newlines\n        separator = '\\n'\n<\/code><\/pre>\n\n<p>Replace the <code>separater = '\\n'<\/code> with <code>separater = r\"\\n\"<\/code> and it seemed to work for me.<\/p>\n"}]}
{"QuestionId":29102165,"AnswerCount":2,"Tags":"<neural-network><theano><deep-learning><lstm>","CreationDate":"2015-03-17T14:45:29.613","AcceptedAnswerId":null,"OwnerUserId":1870498.0,"Title":"How to perform multi-label learning with LSTM using theano?","Body":"<p>I have some text data with multiple labels for each document. I want to train a LSTM network using Theano for this dataset. I came across <a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\" rel=\"noreferrer\">http:\/\/deeplearning.net\/tutorial\/lstm.html<\/a> but it only facilitates a binary classification task. If anyone has any suggestions on which method to proceed with, that will be great. I just need an initial feasible direction, I can work on.<\/p>\n\n<p>thanks,\nAmit<\/p>\n","answers":[{"AnswerId":"34081214","CreationDate":"2015-12-04T05:02:32.193","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>1) Change the last layer of the model. I.e.<\/p>\n\n<pre><code>pred = tensor.nnet.softmax(tensor.dot(proj, tparams['U']) + tparams['b'])\n<\/code><\/pre>\n\n<p>should be replaced by some other layer, e.g. sigmoid:<\/p>\n\n<pre><code>pred = tensor.nnet.sigmoid(tensor.dot(proj, tparams['U']) + tparams['b'])\n<\/code><\/pre>\n\n<p>2) The cost should also be changed. <\/p>\n\n<p>I.e.<\/p>\n\n<pre><code>cost = -tensor.log(pred[tensor.arange(n_samples), y] + off).mean()\n<\/code><\/pre>\n\n<p>should be replaced by some other cost, e.g. cross-entropy:<\/p>\n\n<pre><code>one = np.float32(1.0)\npred = T.clip(pred, 0.0001, 0.9999)  # don't piss off the log\ncost = -T.sum(y * T.log(pred) + (one - y) * T.log(one - pred), axis=1) # Sum over all labels\ncost = T.mean(cost, axis=0) # Compute mean over samples\n<\/code><\/pre>\n\n<p>3) In the function <code>build_model(tparams, options)<\/code>, you should replace:<\/p>\n\n<pre><code>y = tensor.vector('y', dtype='int64')\n<\/code><\/pre>\n\n<p>by <\/p>\n\n<pre><code>y = tensor.matrix('y', dtype='int64') # Each row of y is one sample's label e.g. [1 0 0 1 0]. sklearn.preprocessing.MultiLabelBinarizer() may be handy.\n<\/code><\/pre>\n\n<p>4) Change <code>pred_error()<\/code> so that it supports multilabel (e.g. using some metrics like accuracy or F1 score from scikit-learn).<\/p>\n"},{"AnswerId":"29529916","CreationDate":"2015-04-09T04:41:04.307","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You can change the last layer of the model. It would have a vector of target where each element is 0 or 1, depending if you have the target or not.<\/p>\n"}]}
{"QuestionId":29109991,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-03-17T21:24:02.900","AcceptedAnswerId":null,"OwnerUserId":3268282.0,"Title":"Issues compiling caffe in RHEL-6.5 with Tesla K80","Body":"<p>when I am trying to compile caffe on my RHEL machine I have this errors:\nI did follow the instruction of this link: <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/installation.html<\/a> but it does not look like working. \nany help? thanks<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29206662","CreationDate":"2015-03-23T09:19:54.873","ParentId":null,"OwnerUserId":"4702363","Title":null,"Body":"<p>Proper opencv not installed. Even though you yum'ed opencv, it could be too low version. Check \/usr\/local\/include\/opencv2 if there's core\/core.hpp as the warning says.\nDownload from <a href=\"http:\/\/opencv.org\/downloads.html\" rel=\"nofollow\">http:\/\/opencv.org\/downloads.html<\/a> something like 2.4.10 and install it. <\/p>\n"}]}
{"QuestionId":29124840,"AnswerCount":4,"Tags":"<python><deep-learning><caffe>","CreationDate":"2015-03-18T14:32:37.767","AcceptedAnswerId":31391076.0,"OwnerUserId":1452257.0,"Title":"Prediction in Caffe - Exception: Input blob arguments do not match net inputs","Body":"<p>I'm using Caffe for classifying non-image data using a quite simple CNN structure. I've had no problems training my network on my HDF5-data with dimensions n x 1 x 156 x 12. However, I'm having difficulties classifying new data.<\/p>\n\n<p>How do I do a simple forward pass without any preprocessing? My data has been normalized and have correct dimensions for Caffe (it's already been used to train the net). Below is my code and the CNN structure.<\/p>\n\n<p><strong>EDIT:<\/strong> I've isolated the problem to the function '_Net_forward' in pycaffe.py and found that the issue arises as the self.input dict is empty. Can anyone explain why that is? The set is supposed to be equal to the set coming from the new test data:<\/p>\n\n<pre><\/pre>\n\n<p>My code has changed a bit as I now use the IO methods for converting the data into datum (see below). In that way I've filled the kwargs variable with the correct data.<\/p>\n\n<p>Even small hints would be greatly appreciated!<\/p>\n\n<pre><\/pre>\n\n<p><strong>CNN Prototext<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31391076","CreationDate":"2015-07-13T18:44:38.853","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Here is <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/aojN_bmbg74\">the answer from Evan Shelhamer I got on the Caffe Google Groups<\/a>:<\/p>\n\n<blockquote>\n  <p><code>self._inputs<\/code> is indeed for the manual or \"deploy\" inputs as defined\n  by the input fields in a prototxt. To run a net with data layers in\n  through pycaffe, just call <code>net.forward()<\/code> without arguments. No need\n  to change the definition of your train or test nets.<\/p>\n  \n  <p>See for instance code cell [10] of the <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/tutorial\/examples\/01-learning-lenet.ipynb\">Python LeNet example<\/a>.<\/p>\n<\/blockquote>\n\n<p>In fact I think it's clearer in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\">Instant Recognition with Caffe tutorial<\/a>, cell 6:<\/p>\n\n<pre><code># Feed in the image (with some preprocessing) and classify with a forward pass.\nnet.blobs['data'].data[...] = transformer.preprocess('data', caffe.io.load_image(caffe_root + 'examples\/images\/cat.jpg'))\nout = net.forward()\nprint(\"Predicted class is #{}.\".format(out['prob'].argmax()))\n<\/code><\/pre>\n\n<p>In other words, to generate the predicted outputs as well as their\u00a0probabilities using pycaffe, once you have trained your model, you have to first feed the data layer with your input, then perform a forward pass with <code>net.forward()<\/code>.<\/p>\n\n<hr>\n\n<p>Alternatively, as pointed out in other answers, you can use a deploy prototxt that is similar to the one you use to define the trained network but removing the input and output layers, and add the following at the beginning (obviously adapting according to your input dimension):<\/p>\n\n<pre><code>name: \"your_net\"\ninput: \"data\"\ninput_dim: 1\ninput_dim: 1\ninput_dim: 1\ninput_dim: 250\n<\/code><\/pre>\n\n<p>That's what they use in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cifar10\/cifar10_quick.prototxt\">CIFAR10 tutorial<\/a>.<\/p>\n\n<p><sup>(pycaffe really ought to be better documented\u2026)<\/sup><\/p>\n"},{"AnswerId":"31360213","CreationDate":"2015-07-11T18:04:23.697","ParentId":null,"OwnerUserId":"2434201","Title":null,"Body":"<p>I have exactly the same problem. This is what fixed it. <\/p>\n\n<p>First, take the same prototext file as you used to train, remove the two data layers.<\/p>\n\n<p>Then add the block as Mark's above<\/p>\n\n<pre><code>name: \"Name_of_your_net\"\ninput: \"data\"\ninput_dim: 64 \ninput_dim: 1\ninput_dim: 28\ninput_dim: 28\n<\/code><\/pre>\n\n<p>where my input_dim are for mnist, change them to your dim. <\/p>\n\n<p>Everything works.<\/p>\n"},{"AnswerId":"29223315","CreationDate":"2015-03-24T01:12:48.940","ParentId":null,"OwnerUserId":"2484466","Title":null,"Body":"<p>Only due to my own experimental experience, it's not a very good idea to specify train and test net in one file using {PHASE} clause. I got many weird errors when I used net file like that, but when I used older version of net files which contain two files separately, train and test, it worked. However I was using caffe version in Nov 2014, perhaps there's some bug or compatible issues there.<\/p>\n\n<p>Well, when the model is used for prediction, shouldn't there be a deploy file specifying the net structure? If you look at ImageNet you should find imagenet_deploy.prototxt there. Although deploy file is similar to train\/test file, I heard it's a bit different due to some fillers. I don't know if it's the problem, but any discussion is welcome, I need to learn new caffe schema if there exist too<\/p>\n"},{"AnswerId":"29423913","CreationDate":"2015-04-02T23:03:17.247","ParentId":null,"OwnerUserId":"723090","Title":null,"Body":"<pre><code>Even small hints would be greatly appreciated!\n<\/code><\/pre>\n\n<p>I am stuck too so not much help, sorry. Might want to skip to the end.<\/p>\n\n<p><code>net.inputs<\/code> is a @property function which supposedly generated the names of the input layer(s).<\/p>\n\n<pre><code>@property\ndef _Net_inputs(self):\n    return [list(self.blobs.keys())[i] for i in self._inputs]\n<\/code><\/pre>\n\n<p>Where <code>list(self.blobs.keys())<\/code> for you would be<\/p>\n\n<pre><code>['data', 'feature_conv', 'conv1', 'pool1', 'conv2', 'fc1', 'accuracy', 'loss']\n<\/code><\/pre>\n\n<p>Since <code>inputs<\/code> has to match <code>kwargs.keys() = ['data']<\/code> we can conclude that <code>net._inputs<\/code> should have been <code>[0]<\/code>. Somehow.<\/p>\n\n<p>Since <code>_inputs<\/code> isn't used anywhere else in <code>pycaffe.py<\/code> I have a look at <code>_caffe.cpp<\/code>. Around line 222 it says <\/p>\n\n<pre><code>.add_property(\"_inputs\", p::make_function(&amp;Net&lt;Dtype&gt;::input_blob_indices,\n    bp::return_value_policy&lt;bp::copy_const_reference&gt;()))\n<\/code><\/pre>\n\n<p>So <code>_inputs<\/code> are the <code>input_blob_indices<\/code> and it makes sense that these should be <code>[0]<\/code> for your network.<\/p>\n\n<p><code>input_blob_indices<\/code> in turn is simply a function that returns <code>net_input_blob_indices_<\/code> in <code>include\/caffe\/net.hpp<\/code><\/p>\n\n<pre><code>inline const vector&lt;int&gt;&amp; input_blob_indices() const { return net_input_blob_indices_; }\n<\/code><\/pre>\n\n<p>...which is only used in <code>src\/caffe\/net.cpp<\/code>, but I can't find it being defined or assigned anywhere.<\/p>\n\n<p>I have tried with <code>type: Data<\/code> and <code>type: MemoryData<\/code> but that doesn't make a difference. What does work is using<\/p>\n\n<pre><code>input: \"data\"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 227\ninput_dim: 227\n<\/code><\/pre>\n\n<p>...instead of a layer. In that case <code>net._inputs = [0]<\/code> and <code>net.inputs = ['data']<\/code> (actually <code>net._inputs<\/code> is a <code>caffe._caffe.IntVec object<\/code> but <code>list(net._inputs) = [0]<\/code>).<\/p>\n\n<p><strong>TLDR<\/strong>: It is starting to look a lot like a bug so I submitted it: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2246\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/2246<\/a><\/p>\n\n<p>P.s. it seems like you are converting ndarray to datum and then back again. Does this have a purpose?<\/p>\n"}]}
{"QuestionId":29154372,"AnswerCount":0,"Tags":"<python-2.7><theano>","CreationDate":"2015-03-19T20:29:10.687","AcceptedAnswerId":null,"OwnerUserId":3508639.0,"Title":"Find minimum N elements in theano","Body":"<p>I've got a theano function which computes euclidean distances for 2 matrices\u2014 () and  (). The result is an  matrix of pairwise distances of each vector (or row) in  from each vector (or row) in .<\/p>\n\n<pre><\/pre>\n\n<p>Let's say I change the above function to accept a single vector, an array of vectors, and the number of smallest distances. What I want is a theano function that will find the N smallest distances, similar to below:<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to avoid explicitly sorting all the distances, since the number that I want will generally be much much smaller than the total number of distances.<\/p>\n","answers":[]}
{"QuestionId":29158732,"AnswerCount":2,"Tags":"<deep-learning><caffe>","CreationDate":"2015-03-20T03:08:22.057","AcceptedAnswerId":29208039.0,"OwnerUserId":639973.0,"Title":"functional concept of caffe","Body":"<p>I'm a little lost as to the concept of caffe.<\/p>\n\n<p>Is it for unsupervised feature extraction, for example, by feeding a lot of images without label?<\/p>\n\n<p>Or is it a classifier when the inputs are set of values for certain fixed feature dimension?<\/p>\n","answers":[{"AnswerId":"29346744","CreationDate":"2015-03-30T12:52:07.073","ParentId":null,"OwnerUserId":"4634250","Title":null,"Body":"<p>Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and by community contributors.<\/p>\n\n<p>I'd advice to have look in this link, where some useful doc., codes <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe Algor.<\/a>and examples are found: <\/p>\n"},{"AnswerId":"29208039","CreationDate":"2015-03-23T10:33:49.883","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Caffe is a supervised learning algorithm which extracts the features on a fixed MxN dimensional image. The labels of these images are to be passed through during the training phase. Special care is to be taken to select the training input such that objects of two classes may not be present in the same image.<\/p>\n"}]}
{"QuestionId":29164552,"AnswerCount":1,"Tags":"<theano><autoencoder>","CreationDate":"2015-03-20T10:37:19.930","AcceptedAnswerId":null,"OwnerUserId":4082092.0,"Title":"Implementing a saturating autoencoder in theano","Body":"<p>I am trying to implement an autoencoder, using the regularization method described in this paper: <a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/goroshin-lecun-iclr-13.pdf\" rel=\"nofollow\">\"Saturating Auto-Encoders\", Goroshin et. al., 2013<\/a><\/p>\n\n<p>Essentially, this tries to minimize the difference between the output of the hidden layer, and the flat portions of the non-linear function being used to compute the hidden layer output.<\/p>\n\n<p>Assuming we are using a step function as the nonlinearity, with the step being at 0.5, a simple implementation might be:<\/p>\n\n<pre><\/pre>\n\n<p>Then, the regularization cost can be simply:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to implement this functionality in Theano. Have started off with the <a href=\"http:\/\/deeplearning.net\/tutorial\/dA.html\" rel=\"nofollow\">denoising autoencoder code<\/a> available at the Theano website. Have made some basic modifications to it:<\/p>\n\n<pre><\/pre>\n\n<p>The above loss functions puts an L1 penalty on the hidden layer output, which should (hopefully) drive most of them to 0. In place of this simple L1 penalty, I want to use the saturating penalty as given above.<\/p>\n\n<p>Any idea how to do this? Where do I compute the y_prime? How to do it symbolically?<\/p>\n\n<p>I am a newbie to Theano, and still catching up with the symbolic computation part.<\/p>\n","answers":[{"AnswerId":"29964326","CreationDate":"2015-04-30T09:44:46.570","ParentId":null,"OwnerUserId":"1563927","Title":null,"Body":"<p>The nonlinearities in the paper are applied during coding, i.e. during calculation of the hidden values. Therefore, given your code example, they should be applied inside the <code>get_hidden_values()<\/code> function (NOT in the <code>get_cost_updates()<\/code> function). They should be the last piece of processing that <code>get_hidden_values()<\/code> does before returning.<\/p>\n\n<p>Also, don't use <code>numpy.abs<\/code> in your symbolic expression because that demands that numpy does the calculation. Instead you want Theano to do it, so just use <code>abs<\/code> and I think it should work as needed.<\/p>\n"}]}
{"QuestionId":29164812,"AnswerCount":1,"Tags":"<lua><machine-learning><cluster-analysis><torch>","CreationDate":"2015-03-20T10:51:35.927","AcceptedAnswerId":null,"OwnerUserId":249001.0,"Title":"Clustering in Torch","Body":"<p>I am trying to learn the <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow\">Torch<\/a> library for machine learning.<\/p>\n\n<p>I know that the focus of Torch is neural networks, but just for the sake of it I was trying to run kmeans on it. If nothing, Torch implements fast contiguous storage which should be analogous to numpy arrays, and the <a href=\"https:\/\/github.com\/torch\/torch7\/wiki\/Cheatsheet\" rel=\"nofollow\">Torch cheatsheet<\/a> cites the <a href=\"https:\/\/github.com\/koraykv\/unsup\" rel=\"nofollow\">unsup<\/a> library for unsupervised learning, so why not?<\/p>\n\n<p>I already have <a href=\"https:\/\/github.com\/andreaferretti\/kmeans\" rel=\"nofollow\">a benchmark<\/a> that I use for K-means implementations. Even though all the implementations there are intentionally using an unoptimized algorithm (the README explains why), LuaJIT is able to cluster 100000 points in 611ms. An optimized (or shall I say, not intentionally slowed down) implementation in Nim (not on the repository) runs in 68 ms,so I was expecting something in-between.<\/p>\n\n<p>Unfortunately, things are much worse, so I suspect I am doing something awfully wrong. What I have written is<\/p>\n\n<pre><\/pre>\n\n<p>and the running time is around 6 seconds!<\/p>\n\n<p>Can anyone check if I have done something wrong in using Torch\/unsup?<\/p>\n\n<p>If anyone wants to try it, the file  is in the above repository<\/p>\n","answers":[{"AnswerId":"29182271","CreationDate":"2015-03-21T11:38:54.360","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>Can anyone check if I have done something wrong in using Torch\/unsup?<\/p>\n<\/blockquote>\n\n<p>Everything sounds correct (note: using <code>local<\/code> variables is recommended):<\/p>\n\n<ul>\n<li><code>data<\/code> is a 2-dimensional table and you use the corresponding <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/3585ab5\/doc\/tensor.md#torchtensortable\" rel=\"nofollow\">Torch constructor<\/a>,<\/li>\n<li><code>points<\/code> is a 2-dimensional tensor with nb. rows = nb. of points and nb. cols = points dimension (2 here). This is what <code>unsup.kmeans<\/code> <a href=\"https:\/\/github.com\/koraykv\/unsup\/blob\/605f777\/kmeans.lua#L4\" rel=\"nofollow\">expects as input<\/a>.<\/li>\n<\/ul>\n\n<p>What you can do is change the batch size (4th argument). It may impact the performance. You can also use the verbose mode that will output the average time per iteration:<\/p>\n\n<pre><code>-- batch size = 5000, no callback, verbose mode\ncentroids, counts = unsup.kmeans(points, 10, 15, 5000, nil, true)\n<\/code><\/pre>\n"}]}
{"QuestionId":29173761,"AnswerCount":1,"Tags":"<matlab><neural-network><caffe>","CreationDate":"2015-03-20T18:46:24.210","AcceptedAnswerId":29207862.0,"OwnerUserId":2191652.0,"Title":"How to extract image features using caffe matlab wrapper?","Body":"<p>Does anyone use the matlab wrapper for the caffe framework? Is there a way how to extract an 4096 dimensional feature vector from an image?\nI was already following<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/432\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/432<\/a><\/p>\n\n<p>and also tried to remove the last lines in imagenet_deploy.prototxt to remove layers as suggested in another forum on github.<\/p>\n\n<p>But still when I run \"matcaffe_demo(im, 1)\" I only get a 1000 dim vector of scores (for the image net classes).\nAny help would be appreciated\nKind regards<\/p>\n","answers":[{"AnswerId":"29207862","CreationDate":"2015-03-23T10:25:08.457","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>It seems that you might not be calling the correct prototxt file. If the last layer defined in the prototxt has the top blob of 4096 dimension, there is no way for the output to be 1000 dimension. <\/p>\n\n<p>To be sure, try creating a bug in the prototxt file and see whether the program crashes. If it doesn't then the program indeed is reading some other prototxt file.<\/p>\n"}]}
{"QuestionId":29191808,"AnswerCount":2,"Tags":"<caffe><gflags><glog>","CreationDate":"2015-03-22T06:55:52.843","AcceptedAnswerId":null,"OwnerUserId":1282043.0,"Title":"libgflags bad value error for caffe","Body":"<p>I've linked all the required libraries and the caffee confige ran smoothly. But when I want to make the library I get this error:<\/p>\n\n<p>\/usr\/bin\/ld: \/usr\/local\/lib\/libgflags.a(gflags.cc.o): relocation R_X86_64_32S against `std::basic_string, std::allocator >::_Rep::_S_empty_rep_storage' can not be used when making a shared object; recompile with -fPIC\n\/usr\/local\/lib\/libgflags.a: could not read symbols: Bad value<\/p>\n\n<p>I found a 'workaround' for this problem at the libgflags and glog troubleshooting websites:\n<a href=\"https:\/\/code.google.com\/p\/google-glog\/issues\/detail?id=201\" rel=\"nofollow\">https:\/\/code.google.com\/p\/google-glog\/issues\/detail?id=201<\/a><\/p>\n\n<p>But I tried them and it doesn't seem to work. Am I missing something? Maybe I haven't uncommented a line in my original Makefile.config file? *I am installing caffe on my laptop with no CUDA or parallel computing for now.<\/p>\n","answers":[{"AnswerId":"29205434","CreationDate":"2015-03-23T07:56:24.920","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Try recompiling the gflags library with -fPIC compiler flag.<\/p>\n\n<p>Did the caffe work using gflags shared library instead of using the static one?<\/p>\n"},{"AnswerId":"42661505","CreationDate":"2017-03-08T01:44:02.743","ParentId":null,"OwnerUserId":"2509148","Title":null,"Body":"<p>Try to choose 'BUILD SHARED LIBS' option when using Cmake<\/p>\n"}]}
{"QuestionId":29206224,"AnswerCount":0,"Tags":"<theano><pymc>","CreationDate":"2015-03-23T06:56:08.620","AcceptedAnswerId":null,"OwnerUserId":4701788.0,"Title":"How do I use a FreeRV as a theano function input in pymc3","Body":"<p>I am trying to use  to model my system:  stochastic variables, of unknown distribution, which are related to some observed data  by the function . I have written the function  using theano, however I am not sure how to input the stochastics into my theano code. This is what I have tried:<\/p>\n\n<pre><\/pre>\n\n<p>This returns an error at the z_hat line:<\/p>\n\n<blockquote>\n  <p>ValueError: ('Bad input argument to theano function with name\n  \"scan_pm_v.py:60\"  at index 1(0-based)', 'setting an array element\n  with a sequence.')<\/p>\n<\/blockquote>\n\n<p>I can set the  to be a numpy object, which then gives the error<\/p>\n\n<blockquote>\n  <p>AttributeError: ('Bad input argument to theano function with name\n  \"scan_pm_v.py:60\"  at index 1(0-based)', \"'float' object has no\n  attribute 'dtype'\")<\/p>\n<\/blockquote>\n\n<p>at the same line. I've also tried making the input a theano tensor by adding the line:<\/p>\n\n<pre><\/pre>\n\n<p>Which then returns the error:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name\n  \"scan_pm_v.py:60\"  at index 1(0-based)', 'Expected an array-like\n  object, but found a Variable: maybe you are trying to call a function\n  on a (possibly shared) variable instead of a numeric array?')<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":29209935,"AnswerCount":1,"Tags":"<import><runtime><theano>","CreationDate":"2015-03-23T12:13:26.507","AcceptedAnswerId":null,"OwnerUserId":4690152.0,"Title":"How much time \"import theano\" takes to run?","Body":"<p>I am taking first baby steps with Theano on Windows for deep learning experiments, and I'm surprised how much it takes <em>just loading<\/em> the library.<\/p>\n\n<p>Here is the little test program :<\/p>\n\n<pre><\/pre>\n\n<p>results in my PyCharm console :<\/p>\n\n<pre><\/pre>\n\n<p>:-o<\/p>\n\n<p>Is it normal behaviour ? Should I reinstall everything ?  <\/p>\n\n<p>Here is my configuration :<\/p>\n\n<p>Windows 7. Python 2.7.8 in \"Anaconda\" package. CUDA 6.5. GPU : Nvidia Quadro K2000M.<\/p>\n\n<p>Here is the .theanorc file :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30508871","CreationDate":"2015-05-28T14:03:35.823","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It generally takes a long time to <code>import theano<\/code> when configured to use a GPU. On my machine it takes 0.567 seconds when running on CPU and 13.3 seconds when running on GPU.<\/p>\n\n<p>This may be one reason why to develop on CPU initially and switch to GPU once you've got it running right. The GPU overhead, including the increased Theano startup time, needs to be considered in determining whether running on a GPU is actually worthwhile.<\/p>\n"}]}
{"QuestionId":29213108,"AnswerCount":1,"Tags":"<speech><caffe>","CreationDate":"2015-03-23T14:40:35.877","AcceptedAnswerId":null,"OwnerUserId":1010037.0,"Title":"CNN-based 1-D signal classification using Caffe","Body":"<p>I am looking for an easy and straightforward example of 1-D signal classification (such as speech signal) based on CNN using Caffe.<\/p>\n\n<p>From the <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe<\/a> website, it is possible to follow some examples and tutorial which are Image classification tasks. Instead, I am looking for an example and tutorial on 1-D signals.<\/p>\n\n<p>Your answers are really appreciated. <\/p>\n","answers":[{"AnswerId":"31135832","CreationDate":"2015-06-30T10:30:22.607","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>Conceptually there's no meaningful difference between working with 1D data vs 2D data. You'd need a database where instead of 2D images you'll have 1D \"images\" of shape (channels: 1, height: 1, width: d), and make sure all kernels make use of <em>kernel_w<\/em> and <em>kernel_h<\/em> instead of <em>kernel_size<\/em> (which sets the kernel to a square shape).<\/p>\n\n<p>If you're looking for an example architecture you can use, there's this article you can follow for training a CNN on raw waveform speech data: <a href=\"http:\/\/www.cs.huji.ac.il\/~ydidh\/waveform.pdf\" rel=\"nofollow\">Speech Acoustic Modeling from Raw Multichannel Waveforms<\/a>.<\/p>\n\n<p>There's also <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1570\" rel=\"nofollow\">an open issue on Caffe's Github page<\/a> requesting an example for the speech domain, with more links to potential implementations you can look at.<\/p>\n"}]}
{"QuestionId":29225213,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-03-24T04:47:35.737","AcceptedAnswerId":null,"OwnerUserId":4705902.0,"Title":"In theano, making the matrix of slices from a vector","Body":"<p>I want to do the same thing in <a href=\"https:\/\/stackoverflow.com\/questions\/22159774\/get-matrix-of-vectors-from-a-vector\">get matrix of vectors from a vector<\/a> in theano.<\/p>\n\n<p>Maybe, it can be worked with scan().\nBut i dont know how scan() can be applied in this problem.<\/p>\n\n<p>Following is the code for context.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37475228","CreationDate":"2016-05-27T05:04:32.137","ParentId":null,"OwnerUserId":"4705902","Title":null,"Body":"<pre><code>import theano\nimport theano.tensor as T\n\nself.x = T.vector('x')\nself.i = T.imatrix('i') \n#indices tuple list. ex)[[0,3],[1,4]] means two slices (from 0 to 3 and from 1 to 4)\n\nresults, updates = theano.scan(lambda v:self.x[v[0]:v[1]], sequences=self.i)\nmake_slices = theano.function(inputs=[self.x,self.i], outputs=[results])    \n\nself.slices_list = make_slices(x,i)\n#'x' and 'i' are theano shared variables\n<\/code><\/pre>\n"}]}
{"QuestionId":29248639,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-03-25T06:12:18.017","AcceptedAnswerId":29257879.0,"OwnerUserId":3468673.0,"Title":"torch7 neural network training error","Body":"<p>I'm trying to implement a neural network example in torch7. my data is stored in a text file in this form [19 cols x 10000 rows] :<\/p>\n\n<pre><\/pre>\n\n<p>With labels in last column [100 labels].<\/p>\n\n<p>With this code :<\/p>\n\n<pre><\/pre>\n\n<p>I got this error message:<\/p>\n\n<pre><\/pre>\n\n<p>Can you please help me ?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"29257879","CreationDate":"2015-03-25T14:01:55.420","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>I got this error message [...] Can you please help me?<\/p>\n<\/blockquote>\n\n<p>Within your dataset <code>input<\/code> and <code>output<\/code> should be <code>Tensor<\/code>-s (here <code>input<\/code> is a plain Lua table which is why you obtain this error, i.e there is no <code>dim<\/code> method).<\/p>\n\n<p>To simplify data loading I recommend you to use a <a href=\"http:\/\/luarocks.org\/search?q=csv\" rel=\"nofollow\">csv parser<\/a>, e.g you can use <a href=\"https:\/\/github.com\/willkurt\/csv2tensor\" rel=\"nofollow\">csv2tensor<\/a> to load your data into a <code>Tensor<\/code>.<\/p>\n\n<p>First make sure to add a header (as first line) to your file like:<\/p>\n\n<pre><code>x001,x002,x003,x004,x005,x006,x007,x008,x009,x010,x011,x012,x013,x014,x015,x016,x017,x018,label\n<\/code><\/pre>\n\n<p>Then load your data as follow:<\/p>\n\n<pre><code>local csv2tensor = require 'csv2tensor'\n\nlocal inputs = csv2tensor.load(\"data.csv\", {exclude={\"label\"}})\nlocal labels = csv2tensor.load(\"data.csv\", {include={\"label\"}})\n\nlocal dataset = {}\n\nfor i=1,inputs:size(1) do\n  dataset[i] = {inputs[i], torch.Tensor{labels[i]}}\nend\n\ndataset.size = function(self)\n  return inputs:size(1)\nend\n<\/code><\/pre>\n\n<p>And use this dataset for training:<\/p>\n\n<pre><code>-- ...\ntrainer:train(dataset)\n<\/code><\/pre>\n"}]}
{"QuestionId":29255940,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-03-25T12:40:06.850","AcceptedAnswerId":29256541.0,"OwnerUserId":4695297.0,"Title":"Updating a variable that has been casted with theano.tensor.cast()","Body":"<p>I am trying to update a theano variable in a function, simplified like this:<\/p>\n\n<pre><\/pre>\n\n<p>My problem is that I get the error<\/p>\n\n<pre><\/pre>\n\n<p>The way I get this variable is through using the following (mostly copied from deeplearning.net tutorials) ( is initialized similarly):<\/p>\n\n<pre><\/pre>\n\n<p>which prints<\/p>\n\n<pre><\/pre>\n\n<p>that is, the variable is no longer \"shared\", explaining the error.\nThis makes sense, as I guess the variable is now simply just a casted view of the original shared floats. But how can I update a variable that is casted efficiently?<\/p>\n","answers":[{"AnswerId":"29256541","CreationDate":"2015-03-25T13:06:26.567","ParentId":null,"OwnerUserId":"4695297","Title":null,"Body":"<p>I solved this myself, and the answer was of course the obvious one.<\/p>\n\n<p>Instead of overriding the <code>a_variable<\/code> variable with the casted version, I kept the uncasted version:<\/p>\n\n<pre><code>a_variable_casted = T.cast(a_variable, 'int32')\n<\/code><\/pre>\n\n<p>Updates are now done on <code>a_variable<\/code>, while <code>a_variable_casted<\/code> is used to perform the computations <code>a_variable<\/code> was used for earlier.<\/p>\n\n<p>There might obviously be a more elegant way to do this, in which case I'd love to hear it!<\/p>\n"}]}
{"QuestionId":29274823,"AnswerCount":1,"Tags":"<android><android-ndk><caffe>","CreationDate":"2015-03-26T09:22:48.320","AcceptedAnswerId":null,"OwnerUserId":1485254.0,"Title":"Android caffe built demo shows error","Body":"<p>Being new to <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Android NDK Caffe<\/a>, I would like to use built version in my Android project. I tried to run <a href=\"https:\/\/github.com\/sh1r0\/caffe-android-demo\" rel=\"nofollow\">this built sample demo<\/a>, but while running, it showed the following:<\/p>\n\n<pre><\/pre>\n\n<p>(the app crashed)<\/p>\n","answers":[{"AnswerId":"37325904","CreationDate":"2016-05-19T14:13:14.090","ParentId":null,"OwnerUserId":"2995941","Title":null,"Body":"<p>I can see that the sigsev signal is thrown through android AsyncTask.\nThe problem could come from this function.<\/p>\n\n<pre><code>caffeMobile.predictImage(strings[0])[0]; \/\/line 160 of MainActivity\n<\/code><\/pre>\n\n<p>This signal comes from JNI and it is very difficult to know where is the problem unless you can debug natively (through ndk) the app. The caffe-sample is not configured to debug on native method.<\/p>\n\n<p>Try this issues to manage the error: <\/p>\n\n<blockquote>\n  <p>Ensure that your image path in this string[0] arrays are not empty. and exists.<\/p>\n  \n  <p>Ensure that the other caffeMobile functions are able to exec without\n  problems, for example:<\/p>\n<\/blockquote>\n\n<pre><code>  caffeMobile = new CaffeMobile();\n  caffeMobile.setNumThreads(4);\n  caffeMobile.loadModel(\"\/sdcard\/caffe_mobile\/bvlc_reference_caffenet\/deploy.prototxt\",    \"\/sdcard\/caffe_mobile\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel\");\n<\/code><\/pre>\n\n<p>If you are able to execute the other functions, probably your image path is not correct, check. <\/p>\n\n<p>If you are not able to execute loadModel or setNumThreads function, probably the apk is not loading libjni.so library correctly , or the jni bridge is not able to locate jni functions.<\/p>\n"}]}
{"QuestionId":29286838,"AnswerCount":3,"Tags":"<cuda><ubuntu-14.04><deep-learning><caffe>","CreationDate":"2015-03-26T19:04:52.660","AcceptedAnswerId":null,"OwnerUserId":4717939.0,"Title":"Caffe Installation Issue on Ubuntu 14.04","Body":"<p>I successfully installed caffe on my dual-boot laptop (GTX 860M, Windows 7 + Ubuntu 14.04.2). All the tests were successfully passed. When I <strong>restarted<\/strong>, however, the ubuntu got stuck on the opening screen (the one with ubuntu logo and <strong>five red dots<\/strong>). Don't know what to do with it. <\/p>\n\n<p>Has anyone run into the same issue before? I reckon something is wrong with graphic card driver booting. I installed newest CUDA 7 Toolkit with nvidia drivers built inside. Since all tests were passed before I restarted, it seems that the driver would work once successfully booted. <\/p>\n\n<p>the stuck screen is like this: <a href=\"http:\/\/i.stack.imgur.com\/pRtEF.jpg\" rel=\"nofollow\">http:\/\/i.stack.imgur.com\/pRtEF.jpg<\/a><\/p>\n","answers":[{"AnswerId":"30028621","CreationDate":"2015-05-04T11:09:11.810","ParentId":null,"OwnerUserId":"4841248","Title":null,"Body":"<p>I had a similar issue when trying to install Caffe on my system. The steps below worked for me, but it has at least one known issue (documented below).<\/p>\n\n<p>I'm not sure what precisely caused this problem, but it surely has something to do with the Nvidia Driver and Cuda Toolkit installation and is <strong>not caused by Caffe<\/strong>.<\/p>\n\n<p>After completing the steps below, I've been able to successfully install Caffe on my system with the following tutorials and guides:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow\">Official Install Guide<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM\" rel=\"nofollow\">Github Install Guide<\/a><\/li>\n<\/ul>\n\n<hr>\n\n<h2>Update<\/h2>\n\n<p>Recently, I had the exact same problem trying to <strong>make Cuda 7.5 work on Ubuntu 14.04;<\/strong> this approach also solved that problem. Specs:<\/p>\n\n<ul>\n<li>CPU: Intel Core i7-4700MQ (4x 2.40 GHz with Hyperthreading)<\/li>\n<li>GPU: NVidia GT 940M<\/li>\n<li>RAM: 8 GB<\/li>\n<li>HDD: 52.7 GB (of which 6.7 GB used after installation)<\/li>\n<\/ul>\n\n<hr>\n\n<h1>INSTALL NVIDIA DRIVER AND CUDA ON UBUNTU 14.04<\/h1>\n\n<p>Source: <a href=\"http:\/\/ubuntuforums.org\/showthread.php?t=2246526\" rel=\"nofollow\">ubuntuforums.org\/showthread.php?t=2246526<\/a><\/p>\n\n<p><strong>!! Known Issues !!<\/strong><\/p>\n\n<ul>\n<li>After the system has been suspended (or hibernated, not confirmed), all applications using the Nvidia Driver and Cuda 6.5 Toolkit will freeze. When this happens, the command <code>sudo shutdown -r now<\/code> will print the reboot message but nothing will happen.<\/li>\n<\/ul>\n\n<p><strong>Executed and tested on a fresh 64-bit Ubuntu 14.04 install with the following hardware specifications:<\/strong><\/p>\n\n<ul>\n<li>CPU: Intel Core i5-2410m (2x 2.30 GHz with Hyperthreading)<\/li>\n<li>GPU: NVidia GT 540M<\/li>\n<li>RAM: 6 GB<\/li>\n<li>HDD: 52.7 GB (of which 8.6 GB used after installation)<\/li>\n<\/ul>\n\n<p><strong>The following command was executed before installation:<\/strong><\/p>\n\n<pre><code>sudo apt-get -y build-essential vim git llvm clang\n<\/code><\/pre>\n\n<p><strong>The following steps resulted in a stable system with the latest Nvidia Driver and Cuda 6.5 Toolkit installed:<\/strong><\/p>\n\n<ol>\n<li>Remove all traces of previous\/legacy Nvidia Drivers and Cuda Toolkits or perform a fresh Ubuntu 14.04 install.<\/li>\n<li>Download the latest Nvidia Driver .run file for Ubuntu 14.04 and your system specifications to the <code>~\/Downloads<\/code> directory.\ne.g.: <code>NVIDIA-Linux-x86_64-346.35.run<\/code><\/li>\n<li>Download the latest Cuda 6.5 Toolkit .run file for Ubuntu 14.04 and your system specifications to the <code>~\/Downloads<\/code> directory.\ne.g.: <code>cuda_6.5.14_linux_64.run<\/code><\/li>\n<li><p>Blacklist the 'nouveau' Driver by appending the following lines to <code>\/etc\/modprobe.d\/blacklist.conf<\/code> (nouveau is a free open-source driver for Nvidia cards, it is the default for Ubuntu 14.04):<\/p>\n\n<p><code>blacklist nouveau<\/code><br>\n<code>options nouveau modeset=0<\/code><\/p><\/li>\n<li><p>Reboot the system, do <strong>NOT<\/strong> log in but drop to the terminal with CTRL+ALT+F1<\/p><\/li>\n<li><p>Kill lightdm (replace 'lightdm' with your own Display Manager if you have changed it, lightdm is the default for Ubuntu 14.04):<\/p>\n\n<p><code>sudo service lightdm stop<\/code><\/p><\/li>\n<\/ol>\n\n<p><em><strong>The next step is critical, make sure to check twice before continuing!<\/strong><\/em><\/p>\n\n<ol start=\"7\">\n<li><p>Run the Nvidia Driver installer with the <code>--no-opengl-files<\/code> option (the option prevents OpenGL files from being overwritten; without this option, Unity would not function properly and the screen would freeze after login):<\/p>\n\n<p><code>sudo chmod +x ~\/Downloads\/NVIDIA-Linux-x68_64-346.35.run<\/code><br>\n<code>sudo ~\/Downloads\/NVIDIA-Linux-x68_64-346.35.run --no-opengl-files<\/code><\/p><\/li>\n<li><p>Accept the EULA and acknowledge all further warnings but <strong>deny<\/strong> to install anything extra.<\/p><\/li>\n<li><p>Reboot and login to the desktop, verify with the 'Additional Drivers' (System Settings > Software &amp; Updates > Additional Drivers) utility that the manually installed driver is in use.<\/p><\/li>\n<li><p>Open a terminal and install the Cuda 6.5 Toolkit:<\/p>\n\n<p><code>sudo chmod +x ~\/Downloads\/cuda_6.5.14_linux_64.run<\/code><br>\n<code>sudo ~\/Downloads\/cuda_6.5.14_linux_64.run<\/code><\/p><\/li>\n<li><p>Accept the EULA, do <strong>NOT<\/strong> install the driver, install the Toolkit and the Examples (if you want to), leave all default directories in place.<\/p><\/li>\n<li><p>Add the Cuda 6.5 Toolkit environment variables by appending the following lines to <code>~\/.bashrc<\/code>: <\/p>\n\n<p><code># For 32-bit systems, append these:<\/code><br>\n<code>export PATH=$PATH:\/usr\/local\/cuda-6.5\/bin<\/code><br>\n<code>export LD_LIBRARY_PATH=\/usr\/local\/cuda-6.5\/lib<\/code>  <\/p>\n\n<p><code># For 64-bit systems, append these:<\/code><br>\n<code>export PATH=$PATH:\/usr\/local\/cuda-6.5\/bin<\/code><br>\n<code>export LD_LIBRARY_PATH=\/usr\/local\/cuda-6.5\/lib64<\/code>  <\/p><\/li>\n<\/ol>\n\n<p>The Nvidia Driver and Cuda 6.5 Toolkit should now be correctly installed.<\/p>\n\n<p><strong>Optional:<\/strong> confirm your Nvidia Driver and Cuda 6.5 Toolkit installation.<\/p>\n\n<ol start=\"13\">\n<li><p>Confirm the Nvidia Driver installation by running the following command:<\/p>\n\n<p><code>nvidia-smi<\/code>  <\/p><\/li>\n<li><p>Confirm the Cuda Compiler installation by running the following command:<\/p>\n\n<p><code>nvcc -V<\/code>  <\/p><\/li>\n<li><p>Confirm everything works by building and running the optionally installed Cuda Examples: (build-essential is required to use 'make')<\/p>\n\n<p><code>sudo apt-get install -y build-essential<\/code><br>\n<code>cd ~\/NVIDIA_CUDA-6.5_SAMPLES\/1_Utilities\/deviceQuery<\/code><br>\n<code>make<\/code><br>\n<code>.\/deviceQuery<\/code><br>\n<code>cd ~\/NVIDIA_CUDA-6.5_SAMPLES\/1_Utilities\/bandwidthTest<\/code><br>\n<code>make<\/code><br>\n<code>.\/bandwidthTest<\/code>  <\/p><\/li>\n<\/ol>\n"},{"AnswerId":"29286892","CreationDate":"2015-03-26T19:07:54.530","ParentId":null,"OwnerUserId":"987599","Title":null,"Body":"<p>This problem is not related to caffe.<\/p>\n\n<p>The problem is that the nVidia driver that is installed from the ubuntu software center does not support your card.<\/p>\n\n<p>Uninstall any nvidia package (<code>sudo apt-get purge nvidia-*<\/code>) and install the latest driver version from the nvidia website.<\/p>\n"},{"AnswerId":"35447115","CreationDate":"2016-02-17T03:04:58.767","ParentId":null,"OwnerUserId":"5938204","Title":null,"Body":"<p>I recommend you to change the cuda 7.5 ubuntu 15.04 version. I try it on the ubuntu 14.04, it solves this problem. And when I install cuda 7.5 ubuntu 14.04 version on ubuntu 14.04 I countered the exactly problem.<\/p>\n"}]}
{"QuestionId":29328031,"AnswerCount":1,"Tags":"<python><c++><machine-learning><deep-learning><caffe>","CreationDate":"2015-03-29T10:30:47.060","AcceptedAnswerId":30236484.0,"OwnerUserId":3966933.0,"Title":"Issues while interfacing caffe with c++ or python","Body":"<p>What I have read about the tutorials is that you create your data then write the model using protobuf and then you write the solver file. Finally you train the model and you get your generated file. All this is done though command line. Now there are two questions<\/p>\n\n<p>1) Suppose I have the generated model now how do I load a new image not in the test folder and perform a forward pass. Should it be done though command line or from some language(c++, python) ?<\/p>\n\n<p>2) I guess above was one way of doing it. What is the best way to train the classifier (command line train\/ or though coding) and how to use the generated model file(after training) in your code. <\/p>\n\n<p>I want to interface caffe with my code but I am not able to find a short tutorial which will give me step by step on any database say mnist and the model doesn't need to be as complicated as LeNet but a simple Fully connected layer will also do. But can anyone tell me how to just write a simple code using C++ or python and train any dataset from scratch. <\/p>\n\n<p>A sample C++\/python code for training a classifier and using it to predict new data using caffe would also be appreciated.<\/p>\n","answers":[{"AnswerId":"30236484","CreationDate":"2015-05-14T11:41:42.417","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Training is best done using the command line. See <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">this tutorial<\/a>.<\/p>\n\n<p>Once you trained a model and you have a <code>myModel.caffemodel<\/code> file (a binary file storing the wieghts of the different layers) and a <code>deploy.prototxt<\/code> file (a text file describing your net), you can use python interface to classify images.<\/p>\n\n<p>You can run python script <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/classify.py\" rel=\"nofollow\"><code>classify.py<\/code><\/a> to classify image(s) from command line. This script wraps around <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py\" rel=\"nofollow\"><code>classifier.py<\/code><\/a> - a python object that holds a trained net and allows you to perform forward passes in python.<\/p>\n"}]}
{"QuestionId":29334961,"AnswerCount":1,"Tags":"<lua><neural-network><deep-learning><torch>","CreationDate":"2015-03-29T21:19:08.813","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch, error in StochasticGradient training","Body":"<p>I'm trying to create an autoencoder neural network in Torch, but I got an error when I try to run the training phase.<\/p>\n\n<p>Here's my input data in data.csv:<\/p>\n\n<pre><\/pre>\n\n<p>Here's my code<\/p>\n\n<pre><\/pre>\n\n<p>And here's my error:<\/p>\n\n<pre><\/pre>\n\n<p>Do you have any idea about why I'm gettin' this errors?<\/p>\n","answers":[{"AnswerId":"29335566","CreationDate":"2015-03-29T22:16:44.063","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>dataset[i] = {dnaseIsignals[i-1], dnaseIsignals[i-1]} \n<\/code><\/pre>\n\n<p>This is wrong. What you want to give as input to the neural network is a torch.Tensor, but in this case, you are initializing dataset's elements to be tables. Convert your input to a tensor, and you should be good.<\/p>\n"}]}
{"QuestionId":29338016,"AnswerCount":1,"Tags":"<python-2.7><ubuntu-12.04><theano>","CreationDate":"2015-03-30T03:35:06.437","AcceptedAnswerId":null,"OwnerUserId":1620746.0,"Title":"import theano gets Illegal instruction","Body":"<p>after install the bleeding-edge version and even uninstall Theano, I'm still getting \"Illegal instruction\" from \"import theano\", i'm on ubuntu 12.04 precise<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31951181","CreationDate":"2015-08-11T20:17:25.967","ParentId":null,"OwnerUserId":"1115577","Title":null,"Body":"<p>This can happen when you move from one computer to another with the same <code>.local<\/code> directory. The following (from <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/T0521GHRC1w\" rel=\"noreferrer\">here<\/a>) worked for me:<\/p>\n\n<p>First delete <code>~\/.theano<\/code> which stores some theano compiled files. Then reinstall theano via <code>pip uninstall theano; pip install --user theano<\/code>. It also fixes the gensim install for some reason (which shows the same error upon importing). Perhaps gensim imports theano when it can?<\/p>\n"}]}
{"QuestionId":29354357,"AnswerCount":5,"Tags":"<api><http><lua><torch>","CreationDate":"2015-03-30T19:14:08.427","AcceptedAnswerId":29355228.0,"OwnerUserId":563762.0,"Title":"HTTP Server Using Lua\/ Torch7","Body":"<p>I'm starting to learn Torch7 to get into the machine learning\/ deep learning field and I'm finding it fascinating (and very complicated haha). My main concern, however, is if I can turn this learning into an application - mainly can I turn my Torch7 Lua scripts into a server that an app can use to perform machine learning calculations? And if it's possible, how?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"35467931","CreationDate":"2016-02-17T21:17:53.940","ParentId":null,"OwnerUserId":"3932212","Title":null,"Body":"<p>You can use <a href=\"https:\/\/github.com\/benglard\/waffle\" rel=\"noreferrer\">waffle<\/a>. Here's the hello world example on it's page:<\/p>\n\n<pre><code>local app = require('waffle')\n\napp.get('\/', function(req, res)\n   res.send('Hello World!')\nend)\n\napp.listen()\n<\/code><\/pre>\n\n<p>lets say your algorithm is a simple face detector. The input is an image and the output is the face detections in some json format. You can do the following:<\/p>\n\n<pre><code>local app = require('waffle')\nrequire 'graphicsmagick'\nrequire 'MyAlgorithm'\n\napp.post('\/', function(req, res)\n  local img, detections, outputJson\n  img = req.form.image_file:toImage()\n  detections = MyAlgorithm.detect(img:double())\n  outputJson = {}\n  if (detections ~= nil) then\n    outputJson.faceInPicture = true\n    outputJson.faceDetections = detections\n  else\n    outputJson.faceInPicture = false\n    outputJson.faceDetections = nil\n  end\n  res.json(outputJson)\nend)\n\napp.listen()\n<\/code><\/pre>\n\n<p>That way your algorithm can be used as an independent service.<\/p>\n"},{"AnswerId":"29355228","CreationDate":"2015-03-30T20:01:59.523","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>You should look at Torch as a library (even though you may be accessing it as a standalone executable). That library can be used from some Lua code that is accessible through HTTP. The Lua code may be running inside <a href=\"http:\/\/openresty.org\/\" rel=\"nofollow\">OpenResty<\/a>, which would take care of all HTTP interactions, and you get the same performance as OpenResty can be configured to use LuaJIT.<\/p>\n\n<p>Another option is to use HTTP processing based on luasocket and copas libraries (for examples, <a href=\"http:\/\/keplerproject.github.io\/xavante\/\" rel=\"nofollow\">Xavante<\/a>) or use one of the options listed on <a href=\"http:\/\/lua-users.org\/wiki\/LuaWebserver\" rel=\"nofollow\">LuaWebserver<\/a> page.<\/p>\n"},{"AnswerId":"29356295","CreationDate":"2015-03-30T21:06:49.887","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>You can also use the <a href=\"https:\/\/github.com\/clementfarabet\/async\" rel=\"nofollow\">async<\/a> package that we've tested with torch.<\/p>\n"},{"AnswerId":"30985705","CreationDate":"2015-06-22T16:59:16.197","ParentId":null,"OwnerUserId":"2053898","Title":null,"Body":"<p>Try llserver - minimalistic Lua server. Runs as single coroutine, serves dynamic content via callback function: <a href=\"https:\/\/github.com\/ncp1402\/llserver\" rel=\"nofollow\">https:\/\/github.com\/ncp1402\/llserver<\/a>\nYou can perform other tasks\/calculations in additional coroutines.<\/p>\n"},{"AnswerId":"38837950","CreationDate":"2016-08-08T20:27:42.667","ParentId":null,"OwnerUserId":"213123","Title":null,"Body":"<p>Both <a href=\"https:\/\/github.com\/clementfarabet\/async\" rel=\"nofollow\">async<\/a> and <a href=\"https:\/\/github.com\/benglard\/waffle\" rel=\"nofollow\">waffle<\/a> are great options. Another option is to use <a href=\"http:\/\/zeromq.org\/bindings:lua\" rel=\"nofollow\">ZeroMQ<\/a> + <a href=\"https:\/\/github.com\/djungelorm\/protobuf-lua\" rel=\"nofollow\">Protocol Buffers<\/a>. Whatever your preferred web development environment is, you can send requests to Torch using ZeroMQ in asynchronous manner, possibly serializing messages with protocol buffers, then process each request in Torch and return the result back. <\/p>\n\n<p>This way I managed to get much higher throughput, than <a href=\"https:\/\/github.com\/benglard\/waffle\" rel=\"nofollow\">waffle<\/a>'s 20K fib test.<\/p>\n"}]}
{"QuestionId":29365370,"AnswerCount":3,"Tags":"<python><gradient><theano>","CreationDate":"2015-03-31T09:37:56.733","AcceptedAnswerId":null,"OwnerUserId":1215787.0,"Title":"How to code adagrad in python theano","Body":"<p>To simplify the problem, say when a dimension (or a feature) is already updated n times, the next time I see the feature, I want to set the learning rate to be 1\/n.<\/p>\n\n<p>I came up with these codes:<\/p>\n\n<pre><\/pre>\n\n<p>Theano does not give any error, but the training result give Nan sometimes. Does anybody know how to correct this please ?<\/p>\n\n<p>Thank you for your help<\/p>\n\n<p>PS: I doubt it is the operations in sparse space which creates problems. So I tried to replace * by theano.sparse.mul. This gave the some results as I mentioned before<\/p>\n","answers":[{"AnswerId":"29643793","CreationDate":"2015-04-15T07:12:23.017","ParentId":null,"OwnerUserId":"828368","Title":null,"Body":"<p>Perhaps you can utilize the following <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\" rel=\"noreferrer\">example for implementation of <strong>adadelta<\/strong><\/a>, and use it to derive your own. Please update if you succeeded :-)<\/p>\n"},{"AnswerId":"39567087","CreationDate":"2016-09-19T06:58:27.297","ParentId":null,"OwnerUserId":"217802","Title":null,"Body":"<p>I find <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/updates.py#L385\" rel=\"nofollow\">this implementation from Lasagne<\/a> very concise and readable. You can use it pretty much as it is:<\/p>\n\n<pre><code>for param, grad in zip(params, grads):\n    value = param.get_value(borrow=True)\n    accu = theano.shared(np.zeros(value.shape, dtype=value.dtype),\n                         broadcastable=param.broadcastable)\n    accu_new = accu + grad ** 2\n    updates[accu] = accu_new\n    updates[param] = param - (learning_rate * grad \/\n                              T.sqrt(accu_new + epsilon))\n<\/code><\/pre>\n"},{"AnswerId":"36851284","CreationDate":"2016-04-25T21:20:16.417","ParentId":null,"OwnerUserId":"4295315","Title":null,"Body":"<p>I was looking for the same thing and ended up implementing it myself in the style of the resource zuuz already pointed out. So maybe this helps anyone looking for help here.<\/p>\n\n<pre><code>def adagrad(lr, tparams, grads, inp, cost):\n    # stores the current grads\n    gshared = [theano.shared(np.zeros_like(p.get_value(),\n                                           dtype=theano.config.floatX),\n                             name='%s_grad' % k)\n               for k, p in tparams.iteritems()]\n    grads_updates = zip(gshared, grads)\n    # stores the sum of all grads squared\n    hist_gshared = [theano.shared(np.zeros_like(p.get_value(),\n                                                dtype=theano.config.floatX),\n                                  name='%s_grad' % k)\n                    for k, p in tparams.iteritems()]\n    rgrads_updates = [(rg, rg + T.sqr(g)) for rg, g in zip(hist_gshared, grads)]\n\n    # calculate cost and store grads\n    f_grad_shared = theano.function(inp, cost,\n                                    updates=grads_updates + rgrads_updates,\n                                    on_unused_input='ignore')\n\n    # apply actual update with the initial learning rate lr\n    n = 1e-6\n    updates = [(p, p - (lr\/(T.sqrt(rg) + n))*g)\n               for p, g, rg in zip(tparams.values(), gshared, hist_gshared)]\n\n    f_update = theano.function([lr], [], updates=updates, on_unused_input='ignore')\n\n    return f_grad_shared, f_update\n<\/code><\/pre>\n"}]}
{"QuestionId":29380867,"AnswerCount":3,"Tags":"<theano>","CreationDate":"2015-04-01T00:29:06.957","AcceptedAnswerId":null,"OwnerUserId":4620420.0,"Title":"Call a function from scan in Theano","Body":"<p>I need to execute a theano function a number of times via scan in order to sum-up a cost function and use it in a gradient computation.  I'm familiar with the deep-learning tutorials that do this but my data slicing and some other complications means I need to do this a little different.\nBelow is a much simplified version of what I'm trying to do..<\/p>\n\n<pre><\/pre>\n\n<p>In the scan function, the call to test_fn(curr) is giving the error... \nExpected an array-like object, but found a Variable: maybe you are trying to call a function on a (possibly shared) variable instead of a numeric array?')<\/p>\n\n<p>Even if I pass in an array of values instead of putting the T.arrange(2) in place, I still get the same error.  Is there a reason you can't call a function from scan?  <\/p>\n\n<p>In general I'm wondering if there is a way to call a function like this with a series of indexes so that the output can feed into a T.grad() computation (not shown). <\/p>\n","answers":[{"AnswerId":"29426595","CreationDate":"2015-04-03T05:16:09.377","ParentId":null,"OwnerUserId":"2990135","Title":null,"Body":"<p>Don't make two different <code>theano.functions<\/code>.<\/p>\n\n<p>A <code>theano.function<\/code> takes a symbolic relationship, optimizes it, and compiles it.  What you are doing here is asking <code>theano.scan<\/code> (and thus <code>out_fn<\/code>) to consider a compiled function as a symbolic relationship.  Whether you could <em>technically<\/em> get that to work I'm not sure, but it goes against the idea of Theano.<\/p>\n\n<p>Since I don't know what your cost function does here I can't give an exact example, but here's a quick example which does work and should be similar enough to what I <em>think<\/em> you're trying to do.<\/p>\n\n<pre><code>x = theano.shared(np.asarray([7.1,2.2,3.4], dtype = np.float32))\n\nv = T.vector(\"v\")\ndef fv(v):\n    res,_ = theano.scan(lambda x: x ** 2, v)\n    return T.sum(res)\n\ndef f(i):\n    return fv(x[i:i+2])\n\nouts,_ = theano.scan(\n    f, \n    T.arange(2)\n    )\n\nfn = theano.function(\n    [],\n    outs,\n    )\n\nfn()\n<\/code><\/pre>\n"},{"AnswerId":"40788133","CreationDate":"2016-11-24T13:49:43.883","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<h3>Solution<\/h3>\n\n<p>The standard way is <code>OpFromGraph<\/code> (as of 0.8.2)<\/p>\n\n<pre><code>import theano as th\nimport theano.tensor as T\n\nx = T.scalar('x')\ny = T.scalar('y')\nz = x+y\n# unlike theano.function, must use list for outputs\nop_add = th.OpFromGraph([x,y], [z])\n\ndef my_add(x_, y_):\n    return op_add(x_, y_)[0]\n\nx_list = T.vector('x_li')\nx_sum = th.scan(op_add, sequences=[x_list], outputs_info=[T.constant(0.)])\nfn_sum = th.function([x_list], x_sum)\nfn([1., 2., 3., 4.]) # 10.\n<\/code><\/pre>\n\n<h3>What it does?<\/h3>\n\n<p><code>OpFromGraph<\/code> compiles a function defined from a graph, then pack it into a new Op. Just like defining functions in imperative programming languages.<\/p>\n\n<h3>Pros\/Cons<\/h3>\n\n<ul>\n<li>[+] It can be handy in tricky models.<\/li>\n<li>[+] It saves compilation time. You can compile a commonly used part of a big model into <code>OpFromGraph<\/code>, then directly use it in a bigger model. The final graph will have less nodes than direct implementation.<\/li>\n<li>[-] It cause worse runtime performance. Calling a function has overhead, also the compiler is unable to do a global optimization due to its compiled nature.<\/li>\n<li>[-] It's premature and still under development. The documentation for it is incomplete. Currently does not support <code>updates<\/code> and <code>givens<\/code> as in <code>theano.function<\/code>.<\/li>\n<\/ul>\n\n<h3>Notes<\/h3>\n\n<p>In most cases, you should be defining python functions\/classes to build model. Only use <code>OpFromGraph<\/code> if no workaround is possible or you want to save compilation time.<\/p>\n"},{"AnswerId":"29446969","CreationDate":"2015-04-04T13:25:33.887","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>After some investigation I agree that calling a function from a function is not correct.  The challenge with the code is that following the basic design of the deep-learning tutorials, the first layer of the net has a symbolic variable defined as it's input and the output is propagated up to higher layers until a final cost is computed from the top layer.  The tutorials uses code something like...<\/p>\n\n<pre><code>class layer1(object):\n   def __init__(self):\n      self.x = T.matrix()\n      self.output = activation(T.dot(self.x,self.W) + self.b)\n<\/code><\/pre>\n\n<p>For me the tensor variable (layer1.self.x) needs to change every time scan takes a step to have a new slice of data. The \"givens\" statement in a function does that, but since calling a compiled theano function from inside a \"scan\" doesn't work there are two other solutions I was able to find...<\/p>\n\n<p>1 - Rework the network so that its cost function is based on a series of function calls instead of a propagated variable.  This is technically simple but requires a bit of re-coding to get things organized properly in a multi-layer network.<\/p>\n\n<p>2 - Use theano.clone inside of scan.  That code looks something like...<\/p>\n\n<pre><code>def step(curr):\n    y_in = y[curr]\n    replaces = {tn.layer1.x : x[curr:curr+1]}\n    fn = theano.clone(tn.cost(y_in), replace=replaces)\n    return fn\nouts,_ = theano.scan(step, sequences=[T.arange(batch_start,batch_end)])\n<\/code><\/pre>\n\n<p>Both methods return the same results and appear execute at the same speed.<\/p>\n"}]}
{"QuestionId":29398109,"AnswerCount":2,"Tags":"<python><numpy><theano><deep-learning>","CreationDate":"2015-04-01T18:26:56.050","AcceptedAnswerId":null,"OwnerUserId":3140172.0,"Title":"Using Random Numbers in Theano","Body":"<p>I am a theano newbie. <br><\/p>\n\n<p>Can someone please explain the following code?<\/p>\n\n<pre><\/pre>\n\n<p>Q1. How does nearly_zeros() affect rv_u's generator?<br>\nQ2. Why ?<br> <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29405457","CreationDate":"2015-04-02T05:18:11.813","ParentId":null,"OwnerUserId":"2990135","Title":null,"Body":"<p><strong>Q1<\/strong><\/p>\n\n<p>Looks like only 1 \"value\" (ie a 2x2 matrix) is being generated by rv_u's generator.  You can see this if you use <code>theano.printing.debugprint<\/code> to print the graph of the function.  For reference this is what I got:<\/p>\n\n<pre><code>&gt;&gt;&gt; theano.printing.debugprint(nearly_zeros)\nElemwise{Composite{((i0 + i0) - (i1 * i0))}}[(0, 0)] [@A] ''   1\n |RandomFunction{uniform}.1 [@B] ''   0\n | |&lt;RandomStateType&gt; [@C]\n | |TensorConstant{(2,) of 2} [@D]\n | |TensorConstant{0.0} [@E]\n | |TensorConstant{1.0} [@F]\n |TensorConstant{(1, 1) of 2.0} [@G]\nRandomFunction{uniform}.0 [@B] ''   0\n<\/code><\/pre>\n\n<p>This means the two functions (nearly_zero and f) both grab only 1 value from rv_u, hence why <\/p>\n\n<pre><code>v3 == v1\n<\/code><\/pre>\n\n<p><strong>Q2<\/strong><\/p>\n\n<p>Theano is largely a symbolic math package.  You define relationships between symbolic variables and Theano figures out how to evaluate those relationships.<\/p>\n\n<p>So you could consider rv_u to be a variable which represents a single 2x2 matrix, rather than a rng which generates a new 2x2 matrix each time it's \"called.\"  Given that view, Theano would only need to call the underlying rng once to get the value for the variable rv_u.<\/p>\n"},{"AnswerId":"35075137","CreationDate":"2016-01-29T01:03:17.553","ParentId":null,"OwnerUserId":"5170062","Title":null,"Body":"<h2>Q1<\/h2>\n\n<p>When <code>nearly_zeros()<\/code> is called, it uses the current rng state of the random stream to return a new <code>rv_u<\/code> and increments the state just as it does for the function <code>f<\/code>.  The easiest way to see what is happening to <code>rv_u<\/code> is to add it as an output to the <code>nearly_zeros()<\/code> function.  Although <code>rv_u<\/code> is present three times, the value of <code>rv_u<\/code> is sampled from the random stream only once, hence <code>nearly_zero<\/code> will indeed be nearly zero (up to floating-point quantization error.)<\/p>\n\n<pre><code>nearly_zeros = function([], [rv_u, rv_u + rv_u - 2 * rv_u])\nf() # The return value will be different from the rv_u in the list above\n<\/code><\/pre>\n\n<p>If you do not intend for it to increment, you can specify the <code>no_default_updates=True<\/code> as was done for <code>g<\/code>.<\/p>\n\n<pre><code>nearly_zeros = function([], [rv_u, rv_u + rv_u - 2 * rv_u], \n     no_default_updates=True)\nf() # The return value will be equal to the rv_u in the list above\n<\/code><\/pre>\n\n<h2>Q2<\/h2>\n\n<p>The \"why\" here was a little ambiguous, so I present two possibilities with the answers.<\/p>\n\n<p><h3>Q2(a)<\/h3>Why does using <code>nearly_zeros()<\/code> impact <code>rv_u<\/code> in this way?<\/p>\n\n<p><code>rv_u<\/code> is a uniform srng object which uses a shared variable that is incremented each time it is requested, unless specifically told otherwise (by means of the <code>no_defalut_updates<\/code> parameter.)  This is true whether <code>f<\/code> or <code>nearly_zero<\/code> happens to be the function that requests the value of <code>rv_u<\/code>.<\/p>\n\n<p><h3>Q2(b)<\/h3>Why is the following true:<\/p>\n\n<pre><code>v2 != v1\nv3 == v1\n<\/code><\/pre>\n\n<p>This is because your code saves the state after defining <code>nearly_zeros<\/code> but before it is called.  When the state is set, the first value that is returned is the value used by <code>nearly_zeros<\/code> (which is equal to <code>v2<\/code>.)  The next value requested will be equal to the value of <code>v3<\/code>.  The following is a nearly identical copy of your code, and it should prove illustrative:<\/p>\n\n<pre><code>from theano.tensor.shared_randomstreams import RandomStreams\nfrom theano import function\nsrng = RandomStreams(seed=234)\nrv_u = srng.uniform((2,2))\nrv_n = srng.normal((2,2))\nf = function([], rv_u)\ng = function([], rv_n, no_default_updates=True)    #Not updating rv_n.rng\nnearly_zeros = function([], [rv_u, rv_u + rv_u - 2 * rv_u]) # Give rv_u, too\n\nstate_after_v0 = rv_u.rng.get_value().get_state()\nnearly_zeros()       # this affects rv_u's generator\nv1 = f()\nrng = rv_u.rng.get_value(borrow=True)\nrng.set_state(state_after_v0)\nrv_u.rng.set_value(rng, borrow=True)\nv2 = f()             # v2 != v1, but equal to rv_u used by nearly_zeros\nv3 = f()             # v3 == v1\n<\/code><\/pre>\n\n<p><i>Aside: Although this thread is a bit old, hopefully someone will find these answers helpful.<\/i><\/p>\n"}]}
{"QuestionId":29399067,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-04-01T19:20:41.137","AcceptedAnswerId":29400464.0,"OwnerUserId":563762.0,"Title":"How to get a prediction using Torch7","Body":"<p>I'm still familiarizing myself with Torch and so far so good. I have however hit a dead end that I'm not sure how to get around: how can I get Torch7 (or more specifically the dp library) to evaluate a single input and return the predicted output?<\/p>\n\n<p>Here's my setup (basically the dp demo):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29400464","CreationDate":"2015-04-01T20:45:20.397","ParentId":null,"OwnerUserId":"49985","Title":null,"Body":"<p>You have two options. <\/p>\n\n<p>One. Use the encapsulated <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/module.md#module\" rel=\"nofollow\">nn.Module<\/a> to forward your <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor\" rel=\"nofollow\">torch.Tensor<\/a>:<\/p>\n\n<pre><code>mlp = model:toModule(datasource:trainSet():sub(1,2))\nmlp:float()\ninput = torch.FloatTensor(1, 1, 32, 32) -- replace this with your input\noutput = mlp:forward(input)\n<\/code><\/pre>\n\n<p>Two. Encapsulate your torch.Tensor into a <a href=\"http:\/\/dp.readthedocs.org\/en\/latest\/view\/index.html#dp.ImageView\" rel=\"nofollow\">dp.ImageView<\/a> and forward that through your <a href=\"http:\/\/dp.readthedocs.org\/en\/latest\/model\/index.html#dp.Model\" rel=\"nofollow\">dp.Model<\/a> :<\/p>\n\n<pre><code>input = torch.FloatTensor(1, 1, 32, 32) -- replace with your input\ninputView = dp.ImageView('bchw', input)\noutputView = mlp:forward(inputView, dp.Carry{nSample=1})\noutput = outputView:forward('b')\n<\/code><\/pre>\n"}]}
{"QuestionId":29412658,"AnswerCount":0,"Tags":"<machine-learning><neural-network><gradient-descent><torch><conv-neural-network>","CreationDate":"2015-04-02T12:11:25.753","AcceptedAnswerId":null,"OwnerUserId":1084163.0,"Title":"torch - LookupTable and gradient update","Body":"<p>I am trying to implement a neural network with multiple layers. I am trying to understand if what I have done is correct and if not, how do I debug this. The way I do it is, I define my neural network in the following manner (I initialise the lookuptable layer with some prior learned embeddings):<\/p>\n\n<pre><\/pre>\n\n<p>Now, to train the network, I loop through every training example and for every example I call gradUpdate() which has this code (this is straight from the examples):<\/p>\n\n<pre><\/pre>\n\n<p>The findGrad function is just an implementation of WARP Loss which returns the gradient wrt output. I am wondering if this is all I need? I assume this will backpropagate and update the parameters of all the layers. To check this, I trained this network and saved the model. Then I loaded the model and did:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I checked vector[1] and lookuptable.weight[1] and they were the same. I can't understand why did the weights in the lookup table layer not get updated? What am I missing here?<\/p>\n\n<p>Looking forward to your replies!<\/p>\n","answers":[]}
{"QuestionId":29434671,"AnswerCount":1,"Tags":"<python><neural-network><computer-vision><deep-learning><caffe>","CreationDate":"2015-04-03T15:02:32.747","AcceptedAnswerId":null,"OwnerUserId":2539909.0,"Title":"Caffe network getting very low loss but very bad accuracy in testing","Body":"<p>I'm somewhat new to caffe, and I'm getting some strange behavior.  I'm trying to use fine tuning on the bvlc_reference_caffenet to accomplish an OCR task.<\/p>\n\n<p>I've taken their pretrained net, changed the last FC layer to the number of output classes that I have, and retrained.  After a few thousand iterations I'm getting loss rates of ~.001, and an accuracy over 90 percent when the network tests.  That said, when I try to run my network on data by myself, I get awful results, not exceeding 7 or 8 percent.<\/p>\n\n<p>The code I'm using to run the net is: <\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts on why this performance might be so poor?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>PS: Some additional information which may or not be of use.  When classifying as shown below, the classifier really seems to favor certain classes.  Even though I have a 101 class problem, it seems to only assign a max of 15 different classes<\/p>\n\n<p>PPS: I'm also fairly certain I'm not overfitting.  I've been testing this along the way with snapshots and they all exhibit the same poor results.<\/p>\n","answers":[{"AnswerId":"29776104","CreationDate":"2015-04-21T15:06:11.770","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your code for testing the model you posted seem to miss some components:<\/p>\n\n<ol>\n<li>It seems like you did not subtract the image's mean.<\/li>\n<li>You did not swap channels from RGB to BGR.<\/li>\n<li>You did not scale the inputs to [0..255] range.<\/li>\n<\/ol>\n\n<p>Looking at similar instances of <code>caffe.Classifier<\/code> you may see something like:<\/p>\n\n<pre><code>net = caffe.Classifier('bvlc_reference_caffenet\/deploy.prototxt',\n                       'bvlc_reference_caffenet\/caffenet_train_iter_28000.caffemodel', \n                       mean = NP.load( 'ilsvrc_2012_mean.npy' ),\n                       input_scale=1.0, raw_scale=255,\n                       channel_swap=(2,1,0),\n                       image_dims=(227, 227, 1))\n<\/code><\/pre>\n\n<p>It is <strong>crucial<\/strong> to have the same input transformation in test as in training.<\/p>\n"}]}
{"QuestionId":29434729,"AnswerCount":1,"Tags":"<lua><machine-learning><resize><torch>","CreationDate":"2015-04-03T15:05:25.537","AcceptedAnswerId":29436069.0,"OwnerUserId":470433.0,"Title":"Torch Resize Tensor","Body":"<p>How do I resize a Tensor in Torch? Methods documented in <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#resizing\" rel=\"noreferrer\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#resizing<\/a> do not seem to work.<\/p>\n\n<pre><\/pre>\n\n<p>Why doesn't this approach work? <\/p>\n","answers":[{"AnswerId":"29436069","CreationDate":"2015-04-03T16:32:40.753","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>You need to do:<\/p>\n\n<pre><code>images:resize(...)\n<\/code><\/pre>\n\n<p>What you did:<\/p>\n\n<pre><code>images.resize(...)\n<\/code><\/pre>\n\n<p>images.resize does not pass the current tensor as the first argument.<\/p>\n\n<p><code>images:resize(...)<\/code> is equivalent to <code>images.resize(images, ...)<\/code><\/p>\n"}]}
{"QuestionId":29436556,"AnswerCount":1,"Tags":"<windows><theano>","CreationDate":"2015-04-03T17:04:00.827","AcceptedAnswerId":29449430.0,"OwnerUserId":147530.0,"Title":"How to modify .theanorc so that nvcc uses the -m64 flag during compilation?","Body":"<p>I followed the steps at <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a> to install theano but running into problems. One of them is that by default using the  settings on <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a>, on my machine  tries to compile theano in 32bit mode as I see following when I try to import theano on python shell (note the  below):<\/p>\n\n<pre><\/pre>\n\n<p>It then runs into problems as it cannot find  which indeed does not exist on my machine under 32 bit libs stored under  (does this file exist on another users system in the 32bit folder?). I have  stored under  and would therefore like to compile in 64bit mode. To do that I changed  to:<\/p>\n\n<pre><\/pre>\n\n<p>but this does not give desired effect.  is still trying to compile in 32bit mode:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know the proper syntax how to modify  so that  uses the <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-compiler-driver-nvcc\/#axzz3WB624BPV\" rel=\"nofollow\">-m64 flag<\/a> during compilation?<\/p>\n\n<blockquote>\n  <p>--machine {32|64} -m   Specify 32-bit vs. 64-bit architecture.<\/p>\n  \n  <p>Allowed values for this option: 32, 64.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"29449430","CreationDate":"2015-04-04T17:34:12.860","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The problem is that your PYTHON is a 32bit python. We do not support mixing python, g++ and nvcc bit size. THis mean they all must be 32 bit or 64 bit.<\/p>\n\n<p>Make sure to install all of them as 64 bit.<\/p>\n\n<p>If you want to try to support this mixed case, check in the theano\/sandbox\/cuda\/nvcc_compiler.py, it is there we do the compilation. Here we add the -m32 flags:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/nvcc_compiler.py#L324\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/nvcc_compiler.py#L324<\/a><\/p>\n\n<p>If you make this work, a PR with the requested change to Theano would be welcome.<\/p>\n"}]}
{"QuestionId":29438171,"AnswerCount":1,"Tags":"<windows><theano>","CreationDate":"2015-04-03T19:03:18.313","AcceptedAnswerId":29438391.0,"OwnerUserId":147530.0,"Title":"How to specify a windows directory that has spaces in it in .theanorc","Body":"<p>I need to use the  flag to specify directory that contains  to . If I specify it like this:<\/p>\n\n<pre><\/pre>\n\n<p>I get error while importing theano in python that complains about spaces. What is the proper way to specify this directory in ? Note that I do not want to edit my .<\/p>\n","answers":[{"AnswerId":"29438391","CreationDate":"2015-04-03T19:21:38.480","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<p>Try the following, if you haven't yet, it seems to work for other people:<\/p>\n\n<pre><code>[nvcc]\ncompiler_bindir=C:\\Program Files (x86)\\Microsoft Visual Studio 9.0\\VC\\bin\n<\/code><\/pre>\n\n<p>See also <a href=\"https:\/\/stackoverflow.com\/a\/26073714\">this<\/a> long Stackoverflow guide about installing Theano on Windows.<\/p>\n"}]}
{"QuestionId":29456838,"AnswerCount":1,"Tags":"<image-processing><similarity><caffe>","CreationDate":"2015-04-05T11:36:24.087","AcceptedAnswerId":null,"OwnerUserId":3966933.0,"Title":"Finding similar Images","Body":"<p>I want to find images similar to another image. So after researching i found two methods first was two represent the image by its attributes like <\/p>\n\n<p><\/p>\n\n<p>but the limitation of this method is that I will not be able to get an exhaustive dataset with all the features marked.<\/p>\n\n<p>The second approach I found was to extract features and do feature mapping. \nSo I decided to use deep convolution neural networks with caffe so that by using any of the exsisting models I could learn the features and then perform feature matching or some other operation.  I just wanted to take a general advice what can be the other methods which are good and worth a try. And since I am just starting out with caffe so can anyone give a general guideline how to approach the problem with caffe? \nThanks in advance<\/p>\n\n<p>I looked at phash just was curious that it will find the images which are same like there are minor intensity variations and some other variation wiill it also work to give the same type(semantically) like for a tshirt with blue and red stripes will it give black and white stripe as similar and would it consider things like the length of shirt, collar style etc<\/p>\n","answers":[{"AnswerId":"31133097","CreationDate":"2015-06-30T08:19:58.280","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>It's true, that it's been empirically shown, that the euclidean distance between the features extracted using ConvNets is closer for images of the same class, while farther for images of different classes - but it's important to understand what kind of similarity you're looking for.<\/p>\n\n<p>One can define many types of similarity measures, and the type of features you use (in the case of ConvNets, the type of data it was trained on) affects the kind of similar images you'll get. For instance, maybe given an image of a dog, you want to find other pictures of dogs but not specifically that exact dog, alternatively, maybe you have a picture of a church and you want to find another image of the exact same church but from a different angle - these are two very different problems, with different methods you can use to solve them.<\/p>\n\n<p>One particular kind of convolutional neural networks you can look at, are Siamese Network, which are built to learn similarities between two images, given a dataset of pairs of images with the labels same\/not_same. You can look for implementation in Caffe for this method <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>A different method, is to take a ConvNet trained on ImageNet data (<a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo\" rel=\"nofollow\">see here for options<\/a>), and use the python\/matlab interface to classify images, and then extract the second to last layer, and use that as the representation for that image. Now you can just take the euclidean distance of those representations and this would be your similarity measure.<\/p>\n\n<p>Unrelated to Caffe, you can also use \"old school\" methods of feature matching, included in open source libraries like OpenCV (<a href=\"http:\/\/docs.opencv.org\/doc\/tutorials\/features2d\/feature_flann_matcher\/feature_flann_matcher.html\" rel=\"nofollow\">an example tutorial of such method<\/a>). <\/p>\n"}]}
{"QuestionId":29458330,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-04-05T14:28:50.427","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"Size mismatch in function addmv in Torch7","Body":"<p>I'm working on a small Torch7\/ Lua script to create and train a neural network, but I'm running into errors. Any ideas?<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the error:<\/p>\n\n<pre><\/pre>\n\n<p>And here's a sample of my data:<\/p>\n\n<pre><\/pre>\n\n<p>Basically my aim is to create a deep neural network linking the frequency of words used in a sentence and tie it to the user rating it as either \"positive\", \"negative\" or \"neutral\" (my outputs, which are binary). Please also let me know if my thinking is correct on this.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"29459592","CreationDate":"2015-04-05T16:38:36.680","ParentId":null,"OwnerUserId":"563762","Title":null,"Body":"<p>Found the problem!<\/p>\n\n<p>The issue was that I was giving the wrong sizes when creating the network. I was passing in \"inputs:size(1)\" instead of \"inputs:size(2)\". Here's the fix<\/p>\n\n<pre><code>mlp:add(nn.Linear(inputs:size(2), HUs))\nmlp:add(nn.Tanh())\nmlp:add(nn.Linear(HUs, outputs:size(2)))\n<\/code><\/pre>\n\n<p>Feel like I'm slowly starting to get the hang of Lua\/ Torch! Score<\/p>\n"}]}
{"QuestionId":29474646,"AnswerCount":1,"Tags":"<file><lua><load><torch>","CreationDate":"2015-04-06T15:50:25.323","AcceptedAnswerId":29474911.0,"OwnerUserId":1082019.0,"Title":"How to load a table from file in Torch \/ Lua?","Body":"<p>Very simple operation. I have a file that contains a table made of N rows and 6 columns, and I would like to load it in a table in my Torch \/ Lua script.<\/p>\n\n<p>The data file looks this way:<\/p>\n\n<pre><\/pre>\n\n<p>And I would like to load it in a table, where for example  contains  and so on.<\/p>\n\n<p>How could I do it? Thanks<\/p>\n","answers":[{"AnswerId":"29474911","CreationDate":"2015-04-06T16:07:30.727","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>You can use <code>string.match<\/code> to parse individual line into a table and use <code>io.lines<\/code> to iterate over lines in the file:<\/p>\n\n<pre><code>-- script.lua\nlocal t, patt = {}, (\"(%w+)%s+\"):rep(5)..\"(%w+)\"\nfor line in io.lines() do\n  if not line:find(\"^chromNameA\") then\n    table.insert(t, {line:match(patt)})\n  end\nend\nprint(#t, t[1][1], t[1][6]) -- prints `5 chr22 16678717`\n\n-- file.txt\nchromNameA  startA  endA    chromNameB  startB  endB\nchr22   16867980    16868130    chr22   16669675    16678717\nchr22   16867980    16868130    chr22   16685348    16701095\nchr22   16867980    16868130    chr22   16723869    16739035\nchr22   16867980    16868130    chr22   16748016    16750787\nchr22   16867980    16868130    chr22   16750788    16755877\n\n-- execution: lua script.lua &lt;file.txt\n<\/code><\/pre>\n\n<p>You can then start the script as <code>lua script.lua &lt;file.txt<\/code> and it should produce a table with the structure you want.<\/p>\n"}]}
{"QuestionId":29481458,"AnswerCount":1,"Tags":"<cuda><gpu><caffe>","CreationDate":"2015-04-06T23:46:57.310","AcceptedAnswerId":29481562.0,"OwnerUserId":2297751.0,"Title":"A single program appear on two GPU card","Body":"<p>I have multiple GPU cards(NO.0, NO.1 ...), and every time I run a <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow noreferrer\">caffe<\/a> process on NO.1 or 2 ... (except 0) card, it will use up 73MiB on the NO.0 card.<\/p>\n\n<p>For example, in the fig below, process 11899 will use 73MiB on NO.0 card but it actually run on NO.1 card. <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/lEqay.png\" alt=\"multi_gpus\"><\/p>\n\n<p>Why? Can I disable this feature?<\/p>\n","answers":[{"AnswerId":"29481562","CreationDate":"2015-04-06T23:58:37.810","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>The CUDA driver is like an operating system.  It will reserve memory for various purposes when it is active.  Certain features, such as <a href=\"https:\/\/docs.nvidia.com\/cuda\/cuda-c-programming-guide\/index.html#um-unified-memory-programming-hd\" rel=\"nofollow\">managed memory<\/a>, may cause substantial side-effect allocations to occur (although I don't think this is the case with Caffe).  And its even possible that the application itself may be doing some explicit allocations on those devices, for some reason.<\/p>\n\n<p>If you want to prevent this, one option is to use the <code>CUDA_VISIBLE_DEVICES<\/code> <a href=\"https:\/\/docs.nvidia.com\/cuda\/cuda-c-programming-guide\/index.html#env-vars\" rel=\"nofollow\">environment variable<\/a> when you launch your process.<\/p>\n\n<p>For example, if you want to prevent CUDA from doing anything with card \"0\", you could do something like this (on linux):<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES=\"1,2\" .\/my_application ...\n<\/code><\/pre>\n\n<p>Note that the enumeration used above (the CUDA enumeration) is the same enumeration that would be reported by the <code>deviceQuery<\/code> sample app, but not necessarily the same enumeration reported by <code>nvidia-smi<\/code> (the NVML enumeration).  You may need to experiment or else run <code>deviceQuery<\/code> to determine which GPUs you want to use, and which you want to exclude.<\/p>\n\n<p>Also note that using this option actually affects the devices that are visible to an application, and will cause a re-ordering of device enumeration (the device that was previously \"1\" will appear to be enumerated as device \"0\", for example).  So if your application is multi-GPU aware, and you are selecting specific devices for use, you may need to change the specific devices you (or the application) are selecting, when you use this environment variable.<\/p>\n"}]}
{"QuestionId":29490289,"AnswerCount":0,"Tags":"<lua><caffe><torch>","CreationDate":"2015-04-07T11:16:57.920","AcceptedAnswerId":null,"OwnerUserId":470433.0,"Title":"Torch Caffe Lua - How to get Flickr Style Example to work?","Body":"<p>Using the <a href=\"https:\/\/github.com\/szagoruyko\/torch-caffe-binding\" rel=\"nofollow\">Torch Caffe Binding<\/a>. I want to make a prediction on the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow\">Flickr Style example<\/a>. I've got the trained model and the code below. How do I have to alter the code to make it work?<\/p>\n\n<pre><\/pre>\n\n<p>To output now is a FloatTensor [torch.FloatTensor of dimension 20x1x1] with 20 times NaN values.<\/p>\n","answers":[]}
{"QuestionId":29492848,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><regression><caffe><unsupervised-learning>","CreationDate":"2015-04-07T13:29:15.030","AcceptedAnswerId":null,"OwnerUserId":1913583.0,"Title":"Add regression layer to caffe","Body":"<p>I have implemented a smile detection system based on deep learning. The bottom layer is the output of the system and has 10 output according to the amount of the person's smile.<br>\nI want to convert these ten output with a numeric output in the range of 1 to 10 with a regression layer.<br>\nHow can I do this in caffe?<br>\nThanks<\/p>\n","answers":[{"AnswerId":"32623945","CreationDate":"2015-09-17T06:37:44.003","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In order to convert the 10 outputs you have to a single one you need an <code>\"InnerProduct\"<\/code> layer with 10 inputs and a single output. To train this layer you also need to add a loss layer suitable for regression on top of the 10 output layer you already have.<br>\nSuch loss layers can be, e.g., <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/euclidean_loss_layer.cpp\" rel=\"nofollow\">Euclidean loss layer<\/a> or Ross Girshick's <a href=\"https:\/\/github.com\/rbgirshick\/caffe-fast-rcnn\/blob\/fast-rcnn\/src\/caffe\/layers\/smooth_L1_loss_layer.cpp\" rel=\"nofollow\">smooth L1 loss layer<\/a>.<\/p>\n"}]}
{"QuestionId":29514223,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2015-04-08T12:10:02.487","AcceptedAnswerId":null,"OwnerUserId":1452257.0,"Title":"Dropout layer ignored - \"Warning: had one or more problems upgrading V1LayerParameter\"","Body":"<p>I'm trying to add a dropout layer based on the Imagenet example (see code below). However, it seems to be ignored it's not printed as being part of the network when I train the model and I get the warning message below. I have the newest version of caffe installed. What should I do to include it properly?<\/p>\n\n<p><strong>Dropout layer:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Training log:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32201237","CreationDate":"2015-08-25T10:12:15.157","ParentId":null,"OwnerUserId":"5263750","Title":null,"Body":"<p>I had the exact same problem and stumbled upon this question:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/30253520\/what-does-attempting-to-upgrade-input-file-specified-using-deprecated-transform\">What does &#39;Attempting to upgrade input file specified using deprecated transformation parameters&#39; mean?<\/a><\/p>\n\n<p>As far as I know, in caffe's attempt to upgrade your layer architecture,\ncaffe has already found some layers that are written in the new protobuf syntax. Thus, it skips upgrading your old syntax and therefore ignores the Dropout layer.\nBut I am not 100% sure about this, as it's just my assumption, when taking a short look into the caffe code.<\/p>\n\n<p>For me the solution was to change it as follows:<\/p>\n\n<pre><code>layers {\n  name: \"drop4\"\n  type: DROPOUT\n  bottom: \"fc4\"\n  top: \"fc4\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\n<\/code><\/pre>\n\n<p>Hope it works =)<\/p>\n"}]}
{"QuestionId":29529959,"AnswerCount":2,"Tags":"<caffe><lmdb>","CreationDate":"2015-04-09T04:44:56.380","AcceptedAnswerId":29552948.0,"OwnerUserId":465509.0,"Title":"Caffe: Understanding expected lmdb datastructure for blobs","Body":"<p>I'm trying to understand how data is interpreted in Caffe.\nFor that I've taken a look at the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\">Minst Tutorial<\/a>\nLooking at the input data definition:<\/p>\n\n<pre><\/pre>\n\n<p>I've now looked at the mnist_train_lmdb and taken one of the entries (shown in hex):<\/p>\n\n<pre><\/pre>\n\n<p>(I've added the line breaks here to be able to see the '7' digit.)<\/p>\n\n<p>Now my question is, <strong>where this format is described?<\/strong> Or put differently where is defined that the first 36 bytes are some sort of header and the last 8 bytes have some label correspondence?<\/p>\n\n<p><strong>How would I go about constructing my own data?<\/strong>\nNeither <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\">Blob Tutorial<\/a> nor <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\">Layers Definition<\/a> give much away about required formats. <em>My intention is not to use image data, but time series<\/em><\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"29552948","CreationDate":"2015-04-10T03:40:49.967","ParentId":null,"OwnerUserId":"465509","Title":null,"Body":"<p>I realized that protocol buffers must come into play here. So I tried to deserialize it against some of the types defined in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\">caffe.proto<\/a>.<\/p>\n\n<p><strong>Datum<\/strong> seems to be the perfect fit:<\/p>\n\n<pre><code>{Caffe.Datum}\n    Channels: 1\n    Data: {byte[784]}\n    Encoded: false\n    FloatData: Count = 0\n    Height: 28\n    Label: 7\n    Width: 28\n<\/code><\/pre>\n\n<p>So the answer is simply: <strong>It's a serialized representation of a 'Datum' typed instance as defined per <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\">caffe.proto<\/a><\/strong><\/p>\n\n<p><em>Btw. since english is not my native language I had to first realize that \"Datum\" is a singular form of \"data\"<\/em><\/p>\n\n<p>When it comes to using your own data, it's structured as follows:<\/p>\n\n<blockquote>\n  <p>The conventional blob dimensions for data are number N x channel\n  K x height H x width W. Blob memory is row-major in layout so the last\n  \/ rightmost dimension changes fastest. For example, the value at index\n  (n, k, h, w) is physically located at index ((n * K + k) * H + h) * W\n  + w.<\/p>\n<\/blockquote>\n\n<p>See <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\">Blobs, Layers, and Nets: anatomy of a Caffe model<\/a> for reference<\/p>\n"},{"AnswerId":"30793520","CreationDate":"2015-06-12T00:22:35.943","ParentId":null,"OwnerUserId":"4561745","Title":null,"Body":"<p>I can try to answer your second question. Since Caffe only takes data in a bunch of selected formats like lmdb, hdf5 etc., it is best to convert (or generate - in case of synthetic data) your data to these formats. Following links can help you in this. If you have trouble with <code>import hdf5<\/code>in Python, then you may refer to <a href=\"https:\/\/stackoverflow.com\/questions\/30769048\/error-in-creating-lmdb-database-file-in-python-for-caffe?answertab=votes#tab-top\">this<\/a> page.<\/p>\n\n<p><a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow noreferrer\">Creating an LMDB file in Python<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/5466971\/fastest-way-to-write-hdf5-file-with-python\">Writing an HDF5 file in Python<\/a> <\/p>\n\n<p><a href=\"http:\/\/download.nexusformat.org\/sphinx\/examples\/h5py\/index.html\" rel=\"nofollow noreferrer\">HDF5 more examples<\/a><\/p>\n"}]}
{"QuestionId":29540592,"AnswerCount":2,"Tags":"<cuda><neural-network><theano><deep-learning><conv-neural-network>","CreationDate":"2015-04-09T13:57:04.340","AcceptedAnswerId":null,"OwnerUserId":3778898.0,"Title":"Why does my dropout function in Theano slow down convolution greatly?","Body":"<p>I am learning Theano. I wrote a simple dropout function as below:<\/p>\n\n<pre><\/pre>\n\n<p>When I apply this function to the input of the first two convolutional layers, the average time spent on each image increases from 0.5ms to about 2.5ms! Does anyone know what could be the reason for such drastic slow down?<\/p>\n\n<p>I am using a GTX 980 card with cuDNN installed.<\/p>\n","answers":[{"AnswerId":"30299367","CreationDate":"2015-05-18T09:21:21.447","ParentId":null,"OwnerUserId":"3778898","Title":null,"Body":"<p>RandomStream only works on the CPU. Therefore <code>mask<\/code> has to be copied from CPU to GPU every time drop is called, which is the reason for the drastic slow down. To avoid this, I now use <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#other-implementations\" rel=\"noreferrer\">random stream implementation which works on GPU<\/a>. <\/p>\n"},{"AnswerId":"33594732","CreationDate":"2015-11-08T13:56:58.903","ParentId":null,"OwnerUserId":"5538861","Title":null,"Body":"<p>It seems like a similar problem as mine (<a href=\"https:\/\/stackoverflow.com\/questions\/33592778\/lasagne-dropoutlayer-does-not-utilize-gpu-efficiently\">Lasagne dropoutlayer does not utilize GPU efficiently<\/a>). \nHave you checked that you code sets <code>cuda_enabled = True<\/code> somewhere? otherwise you can manually set it in line 93 of <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/__init__.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/<strong>init<\/strong>.py<\/a>.\nI know this is not a elegant solution but it solved my problem for now. :)<\/p>\n"}]}
{"QuestionId":29564360,"AnswerCount":2,"Tags":"<lua><machine-learning><deep-learning><torch>","CreationDate":"2015-04-10T14:44:48.083","AcceptedAnswerId":29564910.0,"OwnerUserId":1367788.0,"Title":"Bug encountered When running Google's Deep Q Network Code","Body":"<p>Google's Deep Q Network for Atari Games is here.<\/p>\n\n<p><a href=\"https:\/\/github.com\/rahular\/deepmind-dqn\" rel=\"nofollow\">https:\/\/github.com\/rahular\/deepmind-dqn<\/a><\/p>\n\n<p>When I run it with GPU setting<\/p>\n\n<pre><\/pre>\n\n<p>I had this error<\/p>\n\n\n\n<pre class=\"lang-html prettyprint-override\"><\/pre>\n\n<p>The code that caused this issue is in this file <a href=\"https:\/\/github.com\/rahular\/deepmind-dqn\/blob\/master\/dqn\/convnet.lua\" rel=\"nofollow\">https:\/\/github.com\/rahular\/deepmind-dqn\/blob\/master\/dqn\/convnet.lua<\/a><\/p>\n\n<p>and it is in this function <\/p>\n\n\n\n<pre class=\"lang-html prettyprint-override\"><\/pre>\n\n<p>The net:add(convLayer( is 22th line.<\/p>\n\n<p>I used gpu setting so it seems <\/p>\n\n<pre><\/pre>\n\n<p>caused convLayer to be nil.<\/p>\n\n<p>Does anyone know why nn.SpatialConvolutionCUDA returns a nil ?<\/p>\n","answers":[{"AnswerId":"29622297","CreationDate":"2015-04-14T08:14:59.757","ParentId":null,"OwnerUserId":"1367788","Title":null,"Body":"<p>Found the solution.<\/p>\n\n<p>using this github branch<\/p>\n\n<p><a href=\"https:\/\/github.com\/soumith\/deepmind-atari\" rel=\"nofollow\">https:\/\/github.com\/soumith\/deepmind-atari<\/a><\/p>\n\n<p>After cloning this branch, then install cutorch and cunn using luarocks.<\/p>\n\n<p>Now you can run the code.<\/p>\n"},{"AnswerId":"29564910","CreationDate":"2015-04-10T15:11:21.013","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Did the code originally come with GPU support, or did you add it yourself?<\/p>\n\n<p>You should replaced the depreceated layers, i.e. replace:<\/p>\n\n<pre><code>net:add(nn.Transpose({1,2},{2,3},{3,4}))\nconvLayer = nn.SpatialConvolutionCUDA\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>convLayer = nn.SpatialConvolution\n<\/code><\/pre>\n\n<p>Check the documentation for the layers.<\/p>\n\n<p>Edit: Use <a href=\"https:\/\/github.com\/soumith\/deepmind-atari\" rel=\"nofollow\">this branch<\/a>, I fixed it for GPU support.<\/p>\n"}]}
{"QuestionId":29579280,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-04-11T14:57:28.623","AcceptedAnswerId":29579908.0,"OwnerUserId":2161754.0,"Title":"is TensorSharedVariable in theano initilized twice in function?","Body":"<p>In theano, once the sharedvarialbe is initialized in one function, it will never be initialized again even if the function is accessed repeatedly, am I right?<\/p>\n\n<pre><\/pre>\n\n<p>In the code above, the exp_sqr_grads variable and the exp_sqr_ups variable will not be initialized with zeros again when the sgd_updates_adadelta function is called the second time? <\/p>\n","answers":[{"AnswerId":"29579908","CreationDate":"2015-04-11T15:59:16.350","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Shared variables are not static, if that is what you mean. My understanding of your code:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nglobal_list = []\n\ndef f():\n\n    a = np.zeros((4, 5), dtype=theano.config.floatX)\n    b = theano.shared(a)\n    global_list.append(b)\n<\/code><\/pre>\n\n<p>Copy and paste this into an IPython and then try:<\/p>\n\n<pre><code>f()\nf()\n\nprint global_list\n<\/code><\/pre>\n\n<p>The list will contain two items. They are not the same object:<\/p>\n\n<pre><code>In [9]: global_list[0] is global_list[1]\nOut[9]: False\n<\/code><\/pre>\n\n<p>And they don't refer to the same memory: Do<\/p>\n\n<pre><code>global_list[0].set_value(np.arange(20).reshape(4, 5).astype(theano.config.floatX))\n<\/code><\/pre>\n\n<p>Then<\/p>\n\n<pre><code>In [20]: global_list[0].get_value()\nOut[20]: \narray([[  0.,   1.,   2.,   3.,   4.],\n       [  5.,   6.,   7.,   8.,   9.],\n       [ 10.,  11.,  12.,  13.,  14.],\n       [ 15.,  16.,  17.,  18.,  19.]])\n\nIn [21]: global_list[1].get_value()\nOut[21]: \narray([[ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])\n<\/code><\/pre>\n\n<p>Having established that initializing shared variables several times leads to different variables, here is how to update a shared variable using a function. We re-use the established shared variables:<\/p>\n\n<pre><code>s = global_list[1]\nx = T.scalar(dtype=theano.config.floatX)\ng = theano.function([x], [s], updates=[(s, T.inc_subtensor(s[0, 0], x))])\n<\/code><\/pre>\n\n<p><code>g<\/code> now increments the top left value of <code>s<\/code> by <code>x<\/code> at every call:<\/p>\n\n<pre><code>In [7]: s.get_value()\nOut[7]: \narray([[ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])\n\nIn [8]: g(1)\nOut[8]: \n[array([[ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])]\n\nIn [9]: s.get_value()\nOut[9]: \narray([[ 1.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])\n\nIn [10]: g(10)\nOut[10]: \n[array([[ 1.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.]])]\n\nIn [11]: s.get_value()\nOut[11]: \narray([[ 11.,   0.,   0.,   0.,   0.],\n       [  0.,   0.,   0.,   0.,   0.],\n       [  0.,   0.,   0.,   0.,   0.],\n       [  0.,   0.,   0.,   0.,   0.]])\n<\/code><\/pre>\n"}]}
{"QuestionId":29584422,"AnswerCount":2,"Tags":"<macos><lua><ipython><canopy><torch>","CreationDate":"2015-04-11T23:58:09.753","AcceptedAnswerId":null,"OwnerUserId":4611375.0,"Title":"Installing Torch7. iPython installation error (mac)","Body":"<p>I'm trying to install Torch7 on my mac, however the installation halts at this point:<\/p>\n\n<pre><\/pre>\n\n<p>Not sure what it means. Further above I received the following output<\/p>\n\n<pre><\/pre>\n\n<p>However iPython is installed as I can confirm:<\/p>\n\n<pre><\/pre>\n\n<p><em>Attempting to locate the .bashrc file<\/em><\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n","answers":[{"AnswerId":"29595574","CreationDate":"2015-04-12T22:50:44.853","ParentId":null,"OwnerUserId":"726839","Title":null,"Body":"<p>First, IPython may be installed but not seen by the install process.<\/p>\n\n<p>If you enter <code>which ipython<\/code> at a shell prompt it will tell you where it is installed. Then <code>echo $PATH<\/code> will display your PATH variable which should contain the directory that contains IPython. If it isn't then you will have to edit the PATH statement in your .bashrc file to add it.<\/p>\n\n<p>Second, <code>.bashrc<\/code> is a file that is run by the shell when it starts up and should be in your home directory so enter <code>cd<\/code> at a shell prompt and you will be there. Then use a text editor such as Text Edit to edit it. In your case you don't have a <code>.bashrc<\/code> file, instead some things are being set in a file called <code>.profile<\/code>. You should probably check the contents of that.<\/p>\n\n<p>Finally, I wouldn't run any of these commands from within IPython. Only run them (and the Torch install process) from the shell.<\/p>\n\n<p>Further, I notice you have a file with the name \"anaconda\" in it. Have you installed 'Anaconda'?<\/p>\n"},{"AnswerId":"34121829","CreationDate":"2015-12-06T19:44:32.507","ParentId":null,"OwnerUserId":"5647398","Title":null,"Body":"<p>i had a similar problem and solved it, maybe it could help others.<\/p>\n\n<p>Here was the end of the second installation script, and the command \"th\" wasn't working: <\/p>\n\n<pre><code>Not updating your shell profile.\nYou might want to\nadd the following lines to your shell profile:\n\n. \/Users\/myusername\/torch\/install\/bin\/torch-activate\n<\/code><\/pre>\n\n<p>This article explains how your shell profile is organized: <a href=\"https:\/\/serverfault.com\/questions\/110065\/what-profile-is-my-current-shell-using\">https:\/\/serverfault.com\/questions\/110065\/what-profile-is-my-current-shell-using<\/a><\/p>\n\n<p>I realized in my user folder \/Users\/myusername\/ i had a \".bash_profile\" file, i pasted the line \". \/Users\/myusername\/torch\/install\/bin\/torch-activate\" inside but didn't work (command \"th\" no recognized in terminal).<\/p>\n\n<p><strong>So in the same \/Users\/myusername\/ folder i created a \".profile\" file<\/strong> and pasted the line \". \/Users\/myusername\/torch\/install\/bin\/torch-activate\" inside.<\/p>\n\n<p>Then the command \"th\" works fine ;)<\/p>\n"}]}
{"QuestionId":29585037,"AnswerCount":1,"Tags":"<c++><opencv><visual-studio-2013><cuda><caffe>","CreationDate":"2015-04-12T01:43:13.620","AcceptedAnswerId":null,"OwnerUserId":4778446.0,"Title":"Caffe Compilation on Windows - Compiles Successfully But 0xc000007b Error","Body":"<p>I have followed the steps at <a href=\"https:\/\/initialneil.wordpress.com\/2015\/01\/11\/build-caffe-in-windows-with-visual-studio-2013-cuda-6-5-opencv-2-4-9\/\" rel=\"nofollow\">this URL<\/a> to compile Caffe for windows. The compilation succeeds but I am unable to run the generated EXE file. Also, when I downloaded the Git branch, there was already a caffe.exe file listed there. When I tried to run the precompiled file, I also got this error: \"The application was unable to start correctly (0xc000007b). Click OK to close the application\". This is the same error I get on my compiled binary.<\/p>\n\n<p>Please help me. I am running Windows 7 x64. I suspect the problem could be creeping in somewhere that maybe since I have like 32 bit MinGW or something maybe it is trying to use the 32 bit libraries?<\/p>\n\n<p>Right now, I have my configuration set to build x64 bit. I feel like one of the problems could be that maybe the CUDA is trying to build 32 or something? I just don't know what's causing this.. Even stranger, why am I not able to run the precompiled caffe.exe that I found when I downloaded this.. (I get the exact same error, that makes me feel like it isn't my compilation process.. there is something else going on).<\/p>\n\n<p>Thank you for your help<\/p>\n\n<p>OK - I ran the dependency walker. I identified the following issues:<\/p>\n\n<p><strong>Error: At least one module has an unresolved import due to a missing export function in an implicitly dependent module.<\/strong><\/p>\n\n<p><strong>Error: Modules with different CPU types were found.<\/strong><\/p>\n\n<p>LIBGCC_S_DW2-1.DLL<\/p>\n\n<p>LIBGFORTRAN-3.DLL<\/p>\n\n<p>are x86 but the rest of them are x64. Where can I get the 64 bit DLLs from?<\/p>\n\n<p>Also, \nAPI-MS-WIN-APPMODEL-RUNTIME-L1-1-0.DLL<\/p>\n\n<p>API-MS-WIN-CORE-WINRT-ERROR-L1-1-0.DLL<\/p>\n\n<p>API-MS-WIN-CORE-WINRT-L1-1-0.DLL<\/p>\n\n<p>API-MS-WIN-CORE-WINRT-ROBUFFER-L1-1-0.DLL<\/p>\n\n<p>API-MS-WIN-CORE-WINRT-STRING-L1-1-0.DLL<\/p>\n\n<p>API-MS-WIN-SHCORE-SCALING-L1-1-1.DLL<\/p>\n\n<p>DCOMP.DLL<\/p>\n\n<p>IESHIMS.DLL<\/p>\n\n<p>Are listed as being not found (the system cannot find the file specified).<\/p>\n","answers":[{"AnswerId":"29585079","CreationDate":"2015-04-12T01:49:00.903","ParentId":null,"OwnerUserId":"3325075","Title":null,"Body":"<p>This is a dependency error for Windows applications--in others words, you're probably missing a DLL file. Use <a href=\"http:\/\/www.dependencywalker.com\/\" rel=\"nofollow\">dependency walker<\/a> to help find out which DLL files you're missing.<\/p>\n"}]}
{"QuestionId":29593373,"AnswerCount":1,"Tags":"<macos><opencv><cmake><anaconda><caffe>","CreationDate":"2015-04-12T19:00:03.583","AcceptedAnswerId":null,"OwnerUserId":4777983.0,"Title":"caffe make error undefined reference x86_64","Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/29585180\/caffe-make-error-in-mac-os-10-9\">https:\/\/stackoverflow.com\/questions\/29585180\/caffe-make-error-in-mac-os-10-9<\/a><\/p>\n\n<p>Mac OS 10.9, Cuda 6.5, Opencv 2.4.11, Anaconda..<\/p>\n\n<p>By removing -limgcodecs, I am getting the same error as before-<\/p>\n\n<pre><\/pre>\n\n<p>.............................................................................<\/p>\n\n<p>Also I tried  to install opencv 3.0, <\/p>\n\n<p>$ brew install -devel opencv<\/p>\n\n<p>but is giving the foll CMake error-<\/p>\n\n<pre><\/pre>\n\n<p>Any help appreciated.<\/p>\n\n<hr>\n\n<p>Finally I installed opencv3 using cmake and not through brew. But now there seems to the same error- undefined symbols but with leveldb-<\/p>\n\n<pre><\/pre>\n\n<p>I even tried to re-install leveldb but the problem persists. Any help appreciated<\/p>\n","answers":[{"AnswerId":"29657441","CreationDate":"2015-04-15T17:58:18.453","ParentId":null,"OwnerUserId":"4777983","Title":null,"Body":"<p>I had to revert back to opencv 2.4.11. Uninstalled brew and all its dependencies. Loaded Cuda 7.0. Also added the ENV flags in dependencies with -stdlib=libc++ option. This solved the problem for me, although I still don't know the cause.<\/p>\n"}]}
{"QuestionId":29606856,"AnswerCount":2,"Tags":"<python><c++><dllimport><caffe>","CreationDate":"2015-04-13T13:43:07.753","AcceptedAnswerId":null,"OwnerUserId":1972060.0,"Title":"How to import caffe module in Python?","Body":"<p>I have build .dll of _caffe.cpp on Windows (Release, x64).<\/p>\n\n<p>I changed extension .dll to .pyd and trying to import it in python:<\/p>\n\n<pre><\/pre>\n\n<p>What does it mean, some module of dependencies missing which was included in project in Visual Studio, where I build this dll?<\/p>\n","answers":[{"AnswerId":"40053725","CreationDate":"2016-10-15T00:19:53.260","ParentId":null,"OwnerUserId":"6779334","Title":null,"Body":"<p>For windows : <\/p>\n\n<p>Adding <code>\/caffe\/Build\/x64\/Release\/pycaffe<\/code> to system path(<code>path<\/code>) works for me, and I think the best way to do it is :<\/p>\n\n<ol>\n<li>New a system variable : <code>PYTHON_PKG = \/caffe\/Build\/x64\/Release\/pycaffe;<\/code><\/li>\n<li>Include <code>PYTHON_PKG<\/code> in <code>path<\/code> : <code>path = %PYTHON_PKG%; %OtherDirs%<\/code><\/li>\n<\/ol>\n\n<p>After I did this, I get PKG missing <code>google.internal<\/code>, then I did <code>pip install google.internal<\/code> in <code>CMD<\/code>. It works.<\/p>\n"},{"AnswerId":"30886477","CreationDate":"2015-06-17T08:41:16.677","ParentId":null,"OwnerUserId":"4973198","Title":null,"Body":"<p>You need to add Python Caffe to PYTHONPATH. For example:\n<strong>export PYTHONPATH=$PYTHONPATH:\/home\/username\/caffe\/python<\/strong><\/p>\n"}]}
{"QuestionId":29649623,"AnswerCount":1,"Tags":"<python-2.7><memory-management><bigdata><theano>","CreationDate":"2015-04-15T11:59:10.483","AcceptedAnswerId":29650507.0,"OwnerUserId":1357690.0,"Title":"Out of memory when creating a Theano shared variable with borrow=True","Body":"<p>I'm trying to allocate a really big dataset (~28GB of RAM in an ndarray) into theano shared variables, using borrow=True to avoid replicating the memory. In order to do so, I'm using the following function:<\/p>\n\n<pre><\/pre>\n\n<p>In order to avoid data conversions, prior to saving the arrays to disk I already defined them to be in the correct format (afterwards filling them and dumping them into disk with np.save()):<\/p>\n\n<pre><\/pre>\n\n<p>It seems, though, that theano tires to replicate the memory anyway, dumping me the following error:<\/p>\n\n<p><em>Error allocating 25594500000 bytes of device memory (out of memory). Driver report 3775729664 bytes free and 4294639616 bytes total.<\/em><\/p>\n\n<p>Theano is configured to work on the GPU (GTX 970).<\/p>\n","answers":[{"AnswerId":"29650507","CreationDate":"2015-04-15T12:41:20.407","ParentId":null,"OwnerUserId":"1357690","Title":null,"Body":"<p>Instead of using <code>theano.shared<\/code>, it is possible to use <code>theano.tensor._shared<\/code> to force the data to be allocated into CPU memory. The fixed code ends up like this:<\/p>\n\n<pre><code>def load_dataset(path):\n    # Load dataset from memory\n    data_f = np.load(path+'train_f.npy')\n    data_t = np.load(path+'train_t.npy')\n\n    # Split into training and validation\n    return (\n        (\n            theano.tensor._shared(data_f[:-1000, :], borrow=True),\n            theano.tensor._shared(data_t[:-1000, :], borrow=True)\n        ), (\n            theano.tensor._shared(data_f[-1000:, :], borrow=True),\n            theano.tensor._shared(data_t[-1000:, :], borrow=True)\n        )\n    )\n<\/code><\/pre>\n"}]}
{"QuestionId":29658790,"AnswerCount":2,"Tags":"<python><opencv><neural-network><deep-learning><caffe>","CreationDate":"2015-04-15T19:11:19.843","AcceptedAnswerId":null,"OwnerUserId":2631051.0,"Title":"how to test mnist on my own dataset images","Body":"<p>I'm trying to test mnist using my own dataset of digits images.<br>\nI wrote a python script for that but it is giving an error. error is in line no 16 of code. Actually i'm not able to send image for test. give me some suggestions. thanks in advance.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29776555","CreationDate":"2015-04-21T15:24:35.953","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Why don't you use the python wrapper class <code>Classifier<\/code>?<\/p>\n\n<pre><code>net = caffe.Classifier( MODEL_FILE, PRETRAINED )\nnet.predict( [input_image], oversmaple=False )\n<\/code><\/pre>\n\n<p>I'm not 100% sure, but I think LeNeT model expect gray scale image, you might need to read the image<\/p>\n\n<pre><code>input_image = caffe.io.load_image(IMAGE_FILE, color=False)\n<\/code><\/pre>\n"},{"AnswerId":"34608567","CreationDate":"2016-01-05T09:48:37.843","ParentId":null,"OwnerUserId":"988709","Title":null,"Body":"<pre><code>import caffe\nimport os\n\nmodel_file = '..\/examples\/mnist\/lenet.prototxt'\npretrained_file = '..\/examples\/mnist\/lenet_iter_10000.caffemodel'\nnet = caffe.Classifier(model_file, pretrained_file, image_dims=(28, 28), raw_scale=255)\nscore = net.predict([caffe.io.load_image('img\/1.bmp', color=False)], oversample=False)\nprint score\n<\/code><\/pre>\n\n<p>This code work for me, the output is like that:<\/p>\n\n<pre><code>...\n[[ 0.  0.  1.  0.  0.  0.  0.  0.  0.  0.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":29662576,"AnswerCount":1,"Tags":"<lua><statistics><correlation><torch>","CreationDate":"2015-04-15T23:13:33.293","AcceptedAnswerId":29662849.0,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, how to pair together to arrays into a table?","Body":"<p>I need to use the Pearson correlation coefficient in my Torch \/ Lua program.\nThis is the function:<\/p>\n\n<pre><\/pre>\n\n<p>This function wants an input couple table that can be used in the pairs() function.\nI tried to submit to it the right input table but I was not able to get anything working.<\/p>\n\n<p>I tried with:<\/p>\n\n<pre><\/pre>\n\n<p>But unfortunately it does not work. It will compute pairs between the first elements of b, while I want it to compute the correlation between a and b.\nHow could I do?<\/p>\n\n<p>Can you provide a good object that may work as input to the math.pearson() function?<\/p>\n","answers":[{"AnswerId":"29662849","CreationDate":"2015-04-15T23:37:50.427","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>If you are using <a href=\"https:\/\/github.com\/attractivechaos\/klib\/blob\/master\/lua\/klib.lua\" rel=\"nofollow\">this implementation<\/a>, then I think it expects a different parameter structure. You should be passing a table of two-value tables, as in:<\/p>\n\n<pre><code>local z = {\n  {27, 0.0001}, {29, 0.001}, {45, 0.32132}, {98, 0.0001}, {1293, 0.0009}\n}\nprint(math.pearson(z))\n<\/code><\/pre>\n\n<p>This prints <code>-0.25304101592759<\/code> for me.<\/p>\n"}]}
{"QuestionId":29667795,"AnswerCount":2,"Tags":"<cuda><makefile><osx-mavericks><dylib><caffe>","CreationDate":"2015-04-16T07:13:10.263","AcceptedAnswerId":null,"OwnerUserId":4777983.0,"Title":"Segmentation fault in caffe","Body":"<p>Mac OS 10.9, OpenCV 2.4.11, CUDA 7.0, \nAll env flags set to libc++<\/p>\n\n<pre><\/pre>\n\n<p>As this could be related to library environment variables.\nHere are all my env variables-\n$DYLD_LIBRARY_PATH = \/usr\/local\/cuda\/lib<\/p>\n\n<p>$LD_LIBRARY_PATH = \/usr\/local\/cuda\/lib:\/opt\/intel\/composer_xe_2015.2.132\/mkl\/lib\/<\/p>\n\n<p>$DYLD_FALLBACK_LIBRARY_PATH = \n\/usr\/local\/cuda\/lib:\/Developer\/NVIDIA\/CUDA-7.0\/lib:\/Users\/deepsamal\/anaconda\/lib:\/usr\/local\/lib:\/usr\/lib:\/opt\/intel\/composer_xe_2015.2.132\/mkl\/lib\/:<\/p>\n\n<p>As both-<\/p>\n\n<pre><\/pre>\n\n<p>are running without error, this means library is linked but not getting loaded dynamically.<\/p>\n\n<p>Cant figure out what could be the reason? <\/p>\n\n<p>Any help appreciated.<\/p>\n\n<p>EDIT: tried to see the runtime link of the libcudart library.<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: I tried to find all soft links to libcaffe.so, it seems the paths to cuda libs and cudnn are not resolved and that seems to be the problem.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29671149","CreationDate":"2015-04-16T09:51:16.960","ParentId":null,"OwnerUserId":"1662497","Title":null,"Body":"<p>There is a terrible mess with OSX tools at execution time because of the DYLD_LIBRARY_PATH\nI think this kind of hack should work : <\/p>\n\n<pre><code>DYLD_LIBRARY_PATH=''; make runtest\n<\/code><\/pre>\n"},{"AnswerId":"30596127","CreationDate":"2015-06-02T12:29:29.320","ParentId":null,"OwnerUserId":"27943","Title":null,"Body":"<p>Setting the <code>DYLD_FALLBACK_LIBRARY_PATH<\/code> variable fixed this for me at least. I just had to add the <code>\/usr\/local\/cuda\/lib<\/code> as the first path (as you have done)<\/p>\n\n<pre><code>export DYLD_FALLBACK_LIBRARY_PATH=\/usr\/local\/cuda\/lib:\/usr\/local\/lib:\/usr\/lib:\/Developer\/NVIDIA\/CUDA-7.0\/lib:\n<\/code><\/pre>\n\n<p>or as documented <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/BkMspfbYLIE\/Pi6SUVANAKUJ\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":29669774,"AnswerCount":1,"Tags":"<matlab><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-04-16T08:51:54.537","AcceptedAnswerId":null,"OwnerUserId":4795605.0,"Title":"Extract filters and biases using caffe","Body":"<p>I want to extract filters and biases from my own caffemodel, (no need to visualize the feature) and I want to save them into  file for .\nI used  to do this issue: <\/p>\n\n<p>My workaround:<\/p>\n\n<pre><\/pre>\n\n<p>But some errors happened:<\/p>\n\n<pre><\/pre>\n\n<p>I dont know how to solve this problem.\nCan anyone help me? please and thanks.<\/p>\n","answers":[{"AnswerId":"29789189","CreationDate":"2015-04-22T06:13:25.893","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to give an additional argument for <code>'init'<\/code> the phase name.<br>\nI believe in your case<\/p>\n\n<pre><code> caffe('init',...\n       'supe_resolution_train_test.prototxt',...\n       'super_resolution_iter_1000.caffemodel',...\n       'test');\n<\/code><\/pre>\n"}]}
{"QuestionId":29691735,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-04-17T06:18:49.067","AcceptedAnswerId":29705705.0,"OwnerUserId":2161754.0,"Title":"How to find which variable is float64 when trying to use GPU in Theano","Body":"<p>In theano, when using gpu, the variables have to be float32. I checked all my variables to be folat32, but I still get the error below.<\/p>\n\n<pre><\/pre>\n\n<p>Seems that some variables are still float64, my question is how to locate the position I use the a float64 variable.<\/p>\n","answers":[{"AnswerId":"29705705","CreationDate":"2015-04-17T17:23:51.617","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>With Theano 0.7, you can use the Theano flags: warn_float64. You can give him one of those values: 'ignore', 'warn', 'raise', 'pdb'.<\/p>\n\n<p>This allow you to easily find where float64 are created.<\/p>\n"}]}
{"QuestionId":29698136,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-04-17T11:40:40.067","AcceptedAnswerId":29699422.0,"OwnerUserId":894903.0,"Title":"Caffe HDF5 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 HDF5 data 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\n<h1>Update<\/h1>\n\n<p>This seems to run:<\/p>\n\n<pre><\/pre>\n\n<h2>UPDATE2<\/h2>\n\n<p>Virtually identical models: One works the other doesn't. I cant explain why!<\/p>\n\n<p>Works:<\/p>\n\n<pre><\/pre>\n\n<p>Doesn't:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29699422","CreationDate":"2015-04-17T12:36:33.140","ParentId":null,"OwnerUserId":"3391810","Title":null,"Body":"<p>I ran into the same problem yesterday. You might need to check that your caffe version is up-to-date. They changed the protobuf definition quite heavily. In your case, \"type\" used to be an enum, now it just takes a string.<\/p>\n\n<p><strong><em>[Update]<\/em><\/strong> From the discussion in the comments: The answer is to not use \"layers\" but \"layer\". layers was probably present in some old \/ outdated example.<\/p>\n"}]}
{"QuestionId":29703182,"AnswerCount":1,"Tags":"<c++><ios><xcode><caffe>","CreationDate":"2015-04-17T15:16:54.683","AcceptedAnswerId":31537251.0,"OwnerUserId":4733134.0,"Title":"How to build Caffe framework XCode 6.2, iOS 8.3 environment","Body":"<p>I am working on build caffe framework for ios, i used the Caffe master source and make files to build the framework for iOS. <\/p>\n\n<p>I changed the OS target in CMake GUI config as \"\/Applications\/Xcode.app\/Contents\/Developer\/Platforms\/iPhoneOS.platform\/Developer\/SDKs\/iPhoneOS.sdk\"<\/p>\n\n<p>While run Xcode to build project i got the below error messages\n\/Users\/Macpro_ios_v2\/Caffe_iOS\/src\/caffe\/common.cpp:1:10: 'glog\/logging.h' file not found\n\"boost\/thread.h\" file not found\ni included usr\/local\/include and opt\/local\/include to search paths in the build phase  <\/p>\n\n<p>while run the same xcode project for OSX, it works fine and generates the libs perfectly.  If i change the target for iphone OS i got above error.<\/p>\n\n<p>Please help me to fix the above issuse. Please help how to configure  the make list in Caffe master for iphone.<\/p>\n\n<p>I have caffe dylip, lib-a for OSX. Is it possible to link mac osx libaries in iOS project?<\/p>\n","answers":[{"AnswerId":"31537251","CreationDate":"2015-07-21T11:01:38.803","ParentId":null,"OwnerUserId":"1217590","Title":null,"Body":"<p>cloudVision,<\/p>\n\n<p>You should install OpenCV2 from here before you compile the caffe-ios-sample:\n<a href=\"http:\/\/docs.opencv.org\/doc\/tutorials\/introduction\/ios_install\/ios_install.html\" rel=\"nofollow\">http:\/\/docs.opencv.org\/doc\/tutorials\/introduction\/ios_install\/ios_install.html<\/a><\/p>\n"}]}
{"QuestionId":29707174,"AnswerCount":2,"Tags":"<python><machine-learning><theano><deep-learning>","CreationDate":"2015-04-17T18:47:42.160","AcceptedAnswerId":29729742.0,"OwnerUserId":2824962.0,"Title":"Theano\/Lasagne\/Nolearn Neural Network Image Input","Body":"<p>I am working on image classification tasks and decided to use Lasagne + Nolearn for neural networks prototype.\nAll standard examples like MNIST numbers classification run well, but problems appear when I try to work with my own images.<\/p>\n\n<p>I want to use 3-channel images, not grayscale.\nAnd there is the code where I'm trying to get arrays from images:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the code of NN and its fitting:<\/p>\n\n<pre><\/pre>\n\n<p>I recieve exceptions like this one:<\/p>\n\n<pre><\/pre>\n\n<p><strong>So, in which format do you \"feed\" your networks with image data?<\/strong>\nThanks for answers or any tips!<\/p>\n","answers":[{"AnswerId":"30922572","CreationDate":"2015-06-18T18:03:25.130","ParentId":null,"OwnerUserId":"3502360","Title":null,"Body":"<p>If you're doing classification you need to modify a couple of things:<\/p>\n\n<ol>\n<li>In your code you have set <code>regression = True<\/code>. To do classification remove this line.<\/li>\n<li>Ensure that your input shape matches the shape of X if want to input 3 distinct channels<\/li>\n<li><p>Because you are doing classification you need the output to use a softmax nonlinearity (at the moment you have the identity which will not help you with classification)<\/p>\n\n<pre><code>X, y = simple_load(\"new\")\n\nX = np.array(X)\ny = np.array(y)\n\nnet1 = NeuralNet(\n    layers=[  # three layers: one hidden layer\n        ('input', layers.InputLayer),\n        ('hidden', layers.DenseLayer),\n        ('output', layers.DenseLayer),\n        ],\n    # layer parameters:\n    input_shape=(None, 3, 256, 256),  # TODO: change this\n    hidden_num_units=100,  # number of units in hidden layer\n    output_nonlinearity=lasagne.nonlinearities.softmax, # TODO: change this\n    output_num_units=len(y),  # 30 target values\n\n    # optimization method:\n    update=nesterov_momentum,\n    update_learning_rate=0.01,\n    update_momentum=0.9,\n\n    max_epochs=400,  # we want to train this many epochs\n    verbose=1,\n<\/code><\/pre>\n\n<p>)<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"29729742","CreationDate":"2015-04-19T12:18:00.923","ParentId":null,"OwnerUserId":"2824962","Title":null,"Body":"<p>I also asked it in lasagne-users forum and Oliver Duerr helped me a lot with code sample:\n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/lasagne-users\/8ZA7hr2wKfM\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/lasagne-users\/8ZA7hr2wKfM<\/a><\/p>\n"}]}
{"QuestionId":29723803,"AnswerCount":1,"Tags":"<python><sympy><theano>","CreationDate":"2015-04-18T22:50:58.273","AcceptedAnswerId":null,"OwnerUserId":3608005.0,"Title":"How to evaluate and compile such a function using theano or autowrap?","Body":"<p>I would like to evaluate functions of similar form (more complicated) in sympy.<\/p>\n\n<pre><\/pre>\n\n<p>where all variables are vectors of length . The evaluation will take place at every timestep of an optimization routine. As a consequence, I would like to implement it efficiently. Most probably, it would be best to compile those functions but the autowrap module gives me strange errors.<\/p>\n\n<p>What works:<\/p>\n\n<pre><\/pre>\n\n<p>I can evaluate this expression directly in sympy:<\/p>\n\n<pre><\/pre>\n\n<p>Gives me (as expected):<\/p>\n\n<p><\/p>\n\n<p>But I failed at compiling this expression. I am reading blogs and manuals since hours, but either I am too tired or did not find the proper information. Any help is strongly appreciated.<\/p>\n\n<p>EDIT: In response to Eric: I do not now how to implement the sum of a vector in theano or autowrap. I tried different versions using lambda functions and got various errors. Maybe the most reproducible has to do with the dimension of the inputs:<\/p>\n\n<pre><\/pre>\n\n<p>If I try to compile a simple expression using autowrap:<\/p>\n\n<pre><\/pre>\n\n<p>I obtain:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34983147","CreationDate":"2016-01-25T00:01:39.867","ParentId":null,"OwnerUserId":"467314","Title":null,"Body":"<p>There are mulitple issues with your approach. 1) You can't mix in numpy functions with SymPy types and then expect the code generators to work with the mixed types. The code generators in SymPy work with SymPy types only. You can pass in external functions that map to SymPy functions. For example, you can use sympy.Sum() and then write a mapping from SymPy.Sum() to numpy.sum() for the code generator to consume. 2) Eq() is not supported in code generators as far as I know. 3) The indexed types do very specific things when passed to autowrap. You'll need to read the docs carefully about them.<\/p>\n"}]}
{"QuestionId":29735197,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-04-19T20:04:32.660","AcceptedAnswerId":null,"OwnerUserId":1382757.0,"Title":"does Theano's symbolic matrix inverse \"know\" about structured matrices?","Body":"<p>In my application I am computing the inverse of a block tridiagonal matrix A - will Theano's matrix inverse account for that structure (by using a more efficient matrix inverse algorithm)?<\/p>\n\n<p>Further, I only need the diagonal and first off diagonal blocks of the resulting inverse matrix. Is there a way of preventing Theano from computing the remaining blocks? <\/p>\n\n<p>Generally, I'm curious whether it would be worth implementing a forward\/backward block tridagonal matrix inverse algorithm myself.<\/p>\n","answers":[{"AnswerId":"29934743","CreationDate":"2015-04-29T05:01:06.163","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>As of April 2015 Theano matrix inverse function won't do it directly:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nlinalg.html#theano.tensor.nlinalg.MatrixInverse\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nlinalg.html#theano.tensor.nlinalg.MatrixInverse<\/a><\/p>\n\n<p>Theano do not have many optimization and function related to that type of methods. It partially wrap what is under numpy.linalg (most of it) and some of scipy.linalg:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/slinalg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/slinalg.html<\/a><\/p>\n\n<p>So you are better in the short term to do it with numpy\/scipy directly.<\/p>\n\n<p>If you want to add those feature to Theano, this can be done. But it need someone with the time and willingness to do it.<\/p>\n"}]}
{"QuestionId":29739796,"AnswerCount":1,"Tags":"<machine-learning><theano><conv-neural-network><pylearn>","CreationDate":"2015-04-20T05:18:32.140","AcceptedAnswerId":29765165.0,"OwnerUserId":2388116.0,"Title":"How to use leaky ReLus as the activation function in hidden layers in pylearn2","Body":"<p>I am using pylearn2 library to design a CNN. I want to use Leaky ReLus as the activation function in one layer. Is there any possible way to do this using pylearn2? Do I have to write a custom function for it or does pylearn2 have inbuilt funtions for tha? If so, how to write a custom code? Please can anyone help me out here?<\/p>\n","answers":[{"AnswerId":"29765165","CreationDate":"2015-04-21T07:04:21.387","ParentId":null,"OwnerUserId":"1742064","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/models\/mlp.py#L2849\" rel=\"nofollow\" title=\"ConvElemwise\">ConvElemwise<\/a> super-class is a generic convolutional elemwise layer. Among its subclasses <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/models\/mlp.py#L3305\" rel=\"nofollow\" title=\"ConvRectifiedLinear\">ConvRectifiedLinear<\/a> is a convolutional rectified linear layer that uses <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/models\/mlp.py#L2718\" rel=\"nofollow\" title=\"RectifierConvNonlinearity\">RectifierConvNonlinearity<\/a> class.<\/p>\n\n<p>In the <code>apply()<\/code> method:<\/p>\n\n<pre><code>    p = linear_response * (linear_response &gt; 0.) + self.left_slope *\\\n        linear_response * (linear_response &lt; 0.)\n<\/code><\/pre>\n\n<p>As <a href=\"http:\/\/cs231n.github.io\/neural-networks-1\/\" rel=\"nofollow\">this<\/a> gentle review points out:<\/p>\n\n<blockquote>\n  <p>... Maxout neuron (introduced recently by <a href=\"http:\/\/www-etud.iro.umontreal.ca\/~goodfeli\/maxout.html\" rel=\"nofollow\">Goodfellow et al.<\/a>) that generalizes the ReLU and its leaky version.<\/p>\n<\/blockquote>\n\n<p>Examples are <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/models\/maxout.py#L956\" rel=\"nofollow\">MaxoutLocalC01B<\/a> or <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/models\/maxout.py#L515\" rel=\"nofollow\">MaxoutConvC01B<\/a>.<\/p>\n\n<p>The reason for lack of answer in <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/pylearn-users\/JaSbTVNvrWo\" rel=\"nofollow\">pylearn2-user<\/a> may be that <a href=\"http:\/\/deeplearning.net\/software\/pylearn2\/\" rel=\"nofollow\">pylearn2<\/a> is mostly written by researches at <a href=\"https:\/\/sites.google.com\/a\/lisa.iro.umontreal.ca\/mila\/\" rel=\"nofollow\">LISA lab<\/a> and, thus, the threshold for point 13 in <a href=\"http:\/\/deeplearning.net\/software\/pylearn2\/faq.html\" rel=\"nofollow\">FAQ<\/a> may be high.<\/p>\n"}]}
{"QuestionId":29774413,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-04-21T13:59:51.910","AcceptedAnswerId":29774499.0,"OwnerUserId":492372.0,"Title":"What is the borrow parameter in Theano","Body":"<p>I see this following line of code:<\/p>\n\n<pre><\/pre>\n\n<p>In the above line, what is the borrow parameter exactly? What is the advantage of adding that there? FYI, train_set_x is basically a matrix that was generated using the theano.shared method.<\/p>\n","answers":[{"AnswerId":"29774499","CreationDate":"2015-04-21T14:02:56.647","ParentId":null,"OwnerUserId":"28169","Title":null,"Body":"<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/aliasing.html#borrowing-when-creating-shared-variables\">This part of the documentation<\/a> seems relevant:<\/p>\n\n<blockquote>\n  <p>By default (<code>s_default<\/code>) and when explicitly setting <code>borrow=False<\/code>, the shared variable we construct gets a <em>deep<\/em> copy of np_array. So changes we subsequently make to <code>np_array<\/code> have no effect on our shared variable.<\/p>\n<\/blockquote>\n\n<p>Setting it to <code>True<\/code> can then be assumed to make a shallow copy, effectively letting you \"borrow\" access to the memory.<\/p>\n"}]}
{"QuestionId":29774793,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2015-04-21T14:15:15.650","AcceptedAnswerId":29775072.0,"OwnerUserId":894903.0,"Title":"TypeError Python Class","Body":"<p>I have a library (caffe) that has the following definition:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to call it using  I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I checked <a href=\"https:\/\/stackoverflow.com\/questions\/9698614\/super-raises-typeerror-must-be-type-not-classobj-for-new-style-class\">SO-9698614<\/a>,<a href=\"https:\/\/stackoverflow.com\/questions\/489269\/python-super-raises-typeerror-why?lq=1\">SO-576169<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/489269\/python-super-raises-typeerror-why?lq=1\">SO-489269<\/a> but they did not lead to a solution.  My class is a new type class and I could not see why it was not working. <\/p>\n\n<h1>Full trace:<\/h1>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"29775072","CreationDate":"2015-04-21T14:25:49.673","ParentId":null,"OwnerUserId":"100297","Title":null,"Body":"<p>Somehow you managed to bind <code>NetSpec<\/code> to <code>None<\/code> somewhere:<\/p>\n\n<pre><code>&gt;&gt;&gt; super(None, object())\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nTypeError: must be type, not None\n<\/code><\/pre>\n\n<p>The error indicates that the <code>NetSpec<\/code> global is bound to <code>None<\/code>.<\/p>\n\n<p>You could also bypass the <code>NetSpec.__setattr__<\/code> method by going directly to the instance <code>__dict__<\/code> attribute:<\/p>\n\n<pre><code>class NetSpec(object):\n    def __init__(self):\n       self.__dict__['tops'] = OrderedDict()\n<\/code><\/pre>\n\n<p>From the code you shared it could be that this is the culprit:<\/p>\n\n<pre><code>from .layers import layers, params, NetSpec\n<\/code><\/pre>\n\n<p>This imports <code>caffe.layers<\/code> but rebinds <code>caffe.layers<\/code> to the <code>Layers()<\/code> instance. This <em>could<\/em> then trigger Python to delete the module again as there are no other references to it yet (depending on when and how the <code>sys.modules<\/code> reference is created), causing all globals to be rebound to <code>None<\/code> (including <code>NetSpec<\/code>).<\/p>\n"}]}
{"QuestionId":29776502,"AnswerCount":3,"Tags":"<python-2.7><installation><windows-8.1><anaconda><theano>","CreationDate":"2015-04-21T15:22:22.780","AcceptedAnswerId":29855776.0,"OwnerUserId":4815525.0,"Title":"Installing Theano on windows 8.1 with Anaconda: setting the system path configuration script","Body":"<p>I\u2019m trying to install Theano on windows 8.1 64 with Anaconda following step by step the guide provided here: <a href=\"http:\/\/theano.readthedocs.org\/en\/latest\/install_windows.html\" rel=\"nofollow\">http:\/\/theano.readthedocs.org\/en\/latest\/install_windows.html<\/a>.\nI get stuck with the environment configuration script 'env.bat' needed to configure the system path.\nThe example refers to a WinPython distribution but as I\u2019m installing in Anaconda and I don\u2019t know how to configure that specific row. <\/p>\n\n<p>This is the example for WinPython:<\/p>\n\n<pre><\/pre>\n\n<p>Which directory should I set here after the CALL considering I\u2019m using Anaconda? Struggling quite a lot, could anyone please help?<\/p>\n\n<p>EDIT: please note that SCISOFT is the directory where WinPython is installed in the tutorial, the author says: \"The script assumes that you installed WinPython distribution, update the winpython line otherwise.\" and that is what I'm not able to do because it is not specified what to point at.<\/p>\n\n<p>I'm including the whole .bat, though I have no problem with the other settings:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30539474","CreationDate":"2015-05-29T21:38:15.660","ParentId":null,"OwnerUserId":"4853662","Title":null,"Body":"<p>I had similar problems and so I compiled a solid guide on how to install Theano on Windows 8.1 x64, using WinPython x64 and CUDA 7 with MS Visual Studio 2012 - CPU and GPU both set up.<\/p>\n\n<p><a href=\"http:\/\/machinelearning.berlin\/?p=383\" rel=\"nofollow\">http:\/\/machinelearning.berlin\/?p=383<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"29862969","CreationDate":"2015-04-25T08:40:27.653","ParentId":null,"OwnerUserId":"3140336","Title":null,"Body":"<p>next winpython should have theano working out of the box <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/lta_34FXIwg\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/lta_34FXIwg<\/a><\/p>\n"},{"AnswerId":"29855776","CreationDate":"2015-04-24T19:20:56.010","ParentId":null,"OwnerUserId":"161801","Title":null,"Body":"<p>I don't know what is in the WinPython env.bat, but you probably can just delete it. You likely just need to make sure that Anaconda is on the PATH. <\/p>\n"}]}
{"QuestionId":29780804,"AnswerCount":2,"Tags":"<data-structures><lua><torch>","CreationDate":"2015-04-21T18:53:46.230","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"In Torch \/ Lua, is there a command to analyze an object (like str() in R)?","Body":"<p>I have to analyze some objects and their structure in my Torch \/ Lua script.\nI would like to use a command that behave like <a href=\"https:\/\/stat.ethz.ch\/R-manual\/R-devel\/library\/utils\/html\/str.html\" rel=\"nofollow\"> in R<\/a>.<\/p>\n\n<p>Do you have any suggestion?<\/p>\n","answers":[{"AnswerId":"29781146","CreationDate":"2015-04-21T19:13:49.087","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>You may want to use a <a href=\"http:\/\/lua-users.org\/wiki\/TableSerialization\" rel=\"nofollow\">serializer<\/a> to represent complex data structures in a readable way. There is <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/serialization.md\" rel=\"nofollow\">torch.serialize<\/a> function, but it doesn't produce human-readable output. I've written <a href=\"https:\/\/github.com\/pkulchenko\/serpent\" rel=\"nofollow\">Serpent serializer and pretty-printer<\/a> that supports some of the options that <code>str()<\/code> has, like the max nesting level for tables or the max number of elements in a table. It also supports custom formatters, which allows you to modify the output to some degree.<\/p>\n"},{"AnswerId":"29797140","CreationDate":"2015-04-22T12:09:43.590","ParentId":null,"OwnerUserId":"1243636","Title":null,"Body":"<p>I like this module: <a href=\"https:\/\/github.com\/kikito\/inspect.lua\" rel=\"nofollow\">https:\/\/github.com\/kikito\/inspect.lua<\/a><\/p>\n\n<p><code>luarocks install inspect<\/code><\/p>\n\n<p>then import it like this<\/p>\n\n<p><code>local inspect = require 'inspect'<\/code><\/p>\n\n<p>output may be something like this:<\/p>\n\n<pre><code>assert(inspect(setmetatable({a=1}, {b=2}) == [[{\n  a = 1\n  &lt;metatable&gt; = {\n    b = 2\n  }\n}]]))\n<\/code><\/pre>\n\n<p>common usage:<\/p>\n\n<pre><code>print(inspect(myobj))\n<\/code><\/pre>\n"}]}
{"QuestionId":29788075,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><caffe><glog>","CreationDate":"2015-04-22T04:56:59.090","AcceptedAnswerId":29788785.0,"OwnerUserId":1452257.0,"Title":"Setting GLOG_minloglevel=1 to prevent output in shell from Caffe","Body":"<p>I'm using Caffe, which is printing a lot of output to the shell when loading the neural net.<br>\nI'd like to suppress that output, which supposedly can be done by setting  when running the Python script. I've tried doing that using the following code, but I still get all the output from loading the net. How do I suppress the output correctly?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31350273","CreationDate":"2015-07-10T21:06:25.190","ParentId":null,"OwnerUserId":"3006372","Title":null,"Body":"<p>I was able to get <a href=\"https:\/\/stackoverflow.com\/a\/29788785\/1714410\">Shai's solution<\/a> to work, but only by executing that line in Python <em>before<\/em> calling <\/p>\n\n<pre><code>import caffe\n<\/code><\/pre>\n"},{"AnswerId":"29788785","CreationDate":"2015-04-22T05:49:51.057","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>To supress the output level you need to <strong>increase<\/strong> the loglevel to at least 2<\/p>\n\n<pre><code> os.environ['GLOG_minloglevel'] = '2' \n<\/code><\/pre>\n\n<p>The levels are<\/p>\n\n<p>0 - debug<br>\n 1 - info (still a LOT of outputs)<br>\n 2 - warnings<br>\n 3 - errors  <\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong><br>\nSince this flag is <em>global<\/em> to <code>caffe<\/code>, it must be set <em>prior<\/em> to importing of <code>caffe<\/code> package (as pointed out by <a href=\"https:\/\/stackoverflow.com\/a\/31350273\/1714410\">jbum<\/a>). Once the flag is set and <code>caffe<\/code> is imported the behavior of the GLOG tool cannot be changed.<\/p>\n"}]}
{"QuestionId":29802027,"AnswerCount":0,"Tags":"<torch>","CreationDate":"2015-04-22T15:21:17.523","AcceptedAnswerId":null,"OwnerUserId":4820103.0,"Title":"Torch: Model fast when learning\/testing, slow when using it","Body":"<p>I have an issue using a learned model with torch.<\/p>\n\n<p>I followed this howto <a href=\"http:\/\/code.cogbits.com\/wiki\/doku.php?id=tutorial_supervised\" rel=\"nofollow\">http:\/\/code.cogbits.com\/wiki\/doku.php?id=tutorial_supervised<\/a> to train a model. Everything is fine, my model was trained and I have corrects results when I use my model. But it's slow !<\/p>\n\n<p>The testing part for training look like this:<\/p>\n\n<pre><\/pre>\n\n<p>I have the following speed recorded during testing:<\/p>\n\n<p><\/p>\n\n<p>(Of course it vary, but it's ~12ms).<\/p>\n\n<p>I want to use the learned model on others images, so I did this in a simple and new script:<\/p>\n\n<pre><\/pre>\n\n<p>The time spent is much bigger, I have the following output: <\/p>\n\n<p>I tested with more than one image, always with the resizing and normalization outside clock's measurements, and I always have > 200ms \/ image.<\/p>\n\n<p>I don't understand what I'm doing wrong and why my code is much slower than during the training\/testing. <\/p>\n\n<p>Thanks !<\/p>\n","answers":[]}
{"QuestionId":29805179,"AnswerCount":1,"Tags":"<matlab><caffe><matcaffe>","CreationDate":"2015-04-22T17:43:16.757","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Error using caffe Invalid input size","Body":"<p>I tried to train my own neural net using my own imagedatabase as described in <\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html<\/a><\/p>\n\n<p>However when I want to check the neural net after training on some standard images using the matlab wrapper I get the following output \/ error:<\/p>\n\n<pre><\/pre>\n\n<p>I used the matlab wrapper before to extract cnn features based on a pretrained model. It worked. So I don't think the input size of my images is the problem (They are converted to the correct size internally by the function \"prepare_image\").<\/p>\n\n<p>Has anyone an idea what could be the error?<\/p>\n","answers":[{"AnswerId":"29824882","CreationDate":"2015-04-23T13:25:51.700","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>Found the solution: I was referencing the wrong \".prototxt\" file (Its a little bit confusing because the files are quite similar.\nSo for computing features using the matlab wrapper one needs to reference the following to files in \"matcaffe_demo.m\":<\/p>\n\n<pre><code>models\/bvlc_reference_caffenet\/deploy.prototxt\nmodels\/bvlc_reference_caffenet\/MyModel_caffenet_train_iter_450000.caffemodel\n<\/code><\/pre>\n\n<p>where \"MyModel_caffenet_train_iter_450000.caffemodel\" is the only file needed which is created during training.<\/p>\n\n<p>In the beginning I was accidently referencing <\/p>\n\n<pre><code>models\/bvlc_reference_caffenet\/MyModel_train_val.prototxt\n<\/code><\/pre>\n\n<p>which was the \".prototxt\" file used for training.<\/p>\n"}]}
{"QuestionId":29814641,"AnswerCount":2,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-04-23T05:42:00.847","AcceptedAnswerId":29971142.0,"OwnerUserId":4231726.0,"Title":"Regression with lasagne: error","Body":"<p>I am trying to run a regression with lasagne\/nolearn. I am having trouble finding documentation how to do that (new to deep learning in general).<\/p>\n\n<p>Starting off with a simple network (one hidden layer)<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!..<\/p>\n","answers":[{"AnswerId":"29971142","CreationDate":"2015-04-30T15:00:01.150","ParentId":null,"OwnerUserId":"576676","Title":null,"Body":"<p>Make sure that you're using versions of nolearn and Lasagne that are <strong>known to work together<\/strong>.<\/p>\n\n<p>Say you've been following the <a href=\"http:\/\/danielnouri.org\/notes\/2014\/12\/17\/using-convolutional-neural-nets-to-detect-facial-keypoints-tutorial\/\" rel=\"nofollow\">Using convolutional neural nets to detect facial keypoints tutorial<\/a>.  Then the right thing to do is to install the dependencies from <a href=\"https:\/\/raw.githubusercontent.com\/dnouri\/kfkd-tutorial\/master\/requirements.txt\" rel=\"nofollow\">this requirements.txt file<\/a>, like so:<\/p>\n\n<pre><code>pip uninstall Lasagne\npip uninstall nolearn\npip install -r https:\/\/raw.githubusercontent.com\/dnouri\/kfkd-tutorial\/master\/requirements.txt\n<\/code><\/pre>\n\n<p>If, however, you're using nolearn from Git master, then make sure you install the Lasagne version that's in the <a href=\"https:\/\/github.com\/dnouri\/nolearn\/blob\/master\/requirements.txt\" rel=\"nofollow\">requirements.txt file found there<\/a>:<\/p>\n\n<pre><code>pip uninstall Lasagne\npip install -r https:\/\/raw.githubusercontent.com\/dnouri\/nolearn\/master\/requirements.txt\n<\/code><\/pre>\n"},{"AnswerId":"29905390","CreationDate":"2015-04-27T20:31:50.423","ParentId":null,"OwnerUserId":"4839389","Title":null,"Body":"<p>Not sure what version of nolearn and lasagne you are using. I did notice that you have <code>y<\/code> as being of shape <code>(137,)<\/code>. From my usage this needs to be <code>(137, 1)<\/code> to work for your case, and, in general, dim 2 needs to match the <code>output_num_units<\/code>. <\/p>\n\n<p>Try <code>y.reshape((-1, 1))<\/code>.<\/p>\n\n<p>If this doesn't work it may be a Python 3 compatibility issue.<\/p>\n"}]}
{"QuestionId":29818284,"AnswerCount":2,"Tags":"<nlp><convolution><theano><deep-learning>","CreationDate":"2015-04-23T08:49:59.773","AcceptedAnswerId":33602229.0,"OwnerUserId":2161754.0,"Title":"how to make the image_shape dynamic in the convolution in Theano","Body":"<p>I tried to process the tweets dataset using CNN in Theano. Different from images, the lenght of different tweets (corresponding to the image shape) is variable. So the shape of each tweet is different. However, in Theano, the convolution need that the shape information are constant values. So my question is that is there some way to make the image_shape dynamic?<\/p>\n","answers":[{"AnswerId":"29837099","CreationDate":"2015-04-24T01:21:39.723","ParentId":null,"OwnerUserId":"3098048","Title":null,"Body":"<p>Convolutional neural networks are really better suited to processing images.<\/p>\n\n<p>For processing tweets, you might want to read about recursive neural networks.<\/p>\n\n<p><a href=\"http:\/\/nlp.stanford.edu\/~socherr\/EMNLP2013_RNTN.pdf\" rel=\"nofollow\">http:\/\/nlp.stanford.edu\/~socherr\/EMNLP2013_RNTN.pdf<\/a><\/p>\n"},{"AnswerId":"33602229","CreationDate":"2015-11-09T03:45:52.947","ParentId":null,"OwnerUserId":"5149834","Title":null,"Body":"<p>Kalchbrenner et. al (2015) implemented an CNN that accepts dynamic length input and pools them into k elements. If there are less than k elements to begin with, the remaining are zero-padded. Their experiments with sentence classification show that such networks successfully represent grammatical structures.<\/p>\n\n<p>For details check out:<\/p>\n\n<ul>\n<li>the paper (<a href=\"http:\/\/arxiv.org\/pdf\/1404.2188v1.pdf\" rel=\"nofollow\">http:\/\/arxiv.org\/pdf\/1404.2188v1.pdf<\/a>)<\/li>\n<li>Matlab code (link on page 2 of the paper)<\/li>\n<li>suggestion for DCNNs for Theano\/Keras (<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/373\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/issues\/373<\/a>)<\/li>\n<\/ul>\n"}]}
{"QuestionId":29825174,"AnswerCount":1,"Tags":"<python><list><neural-network><caffe>","CreationDate":"2015-04-23T13:37:42.157","AcceptedAnswerId":29825383.0,"OwnerUserId":639973.0,"Title":"python list not showing full elements","Body":"<p>I'm inserting images into Decaf, and want to extract features from 6,7,8th layers. 6th and 7th are supposed to be 4096-dimensions, and 8th is supposed to be 1000.<\/p>\n\n<p>I'm assuming that the generated output functions like a list, and want to record each element in a separate text file as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The f8 file correctly has 1000 files, but f6 and f7 text files have something like the following:<\/p>\n\n<pre><\/pre>\n\n<p>The dots in the middle are literally like that. What happend to all the numbers? Do those dots signify something? some kind of abridgement?\nIs this something that has to do with decaf or python?<\/p>\n","answers":[{"AnswerId":"29825383","CreationDate":"2015-04-23T13:45:02.973","ParentId":null,"OwnerUserId":"190597","Title":null,"Body":"<p>It looks like <code>feat6<\/code> is a NumPy array.\nIf so, instead of <\/p>\n\n<pre><code>f6name = fname+'-f6.txt'\nf6 = open(f6name,'w')\nfor f in feat6:\n    f6.write(str(f))\n    f6.write('\\t')\nf6.close()\n<\/code><\/pre>\n\n<p>use<\/p>\n\n<pre><code>import numpy as np\n\nf6name = fname+'-f6.txt'\nnp.savetxt(f6name, feat6, delimiter='\\t')\n<\/code><\/pre>\n\n<p>This won't include the brackets (<code>[<\/code> and <code>]<\/code>), but that is usually more desireable as it makes parsing the data easier.<\/p>\n\n<hr>\n\n<p>The <code>str<\/code> representation of NumPy arrays includes ellipses when the number of elements in the array exceed <code>threshold<\/code> which by default NumPy sets to 1000. You can change this by <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.set_printoptions.html\" rel=\"nofollow\">setting <code>threshold<\/code> to some higher number<\/a>: <\/p>\n\n<pre><code>import numpy as np\nnp.set_printoptions(threshold=10**6)\n<\/code><\/pre>\n\n<p>With this change, <code>str(f)<\/code> would return a stringified version of <code>f<\/code> without ellipses as long as <code>f.size<\/code> is less than 10**6.<\/p>\n\n<p>While this explains why you are seeing ellipses, I don't recommend using <code>np.set_printoptions<\/code> here since <code>np.savetxt<\/code> solves your problem more simply.<\/p>\n"}]}
{"QuestionId":29825687,"AnswerCount":2,"Tags":"<machine-learning><theano><pylearn><lstm>","CreationDate":"2015-04-23T13:56:22.307","AcceptedAnswerId":null,"OwnerUserId":1259448.0,"Title":"implementing LSTM with theano scan, way slower then using loops","Body":"<p>I am using Theano\/Pylearn2 to implement LSTM model inside my own network. However, I've found that Theano scan is much, much slower than using plain loops. I used the Theano profiler <\/p>\n\n<pre><\/pre>\n\n<p>and the Ops,<\/p>\n\n<pre><\/pre>\n\n<p>So lots and lots of time are spent on Scan (which is kind of as expected, but I didn't expect it to be soo slow). <\/p>\n\n<p>The main body of my code is<\/p>\n\n<pre><\/pre>\n\n<p>And I wrote my scan as:<\/p>\n\n<pre><\/pre>\n\n<p>One thing I've noticed is that the type of Theano scan uses Python implementation (?) is that the reason why this is ridiculously slow? or did I do something wrong? Why is Theano python implementation of Scan instead of C's.<\/p>\n\n<p>(I said using loops is faster, but it's faster at runtime, for large model I didn't manage to compile the version of using loops within reasonable amount of time).<\/p>\n","answers":[{"AnswerId":"38457559","CreationDate":"2016-07-19T11:43:02.800","ParentId":null,"OwnerUserId":"4862162","Title":null,"Body":"<p>It takes time for Theano developers to implement scan and gradient-of-scan using C and GPU, because it is much more complicated than other functions. That is why when you profile it, it shows GpuElemwise, GpuGemv, GpuDot22, etc., but you don't see a GpuScan or GpuGradofScan.<\/p>\n\n<p>Meanwhile, you can only fall back to for loops.<\/p>\n"},{"AnswerId":"31328630","CreationDate":"2015-07-09T21:34:01.867","ParentId":null,"OwnerUserId":"2938232","Title":null,"Body":"<p>This was asked a while ago but I had\/have the same problem.  Answer is that scan is slow on the GPU.<\/p>\n\n<p>See: <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/1168\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/1168<\/a><\/p>\n"}]}
{"QuestionId":29828908,"AnswerCount":2,"Tags":"<python><github><neural-network><theano>","CreationDate":"2015-04-23T16:08:06.533","AcceptedAnswerId":null,"OwnerUserId":2423116.0,"Title":"Theano warning: The same cache key is associated to different modules","Body":"<p>I've been using Lasagne for a while to run Neural Networks. I had installed it by downloading the repo from github and then doing .\nToday I tried to updating to the latest version. This is what I've done:<\/p>\n\n<p>-rename the previous lasagne folder to lasagne_old.<\/p>\n\n<p>-create a new lasagne folder with the new repo<\/p>\n\n<p>-<\/p>\n\n<p>The install completed fine.\nHowever as soon as I try running the usual Neural Networks it starts giving errors:<\/p>\n\n<pre><\/pre>\n\n<p>How could I fix this? And, moving forward, what is the right way to update  a package from a repo?<\/p>\n","answers":[{"AnswerId":"35456839","CreationDate":"2016-02-17T12:32:30.097","ParentId":null,"OwnerUserId":"4530440","Title":null,"Body":"<p>Use the command <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#updating-theano\" rel=\"nofollow\"><code>theano-cache clear<\/code><\/a>. I had a similar problem and it solved it. Hope it's helpful<\/p>\n\n<blockquote>\n  <p>If you installed NumPy\/SciPy with yum\/apt-get, updating NumPy\/SciPy with pip\/easy_install is not always a good idea. This can make Theano crash due to problems with BLAS (but see below). The versions of NumPy\/SciPy in the distribution are sometimes linked against faster versions of BLAS. Installing NumPy\/SciPy with yum\/apt-get\/pip\/easy_install won\u2019t install the development package needed to recompile it with the fast version. This mean that if you don\u2019t install the development packages manually, when you recompile the updated NumPy\/SciPy, it will compile with the slower version. This results in a slower Theano as well. To fix the crash, you can clear the Theano cache like this:<\/p>\n<\/blockquote>\n\n<pre><code>theano-cache clear\n<\/code><\/pre>\n"},{"AnswerId":"42933636","CreationDate":"2017-03-21T16:58:59.777","ParentId":null,"OwnerUserId":"3881080","Title":null,"Body":"<pre><code>$ theano-cache clear\n<\/code><\/pre>\n\n<p>and optionally<\/p>\n\n<pre><code>$ theano-cache purge\n<\/code><\/pre>\n"}]}
{"QuestionId":29842935,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-04-24T08:55:36.763","AcceptedAnswerId":null,"OwnerUserId":4827684.0,"Title":"Mean shape incompatible with input shape -- CAFFE Classification Error in IO.PY","Body":"<p>I'm installed Caffe on an Ubuntu 14.04 virtual server with CUDA installed (without driver) using <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM\" rel=\"nofollow\">this link<\/a>.<\/p>\n\n<p>The classification step: <\/p>\n\n<pre><\/pre>\n\n<p>yields a traceback.  <\/p>\n\n<p>I followed the steps as describe is by user2696499 in <a href=\"https:\/\/stackoverflow.com\/questions\/28692209\/using-gpu-despite-setting-cpu-only-yielding-unexpected-keyword-argument\/28979649#28979649\">this SO thread<\/a>,<br>\nHowever the variable  is not defined. How or where to define it ?<\/p>\n","answers":[{"AnswerId":"29887772","CreationDate":"2015-04-27T05:22:08.210","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><code>ins<\/code> stands for the input shape and is equal to <code>in_shape[1:]<\/code>.<\/p>\n"}]}
{"QuestionId":29871102,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-04-25T21:47:01.440","AcceptedAnswerId":29872347.0,"OwnerUserId":1546029.0,"Title":"In trepl or luajit, how can I find the source code of a library I'm using?","Body":"<p>Let's say I'm working with a lua library I installed using luarocks, and I want to see the definition of a function from that library. In ipython in could use <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>??function_name<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>to see the definition in the terminal, in matlab I could use <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>which function_name<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>then use my editor to look at the path returned by which. How could I do something similar to find the function definition for a lua library?<\/p>\n","answers":[{"AnswerId":"29872347","CreationDate":"2015-04-26T00:35:43.063","ParentId":null,"OwnerUserId":"805875","Title":null,"Body":"<p>In 'plain' Lua\/JIT, you can say <a href=\"http:\/\/www.lua.org\/manual\/5.1\/manual.html#pdf-debug.getinfo\" rel=\"nofollow noreferrer\"><code>debug.getinfo<\/code><\/a><code>( func )<\/code> and will get a table containing (among others) the fields <code>short_src<\/code>, <code>source<\/code> and <code>linedefined<\/code>.<\/p>\n\n<p>For Lua functions, <code>short_src<\/code> will be the filename or <code>stdin<\/code> if it was defined in the REPL. (<code>source<\/code> has a slightly different format, filenames are prefixed with an <code>@<\/code>, a <code>=<\/code> prefix is used for C functions or stuff defined interactively, and for <code>load<\/code>ed functions, it will be the actual string that was loaded.)<\/p>\n\n<p>You can pack that up in a function like<\/p>\n\n<pre><code>function sourceof( f )\n    local info = debug.getinfo( f, \"S\" )\n    return info.short_src, info.linedefined\nend\n<\/code><\/pre>\n\n<p>or maybe even start an editor and point it there, e.g. (for vim)<\/p>\n\n<pre><code>function viewsource( f )\n    -- get info &amp; check it's actually from a file\n    local info = debug.getinfo( f, \"S\" )\n    local src, line = info.source, info.linedefined\n    if src == \"=[C]\"   then  return nil, \"Is a C function.\"  end\n    local path = src:match \"^@(.*)$\"\n    if path then\n        -- start vim (or an other editor if you adapt the format string)\n        return os.execute( (\"vim -fR %q +%d\"):format( path, line ) )\n    end\n    return nil, \"Was defined at run time.\"\nend\n<\/code><\/pre>\n\n<p>And just for fun, here's yet another version that returns the code if it can find it somewhere. (This will also work for functions that have been generated at run time, e.g. by calling <code>load<\/code>, and where no source file exists. You could also work in the other direction by dumping the <code>load<\/code>ed snippet into a temp file and opening that\u2026)<\/p>\n\n<pre><code>-- helper to extract the source block defining the function\nlocal function funclines( str, line1, lineN, filename )\n    -- if linedefined \/ lastlinedefined are 0, this is the main chunk's function\n    if line1 == 0 and lineN == 0 then\n        filename = filename and filename..\" (main chunk)\"\n                             or \"(chunk defined at runtime)\"\n        return \"-- \"..filename..\"\\n\"..str\n    end\n    -- add line info to file name or use placeholder\n    filename = filename and filename..\":\"..line1 or \"(defined at runtime)\"\n    -- get the source block\n    local phase, skip, grab = 1, line1-1, lineN-(line1-1)\n    local ostart, oend -- these will be the start\/end offsets\n    if skip == 0 then  phase, ostart = 2, 0  end -- starts at first line\n    for pos in str:gmatch \"\\n()\" do\n        if phase == 1 then -- find offset of linedefined\n            skip = skip - 1 ; if skip == 0 then  ostart, phase = pos, 2  end \n        else -- phase == 2, find offset of lastlinedefined+1\n            grab = grab - 1 ; if grab == 0 then  oend = pos-2 ; break  end\n        end\n    end\n    return \"-- \"..filename..\"\\n\"..str:sub( ostart, oend )\nend\n\nfunction dumpsource( f )\n    -- get info &amp; line numbers\n    local info = debug.getinfo( f, \"S\" )\n    local src, line, lastline = info.source, info.linedefined, info.lastlinedefined\n    -- can't do anything for a C function\n    if src == \"=[C]\" then  return nil, \"Is a C function.\"  end\n    if src == \"=stdin\" then  return nil, \"Was defined interactively.\"  end\n    -- for files, fetch the definition\n    local path = src:match \"^@(.*)$\"\n    if path then\n        local f = io.open( path )\n        local code = f:read '*a' \n        f:close( )\n        return funclines( code, line, lastline, path )\n    end\n    -- otherwise `load`ed, so `source`\/`src` _is_ the source\n    return funclines( src, line, lastline )\nend\n<\/code><\/pre>\n\n<p>A closing remark:  If you paste code into a Lua\/JIT REPL, <code>local<\/code>s disappear between definitions, because every line (or minimal complete group of lines) is its own chunk.  The common fix (that you probably know) is to wrap everything into a block as <code>do<\/code>*paste*<code>end<\/code>, but an alternative is to <code>load[[<\/code>*paste*<code>]]()<\/code> (possibly with more <code>=<\/code>s like <code>[===[<\/code> and <code>]===]<\/code>.)  If you paste this way, the above <code>dumpsource<\/code> (or any other function using <code>debug.getinfo<\/code>) will then be able to get the source of the function(s). This also means that if you defined a nice function but it's gone from the history and the scroll buffer, you can recover it in this way (<em>if<\/em> you defined it by <code>load<\/code>ing and not directly feeding the interpreter). Saving the source in a file will then also be possible without copy-pasting and not require editing out the <code>&gt;&gt;<\/code> prompts.<\/p>\n"}]}
{"QuestionId":29882271,"AnswerCount":1,"Tags":"<lua><gnuplot><torch>","CreationDate":"2015-04-26T19:13:37.800","AcceptedAnswerId":30022005.0,"OwnerUserId":1546029.0,"Title":"Lua Error: \"Gnuplot terminal is not set\"","Body":"<p>In LuaJIT or in the Torch REPL, I run the commands\n    \nand I get the error \"Gnuplot terminal is not set\".<\/p>\n\n<p>I tried using gnuplot.setterm() with some guesses such as 'x11' and 'qt' as arguments, but get the error \"gnuplot does not seem to have this term\". Is there somewhere I can get a list of terminal emulators\/graphics backends available to gnuplot? Or alternatively, are these errors indicative of some other problem?<\/p>\n","answers":[{"AnswerId":"30022005","CreationDate":"2015-05-04T03:04:19.540","ParentId":null,"OwnerUserId":"1546029","Title":null,"Body":"<p>It turns out that you get this error if you don't have the <a href=\"http:\/\/www.gnuplot.info\/download.html\" rel=\"nofollow\">Gnuplot executable<\/a> installed. <\/p>\n\n<p>I didn't check for this problem before because gnuplot.lua (v. 5.1) has an error check for the case of that executable being unavailable - on line 145 - but for some reason it failed to catch the problem. <\/p>\n"}]}
{"QuestionId":29893793,"AnswerCount":1,"Tags":"<conditional-statements><theano>","CreationDate":"2015-04-27T10:55:00.327","AcceptedAnswerId":null,"OwnerUserId":1215787.0,"Title":"Theano The number of values on the `then` branch should have the same number of variables as the `else` branch","Body":"<p>The title is the error message when I use ifelse in theano. <\/p>\n\n<p>Say, I have a python snippet like this that I want to wirte in theano:<\/p>\n\n<pre><\/pre>\n\n<p>Now if I write in theano:<\/p>\n\n<pre><\/pre>\n\n<p>if gave the error in the title. How to remedy this error, please ?<\/p>\n\n<p>Thank you <\/p>\n","answers":[{"AnswerId":"33180736","CreationDate":"2015-10-16T22:57:32.610","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p><code>ifelse<\/code> needs to evaluate each branch of the condition at compile, your error is because the list you return on the false condition contains zero theano variables but the <code>calculate()<\/code> function returns at least one. I believe there are other checks on the datatype returned, but can't remember off the top of my head.  For the code you posted, if <code>calculate<\/code> always returns a scaler value:<\/p>\n\n<p><code>rval = ifelse(condition, calculate(a,b,c), 0.)<\/code><\/p>\n\n<p>should compile<\/p>\n"}]}
{"QuestionId":29948507,"AnswerCount":1,"Tags":"<lua><luajit><torch>","CreationDate":"2015-04-29T15:41:29.960","AcceptedAnswerId":29949224.0,"OwnerUserId":1546029.0,"Title":"How to specify startup file for Torch REPL","Body":"<p>I'd like to define some command line convenience functions to be run every time I start the Torch REPL. For example,<\/p>\n\n<p><\/p>\n\n<p>and things of that nature. How can I have functions like this added to the namespace every time I start the REPL?<\/p>\n\n<p>I web searched \"luajit|torch|trepl startup|rc file\" but couldn't find any leads on this.<\/p>\n","answers":[{"AnswerId":"29949224","CreationDate":"2015-04-29T16:15:27.313","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>you can alias your th repl to take a default -l parameter:<\/p>\n\n<pre><code>alias thnew='th -lmyadditions '\n<\/code><\/pre>\n\n<p>where myadditions.lua is your file to be executed that is placed in your lua path.<\/p>\n"}]}
{"QuestionId":29951454,"AnswerCount":1,"Tags":"<jpeg><macports><libjpeg><libjpeg-turbo><torch>","CreationDate":"2015-04-29T18:17:17.947","AcceptedAnswerId":null,"OwnerUserId":3870045.0,"Title":"issues installing libjpeg for mac (macports)","Body":"<p>I use a macbook (with Yosemite). I am quite new to macports.<\/p>\n\n<p>However, i need to install libjpeg (so i can use torch7 fully), but get this error When i try:<\/p>\n\n<pre><\/pre>\n\n<p>result:<\/p>\n\n<pre><\/pre>\n\n<p>I dont know quite what to do? as i don't think i can uninstall jpeg without problems<\/p>\n","answers":[{"AnswerId":"29959632","CreationDate":"2015-04-30T05:25:14.020","ParentId":null,"OwnerUserId":"726106","Title":null,"Body":"<p>You can deactivate the conflicting port first with:<\/p>\n\n<pre><code>sudo port -f deactivate jpeg\nsudo port install libjpeg-turbo\n<\/code><\/pre>\n"}]}
{"QuestionId":29955356,"AnswerCount":1,"Tags":"<lua><luajit><torch>","CreationDate":"2015-04-29T22:00:18.860","AcceptedAnswerId":29955728.0,"OwnerUserId":1546029.0,"Title":"How to change working directory in Torch REPL","Body":"<p>Title says it all, how can one change the working directory inside the Torch REPL? I tried using calls to os.execute('cd some_dir') but this doesn't work, as demonstrated here.<\/p>\n\n<p>\n--prints: \/home\/user\/Code<br>\n\n--prints: true    exit    0<br>\n\n-- prints: \/home\/user\/Code<\/p>\n\n<p>where pwd() is a a convenience function that calls os.execute('pwd').<\/p>\n","answers":[{"AnswerId":"29955728","CreationDate":"2015-04-29T22:31:31.700","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Install the lfs package (probably already installed, if not \"luarocks install luafilesystem\")<\/p>\n\n<p>Then,<\/p>\n\n<pre><code>lfs=require 'lfs'\nlfs.chdir(newdir)\n<\/code><\/pre>\n\n<p>Also, in torch REPL, you can execute shell commands with a $ prefix\nExample:<\/p>\n\n<pre><code>th&gt; $ls\n<\/code><\/pre>\n"}]}
{"QuestionId":29962365,"AnswerCount":1,"Tags":"<python-2.7><anaconda><theano>","CreationDate":"2015-04-30T08:08:52.630","AcceptedAnswerId":null,"OwnerUserId":4786629.0,"Title":"Not being able to run Theano on windows XP 32 with Anaconda","Body":"<p>I\u2019m trying to run Theano on windows XP 32 with Anaconda (Python 2.7). I installed Theano following the steps provided in the Anaconda section here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a>. Everything went fine but when I try to run the script import theano I get this error message:<\/p>\n\n<p>ImportError: DLL load failed: The specified procedure could not be found. Struggling quite a bit, could anyone please help?<\/p>\n","answers":[{"AnswerId":"35789344","CreationDate":"2016-03-04T06:24:28.847","ParentId":null,"OwnerUserId":"1653637","Title":null,"Body":"<p>I had the same issue. Now I can use Theano on Windows XP 32bits with Anaconda.<\/p>\n\n<p>With dependency walker, we found when 'import theano', the error 'ImportError' caused by LIBSTDC++6.DLL(a g++ DLL)  cannot import the non-exists symbol 'vsnprintf' in 'MSVCRT.DLL'.<\/p>\n\n<p>The key to fix this issue is, replace the mingw installed by 'conda install mingw' (rev 4.7.0). Install mingw g++ (rev 4.9.3) with '<a href=\"https:\/\/sourceforge.net\/projects\/mingw\/\" rel=\"nofollow\">mingw-get-setup.exe<\/a>' from sourceforge instead. <\/p>\n\n<p>Make user you PATH environment has 'C:\\MinGW\\bin;C:\\MinGW\\lib;' (default install location) to use the correct g++.<\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":29971874,"AnswerCount":2,"Tags":"<python><loops><theano>","CreationDate":"2015-04-30T15:30:48.233","AcceptedAnswerId":null,"OwnerUserId":854585.0,"Title":"Python theano with index computed inside the loop","Body":"<p>I have installed the Theano library for increasing the speed of a computation, so that I can use the power of a GPU.<\/p>\n\n<p>However, inside the inner loop of the computation a new index is calculated, based on the loop index and corresponding values of a couple of arrays.<\/p>\n\n<p>That calculated index is then used to access an element of another array, which, in turn, is used for another calculation.<\/p>\n\n<p>Is this too complicated to expect any significant speedups from Theano?<\/p>\n\n<p>So let me rephrase my question, the other way round.\nHere is an example of GPU code snippet. Some initialisations are left out for reasons of brevity. Can I translate this to Python\/Theano without increasing computation times considerably?<\/p>\n\n<pre><\/pre>\n\n<p>{   <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30126362","CreationDate":"2015-05-08T14:32:53.797","ParentId":null,"OwnerUserId":"1318689","Title":null,"Body":"<p>GPUs aren't great at random access memory when working with their global memory.  I've not used Theano before but if your arrays all fit in local memory - this would be fast as random accesses aren't a problem there.  If it is global memory though it is hard to anticipate what performance would be but it would be a far cry from it's full power.  On another note, is something about this computation even parallelizable? GPUs only really do well when there's alot of these things going on concurrently.<\/p>\n"},{"AnswerId":"30501611","CreationDate":"2015-05-28T08:51:51.717","ParentId":null,"OwnerUserId":"2100935","Title":null,"Body":"<p>No, I see nothing which cannot be done using Tensors instead of a for-loop. This should mean that you might see an increase in speed, but this will really depend on the application. You have an overhead of python+theano as well, especially coming from c-like code.<\/p>\n\n<p>So, instead of<\/p>\n\n<pre><code>for (m = 0; m &lt; M; ++m)\n{\n    unsigned int ind2 = 3 * m;\n\n    float diff_x = x - some_pos[ind2];\n    float diff_y = y - some_pos[ind2 + 1];\n    float diff_z = z - some_pos[ind2 + 2];\n\n    float distance = sqrtf(diff_x * diff_x\n                         + diff_y * diff_y\n                         + diff_z * diff_z);\n\n    unsigned int dist = rintf(distance\/some_factor);\n    ind3 = m * another_factor + dist;\n\n    cuComplex some_element = data[ind3];\n}\n<\/code><\/pre>\n\n<p>You could do something like (of the top of my head)<\/p>\n\n<pre><code>diff_xyz = T.Tensor([x,y,z]).dimshuffle('x',0) - some_pos.reshape(-1,3)\ndistance = T.norm(diff_xyz)\ndist = T.round(distance\/some_factor)\ndata = data.reshape(another_factor,-1)\nsome_elements = data[:,dist]\n<\/code><\/pre>\n\n<p>See? No more loops, therefore a GPU can parallellize this.<\/p>\n\n<blockquote>\n  <p>However, inside the inner loop of the computation a new index is calculated, based on the loop index and corresponding values of a couple of arrays. (...) Is this too complicated to expect any significant speedups from Theano?<\/p>\n<\/blockquote>\n\n<p>In general: this can be optimized, as long as the loop index has a linear relation with the index needed, by using tensors instead of loops. It however needs a bit of creativity and massaging to get right.<\/p>\n\n<p>Non-linear relations are also possible using <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.take\" rel=\"nofollow\">Tensor.take()<\/a>, but I don't dare to vouch for its speed on GPU. My gut-feeling always told me to stay away from it, as it is probably too flexible to optimize nicely. However, it is possible to use when there are no alternatives.<\/p>\n"}]}
{"QuestionId":29985453,"AnswerCount":2,"Tags":"<python><ubuntu-14.04><keras>","CreationDate":"2015-05-01T10:40:45.597","AcceptedAnswerId":29985665.0,"OwnerUserId":4262897.0,"Title":"Linux error when installing Keras","Body":"<p>I am getting this strange to me error when installing Keras on an Ubuntu server:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas how to fix this issue?<\/p>\n\n<p>I've downloaded Keras repository from <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras<\/a>, and used this command to install it:<\/p>\n\n<pre><\/pre>\n\n<p>My Linux specifications are:<\/p>\n\n<ul>\n<li><em>Distributor ID:<\/em> Ubuntu<\/li>\n<li><em>Description:<\/em> Ubuntu 14.04.2 LTS<\/li>\n<li><em>Release:<\/em> 14.04<\/li>\n<li><em>Codename:<\/em> trusty<\/li>\n<\/ul>\n","answers":[{"AnswerId":"29985665","CreationDate":"2015-05-01T10:59:08.237","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>You need to install the <a href=\"https:\/\/launchpad.net\/ubuntu\/+source\/hdf5\" rel=\"nofollow\">hdf5<\/a> package to get the headers you need.<\/p>\n"},{"AnswerId":"42403835","CreationDate":"2017-02-22T22:53:30.930","ParentId":null,"OwnerUserId":"1605761","Title":null,"Body":"<p>Real Error is :<\/p>\n\n<blockquote>\n  <p>\"In file included from \/tmp\/easy_install-qQggXs\/h5py-2.5.0\/h5py\/defs.c:287:0:\n  \/tmp\/easy_install-qQggXs\/h5py-2.5.0\/h5py\/api_compat.h:27:18: fatal error: hdf5.h: No such file or directory\n   #include \"hdf5.h\" \"<\/p>\n<\/blockquote>\n\n<p><em>This error says that header file hdf5.h is missing.<\/em><\/p>\n\n<p>Run the following command to install header file:<\/p>\n\n<pre><code>sudo apt-get install libhdf5-dev\n<\/code><\/pre>\n\n<p>Please note that to install h5py package, run following command :<\/p>\n\n<pre><code>sudo pip install h5py\n<\/code><\/pre>\n\n<p>Hope this solves your problem<\/p>\n"}]}
{"QuestionId":29990173,"AnswerCount":0,"Tags":"<python><ubuntu-14.04><theano><deep-learning>","CreationDate":"2015-05-01T15:45:21.067","AcceptedAnswerId":null,"OwnerUserId":4262897.0,"Title":"Python keras, nolearn, import error","Body":"<p>I am getting this error when I am trying to import nolearn for theano or keras.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n\nWith Keras the same type.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I taught I had this sorted by this <a href=\"https:\/\/stackoverflow.com\/questions\/29985453\/linux-error-when-installing-keras?noredirect=1#comment48090541_29985453\">Linux error when installing Keras<\/a> but reality was different. <\/p>\n\n<p>Any ideas?<\/p>\n\n<p>This is import part of program.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":29993665,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-01T19:22:11.727","AcceptedAnswerId":33098347.0,"OwnerUserId":3979919.0,"Title":"Python theano.scan taps argument","Body":"<p>I'm trying desperately to understand the taps argument\nin the theano.scan function. Unfortunately I'm\nnot able to come up with a specific question.<\/p>\n\n<p>I just don't understand the \"taps\" mechanism.\nWell I ok. I know in which order the sequences\nare passed to  the function, but I don't know \nthe meaning. For example (I borrowed this code from\nanother Question <a href=\"https:\/\/stackoverflow.com\/questions\/26718812\/python-theano-scan-function\">Python - Theano scan() function<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>Setting taps to -1 does make sense to me. As far as I understand\nit's the same as not setting the taps value and the whole vector 'x0'\nis being passed to the addf function. x0 will then be added\nwith the \"step\" parameter (int 2 which will be broadcasted to the same size).\nIn the next iteration the result [4, 5] will be the input and so on\nwhich yields the following output: <\/p>\n\n<pre><\/pre>\n\n<p>Setting taps to -3 however yields the following output:<\/p>\n\n<pre><\/pre>\n\n<p>I don't have any explanation how the scan function creates this\noutput. Why is it just a list now?\nThe \"print(a1)\" turns out to be as expected <\/p>\n\n<pre><\/pre>\n\n<p>Although I know that this is the value that a1 should have, \nI do not know how to interpret it. What is the t-3 th  value \nof x0?\nThe theano documentation\ndoesn't seem to be all to detailed about the taps argument...\nso hopefully one of you guys will be.<\/p>\n\n<p>Thx<\/p>\n","answers":[{"AnswerId":"33098347","CreationDate":"2015-10-13T08:59:52.313","ParentId":null,"OwnerUserId":"642776","Title":null,"Body":"<p>To better understand the use of <code>taps<\/code>, you should first understand how <code>scan<\/code> uses the <code>outputs_info<\/code> argument altogether and how the provided values for it (<code>initial<\/code> to be exact) change the nature of the result.<\/p>\n\n<p><code>scan<\/code> expects you to provide the type of output you expect from this operation (unless of course you dont have any initial values to provide and simply mention <code>None<\/code>, in which case it will start the first round {<code>step<\/code>} and the output is not passed as a parameter to the <code>fn<\/code> in the successive rounds).<\/p>\n\n<p>So <code>scan<\/code> is used for iterative reduction over the provided <code>sequences<\/code>. This means that at <code>step<\/code> <em>n<\/em> (and with no <code>taps<\/code> specified for either <code>sequences<\/code> or <code>outputs_info<\/code>), the given <code>fn<\/code> will be applied to the <em>n<\/em>th elements of each of the <code>sequences<\/code> along with the output(s) generated by the previous (<em>n-1<\/em> th) <code>step<\/code>. Hence the default value of <code>taps<\/code> for <code>sequences<\/code> is <code>0<\/code> and for <code>outputs_info<\/code> is <code>-1<\/code>.<\/p>\n\n<p>Another way to look at it would be to consider all the sequences to consist of slices across their respective first dimension. So for a particular step, the current slice(s) of the <code>sequence(s)<\/code> and the output slice of the previous step are passed to <code>fn<\/code> and the computed output is added to the results as a new slice which would then be used for the next <code>step<\/code>. It is obvious that each of the output slices would be of the same shape. And if you are providing an initial slice as part of <code>outputs_info<\/code> then it should also have the same shape as that produced by the application of <code>fn<\/code>. In your example, if <code>output_info=[dict(initial=x0)]<\/code>, it would take <code>[2, 3]<\/code> as the first slice and use it for the first <code>step<\/code> as the argument <code>a1<\/code> to <code>addf<\/code>.<\/p>\n\n<p>But quite often in signal processing (and elsewhere) you need more than just the last data points in time as causal information. Here I have used time just as a way to represent <code>steps<\/code>. Anyway, this is where <code>taps<\/code> is useful and helps in indicating exactly which data points from the <code>sequences<\/code> and <code>results<\/code> have to be used for the current <code>step<\/code>. In your example, this means that for the current <code>step<\/code> the 3rd last output should be passed to <code>fn<\/code>.<\/p>\n\n<p>And this is where you need to be careful in describing <code>initial<\/code> for <code>outputs_info<\/code>. Because scan will first split the <code>initial<\/code> value into slices along the fist dimension. <strong>Then the first slice among this set of slices would be considered the earliest slice<\/strong> (3rd last in your example) <strong>required to compute the output of the first <code>step<\/code><\/strong>.<\/p>\n\n<p>Lets assume in your example, <code>taps=[-2]<\/code> and <code>input = [2, 3]<\/code>. In this case, scan will split the input into slices and use the first slice (the value 2 here) as the argument <code>a1<\/code> to <code>addf<\/code>. The resulting value 4 would be added to the output and for the next step, the slices would include [2, 3, 4] of which the value 3 is on the second last (-2) tap. And so on. However, with <code>taps=[-3]<\/code> and the same <code>input<\/code>, there is one value missing which is like saying that you had collected the values at times (t-3) and (t-2) but didnt collect the value at (t-1).<\/p>\n\n<p>So if you reckon your output to be of a certain shape, and you require multiple taps of the output beyond -1, then the value of <code>initial<\/code> should be a list of elements of the required output shape <strong>and<\/strong> have <strong>exactly<\/strong> as many such elements as would be required to retrieve the earliest slice.<\/p>\n\n<p>TLDR:\nIn your example, if you want to get 2d vectors as the result of each <code>step<\/code> and are using <code>taps=[-3]<\/code>, then <code>input<\/code> should be a list of 3 such 2d vectors. If you want to get single valued results, then <code>input<\/code> should be a list with 3 integers. A list with 2 integers does not make sense in this context at all. It would only make sense if <code>taps<\/code> is either -2 or -1 or <code>[-2, -1]<\/code>.<\/p>\n"}]}
{"QuestionId":29995397,"AnswerCount":2,"Tags":"<python><c++><windows><anaconda><theano>","CreationDate":"2015-05-01T21:27:34.567","AcceptedAnswerId":30019264.0,"OwnerUserId":4380945.0,"Title":"compilation failure when running theano - windows 8.1 64 bit with Anaconda python distribution","Body":"<p>I am running lasagne\/nolearn, which uses theano.<\/p>\n\n<p>It has been particularly difficult to install and compile theano. The following compilation error happens after installing a 64 bit g++ compiler.<\/p>\n\n<p>Help is much appreciated. Thanks!<\/p>\n\n<p>Problem occurred during compilation with the command line below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35199686","CreationDate":"2016-02-04T11:21:50.617","ParentId":null,"OwnerUserId":"868972","Title":null,"Body":"<p>For those using WinPython and mingw, here some additional information:<\/p>\n\n<p>1) Don't even bother trying the mingw32 package when using a 64bit Windows, immediately go for mingw64. That stops the above shown error from occuring<\/p>\n\n<p>2) For WinPython, the .theanorc or .theanorc.txt file must not be in your home directory, but in the WinPython\/settings directory!<\/p>\n\n<p>3) Make things easier by using linux style path separators such as, for the g++ flag, cxx=d:\/dev\/mingw-w64\/mingw64\/bin\/g++.exe<\/p>\n\n<p>4) nvcc needs the windows path to have included the cl.exe file, which can in general be found in you VS installation under something like (VSPATH)\/VC\/bin<\/p>\n\n<p>Regards,\nG.<\/p>\n"},{"AnswerId":"30019264","CreationDate":"2015-05-03T20:52:36.507","ParentId":null,"OwnerUserId":"4380945","Title":null,"Body":"<p>I found what the problem was and would like to post the solution. This particular problem was caused because the file libpythonxx.a file was missing in the same directory where you find the pythonxx.dll file (in my case python27.dll and thus I created the libpython27.a file.<\/p>\n\n<p>A noble soul posted all steps necessary to install theano at <a href=\"http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano\" rel=\"noreferrer\">http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano<\/a> (in Korean and English). To generate such file, you copy the pythonxx.dll file to a temporary directory and type the following commands in the windows console:<\/p>\n\n<p>gendef pythonXX.dll<\/p>\n\n<p>dlltool --as-flags=--64 -m i386:x86-64 -k --output-lib libpythonXX.a --input-def pythonXX.def<\/p>\n\n<p>Then you paste the generated libpythonxx.a file in the same directory as the pythonxx.dll file.<\/p>\n\n<p>In windows, you usually find this file under C:\\Windows\\System3 but if you are using anaconda as I am, you will find it under?<\/p>\n\n<p>C:\\Users\\xxxxx\\Anaconda\\libs, xxxxx being your user.<\/p>\n"}]}
{"QuestionId":30023117,"AnswerCount":2,"Tags":"<neural-network><torch>","CreationDate":"2015-05-04T05:25:07.223","AcceptedAnswerId":30036454.0,"OwnerUserId":3468673.0,"Title":"torch7 : print matrix in text file with lines longer than 80 characters","Body":"<p>I'm trying to recuperate the parameters of this NN:<\/p>\n\n<pre><\/pre>\n\n<p>using this code :<\/p>\n\n<pre><\/pre>\n\n<p>When saving the output .lua file into text file using this command line:<\/p>\n\n<pre><\/pre>\n\n<p>I get all weight matrix wrapped into segments of six columns each (Columns 1 to 6 ... Columns 193 to 198 ... Columns 199 to 200).<\/p>\n\n<p>Is there anyway to prevent text from being wrapped and displaying weight matrix in only one block?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"30025344","CreationDate":"2015-05-04T08:02:22.897","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>I think what you actually want to do is to save the parameters so that you can load them back in later? In that case, look at this:<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/serialization.md\" rel=\"nofollow\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/serialization.md<\/a><\/p>\n"},{"AnswerId":"30036454","CreationDate":"2015-05-04T17:42:30.233","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>printT = function(t)\n   t = t:view(-1)\n   for i=1,t:nElement() do\n      io.write(t[i] .. ',')\n   end\nend\n\nprintT(mlp:get(1).weight)\nprintT(mlp:get(1).bias)\n\nprintT(mlp:get(3).weight)\nprintT(mlp:get(3).bias)\n\nprintT(mlp:get(5).weight)\nprintT(mlp:get(5).bias)\n<\/code><\/pre>\n"}]}
{"QuestionId":30033096,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2015-05-04T14:47:04.150","AcceptedAnswerId":33711600.0,"OwnerUserId":562769.0,"Title":"What is `lr_policy` in Caffe?","Body":"<p>I just try to find out how I can use <a href=\"http:\/\/caffe.berkeleyvision.org\/\">Caffe<\/a>. To do so, I just took a look at the different  files in the examples folder. There is one option I don't understand:<\/p>\n\n<pre><\/pre>\n\n<p>Possible values seem to be:<\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<li><\/li>\n<li><\/li>\n<li><\/li>\n<li>  <\/li>\n<\/ul>\n\n<p>Could somebody please explain those options?<\/p>\n","answers":[{"AnswerId":"30045244","CreationDate":"2015-05-05T05:55:11.887","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It is a common practice to decrease the learning rate (lr) as the optimization\/learning process progresses. However, it is not clear how exactly the learning rate should be decreased as a function of the iteration number.<\/p>\n\n<p>If you use <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\" rel=\"noreferrer\">DIGITS<\/a> as an interface to Caffe, you will be able to visually see how the different choices affect the learning rate.<\/p>\n\n<p><strong>fixed:<\/strong> the learning rate is kept fixed throughout the learning process.<\/p>\n\n<hr>\n\n<p><strong>inv:<\/strong> the learning rate is decaying as ~<code>1\/T<\/code><br>\n<img src=\"https:\/\/i.stack.imgur.com\/LScLY.png\" alt=\"enter image description here\"><\/p>\n\n<hr>\n\n<p><strong>step:<\/strong> the learning rate is piecewise constant, dropping every X iterations<br>\n<img src=\"https:\/\/i.stack.imgur.com\/W5h6j.png\" alt=\"enter image description here\"> <\/p>\n\n<hr>\n\n<p><strong>multistep:<\/strong> piecewise constant at arbitrary intervals<br>\n<img src=\"https:\/\/i.stack.imgur.com\/DW0qa.png\" alt=\"enter image description here\"><\/p>\n\n<hr>\n\n<p>You can see exactly how the learning rate is computed in the function <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/solvers\/sgd_solver.cpp#L27\" rel=\"noreferrer\"><code>SGDSolver&lt;Dtype&gt;::GetLearningRate<\/code><\/a> (<em>solvers\/sgd_solver.cpp<\/em> line ~30).<\/p>\n\n<hr>\n\n<p>Recently, I came across an interesting and unconventional approach to learning-rate tuning: <a href=\"http:\/\/arxiv.org\/abs\/1506.01186\" rel=\"noreferrer\">Leslie N. Smith's work \"No More Pesky Learning Rate Guessing Games\"<\/a>. In his report, Leslie suggests to use <code>lr_policy<\/code> that alternates between decreasing and <em>increasing<\/em> the learning rate. His work also suggests how to implement this policy in Caffe.<\/p>\n"},{"AnswerId":"33711600","CreationDate":"2015-11-14T18:04:27.833","ParentId":null,"OwnerUserId":"2211907","Title":null,"Body":"<p>If you look inside the <code>\/caffe-master\/src\/caffe\/proto\/caffe.proto<\/code> file (you can find it online <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L157-L172\">here<\/a>) you will see the following descriptions:<\/p>\n\n<pre><code>\/\/ The learning rate decay policy. The currently implemented learning rate\n\/\/ policies are as follows:\n\/\/    - fixed: always return base_lr.\n\/\/    - step: return base_lr * gamma ^ (floor(iter \/ step))\n\/\/    - exp: return base_lr * gamma ^ iter\n\/\/    - inv: return base_lr * (1 + gamma * iter) ^ (- power)\n\/\/    - multistep: similar to step but it allows non uniform steps defined by\n\/\/      stepvalue\n\/\/    - poly: the effective learning rate follows a polynomial decay, to be\n\/\/      zero by the max_iter. return base_lr (1 - iter\/max_iter) ^ (power)\n\/\/    - sigmoid: the effective learning rate follows a sigmod decay\n\/\/      return base_lr ( 1\/(1 + exp(-gamma * (iter - stepsize))))\n\/\/\n\/\/ where base_lr, max_iter, gamma, step, stepvalue and power are defined\n\/\/ in the solver parameter protocol buffer, and iter is the current iteration.\n<\/code><\/pre>\n"}]}
{"QuestionId":30034492,"AnswerCount":1,"Tags":"<python><python-3.x><theano>","CreationDate":"2015-05-04T15:58:06.107","AcceptedAnswerId":30177516.0,"OwnerUserId":1311255.0,"Title":"does nolearn\/lasagne support python 3","Body":"<p>I am working with Neural Net implementation in  as mentioned <a href=\"http:\/\/nbviewer.ipython.org\/github\/ottogroup\/kaggle\/blob\/master\/Otto_Group_Competition.ipynb\">here<\/a><br>\nHowever I get the following error:<br>\n<br>\nI figure out that  is  in python-3<br>\nDoes nolearn\/lasagne support python-3 ? If not, is there any workaround ?<\/p>\n","answers":[{"AnswerId":"30177516","CreationDate":"2015-05-11T20:57:51.337","ParentId":null,"OwnerUserId":"576676","Title":null,"Body":"<p>You seem to be using an older version of nolearn.  Try the current master from Github with these commands:<\/p>\n\n<p><code>pip uninstall nolearn\npip install https:\/\/github.com\/dnouri\/nolearn\/archive\/master.zip#egg=nolearn\n<\/code><\/p>\n\n<p>Here's the tests in master running with both Python 2.7 and 3.4: <a href=\"https:\/\/travis-ci.org\/dnouri\/nolearn\/builds\/61806852\" rel=\"noreferrer\">https:\/\/travis-ci.org\/dnouri\/nolearn\/builds\/61806852<\/a><\/p>\n"}]}
{"QuestionId":30035581,"AnswerCount":2,"Tags":"<macos><caffe>","CreationDate":"2015-05-04T16:55:07.570","AcceptedAnswerId":30051717.0,"OwnerUserId":1348187.0,"Title":"Installing Caffe (deep learning) issues","Body":"<p>I have been able to install Caffe but I had a lot of issues and that's because I didn't follow the instructions very well.<\/p>\n\n<p>I have a Mac OSx and I'm reading the OSx guide for installation.<\/p>\n\n<p>In this point:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/07Akr.png\" alt=\"enter image description here\"><\/p>\n\n<p>when I type  I get: <\/p>\n\n<blockquote>\n  <p>\"hdf5: command not found\"<\/p>\n<\/blockquote>\n\n<p>I've tried to install  by  and  but I'm still getting:<\/p>\n\n<blockquote>\n  <p>\"hdf5: command not found\"<\/p>\n<\/blockquote>\n\n<p>Does anyone have any clue? <\/p>\n\n<p>Thank you very much.<\/p>\n\n<hr>\n\n<p>according to the answer of @mattias, my binaries in  are:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/YVQEk.png\" alt=\"enter image description here\"><\/p>\n","answers":[{"AnswerId":"30036375","CreationDate":"2015-05-04T17:38:04.387","ParentId":null,"OwnerUserId":"1197616","Title":null,"Body":"<p>You can install hdf5 from source. I just tested on OS X 10.9.5.<\/p>\n\n<pre><code>wget http:\/\/www.hdfgroup.org\/ftp\/HDF5\/current\/src\/hdf5-1.8.14.tar\n<\/code><\/pre>\n\n<p>Unpack,<\/p>\n\n<pre><code>tar zxfv hdf5-1.8.14.tar\n<\/code><\/pre>\n\n<p>Enter directory<\/p>\n\n<pre><code>cd hdf5-1.8.14\n<\/code><\/pre>\n\n<p>And then,<\/p>\n\n<pre><code>.\/configure --prefix=\/usr\/local\/hdf5  # or where you want it\nmake\nsudo make install\n<\/code><\/pre>\n\n<p>Then you have it installed in \/usr\/local\/hdf5.<\/p>\n\n<p>Good luck!<\/p>\n"},{"AnswerId":"30051717","CreationDate":"2015-05-05T11:26:51.103","ParentId":null,"OwnerUserId":"1348187","Title":null,"Body":"<p><code>hdf5<\/code> is not a command or anything else. The documentation is just bad, it has to be: <\/p>\n\n<p><code>brew tap homebrew\/science hdf5 opencv<\/code><\/p>\n\n<hr>\n\n<p>So, what I mean is, we have to install <code>hdf5<\/code> and then link it to Caffe. But executing hdf5 is not what the guide meant.<\/p>\n"}]}
{"QuestionId":30045306,"AnswerCount":2,"Tags":"<neural-network><torch>","CreationDate":"2015-05-05T06:00:09.663","AcceptedAnswerId":30053109.0,"OwnerUserId":3468673.0,"Title":"torch7 : how to connect the neurons of the same layer?","Body":"<p>Is it possible to implement, using torch, an architecture that connects the neurons of the same layer?<\/p>\n","answers":[{"AnswerId":"30053109","CreationDate":"2015-05-05T12:35:15.047","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>What you describe is called a recurrent neural network. Note that it needs quite different type of structure, input data, and training algorithms to work well.<\/p>\n\n<p>There is the <a href=\"https:\/\/github.com\/Element-Research\/rnn\" rel=\"nofollow\"><strong>rnn<\/strong><\/a> library for Torch to work with recurrent neural networks.<\/p>\n"},{"AnswerId":"30050671","CreationDate":"2015-05-05T10:37:46.820","ParentId":null,"OwnerUserId":"677824","Title":null,"Body":"<p>Yes, it's possible. Torch has everything that other languages have: logical operations, reading\/writing operations, array operations. That's all what needed for implementing any kind of neural network. If to take into account that torch has usage of CUDA you can even implement neural network which can work faster then some C# or java implementations. Performance improvement can depend from number of if\/else during one iteration<\/p>\n"}]}
{"QuestionId":30055969,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2015-05-05T14:36:27.167","AcceptedAnswerId":null,"OwnerUserId":4866786.0,"Title":"How to handle shape of pymc3 Deterministic variables","Body":"<p>I've been working on getting a hierarchical model of some psychophysical behavioral data up and running in pymc3. I'm incredibly impressed with things overall, but after trying to get up to speed with Theano and pymc3 I have a model that mostly works, however has a couple problems. <\/p>\n\n<p>The code is built to fit a parameterized version of a Weibull to seven sets of data. Each trial is modeled as a binary Bernoulli outcome, while the thresholds (output of thact as the y values which are used to fit a Gaussian function for height, width, and elevation (a, c, and d on a typical Gaussian). <\/p>\n\n<p>Using the parameterized Weibull seems to be working nicely, and is now hierarchical for the slope of the Weibull while the thresholds are fit separately for each chunk of data. However - the output I'm getting from k and y_est leads me to believe they may not be the correct size, and unlike the probability distributions, it doesn't look like I can specify shape (unless there's a theano way to do this that I haven't found - though from what I've read specifying shape in theano is tricky). <\/p>\n\n<p>Ultimately, I'd like to  use y_est to estimate the gaussian height or width, however the output right now results in an incredible mess that I think originates with size problems in y_est and k. Any help would be fantastic - the code below should simulate some data and is followed by the model. The model does a nice job fitting each individual threshold and getting the slopes, but falls apart when dealing with the rest. <\/p>\n\n<p>Thanks for having a look - I'm super impressed with pymc3 so far!<\/p>\n\n<p>EDIT: Okay, so the shape output by y_est.tag.test_value.shape looks like this<\/p>\n\n<pre><\/pre>\n\n<p>I think this is where I'm running into trouble, though it may just be poorly constructed on my part. k has the right shape (one k value per unique_xval). y_est is outputting an entire set of data (101x7) instead of a single estimate (one y_est per unique_xval) for each difficulty level. Is there some way to specify that y_est get specific subsets of df_y_vals to control this?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30130650","CreationDate":"2015-05-08T18:35:08.493","ParentId":null,"OwnerUserId":"359944","Title":null,"Body":"<p>I'm not sure exactly what you want to do when you say \"Is there some way to specify that y_est get specific subsets of df_y_vals to control this\". Can you describe for each y_est value what values of df_y_vals are you supposed to use? What's the shape of df_y_vals? What's the shape of y_est supposed to be? (7,)? <\/p>\n\n<p>I suspect what you want is to index into df_y_vals using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html\" rel=\"nofollow\">numpy advanced indexing<\/a>, which works the same in PyMC as in numpy. Its hard to say exactly without more information.<\/p>\n"}]}
{"QuestionId":30056139,"AnswerCount":1,"Tags":"<c><lua><neural-network><luajit><torch>","CreationDate":"2015-05-05T14:43:03.273","AcceptedAnswerId":30058467.0,"OwnerUserId":4866811.0,"Title":"How to read Torch Tensor from C","Body":"<p>I have to train a convolutional neural network using the Torch framework and then write the same network in C. \nTo do so, I have to read somehow the learned parameters of the net from my C program, but I can't find a way to convert or write to a file the Torch Tensors to make them readable in C. \nIdeally, I want to convert the Tensors into arrays of double in C. <\/p>\n\n<p>Does anyone know how to do that? Thanks in advance :)<\/p>\n","answers":[{"AnswerId":"30058467","CreationDate":"2015-05-05T16:29:50.343","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>I can't find a way to convert or write to a file the Torch Tensors to make them readable in C. Ideally, I want to convert the Tensors into arrays of double in C.<\/p>\n<\/blockquote>\n\n<p>The most basic (and direct) way is to directly <code>fread<\/code> in C the data you have previously written into a binary file. In such a case you would typically concatenate the weights and biases (if any) for each layer.<\/p>\n\n<p>On the Lua\/Torch side you can use the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/file.md\" rel=\"noreferrer\">File<\/a> utilities to literally <code>fwrite<\/code> each tensor data. For example here is a basic function that does that:<\/p>\n\n<pre><code>local fwrite = function(tensor, file)\n  if not tensor then return false end\n  local n = tensor:nElement()\n  local s = tensor:storage()\n  return assert(file:writeDouble(s) == n)\nend\n<\/code><\/pre>\n\n<p>For example if <code>m<\/code> refers to a <code>torch\/nn<\/code> module containing weights you would use it as follow:<\/p>\n\n<pre><code>local file = torch.DiskFile(\"net.bin\", \"w\"):binary()\nfwrite(m.weight, file)\nfwrite(m.bias, file)\n<\/code><\/pre>\n\n<p>Of course you need to write your own logic to make sure you <code>fwrite<\/code> and concatenate all the weights from all your layers. On the C side, in addition to <code>net.bin<\/code>, you also need to know the structure of your network (nb. layers, parameters like kernel size, etc) to know how many block of <code>double<\/code>-s to <code>fread<\/code>.<\/p>\n\n<p>As an example (in Lua) you can have a look at <a href=\"https:\/\/github.com\/jhjin\/overfeat-torch\" rel=\"noreferrer\">overfeat-torch<\/a> (non official project) that illustrates how to read such a plain binary file: see the <a href=\"https:\/\/github.com\/jhjin\/overfeat-torch\/blob\/09e10a6818ee8e079d923c63a9afc55d86ac4515\/run.lua#L83-L100\" rel=\"noreferrer\">ParamBank<\/a> tool.<\/p>\n\n<p>Keep in mind that a robust solution would consist in using a proper binary serialization format like <a href=\"http:\/\/msgpack.org\/\" rel=\"noreferrer\">msgpack<\/a> or <a href=\"https:\/\/developers.google.com\/protocol-buffers\/\" rel=\"noreferrer\">Protocol Buffers<\/a> that would make this export\/import process clean and portable.<\/p>\n\n<p>--<\/p>\n\n<p>Here is a toy example:<\/p>\n\n<pre><code>-- EXPORT\nrequire 'nn'\n\nlocal fwrite = function(tensor, file)\n  if not tensor then return false end\n  local n = tensor:nElement()\n  local s = tensor:storage()\n  return assert(file:writeDouble(s) == n)\nend\n\nlocal m = nn.Linear(2, 2)\n\nprint(m.weight)\nprint(m.bias)\n\nlocal file = torch.DiskFile(\"net.bin\", \"w\"):binary()\nfwrite(m.weight, file)\nfwrite(m.bias, file)\n<\/code><\/pre>\n\n<p>Then in C:<\/p>\n\n<pre><code>\/* IMPORT *\/\n#include &lt;stdio.h&gt;\n#include &lt;stdlib.h&gt;\n#include &lt;assert.h&gt;\n\nint\nmain(void)\n{\n  const int N = 2; \/* nb. neurons *\/\n\n  double *w = malloc(N*N*sizeof(*w)); \/* weights *\/\n  double *b = malloc(N*sizeof(*w));   \/* biases *\/\n\n  FILE *f = fopen(\"net.bin\", \"rb\");\n  assert(fread(w, sizeof(*w), N*N, f) == N*N);\n  assert(fread(b, sizeof(*w), N, f) == N);\n  fclose(f);\n\n  int i, j;\n  for (i = 0; i &lt; N; i++)\n    for (j = 0; j &lt; N; j++)\n      printf(\"w[%d,%d] = %f\\n\", i, j, w[N*i+j]);\n\n  for (i = 0; i &lt; N; i++)\n      printf(\"b[%d] = %f\\n\", i, b[i]);\n\n  free(w);\n  free(b);\n\n  return 0;\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":30074670,"AnswerCount":1,"Tags":"<python><matlab><numpy><matplotlib><theano>","CreationDate":"2015-05-06T10:53:28.913","AcceptedAnswerId":30074835.0,"OwnerUserId":4870118.0,"Title":"Why does matplotlib imshow() display a transposed image?","Body":"<p>I have an matrix of images created in matlab which i will be using as an input to a convolutional neural network i am coding in theano. I've imported the matrix using  and on inspection the matrix appears identical to the one created in matlab. When using  in matlab i get the images displayed corrected, however, when using matplotlib  the images are transposed. Does anyone know the cause of this?<\/p>\n\n<p>Code for matlab:<\/p>\n\n<pre><\/pre>\n\n<p>Code for matplotlib:<\/p>\n\n<pre><\/pre>\n\n<p>I would post the images but im new to stackoverflow and i don't have enough reputation yet.<\/p>\n\n<p>Cheers,<\/p>\n","answers":[{"AnswerId":"30074835","CreationDate":"2015-05-06T11:01:13.367","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Matlab and python has a different way to store arrays in memory. Matlab saves an array column-first, while python uses row-first method.<br>\nConsider, for example, a 2-by-2 matrix <\/p>\n\n<pre><code>M = [1, 2\n     3, 4]\n<\/code><\/pre>\n\n<p>In memory, matlab save the matrix as <code>[1 3 2 4]<\/code> while python's order is <code>[1 2 3 4]<\/code>. This effect causes your image to be transposed.<\/p>\n\n<p>Consider transposing the images in Matlab prior to saving them - this way the data is stored in memory in the same order as in python.<\/p>\n"}]}
{"QuestionId":30085991,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-06T19:36:51.620","AcceptedAnswerId":null,"OwnerUserId":887074.0,"Title":"How do I print the value of a theano variable whenever it is evaluated?","Body":"<p>I'm using lasagne and theano to build a convolutional neural network, and I'm having issues trying to follow the printdebugging examples in <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#how-do-i-step-through-a-compiled-function\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#how-do-i-step-through-a-compiled-function<\/a><\/p>\n\n<p>My function looks like this with G and Y being theano tensors<\/p>\n\n<pre><\/pre>\n\n<p>so the output ave_loss should be a symbolic expression that when compiled and executed with input data will result in computing the average loss over a batch of training images. <\/p>\n\n<p>What I want to do is put a symbolic print expression in here so that whenever the ave_loss is computed it prints the contents of G. <\/p>\n\n<p>But right now I'm stuck just trying to get something to print before and after <\/p>\n\n<pre><\/pre>\n\n<p>The above code does not work, and I'm not really sure how to manipulate theano.function to make it work. <\/p>\n\n<p>What I'm attempting to do is create an identity function that accepts G and returns G without modifying it, but prints pre_func and post_func along the way. <\/p>\n\n<p>How to you use theano.function (or theano.printing.Print) to acomplish this?<\/p>\n","answers":[{"AnswerId":"30157136","CreationDate":"2015-05-10T22:04:49.927","ParentId":null,"OwnerUserId":"3979919","Title":null,"Body":"<p>Unfortunately I can't help you with the printing approach since\nI have never used the print myself. But.. wouldn't it\nbe possible to return G together with the ave_loss.\nThen you can look at the contents...<\/p>\n\n<p>something like: <\/p>\n\n<pre><code>def loss_function(self, G, Y_):\n    ...\n    G = dbgfunc()\n    ...\n    return ave_loss, G\n\n\nG = T.matrix('G') \nY_ = T.matrix('Y')\n\nave_loss, G_prime = loss_function(G, Y_)\n\nf = function([G, _Y], [ave_loss, G_prime])\n\nprint( f(...) )\n<\/code><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>I just saw that the contents of <strong>G<\/strong> don't seem to be changing.\nWhy do you want to print it anyways. Since the print also prevents\nTheano from some optimisations if I remember correctly.<\/p>\n"}]}
{"QuestionId":30099876,"AnswerCount":1,"Tags":"<python><theano><theano-cuda>","CreationDate":"2015-05-07T11:34:07.420","AcceptedAnswerId":null,"OwnerUserId":3155703.0,"Title":"theano installation (windows 64bit python 2)","Body":"<p>I started an installation of theano.<\/p>\n\n<p>My computer SPEC:\n  - OS           : Windows 7 64bit\n  - Graphic card : NVIDIA Geforce GT 630\n  - CPU          : AMD FX-8120<\/p>\n\n<p>I installed theano by installation guide of deeplearning.net .\n(<a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#configure-theano-for-gpu-use\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#configure-theano-for-gpu-use<\/a>)<\/p>\n\n<p>I successfully finished installation process under.<\/p>\n\n<p>Visual Studio 2010 -> Windows Software Development Kit version 7.1 -> CUDA -> Microsoft Visual C++ Complier for Python 2.7 (adding header) -> TDM GCC -> WinPython-64bit-2.7.9.4 -> env.bat -> Theano setup<\/p>\n\n<p>When I create a test file(under) and test it, it successfully execute.<\/p>\n\n<p>-------test file----------------------------------<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>BUT when I add .theanorc.txt<\/strong> \n---.theanorc.txt--------------------------------<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>It gives me an error like this(under)<\/p>\n\n<p><a href=\"https:\/\/www.dropbox.com\/s\/gjspcpaz4hkeep8\/11.PNG?dl=0\" rel=\"nofollow\">https:\/\/www.dropbox.com\/s\/gjspcpaz4hkeep8\/11.PNG?dl=0<\/a><\/p>\n\n<p>I have no problem with CUDA-devicequery &amp; nvidia-smi.exe<\/p>\n\n<p>-------------------------DeviceQuery--------------------------<\/p>\n\n<pre><\/pre>\n\n<p>-------------------------nvidia-smi.exe--------------------------<\/p>\n\n<pre><\/pre>\n\n<p>Please help me..<\/p>\n","answers":[{"AnswerId":"30116189","CreationDate":"2015-05-08T05:06:17.800","ParentId":null,"OwnerUserId":"3140336","Title":null,"Body":"<p>You should use  a \".theanorc\" file, not a \".theanorc.txt\".<\/p>\n\n<p>You may also write your question on the theano-users group: <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/theano-users\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!forum\/theano-users<\/a><\/p>\n\n<p>Next Winpython release will include theano+mingwpy, so it may the reduce complexity of setup.<\/p>\n"}]}
{"QuestionId":30127415,"AnswerCount":1,"Tags":"<cuda><caffe>","CreationDate":"2015-05-08T15:24:08.933","AcceptedAnswerId":null,"OwnerUserId":4879546.0,"Title":"Caffe cuDNN R1 compile error","Body":"<p>I'm trying to compile Caffe with cuDNN-6.5-R1 enabled on Ubuntu 14.04. CUDA version is 7.0. <\/p>\n\n<p>I copied the cudnn.h header file in \/usr\/local\/cuda\/include and the cudnn libraries in \/usr\/local\/cuda\/lib64, then \"sudo ldconfig\", and when I do \"make all\", I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I checked other posts of people who had the same problem as me, but they were all using R2, not R1.<\/p>\n","answers":[{"AnswerId":"30139904","CreationDate":"2015-05-09T12:27:32.957","ParentId":null,"OwnerUserId":"562440","Title":null,"Body":"<p>For detailed instructions on how to install CUDA, cuDNN and Caffe take a look at \"<a href=\"http:\/\/www.joyofdata.de\/blog\/gpu-powered-deeplearning-with-nvidia-digits\/\" rel=\"nofollow\">GPU Powered DeepLearning with NVIDIA DIGITS on EC2<\/a>\".<\/p>\n\n<p>The reason why \"they were all using R2\" BTW is b\/c that is simply a requirement.<\/p>\n"}]}
{"QuestionId":30132036,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano><deep-learning>","CreationDate":"2015-05-08T20:05:58.653","AcceptedAnswerId":30508530.0,"OwnerUserId":588373.0,"Title":"Log Determinant in Theano Loss Function","Body":"<p>I'm using Theano (python package for deep learning), but I'm very new to it and I'm running into an issue with a term in my loss function. The term involves taking the logarithm of the determinant of a matrix; the matrix is a function of a layer of hidden units in my network.\nI import Tensor, and Tensor.nlinalg:<\/p>\n\n<pre><\/pre>\n\n<p>and then stick this term in my loss function:<\/p>\n\n<pre><\/pre>\n\n<p>but when I attempt to train it I get the following exception and traceback:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone offer any advice?\nCheers,\nMike<\/p>\n","answers":[{"AnswerId":"30508530","CreationDate":"2015-05-28T13:50:34.437","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>theano.tensor.nlinalg.Det<\/code> in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nlinalg.html\" rel=\"nofollow\">linear algebra package<\/a> is an operation class, not an operation function. You need to first init an instance of the class and then apply it to the node representing your matrix. For example,<\/p>\n\n<pre><code>import numpy\n\nimport theano\nimport theano.tensor.nlinalg\n\nx = theano.tensor.matrix('x', dtype=theano.config.floatX)\np = theano.shared(numpy.array([[2, 0], [0, 3]], dtype=theano.config.floatX))\ny = theano.dot(x, p)\nc = theano.tensor.log(theano.tensor.nlinalg.Det()(y))\ng = theano.grad(c, x)\n\nprint theano.printing.pp(g)\n<\/code><\/pre>\n\n<p>Note the difference between <code>theano.tensor.nlinalg.Det()(y)<\/code> and <code>theano.tensor.nlinalg.Det(y)<\/code>.<\/p>\n"}]}
{"QuestionId":30135101,"AnswerCount":2,"Tags":"<machine-learning><matrix-multiplication><theano><matrix-inverse><caffe>","CreationDate":"2015-05-09T01:46:44.290","AcceptedAnswerId":31756649.0,"OwnerUserId":3733814.0,"Title":"Whether to use Caffe or Theano for Moore-Penrose Pseudo inverse?","Body":"<p>I need to use(in an application) Extreme Learning Machine(ELM) which is highly optimised for multiple CPUs or GPUs. As ELM main computation involves Moore-Penrose Pseudo inverse and matrix multiplication, what would be the best option to implement ELM among Theano and Caffe ?<\/p>\n\n<p>Secondly, Is it possible to implement a new learning algorithm(ELM) in Caffe using its python interface ?<\/p>\n","answers":[{"AnswerId":"30139872","CreationDate":"2015-05-09T12:24:53.663","ParentId":null,"OwnerUserId":"562440","Title":null,"Body":"<p>As far as Google is concerned Caffe won't help you with \"Extreme Learning Machines\".<\/p>\n\n<blockquote>\n  <p>Secondly, Is it possible to implement a new learning algorithm(ELM) in Caffe using its python interface ?<\/p>\n<\/blockquote>\n\n<p>No, that is not possible. You will have to implement new layers and algorithms in C++. Afterwards you can deal with them via Python.<\/p>\n\n<p>For a primer on Caffe, check out <a href=\"http:\/\/www.joyofdata.de\/blog\/neural-networks-with-caffe-on-the-gpu\/\" rel=\"nofollow\">\"Neural Nets with Caffe Utilizing the GPU\"<\/a>.<\/p>\n"},{"AnswerId":"31756649","CreationDate":"2015-07-31T23:29:27.107","ParentId":null,"OwnerUserId":"1157452","Title":null,"Body":"<p>If I were you I would use Theano, not Caffe.\nCaffe is <strong>not<\/strong> programmed around a general-purpose matrix library so with Caffe  you would be trying to use a screwdriver to open a beer basically. \nIf you definitively feel like using C++ look into MrShadow or any other GPU-based matrix libraries.<\/p>\n\n<p>... or simply use Theano with Python.<\/p>\n\n<p>I'm not a big fan of Python and Theano takes some time to master but it's extremely convenient.  <\/p>\n\n<p>Also there are one or two ELM libraries for Python you can use as a reference, that's a huge plus when you need to test your own implementation.<\/p>\n\n<p><a href=\"https:\/\/github.com\/dclambert\/Python-ELM\" rel=\"nofollow\">https:\/\/github.com\/dclambert\/Python-ELM<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/acba\/elm\" rel=\"nofollow\">https:\/\/github.com\/acba\/elm<\/a><\/p>\n\n<p>I haven't used them so I can't elaborate on their status but something is better than nothing.<\/p>\n\n<p>You can also take a look at Keras and Lasagne, both are neural network libraries built on top of Theano. Just like Caffe, they will not help much with ELMs but they will get you started with Theano+nnets. Then all you have to do is create your own ELM layers.<\/p>\n"}]}
{"QuestionId":30143308,"AnswerCount":4,"Tags":"<macos><lua><torch>","CreationDate":"2015-05-09T18:00:03.727","AcceptedAnswerId":null,"OwnerUserId":3698971.0,"Title":"Torch \/ Lua after installation is not working","Body":"<p>I have followed the following approach in order to install Torch in my machine (Mac).<\/p>\n\n<p><a href=\"http:\/\/torch.ch\/docs\/getting-started.html#_\" rel=\"noreferrer\">http:\/\/torch.ch\/docs\/getting-started.html#_<\/a><\/p>\n\n<p>When I am done with the installation, I type:<\/p>\n\n<p>\nor <\/p>\n\n<p>or  <\/p>\n\n<p>in order to load the  or to make updates on the lua packages. It says  \"command not found\". Do you have any idea how I can resolve this issue? <\/p>\n","answers":[{"AnswerId":"32927986","CreationDate":"2015-10-03T22:08:01.333","ParentId":null,"OwnerUserId":"2479672","Title":null,"Body":"<p>Have you updated your PATH? It should include something like <\/p>\n\n<blockquote>\n  <p>\/home\/user\/torch\/install\/bin<\/p>\n<\/blockquote>\n"},{"AnswerId":"33497300","CreationDate":"2015-11-03T10:54:06.520","ParentId":null,"OwnerUserId":"3698971","Title":null,"Body":"<p>I have resolved the issue. I have deleted torch and I have installed it again. I have updated my PATH, and I have ran the <code>$ luarocks install image<\/code> command. After all of these, I was able to ran <code>$ th<\/code> command and in general torch.<\/p>\n"},{"AnswerId":"34711978","CreationDate":"2016-01-10T23:22:16.113","ParentId":null,"OwnerUserId":"5771204","Title":null,"Body":"<p>I faced the same issue and following this post deleted and reinstalled  everything. However in the end what helped was adding \/home\/user\/torch\/install\/bin\/ to the PATH variable. <\/p>\n"},{"AnswerId":"40624114","CreationDate":"2016-11-16T04:29:46.970","ParentId":null,"OwnerUserId":"4422034","Title":null,"Body":"<p>If you're on a <strong><em>Mac<\/em><\/strong> using the <strong><em>bash terminal<\/em><\/strong>, make sure that you've permanently added <code>\/Users\/you\/torch\/install\/bin<\/code> to your <strong>PATH<\/strong>.<\/p>\n\n<p>To do this:<\/p>\n\n<ol>\n<li><p>Navigate in your terminal to the root directory by running the command:<\/p>\n\n<pre><code>$ cd\n<\/code><\/pre><\/li>\n<li><p>Using the text editor of your choice (emacs, vim, etc.) open the <strong>.bash_profile<\/strong> file for editing. For example:<\/p>\n\n<pre><code>$ emacs .bash_profile\n<\/code><\/pre><\/li>\n<li><p>Add the following line to the end of the file (replacing 'you' with your Mac username):<\/p>\n\n<pre><code>PATH=$PATH\\:\/Users\/you\/torch\/install\/bin ; export PATH\n<\/code><\/pre><\/li>\n<li><p>Save and exit the text editor\n<br>\n<br><\/p><\/li>\n<li><p>Source the changes by running:<\/p>\n\n<pre><code>$ source .bash_profile\n<\/code><\/pre><\/li>\n<li><p>Check that your PATH has been updated (look for <code>\/Users\/you\/torch\/install\/bin<\/code> in the string returned):<\/p>\n\n<pre><code>$ echo $PATH\n<\/code><\/pre><\/li>\n<li><p>To make sure it has been changed permanently, completely quit Terminal, open it and run <code>echo $PATH again<\/code><\/p><\/li>\n<li><p>Now try <code>th<\/code> and it should run Torch!<\/p><\/li>\n<\/ol>\n\n<p><br>\nFor more help on PATH:\n<a href=\"https:\/\/kb.iu.edu\/d\/acar\" rel=\"noreferrer\">https:\/\/kb.iu.edu\/d\/acar<\/a><\/p>\n\n<p>The Torch installation (at least for me) added the line <code>. \/Users\/jb\/torch\/install\/bin\/torch-activate<\/code> to my <strong>.profile<\/strong> file, not <strong>.bash_profile<\/strong>. I tried adding that exact line to .bash_profile but it didn't work, so based on the recommendations here I got rid of the trailing directory and such.<\/p>\n"}]}
{"QuestionId":30170508,"AnswerCount":1,"Tags":"<compilation><cuda><nvcc><caffe>","CreationDate":"2015-05-11T14:35:27.553","AcceptedAnswerId":30186905.0,"OwnerUserId":3532255.0,"Title":"caffe Debug build: stray '\"' character in nvcc command","Body":"<p>I am trying to build my C++ application that uses caffe, in Debug Mode, VS2013 community, x64. To be able to build version that do not need cuda to run, I added to wrapped each .cu file as indicated below:<\/p>\n\n<pre><\/pre>\n\n<p>The project was built and ran fine in CPU_ONLY mode.\nUndefininig the CPU_ONLY flag, the project builds and runs OK in Release mode, but in Debug,\n I am getting the following error when trying to compile the *.cu files:<\/p>\n\n<pre><\/pre>\n\n<p>Where COMMAND is the nvcc compiler call command below, newlined for readability.<\/p>\n\n<pre><\/pre>\n\n<p>The project was able to build successfully in debug mode before adding the CPU_ONLY flags.\nAny ideas?<\/p>\n","answers":[{"AnswerId":"30186905","CreationDate":"2015-05-12T09:33:06.513","ParentId":null,"OwnerUserId":"3532255","Title":null,"Body":"<p>Turns out it was a typo.\nIn project properties->Debug->CUDA C\/C++->Device, instead of\n    <code>compute_30,sm_30<\/code><\/p>\n\n<p>I had<\/p>\n\n<pre><code>`compute_30, sm_30`\n<\/code><\/pre>\n\n<p>that is, with a space separator.<\/p>\n"}]}
{"QuestionId":30173144,"AnswerCount":1,"Tags":"<python><optimization><machine-learning><scipy><theano>","CreationDate":"2015-05-11T16:40:51.863","AcceptedAnswerId":30173316.0,"OwnerUserId":1544186.0,"Title":"SciPy Conjugate Gradient Optimisation not invoking callback method after each iteration","Body":"<p>I followed the tutorial <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_cg.py\" rel=\"nofollow\">here<\/a> in order to implement Logistic Regression using theano. The aforementioned tutorial uses SciPy's  optimisation procedure. Among the important argument to the aforementioned function are:  the object\/cost function to be minimised,  a user supplied initial guess of the parameters,  a function which provides the derivative of the function  at  and  an optional user-supplied function, called after each iteration.<\/p>\n\n<p>The training function is defined as follows:<\/p>\n\n<pre><\/pre>\n\n<p>What the above code does, is basically go through all the minibatches in the training dataset, for each minibatch calculate the average batch cost (i.e. the average of the cost function applied to each of the training samples in the minibatch) and averages the cost over all the batches. It might be worth pointing out that the cost for each individual batch is calculated by  -- a theano function.<\/p>\n\n<p>To me, it seems that the  function is being called arbitrarily, and not after every iteration as the documentation in SciPy claims.<\/p>\n\n<p>Here is the output I received after modifying  and  by adding \"train\" and \"callback\" prints respectively.<\/p>\n\n<pre><\/pre>\n\n<p>My question is, since each call to  is indeed a training epoch, how do I change the behaviour, so that a call to  is invoked after ?<\/p>\n","answers":[{"AnswerId":"30173316","CreationDate":"2015-05-11T16:50:45.197","ParentId":null,"OwnerUserId":"577088","Title":null,"Body":"<p>Each call to <code>train_fn<\/code> is <em>not<\/em> necessarily a single training epoch. I'm not exactly sure how <code>fmin_cg<\/code> is implemented, but in general, <a href=\"http:\/\/en.wikipedia.org\/wiki\/Conjugate_gradient_method\" rel=\"nofollow\">conjugate gradient methods<\/a> may call the cost or gradient function more than once per minimziation step. This is (as far as I understand it) required sometimes to find the conjugate vector relative to the previous step taken.<sup>1<\/sup><\/p>\n\n<p>So your callback is being called every time <code>fmin_cg<\/code> takes a <em>step<\/em>. If you need a function to be called every time the cost or gradient function is called, you can just put the call inside the relevant function.<\/p>\n\n<p><sup>\n1. Edit: At least when they are <em>nonlinear<\/em> methods, as <a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.15.1\/reference\/generated\/scipy.optimize.fmin_cg.html\" rel=\"nofollow\"><code>fmin_cg<\/code><\/a> is. The wikipedia page suggests that vanilla conjugate gradient (CG) methods may not require multiple calls, but I think they aren't as suitable for optimizing nonlinear functions. The CG code that I've seen -- which I guess must have been for nonlinear CG -- definitely involved at least one line search per step. That could certainly call for multiple evaluations of the gradient function.<\/sup><\/p>\n"}]}
{"QuestionId":30184994,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2015-05-12T07:58:11.140","AcceptedAnswerId":30510063.0,"OwnerUserId":1367788.0,"Title":"How can I change device used of theano","Body":"<p>I tried to change the device used in theano-based program. <\/p>\n\n<pre><\/pre>\n\n<p>However I got error<\/p>\n\n<pre><\/pre>\n\n<p>I wonder what is the best way of change gpu to gpu1 in code ?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"36279339","CreationDate":"2016-03-29T08:38:41.243","ParentId":null,"OwnerUserId":"679886","Title":null,"Body":"<p>Another possibility which worked for me was setting the environment variable in the process, before importing theano:<\/p>\n\n<pre><code>import os    \nos.environ['THEANO_FLAGS'] = \"device=gpu1\"    \nimport theano\n<\/code><\/pre>\n"},{"AnswerId":"33579079","CreationDate":"2015-11-07T04:18:40.533","ParentId":null,"OwnerUserId":"160698","Title":null,"Body":"<p>Remove the \"device\" config in .theanorc, then in your code:<\/p>\n\n<pre><code>import theano.sandbox.cuda\ntheano.sandbox.cuda.use(\"gpu0\")\n<\/code><\/pre>\n\n<p>It works for me.<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/woPgxXCEMB4\/l654PPpd5joJ\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/woPgxXCEMB4\/l654PPpd5joJ<\/a><\/p>\n"},{"AnswerId":"30510063","CreationDate":"2015-05-28T14:54:22.620","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There is no way to change this value in code running in the same process. The best you could do is to have a \"parent\" process that alters, for example, the <code>THEANO_FLAGS<\/code> environment variable and spawns children. However, the method of spawning will determine which environment the children operate in.<\/p>\n\n<p>Note also that there is no way to do this in a way that maintains a process's memory through the change. You can't start running on CPU, do some work with values stored in memory then change to running on GPU and continue running using the values still in memory from the earlier (CPU) stage of work. The process must be shutdown and restarted for a change of device to be applied.<\/p>\n\n<p>As soon as you <code>import theano<\/code> the device is fixed and cannot be changed within the process that did the import.<\/p>\n"}]}
{"QuestionId":30195837,"AnswerCount":3,"Tags":"<multithreading><machine-learning><neural-network><caffe><openblas>","CreationDate":"2015-05-12T15:52:00.420","AcceptedAnswerId":null,"OwnerUserId":2751512.0,"Title":"How to use multi CPU cores to train NNs using caffe and OpenBLAS","Body":"<p>I am learning deep learning recently and my friend recommended me caffe. After install it with OpenBLAS, I followed the tutorial, <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">MNIST task<\/a> in the doc. But later I found it was super slow and only one CPU core was working.<\/p>\n\n<p>The problem is that the servers in my lab don't have GPU, so I have to use CPUs instead.<\/p>\n\n<p>I Googled this and got some page like <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/80\" rel=\"nofollow\">this<\/a> . I tried to  and . But caffe still used one core.<\/p>\n\n<p>How can I make caffe use multi CPUs?<\/p>\n\n<p>Many thanks.<\/p>\n","answers":[{"AnswerId":"44172173","CreationDate":"2017-05-25T04:16:20.217","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>While building OpenBLAS, you have to set the flag USE_OPENMP = 1 to enable OpenMP support. Next set Caffe to use OpenBLAS in the Makefile.config, please export the number of threads you want to use during runtime by setting up OMP_NUM_THREADS=n where n is the number of threads you want.<\/p>\n"},{"AnswerId":"34971873","CreationDate":"2016-01-24T03:27:34.973","ParentId":null,"OwnerUserId":"2488716","Title":null,"Body":"<p>I found that this method works:<\/p>\n\n<p>When you build the caffe, in your make command, do use this for 8 cores: \n<code>make all -j8<\/code> and\n<code>make pycaffe -j8<\/code><\/p>\n\n<p>Also, make sure \n        <code>OPENBLAS_NUM_THREADS=8<\/code> \nis set.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/31395729\/how-to-enable-multithreading-with-caffe\">This<\/a> question has a full script for the same.<\/p>\n"},{"AnswerId":"37351642","CreationDate":"2016-05-20T16:34:31.593","ParentId":null,"OwnerUserId":"3701279","Title":null,"Body":"<p>@Karthik. That also works for me. One interesting discovery that I made was that using 4 threads reduces forward\/backward pass during the caffe timing test by a factor of 2. However, increasing the thread count to 8 or even 24 results in f\/b speed that is less than what I get with OPENBLAS_NUM_THREADS=4.\nHere are times for a few thread counts (tested on NetworkInNetwork model).<\/p>\n\n<p>[#threads] [f\/b time in ms]<br>\n1   223<br>\n2   150<br>\n4   113<br>\n8   125<br>\n12  144  <\/p>\n\n<p>For comparison, on a Titan X GPU the f\/b pass took 1.87 ms.<\/p>\n"}]}
{"QuestionId":30197510,"AnswerCount":1,"Tags":"<python><optimization><neural-network><theano>","CreationDate":"2015-05-12T17:17:46.437","AcceptedAnswerId":30222321.0,"OwnerUserId":1563927.0,"Title":"Theano: how to efficiently undo\/reverse max-pooling","Body":"<p>I'm using Theano 0.7 to create a <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">convolutional neural net<\/a> which uses <strong><a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#maxpooling\" rel=\"nofollow\">max-pooling<\/a><\/strong> (i.e. shrinking a matrix down by keeping only the local maxima).<\/p>\n\n<p>In order to \"undo\" or \"reverse\" the max-pooling step, one method is to store the locations of the maxima as auxiliary data, then simply recreate the un-pooled data by making a big array of zeros and using those auxiliary locations to place the maxima in their appropriate locations.<\/p>\n\n<p>Here's how I'm currently doing it:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p><em>(By the way, in this case I have a 3D tensor, and it's only the third axis that gets max-pooled. People who work with image data might expect to see two dimensions getting max-pooled.)<\/em><\/p>\n<\/blockquote>\n\n<p>The output is:<\/p>\n\n<pre><\/pre>\n\n<p>This method <strong>works<\/strong> but it's a <strong>bottleneck<\/strong>, taking most of my computer's time (I think the set_subtensor calls might imply cpu&lt;->gpu data copying). So: can this be implemented more efficiently?<\/p>\n\n<p>I suspect there's a way to express this as a single  call which may be faster, but I don't see how to get the tensor indexing to broadcast properly.<\/p>\n\n<hr>\n\n<p><strong>UPDATE:<\/strong> I thought of a way of doing it in one call, by working on the flattened tensors:<\/p>\n\n<pre><\/pre>\n\n<p>However, this is still not good efficiency-wise because when I run this (added on to the end of the above script) I find out that the Cuda libraries can't currently do the integer index manipulation efficiently:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30222321","CreationDate":"2015-05-13T18:10:49.343","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>I don't know whether this is faster, but it may be a little more concise. See if it is useful for your case.<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\nminibatchsize = 2\nnumfilters = 3\nnumsamples = 4\nupsampfactor = 5\n\ntotalitems = minibatchsize * numfilters * numsamples\n\ncode = np.arange(totalitems).reshape((minibatchsize, numfilters, numsamples))\n\nauxpos = np.arange(totalitems).reshape((minibatchsize, numfilters, numsamples)) % upsampfactor \nauxpos += (np.arange(4) * 5).reshape((1,1,-1))\n\n# first in numpy\nshp = code.shape\nupsampled_np = np.zeros((shp[0], shp[1], shp[2] * upsampfactor))\nupsampled_np[np.arange(shp[0]).reshape(-1, 1, 1), np.arange(shp[1]).reshape(1, -1, 1), auxpos] = code\n\nprint \"numpy output:\"\nprint upsampled_np\n\n# now the same idea in theano\nencoded = T.tensor3()\npositions = T.tensor3(dtype='int64')\nshp = encoded.shape\nupsampled = T.zeros((shp[0], shp[1], shp[2] * upsampfactor))\nupsampled = T.set_subtensor(upsampled[T.arange(shp[0]).reshape((-1, 1, 1)), T.arange(shp[1]).reshape((1, -1, 1)), positions], encoded)\n\nprint \"theano output:\"\nprint upsampled.eval({encoded: code, positions: auxpos})\n<\/code><\/pre>\n"}]}
{"QuestionId":30198926,"AnswerCount":2,"Tags":"<neural-network><computer-vision><deep-learning><caffe><image-segmentation>","CreationDate":"2015-05-12T18:36:33.810","AcceptedAnswerId":32453021.0,"OwnerUserId":562769.0,"Title":"Can Caffe classify pixels of an image directly?","Body":"<p>I would like to classify pixels of an image to \"is street\" or \"is not street\". I have some training data from the <a href=\"http:\/\/www.cvlibs.net\/datasets\/kitti\/eval_road.php\">KITTI dataset<\/a> and I have seen that Caffe has an <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html#images\"><\/a> layer type.\nThe labels are there in form of images of the same size as the input image.<\/p>\n\n<p>Besides Caffe, my first idea to solve this problem was by giving image patches around the pixel which should get classified (e.g. 20 pixels to the top \/ left \/ right \/ bottom, resulting in 41\u00d741=1681 features per pixel I want to classify.<br>\nHowever, if I could tell caffe how to use the labels without having to create those image patches manually (and the layer type  seems to suggest that it is possible) I would prefer that.<\/p>\n\n<p>Can Caffe classify pixels of an image directly? How would such a prototxt network definition look like? How do I give Caffe the information about the labels?<\/p>\n\n<p>I guess the input layer would be something like<\/p>\n\n<pre><\/pre>\n\n<p>However, I am not sure what  exactly means. Is it really centered? How does caffe deal with the corner pixels? What is  and  good for?<\/p>\n","answers":[{"AnswerId":"30207142","CreationDate":"2015-05-13T06:23:57.330","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Can Caffe classify pixels? in theory I think the answer is Yes. I didn't try it myself, but I don't think there is anything stopping you from doing so.<\/p>\n\n<p><strong>Inputs:<\/strong><br>\nYou need two <code>IMAGE_DATA<\/code> layers: one that loads the RGB image and another that loads the <em>corresponding<\/em> label-mask image. Note that if you use <code>convert_imageset<\/code> utility you cannot shuffle each set independently - you won't be able to match an image to its label-mask. <\/p>\n\n<p>An <code>IMAGE_DATA<\/code> layer has two \"tops\" one for \"data\" and one for \"label\" I suggest you set the \"label\"s of both input layers to the index of the image\/label-mask and add a utility layer that verifies that the indices <em>always<\/em> matches, this will prevent you from training on the wrong label-masks ;)<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"ImageData\"\n  top: \"data\"\n  top: \"data-idx\"\n  # paramters...\n}\nlayer {\n  name: \"label-mask\"\n  type: \"ImageData\"\n  top: \"label-mask\"\n  top: \"label-idx\"\n  # paramters...\n}\nlayer {\n  name: \"assert-idx\"\n  type: \"EuclideanLoss\"\n  bottom: \"data-idx\"\n  bottom: \"label-idx\"\n  top: \"this-must-always-be-zero\"\n}\n<\/code><\/pre>\n\n<p><strong>Loss layer:<\/strong><br>\nNow, you can do whatever you like to the input data, but eventually to get pixel-wise labeling you need pixel-wise loss. Therefore, you must have your last layer (before the loss) produce a prediction with the <strong>same<\/strong> width and height as the <code>\"label-mask\"<\/code> Not all loss layers knows how to handle multiple labels, but <code>\"EuclideanLoss\"<\/code> (for example) can, therefore you should have a loss layer something like<\/p>\n\n<pre><code>layer {\n  name: \"loss\"\n  type: \"EuclideanLoss\"\n  bottom: \"prediction\" # size on image\n  bottom: \"label-mask\"\n  top: \"loss\"\n}\n<\/code><\/pre>\n\n<p>I think <code>\"SoftmaxWithLoss\"<\/code> has a newer version that can be used in this scenario, but you'll have to check it our yourself. In that case <code>\"prediction\"<\/code> should be of shape 2-by-h-by-w (since you have 2 labels).<\/p>\n\n<p><strong>Additional notes:<\/strong><br>\nOnce you set the input size in the parameters of the <code>\"ImageData\"<\/code> you fix the sizes of all blobs of the net. You must set the label size to the same size. You must carefully consider how you are going to deal with images of different shape and sizes.<\/p>\n"},{"AnswerId":"32453021","CreationDate":"2015-09-08T08:37:54.717","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Seems you can try <a href=\"http:\/\/www.cs.berkeley.edu\/~jonlong\/long_shelhamer_fcn.pdf\" rel=\"noreferrer\">fully convolutional networks for semantic segmentation<\/a><\/p>\n\n<p>Caffe was cited in this paper: <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Publications\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Publications<\/a><\/p>\n\n<p>Also here is the model:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo#fully-convolutional-semantic-segmentation-models-fcn-xs\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo#fully-convolutional-semantic-segmentation-models-fcn-xs<\/a><\/p>\n\n<p>Also this presentation can be helpfull:\n<a href=\"http:\/\/tutorial.caffe.berkeleyvision.org\/caffe-cvpr15-pixels.pdf\" rel=\"noreferrer\">http:\/\/tutorial.caffe.berkeleyvision.org\/caffe-cvpr15-pixels.pdf<\/a><\/p>\n"}]}
{"QuestionId":30200589,"AnswerCount":2,"Tags":"<python><c++><logging><caffe>","CreationDate":"2015-05-12T20:14:33.703","AcceptedAnswerId":null,"OwnerUserId":1452257.0,"Title":"Disable and renable logging created from C++ module in Python","Body":"<p>I'm using a deep learning library, Caffe, which is written in C++ and has an interface to Python. One of my commands creates a lot of unnecessary output to the log and I would really like to remove that by temporarily disabling logging.<\/p>\n\n<p>Caffe uses GLOG and I've tried using to only log important messages. However, that didn't work. I've also tried using the Python logging module to shut down all logging temporarily using the code below, which didn't work either.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31236798","CreationDate":"2015-07-06T01:51:04.470","ParentId":null,"OwnerUserId":"1170917","Title":null,"Body":"<p>You likely need to set the log level environmental variable before you start Python. Or at leastt this worked for me:<\/p>\n\n<p>GLOG_minloglevel=3 python script.py<\/p>\n\n<p>Which silenced loading messages.<\/p>\n"},{"AnswerId":"34735198","CreationDate":"2016-01-12T03:39:23.500","ParentId":null,"OwnerUserId":"4139842","Title":null,"Body":"<p><code>GLOG_minloglevel=3<\/code> ,only by executing that line in Python before calling<\/p>\n\n<p>so,you can try<\/p>\n\n<pre><code>os.environ[\"GLOG_minloglevel\"] =\"3\"\nimport caffe\n<\/code><\/pre>\n"}]}
{"QuestionId":30205464,"AnswerCount":1,"Tags":"<makefile><protocol-buffers><homebrew><caffe><matcaffe>","CreationDate":"2015-05-13T04:02:42.667","AcceptedAnswerId":null,"OwnerUserId":568145.0,"Title":"protobuf installed using brew but not found in build process","Body":"<p><strong>Background<\/strong><\/p>\n\n<p>Yesterday I built <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">Caffe<\/a> and had no problems with its dependencies.<\/p>\n\n<p>Today I had problems building the Caffe Matlab wrappers due to protobuf dependencies not being found. So I rebuilt Caffe:  followed by a , hoping that would fix the problem.<\/p>\n\n<p>Now the Caffe build is complaining about the protobuf dependency.<\/p>\n\n<p>The error output is given at the bottom of this question.<\/p>\n\n<p>Between the original (successful) build and the failed build, I needed to  to allow the Caffe python wrappers to import protobuf, as python complained about not being able to find the protobuf package. That was the only \"change\" involving protobuf prior to the failed Caffe rebuild.<\/p>\n\n<p>I have tried reinstalling protobuf using brew, but this did not help.<\/p>\n\n<p>So essentially the chronology of events relating to protobuf are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Whenever protobuf was not found,  showed that protobuf (2.6.1) was installed.<\/p>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>Can someone please explain why protobuf is not being found when it is clearly installed?<\/p>\n\n<p>What is particularly confusing is the fact that it was found initially (during the original, successful build) and now it is not being found despite following the same approach.<\/p>\n\n<p>Here is the error output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33409948","CreationDate":"2015-10-29T09:09:13.580","ParentId":null,"OwnerUserId":"3585651","Title":null,"Body":"<p>Did you build protobuf from source? I had a similar problem due to compiling protobuf and my project with differing versions of libc++. <\/p>\n\n<p>I solved it by adding c++11 to CXXFLAGS.<\/p>\n"}]}
{"QuestionId":30225633,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-05-13T21:39:27.067","AcceptedAnswerId":null,"OwnerUserId":3998122.0,"Title":"Cross Entropy for batch with Theano","Body":"<p>I am attempting to implement an RNN and have output predictions p_y of shape (batch_size, time_points, num_classes). I also have a target_output of shape (batch_size, time_points), where the value at a given index of target_output is an integer denoting the class (a value between 0 and num_classes-1). How can I index p_y with target_output to get the probabilities of the given class I need to compute Cross-Entropy?<\/p>\n\n<p>I'm not even sure how to do this in numpy. The expression p_y[target_output] does not give the desired results.<\/p>\n","answers":[{"AnswerId":"30434928","CreationDate":"2015-05-25T09:17:48.823","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You need to use advanced indexing (search for \"advanced indexing\" <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">here<\/a>). But Theano advanced indexing behaves differently to numpy so knowing how to do this in numpy may not be all that helpful!<\/p>\n\n<p>Here's a function which does this for my setup, but note that the order of my dimensions differs from yours. I use (time points, batch_size, num_classes). This also assumes you want to use the 1-of-N categorical cross-entropy variant. You may not want sequence length padding either.<\/p>\n\n<pre><code>def categorical_crossentropy_3d(coding_dist, true_dist, lengths):\n    # Zero out the false probabilities and sum the remaining true probabilities to remove the third dimension.\n    indexes = theano.tensor.arange(coding_dist.shape[2])\n    mask = theano.tensor.neq(indexes, true_dist.reshape((true_dist.shape[0], true_dist.shape[1], 1)))\n    predicted_probabilities = theano.tensor.set_subtensor(coding_dist[theano.tensor.nonzero(mask)], 0.).sum(axis=2)\n\n    # Pad short sequences with 1's (the pad locations are implicitly correct!)\n    indexes = theano.tensor.arange(predicted_probabilities.shape[0]).reshape((predicted_probabilities.shape[0], 1))\n    mask = indexes &gt;= lengths\n    predicted_probabilities = theano.tensor.set_subtensor(predicted_probabilities[theano.tensor.nonzero(mask)], 1.)\n\n    return -theano.tensor.log(predicted_probabilities)\n<\/code><\/pre>\n"}]}
{"QuestionId":30236070,"AnswerCount":2,"Tags":"<python><theano><pymc3>","CreationDate":"2015-05-14T11:21:04.037","AcceptedAnswerId":30241668.0,"OwnerUserId":3568242.0,"Title":"PyMC3 & Theano - Theano code that works stop working after pymc3 import","Body":"<p>Some simple theano code that works perfectly, stop working when I import pymc3<\/p>\n\n<p>Here some snipets in order to reproduce the error:<\/p>\n\n<pre><\/pre>\n\n<p>And I get the following error for each of the previous snippets:<\/p>\n\n<pre><\/pre>\n\n<p>Any Ideas ? \nThanks in advance<\/p>\n","answers":[{"AnswerId":"38948567","CreationDate":"2016-08-15T02:31:58.787","ParentId":null,"OwnerUserId":"1716733","Title":null,"Body":"<p>Solution proposed <a href=\"https:\/\/github.com\/snake-charmer-devs\/snake-charmer\/issues\/18\" rel=\"nofollow\">here<\/a> lasts a bit longer than setting the flag. In your shell type:<\/p>\n\n<pre><code>theano-cache purge\n<\/code><\/pre>\n"},{"AnswerId":"30241668","CreationDate":"2015-05-14T15:49:11.160","ParentId":null,"OwnerUserId":"2288595","Title":null,"Body":"<p>I think this is related to <code>pymc3<\/code> setting <code>theano.config.compute_test_value = 'raise'<\/code>: <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/model.py#L395\" rel=\"noreferrer\">https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/model.py#L395<\/a> <\/p>\n\n<p>You can explicitly set <code>theano.config.compute_test_value<\/code> back to <code>'ignore'<\/code> to get rid of the error.<\/p>\n"}]}
{"QuestionId":30247061,"AnswerCount":3,"Tags":"<python><convolution><theano>","CreationDate":"2015-05-14T20:47:34.070","AcceptedAnswerId":30247267.0,"OwnerUserId":4013571.0,"Title":"How can I get a 1D convolution in theano","Body":"<p>The only function I can find is for 2D convolutions <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/conv.html#theano.tensor.signal.conv.conv2d\" rel=\"noreferrer\">described here<\/a>...<\/p>\n\n<p>Is there any optimised 1D function ?<\/p>\n","answers":[{"AnswerId":"32147403","CreationDate":"2015-08-21T18:46:55.723","ParentId":null,"OwnerUserId":"742616","Title":null,"Body":"<p>Just to be a bit more specific, I found this to work nicely:<\/p>\n\n<pre><code>conv2d = T.signal.conv.conv2d\n\nx = T.dmatrix()\ny = T.dmatrix()\nveclen = x.shape[1]\n\nconv1d_expr = conv2d(x, y, image_shape=(1, veclen), border_mode='full')\n\nconv1d = theano.function([x, y], outputs=conv1d_expr)\n<\/code><\/pre>\n\n<p><code>border_mode = 'full'<\/code> is optional.<\/p>\n"},{"AnswerId":"30247127","CreationDate":"2015-05-14T20:51:58.577","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<p>It looks as though this is <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2028\" rel=\"nofollow\">in development<\/a>.\nI've realised I can use the <code>conv2d()<\/code> function by specifying either width or height as 1...<\/p>\n\n<p>For the function <code>conv2d()<\/code>, the parameter <code>image_shape<\/code> takes a list of length 4 containing:<\/p>\n\n<pre><code>([number_images,] height, width)\n<\/code><\/pre>\n\n<p>by setting <code>height=1<\/code> or <code>width=1<\/code> it forces it to a 1D convolution.<\/p>\n"},{"AnswerId":"30247267","CreationDate":"2015-05-14T21:01:27.147","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>While I believe there's no <code>conv1d<\/code> in theano, Lasagne (a neural network library on top of theano) has several implementations of Conv1D layer. Some are based on <code>conv2d<\/code> function of theano with one of the dimensions equal to 1, some use single or multiple dot products. I would try all of them, may be a dot-product based ones will perform better than <code>conv2d<\/code> with <code>width=1<\/code>.<\/p>\n\n<p><a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/theano_extensions\/conv.py\" rel=\"nofollow\">https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/theano_extensions\/conv.py<\/a><\/p>\n"}]}
{"QuestionId":30253520,"AnswerCount":1,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-05-15T07:05:59.953","AcceptedAnswerId":null,"OwnerUserId":562769.0,"Title":"What does 'Attempting to upgrade input file specified using deprecated transformation parameters' mean?","Body":"<p>I am currently trying to train my first net with Caffe. I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>What does <\/p>\n\n<blockquote>\n  <p> <\/p>\n<\/blockquote>\n\n<p>mean? Where exactly did I use something deprecated? What should I use instead? <\/p>\n","answers":[{"AnswerId":"30285502","CreationDate":"2015-05-17T09:41:34.193","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Recently, input transformation (scaling\/cropping etc.) was separated from the IMAGE_DATA layer into a separate object: data transformer. This change affected the protobuffer syntax and the syntax of the IMAGE_DATA layer.<\/p>\n\n<p>It appears as if your <code>first_net.prototxt<\/code> is in the old format and Caffe converts it for you to the new format. <\/p>\n\n<p>You can do this conversion manually yourself using <code>.\/build\/tools\/upgrade_net_proto_text<\/code> (for prototxt files) and <code>.\/build\/tools\/upgrade_net_proto_binary<\/code> (for binaryproto files).  <\/p>\n"}]}
{"QuestionId":30259911,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><matcaffe>","CreationDate":"2015-05-15T12:46:16.540","AcceptedAnswerId":30285844.0,"OwnerUserId":2191652.0,"Title":"Object categories of pretrained imagenet model in caffe","Body":"<p>I'm using the pretrained imagenet model provided along the caffe (CNN) library (). I can output a 1000 dim vector of object scores for any images using this model.<br>\nHowever I don't know what the actual object categories are. Did someone find a file, where the corresponding object categories are listed?<\/p>\n","answers":[{"AnswerId":"30285844","CreationDate":"2015-05-17T10:24:13.113","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You should look for the file <code>'synset_words.txt'<\/code> it has 1000 line each line provides a description of a different class.<\/p>\n\n<p>For more information on how to get this file (and some others you might need) you can read <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"nofollow\">this<\/a>.<\/p>\n\n<hr>\n\n<p>If you want all the labels to be ready-for-use in Matlab, you can read the txt file into a cell array (a cell per class):<\/p>\n\n<pre><code>C = textread('\/path\/to\/synset_words.txt','%s','delimiter','\\n');\n<\/code><\/pre>\n"}]}
{"QuestionId":30276809,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-05-16T14:22:59.913","AcceptedAnswerId":30277067.0,"OwnerUserId":1546029.0,"Title":"How to permanently add directory to Lua search path?","Body":"<p>What's a simple way to <em>permanently<\/em> add a directory to the Lua search path?<\/p>\n","answers":[{"AnswerId":"30277067","CreationDate":"2015-05-16T14:48:48.967","ParentId":null,"OwnerUserId":"107090","Title":null,"Body":"<p>Set the corresponding <a href=\"http:\/\/www.lua.org\/manual\/5.3\/#env\" rel=\"nofollow\">environment variable<\/a> or rebuild Lua after adding your path to the <a href=\"http:\/\/www.lua.org\/source\/5.3\/luaconf.h.html#LUA_PATH_DEFAULT\" rel=\"nofollow\">source<\/a>. <\/p>\n"}]}
{"QuestionId":30288203,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-05-17T14:58:13.070","AcceptedAnswerId":30288867.0,"OwnerUserId":1715101.0,"Title":"Torch tensor equivalent function to matlab's \"find\"?","Body":"<p>In a nutshell, I would like to know if there is a tensor command in torch that gives me the indices of elements in a tensor that satisfy a certain criteria.<\/p>\n\n<p>Here is matlab code that illustrates what I would like to be able to do in torch:<\/p>\n\n<pre><\/pre>\n\n<p>I understand that I could do this in torch using a for loop, but is there some equivalent to matlab's find command that would allow me to do this more compactly?<\/p>\n","answers":[{"AnswerId":"30288867","CreationDate":"2015-05-17T15:53:41.517","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>x[x:gt(5)] = 0\n<\/code><\/pre>\n\n<p>In general there are x:gt :lt :ge :le :eq<\/p>\n\n<p>There is also the general :apply function tha takes in an anonymous function and applies it to each element.<\/p>\n"}]}
{"QuestionId":30288443,"AnswerCount":1,"Tags":"<python><optimization><neural-network><theano>","CreationDate":"2015-05-17T15:18:50.870","AcceptedAnswerId":null,"OwnerUserId":1606150.0,"Title":"Theano. Flatten all the parameters for later optimization","Body":"<p>I opened the following issue with my problem:\n<a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2920\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2920<\/a><\/p>\n\n<p>Basically, I want to prepare all my model's parameters for future optimization. And I need to flatten and concatenate all my variables in order to find derivative later. But I am stuck.<\/p>\n\n<p>If somebody knows how to solve it, please, share your wisdom :)<\/p>\n","answers":[{"AnswerId":"30415477","CreationDate":"2015-05-23T17:12:04.140","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Parameters are usually Theano shared variables, not symbolic variables. The result of using <code>T.concatenate<\/code> is going to be a symbolic expression but I suspect you don't want that. Try just passing an ordinary Python list of parameters into <code>T.grad<\/code>.<\/p>\n\n<p>So try,<\/p>\n\n<pre><code>self.theta = [self.hidden_layer.theta, self.log_layer.theta]\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>grad_theta = T.grad(cost_var, classifier.theta)\n<\/code><\/pre>\n"}]}
{"QuestionId":30302520,"AnswerCount":1,"Tags":"<python><numpy><parallel-processing><theano>","CreationDate":"2015-05-18T11:52:29.460","AcceptedAnswerId":30302918.0,"OwnerUserId":738017.0,"Title":"Parallelize operations for each cell in a numpy array","Body":"<p>I am trying to figure out which is the best way to parallelize the execution of a single operation for each cell in a 2D numpy array.<\/p>\n\n<p>In particular, I need to do a bitwise operation for each cell in the array.<\/p>\n\n<p>This is what I do using a single  cycle:<\/p>\n\n<pre><\/pre>\n\n<p>I found a way to do the same above using the  method:<\/p>\n\n<pre><\/pre>\n\n<p>However, using vectorize doesn't seem to improve performance.<\/p>\n\n<p>I read about <em>numexpr<\/em> in <a href=\"https:\/\/stackoverflow.com\/a\/11460119\/738017\">this answer on StackOverflow<\/a>, where also <em>Theano<\/em> and <em>Cython<\/em> are cited. <em>Theano<\/em> in particular seems a good solution, but I cannot find examples that fit my case.<\/p>\n\n<p>So my question is: which is the best way to improve the above code, using parallelization and possibly GPU computation? May someone post some sample code to do this?<\/p>\n","answers":[{"AnswerId":"30302918","CreationDate":"2015-05-18T12:12:29.030","ParentId":null,"OwnerUserId":"4367286","Title":null,"Body":"<p>I am not familiar with bitwise operations but this here gives me the same result as your code and is vectorized. <\/p>\n\n<pre><code>import numpy as np\n\n# make sure it is a numpy.array\nv = np.array(v)\n\n# vectorized computation\nN = (v &gt;&gt; 7) &amp; 255\n<\/code><\/pre>\n"}]}
{"QuestionId":30305891,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-05-18T14:27:51.757","AcceptedAnswerId":30309565.0,"OwnerUserId":2213825.0,"Title":"Convert einsum computation to dot product to be used in Theano","Body":"<p>I have just recently learned about  and quickly became addicted to it. But it seems that  doesn't have an equivalent function so I need to convert my  code to theano somehow. How can I write the following computation in theano?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30309565","CreationDate":"2015-05-18T17:35:13.150","ParentId":null,"OwnerUserId":"110026","Title":null,"Body":"<p>You only need to rearrange your axes to get this to work:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; a = np.random.rand(3, 4, 5)\n&gt;&gt;&gt; b = np.random.rand(5, 6)\n&gt;&gt;&gt; np.allclose(np.einsum('ikj,jl-&gt;ikl', a, b), np.dot(a, b))\n<\/code><\/pre>\n\n<p>So with that in mind:<\/p>\n\n<pre><code>&gt;&gt;&gt; a = np.random.rand(3, 5, 4)\n&gt;&gt;&gt; b = np.random.rand(6, 5)\n&gt;&gt;&gt; out_ein = np.einsum('ijk,lj-&gt;ilk', a, b)\n&gt;&gt;&gt; out_dot = np.transpose(np.dot(np.transpose(a, (0, 2, 1)),\n...                               np.transpose(b, (1, 0))),\n...                        (0, 2, 1))\n&gt;&gt;&gt; np.allclose(out_ein, out_dot)\n<\/code><\/pre>\n"}]}
{"QuestionId":30325108,"AnswerCount":2,"Tags":"<c++><layer><deep-learning><caffe>","CreationDate":"2015-05-19T12:04:11.647","AcceptedAnswerId":30394784.0,"OwnerUserId":213615.0,"Title":"Caffe layer creation failure","Body":"<p>I'm trying to load in TEST phase a network configuration which has a memory data layer first and then a convolution layer. The MemoryData layer creation succeeds,\nBut the convolution layer's creation fails at the following location:<\/p>\n\n<pre><\/pre>\n\n<p>Printed error is: <\/p>\n\n<blockquote>\n  <p>F0519 14:54:12.494139 14504 layer_factory.hpp:77] Check failed:\n  registry.count(t ype) == 1 (0 vs. 1) Unknown layer type: Convolution\n  (known types: MemoryData)<\/p>\n<\/blockquote>\n\n<p>registry has one entry only, indeed with MemoryData.\nWhen stepping into the registry creation functions, it looks like it first (and last, since this is a singletone) called from <\/p>\n\n<pre><\/pre>\n\n<p>in memory_data_later.cpp.<\/p>\n\n<p>I see similar  calls for the other supported layers, but it looks like they are never called.\nHow could I solve it?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"30394784","CreationDate":"2015-05-22T10:50:11.197","ParentId":null,"OwnerUserId":"561794","Title":null,"Body":"<p>This error occurs when trying to link caffe statically to an executable. You need to pass extra linker flags to make sure that layer registration code gets included.<\/p>\n\n<p>If you are using cmake take a look at Targets.cmake:<\/p>\n\n<pre><code>###########################################################################################\n# Defines global Caffe_LINK flag, This flag is required to prevent linker from excluding\n# some objects which are not addressed directly but are registered via static constructors\nif(BUILD_SHARED_LIBS)\n  set(Caffe_LINK caffe)\nelse()\n  if(\"${CMAKE_CXX_COMPILER_ID}\" STREQUAL \"Clang\")\n    set(Caffe_LINK -Wl,-force_load caffe)\n  elseif(\"${CMAKE_CXX_COMPILER_ID}\" STREQUAL \"GNU\")\n    set(Caffe_LINK -Wl,--whole-archive caffe -Wl,--no-whole-archive)\n  endif()\nendif()\n<\/code><\/pre>\n\n<p>And then where you create your target:<\/p>\n\n<pre><code># target\nadd_executable(${name} ${source})\ntarget_link_libraries(${name} ${Caffe_LINK})\n<\/code><\/pre>\n\n<p>A quick solution would be to build and link caffe as a shared lib instead of static.<\/p>\n\n<p>Also see <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/Py6IwMQvtqo\/discussion\" rel=\"noreferrer\">this post<\/a>.<\/p>\n\n<p>Just to complete this for MSVC compilation on Windows:\nUse <a href=\"https:\/\/msdn.microsoft.com\/en-US\/library\/bxwfs976(v=vs.120).aspx\" rel=\"noreferrer\">\/OPT:NOREF<\/a> or <a href=\"https:\/\/msdn.microsoft.com\/en-us\/library\/2s3hwbhs(v=VS.100).aspx\" rel=\"noreferrer\">\/INCLUDE<\/a> linker options on the target executable or dll.<\/p>\n"},{"AnswerId":"32119735","CreationDate":"2015-08-20T13:34:15.130","ParentId":null,"OwnerUserId":"2559219","Title":null,"Body":"<p>Replace <code>-l$(PROJECT)<\/code> with <code>$(STATIC_LINK_COMMAND)<\/code> in your Makefile in the appropriate places, and remove the now unnecessary runtime load path: <code>-Wl,-rpath,$(ORIGIN)\/..\/lib<\/code>.  <\/p>\n"}]}
{"QuestionId":30330394,"AnswerCount":1,"Tags":"<command-line><homebrew><caffe>","CreationDate":"2015-05-19T15:51:27.600","AcceptedAnswerId":30389640.0,"OwnerUserId":2438538.0,"Title":"What does the `--fresh` option do in Brew?","Body":"<p>While following installation instructions (e.g., for <a href=\"http:\/\/caffe.berkeleyvision.org\/install_osx.html\" rel=\"nofollow\">caffe<\/a> for os x), I run into the  flag for <a href=\"http:\/\/brew.sh\/\" rel=\"nofollow\">homebrew<\/a>. For example,<\/p>\n\n<pre><\/pre>\n\n<p>However, I see no documentation about what  does, and I don't find it in the source code for homebrew. What does this flag do? (Or what did it used to do?)<\/p>\n","answers":[{"AnswerId":"30389640","CreationDate":"2015-05-22T06:20:14.090","ParentId":null,"OwnerUserId":"2438538","Title":null,"Body":"<p>I found an old <a href=\"https:\/\/github.com\/Homebrew\/homebrew\/issues\/26979\" rel=\"nofollow\">github issue<\/a> describing the behavior of <code>--fresh<\/code>.<\/p>\n\n<p>The flag was meant to ensure packages would be installed without any previously set compile-time options (like <code>--with-python<\/code>), but it was removed because it didn't do anything:<\/p>\n\n<pre><code>commit 64744646e9be93dd758ca5cf202c6605accf4deb\nAuthor: Jack Nagel &lt;jacknagel@gmail.com&gt;\nDate:   Sat Jul 5 19:28:15 2014 -0500\n\n    Remove remaining references to \"--fresh\"\n\n    This option was removed in 8cdf4d8ebf439eb9a9ffcaa0e455ced9459e1e41\n    because it did not do anything.\n<\/code><\/pre>\n"}]}
{"QuestionId":30331476,"AnswerCount":2,"Tags":"<matlab><caffe><matcaffe>","CreationDate":"2015-05-19T16:43:39.957","AcceptedAnswerId":30376249.0,"OwnerUserId":2191652.0,"Title":"How to adapt Caffe Matlab wrapper for a network trained on Mnist?","Body":"<p>I successfully trained my Caffe net on the mnist database following <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html<\/a><\/p>\n\n<p>Now I want to test the network with my own images using the Matlab wrapper.<\/p>\n\n<p>Therefore in \"matcaffe.m\" im loading the file \"lenet.prototxt\" which is not used for training but which seems to be suited for testing. It is  referencing a input size of 28 x 28 pixels:<\/p>\n\n<pre><\/pre>\n\n<p>Therefore I adapted the \"prepare_image\" function in \"matcaffe.m\" accordingly. It now looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>This converts the input image to a [1 x 1 x 28 x 28], 4dim, grayscale image. But still Matlab is complaining:<\/p>\n\n<pre><\/pre>\n\n<p>Does somebody have experience with testing the trained mnist net on his own data?<\/p>\n","answers":[{"AnswerId":"30376249","CreationDate":"2015-05-21T14:08:48.380","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>Finally I found the full solution:\nThis how to predict a digit of your own input image using the matcaffe.m (Matlab wrapper) for Caffe<\/p>\n\n<ol>\n<li>In \"matcaffe.m\": One has to reference the file \"caffe-master\/examples\/mnist\/lenet.prototxt\"<\/li>\n<li>Adapt the file \"lenet.prototxt\" as pointed out by mprat: Change the entry input_dim to  <code>input_dim: 1<\/code><\/li>\n<li>Use the follwing adaptation to the subfunction \"prepare_image\" in matcaffe.m:<\/li>\n<\/ol>\n\n<p>(Input can be an rgb image of any size)<\/p>\n\n<pre><code>function image = prepare_image(im)\n\nIMAGE_DIM = 28;\n\n% If input image is too big , is rgb and of type uint8:\n% -&gt; resize to fixed input size, single channel, type float\n\nim = rgb2gray(im);\nim = imresize(im, [IMAGE_DIM IMAGE_DIM], 'bilinear');\nim = single(im);\n\n% Caffe needs a 4D input matrix which has single precision\n% Data has to be scaled by 1\/256 = 0.00390625 (like during training)\n% In the second last line the image is beeing transposed!\nimages = zeros(1,1,IMAGE_DIM,IMAGE_DIM);\nimages(1,1,:,:) = 0.00390625*im';\nimages = single(images);\n<\/code><\/pre>\n"},{"AnswerId":"30331542","CreationDate":"2015-05-19T16:47:48.007","ParentId":null,"OwnerUserId":"2773607","Title":null,"Body":"<p>The reason you are having that error (input size does not match) is that the network prototxt is expecting a batch of 64 images. The lines<\/p>\n\n<pre><code>input_dim: 64\ninput_dim: 1\ninput_dim: 28\ninput_dim: 28\n<\/code><\/pre>\n\n<p>Mean that the network is expecting a batch of 64 grayscale, 28 by 28 images. If you keep all your MATLAB code the same and change that first line to <\/p>\n\n<pre><code>input_dim: 1\n<\/code><\/pre>\n\n<p>Your problem should go away.<\/p>\n"}]}
{"QuestionId":30356427,"AnswerCount":1,"Tags":"<python><ipython><caffe>","CreationDate":"2015-05-20T17:24:14.643","AcceptedAnswerId":null,"OwnerUserId":562769.0,"Title":"Why do I get 'ImportError: dynamic module does not define init function (PyInit__caffe)' in ipython notebook while it works for Python?","Body":"<p>I try to get the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/classification.ipynb\" rel=\"nofollow\"><\/a> from the <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow\"><\/a> to work. When I Python 2.7.8 via console, it  works. I can  and (after a few seconds) its just finished. No error message. No need to append something to the .<\/p>\n\n<p>When I start the example mentioned above and execute the first Python cell, I get an error. To make it simpler, I added a cell with only  which gives me:<\/p>\n\n<pre><\/pre>\n\n<p>What is the problem here? <\/p>\n\n<p>Do I eventually have the wrong version?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30886398","CreationDate":"2015-06-17T08:37:59.667","ParentId":null,"OwnerUserId":"4973198","Title":null,"Body":"<p>When you work with caffe in the ipython session. You should add caffe root folder to sys.path. From my experience, Ipython doesn't load PYTHONPATH variable like using Python in console.<\/p>\n"}]}
{"QuestionId":30360980,"AnswerCount":2,"Tags":"<lua><neural-network><deep-learning><torch>","CreationDate":"2015-05-20T21:48:49.370","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch, \"size mismatch\" in StochasticGradient function training","Body":"<p>I'm implementing a deep neural network in Torch7 with a dataset made of two torch.Tensor() objects.\nThe first is made of 12 elements (completeTable), the other one is made of 1 element (presentValue).\nEach dataset row is an array of these two tensors:<\/p>\n\n<pre><\/pre>\n\n<p>Everything works for the neural network training and testing.\nBut now I want to switch and use only half of the 12 elements of completeTable, that are only 6 elements (firstChromRegionProfile).<\/p>\n\n<pre><\/pre>\n\n<p>If I run the same neural network architecture with this new dataset, it does not work. <strong>It says that the trainer:train(dataset_firstChromRegion) function cannot work because of \"size mismatch\".<\/strong><\/p>\n\n<p>Here's my neural network function:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the error log:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30402775","CreationDate":"2015-05-22T17:38:57.957","ParentId":null,"OwnerUserId":"1082019","Title":null,"Body":"<p>I was surprisingly able to fix my problem, by eliminating the line: <\/p>\n\n<pre><code>act_function = nn.Tanh();\n<\/code><\/pre>\n\n<p>and consequently by replacing any occurrences of <code>act_function<\/code> with <code>nn.Tanh()<\/code><\/p>\n\n<p>I do not why, but know everything works...\nSo the lesson is: never assign an activation function to a variable (!?).<\/p>\n"},{"AnswerId":"30531799","CreationDate":"2015-05-29T14:10:41.920","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>All of your activation layers share the same <code>nn.Tanh()<\/code> object. That is the problem. Try something like this instead:<\/p>\n\n<pre><code>act_function = nn.Tanh\nperceptron:add( act_function() )\n<\/code><\/pre>\n\n<p>Why?<\/p>\n\n<p>To perform a backward propagation step, we have to compute a gradient of the layer w.r.t. its input. In our case: <\/p>\n\n<blockquote>\n  <p>tanh'(input) = 1 - <strong>tanh(input)<\/strong>^2<\/p>\n<\/blockquote>\n\n<p>One can notice that <strong>tanh(input)<\/strong> = <strong>output<\/strong> of the layer's forward step. You can store this output inside the layer and use it during backward pass to speed up training. This is exactly what happens inside <code>nn<\/code> library:<\/p>\n\n<pre><code>\/\/ torch\/nn\/generic\/Tanh.c\/Tanh_updateGradInput:\n\nfor(i = 0; i &lt; THTensor_(nElement)(gradInput); i++)\n    {\n        real z = ptr_output[i];\n        ptr_gradInput[i] = ptr_gradOutput[i] * (1. - z*z);\n    }\n<\/code><\/pre>\n\n<p>Output sizes of your activation layers don't match, so error occurs. Even if they did, it would lead to wrong result.<\/p>\n\n<p>Sorry about my English.<\/p>\n"}]}
{"QuestionId":30361755,"AnswerCount":1,"Tags":"<lua><openmp><torch>","CreationDate":"2015-05-20T22:56:27.843","AcceptedAnswerId":30368097.0,"OwnerUserId":219603.0,"Title":"How to disable omp in Torch nn package?","Body":"<p>Specifically I would like  to not use omp when the size of the input tensor is small. I have a small script to test the run time.<\/p>\n\n<pre class=\"lang-lua prettyprint-override\"><\/pre>\n\n<p>If  is 10, then my basic log softmax function run much faster:<\/p>\n\n<pre><\/pre>\n\n<p>But once  is 10,000,000, omp really helps a lot:<\/p>\n\n<pre><\/pre>\n\n<p>So I suspect that omp overhead is very high. If my code has to call log softmax several times with small inputs (says tensor size is only 3), it will cost too much time. Is there a way to manually disable omp usage in some cases (but not always)?<\/p>\n","answers":[{"AnswerId":"30368097","CreationDate":"2015-05-21T08:12:31.513","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>Is there a way to manually disable omp usage in some cases (but not always)?<\/p>\n<\/blockquote>\n\n<p>If you really want to do that one possibility is to use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/57ff61e\/utils.c#L196\" rel=\"nofollow\"><code>torch.setnumthreads<\/code><\/a> and <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/57ff61e\/utils.c#L197\" rel=\"nofollow\"><code>torch.getnumthreads<\/code><\/a> like that:<\/p>\n\n<pre><code>local nth = torch.getnumthreads()\ntorch.setnumthreads(1)\n-- do something\ntorch.setnumthreads(nth)\n<\/code><\/pre>\n\n<p>So you can monkey-patch <code>nn.LogSoftMax<\/code> as follow:<\/p>\n\n<pre><code>nn.LogSoftMax.updateOutput = function(self, input)\n  local nth = torch.getnumthreads()\n  torch.setnumthreads(1)\n  local out = input.nn.LogSoftMax_updateOutput(self, input)\n  torch.setnumthreads(nth)\n  return out\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":30362370,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-21T00:02:44.470","AcceptedAnswerId":null,"OwnerUserId":2822004.0,"Title":"What is the purpose of creating symbolic variables in Theano?","Body":"<p>In Theano, variables are written as 'symbols':<\/p>\n\n<pre><\/pre>\n\n<p>From reading the documentation, it is implied that the reason we create these symbols may be due to the fact that these variables are compiled into C code. But I'm not sure if this is the case, much less the only reason for using symbolic variables. <\/p>\n\n<p>What is the purpose of creating symbolic variables in Theano? What can they do that a, out of the box assignment in Python can't do?<\/p>\n","answers":[{"AnswerId":"30500909","CreationDate":"2015-05-28T08:18:13.377","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The Theano web site opens with<\/p>\n\n<blockquote>\n  <p>Theano is a Python library that allows you to define, optimize, and\n  evaluate mathematical expressions involving multi-dimensional arrays\n  efficiently.<\/p>\n<\/blockquote>\n\n<p>which seems like quite a good summary of what it does but perhaps not why it does it.<\/p>\n\n<p>One of the main features of Theano is its symbolic differentiation feature. That is, given a symbolic mathematical expression, Theano can automatically differentiate the expression with respect to some variable within the expression, i.e. it can automatically determine the gradient of the expression along some dimension(s) of interest.<\/p>\n\n<p>For example, if <code>y=x**2<\/code> (where <code>**<\/code> is the power operator) then the gradient of <code>y<\/code> with respect to <code>x<\/code> is <code>dy\/dx = 2*x<\/code>. Theano can do this automatically:<\/p>\n\n<pre><code>import theano\nimport theano.tensor\nx = theano.tensor.scalar('x')\ny = x ** 2\ntheano.printing.pp(y)\ndy_dx = theano.grad(y, x)\ntheano.printing.pp(dy_dx)\n<\/code><\/pre>\n\n<p>If you run this code you'll see the output<\/p>\n\n<pre><code>(x ** TensorConstant{2})\n((fill((x ** TensorConstant{2}), TensorConstant{1.0}) * TensorConstant{2}) * (x ** (TensorConstant{2} - TensorConstant{1})))\n<\/code><\/pre>\n\n<p>The <code>fill<\/code> operation just creates a tensor of the correct shape (a scalar in this case) filled with ones and <code>TensorConstant{a}<\/code> is just the number <code>a<\/code> so this can be simplified to<\/p>\n\n<pre><code>(x ** 2) == x ** 2\n((1 * 2) * (x ** (2 - 1))) == (2 * (x ** 1)) == 2 * x\n<\/code><\/pre>\n\n<p>As expected.<\/p>\n\n<p>Clearly this is not particularly helpful for such a simple mathematical expression, but now imagine you're constructing an arbitrarily large and complex mathematical expression for which the gradient is not so immediately obvious, as is often the case in neural network research.<\/p>\n\n<p>But there's more. Theano can do the above for expressions that involve operations over not just scalars but also vectors, matrices, or any other tensor.<\/p>\n\n<p>Besides symbolic differentiation, Theano's symbolic approach offers other significant benefits:<\/p>\n\n<ul>\n<li>Theano can compile symbolic mathematical expressions into executable code. It uses various techniques that can yield executable code that can run faster than plain Python code.<\/li>\n<li>Theano programs can often be switched between running on CPU and GPU with no code changes whatsoever.<\/li>\n<li>When running on CPU, Theano makes full use of numpy BLAS and OpenMP facilities, when available, to parallelize the execution over multiple CPU cores.<\/li>\n<li>When running on GPU, Theano makes full use of of the many cores on modern GPUs to parallelize the most costly operations (most notably matrix multiplication).<\/li>\n<li>Theano's compiler is an optimizing compiler in that it can change the expression (also known as the computation graph) using various transforms that maintain the semantics of the expression (it still computes the same results given the same inputs) but achieving various performance and, crucially, <a href=\"http:\/\/mathworld.wolfram.com\/NumericalStability.html\" rel=\"nofollow\">numerical stability<\/a> gains.<\/li>\n<\/ul>\n\n<p>Plain Python does not offer any of the above and using numpy alone only offers some of the parallelization features (via BLAS and OpenMP).<\/p>\n\n<p>More on this topic <a href=\"http:\/\/deeplearning.net\/software\/theano\/introduction.html\" rel=\"nofollow\">in the Theano documentation<\/a>.<\/p>\n\n<p>P.S. this is an expansion of @eickenberg brief comment to the question.<\/p>\n"}]}
{"QuestionId":30365455,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-21T05:43:51.617","AcceptedAnswerId":30366957.0,"OwnerUserId":4013571.0,"Title":"Python - theano.scan() - return function values without feeding back into the loop","Body":"<p>Is it possible to return values calculated within the scan function without feeding them back into the scan function.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><\/pre>\n\n<p>which gives<\/p>\n\n<pre><\/pre>\n\n<p>However, I would like to output both of the values in  without feeding the latter value back into the scan function:<\/p>\n\n<pre><\/pre>\n\n<p>to give something like:<\/p>\n\n<pre><\/pre>\n\n<p><em>(I would just like to note that this problem is trivial but I would like to use the idea for debugging scan functions and checking output values)<\/em><\/p>\n","answers":[{"AnswerId":"30366957","CreationDate":"2015-05-21T07:15:38.913","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>You need to specify the <code>outputs_info<\/code> to be <code>None<\/code> for the output which you don't want to feed back into <code>f<\/code>. For further information, see <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">the <code>scan<\/code> documentation.<\/a> Below is an example which should do what you want.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\ntheano.config.exception_verbosity='high'\ntheano.config.optimizer='None'\n\ndef f(seq_v, prev_v):\n    return seq_v*prev_v, seq_v+1\n\na = T.vector('a')\n\nini = T.constant(1, dtype=theano.config.floatX)\n\nresult, updates = theano.scan(fn=f,\n                              outputs_info=[ini,None],\n                              sequences=[a])\n\nfn  = theano.function(inputs=[a], outputs=result)\n\nA = np.arange(1,5, dtype=T.config.floatX)\nout = fn(A)\n\nprint('Values:\\nf:\\t{}'.format(out))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Values:\nf:  [array([  1.,   2.,   6.,  24.], dtype=float32), array([ 2.,  3.,  4.,  5.], dtype=float32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":30367443,"AnswerCount":1,"Tags":"<python><canopy><theano><scikits>","CreationDate":"2015-05-21T07:39:06.387","AcceptedAnswerId":null,"OwnerUserId":2991243.0,"Title":"Theano package of Python (Canopy) returns an error: cannot import name gof","Body":"<p>This this the error returns using   in  :<\/p>\n\n<pre><\/pre>\n\n<p>How can i solve this problem? I installed  and  using package manager and installed  developer version using  in windows 8.1 .<\/p>\n","answers":[{"AnswerId":"30367924","CreationDate":"2015-05-21T08:03:40.933","ParentId":null,"OwnerUserId":"531222","Title":null,"Body":"<p>You can read the history of this bug here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2406\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2406<\/a><\/p>\n\n<p>It says<\/p>\n\n<blockquote>\n  <p>The problem appears to be that the compilation is trying to link with\n  the static library libpython2.7.a, instead of the dynamic version of\n  the library (which should be something like libpython2.7.so).<\/p>\n<\/blockquote>\n\n<p>and later<\/p>\n\n<blockquote>\n  <p>Hey,thanks for your reply, I did remake python and do altinstall\n  ,solved this problem.<\/p>\n<\/blockquote>\n\n<p>You should rebuild Python and it'll be fine.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I think you should get rid of Canopy and install Python manually. Canopy seems to be a ready-built distribution of Python and some scientific packages and I guess Theano will not work in it at this moment.<\/p>\n\n<p>Google for instructions how to install python on windows (I never used windows systems), for example this one: <a href=\"http:\/\/www.anthonydebarros.com\/2014\/02\/16\/setting-up-python-in-windows-8-1\/\" rel=\"nofollow\">http:\/\/www.anthonydebarros.com\/2014\/02\/16\/setting-up-python-in-windows-8-1\/<\/a><\/p>\n\n<p>And then install all required packages via <code>pip<\/code>.<\/p>\n"}]}
{"QuestionId":30369287,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-21T09:09:00.943","AcceptedAnswerId":30423839.0,"OwnerUserId":2213825.0,"Title":"Theano is missing signal.conv module","Body":"<p>My theano doesn't have the  module<\/p>\n\n<pre><\/pre>\n\n<p>My theano version is '0.7.0'. I tried to upgrade doing  and it tells me that I am already up-to-date. How can I get the conv module?<\/p>\n\n<p>PS: I even updated to the dev version by doing  and still no  !!<\/p>\n\n<p>If I do  I get the path to the file  in the same folder I have the file conv.py and downsample.py I can sucessfuly call  but not <\/p>\n\n<p>---- Installing on a Virtualenv ----<\/p>\n\n<p>I tried to reproduce the error on a virtualenv:<\/p>\n\n<pre><\/pre>\n\n<p>I am on a Ubuntu 14.04 64 bits, python 2.7.6<\/p>\n","answers":[{"AnswerId":"30423839","CreationDate":"2015-05-24T13:06:12.330","ParentId":null,"OwnerUserId":"1423333","Title":null,"Body":"<p>as written in the comment above, i think this is caused by <code>tensor<\/code> not implicitly importing <code>signal<\/code> or even <code>signal.conv<\/code>, hence you have to do the import yourself to use it:<\/p>\n\n<pre><code>In [1]: import theano\n\nIn [2]: theano.tensor\nOut[2]: &lt;module 'theano.tensor' from '\/usr\/local\/lib\/python2.7\/site-packages\/theano\/tensor\/__init__.pyc'&gt;\n<\/code><\/pre>\n\n<p>As you can see importing <code>theano<\/code> also gets us the <code>theano.tensor<\/code> module, but as <code>tensor.__init__.py<\/code> doesn't import <code>signal<\/code> for example, the following does not work:<\/p>\n\n<pre><code>In [3]: theano.tensor.signal\n---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-3-53b46c46cb25&gt; in &lt;module&gt;()\n----&gt; 1 theano.tensor.signal\n\nAttributeError: 'module' object has no attribute 'signal'\n\nIn [4]: theano.tensor.signal.conv\n---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-4-b2a3482abaed&gt; in &lt;module&gt;()\n----&gt; 1 theano.tensor.signal.conv\n\nAttributeError: 'module' object has no attribute 'signal'\n<\/code><\/pre>\n\n<p>After importing the submodule it does:<\/p>\n\n<pre><code>In [5]: import theano.tensor.signal.conv\n\nIn [6]: theano.tensor.signal\nOut[6]: &lt;module 'theano.tensor.signal' from '\/usr\/local\/lib\/python2.7\/site-packages\/theano\/tensor\/signal\/__init__.pyc'&gt;\n\nIn [7]: theano.tensor.signal.conv\nOut[7]: &lt;module 'theano.tensor.signal.conv' from '\/usr\/local\/lib\/python2.7\/site-packages\/theano\/tensor\/signal\/conv.pyc'&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":30376730,"AnswerCount":0,"Tags":"<multithreading><mpi><caffe>","CreationDate":"2015-05-21T14:28:26.980","AcceptedAnswerId":null,"OwnerUserId":4925131.0,"Title":"Parallel Caffe installation : This MPI version is NOT support multi-thread! *","Body":"<p>The mpi_train.sh has the following line:<\/p>\n\n<p>mpiexec.hydra -prepend-rank -host node11 -n 16<\/p>\n\n<p>While running examples\/cifar10\/mpi_train_quick.sh in parallel caffe I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have installed mpich-3.1.4 version.\nWhen I do which mpirun, which mpiexec, which mpicc, I get the path to the mpich-3.1.4 folder from all of these\nAlso the system has both Open-Mpi and MPICH-3.1.4 installed since open-mpi gets installed automatically while installing opencv which is a pre-requisite for parallel caffe . How can I resolve this error? <\/p>\n","answers":[]}
{"QuestionId":30383404,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2015-05-21T20:08:14.430","AcceptedAnswerId":30384484.0,"OwnerUserId":2991243.0,"Title":"What is data type for Python Keras deep learning package?","Body":"<p>I didn't find anything about data type that we need to work with  with <a href=\"http:\/\/keras.io\/\" rel=\"noreferrer\">this<\/a> link. I checked array and list but returned error. Any clue?<\/p>\n","answers":[{"AnswerId":"30384484","CreationDate":"2015-05-21T21:16:32.537","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>Keras uses <code>numpy<\/code> arrays containing the <code>theano.config.floatX<\/code> floating point type. This can be configured in your <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANORC\" rel=\"nofollow\"><code>.theanorc<\/code><\/a> file.<\/p>\n\n<p>Typically, it will be <code>float64<\/code> for CPU computations and <code>float32<\/code> for GPU computations, although you can also set it to <code>float32<\/code> when working on the CPU if you prefer.<\/p>\n\n<p>You can create a zero-filled array of the proper type by the command<\/p>\n\n<pre><code>X = numpy.zeros((4,3), dtype=theano.config.floatX)\n<\/code><\/pre>\n"}]}
{"QuestionId":30384908,"AnswerCount":1,"Tags":"<python><numpy><canopy><theano><keras>","CreationDate":"2015-05-21T21:45:01.803","AcceptedAnswerId":30386018.0,"OwnerUserId":2991243.0,"Title":"Python keras neural network (Theano) package returns an error about data dimensions","Body":"<p>I have this code:<\/p>\n\n<pre><\/pre>\n\n<p>Returns this error:<\/p>\n\n<pre><\/pre>\n\n<p>What is the problem?<\/p>\n","answers":[{"AnswerId":"30386018","CreationDate":"2015-05-21T23:27:57.407","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>You specified the wrong output dimensions for your internal layers. See for instance this example from the Keras documentation:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(20, 64, init='uniform'))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, 64, init='uniform'))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, 2, init='uniform'))\nmodel.add(Activation('softmax'))\n<\/code><\/pre>\n\n<p>Note how the output size of one layer matches the input size of the next one:<\/p>\n\n<pre><code>20x64 -&gt; 64x64 -&gt; 64x2\n<\/code><\/pre>\n\n<p>The first number is always the input size (number of neurons on the previous layer), the second number the output size (number of neurons on the next layer). So in this example you have four layers:<\/p>\n\n<ul>\n<li>an input layer with 20 neurons<\/li>\n<li>a hidden layer with 64 neurons<\/li>\n<li>a hidden layer with 64 neurons<\/li>\n<li>an output layer with 2 neurons<\/li>\n<\/ul>\n\n<p>The only hard restriction you have is that the first (input) layer needs to have as many neurons as you have features, and the last (output) layer needs to have as many neurons as you need for your task.<\/p>\n\n<p>For your example, since you have three features, you need to change the input layer size to 3, and you can keep the two output neurons from this example to do binary classification (or use one, as you did, with logistic loss).<\/p>\n"}]}
{"QuestionId":30385351,"AnswerCount":0,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-05-21T22:23:36.170","AcceptedAnswerId":null,"OwnerUserId":2284821.0,"Title":"how to setup Caffe imagenet_solver.prototxt file for fewer jpgs, program exited after iteration 0","Body":"<p>We need help to understand the parameters to use for smaller set of training (6000 jpgs) and val (170 jpgs) jpgs. Our execution was killed and exited after test score 0\/1 in Iteration 0. <\/p>\n\n<p>We are trying to run the imagenet sample on the caffe website tutorial at<\/p>\n\n<pre><\/pre>\n\n<p>Instead of using the full set of ILSVRC2 images in the package, we use our own training set of 6000 jpegs and val set of 170 jpeg images. They are each 256 x 256 jpeg files in the train and val directories as instructed.  We ran the script to get the auxillary data: <\/p>\n\n<pre><\/pre>\n\n<p>The train.txt and val.txt files are setup to describe one of two possible categories for each jpeg file. \nThen we ran the script to compute the mean image data which appeared to run correctly:<\/p>\n\n<pre><\/pre>\n\n<p>We used the model definitions supplied in the tutorial for imagenet_train.prototxt and imagenet_val.prototxt.\nSince we are training on much fewer images we modified the imagenet_solver.prototxt as follows:<\/p>\n\n<pre><\/pre>\n\n<p>when we run it using:<\/p>\n\n<pre><\/pre>\n\n<p>We get the following output where it hangs:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":30398000,"AnswerCount":0,"Tags":"<python><macos><theano>","CreationDate":"2015-05-22T13:29:05.270","AcceptedAnswerId":null,"OwnerUserId":4248384.0,"Title":"How to make Theano operate on Mac Lion?","Body":"<p>I need to have Theano operate on my Mac for a school project. I downloaded sources from <a href=\"https:\/\/pypi.python.org\/pypi\/Theano#downloads\" rel=\"nofollow\">here<\/a> (first line of the table) and put the untared file in the repertory of my school project. I didn't know if this was sufficient, so I also used . Everything seems fine since I got this : <\/p>\n\n<pre><\/pre>\n\n<p>As you can see, I'm using Anaconda. <\/p>\n\n<p>However, when I run a test in a python console, I get theses messages : <\/p>\n\n<pre><\/pre>\n\n<p>Is there any problem ? Are Theano and its compiler well installed ? \nI tried the first example of <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/adding.html\" rel=\"nofollow\">this tutorial<\/a> and had the same results without any bugs. But I'm wondering if there will be any bugs if I use more sophisticated functions and classes. <\/p>\n","answers":[]}
{"QuestionId":30403590,"AnswerCount":2,"Tags":"<computer-vision><neural-network><deep-learning><caffe><matcaffe>","CreationDate":"2015-05-22T18:31:14.680","AcceptedAnswerId":null,"OwnerUserId":242871.0,"Title":"number of parameters in Caffe LENET or Imagenet models","Body":"<p>How to calculate number of parameters in a model e.g. LENET for mnist,  or ConvNet for imagent model etc.\nIs there any specific function in caffe that returns or saves number of parameters in a model.\nregards<\/p>\n","answers":[{"AnswerId":"31399560","CreationDate":"2015-07-14T06:46:59.320","ParentId":null,"OwnerUserId":"4788536","Title":null,"Body":"<p>I can offer an explicit way to do this via the Matlab interface (make sure the matcaffe is installed first).\nBasically, you extract set of parameters from each network layer and count them.\nIn Matlab:<\/p>\n\n<pre><code>% load the network\nnet_model = &lt;path to your *deploy.prototxt file&gt;\nnet_weights = &lt;path to your *.caffemodel file&gt;\nphase = 'test';\ntest_net = caffe.Net(net_model, net_weights, phase);\n\n% get the list of layers\nlayers_list = test_net.layer_names;\n% for those layers which have parameters, count them\ncounter = 0;\nfor j = 1:length(layers_list),\n    if ~isempty(test_net.layers(layers_list{j}).params)\n    feat = test_net.layers(layers_list{j}).params(1).get_data();\n    counter = counter + numel(feat)\n    end\nend\n<\/code><\/pre>\n\n<p>In the end, 'counter' contains the number of parameters. <\/p>\n"},{"AnswerId":"39687866","CreationDate":"2016-09-25T14:19:35.313","ParentId":null,"OwnerUserId":"1621562","Title":null,"Body":"<p>Here is a python snippet to compute the number of parameters in a Caffe model:<\/p>\n\n<pre><code>import caffe\ncaffe.set_mode_cpu()\nimport numpy as np\nfrom numpy import prod, sum\nfrom pprint import pprint\n\ndef print_net_parameters (deploy_file):\n    print \"Net: \" + deploy_file\n    net = caffe.Net(deploy_file, caffe.TEST)\n    print \"Layer-wise parameters: \"\n    pprint([(k, v[0].data.shape) for k, v in net.params.items()])\n    print \"Total number of parameters: \" + str(sum([prod(v[0].data.shape) for k, v in net.params.items()]))\n\ndeploy_file = \"\/home\/ubuntu\/deploy.prototxt\"\nprint_net_parameters(deploy_file)\n\n# Sample output:\n# Net: \/home\/ubuntu\/deploy.prototxt\n# Layer-wise parameters: \n#[('conv1', (96, 3, 11, 11)),\n# ('conv2', (256, 48, 5, 5)),\n# ('conv3', (384, 256, 3, 3)),\n# ('conv4', (384, 192, 3, 3)),\n# ('conv5', (256, 192, 3, 3)),\n# ('fc6', (4096, 9216)),\n# ('fc7', (4096, 4096)),\n# ('fc8', (819, 4096))]\n# Total number of parameters: 60213280\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/gist.github.com\/kaushikpavani\/a6a32bd87fdfe5529f0e908ed743f779\" rel=\"nofollow\">https:\/\/gist.github.com\/kaushikpavani\/a6a32bd87fdfe5529f0e908ed743f779<\/a><\/p>\n"}]}
{"QuestionId":30407358,"AnswerCount":1,"Tags":"<function><lua><lua-table><torch>","CreationDate":"2015-05-22T23:42:37.093","AcceptedAnswerId":null,"OwnerUserId":3113501.0,"Title":"How does this function: input.nn.MSECriterion_updateOutput(self, input, target) work (in Lua\/Torch)?","Body":"<p>I have this function:<\/p>\n\n<pre><\/pre>\n\n<p>Now, <\/p>\n\n<pre><\/pre>\n\n<p>returns a number. I have no idea how it does it. I have walked step by step in the debugger and it seems this just computes a number with no intermediate steps.<\/p>\n\n<pre><\/pre>\n\n<p>I'm confused as to how this can compute a number.<\/p>\n\n<p>I'm confused as to why this is even allowed. The parameter input is a tensor, which doesn't even have any methods called nn.MSE input.nn.MSECriterion_updateOutput.<\/p>\n","answers":[{"AnswerId":"30416207","CreationDate":"2015-05-23T18:30:54.167","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>When you perform <code>require \"nn\"<\/code> this loads <code>init.lua<\/code> which in turns performs a <code>require('libnn')<\/code>. This is the C extension of torch\/nn.<\/p>\n\n<p>If you look at <code>init.c<\/code> you can find <a href=\"https:\/\/github.com\/torch\/nn\/blob\/3dd5d1d\/init.c#L127\"><code>luaopen_libnn<\/code><\/a>: this is the initialization function called when <code>libnn.so<\/code> is <code>require<\/code>-ed.<\/p>\n\n<p>This function takes care to initialize all parts of torch\/nn, including the native parts of <code>MSECriterion<\/code> via <code>nn_FloatMSECriterion_init(L)<\/code> and <code>nn_DoubleMSECriterion_init(L)<\/code>.<\/p>\n\n<p>If you look at <code>generic\/MSECriterion.c<\/code> you can find the <em>generic<\/em> (i.e macro expanded for <code>float<\/code> and <code>double<\/code>) <a href=\"https:\/\/github.com\/torch\/nn\/blob\/3dd5d1d\/generic\/MSECriterion.c#L47-L52\">initialization function<\/a>:<\/p>\n\n<pre><code>static void nn_(MSECriterion_init)(lua_State *L)\n{\n  luaT_pushmetatable(L, torch_Tensor);\n  luaT_registeratname(L, nn_(MSECriterion__), \"nn\");\n  lua_pop(L,1);\n}\n<\/code><\/pre>\n\n<p>This init function modifies the metatable of any <code>torch.FloatTensor<\/code> and <code>torch.DoubleTensor<\/code> so that it is filled with a bunch of functions under the <code>nn<\/code> key (see <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/9b51907\/lib\/luaT\/README.md\">Torch7 Lua C API<\/a> for more details). These functions are defined right before:<\/p>\n\n<pre><code>static const struct luaL_Reg nn_(MSECriterion__) [] = {\n  {\"MSECriterion_updateOutput\", nn_(MSECriterion_updateOutput)},\n  {\"MSECriterion_updateGradInput\", nn_(MSECriterion_updateGradInput)},\n  {NULL, NULL}\n};\n<\/code><\/pre>\n\n<p>In other words any tensor has these functions <em>attached<\/em> thanks to its metatable:<\/p>\n\n<pre><code>luajit -lnn\n&gt; print(torch.Tensor().nn.MSECriterion_updateOutput)\nfunction: 0x40921df8\n&gt; print(torch.Tensor().nn.MSECriterion_updateGradInput)\nfunction: 0x40921e20\n<\/code><\/pre>\n\n<p><em>Note: this mechanism is the same for all torch\/nn modules that have a C native implementation counterpart.<\/em><\/p>\n\n<p>So <code>input.nn.MSECriterion_updateOutput(self, input, target)<\/code> has for effect to call <code>static int nn_(MSECriterion_updateOutput)(lua_State *L)<\/code> as you can see on <a href=\"https:\/\/github.com\/torch\/nn\/blob\/3dd5d1d\/generic\/MSECriterion.c#L5\">generic\/MSECriterion.c<\/a>.<\/p>\n\n<p>This function computes the mean squared error between the input tensors.<\/p>\n"}]}
{"QuestionId":30420807,"AnswerCount":1,"Tags":"<c><lua><torch>","CreationDate":"2015-05-24T06:41:50.947","AcceptedAnswerId":30420916.0,"OwnerUserId":916451.0,"Title":"Strange C syntax in Lua library","Body":"<p>I see functions like <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/lib\/TH\/generic\/THTensor.c#L11\" rel=\"nofollow\">this<\/a> throughout the <a href=\"https:\/\/github.com\/torch\/torch7\" rel=\"nofollow\">torch library<\/a>'s C code:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a preprocessor thing, or something lua specific?  The idea, I think has something to do with the fact that  is a method of sorts on the  \"class\", but I've never seen this kind of syntax.<\/p>\n","answers":[{"AnswerId":"30420916","CreationDate":"2015-05-24T07:01:23.807","ParentId":null,"OwnerUserId":"1322972","Title":null,"Body":"<p>It is a preprocessor macro<\/p>\n\n<pre><code>lib\/TH\/THTensor.h:\n#define THTensor_(NAME)   TH_CONCAT_4(TH,Real,Tensor_,NAME)\n<\/code><\/pre>\n\n<p>which leads to...<\/p>\n\n<pre><code>lib\/TH\/THGeneral.h.in:\n#define TH_CONCAT_4(x,y,z,w) TH_CONCAT_4_EXPAND(x,y,z,w)\n<\/code><\/pre>\n\n<p>and finally...<\/p>\n\n<pre><code>lib\/TH\/THGeneral.h.in:\n#define TH_CONCAT_4_EXPAND(x,y,z,w) x ## y ## z ## w\n<\/code><\/pre>\n\n<p>Therefore, <\/p>\n\n<pre><code>long THTensor_(storageOffset)(const THTensor *self)\n<\/code><\/pre>\n\n<p>ultimately becomes this:<\/p>\n\n<pre><code>long THRealTensor_storageOffset(const THTensor *self)\n<\/code><\/pre>\n\n<p>Aren't preprocessors just <em>grand<\/em> ?<\/p>\n"}]}
{"QuestionId":30426216,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2015-05-24T17:21:23.403","AcceptedAnswerId":null,"OwnerUserId":2213825.0,"Title":"Using Theano shared variables with PyMC3","Body":"<p>My observable data is very big and its likelihood function is quite complex so I load all the observable data in the GPU and then use a theano function to get their likelihood depending on the parameters that I'm trying to estimate.<\/p>\n\n<pre><\/pre>\n\n<p>I don't know if this is the right approach or not. I am getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>----------\/\/-----------\nI tried to make a simple example using the example given by John Salvatier:\nP.S. This model doesn't make any sense.. I am just trying to figure it out how does Theano work with pymc3<\/p>\n\n<pre><\/pre>\n\n<p>And I get this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30426896","CreationDate":"2015-05-24T18:30:21.710","ParentId":null,"OwnerUserId":"359944","Title":null,"Body":"<p>The error you're getting is because you're compiling the logp function into get_logp and then passing it to DensityDist which actually expects a function that returns a theano variable, not a compiled theano function.<\/p>\n\n<p>I think you want something like: <\/p>\n\n<pre><code>with mc.Model() as model:\n    data_values = load_data()\n    X = theano.shared(np.asarray(data_values,\n        dtype=theano.config.floatX),\n         borrow=borrow)\n\n    sigma_r = mc.Uniform('sigma_r',0,1,testval=0.5,dtype='float32')\n    sigma_u = mc.Uniform('sigma_u',0,1,testval=0.5,dtype='float32')\n    B = mc.Uniform('B',0,1,testval=np.array([0.5,0.5,.5,.5],dtype=np.float32),shape=4,dtype='float32')\n\n    def logp(x):\n        #need to pass B sigma r\/u to the function\n        P = complexTheanoFunction(x, B, sigma_r, sigma_u)        \n        return T.sum(T.log(P))\n\n    obs = mc.DensityDist('observations',logp, observed=X)    \n    start = mc.find_MAP()\n    step = mc.NUTS(state=start)\n<\/code><\/pre>\n"}]}
{"QuestionId":30426734,"AnswerCount":1,"Tags":"<linux><caffe><fpic><gflags>","CreationDate":"2015-05-24T18:12:51.783","AcceptedAnswerId":null,"OwnerUserId":1596734.0,"Title":"caffe recompile libgflags.a with -fPIC error","Body":"<p>I'm getting an error when I try to install Caffe on Linux Ubuntu 64.\nThe error is as follows:<\/p>\n\n<blockquote>\n  <p>\/usr\/bin\/ld: \/usr\/local\/lib\/libgflags.a(gflags.cc.o): relocation R_X86_64_32S against `.rodata' can not be used when making a shared object; recompile with -fPIC<br>\n  \/usr\/local\/lib\/libgflags.a: error adding symbols: Bad value<\/p>\n<\/blockquote>\n\n<p>I tried recompile the gflags library with , but the error changed to as follow: <\/p>\n\n<blockquote>\n  <p>src\/caffe\/common.cpp: In function \u2018void caffe::GlobalInit(int*, char***)\u2019:<br>\n  src\/caffe\/common.cpp:35:5: error: \u2018::gflags\u2019 has not been declared<br>\n     ::gflags::ParseCommandLineFlags(pargc, pargv, true);<\/p>\n<\/blockquote>\n\n<p>I also tried to change the CMakeCache.txt of caffe to set the , but do not work either.<\/p>\n","answers":[{"AnswerId":"31129223","CreationDate":"2015-06-30T03:53:33.607","ParentId":null,"OwnerUserId":"3285719","Title":null,"Body":"<p>This error arrises because gflags 2.1 changed the name of the namespace from <code>google<\/code> to <code>gflags<\/code>. There are attempts by members of the caffe community to fix this error albeit they are not finalized. You should reassign the namespace from google to gflags as follows.<\/p>\n\n<p>In files<\/p>\n\n<ul>\n<li>caffe\/include\/caffe\/common.hpp<\/li>\n<li>caffe\/examples\/mnist\/convert_mnist_data.cpp<\/li>\n<\/ul>\n\n<p>Comment out the <code>ifndef<\/code><\/p>\n\n<pre><code>\/\/ #ifndef GFLAGS_GFLAGS_H_\nnamespace gflags = google;\n\/\/ #endif  \/\/ GFLAGS_GFLAGS_H_\n<\/code><\/pre>\n\n<p>This should work temporarily. You should fork and occasionally sync your caffe repo with the BVLC\/caffe repo on github so that you get the latest updates of the code.<\/p>\n"}]}
{"QuestionId":30438900,"AnswerCount":1,"Tags":"<python><slice><pca><theano>","CreationDate":"2015-05-25T12:56:24.413","AcceptedAnswerId":null,"OwnerUserId":1544186.0,"Title":"Slicing a matrix in the givens argument of a theano function","Body":"<p>I have the following piece of code, in which I attempt to apply PCA to the MNIST dataset.<\/p>\n\n<pre><\/pre>\n\n<p>As can be seen in the code above, I try to substitute  with a slice of the training dataset in the  params. The above code, however, results in the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Which implies that I am trying to assign a vector to a matrix, which isn't the behaviour I would expect (I doubled checked using numpy). I also tried a different approach, whereby I index the training dataset X_train with the array of booleans directly instead of using indices are performing the slices my self, but that also didn't work.<\/p>\n\n<pre><\/pre>\n\n<p>Which gives the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The only approach that did work is using disregarding the  param, and using only the inputs and outputs, as such:<\/p>\n\n<pre><\/pre>\n\n<p>None the less, I feel curious as to why my first two approaches do not work, and would appreciate if any one could point me to where I might be going wrong, since I am just starting with Theano.<\/p>\n","answers":[{"AnswerId":"30441790","CreationDate":"2015-05-25T15:43:31.683","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>The <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/library\/compile\/function.html#module-function\" rel=\"nofollow\">theano documentation<\/a> states:<\/p>\n\n<blockquote>\n  <p><code>givens<\/code> (iterable over pairs <code>(Var1, Var2)<\/code> of Variables. List, tuple or\n  dict. The <code>Var1<\/code> and <code>Var2<\/code> in each pair must have the same Type.) \u2013\n  specific substitutions to make in the computation graph (<code>Var2<\/code> replaces\n  <code>Var1<\/code>).<\/p>\n<\/blockquote>\n\n<p>And in the <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/tutorial\/examples.html#basictutexamples\" rel=\"nofollow\">tutorial examples<\/a>, there is the statement (emphasis mine)<\/p>\n\n<blockquote>\n  <p>In practice, a good way of thinking about the <code>givens<\/code> is as a mechanism\n  that allows you to replace any part of your formula with a different\n  expression that evaluates <strong>to a tensor of same shape and dtype<\/strong>.<\/p>\n<\/blockquote>\n\n<p>So, you cannot replace a matrix by a vector by means of the <code>givens<\/code> parameter since they don't have the same shape.<\/p>\n"}]}
{"QuestionId":30470517,"AnswerCount":1,"Tags":"<python><multithreading><machine-learning><theano>","CreationDate":"2015-05-26T23:10:49.067","AcceptedAnswerId":30500400.0,"OwnerUserId":1821593.0,"Title":"Theano\/Pylearn2. How to parallelize training?","Body":"<p>I have Convolutional Neural Network model described in YAML. When I run pylearn2's , I see that only one core of four is used.<\/p>\n\n<p><em>Is there a way to run training multi-threaded?<\/em><\/p>\n\n<p>Yeah, may be it's rather a Theano question. I followed this <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html<\/a> Theano tutorial about multi cores support, and  doesn't work for me - I see only one thread running. And further question:\n<em>can training be parallelized with ?<\/em> Because I can't check it since  doesn't do the thing. <em>Should I bother about my BLAS then?<\/em><\/p>\n\n<p>I have BLAS with LAPACK,  connected to them, python 2.7.9, my system is Ubuntu 15.04 on Core i5 4300U. <\/p>\n\n<p>Thank you, warm wishes!<\/p>\n","answers":[{"AnswerId":"30500400","CreationDate":"2015-05-28T07:51:49.753","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The most direct answer to your question is: you can't parallelize training in the way you desire.<\/p>\n\n<p>BLAS, OpenMP, and\/or running on a GPU only allow certain operations to be parallelized. The training itself can only be parallelized, in the way you want, if the training algorithm is designed to be parallelized. By default PyLearn2 uses the ordinary stochastic gradient descent (SGD) training algorithm which is not parallelizable. There are version of SGD that support parallelization (e.g. <a href=\"http:\/\/papers.nips.cc\/paper\/4687-large-scale-distributed-deep-networks.pdf\" rel=\"nofollow noreferrer\">Google's DistBelief<\/a>) but these are not available in PyLearn2 off-the-shelf. This is mostly because PyLearn2 is built on top of Theano and Theano is very much designed for shared memory operations.<\/p>\n\n<p>If you have a GPU then you'll almost certainly get faster training by switching to the GPU. If that isn't an option you should see more than one core being used some of time as long as your BLAS and OpenMP are set up correctly. The fact that <code>check_blas.py<\/code> doesn't show any improvement when <code>OMP_NUM_THREADS &gt; 2<\/code> suggests you don't have them set up correctly. I suggest opening a new question if you need help with this, providing more information about what you've done, and the settings shown by numpy when you print its config (see <a href=\"https:\/\/stackoverflow.com\/questions\/9000164\/how-to-check-blas-lapack-linkage-in-numpy-scipy\">here<\/a> for example).<\/p>\n"}]}
{"QuestionId":30473429,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-05-27T04:50:16.733","AcceptedAnswerId":30476856.0,"OwnerUserId":2097312.0,"Title":"Bad input argument to theano function","Body":"<p>I am new to theano. I am trying to implement simple linear regression but my program throws following error:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name \"\/home\/akhan\/Theano-Project\/uog\/theano_application\/linear_regression.py:36\"  at index 0(0-based)', 'Expected an array-like object, but found a Variable: maybe you are trying to call a function on a (possibly shared) variable instead of a numeric array?')<\/p>\n<\/blockquote>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>What is the explanation behind this error? (didn't got from the error message). Thanks in Advance.<\/p>\n","answers":[{"AnswerId":"30476856","CreationDate":"2015-05-27T08:19:14.680","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>There's an important distinction in Theano between defining a computation graph and a function which uses such a graph to compute a result.<\/p>\n\n<p>When you define<\/p>\n\n<pre><code>out = T.dot(X, W)\npredict = theano.function(inputs=[X], outputs=out)\n<\/code><\/pre>\n\n<p>you first set up a computation graph for <code>out<\/code> in terms of <code>X<\/code> and <code>W<\/code>. Note that <code>X<\/code> is a purely symbolic variable, it doesn't have any value, but the definition for <code>out<\/code> tells Theano, \"given a value for <code>X<\/code>, this is how to compute <code>out<\/code>\".<\/p>\n\n<p>On the other hand, <code>predict<\/code> is a <code>theano.function<\/code> which takes the computation graph for <code>out<\/code> and actual numeric values for <code>X<\/code> to produce a numeric output. What you pass into a <code>theano.function<\/code> when you call it always has to have an actual numeric value. So it simply makes no sense to do<\/p>\n\n<pre><code>y = predict(X)\n<\/code><\/pre>\n\n<p>because <code>X<\/code> is a symbolic variable and doesn't have an actual value.<\/p>\n\n<p>The reason you want to do this is so that you can use <code>y<\/code> to further build your computation graph. But there is no need to use <code>predict<\/code> for this: the computation graph for <code>predict<\/code> is already available in the variable <code>out<\/code> defined earlier. So you can simply remove the line defining <code>y<\/code> altogether and then define your cost as<\/p>\n\n<pre><code>cost = T.mean(T.sqr(out - Y))\n<\/code><\/pre>\n\n<p>The rest of the code will then work unmodified.<\/p>\n"}]}
{"QuestionId":30475415,"AnswerCount":3,"Tags":"<c++><makefile><deep-learning><caffe><gflags>","CreationDate":"2015-05-27T07:06:46.597","AcceptedAnswerId":null,"OwnerUserId":2163392.0,"Title":"Caffe Compilation Error: gflags.cc' is being linked both statically and dynamically into this executable","Body":"<p>I am trying to install caffe following this <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#compilation\" rel=\"nofollow noreferrer\">tutorial<\/a><\/p>\n\n<p>Basically I have the following error when I type the last make command:<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand how to solve this error. Did anybody find this error before? how can I solve it?<\/p>\n","answers":[{"AnswerId":"31229250","CreationDate":"2015-07-05T10:10:59.927","ParentId":null,"OwnerUserId":"2452553","Title":null,"Body":"<p>Whether or not you've already solved this somewhere else, I'm posting the answer here in-case others run into the same problem.<\/p>\n\n<p>Primarily, this problem seems to have come about because <strong>we don't always read things properly<\/strong> and blindly follow all instructions thinking they all apply to our case. <em>hint: <strong>they don't.<\/em><\/strong><\/p>\n\n<p>In the installation instructions for Caffe (presuming Ubuntu instructions), there is a section which states:<\/p>\n\n<blockquote>\n  <p><strong>Everything is packaged in 14.04.<\/strong><\/p>\n\n<pre><code>sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev protobuf-compiler\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Blindly ignoring the next title, which states clearly:<\/p>\n\n<blockquote>\n  <p><strong>Remaining dependencies, 12.04<\/strong><\/p>\n<\/blockquote>\n\n<p>we go on to install these dependencies, building and installing as required, resulting in the unfortunate side-effect of having 2 versions of libgflags, one dynamic (in <code>\/usr\/lib[\/x86_x64]<\/code> and one static in <code>\/usr\/local\/lib<\/code><\/p>\n\n<p><strong>Resolution<\/strong><\/p>\n\n<ol>\n<li>Promise ourselves failthfully we'll read instructions properly next time around.<\/li>\n<li><p>Uninstall libgflags<\/p>\n\n<pre><code>sudo apt-get remove -y libgflags\n<\/code><\/pre><\/li>\n<li><p>Delete <code>make install<\/code> versions<\/p>\n\n<pre><code>sudo rm -f \/usr\/local\/lib\/libgflags.a \/usr\/local\/lib\/libgflags_nothreads.a\nsudo rm -rf  \/usr\/local\/include\/gflags\n<\/code><\/pre><\/li>\n<li><p>Clean Caffe build<\/p>\n\n<pre><code>cd &lt;path&gt;\/&lt;to&gt;\/caffe\nmake clean\n<\/code><\/pre><\/li>\n<li><p>Re-install libgflags package<\/p>\n\n<pre><code>sudo apt-get install -y libgflags-dev\n<\/code><\/pre><\/li>\n<li><p>Rebuild Caffe<\/p>\n\n<pre><code>make all\nmake test\nmake runtest\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Et Voila. All tests should now run and you're ready to rock the deep-learning boat.<\/p>\n"},{"AnswerId":"56551179","CreationDate":"2019-06-11T20:20:36.067","ParentId":null,"OwnerUserId":"7083698","Title":null,"Body":"<p>I've worked out a way to debug this issue analytically.  In my case, I was cross-compiling for an older ABI, so apt-get wasn't an option and I was compiling all dependencies manually.<\/p>\n\n<p>First let's take a look at what this issue actually is.  In the Google GFlags library, flags are declared through global objects.  When the global object's constructor is run, it calls into the GFlags library to register that command line flag.  If the global constructor gets run multiple times (due to multiple versions of the library containing it being loaded into memory), then the GFlags register method dies with an error.<\/p>\n\n<p>What does GLog have to do with this?  Well, GLog uses GFlags, and it has globally declared flag objects.  Even if GFlags is linked correctly, if the GLog library gets loaded multiple times, you get an error pointing to logging.cc in GLog.<\/p>\n\n<p>Sounds like quite a mess, huh.  Even if GLog and GFlags are linked as shared in most cases, if another library links to a static version or some other version, kaboom.  <\/p>\n\n<p>Luckily, we can debug this issue using GDB and other tools, if you're willing to delve through some tricky symbol analysis.\nFirst, you'll want to run GDB on the Python interpreter when it tries to import caffe: <\/p>\n\n<pre><code>gdb --args python -c 'import caffe'\n<\/code><\/pre>\n\n<p>Now, run the program once through so that GDB can pick up all the libraries it imports:<\/p>\n\n<pre><code>(gdb) r\n<\/code><\/pre>\n\n<p>Now, we can set a breakpoint on the place in the function (<code>FlagRegistry::RegisterFlag()<\/code>) that prints the error message, and run it again.  Note that this line number is from my version of GFlags (2.2.2), you may have to look at the source code of your GFlags version and get the line number.<\/p>\n\n<pre><code>(gdb) break gflags.c:728\n(gdb) r\n<\/code><\/pre>\n\n<p>Hopefully, GDB should then break on the first instance of the error (if not, check that gflags has been built with debugging symbols).\nLook at the backtrace:<\/p>\n\n<pre><code>(gdb) bt\n#0  google::(anonymous namespace)::FlagRegistry::RegisterFlag (this=0xa33b30, flag=0x1249d20) at dev\/gflags-2.2.2\/src\/gflags.cc:728\n#1  0x00007ffff0f3247a in _GLOBAL__sub_I_logging.cc () from prefix\/lib\/libcaffe2.so\n#2  0x00007ffff7de76ca in call_init (l=&lt;optimized out&gt;, argc=argc@entry=3, argv=argv@entry=0x7fffffffdb08, env=env@entry=0x7fffffffdb28) at dl-init.c:72\n#3  0x00007ffff7de77db in call_init (env=0x7fffffffdb28, argv=0x7fffffffdb08, argc=3, l=&lt;optimized out&gt;) at dl-init.c:30\n#4  _dl_init (main_map=main_map@entry=0xd9c2a0, argc=3, argv=0x7fffffffdb08, env=0x7fffffffdb28) at dl-init.c:120\n#5  0x00007ffff7dec8f2 in dl_open_worker (a=a@entry=0x7fffffffcf70) at dl-open.c:575\n#6  0x00007ffff7de7574 in _dl_catch_error (objname=objname@entry=0x7fffffffcf60, errstring=errstring@entry=0x7fffffffcf68, mallocedp=mallocedp@entry=0x7fffffffcf5f, \n    operate=operate@entry=0x7ffff7dec4e0 &lt;dl_open_worker&gt;, args=args@entry=0x7fffffffcf70) at dl-error.c:187\n#7  0x00007ffff7debdb9 in _dl_open (file=0x9aee70 \"prefix\/lib\/python2.7\/site-packages\/caffe2\/python\/caffe2_pybind11_state.so\", mode=-2147483646, \n    caller_dlopen=0x51bb39 &lt;_PyImport_GetDynLoadFunc+233&gt;, nsid=-2, argc=&lt;optimized out&gt;, argv=&lt;optimized out&gt;, env=0x7fffffffdb28) at dl-open.c:660\n#8  0x00007ffff75ecf09 in dlopen_doit (a=a@entry=0x7fffffffd1a0) at dlopen.c:66\n#9  0x00007ffff7de7574 in _dl_catch_error (objname=0xabf9f0, errstring=0xabf9f8, mallocedp=0xabf9e8, operate=0x7ffff75eceb0 &lt;dlopen_doit&gt;, args=0x7fffffffd1a0) at dl-error.c:187\n#10 0x00007ffff75ed571 in _dlerror_run (operate=operate@entry=0x7ffff75eceb0 &lt;dlopen_doit&gt;, args=args@entry=0x7fffffffd1a0) at dlerror.c:163\n#11 0x00007ffff75ecfa1 in __dlopen (file=&lt;optimized out&gt;, mode=&lt;optimized out&gt;) at dlopen.c:87\n#12 0x000000000051bb39 in _PyImport_GetDynLoadFunc ()\n&lt;snip&gt;\n<\/code><\/pre>\n\n<p>Well that's a lot to deal with, but let's focus on the line that's actually important:<\/p>\n\n<pre><code>#1  0x00007ffff0f3247a in _GLOBAL__sub_I_logging.cc () from prefix\/lib\/libcaffe2.so\n<\/code><\/pre>\n\n<p>This is the call to the constructor for the global variables in logging.cc (which is part of GLog).  As you can see, this call is in libcaffe2.so, meaning that GLog has been statically linked to libcaffe2.so [I was using caffe2, but this procedure should be the same for both].<\/p>\n\n<p>You can then set a breakpoint on <code>google::(anonymous namespace)::FlagRegistry::RegisterFlag<\/code> and rerun the program from the start.  Look at each call to RegisterFlag(), and figure out where this particular flag was registered the first time.  If the library providing the flag is a shared library, then it should only ever get registered from that .so file, and nowhere else.<\/p>\n\n<p>To confirm the diagnosis, you can use <\/p>\n\n<pre><code>nm &lt;library&gt; | grep _GLOBAL__sub_I_logging.cc\n<\/code><\/pre>\n\n<p>to check for that init function in a library file.  Once you've found your culprit, you'll need to rebuild it so that it doesn't link to GFlags\/GLog statically.<\/p>\n"},{"AnswerId":"33550606","CreationDate":"2015-11-05T17:06:12.433","ParentId":null,"OwnerUserId":"5530303","Title":null,"Body":"<p>I also had two libraries installed, a shared .so library and a static .a library. I removed them all as well as the \/usr\/local\/include\/glog folder.\nThe .so file I had brought over when I (cross) compiled the system, while the .a was from a native and up-to-date build.\nUltimately it came down to building glog (natively) in such a way that it provided the .so files.\nI started with a clean download:<\/p>\n\n<p><code>git clone git:\/\/github.com\/google\/glog<\/code><\/p>\n\n<p>Then I edited CMakeLists.txt.\nWhere it says:<\/p>\n\n<pre><code>add_library (glog\n  ${GLOG_SRCS}\n)\n<\/code><\/pre>\n\n<p>I changed it to:<\/p>\n\n<pre><code>add_library (glog SHARED\n  ${GLOG_SRCS}\n)\n<\/code><\/pre>\n\n<p>Next you should be able to follow the other instructions. For my particular case I had to use slightly different instructions, not saying you have to do this. For me it was:\nmkdir build\ncd build<\/p>\n\n<pre><code>export CXXFLAGS=\"-fPIC\"\ncmake ..\nmake\nsudo make install\n<\/code><\/pre>\n\n<p>This gave me the .so files and put them in the right place. Then I started over with caffe and it fixed the error for me.<\/p>\n"}]}
{"QuestionId":30486033,"AnswerCount":2,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-05-27T14:52:41.100","AcceptedAnswerId":30497907.0,"OwnerUserId":4841248.0,"Title":"Tackling Class Imbalance: scaling contribution to loss and sgd","Body":"<p><strong>(An update to this question has been added.)<\/strong><\/p>\n\n<p>I am a graduate student at the university of Ghent, Belgium; my research is about emotion recognition with deep convolutional neural networks. I'm using the <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"noreferrer\">Caffe<\/a> framework to implement the CNNs.<\/p>\n\n<p>Recently I've run into a problem concerning class imbalance. I'm using 9216 training samples, approx. 5% are labeled positively (1), the remaining samples are labeled negatively (0).<\/p>\n\n<p>I'm using the <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidCrossEntropyLossLayer.html\" rel=\"noreferrer\">SigmoidCrossEntropyLoss<\/a> layer to calculate the loss. When training, the loss decreases and the accuracy is extremely high after even a few epochs. This is due to the imbalance: the network simply always predicts negative (0). <em>(Precision and recall are both zero, backing this claim)<\/em><\/p>\n\n<p>To solve this problem, I would like to <strong>scale the contribution to the loss depending on the prediction-truth combination<\/strong> (punish false negatives severely). My mentor\/coach has also advised me to <strong>use a scale factor when backpropagating<\/strong> through stochastic gradient descent (sgd): the factor would be correlated to the imbalance in the batch. A batch containing only negative samples would not update the weights at all.<\/p>\n\n<p><em>I have only added one custom-made layer to Caffe: to report other metrics such as precision and recall. My experience with Caffe code is limited but I have a lot of expertise writing C++ code.<\/em><\/p>\n\n<hr>\n\n<p><strong>Could anyone help me or point me in the right direction on how to adjust the <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidCrossEntropyLossLayer.html\" rel=\"noreferrer\">SigmoidCrossEntropyLoss<\/a> and <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidLayer.html\" rel=\"noreferrer\">Sigmoid<\/a> layers to accomodate the following changes:<\/strong><\/p>\n\n<ol>\n<li>adjust the contribution of a sample to the total loss depending on the prediction-truth combination (true positive, false positive, true negative, false negative).<\/li>\n<li>scale the weight update performed by stochastic gradient descent depending on the imbalance in the batch (negatives vs. positives).<\/li>\n<\/ol>\n\n<p>Thanks in advance!<\/p>\n\n<hr>\n\n<h2>Update<\/h2>\n\n<p>I have incorporated the <strong><a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1InfogainLossLayer.html#details\" rel=\"noreferrer\">InfogainLossLayer<\/a> as suggested by <a href=\"https:\/\/stackoverflow.com\/a\/30497907\/1714410\">Shai<\/a><\/strong>. I've also added another custom layer that builds the infogain matrix  based on the imbalance in the current batch.<\/p>\n\n<p>Currently, the matrix is configured as follows:<\/p>\n\n<pre><\/pre>\n\n<p><em>I'm planning on experimenting with different configurations for the matrix in the future.<\/em><\/p>\n\n<p>I have tested this on a 10:1 imbalance. The results have shown that the network is learning useful things now: <em>(results after 30 epochs)<\/em><\/p>\n\n<ul>\n<li>Accuracy is approx. ~70% (down from ~97%);<\/li>\n<li>Precision is approx. ~20% (up from 0%);<\/li>\n<li>Recall is approx. ~60% (up from 0%).<\/li>\n<\/ul>\n\n<p>These numbers were reached at around 20 epochs and didn't change significantly after that.<\/p>\n\n<p><em>!! The results stated above are merely a proof of concept, they were obtained by training a simple network on a 10:1 imbalanced dataset. !!<\/em><\/p>\n","answers":[{"AnswerId":"30497907","CreationDate":"2015-05-28T05:28:51.720","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Why don't you use the <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1InfogainLossLayer.html\" rel=\"nofollow noreferrer\"><strong>InfogainLoss<\/strong><\/a> layer to compensate for the imbalance in your training set?<\/p>\n\n<p>The Infogain loss is defined using a weight matrix <code>H<\/code> (in your case 2-by-2) The meaning of its entries are<\/p>\n\n<pre><code>[cost of predicting 1 when gt is 0,    cost of predicting 0 when gt is 0\n cost of predicting 1 when gt is 1,    cost of predicting 0 when gt is 1]\n<\/code><\/pre>\n\n<p>So, you can set the entries of <code>H<\/code> to reflect the difference between errors in predicting 0 or 1.<\/p>\n\n<p>You can find how to define matrix <code>H<\/code> for caffe in <a href=\"https:\/\/stackoverflow.com\/questions\/27632440\/infogain-loss-layer\">this thread<\/a>.<\/p>\n\n<p>Regarding sample weights, you may find <a href=\"http:\/\/deepdish.io\/2014\/11\/04\/caffe-with-weighted-samples\/\" rel=\"nofollow noreferrer\">this post<\/a> interesting: it shows how to modify the <strong>SoftmaxWithLoss<\/strong> layer to take into account sample weights.<\/p>\n\n<hr>\n\n<p>Recently, a modification to cross-entropy loss was proposed by <a href=\"https:\/\/arxiv.org\/abs\/1708.02002\" rel=\"nofollow noreferrer\"><em>Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, Piotr Doll\u00e1r<\/em> <strong>Focal Loss for Dense Object Detection<\/strong>, (ICCV 2017)<\/a>.<br>\nThe idea behind focal-loss is to assign different weight for each example based on the relative difficulty of predicting this example (rather based on class size etc.). From the brief time I got to experiment with this loss, it feels superior to <code>\"InfogainLoss\"<\/code> with class-size weights. <\/p>\n"},{"AnswerId":"47626265","CreationDate":"2017-12-04T03:55:01.993","ParentId":null,"OwnerUserId":"6064933","Title":null,"Body":"<p>I have also come across this class imbalance problem in my classification task. Right now I am using CrossEntropyLoss with weight (documentation <a href=\"http:\/\/pytorch.org\/docs\/master\/nn.html#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\">here<\/a>) and it works fine. The idea is to give more loss to samples in classes with smaller number of images.<\/p>\n\n<h3>Calculating the weight<\/h3>\n\n<p>weight for each class in inversely proportional to the image number in this class. Here is a snippet to calculate weight for all class using numpy,<\/p>\n\n<pre><code>cls_num = []\n# train_labels is a list of class labels for all training samples\n# the labels are in range [0, n-1] (n classes in total)\ntrain_labels = np.asarray(train_labels)\nnum_cls = np.unique(train_labels).size\n\nfor i in range(num_cls):\n    cls_num.append(len(np.where(train_labels==i)[0]))\n\ncls_num = np.array(cls_num)\n\ncls_num = cls_num.max()\/cls_num\nx = 1.0\/np.sum(cls_num)\n\n# the weight is an array which contains weight to use in CrossEntropyLoss\n# for each class.\nweight = x*cls_num\n<\/code><\/pre>\n"}]}
{"QuestionId":30496978,"AnswerCount":1,"Tags":"<theano><pymc3>","CreationDate":"2015-05-28T03:56:25.587","AcceptedAnswerId":null,"OwnerUserId":4430895.0,"Title":"Is there a workaround for not fusing the observed data into model definition in Pymc3?","Body":"<p>Problem definition: consider the \"Simpletest\" model (from pymc3 examples)which is something similar to the following one:<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to change it so that I'll have a fixed model structure but run the sampling several iterations, each time adding a new data point to the previous (observed) dataset. Since the observed data is somehow embedded inside the model definition, the only way I know to do this is to put the whole model definition inside a loop:<\/p>\n\n<pre><\/pre>\n\n<p>This may produce some unnecessary overhead specially if the model is large. To refrain from the overhead of repeatedly defining the same model, I wonder if there is a solution so that the same results could be achieved with something similar to the following idea:<\/p>\n\n<pre><\/pre>\n\n<p>or even better:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30585998","CreationDate":"2015-06-02T01:42:27.943","ParentId":null,"OwnerUserId":"4430895","Title":null,"Body":"<p>It seems that it's not possible right know. But there is an open issue on this topic with possible solutions here : <a href=\"https:\/\/github.com\/pymc-devs\/pymc\/issues\/10\" rel=\"nofollow\">https:\/\/github.com\/pymc-devs\/pymc\/issues\/10<\/a>  <\/p>\n"}]}
{"QuestionId":30500332,"AnswerCount":1,"Tags":"<python><theano><hessian-matrix>","CreationDate":"2015-05-28T07:48:03.257","AcceptedAnswerId":null,"OwnerUserId":3282028.0,"Title":"How to use theano.gradient.hessian? Example needed","Body":"<p>I tried the code below:<\/p>\n\n<pre><\/pre>\n\n<p>Then I run it with following real values<\/p>\n\n<pre><\/pre>\n\n<p>Then I encountered following error<\/p>\n\n<pre><\/pre>\n\n<p>I am new to python and am exploring Theano for building neural networks.<\/p>\n","answers":[{"AnswerId":"30500423","CreationDate":"2015-05-28T07:53:02.103","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>h<\/code> is a function that accepts two parameters. You are giving it a single parameter which is a list containing two elements.<\/p>\n\n<p>Try changing <code>h([x,y])<\/code> to <code>h(x,y)<\/code>.<\/p>\n"}]}
{"QuestionId":30500977,"AnswerCount":2,"Tags":"<caffe><ubuntu-15.04>","CreationDate":"2015-05-28T08:21:32.407","AcceptedAnswerId":null,"OwnerUserId":601314.0,"Title":"Caffe fails to build on ubuntu 15.04","Body":"<p>Having followed the <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#compilation\" rel=\"nofollow noreferrer\">Caffe build instructions<\/a>, I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>I check the install of  with apt-get:<\/p>\n\n<pre><\/pre>\n\n<p>cuda7 is installed, opencv 3 ...<\/p>\n","answers":[{"AnswerId":"30501122","CreationDate":"2015-05-28T08:28:24.237","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>Maybe try installing the entire <a href=\"https:\/\/launchpad.net\/ubuntu\/+source\/hdf5\" rel=\"nofollow noreferrer\">hdf5<\/a> package, not only the dev portion.<\/p>\n\n<p>If that doesn't work, verify that you have the hdf5.h header on your system and check its path.<\/p>\n\n<p>You can check gcc's include path with the command <a href=\"https:\/\/stackoverflow.com\/a\/6666338\/2929337\">[source]<\/a><\/p>\n\n<pre><code>gcc -xc -E -v -\n<\/code><\/pre>\n"},{"AnswerId":"33025755","CreationDate":"2015-10-08T20:53:11.157","ParentId":null,"OwnerUserId":"1079075","Title":null,"Body":"<p>The steps required to make it build on Ubuntu 15.04 and Debian 8.x can be found in this <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2347\" rel=\"nofollow\">GitHub issue<\/a>.<\/p>\n\n<p>To summarise:<\/p>\n\n<pre><code>#!\/bin\/bash\n# manipulate header path, before building caffe on debian jessie\n# usage:\n# 1. cd root of caffe\n# 2. bash &lt;this_script&gt;\n# 3. build\n\n# transformations :\n#  #include \"hdf5\/serial\/hdf5.h\" -&gt; #include \"hdf5\/serial\/hdf5.h\"\n#  #include \"hdf5_hl.h\" -&gt; #include \"hdf5\/serial\/hdf5_hl.h\"\n\nfind . -type f -exec sed -i -e 's^\"hdf5.h\"^\"hdf5\/serial\/hdf5.h\"^g' -e 's^\"hdf5_hl.h\"^\"hdf5\/serial\/hdf5_hl.h\"^g' '{}' \\;\n<\/code><\/pre>\n\n<p>Followed by<\/p>\n\n<p>Modify <code>INCLUDE_DIRS<\/code> in <code>Makefile.config<\/code><\/p>\n\n<pre><code>INCLUDE_DIRS := $(PYTHON_INCLUDE) \/usr\/local\/include \/usr\/include\/hdf5\/serial\/\n<\/code><\/pre>\n\n<p>And finally, make some simlinks to HD5<\/p>\n\n<pre><code>cd \/usr\/lib\/x86_64-linux-gnu\nsudo ln -s libhdf5_serial.so.8.0.2 libhdf5.so\nsudo ln -s libhdf5_serial_hl.so.8.0.2 libhdf5_hl.so\n<\/code><\/pre>\n"}]}
{"QuestionId":30501043,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-05-28T08:25:12.967","AcceptedAnswerId":null,"OwnerUserId":2213825.0,"Title":"How to apply a Gaussian Blur using theano","Body":"<p>I am not being able to understand   function I want to blur a set of images with a Gaussian kernel:<\/p>\n\n<pre><\/pre>\n\n<p>The above code doesn't work it gives the error:<\/p>\n\n<pre><\/pre>\n\n<p>Either way since the gaussian filter is separable I would rather apply two linear filters one in the  and one in the  of the images. But I don't understand how to pull this off and the documentation of <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/conv.html#theano.tensor.signal.conv.conv2d\" rel=\"nofollow\">conv2d<\/a> is not helping me much.<\/p>\n","answers":[{"AnswerId":"30501266","CreationDate":"2015-05-28T08:35:30.547","ParentId":null,"OwnerUserId":"2100935","Title":null,"Body":"<p>Your filter is bigger than your input image. conv2d cannot handle this situation when <code>border_mode='valid'<\/code>, which is the default border mode.<\/p>\n\n<p>Why? Because the valid region of your image has the size 1 x 1 x 37 x -2. Negative dimensions make no sense.<\/p>\n\n<p>Solution: use <code>R2 = conv2d(M,G_kernel,border_mode='full')<\/code><\/p>\n\n<p>As for the second part of the question, you can use the seperated kernels using <code>R2 = conv2d(conv2d(M,G_x_kernel,border_mode='full'),G_y_kernel,border_mode='full')<\/code>\nHowever, I'm not sure it will increase speed with these small filters and running on a GPU.<\/p>\n"},{"AnswerId":"30505604","CreationDate":"2015-05-28T11:41:53.870","ParentId":null,"OwnerUserId":"2100935","Title":null,"Body":"<p>The pixel dimensions need to be the last dimensions for convolutions in Theano (because those convolutions are usually used for neural nets, not images). In theano, the convention for these feature maps is (batch_size, image channels, x, y). So instead of<\/p>\n\n<pre><code>images = mpimg.imread('Lenna.png')\n<\/code><\/pre>\n\n<p>try<\/p>\n\n<pre><code>images = np.rollaxis(mpimg.imread('Lenna.png'), 2, 0)\n<\/code><\/pre>\n\n<p>To convert your image to 3x512x512 instead of 512x512x3.<\/p>\n"}]}
{"QuestionId":30502464,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-05-28T09:26:21.897","AcceptedAnswerId":30504930.0,"OwnerUserId":1367788.0,"Title":"How is theano dot product broadcasted","Body":"<p>Could anyone example how i theano dot product broadcast. It seems it is different from numpy<\/p>\n\n<pre><\/pre>\n\n<p>Here are the experiments I tried with theano.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30504930","CreationDate":"2015-05-28T11:11:36.810","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano does broadcasting just like numpy.\nTo demonstrate, this code compares Theano and numpy directly:<\/p>\n\n<pre><code>import numpy\n\nimport theano\nimport theano.tensor as T\n\nTENSOR_TYPES = dict([(0, T.scalar), (1, T.vector), (2, T.matrix), (3, T.tensor3), (4, T.tensor4)])\n\nrand = numpy.random.rand\n\n\ndef theano_dot(x, y):\n    sym_x = TENSOR_TYPES[x.ndim]('x')\n    sym_y = TENSOR_TYPES[y.ndim]('y')\n    return theano.function([sym_x, sym_y], theano.dot(sym_x, sym_y))(x, y)\n\n\ndef compare_dot(x, y):\n    print theano_dot(x, y).shape, numpy.dot(x, y).shape\n\n\nprint compare_dot(rand(3, 5, 10), rand(2, 5, 10, 4))\nprint compare_dot(rand(3, 10), rand(2, 5, 10, 4))\nprint compare_dot(rand(3, 10), rand(10, 4))\nprint compare_dot(rand(3, 10), rand(2, 10, 4))\nprint compare_dot(rand(5, 10), rand(2, 10, 10))\n<\/code><\/pre>\n\n<p>The output is<\/p>\n\n<pre><code>(3L, 5L, 2L, 5L, 4L) (3L, 5L, 2L, 5L, 4L)\n(3L, 2L, 5L, 4L) (3L, 2L, 5L, 4L)\n(3L, 4L) (3L, 4L)\n(3L, 2L, 4L) (3L, 2L, 4L)\n(5L, 2L, 10L) (5L, 2L, 10L)\n<\/code><\/pre>\n\n<p>Theano and numpy produce results with the same shape in every case you describe.<\/p>\n"}]}
{"QuestionId":30510722,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-05-28T15:20:54.500","AcceptedAnswerId":30525487.0,"OwnerUserId":4841248.0,"Title":"solver parameter 'test_iter' changes label values during test phase","Body":"<p>I'm using the <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\"><strong>Caffe<\/strong><\/a> framework to construct and research convolutional neural networks.<\/p>\n\n<p>I've discovered (what I believe to be) a bug by accident. <em>(I've already reported it on <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2524\" rel=\"nofollow\">Github<\/a>.)<\/em><\/p>\n\n<p><strong>This is the issue:<\/strong> during the test phase, label values get changed depending on the value of  parameter (defined in the solver  file).<\/p>\n\n<hr>\n\n<p>I'm using 10240 images to train and test a network. Each image has 38 labels, each label can have two (0 or 1) values. I'm using the HDF5 file format to get my image data and labels into Caffe; each file stores 1024 images and their respective labels. <em>(I've checked the HDF5 files, everything is correct there.)<\/em><\/p>\n\n<p>I'm using 9216 (= 9 files) images for training and 1024 (= 1 file) for testing. My Nvidia 540M graphics card merely has 1GB of memory, which means I have to process in batch (usually 32 or 64 images per batch).<\/p>\n\n<p>I'm using the following network to replicate the problem:<\/p>\n\n<pre><\/pre>\n\n<p>This network simply outputs all label values. I'm using the following solver for this network: <em>(Mostly copied from my real network.)<\/em><\/p>\n\n<pre><\/pre>\n\n<p>The following results were obtained by changing the  and  parameters. According to <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html#define-the-mnist-network\" rel=\"nofollow\">this<\/a> tutorial,  of the test data and  in the solver should balance out to make sure all test samples are used during testing. In my case, I'll make sure that .<\/p>\n\n<p><strong>These are my results when changing the values:<\/strong><br>\n: Everything is okay.<br>\n: Labels that were '1' changed to '0.50'.<br>\n: Labels that were '1' changed to '0.50' or '0.25'<br>\n: Labels that were '1' changed to '0.50' or '0.25' or '0.125'<br>\n The pattern continues.  <\/p>\n\n<hr>\n\n<p><strong>What is going on that affects the values of the labels during testing?<\/strong> Am I simply interpreting the use of  and  wrong, or am I missing something else?<\/p>\n","answers":[{"AnswerId":"30525487","CreationDate":"2015-05-29T09:03:45.257","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The results shown in the output log are averages of the iterations, so if you have 2 iterations labels that are one are averaged to 0.5.<\/p>\n\n<p>So, if batch size is 1024, you have 1024 outputs displayed and everything is ok. When batch size is 512, you only have 512 outputs displayed each is an <em>average<\/em> of two labels the <code>i<\/code>-th and the <code>i+512<\/code>-th label, most chances the labels do not co-inside.<\/p>\n\n<p>To verify this, you can arrange your test data such that labels 1 are placed at even places, so when changing the batch_size the labels 1 still coincide and you should get exactly 1 for output. <\/p>\n"}]}
{"QuestionId":30517570,"AnswerCount":1,"Tags":"<for-loop><theano><pymc3>","CreationDate":"2015-05-28T21:41:30.283","AcceptedAnswerId":30585944.0,"OwnerUserId":4484463.0,"Title":"Improving performance of a Theano for loop","Body":"<p>I have the following code that accomplishes what I want to do. But I'm wondering if there's a better way of doing it by avoiding For Loops. Performance is important here since I call these operations many times. <\/p>\n\n<p>I think it could be improved by using \"scan\" and \"function\" but I'm not experienced enough with Theano for it to be obvious to me. I did try putting everything inside a theano.function but it didn't work.  <\/p>\n\n<pre><\/pre>\n\n<p>By the way, this is an implementation of the constrained Probabilistic Matrix Factorization (equation 7 in the paper by Salakhutdinov and Mnih). I'm doing it with pymc3 so \"W\" and \"Y\" are really stochastic pymc3 tensors (which I believe are just theano tensors). <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"30585944","CreationDate":"2015-06-02T01:35:11.793","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>You need to understand how to vectorise your code. For example:<\/p>\n\n<pre><code>Ui=np.zeros(dim)\nfor k in range(m):\n    Ui+=t.dot(I[i,k],W[k,:])\n<\/code><\/pre>\n\n<p>can be implemented as:<\/p>\n\n<pre><code>Ui = I[None, i] * W\n<\/code><\/pre>\n\n<p>Learn on numpy broadcasting. This is a really powerful way of thinking and it do computation faster and with less memory. This work for NumPy and Theano code. <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/numpy.html#broadcasting\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/numpy.html#broadcasting<\/a><\/p>\n\n<p>This can be done at other place I think to speed it up even more.<\/p>\n"}]}
{"QuestionId":30520523,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-05-29T03:16:04.780","AcceptedAnswerId":30529673.0,"OwnerUserId":211450.0,"Title":"Why do I get int64 when creating itensor3 in Theano?","Body":"<p>I'm quite new to Theano, \nI'm trying to create a tensor of int32 using itensor3,but for some reason I get int64 instead of int32.\nDo I need to specify anything in the config file?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30529673","CreationDate":"2015-05-29T12:30:12.333","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>In Theano I believe shapes are always specified in <code>int64<\/code> values.<\/p>\n\n<p>The result of your Theano function, <code>f<\/code>, is a shape size, i.e. <code>l.shape[0]<\/code> so the type of the result returned by <code>f<\/code> is going to be <code>int64<\/code>. This does not change the fact that the input is of type <code>int32<\/code>.<\/p>\n"}]}
{"QuestionId":30521390,"AnswerCount":1,"Tags":"<python><neural-network><lstm><keras>","CreationDate":"2015-05-29T04:53:54.437","AcceptedAnswerId":null,"OwnerUserId":4206428.0,"Title":"Keras LSTM predicting only 1 category, in multi-category classification - how to fix?","Body":"<p>I have a text dataset that has equal number of labels - . I ran the  example  (imdb example) on their website with my dataset and the compile line changed to <\/p>\n\n<pre><\/pre>\n\n<p>But the model predicts only 1 category, that is,  accuracy consistently.<\/p>\n\n<p>Could you please help me fix it \/ change settings as required?<\/p>\n","answers":[{"AnswerId":"34077294","CreationDate":"2015-12-03T22:17:08.203","ParentId":null,"OwnerUserId":"1879926","Title":null,"Body":"<p>You'll need to modify the with<\/p>\n\n<pre><code>model.add(Dense(nb_classes))\n<\/code><\/pre>\n\n<p>where nb_classes corresponds to the number of categorical classes.  <\/p>\n"}]}
{"QuestionId":30521680,"AnswerCount":1,"Tags":"<python><arrays><numpy><theano><mnist>","CreationDate":"2015-05-29T05:22:24.430","AcceptedAnswerId":30524636.0,"OwnerUserId":3748690.0,"Title":"Python Numpy Error: ValueError: setting an array element with a sequence","Body":"<p>I am trying to build a dataset similar to mnist.pkl.gz provided in theano logistic_sgd.py implementation.  Following is my code snippet. <\/p>\n\n<pre><\/pre>\n\n<p>Error Message: \nTraceback (most recent call last):<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>csv file contains two fields.. image name, classification label\nwhen is run this in python interpreter, it seems to be working for me.. as follows.. I dont get error saying setting an array element with a sequence here..<\/p>\n\n<p>---------python interpreter output----------<\/p>\n\n<pre><\/pre>\n\n<p>Even though i am running same set of instruction (logically), when i run sample.py, i get valueError: setting an array element with a sequence.. I trying to understand this behavior.. any help would be great..<\/p>\n","answers":[{"AnswerId":"30524636","CreationDate":"2015-05-29T08:21:16.367","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is probably similar to that of <a href=\"https:\/\/stackoverflow.com\/questions\/4674473\/valueerror-setting-an-array-element-with-a-sequence\">this question<\/a>.<\/p>\n\n<p>You're trying to create a matrix of pixel values with a row per image. But each image has a different size so the number of pixels in each row is different.<\/p>\n\n<p>You can't create a \"jagged\" float typed array in numpy -- every row must be of the same length.<\/p>\n\n<p>You'll need to pad each row to the length of the largest image.<\/p>\n"}]}
{"QuestionId":30548699,"AnswerCount":1,"Tags":"<python><numpy><k-means><theano>","CreationDate":"2015-05-30T16:43:08.150","AcceptedAnswerId":null,"OwnerUserId":1544186.0,"Title":"Theano function: Unused input","Body":"<p>I am trying to implement mini-batch Kmeans. The part that seems to be giving me a really hard time is specifying the minibatches as inputs to theano. I have a  class with an  function, where  is in this case the minibatch, and  is the size of the batch. I also have a  function which takes no arguments, but instead uses the data passed to .<\/p>\n\n<p>My main script consists of the following:<\/p>\n\n<pre><\/pre>\n\n<p>What I intended to do was to initialize a KmeansMiniBatch object with a symbolic variable , which gets replaced by the a  at each iteration. Each of the minibatches is generated by the function  which basically takes as input the entire dataset, and using  returns only a subset of that dataset, which is a  array. Unfortunately, I cannot seem to accomplish what I set out to achieve, as the above code results in the following error message:<\/p>\n\n<blockquote>\n  <p>theano.compile.function_module.UnusedInputError: theano.function was asked to create a function computing outputs given certain inputs, but the provided input variable at index 0 is not part of the computational graph needed to compute the outputs: X.\n  To make this error into a warning, you can pass the parameter on_unused_input='warn' to theano.function. To disable it completely, use on_unused_input='ignore'.<\/p>\n<\/blockquote>\n\n<p>I am not sure why exactly I do get this error, since I do replace the symbolic variable  by the function input . Furthermore, if I do set  I end up with the following error message during evaluating :<\/p>\n\n<blockquote>\n  <p>theano.gof.fg.MissingInputError: (\"An input of the graph, used to compute Shape(X), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.\", X)<\/p>\n<\/blockquote>\n\n<p>Any help would be very much appreciated !<\/p>\n\n<hr>\n\n<p>Edit:<\/p>\n\n<p>So I finally got it working ! My  function used to update a matrix , which is an attribute of the class KmeanMiniBatch, but didn't return it, which apparently caused  to complain, since the input was indeed not used in the output. What I did is, I modified  to return , and that basically solved the issue. Here is my modified <\/p>\n\n<pre><\/pre>\n\n<p>Apparently  functions do not play well with void python functions.<\/p>\n\n<hr>\n\n<p>Edit 2:<\/p>\n\n<p>Just to shed some more light on the problem that I would like to solve. So, the version of  I am implementing is also know as , whereby a dictionary  basically provides a compression for the dataset  into . Initially, the part of  concerning  was as follows:<\/p>\n\n<pre><\/pre>\n\n<p>So basically, at every iteration the dictionary  would be updated, and it would therefore, make no sense to return , which I had to do in order to stop  from complaining:<\/p>\n\n<pre><\/pre>\n\n<p>D is initialized as follows in :<\/p>\n\n<pre><\/pre>\n\n<p>What I would like to achieve is:\n1. Not have to return , but instead update, and evaluate  in place, which I can then retrieve through \n2. I am not about my choice of having  as a symbolic variable? Perhaps a shared variable would be a better choice?\n3. Most importantly, at each of the 30 iterations, I would like to substitute the data  to the  model with a minibatch, and hence, my use of the givens param. Is there a better way to achieve that?<\/p>\n","answers":[{"AnswerId":"30568157","CreationDate":"2015-06-01T07:43:42.610","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>As far as you as a user of Theano are concerned, symbolic variables do not support a concept of \"current value\" or \"updating\". For that you would need a shared variable.<\/p>\n\n<p>You need to be clearer on how you would like to use your <code>KmeansMiniBatch<\/code> class. At the moment it does not encapsulate the D-updating behaviour since the Theano function is compiled and executed outside of <code>KmeansMiniBatch<\/code>. You might prefer a usage like this:<\/p>\n\n<pre><code>kmeans = KmeansMiniBatch()\n\ndata = load_data()\nfor i in xrange(30):\n    kmeans.update(get_batch(data, batch_size=10000))\n\nimage = Image.fromarray(\ntile_raster_images(X=np.transpose(kmeans.get_D()),\n                       img_shape=(12, 12), tile_shape=(10, 30),\n                       tile_spacing=(1, 1)))\n<\/code><\/pre>\n\n<p>Note that there is no Theano functionality visible here, that's all encapsulated into the <code>KmeansMiniBatch<\/code> class. We also don't need to tell <code>KmeansMiniBatch<\/code> what the batch size is because that doesn't change the symbolic expression; instead we tell <code>get_batch<\/code> how large a batch to get.<\/p>\n\n<p>Inside <code>KmeansMiniBatch<\/code> you have two possible approaches.<\/p>\n\n<ol>\n<li><p>Make <code>D<\/code> a shared variable and use <code>updates=...<\/code> in your Theano function to change its contents on each <code>update<\/code>.<\/p>\n\n<pre><code>class KmeansMiniBatch:\n    def __init__(dimensions, K):\n        # ... init srng ...\n        D = srng.normal(size=(dimensions, K))\n        D = D \/ numpy.sqrt(numpy.sum(numpy.sqr(D), axis=0)))\n        self.D = theano.shared(D, 'D')\n        mini_batch = T.matrix('mini_batch', dtype='float64')\n        self.func = theano.function(inputs=[mini_batch], updates=fit_once(mini_batch))\n\n    def update(batch):\n        self.func(batch)\n\n    def fit_once(mini_batch):\n        # ... do work to create S symbolically ...\n        D_update = T.dot(mini_batch, T.transpose(S))\n        D_update = D_update \/ T.sqrt(T.sum(T.sqr(D_update), axis=0))\n        return [(self.D, D_update)]\n\n    def get_D():\n        return self.D.get_value()\n<\/code><\/pre>\n\n<p>Note that the init of D has changed from a Theano operation to a numpy operation.<\/p><\/li>\n<li><p>Make <code>D<\/code> a regular numpy array, pass it in as an input to your Theano function, and change the value to the output of your Theano function, on each <code>update<\/code>.<\/p>\n\n<pre><code>class KmeansMiniBatch:\n    def __init__(dimensions, K):\n        # ... init srng ...\n        self.D = srng.normal(size=(dimensions, K))\n        self.D = self.D \/ numpy.sqrt(numpy.sum(numpy.sqr(self.D), axis=0)))\n        mini_batch = T.matrix('mini_batch', dtype='float64')\n        self.func = theano.function(inputs=[mini_batch], outputs=fit_once())\n\n    def update(batch):\n        self.D = self.func(batch)\n\n    def fit_once(mini_batch):\n        # ... do work to create S symbolically ...\n        D_update = T.dot(mini_batch, T.transpose(S))\n        D_update = D_update \/ T.sqrt(T.sum(T.sqr(D_update), axis=0))\n        return D_update\n\n    def get_D():\n        return self.D\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>As far as I can see there's no need to use <code>givens=...<\/code> at all.<\/p>\n"}]}
{"QuestionId":30556013,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-05-31T09:45:09.610","AcceptedAnswerId":30568365.0,"OwnerUserId":3301357.0,"Title":"How to flatten a calculation graph for a loop with accumulator(symbolic variables)?","Body":"<p>I'm a newbie in theano. But I've already googled,  read official theano documentation &amp; \nI haven't found any clue how to solve my problem.<\/p>\n\n<p>I'm trying to reinvent the wheel: I'm implementing my own batch convolution using theano.\n(I'm doing so to learn this library)<\/p>\n\n<p>So, here's what I'm trying to do:<\/p>\n\n<pre><\/pre>\n\n<p>This results in a very deep calculation graph, because inc_subtensor is added on top of each previous operation:<\/p>\n\n<p>inc_subtensor_stepN(inc_subtensor_stepN-1(inc_subtensor_stepN-2...<\/p>\n\n<p>So I tried to flatten it. As all variables are symbolic, I realised, that I have to substitute them in some way in graph.<\/p>\n\n<p>I tried theano.clone, but it results into the same situation as inc_subtensor.<\/p>\n\n<p>Then I tried to use theano.scan:<\/p>\n\n<pre><\/pre>\n\n<p>But still, I'm getting \"RuntimeError: maximum recursion depth exceeded in comparison\"\nright the first time when sym_im2col_prev_layer_curr_batch = theano.clone  is being executed.<\/p>\n\n<p>The latter code snippet example shows right the thing I'm going to do. But I have no idea why I'm getting a 'maximum recursion depth exceeded'.\nBecause each time I do theano.clone, theano is supposed to substitute sym_im2col_prev_layer_batch_idx (which is already used in scan) with \nit's exact symbolic value - im2col_prev_layer[batch_idx], and give me a copy of this subgraph.\nI might have missed something...<\/p>\n\n<p>How such(or similar) tasks are solved in theano &amp; how can I avoid too deep calculation graphs when\ndoing such tasks?<\/p>\n\n<p>Also I tried such approach:<\/p>\n\n<p>I've tried such approach:<\/p>\n\n<pre><\/pre>\n\n<p>But when trying to print the value of 'convolved' right after the 'for' cycle, I'm getting:<\/p>\n\n<pre><\/pre>\n\n<p>So, the same story.<\/p>\n\n<p>Increasing recursion depth for python is NOT an option.<\/p>\n\n<p>Any ideas how to flatten the calculation graph for my case?<\/p>\n","answers":[{"AnswerId":"30568365","CreationDate":"2015-06-01T07:56:27.487","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>In general <code>theano.scan<\/code> is the solution for recursive situations. In a case like yours <code>theano.scan<\/code> should be used to <em>replace<\/em> a Python <code>for<\/code> loop, not in addition to the <code>for<\/code> loop.<\/p>\n\n<p>It's difficult to see exactly what you're trying to achieve but the extensive use of <code>set_subtensor<\/code> and <code>inc_subtensor<\/code> suggests you're thinking about this in a way that is not well matched with how Theano wants to work. <code>theano.scan<\/code> might allow you to achieve what you want using the approach you're currently taking but, after a quick scan through the code you've provided, it doesn't look like even <code>theano.scan<\/code> is required. If one iteration does not rely on results from a previous iteration, as appears to be the case, then you can probably do this without any loops at all (neither Python <code>for<\/code> loops or a <code>theano.scan<\/code>) by judicious use of Theano tensor operations. The non-loop approach would almost certainly be far more efficient and speedy than doing things via a loop of some kind. Admittedly, these can be more difficult to wrap your head around than sequential, one-row-at-a-time type, operations.<\/p>\n\n<p>If you can't see how your computation might be achieved via plain multi-dimensional tensor operations without loops then I would suggest looking into how you can replace your Python <code>for<\/code> loops with as few <code>theano.scan<\/code> operations as you can get away with.<\/p>\n"}]}
{"QuestionId":30558439,"AnswerCount":1,"Tags":"<for-loop><matrix><lua><deep-learning><torch>","CreationDate":"2015-05-31T14:16:44.340","AcceptedAnswerId":30560653.0,"OwnerUserId":1056274.0,"Title":"Fast way to initialize a tensor in torch7","Body":"<p>I need to initialize a 3D tensor with an index-dependent function in torch7, i.e.<\/p>\n\n<pre><\/pre>\n\n<p>then I initialize a 3D tensor A like this:<\/p>\n\n<pre><\/pre>\n\n<p>But this code runs very slow, and I found it takes up 92% of total running time. Are there any more efficient ways to initialize a 3D tensor in torch7?<\/p>\n","answers":[{"AnswerId":"30560653","CreationDate":"2015-05-31T17:51:03.243","ParentId":null,"OwnerUserId":"2726734","Title":null,"Body":"<p>See the documentation for the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#applying-a-function-to-a-tensor\"><code>Tensor:apply<\/code><\/a><\/p>\n\n<blockquote>\n  <p>These functions apply a function to each element of the tensor on\n  which the method is called (self). These methods are much faster than\n  using a for loop in Lua.<\/p>\n<\/blockquote>\n\n<p>The example in the docs initializes a 2D array based on its index i (in memory). Below is an extended example for 3 dimensions and below that one for N-D tensors. Using the apply method is much, <em>much<\/em> faster on my machine:<\/p>\n\n<pre><code>require 'torch'\n\nA = torch.Tensor(100, 100, 1000)\nB = torch.Tensor(100, 100, 1000)\n\nfunction func(i,j,k) \n    return i*j*k    \nend\n\nt = os.clock()\nfor i=1,A:size(1) do\n    for j=1,A:size(2) do\n        for k=1,A:size(3) do\n            A[{i, j, k}] = i * j * k\n        end\n    end\nend\nprint(\"Original time:\", os.difftime(os.clock(), t))\n\nt = os.clock()\nfunction forindices(A, func)\n  local i = 1\n  local j = 1\n  local k = 0\n  local d3 = A:size(3)\n  local d2 = A:size(2) \n  return function()\n    k = k + 1\n    if k &gt; d3 then\n      k = 1\n      j = j + 1\n      if j &gt; d2 then\n        j = 1\n        i = i + 1\n      end\n    end\n    return func(i, j, k)\n  end\nend\n\nB:apply(forindices(A, func))\nprint(\"Apply method:\", os.difftime(os.clock(), t))\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>This will work for any Tensor object:<\/p>\n\n<pre><code>function tabulate(A, f)\n  local idx = {}\n  local ndims = A:dim()\n  local dim = A:size()\n  idx[ndims] = 0\n  for i=1, (ndims - 1) do\n    idx[i] = 1\n  end\n  return A:apply(function()\n    for i=ndims, 0, -1 do\n      idx[i] = idx[i] + 1\n      if idx[i] &lt;= dim[i] then\n        break\n      end\n      idx[i] = 1\n    end\n    return f(unpack(idx))\n  end)\nend\n\n-- usage for 3D case.\ntabulate(A, function(i, j, k) return i * j * k end)\n<\/code><\/pre>\n"}]}
{"QuestionId":30558465,"AnswerCount":1,"Tags":"<theano><deep-learning>","CreationDate":"2015-05-31T14:19:30.220","AcceptedAnswerId":30567486.0,"OwnerUserId":379539.0,"Title":"Theano multiplying by zero","Body":"<p>Can anybody explain to me what is the meaning behind these two lines of code from here: <a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/4_modern_net.py\" rel=\"nofollow\">https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/4_modern_net.py<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Is it a mistake? Why do we instantiate acc to zero and then multiply it by rho in next line? It looks like it will not achieve anything this way and remain zero. Will there be any difference if we replace \"rho * acc\" by just \"acc\"?<\/p>\n\n<p>The full function is given below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30567486","CreationDate":"2015-06-01T07:01:17.077","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This is just a way to tell Theano \"create a shared variable and initialize its value to be zero <em>in the same shape as p<\/em>.\"<\/p>\n\n<p>This <code>RMSprop<\/code> method is a <em>symbolic<\/em> method. It does not actually compute the RmsProp parameter updates, it only tells Theano how parameter updates should be computed when the eventual Theano function is executed.<\/p>\n\n<p>If you look further down <a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/4_modern_net.py\" rel=\"noreferrer\">the tutorial code you linked to<\/a> you'll see the symbolic execution graph for the parameter updates are constructed by <code>RMSprop<\/code> via a call on line 67. These updates are then compiled into a Theano function called <code>train<\/code> in Python on line 69 and the train function is executed many times on line 74 within the for loops of lines 72 and 73. The Python function <code>RMSprop<\/code> will be called only once, irrespective of how many times the <code>train<\/code> function is called within the for loops on lines 72 and 73.<\/p>\n\n<p>Within <code>RMSprop<\/code>, we are telling Theano that, for each parameter <code>p<\/code>, we need a new Theano variable whose <em>initial<\/em> value has the same shape as <code>p<\/code> and is 0. throughout. We then go on to tell Theano how it should update both this new variable (unnamed as far as Theano is concerned but named <code>acc<\/code> in Python) and how to update the parameter <code>p<\/code> itself. These commands do not alter either <code>p<\/code> or <code>acc<\/code>, they just tell Theano how <code>p<\/code> and <code>acc<\/code> should be updated later, once the function has been compiled (line 69) each time it is executed (line 74).<\/p>\n\n<p>The function executions on line 74 will <em>not<\/em> call the <code>RMSprop<\/code> Python function, they execute a compiled version of <code>RMSprop<\/code>. There will be no initialization inside the compiled version because that already happened in the Python version of RMSprop. Each <code>train<\/code> execution of the line <code>acc_new = rho * acc + (1 - rho) * g ** 2<\/code> will use the <em>current<\/em> value of <code>acc<\/code> not its initial value.<\/p>\n"}]}
{"QuestionId":30566793,"AnswerCount":1,"Tags":"<image><qt><lua><window><torch>","CreationDate":"2015-06-01T06:16:13.520","AcceptedAnswerId":30575161.0,"OwnerUserId":1457196.0,"Title":"Preserving aspect ratio when using torch's image.display","Body":"<p>I have the following very simple script written in lua. I am running it with qlua.<\/p>\n\n<p><\/p>\n\n<p>If the image is large the qt window simply takes the whole screen, which also stretches the image to fit the screen.<\/p>\n\n<p>I can't figure out a way to keep this from happening.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"30575161","CreationDate":"2015-06-01T13:42:45.787","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>If the image is large, resize it down to what you can configure as \"Max height\/Max width\", while preserving the aspect ratio.<\/p>\n\n<p>Sample code:<\/p>\n\n<pre><code>maxSize = 480\n-- find the smaller dimension, and resize it to maxSize (while keeping aspect ratio)\nlocal iW = input:size(3)\nlocal iH = input:size(2)\nif iW &lt; iH then\n   input = image.scale(input, maxSize, maxSize * iH \/ iW)\nelse\n   input = image.scale(input, maxSize * iW \/ iH, maxSize)\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":30569647,"AnswerCount":2,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-06-01T09:10:21.540","AcceptedAnswerId":30800488.0,"OwnerUserId":4924192.0,"Title":"Unable to run the Theano CNN.py file for a different dataset","Body":"<p>I have converted my dataset exactly in the form of the mnist.pkl.gz file and it runs for the logistic_sgd.py and mlp.py programs given in the Theano deeplearning tutorial PDF (University of Montreal). However on running the CNN.py file it gives a huge error which is difficult to understand. Can anyone who has succesfully run the CNN.py program for a different dataset help me out here as I am totally clueless about what the error is.\nMy training set has 1176 entries and my validation and test set has 168 entries each. Maybe the problem is with the batch size. If so can someone please suggest me an appropriate batch size?<\/p>\n\n<p>I am using the Spyder GUI for Python 2.7 that comes with the Anaconda Bundle.<\/p>\n\n<p>Error occurs soon after printing '... building model'<\/p>\n\n<p>Code Snippet:<\/p>\n\n<pre><\/pre>\n\n<p>Prompt at console as soon as error occurs:<\/p>\n\n<pre><\/pre>\n\n<p>Then the following error message appears.<\/p>\n\n<p>Error message:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30571192","CreationDate":"2015-06-01T10:30:37.090","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I encouraged you to add more information (thanks for that) so felt obliged to help as far as I can, but unfortunately that may not be much. Hopefully others can add better answers.<\/p>\n\n<p>I don't think the error is occurring within the code snippet you provided. Yes, it's after the '... building the model' message, but I think it's further on. It's probably generating the error at a <code>theano.function<\/code> call.<\/p>\n\n<p>Your problem appears to be in the C++ toolchain on your system. I would suggest searching for more information with respect to the \"undefined reference to __imp_PyExc_ImportError\" message.<\/p>\n\n<p>For example, <a href=\"https:\/\/stackoverflow.com\/questions\/2842469\/python-undefined-reference-to-imp-py-initmodule4\">this StackOverflow question<\/a> suggests adding <code>-D MS_WIN64<\/code> to the command line. But this is probably an internal command line, not the one you are using originally.<\/p>\n\n<p>I don't understand why it works for one dataset but not another.<\/p>\n"},{"AnswerId":"30800488","CreationDate":"2015-06-12T10:09:10.017","ParentId":null,"OwnerUserId":"4924192","Title":null,"Body":"<p>There was some problem in my laptop only. It is running fine on a different computer.<\/p>\n"}]}
{"QuestionId":30576509,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-06-01T14:45:30.767","AcceptedAnswerId":30596963.0,"OwnerUserId":661935.0,"Title":"Why indexing a Theano tensor variable with a list of `slice` objects is invalid?","Body":"<p>In an <a href=\"https:\/\/github.com\/goodfeli\/theano_exercises\/blob\/master\/01_basics\/03_advanced_expressions\/02_nd_indexing.py\" rel=\"nofollow\">exercise<\/a> of Dr. Goodfellow's Theano tutorial, it's ok to slice  with tuple , but Theano will raise exception for  . <\/p>\n\n<p>Exception info:<\/p>\n\n<blockquote>\n  <p>theano.tensor.var.AsTensorError: ('Cannot convert [slice(, Elemwise{neg,no_inplace}.0, None), slice(, Elemwise{neg,no_inplace}.0, None), slice(, Elemwise{neg,no_inplace}.0, None)] to TensorType', )<\/p>\n<\/blockquote>\n\n<p>Why doesn't it work with ? BTW, slicing a tensor variable with an integer list is ok.\nI've red the <a href=\"https:\/\/theano.readthedocs.org\/en\/latest\/library\/tensor\/basic.html#indexing\" rel=\"nofollow\">document<\/a>, but didn't find the reason.<\/p>\n","answers":[{"AnswerId":"30596963","CreationDate":"2015-06-02T13:04:23.403","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This was a bug in Theano. It has been fixed by Fr\u00e9d\u00e9ric Bastien via <a href=\"https:\/\/github.com\/Theano\/Theano\/pull\/2992\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/pull\/2992<\/a><\/p>\n\n<p>More here: <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/nTRfigJD19w\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/nTRfigJD19w<\/a><\/p>\n"}]}
{"QuestionId":30582388,"AnswerCount":2,"Tags":"<macos><lua><osx-yosemite><libpng><torch>","CreationDate":"2015-06-01T20:11:47.577","AcceptedAnswerId":30620988.0,"OwnerUserId":3993741.0,"Title":"Lua error loading module 'libpng' (Torch, MacOSX)","Body":"<p>How do I make libpng load properly in Lua?  I am running Lua\/Torch in iTorch Notebook in Mac OSX 10.10.3, where other basic functions in Lua work, such as plotting and calculations.  <\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Warning: libpng-1.6.17 already installed<\/p>\n<\/blockquote>\n\n<p>If I run:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>error loading module 'libpng' from file '\/usr\/local\/lib\/lua\/5.1\/libpng.so':\n      dlopen(\/usr\/local\/lib\/lua\/5.1\/libpng.so, 6): Library not loaded: \/usr\/local\/lib\/libpng15.15.dylib\n    Referenced from: \/usr\/local\/lib\/lua\/5.1\/libpng.so\n    Reason: Incompatible library version: libpng.so requires version 33.0.0 or later, but libpng15.15.dylib provides version 29.0.0\n  warning:  could not be loaded (is it installed?)\n  \/usr\/local\/share\/lua\/5.1\/dok\/inline.lua:736:  libpng package not found, please install libpng\n  stack traceback:\n      [C]: in function 'error'\n      \/usr\/local\/share\/lua\/5.1\/dok\/inline.lua:736: in function 'error'\n      \/usr\/local\/share\/lua\/5.1\/image\/init.lua:142: in function 'saver'\n      \/usr\/local\/share\/lua\/5.1\/image\/init.lua:355: in function 'save'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/itorch\/gfx.lua:25: in function 'f'\n      [string \"local f = function() return itorch.image(iii)...\"]:1: in main chunk\n      [C]: in function 'xpcall'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/itorch\/main.lua:177: in function \n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/lzmq\/poller.lua:75: in function 'poll'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/lzmq\/impl\/loop.lua:307: in function 'poll'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/lzmq\/impl\/loop.lua:325: in function 'sleep_ex'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/lzmq\/impl\/loop.lua:370: in function 'start'\n      \/Users\/MY\/torch\/install\/share\/lua\/5.1\/itorch\/main.lua:344: in main chunk\n      [C]: in function 'require'\n      [string \"arg={'\/Users\/MY\/.ipython\/profile_default\/secu...\"]:1: in main chunk<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"30583013","CreationDate":"2015-06-01T20:48:17.250","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Reinstalling the image package as well as forcing the linkage of libpng might fix it:<\/p>\n\n<pre><code>brew link libpng --force\nluarocks install image\n<\/code><\/pre>\n"},{"AnswerId":"30620988","CreationDate":"2015-06-03T13:06:13.423","ParentId":null,"OwnerUserId":"4455483","Title":null,"Body":"<p>I had a similar problem (OSX 10.9.5). You probably have multiple versions of libpng installed, with the one called during install of luarocks having architecture i386 (x86_64 required).<\/p>\n\n<p>To solve this:<\/p>\n\n<ol>\n<li><p>Try installing image again, and reading the log:<\/p>\n\n<p>luarocks install image<\/p><\/li>\n<li><p>Check the log to see if you get a message of type:<\/p>\n\n<p>ld: warning: ignoring file \/Library\/Frameworks\/\/libpng.framework\/libpng, missing required architecture x86_64 in file \/Library\/Frameworks\/\/libpng.framework\/libpng (2 slices)<\/p><\/li>\n<li><p>If this is the case (assuming using brew) remove the libpng framework in \/Library\/Frameworks and do a <\/p>\n\n<p>brew install libpng --universal<\/p><\/li>\n<li><p>Reinstall image and run.<\/p><\/li>\n<\/ol>\n\n<p>This worked for me, I hope it works for you too.<\/p>\n"}]}
{"QuestionId":30591099,"AnswerCount":0,"Tags":"<python-2.7><installation><neural-network><theano>","CreationDate":"2015-06-02T08:34:03.110","AcceptedAnswerId":null,"OwnerUserId":4774186.0,"Title":"Installation nolearn issue","Body":"<p>I want to use nolearn but I am not able to install it. I can' t use pip installation so I tried as in the figure. Any idea how can I solve this? Thank you\n<img src=\"https:\/\/i.stack.imgur.com\/ZeDQs.png\" alt=\"enter image description here\"><\/p>\n","answers":[]}
{"QuestionId":30594858,"AnswerCount":2,"Tags":"<python><installation><theano><nvcc>","CreationDate":"2015-06-02T11:29:35.930","AcceptedAnswerId":30596660.0,"OwnerUserId":4774186.0,"Title":"Theano installation, nvcc not in the path","Body":"<p>I have installed theano on windows7,64bit on  winpython using their guide <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a> and I thought it worked since when I ran their first example I did have the expected results and no errors. I wanted to go ahead and install the part: <strong>Configure Theano for GPU use<\/strong> but when I runned again it  I had this in the python console: <\/p>\n\n<pre><\/pre>\n\n<p>the .theanorc file I am using is:<\/p>\n\n<pre><\/pre>\n\n<p>and I added it in C:\\SciSoft\\WinPython-64bit-2.7.9.4\\settings as I understood from the guide. <\/p>\n\n<p>By the way, I checked the C:\\SciSoft\\env.bat and when I write <strong>where nvcc<\/strong> it says not file found instead I have no problems with the other checkings.  Is that because I haven't an NVIDIA card? I am totaly lost.Any help? Thank you <\/p>\n","answers":[{"AnswerId":"30596660","CreationDate":"2015-06-02T12:51:01.147","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano is designed to work (almost) identically on both CPU and GPU. You don't need a GPU to use Theano and if you don't have a Nvidia GPU then you shouldn't try installing any GPU-specific stuff at all.<\/p>\n"},{"AnswerId":"40499036","CreationDate":"2016-11-09T01:21:54.670","ParentId":null,"OwnerUserId":"6831692","Title":null,"Body":"<p>aleju, if you don't want (or can't) use theano with GPU, you just need to change .theanorc to use only cpu. This will not cause any problem, except for poor performance.<\/p>\n\n<pre><code>[global]\ndevice = cpu\n...\n<\/code><\/pre>\n"}]}
{"QuestionId":30595818,"AnswerCount":1,"Tags":"<numpy><linear-algebra><theano>","CreationDate":"2015-06-02T12:15:37.683","AcceptedAnswerId":30596094.0,"OwnerUserId":1937197.0,"Title":"How to express c[i,j,k] = a[i,j] * b[i,k] in Numpy\/Theano?","Body":"<p>The definition <\/p>\n\n<pre><\/pre>\n\n<p>is an element-wise product with respect to , and an outer product with respect to  and . Is there any way to express this in NumPy\/Theano without loops?<\/p>\n","answers":[{"AnswerId":"30596094","CreationDate":"2015-06-02T12:28:16.050","ParentId":null,"OwnerUserId":"1937197","Title":null,"Body":"<p>I found a solution that works with both Numpy and Theano:<\/p>\n\n<pre><code>c = a[:, :, np.newaxis] * b[:, np.newaxis, :]\n<\/code><\/pre>\n"}]}
{"QuestionId":30627302,"AnswerCount":1,"Tags":"<python-2.7><numpy><theano>","CreationDate":"2015-06-03T17:54:18.383","AcceptedAnswerId":30651282.0,"OwnerUserId":2890103.0,"Title":"Can not update a subset of shared tensor variable after a cast","Body":"<p>I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Which is compiling fine (where  is a numpy array of dtype ).<\/p>\n\n<p>To prevent future type error I want to cast my shared tensor  into  (or  which is equivalent as I have set  to  in the config file).<\/p>\n\n<p>I so add  and then I get the following error: <\/p>\n\n<p>. <\/p>\n\n<p>I do not understand why. According to this <a href=\"https:\/\/stackoverflow.com\/questions\/15917849\/how-can-i-assign-update-subset-of-tensor-shared-variable-in-theano\">question<\/a>, using  should allow me to update a subset of the shared variable.<\/p>\n\n<p>How can I cast my shared tensor while being able to update it?<\/p>\n","answers":[{"AnswerId":"30651282","CreationDate":"2015-06-04T18:19:07.840","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you are trying to update a symbolic variable, not a shared variable.<\/p>\n\n<pre><code>U = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.float64)\nWords = theano.shared(value=U, name='Words')\nzero_vec_tensor = T.vector()\nset_zero = theano.function([zero_vec_tensor], updates=[(Words, T.set_subtensor(Words[0, :], zero_vec_tensor))])\n<\/code><\/pre>\n\n<p>works fine because the thing you are updating, <code>Words<\/code> is a shared variable.<\/p>\n\n<pre><code>U = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.float64)\nWords = theano.shared(value=U, name='Words')\nWords = T.cast(Words, dtype = theano.config.floatX)\nzero_vec_tensor = T.vector()\nset_zero = theano.function([zero_vec_tensor], updates=[(Words, T.set_subtensor(Words[0, :], zero_vec_tensor))])\n<\/code><\/pre>\n\n<p>does not work because now Words is no longer a shared variable, it is a symbolic variable that, when executed, will compute a cast the values in the shared variable to <code>theano.config.floatX<\/code>.<\/p>\n\n<p>The <code>dtype<\/code> of a shared variable is determined by the value assigned to it. So you probably just need to change the type of U:<\/p>\n\n<pre><code>U = np.array([[1, 2, 3], [4, 5, 6]], dtype=theano.config.floatX)\n<\/code><\/pre>\n\n<p>Or cast it using numpy instead of symbolically:<\/p>\n\n<pre><code>U = np.dtype(theano.config.floatX).type(U)\n<\/code><\/pre>\n"}]}
{"QuestionId":30628900,"AnswerCount":1,"Tags":"<python><eclipse><python-2.7><cuda><theano>","CreationDate":"2015-06-03T19:26:14.697","AcceptedAnswerId":null,"OwnerUserId":3878508.0,"Title":"Theano Test File Will Not Compile","Body":"<p>I have been trying to install Theano for windows 7 64 bit machine based off of the tutorial found on the website <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#visual-studio-and-cuda\" rel=\"nofollow\">here<\/a>. I have gotten almost everything to work but after installing CUDA 5.5 and then continuing to on verifying the programs with these commands:<\/p>\n\n<p>\"Please do so, and verify that the following programs are found:<\/p>\n\n<ol>\n<li>where gcc<\/li>\n<li>where gendef<\/li>\n<li>where cl<\/li>\n<li>where nvcc\"<\/li>\n<\/ol>\n\n<p>The first three work fine but the last one returns \"INFO: Could not find files for the given pattern(s).\" I am not sure why because I installed CUDA and nvcc should be found. This is causing a larger problem because when I try to run this test file:<\/p>\n\n<pre><\/pre>\n\n<p>Eclipsed throws the error at the fourth line under config saying \"Undefined variable from import: config\". Then when I run it anyways the error in the console is \"AttributeError: 'module' object has no attribute 'config'\"<\/p>\n\n<p>Any suggestions or advice on any of this is much appreciated.<\/p>\n","answers":[{"AnswerId":"30666848","CreationDate":"2015-06-05T12:42:25.017","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The first problem was fixed in the comments on the question.<\/p>\n\n<p>For the import, I suggest that you uninstall Theano. Do this many times to make sure you remove all version. Depending how you installed python, it could have installed an old version of Theano at the same time.<\/p>\n\n<p>Then install Theano development version.<\/p>\n\n<p>Then it can't find theano.config, most of the time is because there is problem in the installation or you use an old version that had problems related to Windows.<\/p>\n"}]}
{"QuestionId":30632204,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-06-03T22:52:52.313","AcceptedAnswerId":null,"OwnerUserId":1758727.0,"Title":"What is the 'name' field used for in the definition of a theano shared variable?","Body":"<p>I notice that you can define a theano shared variable by:<\/p>\n\n<pre><\/pre>\n\n<p>In theano doc <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/shared.html#module-shared\" rel=\"nofollow\">here<\/a>, it says the name field is \"the name for this variable\". I see in most cases people just pass a exactly same string to the name field as the shared variable's name in Python. For example:<\/p>\n\n<pre><\/pre>\n\n<p>What is the rule of thumb to define the name field? What is the name field used for?<\/p>\n","answers":[{"AnswerId":"30651047","CreationDate":"2015-06-04T18:04:57.397","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Assigning names to Theano objects helps during debugging. For example, <code>theano.printing.debugprint<\/code> will include the name, if present, in its output; this can make understanding a complex computation graph easier.<\/p>\n\n<p>You can also give names to input variables (e.g. <code>x=theano.tensor.scalar('x')<\/code>) and functions (e.g. <code>f=theano.function(inputs, outputs, name='f')<\/code>.<\/p>\n\n<p>The output of <code>theano.printing.debugprint<\/code> can be limited by naming nodes whose content you don't need to see and using the <code>stop_on_name<\/code> parameter.<\/p>\n\n<p>Keep Python names and Theano names in sync can be tedious and error prone. In some circumstances you might like to use functions such as:<\/p>\n\n<pre><code>def name_node(variable, variable_names):\n    for name, query_variable in variable_names.iteritems():\n        if query_variable is variable:\n            variable.name = name\n\n    return variable\n\n\ndef name_nodes(variables, variable_names):\n    if not isinstance(variables, (tuple, list)):\n        return name_node(variables, variable_names)\n\n    for variable in variables:\n        for name, query_variable in variable_names.iteritems():\n            if query_variable is variable:\n                variable.name = name\n\n    return variables\n<\/code><\/pre>\n\n<p>In a symbolic function you can use these like this:<\/p>\n\n<pre><code>def create_graph(w_filename):\n    x = T.scalar()\n    w = theano.shared(numpy.load(w_filename))\n    y = T.tanh(theano.dot(x, w))\n    return name_nodes([x, w, y], locals())\n<\/code><\/pre>\n\n<p>This is a bit of a Python hack and won't be suitable in all circumstances.<\/p>\n"}]}
{"QuestionId":30633181,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-06-04T00:43:07.473","AcceptedAnswerId":30653254.0,"OwnerUserId":3353215.0,"Title":"The output size of theano.tensor.nnet.conv.conv2d","Body":"<p>The function that is currently being used widely on tutorials and other place is of the form:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li><p>If for the first layer of a CNN, I have  as  which is the number of kernals being 20, each 7 X 7, what does the '1' stand for ? My  is .<\/p><\/li>\n<li><p>This convolution now outputs a tensor of shape  which I understand. My next layer now takes the parameters  =  ,  =  and produces a output of shape  . I seem to kind of understand this operation, except, if I want to use a '50' filters layer each working on previous 20 feature maps produced, shouldn't I produce 1000 feature maps in all instead of producing just 50 feature maps ? To restate my question, I have a stack of 20 feature maps each running 50 kernals of convolution, shouldn't my output shape be  instead of  ?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"30653254","CreationDate":"2015-06-04T20:12:42.910","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>To answer your questions:<\/p>\n\n<ol>\n<li><p>The <code>1<\/code> stands for the number of input channels. As you seem to be using gray scale images, this is one. For color images it can be 3. For other convolutional layers as in your second question, it must be equal to the number of outputs that the previous layer generated.<\/p><\/li>\n<li><p>Using a filter of size <code>[50, 20, 5, 5]<\/code> on an input signal of <code>[100, 20, 26, 26]<\/code> is actually a good example for your first question, as well. You have here 50 filters of shape <code>[20, 5, 5]<\/code>. Every image is of shape <code>[20, 26, 26]<\/code>. The convolution uses all the 20 channels each time: Filter 0 gets applied to image channel 0, filter 1 gets applied to image 1, and the whole result gets summed up. Does that make sense?<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":30640102,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-06-04T09:29:39.813","AcceptedAnswerId":30642674.0,"OwnerUserId":1157605.0,"Title":"Theano subtraction when you can't tile","Body":"<p>Say we have a theano matrix X that is nxm, and another one u that is nx1.  We want to do , but if we do that we'll get an input dimension mismatch.  We could try tiling u, but tile only accepts constants and not variables.  How do we do this?<\/p>\n\n<pre><\/pre>\n\n<p>I then get the error <\/p>\n","answers":[{"AnswerId":"30642674","CreationDate":"2015-06-04T11:28:46.863","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p><code>X - u<\/code> should work exactly as you write it by broadcasting:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nn = 10\nm = 20\n\nX = T.arange(n * m).reshape((n, m))\nu = T.arange(0, n * m, m).reshape((n, 1))\n\nr = X - u\n\nr.eval()\n<\/code><\/pre>\n\n<p>Similar to your updated question, you can do<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nX = T.dmatrix()\nu = T.addbroadcast(T.dmatrix(), 1)\n\nr = X - u\n\nf = theano.function([X, u], r)\n\nXX = np.arange(20.).reshape(2, 10)\nuu = np.array([1., 100.]).reshape(2, 1)\n\nf(XX, uu)\n<\/code><\/pre>\n"}]}
{"QuestionId":30663780,"AnswerCount":2,"Tags":"<debugging><gdb><torch>","CreationDate":"2015-06-05T10:03:48.793","AcceptedAnswerId":null,"OwnerUserId":1851417.0,"Title":"how to configure gdb to debug a script not a binary [gdb : file format not recognized]","Body":"<p>I'm trying to use gdb to debug  <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow\">Torch library<\/a> binary file to . When I run from the command line :\ngdb --args th <\/p>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I checked if my current installation of  is 64 bit i installed gdb64 and when i run <\/p>\n\n<pre><\/pre>\n\n<p>I still get the same error, the output of :<\/p>\n\n<pre><\/pre>\n\n<p>is : <\/p>\n\n<pre><\/pre>\n\n<p>I have learned that the problem is that the executable file is not a binary, but a script, so gdb is trying to debug the script instead. <\/p>\n\n<p>My question is how to overcome this and let gdb debug the execution of the command itself. or even replace the Torch installation to be a binary execution instead of a script.<\/p>\n","answers":[{"AnswerId":"30679176","CreationDate":"2015-06-06T05:13:46.517","ParentId":null,"OwnerUserId":"1851417","Title":null,"Body":"<p>with some help from the comments i was able to run gdb over the torch script, through : <\/p>\n\n<pre><code>gdb64 \/bin\/bash    # check your gdb configuration either it's i686 or x86_64 \nrun \/path\/to\/th    # th is the torch running script to be debugged\n<\/code><\/pre>\n"},{"AnswerId":"31819576","CreationDate":"2015-08-04T21:22:10.613","ParentId":null,"OwnerUserId":"55075","Title":null,"Body":"<p>Try <a href=\"https:\/\/en.wikipedia.org\/wiki\/LLDB_(debugger)\" rel=\"nofollow\">LLDB Debugger<\/a> (<code>lldb<\/code>) instead which aims to replace GNU Debugger (<code>gdb<\/code>).<\/p>\n\n<p>It's available by default on BSD\/OS X, on Linux install via: <code>sudo apt-get install lldb<\/code> (or use <code>yum<\/code>). <\/p>\n\n<p>For usage, check <a href=\"http:\/\/lldb.llvm.org\/lldb-gdb.html\" rel=\"nofollow\">gdb to lldb command map<\/a> page.<\/p>\n"}]}
{"QuestionId":30663902,"AnswerCount":1,"Tags":"<matlab><caffe><conv-neural-network><leveldb><matcaffe>","CreationDate":"2015-06-05T10:10:32.150","AcceptedAnswerId":30869166.0,"OwnerUserId":2191652.0,"Title":"Convert data to leveldb for caffe","Body":"<p>I have a bunch of 2D data matrices in Matlab (no image data, but some single precision data). <\/p>\n\n<p>Does anyone know how to convert 2D matlab matrices to the leveldb format which is required by caffe to train a custom neural network?<\/p>\n\n<p>I already did the tutorial on how to train on images (using the imagenet architecture) and on mnist (digit recognition dataset). However in the latter example they didn't show how to create the respective database. In the tutorial the database was already provided.<\/p>\n","answers":[{"AnswerId":"30869166","CreationDate":"2015-06-16T13:34:33.200","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>I still don't know to create a leveldb database of my 2D data matrices for usage in caffe but I finally solved by problem:<br>\nI ended up using <a href=\"https:\/\/stackoverflow.com\/questions\/30663902\/convert-data-to-leveldb-for-caffe#comment49431386_30663902\">Shai's proposal<\/a> to convert the data to HDF5 format. It is quite easy to read and write HDF5 databases in Matlab. You just have to use the functions <code>hdf5info()<\/code>,<code>h5read()<\/code>,<code>h5create()<\/code> and <code>h5write()<\/code> which are already implemented in Matlab.<\/p>\n\n<p><strong>Example:<\/strong><br>\n- Change the data type in your caffe prototxt file to \"hdf5layer\", like this:<\/p>\n\n<pre><code>name: \"LeNet\"\nlayer {\n  name: \"mnist\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TRAIN\n  }\n  hdf5_data_param {\n    source: \"\/path\/to\/your\/database\/myMnist_train.txt\"\n    batch_size: 64\n  }\n}\n<\/code><\/pre>\n\n<p>Use Matlab to create HDF5 databases:<br>\n- Caffe: Your input training data has to be a 4-D matrix where the last two dimensions are equal to the size of your 2D input data matrix in matlab.<br>\n- Example: Take a 2d matrix (image or single precision data) of size 54x24 (#rows x cols)<br>\n- -> transpose it, and stack it into a 24x54x1xN matrix, where N is the number of 2d matrices (training samples)<br>\n- The labels are in a 1xN row vectors in matlab.<br>\n- Now create your hdf5 database:<\/p>\n\n<pre><code>h5create(['train.h5'],'\/data',[24 54 1 length(trainLabels)]);\nh5create(['train.h5'],'\/label',[1 length(trainLabels)]);\nh5write(['train.h5'],'\/data',trainData);\nh5write(['train.h5'],'\/label',trainLabels);\n<\/code><\/pre>\n\n<ul>\n<li>As you can see, caffe expects a hdf5 database with the variables \"data\" and \"label\"<\/li>\n<li>Reading a database:<br>\nUse <code>hdf5info(filename)<\/code> to get the dataset names inside a hdf5 database.\nThen use <code>data = h5read(filename,dataset)<\/code> to read the dataset<\/li>\n<\/ul>\n"}]}
{"QuestionId":30687089,"AnswerCount":2,"Tags":"<c++><armadillo><caffe>","CreationDate":"2015-06-06T20:07:21.983","AcceptedAnswerId":30700276.0,"OwnerUserId":3733814.0,"Title":"Armadillo in Caffe for pseudo-inverse and transpose","Body":"<p>I need pseudo inverse and transpose function to implement a Layer in caffe. So I am using Armadillo library to do that. But How can I convert Caffe Blobs(2-D) to armadillo Mat and vice-versa ?<\/p>\n","answers":[{"AnswerId":"30688162","CreationDate":"2015-06-06T22:24:07.333","ParentId":null,"OwnerUserId":"4092300","Title":null,"Body":"<p>A Caffe blob is a 4-dimensional matrix, I believe armadillo's <code>pinv(A)<\/code> and <code>.t()<\/code> or <code>trans(A)<\/code> are meant to be used with 2 dimensional matrices of type <code>mat<\/code>.<\/p>\n\n<p>You could get a vector of a vector of 2-dimensional armadillo matrices to represent a 4-dimensional Caffe blob. You can do so with something like:<\/p>\n\n<pre><code>using namespace arma;\nusing namespace caffe;\n\nvector&lt;vector&lt;mat&gt;&gt; blob2vvmat (Blob m) {\n\n    vector&lt;vector&lt;mat&gt;&gt; vvm;\n\n    for (int i=0; i&lt;m.shape().at(0); i++) {\n\n        vector&lt;mat&gt; vm;\n\n        for (int j=0; i&lt;m.shape().at(1); i++) {\n\n            mat M (m.shape().at(2), m.shape().at(3));\n\n            for (int k=0; i&lt;m.shape().at(2); i++) {\n\n                for (int l=0; i&lt;m.shape().at(3); i++) {\n\n                     M(k,l) = m.data_at(i, j, k, l);\n\n                }\n\n            }\n\n            vm.push_back(M);\n\n        }\n\n        vvm.push_back(vm);\n\n    }\n\n    return vvm;\n}\n<\/code><\/pre>\n\n<p>I haven't testing the code, but that should theoretically work, unless you're not looking for a vector of vector of matrices.<\/p>\n\n<p>If you know your row, column, height and 4d sizes, you can use the symbolic library on MATLAB or octave, or even better: Mathematica to derive and equation to calculate the pseudo-inverse and transpose of a 4d matrix and port it to your program using the <code>ccode(expr)<\/code> function.<\/p>\n\n<p>Hope this helps!<\/p>\n"},{"AnswerId":"30700276","CreationDate":"2015-06-08T01:53:11.040","ParentId":null,"OwnerUserId":"3733814","Title":null,"Body":"<p>Although, I haven''t found a way of conversion between armadillo and caffe blobs.\nBut I noticed that caffe uses MKL library for CPU computations, which has both of the required functions i.e. pseudo inverse and transpose.\nIn case, caffe is not configured to use MKL, I can easily implement a function for transpose (which would be better than getting it done using armadillo and then converting data types in O(mn) time).<\/p>\n"}]}
{"QuestionId":30689685,"AnswerCount":1,"Tags":"<python><matrix><gpgpu><theano>","CreationDate":"2015-06-07T03:07:23.297","AcceptedAnswerId":null,"OwnerUserId":1293964.0,"Title":"How to toggle theano matrix based on vector of int position","Body":"<p>Using theano tensor operations, how can I toggle one cell on each row of a matrix based on a integer position indicator on the correspond row index of a vector (i.e. |v| = rows of the matrix). For example, given a 100x5 matrix of zeros<\/p>\n\n<pre><\/pre>\n\n<p>and a 100-element vector of integer in the range of [0, 4]. <\/p>\n\n<pre><\/pre>\n\n<p>update (or create another) matrix M to<\/p>\n\n<pre><\/pre>\n\n<p>(I know how to do this iteratively using conventional codes, but I want to run it as part of an algorithm on GPU without complicating my input which is currently vector V, so a direct theano implementation would be great.)<\/p>\n","answers":[{"AnswerId":"30690168","CreationDate":"2015-06-07T04:51:41.450","ParentId":null,"OwnerUserId":"1293964","Title":null,"Body":"<p>I figured out the answer myself. This operation is known as <a href=\"http:\/\/en.wikipedia.org\/wiki\/One-hot\" rel=\"nofollow\">one-hot<\/a> and it is supported as the \"to_one_hot\" in <a href=\"http:\/\/www.deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html\" rel=\"nofollow\">Theano's extra_ops package<\/a>. Code:<\/p>\n\n<pre><code>M_one_hot = theano.tensor.extra_ops.to_one_hot(V, 5, dtype='int32')\n<\/code><\/pre>\n"}]}
{"QuestionId":30692742,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-06-07T11:00:34.527","AcceptedAnswerId":30696815.0,"OwnerUserId":4983163.0,"Title":"Creating square matrix with eigenvalues on the diagonal in theano","Body":"<p>I would like to create a square matrix with the eigenvalues on the diagonal:<\/p>\n\n<pre><\/pre>\n\n<p>However apparently theano does not treat the D matrix I created as a standard matrix, thus I cannot use it in the succeding computations.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30696815","CreationDate":"2015-06-07T18:10:28.147","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You're using an operation class as if it were an operation function.<\/p>\n\n<p>Instead of <\/p>\n\n<pre><code>D = T.nlinalg.AllocDiag(eigen_values)\n<\/code><\/pre>\n\n<p>try<\/p>\n\n<pre><code>D = T.nlinalg.AllocDiag()(eigen_values)\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>D = T.nlinalg.alloc_diag(eigen_values)\n<\/code><\/pre>\n"}]}
{"QuestionId":30700455,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2015-06-08T02:17:34.007","AcceptedAnswerId":null,"OwnerUserId":4984613.0,"Title":"Caffe layer registry error","Body":"<p>I'm new to Caffe and I have a problem running the Caffe mnist example. The error message is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I've searched for solutions and tried linking against dynamic library as suggested in <a href=\"https:\/\/stackoverflow.com\/questions\/30325108\/caffe-layer-creation-failure\">this post<\/a>. However it does not work. I can see the known layers is empty. What could be the cause? Please help me out. Thanks. I'm using Ubuntu 15.04.<\/p>\n","answers":[{"AnswerId":"39996488","CreationDate":"2016-10-12T10:36:56.293","ParentId":null,"OwnerUserId":"7005820","Title":null,"Body":"<p>I'm not sure if you are using the original solver.prototxt, the problem seems to be you define a wrong layer in the prototxt<\/p>\n"},{"AnswerId":"57052580","CreationDate":"2019-07-16T07:56:22.603","ParentId":null,"OwnerUserId":"5495014","Title":null,"Body":"<p>Please use the CMake when build the caffe from source. I also had the different types of layer mismatching. CMake will fix the al issue.<\/p>\n"}]}
{"QuestionId":30710355,"AnswerCount":1,"Tags":"<python><linux><anaconda><caffe>","CreationDate":"2015-06-08T13:25:24.033","AcceptedAnswerId":30713187.0,"OwnerUserId":2191652.0,"Title":"Cannot import caffe into python, libjpeg.so.62 not found","Body":"<p>I cannot import caffe into (anaconda-) python. \nI'm following a <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/hdf5_classification.ipynb\" rel=\"nofollow\">notebook example<\/a> on \"logistic regression on non-image HDF5 data\". When I execute the line<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The library  is definetly installed under . I don't know what is going wrong here or how to tell anacondapython where to look for .<\/p>\n\n<p>I already tried out  but apt-get says <\/p>\n\n<p>I compiled caffe while modifying \"Makefile.config\" such that it was pointing it to the ananconda python path. I also exported the PYTHONPATH and PATH of my anaconda directory:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30713187","CreationDate":"2015-06-08T15:25:47.970","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>Ok I finally found the solution:<\/p>\n\n<p>I had to <code>sudo apt-get install libjpeg62<\/code><\/p>\n\n<p>After that a new error occurred while trying to <code>import caffe<\/code>, namely<\/p>\n\n<pre><code>ImportError: \/home\/myName\/libs\/anaconda\/bin\/..\/lib\/libm.so.6: version `GLIBC_2.15' not found (required by \/usr\/lib\/x86_64-linux-gnu\/libx264.so.142)\n<\/code><\/pre>\n\n<p>That could be solved by removing some buggy anaconda libraries thus resorting to the system libraries,quote shelhamer:\n\"Some versions of Anaconda seem to come with a bad libm. <code>rm ~\/anaconda\/lib\/libm.*<\/code> takes care of this by reverting to the system libm.\"<\/p>\n\n<p>see <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/985\" rel=\"nofollow\">github bvlc<\/a><\/p>\n"}]}
{"QuestionId":30717208,"AnswerCount":2,"Tags":"<python><deep-learning><caffe>","CreationDate":"2015-06-08T19:11:33.030","AcceptedAnswerId":30717972.0,"OwnerUserId":4561745.0,"Title":"Two errors while running Caffe","Body":"<p>I've installed Caffe on my Ubuntu 14.04 machine. the  runs perfectly fine with 581 tests passed. I'm trying to work with the command line and python interface and getting the following two errors:<\/p>\n\n<ol>\n<li><p><strong>Command Line Interface:<\/strong> When I try to run the command , I'm getting the following error:<\/p>\n\n<pre><\/pre><\/li>\n<li><p><strong>Python Interface:<\/strong> When I run the command , I'm getting the following error:<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n","answers":[{"AnswerId":"34112166","CreationDate":"2015-12-05T23:28:00.577","ParentId":null,"OwnerUserId":"3138238","Title":null,"Body":"<p>Regarding the second problem you had, I had the same problem, I solved uncommenting this line inside the Makefile.config:<\/p>\n\n<pre><code># Decomment le line uncommented below:\n# Homebrew installs numpy in a non standard path (keg only)\nPYTHON_INCLUDE += $(dir $(shell python -c 'import numpy.core; print(numpy.core.__file__)'))\/include\n# PYTHON_LIB += $(shell brew --prefix numpy)\/lib\n<\/code><\/pre>\n"},{"AnswerId":"30717972","CreationDate":"2015-06-08T19:58:15.953","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Make sure that numpy is installed correctly and the path is mentioned to detect the newly installed library. The <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#prerequisites\" rel=\"nofollow\">steps<\/a> are provided in Caffe website itself.<\/p>\n\n<p>For the 'Caffe' command to work, you will have to step into the folder where the 'Caffe' executable is created and then try running the executable through the terminal.<\/p>\n"}]}
{"QuestionId":30729071,"AnswerCount":1,"Tags":"<python><amazon-web-services><neural-network><theano>","CreationDate":"2015-06-09T10:09:23.167","AcceptedAnswerId":null,"OwnerUserId":1372512.0,"Title":"Run out of VRAM using Theano on Amazon cluster","Body":"<p>I'm trying to execute the <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"nofollow noreferrer\">logistic_sgd.py<\/a> code on an Amazon cluster running the ami-b141a2f5 (Theano - CUDA 7) image. <\/p>\n\n<p>Instead of the included MNIST database I am using the SD19 database, which requires changing a few dimensional constants, but otherwise no code has been touched. The code runs fine locally, on my CPU, but once I SSH the code and data to the Amazon cluster and run it there, I get this output:\n<img src=\"https:\/\/i.stack.imgur.com\/lakAK.png\" alt=\"enter image description here\"><\/p>\n\n<p>It looks to me like it is running out of VRAM, but it was my understanding that the code should run on a GPU already, without any tinkering on my part necessary. After following the suggestion from the error message, the error persists.<\/p>\n","answers":[{"AnswerId":"33219275","CreationDate":"2015-10-19T16:17:28.783","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There's nothing especially strange here. The error message is almost certainly accurate: there really isn't enough VRAM. Often, a script will run fine on CPU but then fail like this on GPU simply because there is usually much more system memory available than GPU memory, especially since the system memory is virtualized (and can page out to disk if required) while the GPU memory isn't.<\/p>\n\n<p>For this script, there needs to be enough memory to store the training, validation, and testing data sets, the model parameters, and enough working space to store intermediate results of the computation. There are two options available:<\/p>\n\n<ol>\n<li><p>Reduce the amount of memory needed for one or more of these three components. Reducing the amount of training data is usually easiest; reducing the size of the model next. Unfortunately both of those two options will often impair the quality of the result that is being looked for. Reducing the amount of memory needed for intermediate results is usually beyond the developers control -- it is managed by Theano, but there is sometimes scope for altering the computation to achieve this goal once a good understanding of Theano's internals is achieved.<\/p><\/li>\n<li><p>If the model parameters and working memory can fit in GPU memory then the most common solution is to change the code so that the data is no longer stored in GPU memory (i.e. just store it as numpy arrays, not as Theano shared variables) then pass each batch of data in as <code>inputs<\/code> instead of <code>givens<\/code>. The <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\" rel=\"nofollow\">LSTM sample code<\/a> is an example of this approach.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":30729737,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-06-09T10:41:16.797","AcceptedAnswerId":null,"OwnerUserId":2782619.0,"Title":"shuffling the first ith column in a theano matrix","Body":"<p>I need to shuffle the first i column of each row for a theano tensor matrix.<\/p>\n\n<p>The only method I find is raw_random.shuffle_row_elements, but it shuffles every entire row (all columns).<\/p>\n\n<p>Can anyone give me a hint?<\/p>\n","answers":[{"AnswerId":"30793100","CreationDate":"2015-06-11T23:27:31.443","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<p>Use the following<\/p>\n\n<p><code>theano.tensor.basic.swapaxes(y, axis1, axis2)<\/code><\/p>\n\n<p>This is defined as:<\/p>\n\n<pre><code>def swapaxes(y, axis1, axis2):\n    \"swap axes of inputted tensor\"\n    y = as_tensor_variable(y)\n    ndim = y.ndim\n    li = range(0, ndim)\n    li[axis1], li[axis2] = li[axis2], li[axis1]\n    return y.dimshuffle(li)\n<\/code><\/pre>\n"}]}
{"QuestionId":30745647,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2015-06-10T01:37:17.180","AcceptedAnswerId":30858102.0,"OwnerUserId":4561745.0,"Title":"Python interface of Caffe: Error in \"import caffe\"","Body":"<p>I'm trying to run Caffe in it's Python interface. I've already run the command  in the caffe directory and it worked fine. Now, when I run the command  in the python environment in the terminal (Ubuntu 14.04), I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to search my computer for 'caffe.io' but couldn't find any file by that name. Any idea why this error is occurring and how to correct it?<\/p>\n","answers":[{"AnswerId":"30858102","CreationDate":"2015-06-16T03:03:14.510","ParentId":null,"OwnerUserId":"4973198","Title":null,"Body":"<p>You need to add Python Caffe to PYTHONPATH. In your case:<br>\n<strong>export PYTHONPATH=$PYTHONPATH:\/home\/pras\/caffe\/python<\/strong><\/p>\n"}]}
{"QuestionId":30745837,"AnswerCount":2,"Tags":"<c++><gcc><osx-yosemite><caffe>","CreationDate":"2015-06-10T01:59:12.720","AcceptedAnswerId":30751674.0,"OwnerUserId":1871528.0,"Title":"compiling caffe on Yosemite","Body":"<p>I'm trying to install caffe on Yosemite, and my C is not the strongest. Here is my error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm guessing that the problem is with the compiler, so I installed gcc from brew and tried running it using<\/p>\n\n<pre><\/pre>\n\n<p>which still did not help.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"30751674","CreationDate":"2015-06-10T08:58:27.990","ParentId":null,"OwnerUserId":"1871528","Title":null,"Body":"<p>its a boost version problem.<\/p>\n\n<p>If you using brew do the following:<\/p>\n\n<p><a href=\"http:\/\/itinerantbioinformaticist.blogspot.com\/2015\/05\/caffe-incompatible-with-boost-1580.html\" rel=\"noreferrer\">http:\/\/itinerantbioinformaticist.blogspot.com\/2015\/05\/caffe-incompatible-with-boost-1580.html<\/a><\/p>\n"},{"AnswerId":"31306966","CreationDate":"2015-07-09T02:11:57.520","ParentId":null,"OwnerUserId":"3508192","Title":null,"Body":"<p>I had the same problem. As someone else said, it turned out for me to be a problem with boost 1.58.0<\/p>\n\n<p>I fixed it by doing the following (assuming you have brew installed)<\/p>\n\n<pre><code> $ cd \/usr\/local\/Library\/Formula \n $ cd Library\/Formula\/\n $ cp boost.rb boost_backup.rb\n $ cp boost-python.rb boost-python_backup.rb\n $ wget https:\/\/raw.githubusercontent.com\/Homebrew\/homebrew\/6fd6a9b6b2f56139a44dd689d30b7168ac13effb\/Library\/Formula\/boost.rb\n $ mv boost.rb.1 boost.rb\n $ wget https:\/\/raw.githubusercontent.com\/Homebrew\/homebrew\/3141234b3473717e87f3958d4916fe0ada0baba9\/Library\/Formula\/boost-python.rb\n $ mv boost-python.rb.1 boost-python.rb\n $ brew uninstall --force boost\n $ brew install boost\n<\/code><\/pre>\n\n<p>After doing this I was able to make all with GPU support no problem.<\/p>\n\n<p><a href=\"http:\/\/playittodeath.ru\/how-to-install-caffe-on-mac-os-x-yosemite-10-10-4\/\" rel=\"nofollow\">source<\/a><\/p>\n"}]}
{"QuestionId":30761257,"AnswerCount":1,"Tags":"<python><csv><pickle><theano>","CreationDate":"2015-06-10T15:49:03.727","AcceptedAnswerId":null,"OwnerUserId":4995648.0,"Title":"theano csv to pkl file","Body":"<p>I am trying to make a pkl file to be loaded into theano from a csv starting point<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I run the resulting pkl file through Thenao, (as a DBN or SdA) it pretrains just fine, which makes me think the data is stored correctly. <\/p>\n\n<p>However when it comes to finetune I get the following error:<\/p>\n\n<pre>\n\n        epoch 1, minibatch 2775\/2775, validation error 0.000000 %\n\n    Traceback (most recent call last):\n      File \"SdA_custom.py\", line 489, in \n        test_SdA()\n      File \"SdA_custom.py\", line 463, in test_SdA\n        test_losses = test_model()\n      File \"SdA_custom.py\", line 321, in test_score\n        return [test_score_i(i) for i in xrange(n_test_batches)]\n\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/compile\/function_module.py\", line 606, in __call__\n        storage_map=self.fn.storage_map)\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/compile\/function_module.py\", line 595, in __call__\n        outputs = self.fn()\n    ValueError: Input dimension mis-match. (input[0].shape[0] = 10, input[1].shape[0] = 3)\n    Apply node that caused the error: Elemwise{neq,no_inplace}(argmax, Subtensor{int64:int64:}.0)\n    Inputs types: [TensorType(int64, vector), TensorType(int32, vector)]\n    Inputs shapes: [(10,), (3,)]\n    Inputs strides: [(8,), (4,)]\n    Inputs values: ['not shown', array([0, 0, 0], dtype=int32)]\n\n    Backtrace when the node is created:\n      File \"\/home\/dean\/Documents\/DeepLearningRepo\/DeepLearningTutorials-master\/code\/logistic_sgd.py\", line 164, in errors\n        return T.mean(T.neq(self.y_pred, y))\n\n    HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.\n\n<\/pre>\n\n<p>10 is the size of my batch, if I change to a batch size of 1 I get the following:<\/p>\n\n<pre>\n\n    ValueError: Input dimension mis-match. (input[0].shape[0] = 1, input[1].shape[0] = 0)\n\n<\/pre>\n\n<p>I think I am storing the labels wrong when I make a pkl, but I can't seem to spot what is happening or why changing the batch alters the error <\/p>\n\n<p>Hope you can help!<\/p>\n","answers":[{"AnswerId":"36026008","CreationDate":"2016-03-16T02:53:28.697","ParentId":null,"OwnerUserId":"6051033","Title":null,"Body":"<p>Saw this just now as was looking for similar error I was getting. Posting a reply so that it might help someone looking for similar error. For me the error resolved when I changed n_out to 2 from 1 in dbn_test() parameter list. n_out was the number of labels rather than number of output layers.<\/p>\n"}]}
{"QuestionId":30761433,"AnswerCount":3,"Tags":"<python><c++><deep-learning><caffe>","CreationDate":"2015-06-10T15:56:36.377","AcceptedAnswerId":30763062.0,"OwnerUserId":1348187.0,"Title":"[Caffe]: Check failed: ShapeEquals(proto) shape mismatch (reshape not set)","Body":"<p>I have this error and I have tried to take a look in Internet but I got nothing clear.<\/p>\n\n<p>I trained my net with Caffe successfully with around 82% of accuracy.<\/p>\n\n<p>Now I'm trying to try it with an image through this code:<\/p>\n\n<p><\/p>\n\n<p>yes, my images are 64x64, <\/p>\n\n<p>these are the last lines I'm getting:<\/p>\n\n<blockquote>\n  <p>I0610 15:33:44.868100 28657 net.cpp:194] conv3 does not need backward computation.\n  I0610 15:33:44.868110 28657 net.cpp:194] norm2 does not need backward computation.\n  I0610 15:33:44.868120 28657 net.cpp:194] pool2 does not need backward computation.\n  I0610 15:33:44.868130 28657 net.cpp:194] relu2 does not need backward computation.\n  I0610 15:33:44.868142 28657 net.cpp:194] conv2 does not need backward computation.\n  I0610 15:33:44.868152 28657 net.cpp:194] norm1 does not need backward computation.\n  I0610 15:33:44.868162 28657 net.cpp:194] pool1 does not need backward computation.\n  I0610 15:33:44.868173 28657 net.cpp:194] relu1 does not need backward computation.\n  I0610 15:33:44.868182 28657 net.cpp:194] conv1 does not need backward computation.\n  I0610 15:33:44.868192 28657 net.cpp:235] This network produces output fc8_pascal\n  I0610 15:33:44.868214 28657 net.cpp:482] Collecting Learning Rate and Weight Decay.\n  I0610 15:33:44.868238 28657 net.cpp:247] Network initialization done.\n  I0610 15:33:44.868249 28657 net.cpp:248] Memory required for data: 3136120\n  F0610 15:33:45.025965 28657 blob.cpp:458] Check failed: ShapeEquals(proto) shape mismatch (reshape not set)\n  <strong>* Check failure stack trace: *<\/strong>\n  Aborted (core dumped)<\/p>\n<\/blockquote>\n\n<p>I've tried to not setting the --mean_file and more things, but my shots are over. <\/p>\n\n<p>This is my imagenet_deploy.prototxt which I've modified in some parameters to debug, but didn't work anything.<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone could give me a clue?\nThank you very much.<\/p>\n\n<hr>\n\n<p>The same happens with C++ and the <strong>classification bin<\/strong>  they provide:<\/p>\n\n<blockquote>\n  <p>F0610 18:06:14.975601  7906 blob.cpp:455] Check failed: ShapeEquals(proto) shape mismatch (reshape not set)\n  <strong>* Check failure stack trace: *<\/strong>\n     @     0x7f0e3c50761c  google::LogMessage::Fail()\n     @     0x7f0e3c507568  google::LogMessage::SendToLog()\n     @     0x7f0e3c506f6a  google::LogMessage::Flush()\n     @     0x7f0e3c509f01  google::LogMessageFatal::~LogMessageFatal()\n     @     0x7f0e3c964a80  caffe::Blob&lt;>::FromProto()\n     @     0x7f0e3c89576e  caffe::Net&lt;>::CopyTrainedLayersFrom()\n     @     0x7f0e3c8a10d2  caffe::Net&lt;>::CopyTrainedLayersFrom()\n     @           0x406c32  Classifier::Classifier()\n     @           0x403d2b  main\n     @     0x7f0e3b124ec5  (unknown)\n     @           0x4041ce  (unknown)\n  Aborted (core dumped)<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"31251378","CreationDate":"2015-07-06T16:48:53.573","ParentId":null,"OwnerUserId":"1865831","Title":null,"Body":"<p>I just had the same error. In my case my output parameters of the final layer were incorrect: Switching datasets, I changed the number of classes in the train.prototxt and failed to do so in test.prototxt (or deploy.prototxt). Correcting this mistake solved the problem for me.<\/p>\n"},{"AnswerId":"30763062","CreationDate":"2015-06-10T17:16:09.263","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Let me confirm whether the basic steps are correct.<\/p>\n\n<pre><code>input_dim: 10\ninput_dim: 3\ninput_dim: 64\ninput_dim: 64\n<\/code><\/pre>\n\n<p>Have you tried changing the first parameter to 1 as you are only passing a single image.<\/p>\n\n<p>The above mentioned error occurs when the dimensions of the top or bottom blobs are not correct. And there is no where that could go wrong other than the input blobs.<\/p>\n\n<p><strong>Edit 2:<\/strong><\/p>\n\n<p><code>ShapeEquals(proto) shape mismatch (reshape not set)<\/code> error message occurs when 'reshape' parameter is set to false for the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/blob.cpp#L455\" rel=\"nofollow\"><em>fromproto<\/em> function call<\/a>.<\/p>\n\n<p>I did a quick search for the fromproto function call within the library as shown <a href=\"https:\/\/github.com\/BVLC\/caffe\/search?utf8=%E2%9C%93&amp;q=FromProto\" rel=\"nofollow\">here<\/a>. Other than 'CopyTrainedLayersFrom' function no other function actually set the above mentioned parameter as <code>false<\/code>.<\/p>\n\n<p>This is actually confusing. Two methods that I would suggest is:<\/p>\n\n<ol>\n<li>Check whether the caffe source code is updated from the repository.<\/li>\n<li>Try running the <em>test<\/em> portion of <em>caffe.bin<\/em> executable found in \/build\/tools\/.<\/li>\n<\/ol>\n"},{"AnswerId":"32349239","CreationDate":"2015-09-02T09:19:09.917","ParentId":null,"OwnerUserId":"1754513","Title":null,"Body":"<p>In my case, the size of the kernel in the second convolutional layer in my solver file differed from the one in the train file. Changing the size in the solver file solved the problem.<\/p>\n"}]}
{"QuestionId":30762476,"AnswerCount":2,"Tags":"<file-io><lua><torch>","CreationDate":"2015-06-10T16:46:07.373","AcceptedAnswerId":null,"OwnerUserId":3113501.0,"Title":"How to read a Bunch of files in a directory in lua","Body":"<p>I have a path (as a string) to a directory. In that directory, are a bunch of text files. I want to go to that directory open it, and go to each text file and read the data.<\/p>\n\n<p>I've tried<\/p>\n\n<pre><\/pre>\n\n<p>I get the error \"nil    Is a directory\"<\/p>\n\n<p>I've tried:<\/p>\n\n<pre><\/pre>\n\n<p>I get the error: \"Permission denied\"<\/p>\n\n<p>Is it just me, but it seems to be a lot harder than it should be to do basic file io in lua?<\/p>\n","answers":[{"AnswerId":"30762613","CreationDate":"2015-06-10T16:52:56.807","ParentId":null,"OwnerUserId":"258523","Title":null,"Body":"<p>A directory isn't a file. You can't just open it.<\/p>\n\n<p>And yes, lua itself has (intentionally) limited functionality.<\/p>\n\n<p>You can use <a href=\"https:\/\/keplerproject.github.io\/luafilesystem\/\" rel=\"nofollow\">luafilesystem<\/a> or <a href=\"https:\/\/github.com\/luaposix\/luaposix\" rel=\"nofollow\">luaposix<\/a> and similar modules to get more features in this area.<\/p>\n"},{"AnswerId":"30762936","CreationDate":"2015-06-10T17:09:05.360","ParentId":null,"OwnerUserId":"343123","Title":null,"Body":"<p>You can also use the following script to list the names of the files in a given directory (assuming Unix\/Posix):<\/p>\n\n<pre><code>dirname = '.'\nf = io.popen('ls ' .. dirname)\nfor name in f:lines() do print(name) end\n<\/code><\/pre>\n"}]}
{"QuestionId":30766364,"AnswerCount":1,"Tags":"<python><indexing><theano>","CreationDate":"2015-06-10T20:10:43.213","AcceptedAnswerId":30770321.0,"OwnerUserId":3112506.0,"Title":"Remove inf from Theano array","Body":"<p>Supposing I have a vector in Theano and some of the elements are , how do I remove them? Consider the following example:<\/p>\n\n<pre><\/pre>\n\n<p>According to the Theano documentation, this should remove the elements via indexing. However, this is not the case as  returns . <\/p>\n\n<p>How can I do this so that  returns  ?<\/p>\n","answers":[{"AnswerId":"30770321","CreationDate":"2015-06-11T01:58:43.443","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<p>I found an awkward workaround<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nvec = T.vector()\ncompare = T.isinf(vec)\nout = vec[(1-compare).nonzero()]\n\nv = [  1.,   1.,   1.,   1.,  np.inf,   3.,   4.,   5.,   6.,  np.inf]\nv = np.asarray(v)\n\nout.eval({var:v})\narray([ 1.,  1.,  1.,  1.,  3.,  4.,  5.,  6.])\n<\/code><\/pre>\n\n<p>For your example:<\/p>\n\n<pre><code>fin = vec[(1-T.isinf(vec)).nonzero()]\nf = theano.function([vec], fin)\n\nf([1,2,np.inf])\narray([ 1.,  2.])\n<\/code><\/pre>\n"}]}
{"QuestionId":30769048,"AnswerCount":3,"Tags":"<python><numpy><anaconda><caffe><lmdb>","CreationDate":"2015-06-10T23:28:40.987","AcceptedAnswerId":30788664.0,"OwnerUserId":4561745.0,"Title":"Error in creating LMDB database file in Python for Caffe","Body":"<p>I'm trying to create an LMDB data base file in Python to be used with Caffe according to <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"noreferrer\">this<\/a> tutorial. The commands  and  run perfectly fine. However, when I try to run  and , I'm getting the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>I'm running Python 2.7.9 through Anaconda 2.2.0 (64-bit) on Ubuntu 14.04. While installing the dependencies for Caffe according to <a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"noreferrer\">this<\/a> page, I've already installed the lmdb package through .<\/p>\n\n<p>Any ideas why this error might be occuring?<\/p>\n","answers":[{"AnswerId":"37486716","CreationDate":"2016-05-27T15:02:36.787","ParentId":null,"OwnerUserId":"4624901","Title":null,"Body":"<p>If you're using Anaconda, then this can solve your problem (it worked for me):<\/p>\n\n<pre><code>conda install -c https:\/\/conda.binstar.org\/dougal lmdb\n<\/code><\/pre>\n"},{"AnswerId":"52155660","CreationDate":"2018-09-03T20:03:43.910","ParentId":null,"OwnerUserId":"9954163","Title":null,"Body":"<p>For Anaconda users, installing <code>python-lmdb<\/code> package from <code>conda-forge<\/code> should fix the <code>lmdb<\/code> import error:<\/p>\n\n<pre><code>conda install -c conda-forge python-lmdb\n<\/code><\/pre>\n\n<p>This was tested on <code>conda 4.5.11<\/code> on an <code>lxc<\/code>-containerized system running <code>Ubuntu 18.04<\/code>. <\/p>\n\n<p>Note that there is a <code>conda<\/code> package named <code>lmdb<\/code> (without <code>python-<\/code>), installable via:<\/p>\n\n<pre><code>conda install -c conda-forge lmdb\n<\/code><\/pre>\n\n<p>that does not fix the import error.<\/p>\n"},{"AnswerId":"30788664","CreationDate":"2015-06-11T18:28:08.230","ParentId":null,"OwnerUserId":"4561745","Title":null,"Body":"<p>Well, the <code>apt-get install liblmdb-dev<\/code> might work with bash (in the terminal) but apparently it doesn't work with Anaconda Python. I figured Anaconda Python might require it's own module for lmdb and I followed <a href=\"https:\/\/lmdb.readthedocs.org\/en\/release\/#installation-unix\" rel=\"noreferrer\">this<\/a> link. The Python installation for lmdb module can be performed by running the command <code>pip install lmdb<\/code> in the terminal. And then <code>import lmdb<\/code> in Python works like a charm!<\/p>\n\n<p>The above installation commands may require sudo.<\/p>\n"}]}
{"QuestionId":30784588,"AnswerCount":1,"Tags":"<python><debugging><neural-network><theano>","CreationDate":"2015-06-11T15:05:56.003","AcceptedAnswerId":30792985.0,"OwnerUserId":2312926.0,"Title":"Print output of a Theano network","Body":"<p>I am sorry, very newbee question... I trained a neural network with Theano and now I want to see what it outputs for a certain input.<\/p>\n\n<p>So I can say:<\/p>\n\n<pre><\/pre>\n\n<p>where output_layer is my network.\nNow, the last layer happens to be a softmax, so if I say:<\/p>\n\n<pre><\/pre>\n\n<p>I get<\/p>\n\n<pre><\/pre>\n\n<p>I see why I get this I think (namely, because the output is a symbolic tensor variable), but I don't see how I can see the actual values.<\/p>\n\n<p>And just so you know, I did read <a href=\"https:\/\/stackoverflow.com\/questions\/17445280\/theano-print-value-of-tensorvariable\">this post<\/a> and also the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/printing.html\" rel=\"nofollow noreferrer\">documentation on printing<\/a> and <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#how-do-i-print-an-intermediate-value-in-a-function-method\" rel=\"nofollow noreferrer\">FAQ<\/a>, which I am also not fully grasping, I am afraid...<\/p>\n","answers":[{"AnswerId":"30792985","CreationDate":"2015-06-11T23:13:43.360","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<ol>\n<li>Use <code>.eval()<\/code> to evaluate the symbolic expression<\/li>\n<li>Use <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#using-test-values\" rel=\"nofollow\">Test Values<\/a><\/li>\n<\/ol>\n"}]}
{"QuestionId":30785836,"AnswerCount":1,"Tags":"<python><nan><theano><gradient-descent>","CreationDate":"2015-06-11T16:00:45.620","AcceptedAnswerId":null,"OwnerUserId":4999948.0,"Title":"Theano stochastic gradient descent NaN output","Body":"<p>I am using Theano stochastic gradient descent for solving a minimization problem. When running my code, the first iterations seem to work, but after a while and all of a sudden, the optimized parameter (eta) becomes NaNs (as well as the derivatives g_eta). It seems to be a Theano technical issue more than a bug in my code since I have checked it in several different ways. <\/p>\n\n<p>Anyone has an idea of which could be the reason? My code is the following:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"30793051","CreationDate":"2015-06-11T23:21:56.690","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<p>The fact that you are getting the same problem but more slowly with a slow learning rate suggests that you possible have an instability in your function which blows up near where you start SGD.<\/p>\n\n<ol>\n<li>Try different starting values<\/li>\n<li>Adjust your cost function to penalise the nasty area that is blowing up<\/li>\n<li>Try a different gradient descent method<\/li>\n<\/ol>\n"}]}
{"QuestionId":30792665,"AnswerCount":1,"Tags":"<lua><neural-network><backpropagation><training-data><torch>","CreationDate":"2015-06-11T22:43:23.257","AcceptedAnswerId":30818191.0,"OwnerUserId":1082019.0,"Title":"Torch Lua: Why is my gradient descent not optimizing the error?","Body":"<p>I've been trying to implement a siamese neural network in Torch\/Lua, as <a href=\"https:\/\/stackoverflow.com\/questions\/30581199\/torch-lua-is-this-a-good-implementation-of-a-siamese-neural-network\">I already explained here<\/a>.\nNow I have my first implementation, that I suppose to be good.<\/p>\n\n<p>Unfortunately, I'm facing a problem: during training back-propagation, the gradient descent does not update the error. That is, it always computes the same value (that is +1 or -1), without changing it.\nIn a correct implementation, the error should go from +1 to -1 or from -1 to +1. In my case, it's just stuck in the upper value and nothing changes.<\/p>\n\n<p>Why? I'm really looking for someone that could give me some hints.<\/p>\n\n<p>Here's my working code, that you might try to run:<\/p>\n\n<pre><\/pre>\n\n<p>The question is: <strong>why the <em>predictionValue<\/em> variable is always the same? Why doesn't it get updates?<\/strong><\/p>\n\n<p><strong>EDIT<\/strong>: I now realized that the problem was that I was using only 1 output layer dimension. I moved it to 6, but unfortunately I have a new problem. The gradient is not updating in the right direction.\nFor example, here's what happens by using my previous code with output_layer_number=6<\/p>\n\n<pre><\/pre>\n\n<p><strong>That is, the predictionValue never goes towards -1. Why?<\/strong><\/p>\n","answers":[{"AnswerId":"30818191","CreationDate":"2015-06-13T11:34:38.207","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>why the predictionValue variable is always the same? Why doesn't it get updates?<\/p>\n<\/blockquote>\n\n<p>First of all you should perform the backward propagation only if <code>predictionValue*targetValue &lt; 1<\/code> to make sure you back-propagate only if the pairs need to be pushed together (<code>targetValue = 1<\/code>) or pulled apart (<code>targetValue = -1<\/code>).<\/p>\n\n<p><em>See also this torch\/nn <a href=\"https:\/\/github.com\/torch\/nn\/blob\/3bbed8d\/doc\/table.md#cosinedistance\" rel=\"nofollow noreferrer\">official example<\/a> that illustrates this.<\/em><\/p>\n\n<p>That being said you have <strong>only 1 output unit<\/strong> (<code>output_layer_number = 1<\/code>). That means that each branch of your siamese network produces a single scalar, resp. <code>u<\/code> and <code>v<\/code>. This pair of scalars are then compared by the cosine distance:<\/p>\n\n<pre><code>C(u,v) = cosine(u, v) = (u \/ |u|) x (v \/ |v|)\n<\/code><\/pre>\n\n<p><em>Note: this criterion can only take two values here: 1 or -1 (see below in blue).<\/em><\/p>\n\n<p>When it is time to back-propagate you compute the derivatives of this criterion with respect to the inputs, i.e. <code>dC\/du<\/code> and <code>dC\/dv<\/code>. But these <strong>derivatives are null<\/strong> and undefined at 0 (see below in red):<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Guhom.png\" alt=\"enter image description here\"><\/p>\n\n<p>This is why the back-propagation does nothing here, i.e. it remains static (and you can verify this in practice by printing out the norms of these derivatives).<\/p>\n"}]}
{"QuestionId":30797301,"AnswerCount":2,"Tags":"<python-2.7><theano>","CreationDate":"2015-06-12T07:11:37.340","AcceptedAnswerId":null,"OwnerUserId":4680426.0,"Title":"Fail importing theano","Body":"<p>I could import theano ok yesterday. Today, when I want to import theano, it says \"No module named theano\". <\/p>\n\n<p>But question is numpy and scipy could be imported and they are in the same address as theano. I first think theano might error in itself, but after I uninstall and install it again. It still just can't work.<\/p>\n\n<p>I just upgrade my Mac to \"El Capitan preview\" today. Is there any relation between them.<\/p>\n","answers":[{"AnswerId":"31890735","CreationDate":"2015-08-08T07:23:02.170","ParentId":null,"OwnerUserId":"3497273","Title":null,"Body":"<p>In a python shell, type:<\/p>\n\n<pre><code>&gt;&gt;&gt; pip install theano\n<\/code><\/pre>\n\n<p>If you don't have pip, download <a href=\"https:\/\/bootstrap.pypa.io\/get-pip.py\" rel=\"nofollow\">getpip.py<\/a> and execute it (outside of the python shell).<\/p>\n\n<pre><code>python getpip.py\n<\/code><\/pre>\n\n<p>Then,<\/p>\n\n<pre><code>python\n&gt;&gt;&gt; pip install theano\n<\/code><\/pre>\n"},{"AnswerId":"31902313","CreationDate":"2015-08-09T08:24:24.407","ParentId":null,"OwnerUserId":"1196752","Title":null,"Body":"<p>Are you using something like <a href=\"https:\/\/store.continuum.io\/cshop\/anaconda\/\" rel=\"nofollow\">Anaconda<\/a> or non-stock python?<\/p>\n\n<p>You might want to check the <code>~\/.bash_profile<\/code> and add:<\/p>\n\n<pre><code>export PATH = \/path\/to\/anaconda\/bin:$PATH\n<\/code><\/pre>\n\n<p>If you have multiple installations of python, maybe you're executing the one that doesn't have theano installed.<\/p>\n"}]}
{"QuestionId":30808735,"AnswerCount":4,"Tags":"<python><python-2.7><caffe>","CreationDate":"2015-06-12T17:15:04.243","AcceptedAnswerId":30809008.0,"OwnerUserId":2938494.0,"Title":"Error when using classify in caffe","Body":"<p>I am using caffe in python to classify. I get code from <a href=\"http:\/\/www.openu.ac.il\/home\/hassner\/projects\/cnn_agegender\/\">here<\/a>. In here, I just use simple code such as<\/p>\n\n<pre><\/pre>\n\n<p>However, I got error such as<\/p>\n\n<pre><\/pre>\n\n<p>Could you help me to reslove it? Thanks<\/p>\n","answers":[{"AnswerId":"30809008","CreationDate":"2015-06-12T17:31:47.217","ParentId":null,"OwnerUserId":"3755376","Title":null,"Body":"<p>Let go to line 253-254 in caffe\/python\/caffe\/io.py \nReplace <\/p>\n\n<pre><code>if ms != self.inputs[in_][1:]:\n    raise ValueError('Mean shape incompatible with input shape.')\n<\/code><\/pre>\n\n<p>By <\/p>\n\n<pre><code>if ms != self.inputs[in_][1:]:\n    print(self.inputs[in_])\n    in_shape = self.inputs[in_][1:]\n    m_min, m_max = mean.min(), mean.max()\n    normal_mean = (mean - m_min) \/ (m_max - m_min)\n    mean = resize_image(normal_mean.transpose((1,2,0)),in_shape[1:]).transpose((2,0,1)) * (m_max - m_min) + m_min\n    #raise ValueError('Mean shape incompatible with input shape.')\n<\/code><\/pre>\n\n<p>Rebuild. Hope it help<\/p>\n"},{"AnswerId":"47394617","CreationDate":"2017-11-20T14:43:26.910","ParentId":null,"OwnerUserId":"1947377","Title":null,"Body":"<p>Edit deploy_gender.prototxt and set:\ninput_dim: 256\ninput_dim: 256<\/p>\n\n<p>Don't know why it was written wrong...<\/p>\n"},{"AnswerId":"39298492","CreationDate":"2016-09-02T18:11:30.457","ParentId":null,"OwnerUserId":"6788794","Title":null,"Body":"<p>I was having the same problem, based in the imagenet web demo I modified the script using this way to load the mean file in line 95<\/p>\n\n<p><code>mean = np.load(args.mean_file).mean(1).mean(1)<\/code><\/p>\n"},{"AnswerId":"45254585","CreationDate":"2017-07-22T12:31:23.083","ParentId":null,"OwnerUserId":"2496554","Title":null,"Body":"<p>I am pretty scared to rebuild the code as caffe installation did not come easy for me.\nBut to fix, the solution to resize mean require in_shape (user8264's response), which is set internally in caffe\/classifier.py<\/p>\n\n<p>Anyway, I debugged and found the value for in_shape = (3, 227, 227) for age_net.caffemodel<\/p>\n\n<p>So the model used for age and gender prediction would the following change:<\/p>\n\n<pre><code>age_net_pretrained='.\/age_net.caffemodel'\nage_net_model_file='.\/deploy_age.prototxt'\nage_net = caffe.Classifier(age_net_model_file, age_net_pretrained,\n                   mean=mean,\n                   channel_swap=(2,1,0),\n                   raw_scale=255,\n                   image_dims=(227, 227))\n<\/code><\/pre>\n\n<p>But mean needs to be modified first: <\/p>\n\n<pre><code>m_min, m_max = mean.min(), mean.max()\nnormal_mean = (mean - m_min) \/ (m_max - m_min)\nin_shape=(227, 227)\nmean = caffe.io.resize_image(normal_mean.transpose((1,2,0)),in_shape)\n                            .transpose((2,0,1)) * (m_max - m_min) + m_min\n<\/code><\/pre>\n\n<p>This will get rid of \"ValueError: Mean shape incompatible with input shape\". But I am not sure about the accuracy though. Apparently, for me skipping mean parameter gave better age prediction :)<\/p>\n"}]}
{"QuestionId":30812280,"AnswerCount":2,"Tags":"<python><deep-learning><caffe>","CreationDate":"2015-06-12T21:15:22.890","AcceptedAnswerId":31043900.0,"OwnerUserId":2824962.0,"Title":"Caffe net.predict() outputs random results (GoogleNet)","Body":"<p>I used pretrained GoogleNet from <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_googlenet\">https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_googlenet<\/a> and finetuned it with my own data (~ 100k images, 101 classes).\nAfter one day training I achieved 62% in top-1 and 85% in top-5 classification and try to use this network to predict several images. <\/p>\n\n<p>I just followed example from <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/classification.ipynb\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/classification.ipynb<\/a>,<\/p>\n\n<p>Here is my Python code:<\/p>\n\n<pre><\/pre>\n\n<p>In my deploy.prototxt I changed the last layer only to predict my 101 classes.<\/p>\n\n<pre><\/pre>\n\n<p>Here is the distribution of softmax output:<\/p>\n\n<pre><\/pre>\n\n<p>It seems just like random distribution with no sense.<\/p>\n\n<p>Thank you for any help or hint and best regards,\nAlex<\/p>\n","answers":[{"AnswerId":"30826566","CreationDate":"2015-06-14T06:28:56.253","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Please check the image transformation you are using - is it the same in training and test-time?<\/p>\n\n<p>AFAIK bvlc_googlenet subtract image mean with one value per channel, while your python <code>classifier<\/code> uses different mean <code>mean=np.load('ilsvrc_2012_mean.npy').mean(1).mean(1)<\/code>. This might cause the net to be unable to classify your inputs correctly.<\/p>\n"},{"AnswerId":"31043900","CreationDate":"2015-06-25T07:28:50.537","ParentId":null,"OwnerUserId":"2824962","Title":null,"Body":"<p>The solution is really simple: I just forgot to rename the last layer in deploy file:<\/p>\n\n<pre><code>layer {\n  name: \"loss3\/classifier\"\n  type: \"InnerProduct\"\n  bottom: \"pool5\/7x7_s1\"\n  top: \"loss3\/classifier\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n<\/code><\/pre>\n"}]}
{"QuestionId":30815035,"AnswerCount":1,"Tags":"<python><python-2.7><opencv><caffe>","CreationDate":"2015-06-13T04:20:42.443","AcceptedAnswerId":null,"OwnerUserId":2938494.0,"Title":"How to convert Mat from opencv to caffe format","Body":"<p>I am using opencv to crop face from my camera. And then I used caffe to predict that image belongs to male or female. I have a original code that load image from static image. However, I want to use image from camera for it. This is original code in caffe<\/p>\n\n<pre><\/pre>\n\n<p>Now, I will use opencv to capture frame and call predict method<\/p>\n\n<pre><\/pre>\n\n<p>After having resized_image, I will conver it to caffe type such as function<\/p>\n\n<pre><\/pre>\n\n<p>However, when I call that function. I don't know what is self. Could you help me to fix it?<\/p>\n\n<p>Thank you for help!<\/p>\n","answers":[{"AnswerId":"30858059","CreationDate":"2015-06-16T02:56:35.907","ParentId":null,"OwnerUserId":"4973198","Title":null,"Body":"<p>You can use <strong>CVMatToDatum<\/strong> function in caffe.io. More info here:  <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/util\/io.cpp\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/util\/io.cpp<\/a><\/p>\n\n<p>Edit: I think you can use <strong>array_to_datum<\/strong> from\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py<\/a>, \nthough it might be necessary to convert Mat to ndarray first<\/p>\n"}]}
{"QuestionId":30822009,"AnswerCount":2,"Tags":"<python><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-06-13T18:16:28.797","AcceptedAnswerId":32440865.0,"OwnerUserId":2938494.0,"Title":"How to speed up caffe classifer in python","Body":"<p>I am using python to use caffe classifier. I got image from my camera and peform predict image from training set. It work well but the problem is speed very slow. I thinks just 4 frames\/second. Could you suggest to me some way to improve computational time in my code?\nThe problem can be explained as following. I have to reload an network model  that its size about 80MB by following code<\/p>\n\n<pre><\/pre>\n\n<p>And for each input image (), I call the predict function<\/p>\n\n<pre><\/pre>\n\n<p>I think that due to size of network is very large. Then predict function takes long time to predict image. I think the slow time is from it.<br>\nThis is my full reference code. It changed by me. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45878676","CreationDate":"2017-08-25T09:43:52.343","ParentId":null,"OwnerUserId":"4265775","Title":null,"Body":"<p>You can also try <a href=\"https:\/\/arxiv.org\/abs\/1707.06168\" rel=\"nofollow noreferrer\">channel pruning<\/a> your network. It's an algorithm that effectively prune channels in each layer, which could speed up network 2-5x. The github address is : <a href=\"https:\/\/github.com\/yihui-he\/channel-pruning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yihui-he\/channel-pruning<\/a><\/p>\n"},{"AnswerId":"32440865","CreationDate":"2015-09-07T14:19:58.660","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Try calling <code>age_net.predict([caffe_input])<\/code> with <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L47\" rel=\"nofollow\"><code>oversmaple=False<\/code><\/a>:<\/p>\n\n<pre><code>prediction = age_net.predict([caffe_input], oversample=False)\n<\/code><\/pre>\n\n<p>The default behavior of <code>predict<\/code> is to create 10, slightly different, crops of the input image and feed them to the network to classify, by disabling this option you should get a x10 speedup.<\/p>\n"}]}
{"QuestionId":30828095,"AnswerCount":1,"Tags":"<cuda><theano>","CreationDate":"2015-06-14T10:03:59.320","AcceptedAnswerId":30858008.0,"OwnerUserId":5008021.0,"Title":"MAC getting Theano to use the GPU","Body":"<p>I am having quite a bit of trouble setting up Theano to work with my graphics card on a mac, I really hope you can give me some help.<\/p>\n\n<p>WARNING (theano.sandbox.cuda): CUDA is installed, but device gpu0 is not available  (error: Unable to get the number of gpus available: CUDA driver version is insufficient for CUDA runtime version)<\/p>\n","answers":[{"AnswerId":"30858008","CreationDate":"2015-06-16T02:48:22.827","ParentId":null,"OwnerUserId":"4013571","Title":null,"Body":"<p>You need an NVIDIA graphics card to use CUDA as CUDA is an NVIDIA product.<\/p>\n\n<p>To check if your mac has an NVIDIA graphics card:<\/p>\n\n<ul>\n<li>Click on the \uf8ff in the top left of the screen<\/li>\n<li>About This Mac<\/li>\n<li>Check the graphics card under Displays...<\/li>\n<li>If like me you don't see NVIDIA or GeForce anywhere, you can't use CUDA. <\/li>\n<li>See <a href=\"https:\/\/stackoverflow.com\/a\/30701002\/4013571\">this post<\/a> about alternatives for non NVIDIA for deep learning...<\/li>\n<\/ul>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/6WIeU.png\" alt=\"Non Nvidia graphics card\"><\/p>\n"}]}
{"QuestionId":30830987,"AnswerCount":1,"Tags":"<python><python-2.7><face-recognition><caffe>","CreationDate":"2015-06-14T15:12:37.140","AcceptedAnswerId":30966717.0,"OwnerUserId":2938494.0,"Title":"How to create an caffemodel file from training image and its labeled?","Body":"<p>I am working in age classification based on the opensource at <a href=\"http:\/\/www.openu.ac.il\/home\/hassner\/projects\/cnn_agegender\/\" rel=\"nofollow\">here<\/a>\nThe python code has <\/p>\n\n<pre><\/pre>\n\n<p>In which  file is shown as below. I remain one file that is . As the source code, he provided it before. However, I would like to create it again based on my face database. Could you have any tutorial or some way to create it? I assume that I have a folder image that include 100 images and divided belongs to each age groups (1 to 1) such as<\/p>\n\n<pre><\/pre>\n\n<p>This is prototxt file. Thanks in advance<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30966717","CreationDate":"2015-06-21T16:23:48.903","ParentId":null,"OwnerUserId":"1899150","Title":null,"Body":"<p>To get a caffemodel you need to train the network. That prototxt file is only to deploy the model and cannot be used to train it. <\/p>\n\n<p>You need to add a data layer that points to your database. To use a list of files as you mention, the source of the layer should be HDF5. You will probably want to add a transform_param with the mean value. The image files can be replaced by a LMDB or LevelDB database for efficiency purposes.<\/p>\n\n<p>At the end of the network you will have to substitute the 'prob' layer with a 'loss' layer. Something like this:<\/p>\n\n<p>layers {\n  name: \"loss\"\n  type: SoftmaxWithLoss\n  bottom: \"fc8\"\n  top: \"loss\"\n}<\/p>\n\n<p>The layer catalogue can be found here:<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n\n<p>Or, as your network is a well known one... just look at this tutorial :P.<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html<\/a><\/p>\n\n<p>The correct prototxt file for training is included in caffe ('train_val.prototxt').<\/p>\n"}]}
{"QuestionId":30836029,"AnswerCount":1,"Tags":"<cmake><caffe>","CreationDate":"2015-06-15T00:44:51.917","AcceptedAnswerId":null,"OwnerUserId":1282043.0,"Title":"installing caffe cmake error","Body":"<p>I've followed all the steps to install the caffe dependencies. I've bumped into the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not really sure what's going on. It seems as if cmake hasn't been installed properly? What's curious is that I have two cmake folders: One in my caffe folder and the other one inside an extra folder I called \"dependencies\" in which I am compiling all other dependencies for caffe:<\/p>\n\n<p>doing form my \/caffe\/ folder:<\/p>\n\n<pre><\/pre>\n\n<p>doing ls from my \/caffe\/dependencies\/ folder:<\/p>\n\n<pre><\/pre>\n\n<p>Was I supposed to install all the dependencies inside my \/caffe\/ folder?<\/p>\n","answers":[{"AnswerId":"35313238","CreationDate":"2016-02-10T10:55:29.903","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>I suggest you should use the <code>cmake-gui<\/code> for <code>making<\/code> Caffe. There you can set the appropriate paths easily. No need to install all the dependencies in the Caffe root directory. You may install them in your system home directory.<\/p>\n"}]}
{"QuestionId":30850543,"AnswerCount":1,"Tags":"<python><pip><ubuntu-14.04><caffe><lmdb>","CreationDate":"2015-06-15T16:43:33.190","AcceptedAnswerId":30853705.0,"OwnerUserId":4561745.0,"Title":"Error while installing deepdish","Body":"<p>I'm trying to create an LMDB database file to be used with Caffe according to <a href=\"http:\/\/deepdish.io\/\" rel=\"nofollow\">this<\/a> tutorial on an Ubuntu 14.04 machine using Anaconda Python 2.7.9. However, when I do , I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas why this error might be occurring and how to go about correcting it? Any help is much appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"30853705","CreationDate":"2015-06-15T19:44:16.920","ParentId":null,"OwnerUserId":"4561745","Title":null,"Body":"<p>Can't seem to figure out why the error occurred, but after struggling for a while, I did the following and it seemed to work:<\/p>\n\n<ol>\n<li>Download the zip file from Github. <\/li>\n<li>Unzip the file after navigating to the directory where you've downloaded the file using the command <code>unzip deepdish-master.zip<\/code>.<\/li>\n<li>Navigate to the unzipped folder: <code>cd deepdish-master<\/code>.<\/li>\n<li>Make the setup.py file executable: <code>chmod +x setup.py<\/code>.<\/li>\n<li>Run the file with the 'install' option: <code>.\/setup.py install<\/code>.<\/li>\n<li>If you want, you may delete the zip folder and the unzipped folder.<\/li>\n<\/ol>\n\n<p>Now if you go into the python environment in the terminal and do <code>import deepdish as dd<\/code>, it works like a charm!<\/p>\n"}]}
{"QuestionId":30860533,"AnswerCount":2,"Tags":"<python><closures><theano>","CreationDate":"2015-06-16T06:46:10.897","AcceptedAnswerId":null,"OwnerUserId":4491330.0,"Title":"What's the advantage of using this piece of code in python ? (set i = [0] first, and later i[0] += size)","Body":"<p>I'm reading a python script with the module <a href=\"http:\/\/deeplearning.net\/software\/theano\/\" rel=\"nofollow\">Theano<\/a> used. I'm confused by the follow piece of code.  <\/p>\n\n<pre><\/pre>\n\n<p>The function read_param is used in functions beneath. <\/p>\n\n<pre><\/pre>\n\n<p>I was told that it uses the concept of \"closure\" and is advantageous in writing and reading. But I'm not sure why and how it works. <\/p>\n","answers":[{"AnswerId":"30860625","CreationDate":"2015-06-16T06:50:32.440","ParentId":null,"OwnerUserId":"248296","Title":null,"Body":"<p>My only guess is that the author didn't want to define <code>i<\/code> as a global variable:<\/p>\n\n<pre><code>i = 0\ndef read_param(size):\n    global i\n    ret = lparam[i:i+size]\n    i += size\n    return ret\n<\/code><\/pre>\n\n<p>I don't know why (maybe the author doesn't know Python well).<\/p>\n\n<p>There is no advantage in using <code>i[0]<\/code>, only disadvantages -- bad readibility and more CPU used.<\/p>\n\n<p>The trick here is that <code>i<\/code> is a mutable object -- <code>list<\/code> and you mutate it instead of reassigning (which would require to define it as <code>global<\/code>). I would use a more pythonic approach like in @Marius's answer.<\/p>\n\n<p>I think the author considers this a cool trick (though this one is cooler: <code>def read_param(size, i=[0]):<\/code>) and will defend his approach, but <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0020\/\" rel=\"nofollow\">The Zen of Python<\/a> should be followed:<\/p>\n\n<ul>\n<li>Explicit is better than implicit.<\/li>\n<li>Simple is better than complex.<\/li>\n<li>Readability counts.<\/li>\n<li>Special cases aren't special enough to break the rules.<\/li>\n<li>In the face of ambiguity, refuse the temptation to guess.<\/li>\n<li>There should be one-- and preferably only one --obvious way to do it.<\/li>\n<\/ul>\n"},{"AnswerId":"30860735","CreationDate":"2015-06-16T06:55:59.257","ParentId":null,"OwnerUserId":"1222578","Title":null,"Body":"<p>I'm struggling to see the advantages of the current code. <code>i<\/code> is created as a list so it can be mutated within the function, but as warvariuc shows, you could just use <code>global<\/code> to achieve the same thing.<\/p>\n\n<p>Here is how I would do the same thing \"pythonically\":<\/p>\n\n<pre><code>class ParamReader(object):\n    def __init__(self, params):\n        self.params = params\n        self.i = 0\n\n    def read(self, size):\n        ret = self.params[self.i:(self.i + size)]\n        self.i += size\n        return ret\n\n# Dummy values for lparam as I don't have theano\nlparam = list(range(100))\nreader = ParamReader(lparam)\n\nreader.read(5)\nOut[8]: [0, 1, 2, 3, 4]\n\nreader.read(6)\nOut[9]: [5, 6, 7, 8, 9, 10]\n<\/code><\/pre>\n\n<p>The read function has to maintain state, so it seems like an obvious choice to just use a simple class.<\/p>\n"}]}
{"QuestionId":30867336,"AnswerCount":1,"Tags":"<cuda><deep-learning><caffe>","CreationDate":"2015-06-16T12:16:17.737","AcceptedAnswerId":null,"OwnerUserId":2281121.0,"Title":"Caffe Framework Runtest Core dumped error","Body":"<p>I have been installing Caffe Framework with the following GPU:\nGeforce 9500 GT\nCUDA 6.5 (not work with 7.0)<\/p>\n\n<p>when I run:  the following errors appeared and I don't know what are the reasons:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30883960","CreationDate":"2015-06-17T06:38:33.953","ParentId":null,"OwnerUserId":"2281121","Title":null,"Body":"<p>It turns out that (as @Jez suggested) my GPU does not support double precision which used by Caffe math functions. That's the reason for crashes. I have searched for workaround on this issue but haven't found one. Maybe the only solution is to use more modern GPU<\/p>\n"}]}
{"QuestionId":30879506,"AnswerCount":1,"Tags":"<python><variables><shape><theano>","CreationDate":"2015-06-16T22:43:10.687","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Problems on Python Code using Theano","Body":"<p>I am learning the Python Theano Library. I just encountered a block of code using Theano shown below:<\/p>\n\n<pre><\/pre>\n\n<p>My problem is I cannot understand this part of code:<\/p>\n\n<pre><\/pre>\n\n<p>Also, I am wondering if there is any way in Python to check the values as well as dimensions of Theano tensor variables. I really appreciate if anyone can help me solve the problem.<\/p>\n","answers":[{"AnswerId":"30879859","CreationDate":"2015-06-16T23:21:17.500","ParentId":null,"OwnerUserId":"2594119","Title":null,"Body":"<p>According to the docs <code>shape<\/code> returns an lvector representing the shape of x. You can read more about that <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#shaping-and-shuffling\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>In the code block you are referencing <code>n1<\/code> will return the object at index 0 and <code>n2<\/code> will return the object at index 1.<\/p>\n\n<p>You can read a bit more about lists in python <a href=\"http:\/\/www.tutorialspoint.com\/python\/python_lists.htm\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>If you are running this script from the command line, you can use a <code>print<\/code> statement to see what is contained is those variables by adding a line like this:<\/p>\n\n<p><code>n1 = x.shape[0]\nprint n1<\/code><\/p>\n"}]}
{"QuestionId":30885342,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2015-06-17T07:46:23.767","AcceptedAnswerId":30898018.0,"OwnerUserId":1740705.0,"Title":"Optional parameter to theano function","Body":"<p>I have a function  in theano which takes two parameters, one of them optional. When I call the function with the optional parameter being  the check inside  fails. This script reproduces the error:<\/p>\n\n<pre><\/pre>\n\n<p>Fails with the error message<\/p>\n\n<pre><\/pre>\n\n<p>It makes sense that  has no input shapes nor input strides. But I wonder why the  check inside  does not seem to work.<\/p>\n\n<p>How can I make the check inside  work such that the optional parameter  is handled correctly?<\/p>\n","answers":[{"AnswerId":"30898018","CreationDate":"2015-06-17T17:07:09.387","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano does not support optional parameters. By specifying the function's input parameters as <code>ins=[y,c]<\/code> you are telling Theano that the function has two 1-dimensional (vector) parameters. As far as Theano is concerned, both are mandatory. When you try to pass <code>None<\/code> in for <code>c<\/code> Theano checks that the types of the values you pass in match the types declared when you compiled the function (i.e. two vectors) but clearly <code>None<\/code> is not a vector so this exception is raised.<\/p>\n\n<p>A solution is to compile two Theano functions, one that accepts just one parameter and the other that accepts both. You could even use your existing Python function <code>f<\/code> for both.<\/p>\n"},{"AnswerId":"30902424","CreationDate":"2015-06-17T21:07:37.027","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I'm going to try a more complete response.<\/p>\n\n<p>1) the condition \"c is not None\" is run only once when you build the graph. As c is a symbolic variable, the else path will always be executed. If you want to execution condition at run time see this documentation page: <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html<\/a><\/p>\n\n<p>2) Theano have a special Type for None. But I do not recommand that you use it. It is not useful most of the time and it is not documented. So don't use it until you get more familiar with Theano.<\/p>\n\n<p>3) The other answer that tell to use 2 functions will work.<\/p>\n\n<p>4) In that simple case, you could pass a vector of the right size with only one instead of None. That would also work, but be slower.<\/p>\n"}]}
{"QuestionId":30898411,"AnswerCount":1,"Tags":"<nginx><lua><luajit><torch><openresty>","CreationDate":"2015-06-17T17:28:56.177","AcceptedAnswerId":null,"OwnerUserId":5019970.0,"Title":"Openresty torch module loading issue","Body":"<p>I'm trying to use openresty with torch for a Rest api for a neural network. \nFirst query works,  any query after that fails. <\/p>\n\n<h1>Nginx Config<\/h1>\n\n<pre><\/pre>\n\n<h1>testFile.lua<\/h1>\n\n<pre><\/pre>\n\n<h2>The error:<\/h2>\n\n<pre><\/pre>\n\n<p>Would appreciate any help<\/p>\n","answers":[{"AnswerId":"30951592","CreationDate":"2015-06-20T07:46:02.620","ParentId":null,"OwnerUserId":"282536","Title":null,"Body":"<p>You didn't <code>require<\/code> the torch library.\nAdd <code>local torch = require \"torch\"<\/code> at the top.<\/p>\n"}]}
{"QuestionId":30900477,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-06-17T19:17:35.350","AcceptedAnswerId":31148806.0,"OwnerUserId":1357690.0,"Title":"Scan function from Theano replicates non_sequences shared variables","Body":"<p>I'm trying to implement a custom convolutional layer for a CNN network in Theano, and in order to do so I'm using the scan function. The idea is to apply the new convolution mask to each pixel.<\/p>\n\n<p>The  function compiles correctly, but for some reason I get an out-of-memory error. The debug (see below) indicates that the  variables are replicated for each instance of the loop (for each pixel), which of course kills my GPU memory:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the output from the debugger:<\/p>\n\n<pre><\/pre>\n\n<p>The input as you can see is shown as a 1025x2000x3x32x32 tensor, while the original tensor is of size 2000x3x32x32, and the 1025 is the number of iterations of scan + 1.<\/p>\n\n<p>Why are the  variables replicated for each iteration instead of simply being reused, and how can I fix it?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Both  and  are shared variables. Self.input is passed to the class when initialized, while self.b is defined inside the class as follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31148806","CreationDate":"2015-06-30T21:26:35.567","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>It is possible that, when the scan is first created or at some point during the optimization process, a symbolic <code>Alloc<\/code> with that shape is created.\nHowever, it should be optimized at a later stage of the optimization process.<\/p>\n\n<p>We are aware that there was a but related to that recently, which should now be fixed in the development (\"bleeding-edge\") version of Theano. In fact, I just tried your snippet (slightly edited) with a recent development version, and had no memory error. Moreover, there was no 5D tensor anywhere in the computation graph, which would suggest the bug has indeed been fixed.<\/p>\n\n<p>Finally, please be aware that operations such as convolutions, which are not really recurrent, will probably be much slower when expressed with <code>scan<\/code> rather than with one of the existing convolution operations. In particular, <code>scan<\/code> will not be able to parallelize efficiently when the iterations of the loop do not depend on each other.<\/p>\n"}]}
{"QuestionId":30902056,"AnswerCount":0,"Tags":"<python><lua><protocol-buffers><caffe><torch>","CreationDate":"2015-06-17T20:45:28.757","AcceptedAnswerId":null,"OwnerUserId":4468092.0,"Title":"Generate caffemodel file","Body":"<p>I am using both Torch and Caffe for deep learning. I wonder if there is a way to output nn model into caffemodel file. It may involve protobuf, which I am not familiar with.<\/p>\n\n<p>It is kinda easy for the other way, where there are <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\" rel=\"nofollow\">libraries<\/a> to read caffemodel file into torch.<\/p>\n","answers":[]}
{"QuestionId":30919682,"AnswerCount":1,"Tags":"<matlab><deep-learning><leveldb><caffe>","CreationDate":"2015-06-18T15:29:50.903","AcceptedAnswerId":30930801.0,"OwnerUserId":2163392.0,"Title":"Convert a bunch of images to lmdb format in matlab","Body":"<p>I am using caffe deep network library and I have to use googlenet in my data (images).<\/p>\n\n<p>The problem is the fact that I want to do a lot of pre-processing operations, and I don't want to save every image processing operation as image files and then execute create_imagenet.sh included in the library. This library needs text files to indicate where are the train and validation image files. <\/p>\n\n<p>I don't want to save a lot of images and then converting them in LMDB.<\/p>\n\n<p>All I want is to do the image processing operation in a series of images and save them in a row in a lmdb file to be read by caffe later.<\/p>\n\n<p>Is that possible?<\/p>\n","answers":[{"AnswerId":"30930801","CreationDate":"2015-06-19T05:50:59.980","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>This github repository might be the one you are actually looking for:\n<a href=\"https:\/\/github.com\/kyamagu\/matlab-lmdb\" rel=\"nofollow\">https:\/\/github.com\/kyamagu\/matlab-lmdb<\/a><\/p>\n"}]}
{"QuestionId":30921359,"AnswerCount":1,"Tags":"<python><draw><ubuntu-14.04><deep-learning><caffe>","CreationDate":"2015-06-18T16:54:58.637","AcceptedAnswerId":30943006.0,"OwnerUserId":4561745.0,"Title":"Error while drawing net in Caffe","Body":"<p>I'm trying to draw my net in Caffe. The net is defined in  and the desired image path is . When I run the command , I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm running Caffe on Ubuntu 14.04 in CPU mode using the Anaconda Python interface. Any ideas why this error might be occuring and how to go about correcting it?<\/p>\n","answers":[{"AnswerId":"30943006","CreationDate":"2015-06-19T16:30:51.023","ParentId":null,"OwnerUserId":"4561745","Title":null,"Body":"<p>The error log mentions that GraphViz's executables are not found. So, I did the following:<\/p>\n\n<ol>\n<li>Installed GraphViz on Ubuntu: <code>sudo apt-get install GraphViz<\/code>.<\/li>\n<li>Installed GraphViz for Python: <code>pip install GraphViz<\/code>.<\/li>\n<\/ol>\n\n<p>I'm not sure if step 2 is required or not, but step 1 is definitely required. After doing that, the command to draw the net in Caffe works like a charm!<\/p>\n"}]}
{"QuestionId":30927369,"AnswerCount":1,"Tags":"<neural-network><deep-learning><torch>","CreationDate":"2015-06-18T23:13:27.900","AcceptedAnswerId":null,"OwnerUserId":2817198.0,"Title":"What is gradInput and gradOutput in Torch7's 'nn' package?","Body":"<p>Hi I am a starter of using Torch's 'nn' package. In the past two weeks, I am extremely confused about the meaning of gradInput and gradOutput in the Torch's 'nn' library. I believe the 'grad' here means gradient, but what exactly are those two variables refers to?<\/p>\n\n<p>Thanks for anyone's help! <\/p>\n","answers":[{"AnswerId":"30927733","CreationDate":"2015-06-18T23:50:04.897","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>gradOutput: gradient w.r.t. the output of the module. This is passed in either from the loss function, or from the module next to the current module. it is used to compute the gradient w.r.t. the input (gradInput) and gradient w.r.t. the parameters of the module (gradWeight \/ gradBias)<\/p>\n"}]}
{"QuestionId":30942172,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-06-19T15:42:06.863","AcceptedAnswerId":30946985.0,"OwnerUserId":4561745.0,"Title":"How does Caffe determine the number of neurons in each layer?","Body":"<p>Recently, I've been trying to use Caffe for some of the deep learning work that I'm doing. Although writing the model in Caffe is very easy, I've not been able to know the answer to this question. <strong>How does Caffe determine the number of neurons in a hidden layer?<\/strong> I do know that determination of number of neurons in a layer and the number of hidden layers itself are problems that cannot be determined analytically and the use of 'thumb rules' is imperative in this regard. But is there a way to define or know the number of neurons in each layer in Caffe? And by default, how does Caffe inherently determine this?<\/p>\n\n<p>Any help is much appreciated!<\/p>\n","answers":[{"AnswerId":"30946985","CreationDate":"2015-06-19T20:44:12.920","ParentId":null,"OwnerUserId":"5029761","Title":null,"Body":"<p>Caffe doesn't determine the number of neurons--the user does.<br>\nThis is pulled straight from Caffe's website, here: <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n\n<p>For example, this is a convolution layer of 96 nodes (or neurons):<\/p>\n\n<pre><code>layer {\n  name: \"conv1\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"conv1\"\n  # learning rate and decay multipliers for the filters\n  param { lr_mult: 1 decay_mult: 1 }\n  # learning rate and decay multipliers for the biases\n  param { lr_mult: 2 decay_mult: 0 }\n  convolution_param {\n    num_output: 96     # learn 96 filters\n    kernel_size: 11    # each filter is 11x11\n    stride: 4          # step 4 pixels between each filter application\n    weight_filler {\n      type: \"gaussian\" # initialize the filters from a Gaussian\n      std: 0.01        # distribution with stdev 0.01 (default mean: 0)\n    }\n    bias_filler {\n      type: \"constant\" # initialize the biases to zero (0)\n      value: 0\n    }\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":30942693,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-06-19T16:12:21.170","AcceptedAnswerId":30942976.0,"OwnerUserId":2191652.0,"Title":"Caffe for feed forward networks","Body":"<p>Has anyone experience with using caffe as feed forward network instead of convolutional neural network?<\/p>\n\n<p>My input data is 1 dimensional. Everything is ok, but when I want to use layers like max pooling, caffe assumes a square shaped kernel size. I would rather need a 1 dim max pooling kernel.<\/p>\n","answers":[{"AnswerId":"30942976","CreationDate":"2015-06-19T16:29:29.733","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>Found it:<\/p>\n\n<pre><code>layer {\n  name: \"pool1\"\n  type: \"Pooling\"\n  bottom: \"ip1\"\n  top: \"pool1\"\n  pooling_param {\n    pool: MAX\n    kernel_h: 3\n    kernel_w: 1 \n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":30953368,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-06-20T11:18:37.877","AcceptedAnswerId":30955861.0,"OwnerUserId":4961048.0,"Title":"Convolution Neural Network in torch. Error when training the network","Body":"<p>I am trying to base my Convolution neural network upon the following tutorial:<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised<\/a><\/p>\n\n<p>The issue is that my images are of different dimensions than those used in the tutorial. (3x200x200). Also I have only two classes.<\/p>\n\n<p>The following are the changes that I made :<\/p>\n\n<p>Changing the dataset to be loaded in 1_data.lua.<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>in 3_loss.lua and 4_train.lua.<\/p>\n\n<p>My model is the same as that being trained in the tutorial. For Convenience I'll put the code below :<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error when I run the doall.lua file :<\/p>\n\n<pre><\/pre>\n\n<p>I have been stuck at this for over a day now. Please help.<\/p>\n","answers":[{"AnswerId":"30955861","CreationDate":"2015-06-20T15:43:41.190","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>The problem is the convolutional neural network from this tutorial has been made to work with a <strong>fixed size input resolution<\/strong> of 32x32 pixels.<\/p>\n\n<p>Right after the 2 convolutional \/ pooling layers you obtain 64 feature maps with a 5x5 resolution. This gives an input of 64x5x5 = 1,600 elements for the following fully-connected layers.<\/p>\n\n<p>As you can see in the tutorial there is a dedicated <em>reshape<\/em> operation that transforms the 3D input tensor into a 1D tensor with 1,600 elements:<\/p>\n\n<pre><code>-- nstates[2]*filtsize*filtsize = 64x5x5 = 1,600\nmodel:add(nn.Reshape(nstates[2]*filtsize*filtsize))\n<\/code><\/pre>\n\n<p>When you work with a higher resolution input you produce higher resolution output feature maps, here a 200x200 pixels input gives 64 output feature maps of size 47x47. This is why you obtain this <em>wrong size<\/em> error.<\/p>\n\n<p>So you need to adapt the reshape and following linear layers accordingly:<\/p>\n\n<pre><code>model:add(nn.Reshape(nstates[2]*47*47))\nmodel:add(nn.Linear(nstates[2]*47*47, nstates[3]))\n<\/code><\/pre>\n"}]}
{"QuestionId":30959402,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-06-20T22:17:14.300","AcceptedAnswerId":null,"OwnerUserId":1213014.0,"Title":"\"ValueError: Shape mismatch\" error in Lasagne for Neural Network","Body":"<p>I'm trying to create a neural network based on theano\/lasagne that will (essentially) attempt to do a multi-variable regression. <\/p>\n\n<p>The meat of the code is: <\/p>\n\n<pre><\/pre>\n\n<p>Here, train_value is just 1 column of (numerical) data that I want to train my NN to predict, and the following 57 columns (train_data) are all the parameters\/values (all numbers) which should be weighted appropriately to predict the value in the first column.  <\/p>\n\n<p>However, when I run this script, I get the following error: <\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure where it is getting this shape--none of my data has 83 columns or rows. (note: I've tried to adapt this script, which originally was written to look at pictures of faces and guess where different parts were (eyes, nose, mouth, etc)). <\/p>\n\n<p>I have written a much simpler version of this (sans-dropout method) in pybrain, but am trying to migrate to sklearn\/lasagne\/theano as it opens more doors. <\/p>\n","answers":[{"AnswerId":"31931119","CreationDate":"2015-08-11T00:08:26.247","ParentId":null,"OwnerUserId":"1646409","Title":null,"Body":"<p>Since you want to do regression, make sure to set the output type correctly:<\/p>\n\n<pre><code>output_nonlinearity = linear\n<\/code><\/pre>\n\n<p>Are you sure you also have 10 output units? I experienced some weird behavior in Lasagne. I think the API has changed over time and contains some bugs. I succeeded using the latest <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow\">API demo<\/a> and adapting it to my needs.<\/p>\n"}]}
{"QuestionId":30962899,"AnswerCount":2,"Tags":"<image-processing><machine-learning><computer-vision><neural-network><torch>","CreationDate":"2015-06-21T09:06:32.177","AcceptedAnswerId":30982411.0,"OwnerUserId":4961048.0,"Title":"Using Convolution Neural network as Binary classifiers","Body":"<p>Given any image I want my classifier to tell if it is Sunflower or not. How can I go about creating the second class ? Keeping the set of all possible images - {Sunflower} in the second class is an overkill. Is there any research in this direction ? Currently my classifier uses a neural network in the final layer. I have based it upon the following tutorial :<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised<\/a><\/p>\n\n<p>I am taking images with 254x254 as the input.\nWould SVM help in the final layer ? Also I am open to using any other classifier\/features that might help me in this.<\/p>\n","answers":[{"AnswerId":"30982411","CreationDate":"2015-06-22T14:23:35.527","ParentId":null,"OwnerUserId":"3633250","Title":null,"Body":"<p>The standard approach in ML is that:<\/p>\n\n<p>1) Build model\n2) Try to train on some data with positive\\negative examples (start with 50\\50 of pos\\neg in training set)\n3) Validate it on test set (again, try 50\\50 of pos\\neg examples in test set)\nIf results not fine:\na) Try different model?\nb) Get more data<\/p>\n\n<p>For case #b, when deciding which additional data you need the rule of thumb which works for me nicely would be:\n1) If classifier gives lots of false positive (tells that this is a sunflower when it is actually not a sunflower at all) - get more negative examples\n2) If classifier gives lots of false negative (tells that this is not a sunflower when it is actually a sunflower) - get more positive examples<\/p>\n\n<p>Generally, start with <em>some<\/em> reasonable amount of data, check the results, if results on train set or test set are bad - get more data. Stop getting more data when you get the optimal results.<\/p>\n\n<p>And another thing you need to consider, is if your results with current data and current classifier are not good you need to understand if the problem is high bias (well, bad results on train set and test set) or if it is a high variance problem (nice results on train set but bad results on test set). If you have high bias problem - more data or more powerful classifier will definitely help. If you have a high variance problem - more powerful classifier is not needed and you need to thing about the generalization - introduce regularization, remove couple of layers from your ANN maybe. Also possible way of fighting high variance is geting <em>much, MUCH<\/em> more data.<\/p>\n\n<p>So to sum up, you need to use iterative approach and try to increase the amount of data step by step, until you get good results. There is no magic stick classifier and there is no simple answer on how much data you should use.<\/p>\n"},{"AnswerId":"43321129","CreationDate":"2017-04-10T10:45:05.680","ParentId":null,"OwnerUserId":"7434191","Title":null,"Body":"<p>It is a good idea to use CNN as the feature extractor, peel off the original fully connected layer that was used for classification and add a new classifier. This is also known as the transfer learning technique that has being widely used in the Deep Learning research community. For your problem, using the one-class SVM as the added classifier is a good choice. <\/p>\n\n<p>Specifically,<\/p>\n\n<ul>\n<li>a good CNN feature extractor can be trained on a large dataset, e.g. ImageNet,<\/li>\n<li>the one-class SVM can then be trained using your 'sunflower' dataset.<\/li>\n<\/ul>\n\n<p>The essential part of solving your problem is the implementation of the one-class SVM, which is also known as anomaly detection or novelty detection. You may refer <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/outlier_detection.html\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/modules\/outlier_detection.html<\/a> for some insights about the method.<\/p>\n"}]}
{"QuestionId":30978275,"AnswerCount":1,"Tags":"<windows-10><flashlight><torch><camera-flash>","CreationDate":"2015-06-22T11:02:08.230","AcceptedAnswerId":30979837.0,"OwnerUserId":4953112.0,"Title":"Windows 10 Phone app Turn on flash (lamp)","Body":"<p>I have a problem. I m a new windows phone 10 developer and i have a problem.<\/p>\n\n<p>How can i turn on flashlight on a Lumia 930 ? I can t find any answer on internet.<\/p>\n\n<p>Thx alot,\nRegards<\/p>\n","answers":[{"AnswerId":"30979837","CreationDate":"2015-06-22T12:18:08.620","ParentId":null,"OwnerUserId":"61529","Title":null,"Body":"<p>The Windows.Devices.Lights.Lamp namespace is where you need to look, there's an example of Lamp usage in Windows 10 <a href=\"https:\/\/github.com\/Microsoft\/Windows-universal-samples\/tree\/master\/lamp\" rel=\"nofollow\">https:\/\/github.com\/Microsoft\/Windows-universal-samples\/tree\/master\/lamp<\/a> that might be of help. \nHere's a small example based on this as an alternative to the link:<\/p>\n\n<pre><code>using (var lamp = await Lamp.GetDefaultAsync()) \n{ \n    lamp.BrightnessLevel = 1.0F;\n    lamp.IsEnabled = true;\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":30979264,"AnswerCount":1,"Tags":"<parsing><nlp><neural-network><theano><dimensionality-reduction>","CreationDate":"2015-06-22T11:48:52.107","AcceptedAnswerId":null,"OwnerUserId":4038352.0,"Title":"How to deal with different sizes of sentences when giving them as input to a Neural Network?","Body":"<p>I am giving a sentence as input to a tree structured Neural Network, where the leaf nodes will be the word vectors of the words in the sentence. <\/p>\n\n<p>That tree will be a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Branching_(linguistics)\" rel=\"nofollow noreferrer\">binarized constituency<\/a>(see the binary vs n-ary branching section) parse tree. <\/p>\n\n<p>I am trying to develop a semantic representation of the sentence. <\/p>\n\n<p><strong>The problem is<\/strong>, that since each sentence will have a different parse tree and different lengths, each sentence will have a different neural network. Due to this ever changing structure of neural network, I can't train it.<\/p>\n\n<p>But this paper develops a tree structured Neural network using constituency and dependency based parse trees-<\/p>\n\n<p>1) <a href=\"http:\/\/arxiv.org\/pdf\/1503.00075v3.pdf\" rel=\"nofollow noreferrer\">Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks<\/a> by Kai Sheng Tai, Richard Socher, and Christopher Manning<\/p>\n\n<p>And this paper uses CNNs to extract a semantic representation.<\/p>\n\n<p>2) <a href=\"http:\/\/nal.co\/papers\/KalchbrennerBlunsom_EMNLP13\" rel=\"nofollow noreferrer\">Recurrent Continuous Translation Models<\/a> by Nal Kalchbrenner,and  Phil Blunsom. This picture gives a rough idea.<img src=\"https:\/\/i.stack.imgur.com\/qvcI4.png\" alt=\"CSM\"><\/p>\n\n<p><strong>Possible Solution<\/strong>-<br>\nI can map the sentence vector to a fixed number of neurons and then use those neurons to create a tree structure. For example if there is the sentence length is 10 and the max sentence length is 20, I can create a fixed dimensionality layer of 20 neurons and then (in this particular case) the first word to the first 2 neurons, 2nd word to the 3rd and 4th neuron and so on.\nDynamic mapping can be done based on the sentence length.<\/p>\n\n<p>The weight matrix of the sentence layer to the fixed dimensionality layer will be fixed(the weights should be kept 1). No biases.<\/p>\n\n<p>But I think there will be some problems in this representation- for example- If this sentence \"I had a lovely icecream and a pastry for dessert.\" is mapped to the fixed dimensionality layer it will become \"I I had had a a lovely lovely icecream icecream and and a a pastry pastry for for dessert dessert..\". So that means that shorter sentences will have a more profound effect on the neural network compared to the longer sentences. This biasness towards shorter sentences should also create duplicate words in the output(of a sequence generator). Could someone correct me if I am wrong?<\/p>\n\n<p>I would welcome more solutions, especially ones that do not remove the relationships that words have between them in a sentence.<\/p>\n\n<p>I will be implementing this using theano and python. I am considering a layer based approach and using theano.scan to iterate over the layers to finally form the sentence representation.<\/p>\n","answers":[{"AnswerId":"33152671","CreationDate":"2015-10-15T15:33:26.793","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>The papers you cited both use recurrent architectures to deal with variable data lengths but from the graphs you posted and described, I'm not sure if you want to use recurrent structures or not.  If you don't, you'll need to pad the input to some fixed length vector based on your dataset.  (This is likely to be a poor model.)  If you do, I'd look at some of the theano framework (lasagne, keras, blocks) implementations of RNNs.<\/p>\n"}]}
{"QuestionId":30980008,"AnswerCount":1,"Tags":"<matlab><deep-learning><caffe>","CreationDate":"2015-06-22T12:27:49.553","AcceptedAnswerId":30980841.0,"OwnerUserId":2163392.0,"Title":".mat files as input for caffe deep learning network","Body":"<p>Is it possible to use .mat files to store input data for the caffe deep learning framework? if yes, how can I change the 'data' of my train_val.prototxt? I am using the googlenet network.<\/p>\n\n<p>If no, which other file formats can I use rather than mdb and hd5?<\/p>\n\n<p>For me, using .mat files is better because of their size, my .hd5 files are huge (caffe can't read them because of lack of memory) and I can't find how to save my data as .mdb files in Matlab.<\/p>\n","answers":[{"AnswerId":"30980841","CreationDate":"2015-06-22T13:09:35.657","ParentId":null,"OwnerUserId":"2732801","Title":null,"Body":"<p>Mat files version 7.3 are hdf5 files, just make sure to use this format when writing. (check documenting for save) <\/p>\n\n<p>Any hdf5 library which supports gzip compression can read mat files. <\/p>\n"}]}
{"QuestionId":30980338,"AnswerCount":2,"Tags":"<matlab><image-processing><computer-vision><deep-learning><caffe>","CreationDate":"2015-06-22T12:44:36.543","AcceptedAnswerId":30987871.0,"OwnerUserId":2163392.0,"Title":"LMDB files and how they are used for caffe deep learning network","Body":"<p>I am quite new in deep learning and I am having some problems in using the caffe deep learning network. Basically, I didn't find any documentation explaining how I can solve a series of questions and problems I am dealing right now.<\/p>\n\n<p>Please, let me explain my situation first. <\/p>\n\n<p>I have thousands of images and I must do a series of pre-processing operations on them. For each pre-processing operation, I have to save these pre-processed images as 4D matrices and also store a vector with the images labels. I will store this information as LMDB files that will be used as input for the caffe googlenet deep learning. <\/p>\n\n<p>I tried to save my images as .HD5 files, but the final file size is 80GB, which is impossible to process with the memory I have.<\/p>\n\n<p>So, the other option is using LMDB files, right? I am quite newbie in this file format and I appreciate your help in understanding how to create them in Matlab. Basically, my rookie questions are: <\/p>\n\n<p>1- These LMDB files have extension .MDB, right? is this extension the same used by microsoft access? or the right format is .lmdb and they are different? <\/p>\n\n<p>2- I find this solution for creating .mdb files (<a href=\"https:\/\/github.com\/kyamagu\/matlab-leveldb\" rel=\"nofollow\">https:\/\/github.com\/kyamagu\/matlab-leveldb<\/a>), does it create the file format needed by caffe?<\/p>\n\n<p>3- For caffe, should I have to create one .mdb file for labels and other for images or both can be fields of the same .mdb file?<\/p>\n\n<p>4- When I create an .mdb file I have to label the database fields. Can I label one field as image and other as label? does caffe understand which field means?<\/p>\n\n<p>5- what does the function (in <a href=\"https:\/\/github.com\/kyamagu\/matlab-leveldb\" rel=\"nofollow\">https:\/\/github.com\/kyamagu\/matlab-leveldb<\/a>) database.put('key1', 'value1') and database.put('key2', 'value2') do? Should I have to save my 4-d matrices in one field and the label vector in another?<\/p>\n","answers":[{"AnswerId":"30987871","CreationDate":"2015-06-22T19:07:04.557","ParentId":null,"OwnerUserId":"987599","Title":null,"Body":"<p>There is no connection between LMDB files and MS Access files.<\/p>\n\n<p>As I see it you have two options:<\/p>\n\n<ol>\n<li>Use the \"convert_imageset\" tool - it is located in caffe under the tools folder to convert a list of image files and label to lmdb.<\/li>\n<li>Instead of \"data layer\" use \"image data layer\" as an input to the network. This type of layer takes a file with a list of image file names and labels as source so you don't have to build a database (another benefit for training - you can use the shuffle option and get slightly better training results)<\/li>\n<\/ol>\n\n<p>In order to use an image data layer just replace the layer type from Data to ImageData. The source file is the path to a file containing in each line a path of an image file and the label seperated by space. For example:<\/p>\n\n<pre><code>\/path\/to\/filnename.png 23\n<\/code><\/pre>\n\n<p>If you want to do some preprocessing of the data without saving the preprocessed file to disk you can use the transformations available by caffe (mirror and cropping) (see here for information <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/data.html\">http:\/\/caffe.berkeleyvision.org\/tutorial\/data.html<\/a>) or implement your own <code>DataTransformer<\/code>.<\/p>\n"},{"AnswerId":"30987646","CreationDate":"2015-06-22T18:54:38.300","ParentId":null,"OwnerUserId":"1296661","Title":null,"Body":"<p>Caffe doesn't use LevelDB - but <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/50ab52cbbf738cc92b0eda042661b8c0172774c1\/scripts\/travis\/travis_install.sh#L50\" rel=\"nofollow\">it uses<\/a> <a href=\"http:\/\/symas.com\/mdb\/\" rel=\"nofollow\">LMDB 'Lightning' db from Symas<\/a><\/p>\n\n<p>You can try using <a href=\"https:\/\/github.com\/kyamagu\/matlab-lmdb\" rel=\"nofollow\">this<\/a> Matlab LMDB wrapper\nI personally had no experience with using LMDB with Matlab, but there is nice library for doing this from Python: <a href=\"https:\/\/lmdb.readthedocs.org\/en\/release\/\" rel=\"nofollow\">py-lmdb<\/a><\/p>\n\n<p>LMDB database is a Key\/Value db (similar to HashMap in Java or dict in Python). In order to store 4D matrices you need to understand the convention Caffe uses to save images into LMDB format.<\/p>\n\n<p>This means that the best approach to convert images to LMDB for Caffe would be doing this with Caffe.<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/50ab52cbbf738cc92b0eda042661b8c0172774c1\/examples\/mnist\/create_mnist.sh\" rel=\"nofollow\">There are examples in Caffe<\/a> on how to convert images into LMDB - I would try to repeat them and then modify scripts to use your images.<\/p>\n"}]}
{"QuestionId":30982758,"AnswerCount":1,"Tags":"<neural-network><svm><torch>","CreationDate":"2015-06-22T14:37:33.333","AcceptedAnswerId":31038255.0,"OwnerUserId":4961048.0,"Title":"Support Vector Machine in Torch7","Body":"<p>I have based my model upon the following tutorial:<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised<\/a><\/p>\n\n<p>For the last stage a neural network is used upon the features extracted from the CNN. I want to use a SVM in the final layer. How can I add that to my existing model ?<\/p>\n\n<p>It has been shown in some papers that SVM seem to function better than neural network as the final layer in the CNN and therefore I wanted to try them out to increase the accuracy of the model. Also SVM's can be use for one class classification which neural networks lack.I need a one class classifier in the end and so the need for adding an SVM to th CNN.  <\/p>\n\n<p>Kindly help<\/p>\n","answers":[{"AnswerId":"31038255","CreationDate":"2015-06-24T22:30:58.560","ParentId":null,"OwnerUserId":"2113367","Title":null,"Body":"<p><strong><em>Edit:<\/em><\/strong> My old answer was complete rubbish since you <a href=\"https:\/\/stats.stackexchange.com\/questions\/158712\/train-an-svm-via-back-propagation\/158721#158721\">cannot code a (linear) SVM as a complete module<\/a>. Instead, you can think of an SVM as <\/p>\n\n<blockquote>\n  <p>a 1-layer NN with linear activation on the output node and trained via hinge loss <\/p>\n<\/blockquote>\n\n<p>(see accepted answer's comments.)<\/p>\n\n<p>This means that in Torch, you can mock up a (linear) SVM with something like<\/p>\n\n<pre><code>linearSVM = nn.Sequential()\nlinearSVM:add(nn.Linear(ninputs, 1))\ncriterion = nn.MarginCriterion()\n<\/code><\/pre>\n\n<p>see the following question in the <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/torch7\/SVM\/torch7\/2YKxbyVkAs0\/cn_tqsyxG8cJ\" rel=\"nofollow noreferrer\">Torch7 google code mailing list<\/a>...<\/p>\n"}]}
{"QuestionId":30983213,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2015-06-22T14:56:18.570","AcceptedAnswerId":30991590.0,"OwnerUserId":2191652.0,"Title":"How to use 1-dim vector as input for caffe?","Body":"<p>I'd like to train a neural network (NN) on my own 1-dim data, which I stored in a hdf5 database for caffe. According to the documetation this should work. It also works for me as far as I only use \"Fully Connected Layers\", \"Relu\" and \"Dropout\". However I get an error when I try to use \"Convolution\" and \"Max Pooling\" layers in the NN architecture. The error complains about the input dimension of the data.  <\/p>\n\n<pre><\/pre>\n\n<p>This is the error when I only want to use a \"Pooling\" layer behind an \"InnerProduct\" layer:  <\/p>\n\n<pre><\/pre>\n\n<p>However I don't know how to change the input dimensions such that it works. \nThis is the beginning of my prototxt file specifying the network architecture:  <\/p>\n\n<pre><\/pre>\n\n<p>And this is how I output my 4D-database (with two singleton dimensions) using Matlabs h5write function:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"30991590","CreationDate":"2015-06-22T23:33:56.860","ParentId":null,"OwnerUserId":"1452257","Title":null,"Body":"<p>You seem to be outputting your data using the wrong shape. Caffe blobs have the dimensions <code>(n_samples, n_channels, height, width)<\/code> .<\/p>\n\n<p>Other than that your prototxt seems to be fine for doing predictions based on a 1D input.<\/p>\n"},{"AnswerId":"30986393","CreationDate":"2015-06-22T17:39:12.240","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>As I have no experience in using the <code>h5create<\/code> and <code>h5write<\/code> in Matlab, I am not sure on whether the training dataset is generated with the dimensions that you expect it to generate.<\/p>\n\n<p>The error msg for the convolution layer says that <code>shape[i] = -9<\/code>. This means that either the width, height, channels or number of images in a batch is being set to -9.<\/p>\n\n<p>The error msg when using pooling layer alone says that the network could detect only an input of 2D while the network is expecting an input of 4D.<\/p>\n\n<p>The error messages in both the layers are related to reshaping the blobs and this is a clear indication that the dimensions of the input are not as expected.<\/p>\n\n<p>Try debugging the Reshape functions present in blob.cpp &amp; layers\/pooling_layer.cpp to get an insight on which value is actually going rogue.<\/p>\n"}]}
{"QuestionId":30983354,"AnswerCount":3,"Tags":"<lua><torch>","CreationDate":"2015-06-22T15:01:43.873","AcceptedAnswerId":31024198.0,"OwnerUserId":3551854.0,"Title":"Is there special meaning for ()() syntax in Lua","Body":"<p>I see this type of syntax a lot in some Lua source file I was reading lately, what does it mean, especially the second pair of brackets\nAn example, line 8 in\n<a href=\"https:\/\/github.com\/karpathy\/char-rnn\/blob\/master\/model\/LSTM.lua\">https:\/\/github.com\/karpathy\/char-rnn\/blob\/master\/model\/LSTM.lua<\/a><\/p>\n\n<pre><\/pre>\n\n<p>The source code of \n<a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/Identity.lua\">https:\/\/github.com\/torch\/nn\/blob\/master\/Identity.lua<\/a><\/p>\n\n<p>********** UPDATE **************<\/p>\n\n<p>The ()() pattern is used in torch library 'nn' a lot. The first pair of bracket creates an object of the container\/node, and the second pair of bracket references the depending node.<\/p>\n\n<p>For example, y = nn.Linear(2,4)(x) means x connects to y, and the transformation is linear from 1*2 to 1*4.\nI just understand the usage, how it is wired seems to be answered by one of the answers below.<\/p>\n\n<p>Anyway, the usage of the interface is well documented below.\n<a href=\"https:\/\/github.com\/torch\/nngraph\/blob\/master\/README.md\">https:\/\/github.com\/torch\/nngraph\/blob\/master\/README.md<\/a><\/p>\n","answers":[{"AnswerId":"30983647","CreationDate":"2015-06-22T15:14:32.377","ParentId":null,"OwnerUserId":"1009479","Title":null,"Body":"<p>No, <code>()()<\/code> has no special meaning in Lua, it's just two call operators <code>()<\/code> together.<\/p>\n\n<p>The operand is possibly a function that returns a function(or, a table that implements <code>call<\/code> metamethod). For example:<\/p>\n\n<pre><code>function foo()\n  return function() print(42) end\nend\n\nfoo()()   -- 42\n<\/code><\/pre>\n"},{"AnswerId":"31024198","CreationDate":"2015-06-24T10:29:18.583","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>In complement to Yu Hao's answer let me give some Torch related precisions:<\/p>\n\n<ul>\n<li><code>nn.Identity()<\/code> creates an identity module,<\/li>\n<li><code>()<\/code> called on this module triggers <code>nn.Module<\/code> <a href=\"https:\/\/github.com\/torch\/nn\/blob\/b7aa53d96fbb6c0f2eaa1976b28c5cf12edf1ced\/Module.lua#L232-L240\"><code>__call__<\/code><\/a> (thanks to Torch class system that properly hooks up this into the metatable),<\/li>\n<li>by default this <code>__call__<\/code> method performs a forward \/ backward,<\/li>\n<li>but here <a href=\"https:\/\/github.com\/torch\/nngraph\">torch\/nngraph<\/a> is used and <strong>nngraph overrides<\/strong> this method as you can see <a href=\"https:\/\/github.com\/torch\/nngraph\/blob\/eb65b3602cdf3214ff671d5422521ad414f1c3b0\/init.lua#L22-L45\">here<\/a>.<\/li>\n<\/ul>\n\n<p>In consequence every <code>nn.Identity()()<\/code> calls has here for effect to return a <code>nngraph.Node({module=self})<\/code> node where self refers to the current <code>nn.Identity()<\/code> instance.<\/p>\n\n<p>--<\/p>\n\n<p><strong>Update<\/strong>: an illustration of this syntax in the context of <a href=\"https:\/\/en.wikipedia.org\/wiki\/Long_short_term_memory\">LSTM-s<\/a> can be found <a href=\"http:\/\/apaszke.github.io\/lstm-explained.html#computing-gate-values\">here<\/a>:<\/p>\n\n<pre><code>local i2h = nn.Linear(input_size, 4 * rnn_size)(input)  -- input to hidden\n<\/code><\/pre>\n\n<blockquote>\n  <p>If you\u2019re unfamiliar with <code>nngraph<\/code> it probably seems strange that we\u2019re constructing a module and already calling it once more with a graph node. What actually happens is that <strong>the second call converts the <code>nn.Module<\/code> to <code>nngraph.gModule<\/code> and the argument specifies it\u2019s parent in the graph<\/strong>.<\/p>\n<\/blockquote>\n"},{"AnswerId":"37304606","CreationDate":"2016-05-18T15:56:16.840","ParentId":null,"OwnerUserId":"2036809","Title":null,"Body":"<ul>\n<li>The first () calls the init function and the second () calls the call function <\/li>\n<li>If the class doesn't posses either of these functions then the parent functions are called .<\/li>\n<li><p>In the case of nn.Identity()() the nn.Identity has neither init function nor a call function hence the Identity parent nn.Module's init and call functions called .Attaching an illustration <\/p>\n\n<pre><code>require 'torch'\n\n-- define some dummy A class\nlocal A = torch.class('A')\nfunction A:__init(stuff)\n  self.stuff = stuff\n  print('inside __init of A')\nend\n\nfunction A:__call__(arg1)\nprint('inside __call__ of A')\nend\n\n-- define some dummy B class, inheriting from A\nlocal B,parent = torch.class('B', 'A')\n\nfunction B:__init(stuff)\n  self.stuff = stuff\n  print('inside __init of B')\nend\n\nfunction B:__call__(arg1)\nprint('inside __call__ of B')\nend\na=A()()\nb=B()()\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>inside __init of A\ninside __call__ of A\ninside __init of B\ninside __call__ of B\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>Another code sample <\/p>\n\n<pre><code>    require 'torch'\n\n    -- define some dummy A class\n    local A = torch.class('A')\n    function A:__init(stuff)\n      self.stuff = stuff\n      print('inside __init of A')\n    end\n\n    function A:__call__(arg1)\n    print('inside __call__ of A')\n    end\n\n    -- define some dummy B class, inheriting from A\n    local B,parent = torch.class('B', 'A')\n\n    b=B()()\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>    inside __init of A\n    inside __call__ of A\n<\/code><\/pre>\n"}]}
{"QuestionId":30994563,"AnswerCount":2,"Tags":"<artificial-intelligence><neural-network><deep-learning><caffe><deep-dream>","CreationDate":"2015-06-23T05:32:04.760","AcceptedAnswerId":31028871.0,"OwnerUserId":2795733.0,"Title":"Google Deep Dream art: how to pick a layer in a neural network and enhance it","Body":"<p>I am interested in a recent blog post by Google that describes the use of  to make art. <\/p>\n\n<p>I am particularly interested in one technique: <\/p>\n\n<blockquote>\n  <p>'In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at a different level of abstraction, so the complexity of features we generate depends on which layer we choose to enhance. For example, lower layers tend to produce strokes or simple ornament-like patterns, because those layers are sensitive to basic features such as edges and their orientations.' <\/p>\n<\/blockquote>\n\n<p>The post is <a href=\"http:\/\/googleresearch.blogspot.co.uk\/2015\/06\/inceptionism-going-deeper-into-neural.html?m=1\" rel=\"nofollow\">http:\/\/googleresearch.blogspot.co.uk\/2015\/06\/inceptionism-going-deeper-into-neural.html?m=1<\/a>. <\/p>\n\n<p><strong>My question<\/strong>: the post describes this as a 'simple' case--is there an open-source implementation of a nn that could be used for this purpose in a relatively plug-and-play process? \nFor just the technique described, does the network need to be trained? <\/p>\n\n<p>No doubt for other techniques mentioned in the paper one needs a network already trained on a large number of images, but for the one I've described is there already some kind of open-source network layer visualization package?<\/p>\n","answers":[{"AnswerId":"31708507","CreationDate":"2015-07-29T18:34:46.807","ParentId":null,"OwnerUserId":"5086088","Title":null,"Body":"<p>In the link to Ipython notebook Dmitry provided, it says that it does <strong>gradient<\/strong> <strong>ascent<\/strong> with <strong>maximizing<\/strong> L2 normalization. I believe this is what Google means to be enhance the feature from a algorithmic perspective. <\/p>\n\n<p>If you think about it, it's really the case, minimizing L2 would prevent over-fitting, i.e. make the curve looks smoother. If you do the opposite, you are making the feature more obvious.<\/p>\n\n<p>Here is a great link to understand <a href=\"http:\/\/www.onmyphd.com\/?p=gradient.descent\" rel=\"noreferrer\">gradient ascent<\/a>, though it talks about gradient descent mainly.<\/p>\n\n<p>I don't know much about implementation details in caffe, since I use theano mostly. Hope it helps!<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>So I read about the detailed articles [1],[2],[3],[4] today and find out that <a href=\"http:\/\/arxiv.org\/pdf\/1312.6034v2.pdf\" rel=\"noreferrer\">[3]<\/a> actually talks about the algorithm in details<\/p>\n\n<blockquote>\n  <p>A locally-optimal <em>I<\/em> can be found by the back-propagation\n  method. The procedure is related to the ConvNet training procedure, where the back-propagation is\n  used to optimise the layer weights. The difference is that in our case the optimisation is performed\n  with respect to the input image, while the weights are fixed to those found during the training stage.\n  We initialised the optimisation with the zero image (in our case, the ConvNet was trained on the\n  zero-centred image data), and then added the training set mean image to the result.<\/p>\n<\/blockquote>\n\n<p>Therefore, after training the network on classification, you train it again w.r.t to the input image, using gradient ascent in order to get higher score for the class.<\/p>\n"},{"AnswerId":"31028871","CreationDate":"2015-06-24T14:07:25.683","ParentId":null,"OwnerUserId":"2145920","Title":null,"Body":"<p>UPD: Google posted more detail instructions how they implemented it: <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb<\/a><\/p>\n\n<p>There's also another project: <a href=\"https:\/\/317070.github.io\/Dream\/\" rel=\"noreferrer\">https:\/\/317070.github.io\/Dream\/<\/a><\/p>\n\n<p>If you read <a href=\"https:\/\/i.stack.imgur.com\/6NRkY.jpg\" rel=\"noreferrer\">1<\/a>,[2],[3],[4] from your link, you'll see that they used Caffe. This framework already contains the trained networks to play with. You don't need to train anything manually, just download the models using .sh scripts in the <code>models\/<\/code> folder.<\/p>\n\n<p>You want \"plug-and-play process\", it's not so easy because besides the framework, we need the code of the scripts they used and, probably, patch Caffe. I tried to make something using their description. Caffe has Python and Matlab interface but there's more in its internals. <\/p>\n\n<p>The text below describes my thoughts on how it could be possibly implemented. I'm not sure about my words so it's more like an invitation to research with me than the \"plug-and-play process\". But as no one still answered, let me put it here. Maybe someone will fix me.<\/p>\n\n<p>So<\/p>\n\n<p>As far as I understand, they run optimization<\/p>\n\n<p><code>[sum((net.forwardTo(X, n)  - enchanced_layer).^2) + lambda * R(X)] -&gt; min<\/code><\/p>\n\n<p>I.e. look for such input <code>X<\/code> so that the particular layer of the netword would produce the \"enchanced\" data instead of the \"original\" data.<\/p>\n\n<p>There's a regularization constraint <code>R(X)<\/code>: <code>X<\/code> should look like \"natural image\" (without high-frequency noise).<\/p>\n\n<p><code>X<\/code> is our target image. The initial point <code>X0<\/code> is the original image.\n<code>forwardTo(X, n)<\/code> is what our network produces in the layer <code>n<\/code> when we feed the input with X. If speak about Caffe, you can make full-forward pass (<code>net.forward<\/code>) and look at the blob you are interested in (<code>net.blob_vec(n).get_data()<\/code>).<\/p>\n\n<p><code>enchanced_layer<\/code> - we take the original layer blob and \"enchance\" signals in it. What does it mean, I don't know. Maybe they just multiply the values by coefficient, maybe something else.<\/p>\n\n<p>Thus <code>sum((forwardTo(X, n)  - enchanced_net).^2)<\/code>  will become zero when your input image produces exactly what you want in the layer <code>n<\/code>.<\/p>\n\n<p><code>lambda<\/code> is the regularization parameter and <code>R(X)<\/code> is how <code>X<\/code> looks natural. I didn't implement it and my results look very noisy. As for it's formula, you can look for it at [2].<\/p>\n\n<p>I used Matlab and <code>fminlbfgs<\/code> to optimize.<\/p>\n\n<p>The key part was to find the gradient of the formula above because the problem has too many dimensions to calculate the gradient numerically.<\/p>\n\n<p>As I said, I didn't manage to find the gradient of <code>R(X)<\/code>. As for the main part of the formula, I managed to find it this way:<\/p>\n\n<ul>\n<li>Set diff blob at the layer <code>n<\/code> to <code>forwardTo(X, n)  - enchanced_net<\/code>. (see caffe documentation for <code>set_diff<\/code> and <code>set_data<\/code>, <code>set_data<\/code> is used for forward and waits for data and <code>set_diff<\/code> is used for backward propagation and waits for data errors).<\/li>\n<li>Perform <em>partial<\/em> backpropagation from layer <code>n-1<\/code> to the input.<\/li>\n<li>Input diff blob would contain the gradient we need.<\/li>\n<\/ul>\n\n<p>Python and Matlab interfaces do NOT contain partial backward propagation but Caffe C++ internals contain it. I added a patch below to make it available in Matlab.<\/p>\n\n<p>Result of enhancing the 4th layer:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/6NRkY.jpg\" alt=\"Result of enhancing the 4th layer\"><\/p>\n\n<p>I'm not happy with the results but I think there's something in common with the article. <\/p>\n\n<ul>\n<li>Here's the code that produces the picture above \"as is\". The entry point is \"run2.m\", \"fit2.m\" contains the fitness function: <a href=\"https:\/\/github.com\/galchinsky\/caf\" rel=\"noreferrer\">https:\/\/github.com\/galchinsky\/caf<\/a><\/li>\n<li>Here's caffe patch to Matlab interface to make partial backpropagation available: <a href=\"https:\/\/gist.github.com\/anonymous\/53d7cb44c072ae6320ff\" rel=\"noreferrer\">https:\/\/gist.github.com\/anonymous\/53d7cb44c072ae6320ff<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":31005463,"AnswerCount":3,"Tags":"<python><neural-network><caffe>","CreationDate":"2015-06-23T14:22:06.590","AcceptedAnswerId":null,"OwnerUserId":1107562.0,"Title":"How to train new fast-rcnn imageset","Body":"<p>I am using <strong>fast-rcnn<\/strong> and try to train the system for new class (label)\nI followed this: <a href=\"https:\/\/github.com\/EdisonResearch\/fast-rcnn\/tree\/master\/help\/train\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdisonResearch\/fast-rcnn\/tree\/master\/help\/train<\/a> <\/p>\n\n<ol>\n<li><p>Placed the images<\/p><\/li>\n<li><p>Placed the annotations <\/p><\/li>\n<li><p>Prepare the ImageSet with all the image name prefix<\/p><\/li>\n<li><p>Prepared selective search output: train.mat<\/p><\/li>\n<\/ol>\n\n<p><strong>I failed while running the train_net.py with the following error:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>My Questions is:<\/p>\n\n<ol>\n<li>Why am I having this error?<\/li>\n<li>Do I need to rescale the images to fix: 256x256 before training?<\/li>\n<li>Do I need to prepare something in order to set the  class?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"46239725","CreationDate":"2017-09-15T12:34:31.907","ParentId":null,"OwnerUserId":"7596504","Title":null,"Body":"<p>Check out the solution described in the following blog post, Part 4, Issue #4. The solution is to flip the x1 and x2 coordinate values.<\/p>\n\n<p><a href=\"https:\/\/huangying-zhan.github.io\/2016\/09\/22\/detection-faster-rcnn.html\" rel=\"nofollow noreferrer\">https:\/\/huangying-zhan.github.io\/2016\/09\/22\/detection-faster-rcnn.html<\/a><\/p>\n\n<p>Following is copied from link:<\/p>\n\n<p>box [:, 0] > box[:, 2]<\/p>\n\n<p>Solution: add the following code block in imdb.py<\/p>\n\n<pre><code>def append_flipped_images(self):\nnum_images = self.num_images\nwidths = self._get_widths()\nfor i in xrange(num_images):\n    boxes = self.roidb[i]['boxes'].copy()\n    oldx1 = boxes[:, 0].copy()\n    oldx2 = boxes[:, 2].copy()\n    boxes[:, 0] = widths[i] - oldx2\n    boxes[:, 2] = widths[i] - oldx1\n    for b in range(len(boxes)):\n            if boxes[b][2] &lt; boxes[b][0]:\n                boxes[b][0]=0\n    assert (boxes[:, 2] &gt;= boxes[:, 0]).all()\n<\/code><\/pre>\n"},{"AnswerId":"31603956","CreationDate":"2015-07-24T06:47:29.577","ParentId":null,"OwnerUserId":"3985163","Title":null,"Body":"<ol>\n<li>it says that there exists <code>boxes[:,2] &lt; boxes[:, 0]<\/code>, <code>boxes[:, 2]<\/code> is the x-max of bounding box while <code>boxes[:, 0]<\/code> is x-min. So the problem is related to region proposal. I came across with this problem too. I found that it was causes by overflow. I remember that the dtype for boxes is np.uint8(need to check), if the image is too big, you get this error.<\/li>\n<li>rescale is one solution, however this may influence the performance. You can change the dtype from uint8 to float instead.<\/li>\n<li>As far as I know, there is no need for that.<\/li>\n<\/ol>\n"},{"AnswerId":"31678091","CreationDate":"2015-07-28T13:43:07.810","ParentId":null,"OwnerUserId":"1323010","Title":null,"Body":"<p>I'm late to the party, but when I was editing the code this was my yakkity-hack of a solution<\/p>\n\n<pre><code>        for b in range(len(boxes)):\n            if boxes[b][2] &lt; boxes[b][0]:\n                boxes[b][0] = 0\n        assert (boxes[:, 2] &gt;= boxes[:, 0]).all()\n<\/code><\/pre>\n\n<p>There are smarter ways to do this, as <em>every single grad student<\/em> seems to point out, but this works fine.<\/p>\n"}]}
{"QuestionId":31008493,"AnswerCount":2,"Tags":"<python><caffe>","CreationDate":"2015-06-23T16:34:42.337","AcceptedAnswerId":null,"OwnerUserId":286579.0,"Title":"Caffe: Drawing CNN Net","Body":"<p>I used python code to draw Net defined in prototext file as:<\/p>\n\n<pre><\/pre>\n\n<p>It fails to draw. It does not show any error but the results test.png file is white blank image file. Can anyone please help me in fixing it? It would really help to design new nets quickly.<\/p>\n","answers":[{"AnswerId":"35239450","CreationDate":"2016-02-06T09:33:51.583","ParentId":null,"OwnerUserId":"1044366","Title":null,"Body":"<p>Somwhere in mid 2014, Caffe <a href=\"https:\/\/github.com\/BVLC\/caffe\/releases\/tag\/v0.999\" rel=\"nofollow\">changed their proto definition for extensibility<\/a> which causes this problem. As a result of this change, all the proto files have to be updated to the newer definition. <\/p>\n\n<p>To do this, Caffe provides the following tools in the <code>distribute\/bin\/<\/code> or <code>.build_release\/tools<\/code> directory:<\/p>\n\n<ol>\n<li><code>upgrade_net_proto_binary.bin<\/code> <\/li>\n<li><code>upgrade_net_proto_text.bin<\/code><\/li>\n<\/ol>\n\n<p>Here is a simple example of how to convert your proto text file to a newer format:<\/p>\n\n<pre><code>.\/upgrade_net_proto_text.bin \/path\/to\/older_proto_file \/path\/to\/newer_ouput_proto_file\n<\/code><\/pre>\n"},{"AnswerId":"33389706","CreationDate":"2015-10-28T11:21:49.580","ParentId":null,"OwnerUserId":"2528545","Title":null,"Body":"<p>I had same problem. Based on <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/d6fpZTOq5M0\" rel=\"nofollow\">this thread<\/a>, I've managed to solve this by using older Proto syntax as suggested. For instance I had to do this:<\/p>\n\n<p>Rename layers definition from <code>layers<\/code> to <code>layer<\/code>. All layer type rename by caffe documentation (or by example proto files) - i.e. layer <code>type: CONVOLUTION<\/code> to <code>type: \"Convolution\"<\/code>, etc. Substitute newer syntax:<\/p>\n\n<pre><code>blobs_lr: 1         \nblobs_lr: 1         \nweight_decay: 1\nweight_decay: 0\n<\/code><\/pre>\n\n<p>for <\/p>\n\n<pre><code>param {\n    name: \"conv1_w\"\n    lr_mult: 1              \n    decay_mult: 1\n}\nparam {\n    name: \"conv1_b\"\n    lr_mult: 2              \n    decay_mult: 0\n}\n<\/code><\/pre>\n\n<p>Now parsing and new-drawing works just fine. Refer to example .prototxt files in caffe package to get better intuition, how working proto syntax looks like.<\/p>\n"}]}
{"QuestionId":31008673,"AnswerCount":2,"Tags":"<python><exception-handling><pycharm><pydev><theano>","CreationDate":"2015-06-23T16:44:08.180","AcceptedAnswerId":null,"OwnerUserId":5041148.0,"Title":"PyCharm break on exception does not work with Theano","Body":"<p>I cannot get PyCharm to stop on the line of code where an exception is raised, when I import Theano.<\/p>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n\n<p>I expect PyCharm debugger to stop on the  line, but it throws a  and exits the debugger:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34999303","CreationDate":"2016-01-25T18:09:00.103","ParentId":null,"OwnerUserId":"648265","Title":null,"Body":"<p>Looks like a bug in one of the libs (maybe both :^)).<\/p>\n\n<p>For some reason, <code>theano<\/code>'s and <code>PyCharm<\/code>'s excepthooks both think of the other one as its ancestor.<\/p>\n\n<p>Add debug printing into both libs at points where <code>sys.excepthook<\/code> and member variables that point to the previous handler are set to reveal the setting order. Someone appears to be breaking the handler chaining rules.<\/p>\n"},{"AnswerId":"34994645","CreationDate":"2016-01-25T14:10:54.230","ParentId":null,"OwnerUserId":"5041148","Title":null,"Body":"<p>One hack is to comment out this line <code>sys.excepthook = thunk_hook<\/code> in <code>...\/lib\/python2.7\/site-packages\/theano\/gof\/link.py<\/code><\/p>\n"}]}
{"QuestionId":31011485,"AnswerCount":0,"Tags":"<build><lua><compiler-errors><deep-learning><torch>","CreationDate":"2015-06-23T19:12:20.703","AcceptedAnswerId":null,"OwnerUserId":3113501.0,"Title":"Attempting to install FBLuaLib","Body":"<p>FbLuaLib: <a href=\"https:\/\/github.com\/facebook\/fblualib\" rel=\"nofollow\">https:\/\/github.com\/facebook\/fblualib<\/a><\/p>\n\n<p>To install FBLuaLib, I need to install thpp: <a href=\"https:\/\/github.com\/facebook\/thpp\" rel=\"nofollow\">https:\/\/github.com\/facebook\/thpp<\/a><\/p>\n\n<p>I can build folly, fbthrift just fine, but when I run .\/build.sh for thpp I get the following error message:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":31019371,"AnswerCount":1,"Tags":"<machine-learning><artificial-intelligence><regression><theano><pylearn>","CreationDate":"2015-06-24T06:40:26.440","AcceptedAnswerId":31021241.0,"OwnerUserId":3993741.0,"Title":"Pylearn2 example for time series or sequence prediction","Body":"<p>Can Pylearn2 be used for time series or sequence prediction of continuous numerical data?  Can an LSTM recurrent neural network in Pylearn2 be used for this?  If so, can someone post an example code in Pylearn2\/Theano\/Python?  <\/p>\n","answers":[{"AnswerId":"31021241","CreationDate":"2015-06-24T08:15:29.670","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>My understanding is that PyLearn2 is still not great for any kind of recurrent network, though I believe they are intending to improve support for these kinds of models.<\/p>\n\n<p>Having said that, there is experimental support, including an LSTM implementation.<\/p>\n\n<p>Take a look in the PyLearn2 source code in the directory <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/tree\/master\/pylearn2\/sandbox\/rnn\" rel=\"nofollow\">pylearn2\/sandbox\/rnn<\/a>, and in particular at the contents of <a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/sandbox\/rnn\/models\/rnn.py\" rel=\"nofollow\">pylearn2\/sandbox\/rnn\/models\/rnn.py<\/a> where you'll find an LSTM implementation.<\/p>\n\n<p>Because of its experimental nature, this code may not work properly, may not be supported fully, and the documentation may be incomplete or inaccurate.<\/p>\n\n<p>If you're willing to forego the intended ease of use benefits of PyLearn2 and work at a more detailled level then recurrent neural newtworks can be implemented just fine in Theano. There are many tutorials for this, including:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\" rel=\"nofollow\">LSTM Networks for Sentiment Analysis<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/rnnrbm.html\" rel=\"nofollow\">Modeling and generating sequences of polyphonic music with the RNN-RBM<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/rnnslu.html\" rel=\"nofollow\">Recurrent Neural Networks with Word Embeddings<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":31025226,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-06-24T11:23:52.030","AcceptedAnswerId":31044506.0,"OwnerUserId":664456.0,"Title":"Theano set_value for casted shared variable","Body":"<p>In the Theano deep learning tutorial, y is a shared variable that is casted:<\/p>\n\n<pre><\/pre>\n\n<p>I later want to set a new value for y.<\/p>\n\n<p>For GPU this works:<\/p>\n\n<pre><\/pre>\n\n<p>For CPU this works:<\/p>\n\n<pre><\/pre>\n\n<p>Why does this require a different syntax between GPU and CPU? I would like my code to work for both cases, am I doing it wrong?<\/p>\n","answers":[{"AnswerId":"31044506","CreationDate":"2015-06-25T08:00:19.167","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This is a very similar problem to that described in another <a href=\"https:\/\/stackoverflow.com\/questions\/30627302\/can-not-update-a-subset-of-shared-tensor-variable-after-a-cast\/30651282#30651282\">StackOverflow question<\/a>.<\/p>\n\n<p>The problem is that you are using a <em>symbolic<\/em> cast operation which turns the shared variable into a symbolic variable.<\/p>\n\n<p>The solution is to cast the shared variable's value rather than the shared variable itself.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>y = theano.shared(numpy.asarray(data, dtype=theano.config.floatX))\ny = theano.tensor.cast(y, 'int32')\n<\/code><\/pre>\n\n<p>Use<\/p>\n\n<pre><code>y = theano.shared(numpy.asarray(data, dtype='int32'))\n<\/code><\/pre>\n\n<p>Navigating the Theano computational graph via the <code>owner<\/code> attribute is considered bad form. If you want to alter the shared variable's value, maintain a Python reference to the shared variable and set its value directly.<\/p>\n\n<p>So, with y being just a shared variable, and not a symbolic variable, you can now just do:<\/p>\n\n<pre><code>y.set_value(np.asarray(data2, dtype='int32'))\n<\/code><\/pre>\n\n<p>Note that the casting is happening in numpy again, instead of Theano.<\/p>\n"}]}
{"QuestionId":31028519,"AnswerCount":2,"Tags":"<computer-vision><deep-learning><caffe>","CreationDate":"2015-06-24T13:52:36.120","AcceptedAnswerId":null,"OwnerUserId":2163392.0,"Title":"GoogleNet can't read images when ImageData type field is used in train_val.prototxt","Body":"<p>I am trying o use caffe's implementation of GoogleNet. I want to train the deep network according to a list of files and labels in a text file, but the problem is that, when I train the deep network, it can't read the files.<\/p>\n\n<p>Here is the train_val.prototxt definitions, where I use ImageData instead of using big LMDB files with 'Data' type<\/p>\n\n<pre><\/pre>\n\n<p>Here I used ImageData type for the googlenet rather than type Data as suggested here: <a href=\"https:\/\/stackoverflow.com\/questions\/30980338\/lmdb-files-and-how-they-are-used-for-caffe-deep-learning-network\">LMDB files and how they are used for caffe deep learning network<\/a><\/p>\n\n<p>So, I have the text file (file_paths_and_labels.txt) where each line contain the following:<\/p>\n\n<pre><\/pre>\n\n<p>where path to image is the image's address and label is the label of the image (there are 10 different labels).<\/p>\n\n<p>I want to know exactly where I am wrong because when I run the deep network training command<\/p>\n\n<pre><\/pre>\n\n<p>I have the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I think GoogleNet is not finding data in my textfile. What is the problem? the syntax of my train_val.prototxt file?<\/p>\n","answers":[{"AnswerId":"31141286","CreationDate":"2015-06-30T14:37:15.650","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>You're specifying the source using the wrong parameter. For IMAGE_DATA you need to use <strong>image_data_param<\/strong> instead of <strong>data_param<\/strong>. Because you specify your source in data_param, and ImageDataLayer looks at image_data_param, the value of source is the empty string. You can see that in the log here:<\/p>\n\n<pre><code>I0624 10:36:11.525106 15246 image_data_layer.cpp:36] Opening file \n<\/code><\/pre>\n\n<p>The format of this line should be:<\/p>\n\n<pre><code>Opening file &lt;filename&gt;\n<\/code><\/pre>\n\n<p>while in your log there's an empty space following \"Opening file\".<\/p>\n"},{"AnswerId":"33526867","CreationDate":"2015-11-04T16:20:20.513","ParentId":null,"OwnerUserId":"1351629","Title":null,"Body":"<p>That my be helpful for people who will land here in future.<br>\nI had kind of the same problem,   but I was loading images from <code>leveldb<\/code> file.<br>\nAnd this error appeared when I copied <code>leveldb<\/code> files, generated on machine <strong>A<\/strong>, to another machine <strong>B<\/strong> and tried to run caffe on <strong>B<\/strong>.\nThe problem was solved by regenerating <code>leveldb<\/code> files once again on the machine <strong>B<\/strong>.<\/p>\n\n<p>BTW. may be anybody knows why does the machine on which was <code>leveldb<\/code> generated matter?<\/p>\n"}]}
{"QuestionId":31030032,"AnswerCount":1,"Tags":"<lua><deep-learning><torch><conv-neural-network>","CreationDate":"2015-06-24T14:59:12.673","AcceptedAnswerId":null,"OwnerUserId":3805104.0,"Title":"How to prepare data for torch7 deep learning convolutional neural network example?","Body":"<p>I have been trying to use convolutional neural network example in deep learning library of torch7(<a href=\"https:\/\/github.com\/nicholas-leonard\/dp\/blob\/master\/examples\/convolutionneuralnetwork.lua\" rel=\"nofollow\">convolutionalneuralnetwork.lua<\/a>) for my own dataset. I have a dataset of 100x100 binary jpg images and they are in the following directories:<br>\n\/home\/akshay\/project\/data\/train -- training data<br>\n\/home\/akshay\/project\/data\/valid  -- validation data<\/p>\n\n<p>I have changed the dataset to ImageSource and made the other necessary changes as in the code:<\/p>\n\n<pre><\/pre>\n\n<p>But when I ran the code, I got an error as below:<\/p>\n\n<pre><\/pre>\n\n<p>1).How do I prepare the data differently?<br>\n2).Are the parameters passed wrong and How to correct them?<\/p>\n","answers":[{"AnswerId":"32724840","CreationDate":"2015-09-22T19:01:42.470","ParentId":null,"OwnerUserId":"5364826","Title":null,"Body":"<p>I was getting the same error. Change your --loadSize to be '1,100,100' instead of '{1,100,100}'. Similarly for --sampleSize.   <\/p>\n"}]}
{"QuestionId":31032814,"AnswerCount":1,"Tags":"<linear-algebra><torch>","CreationDate":"2015-06-24T17:08:38.730","AcceptedAnswerId":31033424.0,"OwnerUserId":2113367.0,"Title":"Mean Centering a Tensor in torch7","Body":"<p>I was wondering if there was an easier (or more efficient) way to perform a mean centering operation? <\/p>\n\n<p>Currently, I'm doing the following:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31033424","CreationDate":"2015-06-24T17:43:47.670","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>mean = data:mean(1)\ndata:add(-1, mean:expandAs(data))\n<\/code><\/pre>\n"}]}
{"QuestionId":31034293,"AnswerCount":1,"Tags":"<theano><pymc3>","CreationDate":"2015-06-24T18:32:12.050","AcceptedAnswerId":null,"OwnerUserId":1215364.0,"Title":"Is it possible to use NUTS in PyMC3 with a model that involves the eigendecomposition of parameters?","Body":"<p>I have a model where the likelihood involves computing the sum of all the terms in the matrix<\/p>\n\n<p>P = U exp(tD) U^-1<\/p>\n\n<p>Where <\/p>\n\n<p>UDU^-1 = Q<\/p>\n\n<p>and Q is my matrix of parameters.If I wanted to use NUTS in PyMC3, NUTS would have to be able to compute the derivative of all the elements in P with respect to each of the elements in Q. Is this possible using the symbolic differentiator in Theano that PyMC3 uses?<\/p>\n","answers":[{"AnswerId":"31089162","CreationDate":"2015-06-27T13:10:17.100","ParentId":null,"OwnerUserId":"2288595","Title":null,"Body":"<p>PyMC3 uses Theano for computation and autodiff. Theano has very good support for tensor algebra (of which matrix algebra is a subset) so I think that your model should be supported. <\/p>\n"}]}
{"QuestionId":31036340,"AnswerCount":1,"Tags":"<linux><lua><torch>","CreationDate":"2015-06-24T20:26:09.090","AcceptedAnswerId":null,"OwnerUserId":864128.0,"Title":"Torch7 Lua, error loading module 'libpaths' (Linux)","Body":"<p>I am a new user to .\nI have trouble loading module '' (on ). The <strong>error log is<\/strong>:<\/p>\n\n<blockquote>\n  <p>Exception in thread \"main\" com.naef.jnlua.LuaRuntimeException: error\n  loading module 'libpaths' from file\n  '\/usr\/local\/lib\/lua\/5.1\/libpaths.so':\n    \/usr\/local\/lib\/lua\/5.1\/libpaths.so: undefined symbol: lua_gettop    at\n  com.naef.jnlua.LuaState.lua_pcall(Native Method)  at\n  com.naef.jnlua.LuaState.call(LuaState.java:555)   at\n  org.eclipse.koneki.ldt.support.lua51.internal.interpreter.JNLua51Launcher.run(JNLua51Launcher.java:128)\n    at\n  org.eclipse.koneki.ldt.support.lua51.internal.interpreter.JNLua51DebugLauncher.main(JNLua51DebugLauncher.java:24)<\/p>\n<\/blockquote>\n\n<p>What might be the problem? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"31160347","CreationDate":"2015-07-01T11:44:16.743","ParentId":null,"OwnerUserId":"5039693","Title":null,"Body":"<p>This is how to configure torch + eclipse:<\/p>\n\n<p><strong>1) Configure the Lua interpreter with torch<\/strong>:<\/p>\n\n<p>Go to Windows -> Preference -> Lua -> interpreter:<\/p>\n\n<ul>\n<li><p>Interpreter Type   : Lua 5.2<\/p><\/li>\n<li><p>Interpreter executable : \/opt\/torch\/install\/bin\/qlua    (-> this is\nrequired to use qt features)<\/p><\/li>\n<li><p>Interpreter name : Qt + Torch   Interpreter arguments : -lenv -e\n\"io.stdout:setvbuf('no'); if os.getenv('DEBUG_MODE') then require\n'debugger' ; require 'debugger.plugins.ffi'end\"<\/p><\/li>\n<li><p>LinkedExecution argument : Lua 5.2<\/p><\/li>\n<\/ul>\n\n<p><strong>2)<\/strong> Pick this interpreter as a default interpreter <\/p>\n\n<p><strong>3)<\/strong> Also Working with an external interpreter, require that  \"LuaSocket\" packet is installed, \nYou will get a message error of \"libsocket.so not found\" when debugging if it is not installed<\/p>\n\n<p><strong>To install LuaSocket, you may try<\/strong> :<\/p>\n\n<pre><code>sudo luarocks install luasocket --only-server=http:\/\/luarocks.org\/repositories\/rocks-scm\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>sudo luarocks install luasocket\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>luarocks install luasocket\n<\/code><\/pre>\n\n<p>Credits to <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/torch7\/wWwPjL9PgBU\/MkloTN7QqT8J\" rel=\"nofollow\">STRUB Floriab<\/a><\/p>\n"}]}
{"QuestionId":31036680,"AnswerCount":1,"Tags":"<python><numpy><ffi><theano><python-cffi>","CreationDate":"2015-06-24T20:45:33.167","AcceptedAnswerId":null,"OwnerUserId":669329.0,"Title":"Getting a C pointer to a function generated by Theano?","Body":"<p>I would like to use a Theano function from C\/Fortran code (in particular, I want to use an implicit ODE solver written in Fortran with a function created in Theano). Are there any examples\/resources on how to do that?<\/p>\n","answers":[{"AnswerId":"31044667","CreationDate":"2015-06-25T08:08:49.840","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You've tagged your question with ffi\/cffi but that's for calling foreign code from Python. However it sounds like you actually want to call Python\/Theano code from C\/Fortran. For that, the documentation on <a href=\"https:\/\/docs.python.org\/2\/extending\/embedding.html\" rel=\"nofollow\">Embedding Python in Another Application<\/a> might be helpful.<\/p>\n\n<p>In principle you could just run Theano Python code from your C\/Fortran code via facilities in <code>Python.h<\/code>.<\/p>\n\n<p>Although Theano compiles some operations via C code, I don't believe it produces an natively executable function\/library for the entire computation graph that could then be linked in by some other, non-Python, application.<\/p>\n\n<p>Update: via the <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/0hl6O8f260U\" rel=\"nofollow\">thread on the Theano mailing list<\/a>... apparently <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/1408#issuecomment-106029111\" rel=\"nofollow\">a prototype for having Theano create a linkable library<\/a> was done some time ago but isn't currently integrated into Theano.<\/p>\n"}]}
{"QuestionId":31050976,"AnswerCount":2,"Tags":"<python><anaconda><theano>","CreationDate":"2015-06-25T12:53:59.893","AcceptedAnswerId":null,"OwnerUserId":2998813.0,"Title":"python.exe crashes when importing `theano`","Body":"<p>I am using Anaconda(2.2 64bit) on a Windows 7 64-bit machine.\nWhen I try to<\/p>\n\n<pre><\/pre>\n\n<p>Python crashes without infomation. I installed  using Anaconda.<\/p>\n\n<p>Does anyone know where this problem comes from?<\/p>\n","answers":[{"AnswerId":"31051951","CreationDate":"2015-06-25T13:35:30.117","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This could be caused by a missing, or incorrectly configured C compiler. There's a lot of help about installing Theano on Windows <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">in the documentation<\/a>.<\/p>\n\n<p>In particular, make sure you've run<\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n"},{"AnswerId":"31063985","CreationDate":"2015-06-26T02:40:33.773","ParentId":null,"OwnerUserId":"2998813","Title":null,"Body":"<p>I solve this problem by using anaconda 2.10 instead of anaconda 2.20.but I still have no idear about the reason.If anyone can tell me,I will appreiciate<\/p>\n"}]}
{"QuestionId":31055033,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2015-06-25T15:48:16.763","AcceptedAnswerId":31081253.0,"OwnerUserId":4561745.0,"Title":"Caffe: Extremely high loss while learning simple linear functions","Body":"<p>I'm trying to train a neural net to learn the function . The objective is to play around with Caffe in order to learn and understand it better. The data required are synthetically generated in python and written to memory as an lmdb database file.<\/p>\n\n<p>Code for data generation:<\/p>\n\n<pre><\/pre>\n\n<p>Solver.prototext file:<\/p>\n\n<pre><\/pre>\n\n<p>Caffe model:<\/p>\n\n<pre><\/pre>\n\n<p>The loss on the test data that I'm getting is . This is shocking as the loss is three orders of magnitude greater than the numbers in the training and test data sets. Also, the function to be learned is a simple linear function. I can't seem to figure out what is wrong in the code. Any suggestions\/inputs are much appreciated.<\/p>\n","answers":[{"AnswerId":"31081253","CreationDate":"2015-06-26T20:03:38.397","ParentId":null,"OwnerUserId":"4561745","Title":null,"Body":"<p>The loss generated is a lot in this case because Caffe only accepts data (i.e. <code>datum.data<\/code>) in the <code>uint8<\/code> format and labels (<code>datum.label<\/code>) in <code>int32<\/code> format. However, for the labels, <code>numpy.int64<\/code> format also seems to be working. I think <code>datum.data<\/code> is accepted only in <code>uint8<\/code> format because Caffe was primarily developed for Computer Vision tasks where inputs are images, which have RGB values in [0,255] range. <code>uint8<\/code> can capture this using the least amount of memory. I made the following changes to the data generation code:<\/p>\n\n<pre><code>Xtrain = np.uint8(np.random.randint(0,256, size = (Ntrain,K,H,W)))\nXtest = np.uint8(np.random.randint(0,256, size = (Ntest,K,H,W)))\n\nytrain = int(Xtrain[:,0,0,0]) + int(Xtrain[:,1,0,0]) + int(Xtrain[:,2,0,0])\nytest = int(Xtest[:,0,0,0]) + int(Xtest[:,1,0,0]) + int(Xtest[:,2,0,0])\n<\/code><\/pre>\n\n<p>After playing around with the net parameters (learning rate, number of iterations etc.) I'm getting an error of the order of 10^(-6) which I think is pretty good!<\/p>\n"}]}
{"QuestionId":31057219,"AnswerCount":1,"Tags":"<python><python-2.7><gpu><theano><pydot>","CreationDate":"2015-06-25T17:45:08.717","AcceptedAnswerId":null,"OwnerUserId":612837.0,"Title":"Using GPU with Theano","Body":"<p>I'm trying to execute the next code <a href=\"https:\/\/github.com\/erogol\/KLP_KMEANS\/blob\/master\/klp_kmeans.py\" rel=\"nofollow\">https:\/\/github.com\/erogol\/KLP_KMEANS\/blob\/master\/klp_kmeans.py<\/a> using my gpu<\/p>\n\n<p>I execute:<\/p>\n\n<p>THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python klp_kmeans.py <\/p>\n\n<p>But it says:<\/p>\n\n<pre><\/pre>\n\n<p>After doing a little debug, I noticed that it has detected cpu usage due to an instance of class 'Gemm' (checked in line #71)<\/p>\n\n<p>Why is not using the gpu?<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"31069130","CreationDate":"2015-06-26T09:06:14.697","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You really need to contact the author of this script for support. The '<code>Used the cpu<\/code>' message is coming from this script, not from Theano. It's the author's code that is doing the detection, and that detection logic may be faulty.<\/p>\n\n<p>As far as Theano is concerned, given your <code>THEANO_FLAGS<\/code> and the '<code>Using gpu device 0: GeForce GTX 750 Ti<\/code>' message you see at startup, it will use the GPU for all computation graphs that can be run on the GPU.<\/p>\n\n<p>Are you running the script as is? If so, it looks like parts of it are intended to be run on the CPU, and other parts on the GPU (it seems to be doing a speed comparison between the two). Only those calls to <code>klp_kmeans<\/code> where <code>use_gpu=True<\/code> will run on the GPU because of the way the variables are typed (e.g. <code>theano.tensor.dmatrix<\/code> vs. <code>theano.tensor.matrix<\/code>).<\/p>\n"}]}
{"QuestionId":31060970,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-06-25T21:18:06.437","AcceptedAnswerId":31068759.0,"OwnerUserId":3023426.0,"Title":"Customize operation\/function on Theano","Body":"<p>I am new to the Theano library, which is used for deep learning on GPU device. I have noticed that there are several build-in operations which can support gpu computation (I guess they are specially written in a way to support GPU):<\/p>\n\n<pre><\/pre>\n\n<p>1.What's the difference if I use python's build-in function sum() instead of T.sum(). Will sum() still work but maybe slower? <\/p>\n\n<ol start=\"2\">\n<li><p>Suppose sum() doesn't work for gpu computing, then if I need any operation\/function that workable on gpu, I need to implement it in such a way.  e.g. I want calculate sin(x) where x is a vector or matrix, and stored in GPU memory. Is there any hint to implement sin(x) that can operate on gpu device? (this might not be suitable or easy to answer)<\/p><\/li>\n<li><p>I have trouble understanding T.grad(). How T.grad can do the symbolic calculation for any given smooth symbolic function? I am very curious about it.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"31068759","CreationDate":"2015-06-26T08:47:10.033","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>In theano you have to use <code>T.sum(), T.neq(), T.argmax(),T.grad()<\/code> for symbolic computation with theano variables like <code>T.matrix<\/code>. You can't use built in <code>sum()<\/code> for example.\nIf you use theano you have to follow theano's own methods because theano uses a different form of computing to utilize the gpu's architecture.<\/p>\n\n<p>However if you want to use <code>sum()<\/code> you can do the computation with it, and then create a <code>theano.shared<\/code> variable where you can store the results in, this way you storing it in the gpu's memory at runtime.<\/p>\n\n<p>Regarding T.grad() perhaps you should ask the theano developers. :)\nHowever I think when theano is running, it can compute the function's gradient runtime, using the actual variables utilizing the gpu's computing capacity.\nI hope this can help.<\/p>\n\n<p><strong>Glad to have been of help! Feel free to accept my answer if you feel it was useful to you. :-)<\/strong><\/p>\n"}]}
{"QuestionId":31080866,"AnswerCount":1,"Tags":"<python><function><input><theano>","CreationDate":"2015-06-26T19:35:16.763","AcceptedAnswerId":31082580.0,"OwnerUserId":5009112.0,"Title":"Theano function that can take input arrays of different shapes in python","Body":"<p>In theano, I want to make a function that can take several different inputs, such as both matrices and vectors. <\/p>\n\n<p>Normally I would do something like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, then when I enter a vector instead of a matrix, for example:<\/p>\n\n<pre><\/pre>\n\n<p>Then I get an error of dimension mismatch: 'Wrong number of dimensions: expected 2, got 1 with shape (3,).'<\/p>\n\n<p>Is there any way to define a more general input symbol in theano that can take matrices but also different shaped arrays such as vectors or 3-dimensional arrays and still works?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"31082580","CreationDate":"2015-06-26T21:41:20.817","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The number of dimensions must be fixed at the time the Theano function is compiled. Part of the compilation process is to select operation variants that depend on the number of dimensions.<\/p>\n\n<p>You could always compile the function for a high-dimensional tensor and just stack your inputs such that they have the required shape.<\/p>\n\n<p>So<\/p>\n\n<pre><code>x = theano.tensor.tensor3()\ny = 3*x\nf = theano.function([x],y)\n<\/code><\/pre>\n\n<p>will accept and of these<\/p>\n\n<pre><code>f(numpy.array([[[1,2]]]))  # (1,1,3) vector wrapped as a tensor3\nf(numpy.array([[[1,2],[3,4]]]))  # (1,2,2) matrix wrapped as a tensor3\nf(numpy.array([[[1,2],[3,4]],[[5,6],[7,8]]]))  # (2,2,2) tensor3\n<\/code><\/pre>\n"}]}
{"QuestionId":31099233,"AnswerCount":1,"Tags":"<c++><deep-learning><caffe>","CreationDate":"2015-06-28T11:25:13.047","AcceptedAnswerId":31132209.0,"OwnerUserId":1544186.0,"Title":"Euclidean Loss Layer in Caffe","Body":"<p>I am currently trying to implement my own loss layer in caffe, and while attempting to do so, am using other layers as a reference. One thing that puzzles me, however, is the use of  in . I will be using the  as a reference. Here are my questions<\/p>\n\n<ul>\n<li><p>It is my understanding that  holds the error derivative from the next layer, but what if there is no other layer, how is it initialised? since it is used in  without performing any checks:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Again, in the , the derivative for the error with respect to the activations is calculated using the following code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>If my first assumption is correct, and  does indeed hold the error derivative for the layer above, why do we only use the first element i.e.  as opposed to multiplying by the whole vector i.e. ?<\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"31132209","CreationDate":"2015-06-30T07:32:27.687","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>For loss layers, there is no next layer, and so the top diff blob is technically undefined and unused - but Caffe is using this preallocated space to store unrelated data:  Caffe supports multiplying loss layers with a user-defined weight (loss_weight in the prototxt), this information (a single scalar floating point number) is stored in the first element of the diff array of the top blob. That's why you'll see in every loss layer, that they multiply by that amount to support that functionality. This is explained in <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/loss.html\">Caffe's tutorial about the loss layer<\/a>.<\/p>\n\n<p>This weight is usually used to add auxiliary losses to the network. You can read more about it in Google's <a href=\"http:\/\/arxiv.org\/abs\/1409.4842\">Going Deeper with Convoltions<\/a> or in <a href=\"http:\/\/arxiv.org\/abs\/1409.5185\">Deeply-Supervised Nets<\/a>.<\/p>\n"}]}
{"QuestionId":31105548,"AnswerCount":2,"Tags":"<python><visualization><gradient><deep-learning><caffe>","CreationDate":"2015-06-28T22:46:58.017","AcceptedAnswerId":32271804.0,"OwnerUserId":5013415.0,"Title":"Implementation of the paper \"Deep inside convolutional networks: Visualising image classification models and saliency maps\", Simonyan et al","Body":"<p>In visualization of gradient data in Convolutional Neural Networks, employing Caffe framework, having already visualized gradient data with respect to all classes, it is interesting to take gradient regarding a specific class. in deploy.prototxt file in \"bvlc_reference_caffenet\" model, I have set:<\/p>\n\n<pre><\/pre>\n\n<p>and has commented the last part:<\/p>\n\n<pre><\/pre>\n\n<p>,which is before:<\/p>\n\n<pre><\/pre>\n\n<p>, and added instead of it:<\/p>\n\n<pre><\/pre>\n\n<p>, in the python code by calling:<\/p>\n\n<pre><\/pre>\n\n<p>, we forward towards the last layer and afterwards, by calling:<\/p>\n\n<pre><\/pre>\n\n<p>,got the visualization of gradient. Firstly, I'd like to ask this is called saliency map and if I want to do backward with respect to a specific class e.g. 281 is for a cat. what shall I do?<\/p>\n\n<p>thanks in advance for your guidance.<\/p>\n\n<p>P.S. benefited from the code by Yangqing for its notebook in filter visualization.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32271804","CreationDate":"2015-08-28T13:06:36.303","ParentId":null,"OwnerUserId":"5013415","Title":null,"Body":"<p>also for full visualization you can refer to my github, which is more complete and visualize the saliency map as well as the visualization of class models and the gradient visualization in backpropagation.<\/p>\n\n<p><a href=\"https:\/\/github.com\/smajida\/Deep_Inside_Convolutional_Networks\" rel=\"nofollow\">https:\/\/github.com\/smajida\/Deep_Inside_Convolutional_Networks<\/a><\/p>\n"},{"AnswerId":"31374054","CreationDate":"2015-07-13T00:49:17.620","ParentId":null,"OwnerUserId":"5013415","Title":null,"Body":"<p>using the following code it can be done:<\/p>\n\n<pre><code>label_index = 281  # Index for cat class\ncaffe_data = np.random.random((1,3,227,227))\ncaffeLabel = np.zeros((1,1000,1,1))\ncaffeLabel[0,label_index,0,0] = 1;\n\nbw = net.backward(**{net.outputs[0]: caffeLabel})\n<\/code><\/pre>\n"}]}
{"QuestionId":31106263,"AnswerCount":1,"Tags":"<neural-network><convolution><theano><conv-neural-network><keras>","CreationDate":"2015-06-29T00:36:25.100","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Keras \/ Theano: How to add Convolution2D Layers?","Body":"<p>Im having issues on how to understand how Convolution Layers are added.\nIm trying to add Convolution Layers but i get this error : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<hr>\n\n<p>Im trying to understand what is nb_filter, stack_size, nb_row, nb_col are on a convolutional layer.<\/p>\n\n<p>My Objective is to copy the VGG Model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>--\nIm currently using Theano and keras.<\/p>\n\n<p>Please, any tip is appreciated.<\/p>\n","answers":[{"AnswerId":"41734801","CreationDate":"2017-01-19T06:00:26.427","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>You need to correct the output shape for the convolutional layer. Output of a CNN layer depends on many factors such as input size, number of kernels, stride and padding. Generally for an input of size BxCxW1xH1, the output would be BxFxW2xH2 where B is the batch size, C is the input channels, F is the number of output features, W1xH1 is the input size and you can compute the value of W2 and H2 using W1, H1, stride and padding. It is illustrated very well in this tutorial from Stanford: <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/#comp\" rel=\"nofollow noreferrer\">http:\/\/cs231n.github.io\/convolutional-networks\/#comp<\/a><\/p>\n\n<p>Hope it helps!<\/p>\n"}]}
{"QuestionId":31107776,"AnswerCount":3,"Tags":"<python-2.7><graphviz><anaconda><theano><deep-learning>","CreationDate":"2015-06-29T04:26:22.833","AcceptedAnswerId":null,"OwnerUserId":5059640.0,"Title":"Theano Import error","Body":"<p>I am trying to install Theano on CPU machine (running intel HD graphics, without NVIDIA). I get the following import error while testing in python.<\/p>\n\n<pre><\/pre>\n\n<p>I have the g++ installed though.<\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"41894442","CreationDate":"2017-01-27T12:53:35.603","ParentId":null,"OwnerUserId":"7173909","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>ten,<\/p>\n\n<pre><code>from theano import *\n<\/code><\/pre>\n\n<p>It is Worked for me<\/p>\n"},{"AnswerId":"41732135","CreationDate":"2017-01-19T01:08:23.297","ParentId":null,"OwnerUserId":"6936471","Title":null,"Body":"<p>If you use Pycharm on windows, following the following steps:<\/p>\n\n<ul>\n<li>install anaconda<\/li>\n<li>Change your interpreter to anaconda python<\/li>\n<li>install theano<\/li>\n<li>install mingw<\/li>\n<li>install libpython<\/li>\n<\/ul>\n"},{"AnswerId":"31109547","CreationDate":"2015-06-29T06:58:35.620","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\">As described in the documentation<\/a>, make sure you have done this when using Anaconda.<\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n"}]}
{"QuestionId":31108507,"AnswerCount":2,"Tags":"<python><c++><windows><cuda><theano>","CreationDate":"2015-06-29T05:38:37.800","AcceptedAnswerId":31128511.0,"OwnerUserId":891010.0,"Title":"Installing Theano on windows for gpu - suspected nvcc version issue","Body":"<p>I have been following <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">instructions<\/a> to set up Theano to use a GPU on Windows.<\/p>\n\n<p>The issue is I cannot follow these instructions exactly because I have a new graphics card, the GEForce GTX 980M and it only works with cuda 7.0. (The instructions suggest cuda 5.5). Everything works fine except when it gets time to run the GPU then I get an error:<\/p>\n\n<pre><\/pre>\n\n<p>The version of nvcc I have installed does not have version -2008 (that looks to have been deprecated by the lastest cuda 7.0 version, but 2010 is allowed). \nWhat is the best way to fix it? Should I hard code it in Theano into the file cuda\\nvcc_compiler.py? I tried that and it seems to try to use 2008 version anyway.\nIs there a later version of Theano that would use the later nvcc version?<\/p>\n","answers":[{"AnswerId":"31128511","CreationDate":"2015-06-30T02:19:24.140","ParentId":null,"OwnerUserId":"5063541","Title":null,"Body":"<p>I ran into a similar problem when trying to install Theano on Win 8.1 64bit with CUDA 7.0., using a GTX 750Ti graphics card. I was able to get it working by following these <a href=\"https:\/\/my6266blog.wordpress.com\/2015\/01\/21\/installing-theano-pylearn2-and-even-gpu-on-windows\/\">instructions<\/a>.  <\/p>\n"},{"AnswerId":"35959271","CreationDate":"2016-03-12T15:03:07.573","ParentId":null,"OwnerUserId":"5231110","Title":null,"Body":"<p>For me it started working when I replaced<\/p>\n\n<pre><code>[nvcc]\nflags = --use-local-env  --cl-version=2008\n<\/code><\/pre>\n\n<p>by<\/p>\n\n<pre><code>[nvcc]\ncompiler_bindir=C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin\n<\/code><\/pre>\n\n<p>in the .theanorc file.<\/p>\n"}]}
{"QuestionId":31117261,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2015-06-29T13:40:53.537","AcceptedAnswerId":null,"OwnerUserId":3891733.0,"Title":"DeprecationWarning: Module scipy.linalg.blas.fblas is deprecated, use scipy.linalg.blas instead","Body":"<p>I've just now installed theano on my machine but when i try to use it - i.e \nI get this 'DeprecationWarning':<\/p>\n\n<pre><\/pre>\n\n<p>When I comment out the above line the error doesn't show. I've tried updating scipy, numpy, theano.. but nothing works. <\/p>\n\n<p>Any ideas on what's causing this warning and how to get rid of it?<\/p>\n","answers":[{"AnswerId":"33343545","CreationDate":"2015-10-26T10:33:29.450","ParentId":null,"OwnerUserId":"2251058","Title":null,"Body":"<p>You could avoid a Deprecation Warning. If you don't want it to show up you could add the following to your code.<\/p>\n\n<pre><code>import warnings\nwarnings.filterwarnings(\"ignore\", category=DeprecationWarning)\n<\/code><\/pre>\n"}]}
{"QuestionId":31126636,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-06-29T22:29:21.573","AcceptedAnswerId":31129029.0,"OwnerUserId":3113501.0,"Title":"Torch tensors swapping dimensions","Body":"<p>I came across these two lines (back-to-back) of code in a torch project:<\/p>\n\n<pre><\/pre>\n\n<p>What do these two lines do? I assumed they did some sort of swapping.<\/p>\n","answers":[{"AnswerId":"31129029","CreationDate":"2015-06-30T03:27:14.673","ParentId":null,"OwnerUserId":"2726734","Title":null,"Body":"<p>This is covered in indexing in the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"noreferrer\">Torch Tensor Documentation<\/a><\/p>\n\n<p>Indexing using the empty table <code>{}<\/code> is shorthand for all indices in that dimension. Below is a demo which uses <code>{}<\/code> to copy an entire row from one matrix to another:<\/p>\n\n<pre><code>&gt; a = torch.Tensor(3, 3):fill(0)\n     0 0 0\n     0 0 0\n     0 0 0\n\n&gt; b = torch.Tensor(3, 3)\n&gt; for i=1,3 do for j=1,3 do b[i][j] = (i - 1) * 3 + j end end\n&gt; b\n     1 2 3\n     4 5 6\n     7 8 9\n\n&gt; a[{1, {}}] = b[{3, {}}]\n&gt; a\n    7 8 9\n    0 0 0\n    0 0 0\n<\/code><\/pre>\n\n<p>This assignment is equivalent to: <code>a[1] = b[3]<\/code>.<\/p>\n\n<p>Your example is similar:<\/p>\n\n<pre><code> im4[{1,{},{}}] = im3[{3,{},{}}]\n im4[{3,{},{}}] = im3[{1,{},{}}]\n<\/code><\/pre>\n\n<p>which is more clearly stated as:<\/p>\n\n<pre><code> im4[1] = im3[3]\n im4[3] = im3[1]\n<\/code><\/pre>\n\n<p>The first line assigns the values from <code>im3<\/code>'s third row (a 2D sub-matrix) to <code>im4<\/code>'s first row and the second line assigns the first row of <code>im3<\/code> to the third row of <code>im4<\/code>.<\/p>\n\n<p>Note that this is not a swap, as <code>im3<\/code> is never written and <code>im4<\/code> is never read from.<\/p>\n"}]}
{"QuestionId":31126785,"AnswerCount":2,"Tags":"<python><neural-network><gpu><theano>","CreationDate":"2015-06-29T22:42:09.703","AcceptedAnswerId":31166170.0,"OwnerUserId":2938232.0,"Title":"Theano gradient calculation creates float64","Body":"<p>I have some standard NN code on Theano with two separate compiled functions.  One that calculates the cost and one that calculates the cost with AdaGrad updates.<\/p>\n\n<p>For GPU speed, I'm trying to keep everything .  Problem is that I'm getting a warning that the gradient calculation is creating a .  In particular, for the following line of code.<\/p>\n\n<pre><\/pre>\n\n<p>If I comment out the gradient calculation and replace the second line with a placeholder, everything is fine.  Obviously this is a junk update but it helps pinpoint the problem.<\/p>\n\n<pre><\/pre>\n\n<p>For reference, this is the loss function:<\/p>\n\n<pre><\/pre>\n\n<p>The cost only compiled function is:<\/p>\n\n<pre><\/pre>\n\n<p>whereas the cost + update function is:<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I pinned down the problem to a LSTM layer.  Here is minimal code<\/p>\n\n<pre><\/pre>\n\n<p>The problem occurs when I call  to get gradients (here it is a junk update for example).  I've found that commenting out  or replacing the  with something that repeats  to the required shape fixes things. <\/p>\n\n<p><strong>UPDATE 2<\/strong><\/p>\n\n<p>Debugging reveals a mess of  in the  function.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31146833","CreationDate":"2015-06-30T19:28:56.767","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You could use <code>theano.printing.debugprint<\/code> with <code>print_type=True<\/code> to see what the type of the various components are.<\/p>\n\n<p>For example,<\/p>\n\n<pre><code>class TestLSTMLayer(object):\n    def __init__(self, inputs, outputSize, dropout=0.9, inputSize=None, adagradInit=1, forgetGateBias=3, srng=None):\n        self.h0 = theano.shared(np.random.randn(outputSize).astype(floatX))\n\n        self.params = [self.h0]\n\n        def _recurrence(hBelow):\n            print 'hBelow', theano.printing.debugprint(hBelow, print_type=True)    \n            return hBelow\n\n        print 'h0', theano.printing.debugprint(self.h0, print_type=True)    \n        hOutputs, _ = theano.scan(\n            fn=_recurrence,\n            outputs_info=self.h0,\n            n_steps=inputs.shape[0]\n        )\n        self.hOutputs = hOutputs\n\n    def getUpdates(self):\n        print 'sum', theano.printing.debugprint(TT.sum(self.hOutputs), print_type=True)    \n        gradients = TT.grad(TT.sum(self.hOutputs), self.params)\n        paramUpdates = [(self.params[0], self.params[0])]\n        return paramUpdates\n<\/code><\/pre>\n"},{"AnswerId":"31166170","CreationDate":"2015-07-01T15:53:27.097","ParentId":null,"OwnerUserId":"2938232","Title":null,"Body":"<p>I solved this by updating to the bleeding edge (latest version on GitHub) version. <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>Somewhat odd, but a solution nonetheless.<\/p>\n"}]}
{"QuestionId":31143452,"AnswerCount":2,"Tags":"<python><class><theano>","CreationDate":"2015-06-30T16:21:02.827","AcceptedAnswerId":31146656.0,"OwnerUserId":4480756.0,"Title":"Theano Shared Variables on Python","Body":"<p>I am now learning the Theano library, and I am just feeling confused about Theano shared variables. By reading the tutorial, I think I didn't understand its detailed meaning. The following is the definition of the Theano shared variables from the tutorial:<\/p>\n\n<p>\"Variable with Storage that is shared between functions that it appears in. These variables are meant to be created by registered shared constructors.\"<\/p>\n\n<p>Also, I am wondering if the Theano shared variables can be a python class data member. For example:<\/p>\n\n<pre><\/pre>\n\n<p>Can \"data\" be or initialized as a Theano Shared variable? I really appreciate if anyone could help me.<\/p>\n","answers":[{"AnswerId":"31146656","CreationDate":"2015-06-30T19:20:11.807","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano shared variables behave more like ordinary Python variables. They have an explicit value that is persistent. In contrast, symbolic variables are not given an explicit value until one is assigned on the execution of a compiled Theano function.<\/p>\n\n<p>Symbolic variables can be thought of as representing state for the duration of a single execution. Shared variables on the other hand represent state that remains in memory for the lifetime of the Python reference (often similar to the lifetime of the program).<\/p>\n\n<p>Shared variables are usually used to store\/represent neural network weights because we want these values to remain around across many executions of a Theano training or testing function. Often, the purpose of a Theano training function is to update the weights stored in a shared variable. And a testing function needs the current weights to perform the network's forward pass.<\/p>\n\n<p>As far as Python is concerned Theano variables (shared or symbolic) are just objects -- instances of classes defined within the Theano library. So, yes, references to shared variables can be stored in your own classes, just like any other Python object.<\/p>\n"},{"AnswerId":"33017372","CreationDate":"2015-10-08T13:33:14.590","ParentId":null,"OwnerUserId":"3212998","Title":null,"Body":"<p>Shared variable helps in simplifying the operations over a pre-defined variable. An example to @danien-renshaw 's answer, suppose we want to add two matrix, let's say a and b, where the value of b matrix will remain constant throughout the lifetime of the program, we can have the b matrix as the shared variable and do the required operation.<\/p>\n\n<p>Code without using shared variable:<\/p>\n\n<pre><code>a = theano.tensor.matrix('a')\nb = theano.tensor.matrix('b')\nc = a + b\nf = theano.function(inputs = [a, b], outputs = [c])\noutput = f([[1, 2, 3, 4]], [[5, 5, 6, 7]])\n<\/code><\/pre>\n\n<p>Code using shared variable :<\/p>\n\n<pre><code>a = theano.tensor.matrix('a')\nb = theano.tensor.shared( numpy.array([[5, 6, 7, 8]]))\nc = a + b\nf = theano.function(inputs = [a], outputs = [c])\noutput = f([[1, 2, 3, 4]])\n<\/code><\/pre>\n"}]}
{"QuestionId":31144564,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-06-30T17:25:49.497","AcceptedAnswerId":31149198.0,"OwnerUserId":3113501.0,"Title":"What are the differences between doFile and require in Lua","Body":"<p>What are the differences between doFile and require in Lua, especially in Torch? When do you call one but not the other? When will one work but the other won't? (I'm using Lua 5.1, torch7).<\/p>\n","answers":[{"AnswerId":"31149198","CreationDate":"2015-06-30T21:54:03.813","ParentId":null,"OwnerUserId":"646619","Title":null,"Body":"<p><code>dofile<\/code> loads and executes a file right then and there.<\/p>\n\n<p><code>require<\/code> is more complicated; it keeps a table of modules that have already been loaded and their return results, to ensure that the same code isn't loaded twice. It also keeps a list of module loaders that handle loading a module, one of which that can load from <code>dll<\/code>\/<code>so<\/code> files.<\/p>\n\n<p>You probably want <code>require<\/code>, as if you're just loading functions, you don't want to duplicate them.<\/p>\n"}]}
{"QuestionId":31147894,"AnswerCount":1,"Tags":"<emacs><theano><jedi>","CreationDate":"2015-06-30T20:26:36.667","AcceptedAnswerId":31148056.0,"OwnerUserId":1776544.0,"Title":"jedi:complete produces a deferred error when using the theano library","Body":"<p>In emacs, I am trying to get  to work with . To do so, I have the following minimal bit of code.<\/p>\n\n<pre><\/pre>\n\n<p>When I place my cursor at the  sign, and run , I am met with the following error, and no autocompletion is offered.<\/p>\n\n<pre><\/pre>\n\n<p>I wonder if this is an incompatability with a source file in  and . But I am not sure, and I do not know what to do to further resolve the issue.<\/p>\n\n<p>I get a similar error when I try to use .<\/p>\n\n<p>I have installed all of my packages through the  command, and they are updated with the latest versions.<\/p>\n","answers":[{"AnswerId":"31148056","CreationDate":"2015-06-30T20:36:16.410","ParentId":null,"OwnerUserId":"1776544","Title":null,"Body":"<p>I resolved this by looking through <code>m-x packages-list-packages<\/code> and seeing that there was (for some reason) an old deprecated version of jedi installed alongside another version of jedi. I deleted all the deprecated installations I had, and the error has gone away, but <code>jedi<\/code> doesnt seem to be able to autocomplete the above code still. It now just says <code>No completion found<\/code>.<\/p>\n"}]}
{"QuestionId":31148167,"AnswerCount":0,"Tags":"<python><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2015-06-30T20:44:24.920","AcceptedAnswerId":null,"OwnerUserId":4561745.0,"Title":"Caffe: Can't seem to learn y = x^2 function","Body":"<p>I was trying to train a neural network to learn the function y = x^2 in the deep learning framework Caffe. Here is my code:<\/p>\n\n<p>Data generation code:<\/p>\n\n<pre><\/pre>\n\n<p>Solver file:<\/p>\n\n<pre><\/pre>\n\n<p>Caffe Model:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting an error of the order of 10^8, which is unbelievable. The net is supposed to take a single input and produce a single output. The inputs are integers in [0,255] range and the outputs are supposed to be the squares of the respective inputs. Any ideas why such a huge error is obtained?<\/p>\n","answers":[]}
{"QuestionId":31162021,"AnswerCount":1,"Tags":"<python><environment-variables><spyder><theano>","CreationDate":"2015-07-01T12:58:24.673","AcceptedAnswerId":31162799.0,"OwnerUserId":492372.0,"Title":"Why do the environment variables set in command prompt have no effect when I start Spyder","Body":"<p>I am using the Spyder Anaconda IDE for Python. I am writing a code in the Spyder IDE that requires few environment variables to be set ($CPATH, $LIBRARY_PATH and $LD_LIBRARY_PATH) for the Theano library. <\/p>\n\n<p>I am starting Spyder using the command<\/p>\n\n<pre><\/pre>\n\n<p>and it starts fine. Even though I set the environment variables in my <\/p>\n\n<pre><\/pre>\n\n<p>file, the code still fails to accept the path and if I try printing<\/p>\n\n<pre><\/pre>\n\n<p>it raises a KeyError.<\/p>\n\n<p>I tried all the above with a normal user but still it fails. How can I get Spyder IDE to be able to view files in the above paths and where can I set them inside Spyder?<\/p>\n","answers":[{"AnswerId":"31162799","CreationDate":"2015-07-01T13:29:05.577","ParentId":null,"OwnerUserId":"1596068","Title":null,"Body":"<p>You need to tell the <code>sudoers<\/code> file which Environmental Variables to keep when using the <code>sudo<\/code> command.<\/p>\n\n<p>To edit the sudoers file run.<\/p>\n\n<pre><code>sudo visudo\n<\/code><\/pre>\n\n<p>Then add the following line to the end of it.<\/p>\n\n<pre><code>Defaults env_keep = \"LD_LIBRARY_PATH CPATH LIBRARY_PATH\"\n<\/code><\/pre>\n\n<p>Then <code>export<\/code> your variable.<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\"\/path\/to\/library\"\n<\/code><\/pre>\n\n<p>Now you should be able to run it.<\/p>\n\n<p>More info can be found here <a href=\"https:\/\/stackoverflow.com\/questions\/8633461\/how-to-keep-environment-variables-when-using-sudo\">How to keep Environment Variables when Using SUDO<\/a><\/p>\n"}]}
{"QuestionId":31169441,"AnswerCount":3,"Tags":"<amazon-web-services><amazon-ec2><theano><deep-learning><caffe>","CreationDate":"2015-07-01T18:55:23.860","AcceptedAnswerId":31177427.0,"OwnerUserId":5067505.0,"Title":"Library for deep learning on Amazon EC2 with CPU and GPU support for convolutional neural network","Body":"<p>I want to train a CNN on a bunch of images. I want to run it on Amazon EC2 CPU or GPU clusters. For running deep learning on a cluster, I figured that some of the options are:<\/p>\n\n<ol>\n<li>h2o (with Spark)<\/li>\n<li>Caffee<\/li>\n<li>Theano<\/li>\n<\/ol>\n\n<p>I am not sure which of these options suit my needs. I read through <a href=\"http:\/\/h2o-release.s3.amazonaws.com\/h2o\/rel-shannon\/25\/docs-website\/h2o-docs\/index.html#Data%20Science%20Algorithms-Deep%20Learning-Introduction\" rel=\"nofollow\">h2o documentation on deep learning<\/a>, they do not seem to support CNNs. Any ideas on how I should proceed? <\/p>\n\n<p>Another side question: \nHow do I upload my images to the cluster for training the CNN? I am fairly new to cluster computing.<\/p>\n","answers":[{"AnswerId":"53548765","CreationDate":"2018-11-29T22:50:43.797","ParentId":null,"OwnerUserId":"2136741","Title":null,"Body":"<p>AWS provides the Deep Learning AMI with various Deep Learning Frameworks installed into it, which covers your use case since it has Theano as well as Caffe.\n Link to Deep Learning AMI <a href=\"https:\/\/aws.amazon.com\/machine-learning\/amis\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/machine-learning\/amis\/<\/a>.<\/p>\n\n<blockquote>\n  <p>How do I upload my images to the cluster for training the CNN? I am\n  fairly new to cluster computing?<\/p>\n<\/blockquote>\n\n<p>There are many AWS storage services which gives you way to store your training data (images) which will be accessible to your cluster. Few of them are<\/p>\n\n<ol>\n<li>S3<\/li>\n<li>EBS <\/li>\n<li><p>EFS <\/p>\n\n<p>Explore them and see what works best for you. <\/p><\/li>\n<\/ol>\n"},{"AnswerId":"31177427","CreationDate":"2015-07-02T06:52:11.320","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>If you follow the instructions here <a href=\"https:\/\/github.com\/deeplearningparis\/dl-machine\" rel=\"nofollow\">https:\/\/github.com\/deeplearningparis\/dl-machine<\/a> then you can set up an AMI image with Theano and Torch. There is also a PR on the config to have caffe by default as well (if you need it, just checkout the branch and run the install script as soon as the instance is up).<\/p>\n"},{"AnswerId":"41111751","CreationDate":"2016-12-13T00:10:26.380","ParentId":null,"OwnerUserId":"3867406","Title":null,"Body":"<p>Just came accross your question. In this <strong><a href=\"http:\/\/vict0rsch.github.io\/2016\/12\/03\/aws_gpu\/\" rel=\"nofollow noreferrer\">tutorial<\/a><\/strong> you'll also find how to set up an Amazon instance with a GPU to run Deep Learning frameworks.<\/p>\n\n<p>The AMI (~computer model) is pre-configured with:<\/p>\n\n<blockquote>\n  <ul>\n  <li>Ubuntu Server 16.04 as OS<\/li>\n  <li>Anaconda 4.2.0 (scientific Python distribution)<\/li>\n  <li>Python 3.5<\/li>\n  <li>Cuda 8.0 (\u201cparallel computing platform and programming model\u201d, used to send code to the GPU)<\/li>\n  <li>cuDNN 5.1 (Cuda\u2019s library for Deep Learning used by Tensorflow and Theano)<\/li>\n  <li><strong>Tensorflow 0.12<\/strong> for Python 3.5 and GPU-enabled<\/li>\n  <li><strong>Keras 1.1.2<\/strong> (use with Tensorflow backend)<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>I believe you can use this set-up with <a href=\"https:\/\/aws.amazon.com\/fr\/ec2\/Elastic-GPUs\/\" rel=\"nofollow noreferrer\">elastic GPUs<\/a> to scale the system according to your needs or use a <a href=\"https:\/\/aws.amazon.com\/en\/ec2\/instance-types\/p2\/\" rel=\"nofollow noreferrer\">P2 instance<\/a><\/p>\n\n<p>Anyway you can follow the tutorial and use another AMI like <a href=\"https:\/\/aws.amazon.com\/fr\/amazon-ai\/amis\/\" rel=\"nofollow noreferrer\">Amazon's Deep Learning AMI<\/a><\/p>\n"}]}
{"QuestionId":31171277,"AnswerCount":1,"Tags":"<python><numpy><multiprocessing><shared-memory><caffe>","CreationDate":"2015-07-01T20:42:18.650","AcceptedAnswerId":null,"OwnerUserId":1322301.0,"Title":"Sharing contiguous numpy arrays between processes in python","Body":"<p>While I have found numerous answers to questions similar to mine, I don't believe it has been directly addressed here--and I have several additional questions. The motivation for sharing contiguous numpy arrays is as follows:<\/p>\n\n<ul>\n<li>I'm using a convolutional neural network run on Caffe to perform a regression on images to a series of continuous-value labels. <\/li>\n<li>The images require specific preprocessing and data augmentation.<\/li>\n<li>The constraints of (1) the continuous nature of the labels (they're floats) and (2) the data augmentation means that I'm preprocessing the data in python and then serving it up as contiguous numpy arrays using the in-memory data layer in Caffe.<\/li>\n<li>Loading the training data into memory is comparatively slow. I'd like to parallelize it such that:<\/li>\n<\/ul>\n\n<p>(1) The python I'm writing creates a \"data handler\" class which instantiates two contiguous numpy arrays. \n(2) A worker process alternates between those numpy arrays, loading the data from the disk, performing preprocessing, and inserting the data into the numpy array. \n(3) Meanwhile, the python Caffe wrappers send data from the <em>other<\/em> array to the GPU to be run through the net. <\/p>\n\n<p>I have a few questions:<\/p>\n\n<ol>\n<li><p>Is it possible to allocate memory in a contiguous numpy array then wrap it in a shared memory object (I'm not sure if 'object' is the correct term here) using something like the Array class from python's multiprocessing? <\/p><\/li>\n<li><p>Numpy arrays have a .ctypes attribute, I presume this is useful for the instantiation of shared memory arrays from Array(), but can't seem to determine precisely how to use them. <\/p><\/li>\n<li><p>If the shared memory is instantiated <em>without<\/em> the numpy array, does it remain contiguous? If not, is there a way to ensure it does remain contiguous? <\/p><\/li>\n<\/ol>\n\n<p>Is it possible to do something like:<\/p>\n\n<pre><\/pre>\n\n<p>Then instantiate the worker with<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n\n<p>Edit: I'm aware there are a number of libraries that have similar functions in varying states of maintenance. I would prefer to restrict this to pure python and numpy, but if that's not possible I would of course be willing to use one. <\/p>\n","answers":[{"AnswerId":"43680177","CreationDate":"2017-04-28T12:22:52.107","ParentId":null,"OwnerUserId":"794539","Title":null,"Body":"<h2>Wrap numpy's <code>ndarray<\/code> around multiprocessing's <code>RawArray()<\/code><\/h2>\n\n<p>There are multiple ways to share <em>numpy<\/em> arrays in memory across processes. Let's have a look at how you can do it using the <em>multiprocessing<\/em> module.<\/p>\n\n<p>The first important observation is that <em>numpy<\/em> provides the <strong><code>np.frombuffer()<\/code> function to wrap an <em>ndarray<\/em> interface around a preexisting object<\/strong> that supports the buffer protocol (such as <code>bytes()<\/code>, <code>bytearray()<\/code>, <code>array()<\/code> and so on). This creates read-only arrays from read-only objects and writable arrays from writable objects.<\/p>\n\n<p>We can combine that with the <strong>shared memory <code>RawArray()<\/code><\/strong> that <em>multiprocessing<\/em> provides. Note that <code>Array()<\/code> doesn't work for that purpose, as it is a proxy object with a lock and doesn't directly expose the buffer interface. Of course that means that we need to provide for proper synchronization of our <em>numpified RawArrays<\/em> ourselves.<\/p>\n\n<p>There is one complicating issue regarding <em>ndarray<\/em>-wrapped <em>RawArrays<\/em>: When <em>multiprocessing<\/em> sends such an array between processes - and indeed it will need to send our arrays, once created, to both workers - it pickles and then unpickles them. Unfortunately, that results in it creating copies of the <em>ndarrays<\/em> instead of sharing them in memory.<\/p>\n\n<p>The solution, while a bit ugly, is to <strong>keep the <em>RawArrays<\/em> as is<\/strong> until they are transferred to the workers and <strong>only wrap them in <em>ndarrays<\/em> once each worker process has started<\/strong>.<\/p>\n\n<p>Furthermore, it would have been preferable to communicate arrays, be it a plain <em>RawArray<\/em> or an <em>ndarray<\/em>-wrapped one, directly via a <code>multiprocessing.Queue<\/code>, but that doesn't work, either. A <em>RawArray<\/em> cannot be put inside such a <em>Queue<\/em> and an <em>ndarray<\/em>-wrapped one would have been pickled and unpickled, so in effect copied.<\/p>\n\n<p>The workaround is to send a list of all pre-allocated arrays to the worker processes and <strong>communicate indices into that list over the <em>Queues<\/em><\/strong>. It's very much like passing around tokens (the indices) and whoever holds the token is allowed to operate on the associated array.<\/p>\n\n<p>The structure of the main program could look like this:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>#!\/usr\/bin\/env python3\n# -*- coding: utf-8 -*-\n\nimport numpy as np\nimport queue\n\nfrom multiprocessing import freeze_support, set_start_method\nfrom multiprocessing import Event, Process, Queue\nfrom multiprocessing.sharedctypes import RawArray\n\n\ndef create_shared_arrays(size, dtype=np.int32, num=2):\n    dtype = np.dtype(dtype)\n    if dtype.isbuiltin and dtype.char in 'bBhHiIlLfd':\n        typecode = dtype.char\n    else:\n        typecode, size = 'B', size * dtype.itemsize\n\n    return [RawArray(typecode, size) for _ in range(num)]\n\n\ndef main():\n    my_dtype = np.float32\n\n    # 125000000 (size) * 4 (dtype) * 2 (num) ~= 1 GB memory usage\n    arrays = create_shared_arrays(125000000, dtype=my_dtype)\n    q_free = Queue()\n    q_used = Queue()\n    bail = Event()\n\n    for arr_id in range(len(arrays)):\n        q_free.put(arr_id)  # pre-fill free queue with allocated array indices\n\n    pr1 = MyDataLoader(arrays, q_free, q_used, bail,\n                       dtype=my_dtype, step=1024)\n    pr2 = MyDataProcessor(arrays, q_free, q_used, bail,\n                          dtype=my_dtype, step=1024)\n\n    pr1.start()\n    pr2.start()\n\n    pr2.join()\n    print(\"\\n{} joined.\".format(pr2.name))\n\n    pr1.join()\n    print(\"{} joined.\".format(pr1.name))\n\n\nif __name__ == '__main__':\n    freeze_support()\n\n    # On Windows, only \"spawn\" is available.\n    # Also, this tests proper sharing of the arrays without \"cheating\".\n    set_start_method('spawn')\n    main()\n<\/code><\/pre>\n\n<p>This prepares a list of two arrays, two <em>Queues<\/em> - a \"free\" queue where <em>MyDataProcessor<\/em> puts array indices it is done with and <em>MyDataLoader<\/em> fetches them from as well as a \"used\" queue where <em>MyDataLoader<\/em> puts indices of readily filled arrays and <em>MyDataProcessor<\/em> fetches them from - and a <code>multiprocessing.Event<\/code> to start a concerted bail out of all workers. We could do away with the latter for now, as we have only one producer and one consumer of arrays, but it doesn't hurt being prepared for more workers.<\/p>\n\n<p>Then we pre-fill the \"empty\" <em>Queue<\/em> with all indices of our <em>RawArrays<\/em> in the list and instantiate one of each type of workers, passing them the necessary communication objects. We start both of them and just wait for them to <code>join()<\/code>.<\/p>\n\n<p>Here's how <em>MyDataProcessor<\/em> could look like, which consumes array indices from the \"used\" <em>Queue<\/em> and sends the data off to some external black box (<code>debugio.output<\/code> in the example):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyDataProcessor(Process):\n    def __init__(self, arrays, q_free, q_used, bail, dtype=np.int32, step=1):\n        super().__init__()\n        self.arrays = arrays\n        self.q_free = q_free\n        self.q_used = q_used\n        self.bail = bail\n        self.dtype = dtype\n        self.step = step\n\n    def run(self):\n        # wrap RawArrays inside ndarrays\n        arrays = [np.frombuffer(arr, dtype=self.dtype) for arr in self.arrays]\n\n        from debugio import output as writer\n\n        while True:\n            arr_id = self.q_used.get()\n            if arr_id is None:\n                break\n\n            arr = arrays[arr_id]\n\n            print('(', end='', flush=True)          # just visualizing activity\n            for j in range(0, len(arr), self.step):\n                writer.write(str(arr[j]) + '\\n')\n            print(')', end='', flush=True)          # just visualizing activity\n\n            self.q_free.put(arr_id)\n\n            writer.flush()\n\n        self.bail.set()                     # tell loaders to bail out ASAP\n        self.q_free.put(None, timeout=1)    # wake up loader blocking on get()\n\n        try:\n            while True:\n                self.q_used.get_nowait()    # wake up loader blocking on put()\n        except queue.Empty:\n            pass\n<\/code><\/pre>\n\n<p>The first it does is wrap the received <em>RawArrays<\/em> in <em>ndarrays<\/em> using 'np.frombuffer()' and keep the new list, so they are usable as <em>numpy<\/em> arrays during the process' runtime and it doesn't have to wrap them over and over again.<\/p>\n\n<p>Note also that <em>MyDataProcessor<\/em> only ever writes to the <code>self.bail<\/code> <em>Event<\/em>, it never checks it. Instead, if it needs to be told to quit, it will find a <code>None<\/code> mark on the queue instead of an array index. This is done for when a <em>MyDataLoader<\/em> has no more data available and starts the tear down procedure, <em>MyDataProcessor<\/em> can still process all valid arrays that are in the queue without prematurely exiting.<\/p>\n\n<p>This is how <em>MyDataLoader<\/em> could look like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyDataLoader(Process):\n    def __init__(self, arrays, q_free, q_used, bail, dtype=np.int32, step=1):\n        super().__init__()\n        self.arrays = arrays\n        self.q_free = q_free\n        self.q_used = q_used\n        self.bail = bail\n        self.dtype = dtype\n        self.step = step\n\n    def run(self):\n        # wrap RawArrays inside ndarrays\n        arrays = [np.frombuffer(arr, dtype=self.dtype) for arr in self.arrays]\n\n        from debugio import input as reader\n\n        for _ in range(10):  # for testing we end after a set amount of passes\n            if self.bail.is_set():\n                # we were asked to bail out while waiting on put()\n                return\n\n            arr_id = self.q_free.get()\n            if arr_id is None:\n                # we were asked to bail out while waiting on get()\n                self.q_free.put(None, timeout=1)  # put it back for next loader\n                return\n\n            if self.bail.is_set():\n                # we were asked to bail out while we got a normal array\n                return\n\n            arr = arrays[arr_id]\n\n            eof = False\n            print('&lt;', end='', flush=True)          # just visualizing activity\n            for j in range(0, len(arr), self.step):\n                line = reader.readline()\n                if not line:\n                    eof = True\n                    break\n\n                arr[j] = np.fromstring(line, dtype=self.dtype, sep='\\n')\n\n            if eof:\n                print('EOF&gt;', end='', flush=True)   # just visualizing activity\n                break\n\n            print('&gt;', end='', flush=True)          # just visualizing activity\n\n            if self.bail.is_set():\n                # we were asked to bail out while we filled the array\n                return\n\n            self.q_used.put(arr_id)     # tell processor an array is filled\n\n        if not self.bail.is_set():\n            self.bail.set()             # tell other loaders to bail out ASAP\n            # mark end of data for processor as we are the first to bail out\n            self.q_used.put(None)\n<\/code><\/pre>\n\n<p>It is very similar in structure to the other worker. The reason it is bloated up a bit is that it checks the <code>self.bail<\/code> <em>Event<\/em> at many points, so as to reduce the likelihood to get stuck. (It's not completely foolproof, as there is a tiny chance that the <em>Event<\/em> could get set between checking and accessing the <em>Queue<\/em>. If that's a problem, one needs to use some synchronization primitive arbitrating access to both the <em>Event<\/em> and the <em>Queue<\/em> combined.)<\/p>\n\n<p>It also wraps the received <em>RawArrays<\/em> in <em>ndarrays<\/em> at the very beginning and reads data from an external black box (<code>debugio.input<\/code> in the example).<\/p>\n\n<p>Note that by playing around with the <code>step=<\/code> arguments to both workers in the <code>main()<\/code> function, we can change the ratio of how much reading and writing is done (strictly for testing purposes - in a production environment <code>step=<\/code> would be <code>1<\/code>, reading and writing all <em>numpy<\/em> array members).<\/p>\n\n<p>Increasing both values makes the workers only access a few of the values in the <em>numpy<\/em> arrays, thereby significantly speeding everything up, which goes to show that the performance is not limited by the communication between the worker processes. Had we put <em>numpy<\/em> arrays directly onto the <em>Queues<\/em>, copying them forth and back between the processes in whole, increasing the step size would not have significantly improved the performance - it would have remained slow.<\/p>\n\n<p>For reference, here is the <code>debugio<\/code> module I used for testing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#!\/usr\/bin\/env python3\n# -*- coding: utf-8 -*-\n\nfrom ast import literal_eval\nfrom io import RawIOBase, BufferedReader, BufferedWriter, TextIOWrapper\n\n\nclass DebugInput(RawIOBase):\n    def __init__(self, end=None):\n        if end is not None and end &lt; 0:\n            raise ValueError(\"end must be non-negative\")\n\n        super().__init__()\n        self.pos = 0\n        self.end = end\n\n    def readable(self):\n        return True\n\n    def read(self, size=-1):\n        if self.end is None:\n            if size &lt; 0:\n                raise NotImplementedError(\"size must be non-negative\")\n            end = self.pos + size\n        elif size &lt; 0:\n            end = self.end\n        else:\n            end = min(self.pos + size, self.end)\n\n        lines = []\n        while self.pos &lt; end:\n            offset = self.pos % 400\n            pos = self.pos - offset\n            if offset &lt; 18:\n                i = (offset + 2) \/\/ 2\n                pos += i * 2 - 2\n            elif offset &lt; 288:\n                i = (offset + 12) \/\/ 3\n                pos += i * 3 - 12\n            else:\n                i = (offset + 112) \/\/ 4\n                pos += i * 4 - 112\n\n            line = str(i).encode('ascii') + b'\\n'\n            line = line[self.pos - pos:end - pos]\n            self.pos += len(line)\n            size -= len(line)\n            lines.append(line)\n\n        return b''.join(lines)\n\n    def readinto(self, b):\n        data = self.read(len(b))\n        b[:len(data)] = data\n        return len(data)\n\n    def seekable(self):\n        return True\n\n    def seek(self, offset, whence=0):\n        if whence == 0:\n            pos = offset\n        elif whence == 1:\n            pos = self.pos + offset\n        elif whence == 2:\n            if self.end is None:\n                raise ValueError(\"cannot seek to end of infinite stream\")\n            pos = self.end + offset\n        else:\n            raise NotImplementedError(\"unknown whence value\")\n\n        self.pos = max((pos if self.end is None else min(pos, self.end)), 0)\n        return self.pos\n\n\nclass DebugOutput(RawIOBase):\n    def __init__(self):\n        super().__init__()\n        self.buf = b''\n        self.num = 1\n\n    def writable(self):\n        return True\n\n    def write(self, b):\n        *lines, self.buf = (self.buf + b).split(b'\\n')\n\n        for line in lines:\n            value = literal_eval(line.decode('ascii'))\n            if value != int(value) or int(value) &amp; 255 != self.num:\n                raise ValueError(\"expected {}, got {}\".format(self.num, value))\n\n            self.num = self.num % 127 + 1\n\n        return len(b)\n\n\ninput = TextIOWrapper(BufferedReader(DebugInput()), encoding='ascii')\noutput = TextIOWrapper(BufferedWriter(DebugOutput()), encoding='ascii')\n<\/code><\/pre>\n"}]}
{"QuestionId":31186650,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-07-02T13:58:36.130","AcceptedAnswerId":31189561.0,"OwnerUserId":1366749.0,"Title":"What are the most idiomatic ways to combine tensors in torch?","Body":"<p>I'm confronted with concatenating three tensors together so that 3 px1 tensors become one 3px1 tensor. <\/p>\n\n<p>The most succinct I could come up with was:<\/p>\n\n<pre><\/pre>\n\n<p>Are there ways to do this without converting to tables and back to tensors? It seems like there should be a generic way of concatenating tensors along some specified dimension assuming they have compatible shapes.<\/p>\n\n<p>I can see how it would be possible to write such a function, does one not exist?<\/p>\n","answers":[{"AnswerId":"31189561","CreationDate":"2015-07-02T16:08:17.993","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>a = torch.randn(3,1)\nb = torch.randn(3,1)\nc = torch.randn(3,1)\n\nd = torch.cat(a,b,1):cat(c,1)\n\nprint(d)\n<\/code><\/pre>\n"}]}
{"QuestionId":31187549,"AnswerCount":1,"Tags":"<image-processing><face-recognition><torch><conv-neural-network>","CreationDate":"2015-07-02T14:37:54.847","AcceptedAnswerId":null,"OwnerUserId":4961048.0,"Title":"False Positives with Face recognition","Body":"<p>I have a CNN trained upon the images (cropped faces) of Mark Ruffalo. For my positive class I have around 200 images and for the negative datapoints I have sampled 200 random faces.<\/p>\n\n<p>The model has a high recall but a very low precision. How could I increase the precision ?Also I am constrained by the number of positive images that I have. I am ready to compromise the recall in this tradeoff.<\/p>\n\n<p>I have tried increasing the number of negative samples but that introduces a form of bias and the model starts classifying everything as negative to attain a local optima.<\/p>\n\n<p>I have based my CNN upon overfeat:<\/p>\n\n<pre><\/pre>\n\n<p>Kindly Help<\/p>\n","answers":[{"AnswerId":"31192847","CreationDate":"2015-07-02T19:16:32.630","ParentId":null,"OwnerUserId":"4992019","Title":null,"Body":"<p>Try playing with the raw output of the CNN instead of taking the sign() of the output node (since it is a positive and negative class I assume there is only one output in the range <code>[-1,1]<\/code>). <\/p>\n\n<p>For instance, for one sample, the output could be <code>[0.9]<\/code> indicating that the positive class should be picked. But if you play with this values, you can find a specific <code>threshold<\/code> value, <em>hopefully<\/em>, that gives you the precision you need. In other words, if you find that anything greater than <code>[-0.35]<\/code> should actually be chosen as the positive class because it gived you better precision, then <code>-0.35<\/code> should be your <code>threshold<\/code> value.<\/p>\n\n<p>This is where <a href=\"https:\/\/en.wikipedia.org\/wiki\/Receiver_operating_characteristic\" rel=\"nofollow\">ROC analysis<\/a> comes in handy. <\/p>\n\n<p>Let me know if this helps.<\/p>\n"}]}
{"QuestionId":31193977,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2015-07-02T20:22:34.900","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch, which command to insert data in a Torch Tensor?","Body":"<p>I'm using a Torch command to insert data into a simple table, and it works fine:<\/p>\n\n<pre><\/pre>\n\n<p>Now someone made me notice that using Torch Tensors would make everything more efficient.\nSo I replaced the first line with <\/p>\n\n<pre><\/pre>\n\n<p>but unfortunately I cannot find any Tensor function able to replace the table.insert() function.<\/p>\n\n<p>Do you have any idea?<\/p>\n","answers":[{"AnswerId":"31195580","CreationDate":"2015-07-02T22:19:32.600","ParentId":null,"OwnerUserId":"2726734","Title":null,"Body":"<p>There is no function which corresponds to the append functionality of insert since Tensor objects are a fixed size.\nWhat I see your code doing is concatenating three tables into one. If you are using Tensors:<\/p>\n\n<pre><code> firstHalf = torch.Tensor(firstHalf)\n secondHalf = torch.Tensor(secondHalf)\n presentWord = torch.Tensor(presentWord)\n<\/code><\/pre>\n\n<p>then concatenating them together is easy:<\/p>\n\n<pre><code> completeProfile = firstHalf:cat(secondHalf):cat(presentWord)\n<\/code><\/pre>\n\n<p>Another option is to store the last index you inserted to so you know where to \"append\" onto the tensor. The function below creates a closure that will keep track of that last index for you. <\/p>\n\n<pre><code>function appender(t)\n   local last = 0\n   return function(i, v)\n       last = last + 1\n       t[last] = v\n   end\nend\n\ncompleteProfile = torch.Tensor(#firstHalf + #secondHalf + #presentWord)\n\nprofile_append = appender(completeProfile)\n\ntable.foreach(firstHalf, profile_append)\ntable.foreach(secondHalf, profile_append)\ntable.foreach(presentWord, profile_append)\n<\/code><\/pre>\n"}]}
{"QuestionId":31196016,"AnswerCount":0,"Tags":"<gnuplot><torch>","CreationDate":"2015-07-02T23:05:56.483","AcceptedAnswerId":null,"OwnerUserId":3917668.0,"Title":"plotting a 3D+colour scatter with gnuplot (on torch7)","Body":"<p>I'm working with torch7, and I created a PCA function, which gives me an Nx3 tensor which I wish to plot (3D scatter).<\/p>\n\n<p>I stored it in a file ().\nnow I want to plot it, I wrote the following lines<\/p>\n\n<blockquote>\n  <p>NOTE: those lines are in (lua), but you don't really need to know the language, because the command  uses the regular  commands.<\/p>\n  \n  <p>NOTE 2: I followed helpers on this forum to create this part, I probably read a relevant thread you might want to link here. If you do, please explain what's the difference between the linked explanation an what I did<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>cols 1 through 3 in  are the x,y,z coordinates, col 4 is either 1 or 2 (determines colour).<\/p>\n\n<blockquote>\n  <p>LAST NOTE: my script doesn't print an error of any kind, it just doesn't plot the desired 3D scatter.<\/p>\n<\/blockquote>\n\n<p>Thanks ahead<\/p>\n","answers":[]}
{"QuestionId":31198353,"AnswerCount":1,"Tags":"<python><machine-learning><logistic-regression><theano>","CreationDate":"2015-07-03T04:15:26.460","AcceptedAnswerId":31417748.0,"OwnerUserId":4765036.0,"Title":"Python + Theano: Logistic regression weights do not update","Body":"<p>I've compared extensively to existing tutorials but I can't figure out why my weights don't update. Here is the function that return the list of updates:<\/p>\n\n<pre><\/pre>\n\n<p>It is defined at the top level, outside of any classes. This is standard gradient descent for each param. The 'params' parameter here is fed in as mlp.params, which is simply the concatenated lists of the param lists for each layer. I removed every layer except for a logistic regression one to isolate the reason as to why my cost was not decreasing. The following is the definition of mlp.params in MLP's constructor. It follows the definition of each layer and their respective param lists. <\/p>\n\n<pre><\/pre>\n\n<p>The following is the train function, which I call for each minibatch during each epoch:<\/p>\n\n<pre><\/pre>\n\n<p>If you require further details, the entire file is available here: <a href=\"http:\/\/pastebin.com\/EeNmXfGD\" rel=\"nofollow\">http:\/\/pastebin.com\/EeNmXfGD<\/a><\/p>\n\n<p>I don't know how many people use Theano (it doesn't seem like plenty); if you've read to this point, thank you. <\/p>\n\n<p>Fixed: I've determined that I can't use average squared error as the cost function. It works as usual after replacing it with a negative log-likelihood.<\/p>\n","answers":[{"AnswerId":"31417748","CreationDate":"2015-07-14T21:30:23.060","ParentId":null,"OwnerUserId":"5116849","Title":null,"Body":"<p>This behavior it caused by a few things but it comes down to the cost not being properly computed. In your implementation , the output of the LogisticRegression layer is the predicted class for every input digit (obtained with the argmax operation) and you take the squared difference between it and the expected prediction.<\/p>\n\n<p>This will give you gradients of 0s wrt to any parameter in your model because the gradient of the output of the argmax (predicted class) wrt the input of the argmax (class probabilities) will be 0. <\/p>\n\n<p>Instead, the LogisticRegression should output the probabilities of the classes : <\/p>\n\n<pre><code>def output(self, input):\n    input = input.flatten(2)\n    self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W) + self.b)\n    return self.p_y_given_x\n<\/code><\/pre>\n\n<p>And then in the MLP class, you compute the cost. You can used mean squared error between the desired probabilities for each class and the probabilities computed by the model but people tend to use the Negative Log Likelihood of the expected classes and you can implement it as such in the MLP class : <\/p>\n\n<pre><code>def neg_log_likelihood(self, x, y):\n    p_y_given_x = self.output(x)\n    return -T.mean(T.log(p_y_given_x)[T.arange(y.shape[0]), y])\n<\/code><\/pre>\n\n<p>Then you can use this function to compute your cost and the model trains :<\/p>\n\n<pre><code>cost = mlp.neg_log_likelihood(x_, y)\n<\/code><\/pre>\n\n<p>A few additional things: <\/p>\n\n<ul>\n<li>At line 215, when you print your cost, you format it as an integer value but it is a floating point value; this will lose precision in the monitoring.<\/li>\n<li>Initializing all the weights to 0s as you do in your LogisticRegression class is often not recommended. Weights should differ in their original values so as to help break symmetry<\/li>\n<\/ul>\n"}]}
{"QuestionId":31201861,"AnswerCount":1,"Tags":"<python><numpy><anaconda><theano>","CreationDate":"2015-07-03T08:14:59.840","AcceptedAnswerId":31208750.0,"OwnerUserId":3624880.0,"Title":"importing theano on anaconda 3.10.0","Body":"<p>I'm trying to import theano and I'm using the anaconda version 3.10.0, Can anyone give me directions to how to proceed?<\/p>\n\n<p>Thanks in advance<\/p>\n\n<blockquote>\n  <p>Problem occurred during compilation with the command line below: g++\n  -shared -g -D NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION -m64 -DMS_WIN64 -IC:\\Users\\Supreeth\\Anaconda\\lib\\site-packages\\numpy\\core\\include -IC:\\Users\\Supreeth\\Anaconda\\include -o C:\\Users\\Supreeth\\AppData\\Local\\Theano\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-2.7.9-64\\lazylinker_ext\\lazylinker_ext.pyd C:\\Users\\Supreeth\\AppData\\Local\\Theano\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-2.7.9-64\\lazylinker_ext\\mod.cpp\n  -LC:\\Users\\Supreeth\\Anaconda\\libs -LC:\\Users\\Supreeth\\Anaconda -lpython27<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"31208750","CreationDate":"2015-07-03T14:04:59.587","ParentId":null,"OwnerUserId":"3624880","Title":null,"Body":"<p>I find the solution from the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">Theano Installation Document<\/a><\/p>\n\n<blockquote>\n  <p>Specifically for Anaconda users just use the command <\/p>\n<\/blockquote>\n\n<pre><code>$ conda install mingw libpython\n<\/code><\/pre>\n\n<p>in the Anaconda Command Prompt<\/p>\n"}]}
{"QuestionId":31204702,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2015-07-03T10:35:00.550","AcceptedAnswerId":31210528.0,"OwnerUserId":3284343.0,"Title":"Torch7 Access one element from a Tensor as a Tensor","Body":"<p>I'm working on Torch7 to train some neural nets and I've got a Tensor of dim 1 (vector) and I want to access the element i in this vector.\nUnfortunately, it gives me an integer instead of a Tensor of size 1.<\/p>\n\n<p>I got this :<\/p>\n\n<pre><\/pre>\n\n<p>I want this :<\/p>\n\n<pre><\/pre>\n\n<p>I'm obliged to do this :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31210528","CreationDate":"2015-07-03T15:51:55.597","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>You can use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\">torch indexing operator<\/a> as follow:<\/p>\n\n<pre><code>&gt; t = matrix[{ {1} }]\n&gt; = t\n 1\n[torch.DoubleTensor of size 1]\n<\/code><\/pre>\n"}]}
{"QuestionId":31209252,"AnswerCount":2,"Tags":"<linux><makefile><fedora><caffe><gflags>","CreationDate":"2015-07-03T14:33:38.897","AcceptedAnswerId":null,"OwnerUserId":4068560.0,"Title":"Caffe compilation fails: make: *** [.build_release\/src\/caffe\/data_transformer.o] Error 1","Body":"<p>I am trying to build caffe after the instructions on <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#prerequisites\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/installation.html#prerequisites<\/a>\nWhen compiling i get the following error:\n(I use Fedora 22)<\/p>\n\n<pre><\/pre>\n\n<p>What am i doin wrong?<\/p>\n","answers":[{"AnswerId":"41313192","CreationDate":"2016-12-24T12:18:33.797","ParentId":null,"OwnerUserId":"4301883","Title":null,"Body":"<p>You have to install missing dependencies (gflags).  <\/p>\n\n<p>Fedora\/RHEL\/CentOS: <code>sudo yum install gflags-devel<\/code><\/p>\n\n<p>Ubuntu: <code>sudo apt-get install libgflags-dev<\/code><\/p>\n\n<p>There are also instructions for other dependencies: <\/p>\n\n<p>Fedora\/RHEL\/CentOS :<a href=\"http:\/\/caffe.berkeleyvision.org\/install_yum.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/install_yum.html<\/a><\/p>\n\n<p>Ubuntu: <a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/install_apt.html<\/a><\/p>\n"},{"AnswerId":"37355261","CreationDate":"2016-05-20T20:23:21.250","ParentId":null,"OwnerUserId":"2950746","Title":null,"Body":"<p>To install missing gflag dependencies <\/p>\n\n<pre><code>wget https:\/\/github.com\/schuhschuh\/gflags\/archive\/master.zip\nunzip master.zip\ncd gflags-master\nmkdir build &amp;&amp; cd build\nexport CXXFLAGS=\"-fPIC\" &amp;&amp; cmake .. &amp;&amp; make VERBOSE=1\nmake \nsudo make install\n<\/code><\/pre>\n"}]}
{"QuestionId":31214799,"AnswerCount":1,"Tags":"<matlab><deep-learning><caffe>","CreationDate":"2015-07-03T22:18:14.400","AcceptedAnswerId":null,"OwnerUserId":3121945.0,"Title":"Why do the features extracted with matcaffe_demo.m and matcaffe_batch.m for the same input are different?","Body":"<p>I am using Caffe to extract features with matlab wrapper.I have 5011 images as test data set.I chopped all the layers after  in . I found out if you take the same image as input of  and , you will get the different 4096-dim features.<br>\nCould someone tell me why?<br>\nwhat is the differences between you extract features from all these images one by one with  and extract features by listing all these images with ? <\/p>\n","answers":[{"AnswerId":"32183343","CreationDate":"2015-08-24T13:18:00.233","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can find the answer to this question <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2689\" rel=\"nofollow\">at caffe github<\/a>.<br>\nBasically, <code>matcaffe_demo<\/code> is used for classification and it <strong>averages<\/strong> results of 10 crops of the input image, while <code>matcaffe_bathc<\/code> uses only a single input.  <\/p>\n\n<p>Moreover, note that these m-files are no longer available in recent caffe versions.<\/p>\n"}]}
{"QuestionId":31218110,"AnswerCount":1,"Tags":"<ios><iphone><static><caffe>","CreationDate":"2015-07-04T07:24:05.243","AcceptedAnswerId":null,"OwnerUserId":3132189.0,"Title":"How to solve Unknown layer type in Caffe IOS.?","Body":"<p>1.I download source from <a href=\"https:\/\/github.com\/aleph7\/caffe\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aleph7\/caffe\/<\/a> and builded the caffe as static library for IOS and IPhone.\n2.Created sample demo code and Linked Caffe static lib(.a) and execute the code.\n3.Now I got Run time error<\/p>\n\n<p>F0519 14:54:12.494139 14504 layer_factory.hpp:77] Check failed: registry.count(t ype) == 1 (0 vs. 1) Unknown layer type: Convolution (known types: MemoryData)<\/p>\n\n<p>4.I searched a lot and found one solution from below link<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/30325108\/caffe-layer-creation-failure\">Caffe layer creation failure<\/a><\/p>\n\n<p>5.If I create dynamic library instead of static library. It will work.<\/p>\n\n<p>6.I tried to convert static library into Dynamic library.I got Error on Xcode that is cannot open the project I referred below link.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/1349115\/how-do-i-change-an-existing-xcode-target-from-dynamic-to-static\">How do I change an existing XCode target from dynamic to static?<\/a><\/p>\n\n<p>Can you help how to solve this..?<\/p>\n","answers":[{"AnswerId":"42093978","CreationDate":"2017-02-07T15:42:00.683","ParentId":null,"OwnerUserId":"2572084","Title":null,"Body":"<p>Caffe register layer classes through <code>REGISTER_LAYER_CLASS<\/code> macro. Some build tools (e.g. Xcode) will optimize some part of it out (a little complex to explain). You can add <code>-Wl,-force_load path\/to\/libcaffe.a<\/code> to the <code>Other Linker Flags<\/code> build option. It will force Xcode to load all the things in libcaffe.a to the final target. <\/p>\n\n<p>And more, <code>Unknown layer type<\/code> error can also caused by stale code. The caffe code in <a href=\"https:\/\/github.com\/aleph7\/caffe\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aleph7\/caffe\/<\/a> is out of date. You can try my port at <a href=\"https:\/\/github.com\/solrex\/caffe-mobile\" rel=\"nofollow noreferrer\">https:\/\/github.com\/solrex\/caffe-mobile<\/a>. It inlcudes a demo iOS app, works with the newest build tool. The Caffe source is up to date and you can sync the latest caffe code your self.<\/p>\n"}]}
{"QuestionId":31228557,"AnswerCount":2,"Tags":"<python><debugging><theano>","CreationDate":"2015-07-05T08:37:27.760","AcceptedAnswerId":31229234.0,"OwnerUserId":379539.0,"Title":"Printing expressions of variables in Theano","Body":"<p>If I want to print some variable for debugging in theano, it is easy, just write<br>\n , and then use  instead of  in the following computations.\nBut what if I want to print some expression of , for example . How can I do it?<br>\nIf I write  then I will need to insert  into the computation graph later, what is the recommended way to do it?<\/p>\n","answers":[{"AnswerId":"38439306","CreationDate":"2016-07-18T14:29:47.180","ParentId":null,"OwnerUserId":"86430","Title":null,"Body":"<p>This is a nasty hack, but I've resorted to things like:<\/p>\n\n<pre><code>x = 1e-11 * Print(\"mean of x\")(x.mean()) + x\n<\/code><\/pre>\n\n<p>If you make it <code>0 * Print(...)<\/code> then it gets optimised away.<\/p>\n"},{"AnswerId":"31229234","CreationDate":"2015-07-05T10:09:21.107","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The result of a print operation must be reachable (via some path in the computation graph) from an output of the Theano function. If you want to print an expression that is not used then a simple solution is to just include the result of that expression in the outputs of the Theano function.<\/p>\n\n<p>Suppose you are interested in <code>x*y<\/code> but would like to print <code>x+y<\/code>, then<\/p>\n\n<pre><code>x = theano.tensor.scalar()\ny = theano.tensor.scalar()\nz = printing.Print('x+y is: ')(x+y)\nf1 = theano.function([x, y], [x * y]\nf2 = theano.function([x, y], [z]\nf3 = theano.function([x, y], [x * y, z]\n<\/code><\/pre>\n\n<p>f1 will fail to print <code>x+y<\/code> because z is not reachable from an output of the function; f2 will print <code>x+y<\/code> but will not compute <code>x*y<\/code>; f3 will do both.<\/p>\n"}]}
{"QuestionId":31230862,"AnswerCount":1,"Tags":"<arrays><python-3.x><numpy><theano>","CreationDate":"2015-07-05T13:29:34.217","AcceptedAnswerId":31231051.0,"OwnerUserId":3190076.0,"Title":"White spaces in Theano arrays","Body":"<p>I just started playing around with Theano and I got surprised by the result of this code.<\/p>\n\n<pre><\/pre>\n\n<p>Using python3 I get:<\/p>\n\n<pre><\/pre>\n\n<p>The array itself is correct, it contains the right values, however the printed output is odd. I would expect something like this:<\/p>\n\n<pre><\/pre>\n\n<p>or<\/p>\n\n<pre><\/pre>\n\n<p>Why it is so? What are the extra white spaces? Shall I be concerned about?<\/p>\n","answers":[{"AnswerId":"31231051","CreationDate":"2015-07-05T13:51:02.973","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>What you're printing is a <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ndarray.html\" rel=\"nofollow\"><code>numpy.ndarray<\/code><\/a>. By default they format themselves like this when printed.<\/p>\n\n<p>The output array is a floating point array because, by default, Theano uses floating point tensors.<\/p>\n\n<p>If you want to use integer tensors then you need to specify a <code>dtype<\/code>:<\/p>\n\n<pre><code>a = T.vector(dtype='int64')\n<\/code><\/pre>\n\n<p>Or use a bit of syntactic sugar:<\/p>\n\n<pre><code>a = T.lvector()\n<\/code><\/pre>\n\n<p>Compare your output with the output of the following:<\/p>\n\n<pre><code>print numpy.array([0, 2, 1026], dtype=numpy.float64)\nprint numpy.array([0, 2, 1026], dtype=numpy.int64)\n<\/code><\/pre>\n\n<p>You can change the default printing options of numpy using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.set_printoptions.html\" rel=\"nofollow\"><code>numpy.set_printoptions<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":31247616,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-07-06T13:50:27.040","AcceptedAnswerId":null,"OwnerUserId":5085490.0,"Title":"Caffe fails: relu_layer.cu: 29] check failed: error==cudaSuccess <9 vs 0> invalid configuration argument","Body":"<p>When I run my Caffe implement, it fails as follows. <\/p>\n\n<p><strong>...relu_layer.cu: 29] check failed: error==cudaSuccess &lt;9 vs 0> invalid configuration argument<\/strong><\/p>\n\n<p>It is noted that it does not fail in running other Caffe implements (eg. MNIST recognition).\nMy GPU is NVIDIA GeForce GT 620, with computaion ability 2.1, and on Windows 7.<\/p>\n\n<p>If you come across the similar questions, please let me know, and give me some help.<\/p>\n","answers":[{"AnswerId":"39583565","CreationDate":"2016-09-19T23:31:11.473","ParentId":null,"OwnerUserId":"6850895","Title":null,"Body":"<p>The driver is not the most updated. You can use <a href=\"http:\/\/www.nvidia.com\/Download\/Scan.aspx?lang=en-us\" rel=\"nofollow\">http:\/\/www.nvidia.com\/Download\/Scan.aspx?lang=en-us<\/a> \nto test your driver and update it. <\/p>\n"}]}
{"QuestionId":31253870,"AnswerCount":6,"Tags":"<c++><opencv><python-3.x><opencv3.0><caffe>","CreationDate":"2015-07-06T19:17:09.283","AcceptedAnswerId":null,"OwnerUserId":1561108.0,"Title":"Caffe: opencv error","Body":"<p>I've built opencv 3.0 from source and can run a few sample apps, build against the headers ok so I presume it's installed successfully. <\/p>\n\n<p>I'm also using python3 and I now go to install and build caffe. I set a few variables in Makefile.config as I'm using the CPU due to having an AMD GPU and also Anaconda.<\/p>\n\n<p>When I run make all I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>from searching I think this is something to do with using openCV 3 but I'm not sure where to start looking for a solution. Any help?<\/p>\n\n<p>And yes I'm one of the horde of inexperienced users looking to fiddle with the Google Inception learning technique.<\/p>\n","answers":[{"AnswerId":"49849765","CreationDate":"2018-04-16T04:42:07.530","ParentId":null,"OwnerUserId":"7360523","Title":null,"Body":"<p>The problem report is very clear. There is a problem with linking library libraries.The reason may be the difference between 3.0 and 2.x.\nYou need to add<\/p>\n\n<pre><code>opencv_core opencv_highgui opencv_imgproc opencv_imgcodecs\n<\/code><\/pre>\n\n<p>into LIBRARIES +=.<\/p>\n"},{"AnswerId":"56014995","CreationDate":"2019-05-07T03:08:50.383","ParentId":null,"OwnerUserId":"11462352","Title":null,"Body":"<p>You can edit <code>Makefile.config<\/code> with the following 2 lines like this and it worked for me. Note that your opencv path <strong>must be set before default path<\/strong>!<\/p>\n\n<pre><code>INCLUDE_DIRS := $(PYTHON_INCLUDE) \/home\/young\/Soft\/openCV-3.3.1\/include \\\n    \/usr\/local\/include \/usr\/include\/hdf5\/serial\nLIBRARY_DIRS := $(PYTHON_LIB) \/home\/young\/Soft\/openCV-3.3.1\/lib \\\n    \/usr\/local\/lib \/usr\/lib \/usr\/lib\/x86_64-linux-gnu \/usr\/lib\/x86_64-linux-gnu\/hdf5\/serial                                                                                            \n<\/code><\/pre>\n"},{"AnswerId":"31289673","CreationDate":"2015-07-08T10:15:13.747","ParentId":null,"OwnerUserId":"1561108","Title":null,"Body":"<p>I used <code>cmake<\/code> instead with the <code>-DBUILD_TIFF=ON<\/code> flag and got a successful build.<\/p>\n"},{"AnswerId":"41065595","CreationDate":"2016-12-09T17:14:58.483","ParentId":null,"OwnerUserId":"4234099","Title":null,"Body":"<p>It could be that you are using OpenCV version 3.  If yes just uncomment the following line in your <code>Makefile.config<\/code>:<\/p>\n\n<pre><code># OPENCV_VERSION := 3\n<\/code><\/pre>\n\n<p>So it will look like<\/p>\n\n<pre><code>OPENCV_VERSION := 3\n<\/code><\/pre>\n\n<p>You could verify the version currently in use by doing:<\/p>\n\n<pre><code>$ python\n&gt;&gt;&gt; import cv2\n&gt;&gt;&gt; cv2.__version__\n'3.1.0-dev'\n<\/code><\/pre>\n"},{"AnswerId":"58792879","CreationDate":"2019-11-10T20:59:45.993","ParentId":null,"OwnerUserId":"8128190","Title":null,"Body":"<p>A quick workaround is to add <code>-lopencv_imgcodecs<\/code> flag when you're compiling your code.<br>\nThis worked for me:<br>\n<code>g++ test.cpp -o test &lt;Some flags&gt; -lopencv_imgcodecs<\/code><\/p>\n"},{"AnswerId":"32922800","CreationDate":"2015-10-03T12:51:44.833","ParentId":null,"OwnerUserId":"2297751","Title":null,"Body":"<p>You can also add the <code>opencv_imgcodecs<\/code> to the MakeFile in line <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3140\/files#diff-b67911656ef5d18c4ae36cb6741b7965L187\">187<\/a>, see this <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3140\">pull<\/a>.<\/p>\n"}]}
{"QuestionId":31275282,"AnswerCount":1,"Tags":"<python-3.x><neural-network><theano>","CreationDate":"2015-07-07T17:18:30.010","AcceptedAnswerId":31295512.0,"OwnerUserId":2077723.0,"Title":"theano function not updating parameters during gradient optimization in feed forward neural net","Body":"<p>Trying to get my hands wet with theano and deep nets by starting with a very simple implementation of a three layer feed forward neural network and testing it on the mnist data set. <\/p>\n\n<p>I am using a rudimentary implementation of stochastic gradient descent to start out with, and the network is not training properly. The parameters of the network are not being updated. <\/p>\n\n<p>Was wondering if anyone could could point out what I'm doing wrong. <\/p>\n\n<p>The following code is my lstm module. I've called it that because I planned on implementing lstm networks in the future. <\/p>\n\n<pre><\/pre>\n\n<p>The following code is where I create, train, and test a simple three-layer feed forward network on the mnist data set. <\/p>\n\n<pre><\/pre>\n\n<p>The problem I'm facing is that the parameters are not being updated properly. I'm not sure if that's because I'm not calculating the gradient properly, or If I'm not using the theano function correctly. <\/p>\n","answers":[{"AnswerId":"31295512","CreationDate":"2015-07-08T14:21:59.527","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>You have to make more than one pass on the dataset when using stochastic gradient descent.\nIt is not unusual that the classification error and the confusion matrix do not change much during the first epoch, especially if the dataset is small.<\/p>\n\n<p>I made the following change in your code to train for 100 epochs<\/p>\n\n<pre><code>for i in xrange(100):\n    for X, y in zip(X_train, labels_train):\n        c = update(X, y)\n<\/code><\/pre>\n\n<p>The confusion matrix seems to have started improving:<\/p>\n\n<pre><code>[[ 0  0 18  0 13  4  5  0  5  0]\n [ 0 42  0  2  0  0  0  0  2  0]\n [ 0  0 51  0  0  0  0  1  0  0]\n [ 0  0  0 45  0  1  0  1  2  0]\n [ 0  0  0  0 33  0  0  0  0  0]\n [ 0  0  0  0  0 47  0  0  0  0]\n [ 0  0  0  0  0  0 45  0  0  0]\n [ 0  0  0  0  1  0  0 48  0  0]\n [ 0  2  1  0  0  0  0  0 34  0]\n [ 0  1  0 25  0  3  0  2 16  0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":31277820,"AnswerCount":0,"Tags":"<caffe><atlas>","CreationDate":"2015-07-07T19:37:39.583","AcceptedAnswerId":null,"OwnerUserId":5064653.0,"Title":"Error building caffe - undefined reference to cblas_sgemv","Body":"<p>I am trying to build caffe using Atlas, CUDA and cuDNN on Fedora 22 and I'm getting this error.<\/p>\n\n<pre><\/pre>\n\n<p>The fix mentioned <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2348\" rel=\"nofollow\">here<\/a> is to add <em>opencv_imgcodecs<\/em> to the list of libraries but I'm still getting the same error.<\/p>\n\n<p>Is this problem related to Atlas or is the fix in the Makefile? I had to edit FindAtlas.cmake earlier to get it to find the libraries on my system.<\/p>\n","answers":[]}
{"QuestionId":31286024,"AnswerCount":4,"Tags":"<python><neural-network><artificial-intelligence><caffe><deep-dream>","CreationDate":"2015-07-08T07:22:25.923","AcceptedAnswerId":null,"OwnerUserId":3792198.0,"Title":"DeepDream taking too long to render image","Body":"<p>I managed to install #DeepDream in my server.<\/p>\n\n<p>I have duo core and 2gb Ram. but it taking 1min to process a image of size 100kbp.<\/p>\n\n<p>Any advice ?<\/p>\n","answers":[{"AnswerId":"31620288","CreationDate":"2015-07-24T21:46:55.907","ParentId":null,"OwnerUserId":"2353472","Title":null,"Body":"<p>Unless you can move to a better workstation\/get a GPU, you'll have to do with resizing the image.<\/p>\n\n<pre><code>img = PIL.Image.open('sky1024px.jpg')\nimg = np.float32(img.resize( [int(0.5 * s) for s in img.size] ))\n<\/code><\/pre>\n"},{"AnswerId":"36781301","CreationDate":"2016-04-21T22:07:09.870","ParentId":null,"OwnerUserId":"4709102","Title":null,"Body":"<p>Taking 1 minute to process a 100kb image is a sensible turnaround time for #deepdream, and we accept that these renders have an incredibly long baking time. Often, experimental research software will run too slow, hungry for a future of faster computers. That said, there are a couple ways that come to mind about making your setup execute faster.<\/p>\n\n<ul>\n<li><p>Thread! Increase thread count with a command line argument. Here's one way to enable multi-threading in Caffe <a href=\"https:\/\/stackoverflow.com\/questions\/31395729\/how-to-enable-multithreading-with-caffe\">How to enable multithreading with Caffe?<\/a><\/p><\/li>\n<li><p>GPU! Install CUDA and switch from CPU rendering to GPU rendering. If your server doesn't have a special GPU, try getting a GPU instance on amazon ec2. <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Install-Caffe-on-EC2-from-scratch-(Ubuntu,-CUDA-7,-cuDNN)\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Install-Caffe-on-EC2-from-scratch-(Ubuntu,-CUDA-7,-cuDNN)<\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"31430476","CreationDate":"2015-07-15T12:31:56.650","ParentId":null,"OwnerUserId":"3533322","Title":null,"Body":"<p>Do you run it in a Virtual Machine on Windows or OS X? If so, then it's probably not going to work any faster. In a Virtual Machine (I'm using Docker) you're most of the time not able to use CUDA to render the Images. I have the same problem and I'm going to try it by installing Ubuntu and then install the NVidia drivers for CUDA. At the moment I'm rendering 1080p images which are around 300kb and it takes 15 minutes to do 1 image on an Intel core i7 with 8gb of ram. <\/p>\n"},{"AnswerId":"39757574","CreationDate":"2016-09-28T21:00:31.093","ParentId":null,"OwnerUserId":"5256795","Title":null,"Body":"<p>As a rule of thumb deep learning is hard on both compute and memory resources. A 2gb RAM Core Duo machine is just not a good choice for deep learning. Keep in mind a lot of the people who pioneered this field did much of their research using GTX Titan cards because CPU computation even on xeon servers is prohibitivly slow when training deep learning networks.<\/p>\n"}]}
{"QuestionId":31288156,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-07-08T09:07:14.543","AcceptedAnswerId":null,"OwnerUserId":1348187.0,"Title":"Convolutional Neural Networks with Caffe and NEGATIVE IMAGES","Body":"<p>When training a set of classes (let's say #clases <em>(number of classes)<\/em> = N) on Caffe Deep Learning (or any CNN framework) and I make a query to the <em>caffemodel<\/em>, I get a % of probability of that image could be OK.<\/p>\n\n<p>So, let's take a picture of a similar Class 1, and I get the result:<\/p>\n\n<blockquote>\n  <p>1.- 90%<\/p>\n  \n  <p>2.- 10%<\/p>\n  \n  <p>rest... 0%<\/p>\n<\/blockquote>\n\n<p>the problem is: when I take a random picture (for example of my environment), <strong>I keep getting the same result<\/strong>, where one of the class is predominant (>90% probability) but it doesn't belong to any class.<\/p>\n\n<p>So what I'd like to hear is opinions\/answers from people which has experienced this and would have solved how to deal with no-sense inputs to the Neural Network.<\/p>\n\n<p>My purposes are:<\/p>\n\n<ol>\n<li>Train one more extra class with negative images (like with <a href=\"http:\/\/docs.opencv.org\/doc\/user_guide\/ug_traincascade.html\" rel=\"nofollow\">train_cascade<\/a>).<\/li>\n<li>Train one more class extra with all the positive images in the TRAIN set, and the negative on the VAL set.<\/li>\n<\/ol>\n\n<p>But my purposes don't have any scientific base to execute them, that's why I ask you this question.<\/p>\n\n<p>What would you do?<\/p>\n\n<p>Thank you very much in advance.<\/p>\n\n<p>Rafael.<\/p>\n","answers":[]}
{"QuestionId":31297745,"AnswerCount":1,"Tags":"<c++><boost><caffe>","CreationDate":"2015-07-08T15:54:00.313","AcceptedAnswerId":31299048.0,"OwnerUserId":3742823.0,"Title":"Installing caffe on Yosemite Boost Error","Body":"<p>I am trying to install caffe on Yosemite and I am getting the following error:<\/p>\n\n<p><\/p>\n\n<p>As suggested by this blogpost <a href=\"https:\/\/stackoverflow.com\/questions\/30745837\/compiling-caffe-on-yosemite\">compiling caffe on Yosemite<\/a> I downgraded the Boost to v1.57. <\/p>\n\n<p>Any suggestions on how to move forward?<\/p>\n","answers":[{"AnswerId":"31299048","CreationDate":"2015-07-08T16:59:54.200","ParentId":null,"OwnerUserId":"3742823","Title":null,"Body":"<p>This solved the issue:<\/p>\n\n<pre><code>make clean\nmake all\n<\/code><\/pre>\n"}]}
{"QuestionId":31304452,"AnswerCount":0,"Tags":"<python-2.7><numpy><bayesian><theano><pymc3>","CreationDate":"2015-07-08T21:49:51.363","AcceptedAnswerId":null,"OwnerUserId":5095886.0,"Title":"Using the pymc3 likelihood\/posterior outside of pymc3: how?","Body":"<p>For comparison purposes, I want to utilize the posterior density function outside of PyMC3.<\/p>\n\n<p>For my research project, I want to find out how well PyMC3 is performing compared to my own custom made code. As such, I need to compare it to our own in-house samplers and likelihood functions.<\/p>\n\n<p>I think I figured out how to call the internal PyMC3 posterior, but it feels very awkward, and I want to know if there is a better way. Right now I am hand-transforming variables, whereas I should just be able to pass pymc a parameter dictionary and get the posterior density. Is this possible in a straightforward manner?<\/p>\n\n<p>Thanks a lot!<\/p>\n\n<p>Demo code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":31304601,"AnswerCount":1,"Tags":"<arrays><lua><torch>","CreationDate":"2015-07-08T21:59:55.900","AcceptedAnswerId":31305512.0,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, how to select a subset of an array or tensor?","Body":"<p>I'm working on Torch\/Lua and have an array  of 10 elements.<\/p>\n\n<pre><\/pre>\n\n<p>If I write , I can read the structure of the 1st element of the array.<\/p>\n\n<pre><\/pre>\n\n<p>I need to select just 3 elements among all the 10, but I don't know which command to use.\nIf I were working on Matlab, I would write: , but here does not work.<\/p>\n\n<p>Do you have any suggestions?<\/p>\n","answers":[{"AnswerId":"31305512","CreationDate":"2015-07-08T23:20:33.693","ParentId":null,"OwnerUserId":"2726734","Title":null,"Body":"<h1>In Torch<\/h1>\n\n<pre><code>th&gt; x = torch.Tensor{1, 2, 3, 4, 5, 6, 7, 8, 9, 10}\n<\/code><\/pre>\n\n<p>To select a range, like the first three, use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"noreferrer\">the index operator<\/a>:<\/p>\n\n<pre><code>th&gt; x[{{1,3}}]\n1\n2\n3\n<\/code><\/pre>\n\n<p>Where 1 is the 'start' index, and 3 is the 'end' index. <\/p>\n\n<p>See <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#extracting-sub-tensors\" rel=\"noreferrer\">Extracting Sub-tensors<\/a> for more alternatives using Tensor.sub and Tensor.narrow <\/p>\n\n<hr>\n\n<h1>In Lua 5.2 or less<\/h1>\n\n<p>Lua tables, such as your <code>dataset<\/code> variable, do not have a method for selecting sub-ranges.<\/p>\n\n<pre><code>function subrange(t, first, last)\n  local sub = {}\n  for i=first,last do\n    sub[#sub + 1] = t[i]\n  end\n  return sub\nend\n\ndataset = {11,12,13,14,15,16,17,18,19,20}\n\nsub = subrange(dataset, 1, 3)\nprint(unpack(sub))\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>11    12   13\n<\/code><\/pre>\n\n<hr>\n\n<h1>In Lua 5.3<\/h1>\n\n<p>In Lua 5.3 you can use <a href=\"http:\/\/www.lua.org\/manual\/5.3\/manual.html#pdf-table.move\" rel=\"noreferrer\"><code>table.move<\/code><\/a>.<\/p>\n\n<pre><code>function subrange(t, first, last)\n     return table.move(t, first, last, 1, {})\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":31304860,"AnswerCount":1,"Tags":"<lua><torch><mnist>","CreationDate":"2015-07-08T22:23:16.043","AcceptedAnswerId":31320458.0,"OwnerUserId":225814.0,"Title":"Torch: luajit out of memory on simple task","Body":"<p>I am trying to load the MNIST dataset in the  repl and do mean subtraction by the following:<\/p>\n\n<pre><\/pre>\n\n<p>The last line causes the following error:<\/p>\n\n<p><\/p>\n\n<p>I am running this on a laptop with 16GB of RAM. Also MNIST has already been loaded into  so not sure why doing  would cause this issue. Any ideas?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"31320458","CreationDate":"2015-07-09T14:27:53.000","ParentId":null,"OwnerUserId":"225814","Title":null,"Body":"<p>The problem was that it was trying to print the whole matrix (which is large) to the console. <\/p>\n\n<p>This can be overcome by doing either\n<code>data = data:add(-mean)<\/code>\nor\n<code>data:add(-mean);<\/code> - notice the semicolon <\/p>\n\n<p>Answer provided by Soumith Chintala on the torch gitter.<\/p>\n"}]}
{"QuestionId":31307181,"AnswerCount":1,"Tags":"<protocol-buffers><caffe>","CreationDate":"2015-07-09T02:38:18.153","AcceptedAnswerId":null,"OwnerUserId":3742823.0,"Title":"Installation issue with protobuf and caffe","Body":"<p>From the last 2 days, I have been trying to install caffe on OSX 10.10 <\/p>\n\n<p>I was able to run all the installation commands successfully for caffe but when I tried to import caffe in ipython I got the exact same error: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2092\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/2092<\/a> <\/p>\n\n<p>So, as suggested in the thread, I tried to downgrade  from 3.0.0 to 2.6.1. I was successfully able to install  and the new version does say 2.6.1 <\/p>\n\n<p>But now I not able to install  python library. I am following the instructions mentioned here: <a href=\"https:\/\/github.com\/google\/protobuf\/tree\/v2.6.1\/python\" rel=\"nofollow\">https:\/\/github.com\/google\/protobuf\/tree\/v2.6.1\/python<\/a>\nI get the following error while running the command :<\/p>\n\n<blockquote>\n  <p>from google.protobuf import descriptor_pb2<\/p>\n  \n  <p>File \"\/path\/to\/protobuf-2.6.1\/python\/google\/protobuf\/descriptor_pb2.py\",\n  line 21, in module><\/p>\n  \n  <p>80\\x80\\x80\\x02\\\"}\\n\\x10\\x45numValueOptions\\x12\\x19\\n\\ndeprecated\\x18\\x01\n  \\x01(\\x08:\\x05\\x66\\x61lse\\x12\\x43\\n\\x14uninterpreted_option\\x18\\xe7\\x07\n  \\x03(\\x0b\\x32$.google.protobuf.UninterpretedOption*\\t\\x08\\xe8\\x07\\x10\\x80\\x80\\x80\\x80\\x02\\\"{\\n\\x0eServiceOptions\\x12\\x19\\n\\ndeprecated\\x18!\n  \\x01(\\x08:\\x05\\x66\\x61lse\\x12\\x43\\n\\x14uninterpreted_option\\x18\\xe7\\x07\n  \\x03(\\x0b\\x32$.google.protobuf.UninterpretedOption*\\t\\x08\\xe8\\x07\\x10\\x80\\x80\\x80\\x80\\x02\\\"z\\n\\rMethodOptions\\x12\\x19\\n\\ndeprecated\\x18!\n  \\x01(\\x08:\\x05\\x66\\x61lse\\x12\\x43\\n\\x14uninterpreted_option\\x18\\xe7\\x07\n  \\x03(\\x0b\\x32$.google.protobuf.UninterpretedOption*\\t\\x08\\xe8\\x07\\x10\\x80\\x80\\x80\\x80\\x02\\\"\\x9e\\x02\\n\\x13UninterpretedOption\\x12;\\n\\x04name\\x18\\x02\n  \\x03(\\x0b\\x32-.google.protobuf.UninterpretedOption.NamePart\\x12\\x18\\n\\x10identifier_value\\x18\\x03\n  \\x01(\\t\\x12\\x1a\\n\\x12positive_int_value\\x18\\x04\n  \\x01(\\x04\\x12\\x1a\\n\\x12negative_int_value\\x18\\x05\n  \\x01(\\x03\\x12\\x14\\n\\x0c\\x64ouble_value\\x18\\x06\n  \\x01(\\x01\\x12\\x14\\n\\x0cstring_value\\x18\\x07\n  \\x01(\\x0c\\x12\\x17\\n\\x0f\\x61ggregate_value\\x18\\x08\n  \\x01(\\t\\x1a\\x33\\n\\x08NamePart\\x12\\x11\\n\\tname_part\\x18\\x01\n  \\x02(\\t\\x12\\x14\\n\\x0cis_extension\\x18\\x02\n  \\x02(\\x08\\\"\\xb1\\x01\\n\\x0eSourceCodeInfo\\x12:\\n\\x08location\\x18\\x01\n  \\x03(\\x0b\\x32(.google.protobuf.SourceCodeInfo.Location\\x1a\\x63\\n\\x08Location\\x12\\x10\\n\\x04path\\x18\\x01\n  \\x03(\\x05\\x42\\x02\\x10\\x01\\x12\\x10\\n\\x04span\\x18\\x02\n  \\x03(\\x05\\x42\\x02\\x10\\x01\\x12\\x18\\n\\x10leading_comments\\x18\\x03\n  \\x01(\\t\\x12\\x19\\n\\x11trailing_comments\\x18\\x04\n  \\x01(\\tB)\\n\\x13\\x63om.google.protobufB\\x10\\x44\\x65scriptorProtosH\\x01')<\/p>\n  \n  <p>TypeError: <strong>init<\/strong>() got an unexpected keyword argument 'syntax'<\/p>\n<\/blockquote>\n\n<p>There isn't much on Google. Please help.<\/p>\n","answers":[{"AnswerId":"35203808","CreationDate":"2016-02-04T14:35:07.503","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<p>I'm using <a href=\"http:\/\/brew.sh\/\" rel=\"nofollow\">brew<\/a> and pip and things work out fine, try a fresh install for protobuf 2.6.0 as follows :<\/p>\n\n<pre><code># First, uninstall protobuf\n# Then let's install protobuf 2.6.0 for Mac\nbrew install homebrew\/versions\/protobuf260\n# And install the corresponding python library version\npip install protobuf==2.6.0\n<\/code><\/pre>\n\n<p>I used 2.6.0 here because 2.6.1 seems not yet available on brew.<\/p>\n"}]}
{"QuestionId":31316455,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2015-07-09T11:44:03.953","AcceptedAnswerId":31318434.0,"OwnerUserId":2889087.0,"Title":"Deconvolutional autoencoder in theano","Body":"<p>I'm new to theano and trying to use the examples <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">convolutional network<\/a> and <a href=\"http:\/\/deeplearning.net\/tutorial\/dA.html\" rel=\"nofollow\">denoising autoencoder<\/a> to make a denoising convolutional network. I am currently struggling with how to make W', the reverse weights. <a href=\"http:\/\/people.idsia.ch\/~masci\/papers\/2011_icann.pdf\" rel=\"nofollow\">In this paper<\/a> they use tied weights for W' that are flipped in both dimensions.<\/p>\n\n<p>I'm currently working on a 1d signal, so my image shape is (batch_size, 1, 1, 1000) and filter\/W size is (num_kernels, 1, 1, 10)  for example. The output of the convolution is then (batch_size, num_kernels, 1, 991).\nSince I want to W' to be just the flipped in 2 dimensions (or 1d in my case), I'm tempted to do this<\/p>\n\n<pre><\/pre>\n\n<p>where I reverse flip it in the relevant dimension and repeat those weights so that they are the same dimension as the feature maps from the hidden layer.<\/p>\n\n<p>With this setup, do I only have to get the gradients for W to update or should W_prime also be a part of the grad computation?<\/p>\n\n<p>When I do it like this, the MSE drops a lot after the first minibatch and then stops changing. Using cross entropy gives NaN from the first iteration. I don't know if that is related to this issue or if it's one of many other potential bugs I have in my code.<\/p>\n","answers":[{"AnswerId":"31318434","CreationDate":"2015-07-09T13:07:05.673","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I can't comment on the validity of your <code>W_prime<\/code> approach but I can say that you only need to compute the gradient of the cost with respect to each of the original shared variables. Your <code>W_prime<\/code> is a symbolic function of <code>W<\/code>, not a shared variable itself so you don't need to compute gradients with respect to <code>W_prime<\/code>.<\/p>\n\n<p>Whenever you get NaNs, the first thing to try is to reduce the size of the learning rate.<\/p>\n"}]}
{"QuestionId":31319023,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-07-09T13:31:02.553","AcceptedAnswerId":31319619.0,"OwnerUserId":133374.0,"Title":"Theano advanced indexing for tensor, shared index","Body":"<p>I have a tensor  with .<\/p>\n\n<p>And I have a tensor  with  where the values are label indices, i.e. for the third dimension in .<\/p>\n\n<p>Now I want to get a tensor  with  where the third dimension is the index in . Basically<\/p>\n\n<pre><\/pre>\n\n<p>for all .<\/p>\n\n<p>How can I achieve this?<\/p>\n\n<p>A similar problem with solution was posted <a href=\"https:\/\/stackoverflow.com\/a\/31043630\/133374\">here<\/a>.<\/p>\n\n<p>The solution there, if I understand correctly, would be:<\/p>\n\n<pre><\/pre>\n\n<p>But that doesn't seem to work. I get:\n.<\/p>\n\n<p>Also, isn't the creation of the temporal  a bit costly? Esp when I try to workaround by really making it a full dense integer array. There should be a better way.<\/p>\n\n<p>Maybe ? But as far as I understand, that doesn't parallelize the code, so this is also not a solution.<\/p>\n","answers":[{"AnswerId":"31319619","CreationDate":"2015-07-09T13:53:18.760","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>This works for me:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nmax_time, num_batches, num_labels = 3, 4, 6\nmax_seq_len = 5\n\nprobs_ = np.arange(max_time * num_batches * num_labels).reshape(\n    max_time, num_batches, num_labels)\n\ntargets_ = np.arange(num_batches * max_seq_len).reshape(max_seq_len, \n    num_batches) % (num_batches - 1)  # mix stuff up\n\nprobs, targets = map(theano.shared, (probs_, targets_))\n\nprint probs_\nprint targets_\n\nprobs_y = probs[:, T.arange(targets.shape[1])[:, np.newaxis], targets.T]\n\nprint probs_y.eval()\n<\/code><\/pre>\n\n<p>Above used a transposed version of your indices. Your exact proposition also works<\/p>\n\n<pre><code>probs_y2 = probs[:, T.arange(targets.shape[1])[np.newaxis, :], targets]\n\nprint probs_y2.eval()\nprint (probs_y2.dimshuffle(0, 2, 1) - probs_y).eval()\n<\/code><\/pre>\n\n<p>So maybe your problem is somewhere else.<\/p>\n\n<p>As for speed, I am at a loss as to what could be faster than this. <code>map<\/code>, which is a specialization of <code>scan<\/code> almost certainly is not. I do not know to what extent the <code>arange<\/code> is actually built rather than simply iterated over.<\/p>\n"}]}
{"QuestionId":31324404,"AnswerCount":0,"Tags":"<python><c++><caffe>","CreationDate":"2015-07-09T17:30:18.727","AcceptedAnswerId":null,"OwnerUserId":1452257.0,"Title":"Preventing C++ error check from stopping Python","Body":"<p>I'm using <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">caffe<\/a> for training convolutional neural networks. Sometimes the training will diverge, which caffe will find using <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1479\" rel=\"nofollow\">one of the error checks<\/a>. However, this check will stop both the C++ and Python scripts. How can I catch that C++ has stopped in Python (response in log is:  with the previous line refering to the previously mentioned error check) and continue the Python script?<\/p>\n","answers":[]}
{"QuestionId":31324633,"AnswerCount":1,"Tags":"<lua><parallel-processing><neural-network><torch>","CreationDate":"2015-07-09T17:42:46.753","AcceptedAnswerId":31403876.0,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, how to correctly implement minibatch training in a siamese neural network?","Body":"<p>I'm still working on my implementation of a siamese neural network in Torch, as mentioned in some of my previous questions.\nI finally got a good working implementation of it, but now I'd like to add a mini-batch training. That is, I would like to train the siamese neural network with a set of training elements, instead of using just one.<\/p>\n\n<p>Unfortunately, my implementation for 2 minibatches does not work. There's a problem in the back-propagation of the error, that I cannot solve.\nHere's the main architecture:<\/p>\n\n<pre><\/pre>\n\n<p>I've an upper neural network, put together with a lower neural network. They all are insereted into a parallel table. This parallel table is then inserted into a perceptron\nThe same is made for a second parallel table.\nThen the two parallel-table-perceptrons are put together into a general parallel table, that is inserted in a general percepron.<\/p>\n\n<p>I think this architecture is right, but I'm missing something with the gradient_update function.<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>The problem comes with the call to backwards() function.\nPossibly there's a problem in the dimensions...<\/p>\n\n<p>Do you have any ideas on how to solve this?<\/p>\n","answers":[{"AnswerId":"31403876","CreationDate":"2015-07-14T10:19:14.000","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>The problem comes with the call to backwards() function. Possibly there's a problem in the dimensions...<\/p>\n<\/blockquote>\n\n<p>Technically speaking regarding the structure of <code>perceptron_general<\/code> when you perform a backward the 2nd argument (= <code>gradOutput<\/code>) should be <strong>a table made of 2 x 1D tensors<\/strong> (i.e. one <code>gradOutput<\/code> per branch of your top parallel table) which gives something like:<\/p>\n\n<pre><code>gradientWrtOutput = {\n    torch.Tensor{realTarget[1]},\n    torch.Tensor{realTarget[2]}\n}\n<\/code><\/pre>\n\n<p><em>Note: right after there is another error within your main training loop.<\/em><\/p>\n"}]}
{"QuestionId":31324739,"AnswerCount":2,"Tags":"<python><c++><neural-network><deep-learning><caffe>","CreationDate":"2015-07-09T17:48:20.607","AcceptedAnswerId":31847179.0,"OwnerUserId":1452257.0,"Title":"Finding gradient of a Caffe conv-filter with regards to input","Body":"<p>I need to find the gradient with regards to the input layer for a single convolutional filter in a convolutional neural network (CNN) as a way to <a href=\"http:\/\/research.google.com\/pubs\/pub38115.html\" rel=\"nofollow noreferrer\">visualize the filters<\/a>.<br>\nGiven a trained network in the Python interface of <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow noreferrer\">Caffe<\/a> such as the one in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow noreferrer\">this example<\/a>, how can I then find the gradient of a conv-filter with respect to the data in the input layer?<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>Based on the <a href=\"https:\/\/stackoverflow.com\/a\/31349941\/1714410\">answer by cesans<\/a>, I added the code below. The dimensions of my input layer is . My first conv-layer, , has 11 filters with a size of , resulting in the dimensions .<\/p>\n\n<pre><\/pre>\n\n<p>As you can see from the output, the dimensions of the arrays returned by  are equal to the dimensions of my layers in Caffe. After some testing I've found that this output is the gradients of the loss with regards to respectively the  layer and the  layer.<\/p>\n\n<p>However, my question was how to find the gradient of a single conv-filter with respect to the data in the input layer, which is something else. How can I achieve this?<\/p>\n","answers":[{"AnswerId":"31349941","CreationDate":"2015-07-10T20:42:29.847","ParentId":null,"OwnerUserId":"1899150","Title":null,"Body":"<p>You can get the gradients in terms of any layer when you run the <code>backward()<\/code> pass. Just specify the list of layers when calling the function. To show the gradients in terms of the data layer:<\/p>\n\n<pre><code>net.forward()\ndiffs = net.backward(diffs=['data', 'conv1'])`\ndata_point = 16\nplt.imshow(diffs['data'][data_point].squeeze())\n<\/code><\/pre>\n\n<p>In some cases you may want to force all layers to carry out backward, look at the <code>force_backward<\/code> parameter of the model.<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto<\/a><\/p>\n"},{"AnswerId":"31847179","CreationDate":"2015-08-06T05:02:05.603","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe net juggles two \"streams\" of numbers.<br>\nThe first is the data \"stream\": images and labels pushed through the net. As these inputs progress through the net they are converted into high-level representation and eventually into class probabilities vectors (in classification tasks).<br>\nThe second \"stream\" holds the parameters of the different layers, the weights of the convolutions, the biases etc. These numbers\/weights are changed and learned during the train phase of the net.<\/p>\n\n<p>Despite the fundamentally different role these two \"streams\" play, caffe nonetheless use the same data structure, <code>blob<\/code>, to store and manage them.<br>\nHowever, for each layer there are two <strong>different<\/strong> blobs vectors one for each stream.<\/p>\n\n<p>Here's an example that I hope would clarify:<\/p>\n\n<pre><code>import caffe\nsolver = caffe.SGDSolver( PATH_TO_SOLVER_PROTOTXT )\nnet = solver.net\n<\/code><\/pre>\n\n<p>If you now look at<\/p>\n\n<pre><code>net.blobs\n<\/code><\/pre>\n\n<p>You will see a dictionary storing a \"caffe blob\" object for each layer in the net. Each blob has storing room for both data and gradient<\/p>\n\n<pre><code>net.blobs['data'].data.shape    # &gt;&gt; (32, 3, 224, 224)\nnet.blobs['data'].diff.shape    # &gt;&gt; (32, 3, 224, 224)\n<\/code><\/pre>\n\n<p>And for a convolutional layer:<\/p>\n\n<pre><code>net.blobs['conv1\/7x7_s2'].data.shape    # &gt;&gt; (32, 64, 112, 112)\nnet.blobs['conv1\/7x7_s2'].diff.shape    # &gt;&gt; (32, 64, 112, 112)\n<\/code><\/pre>\n\n<p><code>net.blobs<\/code> holds the first data stream, it's shape matches that of the input images up to the resulting class probability vector.<\/p>\n\n<p>On the other hand, you can see another member of <code>net<\/code><\/p>\n\n<pre><code>net.layers\n<\/code><\/pre>\n\n<p>This is a caffe vector storing the parameters of the different layers.<br>\nLooking at the first layer (<code>'data'<\/code> layer):<\/p>\n\n<pre><code>len(net.layers[0].blobs)    # &gt;&gt; 0\n<\/code><\/pre>\n\n<p>There are no parameters to store for an input layer.<br>\nOn the other hand, for the first convolutional layer<\/p>\n\n<pre><code>len(net.layers[1].blobs)    # &gt;&gt; 2\n<\/code><\/pre>\n\n<p>The net stores one blob for the filter weights and another for the constant bias. Here they are<\/p>\n\n<pre><code>net.layers[1].blobs[0].data.shape  # &gt;&gt; (64, 3, 7, 7)\nnet.layers[1].blobs[1].data.shape  # &gt;&gt; (64,)\n<\/code><\/pre>\n\n<p>As you can see, this layer performs 7x7 convolutions on 3-channel input image and has 64 such filters.<\/p>\n\n<p>Now, how to get the gradients? well, as you noted<\/p>\n\n<pre><code>diffs = net.backward(diffs=['data','conv1\/7x7_s2'])\n<\/code><\/pre>\n\n<p>Returns the gradients of the <em>data<\/em> stream. We can verify this by<\/p>\n\n<pre><code>np.all( diffs['data'] == net.blobs['data'].diff )  # &gt;&gt; True\nnp.all( diffs['conv1\/7x7_s2'] == net.blobs['conv1\/7x7_s2'].diff )  # &gt;&gt; True\n<\/code><\/pre>\n\n<p>(<strong>TL;DR<\/strong>) You want the gradients of the parameters, these are stored in the <code>net.layers<\/code> with the parameters:<\/p>\n\n<pre><code>net.layers[1].blobs[0].diff.shape # &gt;&gt; (64, 3, 7, 7)\nnet.layers[1].blobs[1].diff.shape # &gt;&gt; (64,)\n<\/code><\/pre>\n\n<hr>\n\n<p>To help you map between the names of the layers and their indices into <code>net.layers<\/code> vector, you can use <code>net._layer_names<\/code>.  <\/p>\n\n<hr>\n\n<p><strong>Update<\/strong> regarding the use of gradients to visualize filter responses:<br>\nA gradient is normally defined for a <strong>scalar<\/strong> function. The loss is a scalar, and therefore you can speak of a gradient of pixel\/filter weight with respect to the scalar loss. This gradient is a single number per pixel\/filter weight.<br>\nIf you want to get the input that results with maximal activation of a <strong>specific<\/strong> internal hidden node, you need an \"auxiliary\" net which loss is exactly a measure of the activation to the specific hidden node you want to visualize. Once you have this auxiliary net, you can start from an arbitrary input and change this input based on the gradients of the auxilary loss to the input layer: <\/p>\n\n<pre><code>update = prev_in + lr * net.blobs['data'].diff\n<\/code><\/pre>\n"}]}
{"QuestionId":31326015,"AnswerCount":7,"Tags":"<cuda><computer-vision><caffe><conv-neural-network><cudnn>","CreationDate":"2015-07-09T18:58:39.783","AcceptedAnswerId":31349250.0,"OwnerUserId":3785114.0,"Title":"How to verify CuDNN installation?","Body":"<p>I have searched many places but ALL I get is HOW to install it, not how to verify that it is installed. I can verify my NVIDIA driver is installed, and that CUDA is installed, but I don't know how to verify CuDNN is installed. Help will be much appreciated, thanks!<\/p>\n\n<p>PS.<br>\nThis is for a caffe implementation. Currently everything is working without CuDNN enabled.<\/p>\n","answers":[{"AnswerId":"46200018","CreationDate":"2017-09-13T14:20:49.950","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>You first need to find the installed cudnn file and then parse this file. To find the file, you can use:<\/p>\n\n<pre><code>whereis cudnn.h\nCUDNN_H_PATH=$(whereis cudnn.h)\n<\/code><\/pre>\n\n<p>If that doesn't work, see \"Redhat distributions\" below.<\/p>\n\n<p>Once you find this location you can then do the following (replacing <code>${CUDNN_H_PATH}<\/code> with the path):<\/p>\n\n<pre><code>cat ${CUDNN_H_PATH} | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<p>The result should look something like this:<\/p>\n\n<pre><code>#define CUDNN_MAJOR 7\n#define CUDNN_MINOR 5\n#define CUDNN_PATCHLEVEL 0\n--\n#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)\n<\/code><\/pre>\n\n<p>Which means the version is 7.5.0.<\/p>\n\n<h1>Ubuntu 18.04 (via sudo apt install nvidia-cuda-toolkit)<\/h1>\n\n<p>This method of installation installs cuda in \/usr\/include and \/usr\/lib\/cuda\/lib64, hence the file you need to look at is in \/usr\/include\/cudnn.h.<\/p>\n\n<pre><code>CUDNN_H_PATH=\/usr\/include\/cudnn.h\ncat ${CUDNN_H_PATH} | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<h1>Debian and Ubuntu<\/h1>\n\n<p>From CuDNN v5 onwards (at least when you install via <code>sudo dpkg -i &lt;library_name&gt;.deb<\/code> packages), it looks like you might need to use the following:<\/p>\n\n<pre><code>cat \/usr\/include\/x86_64-linux-gnu\/cudnn_v*.h | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>$ cat \/usr\/include\/x86_64-linux-gnu\/cudnn_v*.h | grep CUDNN_MAJOR -A 2                                                         \n#define CUDNN_MAJOR      6\n#define CUDNN_MINOR      0\n#define CUDNN_PATCHLEVEL 21\n--\n#define CUDNN_VERSION    (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)\n\n#include \"driver_types.h\"\n<\/code><\/pre>\n\n<p>indicates that CuDNN version 6.0.21 is installed.<\/p>\n\n<h1>Redhat distributions<\/h1>\n\n<p>On CentOS, I found the location of CUDA with:<\/p>\n\n<pre><code>$ whereis cuda\ncuda: \/usr\/local\/cuda\n<\/code><\/pre>\n\n<p>I then used the procedure about on the cudnn.h file that I found from this location:<\/p>\n\n<pre><code>$ cat \/usr\/local\/cuda\/include\/cudnn.h | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n"},{"AnswerId":"48928917","CreationDate":"2018-02-22T13:41:41.180","ParentId":null,"OwnerUserId":"9277085","Title":null,"Body":"<p>Run <code>.\/mnistCUDNN<\/code> in <code>\/usr\/src\/cudnn_samples_v7\/mnistCUDNN<\/code>\n<p>Here is an example:<\/p>\n\n<pre><code>cudnnGetVersion() : 7005 , CUDNN_VERSION from cudnn.h : 7005 (7.0.5)\nHost compiler version : GCC 5.4.0\nThere are 1 CUDA capable devices on your machine :\ndevice 0 : sms 30  Capabilities 6.1, SmClock 1645.0 Mhz, MemSize (Mb) 24446, MemClock 4513.0 Mhz, Ecc=0,    boardGroupID=0\nUsing device 0\n<\/code><\/pre>\n"},{"AnswerId":"49590972","CreationDate":"2018-03-31T18:20:34.170","ParentId":null,"OwnerUserId":"843442","Title":null,"Body":"<p>When installing on ubuntu via <code>.deb<\/code> you can use <code>sudo apt search cudnn | grep installed<\/code><\/p>\n"},{"AnswerId":"36978616","CreationDate":"2016-05-02T08:56:47.373","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>The installation of CuDNN is just copying some files. Hence to check if CuDNN is installed (and which version you have), you only need to check those files.<\/p>\n\n<h2>Install CuDNN<\/h2>\n\n<p>Step 1: Register an nvidia developer account and <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"noreferrer\">download cudnn here<\/a> (about 80 MB). You might need <code>nvcc --version<\/code> to get your cuda version.<\/p>\n\n<p>Step 2: Check where your cuda installation is. For most people, it will be <code>\/usr\/local\/cuda\/<\/code>. You can check it with <code>which nvcc<\/code>.<\/p>\n\n<p>Step 3: Copy the files:<\/p>\n\n<pre><code>$ cd folder\/extracted\/contents\n$ sudo cp include\/cudnn.h \/usr\/local\/cuda\/include\n$ sudo cp lib64\/libcudnn* \/usr\/local\/cuda\/lib64\n$ sudo chmod a+r \/usr\/local\/cuda\/lib64\/libcudnn*\n<\/code><\/pre>\n\n<h2>Check version<\/h2>\n\n<p>You might have to adjust the path. See step 2 of the installation.<\/p>\n\n<pre><code>$ cat \/usr\/local\/cuda\/include\/cudnn.h | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<h2>Notes<\/h2>\n\n<p>When you get an error like<\/p>\n\n<pre><code>F tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:427] could not set cudnn filter descriptor: CUDNN_STATUS_BAD_PARAM\n<\/code><\/pre>\n\n<p>with TensorFlow, you might consider using CuDNN v4 instead of v5.<\/p>\n\n<p><strong>Ubuntu users who installed it via <code>apt<\/code><\/strong>: <a href=\"https:\/\/askubuntu.com\/a\/767270\/10425\">https:\/\/askubuntu.com\/a\/767270\/10425<\/a><\/p>\n"},{"AnswerId":"31349250","CreationDate":"2015-07-10T19:56:28.280","ParentId":null,"OwnerUserId":"1899150","Title":null,"Body":"<p>Installing CuDNN just involves placing the files in the CUDA directory. If you have specified the routes and the CuDNN option correctly while installing caffe it will be compiled with CuDNN.<\/p>\n\n<p>You can check that using <code>cmake<\/code>. Create a directory <code>caffe\/build<\/code> and run <code>cmake ..<\/code> from there. If the configuration is correct you will see these lines:<\/p>\n\n<pre><code>-- Found cuDNN (include: \/usr\/local\/cuda-7.0\/include, library: \/usr\/local\/cuda-7.0\/lib64\/libcudnn.so)\n\n-- NVIDIA CUDA:\n--   Target GPU(s)     :   Auto\n--   GPU arch(s)       :   sm_30\n--   cuDNN             :   Yes\n<\/code><\/pre>\n\n<p>If everything is correct just run the <code>make<\/code> orders to install caffe from there.<\/p>\n"},{"AnswerId":"47436840","CreationDate":"2017-11-22T14:12:13.637","ParentId":null,"OwnerUserId":"5468983","Title":null,"Body":"<p><strong>To check installation of CUDA, run below command<\/strong>, if it\u2019s installed properly then below command will not throw any error and will print correct version of library.<\/p>\n\n<pre><code>function lib_installed() { \/sbin\/ldconfig -N -v $(sed 's\/:\/ \/' &lt;&lt;&lt; $LD_LIBRARY_PATH) 2&gt;\/dev\/null | grep $1; }\nfunction check() { lib_installed $1 &amp;&amp; echo \"$1 is installed\" || echo \"ERROR: $1 is NOT installed\"; }\ncheck libcuda\ncheck libcudart\n<\/code><\/pre>\n\n<p><strong>To check installation of CuDNN, run below command<\/strong>, if CuDNN is installed properly then you will not get any error.<\/p>\n\n<pre><code>function lib_installed() { \/sbin\/ldconfig -N -v $(sed 's\/:\/ \/' &lt;&lt;&lt; $LD_LIBRARY_PATH) 2&gt;\/dev\/null | grep $1; }\nfunction check() { lib_installed $1 &amp;&amp; echo \"$1 is installed\" || echo \"ERROR: $1 is NOT installed\"; }\ncheck libcudnn \n<\/code><\/pre>\n\n<p><strong>OR<\/strong><\/p>\n\n<p>you can run below command from any directory <\/p>\n\n<pre><code>nvcc -V\n<\/code><\/pre>\n\n<p>it should give output something like this<\/p>\n\n<pre><code> nvcc: NVIDIA (R) Cuda compiler driver\n Copyright (c) 2005-2016 NVIDIA Corporation\n Built on Tue_Jan_10_13:22:03_CST_2017\n Cuda compilation tools, release 8.0, V8.0.61\n<\/code><\/pre>\n"},{"AnswerId":"51202754","CreationDate":"2018-07-06T03:48:53.263","ParentId":null,"OwnerUserId":"207661","Title":null,"Body":"<p><strong>Getting cuDNN Version [Linux]<\/strong><\/p>\n\n<p>Use following to find path for cuDNN:<\/p>\n\n<pre><code>$ whereis cuda\ncuda: \/usr\/local\/cuda\n<\/code><\/pre>\n\n<p>Then use this to get version from header file,<\/p>\n\n<pre><code>$ cat \/usr\/local\/cuda\/include\/cudnn.h | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<p><strong>Getting cuDNN Version [Windows]<\/strong><\/p>\n\n<p>Use following to find path for cuDNN:<\/p>\n\n<pre><code>C:\\&gt;where cudnn*\nC:\\Program Files\\cuDNN6\\cuda\\bin\\cudnn64_6.dll\n<\/code><\/pre>\n\n<p>Then use this to dump version from header file,<\/p>\n\n<pre><code>type \"%PROGRAMFILES%\\cuDNN6\\cuda\\include\\cudnn.h\" | findstr \"CUDNN_MAJOR CUDNN_MINOR CUDNN_PATCHLEVEL\"\n<\/code><\/pre>\n\n<p><strong>Getting CUDA Version<\/strong><\/p>\n\n<p>This works on Linux as well as Windows:<\/p>\n\n<pre><code>nvcc --version\n<\/code><\/pre>\n"}]}
{"QuestionId":31342134,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-07-10T13:33:18.847","AcceptedAnswerId":31343583.0,"OwnerUserId":133374.0,"Title":"argmax result as a subtensor","Body":"<p>I want to use the argmax with kept dimensions as a subtensor. I have:<\/p>\n\n<pre><\/pre>\n\n<p>And I want to set those values to zero in . I.e. I need to use . To use that, I need to specify the subtensor  of  at  but I'm not exactly sure how that looks like.  is wrong for multiple dimensions.<\/p>\n\n<p>This should hold:<\/p>\n\n<pre><\/pre>\n\n<p>In the end, I want to do:<\/p>\n\n<pre><\/pre>\n\n<p>My current solution:<\/p>\n\n<pre><\/pre>\n\n<p>However,  does not work.<\/p>\n","answers":[{"AnswerId":"31343583","CreationDate":"2015-07-10T14:38:22.757","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You need to use Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#indexing\" rel=\"nofollow\">advanced indexing features<\/a> which, unfortunately, differ from <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#advanced-indexing\" rel=\"nofollow\">numpy's advanced indexing<\/a>.<\/p>\n\n<p>Here's an example that does what you want.<\/p>\n\n<p><strong>Update:<\/strong> Now works with parametrized axis but note that <code>axis<\/code> cannot be symbolic.<\/p>\n\n<pre><code>import numpy\n\nimport theano\nimport theano.tensor as tt\n\ntheano.config.compute_test_value = 'raise'\n\naxis = 2\n\nx = tt.tensor3()\nx.tag.test_value = numpy.array([[[3, 2, 6], [5, 1, 4]], [[2, 1, 6], [6, 1, 5]]],\n                               dtype=theano.config.floatX)\n\n# Identify the largest value in each row\nx_argmax = tt.argmax(x, axis=axis, keepdims=True)\n\n# Construct a row of indexes to the length of axis\nindexes = tt.arange(x.shape[axis]).dimshuffle(\n    *(['x' for dim1 in xrange(axis)] + [0] + ['x' for dim2 in xrange(x.ndim - axis - 1)]))\n\n# Create a binary mask indicating where the maximum values appear\nmask = tt.eq(indexes, x_argmax)\n\n# Alter the original matrix only at the places where the maximum values appeared\nx_prime = tt.set_subtensor(x[mask.nonzero()], 0)\n\nprint x_prime.tag.test_value\n<\/code><\/pre>\n"}]}
{"QuestionId":31343525,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-07-10T14:36:08.700","AcceptedAnswerId":31374909.0,"OwnerUserId":395857.0,"Title":"How to generate the predicted labels in Caffe through the CLI?","Body":"<p>I trained a neural network model using Caffe:<\/p>\n\n<pre><\/pre>\n\n<p>I then scored the learned model on the validation set:<\/p>\n\n<pre><\/pre>\n\n<p>But how to get the labels predicted by the trained neural network model in Caffe?<\/p>\n\n<hr>\n\n<p>I know I can use the Python or Matlab bindings for that purpose, but I am curious to know whether we can get the predicted labels in Caffe directly through the command line interface.<\/p>\n\n<p>It does not seem to be mentioned in the <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html\" rel=\"nofollow\">official Caffe's tutorial on interfaces<\/a>, and looking at  's help didn't help:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31374909","CreationDate":"2015-07-13T03:14:08.703","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>If you don't want to go through Python, you can add a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/tutorial\/layers.md#hdf5-output\" rel=\"nofollow\">HDF5_OUTPUT<\/a> layer: it will save the predicted outputs in an HDF5 file.<\/p>\n\n<p>Otherwise if you feel like going in the code, you could print or save\u00a0<code>bottom_data_vector[k].second<\/code>\u00a0at around\u00a0<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/accuracy_layer.cpp#L74\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/accuracy_layer.cpp#L74<\/a><\/p>\n"}]}
{"QuestionId":31345255,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe>","CreationDate":"2015-07-10T15:57:19.757","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"make pycaffe -> \"fatal error: cublas_v2.h: No such file or directory\"","Body":"<p>I compiled  and I know trying to compile . When I run  in the  root folder, I get:<\/p>\n\n<pre><\/pre>\n\n<p>How to fix that?<\/p>\n","answers":[{"AnswerId":"31345318","CreationDate":"2015-07-10T15:59:59.970","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>If you don't plan to use the GPU, you can circumvent the issue by uncommenting <code>CPU_ONLY := 1<\/code> in Makefile.config:<\/p>\n\n<pre><code># CPU-only switch (uncomment to build without GPU support).\nCPU_ONLY := 1\n<\/code><\/pre>\n"}]}
{"QuestionId":31352311,"AnswerCount":2,"Tags":"<python><loops><dictionary><clone><theano>","CreationDate":"2015-07-11T00:47:06.113","AcceptedAnswerId":31352379.0,"OwnerUserId":5009112.0,"Title":"Creating an alternating dictionary with a loop in Python","Body":"<p>I'm trying to create a dictionary with a loop, with alternating entries, although the entries not necessarily need to be alternating, as long as they are all in the dictionary, I just need the simplest solution to get them all in one dictionary. A simple example of what I'm trying to achieve:<\/p>\n\n<p>Normally, for creating a dictionary with a loop I would do this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I'm trying the follwing:<\/p>\n\n<pre><\/pre>\n\n<p>The desired output is:<\/p>\n\n<pre><\/pre>\n\n<p>Or, this output would also be fine, as long as all elements are in the dictionary (order does not matter):<\/p>\n\n<pre><\/pre>\n\n<p>The context is that I'm using <\/p>\n\n<pre><\/pre>\n\n<p>where all items you want to replace must be given in one dictionary. <\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"31352379","CreationDate":"2015-07-11T00:58:36.363","ParentId":null,"OwnerUserId":"2617068","Title":null,"Body":"<p>Comprehensions are cool, but not strictly necessary. Also, standard dictionaries cannot have duplicate keys. There are <a href=\"https:\/\/stackoverflow.com\/questions\/10664856\/make-dictionary-with-duplicate-keys-in-python\">data structures for that<\/a>, but you can also try having a <code>list<\/code> of values for that key:<\/p>\n\n<pre><code>d = {}\nd[8] = []\nfor i in range(3):\n    d[i] = 9\n    d[8].append(i)\n<\/code><\/pre>\n\n<p>&nbsp;<\/p>\n\n<pre><code>&gt;&gt;&gt; d\n{8: [0, 1, 2], 0: 9, 2: 9, 1: 9}\n<\/code><\/pre>\n\n<p>For your new example without duplicate keys:<\/p>\n\n<pre><code>d = {}\nfor i in range(3):\n    d[i] = 9\n    d[i+5] = 8\n<\/code><\/pre>\n\n<p>&nbsp;<\/p>\n\n<pre><code>&gt;&gt;&gt; d\n{0: 9, 1: 9, 2: 9, 5: 8, 6: 8, 7: 8}\n<\/code><\/pre>\n"},{"AnswerId":"31352492","CreationDate":"2015-07-11T01:18:12.530","ParentId":null,"OwnerUserId":"1013719","Title":null,"Body":"<p>You can pull this off in a single line with <code>itertools<\/code><\/p>\n\n<pre><code>dict(itertools.chain.from_iterable(((i, 9), (i+5, 8)) for i in range(3)))\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Explained:<\/strong><\/p>\n\n<p>The inner part creates a bunch of tuples<\/p>\n\n<pre><code>((i, 9), (i+5, 8)) for i in range(3)\n<\/code><\/pre>\n\n<p>which in list form expands to<\/p>\n\n<pre><code>[((0, 9), (5, 8)), ((1, 9), (6, 8)), ((2, 9), (7, 8))]\n<\/code><\/pre>\n\n<p>The <code>chain.from_iterable<\/code> then flattens it by a level to produce<\/p>\n\n<pre><code>[(0, 9), (5, 8), (1, 9), (6, 8), (2, 9), (7, 8)]\n<\/code><\/pre>\n\n<p>This of course works with the <code>dict<\/code> init that takes a sequence of tuples<\/p>\n"}]}
{"QuestionId":31358451,"AnswerCount":4,"Tags":"<python><macos><compiler-errors><neural-network><caffe>","CreationDate":"2015-07-11T14:55:59.287","AcceptedAnswerId":31359361.0,"OwnerUserId":5106073.0,"Title":"make pycaffe fatal error: 'Python.h' file not found","Body":"<p>I compiled caffe on a mac running OSX 10.9.5  and I know trying to compile pycaffe. When I run make pycaffe in the caffe root folder, I get:<\/p>\n\n<pre><\/pre>\n\n<p>how can I fix this?<\/p>\n\n<p>Perhaps is something wrong with Makefile.config. How do I know what is my PYTHONPATH?<\/p>\n","answers":[{"AnswerId":"31359361","CreationDate":"2015-07-11T16:35:28.027","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Looking at the comments, I see that you use Anaconda. In <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/Makefile.config.example\" rel=\"noreferrer\"><code>Makefile.config<\/code><\/a>, you should uncomment the lines dedicated to Anaconda:<\/p>\n\n<pre><code># Anaconda Python distribution is quite popular. Include path:\n# Verify anaconda location, sometimes it's in root.\n# ANACONDA_HOME := $(HOME)\/anaconda\n# PYTHON_INCLUDE := $(ANACONDA_HOME)\/include \\\n        # $(ANACONDA_HOME)\/include\/python2.7 \\\n        # $(ANACONDA_HOME)\/lib\/python2.7\/site-packages\/numpy\/core\/include \\\n\n# We need to be able to find libpythonX.X.so or .dylib.\nPYTHON_LIB := \/usr\/lib\n# PYTHON_LIB := $(ANACONDA_HOME)\/lib\n<\/code><\/pre>\n\n<p><code>Python.h<\/code> is in <code>$(ANACONDA_HOME)\/include\/python2.7<\/code> as you can see running <code>sudo find \/ -name 'Python.h'<\/code>.<\/p>\n"},{"AnswerId":"46387511","CreationDate":"2017-09-24T07:06:03.343","ParentId":null,"OwnerUserId":"6908096","Title":null,"Body":"<p>I uncommented the below code in Makefile.config<\/p>\n\n<pre><code>PYTHON_INCLUDE := \/usr\/include\/python3.5m \\\n                 \/usr\/lib\/python3.5\/dist-packages\/numpy\/core\/include\n<\/code><\/pre>\n\n<p>Then did sudo make pycaffe.<\/p>\n\n<p>It worked.<\/p>\n"},{"AnswerId":"40734269","CreationDate":"2016-11-22T05:18:51.353","ParentId":null,"OwnerUserId":"1904943","Title":null,"Body":"<p>I just finished a tedious Caffe install on Arch Linux; hopefully my install notes (link below) will help others.<\/p>\n\n<p>While specific to my Caffe install, those notes address the \"Python.h\" install error (this Question), as well as a downstream issue mentioned in another SO question,<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/28177298\/import-caffe-error\">Import caffe error<\/a>.<\/p>\n\n<pre><code>https:\/\/stackoverflow.com\/questions\/28177298\/import-caffe-error\n<\/code><\/pre>\n\n<p>My gist file (notes):<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213\" rel=\"nofollow noreferrer\">Caffe Installation Notes<\/a><\/p>\n\n<pre><code>https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213\n<\/code><\/pre>\n"},{"AnswerId":"44726134","CreationDate":"2017-06-23T16:28:47.703","ParentId":null,"OwnerUserId":"8119677","Title":null,"Body":"<p>I met this problem too.\nI have set the <code>PYTHON_INCLUDE<\/code> PATH<\/p>\n\n<pre><code>    PYTHON_INCLUDE := $(ANACONDA_HOME)\/include \\\n    $(ANACONDA_HOME)\/include\/python2.7\n<\/code><\/pre>\n\n<p>But it still can't find the <code>Python.h<\/code><\/p>\n\n<p>So I just give the include path manually to the compiler as follows:<\/p>\n\n<pre><code>    export CPLUS_INCLUDE_PATH=\/home\/woolawren\/anaconda2\/include\/python2.7\/:$CPLUS_INCLUDE_PATH\n<\/code><\/pre>\n\n<p>if you don't use anaconda2, you can use:<\/p>\n\n<pre><code>    export CPLUS_INCLUDE_PATH=\/usr\/include\/python2.7:$CPLUS_INCLUDE_PATH\n<\/code><\/pre>\n\n<p>I have successfully done \"make pycaffe\" by doing this.<\/p>\n"}]}
{"QuestionId":31361377,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2015-07-11T20:13:34.260","AcceptedAnswerId":31362146.0,"OwnerUserId":1293964.0,"Title":"How to get value from a theano tensor variable backed by a shared variable?","Body":"<p>I have a theano tensor variable created from casting a shared variable. How can I extract the original or casted values? (I need that so I don't have to carry the original shared\/numpy values around.)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31362146","CreationDate":"2015-07-11T21:43:50.760","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p><code>get_value<\/code> only works for shared variables. <code>TensorVariables<\/code> are general expressions and thus potentially need extra input in order to be able to determine their value (Imagine you set <code>y = x + z<\/code>, where <code>z<\/code> is another tensor variable. You would need to specify <code>z<\/code> before being able to calculate <code>y<\/code>). You can either create a function to provide this input or provide it in a dictionary using the <code>eval<\/code> method.<\/p>\n\n<p>In your case, <code>y<\/code> only depends on <code>x<\/code>, so you can do<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nx = theano.shared(numpy.asarray([1, 2, 3], dtype='float32'))\ny = T.cast(x, 'int32')\ny.eval()\n<\/code><\/pre>\n\n<p>and you should see the result<\/p>\n\n<pre><code>array([1, 2, 3], dtype=int32)\n<\/code><\/pre>\n\n<p>(And in the case <code>y = x + z<\/code>, you would have to do <code>y.eval({z : 3.})<\/code>, for example)<\/p>\n"}]}
{"QuestionId":31362140,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-07-11T21:43:38.443","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Splitting train_test.prototxt into train.prototxt and test.prototxt","Body":"<p>In Caffe, when configuring a neural neutral architecture, one can either define one single train_test.prototxt, or 2 prototxt files train.prototxt and test.prototxt.<\/p>\n\n<p>For instance, in the examples, <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/hdf5_classification\" rel=\"nofollow\">hdf5_classification<\/a> uses 2 prototxt files (<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/hdf5_classification\/nonlinear_auto_train.prototxt\" rel=\"nofollow\">nonlinear_auto_train.prototxt<\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/hdf5_classification\/nonlinear_auto_test.prototxt\" rel=\"nofollow\">nonlinear_auto_test.prototxt<\/a>), while <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/mnist\" rel=\"nofollow\">mnist<\/a> uses 1 prototxt file (<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">lenet_train_test.prototxt<\/a>).<\/p>\n\n<p>What difference does that make to use 2 prototxt files instead of 1?<\/p>\n","answers":[{"AnswerId":"32092488","CreationDate":"2015-08-19T10:07:22.257","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Merging train and val .prototxt files into a single file is an attempt at reducing the number of files required to drive an experiment. See <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/LpXNE42hfG0\" rel=\"nofollow\">this<\/a> post in the caffe-users group regarding the general intent to merging files.\nI don't know if the plan is to have single file solutions and eliminate any splitting, but can imagine it be kept for advanced purposes, where the user wishes to have vary test conditions while keeping training conditions constant.<\/p>\n"}]}
{"QuestionId":31362914,"AnswerCount":1,"Tags":"<lua><machine-learning><torch>","CreationDate":"2015-07-11T23:36:46.033","AcceptedAnswerId":31370295.0,"OwnerUserId":4747870.0,"Title":"What is correct usage of DataSource with torch\/dp library","Body":"<p>I am new to both programming language lua and torch library. I am trying to get some machine learning algorithms to work ASAP. I tried to get neural nets using dp library using example <a href=\"https:\/\/github.com\/nicholas-leonard\/dp\/blob\/master\/examples\/neuralnetwork.lua\" rel=\"nofollow\">here<\/a>. But I am unable to get my dataset into the form to feed into learning algorithm. I think my best and also initial guess was to do this: <\/p>\n\n<pre><\/pre>\n\n<p>Which gives error:  <\/p>\n\n<p>Where dataset[ 1] is a torch.Tensor containing information about the data and dataset[2] is torch.Tensor of binary information about the data I would like know.<\/p>\n\n<p>Hope it is not a stupid syntax error.<\/p>\n","answers":[{"AnswerId":"31370295","CreationDate":"2015-07-12T17:00:08.670","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>Yep, this is a syntax error. Lua has no named arguments. Lua adepts <a href=\"http:\/\/www.lua.org\/pil\/5.3.html\" rel=\"nofollow\">use table to emulate such a feature<\/a>.<\/p>\n\n<p>So, try this:\n<code>dp.DataSource({train_set=train_set, test_set=test_set})\n<\/code>\nor just\n<code>dp.DataSource{train_set=train_set, test_set=test_set}\n<\/code> (you can remove parenthesis if a function has one parameter).<\/p>\n"}]}
{"QuestionId":31365592,"AnswerCount":5,"Tags":"<python><matlab><cmake><caffe><matcaffe>","CreationDate":"2015-07-12T07:44:27.903","AcceptedAnswerId":null,"OwnerUserId":2711403.0,"Title":"Cannot find -lpython2 : MatCaffe installation error","Body":"<p>During the build of MatCaffe (Matlab wrapper of Caffe), I face the following error :<\/p>\n\n<pre><\/pre>\n\n<p>On closer examination I found using following command that the following file is responsible for the above error : <\/p>\n\n<pre><\/pre>\n\n<p>It revealed to me the following :<\/p>\n\n<pre><\/pre>\n\n<p>So, I changed the corresponding -lpython2 to -lpython2.7 with the hope of resolving the problem. But to no yield.<\/p>\n\n<p>I also tried the following :<\/p>\n\n<ol>\n<li>Delete the CMakeCache.txt and doing the make. But it did not work.<\/li>\n<li><p>I edited the default CMakeLists.txt file in \/cmake-master, to change some default settings. I found that the default python version setting in the CMakeLists.txt in Caffe is 2 .<\/p>\n\n<p>\/\/Specify which python version to use\n    python_version:STRING=2.7<\/p><\/li>\n<\/ol>\n\n<p>I changed it to 2 , and repeated the whole configure-generate-make process in a fresh build folder. But to no yield. Everytime the same matlab\/build.make file shows -lpython2 , and changing that to 2.7 directly does not yield.<\/p>\n\n<ol start=\"3\">\n<li>I tried to look into matlab\/build.make file but could not find anything over there which I could directly connect with this error.<\/li>\n<\/ol>\n\n<p>Any solid help would be deeply appreciated. I use MATLAB 2014a, on Ubuntu 14.04.<\/p>\n","answers":[{"AnswerId":"44247396","CreationDate":"2017-05-29T17:00:28.913","ParentId":null,"OwnerUserId":"6322527","Title":null,"Body":"<p>I had this problem and it seems that it roots in the unability of Cmake to get different versions of python, such as \"libpython.so.1.0\". I changed my CMakeCache.txt file to \"libpython.so\" and the problem solved. It is not only for python, I had this issue with my \"cudnn\" as well, and this solution fixed that. <\/p>\n"},{"AnswerId":"31371971","CreationDate":"2015-07-12T19:59:09.637","ParentId":null,"OwnerUserId":"3440745","Title":null,"Body":"<p>According to error message and command, causes it, it seems that python libary is installed at unusual location, so <code>ld<\/code>(linker) cannot find it in its default paths. As CMake script has found headers, it should also setup <code>mex<\/code> executable for work with library itself, but for some reason it doesn't.<\/p>\n\n<p>The simplest way to make building package work is to set <code>LD_LIBRARY_PATH<\/code> to the directory, where you python library is located, and run <code>make<\/code>. If you want to fix CMake script, this <a href=\"http:\/\/www.cmake.org\/Wiki\/CMake_RPATH_handling\" rel=\"nofollow\">wiki<\/a> may help you.<\/p>\n"},{"AnswerId":"31394301","CreationDate":"2015-07-13T22:07:05.193","ParentId":null,"OwnerUserId":"2711403","Title":null,"Body":"<p>Thanks to @Tsyvarev for the answer. I found a rather simple solution. I just made a symbolic link (libpython2.so) which points to libpython2.7.so in \/usr\/lib folder. This solved the problem. libpython2.7.so was also present in \/usr\/lib so i dont think it was the issue of an unusual install.<\/p>\n"},{"AnswerId":"31472654","CreationDate":"2015-07-17T09:33:07.500","ParentId":null,"OwnerUserId":"5126916","Title":null,"Body":"<p>I had the same problem. In despair I just removed the <code>-lpython2<\/code> from <code>build-matlab\/matlab\/CMakeFiles\/matlab.dir\/build.make<\/code><\/p>\n\n<p>It did compile after that, seems that it found whatever it needed regardless.<\/p>\n"},{"AnswerId":"35286218","CreationDate":"2016-02-09T07:28:11.373","ParentId":null,"OwnerUserId":"1686769","Title":null,"Body":"<p>This is because of a bug in <code>caffe_parse_linker_libs<\/code> function in Utils.cmake which converts something like <code>\/usr\/lib\/x86_64-linux-gnu\/libpython2.7.so<\/code> to <code>-lpython2<\/code><\/p>\n\n<p>This can be fixed by <strong>replacing<\/strong> (in cmake\/Utils.cmake)<\/p>\n\n<pre><code>elseif(IS_ABSOLUTE ${lib})\n  get_filename_component(name_we ${lib} NAME_WE)\n  get_filename_component(folder  ${lib} PATH)\n\n  string(REGEX MATCH \"^lib(.*)\" __match ${name_we})\n  list(APPEND libflags -l${CMAKE_MATCH_1})\n  list(APPEND folders    ${folder})\nelse()\n<\/code><\/pre>\n\n<p><strong>with<\/strong><\/p>\n\n<pre><code>elseif(IS_ABSOLUTE ${lib})\n  get_filename_component(folder  ${lib} PATH)\n  get_filename_component(filename ${lib} NAME)\n  string(REGEX REPLACE \"\\\\.[^.]*$\" \"\" filename_without_shortest_ext ${filename})\n\n  string(REGEX MATCH \"^lib(.*)\" __match ${filename_without_shortest_ext})\n  list(APPEND libflags -l${CMAKE_MATCH_1})\n  list(APPEND folders    ${folder})\nelse()\n<\/code><\/pre>\n\n<p>The updated function correctly converts something like <code>\/usr\/lib\/x86_64-linux-gnu\/libpython2.7.so<\/code> to <code>-lpython2.7<\/code><\/p>\n"}]}
{"QuestionId":31373615,"AnswerCount":0,"Tags":"<opencv><cmake><caffe>","CreationDate":"2015-07-12T23:35:58.977","AcceptedAnswerId":null,"OwnerUserId":5013415.0,"Title":"Error in installing caffe in Ubuntu 14.04 LTS","Body":"<p>Having already required dependencies for caffe installation:\nin cmake step, when I use:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>which in Libraries it has detected wrong path instead of \/home\/majid\/anaconda\/lib\/libpython2.7.so<\/p>\n\n<p>I tried to force to use the correct path by:<\/p>\n\n<pre><\/pre>\n\n<p>but after:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting a couple of time this warning:<\/p>\n\n<pre><\/pre>\n\n<p>in the end it stops by this error: <\/p>\n\n<pre><\/pre>\n\n<p>while when installing opencv I used this cmake:<\/p>\n\n<pre><\/pre>\n\n<p>in which it is: LIBTIFF=ON. <\/p>\n\n<p>after making opencv I set the packages' path by:<\/p>\n\n<pre><\/pre>\n\n<p>and also:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know why I am having this problem from x86-64-linux-gnu while I have the same libopencv-highgui.so in the \/home\/majid\/opencv\/lib<\/p>\n\n<p>I really appreciate it if someone could help me out in this matter.<\/p>\n\n<p>the result of forced cmake:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":31379971,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-07-13T09:39:02.387","AcceptedAnswerId":31384142.0,"OwnerUserId":133374.0,"Title":"`uniq` for 2D Theano tensor","Body":"<p>I have this Numpy code:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I want to extend this to support 2D arrays and make it use Theano. It should be fast on the GPU.<\/p>\n\n<p>I will get an array with multiple sequences as multiple batches in the format (time,batch), and a  which specifies indirectly the length of each sequence.<\/p>\n\n<p>My current try:<\/p>\n\n<pre><\/pre>\n\n<p>How to construct  directly?<\/p>\n\n<p>I would do something like  but I'm not exactly sure how to express that.<\/p>\n","answers":[{"AnswerId":"31384142","CreationDate":"2015-07-13T13:01:13.390","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's a quick answer that only addresses part of your task:<\/p>\n\n<pre><code>def compile_theano_uniq(x):\n    diffs = x[1:] - x[:-1]\n    diffs = tt.concatenate([tt.ones_like([x[0]], dtype=diffs.dtype), diffs])\n    y = diffs.nonzero_values()\n    return theano.function(inputs=[x], outputs=y)\n\ntheano_uniq = compile_theano_uniq(tt.vector(dtype='int32'))\n<\/code><\/pre>\n\n<p>The key is <code>nonzero_values()<\/code>.<\/p>\n\n<p><strong>Update:<\/strong> I can't imagine any way to do this without using <code>theano.scan<\/code>. To be clear, and using 0 as padding, I'm assuming that given the input<\/p>\n\n<pre><code>1 1 2 3 3 4 0\n1 2 2 2 3 3 4\n1 2 3 4 5 0 0\n<\/code><\/pre>\n\n<p>you would want the output to be<\/p>\n\n<pre><code>1 2 3 4 0 0 0\n1 2 3 4 0 0 0\n1 2 3 4 5 0 0\n<\/code><\/pre>\n\n<p>or even<\/p>\n\n<pre><code>1 2 3 4 0\n1 2 3 4 0\n1 2 3 4 5\n<\/code><\/pre>\n\n<p>You could identify the indexes of the items you want to keep without using scan. Then either a new tensor needs to be constructed from scratch or the values you want to keep some how moved to make the sequences contiguous. Neither approaches seem feasible without <code>theano.scan<\/code>.<\/p>\n"}]}
{"QuestionId":31385427,"AnswerCount":1,"Tags":"<python-2.7><caffe><deep-dream>","CreationDate":"2015-07-13T13:58:53.687","AcceptedAnswerId":null,"OwnerUserId":1339643.0,"Title":"How to train a caffe model?","Body":"<p>Has anyone successfully trained a caffe model? I have a training ready image set that I would like to use to create a caffe model for use with Google's Deep Dream. <\/p>\n\n<p>The only resources I've been able to find on how to train a model are these:<br>\n<a href=\"https:\/\/github.com\/jgoode21\/caffe-oxford102\" rel=\"noreferrer\">ImageNet Tutorial<\/a><br>\nEDIT:  Here's another, but it's not creating a deploy.prototxt file. When I try to use one from another model it \"works\" but isn't correct.<br>\n<a href=\"https:\/\/github.com\/jgoode21\/caffe-oxford102\" rel=\"noreferrer\">caffe-oxford 102<\/a><br>\nCan anyone point me in the right direction to training my own model?<\/p>\n","answers":[{"AnswerId":"31659332","CreationDate":"2015-07-27T17:18:24.427","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>I have written a simple example to train a Caffe model on the Iris data set in Python. It also gives the predicted outputs given some user-defined inputs. The network as well as the solver settings need some more tuning but I just wanted to have some code skeleton to get started. Feel free to edit to improve.<\/p>\n\n<p>(<a href=\"https:\/\/github.com\/Franck-Dernoncourt\/caffe_demos\" rel=\"nofollow noreferrer\">GitHub repository<\/a>)<\/p>\n\n<p><code>iris_tuto.py<\/code><\/p>\n\n<pre><code>'''\n\nRequirements:\n - Caffe (script to install Caffe and pycaffe on a new Ubuntu 14.04 LTS x64 or Ubuntu 14.10 x64. \n   CPU only, multi-threaded Caffe. https:\/\/stackoverflow.com\/a\/31396229\/395857)\n - sudo pip install pydot\n - sudo apt-get install -y graphviz\n\nInteresting resources on Caffe:\n - https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\n - http:\/\/nbviewer.ipython.org\/github\/joyofdata\/joyofdata-articles\/blob\/master\/deeplearning-with-caffe\/Neural-Networks-with-Caffe-on-the-GPU.ipynb\n\nInteresting resources on Iris with ANNs:\n - iris data set test bed: http:\/\/deeplearning4j.org\/iris-flower-dataset-tutorial.html\n - http:\/\/se.mathworks.com\/help\/nnet\/examples\/iris-clustering.html\n - http:\/\/lab.fs.uni-lj.si\/lasin\/wp\/IMIT_files\/neural\/doc\/seminar8.pdf\n\nSynonyms:\n - output = label = target\n - input = feature \n'''\n\nimport subprocess\nimport platform\nimport copy\n\nfrom sklearn.datasets import load_iris\nimport sklearn.metrics \nimport numpy as np\nfrom sklearn.cross_validation import StratifiedShuffleSplit\nimport matplotlib.pyplot as plt\nimport h5py\nimport caffe\nimport caffe.draw\n\n\ndef load_data():\n    '''\n    Load Iris Data set\n    '''\n    data = load_iris()\n    print(data.data)\n    print(data.target)\n    targets = np.zeros((len(data.target), 3))\n    for count, target in enumerate(data.target):\n        targets[count][target]= 1    \n    print(targets)\n\n    new_data = {}\n    #new_data['input'] = data.data\n    new_data['input'] = np.reshape(data.data, (150,1,1,4))\n    new_data['output'] = targets\n    #print(new_data['input'].shape)\n    #new_data['input'] = np.random.random((150, 1, 1, 4))\n    #print(new_data['input'].shape)   \n    #new_data['output'] = np.random.random_integers(0, 1, size=(150,3))    \n    #print(new_data['input'])\n\n    return new_data\n\ndef save_data_as_hdf5(hdf5_data_filename, data):\n    '''\n    HDF5 is one of the data formats Caffe accepts\n    '''\n    with h5py.File(hdf5_data_filename, 'w') as f:\n        f['data'] = data['input'].astype(np.float32)\n        f['label'] = data['output'].astype(np.float32)\n\n\ndef train(solver_prototxt_filename):\n    '''\n    Train the ANN\n    '''\n    caffe.set_mode_cpu()\n    solver = caffe.get_solver(solver_prototxt_filename)\n    solver.solve()\n\n\ndef print_network_parameters(net):\n    '''\n    Print the parameters of the network\n    '''\n    print(net)\n    print('net.inputs: {0}'.format(net.inputs))\n    print('net.outputs: {0}'.format(net.outputs))\n    print('net.blobs: {0}'.format(net.blobs))\n    print('net.params: {0}'.format(net.params))    \n\ndef get_predicted_output(deploy_prototxt_filename, caffemodel_filename, input, net = None):\n    '''\n    Get the predicted output, i.e. perform a forward pass\n    '''\n    if net is None:\n        net = caffe.Net(deploy_prototxt_filename,caffemodel_filename, caffe.TEST)\n\n    #input = np.array([[ 5.1,  3.5,  1.4,  0.2]])\n    #input = np.random.random((1, 1, 1))\n    #print(input)\n    #print(input.shape)\n    out = net.forward(data=input)\n    #print('out: {0}'.format(out))\n    return out[net.outputs[0]]\n\n\nimport google.protobuf \ndef print_network(prototxt_filename, caffemodel_filename):\n    '''\n    Draw the ANN architecture\n    '''\n    _net = caffe.proto.caffe_pb2.NetParameter()\n    f = open(prototxt_filename)\n    google.protobuf.text_format.Merge(f.read(), _net)\n    caffe.draw.draw_net_to_file(_net, prototxt_filename + '.png' )\n    print('Draw ANN done!')\n\n\ndef print_network_weights(prototxt_filename, caffemodel_filename):\n    '''\n    For each ANN layer, print weight heatmap and weight histogram \n    '''\n    net = caffe.Net(prototxt_filename,caffemodel_filename, caffe.TEST)\n    for layer_name in net.params: \n        # weights heatmap \n        arr = net.params[layer_name][0].data\n        plt.clf()\n        fig = plt.figure(figsize=(10,10))\n        ax = fig.add_subplot(111)\n        cax = ax.matshow(arr, interpolation='none')\n        fig.colorbar(cax, orientation=\"horizontal\")\n        plt.savefig('{0}_weights_{1}.png'.format(caffemodel_filename, layer_name), dpi=100, format='png', bbox_inches='tight') # use format='svg' or 'pdf' for vectorial pictures\n        plt.close()\n\n        # weights histogram  \n        plt.clf()\n        plt.hist(arr.tolist(), bins=20)\n        plt.savefig('{0}_weights_hist_{1}.png'.format(caffemodel_filename, layer_name), dpi=100, format='png', bbox_inches='tight') # use format='svg' or 'pdf' for vectorial pictures\n        plt.close()\n\n\ndef get_predicted_outputs(deploy_prototxt_filename, caffemodel_filename, inputs):\n    '''\n    Get several predicted outputs\n    '''\n    outputs = []\n    net = caffe.Net(deploy_prototxt_filename,caffemodel_filename, caffe.TEST)\n    for input in inputs:\n        #print(input)\n        outputs.append(copy.deepcopy(get_predicted_output(deploy_prototxt_filename, caffemodel_filename, input, net)))\n    return outputs    \n\n\ndef get_accuracy(true_outputs, predicted_outputs):\n    '''\n\n    '''\n    number_of_samples = true_outputs.shape[0]\n    number_of_outputs = true_outputs.shape[1]\n    threshold = 0.0 # 0 if SigmoidCrossEntropyLoss ; 0.5 if EuclideanLoss\n    for output_number in range(number_of_outputs):\n        predicted_output_binary = []\n        for sample_number in range(number_of_samples):\n            #print(predicted_outputs)\n            #print(predicted_outputs[sample_number][output_number])            \n            if predicted_outputs[sample_number][0][output_number] &lt; threshold:\n                predicted_output = 0\n            else:\n                predicted_output = 1\n            predicted_output_binary.append(predicted_output)\n\n        print('accuracy: {0}'.format(sklearn.metrics.accuracy_score(true_outputs[:, output_number], predicted_output_binary)))\n        print(sklearn.metrics.confusion_matrix(true_outputs[:, output_number], predicted_output_binary))\n\n\ndef main():\n    '''\n    This is the main function\n    '''\n\n    # Set parameters\n    solver_prototxt_filename = 'iris_solver.prototxt'\n    train_test_prototxt_filename = 'iris_train_test.prototxt'\n    deploy_prototxt_filename  = 'iris_deploy.prototxt'\n    deploy_prototxt_filename  = 'iris_deploy.prototxt'\n    deploy_prototxt_batch2_filename  = 'iris_deploy_batchsize2.prototxt'\n    hdf5_train_data_filename = 'iris_train_data.hdf5' \n    hdf5_test_data_filename = 'iris_test_data.hdf5' \n    caffemodel_filename = 'iris__iter_5000.caffemodel' # generated by train()\n\n    # Prepare data\n    data = load_data()\n    print(data)\n    train_data = data\n    test_data = data\n    save_data_as_hdf5(hdf5_train_data_filename, data)\n    save_data_as_hdf5(hdf5_test_data_filename, data)\n\n    # Train network\n    train(solver_prototxt_filename)\n\n    # Print network\n    print_network(deploy_prototxt_filename, caffemodel_filename)\n    print_network(train_test_prototxt_filename, caffemodel_filename)\n    print_network_weights(train_test_prototxt_filename, caffemodel_filename)\n\n    # Compute performance metrics\n    #inputs = input = np.array([[[[ 5.1,  3.5,  1.4,  0.2]]],[[[ 5.9,  3. ,  5.1,  1.8]]]])\n    inputs = data['input']\n    outputs = get_predicted_outputs(deploy_prototxt_filename, caffemodel_filename, inputs)\n    get_accuracy(data['output'], outputs)\n\n\nif __name__ == \"__main__\":\n    main()\n    #cProfile.run('main()') # if you want to do some profiling\n<\/code><\/pre>\n\n<p><code>iris_train_test.prototxt<\/code>:<\/p>\n\n<pre><code>name: \"IrisNet\"\nlayer {\n  name: \"iris\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TRAIN\n  }\n  hdf5_data_param {\n    source: \"iris_train_data.txt\"\n    batch_size: 1\n\n  }\n}\n\nlayer {\n  name: \"iris\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TEST\n  }\n  hdf5_data_param {\n    source: \"iris_test_data.txt\"\n    batch_size: 1\n\n  }\n}\n\n\n\n\nlayer {\n  name: \"ip1\"\n  type: \"InnerProduct\"\n  bottom: \"data\"\n  top: \"ip1\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 50\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"relu1\"\n  type: \"ReLU\"\n  bottom: \"ip1\"\n  top: \"ip1\"\n}\nlayer {\n  name: \"drop1\"\n  type: \"Dropout\"\n  bottom: \"ip1\"\n  top: \"ip1\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\n\n\nlayer {\n  name: \"ip2\"\n  type: \"InnerProduct\"\n  bottom: \"ip1\"\n  top: \"ip2\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 50\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"drop2\"\n  type: \"Dropout\"\n  bottom: \"ip2\"\n  top: \"ip2\"\n  dropout_param {\n    dropout_ratio: 0.4\n  }\n}\n\n\n\nlayer {\n  name: \"ip3\"\n  type: \"InnerProduct\"\n  bottom: \"ip2\"\n  top: \"ip3\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 3\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\n\nlayer {\n  name: \"drop3\"\n  type: \"Dropout\"\n  bottom: \"ip3\"\n  top: \"ip3\"\n  dropout_param {\n    dropout_ratio: 0.3\n  }\n}\n\nlayer {\n  name: \"loss\"\n  type: \"SigmoidCrossEntropyLoss\" \n  # type: \"EuclideanLoss\" \n  # type: \"HingeLoss\"  \n  bottom: \"ip3\"\n  bottom: \"label\"\n  top: \"loss\"\n}\n<\/code><\/pre>\n\n<p><code>iris_deploy.prototxt<\/code>:<\/p>\n\n<pre><code>name: \"IrisNet\"\ninput: \"data\"\ninput_dim: 1 # batch size\ninput_dim: 1\ninput_dim: 1\ninput_dim: 4\n\n\nlayer {\n  name: \"ip1\"\n  type: \"InnerProduct\"\n  bottom: \"data\"\n  top: \"ip1\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 50\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"relu1\"\n  type: \"ReLU\"\n  bottom: \"ip1\"\n  top: \"ip1\"\n}\nlayer {\n  name: \"drop1\"\n  type: \"Dropout\"\n  bottom: \"ip1\"\n  top: \"ip1\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\n\n\nlayer {\n  name: \"ip2\"\n  type: \"InnerProduct\"\n  bottom: \"ip1\"\n  top: \"ip2\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 50\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"drop2\"\n  type: \"Dropout\"\n  bottom: \"ip2\"\n  top: \"ip2\"\n  dropout_param {\n    dropout_ratio: 0.4\n  }\n}\n\n\nlayer {\n  name: \"ip3\"\n  type: \"InnerProduct\"\n  bottom: \"ip2\"\n  top: \"ip3\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 3\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\n\nlayer {\n  name: \"drop3\"\n  type: \"Dropout\"\n  bottom: \"ip3\"\n  top: \"ip3\"\n  dropout_param {\n    dropout_ratio: 0.3\n  }\n}\n<\/code><\/pre>\n\n<p><code>iris_solver.prototxt<\/code>:<\/p>\n\n<pre><code># The train\/test net protocol buffer definition\nnet: \"iris_train_test.prototxt\"\n# test_iter specifies how many forward passes the test should carry out.\ntest_iter: 1\n# Carry out testing every test_interval training iterations.\ntest_interval: 1000\n# The base learning rate, momentum and the weight decay of the network.\nbase_lr: 0.0001\nmomentum: 0.001\nweight_decay: 0.0005\n# The learning rate policy\nlr_policy: \"inv\"\ngamma: 0.0001\npower: 0.75\n# Display every 100 iterations\ndisplay: 1000\n# The maximum number of iterations\nmax_iter: 5000\n# snapshot intermediate results\nsnapshot: 5000\nsnapshot_prefix: \"iris_\"\n# solver mode: CPU or GPU\nsolver_mode: CPU # GPU\n<\/code><\/pre>\n\n<p>FYI: <a href=\"https:\/\/stackoverflow.com\/a\/31396229\/395857\">Script to install Caffe and pycaffe on Ubuntu<\/a>.<\/p>\n"}]}
{"QuestionId":31390427,"AnswerCount":0,"Tags":"<python><neural-network><anaconda><caffe><conda>","CreationDate":"2015-07-13T18:08:20.203","AcceptedAnswerId":null,"OwnerUserId":5106073.0,"Title":"ImportError: No module named caffe","Body":"<p>I am trying to run <a href=\"http:\/\/googleresearch.blogspot.co.uk\/2015\/07\/deepdream-code-example-for-visualizing.html\" rel=\"nofollow\">Google Research's DeepDream code<\/a> on a mac running OSx 10.9.5.<br>\nThere are a few dependencies that I had to install. I am using the Anaconda distribution of python and I made sure that I have all the packages required.<\/p>\n\n<p>The hardest thing was to install Caffe. I have ATLAS installed using fink. Then I have compiled caffe and pycaffe. When I ran 'make runtest' all tests passed. I also ran 'make distribute'.<\/p>\n\n<p>When I run <a href=\"https:\/\/github.com\/google\/deepdream\" rel=\"nofollow\">the notebook released from Google<\/a>, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>How can this be? What can I try to do to fix it?<\/p>\n","answers":[]}
{"QuestionId":31391527,"AnswerCount":1,"Tags":"<python><theano><nolearn><lasagne>","CreationDate":"2015-07-13T19:11:26.030","AcceptedAnswerId":null,"OwnerUserId":1879940.0,"Title":"Neural network for more than one class not working","Body":"<p>I am trying to use a Neural network for a classification problem. I have 6 possible classes and the same input may be in more than one class.<\/p>\n\n<p>The problem is that when I try to train one NN for each class, I set output_num_units = 1 and on train, I pass the first column of y, y[:,0]. I get the following output and error:<\/p>\n\n<pre><\/pre>\n\n<p>If I try to use  (6) and the full y (all six fields), first I get an error of the KStratifiedFold, because it seems that it does not expect y to have multiple rows. If I set , than I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The only configuration that is working is setting more than one output unit and passing only one column to y. Than it trains the NN, but is does not seem to be right as it is giving me 2 output layers, and I have only one Y to compare to.<\/p>\n\n<p><strong>What am I doing wrong? Why can't I use only one output? Should I convert my y classes from a vector of 6 columns to a vector of only one column with a number?<\/strong><\/p>\n\n<p>I use the following code (extract):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31533004","CreationDate":"2015-07-21T07:38:53.153","ParentId":null,"OwnerUserId":"1646409","Title":null,"Body":"<p>I then wanted to use CNNs in Lasagne, but didn't get it to work the same way, as the predictions were always 0...  recommend you to look at <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow\">the MNIST example<\/a>. I find that one much better to use and to extend, as old code snippets didn't fully work due to API changes over time. I've amended the MNIST example, my target vector has labels 0 or 1 and create the output layer for the NN this way:<\/p>\n\n<pre><code># Finally, we'll add the fully-connected output layer, of 2 softmax units:\nl_out = lasagne.layers.DenseLayer(\n        l_hid2_drop, num_units=2,\n        nonlinearity=lasagne.nonlinearities.softmax)\n<\/code><\/pre>\n\n<p>And for the CNN:<\/p>\n\n<pre><code>    layer = lasagne.layers.DenseLayer(\n        lasagne.layers.dropout(layer, p=.5),\n        num_units=2,\n        nonlinearity=lasagne.nonlinearities.softmax)\n<\/code><\/pre>\n"}]}
{"QuestionId":31393897,"AnswerCount":1,"Tags":"<python><macos><neural-network><anaconda><caffe>","CreationDate":"2015-07-13T21:33:17.657","AcceptedAnswerId":null,"OwnerUserId":5106073.0,"Title":"ImportError: dlopen(...) library not open","Body":"<p>I am trying to run <a href=\"http:\/\/googleresearch.blogspot.co.uk\/2015\/07\/deepdream-code-example-for-visualizing.html\" rel=\"nofollow noreferrer\">Google Research's DeepDream code<\/a> on a mac running OSx 10.9.5.<br>\nThere are a few dependencies that I had to install. I am using the Anaconda distribution of python and I made sure that I have all the packages required.<\/p>\n\n<p>The hardest thing was to install Caffe. I have ATLAS installed using fink. Then I have compiled caffe and pycaffe. When I ran 'make runtest' all tests passed. I also ran 'make distribute'.<\/p>\n\n<p>When I run <a href=\"https:\/\/github.com\/google\/deepdream\" rel=\"nofollow noreferrer\">the notebook released from Google<\/a>, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>What can I do to fix this?<\/p>\n\n<hr>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/aGCqO.png\" alt=\"here is the screendump requested in a comment\"><\/p>\n","answers":[{"AnswerId":"31408146","CreationDate":"2015-07-14T13:31:54.857","ParentId":null,"OwnerUserId":"4378596","Title":null,"Body":"<p>libcudart.7.0.dylib is a GPU related library.<\/p>\n\n<p>Does the machine you're running on have a GPU? If not, then you need to specify CPU mode in Makefile.config for caffe.<\/p>\n\n<p>If you do have a GPU, then please take a look here.\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/779\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/779<\/a><\/p>\n"}]}
{"QuestionId":31395027,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-07-13T23:13:01.277","AcceptedAnswerId":null,"OwnerUserId":864128.0,"Title":"how to convert numpy.ndarray object to theano.tensor?","Body":"<p>I have a csv file to change to thenao.tensor, and I plan to do it via 2 steps\n1 csv to ndarray: easy by using <strong>genfromtxt<\/strong> method.\n2 ndarray to theano.tensor: how to do this step? Is there any sample code?<\/p>\n\n<p>Thx!<\/p>\n","answers":[{"AnswerId":"31395847","CreationDate":"2015-07-14T00:47:32.863","ParentId":null,"OwnerUserId":"2156909","Title":null,"Body":"<p>You can use <code>_shared<\/code>, apparently: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#converting-from-python-objects\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#converting-from-python-objects<\/a><\/p>\n\n<pre><code>from theano.tensor import _shared\nimport numpy as np\nx = _shared(np.arange(10))\n<\/code><\/pre>\n"}]}
{"QuestionId":31395729,"AnswerCount":3,"Tags":"<multithreading><ubuntu><neural-network><deep-learning><caffe>","CreationDate":"2015-07-14T00:32:46.923","AcceptedAnswerId":31396229.0,"OwnerUserId":395857.0,"Title":"How to enable multithreading with Caffe?","Body":"<p>I would like to compile \/ configure Caffe so that when I trained an artificial neural network with it, the training is multi-threaded (CPU only, no GPU). How to enable multithreading with Caffe? I use Caffe on Ubuntu 14.04 LTS x64.<\/p>\n","answers":[{"AnswerId":"44172201","CreationDate":"2017-05-25T04:20:12.773","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>While building caffe, you have to add the -fopenmp to the CXXFLAGS and LINKFLAGS to support OPENMP. If you have a flag named OPENMP in the Makefil.config, you can simply set that to 1. You can use either OPENBLAS or Intel MKL BLAS library. While building the OPENBLAS you need to set USE_OPENMP=1 flag so that it supports OPENMP. After building caffe, please export the number of threads you want to use during runtime by setting up OMP_NUM_THREADS=n where n is the number of threads you want. Here is a good discussion related to multi-threading in Caffe: <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/439\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/pull\/439<\/a><\/p>\n"},{"AnswerId":"51347019","CreationDate":"2018-07-15T09:21:23.927","ParentId":null,"OwnerUserId":"5330223","Title":null,"Body":"<p>This is to just extend <strong>Franck's<\/strong> <a href=\"https:\/\/stackoverflow.com\/a\/31396229\/5330223\">answer<\/a> where he used <code>sed<\/code> to modify the <code>config<\/code> file. If you are having problems with that, here is another way to get the same thing done.<\/p>\n\n<p>The difference is that instead of changing the config file you directly change the <code>camke<\/code> flag <code>cmake -DCPU_ONLY=1 -DBLAS=open ..<\/code><\/p>\n\n<pre><code>$sudo apt update &amp;&amp; sudo apt-get install -y libopenblas-dev\n$git clone -b 1.0 --depth 1 https:\/\/github.com\/BVLC\/caffe.git . &amp;&amp; \\\n    pip install --upgrade pip &amp;&amp; \\\n    cd python &amp;&amp; pip install -r requirements.txt &amp;&amp; cd .. &amp;&amp; \\\n    mkdir build &amp;&amp; cd build &amp;&amp; \\\n    cmake -DCPU_ONLY=1 -DBLAS=open .. &amp;&amp; \\\n    make -j\"$(nproc)\"\n<\/code><\/pre>\n"},{"AnswerId":"31396229","CreationDate":"2015-07-14T01:37:41.620","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>One way is to use OpenBLAS instead of the default ATLAS. To do so, <\/p>\n\n<ol>\n<li><code>sudo apt-get install -y libopenblas-dev<\/code><\/li>\n<li>Before compiling Caffe, edit <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/Makefile.config.example\" rel=\"noreferrer\"><code>Makefile.config<\/code><\/a>, replace <code>BLAS := atlas<\/code> by <code>BLAS := open<\/code><\/li>\n<li>After compiling Caffe, running <code>export OPENBLAS_NUM_THREADS=4<\/code> will cause Caffe to use 4 cores.<\/li>\n<\/ol>\n\n<hr>\n\n<p>If interested, here is a script to install Caffe and pycaffe on a new Ubuntu 14.04 LTS x64 or Ubuntu 14.10 x64. CPU only, multi-threaded Caffe. It can probably be improved, but it's good enough for me for now:<\/p>\n\n<pre class=\"lang-bash prettyprint-override\"><code># This script installs Caffe and pycaffe on Ubuntu 14.04 x64 or 14.10 x64. CPU only, multi-threaded Caffe.\n# Usage: \n# 0. Set up here how many cores you want to use during the installation:\n# By default Caffe will use all these cores.\nNUMBER_OF_CORES=4\n# 1. Execute this script, e.g. \"bash compile_caffe_ubuntu_14.04.sh\" (~30 to 60 minutes on a new Ubuntu).\n# 2. Open a new shell (or run \"source ~\/.bash_profile\"). You're done. You can try \n#    running \"import caffe\" from the Python interpreter to test.\n\n#http:\/\/caffe.berkeleyvision.org\/install_apt.html : (general install info: http:\/\/caffe.berkeleyvision.org\/installation.html)\ncd\nsudo apt-get update\n#sudo apt-get upgrade -y # If you are OK getting prompted\nsudo DEBIAN_FRONTEND=noninteractive apt-get upgrade -y -q -o Dpkg::Options::=\"--force-confdef\" -o Dpkg::Options::=\"--force-confold\" # If you are OK with all defaults\n\nsudo apt-get install -y libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev\nsudo apt-get install -y --no-install-recommends libboost-all-dev\nsudo apt-get install -y libatlas-base-dev \nsudo apt-get install -y python-dev \nsudo apt-get install -y python-pip git\n\n# For Ubuntu 14.04\nsudo apt-get install -y libgflags-dev libgoogle-glog-dev liblmdb-dev protobuf-compiler \n\n# LMDB\n# https:\/\/github.com\/BVLC\/caffe\/issues\/2729: Temporarily broken link to the LMDB repository #2729\n#git clone https:\/\/gitorious.org\/mdb\/mdb.git\n#cd mdb\/libraries\/liblmdb\n#make &amp;&amp; make install \n\ngit clone https:\/\/github.com\/LMDB\/lmdb.git \ncd lmdb\/libraries\/liblmdb\nsudo make \nsudo make install\n\n# More pre-requisites \nsudo apt-get install -y cmake unzip doxygen\nsudo apt-get install -y protobuf-compiler\nsudo apt-get install -y libffi-dev python-dev build-essential\nsudo pip install lmdb\nsudo pip install numpy\nsudo apt-get install -y python-numpy\nsudo apt-get install -y gfortran # required by scipy\nsudo pip install scipy # required by scikit-image\nsudo apt-get install -y python-scipy # in case pip failed\nsudo apt-get install -y python-nose\nsudo pip install scikit-image # to fix https:\/\/github.com\/BVLC\/caffe\/issues\/50\n\n\n# Get caffe (http:\/\/caffe.berkeleyvision.org\/installation.html#compilation)\ncd\nmkdir caffe\ncd caffe\nwget https:\/\/github.com\/BVLC\/caffe\/archive\/master.zip\nunzip -o master.zip\ncd caffe-master\n\n# Prepare Python binding (pycaffe)\ncd python\nfor req in $(cat requirements.txt); do sudo pip install $req; done\necho \"export PYTHONPATH=$(pwd):$PYTHONPATH \" &gt;&gt; ~\/.bash_profile # to be able to call \"import caffe\" from Python after reboot\nsource ~\/.bash_profile # Update shell \ncd ..\n\n# Compile caffe and pycaffe\ncp Makefile.config.example Makefile.config\nsed -i '8s\/.*\/CPU_ONLY := 1\/' Makefile.config # Line 8: CPU only\nsudo apt-get install -y libopenblas-dev\nsed -i '33s\/.*\/BLAS := open\/' Makefile.config # Line 33: to use OpenBLAS\n# Note that if one day the Makefile.config changes and these line numbers change, we're screwed\n# Maybe it would be best to simply append those changes at the end of Makefile.config \necho \"export OPENBLAS_NUM_THREADS=($NUMBER_OF_CORES)\" &gt;&gt; ~\/.bash_profile \nmkdir build\ncd build\ncmake ..\ncd ..\nmake all -j$NUMBER_OF_CORES # 4 is the number of parallel threads for compilation: typically equal to number of physical cores\nmake pycaffe -j$NUMBER_OF_CORES\nmake test\nmake runtest\n#make matcaffe\nmake distribute\n\n# Bonus for other work with pycaffe\nsudo pip install pydot\nsudo apt-get install -y graphviz\nsudo pip install scikit-learn\n\n# At the end, you need to run \"source ~\/.bash_profile\" manually or start a new shell to be able to do 'python import caffe', \n# because one cannot source in a bash script. (http:\/\/stackoverflow.com\/questions\/16011245\/source-files-in-a-bash-script)\n<\/code><\/pre>\n\n<p>I have placed this script on GitHub: <br><a href=\"https:\/\/github.com\/Franck-Dernoncourt\/caffe_demos\/tree\/master\/caffe_installation\" rel=\"noreferrer\">https:\/\/github.com\/Franck-Dernoncourt\/caffe_demos\/tree\/master\/caffe_installation<\/a> . <\/p>\n"}]}
{"QuestionId":31404890,"AnswerCount":2,"Tags":"<neural-network><caffe><conv-neural-network>","CreationDate":"2015-07-14T11:07:48.377","AcceptedAnswerId":null,"OwnerUserId":1348187.0,"Title":"Convolutional Neural Networks with Caffe and NEGATIVE or FALSE IMAGES","Body":"<p>When training a set of classes (let's say #clases (number of classes) = N) on Caffe Deep Learning (or any CNN framework) and I make a query to the caffemodel, I get a % of probability of that image could be OK.<\/p>\n\n<p>So, let's take a picture of a similar Class 1, and I get the result:<\/p>\n\n<blockquote>\n  <p>1.- 96%<\/p>\n  \n  <p>2.- 4%<\/p>\n<\/blockquote>\n\n<p>rest... 0%\nthe problem is: when I take a random picture (for example of my environment), I keep getting the same result, where one of the class is predominant (>90% probability) but it doesn't belong to any class.<\/p>\n\n<p>So what I'd like to hear is opinions\/answers from people which has experienced this and would have solved how to deal with no-sense inputs to the Neural Network.<\/p>\n\n<p>My purposes are:<\/p>\n\n<ul>\n<li>Train one more extra class with negative images (like with train_cascade).<\/li>\n<li>Train one more class extra with all the positive images in the TRAIN set, and the negative on the VAL set. But my purposes don't have any scientific base to execute them, that's why I ask you this question.<\/li>\n<\/ul>\n\n<p>What would you do?<\/p>\n\n<p>Thank you very much in advance.<\/p>\n\n<p>Rafael.<\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>After two months, A colleague of mine throw me a clue: <strong>The Activation Function.<\/strong><\/p>\n\n<p>I've seen that I use <strong>ReLU<\/strong> in every layer which means that the value for x is x when x > 0 and 0 otherwise. These are my layers:<\/p>\n\n<pre><\/pre>\n\n<p>if I make ReLU as x for any x (so negative for x &lt; 0) my network converges in accuracy = 0...<\/p>\n\n<p>Is there better way to do it ?<\/p>\n","answers":[{"AnswerId":"31405126","CreationDate":"2015-07-14T11:18:47.690","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>Train an extra class with negative examples.<br>\nOr - this will probably work - use pre-trained network and weights if the network definition satisfies you, for example from ImageNet, and add you classes as additional labels. In that way you have higher chances not to overfit to that additional (the negative) class.  If your network is different you can train it from scratch on a larger dataset instead of using the pre-trained weights.<\/p>\n"},{"AnswerId":"31447289","CreationDate":"2015-07-16T06:56:10.197","ParentId":null,"OwnerUserId":"5122242","Title":null,"Body":"<p>well i am also working on a similar problem , what i dont understand is even if you are to tell the neural network that this is a +ve image or -ve image , i dont understand how thats going to alter the cascade .I think you have to pick out features from the training image .May be you  can build a hybrid system where it alters the XML cascade   <\/p>\n"}]}
{"QuestionId":31414341,"AnswerCount":1,"Tags":"<ubuntu><compilation><neural-network><deep-learning><caffe>","CreationDate":"2015-07-14T18:19:46.523","AcceptedAnswerId":31414383.0,"OwnerUserId":395857.0,"Title":"Compiling Caffe: undefined reference to `PyString_FromString'","Body":"<p>I am trying to compile Caffe from the official GitHub sources + using a couple of layer cpp files added by a user. When compiling I get the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31414383","CreationDate":"2015-07-14T18:22:10.457","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>The compilation error means that at least some cpp files added or altered by the user use Python. To remediate the issue, you should uncomment <code>WITH_PYTHON_LAYER := 1<\/code> in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/Makefile.config.example\" rel=\"nofollow\"><code>Makefile.config<\/code><\/a> before compiling.<\/p>\n"}]}
{"QuestionId":31419921,"AnswerCount":1,"Tags":"<windows><lua><torch><luarocks>","CreationDate":"2015-07-15T01:17:26.080","AcceptedAnswerId":null,"OwnerUserId":3234562.0,"Title":"Installing Torch7 on Win7; cmake and PATH problems","Body":"<p>I'm trying to install Torch 7 on my Win7 system to run an RNN, and it's insane. I installed it easily on my Ubuntu VM, but that can't access my GPU for CUDA acceleration, so either I try experimental PCI passthrough software, or I try to get Torch on Windows. I've managed to install Lua and LuaRocks so far (but I can't run it from anywhere but the C:\\Program Files (x86)\\LuaRocks\\2.2 path). I've installed mingw and cmake. I tried installing Torch using the following command:<\/p>\n\n<pre><\/pre>\n\n<p>(source: <a href=\"https:\/\/stackoverflow.com\/questions\/27276822\/installing-torch7-with-luarocks-on-windows-with-mingw-build-error\">Installing Torch7 with Luarocks on Windows with mingw build error<\/a>)<\/p>\n\n<p>but I get:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know where the cl compiler is, or even if I have it on my system. Regarding my PATH variable, it's apparently a user variable, not a system one (I don't have a system PATH variable). I don't know if that's a problem. It currently looks like this: <\/p>\n\n<pre><\/pre>\n\n<p>I have no clue if that's correct, but if it's meant to let me run lua or luarocks from outside their respective bin directories, it fails at that.\nIf anyone has an easier way of installing Torch on Windows, please let me know (or heck, even a way of enabling GPU acceleration in a VM. Anything to get out of this stuck situation).<\/p>\n","answers":[{"AnswerId":"31420920","CreationDate":"2015-07-15T03:22:12.650","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p><code>cl<\/code> is command line compiler from Visual Studio. CMake is looking for it as its default settings use it. To use mingw that you have, you need to provide an additional option (<a href=\"https:\/\/stackoverflow.com\/a\/28058692\/1442917\">as I described here<\/a>), but I'm not sure how to pass it to luarocks as I usually do it from the command line directly.<\/p>\n\n<p>You may try to follow the steps in the answer I linked; there are details in the torch7 ticket referenced. In short, the steps will involve:<\/p>\n\n<ol>\n<li>Clone, compile and install <a href=\"https:\/\/github.com\/torch\/paths\" rel=\"nofollow noreferrer\">torch\/paths<\/a>;<\/li>\n<li>Clone, compile and install <a href=\"https:\/\/github.com\/torch\/cwrap\" rel=\"nofollow noreferrer\">torch\/cwrap<\/a>;<\/li>\n<li>Clone, compile and install <a href=\"https:\/\/github.com\/torch\/torch7\" rel=\"nofollow noreferrer\">torch\/torch<\/a>; make sure you grab the latest code as it includes the changes I submitted for mingw compilation.<\/li>\n<li>Clone, compile and install <a href=\"https:\/\/github.com\/torch\/nn\" rel=\"nofollow noreferrer\">torch\/nn<\/a>. See the discussion in <a href=\"https:\/\/github.com\/torch\/torch7\/pull\/287\" rel=\"nofollow noreferrer\">this ticket<\/a> for one change you may need to apply.<\/li>\n<\/ol>\n\n<p>The ticket also provides specific commands you can run to compile from the command line.<\/p>\n"}]}
{"QuestionId":31427094,"AnswerCount":1,"Tags":"<image-processing><machine-learning><deep-learning><computer-vision><caffe>","CreationDate":"2015-07-15T09:53:01.303","AcceptedAnswerId":31431716.0,"OwnerUserId":5118798.0,"Title":"A guide to convert_imageset.cpp","Body":"<p>I am relatively new to machine learning\/python\/ubuntu.<\/p>\n\n<p>I have a set of images in .jpg format where half contain a feature I want caffe to learn and half don't. I'm having trouble in finding a way to convert them to the required lmdb format.<\/p>\n\n<p>I have the necessary text input files. <\/p>\n\n<p>My question is can anyone provide a step by step guide on how to use  in the ubuntu terminal?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"31431716","CreationDate":"2015-07-15T13:26:29.693","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h2>A quick guide to Caffe's <code>convert_imageset<\/code><\/h2>\n\n<h3>Build<\/h3>\n\n<p>First thing you must do is build caffe and caffe's tools (<code>convert_imageset<\/code> is one of these tools).<br>\nAfter installing caffe and <code>make<\/code>ing it make sure you ran <code>make tools<\/code> as well.<br>\nVerify that a binary file <code>convert_imageset<\/code> is created in <code>$CAFFE_ROOT\/build\/tools<\/code>.<\/p>\n\n<h3>Prepare your data<\/h3>\n\n<p><em>Images:<\/em> put all images in a folder (I'll call it here <code>\/path\/to\/jpegs\/<\/code>).<br>\n<em>Labels:<\/em> create a text file (e.g., <code>\/path\/to\/labels\/train.txt<\/code>) with a line per input image  . For example:  <\/p>\n\n<blockquote>\n  <p>img_0000.jpeg 1<br>\n  img_0001.jpeg 0<br>\n  img_0002.jpeg 0  <\/p>\n<\/blockquote>\n\n<p>In this example the first image is labeled <code>1<\/code> while the other two are labeled <code>0<\/code>. <\/p>\n\n<h3>Convert the dataset<\/h3>\n\n<p>Run the binary in shell<\/p>\n\n<pre><code>~$ GLOG_logtostderr=1 $CAFFE_ROOT\/build\/tools\/convert_imageset \\\n    --resize_height=200 --resize_width=200 --shuffle  \\\n    \/path\/to\/jpegs\/ \\\n    \/path\/to\/labels\/train.txt \\\n    \/path\/to\/lmdb\/train_lmdb\n<\/code><\/pre>\n\n<p>Command line explained:  <\/p>\n\n<ul>\n<li><code>GLOG_logtostderr<\/code> flag is set to 1 <em>before<\/em> calling <code>convert_imageset<\/code> indicates the logging mechanism to redirect log messages to stderr.  <\/li>\n<li><code>--resize_height<\/code> and <code>--resize_width<\/code> resize <strong>all<\/strong> input images to same size <code>200x200<\/code>.  <\/li>\n<li><code>--shuffle<\/code> randomly change the order of images and does not preserve the order in the <code>\/path\/to\/labels\/train.txt<\/code> file.  <\/li>\n<li>Following are the path to the images folder, the labels text file and the output name. Note that the output name should not exist prior to calling <code>convert_imageset<\/code> otherwise you'll get a scary error message.<\/li>\n<\/ul>\n\n<p>Other flags that might be useful:<\/p>\n\n<ul>\n<li><code>--backend<\/code> - allows you to choose between an <code>lmdb<\/code> dataset or <code>levelDB<\/code>.<\/li>\n<li><code>--gray<\/code>    - convert all images to gray scale.<\/li>\n<li><code>--encoded<\/code> and <code>--encoded_type<\/code>  - keep image data in encoded (jpg\/png) compressed form in the database.<\/li>\n<li><code>--help<\/code>    - shows some help, see all relevant flags under <em>Flags from tools\/convert_imageset.cpp<\/em>  <\/li>\n<\/ul>\n\n<p>You can check out <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/imagenet\/create_imagenet.sh\" rel=\"noreferrer\"><code>$CAFFE_ROOT\/examples\/imagenet\/convert_imagenet.sh<\/code><\/a>\nfor an example how to use <code>convert_imageset<\/code>.<\/p>\n"}]}
{"QuestionId":31433577,"AnswerCount":2,"Tags":"<python><deep-learning><caffe>","CreationDate":"2015-07-15T14:43:13.260","AcceptedAnswerId":31457745.0,"OwnerUserId":894903.0,"Title":"Passing arguments to the forward function in caffe","Body":"<p>I have found this line of code to send inputs to the network in caffe:<\/p>\n\n<pre><\/pre>\n\n<p>I tried adapting this code for my work as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where  (just for testing) but I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am unable to understand why I cant run the forward method this way. I am able to call  regularly though. <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>From the comments I understand that I am supposed to initialize the input array with possibly the  function. <\/p>\n\n<p>I tried the following loop:<\/p>\n\n<pre><\/pre>\n\n<p>But this still causes the same error. <\/p>\n","answers":[{"AnswerId":"31446685","CreationDate":"2015-07-16T06:25:04.330","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>the <code>net<\/code> object is expecting a <strong>list<\/strong> of images as input, you supply only one image.<\/p>\n\n<p>Try<\/p>\n\n<pre><code>out = net.forward(**{net.inputs[0]: [ im_proc ] })\n<\/code><\/pre>\n\n<p>Note the square brackets (<code>[]<\/code>) around <code>im_proc<\/code> converting it into a list containing a single image. <\/p>\n"},{"AnswerId":"31457745","CreationDate":"2015-07-16T14:55:45.797","ParentId":null,"OwnerUserId":"894903","Title":null,"Body":"<p>The solution to this was to switch out the prototext. The C++ code PyCaffe wraps around causes this behavior. This thread has the solution: <a href=\"https:\/\/stackoverflow.com\/questions\/29124840\/prediction-in-caffe-exception-input-blob-arguments-do-not-match-net-inputs\">Prediction in Caffe - Exception: Input blob arguments do not match net inputs<\/a><\/p>\n"}]}
{"QuestionId":31444173,"AnswerCount":4,"Tags":"<python-2.7><theano>","CreationDate":"2015-07-16T01:58:57.050","AcceptedAnswerId":null,"OwnerUserId":5121662.0,"Title":"Error importing theano \"cannot import name gof\"","Body":"<p>I am current getting the error <\/p>\n\n<blockquote>\n  <p>ImportError: cannot import name gof<\/p>\n<\/blockquote>\n\n<p>when importing theano. <\/p>\n\n<pre><\/pre>\n\n<p>I am using python 2.7.10 (). Theano is installed using . \nHope to get you suggestion to solve this problem<\/p>\n","answers":[{"AnswerId":"43197332","CreationDate":"2017-04-04T02:35:05.487","ParentId":null,"OwnerUserId":"395287","Title":null,"Body":"<p>This <code>ImportError<\/code> can be caused because Theano is <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2406\" rel=\"nofollow noreferrer\">unable to compile the <code>gof<\/code> module itself<\/a>. If this is the case, you will see an error message that looks like \"<code>Exception: Compilation Failed (return status=1): C:\\Long\\Path\\...\\mod.cpp:1: sorry, unimplemented: 64-bit mode not compiled in<\/code>\".<\/p>\n\n<h2>Fixing With Conda<\/h2>\n\n<p>If you are installing <code>theano<\/code> into a <code>conda<\/code> environment, make sure that you have a C compiler available to that environment.<\/p>\n\n<p>The command<\/p>\n\n<pre><code>conda install m2w64-toolchain\n<\/code><\/pre>\n\n<p>will provide a C compiler to your environment that's isolated from the rest of the machine. <\/p>\n\n<p>After the <code>m2w64-toolchain<\/code> package is installed, <code>import theano<\/code> should work <\/p>\n\n<h2>Fixing Manually<\/h2>\n\n<p>If you are installing Theano yourself, two points from <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2406\" rel=\"nofollow noreferrer\">these<\/a> <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2732#issuecomment-93009341\" rel=\"nofollow noreferrer\">threads<\/a> may help:<\/p>\n\n<ul>\n<li>Install the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow noreferrer\">bleeding edge version of Theano<\/a><\/li>\n<li>Install <code>libpython<\/code> from <a href=\"http:\/\/www.lfd.uci.edu\/%7Egohlke\/pythonlibs\/\" rel=\"nofollow noreferrer\">http:\/\/www.lfd.uci.edu\/%7Egohlke\/pythonlibs\/<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"31530395","CreationDate":"2015-07-21T04:28:21.183","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Most of the time, when I see this error, it is caused by those 2 errors:<\/p>\n\n<p>1) A syntax error in Theano. Update Theano and make sure to have no local modifcation. I nerver saw this error in the master of Theano, but just in case.<\/p>\n\n<p>2) When there is multiple version of Theano that are installed.<\/p>\n\n<p>In both case, remove all version of Theano. Do it multiple time to be sure there is none left. Then install again.<\/p>\n\n<p>From memory, this always solved the problem when it wasn't a syntax error during development (but not in the master version of Theano that you use)<\/p>\n"},{"AnswerId":"47970697","CreationDate":"2017-12-25T17:24:40.887","ParentId":null,"OwnerUserId":"5108509","Title":null,"Body":"<p>I assume you're using Windows 7 or later.<\/p>\n\n<p>If you have installed Python Anaconda, then open Windows Powershell or Command Prompt and type <code>conda install mingw libpython<\/code> before typing <code>pip install theano<\/code><\/p>\n\n<blockquote>\n  <p>Alternatively, if you don't have Anaconda, download those packages from <\/p>\n  \n  <ul>\n  <li><a href=\"https:\/\/anaconda.org\/anaconda\/mingw\/files\" rel=\"nofollow noreferrer\">anaconda.org\/anaconda\/mingw\/files<\/a><\/li>\n  <li><a href=\"https:\/\/anaconda.org\/anaconda\/libpython\/files\" rel=\"nofollow noreferrer\">anaconda.org\/anaconda\/libpython\/files<\/a><\/li>\n  <li><a href=\"https:\/\/github.com\/Theano\/Theano\" rel=\"nofollow noreferrer\">github.com\/Theano\/Theano<\/a><\/li>\n  <\/ul>\n  \n  <p>Then open Command Prompt, navigate to each folder and type <code>python setup.py install<\/code><\/p>\n<\/blockquote>\n\n<p>Now run Python and <code>import theano<\/code><\/p>\n\n<p><strong>Possible errors:<\/strong><\/p>\n\n<p>If you get the RuntimeError: \"<a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/6507#issuecomment-342519898\" rel=\"nofollow noreferrer\">To use MKL 2018 with Theano you MUST set \"MKL_THREADING_LAYER=GNU\" in your environement<\/a>\" then<\/p>\n\n<ol>\n<li><p>Go to Control Panel > System > Advanced system settings and select \"Environment Variables\".<\/p><\/li>\n<li><p>In the \"System variables\" section, make a new variable name <code>MKL_THREADING_LAYER<\/code> and set its value to <code>GPU<\/code><\/p><\/li>\n<\/ol>\n\n<p>If you get other kinds of errors, then try the following:<\/p>\n\n<ol>\n<li><p>Make an empty file called <code>.theanorc<\/code> (a file extension without a file name) in your home folder C:\\Users\\&lt;username&gt;. If you get the error \"You must type a file name\" then see <a href=\"https:\/\/stackoverflow.com\/q\/5004633\">stackoverflow.com\/q\/5004633<\/a><\/p><\/li>\n<li><p>Open <code>.theanorc<\/code> and write this:<\/p>\n\n<pre><code>[global]\ncxx=C:\\&lt;path to Anaconda&gt;\\Anaconda3\\MinGW\\bin\\g++.exe\n<\/code><\/pre><\/li>\n<li><p>Run Python again and import theano. If it works, then you can probably delete <code>.theanorc<\/code><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"55789385","CreationDate":"2019-04-22T05:07:47.223","ParentId":null,"OwnerUserId":"1682419","Title":null,"Body":"<p>In my case, the fix was to install a python build that's callable as a shared library:<\/p>\n\n<pre><code>PYTHON_CONFIGURE_OPTS=\"--enable-shared\" pyenv install 2.7.15\n<\/code><\/pre>\n"}]}
{"QuestionId":31444272,"AnswerCount":1,"Tags":"<lua><luajit><torch>","CreationDate":"2015-07-16T02:10:43.730","AcceptedAnswerId":31448164.0,"OwnerUserId":3106889.0,"Title":"Read images from local disk using torch7 while those images saved in different subfolders?","Body":"<p>I have images which have saved in the desk. The data saved as follow:  4 main folders (1,2,3 and 4) each folder has 26 subfolders ( these subfolders represent the class of images (A, B, C, D, ..,Z)). Each of these subfolder contains more than 500 images. However, I am looking for file or code in torch that can read these images. In MATLAB I could wrote a code but here I find it confuse. Could you please advise me.<\/p>\n","answers":[{"AnswerId":"31448164","CreationDate":"2015-07-16T07:41:16.587","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>What you can do is use <a href=\"http:\/\/stevedonovan.github.io\/Penlight\/api\/index.html\" rel=\"nofollow\">Penlight<\/a> (the library <a href=\"https:\/\/github.com\/torch\/distro\/blob\/54f060f\/install.sh#L68\" rel=\"nofollow\">is installed<\/a> when you install Torch).<\/p>\n\n<p>Penlight provides <a href=\"http:\/\/stevedonovan.github.io\/Penlight\/api\/libraries\/pl.dir.html\" rel=\"nofollow\"><code>pl.dir<\/code><\/a> that makes it easy to scan files in (sub-)folders. For example what you can do is:<\/p>\n\n<pre><code>local pl = require('pl.import_into')()\nlocal t = {}\nfor i,f in ipairs(pl.dir.getallfiles('\/data\/foo', '*.jpg')) do\n    t[i] = { f, pl.path.basename(pl.path.dirname(f)) }\nend\n<\/code><\/pre>\n\n<p>This creates a list of pairs (filename, class label = \"A\" or \"B\" ...). Of course you are free to change the file pattern (<code>*.jpg<\/code>) or to omit it (in such a case Penlight will simply list all files). You can also load the images on the fly:<\/p>\n\n<pre><code>t[i] = { image.load(f), pl.path.basename(pl.path.dirname(f)) }\n<\/code><\/pre>\n\n<p>Or do that right after when manipulating <code>t<\/code>.<\/p>\n"}]}
{"QuestionId":31454740,"AnswerCount":2,"Tags":"<deep-learning><caffe>","CreationDate":"2015-07-16T12:53:26.693","AcceptedAnswerId":null,"OwnerUserId":5118798.0,"Title":"Caffe: can't open imagenet_mean_test.binaryproto","Body":"<p>Upon running ; creating the layer data, setting up the data, loading the training  file and opening the training lmdb file all works.<\/p>\n\n<p>However when it comes to loading the test  file for the test data I get the following error:<\/p>\n\n<blockquote>\n  <p>Loading mean file from: \/home\/pwhc\/caffe\/Learn\/imagenet_mean_test.binaryproto<br>\n  F0716 13:12:13.917732  3385 db.hpp:109] Check failed: mdb_status == 0 (2 vs. 0) No such file or directory<br>\n  *** Check failure stack trace: ***<br>\n      @     0x7f8337946daa  (unknown)<br>\n      @     0x7f8337946ce4  (unknown)<br>\n      @     0x7f83379466e6  (unknown)<br>\n      @     0x7f8337949687  (unknown)<br>\n      @     0x7f8337cbf5be  caffe::db::LMDB::Open()<br>\n      @     0x7f8337d16b82  caffe::DataLayer&lt;>::DataLayerSetUp()<br>\n      @     0x7f8337d806f9  caffe::BasePrefetchingDataLayer&lt;>::LayerSetUp()<br>\n      @     0x7f8337ca3db3  caffe::Net&lt;>::Init()<br>\n      @     0x7f8337ca5b22  caffe::Net&lt;>::Net()<br>\n      @     0x7f8337cb0a24  caffe::Solver&lt;>::InitTestNets()<br>\n      @     0x7f8337cb111b  caffe::Solver&lt;>::Init()<br>\n      @     0x7f8337cb12e6  caffe::Solver&lt;>::Solver()<br>\n      @           0x40c4c0  caffe::GetSolver&lt;>()<br>\n      @           0x406503  train()<br>\n      @           0x404ab1  main<br>\n      @     0x7f8336e58ec5  (unknown)<br>\n      @           0x40505d  (unknown)<br>\n      @              (nil)  (unknown)<br>\n  Aborted (core dumped)<\/p>\n<\/blockquote>\n\n<p>I modified the  to point the to appropriate  files (using absolute paths) and have checked and double checked to make sure everything matches. <\/p>\n\n<p>Any thoughts would be greatly appreciated. <\/p>\n","answers":[{"AnswerId":"31549397","CreationDate":"2015-07-21T20:52:35.483","ParentId":null,"OwnerUserId":"2404152","Title":null,"Body":"<p>See my answer here:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2780#issuecomment-123385714\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/2780#issuecomment-123385714<\/a><\/p>\n\n<p>Can you post your data layers? It seems like you've switched up <code>data_param.source<\/code> and <code>transform_param.mean_file<\/code>.<\/p>\n"},{"AnswerId":"57052650","CreationDate":"2019-07-16T08:00:17.687","ParentId":null,"OwnerUserId":"5495014","Title":null,"Body":"<p>When you creating new LMDB database from image net, please delete the previous LMDB. This error will come to happen when writing the new image net data for existing LMDB database.<\/p>\n"}]}
{"QuestionId":31458983,"AnswerCount":1,"Tags":"<python><numpy><caffe>","CreationDate":"2015-07-16T15:51:31.077","AcceptedAnswerId":31459123.0,"OwnerUserId":894903.0,"Title":"Correct way to broadcast a 100x9 to a 100x9x1x1 numpy array for computation in Caffe","Body":"<p>I am trying to input my own data into a caffe model using the python wrappers. I read the data from HDF5 as a numpy array with dimension 100x9. But for the input to model, I use the following code:<\/p>\n\n<pre><\/pre>\n\n<p>So basically I need to fill out input_ from a 100x9 array.<\/p>\n","answers":[{"AnswerId":"31459123","CreationDate":"2015-07-16T15:57:26.930","ParentId":null,"OwnerUserId":"3854029","Title":null,"Body":"<p>Heres how you would convert a 100x9 array to a 100x9x1x1 array:<\/p>\n\n<pre><code>x = np.zeros((100,9))\ny = x[:,:,np.newaxis,np.newaxis]\n<\/code><\/pre>\n"}]}
{"QuestionId":31460102,"AnswerCount":0,"Tags":"<python><python-2.7><ipython><caffe>","CreationDate":"2015-07-16T16:47:38.693","AcceptedAnswerId":null,"OwnerUserId":1623856.0,"Title":"Caffe import working with python2 but not python2.7 \/ ipython notebook, despite both being 2.7.10?","Body":"<p>I've successfully compiled the caffe library and the python module.<\/p>\n\n<p>I can do this:<\/p>\n\n<pre><\/pre>\n\n<p>But, bizarrely, this fails:<\/p>\n\n<pre><\/pre>\n\n<p>I cannot understand this at all! Whenever I try to run using iPython notebook I get the same crash.<\/p>\n\n<p>Any ideas as to what may be causing this, and how I might fix it, or at least get iPython Notebook to use the different python version so I can run this thing?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":31475307,"AnswerCount":1,"Tags":"<image-processing><deep-learning><torch>","CreationDate":"2015-07-17T11:56:22.440","AcceptedAnswerId":31875549.0,"OwnerUserId":2598910.0,"Title":"How to load images and labels in Torch for a Convolutional Neural Network","Body":"<p>I'm new to Torch and would like to load some images from two directories (one for each label).  I'm trying to build a convolutional neural network that will classify images as belonging to one class or another (i.e. a binary classifier) but I am unsure how load images, label those images and get the data into the correct format.  I'm using the following tutorial, however the training data is loaded in a different way which I am not familiar with.<\/p>\n\n<p><a href=\"http:\/\/code.madbits.com\/wiki\/doku.php?id=tutorial_supervised\" rel=\"noreferrer\">http:\/\/code.madbits.com\/wiki\/doku.php?id=tutorial_supervised<\/a><\/p>\n\n<p>Hope someone can help me get started and point me in the right direction.  <\/p>\n\n<p>Many thanks in advance.<\/p>\n","answers":[{"AnswerId":"31875549","CreationDate":"2015-08-07T10:29:01.187","ParentId":null,"OwnerUserId":"5072453","Title":null,"Body":"<p>Check  <a href=\"https:\/\/github.com\/torch\/demos\/tree\/master\/train-face-detector\" rel=\"nofollow\">this link<\/a>\nThe data.lua file in that is very simple to understand. They have used torch tensors, and the code is self explanatory. I had achieved a clear understanding of loading the data, post reading this link; hope it helps you too. <\/p>\n"}]}
{"QuestionId":31477724,"AnswerCount":1,"Tags":"<python><artificial-intelligence><neural-network><deep-learning><caffe>","CreationDate":"2015-07-17T14:05:15.817","AcceptedAnswerId":31498402.0,"OwnerUserId":238971.0,"Title":"Setting batch_size in data_param for Caffe has no effect","Body":"<p>When I set  in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt\" rel=\"nofollow\">deploy.prototxt file<\/a> in Google deep dream <em>bvlc_googlenet<\/em> to lower GPU memory requirements, it has no effect on speed nor memory requirements. It's as if it was ignored. I know the file itself (deploy.prototxt) is being used because other changes are reflected in results so that's not the issue. I also tried to set batch_size on all related layers as well (\"inception_4c\/1x1\", etc), again no difference.<\/p>\n\n<p>This is how I'm setting it:<\/p>\n\n<pre><\/pre>\n\n<p>When I time the runtime of the script, it's the same with batch_size 1 and with batch_size 512, there is no difference.<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"31498402","CreationDate":"2015-07-19T05:52:39.730","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><code>data_param<\/code> is a parameter of the input data layer. You can set <code>batch_size<\/code> only for the input and this value propagates through the network.<\/p>\n\n<p>In the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt#L3\" rel=\"nofollow\"><em>deploy.prototxt<\/em><\/a> the batch size is set by the first <code>'input_dim'<\/code> argument (<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt#L3\" rel=\"nofollow\">third line<\/a>), try changing this value and see if it has any effect on the memory consumption of your network.<\/p>\n\n<p>The first few lines of the <em>deploy.prototxt<\/em> file should be interpreted as<\/p>\n\n<pre><code>input: \"data\"  # the \"top\" name of input layer\ninput_dim: 10  # 4th dimension - batch size\ninput_dim: 3   # 3rd dimension - number of channels\ninput_dim: 224 # 2nd dimension - height of input\ninput_dim: 224 # 1st dimension - width of input\n<\/code><\/pre>\n\n<p>Thus, you expect at your fist conv layer (<code>\"conv1\/7x7_s2\"<\/code>) a \"bottom\" named \"data\" with shape <code>10<\/code>-by-<code>3<\/code>-by-<code>224<\/code>-by-<code>224<\/code>.<\/p>\n"}]}
{"QuestionId":31478194,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-07-17T14:28:21.587","AcceptedAnswerId":null,"OwnerUserId":3755060.0,"Title":"Torch Tutorial: meaning of \"trainData.data[{ {},i,{},{} }]:mean()\" in 1_data.lua","Body":"<p>In the torch tutorial, I found the line:<\/p>\n\n<pre><\/pre>\n\n<p>Is there anyone who can explain what the indexing { {},i,{},{} } is doing?\nI could guess, but wanted to know the exact mechanism.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"31479970","CreationDate":"2015-07-17T16:00:19.873","ParentId":null,"OwnerUserId":"1417179","Title":null,"Body":"<p>This is actually a concise syntax for tensor narrowing \/ slicing, detailed <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\">here<\/a> in the documentation.<\/p>\n\n<p>Inside the <code>[{ ... }]<\/code>, you can for each dimension of a tensor:<\/p>\n\n<ul>\n<li>pass a number <code>n<\/code> to only keep the <code>n<\/code>-th component along this dimension,<\/li>\n<li>pass a range <code>{start,end}<\/code> to keep all the components from <code>start<\/code> to <code>end<\/code> along this dimension,<\/li>\n<li>pass <code>{}<\/code> to keep all the components along this dimension.<\/li>\n<\/ul>\n\n<p>In this precise case, it's a narrowing from a <code>u * v * w * x<\/code> tensor to a  <code>u * 1 * w * x<\/code> tensor by keeping only the <code>i<\/code>-th component along the 2nd dimension.<\/p>\n"}]}
{"QuestionId":31492699,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-07-18T15:30:57.023","AcceptedAnswerId":31498732.0,"OwnerUserId":5130560.0,"Title":"Use of None in Array indexing in Python","Body":"<p>I am using the LSTM tutorial for Theano (<a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\">http:\/\/deeplearning.net\/tutorial\/lstm.html<\/a>). In the lstm.py (<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\">http:\/\/deeplearning.net\/tutorial\/code\/lstm.py<\/a>) file, I don't understand the following line:<\/p>\n\n<pre><\/pre>\n\n<p>What does  mean? In this case  is the theano vector while  is a matrix.<\/p>\n","answers":[{"AnswerId":"31498732","CreationDate":"2015-07-19T06:55:06.570","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>This question has been asked and answered on the Theano mailing list, but is actually about the basics of numpy indexing.<\/p>\n\n<p>Here are the question and answer\n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/jq92vNtkYUI\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/jq92vNtkYUI<\/a><\/p>\n\n<p>For completeness, here is another explanation: slicing with <code>None<\/code> adds an axis to your array, see the relevant numpy documentation, because it behaves the same in both numpy and Theano:<\/p>\n\n<p><a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis\">http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis<\/a><\/p>\n\n<p>Note that <code>np.newaxis is None<\/code>:<\/p>\n\n<pre><code>import numpy as np\na = np.arange(30).reshape(5, 6)\n\nprint a.shape  # yields (5, 6)\nprint a[np.newaxis, :, :].shape  # yields (1, 5, 6)\nprint a[:, np.newaxis, :].shape  # yields (5, 1, 6)\nprint a[:, :, np.newaxis].shape  # yields (5, 6, 1)\n<\/code><\/pre>\n\n<p>Typically this is used to adjust shapes to be able to broadcast to higher dimensions. E.g. tiling 7 times in the middle axis can be achieved as<\/p>\n\n<pre><code>b = a[:, np.newaxis] * np.ones((1, 7, 1))\n\nprint b.shape  # yields (5, 7, 6), 7 copies of a along the second axis\n<\/code><\/pre>\n"},{"AnswerId":"31492800","CreationDate":"2015-07-18T15:40:43.960","ParentId":null,"OwnerUserId":"4354477","Title":null,"Body":"<p>I think the Theano vector's <code>__getitem__<\/code> method expects a tuple as an argument! like this:<\/p>\n\n<pre><code>class Vect (object):\n    def __init__(self,data):\n        self.data=list(data)\n\n    def __getitem__(self,key):\n        return self.data[key[0]:key[1]+1]\n\na=Vect('hello')\nprint a[0,2]\n<\/code><\/pre>\n\n<p>Here <code>print a[0,2]<\/code> when <code>a<\/code> is an ordinary list will raise an exception:<\/p>\n\n<pre><code>&gt;&gt;&gt; a=list('hello')\n&gt;&gt;&gt; a[0,2]\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\nTypeError: list indices must be integers, not tuple\n<\/code><\/pre>\n\n<p>But here the <code>__getitem__<\/code> method is different and it accepts a tuple as an argument.<\/p>\n\n<p>You can pass the <code>:<\/code> sign to <code>__getitem__<\/code> like this as <code>:<\/code> means <em>slice<\/em>:<\/p>\n\n<pre><code>class Vect (object):\n    def __init__(self,data):\n        self.data=list(data)\n\n    def __getitem__(self,key):\n        return self.data[0:key[1]+1]+list(key[0].indices(key[1]))\n\na=Vect('hello')\nprint a[:,2]\n<\/code><\/pre>\n\n<p>Speaking about <code>None<\/code>, it can be used when indexing in plain Python as well: <\/p>\n\n<pre><code>&gt;&gt;&gt; 'hello'[None:None]\n'hello'\n<\/code><\/pre>\n"}]}
{"QuestionId":31494687,"AnswerCount":1,"Tags":"<python><gpu><caffe>","CreationDate":"2015-07-18T19:15:31.453","AcceptedAnswerId":null,"OwnerUserId":3211036.0,"Title":"Caffe - inconsistency in the activation feature values - GPU mode","Body":"<p>Hi I am using <strong>caffe<\/strong> on <strong>Ubuntu 14.04<\/strong>, \n<strong>CUDA version 7.0<\/strong>  (latest)\n<strong>cudnn version 2<\/strong> (latest)\n<strong>GPU : NVIDIA GT 730<\/strong><\/p>\n\n<p>In caffe first I get the initialization done and then I load the imagenet model (Alexnet). I also initialize the gpu using \nAfter that I take an image. I copy this image onto the caffe source blob. Then I perform a forward pass for this image by using : \nThen I extract the 4096 dimensional fc7 output.(the activation features of the fc7 layer)<\/p>\n\n<p>The problem I am facing is that when I run the same code multiple times, everytime I obtain a different result. That is, in GPU mode, everytime the activation features are different for the same image. When I am using forward pass, the function of the network is supposed to be deterministic right ? So I should get the same output everytime for the same image. <\/p>\n\n<p>On the other hand, when I run caffe on cpu by using  everything works perfectly, i.e, I get the same output each time\nThe code used and the outputs obtained are shown below. I am not able to understand what the problem is. Is it that the problem is caused due to GPU rounding off ? But the errors are very large. Or is it due to some issues with the latest CUDNN version ? Or is it something else altogether ?<\/p>\n\n<p><strong>Following is the CODE<\/strong><\/p>\n\n<h1>1) IMPORT libraries<\/h1>\n\n<pre><\/pre>\n\n<h1>2) IMPORT Caffe Models and define utility functions<\/h1>\n\n<pre><\/pre>\n\n<h1>3) LOADING Image and setting constants<\/h1>\n\n<pre><\/pre>\n\n<h1>4) Setting the source image and making the forward pass to obtain fc7 activation features<\/h1>\n\n<pre><\/pre>\n\n<p><strong>FOLLOWING is the output that I obtained for 'print dst.data' when I ran the above code multiple times<\/strong><\/p>\n\n<h1>output on 1st execution of code<\/h1>\n\n<pre><\/pre>\n\n<h1>output on 2nd execution of code<\/h1>\n\n<pre><\/pre>\n\n<h1>output on 3rd execution of code<\/h1>\n\n<pre><\/pre>\n\n<h1>output on 4th execution of code<\/h1>\n\n<pre><\/pre>\n\n<p>The output values keep becoming larger and larger and then again become smaller after some time. I am not able to understand the issue.<\/p>\n","answers":[{"AnswerId":"31654008","CreationDate":"2015-07-27T13:10:33.560","ParentId":null,"OwnerUserId":"5160755","Title":null,"Body":"<p>Switch your network to Test mode to prevent the effect of dropout which is non-deterministic and needed for training mode.<\/p>\n\n<p>Add the following line right after initializing your network:<\/p>\n\n<p>net.set_phase_test()<\/p>\n\n<p>So that you'll always have the same results.<\/p>\n\n<p>Soner<\/p>\n"}]}
{"QuestionId":31496020,"AnswerCount":4,"Tags":"<python><opencv><caffe>","CreationDate":"2015-07-18T22:03:02.830","AcceptedAnswerId":null,"OwnerUserId":5125298.0,"Title":"ImportError on cv2.so","Body":"<p>I am trying to run fast-rcnn on a cluster, where cv2.so is not installed for public use. So I directly move the cv2.so into a PATH, but it turns as:<\/p>\n\n<p><strong>\/lib64\/libc.so.6: version `GLIBC_2.14' not found<\/strong> <\/p>\n\n<p>So I have to install the opencv on my local path again, this time it says:<\/p>\n\n<p><strong>ImportError: \/home\/username\/.local\/lib\/python2.7\/site-packages\/cv2.so: undefined symbol: _ZN2cv11arrowedLineERNS_3MatENS_6Point_IiEES3_RKNS_7Scalar_IdEEiiid<\/strong><\/p>\n\n<p>This really confused me, could anyone give me a hand?<\/p>\n","answers":[{"AnswerId":"42237346","CreationDate":"2017-02-14T22:22:38.400","ParentId":null,"OwnerUserId":"3633525","Title":null,"Body":"<p>I know this is a little late, but I just got this same error with python 2.7 and opencv 3.1.0 on Ubuntu.  Turns out I had to reinstall <code>opencv-python<\/code>.  Running <code>sudo pip install opencv-python<\/code> did the trick.<\/p>\n"},{"AnswerId":"32407947","CreationDate":"2015-09-04T23:31:43.283","ParentId":null,"OwnerUserId":"4965916","Title":null,"Body":"<p>I ran into the same issue, but for me <code>PYTHONPATH<\/code> looked something like:<\/p>\n\n<pre><code>PYTHONPATH=\/usr\/local\/lib\/python2.7\/dist-packages:\/opt\/opencv2.4.9\/lib\/python2.7\/dist-packages\n<\/code><\/pre>\n\n<p>Removing <code>\/opt\/opencv2.4.9\/lib\/python2.7\/dist-packages<\/code> from the path provided the fix.<\/p>\n"},{"AnswerId":"31497506","CreationDate":"2015-07-19T02:43:11.973","ParentId":null,"OwnerUserId":"5125298","Title":null,"Body":"<p>The problem has been solved by some tryings.<\/p>\n\n<p>Since I installed under my \/.~local path, it should be noticed that [include],[bin] and [lib] should all point to the local version by modifying the bashrc.<\/p>\n\n<p>I just change the lib path while the other 2 paths remained unchanged, which point to the cluster's opencv version 2.4.9.(Mine is 2.4.11)<\/p>\n"},{"AnswerId":"46825407","CreationDate":"2017-10-19T08:14:03.990","ParentId":null,"OwnerUserId":"3930957","Title":null,"Body":"<p>After struggling with the above solutions, the following one (<a href=\"https:\/\/github.com\/CharlesShang\/FastMaskRCNN\/issues\/111\" rel=\"nofollow noreferrer\">source<\/a>) solved my problem:<\/p>\n\n<pre><code>sudo pip install --upgrade opencv-python\n<\/code><\/pre>\n"}]}
{"QuestionId":31507865,"AnswerCount":1,"Tags":"<image-processing><neural-network><directory-structure><deep-learning><caffe>","CreationDate":"2015-07-20T01:54:50.050","AcceptedAnswerId":31509621.0,"OwnerUserId":4282823.0,"Title":"Preparing image dataset for input into Caffe deep learning","Body":"<p>I know the first step is to create two file lists with the corresponding labels, one for the training and one for the test set. Suppose the former is called train.txt and the latter val.txt. The paths in these file lists should be relative. The labels should start at 0 and look similar to this:<\/p>\n\n<pre><\/pre>\n\n<p>For each of these two sets, we will create a separate LevelDB. Is this formatted as a text file? I thought I would create a directory with several subdirectories for each of my classes. Do I manually have to create a text file?<\/p>\n","answers":[{"AnswerId":"31509621","CreationDate":"2015-07-20T05:54:39.490","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Please see <a href=\"https:\/\/stackoverflow.com\/a\/31431716\/1714410\">this tutorial<\/a> on how to use <code>convert_imageset<\/code> to build <code>levelDb<\/code> or <code>lmdb<\/code> datasets for caffe's training. <\/p>\n\n<p>As you can see from these instruction it does not matter how you arrange the image files on your disk (same folder\/different folders...) as long as you have the correct paths in your <code>'train.txt'<\/code>\/<code>'val.txt'<\/code> files relative to <code>'\/path\/to\/jpegs\/'<\/code> argument. But if you want to use <code>convert_imageset<\/code> tool, you'll have to create a text file listing all the images you want to use.<\/p>\n"}]}
{"QuestionId":31509756,"AnswerCount":1,"Tags":"<image><format><theano><deep-learning>","CreationDate":"2015-07-20T06:05:51.597","AcceptedAnswerId":null,"OwnerUserId":5121855.0,"Title":"How to change a dataset of images to train-images-idx3-ubyte format","Body":"<p>I have 10000 images. I want to convert them to a format like 'train-images-idx3-ubyte'. This format comes from <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow\">here<\/a>. I want them to use the deep learning methods described <a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\" rel=\"nofollow\">here<\/a> <\/p>\n\n<p>I appreciate any help<\/p>\n","answers":[{"AnswerId":"31510287","CreationDate":"2015-07-20T06:49:18.253","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Take a look at how these files are loaded <a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/load.py#L17\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>The use of <code>numpy.fromfile<\/code> indicates that the data are simply saved as raw bytes of a specific dtype. You can achieve this using <code>numpy.tofile<\/code>.<\/p>\n\n<p>However, make sure that this is really what you want to do. If you want to use certain networks on other images, these images will likely need to be of exactly the same size. It is worth digging further into the tutorials - after a while the transposition to other datasets will become easier.<\/p>\n"}]}
{"QuestionId":31509912,"AnswerCount":2,"Tags":"<image><lua><data-visualization><convolution><torch>","CreationDate":"2015-07-20T06:18:53.833","AcceptedAnswerId":31510973.0,"OwnerUserId":4357605.0,"Title":"Visualize images in intermediate layers in torch (lua)","Body":"<p>In the conv-nets model, I know how to visualize the filters, we can do itorch.image(model:get(1).weight)<\/p>\n\n<p>But how could I efficiently visualize the output images after the convolution? especially those images in the second or third layer in a deep neural network?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"31510973","CreationDate":"2015-07-20T07:32:37.617","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Similarly to weight, you can use:<\/p>\n\n<pre><code>itorch.image(model:get(1).output)\n<\/code><\/pre>\n"},{"AnswerId":"32471992","CreationDate":"2015-09-09T06:05:21.330","ParentId":null,"OwnerUserId":"309653","Title":null,"Body":"<p>To visualize the weights:<\/p>\n\n<pre><code>-- visualizing weights\nn = nn.SpatialConvolution(1,64,16,16)\nitorch.image(n.weight)\n<\/code><\/pre>\n\n<p>To visualize the feature maps:<\/p>\n\n<pre><code>-- initialize a simple conv layer\nn = nn.SpatialConvolution(1,16,12,12)\n\n-- push lena through net :)\nres = n:forward(image.rgb2y(image.lena())) \n\n-- res here is a 16x501x501 volume. We view it now as 16 separate sheets of size 1x501x501 using the :view function\nres = res:view(res:size(1), 1, res:size(2), res:size(3))\nitorch.image(res)\n<\/code><\/pre>\n\n<p>For more: <a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/1_get_started.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/1_get_started.ipynb<\/a><\/p>\n"}]}
{"QuestionId":31511115,"AnswerCount":1,"Tags":"<output><convolution><theano><deep-learning>","CreationDate":"2015-07-20T07:40:36.567","AcceptedAnswerId":null,"OwnerUserId":4744673.0,"Title":"How to find the output of convolutional_mlp of Theano tutorial based on python","Body":"<p>I have 10000 images fed to CNN in Theano tutorial described <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>In the classification step, I want to classify those images into 40 classes. So, the number of units in the last layer would be 40. I want to get the predicted value from there. Layer3 calls 'LogisticRegression' function that is available in this package. I think CNN goes to   'LogisticRegression' function to evaluate the predicted values. How can I access to these value?<br>\nThe info related to layers are:<\/p>\n\n<pre><\/pre>\n\n<p>I think, if we could get output from 'layer3', it would be great. I want this value, because I want to count how many samples in each class have been predicted accurately.\nIs there any body to help<\/p>\n","answers":[{"AnswerId":"34728844","CreationDate":"2016-01-11T18:49:37.553","ParentId":null,"OwnerUserId":"5775130","Title":null,"Body":"<p>I'm not too familiar with Theano but the reason you don't have access to layer0 is because when training the convolutional_mlp, you have to not only save layer3, you have to save all the layers. (Alternatively, you can choose to save the parameters from each layer and recreate them).<\/p>\n\n<p>For example, in the while loop you can add the following:<\/p>\n\n<pre><code>with gzip.open('.\/testing\/model.pkl.gz', 'w') as f:\n    cPickle.dump([layer0_input, layer0, layer1, layer2_input, layer2, layer3], f)\n<\/code><\/pre>\n\n<p>Then you can do something like the following as a predict function. <\/p>\n\n<pre><code>  def predict(model='.\/testing\/model.pkl.gz', \n            testset='.\/testing\/testset.pkl.gz',\n            batch_size=5):\n\n      \"\"\" Load a trained model and use it to predict labels.\n\n      :type model: Layers to accept inputs and produce outputs.\n      \"\"\"\n\n      # Load the saved model.\n      classifiers = cPickle.load(gzip.open(model))\n\n      # Pick out the individual layer\n      layer0_input = classifiers[0]\n      layer0 = classifiers[1]\n      layer1 = classifiers[2]\n      layer2_input = classifiers[3]\n      layer2 = classifiers[4]\n      layer3 = classifiers[5]\n\n      # Apply it to our test set\n      testsets = load_data(testset)\n      test_set_x = testsets.get_value()\n\n      # compile a predictor function\n      index = T.lscalar()\n\n      predict_model = theano.function(\n          [layer0_input],\n          layer3.y_pred,\n      )\n\n      predicted_values = predict_model(\n          test_set_x[:batch_size].reshape((batch_size, 1, 28, 23))\n      )\n\n      print('Prediction complete.')\n      return predicted_values\n<\/code><\/pre>\n\n<p>The annoying thing is that by design of the tutorial, you have to pass in batches of the same size as you trained by. Here I only did one batch but you'd want to loop through all the batches in the test_set_x. <\/p>\n"}]}
{"QuestionId":31530360,"AnswerCount":1,"Tags":"<c><lua><luajit><torch>","CreationDate":"2015-07-21T04:25:24.140","AcceptedAnswerId":31530405.0,"OwnerUserId":678392.0,"Title":"How does this C code (from lua library, Torch) even compile\/work?","Body":"<p>See <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/generic\/Tanh.c\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/generic\/Tanh.c<\/a><\/p>\n\n<p>For example,<\/p>\n\n<pre><\/pre>\n\n<p>First, it I don't know how to interpret the first line:  <\/p>\n\n<pre><\/pre>\n\n<p>What are the arguments here? What does Tanh_updateOutput refer to? Does \"nn_\" have special meaning? <\/p>\n\n<p>Second, \"TH_TENSOR_APPLY2\" and \"THTensor_(...)\" are both used but I don't see where they are defined? There are no other includes in this file? <\/p>\n","answers":[{"AnswerId":"31530405","CreationDate":"2015-07-21T04:29:19.847","ParentId":null,"OwnerUserId":"2357112","Title":null,"Body":"<p><code>nn_<\/code> is a macro. You can find the definition by searching the repository for <code>\"#define nn_\"<\/code>; it's in <a href=\"https:\/\/github.com\/torch\/nn\/blob\/fdfcd12d789a885458222e4f7475e74a3ebc516f\/init.c\"><code>init.c<\/code><\/a>:<\/p>\n\n<pre><code>#define nn_(NAME) TH_CONCAT_3(nn_, Real, NAME)\n<\/code><\/pre>\n\n<p>You can keep following the chain of macro definitions, and you'll probably end up with some token pasting thing that makes <code>nn_(Tanh_updateOutput)<\/code> expand to the name of the function.<\/p>\n\n<p>(It's weird that <code>generic\/Tanh.c<\/code> doesn't have any includes; <code>generic\/Tanh.c<\/code> must be included by some other file. That's unusual for <code>.c<\/code> files.)<\/p>\n"}]}
{"QuestionId":31531807,"AnswerCount":1,"Tags":"<python><hash><computer-vision><caffe><locality-sensitive-hash>","CreationDate":"2015-07-21T06:27:51.773","AcceptedAnswerId":null,"OwnerUserId":196048.0,"Title":"Trouble shooting locality sensitive hash","Body":"<p>I am using <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a>, a deep neural network <a href=\"http:\/\/caffe.berkeleyvision.org\" rel=\"nofollow\">library<\/a>, to generate image features for image based retrieval. The particular network I am using generates a 4096 dimensional feature. <\/p>\n\n<p>I am using <a href=\"https:\/\/github.com\/kayzh\/LSHash\" rel=\"nofollow\">LSHash<\/a> to generate hash buckets from the features. When I do a brute for comparison of all available feature, sorting images by euclidean distance, I find the features represent image similarity well. When I use LSHash, however, I find that similar features rarely land in the same bucket. <\/p>\n\n<p>Are the source features too large for use with LSH? Are there other ways to reduce the dimensions of the image features before attempting to hash them?<\/p>\n","answers":[{"AnswerId":"31532224","CreationDate":"2015-07-21T06:55:33.420","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you are looking for intelligent dimensionality reduction, you can simply add another <code>\"InnerProduct\"<\/code> layer on top of your net with lower output dimension.<br>\nTo train only this layer without altering the rest of the weights you can set the <code>lr_mult<\/code> values for all the layers (apart from the new one) to zero thus training (aka \"finetuning\") only the top dim-reduction layer.<\/p>\n"}]}
{"QuestionId":31549129,"AnswerCount":0,"Tags":"<python><cuda><theano>","CreationDate":"2015-07-21T20:36:24.520","AcceptedAnswerId":null,"OwnerUserId":5140643.0,"Title":"Importing theano - nvcc : fatal error : nvcc cannot find a supported version of Microsoft Visual Studio","Body":"<p>I got this warning\/error message whenever I try importing theano-based packages for the first time (it disappear the second time...). I'm wondering how to get rid of it or disable it since it seems not to affect anything except for being ugly... Thanks in advance for your help!<\/p>\n\n<p>I have Visual Studio 2010 and C++ installed and added the path of cl.exe into the system path. Below is the original error message:<\/p>\n\n<p>===============================<\/p>\n\n<blockquote>\n  <p>nvcc : fatal error : nvcc cannot find a supported version of Microsoft\n  Visual Studio. Only the versions 2008, 2010, and 2012 are supported\n  ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ('nvcc\n  return status', -1, 'for cmd', 'nvcc -shared -O3 -Xlinker \/DEBUG -D\n  HAVE_ROUND -m64 -Xcompiler\n  -DCUDA_NDARRAY_CUH=47047b939bf764f5977458e316cf235a,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,\/Zi,\/MD\n  -IC:\\Anaconda\\lib\\site-packages\\theano\\sandbox\\cuda -IC:\\Anaconda\\lib\\site-packages\\numpy\\core\\include -IC:\\Anaconda\\include -IC:\\Anaconda\\lib\\site-packages\\theano\\gof -o C:\\Users\\YDuan\\AppData\\Local\\Theano\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.9-64\\cuda_ndarray\\cuda_ndarray.pyd\n  mod.cu -LC:\\Anaconda\\libs -LC:\\Anaconda -lpython27 -lcublas\n  -lcudart') ERROR:theano.sandbox.cuda:Failed to compile cuda_ndarray.cu: ('nvcc return status', -1, 'for cmd', 'nvcc -shared\n  -O3 -Xlinker \/DEBUG -D HAVE_ROUND -m64 -Xcompiler -DCUDA_NDARRAY_CUH=47047b939bf764f5977458e316cf235a,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,\/Zi,\/MD\n  -IC:\\Anaconda\\lib\\site-packages\\theano\\sandbox\\cuda -IC:\\Anaconda\\lib\\site-packages\\numpy\\core\\include -IC:\\Anaconda\\include -IC:\\Anaconda\\lib\\site-packages\\theano\\gof -o C:\\Users\\YDuan\\AppData\\Local\\Theano\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.9-64\\cuda_ndarray\\cuda_ndarray.pyd\n  mod.cu -LC:\\Anaconda\\libs -LC:\\Anaconda -lpython27 -lcublas\n  -lcudart')<\/p>\n  \n  <p>['nvcc', '-shared', '-O3', '-Xlinker', '\/DEBUG', '-D HAVE_ROUND',\n  '-m64', '-Xcompiler',\n  '-DCUDA_NDARRAY_CUH=47047b939bf764f5977458e316cf235a,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,\/Zi,\/MD',\n  '-IC:\\Anaconda\\lib\\site-packages\\theano\\sandbox\\cuda',\n  '-IC:\\Anaconda\\lib\\site-packages\\numpy\\core\\include',\n  '-IC:\\Anaconda\\include',\n  '-IC:\\Anaconda\\lib\\site-packages\\theano\\gof', '-o',\n  'C:\\Users\\YDuan\\AppData\\Local\\Theano\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.9-64\\cuda_ndarray\\cuda_ndarray.pyd', 'mod.cu', '-LC:\\Anaconda\\libs', '-LC:\\Anaconda', '-lpython27',\n  '-lcublas', '-lcudart']<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":31549838,"AnswerCount":1,"Tags":"<python><anaconda><theano>","CreationDate":"2015-07-21T21:21:49.797","AcceptedAnswerId":31556146.0,"OwnerUserId":3775577.0,"Title":"Import error while using scikit-neuralnetwork","Body":"<p>I'm using Anaconda with Python 3.4 64 bit on windows 8<\/p>\n\n<p>While trying to use the package scikit-neuralnetwork, this line of code raises the following exception:<\/p>\n\n<pre><\/pre>\n\n<p>I installed the GCC and g++ (4.8.1) compilers and added them correctly to PATH, however the following error continues to appear:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31556146","CreationDate":"2015-07-22T07:06:43.333","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This error is often solved by running the following, as mentioned in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">documentation<\/a>.<\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n\n<p>Since you've now installed your own installation of GCC, you may need to play around with environment variables, especially <code>PATH<\/code> to get things working.<\/p>\n"}]}
{"QuestionId":31556268,"AnswerCount":2,"Tags":"<python><neural-network><xor><keras>","CreationDate":"2015-07-22T07:13:16.683","AcceptedAnswerId":null,"OwnerUserId":5142261.0,"Title":"How to use keras for XOR","Body":"<p>I want to practice keras by code a xor, but the result is not right, the followed is my code, thanks for everybody to help me.<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Output<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31884273","CreationDate":"2015-08-07T18:20:16.793","ParentId":null,"OwnerUserId":"12981","Title":null,"Body":"<p>If I increase the number of epochs in your code to 50000 it does often converge to the right answer for me, just takes a little while :)<\/p>\n\n<p>It does often get stuck, though. I get better convergence properties if I change your loss function to 'mean_squared_error', which is a smoother function.<\/p>\n\n<p>I get still faster convergence if I use the Adam or RMSProp optimizers. My final compile line, which works:<\/p>\n\n<pre><code>model.compile(loss='mse', optimizer='adam')\n...\nmodel.fit(train_data, label, nb_epoch = 10000,batch_size = 4,verbose = 1,shuffle=True,show_accuracy = True)\n<\/code><\/pre>\n"},{"AnswerId":"41686065","CreationDate":"2017-01-16T22:31:24.203","ParentId":null,"OwnerUserId":"7427701","Title":null,"Body":"<p>I used a single hidden layer with 4 hidden nodes, and it almost always converges to the right answer within 500 epochs.  I used sigmoid activations.<\/p>\n"}]}
{"QuestionId":31558162,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-07-22T08:42:39.370","AcceptedAnswerId":31558574.0,"OwnerUserId":5130560.0,"Title":"updates argument in theano functions","Body":"<p>What does the \"updates\" argument do when called this way?<\/p>\n\n<pre><\/pre>\n\n<p>All the documentation I have seen about the \"updates\" argument in theano functions talk about pairs of the form (shared variables, expression used to update the shared variable). However, here there is only an expression so how to I know which shared variable is updated?<\/p>\n\n<p>I guess the shared variable is somehow implicit but  and  both depends on different shared variables:<\/p>\n\n<pre><\/pre>\n\n<p>This code comes from  in <a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lstm.html<\/a><\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"31558574","CreationDate":"2015-07-22T09:01:32.750","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It is correct to think that <code>updates<\/code> should be a list (or dictionary) of key value pairs where the key is a shared variable and the value is a symbolic expression describing how to update the corresponding shared variable.<\/p>\n\n<p>These two lines create the pairs:<\/p>\n\n<pre><code>zgup = [(zg, g) for zg, g in zip(zipped_grads, grads)]\nrg2up = [(rg2, 0.95 * rg2 + 0.05 * (g ** 2))\n         for rg2, g in zip(running_grads2, grads)]\n<\/code><\/pre>\n\n<p><code>zipped_grads<\/code> and <code>running_grads2<\/code> were created in the previous lines are each just a list of shared variables. Here, those shared variables are linked to updates using the Python <code>zip<\/code> function, which emits a list of pairs. In fact, the first of these lines could be replaced with<\/p>\n\n<pre><code>zgup = zip(zipped_grads, grads)\n<\/code><\/pre>\n\n<p>This code is quite complex because it is implementing the AdaDelta update mechanism. If you want to see how <code>updates<\/code> works in a simpler setting, take a look at the basic stochastic gradient descent update in the <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">Theano MLP tutorial<\/a>.<\/p>\n\n<pre><code>updates = [\n        (param, param - learning_rate * gparam)\n        for param, gparam in zip(classifier.params, gparams)\n    ]\n<\/code><\/pre>\n"}]}
{"QuestionId":31573856,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-07-22T20:57:46.247","AcceptedAnswerId":31579778.0,"OwnerUserId":87240.0,"Title":"theano: row-wise outer product between two matrices","Body":"<p>I'm trying to compute the row-wise outer-product between two matrices in theano, without using scan. I can do this in numpy by using einsum, which isn't available in theano.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31579778","CreationDate":"2015-07-23T06:34:01.173","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>This should be doable using some reshaping: Many of the simple einsum operations boil down to that. The complicated ones don't.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\na = np.array([[1,1,1],[2,2,2]]).astype('float32')\nb = np.array([[3,3,3,3],[4,4,4,4]]).astype('float32')\n\nA = T.fmatrix()\nB = T.fmatrix()\n\nC = A[:, :, np.newaxis] * B[:, np.newaxis, :]\n\nprint C.eval({A:a, B:b})\n<\/code><\/pre>\n\n<p>results in<\/p>\n\n<pre><code>[[[ 3.  3.  3., 3.]\n  [ 3.  3.  3., 3.]\n  [ 3.  3.  3.. 3.]]\n\n [[ 8.  8.  8., 8.]\n  [ 8.  8.  8., 8.]\n  [ 8.  8.  8., 8.]]]\n<\/code><\/pre>\n"}]}
{"QuestionId":31588281,"AnswerCount":0,"Tags":"<python><gpu><theano><numba><numba-pro>","CreationDate":"2015-07-23T13:07:31.900","AcceptedAnswerId":null,"OwnerUserId":1088305.0,"Title":"What is the main difference between Numba Pro and Theano\/pyautodiff for GPU calculations?","Body":"<p>Both Numba Pro and pyautodiff based on Theano supports conversion of Python code into GPU machine code. Theano will also allow symbolic derivation of the resulted syntax tree, but this is outside the scope of my question.<\/p>\n\n<p>My question is whether there are technical limitations in one or the other framework, which would make the code less efficient.<\/p>\n","answers":[]}
{"QuestionId":31612074,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-07-24T13:39:32.137","AcceptedAnswerId":null,"OwnerUserId":200340.0,"Title":"Theano continue training","Body":"<p>I am looking for some suggestions about how to do continue training in theano. For example, I have the following:<\/p>\n\n<pre>\nclassifier = my_classifier()\n\ncost = ()\nupdates = []\ntrain_model = theano.function(...)\neval_model = theano.function(...)\n\nbest_accuracy = 0\nwhile (epoch &lt; n_epochs):\n\n    train_model()\n\n    current_accuracy = eval_model()\n    if current_accuracy > best_accuracy:\n        save classifier or save theano functions?\n        best_accuracy = current_accuracy\n    else:\n        load saved classifier or save theano functions?\n        if we saved classifier previously, do we need to redefine train_model and eval_model functions?\n\n    epoch+=1\n\n#training is finished\nsave classifier\n<\/pre>\n\n<p>I want to save the current trained model if it has higher accuracy than previously trained models, and load the saved model later if the current trained model accuracy is lower than the best accuracy. <\/p>\n\n<p>My questions are:<\/p>\n\n<p>When saving, should I save the classifier, or theano functions? <\/p>\n\n<p>If the classifier needs to be saved, do I need to redefine theano functions when loading it, since classifier is changed. <\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"31778280","CreationDate":"2015-08-03T01:32:00.797","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>When pickling models, it is always better to save the parameters and when loading re-create the shared variable and rebuild the graph out of this. This allow to swap the device between CPU and GPU.<\/p>\n\n<p>But you can pickle Theano functions. If you do that, pickle all associated function at the same time. Otherwise, they will have each of them a different copy of the shared variable. Each call to load() will create new shared variable if they where pickled. This is a limitation of pickle.<\/p>\n"}]}
{"QuestionId":31615990,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><nvidia-digits>","CreationDate":"2015-07-24T17:00:36.960","AcceptedAnswerId":null,"OwnerUserId":2284821.0,"Title":"Caffe web demo error when running a model trained on Digits","Body":"<p>I trained a neural network model  on Digits and it seemed to run fine there.<br>\nThen i exported the trained model files and  copied them into a different system running the standard caffe web demo. \nI hoped to just be able to plug those files in and have them run in Caffe but i am getting an error.<\/p>\n\n<p>Specifically I copied my model into bvlc_reference_caffenet.caffemodel, the  deploy.prototxt into deploy.prototxt, and the  mean.binaryproto into the  ilsvrc_2012_mean.npy file. \nHowever when I try to run it , it appears to not like the format of the mean.binaryproto file as indicated by the error message:<\/p>\n\n<pre><\/pre>\n\n<p>what am I doing wrong here? Do I need to process the mean.binaryproto file from Digits somehow before I use it with caffe?<\/p>\n","answers":[{"AnswerId":"33775703","CreationDate":"2015-11-18T09:05:26.887","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to convert the <code>.binaryproto<\/code> file to a numpy file.\nThere is a nice example <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/290\" rel=\"nofollow\">here<\/a> using <code>caffe.io<\/code> and <code>caffe.proto<\/code>.<\/p>\n"}]}
{"QuestionId":31617486,"AnswerCount":1,"Tags":"<python><caffe><kaggle>","CreationDate":"2015-07-24T18:34:47.860","AcceptedAnswerId":31634295.0,"OwnerUserId":864128.0,"Title":"use caffe to train Lenet with CSV data","Body":"<p>Excuse me, I have a question on using caffe for hd data? I try to run an example on the Kaggle mnist csv data with the  following steps<\/p>\n\n<ol>\n<li><p>use  to convert it to h5 data. (I use the caffe-example.py to convert)<\/p><\/li>\n<li><p>Then modify the lenet_train_test_prototxt and train it. I am quite at a loss of this step. <\/p><\/li>\n<\/ol>\n\n<p>The only change I made here is<\/p>\n\n<pre><\/pre>\n\n<p>how to change the lenet_train_test_prototxt to suit the data? Or also there are some other files I need to change? The error log is<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31634295","CreationDate":"2015-07-26T06:22:04.260","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I assume you have one hdf5 data file <code>'data\/mnist_train_h5.hd5'<\/code>. <\/p>\n\n<ol>\n<li><p>As you can see from the error message you got, <code>\"HDF5Data\"<\/code> layer does not support data transformation. Specifically, you cannot scale the data by the layer.<br>\nThus, any transformations you wish to have, you must apply them <em>yourself<\/em> during the creation of <code>'data\/mnist_train_h5.hd5'<\/code>.<\/p><\/li>\n<li><p><code>\"HDF5Data\"<\/code> layer does not accept <code>data_param<\/code>, but rather <code>hdf5_data_param<\/code> with a parameter <code>source<\/code> specifying a <em>list<\/em> of hd5 binary files. In your case you should prepare an extra <em>text<\/em> file <code>'data\/mnist_train_h5.txt'<\/code> with a single line:<\/p><\/li>\n<\/ol>\n\n<blockquote>\n  <p>data\/mnist_train_h5.hd5<\/p>\n<\/blockquote>\n\n<p>This text file will tell caffe to read <code>'data\/mnist_train_h5.hd5'<\/code>.<\/p>\n\n<p>The resulting layer should look like:<\/p>\n\n<pre><code>layer {\n  name: \"mnist\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label\"\n  hdf5_data_param {\n    source: \"data\/mnist_train_h5.txt\"\n    batch_size: 64\n  }\n  include {\n    phase: TRAIN\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":31623272,"AnswerCount":0,"Tags":"<python><arrays><numpy><python-imaging-library><caffe>","CreationDate":"2015-07-25T05:26:46.563","AcceptedAnswerId":null,"OwnerUserId":3785114.0,"Title":"How to save a jpegtran JPEG in a numpy same as with PIL?","Body":"<p>I have started using a new library in my Python program to process images due to speed. I was using PIL but now I am using jpegtran. My previous code I would save the image to a numpy array and that worked perfectly fine, but now since the datatype is different I am having problems creating the same numpy array.<\/p>\n\n<p>PIL code:<\/p>\n\n<pre><\/pre>\n\n<p>jpegtran code: <\/p>\n\n<pre><\/pre>\n\n<p>I need the original shape (2136, 3216, 3) for it to work with the rest of my code.<\/p>\n","answers":[]}
{"QuestionId":31627380,"AnswerCount":1,"Tags":"<python><keras><kaggle>","CreationDate":"2015-07-25T14:17:47.200","AcceptedAnswerId":null,"OwnerUserId":5155533.0,"Title":"Trying Kaggle Titanic with keras .. getting loss and valid_loss -0.0000","Body":"<p>Hi I am getting weird results for the following code for the problem posted here (<a href=\"https:\/\/www.kaggle.com\/c\/titanic\" rel=\"nofollow\">https:\/\/www.kaggle.com\/c\/titanic<\/a>) - <\/p>\n\n<pre><\/pre>\n\n<p>I am getting following results :<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to create a simple 3 layer network. Totally basic code.\nI have tried these kind of classification problems before using keras on kaggle. But this time getting this error.<\/p>\n\n<p>Is it overfitting due to less data.\nWhat I am missing ? Can someone help ?<\/p>\n","answers":[{"AnswerId":"40352913","CreationDate":"2016-11-01T01:25:22.950","ParentId":null,"OwnerUserId":"4120005","Title":null,"Body":"<p>Old post, but answering anyway in case someone else attempts Titanic with Keras.<\/p>\n\n<p>Your network may have too many parameters and too little regularization (e.g. dropout).<\/p>\n\n<p>Call model.summary() right before the model.compile and it will show you how many parameters your network has. Just between your two Dense layers you should have 512 X 512 = 262,144 paramters. That's a lot for 762 examples.<\/p>\n\n<p>Also you may want to use a sigmoid activation on the last layer and binary_cross entropy loss as you only have two output classes.<\/p>\n"}]}
{"QuestionId":31639331,"AnswerCount":2,"Tags":"<python><unix><machine-learning><neural-network><caffe>","CreationDate":"2015-07-26T16:31:52.323","AcceptedAnswerId":null,"OwnerUserId":4969461.0,"Title":"Caffe: Almost done but stuck at the last step","Body":"<p>I'm currently setting up caffe with python on my Macbook. I swear all the prerequisites are ok, but it returns error when I tried to build caffe. What's wrong? \nHere is the console:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Ok, Now the aforementioned problem have been solved, but there is another one, could someone help me?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31639789","CreationDate":"2015-07-26T17:17:56.663","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>I guess you have to install cuda dev. kit.<\/p>\n"},{"AnswerId":"46592024","CreationDate":"2017-10-05T17:57:27.453","ParentId":null,"OwnerUserId":"880837","Title":null,"Body":"<p>For anyone else that finds this. The CUDA dev kit is only needed if you have an Nvidia card on the machines you'll use this on. For most Mac users this likely isn't the case (check your hardware specs). If you don't have that kind of graphics card, you can't use CUDA. Just disable it before compiling Caffe:<\/p>\n\n<ul>\n<li>In your caffe dir, edit the Makefile.config<\/li>\n<li>Uncomment this line: <code>CPU_ONLY := 1<\/code> to indicate CUDA won't be needed.<\/li>\n<\/ul>\n"}]}
{"QuestionId":31643409,"AnswerCount":0,"Tags":"<theano><autoencoder>","CreationDate":"2015-07-27T00:29:34.913","AcceptedAnswerId":null,"OwnerUserId":864128.0,"Title":"How can the numerical values of Theano hidden layers be accessed efficiently, i.e. not eval()?","Body":"<p>Excuse me, is there any efficient way of extracting the numerical values of hidden layer in theano? I can do this work with .eval(), which is super slow.<\/p>\n\n<p>The code is (based on modification of dA.py)<\/p>\n\n<pre><\/pre>\n\n<p>It turns out .eval() is super slow, is there anyway I can improve it? I try to write it with theano.function, but I failed. Here is my code<\/p>\n\n<pre><\/pre>\n\n<p>It doesn't work, the error log is<\/p>\n\n<pre><\/pre>\n\n<p>So I have two questions here:<\/p>\n\n<ol>\n<li>How to write the function code here?<\/li>\n<li>Is there any efficient way to output the numerical values of hidden layers in theano?<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":31649216,"AnswerCount":3,"Tags":"<python><caffe><lmdb>","CreationDate":"2015-07-27T09:16:19.880","AcceptedAnswerId":null,"OwnerUserId":1300575.0,"Title":"Writing data to LMDB with Python very slow","Body":"<p>Creating datasets for training with <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"noreferrer\">Caffe<\/a> I both tried using HDF5 and LMDB. However, creating a LMDB is very slow even slower than HDF5. I am trying to write ~20,000 images.<\/p>\n\n<p>Am I doing something terribly wrong? Is there something I am not aware of?<\/p>\n\n<p>This is my code for LMDB creation:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see I am creating a transaction for every 1,000 images, because I thought creating a transaction for each image would create an overhead, but it seems this doesn't influence performance too much.<\/p>\n","answers":[{"AnswerId":"32576493","CreationDate":"2015-09-15T01:56:16.323","ParentId":null,"OwnerUserId":"4323148","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>DB_KEY_FORMAT = \"{:0&gt;10d}\"\ndb = lmdb.open(path, map_size=int(1e12))\n    curr_idx = 0\n    commit_size = 1000\n    with in_db_data.begin(write=True) as in_txn:\n        for curr_commit_idx in range(0, num_data, commit_size):\n            for i in range(curr_commit_idx, min(curr_commit_idx + commit_size, num_data)):\n                d, l = data[i], labels[i]\n                im_dat = caffe.io.array_to_datum(d.astype(float), label=int(l))\n                key = DB_KEY_FORMAT.format(curr_idx)\n                in_txn.put(key, im_dat.SerializeToString())\n                curr_idx += 1\n    db.close()\n<\/code><\/pre>\n\n<p>the code<\/p>\n\n<pre><code>with in_db_data.begin(write=True) as in_txn:\n<\/code><\/pre>\n\n<p>takes much time.<\/p>\n"},{"AnswerId":"37149186","CreationDate":"2016-05-10T21:21:23.713","ParentId":null,"OwnerUserId":"2617351","Title":null,"Body":"<p>In my experience, I've had <strong>50-100 ms writes to LMDB from Python<\/strong> writing Caffe data on ext4 hard disk on Ubuntu. <strong>That's why I use tmpfs<\/strong> (<strong>RAM disk<\/strong> functionality built into Linux) and get these writes done in around <strong>0.07 ms<\/strong>. You can make smaller databases on your ramdisk and copy them to a hard disk and later train on all of them. I'm making around 20-40GB ones as I have 64 GB of RAM.<\/p>\n\n<p>Some pieces of code to help you guys dynamically create, fill and move LMDBs to storage. Feel free to edit it to fit your case. It should save you some time getting your head around how LMDB and file manipulation works in Python.<\/p>\n\n<pre><code>import shutil\nimport lmdb\nimport random\n\n\ndef move_db():\n    global image_db\n    image_db.close();\n    rnd = ''.join(random.choice(string.ascii_uppercase + string.digits) for _ in range(5))\n    shutil.move( fold + 'ram\/train_images',  '\/storage\/lmdb\/'+rnd)\n    open_db()\n\n\ndef open_db():\n    global image_db\n    image_db    = lmdb.open(os.path.join(fold, 'ram\/train_images'),\n            map_async=True,\n            max_dbs=0)\n\ndef write_to_lmdb(db, key, value):\n    \"\"\"\n    Write (key,value) to db\n    \"\"\"\n    success = False\n    while not success:\n        txn = db.begin(write=True)\n        try:\n            txn.put(key, value)\n            txn.commit()\n            success = True\n        except lmdb.MapFullError:\n            txn.abort()\n            # double the map_size\n            curr_limit = db.info()['map_size']\n            new_limit = curr_limit*2\n            print '&gt;&gt;&gt; Doubling LMDB map size to %sMB ...' % (new_limit&gt;&gt;20,)\n            db.set_mapsize(new_limit) # double it\n\n...\n\nimage_datum                 = caffe.io.array_to_datum( transformed_image, label )\nwrite_to_lmdb(image_db, str(itr), image_datum.SerializeToString())\n<\/code><\/pre>\n"},{"AnswerId":"44697800","CreationDate":"2017-06-22T11:15:52.363","ParentId":null,"OwnerUserId":"590335","Title":null,"Body":"<p>LMDB writes are very sensitive to order - If you can sort the data before insertion speed will improve significantly<\/p>\n"}]}
{"QuestionId":31658771,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-07-27T16:46:05.100","AcceptedAnswerId":null,"OwnerUserId":5130560.0,"Title":"Issue with Theano scan function - TypeError: Cannot convert Type TensorType(float64, 3D)","Body":"<p>I am having some trouble with the Theano scan function and the following code:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the variable start is an integer that get a new value after each call of step_ and I want to get the sequences of its values after an arbitrary number of steps n_steps. If I run the code with n_steps = 1, everything works. However, for n_steps > 1, I get this error:<\/p>\n\n<blockquote>\n  <p>TypeError: Cannot convert Type TensorType(float64, 3D) (of Variable IncSubtensor{Set;:int64:}.0) into Type TensorType(float64, (False, True, False)). You can try to manually convert IncSubtensor{Set;:int64:}.0 into a TensorType(float64, (False, True, False)).<\/p>\n<\/blockquote>\n\n<p>I don't get where it comes from as none of my variable is a 3D tensor (I have checked with theano.printing.debugprinting and h and c are rows as expected and sample a scalar).<\/p>\n\n<p>Do you have any clue?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"31673948","CreationDate":"2015-07-28T10:40:23.447","ParentId":null,"OwnerUserId":"5130560","Title":null,"Body":"<p>Actually I have found the solution to my problem. I changed this<\/p>\n\n<pre><code>def _slice(_x, n, dim):\nif _x.ndim == 3:\n    return _x[:, :, n * dim:(n + 1) * dim]\nreturn _x[:, n * dim:(n + 1) * dim]\n<\/code><\/pre>\n\n<p>by this<\/p>\n\n<pre><code>    def _slice(_x, n, dim):\n    if _x.ndim == 3:\n        return _x[:, :, n * dim:(n + 1) * dim]\n    if _x.ndim == 2:\n        return _x[:, n * dim:(n + 1) * dim]\n    return _x[n * dim:(n + 1) * dim]\n<\/code><\/pre>\n\n<p>and this<\/p>\n\n<pre><code>x_ = tensor.dot(emb[None,:], tparams[_p(prefix, 'W')]) + tparams[_p(prefix, 'b')]\n<\/code><\/pre>\n\n<p>by this<\/p>\n\n<pre><code>    x_ = tensor.dot(emb, tparams[_p(prefix, 'W')]) + tparams[_p(prefix, 'b')]\n<\/code><\/pre>\n\n<p>This makes <code>x_, h_ and c_<\/code> theano vector instead of rows as previously and removes the error (though I am not exactly sure why).<\/p>\n\n<p>Of course I've also update the call to <code>scan<\/code><\/p>\n\n<pre><code>    rval, updates = theano.scan(_step,\n                            outputs_info=[start, tensor.alloc(numpy_floatX(0.),\n                                                       dim_proj),\n                                          tensor.alloc(numpy_floatX(0.),\n                                                       dim_proj)],\n                            name=_p(prefix, '_layers'),\n                            n_steps=2)\n<\/code><\/pre>\n"}]}
{"QuestionId":31659899,"AnswerCount":1,"Tags":"<matlab><neural-network><deep-learning><caffe><matcaffe>","CreationDate":"2015-07-27T17:50:10.927","AcceptedAnswerId":31668306.0,"OwnerUserId":5161771.0,"Title":"Trying to use matcaffe for the interface between MATLAB and Caffe, but cannot find caffe_.cpp","Body":"<p>With  being a private function, when I call functions like , there is always an error telling me it cannot find .<\/p>\n\n<p>So how to use that in MATLAB?<\/p>\n","answers":[{"AnswerId":"31668306","CreationDate":"2015-07-28T05:51:25.880","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can find <code>'caffe_.cpp'<\/code> under <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/matlab\/%2Bcaffe\/private\" rel=\"nofollow\"><code>matlab\/+caffe\/private\/<\/code><\/a>.<br>\nMake sure you cloned caffe git properly and that you built the matlab interface:<\/p>\n\n<pre><code>~$ make matcaffe\n<\/code><\/pre>\n"}]}
{"QuestionId":31664988,"AnswerCount":1,"Tags":"<python-2.7><ssh><gnu-screen><theano><keras>","CreationDate":"2015-07-27T23:23:41.007","AcceptedAnswerId":31673239.0,"OwnerUserId":5106953.0,"Title":"Using screen session with Theano - race conditions","Body":"<p>When training a neural net implemented in Keras in a screen session, I appear to be running into race conditions with Theano.<\/p>\n\n<p>I proceed as follows. I ssh into the compute cluster I am using (which I am <em>not<\/em> a root user of).<\/p>\n\n<p>Then I run:<\/p>\n\n<pre><\/pre>\n\n<p>Then, once I'm in this screen session, I run the Python script which trains my model. I detach the screen (Ctrl+A+D), and when I do screen -r, everything is fine. However, if I exit my ssh session before I run screen -r, and run screen -r upon logging back in, then I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know why this happens? It's interesting that it only happens when I logout and try to run screen -r after logging in.<\/p>\n","answers":[{"AnswerId":"31673239","CreationDate":"2015-07-28T10:08:17.247","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>My guess is that your home directory is on a networked filesystem of some kind (e.g. AFS). If so, as soon as you end the session the filesystem security credentials are invalidated and the process, though it continues to run inside the screen, no longer has permission to work with files in the Theano cache directory <code>~\/.theano<\/code>. If this guess is correct then the problem is not a race condition.<\/p>\n\n<p>If the problem relates to AFS credential expiry then a solution is to use a credential cache with the <code>kinit<\/code> command (see the <code>-c<\/code> option in <a href=\"http:\/\/web.mit.edu\/kerberos\/krb5-1.12\/doc\/user\/user_commands\/kinit.html\" rel=\"nofollow\">http:\/\/web.mit.edu\/kerberos\/krb5-1.12\/doc\/user\/user_commands\/kinit.html<\/a>).<\/p>\n"}]}
{"QuestionId":31673350,"AnswerCount":1,"Tags":"<python><theano><pymc3><softmax>","CreationDate":"2015-07-28T10:13:04.867","AcceptedAnswerId":null,"OwnerUserId":5160585.0,"Title":"Softmax choice probabilities with Categorical in PyMC3","Body":"<p>I am trying to perform a parameter estimation for a single parameter of a softmax choice function in the following scenario:<\/p>\n\n<p>In each trial, three option values are given (e.g., [1,2,3]), and a subject makes a choice between the options (0, 1 or 2). The softmax function transforms option values into choice probabilities (vector of 3 probabilities, summing to 1), depending on a temperature parameter (here bound between 0 and 10).<\/p>\n\n<p>The choice in each trial is supposed to be modelled as a Categorical distribution with trial choice probabilities calculated from the softmax. Note that the choice probabilities of the Categorical depend on the option values and are therefore different in each trial.<\/p>\n\n<p>Here's what I came up with:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>This code fails for nTrials bigger than something like 50 with an extremely long warning \/ error message:<\/p>\n\n<p>Warning:<\/p>\n\n<pre><\/pre>\n\n<p>Error:<\/p>\n\n<pre><\/pre>\n\n<p>I am rather new to PyMC (and Theano) and I feel my implementation is really clunky and suboptimal. Any help and advice is strongly appreciated!<\/p>\n\n<p>Felix<\/p>\n\n<p>Edit: I've uploaded the code as a notebook, showing the warnings and error messages in full: <a href=\"http:\/\/nbviewer.ipython.org\/github\/moltaire\/softmaxPyMC\/blob\/master\/softmax_stackoverflow.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.ipython.org\/github\/moltaire\/softmaxPyMC\/blob\/master\/softmax_stackoverflow.ipynb<\/a><\/p>\n","answers":[{"AnswerId":"33740496","CreationDate":"2015-11-16T16:50:10.457","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>I found again this case. Just as a follow up, now not able able to reproduce it now. So I think it got fixed. We fixed related problem that could hve caused in some case this error.<\/p>\n\n<p>It work with g++ 4.5.1. If you have this problem update Theano to the development version. If that don't fix it, try to use a g++ more recent, this could be related to older g++ version.<\/p>\n"}]}
{"QuestionId":31673958,"AnswerCount":4,"Tags":"<cuda><artificial-intelligence><neural-network><nvcc><caffe>","CreationDate":"2015-07-28T10:40:49.407","AcceptedAnswerId":31674451.0,"OwnerUserId":4969461.0,"Title":"Caffe Installation: How to use libc++ instead of clang to power CUDA? (Mac)","Body":"<p>I know it dim-witted to do so, but I have to. 'Cause currently CUDA 7.0 doesn't support clang 7.0 while I'm using Xcode 7 beta, and it'll be virtually impossible for me to roll back to Xcode 6.0<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32642537","CreationDate":"2015-09-18T01:38:53.070","ParentId":null,"OwnerUserId":"4696622","Title":null,"Body":"<p>I had the same issue and I'm guessing lot's more will since Xcode 7.0 is available on App Store now, was going to comment on Jay's but I dont have enough reputation.<br>\nWhat you can do instead of downloading the whole 5gb Xcode 6.x is check Xcode under <\/p>\n\n<p>settings > locations > Command Line Tools <\/p>\n\n<p>if there is an option to use 6.4 just switch to that.  If not, install <em>JUST<\/em> the command line tools 6.4 <a href=\"https:\/\/developer.apple.com\/downloads\/?name=Xcode\" rel=\"nofollow\">here<\/a> and then the option should be available.<br>\nThe way he posted work's as well but then you need two versions of Xcode on your machine.<\/p>\n"},{"AnswerId":"32599934","CreationDate":"2015-09-16T04:41:09.890","ParentId":null,"OwnerUserId":"5158230","Title":null,"Body":"<p>I had the same issue. Instead of rolling back to Xcode 6 you can just install it alongside Xcode 7 by downloading an older release from Apple's <a href=\"https:\/\/developer.apple.com\/downloads\/?name=Xcode\" rel=\"nofollow\">Developer Download Center<\/a>. Give it a name different from Xcode 7's name (e.g. call it \"Xcode 6.4\") and they can live side by side without issue.<\/p>\n\n<p>Just make sure you set the command line tools in your Xcode preferences to the older version when you want to compile with CUDA.<\/p>\n"},{"AnswerId":"31674451","CreationDate":"2015-07-28T11:04:59.873","ParentId":null,"OwnerUserId":"678093","Title":null,"Body":"<p>The <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-getting-started-guide-for-mac-os-x\/\" rel=\"nofollow\">documentation<\/a> for CUDA 7.0 states that for Mac OS X <strong>Clang 6.0 (Xcode 6)<\/strong> is the most recent supported version.<\/p>\n"},{"AnswerId":"32980764","CreationDate":"2015-10-06T22:19:52.547","ParentId":null,"OwnerUserId":"5228839","Title":null,"Body":"<h1><code>nvcc<\/code> actual situation<\/h1>\n\n<p><code>nvcc<\/code> is <strong>parsed from a current clang binary version.<\/strong><br>\ne.g,<\/p>\n\n<pre><code>&gt; export\nPATH=\/usr\/local\/bin:\/usr\/bin:\/usr\/sbin:\/usr\/local\/var\/rbenv\/shims:\/sbin\n\n&gt; clang -v\nclang version 3.8.0 (git@github.com:llvm-mirror\/clang.git\n1082a41a5196e0fdddf1af1aa388af197cfc4514) (git@github.com:llvm\nmirror\/llvm.git 60fe48f86639ab4472d186bc97e7676f269cae18)\nTarget: x86_64-apple-darwin15.0.0\nThread model: posix\nInstalledDir: \/usr\/local\/bin\n\n&gt; make\n\/Developer\/NVIDIA\/CUDA-7.5\/bin\/nvcc -ccbin clang++ -I..\/..\/common\/inc  -m64  -Xcompiler -arch -Xcompiler x86_64  -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o asyncAPI.o -c asyncAPI.cu\nnvcc fatal   : The version ('30800') of the host compiler ('clang') is not supported\nMakefile:229: recipe for target 'asyncAPI.o' failed\nmake: *** [asyncAPI.o] Error 1\n<\/code><\/pre>\n\n<blockquote>\n  <p>nvcc fatal : The version ('30800') of the host compiler ('clang') is not supported<\/p>\n<\/blockquote>\n\n<p>That means...?<\/p>\n\n<pre><code>&gt; export PATH=\/Applications\/Xcode-6.4.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin:\/usr\/local\/bin:\/usr\/bin:\/usr\/sbin:\/usr\/local\/var\/rbenv\/shims:\/sbin\n\n&gt; CC=\/usr\/bin\/clang CXX=\/usr\/bin\/clang++ make\n\/Developer\/NVIDIA\/CUDA-7.5\/bin\/nvcc -ccbin clang++ -I..\/..\/common\/inc  -m64  -Xcompiler -arch -Xcompiler x86_64  -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o asyncAPI.o -c asyncAPI.cu\n\/Developer\/NVIDIA\/CUDA-7.5\/bin\/nvcc -ccbin clang++   -m64  -Xcompiler -arch -Xcompiler x86_64  -Xlinker -rpath -Xlinker \/Developer\/NVIDIA\/CUDA-7.5\/lib  -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o asyncAPI asyncAPI.o\nmkdir -p ..\/..\/bin\/x86_64\/darwin\/release\ncp asyncAPI ..\/..\/bin\/x86_64\/darwin\/release\n<\/code><\/pre>\n\n<p>Success.<br>\n...However, unknown actually what nvcc are using the clang of CXX.<\/p>\n\n<h1>Solution<\/h1>\n\n<p>So, Let's edit <code>clang<\/code> binary.<\/p>\n\n<pre><code>&gt; cp \/Applications\/Xcode-beta.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/clang \/tmp\/clang\n\n&gt; cd \/tmp\n\n&gt; vim .\/clang\n# or nvim, emacs, etc...\n\n# Searching `700.1.75` and replace `602.0.53`.  \n# In case of vim,\n:%s\/700.1.75\/602.0.53\/g\n:wq\n\n&gt; export PATH=\/tmp:\/usr\/local\/bin:\/usr\/bin:\/usr\/sbin:\/usr\/local\/var\/rbenv\/shims:\/sbin\n\n&gt; wihch clang\n\/tmp\/clang\n\n&gt; clang -v\nApple LLVM version 7.0.0 (clang-602.0.53)\nTarget: x86_64-apple-darwin15.0.0\nThread model: posix\n# Faking clang version 602.0.53 (Xcode 6.4's clang)\n# but, Actual implementation clang 7.0.0\n\n&gt; make\n\/Developer\/NVIDIA\/CUDA-7.5\/bin\/nvcc -ccbin clang++ -I..\/..\/common\/inc  -m64  -Xcompiler -arch -Xcompiler x86_64  -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o asyncAPI.o -c asyncAPI.cu\n\/Developer\/NVIDIA\/CUDA-7.5\/bin\/nvcc -ccbin clang++   -m64  -Xcompiler -arch -Xcompiler x86_64  -Xlinker -rpath -Xlinker \/Developer\/NVIDIA\/CUDA-7.5\/lib  -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o asyncAPI asyncAPI.o\nmkdir -p ..\/..\/bin\/x86_64\/darwin\/release\ncp asyncAPI ..\/..\/bin\/x86_64\/darwin\/release\n<\/code><\/pre>\n\n<p>Yeeeeees!<br>\n<code>nvcc<\/code> compiled using <code>clang version 7.0.0<\/code>.  <\/p>\n\n<p>In this way,<\/p>\n\n<pre><code>sudo xcode-select -s \/Applications\/Xcode-6.4.app\/Contents\/Developer\n<\/code><\/pre>\n\n<p>is <strong>not<\/strong> required.  <\/p>\n\n<h1>Be careful.<\/h1>\n\n<p>Now, Your <code>clang<\/code> is remains became <code>\/tmp\/clang<\/code>.<\/p>\n\n<h1>FYI<\/h1>\n\n<p>My gist article,<br>\n<a href=\"https:\/\/gist.github.com\/zchee\/7833bf67013e83523181\" rel=\"nofollow\">gist - How to compiling use CUDA nvcc with Xcode7.0 clang 7.0.0<\/a><\/p>\n"}]}
{"QuestionId":31683098,"AnswerCount":1,"Tags":"<python><numpy><theano><lasagne>","CreationDate":"2015-07-28T17:19:15.630","AcceptedAnswerId":31774478.0,"OwnerUserId":5165960.0,"Title":"numpy array from csv file for lasagne","Body":"<p>I started learning how to use theano with lasagne, and started with the mnist example. Now, I want to try my own example: I have a train.csv file, in which every row starts with 0 or 1 which represents the correct answer, followed by 773 0s and 1s which represent the input. I didn't understand how can I turn this file to the wanted numpy arrays in the load_database() function. this is the part from the original function for the mnist database:<\/p>\n\n<pre><\/pre>\n\n<p>and I need to get the X_train (the input) and the y_train (the beginning of every row) from my csv files.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"31774478","CreationDate":"2015-08-02T17:17:43.177","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You can use <code>numpy.genfromtxt()<\/code> or <code>numpy.loadtxt()<\/code> as follows:<\/p>\n\n<pre><code>from sklearn.cross_validation import KFold\n\nXy = numpy.genfromtxt('yourfile.csv', delimiter=\",\")\n\n# the next section provides the required\n# training-validation set splitting but \n# you can do it manually too, if you want\n\nskf = KFold(len(Xy))\n\nfor train_index, valid_index in skf:\n    ind_train, ind_valid = train_index, valid_index\n    break\n\nXy_train, Xy_valid = Xy[ind_train], Xy[ind_valid]\n\nX_train = Xy_train[:, 1:]\ny_train = Xy_train[:, 0]\n\nX_valid = Xy_valid[:, 1:]\ny_valid = Xy_valid[:, 0]\n\n\n...\n\n# you can simply ignore the test sets in your case\nreturn X_train, y_train, X_val, y_val #, X_test, y_test\n<\/code><\/pre>\n\n<p>In the code snippet we ignored passing the <strong><code>test<\/code><\/strong> set.<\/p>\n\n<p>Now you can import your dataset to the main modul or script or whatever, but be aware to remove all the test part from that too. <\/p>\n\n<p>Or alternatively you can simply pass the valid sets as <strong><code>test<\/code><\/strong> set:<\/p>\n\n<pre><code># you can simply pass the valid sets as `test` set\nreturn X_train, y_train, X_val, y_val, X_val, y_val\n<\/code><\/pre>\n\n<p>In the latter case we don't have to care about the main moduls sections refer to the excepted <strong><code>test<\/code><\/strong> set, but as scores (if have) you will get the the <code>validation scores<\/code> twice i.e. as <code>test scores<\/code>.<\/p>\n\n<p><strong>Note:<\/strong> I don't know, which mnist example is that one, but probably, after you prepared your data as above, you have to make further modifications in your trainer module too to suit to your data. For example: input shape of data, output shape i.e. the number of classes e.g. in your case the former is <code>773<\/code>, the latter is <code>2<\/code>.<\/p>\n"}]}
{"QuestionId":31688759,"AnswerCount":1,"Tags":"<python><macos><cuda><wrapper><caffe>","CreationDate":"2015-07-28T23:20:32.807","AcceptedAnswerId":31757650.0,"OwnerUserId":3543300.0,"Title":"Mac Caffe CUDA driver issue","Body":"<p>I'm trying to build caffe with the python wrapper on Mac OSX 10.0, but keep getting the following error when I execute the command: make runtest (make all -j8 and make test work fine).<\/p>\n\n<blockquote>\n  <p>Check failed: error == cudaSuccess (35 vs. 0) CUDA driver version is insufficient for CUDA runtime version<\/p>\n<\/blockquote>\n\n<p>I have updated the CUDA driver to the latest version online. I also tried uninstalling and reinstalling CUDA and the driver, but the error still persists. How can I solve this?<\/p>\n","answers":[{"AnswerId":"31757650","CreationDate":"2015-08-01T02:39:50.170","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>As was teased out in the comments, the basic problem here is an attempt to use CUDA on a GPU that does not support it (AMD Radeon ...).<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/tags\/cuda\/info\">CUDA<\/a> is a GPU programming technology that only runs on NVIDIA GPUs (ignoring emulators and the like.)<\/p>\n\n<p>To make forward progress, some possibilities might be:<\/p>\n\n<ol>\n<li>Switch to another machine that has an NVIDIA GPU.<\/li>\n<li>Modify the configuration of Caffe so that it does not use or depend on CUDA.<\/li>\n<\/ol>\n"}]}
{"QuestionId":31700179,"AnswerCount":1,"Tags":"<python><machine-learning><nlp><theano><deep-learning>","CreationDate":"2015-07-29T12:17:26.610","AcceptedAnswerId":31706487.0,"OwnerUserId":4763311.0,"Title":"Unable to make sense of how Theano works in RNN NLP for classification","Body":"<pre><\/pre>\n\n<p>Explanation of code:<\/p>\n\n<p>I know this will not be a good thing to do in stackoverflow. But I am struggling for more than a week to decode this code which is used to train  a Recurrent Neural Network. I am newbie to theano first of all.<\/p>\n\n<p>word_batch = array([[ -1,  -1,  -1, 194, 358, 463, 208]], dtype=int32)\nlabel_last_word = 126<\/p>\n\n<p>Thw word_batch is an index for a sentence like the following:<\/p>\n\n<p>'I am going to USA from England'<\/p>\n\n<p>Here the word_batch is a context window associated with one particular word say USA. So, if the context windows is seven the middle ( 194 ) in the word batch represent the index of that word in the dataset. I want to know, when I am passing this as argument to rnn.sentence_train , how the training is happen inside the RNNSLU class. I am confused with the usage of variables like idx, x inside that class. I know how this happens in theory, but unable to decode the theano part explicitly. If my question doesn't make sense, please let me know.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"31706487","CreationDate":"2015-07-29T16:48:10.333","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>rnn.sentence_train<\/code> is Theano function that has <code>updates=sentence_updates<\/code>. This means that on each call to <code>rnn.sentence_train<\/code> all of the shared variables in the <code>sentence_updates<\/code> dictionary's keys will be updated according to the symbolic update expressions in the corresponding <code>sentence_updates<\/code> dictionary values. Those expressions are all classical gradient descent (current parameter value - learning rate * gradient of cost with respect to parameter).<\/p>\n\n<p><code>idxs<\/code> is the symbolic placeholder for the input to the training function. In your example, <code>word_batch<\/code> fills in that placeholder when the training function is called.<\/p>\n"}]}
{"QuestionId":31710586,"AnswerCount":1,"Tags":"<python><lazy-evaluation><symbolic-math><theano>","CreationDate":"2015-07-29T20:35:53.133","AcceptedAnswerId":31711156.0,"OwnerUserId":1815451.0,"Title":"Lazy evaluation of .dot or other theano function","Body":"<p>I am very new at python and theano so this question may be silly. <\/p>\n\n<p>I've read in documentation that .dot produces symbolic tensor. I am debugging some program right now and I can't see TensorVariable without any container or anything and I don't know how to get the values from it. <\/p>\n","answers":[{"AnswerId":"31711156","CreationDate":"2015-07-29T21:09:11.410","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It isn't really possible to get a value from a symbolic variable because they don't have a value; instead they stand in for a value that is provided later.<\/p>\n\n<p>Consider the following example:<\/p>\n\n<pre><code>x = theano.tensor.matrix()\ny = theano.tensor.matrix()\nz = theano.dot(x, y)\nf = theano.function(inputs=[x, y], outputs=z)\na = numpy.array([[1,2,3],[4,5,6]])\nb = numpy.array([[1,2],[3,4],[5,6]])\nc = f(a, b)\n<\/code><\/pre>\n\n<p>Here, <code>x<\/code> and <code>y<\/code> are symbolic matrices. They don't have a value but they stand-in for some value that will be provided after the computation is compiled and executed. <code>a<\/code>, <code>b<\/code>, and <code>c<\/code> are concrete matrices with values. The <code>f = theano.function(...)<\/code> line compiles the computation and the <code>c = f(...)<\/code> executes that function, providing the value called <code>a<\/code> for <code>x<\/code> and providing the value called <code>b<\/code> for <code>y<\/code>; the return value, <code>c<\/code>, takes on the value computed by the symbolic expression <code>z<\/code>.<\/p>\n"}]}
{"QuestionId":31722238,"AnswerCount":2,"Tags":"<python><ubuntu><anaconda><caffe>","CreationDate":"2015-07-30T11:13:00.243","AcceptedAnswerId":31943616.0,"OwnerUserId":3633250.0,"Title":"Installing caffe on ubuntu 15.04 with anaconda 3 for python 3.4 - no module caffe found","Body":"<p>I am trying to install caffe on my ubuntu 15.04 with anaconda 3 (for python 3.4). I managed to install all requirements and I followed the instructions from official website. So I downloaded caffe-master and did:<\/p>\n\n<pre><\/pre>\n\n<p>It completes fine, no errors (finally).\nBut after that if I go into anaconda and do<\/p>\n\n<pre><\/pre>\n\n<p>I get no module caffe is found. What am I doing wrong? Any ideas?<\/p>\n","answers":[{"AnswerId":"55751464","CreationDate":"2019-04-18T17:58:06.533","ParentId":null,"OwnerUserId":"10685311","Title":null,"Body":"<p>you can try these steps:<\/p>\n\n<p>To use caffe within python, export its path as<\/p>\n\n<p>export PYTHONPATH=~\/Home\/<em>username<\/em>\/caffe\/python:$PYTHONPATH<\/p>\n\n<p>Replace username  with the your username in the system.<\/p>\n\n<p>Once you've done this, run the python terminal and import caffe<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import caffe<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>If it throws a 'module not found' error, check if it has been appended in pythonpath properly by typing<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import sys\n      sys.path\n      ['', '\/home\/nikita\/caffe\/python', '\/home\/nikita', '\/usr\/lib\/python2.7', '\/usr\/lib\/python2.7\/plat-x86_64-linux-gnu', '\/usr\/lib\n      \/python2.7\/lib-tk', '\/usr\/lib\/python2.7\/lib-old', '\/usr\/lib\/python2.7\/lib-dynload', '\/home\/nikita\/.local\/lib\/python2.7\/site-\n      packages']<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>If you see that the \/home\/username \/caffe\/python path, isn't there, then do<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>sys.path.append('\/Home\/username\/caffe\/python')<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n"},{"AnswerId":"31943616","CreationDate":"2015-08-11T13:48:57.023","ParentId":null,"OwnerUserId":"3633250","Title":null,"Body":"<p>Finally solved. Honestly the issue was in incorrect makefile.config. I needed to be extremely careful in adjusting it to specify all path to anaconda folders - I have incorrectly specified the path to python3.4 libraries.<\/p>\n\n<p>The point is - when set up caffe with anaconda and facing issues you need to go over makefile.config one more time - you should have misconfigured something<\/p>\n"}]}
{"QuestionId":31730125,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-07-30T17:14:18.183","AcceptedAnswerId":31738101.0,"OwnerUserId":5174743.0,"Title":"theano vector matrix product along third dimension","Body":"<p>I've got  and .<\/p>\n\n<p>Assume w is 1xd and X is 3 x 10 x d.<\/p>\n\n<p>I want as a result a matrix of size 3 x 10 where the i-th row is .<\/p>\n\n<p>Is there a way to do this in theano?<\/p>\n","answers":[{"AnswerId":"31738101","CreationDate":"2015-07-31T04:29:37.793","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>What you are looking for amounts to this in numpy (where I extend the degenerate dimension (1, d) to (6, 5) in order to be general for matrices. If <code>w<\/code> is a vector, then the function may write slightly more simply with 1D arrays)<\/p>\n\n<pre><code>import numpy as np\n\nw = np.arange(6 * 5).reshape(6, 5)\nX = np.arange(3 * 10 * 5).reshape(3, 10, 5)\n\noutput = np.einsum('ij, klj', w, X)\n<\/code><\/pre>\n\n<p>Let's check the zeroth output<\/p>\n\n<pre><code>print w.dot(X[0].T)\nprint output[:, 0]   # same output as above\n<\/code><\/pre>\n\n<p>We can do the same by reshaping the matrices, which will lead us to a valid Theano expression immediately<\/p>\n\n<pre><code>output2 = w.dot(X.reshape(-1, 5).T).reshape((w.shape[0],) + X.shape[:2])\nassert (output2 == output).all()\n<\/code><\/pre>\n\n<p>Now the Theano expression<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nww = T.fmatrix()\nXX = T.tensor3()\n\noutput_expr = ww.dot(XX.reshape((-1, XX.shape[-1])).T).reshape((ww.shape[0], XX.shape[0], XX.shape[1]), ndim=3)\n\nf = theano.function([ww, XX], output_expr)\n\nprint f(w.astype('float32'), X.astype('float32'))[:, 0]\n<\/code><\/pre>\n"}]}
{"QuestionId":31730892,"AnswerCount":1,"Tags":"<c><lua><stanford-nlp><torch><lstm>","CreationDate":"2015-07-30T17:59:56.707","AcceptedAnswerId":null,"OwnerUserId":2827214.0,"Title":"Is this a practical way to resolve 'Not enough memory' from LuaJit with Torch","Body":"<p><a href=\"https:\/\/github.com\/stanfordnlp\/treelstm\">StanfordNLP's TreeLSTM<\/a>, when used with a dataset with > 30K instances, causes LuaJit to error with \"Not Enough Memory.\" I am resolving this by using <a href=\"https:\/\/github.com\/neomantra\/lds\">LuaJit Data Structures<\/a>. In order to get the dataset outside of lua's heap, the trees need to be placed in a LDS.Vector.<\/p>\n\n<p>Since the LDS.Vector holds cdata, the first step was to make the Tree type into a cdata object:<\/p>\n\n<pre><\/pre>\n\n<p>There are also small changes that need to be made in read_data.lua to handle the new cdata CTree type. Using LDS seemed like a reasonable approach to solve the memory limit so far; however, the CTree requires a field named 'composer'. <\/p>\n\n<p>Composer is of the type nn.gModule. To continue with this solution would involve creating a typedef of the nn.gModule as cdata, including creating a typedef for its members. Before continuing, does this seem like the correct direction to follow? Does any one have experience with this problem? <\/p>\n","answers":[{"AnswerId":"31780091","CreationDate":"2015-08-03T05:37:59.833","ParentId":null,"OwnerUserId":"1316615","Title":null,"Body":"<p>As you've discovered, representing structured data in a LuaJIT heap-friendly manner is a bit of a pain at the moment.<\/p>\n\n<p>In the Tree-LSTM implementation, the tree tables each hold a pointer to a composer instance mainly for expediency in implementation.<\/p>\n\n<p>One workaround to avoid typedef-ing <code>nn.gModule<\/code> would be to use the existing <code>idx<\/code> field to index into a table of composer instances. In this approach, the pair (<code>sentence_idx<\/code>, <code>node_idx<\/code>) can be used uniquely identify a composer in a global two-level table of composer instances. To avoid memory issues, the current cleanup code can be replaced with a line that sets the corresponding index in the table to <code>nil<\/code>.<\/p>\n"}]}
{"QuestionId":31731711,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2015-07-30T18:45:46.310","AcceptedAnswerId":null,"OwnerUserId":5174743.0,"Title":"theano: gradient where cost is imag(x)","Body":"<p>If I have a cost that is the imaginary part of a complex number, trying to obtain the gradient with theano I get the following error:<\/p>\n\n<p><\/p>\n\n<p>Is it not possible to use the imaginary part as cost despite it being a real-valued cost?<\/p>\n\n<hr>\n\n<p>Edit. Minimal working example<\/p>\n\n<pre><\/pre>\n\n<p>I would expect this to work as T.imag(a) is a real scalar cost..<\/p>\n","answers":[]}
{"QuestionId":31733166,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2015-07-30T20:13:13.477","AcceptedAnswerId":31747183.0,"OwnerUserId":1150636.0,"Title":"Overlapping iteration over theano tensor","Body":"<p>I am trying to implement a scan loop in theano, which given a tensor will use a \"moving slice\" of the input. It doesn't have to actually be a moving slice, it can be a preprocessed tensor to another tensor that represents the moving slice.<\/p>\n\n<p>Essentially:<\/p>\n\n<pre><\/pre>\n\n<p>where  is the input for each iteration.<\/p>\n\n<p>I am trying to figure out the most efficient way to do it, maybe using some form of referencing or manipulating strides, but I haven't managed to get something to work even for pure numpy.<\/p>\n\n<p>One possible solution I found can be found <a href=\"https:\/\/stackoverflow.com\/questions\/2485669\/consecutive-overlapping-subsets-of-array-numpy-python\/2487551#2487551\">here<\/a>, but I can't figure out how to use strides and I don't see a way to use that with theano.<\/p>\n","answers":[{"AnswerId":"31747183","CreationDate":"2015-07-31T13:19:09.617","ParentId":null,"OwnerUserId":"5116849","Title":null,"Body":"<p>You can build a vector containing the starting index for the slice at each timestep and call Scan with that vector as a sequence and your original vector as a non-sequence. Then, inside Scan, you can obtain the slice you want at every iteration.<\/p>\n\n<p>I included an example in which I also made the size of the slices a symbolic input, in case you want to change it from one call of your Theano function to the next:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\n# Input variables\nx = T.vector(\"x\")\nslice_size = T.iscalar(\"slice_size\")\n\n\ndef step(idx, vect, length):\n\n    # From the idx of the start of the slice, the vector and the length of\n    # the slice, obtain the desired slice.\n    my_slice = vect[idx:idx + length]\n\n    # Do something with the slice here. I don't know what you want to do\n    # to I'll just return the slice itself.\n    output = my_slice\n\n    return output\n\n# Make a vector containing the start idx of every slice\nslice_start_indices = T.arange(x.shape[0] - slice_size + 1)\n\nout, updates = theano.scan(fn=step,\n                        sequences=[slice_start_indices],\n                        non_sequences=[x, slice_size])\n\nfct = theano.function([x, slice_size], out)\n<\/code><\/pre>\n\n<p>Running the function with your parameters produces the output :<\/p>\n\n<pre><code>print fct(range(17), 5)\n\n[[  0.   1.   2.   3.   4.]\n [  1.   2.   3.   4.   5.]\n [  2.   3.   4.   5.   6.]\n [  3.   4.   5.   6.   7.]\n [  4.   5.   6.   7.   8.]\n [  5.   6.   7.   8.   9.]\n [  6.   7.   8.   9.  10.]\n [  7.   8.   9.  10.  11.]\n [  8.   9.  10.  11.  12.]\n [  9.  10.  11.  12.  13.]\n [ 10.  11.  12.  13.  14.]\n [ 11.  12.  13.  14.  15.]\n [ 12.  13.  14.  15.  16.]]\n<\/code><\/pre>\n"},{"AnswerId":"31733984","CreationDate":"2015-07-30T21:02:23.427","ParentId":null,"OwnerUserId":"190597","Title":null,"Body":"<p>You could use <a href=\"https:\/\/stackoverflow.com\/a\/4947453\/190597\">this rolling_window recipe<\/a>:<\/p>\n\n<pre><code>import numpy as np\n\ndef rolling_window_lastaxis(arr, winshape):\n    \"\"\"\n    Directly taken from Erik Rigtorp's post to numpy-discussion.\n    http:\/\/www.mail-archive.com\/numpy-discussion@scipy.org\/msg29450.html\n    (Erik Rigtorp, 2010-12-31)\n\n    See also:\n    http:\/\/mentat.za.net\/numpy\/numpy_advanced_slides\/ (St\u00e9fan van der Walt, 2008-08)\n    https:\/\/stackoverflow.com\/a\/21059308\/190597 (Warren Weckesser, 2011-01-11)\n    https:\/\/stackoverflow.com\/a\/4924433\/190597 (Joe Kington, 2011-02-07)\n    https:\/\/stackoverflow.com\/a\/4947453\/190597 (Joe Kington, 2011-02-09)\n    \"\"\"\n    if winshape &lt; 1:\n        raise ValueError(\"winshape must be at least 1.\")\n    if winshape &gt; arr.shape[-1]:\n        print(winshape, arr.shape)\n        raise ValueError(\"winshape is too long.\")\n    shape = arr.shape[:-1] + (arr.shape[-1] - winshape + 1, winshape)\n    strides = arr.strides + (arr.strides[-1], )\n    return np.lib.stride_tricks.as_strided(arr, shape=shape, strides=strides)\n\nx = np.arange(17)\nprint(rolling_window_lastaxis(x, 5))\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>[[ 0  1  2  3  4]\n [ 1  2  3  4  5]\n [ 2  3  4  5  6]\n [ 3  4  5  6  7]\n [ 4  5  6  7  8]\n [ 5  6  7  8  9]\n [ 6  7  8  9 10]\n [ 7  8  9 10 11]\n [ 8  9 10 11 12]\n [ 9 10 11 12 13]\n [10 11 12 13 14]\n [11 12 13 14 15]\n [12 13 14 15 16]]\n<\/code><\/pre>\n\n<p>Note that there are even fancier extensions of this, such as <a href=\"https:\/\/stackoverflow.com\/a\/4947453\/190597\">Joe Kington's rolling_window<\/a> which can roll over multi-dimensional windows, and <a href=\"https:\/\/gist.github.com\/seberg\/3866040\" rel=\"nofollow noreferrer\">Sebastian Berg's implementation<\/a> which, in addition, can jump by steps.<\/p>\n"}]}
{"QuestionId":31736583,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-07-31T01:09:51.937","AcceptedAnswerId":null,"OwnerUserId":1215364.0,"Title":"Theano Function For Transforming Matrix Into Matrix With Different Dimensions","Body":"<p>I have matrices where the diagonal is the negative of the sum of all other elements in that row. Here is an example<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to write a Theano function that takes in these matrices and return a matrix with the same number of rows, one less column, and the diagonal removed. So for Q this would be<\/p>\n\n<pre><\/pre>\n\n<p>I'd also like to do the reverse i.e. given Q_raw, I'd like to write a Theano function that spits out Q. How can I write these functions in Theano? So far I haven't come up with a solution that takes only the matrix itself as input.<\/p>\n","answers":[{"AnswerId":"31742026","CreationDate":"2015-07-31T08:54:09.500","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's a couple of methods doing what I think you're asking for. There may be more efficient approaches.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef symbolic_remove_diagonal(x):\n    flat_x = x.flatten()\n    indexes = tt.arange(flat_x.shape[0], dtype='int64')\n    diagonal_modulo = indexes % (x.shape[0] + 1)\n    off_diagonal_flat_x = flat_x[tt.neq(diagonal_modulo, 0).nonzero()]\n    return off_diagonal_flat_x.reshape((x.shape[0], x.shape[1] - 1))\n\n\ndef symbolic_add_diagonal(x):\n    diagonal_values = -x.sum(axis=1)\n    flat_x = x.flatten()\n    result_length = flat_x.shape[0] + x.shape[0]\n    indexes = tt.arange(result_length, dtype='int64')\n    diagonal_modulo = indexes % (x.shape[0] + 1)\n    result = tt.zeros((result_length,), dtype=x.dtype)\n    result = tt.set_subtensor(result[tt.eq(diagonal_modulo, 0).nonzero()], diagonal_values)\n    result = tt.set_subtensor(result[tt.neq(diagonal_modulo, 0).nonzero()], flat_x)\n    return result.reshape((x.shape[0], x.shape[1] + 1))\n\n\ndef main():\n    theano.config.compute_test_value = 'raise'\n    x1 = tt.matrix()\n    x1.tag.test_value = numpy.array(\n        [[-6, 2, 2, 1, 1],\n         [1, -4, 0, 1, 2],\n         [1, 0, -4, 2, 1],\n         [2, 1, 0, -3, 0],\n         [1, 1, 1, 1, -4]])\n    x2 = tt.matrix()\n    x2.tag.test_value = numpy.array(\n        [[2, 2, 1, 1],\n         [1, 0, 1, 2],\n         [1, 0, 2, 1],\n         [2, 1, 0, 0],\n         [1, 1, 1, 1]])\n    remove_diagonal = theano.function(inputs=[x1], outputs=symbolic_remove_diagonal(x1))\n    add_diagonal = theano.function(inputs=[x2], outputs=symbolic_add_diagonal(x2))\n    x2_prime = remove_diagonal(x1.tag.test_value)\n    x1_prime = add_diagonal(x2.tag.test_value)\n    print 'Diagonal removed:\\n', x2_prime\n    print 'Diagonal added:\\n', x1_prime\n    assert numpy.all(x2_prime == x2.tag.test_value)\n    assert numpy.all(x1_prime == x1.tag.test_value)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":31750869,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-07-31T16:25:52.323","AcceptedAnswerId":null,"OwnerUserId":2789788.0,"Title":"Theano ImportError and process Warning when compiling function","Body":"<p>I am running Theano with Anaconda on Windows. Ihave pretty much followed the steps in the comments <a href=\"https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\" rel=\"nofollow\">here<\/a>. I can import theano with no problems:<\/p>\n\n<pre><\/pre>\n\n<p>This works fine. But when I do <\/p>\n\n<pre><\/pre>\n\n<p>I get warnings:<\/p>\n\n<pre><\/pre>\n\n<p>and a long error, which ends with:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on how to fix this?<\/p>\n","answers":[{"AnswerId":"31934867","CreationDate":"2015-08-11T06:41:13.360","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Theano on Windows need Theano development version and not last Theano version:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a><\/p>\n\n<p>So just update Theano and it should work.<\/p>\n"}]}
{"QuestionId":31757952,"AnswerCount":1,"Tags":"<python><if-statement><theano>","CreationDate":"2015-08-01T03:39:33.883","AcceptedAnswerId":31759485.0,"OwnerUserId":5179831.0,"Title":"something wrong about T.gt()","Body":"<p>In order to obtain the usage of T.gt(), i wrote a toy code.<\/p>\n\n<pre><\/pre>\n\n<p>I expect the returned value is -4 when a=-4 and -4 when a=4, but it always satisfy the condition and run the return -data.\nI don't know why. Can you help me?<\/p>\n","answers":[{"AnswerId":"31759485","CreationDate":"2015-08-01T07:36:52.927","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>T.gt<\/code> is a <em>symbolic<\/em> function; it doesn't return a Boolean value, instead it returns an object representing a symbolic expression that, when later compiled and executed, will evaluate to a Boolean value.<\/p>\n\n<p>So, in Python, <code>T.gt(...)<\/code> will always be evaluated as <code>True<\/code> because the result is always non-<code>None<\/code>.<\/p>\n\n<p>If you want a conditional expression in Theano then you need to use a symbolic conditional operation. There are two: <code>T.switch<\/code> and <code>theano.ifelse.ifelse<\/code>. The difference is that <code>T.switch<\/code> is an element-wise operation, accepting a tensor condition, while <code>ifelse<\/code> requires a scalar condition.<\/p>\n\n<p>There is another problem with your example. Even if the code was good, it would always return a negative value. In essence your example says, <em>if input is positive return its negative else return the input as-is (which is already negative)<\/em>. I would also recommend using <code>theano.function<\/code> over the <code>eval<\/code> function.<\/p>\n\n<p>Your example could be altered to illustrate the workings of <code>ifelse<\/code> like this:<\/p>\n\n<pre><code>import theano\nimport theano.ifelse\nimport theano.tensor as T\n\n\ndef symbolic_f(x):\n    return theano.ifelse.ifelse(T.gt(x, 0), -x - 1, x + 1)\n\n\ndef main():\n    x = T.scalar()\n    f = theano.function(inputs=[x], outputs=symbolic_f(x))\n\n    print f(-4)\n    print f(4)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":31763838,"AnswerCount":1,"Tags":"<python><numpy><theano><tf-idf>","CreationDate":"2015-08-01T16:37:35.623","AcceptedAnswerId":31763987.0,"OwnerUserId":1196752.0,"Title":"Theano GPU calculation slower than numpy","Body":"<p>I'm learning to use theano. I want to populate a term-document matrix (a numpy sparse matrix) by calculating binary TF-IDF for each element inside it:<\/p>\n\n<pre><\/pre>\n\n<p>But the numpy version is much faster than the theano implementation:<\/p>\n\n<pre><\/pre>\n\n<p>I've read that this can be due to overhead, that for small operations might kill the performance. <\/p>\n\n<p>Is my code bad or should I avoid using GPU because of the overhead?<\/p>\n","answers":[{"AnswerId":"31763987","CreationDate":"2015-08-01T16:52:50.510","ParentId":null,"OwnerUserId":"1428739","Title":null,"Body":"<p>The thing is that you are compiling your Theano function every time. The compilation takes time. Try passing the compiled function like this:<\/p>\n\n<pre><code>def tfidf_gpu(appearance_in_documents,num_documents,document_words,TFIDF):\n    start = perf_counter()\n    ret = TFIDF(num_documents,appearance_in_documents,document_words)\n    end = perf_counter()\n    print(\"\\nTFIDF_GPU \",end-start,\" secs.\")\n    return ret\n\nAPP = T.scalar('APP',dtype='int32')\nN = T.scalar('N',dtype='int32')\nSF = T.scalar('S',dtype='int32')\nF = (T.log(N)-T.log(APP)) \/ SF\nTFIDF = theano.function([N,APP,SF],F)\n\ntfidf_gpu(appearance_in_documents,num_documents,document_words,TFIDF)\n<\/code><\/pre>\n\n<p>Also your TFIDF task is a bandwidth intensive task. Theano, and GPU in general, is best for computation intensive tasks. <\/p>\n\n<p>The current task will considerable overhead taking the data to the GPU and back because in the end you will need to read each element O(1) times. But if you want to do more computation it makes sense to use the GPU. <\/p>\n"}]}
{"QuestionId":31768116,"AnswerCount":1,"Tags":"<python><theano><pycuda>","CreationDate":"2015-08-02T02:47:25.347","AcceptedAnswerId":31771043.0,"OwnerUserId":1103966.0,"Title":"Using a shared variable in a function","Body":"<p>Hi I'm following a neural net tutorial where the author seems to be using shared variables everywhere. From my understanding, a shared variable in theanos simply is a space in memory that can be shared by the gpu and cpu heap. Anyway, I have two matrices which I declare as shared variables and I want to perform some operation on them using function. (Question 1) I'd love it if someone could explain why function is usefull vs regular def function. Anyway, I'm setting up my definition like such:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that I have no idea how to use the shared variables in a regular function-output call. I understand that I can do updates via function([],..update=(shared_var_1, upate_function)). But how do I access them in my regular function?<\/p>\n","answers":[{"AnswerId":"31771043","CreationDate":"2015-08-02T10:43:21.923","ParentId":null,"OwnerUserId":"1196752","Title":null,"Body":"<p>Theano beginner here, so I'm not sure that my answer will cover all the technical aspects.<\/p>\n\n<p>Answering your <strong>first question<\/strong>: you need to declare theano function instead of def function, because theano is like a \"language\" inside python and invoking <code>theano.function<\/code>\u00a0you're compiling some ad-hoc C code performing your task under the hood. This is what makes Theano fast. \nFrom the <a href=\"http:\/\/deeplearning.net\/software\/theano\/introduction.html#sneak-peek\" rel=\"nofollow\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>It is good to think of <code>theano.function<\/code> as the interface to a compiler which builds a callable object from a purely symbolic graph. One of Theano\u2019s most important features is that theano.function can optimize a graph and even compile some or all of it into native machine instructions.<\/p>\n<\/blockquote>\n\n<p>About your <strong>second quetion<\/strong>, in order to access what's stored in your shared variable you should use <\/p>\n\n<pre><code>shared_var.get_value()\n<\/code><\/pre>\n\n<p>Check <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#using-shared-variables\" rel=\"nofollow\">these<\/a> examples:<\/p>\n\n<blockquote>\n  <p>The value can be accessed and modified by the <code>.get_value()<\/code> and\n  <code>.set_value()<\/code> methods.<\/p>\n<\/blockquote>\n\n<p>This code:<\/p>\n\n<pre><code>a = np.array([[1,2],[3,4]], dtype=theano.config.floatX)\nx = theano.shared(a)\nprint(x)\n<\/code><\/pre>\n\n<p>Will output<\/p>\n\n<pre><code>&lt;CudaNdarrayType(float32, matrix)&gt;\n<\/code><\/pre>\n\n<p>But using <code>get_value()<\/code>:<\/p>\n\n<pre><code>print(x.get_value())\n<\/code><\/pre>\n\n<p>It outputs<\/p>\n\n<pre><code>[[ 1.  2.]\n [ 3.  4.]]\n<\/code><\/pre>\n\n<p><strong>Edit:<\/strong> to use shared variables in functions <\/p>\n\n<pre><code>import theano\nimport numpy\na = numpy.int64(2)\ny = theano.tensor.scalar('y',dtype='int64')\nz = theano.tensor.scalar('z',dtype='int64')\nx = theano.shared(a)\nplus = y + z\ntheano_sum = theano.function([y,z],plus)\n# Using shared variable in a function\nprint(theano_sum(x.get_value(),3))\n# Changing shared variable value using a function\nx.set_value(theano_sum(2,2))\nprint(x.get_value())\n# Update shared variable value\nx.set_value(x.get_value(borrow=True)+1)\nprint(x.get_value())\n<\/code><\/pre>\n\n<p>Will output: <\/p>\n\n<pre><code>5\n4\n5\n<\/code><\/pre>\n"}]}
{"QuestionId":31771835,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-02T12:20:47.030","AcceptedAnswerId":31782914.0,"OwnerUserId":5182504.0,"Title":"Plotting Hidden Weights","Body":"<p>I've had an interest for neural networks for a while now and have just started following the deep learning tutorials. I have what I hope is a relatively straight forward question that I am hoping someone may answer. <\/p>\n\n<p>In the multilayer perception tutorial, I am interested in seeing the state of the  network at different layers (something similar to what is seen in this paper: <a href=\"http:\/\/www.iro.umontreal.ca\/~lisa\/publications2\/index.php\/publications\/show\/247\" rel=\"nofollow\">http:\/\/www.iro.umontreal.ca\/~lisa\/publications2\/index.php\/publications\/show\/247<\/a> ). For instance, I am able to write out the weights of the hidden layer using:<\/p>\n\n<pre><\/pre>\n\n<p>When I plot this using the utils.py tile plotting, I get the following pretty plot [edit: pretty plot rmoved as I dont have enough rep]. <\/p>\n\n<p>If I wanted to plot the weights at the logRegressionLayer, such that <\/p>\n\n<pre><\/pre>\n\n<p>what would I actually have to do? The above doesn't seem to work - it returns a 2darray of shape (500,10). I understand that the 500 relates to the number of hidden units. The paragraph on the Miscellaneous page:<\/p>\n\n<blockquote>\n  <p>Plotting the weights is a bit more tricky. We have n_hidden hidden\n  units, each of them corresponding to a column of the weight matrix. A\n  column has the same shape as the visible, where the weight\n  corresponding to the connection with visible unit j is at position j.\n  Therefore, if we reshape every such column, using numpy.reshape, we\n  get a filter image that tells us how this hidden unit is influenced by\n  the input image.<\/p>\n<\/blockquote>\n\n<p>confuses me alittle. I am unsure exactly how I would string it together.<\/p>\n\n<p>Thanks to all - sorry if the question is confusing!<\/p>\n","answers":[{"AnswerId":"31782914","CreationDate":"2015-08-03T08:40:36.443","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You could plot them just the like the weights in the first layer but they will not necessarily make much sense.<\/p>\n\n<p>Consider the weights in the first layer of a neural network. If the inputs have size 784 (e.g. MNIST images) and there are 2000 hidden units in the first layer then the first layer weights are a matrix of size 784x2000 (or maybe the transpose depending on how it's implemented). Those weights can be plotted as either 784 patches of size 2000 or, more usually, 2000 patches of size 784. In this latter case each patch can be plotted as a 28x28 image which directly ties back to the original inputs and thus is interpretable.<\/p>\n\n<p>For you higher level regression layer, you could plot 10 tiles, each of size 500 (e.g. patches of size 22x23 with some padding to make it rectangular), or 500 patches of size 10. Either might illustrate some patterns that are being found but it may be difficult to tie those patterns back to the original inputs.<\/p>\n"}]}
{"QuestionId":31772938,"AnswerCount":1,"Tags":"<python><optimization><machine-learning><theano>","CreationDate":"2015-08-02T14:35:54.803","AcceptedAnswerId":null,"OwnerUserId":1936768.0,"Title":"Theano cost function","Body":"<p>I am trying to learn how to use Theano. I work very frequently with survival analysis and I wanted therefore to try to implement a standard survival model using Theano's automatic differentiation and gradient descent. The model that I am trying to implement is called the Cox model and here is the wikipedia article: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Proportional_hazards_model\" rel=\"nofollow\">https:\/\/en.wikipedia.org\/wiki\/Proportional_hazards_model<\/a><\/p>\n\n<p>Very helpfully, they have written there the partial likelihood function, which is what is maximized when estimating the parameters of a Cox model. I am quite new to Theano and as a result am having difficulties implementing this cost function and so I am looking for some guidance. <\/p>\n\n<p>Here is the code I have written so far. My dataset has 137 records and hence the reason I hard-coded that value. T refers to the tensor module, and W refers to what the wikipedia article calls beta, and status is what wikipedia calls C. The remaining variables are identical to wikipedia's notation.<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, when I run this code, I am unhappily met with an infinite recursion error, which I hoped would not happen. This makes me think that I have not implemented this cost function in the way that Theano would like and so I am hoping to get some guidance on how to improve this code so that it works. <\/p>\n","answers":[{"AnswerId":"31782430","CreationDate":"2015-08-03T08:13:41.483","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You are mixing symbolic and non-symbolic operations but this doesn't work.<\/p>\n\n<p>For example, <code>T.eq<\/code> returns a non-executable <em>symbolic<\/em> expression representing the idea of comparing two things for equality but it doesn't actually do the comparison there and then. <code>T.eq<\/code> actually returns a Python object that represents the equality comparison and since a non-<code>None<\/code> object reference is considered the same as <code>True<\/code> in Python, the execution will always continue inside the if statement.<\/p>\n\n<p>If you need to construct a Theano computation involving conditionals then you need to use one of its two symbolic conditional operations: <code>T.switch<\/code> or <code>theano.ifelse.ifelse<\/code>. <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html\" rel=\"nofollow\">See the documentation for examples and details<\/a>.<\/p>\n\n<p>You are also using Python loops which is probably not what you need. To construct a Theano computation that explicitly loops you need to use the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\"><code>theano.scan<\/code> module<\/a>. However, if you can express your computation in terms of matrix operations (dot products, reductions, etc.) then it will run much, much, faster than something using scans.<\/p>\n\n<p>I suggest you work through some more <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/\" rel=\"nofollow\">Theano tutorials<\/a> before trying to implement something complex from scratch.<\/p>\n"}]}
{"QuestionId":31773204,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2015-08-02T15:05:37.077","AcceptedAnswerId":31774280.0,"OwnerUserId":1020302.0,"Title":"Add bias to Lasagne neural network layers","Body":"<p>I am wondering if there is a way to add bias node to each layer in Lasagne neural network toolkit? I have been trying to find related information in documentation.<\/p>\n\n<p>This is the network I built but i don't know how to add a bias node to each layer.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31774280","CreationDate":"2015-08-02T16:57:33.143","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>Actually you don't have to explicitly create biases, because <code>DenseLayer()<\/code>, and convolution base layers too, has a default keyword argument: <\/p>\n\n<p><code>b=lasagne.init.Constant(0.)<\/code>.<\/p>\n\n<p>Thus you can avoid creating <code>bias<\/code>, if you don't want to have with explicitly pass <code>bias=None<\/code>, but this is not that case.<\/p>\n\n<p>Thus in brief you do have bias parameters while you don't pass <code>None<\/code> to <code>bias<\/code> parameter e.g.:<\/p>\n\n<pre><code>hidden = Denselayer(...bias=None)\n<\/code><\/pre>\n"}]}
{"QuestionId":31774953,"AnswerCount":4,"Tags":"<neural-network><regression><deep-learning><caffe>","CreationDate":"2015-08-02T18:05:20.857","AcceptedAnswerId":31808324.0,"OwnerUserId":3698878.0,"Title":"Test labels for regression caffe, float not allowed?","Body":"<p>I am doing regression using caffe, and my  and  files are like this:<\/p>\n\n<pre><\/pre>\n\n<p>My problem is it seems caffe does not allow float labels like 2.0, when I use float labels while reading, for example the  file caffe only\nrecognizes <\/p>\n\n<blockquote>\n  <p>a total of 1 images<\/p>\n<\/blockquote>\n\n<p>which is wrong.<\/p>\n\n<p>But when I for example change the 2.0 to 2 in the file and the following lines same, caffe now gives <\/p>\n\n<blockquote>\n  <p>a total of 2 images<\/p>\n<\/blockquote>\n\n<p>implying that the float labels are responsible for the problem.<\/p>\n\n<p>Can anyone help me here, to solve this problem, I definitely need to use float labels for regression, so does anyone know about a work around or solution for this? Thanks in advance.<\/p>\n\n<p><b>EDIT<\/b>\nFor anyone facing a similar issue <a href=\"https:\/\/stackoverflow.com\/questions\/31617486\/use-caffe-to-train-lenet-with-csv-data\">use caffe to train Lenet with CSV data<\/a> might be of help. Thanks to @Shai.<\/p>\n","answers":[{"AnswerId":"37291510","CreationDate":"2016-05-18T06:06:02.337","ParentId":null,"OwnerUserId":"6281477","Title":null,"Body":"<p>Besides <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/6281477\">@Shai's answer<\/a> above, I wrote a <strong><a href=\"https:\/\/github.com\/DaleSong89\/caffe-batch-normalization\" rel=\"nofollow noreferrer\">MultiTaskData<\/a><\/strong> layer supporting <code>float<\/code> typed labels. <\/p>\n\n<p>Its main idea is to store the labels in <code>float_data<\/code> field of <code>Datum<\/code>, and the <code>MultiTaskDataLayer<\/code> will parse them as labels for any number of tasks according to the value of <code>task_num<\/code> and <code>label_dimension<\/code> set in <code>net.prototxt<\/code>. The related files include: <code>caffe.proto<\/code>, <code>multitask_data_layer.hpp\/cpp<\/code>, <code>io.hpp\/cpp<\/code>.<\/p>\n\n<p>You can easily add this layer to your own caffe and use it like this (this is an example for face expression label distribution learning task in which the \"exp_label\" can be float typed vectors such as [0.1, 0.1, 0.5, 0.2, 0.1] representing face expressions(5 class)'s probability distribution.):<\/p>\n\n<pre><code>    name: \"xxxNet\"\n    layer {\n        name: \"xxx\"\n        type: \"MultiTaskData\"\n        top: \"data\"\n        top: \"exp_label\"\n        data_param { \n            source: \"expression_ld_train_leveldb\"   \n            batch_size: 60 \n            task_num: 1\n            label_dimension: 8\n        }\n        transform_param {\n            scale: 0.00390625\n            crop_size: 60\n            mirror: true\n        }\n        include:{ phase: TRAIN }\n    }\n    layer { \n        name: \"exp_prob\" \n        type: \"InnerProduct\"\n        bottom: \"data\"  \n        top: \"exp_prob\" \n        param {\n            lr_mult: 1\n            decay_mult: 1\n        }\n        param {\n            lr_mult: 2\n            decay_mult: 0\n        }\n        inner_product_param {\n            num_output: 8\n            weight_filler {\n            type: \"xavier\"\n            }    \n        bias_filler {      \n            type: \"constant\"\n            }  \n        }\n    }\n    layer {  \n        name: \"exp_loss\"  \n        type: \"EuclideanLoss\"  \n        bottom: \"exp_prob\" \n        bottom: \"exp_label\"\n        top: \"exp_loss\"\n        include:{ phase: TRAIN }\n    }\n<\/code><\/pre>\n"},{"AnswerId":"36975172","CreationDate":"2016-05-02T04:48:57.803","ParentId":null,"OwnerUserId":"134077","Title":null,"Body":"<p>I ended up transposing, switching the channel order, and using unsigned ints rather than floats to get results. I suggest reading an image back from your HDF5 file to make sure it displays correctly.<\/p>\n\n<p>First read the image as unsigned ints:<\/p>\n\n<p><code>img = np.array(Image.open('images\/' + image_name))<\/code><\/p>\n\n<p>Then change the channel order from RGB to BGR:<\/p>\n\n<p><code>img = img[:, :, ::-1]<\/code><\/p>\n\n<p>Finally, switch from Height x Width x Channels to Channels x Height x Width:<\/p>\n\n<p><code>img = img.transpose((2, 0, 1))<\/code><\/p>\n\n<p>Merely changing the shape will scramble your image and ruin your data!<\/p>\n\n<p>To read back the image:<\/p>\n\n<pre><code>with h5py.File(h5_filename, 'r') as hf:\n    images_test = hf.get('images')\n    targets_test = hf.get('targets')\n    for i, img in enumerate(images_test):\n        print(targets_test[i])\n        from skimage.viewer import ImageViewer\n        viewer = ImageViewer(img.reshape(SIZE, SIZE, 3))\n        viewer.show()\n<\/code><\/pre>\n\n<p>Here's a script I wrote which deals with two labels  (steer and speed) for a self-driving car task: <a href=\"https:\/\/gist.github.com\/crizCraig\/aa46105d34349543582b177ae79f32f0\" rel=\"nofollow\">https:\/\/gist.github.com\/crizCraig\/aa46105d34349543582b177ae79f32f0<\/a><\/p>\n"},{"AnswerId":"31808324","CreationDate":"2015-08-04T11:39:53.370","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When using the image dataset input layer (with either <code>lmdb<\/code> or <code>leveldb<\/code> backend) caffe only supports one <strong>integer<\/strong> label per input image.<\/p>\n\n<p>If you want to do regression, and use floating point labels, you should try and use the HDF5 data layer. See for example <a href=\"https:\/\/stackoverflow.com\/q\/31617486\/1714410\">this question<\/a>.<\/p>\n\n<p>In python you can use <code>h5py<\/code> package to create hdf5 files.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import h5py, os\nimport caffe\nimport numpy as np\n\nSIZE = 224 # fixed size to all images\nwith open( 'train.txt', 'r' ) as T :\n    lines = T.readlines()\n# If you do not have enough memory split data into\n# multiple batches and generate multiple separate h5 files\nX = np.zeros( (len(lines), 3, SIZE, SIZE), dtype='f4' ) \ny = np.zeros( (len(lines),1), dtype='f4' )\nfor i,l in enumerate(lines):\n    sp = l.split(' ')\n    img = caffe.io.load_image( sp[0] )\n    img = caffe.io.resize( img, (SIZE, SIZE, 3) ) # resize to fixed size\n    # you may apply other input transformations here...\n    # Note that the transformation should take img from size-by-size-by-3 and transpose it to 3-by-size-by-size\n    # for example\n    # transposed_img = img.transpose((2,0,1))[::-1,:,:] # RGB-&gt;BGR\n    X[i] = transposed_img\n    y[i] = float(sp[1])\nwith h5py.File('train.h5','w') as H:\n    H.create_dataset( 'X', data=X ) # note the name X given to the dataset!\n    H.create_dataset( 'y', data=y ) # note the name y given to the dataset!\nwith open('train_h5_list.txt','w') as L:\n    L.write( 'train.h5' ) # list all h5 files you are going to use\n<\/code><\/pre>\n\n<p>Once you have all <code>h5<\/code> files and the corresponding test files listing them you can add an HDF5 input layer to your <code>train_val.prototxt<\/code>:<\/p>\n\n<pre><code> layer {\n   type: \"HDF5Data\"\n   top: \"X\" # same name as given in create_dataset!\n   top: \"y\"\n   hdf5_data_param {\n     source: \"train_h5_list.txt\" # do not give the h5 files directly, but the list.\n     batch_size: 32\n   }\n   include { phase:TRAIN }\n }\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Clarification<\/strong>:<br>\nWhen I say \"caffe only supports one integer label per input image\" I do not mean that the leveldb\/lmdb containers are limited, I meant the tools of caffe, specifically the <a href=\"https:\/\/stackoverflow.com\/a\/31431716\/1714410\"><code>convert_imageset<\/code><\/a> tool.<br>\nAt closer inspection, it seems like caffe stores data of type <code>Datum<\/code> in leveldb\/lmdb and the \"label\" property of this type is defined as integer (see <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L30\" rel=\"nofollow noreferrer\">caffe.proto<\/a>) thus when using caffe interface to leveldb\/lmdb you are restricted to a single int32 label per image.<\/p>\n"},{"AnswerId":"32698722","CreationDate":"2015-09-21T15:12:13.793","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/31808324\/2466336\">Shai's answer<\/a> already covers saving float labels to HDF5 format. In case LMDB is required\/preferred, here's a snippet on how to create an LMDB from float data (adapted from <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issuecomment-70211045\" rel=\"nofollow noreferrer\">this<\/a> github comment):<\/p>\n\n<pre><code>import lmdb\nimport caffe\ndef scalars_to_lmdb(scalars, path_dst):\n\n    db = lmdb.open(path_dst, map_size=int(1e12))\n\n    with db.begin(write=True) as in_txn:    \n        for idx, x in enumerate(scalars):            \n            content_field = np.array([x])\n            # get shape (1,1,1)\n            content_field = np.expand_dims(content_field, axis=0)\n            content_field = np.expand_dims(content_field, axis=0)\n            content_field = content_field.astype(float)\n\n            dat = caffe.io.array_to_datum(content_field)\n            in_txn.put('{:0&gt;10d}'.format(idx) dat.SerializeToString())\n    db.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":31776951,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-02T21:45:46.287","AcceptedAnswerId":null,"OwnerUserId":1103966.0,"Title":"How do I set only a single element in a tensor in theano?","Body":"<p>Currently I have this code which sets the first element of my tensor output.<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunate doing:<\/p>\n\n<pre><\/pre>\n\n<p>Only outputs the first row which is not what I want. Is it possible to create a statement that combines the first 2 lines to give me the matrix that I want?<\/p>\n","answers":[{"AnswerId":"31778583","CreationDate":"2015-08-03T02:23:18.497","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You most probably want to replace<\/p>\n\n<pre><code>output = T.set_subtensor(output[0][0], rotation[0][0])\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>output = T.set_subtensor(output[0, 0], rotation[0, 0])\n<\/code><\/pre>\n\n<p>if possible.<\/p>\n"}]}
{"QuestionId":31777135,"AnswerCount":1,"Tags":"<python><performance><optimization><parallel-processing><theano>","CreationDate":"2015-08-02T22:11:55.293","AcceptedAnswerId":null,"OwnerUserId":1103966.0,"Title":"How do I set many elements in parallel in theano","Body":"<p>Lets say I create a theano function, how do I run operations in parallel elementwise on theano tensors like on matrices?<\/p>\n\n<pre><\/pre>\n\n<p>The question should be reformed, how do I do parallel operations in a Theanos function? I've looked at <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html#parallel-element-wise-ops-with-openmp\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html#parallel-element-wise-ops-with-openmp<\/a> which only talks about adding a setting, but does not explain how an operation is parallelized for element wise operations.<\/p>\n","answers":[{"AnswerId":"31782642","CreationDate":"2015-08-03T08:25:58.837","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>To an extent, Theano expects you to focus more on <em>what<\/em> you want computed rather than on <em>how<\/em> you want it computed. The idea is that the Theano optimizing compiler will automatically parallelize as much as possible (either on GPU or on CPU using OpenMP).<\/p>\n\n<p>The following is an example based on the original post's example. The difference is that the computation is declared symbolically and, crucially, without any loops. Here one is telling Theano that the results should be a stack of tensors where the first tensor is the values in a range modulo the range size and the second tensor is the elements of the same range divided by the range size. We don't say that a loop should occur but clearly at least one will be required. Theano compiles this down to executable code and will parallelize it if it makes sense.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\n\ndef symbolic_range_div_mod(size):\n    r = tt.arange(size)\n    return tt.stack(r % size, r \/ size)\n\n\ndef main():\n    size = tt.dscalar()\n    range_div_mod = theano.function(inputs=[size], outputs=symbolic_range_div_mod(size))\n    print range_div_mod(20)\n\n\nmain()\n<\/code><\/pre>\n\n<p>You need to be able to specify your computation in terms of Theano operations. If those operations can be parallelized on the GPU, they should be parallelized automatically.<\/p>\n"}]}
{"QuestionId":31779920,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-08-03T05:23:25.430","AcceptedAnswerId":null,"OwnerUserId":1103966.0,"Title":"How should I allocate a numpy array inside theano function?","Body":"<p>Let's say I have a theano function:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>How should I allocate tensors\/ arrays or memory within the actual theano to be optimized when compiled by theano.<\/li>\n<li>How should I convert that allocated tensor\/ array whatever to a theano like variable to be returned? I've tried converting it to a shared variable in the function but that didn't work.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"31781266","CreationDate":"2015-08-03T07:04:55.390","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I'm sorry but I don't understand your specific questions but can comment on the code sample you provided.<\/p>\n\n<p>Firstly, your comment above <code>return z<\/code> is incorrect. If <code>x<\/code> and <code>y<\/code> are Theano variables then <code>z<\/code> will also be a Theano variable after <code>z = x + y<\/code>.<\/p>\n\n<p>Secondly, there is no need to pre-allocate memory, using numpy, for return variables. So your <code>my_fun<\/code> can change to simply<\/p>\n\n<pre><code>def my_fun(x, y):\n  z = x + y\n  return z\n<\/code><\/pre>\n\n<p>Thirdly, the output(s) of Theano functions need to be Theano variables, not Python functions. And the output needs to be a function of the inputs. So your <code>theano.function<\/code> call needs to be changed to<\/p>\n\n<pre><code>f = function(\n  inputs=[x, y],\n  outputs=[my_fun(x, y)]\n)\n<\/code><\/pre>\n\n<p>The most important point to grasp about Theano, which can be a little difficult to get one's head around when starting out, is the difference between the symbolic world and the executable world. Tied in to that is the difference between Python expressions and Theano expressions.<\/p>\n\n<p>The modified <code>my_fun<\/code> above could be used like a symbolic function or like a normal executable Python function but it behaves differently for each. If you pass in normal Python inputs then the addition operation occurs immediately and the return value is the result of the computation. So <code>my_fun(1,2)<\/code> returns <code>3<\/code>. If instead you pass in symbolic Theano variables then the addition operation does not take place immediately. Instead the function returns a symbolic expression that after later being compiled and executed will return the result of adding two inputs. So the result of <code>my_fun(theano.tensor.scalar(), theano.tensor.scalar())<\/code> is a Python object that represents a symbolic Theano computation graph. When that result is passed as the output to a <code>theano.function<\/code> it is compiled into something that is executable. Thean when the compiled function is executed, and given some concrete values for the inputs, you actually get the result you were looking for.<\/p>\n"}]}
{"QuestionId":31792416,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2015-08-03T16:33:14.663","AcceptedAnswerId":null,"OwnerUserId":5186172.0,"Title":"does theano have something similar to the python dictionary that is iterable with symbolic keys?","Body":"<p>\nBasically, is the following toy problem possible?:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here the problem is that I have inserted a dictionary for the scan updates, which then fails since it doesn't like theano.tensors as keys.<\/p>\n\n<p>The reason I'm curious is that sometimes based on prior information I sometimes know a dataset contains no knowledge about certain classes, so in these cases I don't want want to \"try\" to learn (when I'm doomed to learn nothing).<\/p>\n\n<h3>Edit:  not really an answer, but a workaround.  Also see comments by  DrBwts below<\/h3>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[]}
{"QuestionId":31796449,"AnswerCount":1,"Tags":"<python><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-03T20:44:44.830","AcceptedAnswerId":null,"OwnerUserId":1204685.0,"Title":"Same fc6 response for different images","Body":"<p>I follow the instruction in <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/filter_visualization.ipynb\" rel=\"nofollow\">filter visualization<\/a> and <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">classification<\/a> example to get the  (fully connected layer6) response to multiple different images in a folder from a pretrained model (bvlc reference model) but for all of the images I get the same vector.\nHere is the code I used:<\/p>\n\n<pre><\/pre>\n\n<p>PS: Is there any simple way to store this data in a file (like txt or csv) that can be used for later and can be read and opened without using Python?<\/p>\n","answers":[{"AnswerId":"31801048","CreationDate":"2015-08-04T05:04:00.107","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are accessing only a <em>single<\/em> element of the <code>fc6<\/code> response (the fourth one). It might be the case that this element in the output is degenerate for the kind of inputs you tested it on. Try looking at the entire <code>fc6<\/code> response.<\/p>\n\n<p>Moreover, I'm not sure what model you are using, but are you certain this specific model expects its <code>mean<\/code> argument to be per-channel mean and not per-pixel?<\/p>\n\n<p>BTW, you are using <code>oversample<\/code> for your input (the default option in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L47\" rel=\"nofollow\"><code>caffe.Classifier.predict<\/code><\/a>) this means the output you are getting is actually an <em>average<\/em> of 10 responses to slightly different input image (different cropping+mirroring). You might want to disable this option using<\/p>\n\n<pre><code>scores = net.predict([input_image], oversample=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":31796909,"AnswerCount":1,"Tags":"<theano><deep-learning><confusion-matrix>","CreationDate":"2015-08-03T21:17:31.887","AcceptedAnswerId":31798877.0,"OwnerUserId":2073392.0,"Title":"How to add a confusion matrix to Theano examples?","Body":"<p>I want to make use of Theano's logistic regression classifier, but I would like to make an apples-to-apples comparison with previous studies I've done to see how deep learning stacks up.  I recognize this is probably a fairly simple task if I was more proficient in Theano, but this is what I have so far.  From the tutorials on the website, I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm pretty sure this is where I need to add the functionality, but I'm not certain how to go about it.  What I need is either access to y_pred and y for each and every run (to update my confusion matrix in python) or to have the C++ code handle the confusion matrix and return it at some point along the way.  I don't think I can do the former, and I'm unsure how to do the latter.  I've done some messing around with an update function along the lines of:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm not completely clear on how to get this to interface with the function in question and give me a numpy array I can work with.\nI'm quite new to Theano, so hopefully this is an easy fix for one of you.  I'd like to use this classifer as my output layer in a number of configurations, so I could use the confusion matrix with other architectures.<\/p>\n","answers":[{"AnswerId":"31798877","CreationDate":"2015-08-04T00:30:09.050","ParentId":null,"OwnerUserId":"3353215","Title":null,"Body":"<p>I suggest using a brute force sort of a way. You need an output for a prediction first. Create a function for it.<\/p>\n\n<pre><code> prediction = theano.function(\n        inputs = [index],\n        outputs = MLPlayers.predicts,\n        givens={\n                x: test_set_x[index * batch_size: (index + 1) * batch_size]})\n<\/code><\/pre>\n\n<p>In your test loop, gather the predictions... <\/p>\n\n<pre><code>labels = labels + test_set_y.eval().tolist() \nfor mini_batch in xrange(n_test_batches):\n    wrong = wrong + int(test_model(mini_batch))   \n    predictions = predictions + prediction(mini_batch).tolist()\n<\/code><\/pre>\n\n<p>Now create confusion matrix this way:<\/p>\n\n<pre><code>    correct = 0\n    confusion = numpy.zeros((outs,outs), dtype = int)\n    for index in xrange(len(predictions)):\n        if labels[index] is predictions[index]:\n            correct = correct + 1\n        confusion[int(predictions[index]),int(labels[index])] = confusion[int(predictions[index]),int(labels[index])] + 1\n<\/code><\/pre>\n\n<p>You can find this kind of an implementation <a href=\"http:\/\/ragavvenkatesan.github.io\/Convolutional-Neural-Networks\/\" rel=\"nofollow\">in this repository<\/a>.<\/p>\n"}]}
{"QuestionId":31798815,"AnswerCount":1,"Tags":"<reshape><theano>","CreationDate":"2015-08-04T00:23:23.953","AcceptedAnswerId":31804637.0,"OwnerUserId":3353215.0,"Title":"Theano Reshaping","Body":"<p>I am unable to clearly comprehend 's . I have an image matrix of shape: <\/p>\n\n<pre><\/pre>\n\n<p>, where there are  stacks of images, each having  of channels. I now want to convert them into the following shape:<\/p>\n\n<pre><\/pre>\n\n<p>such that all the stacks will be combined together into one stack of all channels. I am not sure if reshape will do this for me. I see that reshape seems to not lexicographically order the pixels if they are mixed in dimensions in the middle. I have been trying to achieve this with a combination of , and , but to no avail. I would appreciate some help.<\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"31804637","CreationDate":"2015-08-04T08:46:16.863","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.reshape\" rel=\"noreferrer\">Theano reshape<\/a> works just like <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"noreferrer\">numpy reshape<\/a> with its default <code>order<\/code>, i.e. <code>'C'<\/code>:<\/p>\n\n<blockquote>\n  <p>\u2018C\u2019 means to read \/ write the elements using C-like index order, with\n  the last axis index changing fastest, back to the first axis index\n  changing slowest.<\/p>\n<\/blockquote>\n\n<p>Here's an example showing that the image pixels remain in the same order after a reshape via either numpy or Theano.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor\n\n\ndef main():\n    batch_size = 2\n    stack1_size = 3\n    stack2_size = 4\n    height = 5\n    width = 6\n    data = numpy.arange(batch_size * stack1_size * stack2_size * height * width).reshape(\n        (batch_size, stack1_size, stack2_size, height, width))\n    reshaped_data = data.reshape([batch_size, stack1_size * stack2_size, 1, height, width])\n    print data[0, 0, 0]\n    print reshaped_data[0, 0, 0]\n\n    x = theano.tensor.TensorType('int64', (False,) * 5)()\n    reshaped_x = x.reshape((x.shape[0], x.shape[1] * x.shape[2], 1, x.shape[3], x.shape[4]))\n    f = theano.function(inputs=[x], outputs=reshaped_x)\n    print f(data)[0, 0, 0]\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":31804340,"AnswerCount":2,"Tags":"<theano>","CreationDate":"2015-08-04T08:30:01.340","AcceptedAnswerId":null,"OwnerUserId":5179831.0,"Title":"A weird error with updates in theano","Body":"<p>I designed a variable net, but it occurred some problems with theano. The general idea is that different input will get different net with same parameters, something like a recursive neural network with auto-encoder.\nThere are two cases in my code, one case is run  if , the other case is run .<\/p>\n\n<p>It is weird that the code can run without bugs if I comment updates=updates, which is not my expected (train_test theano function in code). However, if I uncomment updates=updates, an error occurred (train_test_bug theano function in code). The later one is that I'd like to implement.<\/p>\n\n<p>I have been already spend some days on this bug. Who can help me? I will appreciate that.<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT (by @danielrenshaw)<\/strong><\/p>\n\n<p>I've cut the code down to a simpler demonstration of the problem.<\/p>\n\n<p>The cause is in the gradient computation of a double-nested scan expression. The problem disappears when a modified inner-most recursive expression is used (see comments in first function below).<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31822498","CreationDate":"2015-08-05T02:37:08.400","ParentId":null,"OwnerUserId":"5179831","Title":null,"Body":"<p>I solved this problem edited by danielrenshaw. When I add h0 as outputs_info, it work. Before that I used first element of sequence as outputs_info, I think it caused the error. But I still cannot solve my original problem.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\nimport theano.ifelse\n\n\ndef inner_scan_step(x_t_t, h_tm1, w):\n    # Fails when using this recursive expression\n    h_t = tt.dot(h_tm1, w) + x_t_t\n\n    # No failure when using this recursive expression\n    # h_t = h_tm1 + tt.dot(x_t_t, w)\n\n    return h_t\n\n\ndef outer_scan_step(x_t, w, h0):\n    h, _ = theano.scan(inner_scan_step,\n                       sequences=[x_t],\n                       outputs_info=[h0],\n                       non_sequences=[w],\n                       strict=True)\n    return h[-1]\n\n\ndef get_outputs(x, w, h0):\n    features, _ = theano.scan(outer_scan_step,\n                              sequences=[x],\n                              non_sequences=[w, h0],\n                              strict=True)\n    return tt.grad(features.sum(), w)\n\n\ndef main():\n    theano.config.compute_test_value = 'raise'\n\n    x_value = numpy.arange(12, dtype=theano.config.floatX).reshape((2, 2, 3))\n\n    x = tt.tensor3()\n    x.tag.test_value = x_value\n\n    w = theano.shared(value=numpy.ones((3, 3), dtype=theano.config.floatX), borrow=True)\n    h0 = theano.shared(value=numpy.zeros(3, dtype=theano.config.floatX), borrow=True)\n\n    f = theano.function(inputs=[x], outputs=get_outputs(x, w, h0))\n\n    print f(x_value)\n\n\nif __name__ == \"__main__\":\n    main()\n<\/code><\/pre>\n"},{"AnswerId":"36077074","CreationDate":"2016-03-18T05:50:01.230","ParentId":null,"OwnerUserId":"6080844","Title":null,"Body":"<p>I've encountered the same issue and I fixed it by letting optimizer=fast_compile in theano_flags. Guess that is a bug of theano.<\/p>\n"}]}
{"QuestionId":31807155,"AnswerCount":0,"Tags":"<python><numpy><pycharm><theano>","CreationDate":"2015-08-04T10:42:31.137","AcceptedAnswerId":null,"OwnerUserId":5173166.0,"Title":"Errors importing Theano in Pycharm","Body":"<p>I have a project set up in PyCharm that uses imports from Theano.\nI believe I have completed all the steps in the Theano installation guide and have included the necessary libraries in the Project Interpreter (Theano 0.7.0, numpy 1.9.2, and scipu 0.14.0, among others).<\/p>\n\n<p>Yet, when compiling I get the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Would appreciate it if someone can point out what else I'm missing.<\/p>\n","answers":[]}
{"QuestionId":31812660,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-08-04T14:56:04.513","AcceptedAnswerId":null,"OwnerUserId":3673925.0,"Title":"Error message during installing caffe command 'make all'","Body":"<p>I ran<\/p>\n\n<pre><\/pre>\n\n<p>as suggested on the website to complete the installation.\nI use Ubuntu 14.04 with CUDA and OpenBlas.<\/p>\n\n<p>The error messages showed as follows<\/p>\n\n<blockquote>\n  <p>CXX\/LD -o .build_release\/tools\/upgrade_net_proto_text.bin\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::curandGetErrorString(curandStatus)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::weight_gpu_gemm(double const*, double const*, double*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::forward_gpu_bias(double*, double const*)<br>\n          .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::forward_gpu_bias(float*, float const*)<br>\n          .build_release\/lib\/libcaffe.so: undefined reference to caffe::cudnn::dataType::zero<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::cudnn::dataType::one<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::backward_gpu_gemm(float const*, float const*, float*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::cublasGetErrorString(cublasStatus_t)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::forward_gpu_gemm(double const*, double const*, double*, bool)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::backward_gpu_gemm(double const*, double const*, double*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::backward_gpu_bias(double*, double const*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::forward_gpu_gemm(float const*, float const*, float*, bool)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::cudnn::dataType::zero<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::weight_gpu_gemm(float const*, float const*, float*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::BaseConvolutionLayer::backward_gpu_bias(float*, float const*)<br>\n      .build_release\/lib\/libcaffe.so: undefined reference to caffe::cudnn::dataType::one<br>\n      collect2: error: ld returned 1 exit status<br>\n      make: *** [.build_release\/tools\/upgrade_net_proto_text.bin] Error 1<\/p>\n<\/blockquote>\n\n<p>I only modified Makefile.config.\nThe modified Makefile.config shown as follows<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31825926","CreationDate":"2015-08-05T07:22:54.087","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to change the <code>BLAS<\/code> settings in <code>Makefile.config<\/code> to<\/p>\n\n<pre><code>BLAS := open\n<\/code><\/pre>\n\n<p>Rather than <code>'OpenBlas'<\/code>.<\/p>\n"}]}
{"QuestionId":31817007,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2015-08-04T18:51:10.763","AcceptedAnswerId":null,"OwnerUserId":4562171.0,"Title":"Why does theano.tensor.stack() make function compilation time so much faster?","Body":"<p>I'm writing a function which takes in an array of enzyme and substrate concentrations, and returns an array of maximum reaction rates.<\/p>\n\n<pre><\/pre>\n\n<p>enzyme_vars_array and substrate_vars array are lists of theanor tensors, built using T.dscalar('name_of_thing').<\/p>\n\n<p>rateExpressionsArray is a list of python expressions which gives the kinetic rate of a reaction in terms of the theano tensors in \nenzyme_vars_array and substrate_vars array. Because some enzyme and substrate concentrations may not be used, I've included the on_unused_input='ignore' flag.<\/p>\n\n<p>The size of these inputs are:<\/p>\n\n<p>len(enzyme_vars_array) = 132<\/p>\n\n<p>len(substrate_vars array) = 17<\/p>\n\n<p>len(rateExpressionsArray) = 2402<\/p>\n\n<p>Running this line of code as written causes an extremely long compilation, at least 5 minutes (I gave up waiting eventually).<\/p>\n\n<p>However this single change reduced the compilation time to only a few seconds:<\/p>\n\n<pre><\/pre>\n\n<p>What is theano.tensor.stack() doing which causes this massive change in compilation time? And what about the original formulation made it so slow?<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>While writing up example code which can be run, this problem got stranger. This code generates examples of enzyme_vars_array, substrate_vars_array, and rateExpressionArray:<\/p>\n\n<pre><\/pre>\n\n<p>This is the code I expected to run slowly (as it runs without T.stack). However in this example it now runs at a perfectly reasonable pace. When I replace the last line with this however:<\/p>\n\n<pre><\/pre>\n\n<p>It runs extremely slowly again, the opposite of my observed behavior in my actual code!<\/p>\n\n<p>I considered possible differences between this example and my actual code and realized that my real rateExpressionsArray is very sparse, it mostly consists of a placeholder which just returns a default value, with only a few real expressions. So rewriting my example code to look more like this, I have:<\/p>\n\n<pre><\/pre>\n\n<p>This produces the original behavior - when only a few indexes in rateExpressionArray are nontrivial expressions, the code runs fast with T.stack() and slowly without T.stack(). To replicate the slow version, replace the last line with:<\/p>\n\n<pre><\/pre>\n\n<p>So it seems that T.stack produces different behavior on sparse versus dense inputs? I'm even more confused as to the function and mechanism of stack() now. Any insights into what's happening here would be very much appreciated!<\/p>\n","answers":[]}
{"QuestionId":31820027,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-08-04T21:52:45.130","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"How can I modify the batch size on the fly, i.e. after loading an ANN in pycaffe?","Body":"<p>I trained a neural network using Caffe. I then use it to predict outputs given some new inputs. To do so, I load the trained neural network in pycaffe along with a deploy.prototxt, which specifies the inputs:<\/p>\n\n<pre><\/pre>\n\n<p>I load the neural network using:<\/p>\n\n<pre><\/pre>\n\n<p>Since I don't know in advance how many inputs I will have, I would like to be able to change the batch size after I loaded the neural network (i.e. after calling ). How to do so?<\/p>\n","answers":[{"AnswerId":"31820028","CreationDate":"2015-08-04T21:52:45.130","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>You can use <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/tutorial\/interfaces.md#reshape\" rel=\"nofollow\"><code>reshape<\/code><\/a>:<\/p>\n\n<pre><code>net.blobs['data'].reshape(data_batch_size, 1, 1, data_of_features)\nnet.blobs['adbeoption'].reshape(adbeoption_batch_size, 1, 1, 1)\n<\/code><\/pre>\n\n<p>Then you can call <code>net.forward()<\/code>.<\/p>\n\n<p>It will modify the batch size on the fly, without having to reload the ANN.<\/p>\n"}]}
{"QuestionId":31823898,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-08-05T05:11:01.307","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Changing the solver parameters in Caffe through pycaffe","Body":"<p>How can I change the solver parameter in Caffe through pycaffe?<\/p>\n\n<p>E.g. right after calling  I would like to change the solver's parameters (learning rate, stepsize, gamma, momentum, base_lr, power, etc.), without having to change .<\/p>\n","answers":[{"AnswerId":"32361005","CreationDate":"2015-09-02T18:54:55.657","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>Maybe you can create a temporary file.<\/p>\n\n<p>First of all, load your solver parameters with<\/p>\n\n<pre><code>from caffe.proto import caffe_pb2\nfrom google.protobuf import text_format\nsolver_config = caffe_pb2.SolverParameter()\nwith open('\/your\/solver\/path') as f:\n    text_format.Merge(str(f.read()), solver_config)\n<\/code><\/pre>\n\n<p>You can modify any solver parameter just setting the desired value in <code>solver_config<\/code> (e.g. <code>solver_config.test_interval = 15<\/code>). Then, it's just creating a temp file and load your solver from it:<\/p>\n\n<pre><code>new_solver_config = text_format.MessageToString(solver_config)\nwith open('temp.prototxt', 'w') as f:\n    f.write(new_solver_config) \nsolver = caffe.get_solver('temp.prototxt')\nsolver.step(1)\n<\/code><\/pre>\n"}]}
{"QuestionId":31830189,"AnswerCount":1,"Tags":"<python><macos><caffe><pycaffe>","CreationDate":"2015-08-05T10:40:40.153","AcceptedAnswerId":null,"OwnerUserId":342323.0,"Title":"Prebuilt Python Caffe for OSX","Body":"<p>Is there any pre-built PyCaffe out there for OSX? I do see instructions on how to build it but I'm sure I'll have a lot of difficulties trying to build all of its dependencies. So, I'd appreciate it anyone knows where I can get the prebuilt PyCaffe module? Or is it necessary that it gets fully built on the machine?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"32004200","CreationDate":"2015-08-14T07:06:06.917","ParentId":null,"OwnerUserId":"627806","Title":null,"Body":"<p>In case you're familiar with <a href=\"https:\/\/www.docker.com\/\" rel=\"nofollow\">docker<\/a>, you could try: <a href=\"https:\/\/github.com\/ryankennedyio\/deep-dream-generator\" rel=\"nofollow\">https:\/\/github.com\/ryankennedyio\/deep-dream-generator<\/a><\/p>\n\n<p>On OS X, you'll need to have <a href=\"http:\/\/boot2docker.io\/\" rel=\"nofollow\">boot2docker<\/a> or some such.<\/p>\n\n<p>Anyway, I was able to get up and running in under 5 minutes or so with this approach, after spending an hour or so (and giving up) downloading\/building\/installing dependencies directly.<\/p>\n"}]}
{"QuestionId":31832157,"AnswerCount":1,"Tags":"<numpy><scikit-learn><theano>","CreationDate":"2015-08-05T12:11:13.643","AcceptedAnswerId":31832445.0,"OwnerUserId":1584115.0,"Title":"Cant fit scikit-neuralnetwork classifier because of tuple index out of range","Body":"<p>I am trying to get this <a href=\"https:\/\/scikit-neuralnetwork.readthedocs.org\/en\/latest\/guide_beginners.html#classification\" rel=\"nofollow\">classifier<\/a> working. It is a extension for scikit learn with dependencies to Theano.<\/p>\n\n<p>My goal was to fit a neural network with a list of years and teach it to know if it is a leap year or not (later I would increase the range). But I run in an error if I want to test this example.<\/p>\n\n<p>My code looks like this:<\/p>\n\n<p><strong>leapyear.py<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>requirements.txt<\/p>\n\n<pre><\/pre>\n\n<p>my output with error:<\/p>\n\n<pre><\/pre>\n\n<p>I looked into the sources of mlp.py, but I dont know how to fix it. What has to be changed that I can fit my network?<\/p>\n\n<p>Update not question related:\nI just wanted to add, that I need to convert the year to a binary representation, after this the neural network will work.<\/p>\n","answers":[{"AnswerId":"31832445","CreationDate":"2015-08-05T12:24:32.490","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that the classifier requires the data to be presented as a 2 dimensional numpy array, with the first axis being the samples and the second axis being the features.<\/p>\n\n<p>In your case you have only one \"feature\" (the year) so you need to turn the years data into a Nx1 2D numpy array. This can be achieved by adding the following line just before the data split statement:<\/p>\n\n<pre><code>years = np.array([[year] for year in years])\n<\/code><\/pre>\n"}]}
{"QuestionId":31840488,"AnswerCount":1,"Tags":"<python><neural-network><euclidean-distance><caffe><pycaffe>","CreationDate":"2015-08-05T18:48:26.940","AcceptedAnswerId":null,"OwnerUserId":3543300.0,"Title":"Caffe Iteration loss versus Train Net loss","Body":"<p>I'm using caffe to train a CNN with a Euclidean loss layer at the bottom, and my solver.prototxt file configured to display every 100 iterations. I see something like this,<\/p>\n\n<pre><\/pre>\n\n<p>I'm confused as to what the difference between the Iteration loss and Train net loss is. Usually the iteration loss is very small (around 0) and the Train net output loss is a bit larger. Can somebody please clarify?<\/p>\n","answers":[{"AnswerId":"34014534","CreationDate":"2015-12-01T07:07:23.637","ParentId":null,"OwnerUserId":"1495615","Title":null,"Body":"<p>Evan Shelhamer already gave his answer on <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/WEhQ92s9Vus\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/WEhQ92s9Vus<\/a>. <\/p>\n\n<p>As he pointe out, The <code>net output #k<\/code> result is the output of the net for that particular iteration \/ batch while the <code>Iteration T, loss = X<\/code> output is smoothed across iterations according to the <code>average_loss<\/code> field.<\/p>\n"}]}
{"QuestionId":31849778,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-06T07:46:44.880","AcceptedAnswerId":31852207.0,"OwnerUserId":5197007.0,"Title":"TypeError \"Bad input argument to theano function\"","Body":"<p>The error:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name\n  \"c2.py:77\"  at index 1(0-based)', 'Wrong number of dimensions:\n  expected 2, got 1 with shape (128L,).')<\/p>\n<\/blockquote>\n\n<p>Please advise how to fix? <\/p>\n\n<p>The code and data can be downloaded on this link:\n <a href=\"http:\/\/u.163.com\/axfWJ81e\" rel=\"noreferrer\">http:\/\/u.163.com\/axfWJ81e<\/a>\n and enter this code: QU90WxTZ<\/p>\n\n<p>And here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31852207","CreationDate":"2015-08-06T09:39:17.887","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you tell Theano Y is a matrix of floating point values but the value you provide for Y is a vector of integers.<\/p>\n\n<p>It's not entirely clear which is correct, but I suspect you intend Y to be a vector of integers and to use the 1-hot variant of cross entropy. If so, the problem might be fixed by changing the Theano definition of Y to<\/p>\n\n<pre><code>Y = T.lvector()\n<\/code><\/pre>\n"}]}
{"QuestionId":31864585,"AnswerCount":2,"Tags":"<macos><cuda><gpu><theano>","CreationDate":"2015-08-06T19:51:53.350","AcceptedAnswerId":null,"OwnerUserId":865662.0,"Title":"Can't get Theano to link against CUDNN on OSX","Body":"<p>After a big battle I was finally able to get Theano to use the GPU in OSX.\nBut now, I've tried everything I can remember and Theano still can't use CuDNN.<\/p>\n\n<p>I installed CUDA version 7 and CUDNN version 3.<\/p>\n\n<p>I tried copying the libraries to  and also to , the include file was copied to <\/p>\n\n<p>My .theanorc is<\/p>\n\n<pre><\/pre>\n\n<p>And my .profile has the relevant parts:<\/p>\n\n<pre><\/pre>\n\n<p>But still when I try to get Theano to use CUDNN the further I get (with the files in lib64) gives me the error:<\/p>\n\n<pre><\/pre>\n\n<p>Seems like clang is not getting the cudnn library even when I specifically told it to check that path.<\/p>\n","answers":[{"AnswerId":"32122669","CreationDate":"2015-08-20T15:44:23.810","ParentId":null,"OwnerUserId":"5248131","Title":null,"Body":"<p>In your profile add:  <\/p>\n\n<pre><code>export LIBRARY_PATH=$CUDA_ROOT\/lib:$CUDA_ROOT\/lib64:$LIBRARY_PATH\n<\/code><\/pre>\n\n<p>Including the lib on LIBRARY_PATH worked for me. <\/p>\n"},{"AnswerId":"33785706","CreationDate":"2015-11-18T16:47:13.113","ParentId":null,"OwnerUserId":"636626","Title":null,"Body":"<p>I had the same problem and could fix it by setting<\/p>\n\n<pre><code>export DYLD_LIBRARY_PATH=\/Developer\/NVIDIA\/CUDA-7.5\/lib:$DYLD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>in my <code>.bashrc<\/code> and running<\/p>\n\n<pre><code>sudo update_dyld_shared_cache\n<\/code><\/pre>\n\n<p>afterwards.<\/p>\n"}]}
{"QuestionId":31868441,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-07T01:43:34.557","AcceptedAnswerId":null,"OwnerUserId":4029252.0,"Title":"\"BUG IN FGRAPH.REPLACE OR A LISTENER\" error reported when compile theano function","Body":"<p>I encountered an stranger error as below, when compile my theano function. I am using the version 0.7 of theano. I hope a quick work around is available. The function dump is <a href=\"http:\/\/pan.baidu.com\/s\/1eQGsjIE\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34083971","CreationDate":"2015-12-04T08:35:27.977","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This error message appears when a bug in a Theano optimization causes an invalid graph modification.<\/p>\n\n<p>If you ever see \"Optimization failure due to: &lt;something&gt;\", try the following:<\/p>\n\n<ol>\n<li><p>Search the internet, and the <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/theano-users\" rel=\"nofollow\">theano-users mailing list<\/a> in particular, for the message including the specific &lt;something&gt; (in this case &lt;something&gt; is \"local_shape_to_shape_i\"). You may find a message indicating that the bug has already been identified. If it's been reported to the Theano developers then it may have already been fixed though you may need to update to the bleeding edge version of Theano direct from GitHub (i.e. pip install --upgrade may not be sufficient).<\/p><\/li>\n<li><p>Even if you can't find any mention online, try updating to the bleeding edge version if that's possible for you. It may have already been fixed.<\/p><\/li>\n<li><p>If the latest bleeding edge version still exhibits the bug then report it on the <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/theano-users\" rel=\"nofollow\">theano-users mailing list<\/a>.<\/p><\/li>\n<li><p>Ignore it. Optimization failures do not cause invalid computations. The only side effect (at least in theory) is that the computation may not be as efficient as it might otherwise be.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":31871083,"AnswerCount":1,"Tags":"<cuda><deep-learning><caffe>","CreationDate":"2015-08-07T06:29:56.980","AcceptedAnswerId":31893413.0,"OwnerUserId":2467772.0,"Title":"Errors in building caffe on Jetson-TK1 board","Body":"<p>I am building caffe on my Jetson-TK1 board. The board runs Ubuntu Linux 32bit. My Makefile.config is as follow<\/p>\n\n<pre><\/pre>\n\n<p>I can do  with success.\nThe errors came when I run .\nThe errors are<\/p>\n\n<pre><\/pre>\n\n<p>I use CUDA-6.5.<\/p>\n\n<p>What could be wrong with this build?<\/p>\n\n<p><strong>EDIT 1:<\/strong>\nThe <a href=\"http:\/\/petewarden.com\/2014\/10\/25\/how-to-run-the-caffe-deep-learning-vision-library-on-nvidias-jetson-mobile-gpu-board\/\" rel=\"nofollow\">link<\/a> what  @Klaus Prinoth mentioned is useful. Now I can build. I can also test  for both CPU and GPU. But when I do , I got the message as . I am not sure what is wrong. The message is <\/p>\n\n<pre><\/pre>\n\n<p>What does the message mean?<\/p>\n","answers":[{"AnswerId":"31893413","CreationDate":"2015-08-08T13:04:11.983","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>I solved the problems following the steps below. These are the steps mentioned in this <a href=\"http:\/\/planspace.org\/20150614-the_nvidia_jetson_tk1_with_caffe_on_mnist\/\" rel=\"nofollow\">link<\/a>. <\/p>\n\n<p>(1)Need to make sure all the dependencies are installed. They are<\/p>\n\n<pre><code>    sudo apt-get install \\\n    libprotobuf-dev protobuf-compiler gfortran \\\n    libboost-dev cmake libleveldb-dev libsnappy-dev \\\n    libboost-thread-dev libboost-system-dev \\\n    libatlas-base-dev libhdf5-serial-dev libgflags-dev \\\n    libgoogle-glog-dev liblmdb-dev gcc-4.7 g++-4.7\n<\/code><\/pre>\n\n<p>Since I don't use Python, I skip steps necessary for Python interfaces.<\/p>\n\n<p>(2)Get the caffe sources<\/p>\n\n<pre><code>sudo apt-get install git\ngit clone https:\/\/github.com\/BVLC\/caffe.git\ncd caffe\ncp Makefile.config.example Makefile.config\n<\/code><\/pre>\n\n<p>(3) Need to change 1099511627776 to 536870912 in src\/caffe\/util\/db.cpp before <code>make -j 8 runtest<\/code>, without that it will lead to <code>MDB_MAP_FULL error in runtest<\/code>.\nMy Makefile.config is shown in the original post.\nThen you are ready for <\/p>\n\n<pre><code>make -j 8 all\nmake -j 8 test\nmake -j 8 runtest\n<\/code><\/pre>\n\n<p>Performance differences on CPU and GPU processing can be tested with <\/p>\n\n<p>For GPU: \"<code>run build\/tools\/caffe time --model=models\/bvlc_alexnet\/deploy.prototxt --gpu=0<\/code>\"<\/p>\n\n<p>For CPU: \"<code>run build\/tools\/caffe time --model=models\/bvlc_alexnet\/deploy.prototxt<\/code>\"\nThanks to @Klaus Prinoth, for giving me the link.<\/p>\n"}]}
{"QuestionId":31880475,"AnswerCount":1,"Tags":"<python><debugging><reshape><theano>","CreationDate":"2015-08-07T14:41:21.063","AcceptedAnswerId":31891337.0,"OwnerUserId":1799871.0,"Title":"Theano reshape \u2013 index out ouf bounds","Body":"<p>I can't seem to get Theano to reshape my tensors as want it to. The reshaping in the code bellow is supposed to keep  dimensions and flatten all remaining ones into a single array.<\/p>\n\n<p>The code fails with  on the  line if I run it with a test value. Otherwise, the function seems to compile, but fails upon first real input with .<\/p>\n\n<p>When I tried using just numpy for an equivalent code, it worked normally. Is there anything I am doing wrong? Or is there any easy way to see the resulting dimensions that are used for the reshaping ( does not help since everything is a Theano variable)?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31891337","CreationDate":"2015-08-08T08:45:11.993","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem relates to how you're combining the two components of the new shape. The <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.reshape\" rel=\"nofollow\"><code>reshape<\/code> command requires an <code>lvector<\/code> for the new shape<\/a>.<\/p>\n\n<p>Since you're using the test values mechanism you can debug this problem by simply printing test value bits and pieces. For example, I used<\/p>\n\n<pre><code>print inputs.shape.tag.test_value\nprint inputs.shape[0:keep_dims].tag.test_value\nprint inputs.shape[keep_dims:].tag.test_value\nprint T.prod(inputs.shape[keep_dims:]).tag.test_value\nprint (inputs.shape[0:keep_dims] + (T.prod(inputs.shape[keep_dims:]),)).tag.test_value\nprint T.concatenate([inputs.shape[0:keep_dims], [T.prod(inputs.shape[keep_dims:])]]).tag.test_value\n<\/code><\/pre>\n\n<p>This shows a fix to the problem: using <code>T.concatenate<\/code> to combine the <code>keep_dim<\/code>s and the product of the remaining dims.<\/p>\n"}]}
{"QuestionId":31880720,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2015-08-07T14:52:07.220","AcceptedAnswerId":31884034.0,"OwnerUserId":3828416.0,"Title":"How to prepare a dataset for Keras?","Body":"<h2>Motivation<\/h2>\n\n<p>To run a set of labeled vectors through <a href=\"http:\/\/keras.io\/\">Keras<\/a> neural network.<\/p>\n\n<h2>Example<\/h2>\n\n<p>Looking at Keras dataset example mnist:<\/p>\n\n<pre><\/pre>\n\n<p>It seem to be a 3 dimensional numpy array:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>1st dimension is for the samples<\/li>\n<li>2nd and 3rd for each sample features<\/li>\n<\/ul>\n\n<h2>Attempt<\/h2>\n\n<p>Building the labeled vectors:<\/p>\n\n<pre><\/pre>\n\n<h2>The training code<\/h2>\n\n<pre><\/pre>\n\n<h2>Result<\/h2>\n\n<pre><\/pre>\n\n<p>Why do I get such a bad result for such a simple dataset?\nIs my dataset malformed?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"31884034","CreationDate":"2015-08-07T18:03:12.217","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<p>A softmax over just one output node doesn't make much sense. If you change <code>model.add(Activation('softmax'))<\/code> to <code>model.add(Activation('sigmoid'))<\/code>, your network performs well.<\/p>\n\n<p>Alternatively you can also use two output nodes, where <code>1, 0<\/code> represents the case of <code>True<\/code> and <code>0, 1<\/code> represents the case of <code>False<\/code>. Then you can use a softmax layer. You just have to change your <code>Y_train<\/code> and <code>Y_test<\/code> accordingly.<\/p>\n"}]}
{"QuestionId":31889087,"AnswerCount":0,"Tags":"<python><amazon-web-services><amazon-ec2><apache-spark><theano>","CreationDate":"2015-08-08T02:40:28.633","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Apache-Spark with Python Theano","Body":"<p>I have written a binary classifier using Python Theano Library. Based on different data files, I would like to parallelize the classifier based on Amazon Web Service (AWS) EC2 with one master node and several slave nodes using Apache-Spark. I've tested my code with \"local\" mode on the AWS-EC2 master node, and classifier works very well except the long running time. However, when I run my same code on AWS-EC2 cluster with one master node and two slave nodes (all three nodes are  type), I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am sure that I correctly set-up my cluster on AWS-EC2 as well as the master node url in my code, and I'm also sure that all my codes have no bugs. But I have no idea about the problem. I really appreciate if anyone helps me solve the problem.<\/p>\n","answers":[]}
{"QuestionId":31892519,"AnswerCount":2,"Tags":"<python-2.7><cuda><gpu><theano><nvcc>","CreationDate":"2015-08-08T11:16:18.523","AcceptedAnswerId":31892615.0,"OwnerUserId":4899439.0,"Title":"Link error with CUDA 7.5 in Windows 10 (from Theano project): MSVCRT.lib error LNK2019: unresolved external symbol","Body":"<p>I am trying to properly setup CUDA in order to take advantage of the GPU in Theano.<\/p>\n\n<p>After fixing many compilation problems by tuning my  and  files, I am struggling to fix this linking error:<\/p>\n\n<pre><\/pre>\n\n<p>Here's my file:<\/p>\n\n<pre><\/pre>\n\n<p>And here is my  file:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that this is not an uncommon error, but generally fixes involve <a href=\"https:\/\/stackoverflow.com\/questions\/2098627\/how-to-integrate-cuda-cu-code-with-c-app\">changing some setting in the Visual Studio project<\/a>. However, here I don't have a Visual Studio project. The code is dynamically generated by Theano and compiled at runtime.<\/p>\n\n<p>Relevant system settings:<\/p>\n\n<ul>\n<li>Windows 10 (yes...)<\/li>\n<li>Python 2.7.10 64bits (Anaconda distrib)<\/li>\n<li>CUDA 7.5 \/ NVIDIA driver 353.54 \/ GeForce GTX 760<\/li>\n<li>Visual Studio Community 2013<\/li>\n<\/ul>\n","answers":[{"AnswerId":"36092366","CreationDate":"2016-03-18T19:00:22.990","ParentId":null,"OwnerUserId":"5473273","Title":null,"Body":"<p>I would also like to thank you. I have been trying to get this working for hours, and this was the post that put me over the edge. My config was slightly different, so my actual links were different. I am posting them in case it helps anyone else<\/p>\n\n<p>.theanorc<\/p>\n\n<pre><code>[global]\ndevice = gpu\nfloatX = float32\n\n[nvcc]\nflags = --use-local-env  --cl-version=2008\n<\/code><\/pre>\n\n<p>nvcc.profile <\/p>\n\n<pre><code>TOP              = $(_HERE_)\/..\n\nNVVMIR_LIBRARY_DIR = $(TOP)\/nvvm\/libdevice\n\nPATH            += $(TOP)\/open64\/bin;$(TOP)\/nvvm\/bin;$(_HERE_);$(TOP)\/lib;\n\nINCLUDES        +=  \"-I$(TOP)\/include\" \"-I$(TOP)\/include\/cudart\" \"-IC:\/Program Files (x86)\/Common Files\/Microsoft\/Visual C++ for Python\/9.0\/VC\/include\" \"-IC:\\Program Files\\Microsoft SDKs\\Windows\\v7.1\\Include\"$(_SPACE_)\n\nLIBRARIES        =+ $(_SPACE_) \"\/LIBPATH:$(TOP)\/lib\/$(_WIN_PLATFORM_)\" \"\/LIBPATH:C:\/Program Files (x86)\/Common Files\/Microsoft\/Visual C++ for Python\/9.0\/VC\/lib\/amd64\" \"\/LIBPATH:C:\\Program Files\\Microsoft SDKs\\Windows\\v7.1\\Lib\\x64\"\n\nCUDAFE_FLAGS    +=\nOPENCC_FLAGS    +=\nPTXAS_FLAGS     +=\n<\/code><\/pre>\n\n<p>using:\nwindows 7, 64bit\ncuda 5.5\npython 2.7\nwindows SDK 7.1\nMicrosoft Visual C++ Compiler for Python 2.7<\/p>\n"},{"AnswerId":"31892615","CreationDate":"2015-08-08T11:29:23.583","ParentId":null,"OwnerUserId":"4899439","Title":null,"Body":"<p>Damn it! I figured it out just after posting the question. The solution: slightly different include and library folders:<\/p>\n\n<pre><code>TOP              = $(_HERE_)\/..\n\nNVVMIR_LIBRARY_DIR = $(TOP)\/nvvm\/libdevice\n\nPATH            += $(TOP)\/open64\/bin;$(TOP)\/nvvm\/bin;$(_HERE_);$(TOP)\/lib;\n\nINCLUDES        +=  \"-I$(TOP)\/include\" \"-IC:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\include\" \"-IC:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\\Include\" $(_SPACE_)\n\nLIBRARIES        =+ $(_SPACE_) \"\/LIBPATH:$(TOP)\/lib\/$(_WIN_PLATFORM_)\" \"\/LIBPATH:C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\lib\\amd64\" \"\/LIBPATH:C:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\\Lib\\x64\"\n\nCUDAFE_FLAGS    +=\nPTXAS_FLAGS     +=\n<\/code><\/pre>\n\n<p>In particular I switched from:<\/p>\n\n<pre><code>C:\\Program Files\\Microsoft SDKs\\Windows\\v6.0A\\\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>C:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v7.1A\n<\/code><\/pre>\n\n<p>(I thought that I had to use the <code>Program Files<\/code> ones because of my 64bits project, but in fact the 64bits files are also included in <code>Program Files (x86)<\/code>)<\/p>\n"}]}
{"QuestionId":31896518,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-08-08T17:12:43.923","AcceptedAnswerId":35181529.0,"OwnerUserId":3106889.0,"Title":"Calculating Recall and precision for the Confusion matrix using torch7","Body":"<p>I am using the tutorial <a href=\"https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised\" rel=\"nofollow\">supervised<\/a> and I would like to calculate the Recall and Precision. Is there way to calculate them in the tutorial?<\/p>\n","answers":[{"AnswerId":"35181529","CreationDate":"2016-02-03T15:49:19.160","ParentId":null,"OwnerUserId":"3106889","Title":null,"Body":"<p>I saved my data in desk then I used python<\/p>\n"}]}
{"QuestionId":31900850,"AnswerCount":1,"Tags":"<python><memory><amazon-web-services><apache-spark><theano>","CreationDate":"2015-08-09T04:15:36.280","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"PySpark OSError: [Errno 12] Cannot allocate memory","Body":"<p>I've written a binary classifier using Python Theano Library. Based on different data files, I would like to parallelize the classifier based on Amazon Web Service (AWS) EC2 with one master node and several slave nodes using Apache-Spark. When I tested my code with \"local\" mode on the AWS-EC2 master node which is t2.micro type (1 GB RAM), I got the following error on the memory:<\/p>\n\n<pre><\/pre>\n\n<p>I'm sure that all my code has no bugs, and it seems that the problem is running out of memory or other problems related to memory. But t2.micro on AWS has 1 GB memory, data that the code reads for training the classifier is just about 1.7 KB, so I think the memory is big enough. I really have no idea about the error as well as how to fix it, and I really appreciate if anyone helps me.<\/p>\n","answers":[{"AnswerId":"33621839","CreationDate":"2015-11-10T02:26:10.783","ParentId":null,"OwnerUserId":"5461201","Title":null,"Body":"<p>maybe you can use subprocess.Open(\"cmd\" , stdout= subprocess.PIPE, shell =True ,close_fds = True )<\/p>\n"}]}
{"QuestionId":31908698,"AnswerCount":1,"Tags":"<installation><theano>","CreationDate":"2015-08-09T20:48:47.500","AcceptedAnswerId":null,"OwnerUserId":4859843.0,"Title":"gendef returning invalid syntax error","Body":"<p>I am trying to install Theano for machine learning on my Windows 7 computer.<\/p>\n\n<p>One of the last steps in installing the dependencies is to 'create a link library for GCC' by 'Opening up the Python shell and cd to C:\\SciSoft. Then execute:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried doing this but I get a invalid syntax error highlighted on 'WinPython'. I tried changing directory to go deeper and running gendef again and it also returned the same error. This is a copy and paste job from <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a><\/p>\n","answers":[{"AnswerId":"35839056","CreationDate":"2016-03-07T08:18:12.260","ParentId":null,"OwnerUserId":"1762932","Title":null,"Body":"<p>I also followed the tutorial at the link to install Theano. <\/p>\n\n<p>The line \"<em>Finally we need to create a link library for GCC. Open up the Python shell and cd to c:\\SciSoft<\/em>\" is probably an error; \"the Python shell\" should be modified to \"cmd.exe\".<\/p>\n\n<p>The two-line scripts are not python scripts, and can be successfully run on cmd.exe after changing directory to c:\\SciSoft.<\/p>\n"}]}
{"QuestionId":31919818,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-08-10T12:39:20.993","AcceptedAnswerId":31923358.0,"OwnerUserId":1196752.0,"Title":"Theano sqrt returning NaN values","Body":"<p>In my code I'm using theano to calculate an euclidean distance matrix (code from <a href=\"https:\/\/stackoverflow.com\/questions\/25886374\/pdist-for-theano-tensor\">here<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>But the following code causes some values of the matrix to be . I've read that this happens when calculating  and <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/IXfxAU35MiE\" rel=\"nofollow noreferrer\">here<\/a> it's suggested to <\/p>\n\n<blockquote>\n  <p>Add an eps inside the sqrt (or max(x,EPs))<\/p>\n<\/blockquote>\n\n<p>So I've added an eps to my code:<\/p>\n\n<pre><\/pre>\n\n<p>And I'm adding it before performing . I'm getting less s, but I'm still getting them. What is the proper solution to the problem? I've also noticed that if \u00a0is \u00a0there are no <\/p>\n","answers":[{"AnswerId":"31924111","CreationDate":"2015-08-10T15:58:41.733","ParentId":null,"OwnerUserId":"3497273","Title":null,"Body":"<h2>Just checking<\/h2>\n\n<p>In <code>squared_euclidian_distances<\/code> you're adding a column, a row, and a matrix. Are you sure this is what you want?<\/p>\n\n<p>More precisely, if <code>MAT<\/code> is of shape (n, p), you're adding matrices of shapes (n, 1), (1, n) and (n, n).<\/p>\n\n<p>Theano seems to silently repeat the rows (resp. the columns) of each one-dimensional member to match the number of rows and columns of the dot product.<\/p>\n\n<h2>If this is what you want<\/h2>\n\n<p>In reshape, you should probably specify <code>ndim=2<\/code> according to <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.reshape\" rel=\"nofollow\">basic tensor functionality : reshape<\/a>.<\/p>\n\n<blockquote>\n  <p>If the shape is a Variable argument, then you might need to use the optional ndim parameter to declare how many elements the shape has, and therefore how many dimensions the reshaped Variable will have.<\/p>\n<\/blockquote>\n\n<p>Also, it seems that <code>squared_euclidean_distances<\/code> should always be positive, unless imprecision errors in the difference change zero values into small negative values. If this is true, and if negative values are responsible for the NaNs you're seeing, you could indeed get rid of them without corrupting your result by surrounding <code>squared_euclidean_distances<\/code> with <code>abs(...)<\/code>.<\/p>\n"},{"AnswerId":"31923358","CreationDate":"2015-08-10T15:20:03.783","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There are two likely sources of NaNs when computing Euclidean distances.<\/p>\n\n<ol>\n<li><p>Floating point representation approximation issues causing negative distances when it's really just zero. The square root of a negative number is undefined (assuming you're not interested in the complex solution).<\/p>\n\n<p>Imagine <code>MAT<\/code> has the value<\/p>\n\n<pre><code>[[ 1.62434536 -0.61175641 -0.52817175 -1.07296862  0.86540763]\n [-2.3015387   1.74481176 -0.7612069   0.3190391  -0.24937038]\n [ 1.46210794 -2.06014071 -0.3224172  -0.38405435  1.13376944]\n [-1.09989127 -0.17242821 -0.87785842  0.04221375  0.58281521]]\n<\/code><\/pre>\n\n<p>Now, if we break down the computation we see that <code>(MAT ** 2).sum(1).reshape((MAT.shape[0], 1)) + (MAT ** 2).sum(1).reshape((1, MAT.shape[0]))<\/code> has value<\/p>\n\n<pre><code>[[ 10.3838024   -9.92394296  10.39763039  -1.51676099]\n [ -9.92394296  18.16971188 -14.23897281   5.53390084]\n [ 10.39763039 -14.23897281  15.83764622  -0.65066204]\n [ -1.51676099   5.53390084  -0.65066204   4.70316652]]\n<\/code><\/pre>\n\n<p>and <code>2 * MAT.dot(MAT.T)<\/code> has value<\/p>\n\n<pre><code>[[ 10.3838024   14.27675714  13.11072431   7.54348446]\n [ 14.27675714  18.16971188  17.00367905  11.4364392 ]\n [ 13.11072431  17.00367905  15.83764622  10.27040637]\n [  7.54348446  11.4364392   10.27040637   4.70316652]]\n<\/code><\/pre>\n\n<p>The diagonal of these two values should be equal (the distance between a vector and itself is zero) and from this textual representation it looks like that is true, but in fact they are slightly different -- the differences are too small to show up when we print the floating point values like this<\/p>\n\n<p>This becomes apparent when we print the value of the full expression (the second of the matrices above subtracted from the first)<\/p>\n\n<pre><code>[[  0.00000000e+00   2.42007001e+01   2.71309392e+00   9.06024545e+00]\n [  2.42007001e+01  -7.10542736e-15   3.12426519e+01   5.90253836e+00]\n [  2.71309392e+00   3.12426519e+01   0.00000000e+00   1.09210684e+01]\n [  9.06024545e+00   5.90253836e+00   1.09210684e+01   0.00000000e+00]]\n<\/code><\/pre>\n\n<p>The diagonal is almost composed of zeros but the item in the second row, second column is now a very small negative value. When you then compute the square root of all these values you get <code>NaN<\/code> in that position because the square root of a negative number is undefined (for real numbers).<\/p>\n\n<pre><code>[[ 0.          4.91942071  1.64714721  3.01002416]\n [ 4.91942071         nan  5.58951267  2.42951402]\n [ 1.64714721  5.58951267  0.          3.30470398]\n [ 3.01002416  2.42951402  3.30470398  0.        ]]\n<\/code><\/pre><\/li>\n<li><p>Computing the gradient of a Euclidean distance expression with respect to a variable inside the input to the function. This can happen not only if a negative number of generated due to floating point approximations, as above, but also if any of the inputs are zero length.<\/p>\n\n<p>If <code>y = sqrt(x)<\/code> then <code>dy\/dx = 1\/(2 * sqrt(x))<\/code>. So if <code>x=0<\/code> or, for your purposes, if <code>squared_euclidean_distances=0<\/code> then the gradient will be <code>NaN<\/code> because <code>2 * sqrt(0) = 0<\/code> and dividing by zero is undefined.<\/p><\/li>\n<\/ol>\n\n<p>The solution to the first problem can be achieved by ensuring squared distances are never negative by forcing them to be no less than zero:<\/p>\n\n<pre><code>T.sqrt(T.maximum(squared_euclidean_distances, 0.))\n<\/code><\/pre>\n\n<p>To solve both problems (if you need gradients) then you need to make sure the squared distances are never negative or zero, so bound with a small positive epsilon:<\/p>\n\n<pre><code>T.sqrt(T.maximum(squared_euclidean_distances, eps))\n<\/code><\/pre>\n\n<p>The first solution makes sense since the problem only arises from approximate representations. The second is a bit more questionable because the true distance is zero so, in a sense, the gradient should be undefined. Your specific use case may yield some alternative solution that is maintains the semantics without an artificial bound (e.g. by ensuring that gradients are never computed\/used for zero-length vectors). But <code>NaN<\/code> values can be pernicious: they can spread like weeds.<\/p>\n"}]}
{"QuestionId":31921084,"AnswerCount":2,"Tags":"<python><serialization><save><loading><theano>","CreationDate":"2015-08-10T13:38:47.710","AcceptedAnswerId":31922453.0,"OwnerUserId":3497273.0,"Title":"How to save \/ serialize a trained model in theano?","Body":"<p>I saved the model as documented on <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loading_and_saving.html\" rel=\"noreferrer\">loading and saving<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p> is a trained auto-encoder. It's a instance of class <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/cA.py\" rel=\"noreferrer\"><\/a>. From the script in which I build and save the model I can call  and  without any problem.<\/p>\n\n<p>In a different script I try to load the trained model.<\/p>\n\n<pre><\/pre>\n\n<p>I receive the following error.<\/p>\n\n<blockquote>\n<pre><\/pre>\n  \n  <p>AttributeError: 'module' object has no attribute 'cA'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"31922453","CreationDate":"2015-08-10T14:38:26.183","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>All the class definitions of the pickled objects need to be known by the script that does the unpickling. There is more on this in other StackOverflow questions (e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/9127714\/attributeerror-module-object-has-no-attribute-newperson\">AttributeError: &#39;module&#39; object has no attribute &#39;newperson&#39;<\/a>).<\/p>\n\n<p>Your code is correct as long as you properly import <code>cA<\/code>. Given the error you're getting it may not be the case. Make sure you're using <code>from cA import cA<\/code> and not just <code>import cA<\/code>.<\/p>\n\n<p>Alternatively, your model is defined by its parameters so you could instead just pickle the parameter values). This could be done in two ways depending on what you point of view.<\/p>\n\n<ol>\n<li><p>Save the Theano shared variables. Here we assume that <code>ca.params<\/code> is a regular Python list of Theano shared variable instances.<\/p>\n\n<pre><code>cPickle.dump(ca.params, f, protocol=cPickle.HIGHEST_PROTOCOL)\n<\/code><\/pre><\/li>\n<li><p>Save the numpy arrays stored inside the Theano shared variables.<\/p>\n\n<pre><code>cPickle.dump([param.get_value() for param in ca.params], f, protocol=cPickle.HIGHEST_PROTOCOL)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>When you want to load the model you'll need to reinitialize the parameters. For example, create a new instance of the <code>cA<\/code> class then either<\/p>\n\n<pre><code>ca.params = cPickle.load(f)\nca.W, ca.b, ca.b_prime = ca.params\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>ca.params = [theano.shared(param) for param in cPickle.load(f)]\nca.W, ca.b, ca.b_prime = ca.params\n<\/code><\/pre>\n\n<p>Note that you need to set both the <code>params<\/code> field and the separate parameters fields.<\/p>\n"},{"AnswerId":"40000726","CreationDate":"2016-10-12T14:04:51.743","ParentId":null,"OwnerUserId":"5637569","Title":null,"Body":"<p>One alternative way to save model is to save its weights and architecture and then load the same, the way we do for pre-train CNN:<\/p>\n\n<pre><code>def save_model(model):\n\n\n   model_json = model.to_json()\n   open('cifar10_architecture.json', 'w').write(model_json)\n   model.save_weights('cifar10_weights.h5', overwrite=True)\n<\/code><\/pre>\n\n<p>source\/ref : <a href=\"https:\/\/blog.rescale.com\/neural-networks-using-keras-on-\" rel=\"nofollow\">https:\/\/blog.rescale.com\/neural-networks-using-keras-on-<\/a>   rescale\/<\/p>\n"}]}
{"QuestionId":31923625,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2015-08-10T15:33:54.917","AcceptedAnswerId":36524052.0,"OwnerUserId":3828416.0,"Title":"Recurrent neural layers in Keras","Body":"<p>I'm learning neural networks through Keras and would like to explore my sequential dataset on a recurrent neural network.\nI was <a href=\"http:\/\/keras.io\/layers\/recurrent\/\" rel=\"noreferrer\">reading the docs<\/a> and trying to make sense of the <a href=\"http:\/\/keras.io\/examples\/\" rel=\"noreferrer\">LSTM example<\/a>.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>What are the  that are required for both layers?<\/li>\n<li>How do I prepare a sequential dataset that works with  as an input for those recurrent layers?<\/li>\n<li>What does the  layer do?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"36524052","CreationDate":"2016-04-09T23:38:55.190","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<ol>\n<li><p>Timesteps are a pretty bothering thing about Keras. Due to the fact that data you provide as an input to your LSTM must be a numpy array it is needed (at least for Keras version &lt;= 0.3.3) to have a specified shape of data - even with a \"time\" dimension. You can only put a sequences which have a specified length as an input - and in case your inputs vary in a length - you should use either an artificial data to \"fill\" your sequences or use a \"stateful\" mode (please read carefully Keras documentation to understand what this approach means). Both solutions might be unpleasent - but it's a cost you pay that Keras is so simple :) I hope that in version 1.0.0 they will do something with that.<\/p><\/li>\n<li><p>There are two ways to apply norecurrent layers after LSTM ones: <\/p>\n\n<ul>\n<li>you could set an argument return_sequences to False - then only the last activations from every sequence will be passed to a \"static\" layer.<\/li>\n<li>you could use one of \"time distributed\" layers - to get more flexibility with what you want to do with your data.<\/li>\n<\/ul><\/li>\n<li><p><a href=\"https:\/\/stats.stackexchange.com\/questions\/182775\/what-is-an-embedding-layer-in-a-neural-network\">https:\/\/stats.stackexchange.com\/questions\/182775\/what-is-an-embedding-layer-in-a-neural-network<\/a> :)<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":31927565,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2015-08-10T19:27:56.533","AcceptedAnswerId":null,"OwnerUserId":2293069.0,"Title":"Torch.gesv B should be 2 dimensional","Body":"<p>I just start taking the oxford machine learning course and I am new in lua and torch.<\/p>\n\n<p>I am trying to solve a simple linear equation problem with torch.\nThe problem is like AX = B<\/p>\n\n<p>However, I am not able to this because B is just a 1-D tensor (a vector). I think the case of B being a vector should be a common. And duplicating B into a 2-D tensor is wasteful. <\/p>\n\n<pre><\/pre>\n\n<p>And I will get:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions?<\/p>\n","answers":[]}
{"QuestionId":31935294,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-11T07:06:23.050","AcceptedAnswerId":31940737.0,"OwnerUserId":5197007.0,"Title":"How to get the predictation result in csv format from shared variables","Body":"<p>Data is shared variables. I want to get the predictation result in csv format. Below is the code. \nIt throws an error. How to fix?  Thank you for your help!<\/p>\n\n<pre><\/pre>\n\n<p>And more information about \"index\".<\/p>\n\n<pre><\/pre>\n\n<p>when enter:<\/p>\n\n<pre><\/pre>\n\n<p>The console shows:    <\/p>\n\n<pre><\/pre>\n\n<p>When enter:<\/p>\n\n<pre><\/pre>\n\n<p>It throws an error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"31940737","CreationDate":"2015-08-11T11:38:20.510","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Your <code>test_model<\/code> function has a single input value,<\/p>\n\n<pre><code>inputs=[index],\n<\/code><\/pre>\n\n<p>Your pasted code doesn't show the creation of the variable <code>index<\/code> but my guess is that it's a Theano symbolic scalar with an integer type. If so, you need to call the compiled function with a single integer input, for example<\/p>\n\n<pre><code>test_model(1)\n<\/code><\/pre>\n\n<p>You are trying to call <code>test_model(test_set_x)<\/code> which doesn't work because <code>test_set_x<\/code> is (again probably) a shared variable, not the integer index the function is expecting.<\/p>\n\n<p>Note that the <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"nofollow\">tutorial code<\/a> does this:<\/p>\n\n<pre><code>test_losses = [test_model(i) for i in xrange(n_test_batches)]\n<\/code><\/pre>\n"}]}
{"QuestionId":31936080,"AnswerCount":2,"Tags":"<python><ipython><caffe>","CreationDate":"2015-08-11T07:50:15.417","AcceptedAnswerId":32722501.0,"OwnerUserId":148346.0,"Title":"Construct caffe.Net object using NetParameter","Body":"<p>From the <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1Net.html\" rel=\"noreferrer\">documentation<\/a> I thought there was a constructor taking a NetParameter argument,<\/p>\n\n<blockquote>\n  <p>explicit Net(const NetParameter&amp; param);<\/p>\n<\/blockquote>\n\n<p>but when I try to use it like this:<\/p>\n\n<pre><\/pre>\n\n<p>The below error occurs in ipython<\/p>\n\n<pre><\/pre>\n\n<p>So what am I doing wrong here?  How do I use this constructor?<\/p>\n\n<p>Note: I know how use the \"read file from disk constructor\" already, I want to use the NetParameter one \/ or understand why it doesn't work.<\/p>\n\n<p>Edit after Shai's comment:<\/p>\n\n<p>I acquired caffe using this command on Jul 26, 2015:\ngit clone <a href=\"https:\/\/github.com\/BVLC\/caffe.git\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe.git<\/a><\/p>\n\n<p>Here's the file on my disk:<\/p>\n\n<pre><\/pre>\n\n<p>The -version switch appears to do nothing.  I grepped through the source and was unable to find a version number.<\/p>\n","answers":[{"AnswerId":"32722501","CreationDate":"2015-09-22T16:51:49.453","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Nothing wrong with your code. Indeed there's an overloaded constructor of the Net class in C++, but it's currently not exposed by the python interface. The python interface is limited to the constructor with the file param.<\/p>\n\n<p>I'm not sure if simply exposing it in python\/caffe\/_caffe.cpp is the only thing keeping us from constructing a python Net object with NetParameter or if more elaborate changes are needed.<\/p>\n"},{"AnswerId":"34172374","CreationDate":"2015-12-09T06:47:09.053","ParentId":null,"OwnerUserId":"2932001","Title":null,"Body":"<p>I came across the same problem and I get a solution <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/OGLXn8qQzGI\" rel=\"nofollow\">on google user group<\/a>, which explains that your c++ boost lib is too old, you may need to update it.<\/p>\n"}]}
{"QuestionId":31962975,"AnswerCount":3,"Tags":"<python><opencv><ubuntu><anaconda><caffe>","CreationDate":"2015-08-12T10:46:26.110","AcceptedAnswerId":32010074.0,"OwnerUserId":3633250.0,"Title":"Caffe install on ubuntu for anaconda with python 2.7 fails with libpng16.so.16 not found","Body":"<p>So I have installed anaconda with python 2.7 and installed all of the requirements for Caffe library. I ensured that opencv is installed by<\/p>\n\n<pre><\/pre>\n\n<p>And checking that I can run couple of examples from docs.<\/p>\n\n<p>Now I download caffe, configure makefile.config properly and run make all. I get very odd error:<\/p>\n\n<pre><\/pre>\n\n<p>What's wrong with that guy? Notice that I originally had anaconda3 and compiled caffe for it but successfully but I faced tons of issues with caffe under python3, so I had to remove it and try to set it up for anaconda with python 2.7.<\/p>\n\n<p>And of course I have ensured that libpng16.so.16 is in anaconda:<\/p>\n\n<pre><\/pre>\n\n<p>I googled the error, but haven't found anything in relation to caffe.<\/p>\n","answers":[{"AnswerId":"32010074","CreationDate":"2015-08-14T12:25:49.790","ParentId":null,"OwnerUserId":"3633250","Title":null,"Body":"<p>Per @cel suggestion - <\/p>\n\n<pre><code>ldd libopencv_highgui.so \n<\/code><\/pre>\n\n<p>shows the files on which this lib depends. Couple of them (not the libpng!) were located in folder which I haven't included into the makefile.config. After including their folder into MakeFile build succeeded. Notice: after building the caffe you may won't to go in Spyder into the PythonPath manager and add the caffe's folder into it (or just include it into pythonpath if you are not using anaconda\\spyder).<\/p>\n"},{"AnswerId":"32438450","CreationDate":"2015-09-07T12:11:31.137","ParentId":null,"OwnerUserId":"116067","Title":null,"Body":"<p>I ran into the same problem and I fixed it by adding an <code>-rpath<\/code> in my Makefile.config :<\/p>\n\n<p><code>LINKFLAGS := -Wl,-rpath,$(HOME)\/anaconda\/lib<\/code><\/p>\n\n<p>I think this is the correct fix because it (-rpath) tells GCC where it can find libraries (libjpeg, libpng) that other libraries (in this case opencv) depend on.<\/p>\n"},{"AnswerId":"33126417","CreationDate":"2015-10-14T13:17:19.357","ParentId":null,"OwnerUserId":"2686319","Title":null,"Body":"<p>Adding <\/p>\n\n<pre><code>LINKFLAGS := -Wl,-rpath,$(HOME)\/anaconda\/lib\n<\/code><\/pre>\n\n<p>in to Makefile.config worked.<\/p>\n"}]}
{"QuestionId":31964764,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-12T12:07:18.367","AcceptedAnswerId":31966758.0,"OwnerUserId":3054356.0,"Title":"Theano function analyse givens","Body":"<p>How can we monitor\/analyse the  of a theano function? <\/p>\n\n<p>As an example consider the following function:<\/p>\n\n<pre><\/pre>\n\n<p>What would be a way to monitor\/analyse the shared variables x and y?<\/p>\n","answers":[{"AnswerId":"31966758","CreationDate":"2015-08-12T13:33:38.803","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you're following\/using the code from the Theano tutorials, as appears to be the case, then <code>x_train<\/code> and <code>y_train<\/code> are shared variables containing your training data (<code>x_train<\/code> is the input and <code>y_train<\/code> is the true\/actual output that you want your model to predict when correct).<\/p>\n\n<p>The contents of these shared variables never (or, at least, shouldn't) change because your training data is normally static while a model is training.<\/p>\n\n<p>So, looking at the contents of shared variables <code>train_x<\/code> and <code>train_y<\/code> is just the same as looking at your training data. You can presumably just go look at the data wherever you load it from (e.g. maybe CSV data files, or numpy saved arrays, etc.)<\/p>\n\n<p>If you really want to look at the contents of a shared variable then you can do this using the <code>get_value()<\/code> method which returns the underlying numpy array:<\/p>\n\n<pre><code>x_data = X_train.get_value()\nprint x_data.shape\n# etc.\n<\/code><\/pre>\n\n<p>Theano is not involved at all here. Nothing is symbolic, it's just concrete numpy arrays.<\/p>\n"}]}
{"QuestionId":31966126,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2015-08-12T13:04:37.440","AcceptedAnswerId":31966191.0,"OwnerUserId":2711403.0,"Title":"what is the difference between x.type and type(x) in Python?","Body":"<p>Consider the following lines <\/p>\n\n<pre><\/pre>\n\n<p>And then,<\/p>\n\n<pre><\/pre>\n\n<p>while,<\/p>\n\n<pre><\/pre>\n\n<p>Why type(x) and x.type give two different pieces of information ? What information is conveyed by them ?<\/p>\n\n<p>I also see that referring to <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/adding.html\" rel=\"nofollow\">Theano tutorial<\/a> ,<\/p>\n\n<pre><\/pre>\n\n<p>Why type(x) output is different in my case ? Are these caused by version specific implementation differences and what is signified by this difference ?<\/p>\n","answers":[{"AnswerId":"31966191","CreationDate":"2015-08-12T13:07:32.897","ParentId":null,"OwnerUserId":"2296458","Title":null,"Body":"<p><code>theano.tensor<\/code> has an attribute <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.type\" rel=\"nofollow\"><code>type<\/code><\/a> which you are looking at when you say<\/p>\n\n<pre><code>x.type\n<\/code><\/pre>\n\n<p>This is analagous to numpy objects <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.dtype.html\" rel=\"nofollow\"><code>dtype<\/code><\/a> attribute that many of their objects carry (if you are familiar with that library).<\/p>\n\n<p>On the other hand <a href=\"https:\/\/docs.python.org\/3\/library\/functions.html#type\" rel=\"nofollow\"><code>type<\/code><\/a> is a Python function that looks at the actual type of the object you pass in, which for <code>type(x)<\/code> is indeed a <\/p>\n\n<pre><code>theano.tensor.var.TensorVariable\n<\/code><\/pre>\n\n<p>So long story short, you are comparing an attribute to the actual object type.<\/p>\n"},{"AnswerId":"31966200","CreationDate":"2015-08-12T13:07:51.483","ParentId":null,"OwnerUserId":"3398583","Title":null,"Body":"<p><code>type(x)<\/code> is a builtin.<\/p>\n\n<p><code>x.type<\/code> is an attribute that's defined in your object.<\/p>\n\n<p>They are completely seperate, <code>type(x)<\/code> returns what type of object <code>x<\/code> is and <code>x.type<\/code> does whatever the object wants it to. In this case, it returns some information on the type of object it is<\/p>\n"},{"AnswerId":"32140234","CreationDate":"2015-08-21T12:21:57.730","ParentId":null,"OwnerUserId":"650654","Title":null,"Body":"<p>As others have mentioned, <code>type(x)<\/code> is Python's <a href=\"https:\/\/docs.python.org\/2\/library\/functions.html#type\" rel=\"nofollow\">builtin function<\/a> that returns the type of the object.  It has nothing to do with Theano per se.  This builtin function can be applied to any Python object (and everything in Python is an object).  For example,<\/p>\n\n<ul>\n<li><code>type(1)<\/code> is <code>int<\/code>,<\/li>\n<li><code>type(True)<\/code> is <code>bool<\/code>,<\/li>\n<li><code>type(lambda x: x * x)<\/code> is <code>function<\/code>, etc.<\/li>\n<\/ul>\n\n<p>Interestingly, you can call <code>type<\/code> on <code>type<\/code> itself (everything, including <code>type<\/code>, is an object) - <code>type(type)<\/code> is <code>type<\/code>.<\/p>\n\n<p>Incidentally, <code>type(T.dscalar)<\/code> is <code>TensorType<\/code> (<code>theano.tensor.type.TensorType<\/code> to be precise).<\/p>\n\n<p><code>x.type<\/code>, as others have mentioned, is an attribute of the object <code>x<\/code>.  It points back to <code>type(T.dscalar)<\/code>.  <code>x.type<\/code> returns <code>TensorType(float64, scalar)<\/code> - this not only shows you the type of <code>T.dscalar<\/code>, it also tells you that <code>x<\/code> is scalar and it is 64-bit float.<\/p>\n\n<p>Other examples of the type attribute:<\/p>\n\n<pre><code>&gt;&gt;&gt; iv = T.ivector()\n&gt;&gt;&gt; iv.type\nTensorType(int32, vector)    # iv is a vector of 32-bit ints\n&gt;&gt;&gt; fm = T.fmatrix()\n&gt;&gt;&gt; fm.type\nTensorType(float32, matrix)   # fm is a matrix of 32-bit floats\n&gt;&gt;&gt; lt3 = T.ltensor3()\n&gt;&gt;&gt; lt3.type\nTensorType(int64, 3D)    # lt3 is a 3D array of 64-bit ints\n<\/code><\/pre>\n"}]}
{"QuestionId":31971462,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-12T17:10:13.843","AcceptedAnswerId":null,"OwnerUserId":3697238.0,"Title":"Bernoulli in theano","Body":"<p>I have a question about Bernoulli mask. <\/p>\n\n<p>As far as I understand the mask should be depending on the rate (probability) p. ( e.g. if p= 0.5 and the mask is an array A with size=2, then the mask should be something like:  [0,1] or [1,0]). \nThe most of theano codes use bernoulli like :<\/p>\n\n<p>rs = np.random.RandomState(1234)<\/p>\n\n<p>rng = theano.tensor.shared_randomstreams.RandomStreams(rs.randint(999999))<\/p>\n\n<p>mask = rng.binomial(n=1, p=(0.5), size=A.shape)<\/p>\n\n<p>but when I test this I find out that the mask could also be [0,0] or [1,1], which seems for me to be not logic. Because I want to randomly set exactly the half of the array to zeros. \nIs there maybe a bug? or maybe there is an alternative that theano provides for that purpose. \nThank you in advance!<\/p>\n","answers":[{"AnswerId":"31971699","CreationDate":"2015-08-12T17:25:05.797","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>rng.binomial(n=1, p=0.5, shape=A.shape)<\/code> will sample from a Bernoulli distribution independently for each element of the result tensor with shape <code>A.shape<\/code>. Because each sample is independent each element of the result tensor will be zero with probability 0.5 and 1 with probability 0.5. Consequently, if the result tensor should have shape <code>(2, )<\/code> (i.e. a vector of length 2), there are four possible outcomes and each will be obtained with probability 0.25:<\/p>\n\n<pre><code>[0, 0]\n[0, 1]\n[1, 0]\n[1, 1]\n<\/code><\/pre>\n\n<p>It's not clear what your use-case is but if it is for a denoising autoencoder for example, this is the usual approach; sometimes you drop more features than at other times. Dropping all features is rather extreme but this is an unlikely outcome when the size of the result tensor is much larger than 2.<\/p>\n\n<p>If you really need to mask exactly half the elements then you might be able to use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/raw_random.html\" rel=\"nofollow\"><code>theano.tensor.raw_random.shuffle_row_elements<\/code><\/a>. I've not tried this but the idea would be to symbolically shuffle a list of indexes using <code>shuffle_row_elements<\/code>, select the first half of the resulting list, then use <code>set_subtensor<\/code> to mask just those elements in the original tensors at the selected indexes.<\/p>\n"}]}
{"QuestionId":31972448,"AnswerCount":1,"Tags":"<c++><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-12T18:09:57.523","AcceptedAnswerId":31984228.0,"OwnerUserId":2467772.0,"Title":"How to understand the Cifar10 prediction output?","Body":"<p>I have trained (<a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">caffe<\/a>) model for two classes classification. Pedestrian and non-pedestrian. Training looks fine, I have updated weights in a  file. I used two labels 1 for pedestrians and 2 for non-pedestrians, together with images for pedestrians (64 x 160) and background images (64 x 160).\nAfter training, I do testing with positive image(pedestrian image) and negative image (background image). My testing  file is as shown below<\/p>\n\n<pre><\/pre>\n\n<p>For testing, I used  and did some modifications especially for paths and image size.<\/p>\n\n<p>I can't figure out the test output. When I test with positive image, I got the output as<\/p>\n\n<pre><\/pre>\n\n<p>When I test with a negative image, I got the output as<\/p>\n\n<pre><\/pre>\n\n<p>How to understand the testing output?<\/p>\n\n<p>Is there any more effecient testing algorithm for testing the model from video feed (frame by frame from video clip)?<\/p>\n","answers":[{"AnswerId":"31984228","CreationDate":"2015-08-13T09:20:07.437","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Why do you have <code>num_output: 10<\/code> for the last layer <code>ip2<\/code>? you only need 2-way classifier? Why are you using labels 1 and 2 instead of 0 and 1?<\/p>\n\n<p><strong>What you got<\/strong>: You have 11 outputs: one is the <code>\"label\"<\/code> output from the data layer, and the other 10 outputs are the 10-vector output of the softmax layer. It is unclear what the values of the 10-vector are since you only trained using two labels, thus 8 out of 10 entries were not supervised at all. Moreover, judging by the first output it seems both tests were samples with label <code>1<\/code> and not <code>2<\/code>.<\/p>\n\n<p><strong>What you should do:<\/strong><br>\n1. Change the topmost fully connected layer to have only two outputs (I also changed the format to match the new version protobuff)<\/p>\n\n<pre><code>layer {\n  name: \"ip2\/pedestrains\"\n  type: \"InnerProduct\"\n  bottom: \"ip1\"\n  top: \"ip2\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 2 # This is what you need changing\n  }\n}\n<\/code><\/pre>\n\n<p>2. Change the binary labels in your training data to 0\/1 rather that 1\/2.  <\/p>\n\n<p>Now you can train again and see what you get.<\/p>\n"}]}
{"QuestionId":31978186,"AnswerCount":3,"Tags":"<c++><classification><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-13T01:43:08.970","AcceptedAnswerId":32006840.0,"OwnerUserId":2281121.0,"Title":"Monitor training\/validation process in Caffe","Body":"<p>I'm training Caffe Reference Model for classifying images.\nMy work requires me to monitor the training process by drawing graph of accuracy of the model after every 1000 iterations on entire training set and validation set which has 100K and 50K images respectively. \nRight now, Im taking the naive approach, make snapshots after every 1000 iterations, run the C++ classififcation code which reads raw JPEG image and forward to the net and output the predicted labels. However, this takes too much time on my machine (with a Geforce GTX 560 Ti)<\/p>\n\n<p>Is there any faster way that I can do to have the graph of accuracy of the snapshot models on both training and validation sets?<\/p>\n\n<p>I was thinking about using LMDB format instead of raw images. However, I cannot find documentation\/code about doing classification in C++ using LMDB format.<\/p>\n","answers":[{"AnswerId":"38478071","CreationDate":"2016-07-20T09:58:07.317","ParentId":null,"OwnerUserId":"2736559","Title":null,"Body":"<p>Caffe creates logs each time you try to train something, and its located in the tmp folder (both linux and windows).<br>\nI also wrote a plotting script in python which you can easily use to visualize your loss\/accuracy.<br>\nJust place your training logs with <code>.log<\/code> extension next to the script and double click on it.\nYou can use command prompts as well, but for ease of use, when executed it loads all logs (*.log) it can find in the current directory. \nit also shows the top 4 accuracies and at-which accuracy they were achieved. <\/p>\n\n<p>you can find it here : <a href=\"https:\/\/gist.github.com\/Coderx7\/03f46cb24dcf4127d6fa66d08126fa3b\" rel=\"nofollow\">https:\/\/gist.github.com\/Coderx7\/03f46cb24dcf4127d6fa66d08126fa3b<\/a> <\/p>\n"},{"AnswerId":"44223960","CreationDate":"2017-05-28T04:59:48.543","ParentId":null,"OwnerUserId":"8076417","Title":null,"Body":"<pre><code>python \/pathtocaffe\/tools\/extra\/parse_log.py lenet_train.log\n<\/code><\/pre>\n\n<p>command produces the following error:<\/p>\n\n<pre><code>usage: parse_log.py [-h] [--verbose] [--delimiter DELIMITER]\n                logfile_path output_dir\nparse_log.py: error: too few arguments\n<\/code><\/pre>\n\n<p>Solution:<\/p>\n\n<p>For successful execution of \"parse_log.py\" command, we should pass the two arguments:<\/p>\n\n<ol>\n<li>log file<\/li>\n<li>path of output directory<\/li>\n<\/ol>\n\n<p>So the correct command is as follows:<\/p>\n\n<pre><code>python \/pathtocaffe\/tools\/extra\/parse_log.py lenet_train.log output_dir\n<\/code><\/pre>\n"},{"AnswerId":"32006840","CreationDate":"2015-08-14T09:32:54.823","ParentId":null,"OwnerUserId":"2347543","Title":null,"Body":"<p>1) You can use the <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\" rel=\"noreferrer\">NVIDIA-DIGITS<\/a> app to monitor your networks. They provide a GUI including dataset preparation, model selection, and learning curve visualization. More, they use a caffe distribution allowing <a href=\"https:\/\/developer.nvidia.com\/digits\" rel=\"noreferrer\">multi-GPU training<\/a>.<\/p>\n\n<p>2) Or, you can simply use the log-parser inside caffe.<\/p>\n\n<pre><code>\/pathtocaffe\/build\/tools\/caffe train --solver=solver.prototxt 2&gt;&amp;1 | tee lenet_train.log\n<\/code><\/pre>\n\n<p>This allows you to save train log into \"lenet_train.log\". Then by using:<\/p>\n\n<pre><code>python \/pathtocaffe\/tools\/extra\/parse_log.py lenet_train.log .\n<\/code><\/pre>\n\n<p>you parse your train log into two csv files, containing train and test loss. You can then plot them using the following python script<\/p>\n\n<pre><code>import pandas as pd\nfrom matplotlib import *\nfrom matplotlib.pyplot import *\n\ntrain_log = pd.read_csv(\".\/lenet_train.log.train\")\ntest_log = pd.read_csv(\".\/lenet_train.log.test\")\n_, ax1 = subplots(figsize=(15, 10))\nax2 = ax1.twinx()\nax1.plot(train_log[\"NumIters\"], train_log[\"loss\"], alpha=0.4)\nax1.plot(test_log[\"NumIters\"], test_log[\"loss\"], 'g')\nax2.plot(test_log[\"NumIters\"], test_log[\"acc\"], 'r')\nax1.set_xlabel('iteration')\nax1.set_ylabel('train loss')\nax2.set_ylabel('test accuracy')\nsavefig(\".\/train_test_image.png\") #save image as png\n<\/code><\/pre>\n"}]}
{"QuestionId":31983794,"AnswerCount":2,"Tags":"<python><gpu><theano><lasagne>","CreationDate":"2015-08-13T09:00:17.930","AcceptedAnswerId":31983939.0,"OwnerUserId":2902280.0,"Title":"Select GPU during execution in Theano","Body":"<p>I am training neural nets with theano and lasagne on a 4 GPU machine. My  contains the following lines:<\/p>\n\n<pre><\/pre>\n\n<p>So when in python I execute  , I get <\/p>\n\n<p>What if, after importing theano, I chose to use say gpu1? I'd like to do this dynamically, that is, without editing  is it possible? Or even to choose it at runtime?<\/p>\n","answers":[{"AnswerId":"31983939","CreationDate":"2015-08-13T09:07:18.683","ParentId":null,"OwnerUserId":"1692028","Title":null,"Body":"<p>I'm afraid it's not possible to change the execution device after Theano has been imported. From the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>config.device<\/p>\n  \n  <p>String value: either 'cpu', 'gpu', 'gpu0', 'gpu1',\n  'gpu2', or 'gpu3'<\/p>\n  \n  <p>[...]<\/p>\n  \n  <p><em>This flag\u2019s value cannot be modified during\n  the program execution.<\/em><\/p>\n<\/blockquote>\n\n<p><strong>Bonus<\/strong>: however, let's say you wanted to have two Python processes each running on a separate GPU (is that what you want?), then you could do something like:<\/p>\n\n<pre><code>import os\nos.system(\"THEANO_FLAGS='device=gpu0' python myscript.py\")\nos.system(\"THEANO_FLAGS='device=gpu1' python myscript.py\")\n<\/code><\/pre>\n\n<p>or hack into\/extend Python's <a href=\"https:\/\/docs.python.org\/3.5\/library\/multiprocessing.html\">multiprocessing<\/a> module (which works by spawning subprocesses) to ensure the flag is set before a child process is spawned.<\/p>\n"},{"AnswerId":"36829078","CreationDate":"2016-04-24T21:27:28.603","ParentId":null,"OwnerUserId":"1667840","Title":null,"Body":"<p>EDIT:\nTheano is now based on the GPU array backend and the following API is no longer available. <\/p>\n\n<p>As @EelkeSpaak mentioned, you can't change the GPU device after theano was imported. But if you want to choose it programmatically before that's possible without changing environment variables. <\/p>\n\n<ol>\n<li><p>Make sure you're not choosing a device in your .theanorc file. \nSo nothing like: <\/p>\n\n<p><code>device=gpu<\/code><\/p><\/li>\n<li><p>before calling <code>import theano<\/code> choose the GPU device as follows:<\/p>\n\n<pre><code>import theano.sandbox.cuda\ntheano.sandbox.cuda.use('gpu1')\n\n#Results in Using gpu device 1: Tesla K80\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":31988629,"AnswerCount":1,"Tags":"<python><theano><conv-neural-network>","CreationDate":"2015-08-13T12:44:01.263","AcceptedAnswerId":31989059.0,"OwnerUserId":4899439.0,"Title":"Theano - Shared variable as input of function for large dataset","Body":"<p>I am new to Theano... My apologies if this is obvious.<\/p>\n\n<p>I am trying to train a CNN, based on the <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">LeNet tutorial<\/a>. A major difference with the tutorial is that my dataset is too large to fit in memory, so I have to load each batch during training.<\/p>\n\n<p>The original model has this:<\/p>\n\n<pre><\/pre>\n\n<p>...Which does not work for me, as it assumes that  is entirely loaded in memory.<\/p>\n\n<p>So I switched to this:<\/p>\n\n<pre><\/pre>\n\n<p>And tried to call it with:<\/p>\n\n<pre><\/pre>\n\n<p>But got:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name \":103\"  at index 0(0-based)', 'Expected an array-like object, but found a Variable: maybe you are trying to call a function on a (possibly shared) variable instead of a numeric array?')<\/p>\n<\/blockquote>\n\n<p>So I guess I can't use a shared variable as an input to a Theano function. But then, how should I proceed...?<\/p>\n","answers":[{"AnswerId":"31989059","CreationDate":"2015-08-13T13:03:36.290","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>All inputs to compiled Theano functions (i.e the output of a call to <code>theano.function(...)<\/code>) should always be concrete values, typically scalars or numpy arrays. Shared variables are a way to wrap a numpy array and treat it like a symbolic variable but this is not necessary when the data is being passed as an input.<\/p>\n\n<p>So you should be able to just omit wrapping your data and target values as shared variables and do the following instead:<\/p>\n\n<pre><code>cost_ij = train_model(data, target)\n<\/code><\/pre>\n\n<p>Note that, if you're using a GPU, this means your data will reside in the computer's main memory, and each part you pass as input will need to be copied to the GPU memory separately, increasing overhead and slowing it down. Also note that you will have to divide your data up and only pass part of it; this change of approach won't allow you to do GPU computations over your whole dataset at once if the whole dataset won't fit in GPU memory.<\/p>\n"}]}
{"QuestionId":31992971,"AnswerCount":0,"Tags":"<deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-13T15:49:43.417","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Detecting pedestrian using CaffeNet model at moderate framerate","Body":"<p>I train the CaffeNet (<a href=\"https:\/\/stackoverflow.com\/questions\/31972448\/how-to-understand-the-cifar10-prediction-output\">more precisely Cifar10 model for two classes classification<\/a>) model. Now the model is ready for detection. For the model testing with a single image, I use . I haven't tested how fast the code can run for 640 x 480 image. My target is I like to have 5~10 frame\/sec is just nice for offline detection. I understand that I need to implement multi-size detection (i.e. something like we do in face detection, original image size is re-sized for different smaller sizes) so that I don't miss the pedestrian in each frame.                    <\/p>\n\n<p>According to this <a href=\"http:\/\/arxiv.org\/pdf\/1501.05790.pdf\" rel=\"nofollow noreferrer\">paper<\/a>, they use 64 x 128 image size in training and detection takes 3ms\/window and for 100 windows\/image, it takes 300msec\/frame. Not sure they implement multi-size detection approach. If multi-size is implemented, it will take much longer.  <\/p>\n\n<p>At this moment, I have only knowledge of implementing  method for multi-size detection. I know it will be very slow.\nIs there any more efficient way of detection using CaffeNet model? My target is just nice for 5~10 frame\/sec rate. Thanks<\/p>\n","answers":[]}
{"QuestionId":31997124,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-08-13T19:46:25.900","AcceptedAnswerId":31999155.0,"OwnerUserId":2293069.0,"Title":"Grid search for hyper-paramerters in torch \/ lua","Body":"<p>I am new in torch\/lua and am trying evaluate some different optimization algorithms and different parameters for each of them.<\/p>\n\n<p>Algo: optim.sgd optim.lbfgs<\/p>\n\n<p>Parameters:<\/p>\n\n<ul>\n<li>learning_rate: {1e-1, 1e-2, 1e-3}<\/li>\n<li>weight_decay: {1e-1, 1e-2}<\/li>\n<\/ul>\n\n<p>So what I am trying to achieve is try every combination of the hyper-parameters and get the optimal parameter set for each of the algorithm.<\/p>\n\n<p>So is there something like:<\/p>\n\n<pre><\/pre>\n\n<p>as in <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/grid_search.html\" rel=\"nofollow\">http:\/\/scikit-learn.org\/stable\/modules\/grid_search.html<\/a> available in torch to deal with it?<\/p>\n\n<p>Any suggestions would be nice!<\/p>\n","answers":[{"AnswerId":"31999155","CreationDate":"2015-08-13T22:01:39.327","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Try this hyper-optimization library that is being worked on:\n<a href=\"https:\/\/github.com\/nicholas-leonard\/hypero\" rel=\"nofollow\">https:\/\/github.com\/nicholas-leonard\/hypero<\/a><\/p>\n"}]}
{"QuestionId":31997366,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2015-08-13T20:00:30.607","AcceptedAnswerId":32938020.0,"OwnerUserId":3054356.0,"Title":"Python: keras shape mismatch error","Body":"<p>I am trying to build a very simple multilayer perceptron (MLP) in :<\/p>\n\n<pre><\/pre>\n\n<p>My training data shape:  gives <\/p>\n\n<p>The labels belong to binary class with shape:  gives <\/p>\n\n<p>So my  code should produce the network with following connection: <\/p>\n\n<p>which produces the shape mismatch error:<\/p>\n\n<pre><\/pre>\n\n<p>At  at line . Am I overseeing something obvious in Keras?<\/p>\n\n<p><strong>EDIT:<\/strong> I have gone through the question <a href=\"https:\/\/stackoverflow.com\/questions\/30384908\/python-keras-neural-network-theano-package-returns-an-error-about-data-dimensi\/30386018#30386018\">here<\/a> but does not solve my problem<\/p>\n","answers":[{"AnswerId":"32938020","CreationDate":"2015-10-04T20:26:27.393","ParentId":null,"OwnerUserId":"5407700","Title":null,"Body":"<p>I had the same problem and then found this thread; <\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/68\">https:\/\/github.com\/fchollet\/keras\/issues\/68<\/a><\/p>\n\n<p>It appears for you to state a final output layer of 2 or for any number of categories the labels need to be of a categorical type where essentially this is a binary vector for each observation e.g a 3 class output vector [0,2,1,0,1,0] becomes [[1,0,0],[0,0,1],[0,1,0],[1,0,0],[0,1,0],[1,0,0]].<\/p>\n\n<p>The np_utils.to_categorical function solved this for me;<\/p>\n\n<pre class=\"lang-php prettyprint-override\"><code>from keras.utils import np_utils, generic_utils\n\ny_train, y_test = [np_utils.to_categorical(x) for x in (y_train, y_test)]\n<\/code><\/pre>\n\n"}]}
{"QuestionId":31998824,"AnswerCount":1,"Tags":"<neural-network><caffe>","CreationDate":"2015-08-13T21:34:15.877","AcceptedAnswerId":32013403.0,"OwnerUserId":4080543.0,"Title":"Parser in Caffe","Body":"<p>I am trying to find the parser in Caffe. By parser, I mean the part of the code that reads the network configuration from a file and parses it. I was wondering if anyone knows where in the Caffe codebase I should look for this specific piece of code. <\/p>\n","answers":[{"AnswerId":"32013403","CreationDate":"2015-08-14T15:18:07.097","ParentId":null,"OwnerUserId":"2014584","Title":null,"Body":"<p>Caffe's text file format for specifying models uses the <a href=\"https:\/\/developers.google.com\/protocol-buffers\/?hl=en\" rel=\"nofollow\">Google Protocol Buffer<\/a> format.<\/p>\n\n<p>You can see the code that reads a model in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/fc77ef3d23423e57d83e16c32844bbbe8d4d9f0c\/src\/caffe\/util\/io.cpp#L32\" rel=\"nofollow\">src\/caffe\/util\/io.cpp<\/a>:<\/p>\n\n<pre><code>bool ReadProtoFromTextFile(const char* filename, Message* proto) {\n  int fd = open(filename, O_RDONLY);\n  CHECK_NE(fd, -1) &lt;&lt; \"File not found: \" &lt;&lt; filename;\n  FileInputStream* input = new FileInputStream(fd);\n  bool success = google::protobuf::TextFormat::Parse(input, proto);\n  delete input;\n  close(fd);\n  return success;\n}\n<\/code><\/pre>\n\n<p>Try using GitHub's search to see places in the code that call this function.<\/p>\n"}]}
{"QuestionId":31999536,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-13T22:36:49.513","AcceptedAnswerId":32004645.0,"OwnerUserId":1157605.0,"Title":"Theano scan: how do I have a tuple output as input into next step?","Body":"<p>I want to do this sort of loop in Theano:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I do<\/p>\n\n<pre><\/pre>\n\n<p>I get .  If I change it so that the function takes a single tuple instead, <\/p>\n\n<p>In particular, I eventually want to differentiate the entire result with respect to k.<\/p>\n","answers":[{"AnswerId":"32004645","CreationDate":"2015-08-14T07:32:46.320","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The first error is because your <code>add_multiply<\/code> function takes 3 arguments but, by having only one element in the <code>outputs_info<\/code> list, you're only providing a single argument. It's not clear if you intended the <code>x0<\/code> vector to be the initial value for just <code>a<\/code> or were expecting it to be spread over <code>a<\/code>, <code>b<\/code>, and <code>k<\/code>. The latter isn't supported by Theano and, in general, tuples are not supported by Theano. In Theano, everything needs to be a tensor (e.g. scalars are just special types of tensors with zero dimensions).<\/p>\n\n<p>You can achieve a replica of the Python implementation in Theano as follows.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\n\ndef add_multiply(a, b, k):\n    return a + b + k, a * b * k\n\n\ndef python_main():\n    x = 1\n    y = 2\n    k = 1\n    tuples = []\n    for i in range(5):\n        x, y = add_multiply(x, y, k)\n        tuples.append((x, y, k))\n    return tuples\n\n\ndef theano_main():\n    x = tt.constant(1, dtype='uint32')\n    y = tt.constant(2, dtype='uint32')\n    k = tt.scalar(dtype='uint32')\n    outputs, _ = theano.scan(add_multiply, outputs_info=[x, y], non_sequences=[k], n_steps=5)\n    g = theano.grad(tt.sum(outputs), k)\n    f = theano.function(inputs=[k], outputs=outputs + [g])\n    tuples = []\n    xvs, yvs, _ = f(1)\n    for xv, yv in zip(xvs, yvs):\n        tuples.append((xv, yv, 1))\n    return tuples\n\n\nprint 'Python:', python_main()\nprint 'Theano:', theano_main()\n<\/code><\/pre>\n\n<p>Note that in the Theano version, all the tuple handling happens outside Theano; Python has to convert from the three tensors returned by the Theano function into a list of tuples.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>It's unclear what \"the entire result\" should refer to but the code has been updated to show how you might differentiate with respect to <code>k<\/code>. Note that in Theano the symbolic differentiation only works with scalar expressions, but can differentiate with respect to multi-dimensional tensors.<\/p>\n\n<p>In this update the <code>add_multiply<\/code> method no longer returns <code>k<\/code> since that is constant. For similar reasons, the Theano version now accepts <code>k<\/code> as a <code>non_sequence<\/code>.<\/p>\n"}]}
{"QuestionId":32001363,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2015-08-14T02:27:03.590","AcceptedAnswerId":32004904.0,"OwnerUserId":2789788.0,"Title":"Trying to understand variables... can I print them?","Body":"<p>If we use one of the tutorial examples:<\/p>\n\n<pre><\/pre>\n\n<p>Does  now hold the array ? <\/p>\n\n<p>How can I print out  after my call yo ? <\/p>\n\n<p>I'm having trouble understanding this functional style of usage. Can anyone explain?<\/p>\n","answers":[{"AnswerId":"32007867","CreationDate":"2015-08-14T10:25:57.913","ParentId":null,"OwnerUserId":"5226958","Title":null,"Body":"<p>We can not print a symbolic var since it remembers no real value.<\/p>\n\n<p>However, if you claim diff to be a shared var, u can do the following:\nprint diff.get_value()<\/p>\n"},{"AnswerId":"32004904","CreationDate":"2015-08-14T07:48:41.207","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I think you're hitting the most difficult conceptual challenge of learning Theano.<\/p>\n\n<p>To answer your specific questions first,<\/p>\n\n<ol>\n<li><p>No, <code>diff<\/code> does not hold the array <code>[[ 1.,  0.],[-1., -2.]]<\/code>. As far as Python is concerned <code>diff<\/code> is a reference to a Theano object. That object is a symbolic variable and never, itself, has or holds a value.<\/p><\/li>\n<li><p>When you ask how to print out <code>diff<\/code>, I take that to mean \"how can I print out the result of the computation that <code>diff<\/code> is a reference to?\" Well, in one sense, you're already doing it: <code>diff<\/code> is the output of the Theano function <code>f<\/code> so after <code>f<\/code> is compiled and then executed via <code>f([[1, 1], [1, 1]], [[0, 1], [2, 3]])<\/code> you see the result is <code>[array([[ 1.,  0.], [-1., -2.]])<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>What's difficult is separating the symbolic world from the concrete\/executable world. <code>diff<\/code> exists in the symbolic world and so doesn't really ever have a value. Everything that happens after compiling <code>f<\/code> via <code>theano.function(...)<\/code> is in the concrete\/executable world. With some caveats, nothing in the symbolic world is affected by what happens in the concrete\/executable world.<\/p>\n\n<p>It's also important to note that <code>diff<\/code> cannot \"have a value\" (in any sense) until the symbolic variables on which it depends, <code>a<\/code> and <code>b<\/code> have been given a value. If I posed a mathematics problem, telling you that <code>y = x * z<\/code> and asked you what value <code>y<\/code> has, you couldn't answer unless I also told you what value <code>x<\/code> and <code>z<\/code> have. The same is true here, the symbolic world operates like symbolic mathematics: it is an abstract definition of some computation that doesn't and can't have a value until compiled and input values are provided.<\/p>\n\n<p>When developing Theano code it can be helpful to separate the symbolic code from the concrete\/executable code into two separate areas. Without doing this you will often find, for example, wanting to call a symbolic variable <code>x<\/code> but then also wanting to give a concrete value that is to be provided for <code>x<\/code> the same name.<\/p>\n"},{"AnswerId":"32001571","CreationDate":"2015-08-14T02:54:16.767","ParentId":null,"OwnerUserId":"2957601","Title":null,"Body":"<p>In the console, just type the var name<\/p>\n\n<pre><code>&gt;&gt;&gt; diff\n<\/code><\/pre>\n\n<p>In Python 2, use <code>print diff<\/code>, and in Python 3, use <code>print(diff)<\/code><\/p>\n"}]}
{"QuestionId":32013927,"AnswerCount":2,"Tags":"<indexing><torch><one-hot-encoding>","CreationDate":"2015-08-14T15:46:02.190","AcceptedAnswerId":32015143.0,"OwnerUserId":1120370.0,"Title":"In Torch how do I create a 1-hot tensor from a list of integer labels?","Body":"<p>I have a byte tensor of integer class labels, e.g. from the MNIST data set.<\/p>\n\n<pre><\/pre>\n\n<p>How do use it to create a tensor of 1-hot vectors?<\/p>\n\n<pre><\/pre>\n\n<p>I know I could do this with a loop, but I'm wondering if there's any clever Torch indexing that will get it for me in a single line.<\/p>\n","answers":[{"AnswerId":"32015143","CreationDate":"2015-08-14T16:55:55.833","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>indices = torch.LongTensor{1,7,5}:view(-1,1)\none_hot = torch.zeros(3, 10)\none_hot:scatter(2, indices, 1)\n<\/code><\/pre>\n\n<p>You can find the documentation for <code>scatter<\/code> in the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-scatterdim-index-srcval\" rel=\"nofollow noreferrer\">torch\/torch7 github readme<\/a> (in the master branch).<\/p>\n"},{"AnswerId":"40506828","CreationDate":"2016-11-09T12:16:33.423","ParentId":null,"OwnerUserId":"4412981","Title":null,"Body":"<p>An alternate method is to shuffle rows from an identity matrix:<\/p>\n\n<pre><code>indicies = torch.LongTensor{1,7,5}\none_hot = torch.eye(10):index(1, indicies)\n<\/code><\/pre>\n\n<p>This was not my idea, I found it in <a href=\"https:\/\/github.com\/karpathy\/char-rnn\/blob\/master\/util\/OneHot.lua\" rel=\"nofollow noreferrer\">karpathy\/char-rnn<\/a>.<\/p>\n"}]}
{"QuestionId":32014110,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2015-08-14T15:55:58.447","AcceptedAnswerId":null,"OwnerUserId":2789788.0,"Title":"mingw seems to crash python (Anaconda on windows)","Body":"<p>I'm trying to install theano on a windows computer. When I do  python crashes. I think this is a mingw issue because when I uninstall mingw libpython and try to import theano, I get a theano warning, but python does not crash.<\/p>\n\n<p>I have tried Anaconda 2.10 and 2.30. I follow the installation up with , and then install the latest theano from github.  still crashes. Any ideas?<\/p>\n\n<p>Here are the details of the \"python.exe has stopped working\" message:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":32018923,"AnswerCount":1,"Tags":"<python><windows-7><64-bit><python-import><theano>","CreationDate":"2015-08-14T21:29:11.670","AcceptedAnswerId":54008530.0,"OwnerUserId":5228890.0,"Title":"Theano import Error-windows 7","Body":"<p>I have a problem in importing theano in python. When I import Theano in the python 27 32 bit, in Windows 7 64 bit, I get the following errors and warning:\nI also should add that currently I have installed GCC 4.8.1.\nWhat I have to do in order to fix it. <\/p>\n\n<p>Thanks,\nAfshin<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"54008530","CreationDate":"2019-01-02T14:58:46.637","ParentId":null,"OwnerUserId":"4152633","Title":null,"Body":"<p>You should check your windows path and make sure they are all right.<\/p>\n"}]}
{"QuestionId":32020815,"AnswerCount":1,"Tags":"<windows><import><theano>","CreationDate":"2015-08-15T02:07:54.157","AcceptedAnswerId":32035083.0,"OwnerUserId":5228890.0,"Title":"Theano Import Error-Windows 7-64b","Body":"<p>I just installed the Theano. \nI have a problem in importing theano in python. When I import Theano in the python 27 32 bit, in Windows 7 64 bit, I get the following errors and warning: I also should add that currently I have installed GCC 4.8.1. What I have to do in order to fix it.<\/p>\n\n<p>Thanks, Afshin<\/p>\n\n<pre><\/pre>\n\n<p>......\n(I just removed some codes, because Stachoverflow does not allow more than 30000 character......\n......<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32035083","CreationDate":"2015-08-16T12:28:52.350","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This was also discussed on the <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/jSG_xK2IPRM\" rel=\"nofollow\">Theano mailing list<\/a>. For the benefit of StackOverflow users, the answer was to install and use <a href=\"http:\/\/www.mingw.org\/\" rel=\"nofollow\">MinGW<\/a> in preference to Cygwin.<\/p>\n\n<p>If both are installed, MinGW should be before Cygwin in the <code>PATH<\/code>.<\/p>\n"}]}
{"QuestionId":32021975,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2015-08-15T05:48:01.197","AcceptedAnswerId":null,"OwnerUserId":1263702.0,"Title":"Python Caffe cpu & gpu mode simultaneously","Body":"<p>Is it possible to run Caffe in both CPU and GPU mode? I have several Caffe models, but my GPU resources are limited, so that I can't put all models into GPU memory. I want to use e.g. 3 models with GPU mode and 2 models with CPU mode, but  and  commands just switch the mode for the whole library.<\/p>\n","answers":[{"AnswerId":"35031858","CreationDate":"2016-01-27T08:10:24.187","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>In my opinion, you can write multiple python scripts one for each task.\nIn each script you can choose whether use CPU or GPU (and GPU device).\nThen you can run these multiple scripts at the same time.\nBut run multiple tasks at one GPU card with slow down the speed severely in my experience. Good luck!<\/p>\n"}]}
{"QuestionId":32034231,"AnswerCount":1,"Tags":"<python><theano><lasagne>","CreationDate":"2015-08-16T10:50:42.140","AcceptedAnswerId":32035049.0,"OwnerUserId":127400.0,"Title":"Lasagne\/Theano wrong number of dimensions","Body":"<p>Headed into Lasagne and Theano with a modified mnist.py (the primary example of Lasagne) to train a very simple XOR.<\/p>\n\n<pre><\/pre>\n\n<p>Defined a training set at (1), modified the input to the new dimension at (2) and get an exception at (3):<\/p>\n\n<pre><\/pre>\n\n<p>And I have no clue what I did wrong. When I print the dimension (or the output of the program until the exception) I get this<\/p>\n\n<pre><\/pre>\n\n<p>Which seem to be perfect. What I'm doing wrong and how must the array be formed to work?<\/p>\n","answers":[{"AnswerId":"32035049","CreationDate":"2015-08-16T12:25:21.423","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is with the second input, <code>targets<\/code>. Note that the error message indicated this by saying \"...at index 1(0-based)...\", i.e. the second parameter.<\/p>\n\n<p><code>target_var<\/code> is an <code>ivector<\/code> but you're providing a 4-dimensional tensor for <code>targets<\/code>. The solution is to alter your <code>y_train<\/code> dataset so that it is 1-dimensional:<\/p>\n\n<pre><code>y_train = [0, 1, 1, 0]\n<\/code><\/pre>\n\n<p>This will cause another error because you currently assert that the first dimension of the inputs and targets should match, but changing<\/p>\n\n<pre><code>assert len(inputs) == len(targets)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>assert inputs.shape[2] == len(targets)\n<\/code><\/pre>\n\n<p>will solve the second problem and allow the script to run successfully.<\/p>\n"}]}
{"QuestionId":32036386,"AnswerCount":1,"Tags":"<deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-16T14:51:21.567","AcceptedAnswerId":32055436.0,"OwnerUserId":2467772.0,"Title":"Error in testing Caffe's Alexnet caffe model","Body":"<p>I trained caffe's Alexnet model for testing with more efficient model. Since my training is for pedestrians my image size is 64 x 80 images. I made changes to prototxt files to match to my trained image size. According to this <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow\">tutorial<\/a>, it will be better to set the convolution filter size to match the input image size. So my filter sizes have slight changes from the original Alexnet's provided prototxt files (I trained and tested with Alexnet's original prototxt files and get the same error at the same line mentioned below).<\/p>\n\n<p>According to my calculation, image sizes after passing each layer will be<\/p>\n\n<p>80x64x3 -> Conv1 -> 38x30x96<br>\n38x30x96 -> Pools -> 18x14x96<br>\n18x14x96 -> Conv2 -> 19x15x256<br>\n19x15x256 -> Pool2 -> 9x7x256<br>\n9x7x256 -> Conv3 -> 9x7x384<br>\n9x7x384 -> Conv4 -> 9x7x384<br>\n9x7x384 -> Conv5 -> 9x7x256<br>\n9x7x256 -> Pool5 -> 4x3x256<\/p>\n\n<p>The error is at fc6 layer and line number 714 of . I use  file to test the model.<\/p>\n\n<pre><\/pre>\n\n<p>The error is <\/p>\n\n<pre><\/pre>\n\n<p>I don't understand why it is like that.<\/p>\n\n<p>My two prototxt files are shown below.<\/p>\n\n<pre><\/pre>\n\n<p>This is the testing file for the model.<\/p>\n\n<pre><\/pre>\n\n<p>What is wrong with this error?<\/p>\n","answers":[{"AnswerId":"32055436","CreationDate":"2015-08-17T16:33:56.437","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>Those who have the same problem as i faced, please look at the prototxt files shown below. There are some modifications made compared to the original prototxt files provided in the downloaded folders. I used 80x64 image sizes for the input in training and testing.<\/p>\n\n<pre><code>Train_val.prototxt\nname: \"AlexNet\"\nlayers {\n  name: \"data\"\n  type: DATA\n  top: \"data\"\n  top: \"label\"\n  data_param {\n    source: \"..\/..\/examples\/Alexnet_2\/Alexnet_train_leveldb\"\n    batch_size: 100\n  }\n  transform_param {\n    mean_file: \"..\/..\/examples\/Alexnet_2\/mean.binaryproto\"\n\n  }\n  include: { phase: TRAIN }\n}\nlayers {\n  name: \"data\"\n  type: DATA\n  top: \"data\"\n  top: \"label\"\n  data_param {\n    source: \"..\/..\/examples\/Alexnet_2\/Alexnet_test_leveldb\"\n    batch_size: 100\n  }\n  transform_param {\n        mean_file: \"..\/..\/examples\/Alexnet_2\/mean.binaryproto\"\n      }\n  include: { phase: TEST }\n}\nlayers {\n  name: \"conv1\"\n  type: CONVOLUTION\n  bottom: \"data\"\n  top: \"conv1\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  convolution_param {\n    num_output: 96\n    kernel_size: 11\n    stride: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayers {\n  name: \"relu1\"\n  type: RELU\n  bottom: \"conv1\"\n  top: \"conv1\"\n}\nlayers {\n  name: \"pool1\"\n  type: POOLING\n  bottom: \"conv1\"\n  top: \"pool1\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"norm1\"\n  type: LRN\n  bottom: \"pool1\"\n  top: \"norm1\"\n  lrn_param {\n    local_size: 5\n    alpha: 0.0001\n    beta: 0.75\n  }\n}\nlayers {\n  name: \"conv2\"\n  type: CONVOLUTION\n  bottom: \"norm1\"\n  top: \"conv2\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  convolution_param {\n    num_output: 256\n    pad: 2\n    kernel_size: 5\n    group: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 1\n    }\n  }\n}\nlayers {\n  name: \"relu2\"\n  type: RELU\n  bottom: \"conv2\"\n  top: \"conv2\"\n}\nlayers {\n  name: \"pool2\"\n  type: POOLING\n  bottom: \"conv2\"\n  top: \"pool2\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"norm2\"\n  type: LRN\n  bottom: \"pool2\"\n  top: \"norm2\"\n  lrn_param {\n    local_size: 5\n    alpha: 0.0001\n    beta: 0.75\n  }\n}\nlayers {\n  name: \"conv3\"\n  type: CONVOLUTION\n  bottom: \"norm2\"\n  top: \"conv3\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  convolution_param {\n    num_output: 384\n    pad: 1\n    kernel_size: 3\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayers {\n  name: \"relu3\"\n  type: RELU\n  bottom: \"conv3\"\n  top: \"conv3\"\n}\nlayers {\n  name: \"conv4\"\n  type: CONVOLUTION\n  bottom: \"conv3\"\n  top: \"conv4\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  convolution_param {\n    num_output: 384\n    pad: 1\n    kernel_size: 3\n    group: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 1\n    }\n  }\n}\nlayers {\n  name: \"relu4\"\n  type: RELU\n  bottom: \"conv4\"\n  top: \"conv4\"\n}\nlayers {\n  name: \"conv5\"\n  type: CONVOLUTION\n  bottom: \"conv4\"\n  top: \"conv5\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  convolution_param {\n    num_output: 256\n    pad: 1\n    kernel_size: 3\n    group: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 1\n    }\n  }\n}\nlayers {\n  name: \"relu5\"\n  type: RELU\n  bottom: \"conv5\"\n  top: \"conv5\"\n}\nlayers {\n  name: \"pool5\"\n  type: POOLING\n  bottom: \"conv5\"\n  top: \"pool5\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"fc6\"\n  type: INNER_PRODUCT\n  bottom: \"pool5\"\n  top: \"fc6\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  inner_product_param {\n    num_output: 4096\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.005\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 1\n    }\n  }\n}\nlayers {\n  name: \"relu6\"\n  type: RELU\n  bottom: \"fc6\"\n  top: \"fc6\"\n}\nlayers {\n  name: \"drop6\"\n  type: DROPOUT\n  bottom: \"fc6\"\n  top: \"fc6\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayers {\n  name: \"fc7\"\n  type: INNER_PRODUCT\n  bottom: \"fc6\"\n  top: \"fc7\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  inner_product_param {\n    num_output: 4096\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.005\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 1\n    }\n  }\n}\nlayers {\n  name: \"relu7\"\n  type: RELU\n  bottom: \"fc7\"\n  top: \"fc7\"\n}\nlayers {\n  name: \"drop7\"\n  type: DROPOUT\n  bottom: \"fc7\"\n  top: \"fc7\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayers {\n  name: \"fc8\"\n  type: INNER_PRODUCT\n  bottom: \"fc7\"\n  top: \"fc8\"\n  blobs_lr: 1\n  blobs_lr: 2\n  weight_decay: 1\n  weight_decay: 0\n  inner_product_param {\n    num_output: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayers {\n  name: \"accuracy\"\n  type: ACCURACY\n  bottom: \"fc8\"\n  bottom: \"label\"\n  top: \"accuracy\"\n  include: { phase: TEST }\n}\nlayers {\n  name: \"loss\"\n  type: SOFTMAX_LOSS\n  bottom: \"fc8\"\n  bottom: \"label\"\n  top: \"loss\"\n}\n\ntest.prototxt\nname: \"CaffeNet\"\nlayers \n{\n  name: \"data\"\n  type: MEMORY_DATA\n  top: \"data\"\n  top: \"label\"\n  memory_data_param \n  {\n    batch_size: 1\n    channels: 3\n    height: 80\n    width: 64\n  }\n  transform_param \n  {\n    crop_size: 64\n    mirror: false\n    mean_file: \"..\/..\/examples\/Alexnet_2\/mean.binaryproto\"\n  }\n}\nlayers {\n  name: \"conv1\"\n  type: CONVOLUTION\n  bottom: \"data\"\n  top: \"conv1\"\n  convolution_param {\n    num_output: 96\n    kernel_size: 11\n    stride: 2\n  }\n}\nlayers {\n  name: \"relu1\"\n  type: RELU\n  bottom: \"conv1\"\n  top: \"conv1\"\n}\nlayers {\n  name: \"pool1\"\n  type: POOLING\n  bottom: \"conv1\"\n  top: \"pool1\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"norm1\"\n  type: LRN\n  bottom: \"pool1\"\n  top: \"norm1\"\n  lrn_param {\n    local_size: 5\n    alpha: 0.0001\n    beta: 0.75\n  }\n}\nlayers {\n  name: \"conv2\"\n  type: CONVOLUTION\n  bottom: \"norm1\"\n  top: \"conv2\"\n  convolution_param {\n    num_output: 256\n    pad: 2\n    kernel_size: 5\n    group: 2\n  }\n}\nlayers {\n  name: \"relu2\"\n  type: RELU\n  bottom: \"conv2\"\n  top: \"conv2\"\n}\nlayers {\n  name: \"pool2\"\n  type: POOLING\n  bottom: \"conv2\"\n  top: \"pool2\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"norm2\"\n  type: LRN\n  bottom: \"pool2\"\n  top: \"norm2\"\n  lrn_param {\n    local_size: 5\n    alpha: 0.0001\n    beta: 0.75\n  }\n}\nlayers {\n  name: \"conv3\"\n  type: CONVOLUTION\n  bottom: \"norm2\"\n  top: \"conv3\"\n  convolution_param {\n    num_output: 384\n    pad: 1\n    kernel_size: 3\n  }\n}\nlayers {\n  name: \"relu3\"\n  type: RELU\n  bottom: \"conv3\"\n  top: \"conv3\"\n}\nlayers {\n  name: \"conv4\"\n  type: CONVOLUTION\n  bottom: \"conv3\"\n  top: \"conv4\"\n  convolution_param {\n    num_output: 384\n    pad: 1\n    kernel_size: 3\n    group: 2\n  }\n}\nlayers {\n  name: \"relu4\"\n  type: RELU\n  bottom: \"conv4\"\n  top: \"conv4\"\n}\nlayers {\n  name: \"conv5\"\n  type: CONVOLUTION\n  bottom: \"conv4\"\n  top: \"conv5\"\n  convolution_param {\n    num_output: 256\n    pad: 1\n    kernel_size: 3\n    group: 2\n  }\n}\nlayers {\n  name: \"relu5\"\n  type: RELU\n  bottom: \"conv5\"\n  top: \"conv5\"\n}\nlayers {\n  name: \"pool5\"\n  type: POOLING\n  bottom: \"conv5\"\n  top: \"pool5\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 3\n    stride: 2\n  }\n}\nlayers {\n  name: \"fc6\"\n  type: INNER_PRODUCT\n  bottom: \"pool5\"\n  top: \"fc6\"\n  inner_product_param {\n    num_output: 4096\n  }\n}\nlayers {\n  name: \"relu6\"\n  type: RELU\n  bottom: \"fc6\"\n  top: \"fc6\"\n}\nlayers {\n  name: \"drop6\"\n  type: DROPOUT\n  bottom: \"fc6\"\n  top: \"fc6\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayers {\n  name: \"fc7\"\n  type: INNER_PRODUCT\n  bottom: \"fc6\"\n  top: \"fc7\"\n  inner_product_param {\n    num_output: 4096\n  }\n}\nlayers {\n  name: \"relu7\"\n  type: RELU\n  bottom: \"fc7\"\n  top: \"fc7\"\n}\nlayers {\n  name: \"drop7\"\n  type: DROPOUT\n  bottom: \"fc7\"\n  top: \"fc7\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayers {\n  name: \"fc8\"\n  type: INNER_PRODUCT\n  bottom: \"fc7\"\n  top: \"fc8\"\n  inner_product_param {\n    num_output: 2\n  }\n}\nlayers {\n  name: \"prob\"\n  type: SOFTMAX\n  bottom: \"fc8\"\n  top: \"prob\"\n}\nlayers {\n  name: \"output\"\n  type: ARGMAX\n  bottom: \"prob\"\n  top: \"output\"\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":32038757,"AnswerCount":1,"Tags":"<python><windows><theano><canopy>","CreationDate":"2015-08-16T18:54:53.610","AcceptedAnswerId":null,"OwnerUserId":3656049.0,"Title":"Installing Theano with Canopy EPD on windows 7, 64 bit","Body":"<p>I have successfully installed theano on Canopy EPD, windows 7, 64 bit. While importing theano (for testing at first time), I am getting this error. Can anybody help. Thanks.<\/p>\n\n<p>It is similar to this question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/10270871\/installing-theano-on-epd-windows-x64\">Installing Theano on EPD (Windows x64)<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37845271","CreationDate":"2016-06-15T20:34:50.737","ParentId":null,"OwnerUserId":"2280645","Title":null,"Body":"<p>Since this question wasnt properly closed by the author and it's also relevant to most EPD users, including me when i first started with EPD + theano, i will answer to it.<\/p>\n\n<p>First, remove any other python environment if possible.\nBe sure there is no other conflicting environment path of python, just in case.<\/p>\n\n<p><strong>1.Install EPD CANOPY<\/strong><\/p>\n\n<p><strong>2.Install the MinGW package that is on package installer available on Canopy application<\/strong><\/p>\n\n<p>That should be easy to do, you just open the package manager from inside the Canopy application.<\/p>\n\n<p><strong>3.Install Theano from inside EPD<\/strong><\/p>\n\n<p><strong>Before<\/strong> you go spamming <\/p>\n\n<p>pip install --upgrade  --no-deps git+git:\/\/github.com\/Theano\/Theano.git<\/p>\n\n<p>in every prompt of command you see, <strong>DONT DO IT<\/strong>.<\/p>\n\n<p>--\n<strong>Instead, Open Canopy.<\/strong><\/p>\n\n<p>In Canopy screen, go to <strong>\"Tools\"<\/strong>, and then <strong>open \"Canopy Command Prompt\"<\/strong><\/p>\n\n<p>A screen exactly like the CMD from windows will open.<\/p>\n\n<p>On that screen, <strong>execute<\/strong> :<\/p>\n\n<p><strong>pip install --upgrade  --no-deps git+git:\/\/github.com\/Theano\/Theano.git<\/strong><\/p>\n\n<p>Should work fine now.<\/p>\n\n<p>--<\/p>\n\n<p>However, dont forget you will need , obviously the Windows SDK , depending on the windows version of course.\nYou can try installing the last Visual Studio with comes with everything you need.<\/p>\n\n<p>If you need more info on this, check this related Stackoverflow topic :<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/35630020\/installing-theano\/37780769#37780769\">Installing theano<\/a><\/p>\n"}]}
{"QuestionId":32039549,"AnswerCount":1,"Tags":"<protocol-buffers><caffe>","CreationDate":"2015-08-16T20:22:16.110","AcceptedAnswerId":null,"OwnerUserId":5074889.0,"Title":"protobuf common.h \"No such file\"","Body":"<p>I am trying to install Caffe, and I am running into this frustrating error. When I run  I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>I am using the Google protocol buffer 2.6.1 (<a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/downloads\" rel=\"noreferrer\">https:\/\/developers.google.com\/protocol-buffers\/docs\/downloads<\/a>), and I have indeed added the directory to the PATH. The common.h file is definitely there in the directory (I see it with my own eyes), but somehow it is unable to detect it. I have no clue what to do, and all the solutions from <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/19\" rel=\"noreferrer\">this issue<\/a> don't seem to work for me.<\/p>\n\n<p>Any insight would be appreciated. I suspect I am neglecting a step somewhere, as I am rather new to Linux.<\/p>\n\n<p>Thank you very much.<\/p>\n","answers":[{"AnswerId":"32042334","CreationDate":"2015-08-17T03:24:42.990","ParentId":null,"OwnerUserId":"2686899","Title":null,"Body":"<p><code>PATH<\/code> tells your shell where to search for commands. It does not tell your compiler where to search for headers. To tell your compiler to find headers in a particular directory, you need to use the <code>-I<\/code> flag. For example:<\/p>\n\n<pre><code>g++ -I\/path\/to\/protobuf\/include -c my-source.cc\n<\/code><\/pre>\n\n<p>You will need to convince your build system to add this flag to the compiler command-line. All reasonable build systems have some way to do this, but the details vary. For autoconf you can specify when you run configure:<\/p>\n\n<pre><code>.\/configure CXXFLAGS=-I\/path\/to\/protobuf\/include\n<\/code><\/pre>\n\n<p>For cmake I think you can do something like this (not tested):<\/p>\n\n<pre><code>cmake -DCMAKE_CXX_FLAGS=-I\/path\/to\/protobuf\/include\n<\/code><\/pre>\n\n<p>Alternatively, you would probably not have this problem if you installed protobuf to the standard location -- either <code>\/usr<\/code> or <code>\/usr\/local<\/code> (hence placing the headers in <code>\/usr\/include\/google\/protobuf<\/code> or <code>\/usr\/local\/include\/google\/protobuf<\/code>).<\/p>\n\n<p>Also note that almost all Linux distributions have a Protobuf package, and you should probably use that rather than installing Protobuf from source code. You will need the <code>-dev<\/code> or <code>-devel<\/code> package in order to get headers. On Debian\/Ubuntu:<\/p>\n\n<pre><code>sudo apt-get install libprotobuf-dev protobuf-compiler\n<\/code><\/pre>\n"}]}
{"QuestionId":32045279,"AnswerCount":1,"Tags":"<macos><gpu><theano>","CreationDate":"2015-08-17T07:55:09.117","AcceptedAnswerId":32046938.0,"OwnerUserId":2394303.0,"Title":"How can I enable my MacBook Pro GPU optimization for theano?","Body":"<ul>\n<li>MacBook Pro (Retina, 13-inch, Mid 2014)<\/li>\n<li>OS X Yosemite 10.10.3<\/li>\n<li>Intel Iris 1536MB<\/li>\n<\/ul>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#using-the-gpu\" rel=\"nofollow\">Installing Theano<\/a> shows I need CUDA, but I do not have NVIDIA, that means I can never enable GPU optimization? <\/p>\n","answers":[{"AnswerId":"32046938","CreationDate":"2015-08-17T09:28:44.983","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You are correct: an NVIDIA graphics processor is currently required to enable Theano's GPU operation. However, that does not prevent you from running Theano -- it works just fine on the CPU.<\/p>\n\n<p>Theano's current GPU implementation is based on CUDA and thus requires an NVIDIA GPU. A new implementation based on OpenCL is in development which should enable operation on non-NVIDIA GPUs, but this implementation is incomplete and not yet practically useful.<\/p>\n\n<p>The CPU implementation will work fine. In many ways it is easier to use than the GPU implementation and, if you use OpenMP, can still perform reasonably well by utilizing many CPU cores.<\/p>\n"}]}
{"QuestionId":32068712,"AnswerCount":1,"Tags":"<matplotlib><ipython-notebook><theano><lasagne>","CreationDate":"2015-08-18T09:34:34.703","AcceptedAnswerId":32076978.0,"OwnerUserId":2902280.0,"Title":"Live plot losses while training neural net in Theano\/lasagne","Body":"<p>I'm training a neural net in Theano and lasagne, running the code in an iPython notebook. I like having the train and valid loss displayed at each iteration, like this:<\/p>\n\n<pre><\/pre>\n\n<p>but I would also like to see the live\/dynamic plot of the two losses. Is there a built-in way to do so?<\/p>\n\n<p>I have tried creating a custom class and adding it to my net's , but either I get a new plot at each iteration (I'd like a single one, updating), either I have to erase the previous output at each iteration, and thus cannot see the text output (which I want to keep).<\/p>\n","answers":[{"AnswerId":"32076978","CreationDate":"2015-08-18T15:47:17.567","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>I finally managed to update the loss plot as I wanted, using the following class:<\/p>\n\n<pre><code>from IPython import display\nfrom matplotlib import pyplot as plt\nimport numpy as np\nfrom lasagne import layers\nfrom lasagne.updates import nesterov_momentum\nfrom nolearn.lasagne import NeuralNet\nfrom lasagne import nonlinearities\n\nclass PlotLosses(object):\n    def __init__(self, figsize=(8,6)):\n        plt.plot([], []) \n\n    def __call__(self, nn, train_history):\n        train_loss = np.array([i[\"train_loss\"] for i in nn.train_history_])\n        valid_loss = np.array([i[\"valid_loss\"] for i in nn.train_history_])\n\n        plt.gca().cla()\n        plt.plot(train_loss, label=\"train\") \n        plt.plot(valid_loss, label=\"test\")\n\n        plt.legend()\n        plt.draw()\n<\/code><\/pre>\n\n<p>and the code sample to reproduce:<\/p>\n\n<pre><code>net_SO = NeuralNet(\n    layers=[(layers.InputLayer, {\"name\": 'input', 'shape': (None, 1, 28, 28)}),\n            (layers.Conv2DLayer, {\"name\": 'conv1', 'filter_size': (3,3,), 'num_filters': 5}),\n            (layers.DropoutLayer, {'name': 'dropout1', 'p': 0.2}),\n            (layers.DenseLayer, {\"name\": 'hidden1', 'num_units': 50}),\n            (layers.DropoutLayer, {'name': 'dropout2', 'p': 0.2}),\n            (layers.DenseLayer, {\"name\": 'output', 'nonlinearity': nonlinearities.softmax, 'num_units': 10})],\n    # optimization method:\n    update=nesterov_momentum,\n    update_learning_rate=10**(-2),\n    update_momentum=0.9,\n\n    regression=False,  \n    max_epochs=200, \n    verbose=1,\n\n    on_epoch_finished=[PlotLosses(figsize=(8,6))], #this is the important line\n    )\n\nnet_SO.fit(X, y) #X and y from the MNIST dataset\n<\/code><\/pre>\n"}]}
{"QuestionId":32079552,"AnswerCount":1,"Tags":"<protocol-buffers><caffe>","CreationDate":"2015-08-18T18:02:27.083","AcceptedAnswerId":null,"OwnerUserId":4996964.0,"Title":"Error during \"make all\" in Caffe","Body":"<p>I had an error after running the  command while compiling Caffe. Here is what I got (it is a snippet):<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix this issue?<\/p>\n\n<p>EDIT: I'm installing it in Ubuntu 12.04.<\/p>\n","answers":[{"AnswerId":"38563673","CreationDate":"2016-07-25T09:07:08.233","ParentId":null,"OwnerUserId":"4265775","Title":null,"Body":"<p>comment out line 113 in<\/p>\n\n<pre><code>\/usr\/local\/cuda\/include\/host_config.h.\n<\/code><\/pre>\n\n<p>it is because gcc 5 error<\/p>\n"}]}
{"QuestionId":32080017,"AnswerCount":2,"Tags":"<python><numpy><attributes><caffe><pycaffe>","CreationDate":"2015-08-18T18:25:44.323","AcceptedAnswerId":32080111.0,"OwnerUserId":1245262.0,"Title":"Why does assigning an ndarray to an ndarray in PyCaffe raise an Attribute Error?","Body":"<p>When reading through a Caffe tutorial (<a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a>), I came across the following statement:<\/p>\n\n<pre><\/pre>\n\n<p>It basically serves to assign a single image to . <\/p>\n\n<p> is a 4D ndarray and returns a 3D ndarray, so the ellipsis serve to copy the 3D array over the 0th axis. This made me think I should be able to rewrite the code to avoid the ellipsis as follows:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I do, I get <\/p>\n\n<pre><\/pre>\n\n<p>Even though,<\/p>\n\n<pre><\/pre>\n\n<p>works fine. Does this make sense to anyone?<\/p>\n\n<p>I've verified the shapes and type of my variables as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Why does  cause problems?<\/p>\n","answers":[{"AnswerId":"37271679","CreationDate":"2016-05-17T09:04:34.873","ParentId":null,"OwnerUserId":"6064933","Title":null,"Body":"<p>The problem with the first method(<code>net.blobs['data'].data = z4<\/code>) is that 'data' is an attribute of net.blobs['data'](which is Caffe <strong>Blob<\/strong> object) which can not be assigned. If you assign numpy array to the data attribute, you mean \"instead of using the memory allocated for data, use the memory of the numpy array\", which is <strong>not acceptable<\/strong>.<\/p>\n\n<p>But if you use <code>net.blobs['data'].data[...] = z4<\/code>, you mean \"copy the data from the numpy array to the memory allocated for the data attribute\", which is <strong>acceptable<\/strong>.<\/p>\n\n<hr>\n\n<p>For more information, you can read <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/1TqHBsfj1Zc\" rel=\"nofollow\">a similar question<\/a> in the Caffe Users Group.<\/p>\n"},{"AnswerId":"32080111","CreationDate":"2015-08-18T18:31:11.597","ParentId":null,"OwnerUserId":"1427416","Title":null,"Body":"<p>Doing <code>obj.attr = blah<\/code> is setting an attribute on the object <code>obj<\/code>, so <code>obj<\/code> controls this.  Doing <code>obj.attr[...] = blah<\/code> is setting an <em>item<\/em> (e.g., the \"contents\" of some array-like object) on the object referred to by <code>obj.attr<\/code>, so the object <code>obj.attr<\/code> controls this.<\/p>\n\n<p>In your example, <code>net.blobs['data']<\/code> is some kind of object that won't allow its <code>data<\/code> attribute to be set, so you can't do <code>net.blobs['data'].data = blah<\/code>.  But <code>net.blobs['data'].data<\/code> is an array that <em>does<\/em> allow you to change its contents, so you can do <code>net.blobs['data'].data[...] = stuff<\/code>.  You're operating on two different objects with those two syntaxes (<code>net.blobs['data']<\/code> in one case and <code>net.blobs['data'].data<\/code> in the other).<\/p>\n"}]}
{"QuestionId":32082506,"AnswerCount":0,"Tags":"<neural-network><reinforcement-learning><torch><lstm><temporal-difference>","CreationDate":"2015-08-18T21:03:48.150","AcceptedAnswerId":null,"OwnerUserId":5240885.0,"Title":"Neural Network Reinforcement Learning Requiring Next-State Propagation For Backpropagation","Body":"<p>I am attempting to construct a neural network incorporating convolution and LSTM (using the Torch library) to be trained by Q-learning or Advantage-learning, both of which require propagating state T+1 through the network before updating the weights for state T.<\/p>\n\n<p>Having to do an extra propagation would cut performance and that's bad, but not <em>too<\/em> bad; However, the problem is that there is all kinds of state bound up in this. First of all, the Torch implementation of backpropagation has some efficiency shortcuts that rely on the back propagation happening immediately after the forward propagation, which an additional propagation would mess up. I could possibly get around this by having a secondary cloned network sharing the weight values, but we come to the second problem.<\/p>\n\n<p>Every forward propagation involving LSTMs is stateful. How can I update the weights at T+1 when propagating network(T+1) may have changed the contents of the LSTMs? I have tried to look at the discussion of TD weight updates as done in TD-Gammon, but it's obtuse to me and that's for feedforward anyway, not recurrent.<\/p>\n\n<p>How can I update the weights of a network at T without having to advance the network to T+1, or how do I advance the network to T+1 and then go back and adjust the weights as if it were still T?<\/p>\n","answers":[]}
{"QuestionId":32082696,"AnswerCount":1,"Tags":"<python><gpu><pickle><theano>","CreationDate":"2015-08-18T21:16:40.267","AcceptedAnswerId":32084018.0,"OwnerUserId":4899439.0,"Title":"Theano - NameError: name 'register_gpu_opt' is not defined","Body":"<p>I have trained a Theano model on a GPU and now want to set it up to run on a server (with no GPU).<\/p>\n\n<p>First I faced the problem that my model could not be unpickled, due to the missing  type. Then, following a recommendation from <a href=\"https:\/\/stackoverflow.com\/questions\/25237039\/converting-a-theano-model-built-on-gpu-to-cpu\">this post<\/a>, I set the option  to .<\/p>\n\n<p>But then I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>The  file was produced the following way:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32084018","CreationDate":"2015-08-18T23:05:57.640","ParentId":null,"OwnerUserId":"5226958","Title":null,"Body":"<p>Yes that is a typical problem.\nWhy does it happen? Because theano compiles the model into c\/cuda c codes, so an error will happen when the model finds  codes from inconsistent compiler.<\/p>\n\n<p>How to handler is problem? I would choose to save all the parameters as numpy value. For example<\/p>\n\n<pre><code>values_to_pickle = [p.get_value() for p in model.all_parameters()]\n<\/code><\/pre>\n"}]}
{"QuestionId":32086106,"AnswerCount":2,"Tags":"<neural-network><torch>","CreationDate":"2015-08-19T03:35:49.427","AcceptedAnswerId":32086325.0,"OwnerUserId":5118777.0,"Title":"[torch]how to read weights in nn model","Body":"<p>I constructed the nn model using itorch notebook.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Input data to the model<\/p>\n\n<pre><\/pre>\n\n<p>Then, I print the model and got this.<\/p>\n\n<pre><\/pre>\n\n<p>how to read the weight in nn.linear ?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"32086325","CreationDate":"2015-08-19T04:00:16.477","ParentId":null,"OwnerUserId":"5118777","Title":null,"Body":"<p>Oh, it is similar to php<\/p>\n\n<pre><code>model.modules[2].weight\n<\/code><\/pre>\n"},{"AnswerId":"41345372","CreationDate":"2016-12-27T12:38:06.743","ParentId":null,"OwnerUserId":"7341318","Title":null,"Body":"<p>I find that <code>model.modules[1].weight<\/code> is similar to <code>model:get(1).weight<\/code>, but both can't get the parameters from the table layer like residual block. In this way the residual block as a layer.<\/p>\n\n<p>however, we can use <code>params, gradParams = model:parameters()<\/code> to get the parameters for each layer even in the table layer. <\/p>\n\n<p>It is worth noting that, in the second way, each layer of the network parameters are divided into two layers and arranged in layers<\/p>\n"}]}
{"QuestionId":32087747,"AnswerCount":0,"Tags":"<python><c++><machine-learning><neural-network><theano>","CreationDate":"2015-08-19T06:11:28.027","AcceptedAnswerId":null,"OwnerUserId":1377127.0,"Title":"Can I use Theano with C++?","Body":"<p>I've been using the Theano library in Python and had some success, although my main program where I'm extracting sections of images to classify is written in C++. Is there any way to call Theano via C++ or is it only built for Python?<\/p>\n","answers":[]}
{"QuestionId":32096840,"AnswerCount":1,"Tags":"<cuda><gpu><nvidia><caffe>","CreationDate":"2015-08-19T13:25:42.123","AcceptedAnswerId":null,"OwnerUserId":2778860.0,"Title":"Caffe using GPU with NVidia Quadro 2200","Body":"<p>I'm using the deep learning framework <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe<\/a> on a Ubuntu 14.04 machine. I compiled CAFE with  option, i.e. I disabled GPU and CUDA usage. I have an NVidia Quadro K2200 graphics card and CUDA version 5.5. <\/p>\n\n<p>I would like to know if it is possible to use Caffe with CUDA enabled with my GPU. On NVidia page, it is written that Quadro K2200 has a compute capability of 5.0. Does it mean that I can use it with CUDA versions up to release 5.0? When it is possible to use Caffe with GPU-enabled with Quadro K2200, how can I choose the appropriate CUDA version for that?<\/p>\n","answers":[{"AnswerId":"32097095","CreationDate":"2015-08-19T13:35:26.207","ParentId":null,"OwnerUserId":"2771245","Title":null,"Body":"<p>CUDA version is not the same thing as Compute Capability. For one, CUDA  is current (7.5 prerelease), while CC is only at 5.2. K2200 supports CC 5.0.<\/p>\n\n<p>The difference:<\/p>\n\n<p>CUDA version means the library\/toolkit\/SDK\/etc version. You should always use the highest one available.<\/p>\n\n<p>Compute Capability is your GPU's capability to perform certain instructions, etc. Every CUDA function has a minimum CC requirement. When you write a CUDA program, it's CC requirement is the maximum of the requirements of all the features you used.<\/p>\n\n<hr>\n\n<p>That said, I've no idea what Caffe is, but a quick search shows they require CC of 2.0, so you should be good to go. CC 5.0 is pretty recent, so very few things won't work on it.<\/p>\n"}]}
{"QuestionId":32101249,"AnswerCount":2,"Tags":"<macos><theano><osx-yosemite>","CreationDate":"2015-08-19T16:46:13.547","AcceptedAnswerId":32110990.0,"OwnerUserId":5244344.0,"Title":"Use theano in a macbook pro without NVIDIA card","Body":"<p>I have:<\/p>\n\n<ul>\n<li>MacBook Pro (Retina, 13-inch, Mid 2014)<\/li>\n<li>OS X Yosemite<\/li>\n<li>Intel Iris 1536MB<\/li>\n<\/ul>\n\n<p>I heard that I can not use GPU theano but CPU I do, but I want to know if the programming will be the same and internaly theano will work with CPU or in anothe case with GPU. Or if when I programming I have one way to program for each one.<\/p>\n\n<p>Thanks a lot<\/p>\n","answers":[{"AnswerId":"33896455","CreationDate":"2015-11-24T14:39:13.543","ParentId":null,"OwnerUserId":"5244344","Title":null,"Body":"<p>Yes, effectively Theano understand if you have GPU or not, and decide to use CPU or GPU to create variables, the only difficult is that when you create a model and variables with Theano with or without GPU config of variables change, in other words if you create a model with GPU (or CPU) and save them in a *.pickle (e.g.) and you go to another pc without CPU (or GPU respectively) this model and their variables saved will not work.<\/p>\n"},{"AnswerId":"32110990","CreationDate":"2015-08-20T06:28:46.980","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>For the most part a Theano program that runs well on a CPU will also run well on a GPU, with no changes to the code required. There are, however, a few things to bear in mind.<\/p>\n\n<ul>\n<li><p>Not all operations have GPU versions so it's possible to create a computation that contains a component that cannot run on the GPU at all. When you run one of these computations on a GPU it will silently fall back to the CPU, so the program will run without failing and the result will be correct, but the computation is not running as efficiently as it might and will be slower because it has to copy data backwards and forwards between main and GPU memory.<\/p><\/li>\n<li><p>How you access your data can affect GPU performance quite a lot but has little impact on CPU performance. Main memory tends to be larger than GPU memory so it's often the case that your entire dataset can fit in main memory but not in GPU memory. There are techniques that can be used to avoid this problem but you need to bear in mind the GPU limitations in advance.<\/p><\/li>\n<\/ul>\n\n<p>If you stick to conventional neural network techniques, and follow the patterns used in the Theano sample\/tutorial code, then it will probably run fine on the GPU since this is the primary use-case for Theano.<\/p>\n"}]}
{"QuestionId":32111311,"AnswerCount":1,"Tags":"<caffe><conv-neural-network>","CreationDate":"2015-08-20T06:47:40.933","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Parameter fine tuning for training Alexnet with smaller image size","Body":"<p>Alexnet is intended to use 227x227x3 image size.\nIf I like to train the image size smaller like 32x80x3, what are the parameters to be fine tuned.<\/p>\n\n<p>I initially trained with 64x80x3 image sizes with all parameters same as provided except the stride in the first Conv1 layer, it was changed to 2.\nI achieved the testing accuracy very high and as high as 0.999. Then in real use also, I get reasonable high in accuracy in detection.<\/p>\n\n<p>Then I prefer to use the smaller image size 32x80x3. I used the same parameters as trained in 64x80x3 image size, but accuracy is as low as 0.9671.\nI tried to fine tune parameters like Conv1 layer's filer size to 5. Gaussian weight filter's std size to 10 times and 100 times smaller. But none of them can help achieve the accuracy achieved in training 64x80x3 images.<\/p>\n\n<p>For smaller image sizes to train, what are the parameters to be fine tuned to achieve the higher accuracy?\nI used 24000 dataset. 20000 is for training and 4000 is for testing.<\/p>\n\n<p>For both 32x80x3 and  64x80x3, I used same images, just that image size is edited to be 32x80 and 64x80.<\/p>\n","answers":[{"AnswerId":"33314545","CreationDate":"2015-10-24T04:26:21.580","ParentId":null,"OwnerUserId":"5458952","Title":null,"Body":"<p>Maybe you can train to resize the <code>32x80x3 images to 64x80x3<\/code> and then use the similar parameter settings.\nAlso, maybe you could find some thing useful here <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/cifar10\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/cifar10<\/a>.\nThere are some solver and train_val files for fine tuning over <code>CIFAR-10<\/code>, which is a dataset consists of small images too.<\/p>\n"}]}
{"QuestionId":32113915,"AnswerCount":1,"Tags":"<theano><lasagne><nolearn>","CreationDate":"2015-08-20T09:00:17.167","AcceptedAnswerId":32115199.0,"OwnerUserId":2902280.0,"Title":"Aggregate predictions with data augmentation in lasagne","Body":"<p>I am working on the MNIST dataset and using data augmentation to train a neural network. I have a BatchIterator which randomly extracts from each picture a  24, 24 subimage, and uses it as input for the NN. <\/p>\n\n<p>As far as training is concerned, everything goes fines. But for prediction, I want to extract 5 sub-images from a given image, and average the predictions, but I cannot get it to work:<\/p>\n\n<p>Here's my BatchIterator:<\/p>\n\n<pre><\/pre>\n\n<p>Fitting my net to the training data works, but when I do , I get an error because  is, I believe, called with  equal to .<\/p>\n\n<p>here's the full call stack:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea on how to fix it in the testing part of  ?<\/p>\n","answers":[{"AnswerId":"32115199","CreationDate":"2015-08-20T10:00:05.627","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Looking at the <a href=\"https:\/\/github.com\/dnouri\/nolearn\/blob\/f8e8eb7837b0d620a66a4537266d2639efcf34f6\/nolearn\/lasagne\/base.py#L63\" rel=\"nofollow\">code for <code>nolearn.lasagne.BatchIterator<\/code><\/a> and <a href=\"https:\/\/github.com\/dnouri\/nolearn\/blob\/f8e8eb7837b0d620a66a4537266d2639efcf34f6\/nolearn\/lasagne\/base.py#L520\" rel=\"nofollow\">how it is used<\/a> by the <code>nolearn.lasagne.NeuralNet<\/code> class, it looks like <code>BatchIterator<\/code>s need to work when <code>y<\/code> is not provided, i.e. in prediction mode. Note the call at <a href=\"https:\/\/github.com\/dnouri\/nolearn\/blob\/f8e8eb7837b0d620a66a4537266d2639efcf34f6\/nolearn\/lasagne\/base.py#L520\" rel=\"nofollow\">line 520<\/a> where <code>X<\/code> is provided but no value is given for <code>y<\/code> so it defaults to <code>None<\/code>.<\/p>\n\n<p>Your <code>CropIterator<\/code> currently assumes that <code>yb<\/code> is always a non-<code>None<\/code> value. I don't know if it makes sense to do anything useful when <code>yb<\/code> is not provided but I assume you could just transform <code>Xb<\/code> and return <code>None<\/code> for <code>y_new<\/code> if <code>yb<\/code> is <code>None<\/code>.<\/p>\n"}]}
{"QuestionId":32118273,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-20T12:27:29.540","AcceptedAnswerId":null,"OwnerUserId":650654.0,"Title":"Theano: Cannot allocate memory error for \"shape.eval()\" but not \"np.asarray\"","Body":"<p>When I try to print the shape of a shared variable using shape.eval(), I get a \"Cannot allocate memory\" error.  But when I convert it to a numpy array (thereby moving it to CPU), I can not only get the shape but also the array itself.  See below where it is stopped in pdb, data has been loaded into CPU\/RAM and then I am trying to print the shape of p1:<\/p>\n\n<pre><\/pre>\n\n<p>borrow is set to True for p1, but this should not matter since I am using GPU, right?<\/p>\n\n<p>There is enough free space in memory:<\/p>\n\n<pre><\/pre>\n\n<p>This error is a RAM (as opposed to GPU memory) error, right?  Why am I able to print the shape after getting the variable to CPU (using np.asarray), but shape.eval() prints an error?<\/p>\n\n<p>What's a good way to address this issue (besides, may be, getting more RAM)?<\/p>\n","answers":[{"AnswerId":"32163446","CreationDate":"2015-08-23T05:20:37.060","ParentId":null,"OwnerUserId":"650654","Title":null,"Body":"<p>Found the issue.  See <a href=\"https:\/\/stackoverflow.com\/a\/13329386\/650654\">here<\/a>.<\/p>\n\n<p><code>eval()<\/code> invokes <code>subprocess()<\/code> (for compiling) which in turn invokes <code>fork()<\/code>.<\/p>\n\n<p>This process had a very large amount of virtual memory (<code>vsize<\/code>) allocated to it.  Also, there was no swap configured on this host.  Even though the actual memory needed to execute this <code>eval()<\/code> would have been tiny, I believe the <code>fork<\/code> failed in allocating virtual memory and hence the error.<\/p>\n\n<p>Adding swap space made the issue go away.  The swap space never got used, but I believe was sufficient to satisfy the virtual memory allocation.<\/p>\n"}]}
{"QuestionId":32130304,"AnswerCount":1,"Tags":"<theano><deep-learning><mnist>","CreationDate":"2015-08-20T23:55:54.547","AcceptedAnswerId":32134434.0,"OwnerUserId":5249666.0,"Title":"How do I use a trained Theano artificial neural network on single examples?","Body":"<p>I have been following the  tutorial on how to train an ANN to classify the MNIST numbers. I am now at the \"Convolutional Neural Networks\" chapter. I want to use the trained network on single examples (MNIST images) and get the predictions. Is there a way to do that?<\/p>\n\n<p>I have looked ahead in the tutorial and on google but can't find anything.<\/p>\n\n<p>Thanks a lot in advance for any kind of help!<\/p>\n","answers":[{"AnswerId":"32134434","CreationDate":"2015-08-21T07:14:43.627","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The material in the Theano tutorial in the earlier chapters, before reaching the Convolutional Neural Networks (CNN) chapter, give a good overview of how Theano works and some of the components the CNN sample code uses. It might be reasonable to assume that students reaching this point have developed their understanding of Theano sufficiently to figure out how to modify the code to extract the model's predictions. Here's a few hints.<\/p>\n\n<p>The CNN's output layer, called <code>layer3<\/code>, is an instance of the <code>LogisticRegression<\/code> class, introduced in an earlier chapter.<\/p>\n\n<p>The <code>LogisticRegression<\/code> class has an attribute called <code>y_pred<\/code>. The comments next to the code which assigns that attribute's values says<\/p>\n\n<blockquote>\n  <p>symbolic description of how to compute prediction as class whose\n  probability is maximal<\/p>\n<\/blockquote>\n\n<p>Looking for places where <code>y_pred<\/code> is used in the logistic regression sample will highlight a function called <code>predict()<\/code>. This does for the logistic regression sample what is desired of the CNN example.<\/p>\n\n<p>If one follows the same approach, using <code>layer3.y_pred<\/code> as the output of a new Theano function, the model's predictions will become apparent.<\/p>\n"}]}
{"QuestionId":32132388,"AnswerCount":1,"Tags":"<arrays><python-2.7><numpy><slice><theano>","CreationDate":"2015-08-21T04:32:22.717","AcceptedAnswerId":32134610.0,"OwnerUserId":5106953.0,"Title":"Looping through slices of Theano tensor","Body":"<p>I have two 2D Theano tensors, call them  and , and suppose for the sake of example, both  and  have shape (1, 50). Now, to compute their mean squared error, I simply run:<\/p>\n\n<pre><\/pre>\n\n<p>However, what I wanted to do was construct a new tensor that consists of their mean squared error in chunks of 10. In other words, since I'm more familiar with NumPy, what I had in mind was to create the following tensor M in Theano:<\/p>\n\n<pre><\/pre>\n\n<p>Now, since Theano doesn't have for loops, but instead uses scan (which map is a special case of), I thought I would try the following:<\/p>\n\n<pre><\/pre>\n\n<p>However, this does not seem to work, as I get the error:<\/p>\n\n<blockquote>\n  <p>only integers, slices (), ellipsis (), numpy.newaxis () and integer or boolean arrays are valid indices<\/p>\n<\/blockquote>\n\n<p>Is there a way to loop through the slices using theano.scan (or map)? Thanks in advance, as I'm new to Theano!<\/p>\n","answers":[{"AnswerId":"32134610","CreationDate":"2015-08-21T07:24:42.720","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>Similar to what can be done in <code>numpy<\/code>, a solution would be to reshape your (1, 50) tensor to a (1, 10, 5) tensor (or even a (10, 5) tensor), and then to compute the mean along the second axis.<\/p>\n\n<p>To illustrate this with numpy, suppose I want to compute means by slices of 2<\/p>\n\n<pre><code>x = np.array([0, 2, 0, 4, 0, 6])\nx = x.reshape([3, 2])\nnp.mean(x, axis=1)\n<\/code><\/pre>\n\n<p>outputs<\/p>\n\n<pre><code>array([ 1.,  2.,  3.])\n<\/code><\/pre>\n"}]}
{"QuestionId":32141732,"AnswerCount":0,"Tags":"<matlab><regression><caffe><matcaffe>","CreationDate":"2015-08-21T13:36:03.877","AcceptedAnswerId":null,"OwnerUserId":3698878.0,"Title":"Caffe regression wrong no. of outputs in final layer","Body":"<p>I am doing regression using a fine-tuned net, on the lines of caffe flickr-style example, I have changed the num outputs in the last layer to 1, but when testing it on an image using the matlab wrapper function . It returns 10 outputs corresponding to a single image while it should return only 1. <\/p>\n\n<p>This is my final layer in <\/p>\n\n<pre><\/pre>\n\n<p>As one can see, the  is 1, the problem is when I test the finetuned net on an image using , it gives 10 output labels instead of 1. Can anyone help me understand what is happening here. Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":32142970,"AnswerCount":2,"Tags":"<machine-learning><neural-network><theano><deep-learning><shogun>","CreationDate":"2015-08-21T14:31:16.440","AcceptedAnswerId":32161588.0,"OwnerUserId":1377127.0,"Title":"How does the SHOGUN Toolbox convolutional neural network compare to Caffe and Theano?","Body":"<p>I'm interested in implementing a convolutional neural network in my C++ program where I'm tracking tagged insects (I'm also using OpenCV). I see people mention Caffe, Torch and Theano a lot but I haven't heard the CNN in the SHOGUN Toolbox discussed. Does this CNN work well and would anyone recommend it if you're working in C++? I've used Theano via scikit-neuralnetwork in Python to test out some images and that worked really well, except unfortunately Theano is Python-only.<\/p>\n","answers":[{"AnswerId":"32161588","CreationDate":"2015-08-22T23:00:39.263","ParentId":null,"OwnerUserId":"5226958","Title":null,"Body":"<p>The difference lies in the speed. cnn is computationally expensive, so a GPU implementation is at least 10 times faster than CPU. caffe and theano provide seamless  integration of calling either CPU or GPU, which may not be easy for you to implement without much GPU programming experience.<\/p>\n\n<p>Other factors may exist including a unified interface for multiplayer, stochastic gradient descent, and etc. but I think speed issue is most crucial among all these factors.<\/p>\n"},{"AnswerId":"34257516","CreationDate":"2015-12-13T23:07:24.527","ParentId":null,"OwnerUserId":"5675011","Title":null,"Body":"<p>Shogun also has GPU support of some of the operations used in the NN code. This is work in progress though. At this point in time, other libraries might be faster. We mostly built these networks in there in order to be able to easily compare them to the other algorithms in the toolbox.<\/p>\n\n<p>The advantage, however, is that you can use it from a large number of languages (while internally, C++ code is executed) -- useful if you don't want to use python.<\/p>\n\n<p>Here are some IPython notebooks that you could use as a basis to compare:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/www.shogun-toolbox.org\/static\/notebook\/current\/autoencoders.html\" rel=\"nofollow\">autoencoders for denoising and classification<\/a><\/li>\n<li><a href=\"http:\/\/www.shogun-toolbox.org\/static\/notebook\/current\/neuralnets_digits.html\" rel=\"nofollow\">(convolution) networks for digit classification<\/a><\/li>\n<\/ul>\n\n<p>We appreciate any experience to be shared. Shogun is in constant development and especially the NNs attract a lot of people to work on them, so expect things to change. If you are interested in helping GPU-fying Shogun, please let us know.<\/p>\n"}]}
{"QuestionId":32148440,"AnswerCount":1,"Tags":"<centos><makefile><protocol-buffers><caffe>","CreationDate":"2015-08-21T20:02:19.540","AcceptedAnswerId":null,"OwnerUserId":1556092.0,"Title":"how to address \"make: protoc: Command not found\"","Body":"<p>I'm installing Caffe on a CentOS system over which I do not have administrative privileges. When I attempt to compile, I encounter the following message:<\/p>\n\n<pre><\/pre>\n\n<p>What I have done so far is the following:<\/p>\n\n<pre><\/pre>\n\n<p>How should I address this error?<\/p>\n","answers":[{"AnswerId":"32150726","CreationDate":"2015-08-21T23:34:12.630","ParentId":null,"OwnerUserId":"4518274","Title":null,"Body":"<p>Since you lack administrative privileges you can either<\/p>\n\n<ul>\n<li>ask your admin to install the <a href=\"http:\/\/rpm.pbone.net\/index.php3\/stat\/4\/idpl\/23552166\/dir\/centos_6\/com\/protobuf-2.5.0-16.1.x86_64.rpm.html\" rel=\"nofollow\">protobuf<\/a> and <a href=\"http:\/\/rpm.pbone.net\/index.php3\/stat\/4\/idpl\/23552167\/dir\/centos_6\/com\/protobuf-compiler-2.5.0-16.1.x86_64.rpm.html\" rel=\"nofollow\">protobuf-compiler<\/a> packages.<\/li>\n<li>compile it yourself and install the binaries in your <code>~\/bin<\/code> directory.<\/li>\n<\/ul>\n\n<p>For the latter, this page (<em><a href=\"http:\/\/tech.yipp.ca\/linux\/install-google-protocol-buffers-linux\/\" rel=\"nofollow\">Install google protocol buffers (protoc, protobuf) on CentOS 6 (linux)<\/a><\/em>) hints that it may be as simple as using the <code>--prefix=$HOME<\/code> option on the configure script.<\/p>\n"}]}
{"QuestionId":32149975,"AnswerCount":1,"Tags":"<python><numpy><fft><theano><ifft>","CreationDate":"2015-08-21T22:10:30.477","AcceptedAnswerId":32165217.0,"OwnerUserId":5106953.0,"Title":"Inverse FFT in Theano","Body":"<p>I know that <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/fourier.py\" rel=\"nofollow\"> is essentially <\/a>. However, I was wondering if the inverse FFT was implemented? Namely, is there something like a , which is equivalent to ?<\/p>\n\n<p>I noticed that <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/fourier.py\" rel=\"nofollow\">this<\/a> has it, but I'm not sure how complete or reliable it is for doing what I want. Perhaps someone with a better understanding of Theano can weigh in here.<\/p>\n\n<p>Also, if I were to use this sandbox Fourier, how would I go about doing it? Simply calling , where  is a 1D tensor, returns the error:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to fix this?<\/p>\n","answers":[{"AnswerId":"32165217","CreationDate":"2015-08-23T09:38:46.893","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I can't comment on the robustness of the code but Theano as a whole is still in development (version 0.7) and this code is in the <code>sandbox<\/code> which should, I believe, be considered even less robust than the rest of Theano.<\/p>\n\n<p>It's clear that this FFT operation is incomplete because it is currently incapable of computing gradients (note the TODO comments). If you need gradients then sorry, this operation isn't going to help (maybe you could finish it off and submit the enhancement?)<\/p>\n\n<p>This implementation is just a shim around numpy's implementation so if numpy's implementation is sufficiently complete and reliable for doing what you want then this Theano shim probably is as well.<\/p>\n\n<p>Note that because this just wraps numpy, it won't run on the GPU and if you're mixing this operation with other GPU enabled operations and running on a GPU then you're going to have a slowdown due to data being copied backwards and forwards between main and GPU memories.<\/p>\n\n<p>To use this operation you'd do this:<\/p>\n\n<pre><code>import theano\nimport theano.sandbox.fourier as tsf\n\ntsf.ifft(frames=..., n=..., axis=...)\n<\/code><\/pre>\n"}]}
{"QuestionId":32151251,"AnswerCount":3,"Tags":"<python><theano><lasagne><nolearn>","CreationDate":"2015-08-22T01:01:15.150","AcceptedAnswerId":null,"OwnerUserId":865662.0,"Title":"How to define a cost function in nolearn, lasagne?","Body":"<p>I'm doing a neural network in nolearn, a Theano based library that uses lasagne.<\/p>\n\n<p>I'm not understanding how do I define my own cost function.<\/p>\n\n<p>The output layer is only 3 neurons  and I want it to be mostly sure when it gives 1 or 2, but otherwise - if it isn't really sure of 1, 2 - to give back simply 0.<\/p>\n\n<p>So, I came up with a cost function (will need tuning) where the cost is double for 1 and 2 than for 0, but I can't understand how to tell this to the network.<\/p>\n\n<pre><\/pre>\n\n<p>This is the code for the update, but how to I tell SGD to use my cost function instead of it's own?<\/p>\n\n<p><strong>EDIT:<\/strong>\nThe full net code is:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT<\/strong>\nError when using <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38563992","CreationDate":"2016-07-25T09:22:00.583","ParentId":null,"OwnerUserId":"6632332","Title":null,"Body":"<p>I used a custom loss function in a classification task and thought i'd share that with you too. I basically wanted different emphasis on training data depending on the label.<\/p>\n\n<pre><code>import lasagne\nimport theano.tensor as T\nimport theano\n\ndef weighted_crossentropy(predictions, targets):\n\n  weights_per_label = theano.shared(lasagne.utils.floatX([0.2, 0.4, 0.4]))\n  weights = weights_per_label[targets]  #returns a targets-shaped weight matrix\n  loss = lasagne.objectives.aggregate(T.nnet.categorical_crossentropy(predictions, targets), weights=weights)\n  return loss\n\nnet = NeuralNet(\n    # layers and parameters\n    objective_loss_function=weighted_crossentropy,\n    # ...\n    )\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/dnouri\/nolearn\/issues\/109\" rel=\"nofollow\" title=\"Weighted loss\">This<\/a> is where I found how to implement it.<\/p>\n"},{"AnswerId":"32198644","CreationDate":"2015-08-25T08:03:29.400","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>When you instantiate your neural network, you can pass a custom loss function that you've defined previously:<\/p>\n\n<pre><code>import theano.tensor as T\nimport numpy as np\nfrom nolearn.lasagne import NeuralNet\n# I'm skipping other inputs for the sake of concision\n\ndef multilabel_objective(predictions, targets):\n    epsilon = np.float32(1.0e-6)\n    one = np.float32(1.0)\n    pred = T.clip(predictions, epsilon, one - epsilon)\n    return -T.sum(targets * T.log(pred) + (one - targets) * T.log(one - pred), axis=1)\n\nnet = NeuralNet(\n    # your other parameters here (layers, update, max_epochs...)\n    # here are the one you're interested in:\n    objective_loss_function=multilabel_objective,\n    custom_score=(\"validation score\", lambda x, y: np.mean(np.abs(x - y)))\n    )\n<\/code><\/pre>\n\n<p>As you can see, it's also possible to define a custom score (using the keyword <code>custom_score<\/code>)<\/p>\n"},{"AnswerId":"32163561","CreationDate":"2015-08-23T05:40:31.227","ParentId":null,"OwnerUserId":"650654","Title":null,"Body":"<p>See the following example (taken from <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/updates.py\" rel=\"nofollow\">here<\/a>) that specifies its own loss function:<\/p>\n\n<pre><code>import lasagne\nimport theano.tensor as T\nimport theano\nfrom lasagne.nonlinearities import softmax\nfrom lasagne.layers import InputLayer, DenseLayer, get_output\nfrom lasagne.updates import sgd, apply_momentum\nl_in = InputLayer((100, 20))\nl1 = DenseLayer(l_in, num_units=3, nonlinearity=softmax)\nx = T.matrix('x')  # shp: num_batch x num_features\ny = T.ivector('y') # shp: num_batch\nl_out = get_output(l1, x)\nparams = lasagne.layers.get_all_params(l1)\nloss = T.mean(T.nnet.categorical_crossentropy(l_out, y))\nupdates_sgd = sgd(loss, params, learning_rate=0.0001)\nupdates = apply_momentum(updates_sgd, params, momentum=0.9)\ntrain_function = theano.function([x, y], updates=updates)\n<\/code><\/pre>\n\n<p>Coincidentally, this code also has three units in the output layer.<\/p>\n"}]}
{"QuestionId":32158870,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-22T17:26:45.363","AcceptedAnswerId":32165112.0,"OwnerUserId":4305160.0,"Title":"Theano shared updating last element in python","Body":"<p>I have a shared variable persistent_vis_chain which is being updated by a theano function where it gets its function from a theano.scan, But thats not the problem just back story.<\/p>\n\n<p>My shared variable looks like D = [image1, ... , imageN] where each images is [x1,x2,...,x784].<\/p>\n\n<p>What I want to do is take the average of all the images and put them into the last imageN. That is I want to sum all the values in each image except the last 1, which will result in [s1,s2,...,s784] then I want to set imageN = [s1\/len(D),s2\/len(D),...s784\/len(D)]<\/p>\n\n<p>So my problem is I do not know how to do this with theano.shared and may be with my understanding of  theano functions and doing this computation with symbolic variables. Any help would be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"32165112","CreationDate":"2015-08-23T09:27:13.570","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you have <code>N<\/code> images, each of shape <code>28x28=784<\/code> then, presumably, your shared variable has shape <code>(N,28,28)<\/code> or <code>(N,784)<\/code>? This method should work with either shape.<\/p>\n\n<p>Given <code>D<\/code> is your shared variable containing your image data. If you want to get the average image then <code>D.mean(keepdims=True)<\/code> will give it to you symbolically.<\/p>\n\n<p>It's unclear if you want to change the final image to equal the mean image (sounds like a strange thing to do), or if you want to add a further <code>N+1<\/code>'th image to the shared variable. For the former you could do something like this:<\/p>\n\n<pre><code>D = theano.shared(load_D_data())\nD_update_expression = do_something_with_scan_to_get_D_update_expression(D)\nupdates = [(D, T.concatenate(D_update_expression[:-1],\n                             D_update_expression.mean(keepdims=True)))]\nf = theano.function(..., updates=updates)\n<\/code><\/pre>\n\n<p>If you want to do the latter (add an additional image), change the <code>updates<\/code> line as follows:<\/p>\n\n<pre><code>updates = [(D, T.concatenate(D_update_expression,\n                             D_update_expression.mean(keepdims=True)))]\n<\/code><\/pre>\n\n<p>Note that this code is intended as a guide. It may not work as it stands (e.g. you may need to mess with the <code>axis=<\/code> parameter in the <code>T.concatenate<\/code> command).<\/p>\n\n<p>The point is that you need to construct a symbolic expression explaining what the new value for D looks like. You want it to be a combination of the updates from scan plus this additional average thing. <code>T.concatenate<\/code> allows you to combine those two parts together.<\/p>\n"}]}
{"QuestionId":32159129,"AnswerCount":1,"Tags":"<python><python-2.7><keras>","CreationDate":"2015-08-22T17:55:45.833","AcceptedAnswerId":null,"OwnerUserId":5106953.0,"Title":"Error importing Keras layer","Body":"<p>I am having issues importing a new layer (let's call it \"newlayer\") for the sake of simplicity, in Keras.<\/p>\n\n<p>I recently upgraded my Keras version using:<\/p>\n\n<pre><\/pre>\n\n<p>because my older install of Keras did not have newlayer. The interesting thing I notice though is when I upgrade, Keras gets installed in the .\/Python\/2.7\/site-packages directory. So when I cd to that directory, and import newlayer, it works fine.<\/p>\n\n<p>However, when I am in my home directory and I import newlayer, it does not work (I get \"ImportError: cannot import name newlayer\").<\/p>\n\n<p>Is there a reason for this? Maybe I installed Keras somehow to my home directory a while back and it is using that version? I tried searching my home directory for a Keras installation and it's not installed there at all. More importantly, is there a way to fix this instead of having to cd into .\/Python\/2.7\/site-packages each time?<\/p>\n","answers":[{"AnswerId":"35438432","CreationDate":"2016-02-16T16:55:01.083","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>Make sure that pip is setup properly for the version of python which you are using. <\/p>\n\n<p>You can do for example<\/p>\n\n<pre><code>curl -O https:\/\/bootstrap.pypa.io\/get-pip.py\npython2.7 get-pip.py\n<\/code><\/pre>\n\n<p>to re-install pip.<\/p>\n\n<p>and then:<\/p>\n\n<pre><code>pip-2.7 install --upgrade git+git:\/\/github.com\/fchollet\/keras.git\n<\/code><\/pre>\n"}]}
{"QuestionId":32169085,"AnswerCount":1,"Tags":"<matlab><image-processing><caffe><nvidia-digits>","CreationDate":"2015-08-23T16:50:59.153","AcceptedAnswerId":32218651.0,"OwnerUserId":1661607.0,"Title":"Input must have 4 axes, corresponding to (num, channels, height, width)","Body":"<p>Not sure if this is an issue or not - but I have been searching for days and cannot seem to figure it out!<\/p>\n\n<p>Any image I try to classify individually using digits seems to run okay. However, when using the \"classify many images\" button, the network crashes because of the aforementioned title\/bug\/I don't even know what the hell it is.<\/p>\n\n<p>I'm entirely new to caffe and DIGITS, and as I said I've spent days googling this problem - and cant seem to figure it out. What is the 5th dimension on the image, and if I do actually have 5D images, how do I convert them to 4D?<\/p>\n","answers":[{"AnswerId":"32218651","CreationDate":"2015-08-26T05:31:33.063","ParentId":null,"OwnerUserId":"1661607","Title":null,"Body":"<p>Turns out it was a bug - Posted on git and they flagged it officially and patched - to anyone else experiencing this problem, update your DIGITS files<\/p>\n"}]}
{"QuestionId":32171043,"AnswerCount":3,"Tags":"<python><gcc><theano>","CreationDate":"2015-08-23T20:12:35.580","AcceptedAnswerId":32204900.0,"OwnerUserId":2430739.0,"Title":"Error running theano: LONG_BIT definition appears wrong for platform (bad gcc\/glibc config?)","Body":"<p>I followed the directions on <a href=\"https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\/87880\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\/87880<\/a> (with OpenBLAS) to install Theano with Python 3.4, on 64-bit Windows 7.<\/p>\n\n<p>Theano seemed to install without error, but when I try to run a test program (or just \"import theano\" in python) I get an error the core of which seems to be:<\/p>\n\n<pre><\/pre>\n\n<p>How do I \"configure\" gcc\/glibc correctly?<\/p>\n\n<p>I looked at several other questions on this error but haven't found a solution.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/648482\/a-trivial-python-swig-error-question\">A trivial Python SWIG error question<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/26879158\/mingw-compiler-for-pip-after-cannot-find-vcvarsall-bat-error-still-does-not-w\/27220505#27220505\">MinGW compiler for pip after &quot;cannot find vcvarsall.bat&quot; error, still does not work<\/a><\/li>\n<\/ul>\n","answers":[{"AnswerId":"32204900","CreationDate":"2015-08-25T13:08:53.150","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This error message is strongly indicative of using Theano with Cygwin. The solution is to use MinGW instead. If you have both installed then make sure MinGW appears before Cygwin in the <code>PATH<\/code> environment variable.<\/p>\n"},{"AnswerId":"43254960","CreationDate":"2017-04-06T12:15:31.727","ParentId":null,"OwnerUserId":"3257826","Title":null,"Body":"<p><code>conda install theano<\/code> is all you need to do now.<\/p>\n"},{"AnswerId":"45971053","CreationDate":"2017-08-30T23:18:19.533","ParentId":null,"OwnerUserId":"5057446","Title":null,"Body":"<p>Check if gcc is installed first.If not make sure you install it.<\/p>\n\n<p>If gcc already exists abd still facing the issue, make sure you are using the right bits for Theano and Python, like 64-bit in your case.\nIn my case I installed 32-bit anaconda python on 64-bit OS which caused the issue. re-installing the right version fixed it. <\/p>\n"}]}
{"QuestionId":32171378,"AnswerCount":1,"Tags":"<c++><caffe>","CreationDate":"2015-08-23T20:47:33.480","AcceptedAnswerId":32171424.0,"OwnerUserId":2439854.0,"Title":"error while loading shared libraries: libcaffe.so","Body":"<p>I am trying to write a simple c++ app which uses caffee <\/p>\n\n<p>This is part of my makefile:<\/p>\n\n<pre><\/pre>\n\n<p>the program compiles succesfully but when I try to run the result I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>But the file is clearly at the location: ..\/Caffe\/caffe\/build\/lib which I have included. Can anyone help me out here?<\/p>\n","answers":[{"AnswerId":"32171424","CreationDate":"2015-08-23T20:53:20.737","ParentId":null,"OwnerUserId":"200291","Title":null,"Body":"<p>When you link, it includes a little note in the executable to the dynamic linker that \u201chey, I need <code>libcaffe.so<\/code>!\u201d but it doesn\u2019t say where to find it. When you run the program, you may need to give the dynamic linker some extra information, saying \u201chey, when you\u2019re looking for libraries, check here too!\u201d, and you can do by setting the <code>LD_LIBRARY_PATH<\/code> environment variable to the directory containing <code>libcaffe.so<\/code> before running your program.<\/p>\n"}]}
{"QuestionId":32171454,"AnswerCount":1,"Tags":"<python><image><caffe>","CreationDate":"2015-08-23T20:56:03.700","AcceptedAnswerId":32171560.0,"OwnerUserId":2539274.0,"Title":"Caffe, how to run classify.py for a set of images","Body":"<p>I installed Caffe on Linux successfully. Then I failed to make it work with Matlab. So I installed it with Python following the tutorial of <a href=\"http:\/\/radar.oreilly.com\/2014\/07\/how-to-build-and-run-your-first-deep-learning-network.html\" rel=\"nofollow\">Pete Warden<\/a>. However, I never used Python before I just run the command  and it works. <\/p>\n\n<p>My question is how can I test  for a set images rather than a single image? I tried to read images from test directory as following <\/p>\n\n<pre><\/pre>\n\n<p>but it returns each time<\/p>\n\n<blockquote>\n  <p>Error; Syntax inccorct <\/p>\n<\/blockquote>\n\n<p>I just work intuitively as in Matlab. Do I need to compile  before using it? Is the passage of arguments correct?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"32171560","CreationDate":"2015-08-23T21:08:55.940","ParentId":null,"OwnerUserId":"5113071","Title":null,"Body":"<p>well, this works for files in a dir<\/p>\n\n<pre><code>mypath = '.\/'\nfiles = [ f for f in listdir(mypath) if isfile(join(mypath,f)) ]\nfor f in files:\n    print join(mypath,f)\n<\/code><\/pre>\n\n<p>so perhaps you should modify yours to something like<\/p>\n\n<pre><code>import os\nfrom os.path import isfile, join\n\nmypath = '.\/example\/images\/'\nfiles = [ f for f in listdir(mypath) if isfile(join(mypath,f)) ]\nfor f in files:\n    cmd = \"python python\/classify.py --print_results %s foo\" % join(mypath,f)\n    os.system(cmd)\n<\/code><\/pre>\n"}]}
{"QuestionId":32177764,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2015-08-24T08:33:15.017","AcceptedAnswerId":32178158.0,"OwnerUserId":1714410.0,"Title":"What is `weight_decay` meta parameter in Caffe?","Body":"<p>Looking at an example <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/tutorial\/examples\/cifar10\/cifar10_full_solver.prototxt#L15\" rel=\"noreferrer\"><\/a>, posted on BVLC\/caffe git, there is a training meta parameter<\/p>\n\n<pre><\/pre>\n\n<p>What does this meta parameter mean? And what value should I assign to it?<\/p>\n","answers":[{"AnswerId":"32178158","CreationDate":"2015-08-24T08:55:19.120","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The <code>weight_decay<\/code> meta parameter govern the regularization term of the neural net.<\/p>\n\n<p>During training a regularization term is added to the network's loss to compute the backprop gradient. The <code>weight_decay<\/code> value determines how dominant this regularization term will be in the gradient computation.  <\/p>\n\n<p>As a rule of thumb, the more training examples you have, the weaker this term should be. The more parameters you have (i.e., deeper net, larger filters, larger InnerProduct layers etc.) the higher this term should be.<\/p>\n\n<p>Caffe also allows you to choose between <code>L2<\/code> regularization (default) and <code>L1<\/code> regularization, by setting<\/p>\n\n<pre><code>regularization_type: \"L1\"\n<\/code><\/pre>\n\n<p>However, since in most cases weights are small numbers (i.e., <code>-1&lt;w&lt;1<\/code>), the <code>L2<\/code> norm of the weights is significantly smaller than their <code>L1<\/code> norm. Thus, if you choose to use <code>regularization_type: \"L1\"<\/code> you might need to tune <code>weight_decay<\/code> to a significantly smaller value.<\/p>\n\n<p>While learning rate may (and usually does) change during training, the regularization weight is fixed throughout.<\/p>\n"},{"AnswerId":"32178151","CreationDate":"2015-08-24T08:54:54.623","ParentId":null,"OwnerUserId":"987599","Title":null,"Body":"<p>Weight decay is a regularization term that penalizes big weights.\nWhen the weight decay coefficient is big the penalty for big weights is also big, when it is small weights can freely grow.<\/p>\n\n<p>Look at this answer (not specific to caffe) for a better explanation:\n<a href=\"https:\/\/stats.stackexchange.com\/questions\/29130\/difference-between-neural-net-weight-decay-and-learning-rate\">Difference between neural net <em>\"weight decay\"<\/em> and <em>\"learning rate\"<\/em><\/a>.<\/p>\n"}]}
{"QuestionId":32179981,"AnswerCount":1,"Tags":"<predict><deep-learning><mse><keras>","CreationDate":"2015-08-24T10:24:58.630","AcceptedAnswerId":null,"OwnerUserId":5142261.0,"Title":"the loss of mse always be 0 when keras for topic predict","Body":"<p>my input is a 200 dims vector, which is generated by mean of the word2vector of all words of a article,\nmy output is a 50 dims vector,which is generated by the LDA results of a article\nI want to use mse as the loss function,but the value of the loss always be 0\nmy code as follows:<\/p>\n\n<pre><\/pre>\n\n<p>the screen output as follows:\n<a href=\"https:\/\/i.stack.imgur.com\/euVdr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/euVdr.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>who can tell me why ,thanks!<\/p>\n","answers":[{"AnswerId":"32272087","CreationDate":"2015-08-28T13:21:38.377","ParentId":null,"OwnerUserId":"3403018","Title":null,"Body":"<p>First, is your output a one-hot vector of predicted classes? IE: class one is [1, 0, 0, ...] and class two is [0, 1, 0, 0, ...].<\/p>\n\n<p>If so, then using softmax activation at the output layer is acceptable and you are doing a classification problem. If you are doing a classification problem (one-hot output) you cannot use MSE as the loss, use categorical cross-entropy. <\/p>\n\n<p>Softmax scales the output so that the number given is a predicted probability of a certain class. Wikipedia here: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Softmax_function\" rel=\"noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Softmax_function<\/a><\/p>\n\n<p>If you are expecting the output vector to be real numbers then you need to use linear activation on your output neurons.<\/p>\n"}]}
{"QuestionId":32185649,"AnswerCount":1,"Tags":"<python><executable><theano>","CreationDate":"2015-08-24T15:06:13.623","AcceptedAnswerId":32334133.0,"OwnerUserId":4305160.0,"Title":"How do I compile a theano network into an self-contained executable?","Body":"<p>We would like to enter a competition where we would need to submit our network in the form of an executable. In one of the subcompetitions, the network will be trained by the judges, e.g.<\/p>\n\n<p>CASE 1:<\/p>\n\n<pre><\/pre>\n\n<p>CASE 2:\nand in another, we would submit a trained network that will just receive an input file, e.g.<\/p>\n\n<pre><\/pre>\n\n<p>How do I compile our theano-built network (whether it's an RBM or other variation) into something that can be packaged in this way without requiring the people on the other end to actually install theano?<\/p>\n","answers":[{"AnswerId":"32334133","CreationDate":"2015-09-01T14:23:35.270","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Can the judge install a VM or a Docker image with Theano and your stuff installed? This would be easier in the short term.<\/p>\n\n<p>We don't have exactly what you ask, but there is work to generate a shared library for a Theano function. This is not yet merged in Theano itself. I started this and Alberto Orlandi is continuing this. He have a working version on a branch. See this branch:<\/p>\n\n<p><a href=\"https:\/\/github.com\/AlOa\/Theano\/commits\/embed_theano\" rel=\"nofollow\">https:\/\/github.com\/AlOa\/Theano\/commits\/embed_theano<\/a><\/p>\n\n<p>Also this mailing list discussion:<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/d\/topic\/theano-users\/JtxIequsHNg\/discussion\" rel=\"nofollow\">https:\/\/groups.google.com\/d\/topic\/theano-users\/JtxIequsHNg\/discussion<\/a><\/p>\n\n<p>He is working to make a PR for this. But the current version do not do exactly what you want and need some work from you for the shared variables and to make this work.<\/p>\n\n<p>What this branch do is generate a C++ shared library that depend on Python and NumPy, but not Theano.<\/p>\n"}]}
{"QuestionId":32188392,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-08-24T17:41:09.487","AcceptedAnswerId":32231558.0,"OwnerUserId":5260975.0,"Title":"Expecting a contiguous tensor error with nn.Sum","Body":"<p>I have a 2x16x3x10x10 tensor that I feed into my network. My network has two parts that work in parallel. The first part takes the 16x3x10x10 matrix and computes the sum over the last two dimensions, returning a 16x3 tensor.\nThe second part is a convolutional neural network that produces a 16x160 tensor.\nWhenever I try to run this model, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the relevant part of the model:<\/p>\n\n<pre><\/pre>\n\n<p>The code works when the input tensor is 2x1x3x10x10, but not when the tensor is 2x16x3x10x10.<\/p>\n\n<p>Edit: I only just realized that this happens when I do model:backward and not model:forward. Here is the relevant code:<\/p>\n\n<pre><\/pre>\n\n<p>x is a 2x16x3x10x10 tensor and dE_dy is 16x2.<\/p>\n","answers":[{"AnswerId":"32231558","CreationDate":"2015-08-26T16:07:35.697","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>This is a flaw in <code>torch.nn<\/code> library. To perform a backward step, <code>nn.Parallel<\/code> splits <code>gradOutput<\/code> it receives from higher module into pieces and sends them to its parallel submodules. Splitting are done effectively without copying memory, and thus those pieces are <a href=\"https:\/\/stackoverflow.com\/a\/4059454\">non-contiguous<\/a> (unless you split on the 1st dimension).<\/p>\n\n<pre><code>local first_part = nn.Parallel(1,2)\n--                               ^\n--                 Merging on the 2nd dimension; \n--       Chunks of splitted gradOutput will not be contiguous\n<\/code><\/pre>\n\n<p>The problem is that <code>nn.Sum<\/code> cannot work with non-contiguous <code>gradOutput<\/code>. I haven't got a better idea than to make changes to it: <\/p>\n\n<pre><code>Sum_nc, _ = torch.class('nn.Sum_nc', 'nn.Sum')\nfunction Sum_nc:updateGradInput(input, gradOutput)\n    local size = input:size()\n    size[self.dimension] = 1\n    -- modified code:\n    if gradOutput:isContiguous() then\n        gradOutput = gradOutput:view(size) -- doesn't work with non-contiguous tensors\n    else\n        gradOutput = gradOutput:resize(size) -- slower because of memory reallocation and changes gradOutput\n        -- gradOutput = gradOutput:clone():resize(size) -- doesn't change gradOutput; safer and even slower\n    end\n    --\n    self.gradInput:resizeAs(input)\n    self.gradInput:copy(gradOutput:expandAs(input))\n    return self.gradInput\nend \n\n[...]\n\nsums = nn.Sequential()\nsums:add(nn.Sum_nc(3)) -- &lt;- will use torch.view\nsums:add(nn.Sum_nc(3)) -- &lt;- will use torch.resize\n<\/code><\/pre>\n"}]}
{"QuestionId":32202249,"AnswerCount":0,"Tags":"<python><numpy><theano>","CreationDate":"2015-08-25T10:59:46.213","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"Different results of cost function in theano and numpy","Body":"<p>I'm getting different results when calculating a negative log likelihood of a simple two layer neural net in theano and numpy.<\/p>\n\n<p>This is the numpy code:<\/p>\n\n<pre><\/pre>\n\n<p>where model is a function that generates initial parameters and X is the input array.<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting a result of 1.3819194609246772, which is the correct value for the loss function. However my Theano code yields a value of 1.3715655944645178.<\/p>\n\n<pre><\/pre>\n\n<p>I'm already getting wrong results when calculating the values in the output layer. I'm not really sure what the problem could be. Does the shared function keep the type of numpy array that is used as the value argument or is that converted to float32? Can anybody tell me what I'm doing wrong in the theano code?<\/p>\n\n<p>EDIT: The problem seems to occur in the hidden layer after applying the ReLU function: Here's the comparison in the results between theano and numpy in each layer:<\/p>\n\n<pre><\/pre>\n\n<p>I have the idea of using the switch() function for the ReLU layer from this post: <a href=\"https:\/\/stackoverflow.com\/questions\/26497564\/theano-hiddenlayer-activation-function\">Theano HiddenLayer Activation Function<\/a> and I don't really see how that function is different from the equivalent numpy code: z_3 = np.maximum(0, z_2)?! \nSolution to first problem: T.switch(first_layer > 0,0,first_layer) sets all the values greater than 0 to 0 => it should be T.switch(first_layer &lt; 0,0,first_layer).<\/p>\n\n<p>EDIT2: The gradients that theano calculates significantly differ from the numerical gradients I was given, this is my implementation:<\/p>\n\n<pre><\/pre>\n\n<p>This is an assignment for the Convolutional Neural Networks class that was offered by Stanford earlier this year and I think it's safe to say that their numerical gradients are probably correct. I could post the code to their numerical implementation though if required.<\/p>\n\n<p>Using a relative error the following way:<\/p>\n\n<pre><\/pre>\n\n<p>Calculating the numerical gradients using the eval_numerical_gradient method that was provided by the course get the following relative errors for the gradients:<\/p>\n\n<pre><\/pre>\n\n<p>Which are too large for W1 and W2, the relative error should be less than 1e-8. Can anybody explain this or help in any way?<\/p>\n","answers":[]}
{"QuestionId":32203687,"AnswerCount":1,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-08-25T12:11:20.253","AcceptedAnswerId":null,"OwnerUserId":1377127.0,"Title":"What does the \"TEST\" variable do in Caffe example code?","Body":"<p>In the C++ <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\" rel=\"nofollow\">example provided by Caffe<\/a> they use the variable \"TEST\" on line 56. I haven't seen it referenced before or after so I was wondering if anyone knew what the variable does?<\/p>\n","answers":[{"AnswerId":"32203950","CreationDate":"2015-08-25T12:23:44.617","ParentId":null,"OwnerUserId":"5089383","Title":null,"Body":"<p>I didn't download the whole project to find the definition of that constant (not variable) but the comments say there are two constants TEST and TRAIN indicating the \"phase\".<\/p>\n\n<p>They must be defined (and maybe commented further) in one of the many Caffe hpp files.<\/p>\n"}]}
{"QuestionId":32204866,"AnswerCount":1,"Tags":"<c++><opencv><machine-learning><deep-learning><caffe>","CreationDate":"2015-08-25T13:06:53.727","AcceptedAnswerId":32206293.0,"OwnerUserId":1377127.0,"Title":"Converting OpenCV grayscale Mat to Caffe blob","Body":"<p>I've been <a href=\"https:\/\/github.com\/jyegerlehner\/caffe\/blob\/nat_img_autoenc_parallel\/tools\/prop-patches.cpp\" rel=\"nofollow\">following an example<\/a> I was referred to on how to convert an OpenCV  into a Caffe object I could make predictions from. From what I understand, the first section scales the image and then initialises the caffe class :<\/p>\n\n<pre><\/pre>\n\n<p>Then, the OpenCV Mat \"patch\" is converted into \"input_blob\". I've changed this part because I've loaded in my image in grayscale instead of colour. <\/p>\n\n<pre><\/pre>\n\n<p>Finally, I'm not too sure what this section does - if I already have my OpenCV Mat converted to a Caffe blob, why do I need to push back on the \"input\" vector and pass it to the net? Can't I pass input_blob directly into the net to get my prediction back?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32206293","CreationDate":"2015-08-25T14:08:52.813","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to <code>push_back<\/code> your <code>input_blob<\/code> in order to pass it to <code>net<\/code> since <code>net<\/code> is expecting its input as a <em><code>std::vector<\/code><\/em> of <code>Blobs<\/code> (in principle, there may be <code>net<\/code>s that need more than a single input blob to produce output).<br>\nNote that you are not copying <code>input_blob<\/code> into the input vector, but rather passing a pointer to it.<\/p>\n"}]}
{"QuestionId":32208193,"AnswerCount":1,"Tags":"<opencv><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-08-25T15:33:43.590","AcceptedAnswerId":32211796.0,"OwnerUserId":1377127.0,"Title":"How to modify Caffe network input for C++ API?","Body":"<p>I'm trying to use the MINST Caffe example via the C++ API, but I'm having a bit of trouble working out how to restructure the network prototxt file I'll deploy after training. I've trained and tested the model with the original file (<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">lenet_train_test.prototxt<\/a>), but when I want to deploy it and make predictions like in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\" rel=\"nofollow\">C++ and OpenCV example<\/a>, I realise I have to modify the input section to make it similar to the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt\" rel=\"nofollow\">deploy.prototxt<\/a> file they have.<\/p>\n\n<p>Can I replace the information in the training and testing layers of the lenet_train_test.prototxt with this section of the deploy.prototxt file?<\/p>\n\n<pre><\/pre>\n\n<p>The images I'll be passing for classification to the network will be grayscale and 24*24 pixels, and I'll also want to scale it like was done with the MINST dataset, so could I modify the section to this?<\/p>\n\n<pre><\/pre>\n\n<p>I'm not entirely sure what the \"\" is coming from though.<\/p>\n","answers":[{"AnswerId":"32211796","CreationDate":"2015-08-25T18:49:25.470","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In order to \"convert\" you <em>train_val<\/em> prototxt to a <em>deploy<\/em> one you remove the input data layers (reading your train\/val data) and replacing them with the declaration<\/p>\n\n<pre><code>name: \"CaffeNet\"\ninput: \"data\"\ninput_shape {\n  dim: 10\n  dim: 1\n  dim: 24\n  dim: 24\n}\n<\/code><\/pre>\n\n<p>Note that the <em>deploy<\/em> prototxt does not have two phases for train and test only a single flavor.<br>\nReplacing the input data  layer with this declaration basically tells caffe that you are responsible of supplying the data, and the net should allocate space for inputs of this size.<\/p>\n\n<p>Regarding scale: once you <em>deploy<\/em> your net, the net has no control over the inputs - it does not read the data for you as the input data layers in the <em>train_val<\/em> net. Therefore, you'll have to scale the input data yourself <strong>before<\/strong> feeding it to the network. You can use the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/data_transformer.hpp\" rel=\"nofollow\">DataTransformer<\/a> class to help you transform your input blobs in the same way they were transformed during training.<\/p>\n\n<p>Regarding the first <code>dim: 10<\/code>: every Blob (i.e., data\/parameters storage unit) in caffe has 4 dimensions: batch-size, channels, height and width. This parameter actually means the net should allocate space for batches of 10 inputs at a time.<br>\nThe \"magic\" number 10 comes from the way googlenet and other competitors in ILSVRC challenge used to classify images: they classified 10 crops from each image and averaged the outputs to produce better classification results.<\/p>\n"}]}
{"QuestionId":32208646,"AnswerCount":1,"Tags":"<python-2.7><python-import><theano><lasagne><nolearn>","CreationDate":"2015-08-25T15:56:31.980","AcceptedAnswerId":null,"OwnerUserId":4801125.0,"Title":"nolearn 0.5 are not compatible with lasagne 0.1 or 0.2?","Body":"<p>When I want to import:<\/p>\n\n<pre><\/pre>\n\n<p>I always got this error <\/p>\n\n<pre><\/pre>\n\n<p>My Theano version is 0.7.0.<\/p>\n","answers":[{"AnswerId":"32209459","CreationDate":"2015-08-25T16:37:00.723","ParentId":null,"OwnerUserId":"4801125","Title":null,"Body":"<p>I uninstalled nolearn and lasagne:<\/p>\n\n<pre><code>pip uninstall nolearn\npip uninstall lasagne\n<\/code><\/pre>\n\n<p>And then run following in command line:<\/p>\n\n<pre><code>pip install -r https:\/\/raw.githubusercontent.com\/dnouri\/nolearn\/master\/requirements.txt https:\/\/github.com\/dnouri\/nolearn\/archive\/master.zip#egg=nolearn\n<\/code><\/pre>\n\n<p>Now it works. <\/p>\n"}]}
{"QuestionId":32211530,"AnswerCount":2,"Tags":"<python-2.7><theano><nolearn>","CreationDate":"2015-08-25T18:33:46.427","AcceptedAnswerId":null,"OwnerUserId":4801125.0,"Title":"Theano TensorType error","Body":"<p>When I am using nolearn to implement multi-label classification, I got this error:<\/p>\n\n<blockquote>\n  <p>'Bad input argument to theano function with name \"\/Users\/lm\/Documents\/anaconda\/lib\/python2.7\/site-packages\/nolearn\/lasagne\/base.p\u200c\u200by:391\"\n  at index 1(0-based)', 'TensorType(float32, matrix) cannot store a\n  value of dtype int64 without risking loss of precision. If you do not\n  mind this loss, you can: 1) explicitly cast your data to float32, or\n  2) set \"allow_input_downcast=True\" when calling \"function\".',\n  array([[0, 0, 0, ..., 0, 0, 1],<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"44017652","CreationDate":"2017-05-17T06:55:38.230","ParentId":null,"OwnerUserId":"5695374","Title":null,"Body":"<p>In my case all I did was change the <code>floatX<\/code> flag (under <code>[global]<\/code>) to on the .theanorc file from : <\/p>\n\n<pre><code>[global]\nfloatX = float64\n<\/code><\/pre>\n\n<p>to: <\/p>\n\n<pre><code>[global]\nfloatX = float32\n<\/code><\/pre>\n\n<p><em>Notice that the 64 at the end was replaced by the 32.<\/em><\/p>\n"},{"AnswerId":"32211597","CreationDate":"2015-08-25T18:37:45.730","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>As told in the error message, you need to convert your input and output to the appropriate type (if you do not fear losing precision).<\/p>\n\n<pre><code>input = input.astype(np.float32)\noutput = output.astype(np.float32)\n<\/code><\/pre>\n\n<p>should work<\/p>\n\n<p>Note: even if you do this, the error might remain if you have a <code>BatchIterator<\/code> which transforms your data (and  by inadvertance uses <code>float64<\/code> again). The solution is the same: inside the <code>BatchIterator<\/code>, cast the data to <code>float32<\/code> right before returning it.<\/p>\n"}]}
{"QuestionId":32216638,"AnswerCount":1,"Tags":"<opencv><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-08-26T01:33:29.647","AcceptedAnswerId":32217915.0,"OwnerUserId":1377127.0,"Title":"Scaling OpenCV Mat for Caffe","Body":"<p>I've been following the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">Caffe MINST example<\/a> and trying to deploy a test of the trained model with C++ where I use OpenCV to read in the images. In the example, they mention how for the training and test images they<\/p>\n\n<blockquote>\n  <p>scale the incoming pixels so that they are in the range [0,1). Why\n  0.00390625? It is 1 divided by 256.<\/p>\n<\/blockquote>\n\n<p>I've heard how there's a DataTransformer class in Caffe you can use to scale your images, but if I multiplied each pixel in the OpenCV Mat object by 0.00390625 would this give the same result?<\/p>\n","answers":[{"AnswerId":"32217915","CreationDate":"2015-08-26T04:17:23.577","ParentId":null,"OwnerUserId":"2589776","Title":null,"Body":"<p>The idea is right. But remember to convert your OpenCV Mats to float or double type before scaling.<\/p>\n\n<p>Something like:<\/p>\n\n<pre><code>cv::Mat mat; \/\/ assume this is one of your images (grayscale)\n\n\/* convert it to float *\/\nmat.convertTo(mat, CV_32FC1); \/\/ use CV_32FC3 for color images\n\n\/* scaling here *\/\nmat = mat * 0.00390625;\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Update #1<\/strong>: Converting and scaling can also simply be done in one line, i.e. <\/p>\n\n<pre><code>cv::Mat mat; \/\/ assume this is one of your images (grayscale)\n\n\/* convert and scale here *\/\nmat.convertTo(mat, CV_32FC1, 0.00390625);\n<\/code><\/pre>\n"}]}
{"QuestionId":32218132,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-08-26T04:41:12.587","AcceptedAnswerId":32224804.0,"OwnerUserId":1401278.0,"Title":"theano finding the indices of a tensor elements in a second tensor","Body":"<p>I can't seem to find a solution for this. Given two theano tensors a and b, I want to find the indices of elements in b within the tensor a. This example will help, say a = [1, 5, 10, 17, 23, 39] and b = [1, 10, 39], I want the result to be the indices of the b values in tensor a, i.e. [0, 2, 5].<\/p>\n\n<p>After spending some time, I thought the best way would be to use scan; here is my shot at the minimal example.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>in the return statement line. Further, the output needs to be a single tensor of shape b and I am not sure if this would get the desired result. Any suggestion would be helpful.<\/p>\n","answers":[{"AnswerId":"32224804","CreationDate":"2015-08-26T10:58:04.203","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>It all depends on how big your arrays will be. As long as it fits in memory you can proceed as follows<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\naa = T.ivector()\nbb = T.ivector()\n\nequality = T.eq(aa, bb[:, np.newaxis])\nindices = equality.nonzero()[1]\n\nf = theano.function([aa, bb], indices)\n\na = np.array([1, 5, 10, 17, 23, 39], dtype=np.int32)\nb = np.array([1, 10, 39], dtype=np.int32)\n\nf(a, b)\n\n# outputs [0, 2, 5]\n<\/code><\/pre>\n"}]}
{"QuestionId":32218466,"AnswerCount":4,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-08-26T05:15:23.917","AcceptedAnswerId":32218586.0,"OwnerUserId":1377127.0,"Title":"Compiling Caffe C++ Classification Example","Body":"<p>I recently modified the Caffe <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\" rel=\"noreferrer\">C++ classification example file<\/a> and I am trying to recompile it. However, I'm having trouble linking a simple g++ compilation to the .hpp files in the include directory. I know this is a basic question but I can't seem to work it out - can someone help me work out how to compile this program? The compilation looks like this now:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm running this on a machine without Nvidia GPUs so when I looked at the device_alternate.hpp file I realised this is calling a lot of cuda-related .hpp files as well which don't exist.<\/p>\n","answers":[{"AnswerId":"32297533","CreationDate":"2015-08-30T14:23:20.910","ParentId":null,"OwnerUserId":"5282027","Title":null,"Body":"<p>Change Makefile.config to set <code>CPU_ONLY := 1<\/code><\/p>\n\n<pre><code># CPU-only switch (uncomment to build without GPU support).\nCPU_ONLY := 1\n<\/code><\/pre>\n"},{"AnswerId":"32218586","CreationDate":"2015-08-26T05:25:55.033","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Usually, in order to help the compiler locate header files you need to add <a href=\"http:\/\/www.rapidtables.com\/code\/linux\/gcc\/gcc-i.htm\" rel=\"noreferrer\"><code>-I \/path\/to\/include\/folder<\/code><\/a> option to the compilation line:<\/p>\n\n<pre><code>~$ g++ -I \/path\/to\/caffe\/include myfile.cpp\n<\/code><\/pre>\n"},{"AnswerId":"32295890","CreationDate":"2015-08-30T11:08:25.693","ParentId":null,"OwnerUserId":"2430986","Title":null,"Body":"<p>If you want to build custom files in caffe, there are two ways<\/p>\n\n<p><strong>The easy way<\/strong> <\/p>\n\n<ul>\n<li>Make the necessary changes and keep the file ( in your case - classification.cpp ) inside a directory ( say test ) in examples folder in th\ne caffe root directory.<\/li>\n<li>run <code>make<\/code>. This will automatically add the necessary cxxflags and ldflags and compile your code and place the executable in the build\/examples\/test folder. This also ensure the flag <strong>CPU_ONLY<\/strong> is set ( as mentioned in the <em>Makefile.config<\/em> )<\/li>\n<\/ul>\n\n<p><strong>The Hard way<\/strong><\/p>\n\n<ul>\n<li>Run make without the pretty print option ( mentioned inside <em>Makefile.config<\/em> ). You will be able to see the compile and link options used to build the examples and tools. You can copy and paste these options ( and make necessary changes to relative paths if used ) to compile your file<\/li>\n<\/ul>\n\n<p>Hope this helps<\/p>\n\n<p><strong>Edit<\/strong>\nAs the op requested an easy way, it can be done as follows<\/p>\n\n<p>This is a very <strong>minimal example<\/strong> and I encourage the OP to refer to full online documentation and example of cmake usage.<\/p>\n\n<ul>\n<li>Requirements\n\n<ul>\n<li>Caffe needs to be built with <strong>cmake<\/strong> - Relatively easy as the current master branch has CMakeLists and everything defined. Use the Cmake-gui or ccmake to set your options<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Now, I am assuming you have a project structure as follows.<\/p>\n\n<pre><code>-project  \n    - src  \n         - class1.cpp\n         - CMakeLists.txt ( to be added )\n    - include\n         - class1.hpp\n\n    - main.cpp\n    - CMakeLists.txt ( to be added )\n<\/code><\/pre>\n\n<p>The CMakeLists.txt ( src ) needs to contain (<em>at minimum<\/em>) the following lines,<\/p>\n\n<pre><code>cmake_minimum_required(VERSION 2.8)\nfind_package(OpenCV REQUIRED) # Optional in case of dependency on opencv \nadd_library( c1 class1.cpp )\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> In case class1 depends on other external libraries, the path to headers must be included using <code>include_directories<\/code>.<\/p>\n\n<p>The CMakeLists.txt ( outermost ) needs to contain the following at <em>minimum<\/em><\/p>\n\n<pre><code>cmake_minimum_required(VERSION 2.8)\nPROJECT(MyProject)\n\nfind_package(OpenCV REQUIRED)\nfind_package(Caffe REQUIRED)\n\ninclude_directories( \"${PROJECT_SOURCE_DIR}\/include\" )\nadd_subdirectory( src )\n\ninclude_directories( \"$Caffe_INCLUDE_DIRS}\" )\nadd_executable(MyProject main.cpp)\n\ntarget_link_libraries( MyProject ${OpenCV_LIBS} c1 ${Caffe_LIBRARIES} )    \n<\/code><\/pre>\n\n<p>Now, the following commands from inside the project directory will create the executable <code>MyProject<\/code> inside the <code>build<\/code> folder.<\/p>\n\n<pre><code>mkdir build\ncd build\ncmake ..\nmake\n<\/code><\/pre>\n\n<p>You can then run your program with <code>.\/MyProject (arguments)<\/code><\/p>\n\n<p><strong>EDIT 2<\/strong><\/p>\n\n<p>Satisfying the requirement of building caffe with CMake is very important for this to work. You need to configure and generate the Makefiles using CMake. Use <code>cmake-gui<\/code> or <code>ccmake<\/code> for this purpose so that you can set your options like CPU_ONLY,  etc.<\/p>\n\n<p>You should create a build directory inside caffe and execute the following for a basic setup<\/p>\n\n<pre><code>mkdir build\ncd build\ncmake ..\nmake -jX #X is the number of threads your CPU can handle\n<\/code><\/pre>\n\n<p>Now, the .cmake directory in your $HOME folder consists of the following\n<code>\/home\/user\/.cmake\/packages\/Caffe\/&lt;random_string&gt;<\/code> file. This file points to the install location of caffe ( which is our build directory )<\/p>\n\n<p>Now, the find_package command should run without errors for your other projects. And since you are using CMake you can keep your project folder outside the Caffe folder ( and it is better to keep it outside as the make process of caffe will try to build your files but it will fail )<\/p>\n\n<p><strong>Note<\/strong>: In case that the error persists, you can manually set the <em>Caffe_DIR<\/em> during  cmake configuration.<\/p>\n"},{"AnswerId":"54447936","CreationDate":"2019-01-30T19:13:43.140","ParentId":null,"OwnerUserId":"6327658","Title":null,"Body":"<p>In <code>.\/caffe\/include\/caffe\/util\/device_alternate.hpp<\/code> you have<\/p>\n\n<pre><code>...\n\n#ifdef CPU_ONLY  \/\/ CPU-only Caffe.\n\n...\n\n#else  \/\/ Normal GPU + CPU Caffe.\n\n#include &lt;cublas_v2.h&gt;\n\n...\n<\/code><\/pre>\n\n<p>This means it will try to include <code>cublas_v2.h<\/code> unless <code>CPU_ONLY<\/code> flag is defined.<\/p>\n\n<p>Since your machine doesn't have Nvidia GPU, you have to define <code>CPU_ONLY<\/code> flag during compilation with <code>-DCPU_ONLY=1<\/code><\/p>\n\n<p>So your full compilation command should look like<\/p>\n\n<pre><code>g++ -DCPU_ONLY=1 -I \/home\/jack\/caffe\/include classification.cpp -o classify\n<\/code><\/pre>\n"}]}
{"QuestionId":32218712,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><neural-network><keras>","CreationDate":"2015-08-26T05:37:10.530","AcceptedAnswerId":32389326.0,"OwnerUserId":2605604.0,"Title":"Neural network dimension mis-match","Body":"<p>I have a neural network setup for the MNIST digits dataset in Keras that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>features_train is of shape (1000,784), labels_train is (1000,1), and both are numpy arrays. I want 784 input nodes, 200 hidden, and 9 output to classify the digits<\/p>\n\n<p>I keep getting an input dimension mismatch error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to identify where my dimensions may be incorrect but I'm not seeing it. Can anyone see the problem?<\/p>\n","answers":[{"AnswerId":"32389326","CreationDate":"2015-09-04T02:50:28.077","ParentId":null,"OwnerUserId":"2605604","Title":null,"Body":"<p>I've been training 2 class classification models for so long that I'm used to dealing with labels that are just single values. For this problem (classifying more than 1 outcome) I just had to change the labels to be vectors themselves.<\/p>\n\n<p>This solved my problem:<\/p>\n\n<pre><code>from keras.utils.np_utils import to_categorical\n\nlabels_train = to_categorical(labels_train)\n<\/code><\/pre>\n"}]}
{"QuestionId":32225388,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><lstm><recurrent-neural-network>","CreationDate":"2015-08-26T11:27:22.137","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"LSTM module for Caffe","Body":"<p>Does anyone know if there exists a nice LSTM module for Caffe? I found one from a github account by russel91 but apparantly the webpage containing examples and explanations disappeared (Formerly <a href=\"http:\/\/apollo.deepmatter.io\/\" rel=\"noreferrer\">http:\/\/apollo.deepmatter.io\/<\/a> --> it now redirects only to the <a href=\"https:\/\/github.com\/russell91\/apollocaffe\" rel=\"noreferrer\">github page<\/a> which has no examples or explanations anymore).<\/p>\n","answers":[{"AnswerId":"32440448","CreationDate":"2015-09-07T13:58:32.040","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I know <a href=\"http:\/\/jeffdonahue.com\/\" rel=\"noreferrer\">Jeff Donahue<\/a> worked on LSTM models using Caffe. He also gave a nice <a href=\"http:\/\/tutorial.caffe.berkeleyvision.org\/caffe-cvpr15-sequences.pdf\" rel=\"noreferrer\">tutorial<\/a> during CVPR 2015. He has a <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2033\" rel=\"noreferrer\">pull-request<\/a> with RNN and LSTM. <\/p>\n\n<p><strong>Update<\/strong>: there is a <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3948\" rel=\"noreferrer\">new PR<\/a> by Jeff Donahue including RNN and LSTM. This PR was merged on June 2016 to master.<\/p>\n"},{"AnswerId":"35967589","CreationDate":"2016-03-13T07:10:34.850","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In fact, training recurrent nets is often done by unrolling the net. That is, replicating the net over the temporal steps (sharing weights across the temporal steps) and simply doing forward-backward passes on the unrolled model.<\/p>\n\n<p>To unroll LSTM (or any other unit) you don't have to use <a href=\"http:\/\/jeffdonahue.com\/\" rel=\"noreferrer\">Jeff Donahue<\/a>'s recurrent branch, but rather use <code>NetSpec()<\/code> to explicitly unroll the model.<\/p>\n\n<p>Here's a simple example:<\/p>\n\n<pre><code>from caffe import layers as L, params as P, to_proto\nimport caffe\n\n# some utility functions\ndef add_layer_to_net_spec(ns, caffe_layer, name, *args, **kwargs):\n  kwargs.update({'name':name})\n  l = caffe_layer(*args, **kwargs)\n  ns.__setattr__(name, l)\n  return ns.__getattr__(name)\ndef add_layer_with_multiple_tops(ns, caffe_layer, lname, ntop, *args, **kwargs):    \n  kwargs.update({'name':lname,'ntop':ntop})\n  num_in = len(args)-ntop # number of input blobs\n  tops = caffe_layer(*args[:num_in], **kwargs)\n  for i in xrange(ntop):\n      ns.__setattr__(args[num_in+i],tops[i])\n  return tops\n\n# implement single time step LSTM unit\ndef single_time_step_lstm( ns, h0, c0, x, prefix, num_output, weight_names=None):\n  \"\"\"\n  see arXiv:1511.04119v1\n  \"\"\"\n  if weight_names is None:\n      weight_names = ['w_'+prefix+nm for nm in ['Mxw','Mxb','Mhw']]\n  # full InnerProduct (incl. bias) for x input\n  Mx = add_layer_to_net_spec(ns, L.InnerProduct, prefix+'lstm\/Mx', x,\n                    inner_product_param={'num_output':4*num_output,'axis':2,\n                                           'weight_filler':{'type':'uniform','min':-0.05,'max':0.05},\n                                           'bias_filler':{'type':'constant','value':0}},\n                    param=[{'lr_mult':1,'decay_mult':1,'name':weight_names[0]},\n                           {'lr_mult':2,'decay_mult':0,'name':weight_names[1]}])\n  Mh = add_layer_to_net_spec(ns, L.InnerProduct, prefix+'lstm\/Mh', h0,\n                    inner_product_param={'num_output':4*num_output, 'axis':2, 'bias_term': False,\n                                       'weight_filler':{'type':'uniform','min':-0.05,'max':0.05},\n                                       'bias_filler':{'type':'constant','value':0}},\n                    param={'lr_mult':1,'decay_mult':1,'name':weight_names[2]})\n  M = add_layer_to_net_spec(ns, L.Eltwise, prefix+'lstm\/Mx+Mh', Mx, Mh,\n                          eltwise_param={'operation':P.Eltwise.SUM})\n  raw_i1, raw_f1, raw_o1, raw_g1 = \\\n  add_layer_with_multiple_tops(ns, L.Slice, prefix+'lstm\/slice', 4, M,\n                             prefix+'lstm\/raw_i', prefix+'lstm\/raw_f', prefix+'lstm\/raw_o', prefix+'lstm\/raw_g',\n                             slice_param={'axis':2,'slice_point':[num_output,2*num_output,3*num_output]})\n  i1 = add_layer_to_net_spec(ns, L.Sigmoid, prefix+'lstm\/i', raw_i1, in_place=True)\n  f1 = add_layer_to_net_spec(ns, L.Sigmoid, prefix+'lstm\/f', raw_f1, in_place=True)\n  o1 = add_layer_to_net_spec(ns, L.Sigmoid, prefix+'lstm\/o', raw_o1, in_place=True)\n  g1 = add_layer_to_net_spec(ns, L.TanH, prefix+'lstm\/g', raw_g1, in_place=True)\n  c1_f = add_layer_to_net_spec(ns, L.Eltwise, prefix+'lstm\/c_f', f1, c0, eltwise_param={'operation':P.Eltwise.PROD})\n  c1_i = add_layer_to_net_spec(ns, L.Eltwise, prefix+'lstm\/c_i', i1, g1, eltwise_param={'operation':P.Eltwise.PROD})\n  c1 = add_layer_to_net_spec(ns, L.Eltwise, prefix+'lstm\/c', c1_f, c1_i, eltwise_param={'operation':P.Eltwise.SUM})\n  act_c = add_layer_to_net_spec(ns, L.TanH, prefix+'lstm\/act_c', c1, in_place=False) # cannot override c - it MUST be preserved for next time step!!!\n  h1 = add_layer_to_net_spec(ns, L.Eltwise, prefix+'lstm\/h', o1, act_c, eltwise_param={'operation':P.Eltwise.PROD})\n  return c1, h1, weight_names\n<\/code><\/pre>\n\n<p>Once you have the single time step, you can unroll it as many times you want...<\/p>\n\n<pre><code>def exmaple_use_of_lstm():\n  T = 3 # number of time steps\n  B = 10 # batch size\n  lstm_output = 500 # dimension of LSTM unit\n\n  # use net spec\n  ns = caffe.NetSpec()\n\n  # we need initial values for h and c\n  ns.h0 = L.DummyData(name='h0', dummy_data_param={'shape':{'dim':[1,B,lstm_output]},\n                               'data_filler':{'type':'constant','value':0}})\n\n  ns.c0 = L.DummyData(name='c0', dummy_data_param={'shape':{'dim':[1,B,lstm_output]},\n                                   'data_filler':{'type':'constant','value':0}})\n\n  # simulate input X over T time steps and B sequences (batch size)\n  ns.X = L.DummyData(name='X', dummy_data_param={'shape': {'dim':[T,B,128,10,10]}} )\n  # slice X for T time steps\n  xt = L.Slice(ns.X, name='slice_X',ntop=T,slice_param={'axis':0,'slice_point':range(1,T)})\n  # unroling\n  h = ns.h0\n  c = ns.c0\n  lstm_weights = None\n  tops = []\n  for t in xrange(T):\n    c, h, lstm_weights = single_time_step_lstm( ns, h, c, xt[t], 't'+str(t)+'\/', lstm_output, lstm_weights)\n    tops.append(h)\n    ns.__setattr__('c'+str(t),c)\n    ns.__setattr__('h'+str(t),h)\n  # concat all LSTM tops (h[t]) to a single layer\n  ns.H = L.Concat( *tops, name='concat_h',concat_param={'axis':0} )\n  return ns\n<\/code><\/pre>\n\n<p>Writing the prototxt:<\/p>\n\n<pre><code>ns = exmaple_use_of_lstm()\nwith open('lstm_demo.prototxt','w') as W:\n  W.write('name: \"LSTM using NetSpec example\"\\n')\n  W.write('%s\\n' % ns.to_proto())\n<\/code><\/pre>\n\n<p>The resulting unrolled net (for three time steps) looks like<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/K11tK.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/K11tK.png\" alt=\"LSTM\"><\/a><\/p>\n"}]}
{"QuestionId":32226362,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-08-26T12:14:25.750","AcceptedAnswerId":32227610.0,"OwnerUserId":1139393.0,"Title":"Advanced 2d indexing in Theano to extract multiple pixels from an image","Body":"<h2>Background<\/h2>\n\n<p>I have an image that I want to sample at a number (P) of x,y coordinates.<\/p>\n\n<p>In Numpy I can use advanced indexing to do this via:<\/p>\n\n<pre><\/pre>\n\n<p>This returns a vector of P pixels sampled from the image.<\/p>\n\n<h2>Question<\/h2>\n\n<p>How can I do this advanced indexing in Theano?<\/p>\n\n<h2>What I've tried<\/h2>\n\n<p>I've tried the corresponding code in theano:<\/p>\n\n<pre><\/pre>\n\n<p>this compiles and executes without any warning messages, but results in an output tensor of shape (2,8,100), so it looks like it is doing some variant of basic indexing returning lots of rows of the image, instead of extracting pixels.<\/p>\n\n<h2>Full code<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This prints (8,) for Numpy, and (2,8,100) for Theano.<\/p>\n\n<p>My Theano version is 0.7.0.dev-RELEASE<\/p>\n","answers":[{"AnswerId":"32227610","CreationDate":"2015-08-26T13:11:27.493","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This appears to be a subtle difference between Theano and numpy advanced indexing (there are other differences that don't apply here).<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>t_points = t_image[ [t_pos[:,1],t_pos[:,0]] ]\n<\/code><\/pre>\n\n<p>you need to use<\/p>\n\n<pre><code>t_points = t_image[ (t_pos[:,1],t_pos[:,0]) ]\n<\/code><\/pre>\n\n<p>Note the change from a list of lists to a tuple of lists. This variant also works in numpy so it's probably best to just use a tuple instead of a list there too.<\/p>\n"}]}
{"QuestionId":32229486,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-08-26T14:31:29.103","AcceptedAnswerId":32230092.0,"OwnerUserId":3042790.0,"Title":"verify_grad function: 'TensorVariable' object is not callable","Body":"<p>I'd like to use the verify_grad function, but I'm getting errors of the form \"'TensorVariable' object is not callable\". <\/p>\n\n<pre><\/pre>\n\n<p>In the doc it says that fun is \"a Python function that takes Theano variables as inputs, and returns a Theano variable. For instance, an Op instance with a single output.\"<\/p>\n\n<p>I have gone through the graph structures section in the docs and I thought I understood what an op node is, but apparently I don't. <\/p>\n\n<p>E.g. if I have two TensorVariables x and y and I'd like to take the product of them, then * is the op node, correct? But if I declare z=x*y, then z is again a TensorVariable, right?<\/p>\n\n<p>So is there any way how to define an op for e.g. a negative log likelihood function in order to evaluate the correctness of the gradient for that function? Or is there any other way to get the numerical gradient in theano for a function that you constructed?<\/p>\n","answers":[{"AnswerId":"32230092","CreationDate":"2015-08-26T14:58:51.553","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's an example of <code>verify_grad<\/code> in use:<\/p>\n\n<pre><code>import numpy\nimport theano\n\ndef something_complicated(x, y):\n    z = x * y\n    return z\n\nx_value = numpy.array([[1., 2., 3.], [4., 5., 6.]], dtype=theano.config.floatX)\ny_value = numpy.array([[7., 8., 9.], [10., 11., 12.]], dtype=theano.config.floatX)\ntheano.gradient.verify_grad(something_complicated, (x_value, y_value), rng=numpy.random)\n<\/code><\/pre>\n\n<p>As required, <code>something_complicated<\/code> is \"a Python function that takes Theano variables as inputs [<code>x<\/code> and <code>y<\/code> in this case], and returns a Theano variable [<code>z<\/code> in this case].\"<\/p>\n\n<p>You can construct any symbolic expression inside <code>something_complicated<\/code>, such as the computation for a negative log likelihood.<\/p>\n\n<p>A Theano operation can be anything as long as<\/p>\n\n<ol>\n<li>It is <em>callable<\/em> (objects are callable if they implement the special <a href=\"https:\/\/docs.python.org\/2\/reference\/datamodel.html#object.__call__\" rel=\"nofollow\"><code>__call__<\/code> function<\/a>)<\/li>\n<li>When it is called, it treats all inputs as Theano variables<\/li>\n<li>When it is called, it only returns Theano variables.<\/li>\n<\/ol>\n\n<p><code>something_complicated<\/code> clearly meets these requirements. It is callable by virtue of being a Python function, it assumes <code>x<\/code> and <code>y<\/code> are Theano variables, and its return value <code>z<\/code> is also a Theano variable.<\/p>\n"}]}
{"QuestionId":32229882,"AnswerCount":1,"Tags":"<python><gpu><gpgpu><theano>","CreationDate":"2015-08-26T14:50:06.197","AcceptedAnswerId":null,"OwnerUserId":5186172.0,"Title":"Are int operations possible on GPU in Theano?","Body":"<p>So I have read that theano cannot do gpu computations with  and to store ints as shared variables on the gpu they have to be initialised as shared  data, and then recast to ints (like in the \"little hack\" in the logistic regression <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">example<\/a>)...but after such a recast, can theano do gpu computations on <strong>ints<\/strong>? and is storage a precondition for computation?  In other words, are the following two scenarios possible?<\/p>\n\n<p>Scenario 1.  I want to do a dot product on two large vectors of ints.  I therefore make them shared as  and recast them to int before the dot product, is this dot product then done on the gpu(regardless of int type)?<\/p>\n\n<p>Scenario 2.  If scenario 1 is possible, would it still be possible to do the computation on the gpu without storing them as shared float32 first?  (I understand that sharing variables might mitigate gpu-cpu communication, but would the dot product still be possible?  Is storage a precondition for computation on gpu?)<\/p>\n","answers":[{"AnswerId":"32231011","CreationDate":"2015-08-26T15:41:27.880","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>No, there is (currently) no way to do any operations on the GPU with any type other than <code>float32<\/code>.<\/p>\n\n<p>This can be seen with this little demo code:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\nx = theano.shared(numpy.arange(9 * 10).reshape((9, 10)).astype(numpy.float32))\ny = theano.shared(numpy.arange(10 * 11).reshape((10, 11)).astype(numpy.float32))\nz = theano.dot(tt.cast(x, 'int32'), tt.cast(y, 'int32'))\nf = theano.function([], outputs=z)\ntheano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>When run on a GPU it will print the following computation graph:<\/p>\n\n<pre><code>dot [@A] ''   4\n |Elemwise{Cast{int32}} [@B] ''   3\n | |HostFromGpu [@C] ''   1\n |   |&lt;CudaNdarrayType(float32, matrix)&gt; [@D]\n |Elemwise{Cast{int32}} [@E] ''   2\n   |HostFromGpu [@F] ''   0\n     |&lt;CudaNdarrayType(float32, matrix)&gt; [@G]\n<\/code><\/pre>\n\n<p>Here you can see that the two shared variables are indeed stored in GPU memory (the two <code>CudaNdarrayType<\/code>s) but they are moved to the host (i.e. CPU\/main memory) from the GPU (the <code>HostFromGpu<\/code> operations) before being cast to ints and a regular <code>dot<\/code> operation being used.<\/p>\n\n<p>If the casts are omitted then you would see<\/p>\n\n<pre><code>HostFromGpu [@A] ''   1\n |GpuDot22 [@B] ''   0\n   |&lt;CudaNdarrayType(float32, matrix)&gt; [@C]\n   |&lt;CudaNdarrayType(float32, matrix)&gt; [@D]\n<\/code><\/pre>\n\n<p>Showing that the GPU is performing the dot product (the <code>GpuDot22<\/code> operation) but on floating point data, not integer data.<\/p>\n"}]}
{"QuestionId":32233262,"AnswerCount":1,"Tags":"<matlab><machine-learning><computer-vision><neural-network><caffe>","CreationDate":"2015-08-26T17:43:51.457","AcceptedAnswerId":32240964.0,"OwnerUserId":5269749.0,"Title":"How to extract trained caffe kernels filter of the first layer","Body":"<p>I am using caffe and I wonder if I just can use one of the filters separately.\nSo basically I just need the trained kernel of that filters (using in the first layer).<br>\nI could not find the formula of the kernels in the paper.<br>\nSo I really appreciate it if someone can help me out please.<br>\nIf you can also tell me how to extract them in matlab version I would be so grateful.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"32240964","CreationDate":"2015-08-27T04:52:50.310","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Suppose you have a trained net with its <code>'deploy.prototxt'<\/code> file defining the net, and the trained parameters at <code>'my_weights.caffemodel'<\/code> file.<br>\nSuppose the layer you are interested in is defined like this in <code>'deploy.prototxt<\/code>':<\/p>\n\n<pre><code>layer {\n  name: \"conv1\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"conv1\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  convolution_param {\n    num_output: 32\n    pad: 2\n    kernel_size: 5\n    stride: 1\n  }\n}\n<\/code><\/pre>\n\n<p>As you can see the layer's name is <code>\"conv1\"<\/code> it has 32 filters of size 5-by-5.<\/p>\n\n<p>First, you need to load the net in Matlab<\/p>\n\n<pre><code> &gt;&gt; net = get_net( 'path\/to\/deploy.prototxt', 'path\/to\/my_weights.caffemodel', 'test' );\n<\/code><\/pre>\n\n<p>Once you load the net, you can access its parameters using the layer's name<\/p>\n\n<pre><code>&gt;&gt; w = net.params( 'conv1', 1 ).get(); \n<\/code><\/pre>\n"}]}
{"QuestionId":32240380,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano>","CreationDate":"2015-08-27T03:49:43.540","AcceptedAnswerId":32311521.0,"OwnerUserId":4013571.0,"Title":"Best way to re-initialise a compiled Theano function","Body":"<p>I want to refresh my compiled MLP model in Theano as I want to repeat a model with different hyper-parameters.<\/p>\n\n<p>I am aware that I can redefine all the functions, however, compile time for each function is significant.<\/p>\n\n<p>I want to define a function which will refresh the models of their <\/p>\n\n<p>The following code below is shown for demonstration.<\/p>\n\n<pre><\/pre>\n\n<p>My instinct suggests that I can simply re-define the  class without any implications on the other compiled functions.<\/p>\n\n<p>Is this correct?<\/p>\n\n<p>I was thinking that if this is the case I could define a function of  which re-instanciated parameters for each component of the  class <\/p>\n","answers":[{"AnswerId":"32311521","CreationDate":"2015-08-31T12:39:52.483","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It's unclear how the <code>MLP<\/code> class is working but you can re-use a previously compiled computation as long as the number of dimensions of the shared variables does not change.<\/p>\n\n<p>In the following example the <code>compile_model<\/code> function creates a simple neural network with randomly initialized parameters. After training with those parameters, the shared variables are re-initialized to new random values, but this time the network's hidden layer size is increased. Despite this change in size, the original training function is re-used.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile_model(input_size, hidden_size, output_size, learning_rate):\n    w_h = theano.shared(numpy.random.randn(input_size, hidden_size).astype(theano.config.floatX))\n    b_h = theano.shared(numpy.zeros(hidden_size, dtype=theano.config.floatX))\n    w_y = theano.shared(numpy.random.randn(hidden_size, output_size).astype(theano.config.floatX))\n    b_y = theano.shared(numpy.zeros(output_size, dtype=theano.config.floatX))\n    parameters = (w_h, b_h, w_y, b_y)\n    x = tt.matrix()\n    z = tt.lvector()\n    h = tt.tanh(theano.dot(x, w_h) + b_h)\n    y = tt.nnet.softmax(theano.dot(h, w_y) + b_y)\n    c = tt.nnet.categorical_crossentropy(y, z).mean()\n    u = [(p, p - learning_rate * theano.grad(c, p)) for p in parameters]\n    trainer = theano.function([x, z], outputs=[c], updates=u)\n    tester = theano.function([x], outputs=[y])\n    return trainer, tester, parameters\n\n\ndef refresh_model(parameters, input_size, hidden_size, output_size):\n    w_h, b_h, w_y, b_y = parameters\n    w_h.set_value(numpy.random.randn(input_size, hidden_size).astype(theano.config.floatX))\n    b_h.set_value(numpy.zeros(hidden_size, dtype=theano.config.floatX))\n    w_y.set_value(numpy.random.randn(hidden_size, output_size).astype(theano.config.floatX))\n    b_y.set_value(numpy.zeros(output_size, dtype=theano.config.floatX))\n\n\ndef main():\n    input_size = 30\n    hidden_size = 10\n    output_size = 20\n    learning_rate = 0.01\n    batch_size = 40\n    epoch_count = 50\n\n    trainer, tester, parameters = compile_model(input_size, hidden_size, output_size, learning_rate)\n    x = numpy.random.randn(batch_size, input_size)\n    z = numpy.random.randint(output_size, size=(batch_size,))\n    print 'Training model with hidden size', hidden_size\n\n    for _ in xrange(epoch_count):\n        print trainer(x, z)\n\n    hidden_size = 15\n    refresh_model(parameters, input_size, hidden_size, output_size)\n    print 'Training model with hidden size', hidden_size\n\n    for _ in xrange(epoch_count):\n        print trainer(x, z)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":32241193,"AnswerCount":1,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-08-27T05:14:39.157","AcceptedAnswerId":32243571.0,"OwnerUserId":1377127.0,"Title":"Error with Caffe C++ example with different deploy.prototxt file","Body":"<p>I <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow noreferrer\">trained<\/a> a model using the MNIST <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow noreferrer\">example architecture<\/a> (but on my own set of 3 image classes) and have been trying to integrate it into the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\" rel=\"nofollow noreferrer\">C++ example<\/a>. I modified the MNIST architecture file to make it similar to the deploy.prototxt file for the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt\" rel=\"nofollow noreferrer\">C++ example<\/a> (<a href=\"https:\/\/stackoverflow.com\/questions\/32208193\/how-to-modify-caffe-network-input-for-c-api\/32211796?noredirect=1#comment52319229_32211796\">replacing the train and test layers with the input layer<\/a>).  <\/p>\n\n<p>Unfortunately, when I run the C++ program it gives me the following error:<\/p>\n\n<blockquote>\n  <p>F0827 14:57:28.427697 25511 insert_splits.cpp:35] Unknown bottom blob\n   'label' (layer 'accuracy', bottom index 1)<\/p>\n<\/blockquote>\n\n<p>I tried to Google it and I think there's some difference between the layers in the files for the MNIST and C++ examples but can't work out what I can change to make this work.<\/p>\n","answers":[{"AnswerId":"32243571","CreationDate":"2015-08-27T07:49:45.647","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>As pointed out by <a href=\"https:\/\/stackoverflow.com\/questions\/32241193\/error-with-caffe-c-example-with-different-deploy-prototxt-file#comment52364300_32241193\">AbdulRahman AlHamali's comment<\/a> it seems like you left in your <code>deploy.prototxt<\/code> file the loss and accuracy layers that expects as inputs (\"bottom\"s) <code>\"label\"<\/code>.<br>\nRemoving these loss layers from <code>deploy.prototxt<\/code> should solve the poroblem.<\/p>\n\n<p>Note, that if you used <code>\"SoftmaxWithLoss\"<\/code> layer as a loss, you need to <em>replace<\/em> it with a <code>\"Softmax\"<\/code> layer to get class probabilities as the net outputs. <code>\"Softmax\"<\/code> layer takes only one <code>\"bottom\"<\/code> and does not require <code>bottom: \"label\"<\/code>.<\/p>\n"}]}
{"QuestionId":32243285,"AnswerCount":2,"Tags":"<autoencoder><keras>","CreationDate":"2015-08-27T07:34:05.283","AcceptedAnswerId":null,"OwnerUserId":4535321.0,"Title":"keras autoencoder not converging","Body":"<p>Could someone please explain to me why the autoencoder is not converging? To me the results of the two networks below should be the same. However, the autoencoder below is not converging, whereas, the network beneath it is. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<hr>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"33327770","CreationDate":"2015-10-25T08:44:35.913","ParentId":null,"OwnerUserId":"2616754","Title":null,"Body":"<p>I think Keras's Autoencoder implementation ties the weights of the encoder and decoder, whereas in your implementation, the encoder and decoder have separate weights. If your implementation is leading to much better performance on the test data, then it may indicate that un-tied weights may be needed for your problem. <\/p>\n"},{"AnswerId":"34306306","CreationDate":"2015-12-16T07:35:15.470","ParentId":null,"OwnerUserId":"2031688","Title":null,"Body":"<p>The new version (0.3.0) of Keras no longer has tied weights in AutoEncoder, and it still shows different convergence. This is because weights are initialized differently.<\/p>\n\n<p>In the non-AE example, Dense(32,16) weights are initialized first, followed by Dense(16,32). In the AE example, Dense(32,16) weights are initialized first, followed by Dense(16,32), and then when you create the AutoEncoder instance, Dense(32,16) weights are initialized again (self.encoder.set_previous(node) will call build() to initialize weights).<\/p>\n\n<p>Now the following two NNs converge exactly the same:<\/p>\n\n<pre><code>autoencoder = Sequential()\nencoder = containers.Sequential([Dense(32,16,activation='tanh')]) \ndecoder = containers.Sequential([Dense(16,32)])\nautoencoder.add(AutoEncoder(encoder=encoder, decoder=decoder, \n                        output_reconstruction=True))\nrms = RMSprop()\nautoencoder.compile(loss='mean_squared_error', optimizer=rms)\nnp.random.seed(0)\nautoencoder.fit(trainData,trainData, nb_epoch=20, batch_size=64,\n            validation_data=(testData, testData), show_accuracy=False)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code># non-autoencoder\nmodel = Sequential()\nmodel.add(Dense(32,16,activation='tanh')) \nmodel.add(Dense(16,32))\nmodel.set_weights(autoencoder.get_weights())\nmodel.compile(loss='mean_squared_error', optimizer=rms)\nnp.random.seed(0)\nmodel.fit(trainData,trainData, nb_epoch=numEpochs, batch_size=batch_size,\n            validation_data=(testData, testData), show_accuracy=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":32247374,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-08-27T10:49:25.803","AcceptedAnswerId":32248352.0,"OwnerUserId":1377127.0,"Title":"Caffe output layer number accuracy","Body":"<p>I've modified the Caffe <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\">MNIST example<\/a> to classify 3 classes of image. One thing I noticed was that if I specify the number of output layers as 3, then my test accuracy drops horribly - down to the low 40% range. However, if I +1 and have 4 output layers, the result is in the 95% range.<br>\nI added an extra class of images to my dataset (so 4 classes) and noticed the same thing - if the number of output layers were the same as the number of classes, then the result was horrible, if it was the same +1, then it worked really well.<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know why this is? I've noticed that when I use the model I train with the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\">C++ example code<\/a> on an image from my test set then it will complain that I've told it that there are 4 classes present and I've only supplied labels for 3 in my labels file. If I invent a label and add it to the file, I can get the program to run, but then it just returns one of the classes with a probability of 1.0 no matter what image I give it.<\/p>\n","answers":[{"AnswerId":"32248352","CreationDate":"2015-08-27T11:36:40.057","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It is important to notice that when fine-tuning and\/or changing the number of labels the input labels <strong>must<\/strong> always start from 0, as they are used as indices into the output probability vector when computing the loss.<br>\nThus, if you have<\/p>\n\n<pre><code> inner_product_params {\n   num_output: 3\n }\n<\/code><\/pre>\n\n<p>You must have training labels 0,1 and 2 only.<\/p>\n\n<p>If you use <code>num_output: 3<\/code> with labels 1,2,3 caffe is unable to represent label 3 and in fact has a redundant line corresponding to label 0 that is left unused.<br>\nAs you observed, when changing to <code>num_output: 4<\/code> caffe is again able to represent label 3 and the results improved, but still you have an unused row in the parameters matrix.<\/p>\n"}]}
{"QuestionId":32249861,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-08-27T12:49:54.467","AcceptedAnswerId":32292853.0,"OwnerUserId":1377127.0,"Title":"Caffe predicts same class regardless of image","Body":"<p>I <a href=\"https:\/\/github.com\/jackbrucesimpson\/caffe_tests\/blob\/master\/lenet_train_test.prototxt\" rel=\"nofollow\">modified the MNIST example<\/a> and when I train it with my 3 image classes it returns an accuracy of 91%. However, when I modify the <a href=\"https:\/\/github.com\/jackbrucesimpson\/caffe_tests\/blob\/master\/classification.cpp\" rel=\"nofollow\">C++ example<\/a> with a <a href=\"https:\/\/github.com\/jackbrucesimpson\/caffe_tests\/blob\/master\/deploy_arch.prototxt\" rel=\"nofollow\">deploy prototxt<\/a> file and <a href=\"https:\/\/github.com\/jackbrucesimpson\/caffe_tests\/blob\/master\/labels.txt\" rel=\"nofollow\">labels file<\/a>, and try to test it on some images it returns a prediction of the second class (1 circle) with a probability of 1.0 no matter what image I give it - even if it's images that were used in the training set. I've tried a dozen images and it consistently just predicts the one class.<\/p>\n\n<p>To clarify things, in the C++ example I modified I did scale the image to be predicted just like the images were scaled in the training stage:<\/p>\n\n<pre><\/pre>\n\n<p>If that was the right thing to do, then it makes me wonder if I've done something wrong with the output layers that calculate probability in my <a href=\"https:\/\/github.com\/jackbrucesimpson\/caffe_tests\/blob\/master\/deploy_arch.prototxt\" rel=\"nofollow\">deploy_arch.prototxt<\/a> file.<\/p>\n","answers":[{"AnswerId":"32292853","CreationDate":"2015-08-30T03:25:17.773","ParentId":null,"OwnerUserId":"2376331","Title":null,"Body":"<p>I think you have forgotten to scale the input image during classification time, as can be seen in line 11 of the train_test.prototxt file. You should probably multiply by that factor somewhere in your C++ code, or alternatively use a Caffe layer to scale the input (look into ELTWISE or POWER layers for this).<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>After a conversation in the comments, it turned out that the image mean was mistakenly being subtracted in the classification.cpp file whereas it was not being subtracted in the original training\/testing pipeline.<\/p>\n"},{"AnswerId":"32293302","CreationDate":"2015-08-30T04:55:54.373","ParentId":null,"OwnerUserId":"4329255","Title":null,"Body":"<p>Are your train classes balanced?\nYou may get to a stacked network on a prediction of one major class.\nIn order to find the issue I suggest to output the train prediction during training compared to predictions with the forward example on same train images from a different class.<\/p>\n"}]}
{"QuestionId":32268923,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2015-08-28T10:30:05.277","AcceptedAnswerId":null,"OwnerUserId":2537088.0,"Title":"In Lasagne\/Theano, do I need a 4d numpy array for a 4d Theano tensor?","Body":"<p>I'm building a neural network with lasagne and am <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\" rel=\"nofollow\">following the example from the github.<\/a> I'm curious on how exactly to input into the network. In the example they state that the input layer is 4 dimensions and indeed it is a theano tensor4. Does this mean I have to give the network a 4 dimensional numpy array? Is that even possible? How would you build one from a 4 d vector of lists?<\/p>\n","answers":[{"AnswerId":"32270198","CreationDate":"2015-08-28T11:37:48.033","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>In the MNIST example provided by Lasagne, you need to input a 4D tensor.<\/p>\n\n<p>Generally speaking, if your data is 2 dimensional (images for example), the shape of your input must be <code>(n_samples, n_channels, height, width)<\/code>. In the MNIST dataset <code>n_channel<\/code> is 1 (might be something else, e.g. 3 for RGB images), and <code>height<\/code> and <code>width<\/code> are both 28.<\/p>\n\n<p>If your data is only 1 dimensional, then you must input a 3D tensor, of shape <code>(n_samples, n_channel, n_features)<\/code>.<\/p>\n\n<p>Note that this might be problematic if you want to predict a label for a single image ((28, 28) ndarray as in <a href=\"https:\/\/stackoverflow.com\/questions\/31499761\/get-output-from-lasagne-python-deep-neural-network-framework\/32007540#32007540\">this question<\/a>, because you need to make the input 4 dimensional. In this case, you can add axis using <code>data = data[None, None, :, :]<\/code>.<\/p>\n"}]}
{"QuestionId":32274543,"AnswerCount":2,"Tags":"<python><machine-learning><theano>","CreationDate":"2015-08-28T15:23:33.687","AcceptedAnswerId":32275233.0,"OwnerUserId":5197007.0,"Title":"In Python, can I call the variable from main function - use global variable?","Body":"<p>In Python, can I call a variable from main function? Use global variable? Any help appreciated! <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32275188","CreationDate":"2015-08-28T15:59:26.380","ParentId":null,"OwnerUserId":"678611","Title":null,"Body":"<p>Try with this.<\/p>\n\n<p><strong>main.py<\/strong><\/p>\n\n<pre><code>__dataset__ = main(dataset, n_h, n_y, batch_size, dev_split, 5000)\n<\/code><\/pre>\n\n<p><strong>sub.py<\/strong><\/p>\n\n<pre><code>import sys, main\n__dataset__ = sys.modules['__main__'].__dataset__\n<\/code><\/pre>\n\n<p><br>\n<strong>EDIT:<\/strong><br>\nAnother method is to use a class with static variables.<\/p>\n\n<p><strong>mclass.py<\/strong><\/p>\n\n<pre><code>class MClass:\n    i = 0\n\nMClass.i = 1\n<\/code><\/pre>\n\n<p><strong>main.py<\/strong><\/p>\n\n<pre><code>import sub\nfrom mclass import MClass\n\n# In the main file\nprint(MClass.i) # Outputs 1\nMClass.i = 3\nprint(MClass.i) # Outputs 3\n\n# In a subfile\nsub.mPrint() # Outputs 3\nsub.set(10)\nsub.mPrint() # Outputs 10\n\n# And back in the main\nprint(MClass.i) # Outputs 10\n<\/code><\/pre>\n\n<p><strong>sub.py<\/strong><\/p>\n\n<pre><code>from mclass import MClass\n\ndef mPrint():\n   print(MClass.i)\n\ndef set(n):\n   MClass.i = n\n<\/code><\/pre>\n"},{"AnswerId":"32275233","CreationDate":"2015-08-28T16:01:43.670","ParentId":null,"OwnerUserId":"5215951","Title":null,"Body":"<p>You have to define 'input_to_state' and 'RNN' outside the 'main' function, and then modify them afterwards. Like this:<\/p>\n\n<pre><code>input_to_state = None\nRNN = None\ndef main(dataset, n_h, n_y, batch_size, dev_split, n_epochs):\n    # Calling 'global' allows you to modify these variables\n    global input_to_state\n    global RNN\n    input_to_state = Linear(name='input_to_state',\n                            input_dim=seq_u.shape[-1],\n                            output_dim=n_h)\n    RNN = SimpleRecurrent(activation=Tanh(),\n                          dim=n_h, name=\"RNN\")\n\n\ndef predict(dev_X):\n    dev_transform = input_to_state.apply(dev_X)\n    dev_h = RNN.apply(dev_transform)\n\nif __name__ == \"__main__\":   \n    main(args) \n    predict(dev_X)\n<\/code><\/pre>\n\n<p>Howerver, I would not recommend this, global variables should be used as little as possible. <a href=\"https:\/\/google-styleguide.googlecode.com\/svn\/trunk\/pyguide.html#Global_variables\" rel=\"nofollow\">more detail here<\/a>.<\/p>\n\n<p>A better solution would be to return 'input_to_state' and 'RNN' at the end of the main function, like this:<\/p>\n\n<pre><code>def main(dataset, n_h, n_y, batch_size, dev_split, n_epochs):\n    input_to_state = Linear(name='input_to_state',\n                            input_dim=seq_u.shape[-1],\n                            output_dim=n_h)\n    RNN = SimpleRecurrent(activation=Tanh(),\n                          dim=n_h, name=\"RNN\")\n    return input_to_state, RNN\n\ndef predict(dev_X, input_to_state, RNN):\n    dev_transform = input_to_state.apply(dev_X)\n    dev_h = RNN.apply(dev_transform)\n\nif __name__ == \"__main__\":   \n    input_to_state, RNN = main(args) \n    predict(dev_X, input_to_state, RNN)\n<\/code><\/pre>\n"}]}
{"QuestionId":32278568,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-08-28T19:32:29.030","AcceptedAnswerId":32293404.0,"OwnerUserId":2832033.0,"Title":"How to use convert_imageset in caffe for images which are not put in one folder?","Body":"<p>I'm trying to train a CNN on my own dataset using Caffe framework, and it is highly recommended that the dataset be converted to the lmdb or leveldb formats due to speed efficiency. To do so, all images must be put into a <strong>single folder<\/strong> and the  must be prepared accordingly. My own dataset is so huge and in so many folders and subfolders so that it would be so laborious copy all of them into a single folder. Therefore\u0648 I'm wondering to know whether there exist any alternative way to generate the lmdb file without need to copy all images into a single folder.<\/p>\n","answers":[{"AnswerId":"32293404","CreationDate":"2015-08-30T05:14:53.450","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>There are (at least) two solutions to your problem.<\/p>\n\n<ol>\n<li><p>Don't copy the files to a single folder, just create <a href=\"http:\/\/www.cyberciti.biz\/faq\/creating-soft-link-or-symbolic-link\/\" rel=\"nofollow\">symbolic links<\/a>.<\/p><\/li>\n<li><p>All the images do not have to be in the same folder. You can have full paths in <code>'list.txt'<\/code> file. For example:<\/p><\/li>\n<\/ol>\n\n<blockquote>\n  <p>\/path\/to\/image.jpg 0<br>\n  \/another\/path\/class01.jpg 1<br>\n  \/yet\/another\/path\/class0.jpg 0  <\/p>\n<\/blockquote>\n\n<p>And so on...<\/p>\n"}]}
{"QuestionId":32280071,"AnswerCount":2,"Tags":"<python><arrays><numpy><matrix><theano>","CreationDate":"2015-08-28T21:34:05.073","AcceptedAnswerId":32280982.0,"OwnerUserId":4491330.0,"Title":"How to assign values elementwise to theano matrix ? Difference between Numpy and Theano?","Body":"<p>I'm new to theano. I would like to replace the numpy functions in my scripts with theano functions in order to speed up the calculation process. I'm not sure how to do it. <\/p>\n\n<p>My final goal is to apply affine transformation to 3D rigid body, assign a score to the conformation after each transformation, and do some optimization on the parameters determining the scores. <\/p>\n\n<p>Here's an example of what I'm trying to do. <\/p>\n\n<pre><\/pre>\n\n<p>The above theano function didn't pass compilation. I've got the following error message. <\/p>\n\n<pre><\/pre>\n\n<p>In general, my questions are <\/p>\n\n<p>(1) is there a way to assign or update or initialize a theano matrix with a specific shape elementwise, <\/p>\n\n<p>(2) as theano is closely related with numpy, what's the difference between theano and numpy in defining, optimizing, and evaluating mathematical expressions, <\/p>\n\n<p>and (3) can theano replace numpy in the sense that we can use theano functions solely in defining, optimizing, and evaluating mathematical expressions without calling numpy functions. <\/p>\n","answers":[{"AnswerId":"32282512","CreationDate":"2015-08-29T03:47:41.500","ParentId":null,"OwnerUserId":"4491330","Title":null,"Body":"<p>Just for getting the theano function posted above work, my version is: <\/p>\n\n<pre><code>angle_var = T.dscalar()\n\ndef rotate_x_axis_expr(angle):\n    a    = T.deg2rad(angle)\n    cosa = T.cos(a)\n    sina = T.sin(a)   \n\n    R = theano.shared(np.identity(4))\n    R = T.set_subtensor(R[1,1],  cosa)\n    R = T.set_subtensor(R[1,2], -sina)\n    R = T.set_subtensor(R[2,1],  sina)\n    R = T.set_subtensor(R[2,2],  cosa)\n\n    return R\n\nrotate_x_axis = theano.function([angle_var],rotate_x_axis_expr(angle_var))\n<\/code><\/pre>\n"},{"AnswerId":"32280982","CreationDate":"2015-08-28T23:03:10.467","ParentId":null,"OwnerUserId":"1730674","Title":null,"Body":"<p>I can't answer your questions 1, 2, 3, since I haven't used theano before ten minutes ago. But, to define the function in theano, you don't seem to use the <code>def<\/code> construction; you want to do something more like this:<\/p>\n\n<pre><code>angle_var = T.dscalar('angle_var')\na    = T.deg2rad(angle_var)\ncosa = T.cos(a)\nsina = T.sin(a)   \n\nR = theano.shared(np.identity(4))\nR = T.set_subtensor(R[1,1],  cosa)\nR = T.set_subtensor(R[1,2], -sina)\nR = T.set_subtensor(R[2,1],  sina)\nR = T.set_subtensor(R[2,2],  cosa)\n\nrotate_x_axis_theano = theano.function([angle_var], R)\n<\/code><\/pre>\n\n<p>Doesn't help much with speed though, for a scalar angle at least:<\/p>\n\n<pre><code>In [368]: timeit rotate_x_axis_theano(10)\n10000 loops, best of 3: 67.7 \u00b5s per loop\n\nIn [369]: timeit rotate_x_axis_numpy(10)\nThe slowest run took 4.23 times longer than the fastest. This could mean that an intermediate result is being cached\n10000 loops, best of 3: 22.7 \u00b5s per loop\n\nIn [370]: np.allclose(rotate_x_axis_theano(10), rotate_x_axis_numpy(10))\nOut[370]: True\n<\/code><\/pre>\n"}]}
{"QuestionId":32283364,"AnswerCount":2,"Tags":"<python><theano><lstm><lasagne>","CreationDate":"2015-08-29T06:34:23.860","AcceptedAnswerId":null,"OwnerUserId":2537088.0,"Title":"Can't figure out the issue with my Lasagne LSTM","Body":"<p>I'm curious if anyone has any insights. Even if you can't figure out the issue how can I begin to debug it. I must say, I am not strong in theano. <\/p>\n\n<p>The input data is a numpy tensor of shape (10,15,10)<\/p>\n\n<p>Here it is. It ran when I just hooked up the input to the dense layer. <\/p>\n\n<pre><\/pre>\n\n<p>Here is the error. It is kind of a doozy. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36096190","CreationDate":"2016-03-19T00:00:31.357","ParentId":null,"OwnerUserId":"3858735","Title":null,"Body":"<p>There are a few mistakes in this: <\/p>\n\n<p>Your input is a vector. Your output has 10 units per entry in your batch:<\/p>\n\n<pre><code>target_var = T.ivector('targets')\nl_out = lasagne.layers.DenseLayer(l_shp, num_units=10)\n<\/code><\/pre>\n\n<p>Your output shape is (150, 10) which you want to compare to your label input (10,):<\/p>\n\n<pre><code>l_shp =lasagne.layers.ReshapeLayer(l_lstm, (10*15, 10))\nprint 'l_shp shape:', l_shp.output_shape\nl_out = lasagne.layers.DenseLayer(l_shp, num_units=10)\n<\/code><\/pre>\n\n<p>You probably want to use a (10, 10) matrix for your labels. This way you are using 10 classes for your batch size of 10<\/p>\n\n<pre><code>target_var = T.matrix('targets')\n<\/code><\/pre>\n\n<p>You also need to change your network so that it outputs a (10,10) shaped result. You can do this by reshaping your l_shp differently by reshaping directly to (10, something). There are lots of other options. <\/p>\n\n<p>Btw, you can look at layer shapes with their 'output_shape' property.<\/p>\n"},{"AnswerId":"32401611","CreationDate":"2015-09-04T15:26:38.513","ParentId":null,"OwnerUserId":"1139393","Title":null,"Body":"<p>Is it possible that ResultsBatch contains numbers from 1 to 10?<\/p>\n\n<p>If so, you could try testing with <\/p>\n\n<pre><code>train_fn(TextBatch,ResultsBatch - 1)\n<\/code><\/pre>\n\n<p>to convert the targets to 0,1,2,..,9 instead.<\/p>\n"}]}
{"QuestionId":32288399,"AnswerCount":0,"Tags":"<machine-learning><computer-vision><svm><libsvm><caffe>","CreationDate":"2015-08-29T16:44:54.990","AcceptedAnswerId":null,"OwnerUserId":2832033.0,"Title":"How can I train an SVM classifier with leveldb datatype, generated from caffe framework?","Body":"<p>I've fed a bunch of images to the caffe framework (AlexNet) and the last feature descriptors have been extracted and stored in LEVELDB. Now, I want to train a linear SVM classifier on these descriptors. Therefore, I'm wondering to know whether there is any way to train the SVM with LEVELDB data or not. \nThanks<\/p>\n\n<p>EDIT:\nWorking with <a href=\"https:\/\/github.com\/jostosh\/caffe-utils\" rel=\"nofollow\">this tool<\/a> solved the problem<\/p>\n","answers":[]}
{"QuestionId":32302599,"AnswerCount":1,"Tags":"<deep-learning><theano><lasagne>","CreationDate":"2015-08-31T00:43:25.743","AcceptedAnswerId":32310274.0,"OwnerUserId":5081447.0,"Title":"Modifying perform function in Theano.tensor.nnet.softmax","Body":"<p>I have just begun using lasagne and Theano to do some machine learning on Python.<\/p>\n\n<p>I am trying to modify the softmax class in Theano. I want to change how the activation function(softmax) is calculated. Instead of dividing e_x by e_x.sum(axis=1), I want to divide e_x by sum of three consecutive numbers. <\/p>\n\n<p>For instance, the result will be as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and so on...<\/p>\n\n<p>The problem is that I cannot quite grasp how theano carries out the computation.<\/p>\n\n<p>Here is my main question. Does it suffice to just change the perform() function in the softmax class? <\/p>\n\n<p>Here is the original perform() function:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my modified perform()<\/p>\n\n<pre><\/pre>\n\n<p>With the current code, I am getting 'unorderable types:int()>str()' error when I use the predict method in lasagne.<\/p>\n","answers":[{"AnswerId":"32310274","CreationDate":"2015-08-31T11:34:19.807","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>For something like this you're probably better off constructing a custom softmax via symbolic expressions rather than creating (or modifying) an operation.<\/p>\n\n<p>Your custom softmax can be defined in terms of symbolic expressions. Doing it this way will give you gradients (and other Theano operation bits and pieces) \"for free\" but might run slightly slower than a custom operation could.<\/p>\n\n<p>Here's an example:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\nx = tt.matrix()\n\n# Use the built in softmax operation\ny1 = tt.nnet.softmax(x)\n\n# A regular softmax operation defined via ordinary Theano symbolic expressions\ny2 = tt.exp(x)\ny2 = y2 \/ y2.sum(axis=1)[:, None]\n\n# Custom softmax operation\ndef custom_softmax(a):\n    b = tt.exp(a)\n    b1 = b[:, :3] \/ b[:, :3].sum(axis=1)[:, None]\n    b2 = b[:, 3:] \/ b[:, 3:].sum(axis=1)[:, None]\n    return tt.concatenate([b1, b2], axis=1)\ny3 = custom_softmax(x)\n\nf = theano.function([x], outputs=[y1, y2, y3])\n\nx_value = [[.1, .2, .3, .4, .5, .6], [.1, .3, .5, .2, .4, .6]]\ny1_value, y2_value, y3_value = f(x_value)\nassert numpy.allclose(y1_value, y2_value)\nassert y3_value.shape == y1_value.shape\na = numpy.exp(.1) + numpy.exp(.2) + numpy.exp(.3)\nb = numpy.exp(.4) + numpy.exp(.5) + numpy.exp(.6)\nc = numpy.exp(.1) + numpy.exp(.3) + numpy.exp(.5)\nd = numpy.exp(.2) + numpy.exp(.4) + numpy.exp(.6)\nassert numpy.allclose(y3_value, [\n    [numpy.exp(.1) \/ a, numpy.exp(.2) \/ a, numpy.exp(.3) \/ a, numpy.exp(.4) \/ b, numpy.exp(.5) \/ b, numpy.exp(.6) \/ b],\n    [numpy.exp(.1) \/ c, numpy.exp(.3) \/ c, numpy.exp(.5) \/ c, numpy.exp(.2) \/ d, numpy.exp(.4) \/ d, numpy.exp(.6) \/ d]\n]), y3_value\n<\/code><\/pre>\n"}]}
{"QuestionId":32320919,"AnswerCount":0,"Tags":"<gpu><nvidia><deep-learning><caffe><nvidia-digits>","CreationDate":"2015-08-31T22:32:45.983","AcceptedAnswerId":null,"OwnerUserId":4058368.0,"Title":"Running DIGITS in CPU mode only","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/z2nR8.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/z2nR8.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am trying to build a model in DIGITS. I am using CPU only to do the learning.. However, I get a CUDA driver version error although I am not using GPU. What could this problem be? I have attached my solver.prototxt below. <a href=\"https:\/\/i.stack.imgur.com\/3dLOS.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3dLOS.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
{"QuestionId":32321943,"AnswerCount":1,"Tags":"<lua><neural-network><convolution><torch>","CreationDate":"2015-09-01T00:44:43.373","AcceptedAnswerId":32357456.0,"OwnerUserId":5240885.0,"Title":"Custom Spatial Convolution In Torch","Body":"<p>I need to perform a custom spatial convolution in Torch. Rather than simply multiplying each input pixel by a weight for that pixel and adding them together with the filter's bias to form each output pixel, I need to do a more complex mathematical function to the input pixels before adding them together.<\/p>\n\n<p>I know how to do this, but I do not know a GOOD way to do this. The best way I've come up with is to take the full input tensor, create a bunch of secondary tensors that are \"views\" of the original without allocating additional memory, putting those into a Replicate layer (the output filter count being the replication count), and feeding that into a ParallelTable layer containing a bunch of regular layers that have their parameters shared between filters.<\/p>\n\n<p>The trouble is, even though this is fine memory-wise with a very manageable overhead, we're talking <strong>inputwidth<\/strong>^<strong>inputheight<\/strong>^<strong>inputdepth<\/strong>^<strong>outputdepth<\/strong> mini-networks, here. Maybe there's some way to create massive \"long and tall\" networks that work on the entire replicated input set at once, but how do I create layers that are partially-connected (like convolutions) instead of fully-connected?<\/p>\n\n<p>I would have liked to just use inheritance to create a special copy of the regular <strong>SpatialConvolution \"class\"<\/strong> and modify it, but I can't even try because it's implemented in an external C library. I can't just use regular layers before a regular SpatialConvolution layer because I need to do my math with different weights and biases for each filter (shared between applications of the same filter to different input coordinates).<\/p>\n","answers":[{"AnswerId":"32357456","CreationDate":"2015-09-02T15:39:43.890","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>Good question. You made me give some serious thought.\nYour approach has a flaw: it does not allow to take advantage of vectorized computations since each mini-network works independently. <\/p>\n\n<p><strong>My idea is as follows:<\/strong><\/p>\n\n<p>Suppose network's <code>input<\/code> and <code>output<\/code> are 2D tensors. We can produce (efficiently, without memory copying) an auxiliary 4D tensor \n<code>rf_input (kernel_size x kernel_size x output_h x output_w)<\/code> \nsuch that <code>rf_input[:, :, k, l]<\/code> is a 2D tensor of size <code>kernel_size x kernel_size<\/code> containing a receptive field which <code>output[k, l]<\/code> will be gotten from. Then we iterate over positions inside the kernel <code>rf_input[i, j, :, :]<\/code> getting pixels at position <code>(i, j)<\/code> inside all receptive fields and computing their contribution to each <code>output[k, l]<\/code> at once using vectorization.<\/p>\n\n<p><strong>Example:<\/strong><\/p>\n\n<p>Let our \"convolving\" function be, for example, a product of tangents of sums:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hRSLt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hRSLt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Then its partial derivative w.r.t. the <code>input<\/code> pixel at position <code>(s,t)<\/code> in its receptive field is<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pP8U1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pP8U1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Derivative w.r.t. <code>weight<\/code> is the same.<\/p>\n\n<p>At the end, of course, we must sum up gradients from different <code>output[k,l]<\/code> points. For example, each <code>input[m, n]<\/code> contributes to at most <code>kernel_size^2<\/code> outputs as a part of their receptive fields, and each <code>weight[i, j]<\/code> contributes to all <code>output_h x output_w<\/code> outputs.<\/p>\n\n<p><strong>Simple implementation may look like this:<\/strong><\/p>\n\n<pre><code>require 'nn'\nlocal CustomConv, parent = torch.class('nn.CustomConv', 'nn.Module')\n\n-- This module takes and produces a 2D map. \n-- To work with multiple input\/output feature maps and batches, \n-- you have to iterate over them or further vectorize computations inside the loops.\n\nfunction CustomConv:__init(ker_size)\n    parent.__init(self)\n\n    self.ker_size = ker_size\n    self.weight = torch.rand(self.ker_size, self.ker_size):add(-0.5)\n    self.gradWeight = torch.Tensor(self.weight:size()):zero()\nend\n\nfunction CustomConv:_get_recfield_input(input)\n    local rf_input = {}\n    for i = 1, self.ker_size do\n        rf_input[i] = {}\n        for j = 1, self.ker_size do\n            rf_input[i][j] = input[{{i, i - self.ker_size - 1}, {j, j - self.ker_size - 1}}]\n        end\n    end\n    return rf_input\nend\n\nfunction CustomConv:updateOutput(_)\n    local output = torch.Tensor(self.rf_input[1][1]:size())\n    --  Kernel-specific: our kernel is multiplicative, so we start with ones\n    output:fill(1)                                              \n    --\n    for i = 1, self.ker_size do\n        for j = 1, self.ker_size do\n            local ker_pt = self.rf_input[i][j]:clone()\n            local w = self.weight[i][j]\n            -- Kernel-specific\n            output:cmul(ker_pt:add(w):tan())\n            --\n        end\n    end\n    return output\nend\n\nfunction CustomConv:updateGradInput_and_accGradParameters(_, gradOutput)\n    local gradInput = torch.Tensor(self.input:size()):zero()\n    for i = 1, self.ker_size do\n        for j = 1, self.ker_size do\n            local ker_pt = self.rf_input[i][j]:clone()\n            local w = self.weight[i][j]\n            -- Kernel-specific\n            local subGradInput = torch.cmul(gradOutput, torch.cdiv(self.output, ker_pt:add(w):tan():cmul(ker_pt:add(w):cos():pow(2))))\n            local subGradWeight = subGradInput\n            --\n            gradInput[{{i, i - self.ker_size - 1}, {j, j - self.ker_size - 1}}]:add(subGradInput)\n            self.gradWeight[{i, j}] = self.gradWeight[{i, j}] + torch.sum(subGradWeight)\n        end\n    end\n    return gradInput\nend\n\nfunction CustomConv:forward(input)\n    self.input = input\n    self.rf_input = self:_get_recfield_input(input)\n    self.output = self:updateOutput(_)\n    return self.output\nend\n\nfunction CustomConv:backward(input, gradOutput)\n    gradInput = self:updateGradInput_and_accGradParameters(_, gradOutput)\n    return gradInput\nend\n<\/code><\/pre>\n\n<p>If you change this code a bit: <\/p>\n\n<pre><code>updateOutput:                                             \n    output:fill(0)\n    [...]\n    output:add(ker_pt:mul(w))\n\nupdateGradInput_and_accGradParameters:\n    local subGradInput = torch.mul(gradOutput, w)\n    local subGradWeight = torch.cmul(gradOutput, ker_pt)\n<\/code><\/pre>\n\n<p>then it will work exactly as <code>nn.SpatialConvolutionMM<\/code> with zero <code>bias<\/code> (I've tested it).<\/p>\n"}]}
{"QuestionId":32324961,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-09-01T06:30:59.643","AcceptedAnswerId":null,"OwnerUserId":2293069.0,"Title":"Torch nngraph building nodes with nn.Linear","Body":"<p>Hi I am new in torch\/lua and I am doing the <a href=\"http:\/\/(https:\/\/www.cs.ox.ac.uk\/people\/nando.defreitas\/machinelearning\/practicals\/practical5.pdf\" rel=\"nofollow\">practical5<\/a> of the oxford machine learning course: <\/p>\n\n<p>What I am trying to implement is a simple layer:\nm = x1 + x2 cmul linear(x3)\nwhere cmul is the element multiply and linear is just a linear layer.<\/p>\n\n<p>My code looks like:<\/p>\n\n<pre><\/pre>\n\n<p>However, I got the error message:<\/p>\n\n<pre><\/pre>\n\n<p>And if I only want\n    aa = nn.Linear(10,20)()<\/p>\n\n<p>I got same error as above.<\/p>\n\n<p>Even if I follow the <a href=\"https:\/\/github.com\/torch\/nngraph\/#two-hidden-layers-mlp\" rel=\"nofollow\">example<\/a> on the torch github: <\/p>\n\n<p>I got the same error.<\/p>\n\n<p><strong>UPDATE<\/strong> and SOLVED:<\/p>\n\n<p>I miss import the package. Although  and  are both referred to as  in the code, they are actually different packages.<\/p>\n\n<p>Should do<\/p>\n\n<pre><\/pre>\n\n<p>And I only did<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39314203","CreationDate":"2016-09-04T06:36:28.430","ParentId":null,"OwnerUserId":"6792483","Title":null,"Body":"<p>Last argument in this line should be x3 instead of x1:<\/p>\n\n<pre><code>l3 = nn.Linear(params.x3_size1, params.x3_size2)(x3)\n<\/code><\/pre>\n"}]}
{"QuestionId":32337591,"AnswerCount":4,"Tags":"<lua><luajit><torch>","CreationDate":"2015-09-01T17:27:44.140","AcceptedAnswerId":null,"OwnerUserId":3113501.0,"Title":"How catch ctrl-c in lua when ctrl-c is sent via the command line","Body":"<p>I would like to know when the user from a command line presses control-c so I can save some stuff. <\/p>\n\n<p>How do I do this? I've looked but haven't really seen anything.<\/p>\n\n<p>Note: I'm somewhat familiar with lua, but I'm no expert. I mostly use lua to use the library Torch (<a href=\"http:\/\/torch.ch\/\">http:\/\/torch.ch\/<\/a>)<\/p>\n","answers":[{"AnswerId":"32409607","CreationDate":"2015-09-05T04:50:48.193","ParentId":null,"OwnerUserId":"5129715","Title":null,"Body":"<p><a href=\"https:\/\/msdn.microsoft.com\/en-us\/library\/windows\/desktop\/ms685049(v=vs.85).aspx\" rel=\"nofollow\">windows : SetConsoleCtrlHandler<\/a><\/p>\n\n<p><a href=\"http:\/\/www.gnu.org\/software\/libc\/manual\/html_node\/Basic-Signal-Handling.html\" rel=\"nofollow\">linux : signal<\/a><\/p>\n\n<p>There are two behaviors of the signal which are undesirable, which will cause complexities in the code.<\/p>\n\n<ol>\n<li>Program termination<\/li>\n<li>Broken IO<\/li>\n<\/ol>\n\n<p>The first behavior can be caught and remembered in a C program by using SetConsoleCtrlHandler\/signal.  This will allow your function to be called, and you can remember that the system needs to shutdown.  Then at some point in the lua code you see it has happened (call to check), and perform your tidy up and shutdown.<\/p>\n\n<p>The second behavior, is that a blocking operation (read\/write) will be cancelled by the signal, and the operation will be unfinished.  That would need to be checked at each IO event, and then re-started, or cancelled as appropriate.<\/p>\n"},{"AnswerId":"32471856","CreationDate":"2015-09-09T05:54:17.147","ParentId":null,"OwnerUserId":"2328287","Title":null,"Body":"<p>There exists io libraries that support this.\nI know zmq and libuv<\/p>\n\n<p>Libuv example with lluv binding - <a href=\"https:\/\/github.com\/moteus\/lua-lluv\/blob\/master\/examples\/sig.lua\" rel=\"nofollow\">https:\/\/github.com\/moteus\/lua-lluv\/blob\/master\/examples\/sig.lua<\/a><\/p>\n\n<p>ZeroMQ return EINTR from poll function when user press Ctrl-C<\/p>\n\n<p>But I do not handle thi byself<\/p>\n"},{"AnswerId":"32598961","CreationDate":"2015-09-16T02:51:33.307","ParentId":null,"OwnerUserId":"4808555","Title":null,"Body":"<pre><code>require('sys')\nsys.catch_ctrl_c()\n<\/code><\/pre>\n\n<p>I use this to catch the exit from cli.<\/p>\n"},{"AnswerId":"34409274","CreationDate":"2015-12-22T05:52:06.407","ParentId":null,"OwnerUserId":"320911","Title":null,"Body":"<p>Implementing a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Unix_signal#SIGINT\" rel=\"noreferrer\"><code>SIGINT<\/code><\/a> handler is straightforward using the excellent <a href=\"https:\/\/luarocks.org\/modules\/gvvaughan\/luaposix\" rel=\"noreferrer\">luaposix<\/a> library:<\/p>\n\n<pre><code>local signal = require(\"posix.signal\")\n\nsignal.signal(signal.SIGINT, function(signum)\n  io.write(\"\\n\")\n  -- put code to save some stuff here\n  os.exit(128 + signum)\nend)\n<\/code><\/pre>\n\n<p>Refer to the <a href=\"https:\/\/luaposix.github.io\/luaposix\/modules\/posix.signal.html\" rel=\"noreferrer\">posix.signal<\/a> module's API documentation for more information.<\/p>\n"}]}
{"QuestionId":32339193,"AnswerCount":0,"Tags":"<python><python-2.7><theano><keras>","CreationDate":"2015-09-01T19:11:03.393","AcceptedAnswerId":null,"OwnerUserId":4513062.0,"Title":"cannot import name Cop from keras.models python","Body":"<p>Greeting Dear Community,<\/p>\n\n<p>I try to use the python keras package and I got this error:\nI am running this on a oracle-linux virtual box. Is it something to do \nit expects some kind of GPU box ?<\/p>\n\n<p>Thx for the  help.<\/p>\n\n<pre><\/pre>\n\n<p>The line from dnn.py<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":32341493,"AnswerCount":1,"Tags":"<python><machine-learning><theano><cox-regression>","CreationDate":"2015-09-01T21:53:55.767","AcceptedAnswerId":null,"OwnerUserId":2723734.0,"Title":"Negative log likelihood in theano (cox regression)","Body":"<p>I'm trying to implement cox regression in theano.    <\/p>\n\n<p>I'm using the logistic regression tutorial (<a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a>) as a framework and replacing the logistic log likelihood (LL) function by the cox regression LL function (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Proportional_hazards_model#The_partial_likelihood\" rel=\"nofollow\">https:\/\/en.wikipedia.org\/wiki\/Proportional_hazards_model#The_partial_likelihood<\/a>).  <\/p>\n\n<p>Here's what I have so far:<\/p>\n\n<pre><\/pre>\n\n<p>Basically, I need to sum over LL_i (where i is 0 to ytime.shape - 1).  But I'm not sure how to do this.  Should I use the scan function?  <\/p>\n","answers":[{"AnswerId":"32343254","CreationDate":"2015-09-02T01:20:56.750","ParentId":null,"OwnerUserId":"2723734","Title":null,"Body":"<p>Figured it out.  The trick was not to use the scan function, but to convert the double summation to a pure matrix operation.  <\/p>\n"}]}
{"QuestionId":32342672,"AnswerCount":1,"Tags":"<machine-learning><image-recognition><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-09-01T23:59:53.453","AcceptedAnswerId":null,"OwnerUserId":933728.0,"Title":"Convolutional Neural Network - Pretrained model with fast feature extraction","Body":"<p>What is the fastest pretrained network I can get for image recognition?<br>\nThe biggest source is probably <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo\" rel=\"nofollow\">Model Zoo<\/a>, but I cannot get precise information about feature extraction time.<\/p>\n\n<p>Right now I am using <a href=\"https:\/\/gist.github.com\/ksimonyan\/fd8800eeb36e276cd6f9#file-readme-md\" rel=\"nofollow\">VGG_CNN_S<\/a> model which gives mi around 80ms for processing 1 image.<\/p>\n\n<p>As far as I know you can get down to 1 ms for one image.<\/p>\n\n<p>Which pretrained network available for download is the fastest in extracting features?<\/p>\n","answers":[{"AnswerId":"32344752","CreationDate":"2015-09-02T04:43:28.227","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Good and efficient deep architectures are introduced quite frequently, so I suppose any answer you'll get here will have a short expiration date.<\/p>\n\n<p>Regarding \"feature extraction time\", I suppose you mean the duration of the forward pass only - you are not interested in training time. This time will also depend on the layer which you want to extract: the deeper the layer, the longer the time (for the same net) since it requires more computations to get deeper into any specific net. However, for different nets it often takes different times to reach the same \"depth\" since the computations at each depth are different.<\/p>\n\n<p>Nevertheless, roughly when Oxford VGG lab intorduced VGG_CNN_S, Google labs came up with <a href=\"http:\/\/vision.princeton.edu\/pvt\/GoogLeNet\/\" rel=\"nofollow\">GoogLeNet<\/a>: it is a very deep architecture for recognition, but with an extra effort on keeping the computational burden within reason. It's worth giving it a try. <\/p>\n"}]}
{"QuestionId":32343282,"AnswerCount":1,"Tags":"<c++><machine-learning><deep-learning><caffe><glog>","CreationDate":"2015-09-02T01:24:02.350","AcceptedAnswerId":32346978.0,"OwnerUserId":1377127.0,"Title":"What does \"InitGoogleLogging\" do?","Body":"<p>I've been modifying an <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\" rel=\"noreferrer\">example C++ program<\/a> from the Caffe deep learning library and I noticed this code on <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp#L234\" rel=\"noreferrer\">line 234<\/a> that doesn't appear to be referenced again.<\/p>\n\n<pre><\/pre>\n\n<p>The argument provided is a prototxt file which defines the parameters of the deep learning model I'm calling. The thing that is confusing me is where the results from this line go? I know they end up being used in the program because if I make a mistake in the prototxt file then the program will crash. However I'm struggling to see how the data is passed to the class performing the classification tasks.<\/p>\n","answers":[{"AnswerId":"32346978","CreationDate":"2015-09-02T07:25:28.993","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>First of all, <code>argv[0]<\/code> is <em>not<\/em> the first argument you pass to your executable, but rather the <a href=\"https:\/\/stackoverflow.com\/a\/2051031\/1714410\">executable name<\/a>. So you are passing to <code>::google::InitGoogleLogging<\/code> the executable name and not the prototxt file.<br>\n<code>'glog'<\/code> module (google logging) is using this name to decorate the log entries it outputs.<\/p>\n\n<p>Second, caffe is using google logging (<a href=\"https:\/\/google-glog.googlecode.com\/svn\/trunk\/\" rel=\"nofollow noreferrer\">aka <code>'glog'<\/code><\/a>) as its logging module, and hence this module must be initialized once when running caffe. This is why you have this <\/p>\n\n<pre><code>::google::InitGoogleLogging(argv[0]);\n<\/code><\/pre>\n\n<p>in your code.<\/p>\n"}]}
{"QuestionId":32346763,"AnswerCount":1,"Tags":"<python-2.7><cuda><gpu><theano><lasagne>","CreationDate":"2015-09-02T07:14:06.923","AcceptedAnswerId":32347299.0,"OwnerUserId":3006092.0,"Title":"Running Python lasagne with CUDNN in Ubuntu 14.04 Linux","Body":"<p>I am on a Linux 3.16.0-30-generic #40~14.04.1-Ubuntu x86_64 GNU\/Linux machine with NVIDIA Corporation GF108 [GeForce GT 430] [10de:0de1] (rev a1) graphic card. <\/p>\n\n<p>I am trying to run  package with GPU enabled, however by running \n, I get an ImportError:<\/p>\n\n<pre><\/pre>\n\n<p>Currently, I have installed and successfully compiled all the  samples as well as .<\/p>\n\n<pre><\/pre>\n\n<p>I get this output:<\/p>\n\n<pre><\/pre>\n\n<p>I am using  and used  to install all these packages from their  files, which are downloaded from their repositories.<\/p>\n","answers":[{"AnswerId":"32347299","CreationDate":"2015-09-02T07:43:10.537","ParentId":null,"OwnerUserId":"3006092","Title":null,"Body":"<p>I finally figured out the problem! My system's GPU cannot\/is not supported by <code>cudNN<\/code>, unfortunately. :(<\/p>\n"}]}
{"QuestionId":32348618,"AnswerCount":0,"Tags":"<pandas><machine-learning><neural-network><theano><deep-learning>","CreationDate":"2015-09-02T08:49:55.473","AcceptedAnswerId":null,"OwnerUserId":5197007.0,"Title":"Load data for RNN","Body":"<p>In RNN training examples, I noticed input data and target data are all 3-dimension arrays, and need to define time-step delay between input and output.<\/p>\n\n<pre><\/pre>\n\n<p>I wanted to load custom data for RNN training - input vector=1, output vector =1, time_steps =1 (see attached data1a.csv). Reshape doesn't work here. Could anyone please illustrate how to do it?<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n\n<p>data link:\n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/--X2zGSd8rw\" rel=\"nofollow\">links<\/a>\n<a href=\"https:\/\/groups.google.com\/group\/theano-users\/attach\/28565ada37494d\/data1a.csv?part=0.1&amp;authuser=0\" rel=\"nofollow\">data1a.csv<\/a><\/p>\n\n<p>I'm just getting some ideas about it, but don't know how to continue:<\/p>\n\n<pre><\/pre>\n\n<p>What is the next move? <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":32352820,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2015-09-02T12:12:09.343","AcceptedAnswerId":null,"OwnerUserId":416645.0,"Title":"Caffe MemoryData Layer and Solver inteface","Body":"<p>I am trying to train a network but rather using lmdb or leveldb, I am feeding data to my network on the fly. So I am following procedure as outlined below<\/p>\n\n<ol>\n<li>My data is loaded in Memory Data Layer. <\/li>\n<li>I create a mini batch using a python script. <\/li>\n<li>Set data and label as  <\/li>\n<li>After that I call <\/li>\n<\/ol>\n\n<p>Here solver is of type SGDSolver. Now my question is what is the difference between  and ? <\/p>\n\n<p>2ndly this approach doesn't let me have a memory data layer for test network. Is there any work around for that?<\/p>\n\n<p>My solver.prototxt looks like<\/p>\n\n<pre><\/pre>\n\n<p>With my approach every 20th iteration network displays some output loss etc. And somehow loss stays constant over some numbert of iterations, what could be the reason for that.<\/p>\n","answers":[{"AnswerId":"39089447","CreationDate":"2016-08-22T22:05:15.367","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<ol>\n<li>What is the difference between solver.solve and solver.step?<\/li>\n<\/ol>\n\n<p><strong>solve<\/strong> does the entire training run, to whatever limits you've set -- usually the iteration limit.  <strong>step<\/strong> does only the specified number of iterations.<\/p>\n\n<ol start=\"2\">\n<li>How do I get a memory data layer for test network?<\/li>\n<\/ol>\n\n<p>If you're not reading from a supported data channel \/ format, I <em>think<\/em> you have to write a customer input routine (your <strong>data<\/strong> package).<\/p>\n\n<ol start=\"3\">\n<li>Loss stays constant over multiple iterations, what could be the reason?<\/li>\n<\/ol>\n\n<p>There are several possibilities, depending on the surrounding effects.  If the loss shows only one value ever, then your back propagation is a likely culprit.  Perhaps you're not properly connected to the data set, and you're not getting the expected classifications fed in.<\/p>\n\n<p>If the loss has a temporary stability but then converges decently, don't worry about it; this is likely an effect of training ordering.<\/p>\n\n<p>If the loss declines decently and then settles at a fixed value, then you're also doing well: the training converged before it ran out of iterations.<\/p>\n"}]}
{"QuestionId":32353509,"AnswerCount":3,"Tags":"<caffe><openblas>","CreationDate":"2015-09-02T12:43:22.147","AcceptedAnswerId":null,"OwnerUserId":4996964.0,"Title":"\"\/usr\/bin\/ld: cannot find -lopenblas\" error in Caffe compilation","Body":"<p>When I was compiling Caffe, I had this error, despite OpenBLAS is installed:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a solution for it?<\/p>\n","answers":[{"AnswerId":"36215963","CreationDate":"2016-03-25T07:44:15.020","ParentId":null,"OwnerUserId":"1511337","Title":null,"Body":"<p>I saw the similar problem (I'm compiling caffe again for some reason).\nI found the library file the builder is looking for (-lcblas or -latlas means libcblas.so and libatlas.so) are under \/usr\/lib64\/atlas. So just added symbolic links under \/usr\/lib64 like this.<\/p>\n\n<pre><code>sudo ln \/usr\/lib64\/atlas\/libcblas.so.3.0 \/usr\/lib64\/libcblas.so\nsudo ln -s \/usr\/lib64\/atlas\/libatlas.so.3.0 \/usr\/lib64\/libatlas.so\n<\/code><\/pre>\n\n<p>But I guess more proper method is to set Makefile.config (the CBLAS path). (I thought the default path will do away with it reading the comment saying so, but it did not.) Hope this helps anyone.<\/p>\n"},{"AnswerId":"37448511","CreationDate":"2016-05-25T22:11:36.693","ParentId":null,"OwnerUserId":"5043076","Title":null,"Body":"<p>Including the base packs even after cloning OpenBlas and making will link the appropriate libraries in 14.04 and 16.<\/p>\n\n<pre><code>apt install liblapack-dev liblapack3 libopenblas-base libopenblas-dev\n<\/code><\/pre>\n\n<p>apt install liblapack-dev liblapack3 libopenblas-base libopenblas-dev<\/p>\n"},{"AnswerId":"34243307","CreationDate":"2015-12-12T18:17:05.440","ParentId":null,"OwnerUserId":"2807033","Title":null,"Body":"<p>I faced the same problem. Even adding library directory \"\/opt\/OpenBLAS\/lib\/\" to ldconfig cache didn't help (as my libopenblas.so is at \"\/opt\/OpenBLAS\/lib\/libopenblas.so\").<\/p>\n\n<p>Using cmake helped me. Try this from caffe root directory:<\/p>\n\n<p><code>mkdir build\ncd build\ncmake -DBLAS=open ..\nmake all\nmake runtest<\/code><\/p>\n\n<p>If you need to use make, add the symlink of libopenblas.so to \/usr\/lib. I did the following:<\/p>\n\n<p><code>ln -s \/opt\/OpenBLAS\/lib\/libopenblas.so \/usr\/lib\/libopenblas.so<\/code><\/p>\n"}]}
{"QuestionId":32368577,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-09-03T06:39:38.553","AcceptedAnswerId":null,"OwnerUserId":5295427.0,"Title":"How do I use a pre-trained Caffe model?","Body":"<p>I have some questions about how to actually interact with a pre-trained Caffe model. In my case I'm using a model for <a href=\"http:\/\/places.csail.mit.edu\" rel=\"nofollow\">scene recognition<\/a>.<\/p>\n\n<p>In the caffe git repository, there are some code examples in Python and C++ on the implementations of Image Classifiers. However, those do not apply to my use case (since they only classify the input image as ONE class).<\/p>\n\n<p>My goal is an application that takes an input image (jpg) and outputs the highest predicted class label for each pixel in the input image (e.i., indices for sky, beach, road, car).<\/p>\n\n<p>Could anyone give me some pointers on how to proceed? <\/p>\n\n<p>There already seem to exist implementations for this. This demo (<a href=\"http:\/\/places.csail.mit.edu\/demo.html\" rel=\"nofollow\">http:\/\/places.csail.mit.edu\/demo.html<\/a>) is kind of what I what. <\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"32373937","CreationDate":"2015-09-03T11:07:32.110","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>What you are looking for is not image classification, but rather <em>semantic segmentation<\/em>.<\/p>\n\n<p>A recent work, by Jonathan Long, Evan Shelhamer and Trevor Darrell is based on Caffe, and can be found <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo#fcn\" rel=\"nofollow\">here<\/a>. It uses fully convolutional network, that is, a network with no <code>\"InnerProduct\"<\/code> layers only convolutional layers, thus capable of producing outputs with different sizes for different sizes of inputs.<\/p>\n"}]}
{"QuestionId":32379841,"AnswerCount":2,"Tags":"<lua><torch><cmdline-args>","CreationDate":"2015-09-03T15:30:23.690","AcceptedAnswerId":32380948.0,"OwnerUserId":49985.0,"Title":"Lua cmd-line string","Body":"<p>Say I call Lua with this cmd: <\/p>\n\n<pre class=\"lang-bsh prettyprint-override\"><\/pre>\n\n<p>How can I access this same cmd-line string from Lua?<\/p>\n\n<p>I know about the  table, but it removes all quotes from the original command string making it difficult to reconstruct:<\/p>\n\n<pre><\/pre>\n\n<p>If I can save the exact string to a file from within Lua, I can easily call it again later.<\/p>\n","answers":[{"AnswerId":"32380948","CreationDate":"2015-09-03T16:26:55.833","ParentId":null,"OwnerUserId":"2726734","Title":null,"Body":"<p>@peterpi is correct that the shell is interpreting the command and as a result stripping away the quotes. However, reconstructing the command exactly is not really necessary to have the shell interpret the command the same way as before.<\/p>\n\n<p>For simple cases concatenating the arguments to the script is often enough:<\/p>\n\n<pre><code>local command = table.concat(arg, ' ', -1, #arg)\n<\/code><\/pre>\n\n<p>This will fail if the quotes are actually necessary, most commonly  when an argument contains a space or shell character, so quoting everything is easy and somewhat more robust, but not pretty.<\/p>\n\n<p>Here is an example with a Lua pattern to check for special (bash) shell characters and spaces to decide if and which quotes are necessary. It may not be complete but it handles filenames, most strings, and numbers as arguments.<\/p>\n\n<pre><code>local mod_arg = { }\nfor k, v in pairs(arg) do\n    if v:find\"'\" then\n      mod_arg[k] = '\"'..v..'\"'\n    elseif v:find'[%s$`&gt;&lt;|#]' then\n      mod_arg[k] = \"'\"..v..\"'\"         \n    else\n      mod_arg[k] = v\n    end\nend \nlocal command = table.concat(mod_arg, ' ', -1, #mod_arg)\nprint(command)\n<\/code><\/pre>\n"},{"AnswerId":"32380157","CreationDate":"2015-09-03T15:46:34.640","ParentId":null,"OwnerUserId":"819046","Title":null,"Body":"<p>No doubt somebody will prove me wrong, but generally I don't think this is possible.  It's the <em>shell<\/em> rather than luajit that takes the quotes away and chops the line up into individual tokens.<\/p>\n"}]}
{"QuestionId":32379878,"AnswerCount":2,"Tags":"<ipython-notebook><caffe><pycaffe>","CreationDate":"2015-09-03T15:32:43.747","AcceptedAnswerId":32509003.0,"OwnerUserId":2191652.0,"Title":"Cheat sheet for caffe \/ pycaffe?","Body":"<p>Does anyone know whether there is a cheat sheet for all important pycaffe commands?\nI was so far using caffe only via Matlab interface and terminal + bash scripts.<\/p>\n\n<p>I wanted to shift towards using ipython and work through the ipython notebook examples. However I find it hard to get an overview of all the functions that are inside the caffe module for python. (I'm also quite new to python).<\/p>\n","answers":[{"AnswerId":"32509003","CreationDate":"2015-09-10T18:16:05.333","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/python\/caffe\/test\">pycaffe tests<\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/_caffe.cpp\">this file<\/a> are the main gateway to the python coding interface.<\/p>\n\n<p>First of all, you would like to choose whether to use Caffe with CPU or GPU. It is sufficient to call <code>caffe.set_mode_cpu()<\/code> or <code>caffe.set_mode_gpu()<\/code>, respectively.<\/p>\n\n<h2>Net<\/h2>\n\n<p>The main class that the pycaffe interface exposes is the <code>Net<\/code>. It has two constructors:<\/p>\n\n<pre><code>net = caffe.Net('\/path\/prototxt\/descriptor\/file', caffe.TRAIN)\n<\/code><\/pre>\n\n<p>which simply create a <code>Net<\/code> (in this case using the <em>Data Layer<\/em> specified for training), or<\/p>\n\n<pre><code>net = caffe.Net('\/path\/prototxt\/descriptor\/file', '\/path\/caffemodel\/weights\/file', caffe.TEST)\n<\/code><\/pre>\n\n<p>which creates a <code>Net<\/code> and automatically loads the weights as saved in the provided <em>caffemodel<\/em> file - in this case using the <em>Data Layer<\/em> specified for testing.<\/p>\n\n<p>A <code>Net<\/code> object has several attributes and methods. They can be found <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/pycaffe.py\">here<\/a>. I will cite just the ones I use more often.<\/p>\n\n<p>You can access the network blobs by means of <code>Net.blobs<\/code>. E.g.<\/p>\n\n<pre><code>data = net.blobs['data'].data\nnet.blobs['data'].data[...] = my_image\nfc7_activations = net.blobs['fc7'].data\n<\/code><\/pre>\n\n<p>You can access the parameters (weights) too, in a similar way. E.g.<\/p>\n\n<pre><code>nice_edge_detectors = net.params['conv1'].data\nhigher_level_filter = net.params['fc7'].data\n<\/code><\/pre>\n\n<p>Ok, now it's time to actually feed the net with some data. So, you will use <code>backward()<\/code> and <code>forward()<\/code> methods. So, if you want to classify a single image<\/p>\n\n<pre><code>net.blobs['data'].data[...] = my_image\nnet.forward() # equivalent to net.forward_all()\nsoftmax_probabilities = net.blobs['prob'].data\n<\/code><\/pre>\n\n<p>The <code>backward()<\/code> method is equivalent, if one is interested in computing gradients.<\/p>\n\n<p>You can save the net weights to subsequently reuse them. It's just a matter of<\/p>\n\n<pre><code> net.save('\/path\/to\/new\/caffemodel\/file')\n<\/code><\/pre>\n\n<h2>Solver<\/h2>\n\n<p>The other core component exposed by pycaffe is the <code>Solver<\/code>. There are several types of solver, but I'm going to use only <code>SGDSolver<\/code> for the sake of clarity. It is needed in order to train a caffe model.\nYou can instantiate the solver with<\/p>\n\n<pre><code>solver = caffe.SGDSolver('\/path\/to\/solver\/prototxt\/file')\n<\/code><\/pre>\n\n<p>The <code>Solver<\/code> will encapsulate the network you are training and, if present, the network used for testing. Note that they are usually the same network, only with a different <em>Data Layer<\/em>. The networks are accessible with<\/p>\n\n<pre><code> training_net = solver.net\n test_net = solver.test_nets[0] # more than one test net is supported\n<\/code><\/pre>\n\n<p>Then, you can perform a solver iteration, that is, a forward\/backward pass with weight update, typing just<\/p>\n\n<pre><code> solver.step(1)\n<\/code><\/pre>\n\n<p>or run the solver until the last iteration, with<\/p>\n\n<pre><code> solver.solve()\n<\/code><\/pre>\n\n<h2>Other features<\/h2>\n\n<p>Note that pycaffe allows you to do more stuff, such as <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_net_spec.py\">specifying the network architecture through a Python class<\/a> or <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_python_layer.py\">creating a new <em>Layer<\/em> type<\/a>.\nThese features are less often used, but they are pretty easy to understand by reading the test cases.<\/p>\n"},{"AnswerId":"38427997","CreationDate":"2016-07-18T02:59:08.387","ParentId":null,"OwnerUserId":"4481686","Title":null,"Body":"<p>Please note that the answer by Flavio Ferrara has a litte problem which may cause you waste a lot of time:<\/p>\n\n<pre><code>net.blobs['data'].data[...] = my_image\nnet.forward()\n<\/code><\/pre>\n\n<p>The code above is noneffective if your first layer is a Data type layer, because when <code>net.forward()<\/code> is called, it will begin from the first layer, and then your inserted data <code>my_image<\/code> will be covered. So it will show no error but give you totally irrelevant output. The correct way is to assign the start and end layer, for example:<\/p>\n\n<p><code>net.forward(start='conv1', end='fc')<\/code><\/p>\n\n<p>Here is a Github repository of Face Verification Experiment on LFW Dataset, using pycaffe and some matlab code. I guess it could help a lot, especially the <code>caffe_ftr.py<\/code> file.<\/p>\n\n<p><a href=\"https:\/\/github.com\/AlfredXiangWu\/face_verification_experiment\">https:\/\/github.com\/AlfredXiangWu\/face_verification_experiment<\/a><\/p>\n\n<p>Besides, here are some short example code of using pycaffe for image classification\uff1a<\/p>\n\n<p><a href=\"http:\/\/codrspace.com\/Jaleyhd\/caffe-python-tutorial\/\">http:\/\/codrspace.com\/Jaleyhd\/caffe-python-tutorial\/<\/a>\n<a href=\"http:\/\/prog3.com\/sbdm\/blog\/u011762313\/article\/details\/48342495\">http:\/\/prog3.com\/sbdm\/blog\/u011762313\/article\/details\/48342495<\/a><\/p>\n"}]}
{"QuestionId":32389905,"AnswerCount":2,"Tags":"<c++><linux><posix><caffe><sigaction>","CreationDate":"2015-09-04T04:08:40.777","AcceptedAnswerId":32390101.0,"OwnerUserId":2467772.0,"Title":"Sigaction and porting Linux code to Windows","Body":"<p>I am trying to port <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"noreferrer\">caffe<\/a> (developed for Linux) source code to Windows environment. The problem is at  structure at  and . The source codes are shown below.\nMy query is which library or code replacement can be done to make this  works in Windows.<\/p>\n\n<p>\/\/\/Header file       <\/p>\n\n<pre><\/pre>\n\n<p>\/\/\/Source file                  <\/p>\n\n<pre><\/pre>\n\n<p>The errors are <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34414651","CreationDate":"2015-12-22T11:17:10.290","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>Based on @nneonneo:<\/p>\n\n<pre><code>void handle_signal(int signal) {\n    switch (signal) {\n#ifdef _WIN32\n    case SIGTERM:\n    case SIGABRT:\n    case SIGBREAK:\n#else\n    case SIGHUP:\n#endif\n      got_sighup = true;\n      break;\n    case SIGINT:\n      got_sigint = true;\n      break;\n    }\n  }\nvoid HookupHandler() {\n    if (already_hooked_up) {\n      LOG(FATAL) &lt;&lt; \"Tried to hookup signal handlers more than once.\";\n    }\n    already_hooked_up = true;\n#ifdef _WIN32\n    signal(SIGINT, handle_signal);\n    signal(SIGTERM, handle_signal);\n    signal(SIGABRT, handle_signal);\n#else\n    struct sigaction sa;\n    \/\/ Setup the handler\n    sa.sa_handler = &amp;handle_signal;\n    \/\/ Restart the system call, if at all possible\n    sa.sa_flags = SA_RESTART;\n    \/\/ Block every signal during the handler\n    sigfillset(&amp;sa.sa_mask);\n    \/\/ Intercept SIGHUP and SIGINT\n    if (sigaction(SIGHUP, &amp;sa, NULL) == -1) {\n      LOG(FATAL) &lt;&lt; \"Cannot install SIGHUP handler.\";\n    }\n    if (sigaction(SIGINT, &amp;sa, NULL) == -1) {\n      LOG(FATAL) &lt;&lt; \"Cannot install SIGINT handler.\";\n    }\n#endif\n  }\n void UnhookHandler() {\n    if (already_hooked_up) {\n#ifdef _WIN32\n        signal(SIGINT, SIG_DFL);\n        signal(SIGTERM, SIG_DFL);\n        signal(SIGABRT, SIG_DFL);\n#else\n      struct sigaction sa;\n      \/\/ Setup the sighub handler\n      sa.sa_handler = SIG_DFL;\n      \/\/ Restart the system call, if at all possible\n      sa.sa_flags = SA_RESTART;\n      \/\/ Block every signal during the handler\n      sigfillset(&amp;sa.sa_mask);\n      \/\/ Intercept SIGHUP and SIGINT\n      if (sigaction(SIGHUP, &amp;sa, NULL) == -1) {\n        LOG(FATAL) &lt;&lt; \"Cannot uninstall SIGHUP handler.\";\n      }\n      if (sigaction(SIGINT, &amp;sa, NULL) == -1) {\n        LOG(FATAL) &lt;&lt; \"Cannot uninstall SIGINT handler.\";\n      }\n#endif\n\n      already_hooked_up = false;\n    }\n  }\n<\/code><\/pre>\n"},{"AnswerId":"32390101","CreationDate":"2015-09-04T04:30:06.827","ParentId":null,"OwnerUserId":"1204143","Title":null,"Body":"<p><code>sigaction<\/code> is part of the UNIX signals API. Windows provides only <code>signal<\/code>, which doesn't support <code>SIGHUP<\/code> or any flags (such as <code>SA_RESTART<\/code>). However, the very basic support is still there, so the code should still work reasonably correctly if you use just <code>signal<\/code> (and not <code>sigaction<\/code>).<\/p>\n"}]}
{"QuestionId":32397226,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-09-04T11:42:07.813","AcceptedAnswerId":32397924.0,"OwnerUserId":2625087.0,"Title":"Theano not running on Windows","Body":"<p>I am trying to install theano on windows using python 3.4. I am following this instruction <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow noreferrer\">Theano on Windows<\/a><\/p>\n\n<p>Its creating one file called <b>Theano.egg-link<\/b> inside python <b>Lib\/site-packages<\/b><\/p>\n\n<p>but i am getting this error when trying to call <i>import theano<\/i> <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sTUnH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sTUnH.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>i used  and it gave me this window. which indicating everything installed fine. any help?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5aEps.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5aEps.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"32397924","CreationDate":"2015-09-04T12:19:59.753","ParentId":null,"OwnerUserId":"1843331","Title":null,"Body":"<p>This can most likely be fixed by <a href=\"https:\/\/github.com\/Theano\/Theano\/archive\/master.zip\" rel=\"nofollow\">redownloading the theano project<\/a>. <\/p>\n\n<p>As you can see <a href=\"https:\/\/github.com\/Theano\/Theano\/search?utf8=%E2%9C%93&amp;q=%2C+e0\" rel=\"nofollow\">here<\/a>, the code that is giving you that error is not in the current codebase anymore. It now looks like this<\/p>\n\n<pre><code>def dot(l, r):\n    \"\"\"Return a symbolic matrix\/dot product between l and r \"\"\"\n    rval = NotImplemented\n    e0, e1 = None, None\n\n    if rval == NotImplemented and hasattr(l, '__dot__'):\n        try:\n            rval = l.__dot__(r)\n        except Exception as e0:\n            rval = NotImplemented\n    if rval == NotImplemented and hasattr(r, '__rdot__'):\n        try:\n            rval = r.__rdot__(l)\n        except Exception as e1:\n            rval = NotImplemented\n    if rval == NotImplemented:\n        raise NotImplementedError(\"Dot failed for the following reasons:\",\n                                  (e0, e1))\n    return rval\n<\/code><\/pre>\n"}]}
{"QuestionId":32405035,"AnswerCount":3,"Tags":"<python><opencv><ubuntu><anaconda><caffe>","CreationDate":"2015-09-04T19:09:21.177","AcceptedAnswerId":32514285.0,"OwnerUserId":1391376.0,"Title":"caffe installation : opencv libpng16.so.16 linkage issues","Body":"<p>I am trying to compile caffe with python interface on an Ubuntu 14.04 machine.<\/p>\n\n<p>I have installed Anaconda and opencv with . I have also installed all the requirement stipulated in the coffee and changed the commentary blocks in  so that PYTHON_LIB and PYTHON_INCLUDE point towards Anaconda distributions. <\/p>\n\n<p>When I am calling , the following command is issued: <\/p>\n\n<pre><\/pre>\n\n<p>However, it is stopped by the following set of errors: <\/p>\n\n<pre><\/pre>\n\n<p>Following the advice from this question: <a href=\"https:\/\/stackoverflow.com\/questions\/31962975\/caffe-install-on-ubuntu-for-anaconda-with-python-2-7-fails-with-libpng16-so-16-n\">Caffe install on ubuntu for anaconda with python 2.7 fails with libpng16.so.16 not found<\/a>, I tired running the , and obtained the following output:<\/p>\n\n<pre><\/pre>\n\n<p>I then proceeded to add all the required directories into the  and  of the  file (hence additional  and  in the  call above), including the explicit link to  already present in anaconda libraries. This have however not resolved the issue. I also tried adding the file to  and , but without any effect.<\/p>\n\n<p>What might the problem be and how could I resolve it?<\/p>\n","answers":[{"AnswerId":"32436803","CreationDate":"2015-09-07T10:47:13.690","ParentId":null,"OwnerUserId":"5308701","Title":null,"Body":"<p>I'm guessing you've added <code>\/home\/andrei\/anaconda\/bin<\/code> to the <code>PATH<\/code> environment variable, so that <code>libpng-config<\/code> resolves to <code>\/home\/andrei\/anaconda\/bin\/libpng16-config<\/code>, which is what is causing <code>cmake<\/code> to try and link with libpng v1.6. <\/p>\n\n<p>Strip the <code>anaconda<\/code> dir from your <code>PATH<\/code> environment variable and <code>libpng-config<\/code> should default to libpng v1.2 in <code>\/usr\/lib<\/code> or similar.<\/p>\n"},{"AnswerId":"32514285","CreationDate":"2015-09-11T01:49:40.003","ParentId":null,"OwnerUserId":"1319871","Title":null,"Body":"<p>I came across the same problem. I found it similar to <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2007\">https:\/\/github.com\/BVLC\/caffe\/issues\/2007<\/a>, and I solved it by<\/p>\n\n<pre><code>cd \/usr\/lib\/x86_64-linux-gnu\nsudo ln -s ~\/anaconda\/lib\/libpng16.so.16 libpng16.so.16\nsudo ldconfig\n<\/code><\/pre>\n"},{"AnswerId":"46981654","CreationDate":"2017-10-27T18:50:23.593","ParentId":null,"OwnerUserId":"871418","Title":null,"Body":"<p>It works with me based on the solution here <a href=\"https:\/\/github.com\/hashdist\/hashstack\/issues\/670\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hashdist\/hashstack\/issues\/670<\/a><\/p>\n\n<pre><code>export LD_LIBRARY_PATH=~\/anaconda2\/lib:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>This shall be added in <code>~\/.bashrc<\/code> or <code>~\/.bash_profile<\/code><\/p>\n"}]}
{"QuestionId":32409231,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-09-05T03:29:55.863","AcceptedAnswerId":32410831.0,"OwnerUserId":588495.0,"Title":"Emulating boolean masks in Theano","Body":"<p>I'm porting a numpy expression to theano.  The expression finds the number of true positive predictions for each class, given a one-hot matrix  of ground truth classes and a one-hot matrix  of predicted classes.  The numpy code is:<\/p>\n\n<pre><\/pre>\n\n<p>The last expression yields .  I've tried using theano's nonzero:<\/p>\n\n<pre><\/pre>\n\n<p>The eval results in , which is a 0-1 mask of the rows of  where the prediction is correct.  Any suggestions for how to make this work?  For different ways of doing it?  Thanks.<\/p>\n","answers":[{"AnswerId":"32410831","CreationDate":"2015-09-05T07:59:54.380","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here are three variants demonstrating how to re-implement parts of your numpy code in Theano.<\/p>\n\n<p>Note that Theano's <code>Unique<\/code> operation does not support running on the GPU and does not appear to support gradients either. As a result version 3 many not be of much use. Version 2 provides a workaround: compute the unique values outside Theano and pass them in. Version 1 is a Theano implementation of the final line of your numpy code only.<\/p>\n\n<p>To address your specific issue: there is no need to use <code>nonzero<\/code>; in this case the indexing works in Theano just like it works in numpy. Maybe you were getting confused between <code>y<\/code> and <code>Y<\/code>? (common Python style is to stick with lower case for all variable and parameter names).<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as tt\nimport theano.tensor.extra_ops\n\n\ndef numpy_ver(y, y_hat):\n    Y = np.zeros(shape=(len(y), len(np.unique(y))), dtype=np.int64)\n    Y_hat = np.zeros_like(Y, dtype=np.int64)\n    rows = np.arange(len(y), dtype=np.int64)\n    Y[rows, y] = 1\n    Y_hat[rows, y_hat] = 1\n    return ((Y_hat == Y) &amp; (Y == 1)).sum(axis=0), Y, Y_hat\n\n\ndef compile_theano_ver1():\n    Y = tt.matrix(dtype='int64')\n    Y_hat = tt.matrix(dtype='int64')\n    z = (tt.eq(Y_hat, Y) &amp; tt.eq(Y, 1)).sum(axis=0)\n    return theano.function([Y, Y_hat], outputs=z)\n\n\ndef compile_theano_ver2():\n    y = tt.vector(dtype='int64')\n    y_hat = tt.vector(dtype='int64')\n    y_uniq = tt.vector(dtype='int64')\n    Y = tt.zeros(shape=(y.shape[0], y_uniq.shape[0]), dtype='int64')\n    Y_hat = tt.zeros_like(Y, dtype='int64')\n    rows = tt.arange(y.shape[0], dtype='int64')\n    Y = tt.set_subtensor(Y[rows, y], 1)\n    Y_hat = tt.set_subtensor(Y_hat[rows, y_hat], 1)\n    z = (tt.eq(Y_hat, Y) &amp; tt.eq(Y, 1)).sum(axis=0)\n    return theano.function([y, y_hat, y_uniq], outputs=z)\n\n\ndef compile_theano_ver3():\n    y = tt.vector(dtype='int64')\n    y_hat = tt.vector(dtype='int64')\n    y_uniq = tt.extra_ops.Unique()(y)\n    Y = tt.zeros(shape=(y.shape[0], y_uniq.shape[0]), dtype='int64')\n    Y_hat = tt.zeros_like(Y, dtype='int64')\n    rows = tt.arange(y.shape[0], dtype='int64')\n    Y = tt.set_subtensor(Y[rows, y], 1)\n    Y_hat = tt.set_subtensor(Y_hat[rows, y_hat], 1)\n    z = (tt.eq(Y_hat, Y) &amp; tt.eq(Y, 1)).sum(axis=0)\n    return theano.function([y, y_hat], outputs=z)\n\n\ndef main():\n    y = np.array([1, 0, 1, 2, 2], dtype=np.int64)\n    y_hat = np.array([2, 0, 1, 1, 0], dtype=np.int64)\n    y_uniq = np.unique(y)\n    result, Y, Y_hat = numpy_ver(y, y_hat)\n    print result\n    theano_ver1 = compile_theano_ver1()\n    print theano_ver1(Y, Y_hat)\n    theano_ver2 = compile_theano_ver2()\n    print theano_ver2(y, y_hat, y_uniq)\n    theano_ver3 = compile_theano_ver3()\n    print theano_ver3(y, y_hat)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":32409825,"AnswerCount":0,"Tags":"<c++><visual-studio-2012><caffe>","CreationDate":"2015-09-05T05:30:30.210","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"LNK2005 error between cpp file and cu file","Body":"<p>I am building dll for <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">caffe<\/a> library with Visual Studio 2013. I have Link error 2005, I guess a conflict between  and .\nThe errors are<\/p>\n\n<pre><\/pre>\n\n<p>This  is not in both  and . Why this Linker error is there? I use Visual Studio 2013 to compile.<\/p>\n\n<p>My  and  are as follows.<\/p>\n\n<p>flatten_layer.cpp<\/p>\n\n<pre><\/pre>\n\n<p>flatten_layer.cu<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":32416226,"AnswerCount":2,"Tags":"<c++><visual-studio><caffe>","CreationDate":"2015-09-05T18:12:48.247","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Create layer error at layer_factory.hpp","Body":"<p>I am trying to extract features using . I implement  in Visual Studio. My caffe library is also build as static library and link to . When I run the code, I have error as <\/p>\n\n<pre><\/pre>\n\n<p>The error happens at <\/p>\n\n<pre><\/pre>\n\n<p>Initially was I thought Link error. Now I look more carefully and realized that not linker issue. What could be the error?<\/p>\n","answers":[{"AnswerId":"38584507","CreationDate":"2016-07-26T08:15:59.170","ParentId":null,"OwnerUserId":"4265775","Title":null,"Body":"<p>make sure your caffe is compiled with opencv<\/p>\n"},{"AnswerId":"33351175","CreationDate":"2015-10-26T16:48:07.363","ParentId":null,"OwnerUserId":"5490165","Title":null,"Body":"<p>I've met recently familiar problem to run my applicatin that has been linked with static library of Caffe (compiled in Visual Studio). There I've found 2 different solutions:<\/p>\n\n<ol>\n<li><p>Add Caffe project to your solution and set the next option in your main project:<\/p>\n\n<p>Project properties -> Common Properties -> Framework and References -> Caffe -> Use Library Dependency Inputs -> True<\/p><\/li>\n<\/ol>\n\n<p>This method is simple, but sometimes we want to use only caffe.lib without project and here comes the 2nd method.<\/p>\n\n<ol start=\"2\">\n<li>Create header files in your project and add there all layer classes declarations externally to oblige a linker to use their symbols. See an example below:<\/li>\n<\/ol>\n\n<h1>Example<\/h1>\n\n<pre><code>#include \"caffe\/common.hpp\"\nnamespace caffe\n{\n    extern INSTANTIATE_CLASS(ConvolutionLayer);\n    extern INSTANTIATE_CLASS(PoolingLayer);\n    extern INSTANTIATE_CLASS(ReLULayer);\n    extern INSTANTIATE_CLASS(TanHLayer);\n}\n<\/code><\/pre>\n\n<p>Finally include the very header file in your application where you're using caffe.<\/p>\n\n<p>Also check the Layer you met mentioned in your problem, for instance, in your case it is \"ImageData\" (or to be more corrected ImageDataLayer), open \"image_data_layer.cpp\" file in VS and check there that \"REGISTER_LAYER_CLASS(ImageData);\" is available there.  <\/p>\n\n<p>Hope it will help to solve the problem.<\/p>\n"}]}
{"QuestionId":32419510,"AnswerCount":8,"Tags":"<python><numpy><theano><keras>","CreationDate":"2015-09-06T02:41:11.500","AcceptedAnswerId":null,"OwnerUserId":579980.0,"Title":"How to get reproducible results in keras","Body":"<p>I get different results (test accuracy) every time I run the  example from Keras framework (<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_lstm.py\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_lstm.py<\/a>)\nThe code contains  in the top, before any keras imports. It should prevent it from generating different numbers for every run. What am I missing?  <\/p>\n\n<p>UPDATE: How to repro:  <\/p>\n\n<ol>\n<li>Install Keras (<a href=\"http:\/\/keras.io\/\">http:\/\/keras.io\/<\/a>)   <\/li>\n<li>Execute <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_lstm.py\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_lstm.py<\/a>  a few times. It will train the model and output test accuracy.<br>\nExpected result: Test accuracy is the same on every run.<br>\nActual result: Test accuracy is different on every run.<\/li>\n<\/ol>\n\n<p>UPDATE2: I'm running it on Windows 8.1 with MinGW\/msys, module versions:<br>\ntheano 0.7.0<br>\nnumpy 1.8.1<br>\nscipy 0.14.0c1<\/p>\n\n<p>UPDATE3: I narrowed the problem down a bit. If I run the example with GPU (set theano flag device=gpu0) then I get different test accuracy every time, but if I run it on CPU then everything works as expected. My graphics card: NVIDIA GeForce GT 635)<\/p>\n","answers":[{"AnswerId":"59076062","CreationDate":"2019-11-27T18:03:28.080","ParentId":null,"OwnerUserId":"2543623","Title":null,"Body":"<p>The problem is now solved in Tensorflow 2.0 ! I had the same issue with TF 1.x (see  <a href=\"https:\/\/stackoverflow.com\/questions\/59075244\/if-keras-results-are-not-reproducible-whats-the-best-practice-for-comparing-mo\">If Keras results are not reproducible, what&#39;s the best practice for comparing models and choosing hyper parameters?<\/a> ) but <\/p>\n\n<pre><code>import os\n####*IMPORANT*: Have to do this line *before* importing tensorflow\nos.environ['PYTHONHASHSEED']=str(1)\n\nimport tensorflow as tf\nimport tensorflow.keras as keras\nimport tensorflow.keras.layers \nimport random\nimport pandas as pd\nimport numpy as np\n\ndef reset_random_seeds():\n   os.environ['PYTHONHASHSEED']=str(1)\n   tf.random.set_seed(1)\n   np.random.seed(1)\n   random.seed(1)\n\n#make some random data\nreset_random_seeds()\nNUM_ROWS = 1000\nNUM_FEATURES = 10\nrandom_data = np.random.normal(size=(NUM_ROWS, NUM_FEATURES))\ndf = pd.DataFrame(data=random_data, columns=['x_' + str(ii) for ii in range(NUM_FEATURES)])\ny = df.sum(axis=1) + np.random.normal(size=(NUM_ROWS))\n\ndef run(x, y):\n    reset_random_seeds()\n\n    model = keras.Sequential([\n            keras.layers.Dense(40, input_dim=df.shape[1], activation='relu'),\n            keras.layers.Dense(20, activation='relu'),\n            keras.layers.Dense(10, activation='relu'),\n            keras.layers.Dense(1, activation='linear')\n        ])\n    NUM_EPOCHS = 500\n    model.compile(optimizer='adam', loss='mean_squared_error')\n    model.fit(x, y, epochs=NUM_EPOCHS, verbose=0)\n    predictions = model.predict(x).flatten()\n    loss = model.evaluate(x,  y) #This prints out the loss by side-effect\n\n#With Tensorflow 2.0 this is now reproducible! \nrun(df, y)\nrun(df, y)\nrun(df, y)\n<\/code><\/pre>\n"},{"AnswerId":"40151338","CreationDate":"2016-10-20T10:05:08.103","ParentId":null,"OwnerUserId":"36195","Title":null,"Body":"<p>I finally got reproducible results with my code. It's a combination of answers I saw around the web. The first thing is doing what @alex says:<\/p>\n\n<ol>\n<li>Set <code>numpy.random.seed<\/code>;<\/li>\n<li>Use <code>PYTHONHASHSEED=0<\/code> for Python 3.<\/li>\n<\/ol>\n\n<p>Then you have to solve the issue noted by @user2805751 regarding cuDNN by calling your Keras code with the following additional <code>THEANO_FLAGS<\/code>:<\/p>\n\n<ol start=\"3\">\n<li><code>dnn.conv.algo_bwd_filter=deterministic,dnn.conv.algo_bwd_data=deterministic<\/code><\/li>\n<\/ol>\n\n<p>And finally, you have to patch your Theano installation as per <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1935#issuecomment-194359606\" rel=\"noreferrer\">this comment<\/a>, which basically consists in:<\/p>\n\n<ol start=\"4\">\n<li>replacing all calls to <code>*_dev20<\/code> operator by its regular version in <code>theano\/sandbox\/cuda\/opt.py<\/code>.<\/li>\n<\/ol>\n\n<p>This should get you the same results for the same seed.<\/p>\n\n<p>Note that there might be a slowdown. I saw a running time increase of about 10%.<\/p>\n"},{"AnswerId":"36888790","CreationDate":"2016-04-27T11:44:02.840","ParentId":null,"OwnerUserId":"3667840","Title":null,"Body":"<p>I agree with the previous comment, but reproducible results sometimes needs the same environment(e.g. installed packages, machine characteristics and so on). So that, I recommend to copy your environment to other place in case to have reproducible results. Try to use one of the next technologies:<\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.docker.com\/\" rel=\"nofollow\">Docker<\/a>. If you have a Linux this very easy to move your environment to other place. Also you can try to use <a href=\"https:\/\/hub.docker.com\/\" rel=\"nofollow\">DockerHub<\/a>. <\/li>\n<li><a href=\"http:\/\/mybinder.org\/\" rel=\"nofollow\">Binder<\/a>. This is a cloud platform for reproducing scientific experiments.<\/li>\n<li><a href=\"http:\/\/everware.xyz\/\" rel=\"nofollow\">Everware<\/a>. This is yet another cloud platform for \"reusable science\". See the <a href=\"https:\/\/github.com\/everware\" rel=\"nofollow\">project repository<\/a> on Github.<\/li>\n<\/ol>\n"},{"AnswerId":"56606207","CreationDate":"2019-06-14T23:40:46.947","ParentId":null,"OwnerUserId":"7350637","Title":null,"Body":"<p>This works for me:<\/p>\n\n<pre><code>SEED = 123456\nimport os\nimport random as rn\nimport numpy as np\nfrom tensorflow import set_random_seed\n\nos.environ['PYTHONHASHSEED']=str(SEED)\nnp.random.seed(SEED)\nset_random_seed(SEED)\nrn.seed(SEED)\n<\/code><\/pre>\n"},{"AnswerId":"38371558","CreationDate":"2016-07-14T10:17:14.887","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>I have trained and tested <code>Sequential()<\/code> kind of neural networks using Keras. I performed non linear regression on noisy speech data. I used the following code to generate random seed : <\/p>\n\n<pre><code>import numpy as np\nseed = 7\nnp.random.seed(seed)\n<\/code><\/pre>\n\n<p>I get the exact same results of <code>val_loss<\/code> each time I train and test on the same data. <\/p>\n"},{"AnswerId":"52897216","CreationDate":"2018-10-19T17:23:30.617","ParentId":null,"OwnerUserId":"9024698","Title":null,"Body":"<p>You can find the answer at the Keras docs: <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-reproducible-results-using-keras-during-development\" rel=\"noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-reproducible-results-using-keras-during-development<\/a>.<\/p>\n\n<p>In short, to be absolutely sure that you will get reproducible results with your python script <strong>on one computer's\/laptop's CPU<\/strong> then you will have to do the following:<\/p>\n\n<ol>\n<li>Set the <code>PYTHONHASHSEED<\/code> environment variable at a fixed value<\/li>\n<li>Set the <code>python<\/code> built-in pseudo-random generator at a fixed value<\/li>\n<li>Set the <code>numpy<\/code> pseudo-random generator at a fixed value<\/li>\n<li>Set the <code>tensorflow<\/code> pseudo-random generator at a fixed value<\/li>\n<li>Configure a new global <code>tensorflow<\/code> session<\/li>\n<\/ol>\n\n<p>Following the <code>Keras<\/code> link at the top, the source code I am using is the following:<\/p>\n\n<pre><code># Seed value\n# Apparently you may use different seed values at each stage\nseed_value= 0\n\n# 1. Set the `PYTHONHASHSEED` environment variable at a fixed value\nimport os\nos.environ['PYTHONHASHSEED']=str(seed_value)\n\n# 2. Set the `python` built-in pseudo-random generator at a fixed value\nimport random\nrandom.seed(seed_value)\n\n# 3. Set the `numpy` pseudo-random generator at a fixed value\nimport numpy as np\nnp.random.seed(seed_value)\n\n# 4. Set the `tensorflow` pseudo-random generator at a fixed value\nimport tensorflow as tf\ntf.set_random_seed(seed_value)\n\n# 5. Configure a new global `tensorflow` session\nfrom keras import backend as K\nsession_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)\nsess = tf.Session(graph=tf.get_default_graph(), config=session_conf)\nK.set_session(sess)\n<\/code><\/pre>\n\n<p>It is needless to say that you do not have to to specify any <code>seed<\/code> or <code>random_state<\/code> at the <code>numpy<\/code>, <code>scikit-learn<\/code> or <code>tensorflow<\/code>\/<code>keras<\/code> functions that you are using in your python script exactly because with the source code above we set globally their pseudo-random generators at a fixed value.<\/p>\n"},{"AnswerId":"38950594","CreationDate":"2016-08-15T06:57:53.947","ParentId":null,"OwnerUserId":"2004627","Title":null,"Body":"<p>I would like to add something to the previous answers. If you use <strong>python 3<\/strong> and you want to get reproducible results for every run, you have to<\/p>\n\n<ol>\n<li>set numpy.random.seed in the beginning of your code<\/li>\n<li>give PYTHONHASHSEED=0 as a parameter to the python interpreter<\/li>\n<\/ol>\n"},{"AnswerId":"32687043","CreationDate":"2015-09-21T03:45:27.923","ParentId":null,"OwnerUserId":"3299394","Title":null,"Body":"<p>Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/sandbox\/randomnumbers.html\">documentation<\/a> talks about the difficulties of seeding random variables and why they seed each graph instance with its own random number generator. <\/p>\n\n<blockquote>\n  <p>Sharing a random number generator between different {{{RandomOp}}}\n  instances makes it difficult to producing the same stream regardless\n  of other ops in graph, and to keep {{{RandomOps}}} isolated.\n  Therefore, each {{{RandomOp}}} instance in a graph will have its very\n  own random number generator. That random number generator is an input\n  to the function. In typical usage, we will use the new features of\n  function inputs ({{{value}}}, {{{update}}}) to pass and update the rng\n  for each {{{RandomOp}}}. By passing RNGs as inputs, it is possible to\n  use the normal methods of accessing function inputs to access each\n  {{{RandomOp}}}\u2019s rng. In this approach it there is no pre-existing\n  mechanism to work with the combined random number state of an entire\n  graph. So the proposal is to provide the missing functionality (the\n  last three requirements) via auxiliary functions: {{{seed, getstate,\n  setstate}}}.<\/p>\n<\/blockquote>\n\n<p>They also provide <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html\">examples<\/a> on how to seed all the random number generators. <\/p>\n\n<blockquote>\n  <p>You can also seed all of the random variables allocated by a\n  RandomStreams object by that object\u2019s seed method. This seed will be\n  used to seed a temporary random number generator, that will in turn\n  generate seeds for each of the random variables.<\/p>\n<\/blockquote>\n\n<pre><code>&gt;&gt;&gt; srng.seed(902340)  # seeds rv_u and rv_n with different seeds each\n<\/code><\/pre>\n"}]}
{"QuestionId":32426221,"AnswerCount":3,"Tags":"<python><theano><dimensionality-reduction><autoencoder>","CreationDate":"2015-09-06T17:19:17.993","AcceptedAnswerId":null,"OwnerUserId":1081942.0,"Title":"Autoencoders for high dimensional data","Body":"<p>I'm working on a project where I need to reduce the dimensionality of my observations and still have a significative representation of them. The use of Autoencoders was strongly suggested for many reasons but I'm not quite sure it's the best approach.<\/p>\n\n<p>I have 1400 samples of dimension ~60,000 which is far too high, I am trying to reduce their dimensionality to a 10% of the original one. I'm using <strong>theano autoencoders<\/strong> [<a href=\"http:\/\/deeplearning.net\/tutorial\/dA.html\" rel=\"nofollow noreferrer\">Link<\/a>] and it seems like the cost keeps being around 30,000 (which is very high). I tried raising the number of epochs or lowering the learning rate with no success. I'm not a big expert on autoencoders so I'm not sure how to proceed from here or when to just stop trying.<\/p>\n\n<p>There are other tests I can run but, before going any further, I'd like to have an input from you. <\/p>\n\n<ul>\n<li><p>Do you think the dataset is too small (I can add another 600 samples for a total of ~2000) ? <\/p><\/li>\n<li><p>Do you think using stacked autoenoders could help ? <\/p><\/li>\n<li><p>Should I keep tweaking the parameters (epochs and learning rate) ?<\/p><\/li>\n<\/ul>\n\n<p>Since the dataset is an ensamble of pictures I tried to visualize the reconstructions from the autoencoders and all I got was the same output for every sample. This means that given the input the autoencoder tries to rebuild the input but what I get instead is the same (almost exactly) image for any input(which kind of looks like an average of all the images in the dataset). This means that the inner representation is not good enough since the autoencoder can't reconstruct the image from it.<\/p>\n\n<p><strong>The dataset:<\/strong> 1400 - 2000 images of scanned books (covers included) of around ~60.000 pixels each (which translates to a feature vector of 60.000 elements). Each feature vector has been normalized in [0,1] and originally had values in [0,255].<\/p>\n\n<p><strong>The problem<\/strong>: Reduce their dimensionality with Autoencoders (if possible)<\/p>\n\n<p>If you need any extra info or if I missed something that might be useful to better understand the problem, please add a comment and I will happily help you help me =).<\/p>\n\n<p>Note: I'm currently running a test with a higher number of epochs on the whole dataset and I will update my post accordingly to the result, it might take a while though.<\/p>\n","answers":[{"AnswerId":"32443030","CreationDate":"2015-09-07T16:42:12.120","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There's no reason to necessarily consider a cost of 30,000 as \"high\" unless more is known about the situation than described in the question. The globally minimal cost might actually be around 30,000 if, for example, the size of the hidden layer is particularly small and there is little redundancy in the data.<\/p>\n\n<p>If the cost is 30,000 before training (i.e. with random encoder and decoder weights) and remains around that level even after some training then something probably is wrong.<\/p>\n\n<p>You should expect the cost to decrease after the first update (you'll have many updates per epoch if you're using minibatch stochastic gradient descent). You should also expect the convergence cost to decrease as the size of the hidden layer is increased.<\/p>\n\n<p>Other techniques that might help in a situation like this include the <a href=\"http:\/\/ace.cs.ohio.edu\/~razvan\/courses\/dl6890\/papers\/vincent08.pdf\" rel=\"nofollow noreferrer\">denoising autoencoder<\/a> (which can be thought of as artificially increasing the size of your training dataset by repeated application of random noise) and the <a href=\"http:\/\/www.icml-2011.org\/papers\/455_icmlpaper.pdf\" rel=\"nofollow noreferrer\">contractive autoencoder<\/a> which focusses its regularization power on the encoder, the part you care about. Both can be implemented in Theano and the first is the <a href=\"http:\/\/deeplearning.net\/tutorial\/dA.html\" rel=\"nofollow noreferrer\">subject of this tutorial<\/a> (<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/dA.py\" rel=\"nofollow noreferrer\">with code<\/a>).<\/p>\n"},{"AnswerId":"35137864","CreationDate":"2016-02-01T18:39:08.010","ParentId":null,"OwnerUserId":"1268117","Title":null,"Body":"<p>Autoencoders are useful in part because they can learn nonlinear dimensionality reductions. There are other dimensionality reduction techniques, however, which are much faster than autoencoders. Diffusion maps is a popular one; locally-linear embedding is another. I've used diffusion maps on >2000 60k-dimensional data (also images) and it works in under a minute.<\/p>\n\n<p>Here's a straightforward Python implementation using numpy et al:<\/p>\n\n<pre><code>def diffusion_maps(data, d, eps=-1, t=1):\n    \"\"\"\n    data is organized such that columns are points. so it's 60k x 2k for you\n    d is the target dimension\n    eps is the kernel bandwidth, estimated automatically if == -1\n    t is the diffusion time, 1 is usually fine\n    \"\"\"\n\n    from scipy.spatial import pdist, squareform\n    from scipy import linalg as la\n    import numpy as np\n\n    distances = squareform(pdist(data.T))\n\n    if eps == -1:\n        # if a kernel bandwidth was not supplied,\n        # just use the distance to the tenth-nearest neighbor\n        k = 10\n        nn = np.sort(distances)\n        eps = np.mean(nn[:, k + 1])\n\n    kernel = np.exp(-distances ** 2 \/ eps ** 2)\n    one = np.ones(n_samples)\n    p_a = np.dot(kernel, one)\n    kernel_p = walk \/ np.outer(p_a, p_a)\n    dd = np.dot(kernel_p, one) ** 0.5\n    walk = kernel_p \/ np.outer(dd, dd)\n\n    vecs, eigs, _ = la.svd(walk, full_matrices=False)\n    vecs = vecs \/ vecs[:, 0][:, None]\n    diffusion_coordinates = vecs[:, 1:d + 1].T * (eigs[1:d + 1][:, None] ** t)\n\n    return diffusion_coordinates\n<\/code><\/pre>\n\n<p>The gist of diffusion maps is that you form a random walk on your data such that you're much much more likely to visit nearby points than far-away ones. Then you can define a distance between points (the diffusion distance), which is in essence an average probability of moving between two points over all possible paths. The trick is this is actually extremely easy to compute; all you need to do is diagonalize a matrix, and then embed your data in low-dimensional space using its eigenvectors. In this embedding the Euclidean distance is the diffusion distance, up to an approximation error.<\/p>\n"},{"AnswerId":"48127709","CreationDate":"2018-01-06T13:15:27.930","ParentId":null,"OwnerUserId":"9175058","Title":null,"Body":"<p><strong>Simple things first...<\/strong>  Note that if you have only 1400 points in 60,000 dimensional space, then you can <strong>without loss<\/strong>, reduce dimensionality to size &lt;=1400.  That is a simple mathematical fact: your data matrix is 1400x60,000, so its rank (dimensionality) is at most 1400. Thus, Principal Components Analysis (PCA) will produce 1400 points in 1400 dimensional space, without loss.  I strongly suggest using PCA to reduce the dimensionality of your data before considering anything else.  <\/p>\n"}]}
{"QuestionId":32433515,"AnswerCount":1,"Tags":"<python><python-2.7><python-3.x><subprocess><theano>","CreationDate":"2015-09-07T07:45:22.720","AcceptedAnswerId":null,"OwnerUserId":4232441.0,"Title":"Multiple class call with same output","Body":"<p>I have this  defined in , Now in the same file I have another class and a separate function, in both of that I call the .<\/p>\n\n<p>In  I have 2 classes -  and  and one separate . In class MLP, I use this statement  which returns a 2 dimensional array(eg. 5000 * 2100). Then In the another module  I call the   In which I want the Same value of  to be passed as input to the function(iter_minibatches) or I want the same value to get assigned in the beginning of the function(iter_minibatches), so I tried calling the same class in , but since I use random function for permutation am getting different output(eg. 5000 * 2102) but my requirement is to get same value in the function iter_minibatches as well.<\/p>\n\n<p>Now I want same value of <strong>X<\/strong> should be returned from both the call. So how can I do that?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32441573","CreationDate":"2015-09-07T15:01:10.743","ParentId":null,"OwnerUserId":"1405065","Title":null,"Body":"<p>Your question seems mostly to be about how you can share a piece of data between several functions located in different modules. There are two good ways to go about it:<\/p>\n\n<p>First, you can store the data in a global variable and access it there any time you need it. For example:<\/p>\n\n<pre><code># top level code of module1.py\ninp_x = Input(traindata).X  # I'm assuming traindata is also a global\n\n# later code in the same module\ndo_stuff(inp_x)\n\n# code in other modules can do:\nimport module1\ndo_other_stuff(module1.inp_x)  # a module's global variables are attributes on the module\n<\/code><\/pre>\n\n<p>The second option is to create the data in some specific part of your program and store it locally, then pass it to each of the other places you need to use it. This lets you use a more general structure to your code, and you can pass different values at different times. Here's what that might look like:<\/p>\n\n<pre><code># functions that use an X value:\ndef func1(arg1, arg2, X):\n    do_stuff(X)\n\ndef func2(X):\n    do_other_stuff(X)\n\n# main module (or wherever you call func1 and func2 from)\nfrom module1 import func1\nfrom module2 import func2\n\ndef main():\n    x = Input(traindata).X\n    func1(\"foo\", \"bar\", x)\n    func2(x)\n<\/code><\/pre>\n\n<p>In these examples, I'm only saving (or passing as an argument) the <code>X<\/code> value that's calculated in the <code>Input<\/code> class, which seems to be how you're using that class too.<\/p>\n\n<p>That is a bit silly. If you don't need to keep the instance of <code>Input<\/code>, you probably shouldn't make <code>Input<\/code> a class in the first place. Instead, make it a function that returns the <code>X<\/code> value at the end. You'll have a few more arguments to pass between the <code>realize<\/code> and <code>expand<\/code> functions, but the code will likely be cleaner overall.<\/p>\n\n<p>On the other hand, if the instances of <code>Input<\/code> have some other uses (which you simply haven't shown in your example), it might make sense to save the instance you're creating instead of its <code>X<\/code> attribute:<\/p>\n\n<pre><code>inp = Input(traindata)  # save the Input instance, not only X\n\n# later code:\ndo_stuff(inp.X)  # use the X attribute, as above\n\n# other code\ndo_other_stuff(inp)  # pass the instance, so you can use other attributes or methods\n<\/code><\/pre>\n"}]}
{"QuestionId":32444016,"AnswerCount":2,"Tags":"<compilation><cuda><automated-tests><caffe>","CreationDate":"2015-09-07T18:03:59.283","AcceptedAnswerId":32445580.0,"OwnerUserId":3102241.0,"Title":"Caffe compiled fine with cudnn however runtest fails with error: CUDNN_STATUS_ARCH_MISMATCH","Body":"<p>When running make runtest using caffe I get the following output, it all compiles fine with Cudnn no errors are provided, I have also included the output of build_release\/tools\/caffe device_query -gpu &lt;0,1> for both NVidia Tesla GPUs running Cuda driver and runtime version 7.0. Could anyone help?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32445580","CreationDate":"2015-09-07T20:22:59.800","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>The cuDNN library <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"noreferrer\">requires a GPU of compute capability 3.0 or higher<\/a>:<\/p>\n\n<blockquote>\n  <p>Supported on Windows, Linux and MacOS systems with Kepler, Maxwell or Tegra K1 GPUs.<\/p>\n<\/blockquote>\n\n<p>Your Fermi M2090 is a compute capability 2.0 GPU:<\/p>\n\n<pre><code>I0907 18:55:05.037195   729 common.cpp:169] Major revision number:         2\nI0907 18:55:05.037201   729 common.cpp:170] Minor revision number:         0\nI0907 18:55:05.037207   729 common.cpp:171] Name:                          Tesla M2090\n<\/code><\/pre>\n"},{"AnswerId":"44433724","CreationDate":"2017-06-08T10:52:24.880","ParentId":null,"OwnerUserId":"2091453","Title":null,"Body":"<p>Firstly check type this command <code>nvidia-smi<\/code> using  the terminal.\nThen go to this <a href=\"https:\/\/en.wikipedia.org\/wiki\/CUDA\" rel=\"nofollow noreferrer\">link<\/a> and try to find your GPU in the table. You will probably find out that your gpu have lower compute capability than 3.0 which is not supported for cuDNN.<\/p>\n\n<p>Based on your output and TESLA M2090 you probably have one of the following GPUs:<\/p>\n\n<blockquote>\n  <p>GeForce GTX 590, GeForce GTX 580, GeForce GTX 570, GeForce GTX 480,\n  GeForce GTX 470, GeForce GTX 465, GeForce GTX 480M<\/p>\n<\/blockquote>\n\n<p>And  the compute capability of above GPUs is 2.0. So my suggestion is  try to install caffe without cuDNN , and do not use it at least in current machine.<\/p>\n"}]}
{"QuestionId":32451934,"AnswerCount":1,"Tags":"<image-processing><neural-network><deep-learning><caffe><labeling>","CreationDate":"2015-09-08T07:39:54.857","AcceptedAnswerId":32471602.0,"OwnerUserId":5295427.0,"Title":"Image per-pixel Scene labeling output issue (using FCN-32s Semantic Segmentation)","Body":"<p>I'm looking for a way that, given an input image and a neural network, it will output a labeled class for each pixel in the image (sky, grass, mountain, person, car etc).<\/p>\n\n<p>I've set up Caffe (the future-branch) and successfully run the <a href=\"https:\/\/gist.github.com\/shelhamer\/80667189b218ad570e82\/\" rel=\"nofollow noreferrer\">FCN-32s Fully Convolutional Semantic Segmentation on PASCAL-Context<\/a> model. However, I'm unable to produce clear labeled images with it.<\/p>\n\n<p>Images that visualizes my problem:<br>\nInput image<br>\n<img src=\"https:\/\/i.imgur.com\/gYuCHMj.jpg\" alt=\"\"><br>\nground truth<br>\n<img src=\"https:\/\/i.imgur.com\/zalGt9v.png\" alt=\"\"><br>\nAnd my result:<br>\n<img src=\"https:\/\/i.imgur.com\/TN9juhO.png\" alt=\"\"><\/p>\n\n<p>This might be some resolution issue. Any idea of where I'm going wrong?<\/p>\n","answers":[{"AnswerId":"32471602","CreationDate":"2015-09-09T05:35:01.157","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like the 32s model is making large strides and thus working at a coarse resolution. Can you try the <a href=\"https:\/\/gist.github.com\/shelhamer\/91eece041c19ff8968ee\" rel=\"noreferrer\">8s model<\/a> that seems to perform less resolution reduction.<br>\nLooking at <a href=\"http:\/\/www.cs.berkeley.edu\/~jonlong\/long_shelhamer_fcn.pdf\" rel=\"noreferrer\">J Long, E Shelhamer, T Darrell <em>Fully Convolutional Networks for Semantic Segmentation<\/em>, CVPR 2015<\/a> (especially at figure 4) it seems like the 32s model is not designed for capturing fine details of the segmentation.<\/p>\n"}]}
{"QuestionId":32452265,"AnswerCount":1,"Tags":"<computer-vision><neural-network><convolution><deep-learning><caffe>","CreationDate":"2015-09-08T07:59:01.317","AcceptedAnswerId":null,"OwnerUserId":639973.0,"Title":"Multi-label Input to Single-label output","Body":"<p>Is there a research paper where inputs are multi-labeled but the output (classifier) is a single-labeled? Preferably in computer vision field.<\/p>\n","answers":[{"AnswerId":"32456454","CreationDate":"2015-09-08T11:20:39.090","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your question is not very clear, but were you thinking of something like <a href=\"http:\/\/lear.imag.fr\/pubs\/2007\/NJ07\/NJ07.pdf\" rel=\"nofollow\">similarity learning<\/a> where you are trying to learn a binary label (\"same\"\/\"not-same\") given that the inputs are labeled with class labels (0..<code>L<\/code>)?<\/p>\n\n<p>In computer vision you can find similar task for faces: given an input identity of faces you want a classifier to output (for a pair of faces) if they are of the same person or not. You can find more information regarding this research field on the web page of the <a href=\"http:\/\/vis-www.cs.umass.edu\/lfw\/\" rel=\"nofollow\">labeled faces in the wild<\/a> benchmark.<\/p>\n"}]}
{"QuestionId":32457236,"AnswerCount":3,"Tags":"<python><theano>","CreationDate":"2015-09-08T11:57:37.417","AcceptedAnswerId":null,"OwnerUserId":4476617.0,"Title":"Theano stack matrices programmatically?","Body":"<p>I have the following code which stacks 2 matrices into a 3D tensor.<\/p>\n\n<pre><\/pre>\n\n<p>However, I do not know how many times I need to stack the matrix in advance. For example the fourth line of code may be:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a theano function to duplicate a matrix n times:<\/p>\n\n<pre><\/pre>\n\n<p>Then I can pass that 3, as an argument to the theano function f. Is this possible? I looked into broadcasting but broadcasting does not explicitly change dimensionality\/stack.<\/p>\n","answers":[{"AnswerId":"32457538","CreationDate":"2015-09-08T12:12:34.727","ParentId":null,"OwnerUserId":"3849224","Title":null,"Body":"<p>I don't know about theano, but you could accomplish this using list comprehension and unpacking argument list:<\/p>\n\n<pre><code>n = 5\nB = theano.tensor.stack(*[A for dummy in range(n)])\n<\/code><\/pre>\n\n<p>which is equivalent to:<\/p>\n\n<pre><code>B = theano.tensor.stack(A, A, A, A, A)\n<\/code><\/pre>\n\n<p>What this does is, it first constructs a list with <code>n<\/code> copies of <code>A<\/code> and then unpacks this list into separate arguments (see <a href=\"https:\/\/docs.python.org\/2\/tutorial\/controlflow.html#unpacking-argument-lists\" rel=\"nofollow\">https:\/\/docs.python.org\/2\/tutorial\/controlflow.html#unpacking-argument-lists<\/a>).<\/p>\n"},{"AnswerId":"32462494","CreationDate":"2015-09-08T16:08:01.990","ParentId":null,"OwnerUserId":"4476617","Title":null,"Body":"<p>After digging long and hard through the theano documentation I have found the solution:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nA = T.matrix(\"A\")\nB = [A]\nC = theano.tensor.extra_ops.repeat(B, 3, axis=0)\nf = theano.function(inputs=[A], outputs=C)\nprint f([range(10)]*2)\n<\/code><\/pre>\n\n<p>is equivalent to:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nA = T.matrix(\"A\")\nB = theano.tensor.stack(A, A, A)\nf = theano.function(inputs=[A], outputs=B)\nprint f([range(10)]*2)\n<\/code><\/pre>\n\n<p>except we can now choose the number of repeats programatically as the second argument to: theano.tensor.extra_ops.repeat<\/p>\n"},{"AnswerId":"32478228","CreationDate":"2015-09-09T11:23:04.733","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Here is an example using broadcasting<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nA = T.fmatrix()\nn = T.iscalar()\n\nones = T.ones((n, 1, 1))\n\nstackedA = ones * A[np.newaxis, :, :]\n\nf = theano.function([A, n], stackedA)\n\na = np.arange(30).reshape(5, 6).astype('float32')\nnn = 3\n\nr = f(a, nn)\n\nprint r.shape  # outputs (3, 4, 5)\nprint (r == a[np.newaxis]).all()  # outputs True\n<\/code><\/pre>\n\n<p>This approach can help the compiler avoid tiling if it can optimize that away.<\/p>\n"}]}
{"QuestionId":32462036,"AnswerCount":2,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2015-09-08T15:45:17.510","AcceptedAnswerId":32462902.0,"OwnerUserId":3888963.0,"Title":"python theano Optimization failure due to: local_dot_to_dot22","Body":"<p>I just pip installed theano and tried to run theano.test(). It produced a very long log of errors and I copied the first part. I also tried a couple other examples - I have seen <\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>several times. <\/p>\n\n<p>I'm using python 2.7 (canopy), scipy 0.15.1-2 and numpy 1.9.2-1. I am very new to theano. I appreciate if you can point me to the right direction. Thanks!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43199865","CreationDate":"2017-04-04T06:30:28.040","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In case you don't want to reinstall things, if they're heavy programs, for instance, affecting Window's registry and so, you can try <strong>symbolic links<\/strong>.    <\/p>\n\n<p>A symbolic link will create something similar to a shortcut to a folder, but seen as an actual folder by other applications.   <\/p>\n\n<p>So, you can do something like this:<\/p>\n\n<ul>\n<li>Run <code>cmd<\/code> as administrator    <\/li>\n<li>User this command: <code>mklink \/D \"C:\\LinkToProgramFiles\" \"C:\\Program Files\"<\/code><\/li>\n<\/ul>\n\n<p>And then, you start using \"C:\\LinkToProgramFiles\" in your ldflags var. <\/p>\n"},{"AnswerId":"32462902","CreationDate":"2015-09-08T16:33:39.487","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem here is problem caused by having spaces in your path, i.e. Canopy is installed in <code>C:\\Program Files\\Enthought\\Canopy<\/code> but the Theano scripts don't work well with the space between <code>Program<\/code> and <code>Files<\/code>. Try uninstalling Canopy and reinstall in a directory with no space in the path.<\/p>\n\n<p>You should also follow the other instructions for <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">installing Theano on Windows<\/a>. Unfortunately it's not as simple as just <code>pip install theano<\/code>.<\/p>\n"}]}
{"QuestionId":32471395,"AnswerCount":0,"Tags":"<fft><convolution><theano>","CreationDate":"2015-09-09T05:15:33.650","AcceptedAnswerId":null,"OwnerUserId":5307788.0,"Title":"can you modify this theano code for fft-convolution available?","Body":"<p>I'm searching for the way to use fft-convolution in theano.\nI wrote simple convolution code with theano. \nBut this code doesn't work if i set  though simple convolution works with <\/p>\n\n<p>Please tell me what is wrong with this code?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The error message is like below: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":32471670,"AnswerCount":1,"Tags":"<memory><gpu><theano>","CreationDate":"2015-09-09T05:40:42.667","AcceptedAnswerId":40531639.0,"OwnerUserId":1401278.0,"Title":"Storing specific shared variables in CPU","Body":"<p>Is it possible in theano to selectively choose some shared variables in the CPU? I have a huge matrix in the output layer over entire vocabulary (~2M) that wouldn't fit in the GPU memory. I have experimented with reducing its size thro' sampling, but I want to see if I can use the entire matrix. One way I could do is to use  in theano flags. But, this seem to use GPU only on a need basis. I checked the tutorial and it doesn't seem to have more details.<\/p>\n\n<p>I wonder if it is possible to specify one or few shared variables to be stored in cpu. One can do this when creating the shared variable I guess. Having some of the variables in GPU will be faster than having everything in CPU right? Or does theano somehow figure out which ones to implicitly keep\/move automatically? Would appreciate some explanation.<\/p>\n","answers":[{"AnswerId":"40531639","CreationDate":"2016-11-10T15:48:50.020","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>In newer Theano (I forgot Theano 0.8.2 or the dev version of Theano 0.9), there is a different interface. You can do theano.shared(data, target='cpu')<\/p>\n\n<p>Continue to initialize the GPU as you did before.<\/p>\n"}]}
{"QuestionId":32472524,"AnswerCount":1,"Tags":"<image-processing><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-09-09T06:39:57.640","AcceptedAnswerId":null,"OwnerUserId":1377127.0,"Title":"Caffe error: no field named \"net\"","Body":"<p>I had the Caffe C++ example program working on my computer, but after recently recompiling Caffe, I've encountered this error when I try to run the program:<\/p>\n\n<blockquote>\n  <p>[libprotobuf ERROR google\/protobuf\/text_format.cc:245] Error parsing\n   text-format caffe.NetParameter: 2:4: Message type \"caffe.NetParameter\"\n   has no field named \"net\".<br>\n   upgrade_proto.cpp:928] Check failed: ReadProtoFromTextFile(param_file,\n   param) Failed to parse NetParameter file:\n   \/home\/jack\/Desktop\/beeshiny\/deploy.prototxt<\/p>\n<\/blockquote>\n\n<p>Am I missing something or has the syntax of the prototxt files been changed? My deploy.prototxt file (that I pass to the C++ program) looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>The contents of the deploy_arch.prototxt file referenced in the prototxt file above:<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand why this has stopped working all of a sudden, unless an update has made my prototxt file obsolete?<\/p>\n","answers":[{"AnswerId":"37581277","CreationDate":"2016-06-02T01:30:12.567","ParentId":null,"OwnerUserId":"3835707","Title":null,"Body":"<p>I solved my problem by adding <code>caffe\/python<\/code> in <code>$PYTHONPATH<\/code>.<\/p>\n"}]}
{"QuestionId":32487164,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-09-09T18:51:57.533","AcceptedAnswerId":null,"OwnerUserId":5285204.0,"Title":"Calling seperate theano functions with same inputs?","Body":"<p>I have something like this:<\/p>\n\n<pre><\/pre>\n\n<p>This is working fine. I would now like to see what the mean of the hidden units is. I tried adding this before the line where  is declared:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<p><\/p>\n\n<p>I really have two questions: 1) Why can't I do this? 2) What is the correct way to achieve what I want?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"32496887","CreationDate":"2015-09-10T08:32:58.287","ParentId":null,"OwnerUserId":"4511818","Title":null,"Body":"<p>As far as I can see, you are giving him the function as input. If you use your array\/matrix of hidden units instead the code should work.<\/p>\n\n<pre><code>hidden_mean_func = th.function(inputs=[hidden], outputs=[hm], name=\"hidden_mean_function_printer\")\nprint hidden_mean_func(hidden)\n<\/code><\/pre>\n"}]}
{"QuestionId":32493904,"AnswerCount":5,"Tags":"<linux><ubuntu><cuda><nvidia><caffe>","CreationDate":"2015-09-10T05:35:21.713","AcceptedAnswerId":51078813.0,"OwnerUserId":2467772.0,"Title":"Could not insert 'nvidia_352': No such device","Body":"<p>I am trying to run <a href=\"http:\/\/caffe.berkeleyvision.org\/\">caffe<\/a> on . \nAfter installation, I run caffe in gpu and the error is <\/p>\n\n<pre><\/pre>\n\n<p>My NVIDIA driver is 352.41.\nI installed 352 and it is installed latest version.<\/p>\n\n<pre><\/pre>\n\n<p>My Ubuntu has NVIDIA driver 352 and why I have error like<\/p>\n\n<pre><\/pre>\n\n<p>I checked whether I have CUDA capable device like<\/p>\n\n<pre><\/pre>\n\n<p>I have CUDA capable device and why I get the error?<\/p>\n\n<p>EDIT 1:\nYeah my test with .\/deviceQuery failed.<\/p>\n\n<pre><\/pre>\n\n<p>I checked in the dev\/ folder, I have nvidia0.<\/p>\n\n<pre><\/pre>\n\n<p>My nvcc -V check gave me<\/p>\n\n<pre><\/pre>\n\n<p>Then my version check<\/p>\n\n<pre><\/pre>\n\n<p>What could be wrong?<\/p>\n","answers":[{"AnswerId":"37337328","CreationDate":"2016-05-20T03:08:38.730","ParentId":null,"OwnerUserId":"2492724","Title":null,"Body":"<p>I also had this problem. The above answers didn't work for me. When I installed latest driver(<code>nvidia-364<\/code>), it worked. Commands to run:<\/p>\n\n<pre><code>sudo add-apt-repository ppa:xorg-edgers\/ppa \nsudo apt-get update \nsudo apt-get install nvidia-364\n<\/code><\/pre>\n\n<p>I think the problem occurs when we have different version of <code>gcc<\/code> used to compile driver modules and the Linux kernel. <\/p>\n"},{"AnswerId":"56833148","CreationDate":"2019-07-01T09:27:53.007","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>If you are showing video from non-nvidia device but have driver installed, you have to install it with \u201c--no-opengl-files\u201d flag, for Gnome to work.<\/p>\n\n<p>I suggest to download a separate driver and install it manually by logging to console:<\/p>\n\n<pre><code>1. Alt Ctrl F2\/f3\/f4\/f5 to get to console.\n2. \u201cinit 3\u201d  to kill UI\n3. relogin if necessary to console\n4. wget http:\/\/us.download.nvidia.com\/tesla\/418.67\/NVIDIA-Linux-\n<\/code><\/pre>\n\n<p>driver x86_64-418.67.run<\/p>\n\n<pre><code>5. sh NVIDIA-Linux-x86_64-418.67.run --no-opengl-files\n6. After installation - reboot\n<\/code><\/pre>\n"},{"AnswerId":"32556866","CreationDate":"2015-09-14T03:24:42.460","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>Now the problem is solved.\nI checked <code>sudo dpkg --list | grep nvidia<\/code>\nI found as my kernel has 352.41, but the client has 304.12.\nSo I did <code>sudo apt-get remove --purge nvidia-*<\/code>. It removed all packages.\nThen, install 352.41 as<\/p>\n\n<pre><code>$ sudo add-apt-repository ppa:xorg-edgers\/ppa -y\n$ sudo apt-get update\n$ sudo apt-get install nvidia-352\n<\/code><\/pre>\n\n<p>After that <\/p>\n\n<pre><code>$ sudo dpkg --list | grep nvidia\nrc nvidia-304 304.128-0ubuntu0~gpu14.04.2 amd64 NVIDIA legacy binary driver - version 304.128\nrc nvidia-304-updates 304.125-0ubuntu0.0.2 amd64 NVIDIA legacy binary driver - version 304.125\nii nvidia-352 352.41-0ubuntu0~gpu14.04.1 amd64 NVIDIA binary driver - version 352.41\nrc nvidia-opencl-icd-304 304.128-0ubuntu0~gpu14.04.2 amd64 NVIDIA OpenCL ICD\nrc nvidia-opencl-icd-304-updates 304.125-0ubuntu0.0.2 amd64 NVIDIA OpenCL ICD\nii nvidia-opencl-icd-352 352.41-0ubuntu0~gpu14.04.1 amd64 NVIDIA OpenCL ICD\nii nvidia-prime 0.6.2 amd64 Tools to enable NVIDIA's Prime\nii nvidia-settings 355.11-0ubuntu0~gpu14.04.1 amd64 Tool for configuring the NVIDIA graphics driver\n<\/code><\/pre>\n\n<p>Now version matches.\nThen .\/deviceQuery and all work as expected.\nThanks<\/p>\n"},{"AnswerId":"32942278","CreationDate":"2015-10-05T06:13:28.900","ParentId":null,"OwnerUserId":"5408646","Title":null,"Body":"<p>I have this problem too. And re-installing the nvidia drivers didn't solve the issue. <\/p>\n\n<p>Finally, I solved this problem by add two kernel parameters with grub.<\/p>\n\n<p>add in:<\/p>\n\n<pre><code>GRUB_CMDLINE_LINUX_DEFAULT\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>pci=nocrs pci=realloc\n<\/code><\/pre>\n\n<p>I think this is a collision between <code>cuda7.5<\/code> and <code>kernel3.19<\/code>.<\/p>\n"},{"AnswerId":"51078813","CreationDate":"2018-06-28T09:04:04.137","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>Another way I can do is install using .run file.\nThat needs to kill X server first.\nX server is killed as follow.<\/p>\n\n<pre><code>Make sure you are logged out.\nHit CTRL+ALT+F1 and login using your credentials.\nkill your current X server session by typing sudo service lightdm stop or sudo stop lightdm\nEnter runlevel 3 (or 5) by typing sudo init 3 (or sudo init 5) and install your .run file.\nYou might be required to reboot when the installation finishes. If not, run sudo service start lightdm or sudo start lightdm to start your X server again.\n<\/code><\/pre>\n\n<p>Then <code>run .run file as sudo sh xxxxx.run<\/code><\/p>\n\n<p>You may get error as <code>The distribution-provided pre-install script failed! Are you sure you want to continue?<\/code>. Then abort the installation and  <\/p>\n\n<pre><code>disable the \"Nouveau kernel driver\" as sudo update-initramfs -u\n<\/code><\/pre>\n\n<p>Then reboot the system and <code>redo stop X server, enter runlevel 3 and do sudo sh xxxx.run again.<\/code><\/p>\n\n<p>This time you can ignore the message and continue for that prescript fail message.\nThen you will be able to install Nvidia Driver from .run file.<\/p>\n"}]}
{"QuestionId":32499160,"AnswerCount":1,"Tags":"<permissions><scipy><permission-denied><theano>","CreationDate":"2015-09-10T10:19:00.270","AcceptedAnswerId":32500869.0,"OwnerUserId":639973.0,"Title":"Permission denied due to scipy while installing Theano","Body":"<p>I'm installing Theano on a server where I'm not the root.<\/p>\n\n<p>I ran<\/p>\n\n<pre><\/pre>\n\n<p>which returns the following error<\/p>\n\n<pre><\/pre>\n\n<p>so apparently, Theano wants to install scipy, but it's already installed, so it attempts to uninstall it first, which brings the permission issue.<\/p>\n\n<p>How can I go around it so as to not uninstall scipy, but use the existing one?<\/p>\n","answers":[{"AnswerId":"32500869","CreationDate":"2015-09-10T11:38:02.227","ParentId":null,"OwnerUserId":"4511818","Title":null,"Body":"<p>The problem is that the scipy-version you have installed is not recommended. Theano usually needs at least version 0.11 to work. It seems that your version is also working but has some known bugs. (<a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow\">Installation Instructions<\/a>) If you wish to use your old version and risk the bugs, you should be able to use:<\/p>\n\n<pre><code>pip install Theano --user --no-dependencies\n<\/code><\/pre>\n\n<p>Note that the other two requirements numpy and six will also not be checked and updated<\/p>\n"}]}
{"QuestionId":32504394,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2015-09-10T14:16:50.080","AcceptedAnswerId":32504940.0,"OwnerUserId":1103412.0,"Title":"Batch processing mode in Caffe","Body":"<p>I'd like to use the Caffe library to extract image features but I'm having performance issues. I can only use the CPU mode. I was told Caffe supported batch processing mode, in which the average time required to process one image was much slower.<\/p>\n\n<p>I'm calling the following method:<\/p>\n\n<pre><\/pre>\n\n<p>and I'm putting in a vector of size 1, containing a single blob of the following dimensions - (num: 10, channels: 3, width: 227, height: 227). It represents a single image oversampled in the same way as in the official python wrapper.<\/p>\n\n<p>This works and gives correct results. It is, however, too slow.<\/p>\n\n<p>Whenever I try to send in a vector containing more than one blob (of the same dimensions), I get the following error:<\/p>\n\n<blockquote>\n  <p>F0910 16:10:14.848492 15615 blob.cpp:355] Trying to copy blobs of different sizes.<br>\n  Check failure stack trace:<\/p>\n<\/blockquote>\n\n<p>How do I make Caffe process my images in a batch?<\/p>\n","answers":[{"AnswerId":"32504940","CreationDate":"2015-09-10T14:39:40.497","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you want to feed larger batches you need the first (and only) blob in <code>bottom<\/code> to have <code>num&gt;10<\/code>. Feeding a blob with <code>num=20<\/code> is the same as feeding two inputs with <code>oversample=10<\/code>. You will, of course, have to perform the averaging manually according to the <code>oversampling<\/code> you are using.<\/p>\n\n<p>Furthermore, you might want to change the first input dimension in your <code>deploy.prototxt<\/code> file from 10 to some larger value (depending on your machine's memory capacity)<\/p>\n"}]}
{"QuestionId":32504728,"AnswerCount":1,"Tags":"<python><python-3.x><theano><ode>","CreationDate":"2015-09-10T14:31:12.033","AcceptedAnswerId":42079900.0,"OwnerUserId":99989.0,"Title":"How do I use Theano to solve an ordinary differential equation?","Body":"<p>Here is my Python code:<\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to use Theano to solve the ODE?<\/p>\n","answers":[{"AnswerId":"42079900","CreationDate":"2017-02-07T00:51:51.727","ParentId":null,"OwnerUserId":"3583290","Title":null,"Body":"<p>Here is a VERY simple ode solver in theano:<\/p>\n\n<pre><code>import numpy\nimport theano\n\n# the right-hand side\ndef f(x, t):\n    return x*(1-x)\n\nx = theano.tensor.matrix() # why not a matrix\ndt = theano.tensor.scalar()\nt = theano.tensor.scalar()\n\nx_next = x + f(x, t)*dt # implement your favourite RK method here!\n\n# matrix of random initial values\n# store it on the device\nx_shared = theano.shared(numpy.random.rand(10, 10))\n\nstep = theano.function([t, dt], [],\n        givens=[(x, x_shared)],\n        updates=[(x_shared, x_next)],\n        on_unused_input='warn')\n\nt = 0.0\ndt = 0.01\n\nwhile t &lt; 10:\n    step(t, dt)\n    t += dt\n    # test halt condition here\n\nprint(x_shared.get_value()) # read back the result\n<\/code><\/pre>\n\n<p>I hope it helps.\nBasically you have to implement your Runge-Kutta method.<\/p>\n\n<p>And mind that the strength of theano is in vectorization, so I won't bother implementing the time loop in theano. That's why I used a simple python while loop, although I could use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow noreferrer\">theano scan<\/a>. Anyway, depending on your goal, the optimization can be tricky. I'm not 100% convinced that theano is a good choice for an ODE solver. Numpy does the vectorizations on your force and position matrices anyway, at least on CPU. With a theano implementation you can utilize GPU, but that's not a guarantee for a speedup.<\/p>\n"}]}
{"QuestionId":32505458,"AnswerCount":3,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2015-09-10T15:04:23.923","AcceptedAnswerId":32505913.0,"OwnerUserId":1103412.0,"Title":"Reduce a Caffe network model","Body":"<p>I'd like to use Caffe to extract image features. However, it takes too long to process an image, so I'm looking for ways to optimize for speed.<\/p>\n\n<p>One thing I noticed is that the network definition I'm using has four extra layers on top the one from which I'm reading a result (and there are no feedback signals, so they should be safe to delete).<\/p>\n\n<p>I tried to delete them from the definition file but it had no effect at all. I guess I might need to remove the corresponding part of the file that contains pre-trained weights, too. That is, however, a binary file (a protobuffer) so editing it is not that easy.<\/p>\n\n<p>Do you think that removing the four layers might have a profound effect of the net performance?<\/p>\n\n<p>If so then how do I get familiar with the file contents so that I could edit it and how do I know which parts to remove?<\/p>\n","answers":[{"AnswerId":"32505913","CreationDate":"2015-09-10T15:25:13.650","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>first, I don't think removing the binary weights will have any effect.<br>\nSecond, you can do it easily using the python interface: see <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow noreferrer\">this tutorial<\/a>.<br>\nLast but not least, have you tried running <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html\" rel=\"nofollow noreferrer\"><code>caffe time<\/code><\/a> to measure the performance of your net? this may help you identify the bottlenecks of your computations.<\/p>\n\n<p>PS,\nYou might find <a href=\"https:\/\/stackoverflow.com\/q\/30822009\/1714410\">this thread<\/a> relevant as well.<\/p>\n"},{"AnswerId":"49224090","CreationDate":"2018-03-11T19:21:34.137","ParentId":null,"OwnerUserId":"1060382","Title":null,"Body":"<p>I would retrain on a smaller input size, change strides, etc. However if you want to reduce file size, I'd suggest quantizing the weights <a href=\"https:\/\/github.com\/yuanyuanli85\/CaffeModelCompression\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yuanyuanli85\/CaffeModelCompression<\/a> and then using something like lzma compression (xz for unix). We do this so we can deploy to mobile devices. 8 bit weights compress nicely.<\/p>\n"},{"AnswerId":"37579452","CreationDate":"2016-06-01T21:55:51.403","ParentId":null,"OwnerUserId":"5326903","Title":null,"Body":"<p>Caffemodel stores data as key-value pair. Caffe only copies weight for those layers (in train.prototxt) having exactly same name as caffemodel. Hence I don't think removing binary weights will work. If you want to change network structure, just modify train.prototxt and deploy.txt.<\/p>\n\n<p>If you insist to remove weights from binary file, follow this <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">caffe example<\/a>.<\/p>\n\n<p>And to make sure you delete right part, this <a href=\"http:\/\/ethereon.github.io\/netscope\/#\/editor\" rel=\"nofollow\">visualizing tool<\/a> should help.<\/p>\n"}]}
{"QuestionId":32507202,"AnswerCount":1,"Tags":"<machine-learning><neural-network><classification><theano><lasagne>","CreationDate":"2015-09-10T16:30:58.843","AcceptedAnswerId":null,"OwnerUserId":5321075.0,"Title":"neural network with lasagne accuracy","Body":"<p>I'm trying to construct a binary classifier with a neural network on some images using Lasagne. The training and validation loss fluctuate wildly (and do not settle) and the validation accuracy is always at . Furthermore, the network always predicts the target as  for the test set.<\/p>\n\n<p>The network I am using is basically just a copy of Lasagne's example for the mnist dataset found <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow\">here<\/a>, but adapted for my images which are quite a bit larger () with around  images in the training set. I am wondering if this is a problem, and if the network may need to be deeper \/ have more neurons? <\/p>\n\n<p>Do I need a larger training set for this size of image? Or should I be seeing some, albeit inaccurate, set of predictions for my test set?<\/p>\n","answers":[{"AnswerId":"32515310","CreationDate":"2015-09-11T04:11:06.493","ParentId":null,"OwnerUserId":"5226958","Title":null,"Body":"<p>I'd resize the images into smaller ones. Since your training examples are so limited you probably do not want to train a big model which easily overfits.<\/p>\n\n<p>The following tricks may also be useful for you:<\/p>\n\n<ul>\n<li><p>check whether your images are subtracted some mean value. If your input values are raw pixels between [0,255], that will be too big. <\/p><\/li>\n<li><p>try different learning rates. If your result fluctuates it is possible your learning rate is too high.<\/p><\/li>\n<li><p>use data augmentation. You may flip your images, or move it up\/down\/left\/right some pixels. Then you can get more training examples.<\/p><\/li>\n<li><p>look at training set. See where your model makes mistakes. If your training error is bad, the there must be something wrong.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":32519272,"AnswerCount":1,"Tags":"<theano><keras>","CreationDate":"2015-09-11T08:46:43.920","AcceptedAnswerId":null,"OwnerUserId":417357.0,"Title":"Add AUC as loss function for keras","Body":"<p>Has anyone had any luck with writing a custom AUC loss function for Keras using Theano?<\/p>\n\n<p>The documentation is here: <a href=\"http:\/\/keras.io\/objectives\/\" rel=\"noreferrer\">http:\/\/keras.io\/objectives\/<\/a><\/p>\n\n<p>Sample code is here: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/objectives.py\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/objectives.py<\/a><\/p>\n\n<p>I saw there is an implementation in pylearn2, (which is really a wrapper around sklearn) but was unable to port this to use in Keras<\/p>\n\n<p><a href=\"https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/train_extensions\/roc_auc.py\" rel=\"noreferrer\">https:\/\/github.com\/lisa-lab\/pylearn2\/blob\/master\/pylearn2\/train_extensions\/roc_auc.py<\/a><\/p>\n\n<p>So i guess my question is, has anybody been able to write this function? and would you be willing to share?<\/p>\n","answers":[{"AnswerId":"36970788","CreationDate":"2016-05-01T19:10:03.990","ParentId":null,"OwnerUserId":"1397061","Title":null,"Body":"<p>AUC is not differentiable, so you can't use it as a loss function without some modification.  There's been <a href=\"http:\/\/www.icml-2011.org\/papers\/198_icmlpaper.pdf\">some work<\/a> on algorithms to maximize AUC, but I'd recommend just using the regular cross-entropy \/ log likelihood loss.<\/p>\n"}]}
{"QuestionId":32520049,"AnswerCount":1,"Tags":"<macos><python-3.x><theano>","CreationDate":"2015-09-11T09:25:01.703","AcceptedAnswerId":32522028.0,"OwnerUserId":5324587.0,"Title":"I can't install Theano on Mac OS X 10.10.5","Body":"<p>I'm trying to install Theano (and subsequently pylearn2) with Python 3.4 on my MacBookPro, with Mac OS X 10.10.5. I have Anaconda and I follow the instructions reported on Theano's documentation (<a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install.html<\/a>). I have to use the sudo -H pip command and Theano gets downloaded, a Theano folder is created in anaconda and all the required dependencies satisfying the versions requisites. However, when i write the<\/p>\n\n<blockquote>\n  <p>import theano<\/p>\n<\/blockquote>\n\n<p>command in Python, I get (both in the terminal and in ipython-qtconsole) the following:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 1, in <\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/<strong>init<\/strong>.py\", line 44, in \n      from theano.gof import \\<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/<strong>init<\/strong>.py\", line 38, in \n      from theano.gof.cc import \\<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/cc.py\", line 64, in \n      from theano.gof import link<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/link.py\", line 12, in \n      from theano.gof.type import Type<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/type.py\", line 14, in \n      from theano.gof.op import CLinkerObject<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/op.py\", line 31, in \n      from theano.gof.cmodule import GCC_compiler<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/cmodule.py\", line 37, in \n      from theano.gof.compiledir import gcc_version_str, local_bitwidth<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/compiledir.py\", line 259, in \n      in_c_key=False)<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/sitepackages\/theano\/configparser.py\", line 237, in AddConfigVar configparam.<strong>get<\/strong>()<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/configparser.py\", line 279, in <strong>get<\/strong>\n      self.<strong>set<\/strong>(None, val_str)<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/configparser.py\", line 290, in <strong>set<\/strong>\n      self.val = self.filter(val)<\/p>\n  \n  <p>File \"\/Users\/davidefloriello\/anaconda\/lib\/python3.4\/site-packages\/theano\/gof\/compiledir.py\", line 185, in filter_compiledir\n      \" or listing permissions.\" % path)\n  ValueError: compiledir '\/Users\/davidefloriello\/.theano\/compiledir_Darwin-14.5.0-x86_64-i386-64bit-i386-3.4.3-64' exists but you don't have read, write or listing permissions.<\/p>\n<\/blockquote>\n\n<p>This last line is quite surprising since I'm the only admin. I gave a look at this other question <a href=\"https:\/\/stackoverflow.com\/questions\/24728255\/how-to-install-theano-library-on-os-x\">How to install theano library on OS X?<\/a> but I haven't found it very useful, as our Python versions are different and it seems to me that my problem is a bit less vague.<\/p>\n\n<p>Any help is very much appreciated!<\/p>\n\n<p>Thanks a lot,<\/p>\n\n<p>Davide <\/p>\n","answers":[{"AnswerId":"32522028","CreationDate":"2015-09-11T11:09:40.003","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>A similar problem was found reported <a href=\"https:\/\/github.com\/dnouri\/kfkd-tutorial\/issues\/9\" rel=\"nofollow\">here<\/a>. The solution, updated to account for different user, is:<\/p>\n\n<blockquote>\n  <p>Try and remove that '\/Users\/davidefloriello\/.theano\/' directory and\n  run Python again without sudo.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":32522469,"AnswerCount":1,"Tags":"<c++><image-processing><deep-learning><caffe>","CreationDate":"2015-09-11T11:32:43.670","AcceptedAnswerId":32531407.0,"OwnerUserId":1103412.0,"Title":"Batch processing mode in Caffe - no performance gains","Body":"<p>Following on <a href=\"https:\/\/stackoverflow.com\/q\/32504394\/1103412\">this thread<\/a> I reimplemented my image processing code to send in 10 images at once (i.e. I now have the num property of the input blob set to 100 instead of 10).<\/p>\n\n<p>However, the time required to process this batch is 10 times bigger than originally. Which means that I did not get any performance increase.<\/p>\n\n<p>Is that reasonable or did I make something wrong?<\/p>\n\n<p>I am running Caffe in CPU mode. Unfortunately GPU mode is not an option for me.<\/p>\n","answers":[{"AnswerId":"32531407","CreationDate":"2015-09-11T20:10:48.963","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>Update: Caffe now natively supports parallel processing of multiple images when using multiple GPUs. Though it seems relatively simple to implement base on the current implementation of GPU parallelism, at the moment there's no similar support for parallel processing on multiple CPUs.<\/p>\n\n<p>Considering that the main problem with implementing parallelism is the syncing you need during training If you just want to process your images in parallel (as opposed to training the model), then you could load several copies of the same network to memory (whether through python with multiprocessing or c++ with multi-threading), and process each image on a different network. It would be simple and quite effective, especially if you load the networks once and then just process a large amount of images. Nevertheless, GPUs are much faster :)<\/p>\n\n<hr>\n\n<p>Caffe doesn't process multiple images in parallel, the only saving you get by batch processing several images is in the time it takes to transfer the image data back and forth between Caffe's framework, which could be significant when dealing with the GPU.<\/p>\n\n<p>IIRC there are several attempts to make Caffe process images in parallel, but most focus on the GPU implementation (CUDNN, CUDA Streams etc.), with few attempts to add parallelism to the CPU code (OpenBLAS's multithread mode, or simply running on multiple threads).  Of those I believe only the CUDNN option is currently part of the stable version of Caffe, but obviously requires a GPU. You can try to look at one of the pull requests about this matter on Caffe's github page and see if it works for you, but note that it might cause compatibilities issue with your current version.<\/p>\n\n<p>This is one such version that in the past I've used, though it's no longer maintained: <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/439\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/439<\/a><\/p>\n\n<p>I've also noticed in the last comment of the above issue that there's some speed up to the CPU code on this pull request as well, though I've never tried it myself: <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2610\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/2610<\/a><\/p>\n"}]}
{"QuestionId":32527603,"AnswerCount":0,"Tags":"<theano><lasagne>","CreationDate":"2015-09-11T16:01:43.397","AcceptedAnswerId":null,"OwnerUserId":1001019.0,"Title":"LSTMLayer produces NaN values even before training it","Body":"<p>I'm currently trying to construct a LSTM network with Lasagne to predict the next step of noisy sequences. I first trained a stack of 2 LSTM layers for a while, but had to use an abysmally small learning rate (1e-6) because of divergence issues (that ultimately produced NaN values). The results were kind of disappointing,\u00a0as the network produced smooth, out-of-phase versions of the input.<\/p>\n\n<p>I then came to the conclusion I should use better parameter initialization than what is given by default. The goal was to start from a network that just mimics identity, since for strongly auto-correlated signal it should be a good first estimation of the next step (x(t) ~ x(t+1)), and to sprinkle a bit of noise on top of it.<\/p>\n\n<pre><\/pre>\n\n<p>I then use this lstm generation code to generate the following network:<\/p>\n\n<pre><\/pre>\n\n<p>Problem is, even without any training, this network produces garbage values and sometimes even a bunch of NaNs, right from the very first LSTM layer:<\/p>\n\n<pre><\/pre>\n\n<p>I don't get why, I checked each matrices and their values are fine, like I wanted them to be. I even tried to recreate each gate activations and the resulting hidden activations using the actual numpy arrays and they reproduce the input just fine. What did I do wrong there??<\/p>\n","answers":[]}
{"QuestionId":32529056,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-09-11T17:30:40.013","AcceptedAnswerId":33463358.0,"OwnerUserId":3261934.0,"Title":"Why theano scan works differently when the codes are nearly the same?","Body":"<p>The code below:<\/p>\n\n<pre><\/pre>\n\n<p>have an error:<\/p>\n\n<pre><\/pre>\n\n<p>But when I change 1 to any other number, for example:<\/p>\n\n<pre><\/pre>\n\n<p>It works fine.<\/p>\n\n<p>I don't know what happened here. Could someone help me?<\/p>\n","answers":[{"AnswerId":"33463358","CreationDate":"2015-11-01T14:41:15.117","ParentId":null,"OwnerUserId":"2859669","Title":null,"Body":"<p>Please follow this post: <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2985\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2985<\/a><\/p>\n\n<p>Passing a tensor whose shape includes 1 as a part of outputs_info when calling theano.scan fails unless those axes of shape 1 are unbroadcasted manually using tensor.unbroadcast. This is due to the different broadcasting pattern between the actual return from the inner function of scan and the corresponding one passed via outputs_info.<\/p>\n\n<p>Try:<\/p>\n\n<pre><code>h1=T.unbroadcast(T.as_tensor_variable(np.zeros((1, 20), dtype=theano.config.floatX)), 0)\ns1=T.unbroadcast(T.as_tensor_variable(np.zeros((1, 20), dtype=theano.config.floatX)), 0)\n<\/code><\/pre>\n\n<p>to make the first dimension unbroadcastable. <\/p>\n"}]}
{"QuestionId":32536426,"AnswerCount":2,"Tags":"<computer-vision><neural-network><convolution><deep-learning><caffe>","CreationDate":"2015-09-12T07:40:09.220","AcceptedAnswerId":32547133.0,"OwnerUserId":639973.0,"Title":"fine-tuning a CNN from a lower fc layer","Body":"<p>I've noticed that most fine-tuning of CNN over new dataset is done only on the \"last\" fully connected (fc) layer.<\/p>\n\n<p>I'm interested in fine-tuning from the \"first\" fully connected layer: that is, I want to use mid-level features from convolution and pooling layers as they are, (supposing it's trained on ImageNet) but then fit all fc layers to my new dataset.<\/p>\n\n<p>Theoretically and in practice, what is the supposed effect of this? Is it likely to learn a more proper set of parameters for my new dataset?<\/p>\n","answers":[{"AnswerId":"32547133","CreationDate":"2015-09-13T06:58:58.803","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Theoretically, the deeper you fine-tune, the better your model fits your data. So, if you could fine-tune the whole model - the better.  <\/p>\n\n<p>So, what's the catch, you must be asking, why don't everyone fine-tune the whole model?<br>\nFirst, fine-tuning the whole model involves lots and lots of parameters, in order to train properly millions of parameters without the risk of overfitting, you must have a LOT of new training examples. In most cases, when fine-tuning you only have a very few annotated samples for the new task and therefore you are unable to afford fine-tuning of the whole model.<br>\nSecond, fine-tuning the whole model takes much longer than training just the top fc layer. Thus, if you have little time and budget you only fine-tune the top fc layer.<\/p>\n\n<p>In your case, if you have enough samples you may fine-tune the top two fc layers. From my experience it is better to fine tune the top layer first and then fine-tune the top two together after some iterations are done on the top layer alone. <\/p>\n"},{"AnswerId":"42738900","CreationDate":"2017-03-11T18:20:30.483","ParentId":null,"OwnerUserId":"3718030","Title":null,"Body":"<p>The purpose of FC layers in a ConvNet is only to perform classification for your problem. You could just use the final flattened output from your last Conv\/Pooling layer as engineered features and put it in another machine learning model and it would have the same effect.<\/p>\n\n<p>This implies that the parameters learned by FC layers in most cases are very problem specific (depends on the data) and, in most cases not transferable. <\/p>\n\n<p>So whenever people fine tune on a pre-trained model they almost always dump the FC layers on the top.<\/p>\n\n<p>Now you can go 2 ways from here.<\/p>\n\n<ol>\n<li>Use the final flattened output from the last Conv\/Pooling layer as extracted features for your problem and train a ML model on it. This method is commonly used if your data set is small or not similar to the pre-trained model.<\/li>\n<li>Get the extracted features using the above method and then use them to train a FC neural network. Once you achieved a decent accuracy stack it on top of last conv\/pooling layer of the pre-trained model (dont forget to remove the orignal FC layers). Now freeze(parameters are made constant and dont change on training) most of the pre-trained model and only allow the last few conv layers to be trained. Now train the entire network to train with a very small learning rate.<\/li>\n<\/ol>\n\n<p>The whole point of freezing most of the model is that we assume that the model already knows basic stuff like edge detection and color from the earlier conv layers. Now we fine tune the last few layers for the our problem. We chose a small learning rate so that we don't end up messing up what the model has already learned. <\/p>\n\n<p>The reason we train the FC layers before we fit them on to the pre-trained model is just to save training time and more importantly to ensure that we don't make to much changes to the Conv layers and end up over fitting. <\/p>\n"}]}
{"QuestionId":32538758,"AnswerCount":2,"Tags":"<python><ipython><caffe><pycaffe>","CreationDate":"2015-09-12T12:25:27.750","AcceptedAnswerId":32539282.0,"OwnerUserId":2452617.0,"Title":"NameError: name 'get_ipython' is not defined","Body":"<p>I am working on Caffe framework and using PyCaffe interface. I am using a Python script obtained from converting the IPython Notebook <strong>00-classification.ipynb<\/strong> for testing the classification by a trained model for ImageNet. But any <strong>get_ipython()<\/strong> statement in the script is giving the following error:<\/p>\n\n<pre><\/pre>\n\n<p>In the script, I'm importing the following:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone please help me to resolve this error?<\/p>\n","answers":[{"AnswerId":"32539282","CreationDate":"2015-09-12T13:17:25.860","ParentId":null,"OwnerUserId":"1048100","Title":null,"Body":"<p>You have to run your script with ipython:<\/p>\n\n<pre><code>$ ipython python\/my_test_imagenet.py\n<\/code><\/pre>\n\n<p>Then <code>get_ipython<\/code> will be already in global context.<\/p>\n\n<p>Note: Importing it via <code>from IPython import get_ipython<\/code> in ordinary shell <code>python<\/code> will not work as you really need <code>ipython<\/code> running.<\/p>\n"},{"AnswerId":"48032629","CreationDate":"2017-12-30T09:27:33.323","ParentId":null,"OwnerUserId":"207661","Title":null,"Body":"<p>If your intention is to run converted .py file notebook then you should just comment out <code>get_ipython()<\/code> statements. The matlibplot output can't be shown inside console so you would have some work to do . Ideally, iPython shouldn't have generated these statements. You can use following to show plots:<\/p>\n\n<pre><code>plt.show(block=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":32553374,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><theano>","CreationDate":"2015-09-13T18:55:14.593","AcceptedAnswerId":null,"OwnerUserId":3301357.0,"Title":"How can I get not only an unrolled for k steps truncated-BPTT grad in theano 'scan', but also an each(1st through k-th) of components of that grad?","Body":"<p>If I use theano.scan with a truncate_gradient=k(int >0), I can obtain a truncated-BPTT gradient for a recurrent neural network. And that gradient will be calculated for an unrolled for k steps backwards in time RNN(which becomes k-layered feed forward neural network FFNN).\nBasically(according to theory), that truncated gradient for RNN(calculated by unrolling RNN to k steps backwards in time) should consist of sum of gradients for all k layers of unrolled FFNN(obtained from unrolling RNN to k steps backwards).<\/p>\n\n<p>Basically, I'm looking how to get each component of that sum in theano...\nAs it seems to me, that you can easily get the resulting truncated grad of  an RNN(which is represented by a scan op with truncate_gradient=k)(this functionality works out of the box), but it's quite tricky to get each component of that unrolled sum for truncated gradients for scan op\/RNN.<\/p>\n\n<p><strong>What I already tried:<\/strong><\/p>\n\n<p>I looked through the theano scan op internals, especially through the 'grad' method... And it's code is quite complicated.<\/p>\n\n<p>Also I tried to print graphs for resulting gradients - i tried to print graphs for the same RNN with truncate_gradients=1 and truncate_gradients=10 (using theano.printing.pydotprint) - These graphs were the same :(<\/p>\n\n<p>Then I used theano.printing.pydotprint with scan_graphs=True\nThat had to print internals of scan. But I failed with such exception:<\/p>\n\n<p>AttributeError: 'Scan' object has no attribute 'fn'.<\/p>\n\n<p>Also I googled for these questions - nothing.\nAsked in theano-users - haven't got any response yet<\/p>\n","answers":[]}
{"QuestionId":32563651,"AnswerCount":1,"Tags":"<lua><neural-network><batch-processing><torch>","CreationDate":"2015-09-14T11:27:08.540","AcceptedAnswerId":32570711.0,"OwnerUserId":350605.0,"Title":"Batch processing in Torch with ClassNLLCriterion","Body":"<p>I'm trying to implement a simple NN in Torch to learn more about it. I created a very simple dataset: binary numbers from 0 to 15 and my goal is to classify the numbers into two classes - class 1 are numbers 0-3 and 12-15, class 2 are the remaining ones. The following code is what i have now (i have removed the data loading routine only):<\/p>\n\n<pre><\/pre>\n\n<p>This is how the data and class Tensors look like:<\/p>\n\n<pre><\/pre>\n\n<p>Which is what I expect it to be. However when running this code, i get the following error on line <strong>loss = criterion:forward( prediction, class )<\/strong>:<\/p>\n\n<blockquote>\n  <p>torch\/install\/share\/lua\/5.1\/nn\/ClassNLLCriterion.lua:69: attempt to\n  perform arithmetic on a nil value<\/p>\n<\/blockquote>\n\n<p>When i modify the training routine like this (processing a single data point at a time instead of all 16 in a batch) it works and the network successfully learns to recognize the two classes:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure what might be wrong with the \"batch processing\" i'm trying to do. A brief look at the ClassNLLCriterion didn't help, it seems i'm giving it the expected input (see below), but it still fails. The input it receives (prediction and class Tensors) looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone help me out here? Thanks.<\/p>\n","answers":[{"AnswerId":"32570711","CreationDate":"2015-09-14T17:42:17.660","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>Experience has shown that <code>nn.ClassNLLCriterion<\/code> expects target to be a <strong>1D tensor<\/strong> of size <code>batch_size<\/code> or a <strong>scalar<\/strong>. Your <code>class<\/code> is a 2D one (<code>batch_size x 1<\/code>) but <code>class[i]<\/code> is 1D, that's why your non-batch version works. <\/p>\n\n<p>So, this will solve your problem:<\/p>\n\n<pre><code>class = class:view(-1)\n<\/code><\/pre>\n\n<p>Alternatively, you can replace<\/p>\n\n<pre><code>network:add( nn.LogSoftMax() )\ncriterion = nn.ClassNLLCriterion()\n<\/code><\/pre>\n\n<p>with the equivalent:<\/p>\n\n<pre><code>criterion = nn.CrossEntropyCriterion()\n<\/code><\/pre>\n\n<p>The interesting thing is that <code>nn.CrossEntropyCriterion<\/code> is also able to take a <strong>2D tensor<\/strong>. Why is <code>nn.ClassNLLCriterion<\/code> not?<\/p>\n"}]}
{"QuestionId":32565102,"AnswerCount":0,"Tags":"<python><theano><lasagne><nolearn>","CreationDate":"2015-09-14T12:42:13.403","AcceptedAnswerId":null,"OwnerUserId":2902280.0,"Title":"Nolearn\/Lasagne neural network not starting training","Body":"<p>I am using ,  and  to train a neural net on GPU (in an  notebook). However, because of the first layer (), the following network does not start training, aven after waiting a few hours:<\/p>\n\n<pre><\/pre>\n\n<p>If I comment out the first layer, the training starts in a few seconds. Training more complicated networks is not a problem either. Any idea of what's causing the issue?<\/p>\n\n<p><strong>Edit<\/strong>: oddly enough, if I remove  and , the training also starts within a reasonable time.<\/p>\n\n<p><strong>Edit2<\/strong>: what's even stranger is, if I change the size of the filters to 10 in the layer , then the training starts in a reasonable amount of time. If after that I stop the cell's execution, change this value to 1, and reexecute the cell, the training goes fine...<\/p>\n\n<p>Finally I've started using another framework, but if someone's interested, <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/lasagne-users\/not$20starting\/lasagne-users\/XR-jmU1pMWc\/Erym1kOhBQAJ\" rel=\"nofollow\">here's the link to the thread<\/a> I started on the lasagne user group.<\/p>\n","answers":[]}
{"QuestionId":32574251,"AnswerCount":0,"Tags":"<c++><protocol-buffers><deep-learning><caffe><gradient-descent>","CreationDate":"2015-09-14T21:29:10.757","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"Adjusting proto file for Caffe","Body":"<p>I'm trying to modify  in order to add 2 new fields to SolverParameter. the two lines I add, at the very end of the SolverParameter message are:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I rerun training nets that worked before, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The full protobuf message is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Also, surprisingly (to me), if the two new lines:<\/p>\n\n<pre><\/pre>\n\n<p>Are inserted just after <\/p>\n\n<pre><\/pre>\n\n<p>Then, the error I get is connected with a misreading\/parsing of the  variable (key id = 30).<\/p>\n\n<p>Does this make sense to anyone? Is there something obviously wrong with the way I'm inserting my 2 new fields? Is there some other code section I need to modify?<\/p>\n\n<p>Thanks....<\/p>\n","answers":[]}
{"QuestionId":32576858,"AnswerCount":1,"Tags":"<backpropagation><torch>","CreationDate":"2015-09-15T02:40:50.587","AcceptedAnswerId":32589478.0,"OwnerUserId":5118777.0,"Title":"How backpropagation works in torch 7?","Body":"<p>I tried to understand supervised learning by torch tutorial.<\/p>\n\n<p><a href=\"http:\/\/code.madbits.com\/wiki\/doku.php?id=tutorial_supervised\" rel=\"nofollow\">http:\/\/code.madbits.com\/wiki\/doku.php?id=tutorial_supervised<\/a><\/p>\n\n<p>And backpropagation :<\/p>\n\n<p><a href=\"http:\/\/home.agh.edu.pl\/~vlsi\/AI\/backp_t_en\/backprop.html\" rel=\"nofollow\">http:\/\/home.agh.edu.pl\/~vlsi\/AI\/backp_t_en\/backprop.html<\/a><\/p>\n\n<p>As I know, parameter update in this torch tutorial is in Step 4 Training Procedure,<\/p>\n\n<pre><\/pre>\n\n<p>For example, I got this<\/p>\n\n<pre><\/pre>\n\n<p>df_do is this ?<\/p>\n\n<pre><\/pre>\n\n<p>I know the target is 9 and the output is 4 in this example, so the result is wrong and give the 9-th element of df_do \"-1\".<\/p>\n\n<p>But why ?<\/p>\n\n<p>According to <a href=\"http:\/\/home.agh.edu.pl\/~vlsi\/AI\/backp_t_en\/backprop.html\" rel=\"nofollow\">http:\/\/home.agh.edu.pl\/~vlsi\/AI\/backp_t_en\/backprop.html<\/a>,<\/p>\n\n<p>df_do is [ target (desired output) - output ].<\/p>\n","answers":[{"AnswerId":"32589478","CreationDate":"2015-09-15T15:05:36.557","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>In Torch backprop works exactly as it does in mathematics. <code>df_do<\/code> is a <strong>derivative of loss w.r.t. prediction<\/strong>, and therefore entirely defined by your <a href=\"http:\/\/code.madbits.com\/wiki\/doku.php?id=tutorial_supervised_3_loss\" rel=\"nofollow noreferrer\">loss function<\/a>, i.e. <code>nn.Criterion<\/code>. \nThe most famous one is Mean Square Error (<code>nn.MSECriterion<\/code>):\n<a href=\"https:\/\/i.stack.imgur.com\/fsfI0.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fsfI0.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Note that MSE criterion expects target to have the same size as prediction (a one-hot vector for classification). If you choose MSE, your derivative vector <code>df_do<\/code> will be computed as: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Z2dr2.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Z2dr2.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The MSE criterion, however, is typically not very good for classification. The more suitable one is Likelihood criterion, which takes a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Probability_vector\" rel=\"nofollow noreferrer\">probability vector<\/a> as prediction and a scalar index of the true class as target. The aim is to simply maximize probability of the true class, that equals to minimization of its negative:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qEK3O.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qEK3O.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If we give it log-probability vector qua prediction (it is a monotone transformation and thus doesn't affect the optimization result but more computationally stable), we'll get the Negative Log Likelihood loss function (<code>nn.ClassNLLCriterion<\/code>):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9ME9n.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9ME9n.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In that case, <code>df_do<\/code> is as follows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/glQ7R.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/glQ7R.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In the torch tutorial <a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/3_loss.lua\" rel=\"nofollow noreferrer\">NLL criterion is used by default<\/a>.<\/p>\n"}]}
{"QuestionId":32578951,"AnswerCount":2,"Tags":"<input><deep-learning><classification><theano>","CreationDate":"2015-09-15T06:17:19.607","AcceptedAnswerId":32590548.0,"OwnerUserId":5121855.0,"Title":"How to get Elemwise{tanh,no_inplace}.0 value","Body":"<p>I am using <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\" rel=\"nofollow\">Deep learning Theano<\/a>. How can I see the content of a variable like this: . It is the input data of <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/blob\/master\/code\/logistic_sgd.py\" rel=\"nofollow\">logistic layer<\/a>. <\/p>\n","answers":[{"AnswerId":"32583060","CreationDate":"2015-09-15T09:57:17.340","ParentId":null,"OwnerUserId":"4511818","Title":null,"Body":"<p>Right now, you don't seem to print values but operations. The output <code>Elemwise{tanh,no_inplace}.0<\/code> means, that you have an element wise operation of tanh, that is not done in place. You still need to create a function that takes input and executes your operation. Then you need to call that function and print the result. You can read more about that in the graph-structure part of their <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/symbolic_graphs.html\" rel=\"nofollow\">tutorial<\/a>.<\/p>\n"},{"AnswerId":"32590548","CreationDate":"2015-09-15T15:56:25.320","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Suppose your variable is called <code>t<\/code>. Then you can evaluate it by calling <code>t.eval()<\/code>. This may fail if input data are needed. In that case you need to supply them by providing a dictionary like this <code>t.eval({input_var1: value1, input_var2: value2})<\/code>. This is the ad-hoc way of evaluating a theano-expression.<\/p>\n\n<p>The way it works in real programs is to create a function taking the necessary input, for example: <code>f = theano.function([input_var1, input_var2], t)<\/code>, will yield a function that takes two input variables, calculates <code>t<\/code> from them and outputs the result.<\/p>\n"}]}
{"QuestionId":32581184,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-09-15T08:23:35.313","AcceptedAnswerId":null,"OwnerUserId":4565947.0,"Title":"No module named nanguardmode","Body":"<p>I have a bug in my theano program leading to NaN values. The doc recommends using  to track down the source of the problem.<\/p>\n\n<p>When I copy\/paste this line from the doc webpage:<\/p>\n\n<pre><\/pre>\n\n<p>I get:<\/p>\n\n<pre><\/pre>\n\n<p>Can't find any sign of  when I type:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea why  is absent? How can I fix this?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Thanks for your replies.<\/p>\n\n<p>Concerning my Theano version, I'm couldn't find how to check it. But I assume it is the latest: I installed it form the install webpage about about a month ago. I'm on Windows 64bit.<\/p>\n\n<p>Concerning detect_nan hack: things just get weirder!<\/p>\n\n<p>First: if I try to use:<\/p>\n\n<pre><\/pre>\n\n<p>I get:  <\/p>\n\n<pre><\/pre>\n\n<p>Indeed, numpy was not imported in the monitormode module... Is that a known bug?<\/p>\n\n<p>Second: if I try to use a copy\/paste of detect_nan, the NaNs magically go away. Everything else remaining the same, without detect_nan in my theano function (that trains a model iteratively), I get NaNs at iteration 5:<\/p>\n\n<pre><\/pre>\n\n<p>(the last figure is the cost value)<\/p>\n\n<p>When I do add<\/p>\n\n<pre><\/pre>\n\n<p>to the function, no NaNs appear up to at least iteration 100 (and probably more).<\/p>\n\n<pre><\/pre>\n\n<p>What's going on here??? <\/p>\n","answers":[{"AnswerId":"32595409","CreationDate":"2015-09-15T20:52:40.590","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>NanGuardMode<\/code> was moved to Theano's bleeding edge version (from PyLearn2) on May 1st. This was after the release of version 0.7 on March 26th so you'll need to <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">upgrade to the bleeding edge version<\/a> from GitHub to use NanGuardMode.<\/p>\n\n<p>Alternatively you could use the <code>detect_nan<\/code> sample found in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#how-do-i-step-through-a-compiled-function\" rel=\"nofollow\">debug FAQ<\/a>:<\/p>\n\n<pre><code>import numpy\n\nimport theano\n\n# This is the current suggested detect_nan implementation to\n# show you how it work.  That way, you can modify it for your\n# need.  If you want exactly this method, you can use\n# ``theano.compile.monitormode.detect_nan`` that will always\n# contain the current suggested version.\n\ndef detect_nan(i, node, fn):\n    for output in fn.outputs:\n        if (not isinstance(output[0], numpy.random.RandomState) and\n            numpy.isnan(output[0]).any()):\n            print '*** NaN detected ***'\n            theano.printing.debugprint(node)\n            print 'Inputs : %s' % [input[0] for input in fn.inputs]\n            print 'Outputs: %s' % [output[0] for output in fn.outputs]\n            break\n\nx = theano.tensor.dscalar('x')\nf = theano.function([x], [theano.tensor.log(x) * x],\n                    mode=theano.compile.MonitorMode(\n                        post_func=detect_nan))\n<\/code><\/pre>\n"}]}
{"QuestionId":32587927,"AnswerCount":1,"Tags":"<python><deep-learning><caffe>","CreationDate":"2015-09-15T13:52:37.517","AcceptedAnswerId":32600424.0,"OwnerUserId":1103412.0,"Title":"Caffe - draw_net_to_file - 'Classifier' object has no attribute 'name'","Body":"<p>I found the  method in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/draw.py\" rel=\"nofollow\"><\/a> and want to use it to understand the Caffe network I was given to work with better.<\/p>\n\n<p>The problem is that the following code<\/p>\n\n<pre><\/pre>\n\n<p>fails with the following error<\/p>\n\n<pre><\/pre>\n\n<p>Upon closer investigation, the  object really does not expose many methods of the underlying  object, such as . How do I instantiate a correctly working  instance for this case?<\/p>\n\n<p>I'm using Caffe built from revision .<\/p>\n","answers":[{"AnswerId":"32600424","CreationDate":"2015-09-16T05:29:18.090","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Take a look at the script <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/draw_net.py\" rel=\"nofollow\"><code>draw_net.py<\/code><\/a> where you can see an example of how to use the functions of <code>draw.py<\/code>. The <code>net<\/code> argument is not exactly the same as the <code>caffe.Net<\/code> object but rather a parsed prototxt:<\/p>\n\n<pre><code>from google.protobuf import text_format\nimport caffe.draw\nfrom caffe.proto import caffe_pb2\n\nnet = caffe_pb2.NetParameter()\ntext_format.Merge(open(args.input_net_proto_file).read(), net)\n<\/code><\/pre>\n"}]}
{"QuestionId":32589349,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-09-15T14:59:39.107","AcceptedAnswerId":32600303.0,"OwnerUserId":1377127.0,"Title":"How many images can you pass to Caffe at a time?","Body":"<p>I noticed how the Caffe MNIST <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet.prototxt\" rel=\"nofollow\">example prototxt file<\/a> allows for up to 64 images to be passed to the network at a time.<br>\nIs there a limit for how high I can set this number?<br>\nCould I (for example) set this number to 200 or even 500 so that I can accept up to 200\/500 images at a time without it impacting the predictions negatively?<\/p>\n","answers":[{"AnswerId":"32600303","CreationDate":"2015-09-16T05:18:31.433","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The only limit is your machine's memory: When caffe loads the model it allocates memory for all the parameters <em>and<\/em> all the intermediate data blobs. The more images you process concurrently, the larger the memory you need to allocate in advance.<br>\nThe easiest (and crudest) way of determining this number is simply trail-and-error, try setting it to 200 and see if you get an \"out of memory\" error when loading the model.<br>\nNote that the number of images you can process at the same time depends also on whether you are using GPU or CPU: usually GPU memory is smaller than CPU memory and thus allows you to process fewer images.<\/p>\n"}]}
{"QuestionId":32601546,"AnswerCount":1,"Tags":"<ubuntu><g++><shared-libraries><ld><caffe>","CreationDate":"2015-09-16T06:50:35.897","AcceptedAnswerId":null,"OwnerUserId":5340867.0,"Title":"ld error while compiling caffe: libpng and libgfortran not found, shared libraries","Body":"<p>I'm trying to compile caffe on ubuntu 14.04 LTS with Anaconda, in CPU-only mode and with OpenBLAS. Unfortunately, I get an ld error.<\/p>\n\n<p>I followed the instructions, added the dependencies that didn't come with Anaconda and adjusted the Makefile.config accordingly, especially included the Anaconda path.\nWhen I do  I get the error you see below (I also included the Makefile.config), even though the two files that were not found are in the anaconda\/lib folder and in their respective pkgs folders as shared libraries.<\/p>\n\n<p>Thank you very much for your help!<\/p>\n\n<p>Terminal:<\/p>\n\n<pre><\/pre>\n\n<p>Makefile.config:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32604843","CreationDate":"2015-09-16T09:35:33.360","ParentId":null,"OwnerUserId":"5340867","Title":null,"Body":"<p>Okay it looks like i was able to solve this by myself. I don't know whether it's a good solution though, so I appreciate any comment: In the folder \/etc\/ld.so.conf.d I created a file fooLibrary.conf. This file contains only the full path to Anaconda's lib-folder. After <code>sudo ldconfig<\/code>, ld was able to find the relevant packages. Once I was done I deleted the file again, but I doubt the problem is solved conclusively. Hence I'll be glad about comments or better solutions.<\/p>\n"}]}
{"QuestionId":32614006,"AnswerCount":1,"Tags":"<theano><deep-learning>","CreationDate":"2015-09-16T16:28:55.600","AcceptedAnswerId":null,"OwnerUserId":3511203.0,"Title":"Is it normal that memory usage is fluctuating in Theano?","Body":"<p>I started to use Theano library, because I am seek of compile &amp; debugging with C++ with caffe ( though, it was a really great library :) ) <\/p>\n\n<p>Anyway, I made deep network(almost like CNN), with lasagne, and I started learning my network. However, my nvidia-smi shows that the memory usage keep fluctuating and I feel bad about it. It was not shown when I used caffe, and, because of this, learning could be slow.<\/p>\n\n<p>I used multiprocess module to fetch dataset in advance, and my queue status seems right, so loading a dataset could not be the case for my slow training.<\/p>\n\n<p>I used T.shared to allocate memory in GPU, in advance and make function with given variables. <\/p>\n\n<p>Any ideas?<\/p>\n\n<p>Thanks!\nHappy learning!<\/p>\n","answers":[{"AnswerId":"32615028","CreationDate":"2015-09-16T17:28:04.913","ParentId":null,"OwnerUserId":"3511203","Title":null,"Body":"<p>I find out allow_gc option, and after I turn off the option with allow_gc = False, everything goes fine. <\/p>\n\n<p>My GPU-Utilization shows better, and learning is faster now. <\/p>\n"}]}
{"QuestionId":32628838,"AnswerCount":1,"Tags":"<python><image-processing><neural-network><theano><lasagne>","CreationDate":"2015-09-17T11:06:04.870","AcceptedAnswerId":null,"OwnerUserId":5321075.0,"Title":"Convolutional Net With Lasange - image resizing","Body":"<p>I've been training some neural and convolutional nets with Lasagne and was doing most of my data\/image preprocessing using Python. However I would like to incorporate some of this into my Lasagne layers to make my code more flexible. <\/p>\n\n<p>Is there a Lasagne layer which can resize the input images?<\/p>\n","answers":[{"AnswerId":"32628956","CreationDate":"2015-09-17T11:13:21.377","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>Rather than doing that in a layer, you can use a <code>nolearn.lasagne.BatchIterator<\/code>; in the following snippet I resample my original 1D signal to a 1000 points signal:<\/p>\n\n<pre><code>from nolearn.lasagne import BatchIterator\nfrom scipy.signal import resample\nimport numpy as np\n\nclass ResampleIterator(BatchIterator):\n\n    def __init__(self, batch_size, newSize):\n        super(ResampleIterator, self).__init__(batch_size)\n        self.newSize = newSize\n\n    def transform(self, Xb, yb):\n        X_new = resample(Xb, self.newSize, axis=2).astype(np.float32)    \n        return X_new, yb\n\n\n myNet = NeuralNet(\n # define your usual other parameters (layers, etc) here\n\n    # and these are the lines you are interested in:\n    batch_iterator_train=CropIterator(batch_size=128, newSize=1000),\n    batch_iterator_test=CropIterator(batch_size=128, newSize=1000),\n    )\n<\/code><\/pre>\n\n<p>I don't know if you've used <code>nolearn<\/code> yet, you can read more about it (installation, examples) <a href=\"https:\/\/github.com\/dnouri\/nolearn\" rel=\"nofollow\">here<\/a><\/p>\n"}]}
{"QuestionId":32630635,"AnswerCount":1,"Tags":"<neural-network><torch><conv-neural-network>","CreationDate":"2015-09-17T12:45:44.197","AcceptedAnswerId":32636004.0,"OwnerUserId":350605.0,"Title":"Torch: NN handling text and numeric input","Body":"<p>I have the following NN architecture:<\/p>\n\n<p>Part 1:<\/p>\n\n<pre><\/pre>\n\n<p>Part 2:<\/p>\n\n<pre><\/pre>\n\n<p>What i would like to do is to use the output of these two parts as input to another part:<\/p>\n\n<pre><\/pre>\n\n<p>Notice how part 1 has 128 outputs and part 2 has 4 and finally part 3 has 132 inputs. So basically what i want is a network that takes two types of input (part 1 is for text, part 2 for a number vector) and to use both these information in the third layer for a 2 class classification.<\/p>\n\n<p>I have looked at the various containers but nothing seems like it is what i need. Specifically i have looked at nn.Parallel but from the docs it looks like it does something completely different (same input for two different modules). The first problem is what should the input look like for the network (since each part takes a different type of Tensor, i thought that a simple table (array) would be fine, with its first element being a 2D tensor and second a 1D tensor) and how to plug its outputs into another network so i can use forward\/backward calls as usual.<\/p>\n\n<p>Is there any way how to do this?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"32636004","CreationDate":"2015-09-17T16:59:30.807","ParentId":null,"OwnerUserId":"1417179","Title":null,"Body":"<p>What you need is <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/table.md#nn.ParallelTable\" rel=\"noreferrer\">nn.ParallelTable<\/a> and <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/table.md#nn.JoinTable\" rel=\"noreferrer\">nn.JoinTable<\/a>.<\/p>\n\n<pre><code>local parallel = nn.ParallelTable()\nparallel:add(part1)\nparallel:add(part2)\n\nlocal net = nn.Sequential()\nnet:add(parallel)                   -- (A)\nnet:add(nn.JoinTable(1))            -- (B)\nnet:add(part3)                      -- (C)\n<\/code><\/pre>\n\n<p>(A):<\/p>\n\n<p><code>parallel<\/code> will take a table of 2 tensors (in your case, text and numbers), forward first tensor to <code>part1<\/code>, second tensor to <code>part2<\/code> and output both results in another table of 2 tensors. <\/p>\n\n<p>(B):<\/p>\n\n<p>The following <code>nn.JoinTable<\/code> takes this table as input and concatenates the 2 tensors in a single one. You may have to play with the parameter handling the concatenation dimension (<code>1<\/code> in my example) depending on the shape of your tensors.<\/p>\n\n<p>(C):<\/p>\n\n<p>Finally you can add the third part of your network taking the concatenated tensor as input.<\/p>\n"}]}
{"QuestionId":32636382,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2015-09-17T17:23:36.983","AcceptedAnswerId":null,"OwnerUserId":1366749.0,"Title":"How can I force a dependency to reload in torch after recompiling it?","Body":"<p>I am working on a torch package that is a luarock. During testing I want to   and get the latest code without exiting my  session.<\/p>\n\n<p>However, as I understand it,  caches the loaded packages so they're not loaded a second time. The <a href=\"http:\/\/www.lua.org\/pil\/8.1.html\" rel=\"nofollow\">Lua manual<\/a> seems to suggest that there is global variable  that stores this cached information and I could simply set the value for my package to :<\/p>\n\n<pre><\/pre>\n\n<p>However, in torch, there doesn't seem to be a  global, even after I have required some packages.<\/p>\n\n<p>Thoughts? <\/p>\n","answers":[]}
{"QuestionId":32642298,"AnswerCount":1,"Tags":"<deep-learning><classification><theano><convolution>","CreationDate":"2015-09-18T01:02:40.353","AcceptedAnswerId":null,"OwnerUserId":5121855.0,"Title":"How to change the size of convolutional layer in CNN in Theano","Body":"<p>I am using classifcation with deep learning using <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\" rel=\"nofollow noreferrer\">Theano library<\/a>. In <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/blob\/master\/code\/convolutional_mlp.py\" rel=\"nofollow noreferrer\">convolutional_mlp<\/a> code, it uses two convolutional layers with size of , and max_pool layer with  along with a fully hidden layer connected like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The question is that if I want to change the size of convolution layer from  to , what would be the architecture of CNN?<\/p>\n","answers":[{"AnswerId":"32768488","CreationDate":"2015-09-24T18:31:48.027","ParentId":null,"OwnerUserId":"1662069","Title":null,"Body":"<p>with 3*3 filter 2*2 pooling of 28*28 image size:<\/p>\n\n<p>filtering reduces the image size to <code>(28-3+1, 28-3+1) = (26, 26)<\/code>\nand pooling reduces this further to <code>(26\/2, 26\/2) = (13, 13)<\/code> for first CNN layer and so on.<\/p>\n"}]}
{"QuestionId":32646052,"AnswerCount":1,"Tags":"<c++><deep-learning><caffe>","CreationDate":"2015-09-18T07:17:52.350","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Saving caffe Model C++ source code?","Body":"<p>I am trying to save caffe net model to .caffemodel format using following code. But I am getting segmentation fault.<\/p>\n\n<pre><\/pre>\n\n<p>where caffe_net_ is a trained caffe net, wgich containes blobs parameters. <\/p>\n","answers":[{"AnswerId":"32675410","CreationDate":"2015-09-20T03:13:52.863","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>No idea why you would get a seg fault from that. That's exactly how you would manually save a network state. Have you tried stepping through and seeing at what point it faults? Otherwise, just snapshot it during training?<\/p>\n"}]}
{"QuestionId":32649369,"AnswerCount":1,"Tags":"<numpy><cuda><theano><convolution>","CreationDate":"2015-09-18T10:08:00.837","AcceptedAnswerId":null,"OwnerUserId":3778898.0,"Title":"Why are the convolution outputs calculated with theano and numpy not the same?","Body":"<p>I made a <a href=\"https:\/\/gist.github.com\/entron\/e4778d9c0b0d241d8c91\" rel=\"nofollow\">simple example ipython notebook to calculate convolution with theano and with numpy<\/a>, however the results are different. Does anybody know where is the mistake?<\/p>\n\n<pre><\/pre>\n\n<p>Define the input image x0:<\/p>\n\n<pre><\/pre>\n\n<p>Define the convolution kernel:<\/p>\n\n<pre><\/pre>\n\n<p>Calculate the convolution with theano and cudnn:<\/p>\n\n<pre><\/pre>\n\n<p>Calculate convolution with numpy (note the result is different):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32649592","CreationDate":"2015-09-18T10:19:37.740","ParentId":null,"OwnerUserId":"1199693","Title":null,"Body":"<p>I'm not exactly sure what kind of convolution you are trying to compute, but it seems to me that <code>numpy.sum(x0*w0)<\/code> might not be the way to do it. Does this help?<\/p>\n\n<pre><code>import numpy as np\n# ... define x0 and w0 like in your example ...\nnp_convolution = np.fft.irfftn(np.fft.rfftn(x0) * np.fft.rfftn(w0))\n<\/code><\/pre>\n\n<p>The last element of the resulting array, i.e. <code>np_convolution[-1,-1,-1,-1]<\/code> is <code>-0.047490807560833327<\/code>, which seems to be the answer you're looking for in your notebook.<\/p>\n"}]}
{"QuestionId":32656676,"AnswerCount":1,"Tags":"<cuda><scikit-learn><opencl><neural-network><theano>","CreationDate":"2015-09-18T16:18:02.180","AcceptedAnswerId":null,"OwnerUserId":5028929.0,"Title":"Can pylearn2 and Theano run on AMD GPU based platform?","Body":"<p>I'd like to use pylearn2, theano and scikit-neuralnetwork to build neural network models. But my friend told me that all this module can only run on NVIDIA GPU based platform (because they would import the pycuda module). But I only have an AMD GPU(R9 270,and an AMD FX-8300 CPU),and I wish to take advantage of AMD GPU to speed up computing. Can I use all of the modules metioned above? Or is there any substitutes I can use to build neural network models ? Thanks!<\/p>\n","answers":[{"AnswerId":"32656856","CreationDate":"2015-09-18T16:28:51.993","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Currently, Theano only supports nvidia GPUs. There is a partial implementation of an OpenCL backend that would support AMD GPUs but it is incomplete and unsupported.<\/p>\n\n<p>scikit-neuralnetwork builds on PyLearn2 and PyLearn2 builds on Theano so none of those packages can operate on AMD GPUs.<\/p>\n\n<p>Torch appears to already have some <a href=\"https:\/\/github.com\/torch\/torch7\/wiki\/Cheatsheet#opencl\" rel=\"nofollow\">OpenCL support<\/a>. Caffe's OpenCL support appears to be <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2195\" rel=\"nofollow\">under development<\/a>.<\/p>\n"}]}
{"QuestionId":32658142,"AnswerCount":1,"Tags":"<lua><torch><gfx>","CreationDate":"2015-09-18T17:49:03.510","AcceptedAnswerId":32661031.0,"OwnerUserId":1082019.0,"Title":"gfx.js on Torch \/ Lua, how to visualize an image?","Body":"<p>I would like to use gfx.js in my Torch \/ Lua code, but it does not seem to work.\nI tried to install it and to run it by following <a href=\"https:\/\/github.com\/clementfarabet\/gfx.js\/\" rel=\"nofollow\">the instructions here<\/a>.<\/p>\n\n<p>The installation worked fine. Then I ran:<\/p>\n\n<pre><\/pre>\n\n<p>But then no image is shown. Instead, this string is printed:<\/p>\n\n<pre><\/pre>\n\n<p>I've no idea about what this means and why I got it. \nCan someone give me some suggestions?<\/p>\n\n<p>I'm working on CentOS Linux release 7.1.1503 (Core).<\/p>\n","answers":[{"AnswerId":"32661031","CreationDate":"2015-09-18T21:06:02.570","ParentId":null,"OwnerUserId":"258523","Title":null,"Body":"<p>The way the documentation reads it sounds like the <code>gfx.image<\/code> call will create the new image window in the browser accessed \"page\" that the first call started the server for.<\/p>\n\n<p>That message looks very much like a \"success\" status message about <code>gfx.js<\/code> having done that.<\/p>\n\n<p>Do you have that page loaded in your browser when you try this?<\/p>\n"}]}
{"QuestionId":32667080,"AnswerCount":1,"Tags":"<image><matlab><shell><caffe><lmdb>","CreationDate":"2015-09-19T10:24:06.990","AcceptedAnswerId":null,"OwnerUserId":1980557.0,"Title":"Error creating LMDB database file in MATLAB for Caffe","Body":"<p>I am trying to convert a dataset of images to LMDB format for use with Caffe, and I need to call the convert_imageset function for applying this conversion from inside Matlab.<\/p>\n\n<p>I am using Linux, and I have created a shell (.sh) script with the needed parameters for running the conversion. Here is an example of how does my shell file look like:<\/p>\n\n<pre><\/pre>\n\n<p>When I simply run my script from the terminal like this:<\/p>\n\n<pre><\/pre>\n\n<p>it works without any problem. \nBut when I try to do it from Matlab using the  function:<\/p>\n\n<pre><\/pre>\n\n<p>it seems it is not able to open\/find my files, rising the following error for each image in train.txt:<\/p>\n\n<pre><\/pre>\n\n<p>Here are some sample lines from train.txt file (do not mind the 0s, they are just dummy labels):<\/p>\n\n<pre><\/pre>\n\n<p>They are absolute paths, so there should be no problem.<\/p>\n\n<p>Any idea you might have about what could be happening can be very helpful for me!<\/p>\n\n<p>Thank you,<\/p>\n\n<p>Marc<\/p>\n","answers":[{"AnswerId":"32692753","CreationDate":"2015-09-21T10:20:25.867","ParentId":null,"OwnerUserId":"1980557","Title":null,"Body":"<p>I have not been able to solve the specific problem with Matlab, but I have managed to do the following (weird) workaround by using .txt files for communication:<\/p>\n\n<ol>\n<li>Call the main Matlab's program from Python.<\/li>\n<li>Check whenever Matlab needs to call the <code>.\/example_shell.sh<\/code> script. <\/li>\n<li>Python does the conversion calling <code>.\/example_shell.sh<\/code>.<\/li>\n<li>Matlab execution continues.<\/li>\n<\/ol>\n"}]}
{"QuestionId":32668699,"AnswerCount":2,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-09-19T13:19:11.430","AcceptedAnswerId":32757561.0,"OwnerUserId":1377127.0,"Title":"Modifying the Caffe C++ prediction code for multiple inputs","Body":"<p>I implemented a modified version of the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cpp_classification\/classification.cpp\">Caffe C++ example<\/a> and while it works really well, it's incredibly slow because it only accepts images one by one. Ideally I'd like to pass Caffe a vector of 200 images and return the best prediction for each one. I received some <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/oh-ehSXxDR8\">great help from Fanglin Wang<\/a> and implemented some of his recommendations, but am still having some trouble working out how to retrieve the best result from each image. <\/p>\n\n<p>The Classify method is now passed a vector of  objects (variable ) which is a vector of grayscale floating point images. I've eliminated the preprocessing method in the code because I don't need to convert these images to floating point or subtract the mean image. I've also been trying to get rid of the  variable because I only want to return the top prediction and probability for each image.<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Thank-you so much for your help Shai, I made the changes you recommended but seem to be getting some strange compilation issues I can't work out (I managed to sort out a few of the issues).<\/p>\n\n<p>These are the changes I made:<\/p>\n\n<p>Header File:<\/p>\n\n<pre><\/pre>\n\n<p>Class File:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32675354","CreationDate":"2015-09-20T03:06:07.673","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>Unfortunately, I don't believe a parallelization of network Forward passes has been implemented. However, if you'd like you could simply implement your own wrapper to repeatedly run data through copies of your network, in parallel?<\/p>\n\n<p>Have a look at <a href=\"https:\/\/stackoverflow.com\/questions\/32589349\/how-many-images-can-you-pass-to-caffe-at-a-time\">How many images can you pass to Caffe at a time?<\/a><\/p>\n\n<p>In the linked prototxt all you have to define is<\/p>\n\n<pre><code>input_shape {\n  dim: 64 \/\/ num of images\n  dim: 1\n  dim: 28 \/\/ height\n  dim: 28 \/\/ width\n}\n<\/code><\/pre>\n\n<p>The existing implementation evaluates a batch of 64 images but not necessarily in parallel. However, if running on a GPU, processing a batch of 64 will be faster than 64 single-image batches.<\/p>\n"},{"AnswerId":"32757561","CreationDate":"2015-09-24T09:12:15.823","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If I understand your problem correctly, you input <code>n<\/code> images, expecting <code>n<\/code> pairs of <code>(label, prob)<\/code>, but getting only one such pair.<\/p>\n\n<p>I believe these modifications should do the trick for you:<\/p>\n\n<ol>\n<li><p><code>Classifier::Predict<\/code> should return a <code>vector&lt; vector&lt;float&gt; &gt;<\/code>, that is a <em>vector<\/em> of probabilities per input image. That is a <code>vector<\/code> of size <code>n<\/code> of vectors of size <code>output_layer-&gt;channels()<\/code>:<\/p>\n\n<pre><code>std::vector&lt; std::vecot&lt;float&gt; &gt; \nClassifier::Predict(const std::vector&lt;cv::Mat&gt; &amp;input_channels, \n                    int num_images) {\n  \/\/ same code here...\n\n  \/* changes here: Copy the output layer to a std::vector *\/\n  Blob&lt;float&gt;* output_layer = net_-&gt;output_blobs()[0];\n  std::vector&lt; std::vector&lt;float&gt; &gt; ret;\n  for ( int i = 0 ; i &lt; num_images ; i++ ) {\n      const float* begin = output_layer-&gt;cpu_data() + i*output_layer-&gt;channels();\n      const float* end = begin + output_layer-&gt;channels();\n      ret.push_back( std::vector&lt;float&gt;(begin, end) );\n  }\n  return ret;\n}\n<\/code><\/pre><\/li>\n<li><p>In <code>Classifier::Classify<\/code> you need to process each <code>vector&lt;float&gt;<\/code> through <code>Argmax<\/code> independantly:<\/p>\n\n<pre><code> std::vector&lt; std::pair&lt;int,float&gt; &gt; \n Classifier::Classify(const std::vector&lt;cv::Mat&gt; &amp;input_channels) {\n\n   std::vector&lt; std::vector&lt;float&gt; &gt; output = Predict(input_channels);\n\n   std::vector&lt; std::pair&lt;int,float&gt; &gt; predictions;\n   for ( int i = 0 ; i &lt; output.size(); i++ ) {\n       std::vector&lt;int&gt; maxN = Argmax(output[i], 1);\n       int idx = maxN[0];\n       predictions.push_back(std::make_pair(labels_[idx], output[idx]));\n   }\n   return predictions;\n }\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":32673654,"AnswerCount":0,"Tags":"<python><machine-learning><torch>","CreationDate":"2015-09-19T08:16:02.440","AcceptedAnswerId":null,"OwnerUserId":4031374.0,"Title":"Torch: Confusion on weights and optimization","Body":"<p>If I want to convert the code snippet from Python to Torch, are these two equivalent?<\/p>\n\n<pre><\/pre>\n\n<p>And code for Torch:<\/p>\n\n<pre><\/pre>\n\n<p>, where  gives hidden states and input a weighted linear transformation. But where are the weight matrices kept for later's optimization? Similar use of Linear() function can be found in <a href=\"https:\/\/github.com\/wojzaremba\/lstm\/blob\/master\/main.lua#L65\" rel=\"nofollow\">LSTM's code<\/a> on GitHub:<\/p>\n\n<pre><\/pre>\n\n<p>It seemed each time input (, , ) pushed through the LSTM unit, the weights inside are generated once again. So how could this model be optimized and weights decay during training? And yet another question about  is that there's a  <a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/4_train.lua#L145\" rel=\"nofollow\">passed to the optimization functions<\/a>, which outputs  and . What does this  mean? If itself represents the error estimation (seemed to have done all necessary calculations), what is the use of an  or  function? <\/p>\n","answers":[]}
{"QuestionId":32679255,"AnswerCount":1,"Tags":"<c++><opencv><cuda><theano>","CreationDate":"2015-09-20T12:11:28.673","AcceptedAnswerId":32679934.0,"OwnerUserId":4431798.0,"Title":"Neural Network Image Classification, The Most Efficient Solution \/ Suggestion","Body":"<p>I have already built a deep neural network image classifier program in Matlab (gives 1 output for each example, such as is it a car or not), using gradient descent and back propagation algorithms. It is a simple feed forward network, with 1 or 2 hidden layers. I'm using the obtained weights in a nvcc C++ for real time object detection.<\/p>\n\n<p>NN training results have quite a good accuracy (more than %99.9, but not enough), and can process more than 100,000 image files of size 32x32. But only problem with the Matlab code is: it ends up in local minimums in each training, so requires many different trainings but its training speed is quite slow.<\/p>\n\n<p>Other than my slowly working Matlab NN training code, I have tried:<\/p>\n\n<p>1) OpenCV 3.0.0, it \"probably\" has a <a href=\"http:\/\/answers.opencv.org\/question\/71281\/cvmlstatmodelpredict-not-working-neural-networks\/\" rel=\"nofollow\">bug<\/a> in virtual float cv::ml::StatModel::predict function at the moment. So I weren't able to use it properly. <\/p>\n\n<p>2) Tried <a href=\"http:\/\/www.artelnics.com\/opennn\" rel=\"nofollow\">OpenNN with gui<\/a>, but it even gets stuck during loading and training. I'm still working on to fix that.<\/p>\n\n<p>3) Checked <a href=\"http:\/\/leenissen.dk\/fann\/wp\/\" rel=\"nofollow\">FANN<\/a>, but could only find \"one\" tutorial code written in C++. May take quite a time for me to master it with out examples.<\/p>\n\n<p>4) I had tried Theano in Python a few months ago, it was quite customizable, and has quite many tutorials. But had never tried training image files with it.<\/p>\n\n<p>5) I can also transfer my Matlab code to nvcc C++, and try conjugate gradients method to speed up further. But didn't try this yet, it is the last choice for me.<\/p>\n\n<p>Mastering in each path may take quite a big time, and I have many more different works to do too. Which path should I take, or do you have another suggestion? Thank you<\/p>\n","answers":[{"AnswerId":"32679934","CreationDate":"2015-09-20T13:29:39.260","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>If you have experience with Matlab, the easiest path is to go through the \"VGG Convolutional Neural Networks Practical\" and use their open source MatConvNet toolbox for Matlab: <a href=\"http:\/\/www.vlfeat.org\/matconvnet\/\" rel=\"nofollow\">http:\/\/www.vlfeat.org\/matconvnet\/<\/a>.<\/p>\n"}]}
{"QuestionId":32679476,"AnswerCount":2,"Tags":"<python><python-2.7><neural-network><theano><conv-neural-network>","CreationDate":"2015-09-20T12:34:49.507","AcceptedAnswerId":null,"OwnerUserId":4969461.0,"Title":"Theano import error: No module named cPickle","Body":"<pre><\/pre>\n\n<p>I'm pretty sure <strong>there is no problem with cPickle<\/strong>. Can it be the problem of other dependencies? It's true that I upgraded all the packages this morning, so that a conflict of version may occur. Nonetheless, the problem still exists after I switched to the bleeding-edge version of theano. Would someone help me figure it out?<\/p>\n\n<p>PS: My developing device is Macbook Pro 13(early 2015); my system version is OS X 10.10.5; python version is 2.7.10<\/p>\n\n<p>=========================== Second Edit ===================================<\/p>\n\n<p>It seems I do have latest version of six installed on my Mac.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38034287","CreationDate":"2016-06-26T00:51:46.110","ParentId":null,"OwnerUserId":"3765905","Title":null,"Body":"<p>Try to avoid using <code>sudo<\/code> at all costs. Open a python session and run<\/p>\n\n<pre><code>import six\nprint six.__version__\nprint six.__file__\n<\/code><\/pre>\n\n<p>The version will likely not be most recent (1.10.0), so go manually delete the six.py and six.pyc files at the imported path, and then run <code>pip install six<\/code>. Python should now be importing the latest version of six, compatible with Theano :)<\/p>\n"},{"AnswerId":"36827555","CreationDate":"2016-04-24T18:59:42.877","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>I had the same problem and <em>upgrading<\/em> six solved issue:<\/p>\n\n<pre><code>sudo easy_install --upgrade six\n<\/code><\/pre>\n"}]}
{"QuestionId":32679918,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-09-20T13:28:34.360","AcceptedAnswerId":null,"OwnerUserId":375986.0,"Title":"How to ignore solid-color image backgrounds when working with convolutional neural networks (CNNs)?","Body":"<p>Given an image of an object with a known, solid background color, how can I influence a CNN to ignore\/discount the features of the background, thereby emphasizing the object?<\/p>\n\n<p>For your information, my scenario is feature extraction (e.g. neural codes) for the purposes of content-based image retrieval (CBIR). I am using Caffe.<\/p>\n","answers":[{"AnswerId":"32687910","CreationDate":"2015-09-21T05:26:02.010","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>One option I can think of is creating two inputs per image: the first is your 3xHxW color image and the other is 1xHxW mask with zeros on the background (the \"solid-color\" pixels) and ones otherwise. Then you can do element-wise multiplication of the mask with the output of your first conv layer, thus forcing all features of \"solid-color\" pixels to be zero. <\/p>\n"}]}
{"QuestionId":32680860,"AnswerCount":3,"Tags":"<neural-network><deep-learning><caffe><multilabel-classification>","CreationDate":"2015-09-20T15:09:18.793","AcceptedAnswerId":32697800.0,"OwnerUserId":639973.0,"Title":"caffe with multi-label images","Body":"<p>I have a dataset of images that have multiple labels; There are 100 classes in the dataset, and each image has 1 to 5 labels associated with them.<\/p>\n\n<p>I'm following the instruction in the following URL:<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/550\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/550<\/a><\/p>\n\n<p>It says that I need to generate a text file listing the images and its labels as in <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>In my case, since I have multi-label images, does it work to simply add labels as in following?<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I have a feeling that it's probably not going to be that simple, and if I'm right, in what step and how should I integrate the multi-label-ness of the dataset in the process of setting up Caffe?<\/p>\n","answers":[{"AnswerId":"32687772","CreationDate":"2015-09-21T05:10:58.450","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>AFAIK, current Caffe version does not support lmdb\/leveldb datasets for images with multilabels. However, you can (and probably should) prepare your inputs in HDF5 format. Caffe HDF5 input layer is much more flexible and will allow you to have multiple labels per input.<br>\n<a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">This answer<\/a> gives a brief description of how to create HDF5 input for caffe.<\/p>\n\n<p>Another issue you must address is the fact that you are interested not only in <strong>multi<\/strong>-label per image, but also with <strong>varying<\/strong> number of labels per image. How do you define your loss per image, per label? it might be the case that you would have to write your own loss layer.<br>\nThere are some loss layers that supports \"ignore label\": that is, if a specific input label is assigned to the image, no loss is computed for the respective image. See, e.g. <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/loss_layers.hpp#L90\" rel=\"nofollow noreferrer\"><code>AccuracyLayer<\/code><\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/loss_layers.hpp#L696\" rel=\"nofollow noreferrer\"><code>SoftmaxWithLossLayer<\/code><\/a>.<\/p>\n"},{"AnswerId":"32697800","CreationDate":"2015-09-21T14:28:55.800","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>I believe Shai's answer is no longer up-to-date.\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issue-53768814\" rel=\"nofollow noreferrer\">Caffe supports multi-label\/matrix ground truth<\/a> for HDF5 and LMDB formats.\nThe python snippet in <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issuecomment-70211045\" rel=\"nofollow noreferrer\">this<\/a> github comment demonstrates how to construct multi-label LMDB ground truth (see <a href=\"https:\/\/stackoverflow.com\/a\/32687772\/2466336\">Shai's answer<\/a> for HDF5 format).\nDifferent from the construction of single-label image datasets, an lmdb is constructed for the images while a second separate lmdb is constructed for the multi-label ground truth data. The snippet deals with spatial multi-label ground truth useful for pixel-wise labeling of images.<\/p>\n\n<p>The order in which data is written to the lmdb is crucial. The order of the ground truth must match the order of the images.<\/p>\n\n<p>Loss layers such as SOFTMAX_LOSS, EUCLIDEAN_LOSS, SIGMOID_CROSS_ENTROPY_LOSS also support multi-label data.\nHowever, the Accuracy layer is still limited to single-label data. You might want to follow <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2188\" rel=\"nofollow noreferrer\">this github issue<\/a> to keep track of when this feature is added to Caffe.<\/p>\n"},{"AnswerId":"37902054","CreationDate":"2016-06-18T22:05:04.973","ParentId":null,"OwnerUserId":"3817456","Title":null,"Body":"<p>caffe supports multilabel. You can put the labels into n-hot vectors e.g. [0,1,1,0,0,1,...]  .  You need to reshape the labels to n*k*1*1 tensors and use sigmoid cross-entropy or euclidean, not softmax (which forces sum(outputs)=1 ) <\/p>\n"}]}
{"QuestionId":32707393,"AnswerCount":2,"Tags":"<c++><caffe><leveldb>","CreationDate":"2015-09-22T02:19:26.167","AcceptedAnswerId":32716439.0,"OwnerUserId":290163.0,"Title":"What's caffe's input format?","Body":"<p>I'm try to use  for audio recognition, but can't find a document for its input format.<\/p>\n\n<p>I want to use , thus I must create a key and a value for each record, which is a pair of label string and data byte array.<\/p>\n\n<p>It seems that no document describes this, and after I found the value is written by , I can't find where  is and then lost.<\/p>\n\n<p>Does anyone know how to convert non-image records into  records for ? Thanks!<\/p>\n","answers":[{"AnswerId":"32722243","CreationDate":"2015-09-22T16:37:05.163","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p><code>leveldb<\/code>, <code>lmdb<\/code> and <code>HDF5<\/code> are currently the main formats for feeding data into Caffe. The <code>MemoryData<\/code> layer enable in-memory input as well, so it's possible to use whatever input format and and use Caffe's python or c++ interfaces to populate the data blobs.<\/p>\n\n<p>If you're already set on <code>leveldb<\/code>, <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/745\" rel=\"nofollow\">this<\/a> discussion on caffe issues could be useful.<\/p>\n\n<p>Below is an example for populating a <code>leveldb<\/code> with python. It requires pycaffe and plyvel. It's adapted from <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/745\" rel=\"nofollow\">caffe's github issues<\/a> posted by <a href=\"https:\/\/github.com\/Zackory\" rel=\"nofollow\">Zackory<\/a>. It's not specific to images as long as you represent each example in the form of a CxHxW where any or all can be equal to 1:<\/p>\n\n<pre><code>import caffe\n\ndb = plyvel.DB('train_leveldb\/', create_if_missing=True, error_if_exists=True, write_buffer_size=268435456)\nwb = db.write_batch()\ncount = 0\nfor file in dataset:\n    mat = # load numpy array from file\n\n    # Load matrix into datum object\n    datum = caffe.io.array_to_datum(mat)\n\n    wb.put('%08d_%s' % (count, file), datum.SerializeToString())\n\n    count += 1\n\n    # Write to db in regular intervals\n    if count % 1000 == 0:\n        # Write batch of images to database\n        wb.write()\n        del wb\n        wb = db.write_batch()\n\n    # Write last batch of images\n    if count % 1000 != 0:\n        wb.write()\n<\/code><\/pre>\n\n<p>I find constructing lmdb a lot simpler. <code>lmdb<\/code> example <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"32716439","CreationDate":"2015-09-22T12:05:17.817","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>The Datum object is defined with protobuf. See here:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L30-L41\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L30-L41<\/a><\/p>\n\n<p>It generates a file <code>caffe.pb.h<\/code> in <code>.build_release\/src\/caffe\/proto<\/code> with the class <code>Datum<\/code>. You can have a look there to understand how this object works. <\/p>\n"}]}
{"QuestionId":32717054,"AnswerCount":1,"Tags":"<neural-network><deep-learning><computer-vision><caffe><vgg-net>","CreationDate":"2015-09-22T12:36:40.770","AcceptedAnswerId":32720518.0,"OwnerUserId":639973.0,"Title":"fine-tuning with VGG on caffe","Body":"<p>I'm replicating the steps in\n<a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html<\/a><\/p>\n\n<p>I want to change the network to VGG model which is obtained at\n<a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/very_deep\/caffe\/VGG_ILSVRC_16_layers.caffemodel\" rel=\"nofollow\">http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/very_deep\/caffe\/VGG_ILSVRC_16_layers.caffemodel<\/a><\/p>\n\n<p>does it suffice to simply substitute the model parameter as following?<\/p>\n\n<pre><\/pre>\n\n<p>Or do I need to adjust learning rates, iterations, i.e. does it come with separate prototxt files?<\/p>\n","answers":[{"AnswerId":"32720518","CreationDate":"2015-09-22T15:10:00.303","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>There needs to be a 1-1 correspondence between the weights of the network you want to train and the weights you use for initializing\/fine-tuning. The architecture of the old and new model have to match.<\/p>\n\n<p>VGG-16 has a different architecture than the model described by models\/finetune_flickr_style\/train_val.prototxt (FlickrStyleCaffeNet). This is the network that the solver will try to optimize. Even if it doesn't crash, the weights you've loaded don't have any meaning in the new network.<\/p>\n\n<p>The VGG-16 network is described in the deploy.prototxt file on <a href=\"https:\/\/gist.github.com\/ksimonyan\/211839e770f7b538e2d8#file-readme-md\" rel=\"noreferrer\">this<\/a> page in Caffe's Model Zoo.<\/p>\n"}]}
{"QuestionId":32717102,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-09-22T12:38:38.937","AcceptedAnswerId":32724857.0,"OwnerUserId":2191652.0,"Title":"Fusing different input channels in caffe?","Body":"<p>I'd like to process different types of data seperately first and then fuse them in a common layer. Is this possible in Caffe and if yes what would be the best way to do it?<\/p>\n\n<p>I've read that one can define several data layers in the same prototxt file. But how to fuse them?<\/p>\n\n<p>Can I just create a  layer and specify several  layers? Or do I have to concatenate the individual layers first using a  layer?<\/p>\n\n<p>For any small code example I would be very thankful!<\/p>\n","answers":[{"AnswerId":"32724857","CreationDate":"2015-09-22T19:02:20.113","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>As discussed in the comments above, <code>InnerProduct<\/code> works with a single input. The fusion (concatenation) can then be done in a specific <code>Concat<\/code> layer with a configuration like this:<\/p>\n\n<pre><code>layer {\n  name: \"concat\"\n  bottom: \"in1\"\n  bottom: \"in2\"\n  top: \"out\"\n  type: \"Concat\"\n  concat_param {\n    axis: 1\n  }\n}\n<\/code><\/pre>\n\n<p>The official documentation has more details about that layer: <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n"}]}
{"QuestionId":32717528,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-09-22T12:57:41.657","AcceptedAnswerId":null,"OwnerUserId":5285204.0,"Title":"Is my learning_rate actually getting changed in theano?","Body":"<p>I'm trying to adjust the learning rate of the gradient descent algorithm. I would like to be able to confirm whether or not my changes to  are actually having an effect on my theano training function.<\/p>\n\n<p>Sample code:<\/p>\n\n<pre><\/pre>\n\n<p>Is this going to work as I desire? How can I confirm?<\/p>\n\n<h1>Edit:<\/h1>\n\n<p>It seems like this will not work, based on this little test:<\/p>\n\n<pre><\/pre>\n\n<p>What is the correct way to go about this?<\/p>\n","answers":[{"AnswerId":"35572763","CreationDate":"2016-02-23T08:47:23.220","ParentId":null,"OwnerUserId":"3302563","Title":null,"Body":"<p>Inspired with the answer of @Daniel Renshaw you can try the following:<\/p>\n\n<pre><code>learning_rate = theano.shared(0.01)\nfor param in params:\n    updates.append((param, param-learning_rate*T.grad(cost, param)))\n#set up the training function\ntrain = theano.function(inputs=[index], outputs=[cost], updates=updates, givens={x:self.X[index:index+mini_batch_size,:]})\n\n#run through the minibatches\nfor epoch in range(n_epochs):\n    for row in range(0,self.m, mini_batch_size):\n        cost = train(row)\n    #occasionally adjust the learning rate\n    learning_rate.set_value(learning_rate.get_value()\/ 2)\n<\/code><\/pre>\n\n<p>Basically you use a shared variable and update it manually every iteration.<\/p>\n"},{"AnswerId":"32768967","CreationDate":"2015-09-24T19:01:41.903","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that your code is compiling the learning rate into the computation graph as a constant. If you want to change the rate you'll need to use a Theano variable to represent it in the computation graph and then provide a value when the function is executed. This can be done in two ways:<\/p>\n\n<ol>\n<li><p>Pass the rate each time the function is executed by treating it as an input value and representing it in the computation graph as a scalar tensor.<\/p><\/li>\n<li><p>Store the rate in a Theano shared variable. Change the variable manually before executing the function.<\/p><\/li>\n<\/ol>\n\n<p>There are two variations of the second approach. In the first you manually adjust the rate value before the execution. In the second you specify a symbolic expression explaining how the rate should be updated on each execution.<\/p>\n\n<p>These three approaches are demonstrated in this sample code based the edited portion of the question.<\/p>\n\n<pre><code>import theano as th\nimport theano.tensor\n\n# Original version (changing rate doesn't affect theano function output)\nx = th.tensor.dscalar()\nrate=5.0\nf = th.function(inputs=[x], outputs=2*x*rate)\nprint(f(10))\nrate=0.0\nprint(f(10))\n\n\n# New version using an input value\nx = th.tensor.dscalar()\nrate=th.tensor.scalar()\nf = th.function(inputs=[x, rate], outputs=2*x*rate)\nprint(f(10, 5.0))\nprint(f(10, 0.0))\n\n\n# New version using a shared variable with manual update\nx = th.tensor.dscalar()\nrate=th.shared(5.0)\nf = th.function(inputs=[x], outputs=2*x*rate)\nprint(f(10))\nrate.set_value(0.0)\nprint(f(10))\n\n\n# New version using a shared variable with automatic update\nx = th.tensor.dscalar()\nrate=th.shared(5.0)\nupdates=[(rate, rate \/ 2.0)]\nf = th.function(inputs=[x], outputs=2*x*rate, updates=updates)\nprint(f(10))\nprint(f(10))\nprint(f(10))\nprint(f(10))\n<\/code><\/pre>\n"}]}
{"QuestionId":32744542,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-09-23T16:11:31.567","AcceptedAnswerId":null,"OwnerUserId":5368537.0,"Title":"what is 'top' parameter in caffe","Body":"<p>I am trying to run a caffe Experiment.I am using the following loss layer in my Train.prototxt,<\/p>\n\n<pre><\/pre>\n\n<p>I see the following configuration being displayed when the training is started,<\/p>\n\n<blockquote>\n  <p>I0923 21:19:13.101313 26423 net.cpp:410] loss &lt;- ip2<br>\n  I0923 21:19:13.101323 26423 net.cpp:410] loss &lt;- label<br>\n  I0923 21:19:13.101339 26423 net.cpp:368] loss -> (automatic)<\/p>\n<\/blockquote>\n\n<p>I have not given  parameter in the loss layer.<br>\nWhat exactly the automatic(loss -> (automatic)) means here?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"32746933","CreationDate":"2015-09-23T18:31:28.393","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>Caffe layers, including Loss layers, produce Blob (4-D arrays) as output of their computations. If you don't set a Blob name through the <code>top<\/code> parameter, the corresponding Blob will be added to the \"output\" of the net.<\/p>\n\n<p>This means that, if you call the <code>Net::forward()<\/code> method, it will return a list of Blobs, i.e., the ones that are unbounded to be the input for another layer.<\/p>\n\n<p>When you call the Caffe training tool, it automatically print to screen such Blobs. This way you can follow the value of loss or accuracy during training.<\/p>\n"}]}
{"QuestionId":32754110,"AnswerCount":1,"Tags":"<deep-learning><torch><unsupervised-learning>","CreationDate":"2015-09-24T05:39:30.617","AcceptedAnswerId":null,"OwnerUserId":5118777.0,"Title":"why unsupervised model needs to implement nn.diag?","Body":"<p>I am trying to learn deep learning.<\/p>\n\n<p>In torch tutorial, <\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/2_model.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/2_model.lua<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/3_unsupervised\/2_models.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/3_unsupervised\/2_models.lua<\/a><\/p>\n\n<p><strong>Supervised model<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Unsupervised model<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>why unsupervised model needs to implement nn.Diag ?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"36943277","CreationDate":"2016-04-29T16:25:09.680","ParentId":null,"OwnerUserId":"2847699","Title":null,"Body":"<p>It is in fact a scaling by a learnable vector (the diagonal of a matrix). This is mentioned in the section 3.1 of the paper <a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/gregor-icml-10.pdf\" rel=\"nofollow\">Learning Fast Approximations of Sparse Coding<\/a>. It is multiplied by the tanh and together form the non linearity.<\/p>\n"}]}
{"QuestionId":32761243,"AnswerCount":2,"Tags":"<python><symlink><symbolic-math><theano>","CreationDate":"2015-09-24T12:23:51.043","AcceptedAnswerId":32761395.0,"OwnerUserId":3867406.0,"Title":"What does \"symbolic\" mean in this sentence?","Body":"<p>I'm reading through Bergstra <em>et al<\/em> <a href=\"http:\/\/arxiv.org\/abs\/1211.5590\" rel=\"nofollow\">\"Theano: new features and speed improvements\"<\/a> [2012] and do not understand the meaning of \"<strong>symbolic<\/strong>\" in this context (versus for instance a symbolic link to a directory) :<\/p>\n\n<blockquote>\n  <p>Using Theano\u2019s <strong>symbolically<\/strong>-defined implementations within a Python\n  loop prevents <strong>symbolic<\/strong> differentiation through the iterative process,\n  and prevents certain graph optimizations from being applied.\n  Completely unrolling the loop into a <strong>symbolic<\/strong> chain often leads to an\n  unmanageably large graph and does not allow for \u201cwhile\u201d-style loops\n  with a variable number of iterations.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"32761395","CreationDate":"2015-09-24T12:31:20.967","ParentId":null,"OwnerUserId":"2296458","Title":null,"Body":"<p>Symbolic in this context refers to <a href=\"https:\/\/en.wikipedia.org\/wiki\/Symbolic_computation\">symbolic math<\/a>.<\/p>\n\n<p>This means that you can perform mathematical programming using symbols, instead of needing variables that require numeric values, for example<\/p>\n\n<pre><code>y = x**2 + 5*x + 7\n<\/code><\/pre>\n\n<p>For the above to be valid in non-symbolic math, the variable <code>x<\/code> would need to have a numerical value, and the result of the right-hand side would be assigned to <code>y<\/code>, which would then have a numerical value.<\/p>\n\n<p>Conversely in symbolic math, the variables <code>x<\/code> and <code>y<\/code> could be <em>symbols<\/em> so you could then do something like (in psuedocode)<\/p>\n\n<pre><code>diff(y, x)  # meaning differentiate y with respect to x\n<\/code><\/pre>\n\n<p>and it would result in<\/p>\n\n<pre><code>dy\/dx = 2*x +5\n<\/code><\/pre>\n"},{"AnswerId":"32761527","CreationDate":"2015-09-24T12:37:31.933","ParentId":null,"OwnerUserId":"860196","Title":null,"Body":"<p>In this context it means that for a variable with <code>x = 10<\/code> only the 'symbol x' of the variable is used for computations. <code>x+x = 2x<\/code> but the value 20 is actually not computed, only the algebraic expression. That way you can e.g. have a series of math operations optimized before inserting the actual values to save computation time or potential other issues like null division. <\/p>\n\n<p>Not to be confused with but somewhat related is <a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation\" rel=\"nofollow\">automatic differentiation<\/a>, also used for neural networks, <a href=\"https:\/\/justindomke.wordpress.com\/2009\/02\/17\/automatic-differentiation-the-most-criminally-underused-tool-in-the-potential-machine-learning-toolbox\/\" rel=\"nofollow\">here's<\/a> a good article introducing the concept.<\/p>\n"}]}
{"QuestionId":32771305,"AnswerCount":1,"Tags":"<python><theano><pycuda>","CreationDate":"2015-09-24T21:35:00.850","AcceptedAnswerId":37988153.0,"OwnerUserId":1401278.0,"Title":"CUDA runtime gpu initialization with theano","Body":"<p>I am trying to parallelize my NN across two GPUs following <a href=\"https:\/\/github.com\/uoguelph-mlrg\/theano_multi_gpu\" rel=\"nofollow\">https:\/\/github.com\/uoguelph-mlrg\/theano_multi_gpu<\/a>. I have all the dependencies, but the cuda runtime initialization fails with the following message.<\/p>\n\n<pre><\/pre>\n\n<p>The relevant part of the code snippet is here:<\/p>\n\n<pre><\/pre>\n\n<p>The error is triggered when it imports theano.sandbox.cuda. And this is where, I launch the training function as two processes.<\/p>\n\n<pre><\/pre>\n\n<p>The import statement however seem to work if I try to initialize the gpu interactively in Python. I executed the first 20 lines of the train() and it worked fine there and also correctly assigned me to gpu0 as I requested.<\/p>\n","answers":[{"AnswerId":"37988153","CreationDate":"2016-06-23T09:52:09.663","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>After digging around and running pdb, the original poster found the issue.<\/p>\n\n<p>Basically theano and pycuda were both competing to initialize the gpu, causing the problem. The solution is to first 'import theano', which would get a gpu and then attach to the specific <code>context<\/code> in pycuda. So, the import sections within <code>train<\/code> function would look like this:<\/p>\n\n<pre><code>def train(private_args, process_env, &lt;some other args&gt;)\n    if process_env is not None:\n       os.environ = process_env\n\n    ####\n    # import theano related\n    # We need global imports and so we make them as such\n    theano = __import__('theano')\n    _t_tensor = __import__('theano', globals(), locals(), ['tensor'], -1)\n    tensor = _t_tensor.tensor\n\n    import theano.sandbox.cuda\n    import theano.misc.pycuda_utils\n\n    ####\n    # pycuda and zmq environment\n    import zmq\n    import pycuda.driver as drv\n    import pycuda.gpuarray as gpuarray\n\n    drv.init()\n    # Attach the existing context (already initialized by theano import statement)\n    ctx = drv.Context.attach()\n    sock = zmq.Context().socket(zmq.PAIR)\n\n    if private_args['flag_client']:\n        sock.connect('tcp:\/\/localhost:5000')\n    else:\n        sock.bind('tcp:\/\/*:5000')\n<\/code><\/pre>\n\n<p>[This answer was added as a community wiki entry from an edit made by the OP in a attempt to get this question off the unaswered list].<\/p>\n"}]}
{"QuestionId":32772376,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2015-09-24T23:09:55.913","AcceptedAnswerId":32810517.0,"OwnerUserId":2280645.0,"Title":"Getting output classification with lasagne","Body":"<p>Getting output classification with Lasagne\/Theano<\/p>\n\n<p>I am migrating my code from pure Theano to Lasagne. \nI had this certain code from a tutorial to get the result of a prediction with a certain data and I would generate a csv file to send to kaggle.\nBut with lasagne, it doesn't work.\nI have tried several things but they all give errors.<\/p>\n\n<p>I would love if anyone could help me figure what's wrong!<\/p>\n\n<p>I pasted the whole code here :\n<a href=\"http:\/\/pastebin.com\/e7ry3280\" rel=\"nofollow\">http:\/\/pastebin.com\/e7ry3280<\/a><\/p>\n\n<pre><\/pre>\n\n<p>The code fails on the line that starts with :<\/p>\n\n<blockquote>\n  <p>TypeError: Cannot convert Type TensorType(float32, 3D) (of Variable Subtensor{int64:int64:}.0) into Type TensorType(float32, 4D). You can try to manually convert Subtensor{int64:int64:}.0 into a TensorType(float32, 4D).<\/p>\n<\/blockquote>\n\n<p>I'm just trying to input the test data to the CNN and get the results to a CSV file. How can I do it? I know I must use minibatches because the whole test data wont fit on the GPU.<\/p>\n","answers":[{"AnswerId":"32810517","CreationDate":"2015-09-27T17:04:43.480","ParentId":null,"OwnerUserId":"817766","Title":null,"Body":"<p>As pointed out by the error message and Daniel Renshaw in the comments, the problem is a mismatch of dimensions between <code>test_data<\/code> and <code>input_var<\/code>. On the first line on the loop, you write:<\/p>\n\n<pre><code>test_data = test_data[it]\n<\/code><\/pre>\n\n<p>Which turns the 4D array <code>test_data<\/code> into a 3D array with the same name (that is why using the same variable name for different types is never recommended :) ). After that you encapsulate it in a shared variable which doesn't change the dimension, and then you slice it to assign it to <code>input_var<\/code>, which again doesn't change the dimension.<\/p>\n\n<p>If I understand your code, I think you should just remove that first line. That way <code>test_data<\/code> remains a list of examples, and you can slice it to make a batch.<\/p>\n"}]}
{"QuestionId":32781847,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-09-25T12:06:03.377","AcceptedAnswerId":32783827.0,"OwnerUserId":2711645.0,"Title":"Python theano: how does the R-operation work?","Body":"<p>I am executing the following on theano:<\/p>\n\n<pre><\/pre>\n\n<p>However, if I calculate a simple example by hand I have:<\/p>\n\n<pre><\/pre>\n\n<p>Thus I would expect a result of , but instead I get the sum of these. What is the  calculating differently?<\/p>\n","answers":[{"AnswerId":"32783827","CreationDate":"2015-09-25T13:55:56.147","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The difference is caused by a misunderstanding of what <code>theano.dot<\/code> is doing in your sample code.<\/p>\n\n<p><code>theano.dot(x, y)<\/code> is equal to the scalar <code>34000<\/code> given your example inputs because it is computing the vector inner product not the element-wise product, as your hand example suggests you expect.<\/p>\n\n<p>One can get the result you expect by changing<\/p>\n\n<pre><code>f = T.dot(x,y)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>f = x * y\n<\/code><\/pre>\n"}]}
{"QuestionId":32783086,"AnswerCount":1,"Tags":"<computer-vision><neural-network><hdf5><deep-learning><caffe>","CreationDate":"2015-09-25T13:15:32.943","AcceptedAnswerId":32784038.0,"OwnerUserId":639973.0,"Title":"Caffe: Multi-Label Images with Varying Number of Labels","Body":"<p>I have a dataset where the images have <strong>VARYING<\/strong> number of labels. The number of labels is between 1 and 5. There are 100 classes.<\/p>\n\n<p>After googling, it seems like HDF5 db with slice layer can deal with multiple labels, as in the following <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/caffe-users\/james$20guo\/caffe-users\/_NdX1tpHt5Y\/4A7ln18nr3YJ\" rel=\"nofollow\">URL<\/a>.<\/p>\n\n<p>The only problem is that it supposes a fixed number of labels. Following this, I would have to create a 1x100 matrix, where entry value is 1 for the labeled classes, and 0 for non-label classes, as in the following definition:<\/p>\n\n<pre><\/pre>\n\n<p>where each image contains a a label looking like (1,0,0,...1,...0,....,0,1) where the vector size is 100 dimension. <\/p>\n\n<p>Now, I apologize that my question becomes somehow vague, but is this a feasible idea? I.e., is there a better approach to this problem?<\/p>\n","answers":[{"AnswerId":"32784038","CreationDate":"2015-09-25T14:06:20.610","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>I get that you have 5 types of labels that are not always present for each data point. 1 of the 5 labels is for 100-way classification. Correct so far?<\/p>\n\n<p>I would suggest always writing all 5 labels into your HDF5 and use a special value for when the label is missing. You can then use the <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1654\" rel=\"nofollow\">missing_value<\/a> option to skip computing the loss for that layer for that iteration. Using it requires add <code>loss_param{ ignore_label = Y }<\/code> to the loss layer in your network prototxt definition where Y is a scalar.<\/p>\n\n<p>The backpropagated error will only be a function of labels that are present. If input X does not have a valid value for a label, the network will still produce an estimate for that label. But it will not be penalized for it. The output is produced without any effect on how the weights are updated in that iteration. Only outputs for non-missing labels contribute to the error signal and the weight gradients.<\/p>\n\n<p>It seems that only the Accuracy and SoftmaxWithLossLayer layers support missing_values.<\/p>\n\n<p>Each label is a 1x5 matrix. The first entry can be for the 100-way classification (e.g. [0-99]) and entries 2:5 have scalars that reflect the values that the other labels can take. The order of the columns is the same for all entries in your dataset. A missing label is marked by a special value of your choosing. This special value has to lie outside the set of valid label values. This will depend on what those labels represent. If a label value of -1 never occurs you can use this to flag a missing label.<\/p>\n"}]}
{"QuestionId":32785019,"AnswerCount":0,"Tags":"<caffe><pycaffe>","CreationDate":"2015-09-25T14:56:42.823","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"\"ImportError: cannot import name layers\" in caffe \/ pycaffe","Body":"<p>I'm following the ipython notebook example on training caffe via its python interface. However the caffe module seems not to contain the modules described in the <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow\">tutorial<\/a><\/p>\n\n<p>When I enter <\/p>\n\n<pre><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>when I do dir(caffe) I get:<\/p>\n\n<pre><\/pre>\n\n<p>so it doesn't contain a module called \"layers\".<\/p>\n\n<p>Is the tutorial somehow outdated? Is there a newer version which works for my version of caffe?<\/p>\n","answers":[]}
{"QuestionId":32794636,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-09-26T07:14:56.090","AcceptedAnswerId":32794839.0,"OwnerUserId":445114.0,"Title":"theano - how to efficiently replicate and add tensors?","Body":"<p>I have one tensor in shape (2, G) and another in shape (N, 2).<\/p>\n\n<p>I need to add them in such a way that the output is (N, 2, G), meaning that the first tensor is replicated to (N, 2, G) and then the second tensor is added to each matrix along the third dimension. (or vice versa: the second tensor is replicated to (N, 2, G) and the first one is added to every sub-tensor along the first dimension).<\/p>\n\n<p>How can this done efficiently in Theano?\nThanks.<\/p>\n","answers":[{"AnswerId":"32794839","CreationDate":"2015-09-26T07:41:15.870","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>In an attempt to understand the problem, the following example is assumed to be representative.<\/p>\n\n<p>If<\/p>\n\n<pre><code>A = [[1, 2, 3],\n     [4, 5, 6]]\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>B = [[1, 2],\n     [3, 4],\n     [5, 6],\n     [7, 8]]\n<\/code><\/pre>\n\n<p>then the result should be<\/p>\n\n<pre><code>C = [[[  2.   3.   4.]\n      [  6.   7.   8.]]\n     [[  4.   5.   6.]\n      [  8.   9.  10.]]\n     [[  6.   7.   8.]\n      [ 10.  11.  12.]]\n     [[  8.   9.  10.]\n      [ 12.  13.  14.]]]\n<\/code><\/pre>\n\n<p>Here <code>G=3<\/code> and <code>N=4<\/code>.<\/p>\n\n<p>To achieve this in Theano, one need only add new broadcastable dimensions and rely on broadcasting to get the desired result.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\nx = tt.matrix()\ny = tt.matrix()\nz = x.dimshuffle('x', 0, 1) + y.dimshuffle(0, 1, 'x')\nf = theano.function([x, y], outputs=z)\nprint f(numpy.array([[1, 2, 3], [4, 5, 6]]), numpy.array([[1, 2], [3, 4], [5, 6], [7, 8]]))\n<\/code><\/pre>\n"}]}
{"QuestionId":32799733,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-09-26T17:10:23.020","AcceptedAnswerId":32805437.0,"OwnerUserId":445114.0,"Title":"Clip parts of a tensor","Body":"<p>I have a theano tensor and I would like to clip its values, but each index to a different range.<\/p>\n\n<p>For example, if I have a vector [a,b,c] , I want to clip a to [0,1] , clip b to [2,3] and c to [3,5].<\/p>\n\n<p>How can I do that efficiently?\nThanks!<\/p>\n","answers":[{"AnswerId":"32805437","CreationDate":"2015-09-27T07:06:47.150","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The <code>theano.tensor.clip<\/code> operation supports symbolic minimum and maximum values so you can pass three tensors, all of the same shape, and it will perform an element-wise clip of the first with respect to the second (minimum) and third (maximum).<\/p>\n\n<p>This code shows two variations on this theme. <code>v1<\/code> requires the minimum and maximum values to be passed as separate vectors while <code>v2<\/code> allows the minimum and maximum values to be passed more like a list of pairs, represented as a two column matrix.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\n\ndef v1():\n    x = tt.vector()\n    min_x = tt.vector()\n    max_x = tt.vector()\n    y = tt.clip(x, min_x, max_x)\n    f = theano.function([x, min_x, max_x], outputs=y)\n    print f([2, 1, 4], [0, 2, 3], [1, 3, 5])\n\n\ndef v2():\n    x = tt.vector()\n    min_max = tt.matrix()\n    y = tt.clip(x, min_max[:, 0], min_max[:, 1])\n    f = theano.function([x, min_max], outputs=y)\n    print f([2, 1, 4], [[0, 1], [2, 3], [3, 5]])\n\n\ndef main():\n    v1()\n    v2()\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":32804790,"AnswerCount":1,"Tags":"<gpu><cpu><shared><theano>","CreationDate":"2015-09-27T05:07:14.033","AcceptedAnswerId":32805391.0,"OwnerUserId":1282982.0,"Title":"Different results of CPU and GPU with Theano","Body":"<p>I have the following piece of code:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Run with CPU<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Run with GPU<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>How can I get the same results as CPU?<\/p>\n","answers":[{"AnswerId":"32805391","CreationDate":"2015-09-27T06:59:56.923","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The method you are using to access child nodes in the computation graph, <code>.owner.inputs[0]<\/code>, is not appropriate for cross-platform code in general.<\/p>\n\n<p>You call the shared variable <code>x<\/code> so the correct way to access <code>x<\/code>'s value is to use <code>x.get_value()<\/code>.<\/p>\n\n<p>This code should run the same on CPU and GPU:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nx = theano.shared(np.asarray([1, 2, 3], dtype=theano.config.floatX), borrow=True)\ny = T.cast(x, 'int32')\nprint 'type of y: ', type(y)\nprint 'type of x: ', type(x)\nprint 'value of x: ', x.get_value(borrow=True)\n<\/code><\/pre>\n\n<p>If you want to see the result of applying the symbolic cast operation to <code>x<\/code>, the symbolic result of which you call <code>y<\/code>, then you could do this:<\/p>\n\n<pre><code>print 'value of y: ', y.eval()\n<\/code><\/pre>\n"}]}
{"QuestionId":32808855,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-09-27T14:16:01.007","AcceptedAnswerId":32810143.0,"OwnerUserId":639973.0,"Title":"Data Layer Prefetch Queue Empty followed by loss 0","Body":"<p>I'm fine-tuning caffenet on my own dataset.\nI've successfully done it on flickr style dataset,\nso I've modified solver and train_val prototxt files to point to my dataset, with new output dimension (=100 classes).<\/p>\n\n<p>The size of my dataset is 1.88M images for training, and .48M images for testing.<\/p>\n\n<p>When I started training, it paused for a while showing<\/p>\n\n<pre><\/pre>\n\n<p>and then it proceeded as following:<\/p>\n\n<pre><\/pre>\n\n<p>and so on.\nSo apparently, no loss is being calculated at all. Weirdly, the initial loss at iteration 0 seems fair enough though.<\/p>\n\n<p>So apparently, data were not loaded properly. I'm using text files to list the file paths and their labels. <\/p>\n\n<p>I'm using softmax loss.<\/p>\n\n<p>I've tried reducing initial learning rate to 1\/10 but no change.<\/p>\n\n<p>**edit: I've tried initial learning rate of 1\/1000. Initially it seemed to work. After about 300 iterations, its loss quickly declined again and accuracy reached 100%, signifying that something might've gone wrong again. The symptoms appeared with both caffenet and VGG.<\/p>\n\n<p>Any idea what might have possibly gone wrong?<\/p>\n","answers":[{"AnswerId":"32810143","CreationDate":"2015-09-27T16:25:41.463","ParentId":null,"OwnerUserId":"639973","Title":null,"Body":"<p>The problem was that the text files for training and test data were not shuffled, but sorted. Thus, test was done only on the images of same class, and parameters were learned from the images of that same class as well, leading to 100% accuracy very quickly. <\/p>\n"}]}
{"QuestionId":32812159,"AnswerCount":0,"Tags":"<python><compilation><server><theano><gensim>","CreationDate":"2015-09-27T19:48:10.447","AcceptedAnswerId":null,"OwnerUserId":4801125.0,"Title":"Cannot find -lInclude on server side","Body":"<p>After I installed the Theano and genesim by pip install xxx --user on server side, I tried to import them in python code, but I had this error: <\/p>\n\n<p>Exception: Compilation failed (return status=1): \/usr\/bin\/ld: cannot find -lInclude. collect2: ld returned 1 exit status. <\/p>\n\n<p>In addition, I have no root permission, how can I make it?<\/p>\n","answers":[]}
{"QuestionId":32813590,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-09-27T22:41:55.470","AcceptedAnswerId":32816969.0,"OwnerUserId":5382728.0,"Title":"Differences between index-assignment in Numpy and Theano's set_subtensor()","Body":"<p>I am trying to do index-assignment in Theano using set_subtensor(), but it is giving different results to Numpy's index-assignment. Am I doing something wrong, or is this a difference in how set_subtensor and Numpy's index-assignment work?<\/p>\n\n<p>What I want to do:<\/p>\n\n<pre><\/pre>\n\n<p>Trying to do the same thing in Theano:<\/p>\n\n<pre><\/pre>\n\n<p>Raises this error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32816969","CreationDate":"2015-09-28T06:36:34.007","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This highlights a subtle difference between numpy and Theano but it can be worked around easily.<\/p>\n\n<p>Advanced indexing can be enabled in numpy by using a list of positions or a tuple of positions. In Theano, one can only use a tuple of positions.<\/p>\n\n<p>So changing<\/p>\n\n<pre><code>X = T.set_subtensor(X[[[0, 1], [0, 1]]], np.array([1, 2]))\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>X = T.set_subtensor(X[([0, 1], [0, 1])], np.array([1, 2]))\n<\/code><\/pre>\n\n<p>solves the problem in Theano.<\/p>\n\n<p>One continues to get the same result in numpy if one changes<\/p>\n\n<pre><code>X[[[0, 1], [0, 1]]] = np.array([1, 2])\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>X[([0, 1], [0, 1])] = np.array([1, 2])\n<\/code><\/pre>\n"}]}
{"QuestionId":32817496,"AnswerCount":1,"Tags":"<neural-network><theano>","CreationDate":"2015-09-28T07:13:31.103","AcceptedAnswerId":32820816.0,"OwnerUserId":1244924.0,"Title":"Extraction of neural network activation values from Python Theano+Keras","Body":"<p>Say I have simple MLP network for 2 class problem:<\/p>\n\n<pre><\/pre>\n\n<p>after training this network I was unable to see any values in W object when observed it in debug mode.<\/p>\n\n<p>Are they stored somewhere in Theano's computational graph, if so is it possible to get them? If not why all values in activation layer of model are None?<\/p>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>Sorry for being too quick. Tensor object holding weights from Dense layer can be perfectly found. But invoking:<\/p>\n\n<pre><\/pre>\n\n<p>gives me Activation layer. Where I would like to see activation levels. Instead I see only:<\/p>\n\n<pre><\/pre>\n\n<p>I assume keras just clears all these values after model evaluation - is it true?\nIf so the only way to record activations from neurons is to create custom Activation layer which would record needed info - right? <\/p>\n","answers":[{"AnswerId":"32820816","CreationDate":"2015-09-28T10:29:02.310","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I am not familiar with Keras but if it is building a conventional Theano neural network computation graph then it is not possible to view the activation values in the way you propose.<\/p>\n\n<p>Conventionally, only weights are stored persistently, as shared variables. Values computed in intermediate stages of a Theano computation are transient and can only be viewed via debugging the execution of the compiled Theano function (note that this is not easy to do -- debugging the host Python application is not sufficient).<\/p>\n\n<p>If you were building the computation directly instead of using Keras I would advise including the intermediate activation values that you are interested in in the output values list of the Theano function. I cannot comment on how this can be achieved via Keras.<\/p>\n"}]}
{"QuestionId":32818013,"AnswerCount":0,"Tags":"<macos><luajit><torch>","CreationDate":"2015-09-28T07:47:14.640","AcceptedAnswerId":null,"OwnerUserId":3063834.0,"Title":"Getting error when installing Torch on OS X 10.11","Body":"<p>I am trying to install Torch on my Mac (OS X 10.11 GM) by following the guide on <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">http:\/\/torch.ch\/docs\/getting-started.html<\/a>.<\/p>\n\n<p>However, when I run <\/p>\n\n<pre><\/pre>\n\n<p>I always get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I cannot find other ways to install Torch.<\/p>\n","answers":[]}
{"QuestionId":32818091,"AnswerCount":1,"Tags":"<python><segmentation-fault><theano>","CreationDate":"2015-09-28T07:51:38.573","AcceptedAnswerId":null,"OwnerUserId":1215889.0,"Title":"Theano is throwing Segmentation fault","Body":"<p>I have a  code which is using batches of data. When I am increasing the batch size, it is throwing me the SegFault error. It is working perfectly fine upto . But  it to  is throwing me error. I also checked that it is using only  at any time. But I have  on my system. Can anyone help me here to debug the issue.<\/p>\n","answers":[{"AnswerId":"32819347","CreationDate":"2015-09-28T09:11:27.783","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can profile Theano's memory use by enabling the <code>profile<\/code> and <code>profile_memory<\/code> configuration flags. See the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/profiling.html\" rel=\"nofollow\">documentation<\/a> for more details.<\/p>\n\n<p>There is also more information on Python\/Theano memory management <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/python-memory-management.html\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":32820472,"AnswerCount":2,"Tags":"<neural-network><torch>","CreationDate":"2015-09-28T10:11:02.090","AcceptedAnswerId":null,"OwnerUserId":5368537.0,"Title":"Single Classifier network in Torch","Body":"<p>I am new to Torch.\nI am trying to run single classifier Experiment in Torch.But,I am getting the following error when Training is started,<\/p>\n\n<p><strong>\/torch\/install\/bin\/luajit: bad argument #2 to '?' (out of range at \/torch\/pkg\/torch\/generic\/Tensor.c:853)<\/strong><\/p>\n\n<p><strong>stack traceback:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to train single classifier network in Torch?<br>\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"48023004","CreationDate":"2017-12-29T12:58:25.830","ParentId":null,"OwnerUserId":"4360306","Title":null,"Body":"<p>Yes, you can train in one class.<\/p>\n\n<p>The error you got is pointing towards confusion matrix.\nFor one class it should be in the following way:\n-- classes<\/p>\n\n<pre><code>classes = {1}\n<\/code><\/pre>\n\n<p>-- This matrix records the current confusion across classes<\/p>\n\n<pre><code>confusion = optim.ConfusionMatrix(classes)\n<\/code><\/pre>\n\n<p>Your train labels or test labels should never contain '0' number as the label.<\/p>\n"},{"AnswerId":"32947531","CreationDate":"2015-10-05T11:29:53.057","ParentId":null,"OwnerUserId":"5368537","Title":null,"Body":"<p>Thanks for your reply.\nMy model contains,<\/p>\n\n<p>classifier = nn.Sequential()<\/p>\n\n<p>classifier:add(nn.Dropout(0.5))<\/p>\n\n<p>classifier:add(nn.Linear(512,512))<\/p>\n\n<p>classifier:add(nn.BatchNormalization(512))<\/p>\n\n<p>classifier:add(nn.ReLU(true))<\/p>\n\n<p>classifier:add(nn.Dropout(0.5))<\/p>\n\n<p>classifier:add(nn.Linear(512,1))<\/p>\n\n<p>classifier:add(nn.LogSoftMax())<\/p>\n\n<p>and I am using  <strong>nn.CrossEntropyCriterion()<\/strong> for Loss in the network.<\/p>\n\n<p>Is it possible to run single classifier experiment?<\/p>\n"}]}
{"QuestionId":32820679,"AnswerCount":1,"Tags":"<python><loops><iteration><theano>","CreationDate":"2015-09-28T10:21:34.670","AcceptedAnswerId":32821367.0,"OwnerUserId":1861627.0,"Title":"Scanning over different dimensions of tensors in theano","Body":"<p>I'm moving my first steps with  and I cannot figure out how to solve this problem which could be actually very easy. <\/p>\n\n<p>I have a  tensor, like the following:<\/p>\n\n<pre><\/pre>\n\n<p>So I have  sequences, each of them of length  with their elements that are vectors of dimension . Actually, the sequences can be of different length but I could think of padding them with  vectors, as shown above. <\/p>\n\n<p>What I want to do is, first scan through the first axis of the tensor and sum up all the vector in the lists <strong>up to the the first [0 0] vector<\/strong> -- that's why I added the [9 9] at the end of the first tensor slice, in order to check the sum exit condition [1]. I should end up in . I tried in many ways to solve this problem which seems to me <em>just<\/em> a nested looping problem... but always got some weird errors[2].<\/p>\n\n<p>Thanks,<br>\nGiulio  <\/p>\n\n<p>--<br>\n[1]: the actual problem is the training of a recurrent neural network for NLP purposes, with  the dimension of the batch,  the max length of a sentence in the batch and  the dimension of the representation of each word. I omitted the problem so that I could focus on the simplest coding aspect.<br>\n[2] I omit the history of my failures, maybe I could add them later.<\/p>\n","answers":[{"AnswerId":"32821367","CreationDate":"2015-09-28T10:59:45.313","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If your sequences are always zero padded then you can just sum along the axis of interest since the padding regions will not change the sum. However, if the padding regions may contain non-zero values there are two approaches.<\/p>\n\n<ol>\n<li>Use scan. This is slow and should be avoided if possible. In fact it can be avoided because,<\/li>\n<li>Create a binary mask and multiply out the padding region.<\/li>\n<\/ol>\n\n<p>Here's some code that illustrates these three approaches. For the two approaches that allow for non-zero padding regions (<code>v2<\/code> and <code>v3<\/code>) the computation needs an additional input: a vector giving the lengths of the sequences within the batch.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef v1():\n    # NOTE: [9, 9] element changed to [0, 0] \n    # since zero padding must be used for\n    # this method\n    x_data = [[[1, 1], [1, 1], [0, 0], [0, 0]],\n              [[2, 2], [2, 2], [2, 2], [0, 0]],\n              [[3, 3], [3, 3], [3, 3], [3, 3]]]\n    x = tt.tensor3()\n    x.tag.test_value = x_data\n    y = x.sum(axis=1)\n    f = theano.function([x], outputs=y)\n    print f(x_data)\n\n\ndef v2_step(i_t, s_tm1, x, l):\n    in_sequence = tt.lt(i_t, l).dimshuffle(0, 'x')\n    s_t = s_tm1 + tt.switch(in_sequence, x[i_t], 0)\n    return s_t\n\n\ndef v2():\n    x_data = [[[1, 1], [1, 1], [0, 0], [9, 9]],\n              [[2, 2], [2, 2], [2, 2], [0, 0]],\n              [[3, 3], [3, 3], [3, 3], [3, 3]]]\n    l_data = [2, 3, 4]\n    x = tt.tensor3()\n    x.tag.test_value = x_data\n    l = tt.lvector()\n    l.tag.test_value = l_data\n    # Must dimshuffle first because scan can only iterate over first (0'th) axis.\n    x_hat = x.dimshuffle(1, 0, 2)\n    y, _ = theano.scan(v2_step, sequences=[tt.arange(x_hat.shape[0])],\n                       outputs_info=[tt.zeros_like(x_hat[0])],\n                       non_sequences=[x_hat, l], strict=True)\n    f = theano.function([x, l], outputs=y[-1])\n    print f(x_data, l_data)\n\n\ndef v3():\n    x_data = [[[1, 1], [1, 1], [0, 0], [9, 9]],\n              [[2, 2], [2, 2], [2, 2], [0, 0]],\n              [[3, 3], [3, 3], [3, 3], [3, 3]]]\n    l_data = [2, 3, 4]\n    x = tt.tensor3()\n    x.tag.test_value = x_data\n    l = tt.lvector()\n    l.tag.test_value = l_data\n    indexes = tt.arange(x.shape[1]).dimshuffle('x', 0)\n    mask = tt.lt(indexes, l.dimshuffle(0, 'x')).dimshuffle(0, 1, 'x')\n    y = (mask * x).sum(axis=1)\n    f = theano.function([x, l], outputs=y)\n    print f(x_data, l_data)\n\n\ndef main():\n    theano.config.compute_test_value = 'raise'\n    v1()\n    v2()\n    v3()\n\n\nmain()\n<\/code><\/pre>\n\n<p>In general, <strong>if your step function is dependent on the output of a previous step then you need to use<\/strong> <code>scan<\/code>.  <\/p>\n\n<p>If every step\/iteration could, in principle, be executed concurrently (i.e. they don't rely on each other at all) then there is often a much more efficient way to do this without using scan<\/p>\n"}]}
{"QuestionId":32824441,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-09-28T13:35:46.040","AcceptedAnswerId":null,"OwnerUserId":2971693.0,"Title":"RNN with dynamic inputs at each time step in Theano","Body":"<p>I was wondering if it was possible to implement a recurrent network in Theano in the case where inputs are not known initially.\nSpecifically, I have in mind the 'Recurrent Models of Visual Attention' paper (<a href=\"http:\/\/arxiv.org\/abs\/1406.6247\" rel=\"nofollow\">http:\/\/arxiv.org\/abs\/1406.6247<\/a>) and the part concerning game playing. In this case, each game image is only available after the network has outputted an action.<\/p>\n\n<p>As I understand, RNN's in Theano are implemented using theano.scan function, which expects a sequence as an input. Obviously, I can't produce such a sequence of game images without running the full recurrent loop and recording the actions that would be generated. And I can't run the loop and generate the sequence of actions, since I don't have the sequence of game images to pass as an input.<\/p>\n\n<p>So, it would seem that under those conditions I can't use a proper back propagation and train the network correctly. I could run each iteration of the loop manually, but then there would be no BPTT.<\/p>\n\n<p>What am I missing here?\nIs it possible to implement the algorithm in the paper describing the game playing part in Theano (I've seen implementations of digit classification part, but it's easier, since the input never changes)?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"32825455","CreationDate":"2015-09-28T14:25:23.953","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If I understand it correctly (assuming you're referring to the simple game of \"catch\"), I don't see any problem.<\/p>\n\n<p>There's an initial state for the game (i.e. the initial position of the paddle and the initial position of the ball) that can be provided to the network in the first time step. The network predicts an action to be performed and the game state is updated based on the chosen action. The updated game state is then provided as input to the network in the second time step.<\/p>\n\n<h2>Update<\/h2>\n\n<p>Here's some sample code showing how to use the output of an earlier time step to update a state within a <code>theano.scan<\/code> operation.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\n\ndef choose_action(s):\n    # TODO: Given the game state s, choose which action to perform\n    return s.argmax()\n\n\ndef update_state(s, y):\n    # TODO: Update game state s given action y\n    return s + y\n\n\ndef is_end_state(s):\n    # TODO: Determine whether game state s is an end-game state\n    return s.max() &gt; 100\n\n\ndef step(s_tm1):\n    y_tm1 = choose_action(s_tm1)\n    s_t = update_state(s_tm1, y_tm1)\n    return (y_tm1, s_t), theano.scan_module.until(is_end_state(s_t))\n\n\ndef main():\n    theano.config.compute_test_value = 'raise'\n    initial_state = tt.matrix()\n    initial_state.tag.test_value = [[0, 2, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 1]]\n    (action_sequence, state_sequence), _ = theano.scan(step, outputs_info=[None, initial_state], n_steps=1000)\n    state_sequence = tt.concatenate([[initial_state], state_sequence])\n    f = theano.function([initial_state], outputs=[action_sequence, state_sequence])\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":32833458,"AnswerCount":0,"Tags":"<lua><deep-learning><luajit><torch>","CreationDate":"2015-09-28T23:42:15.550","AcceptedAnswerId":null,"OwnerUserId":132002.0,"Title":"Loop in torch seems much slower than it should be","Body":"<p>I've the following piece of code<\/p>\n\n<pre><\/pre>\n\n<p>in a custom module in torch.<\/p>\n\n<p>For values of A = 18K, B=30, this loop is taking ~4 s (measured using torch.Timer) which seems extremely high to me.<\/p>\n\n<p>For context, C is a FloatTensor of size (A, B) and gradOutput of (2, A, B). The processor is AMD Opteron 4386 having 16 cores each with cache size 2MB.<\/p>\n\n<p>So, it seems to me that C, gradOutput, gradInput should already be cached so 4s for such a short loop seems insanely high to me.<\/p>\n\n<p>Anyone have any ideas why it's so slow and if\/how I can improve the performance?<\/p>\n","answers":[]}
{"QuestionId":32835231,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2015-09-29T03:30:52.977","AcceptedAnswerId":null,"OwnerUserId":499372.0,"Title":"Transferring to GPU in a for-loop","Body":"<p>We define a tensor as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Then we assign a value to  and call  in a loop. It takes about 10 seconds in the first iteration, but it takes 0.01 second in the next iterations. I would appreciate it if someone explains why this happens and how we can fix the problem.<\/p>\n","answers":[]}
{"QuestionId":32846518,"AnswerCount":1,"Tags":"<image-processing><neural-network><caffe><conv-neural-network>","CreationDate":"2015-09-29T14:24:10.067","AcceptedAnswerId":32867356.0,"OwnerUserId":5389199.0,"Title":"How to use caffe convnet library to detect facial expressions?","Body":"<p>How can I use caffe convnet to detect facial expressions? <\/p>\n\n<p>I have a image dataset, Cohn Kanade, and I want to train caffe convnet with this dataset. <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe<\/a> has a documentation site, but its not explain how to train my own data. Just with pre trained data. <\/p>\n\n<p>Can someone teach me how to do it? <\/p>\n","answers":[{"AnswerId":"32867356","CreationDate":"2015-09-30T13:26:04.037","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Caffe supports multiple formats for the input data (HDF5\/lmdb\/leveldb). It's just a matter of picking the one you feel most comfortable with. Here are a couple of options:<\/p>\n\n<ol>\n<li><strong>caffe\/build\/tools\/convert_imageset:<\/strong><\/li>\n<\/ol>\n\n<p><code>convert_imageset<\/code> is one of the command line tools you get from building caffe.<\/p>\n\n<p>Usage is along the lines of:<\/p>\n\n<ul>\n<li>specifying a list of images and label pairs in a text file. 1 row per pair.<\/li>\n<li>specifying where the images are located.<\/li>\n<li>Choosing a backend db (which format). Default is lmdb which should be fine.<\/li>\n<\/ul>\n\n<p>You need to write up a text file where each line starts with the filename of the image followed by a scalar label (e.g. 0, 1, 2,...)<\/p>\n\n<ol start=\"2\">\n<li><strong>Construct your lmdb in python using Caffe's <code>Datum<\/code> class:<\/strong><\/li>\n<\/ol>\n\n<p>This requires building caffe's python interface. Here you write some python code that:<\/p>\n\n<ul>\n<li>iterates through a list of images<\/li>\n<li>loads the images into a <code>numpy<\/code> array.<\/li>\n<li>Constructs a caffe <code>Datum<\/code> object<\/li>\n<li>Assigns the image data to the <code>Datum<\/code> object.<\/li>\n<li>The <code>Datum<\/code> class has a member called <code>label<\/code> you can set it to the AU class from your CK dataset, if that is what you want your network to classify.<\/li>\n<li>Writes the <code>Datum<\/code> object to the db and moves on to the next image.<\/li>\n<\/ul>\n\n<p><a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow\">Here's<\/a> a code snippet of converting images to an lmdb from a blog post by Gustav Larsson. In his example he constructs an lmdb of images and label pairs for image classification.<\/p>\n\n<p><strong>Loading the lmdb into your network:<\/strong><\/p>\n\n<p>This is done exactly like in the LeNet example. This Data layer at the beginning of the network prototxt that describes the LeNet model.<\/p>\n\n<pre><code>layer {\n  name: \"mnist\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TRAIN\n  }\n  transform_param {\n    scale: 0.00390625\n  }\n  data_param {\n    source: \"examples\/mnist\/mnist_train_lmdb\"\n    batch_size: 64\n    backend: LMDB\n  }\n}\n<\/code><\/pre>\n\n<p>The source field is where you point caffe to the location of the lmdb you just created.<\/p>\n\n<p>Something more related to performance and not critical to getting this to work is specifying how to normalize the input features. This is done through the <code>transform_param<\/code> field. CK+ has fixed size images, so no need for resizing. One thing you do need though is normalize the grayscale values. You can do this through mean subtraction. A simple of doing this is to replace the value of <code>transform_param:scale<\/code> with the mean value of the gray scale intensities in your CK+ dataset.<\/p>\n"}]}
{"QuestionId":32856173,"AnswerCount":1,"Tags":"<c++><vector><compiler-errors><machine-learning><caffe>","CreationDate":"2015-09-30T01:06:18.550","AcceptedAnswerId":32856416.0,"OwnerUserId":1377127.0,"Title":"Error when I push back a pair in C++","Body":"<p>I've been trying to compile my program which should push a string and a float pair back on a vector:<\/p>\n\n<pre><\/pre>\n\n<p>However, every time I try to compile this, I get this error:<\/p>\n\n<blockquote>\n  <p>error: no matching member function for call to 'push_back'\n  predictions.push_back(std::make_pair(labels_[idx], output[idx]));<\/p>\n<\/blockquote>\n\n<p>I also get a few other warnings saying things like<\/p>\n\n<blockquote>\n  <p>candidate function not viable: no known conversion from 'pair&lt;[...],\n  typename __make_pair_return >\n  &amp;>::type>' to 'const pair&lt;[...], float>' for 1st argument\n  _LIBCPP_INLINE_VISIBILITY void push_back(const_reference __x);<\/p>\n<\/blockquote>\n\n<p>and<\/p>\n\n<blockquote>\n  <p>candidate function not viable: no known conversion from 'pair&lt;[...],\n  typename __make_pair_return >\n  &amp;>::type>' to 'pair&lt;[...], float>' for 1st argument\n  _LIBCPP_INLINE_VISIBILITY void push_back(value_type&amp;&amp; __x);<\/p>\n<\/blockquote>\n\n<p>I've been trying to rewrite things and modify my functions but I can't work out why this error remains, does anyone know what I can do to fix this?<\/p>\n\n<p>Here is the code in context if that helps, the header file:<\/p>\n\n<pre><\/pre>\n\n<p>Class File:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks so much!<\/p>\n","answers":[{"AnswerId":"32856416","CreationDate":"2015-09-30T01:40:42.863","ParentId":null,"OwnerUserId":"293647","Title":null,"Body":"<pre><code>typedef std::pair&lt;string, float&gt; Prediction;\nstd::vector&lt;Prediction&gt; predictions;\nstd::vector&lt; std::vector&lt;float&gt; &gt; output = Predict(img, img.size());\n<\/code><\/pre>\n\n<p>make_pair expects a string and a <strong>float<\/strong>. output[idx] gives a <strong>vector of floats<\/strong>. So you need output[i][idx] for only a <strong>float<\/strong>.<\/p>\n"}]}
{"QuestionId":32856606,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-09-30T02:04:54.687","AcceptedAnswerId":32954705.0,"OwnerUserId":639973.0,"Title":"Google Inceptionism: obtain images by class","Body":"<p>In the famous Google Inceptionism article,\n<a href=\"http:\/\/googleresearch.blogspot.jp\/2015\/06\/inceptionism-going-deeper-into-neural.html\" rel=\"noreferrer\">http:\/\/googleresearch.blogspot.jp\/2015\/06\/inceptionism-going-deeper-into-neural.html<\/a>\nthey show images obtained for each class, such as banana or ant. I want to do the same for other datasets.<\/p>\n\n<p>The article does describe how it was obtained, but I feel that the explanation is insufficient.<\/p>\n\n<p>There's a related code\n<a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb<\/a><\/p>\n\n<p>but what it does is to produce a random dreamy image, rather than specifying a class and learn what it looks like in the network, as shown in the article above.<\/p>\n\n<p>Could anyone give a more concrete overview, or code\/tutorial on how to generate images for specific class? (preferably assuming caffe framework)<\/p>\n","answers":[{"AnswerId":"32954705","CreationDate":"2015-10-05T17:45:12.980","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>I think <a href=\"https:\/\/github.com\/artvandelay\/Deep_Inside_Convolutional_Networks\/blob\/master\/visualize.py\" rel=\"nofollow\">this code<\/a> is a good starting point to reproduce the images Google team published. The procedure looks clear:<\/p>\n\n<ol>\n<li>Start with a pure noise image and a class (say \"cat\")<\/li>\n<li>Perform a forward pass and backpropagate the error wrt the imposed class label<\/li>\n<li>Update the initial image with the gradient computed at the data layer<\/li>\n<\/ol>\n\n<p>There are some tricks involved, that can be found in the <a href=\"http:\/\/arxiv.org\/abs\/1312.6034\" rel=\"nofollow\">original paper<\/a>.<\/p>\n\n<p>It seems that the main difference is that Google folks tried to get a more \"realistic\" image:<\/p>\n\n<blockquote>\n  <p>By itself, that doesn\u2019t work very well, but it does if we impose a prior constraint that the image should have similar statistics to natural images, such as neighboring pixels needing to be correlated. <\/p>\n<\/blockquote>\n"}]}
{"QuestionId":32859643,"AnswerCount":1,"Tags":"<python><loops><iteration><theano>","CreationDate":"2015-09-30T06:52:42.783","AcceptedAnswerId":32862567.0,"OwnerUserId":1861627.0,"Title":"Scanning without sequences in theano? (emulating range())","Body":"<p>Theano newbie here. I am doing some experiments in order to generate variable length sequences. I started with the simplest thing coming to my mind: emulating range(). Here is the simple code I wrote:<\/p>\n\n<pre><\/pre>\n\n<p>So I had to use  as the length of a range to be used as input of the  theano . So, my main question is this: is there any way to use a  without an input sequence? And, as I suppose the answer is <em>no<\/em>, the next question is: is this the correct (most efficient, ecc) way to do what I am traying to do?<\/p>\n","answers":[{"AnswerId":"32862567","CreationDate":"2015-09-30T09:28:35.640","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There is no need to provide an input sequence to scan. You can instead specify the number of iterations via scan's <code>n_steps<\/code> parameter. Optionally, you can also specify a condition under which the scan should stop early via <code>theano.scan_module.until<\/code>.<\/p>\n\n<p>So Python's <code>range<\/code> function can be emulated using Theano's <code>scan<\/code> without specifying an input sequence by figuring out how many iterations would be required to construct the requested sequence.<\/p>\n\n<p>Here's an implementation of the range function based on Theano's <code>scan<\/code>. The only complicated part is figuring out how many steps are required.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\nimport theano.ifelse\n\n\ndef scan_range_step(x_tm1, step):\n    return x_tm1 + step\n\n\ndef compile_theano_range():\n    tt.arange\n    symbolic_start = tt.lscalar()\n    symbolic_stop = tt.lscalar()\n    symbolic_step = tt.lscalar()\n    n_steps = tt.cast(\n        tt.ceil(tt.abs_(symbolic_stop - symbolic_start) \/ tt.cast(tt.abs_(symbolic_step), theano.config.floatX)),\n        'int64') - 1\n    outputs, _ = theano.scan(scan_range_step, outputs_info=[symbolic_start], n_steps=n_steps,\n                             non_sequences=[symbolic_step], strict=True)\n    outputs = theano.ifelse.ifelse(tt.eq(n_steps, 0), tt.stack(symbolic_start), outputs)\n    f = theano.function([symbolic_start, symbolic_stop, symbolic_step],\n                        outputs=tt.concatenate([[symbolic_start], outputs]))\n\n    def theano_range(start, stop=None, step=1):\n        assert isinstance(start, int)\n        assert isinstance(step, int)\n        if step == 0:\n            raise ValueError()\n        if stop is None:\n            stop = start\n            start = 0\n        else:\n            assert isinstance(stop, int)\n        if start == stop:\n            return []\n        if stop &lt; start and step &gt; 0:\n            return []\n        if stop &gt; start and step &lt; 0:\n            return []\n        return f(start, stop, step)\n\n    return theano_range\n\n\ndef main():\n    theano_range = compile_theano_range()\n    python_range = range\n\n    for start in [-10, -5, -1, 0, 1, 5, 10]:\n        for stop in [-10, -5, -1, 0, 1, 5, 10]:\n            for step in [-3, -2, -1, 1, 2, 3]:\n                a = theano_range(start, stop, step)\n                b = python_range(start, stop, step)\n                assert numpy.all(numpy.equal(a, b)), (start, stop, step, a, b)\n\n\nmain()\n<\/code><\/pre>\n\n<p>Clearly this is a daft thing to do\/use for real since Theano already provides a symbolic version of Python's <code>range<\/code> function, i.e. <code>theano.tensor.arange<\/code>. The built in implementation is also far more efficient than our <code>scan<\/code> version because it doesn't use <code>scan<\/code>, it uses a custom operation instead.  <\/p>\n\n<p>As a rule of thumb: you <em>have to<\/em> set a maximum number of iteration steps via the <code>range<\/code> or the the <code>n_steps<\/code> argument. You can set it to a very large number and then use <code>theano.scan_module.until<\/code> to stop the iteration at an earlier stage if your stop condition is met.<\/p>\n"}]}
{"QuestionId":32867030,"AnswerCount":3,"Tags":"<theano>","CreationDate":"2015-09-30T13:10:46.480","AcceptedAnswerId":32867775.0,"OwnerUserId":133374.0,"Title":"1-of-k (one-hot) encoding in Theano","Body":"<p>I'm doing <a href=\"https:\/\/stackoverflow.com\/questions\/30604907\/generate-sequence-by-indices\">this<\/a> for Numpy.  is a list with indices. I.e. this implements a 1-of-k encoding (also called one-hot).<\/p>\n\n<pre><\/pre>\n\n<p>How would I do the same thing in Theano? (Most efficient solution which also is efficient for CUDA.)<\/p>\n","answers":[{"AnswerId":"32869822","CreationDate":"2015-09-30T15:18:13.280","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>My solution:<\/p>\n\n<pre><code>def class_idx_seq_to_1_of_k(seq, num_classes, dtype=\"float32\"):\n  shape = [seq.shape[i] for i in range(seq.ndim)] + [num_classes]\n  eye = T.eye(num_classes, dtype=dtype)\n  m = eye[T.cast(seq, 'int32')].reshape(shape)\n  return m\n<\/code><\/pre>\n"},{"AnswerId":"32898087","CreationDate":"2015-10-01T22:58:13.423","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Isn't this a simple broadcast?<\/p>\n\n<pre><code>import numpy as np\nseq = np.random.randint(0, 5, 10)\none_hot = seq[:, np.newaxis] == np.arange(seq.max())  # class ids contiguous, maximum class seen (otherwise set variable n_classes)\n\nprint one_hot\n\n\nimport theano\nimport theano.tensor as T\n\nt_seq = T.ivector()\nt_one_hot = T.eq(t_seq.reshape((-1, 1)), T.arange(t_seq.max()))\n\nf = theano.function([t_seq], [t_one_hot])\n\nprint f(seq.astype('int32'))\n<\/code><\/pre>\n"},{"AnswerId":"32867775","CreationDate":"2015-09-30T13:46:27.310","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There is a built in function to do this (<a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html#theano.tensor.extra_ops.to_one_hot\" rel=\"nofollow\"><code>theano.tensor.extra_ops.to_one_hot<\/code><\/a>) but it is still much slower then doing it in numpy. If feasible for your task, you might be better off computing this outside Theano and passing the dense result in as an input instead of passing just the indices.<\/p>\n\n<p>Here's some code illustrating three numpy methods, and four Theano methods. This code includes the answers provided by Albert (<code>numpy_1_of_k_3<\/code>\/<code>compile_theano_1_of_k_3<\/code>) and eickenberg (<code>numpy_1_of_k_2<\/code>\/<code>compile_theano_1_of_k_4<\/code>) for comparison.<\/p>\n\n<p>It turns out the the built in Theano method (<code>compile_theano_1_of_k_2<\/code>) uses roughly the same code as my own attempt (<code>numpy_1_of_k_1<\/code>\/<code>compile_theano_1_of_k_1<\/code>).<\/p>\n\n<pre><code>import timeit\nimport numpy as np\nimport theano\nimport theano.tensor as tt\nimport theano.tensor.extra_ops\n\n\ndef numpy_1_of_k_1(seq, num_classes):\n    num_frames = len(seq)\n    m = np.zeros((num_frames, num_classes))\n    m[np.arange(num_frames), seq] = 1\n    return m\n\n\ndef numpy_1_of_k_2(seq, num_classes):\n    return seq[:, np.newaxis] == np.arange(num_classes)\n\n\ndef numpy_1_of_k_3(seq, num_classes):\n    shape = [seq.shape[i] for i in range(seq.ndim)] + [num_classes]\n    eye = np.eye(num_classes)\n    return eye[seq].reshape(shape)\n\n\ndef compile_theano_1_of_k_1():\n    seq = tt.lvector()\n    num_classes = tt.lscalar()\n    num_frames = seq.shape[0]\n    m = tt.zeros((num_frames, num_classes))\n    m = tt.set_subtensor(m[tt.arange(num_frames), seq], 1)\n    return theano.function([seq, num_classes], outputs=m)\n\n\ndef compile_theano_1_of_k_2():\n    seq = tt.lvector()\n    num_classes = tt.lscalar()\n    return theano.function([seq, num_classes], outputs=theano.tensor.extra_ops.to_one_hot(seq, num_classes))\n\n\ndef compile_theano_1_of_k_3():\n    seq = tt.lvector()\n    num_classes = tt.lscalar()\n    shape = [seq.shape[i] for i in range(seq.ndim)] + [num_classes]\n    eye = tt.eye(num_classes)\n    m = eye[seq].reshape(shape)\n    return theano.function([seq, num_classes], outputs=m)\n\n\ndef compile_theano_1_of_k_4():\n    seq = tt.lvector()\n    num_classes = tt.lscalar()\n    one_hot = tt.eq(seq.reshape((-1, 1)), tt.arange(num_classes))\n    return theano.function([seq, num_classes], outputs=one_hot)\n\n\ndef main(iterations):\n    theano_1_of_k_1 = compile_theano_1_of_k_1()\n    theano_1_of_k_2 = compile_theano_1_of_k_2()\n    theano_1_of_k_3 = compile_theano_1_of_k_3()\n    theano_1_of_k_4 = compile_theano_1_of_k_4()\n\n    test_seq = np.array([0, 1, 2, 0, 1, 2])\n    test_num_classes = 4\n    test_functions = [numpy_1_of_k_1, numpy_1_of_k_2, numpy_1_of_k_3, theano_1_of_k_1, theano_1_of_k_2, theano_1_of_k_3,\n                      theano_1_of_k_4]\n    test_results = [test_function(test_seq, test_num_classes) for test_function in test_functions]\n\n    for a, b in zip(test_results[:-1], test_results[1:]):\n        assert np.all(np.equal(a, b)), (a, b)\n\n    data = []\n    for _ in xrange(iterations):\n        num_classes = np.random.randint(100) + 1\n        seq = np.random.randint(num_classes, size=(np.random.randint(100) + 1))\n        data.append((seq, num_classes))\n\n    for test_function in test_functions:\n        start = timeit.default_timer()\n        total = 0\n        for seq, num_classes in data:\n            total += test_function(seq, num_classes).sum()\n        print timeit.default_timer() - start, total\n\n\nmain(100000)\n<\/code><\/pre>\n\n<p>Using a laptop and running the Theano code on CPU, I get the following timings in seconds:<\/p>\n\n<pre><code>numpy_1_of_k_1    1.0645\nnumpy_1_of_k_2    1.4018\nnumpy_1_of_k_3    1.6131\ntheano_1_of_k_1   6.3542\ntheano_1_of_k_2   6.4628\ntheano_1_of_k_3   6.5637\ntheano_1_of_k_4   5.4588\n<\/code><\/pre>\n\n<p>So in numpy, the identity approach is slower than the simple broadcast which is slower than the set from zeros. In Theano however the relative performance order differs; here the simple broadcast approach is fastest.<\/p>\n\n<p>These are quite small test cases so the relative performances may differ which much larger matrices, or when running on GPU.<\/p>\n"}]}
{"QuestionId":32879868,"AnswerCount":2,"Tags":"<neural-network><caffe><lmdb>","CreationDate":"2015-10-01T05:10:43.643","AcceptedAnswerId":null,"OwnerUserId":639973.0,"Title":"extract features to a text file","Body":"<p>Following the steps in \n<a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/feature_extraction.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/feature_extraction.html<\/a>\nto extract feature from new images after training,<\/p>\n\n<p>you end up getting a data.mdb file.<\/p>\n\n<p>I would much prefer to write features to a txt file so I can easily manipulate it.<\/p>\n\n<p>I did some googling and found some code, but did not work. Furthermore, the generated data.mdb files, when opened using mdb opener apps on mac, do not show any tables at all.<\/p>\n\n<p>Is there a easy way to write extracted features to a text file, or alternatively, easy way to manipulate mdb file so we can check the actual values per image?<\/p>\n","answers":[{"AnswerId":"33113291","CreationDate":"2015-10-13T21:52:43.470","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>The c++ interface for Caffe provides some lmdb functionality to read in data and then write it back out.<\/p>\n\n<p>You could take a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/fc77ef3d23423e57d83e16c32844bbbe8d4d9f0c\/src\/caffe\/layers\/hdf5_data_layer.cpp\" rel=\"nofollow\">look yourself<\/a> at how the LMDBDataLayer performs it's reads and then use that as reference to write a program to manipulate\/write to text the data from an LMDB database.<\/p>\n"},{"AnswerId":"37257159","CreationDate":"2016-05-16T15:06:27.693","ParentId":null,"OwnerUserId":"4033651","Title":null,"Body":"<p>I don\u2019t know if you have already figured out how to solve this problem, but here is how I\u2019ve just found out recently.<\/p>\n\n<p>I am new to Caffe and I find it hard to understand the whole Caffe's architecture. I just wanted a quick method to extract the features from Caffe\u2019s CNN  and to manipulate them later. Moreover, I work on OSX and hadn't install Caffe from source. I installed it through 'port', and the installation seems incomplete. So I ran Caffe's 'feature_extractor' on another machine where Caffe's properly installed and copied the output file to my machine to further process.<\/p>\n\n<p>To this end, you'll have to install LMDB and Google's Protobuf on your machine. You'll have to link the C\/C++ program with liblmdb and libprotobuf.<\/p>\n\n<p>I followed the Caffe\u2019s tutorial to save the output from the layer \u2018fc7\u2019 of AlexNet into a file in format LMDB. Then I wrote a simple C\/C++ program to read it. This can be done using the following code:<\/p>\n\n<pre><code>#include &lt;fstream&gt;\n#include &lt;iostream&gt;\n#include &lt;lmdb.h&gt;\n\nusing namespace std;\n\nint main( int argc, char *argv[] )\n{\n    if( argc!=2 )\n    {\n        cerr&lt;&lt; \"Error\"&lt;&lt; endl\n            &lt;&lt; \"Usage : \"&lt;&lt; argv[0]&lt;&lt; \" mdb_dirname\"&lt;&lt; endl;\n        return 0;\n    }\n    char *mdb_dirname     = argv[1];\n\n    int rc;\n    MDB_env *env;\n    MDB_dbi dbi;\n    MDB_val key, data;\n    MDB_txn *txn;\n    MDB_cursor *cursor;\n    char sval[32];\n\n    rc = mdb_env_create(&amp;env);\n    rc = mdb_env_open(env, mdb_dirname, 0, 0664);\n    rc = mdb_txn_begin(env, NULL, 0, &amp;txn);\n    rc = mdb_open(txn, NULL, 0, &amp;dbi);\n    rc = mdb_cursor_open(txn, dbi, &amp;cursor);\n\n    key.mv_size  = sizeof(int);\n    key.mv_data  = sval;\n    data.mv_size = sizeof(sval);\n    data.mv_data = sval;\n\n    while ((rc = mdb_cursor_get(cursor, &amp;key, &amp;data, MDB_NEXT)) == 0)\n    {\n        printf(\"key: %p %d %.*s, data: %p %d  %.*s\\n\",\n               key.mv_data,  (int) key.mv_size,  (int) key.mv_size,  (char *) key.mv_data,\n               data.mv_data, (int) data.mv_size, (int) data.mv_size, (char *) data.mv_data);\n    }\n    mdb_cursor_close(cursor);\n    mdb_txn_abort(txn);\n\n    mdb_close(env, dbi);\n    mdb_env_close(env);\n\n    return 0;\n}\n<\/code><\/pre>\n\n<p>'mdb_dirname\u2019 is the directory created by 'feature_extractor'. It contains 'data.mdb' and 'lock.mdb'.<\/p>\n\n<p>Noted that I am also new to LMDB. I don\u2019t really understand every lines of code above. However, I does work :)<\/p>\n\n<p>If you process your LMDB file, you may observe that the \u2018key.mv_data\u2019 is indeed the index of your example. So the \u2018data.mv_data\u2019 should contain the feature vector for this example. I looked into the source code of Caffe\u2019s feature_extractor and found that the string \u2018data.mv_data\u2019 is obtained from the serialization of a \u2018Datum\u2019 object. This Datum is indeed constructed using Google\u2019s Protocol Buffer or Protobuf. You may find \u2018caffe.proto\u2019 somewhere in the Caffe directory. This .proto file is processed by the compiler \u2018protoc\u2019 and produce \u2018caffe.pb.h\u2019 and \u2018caffe.pb.cc\u2019 that must be included in your project. If you cannot find it, this is how it looks like<\/p>\n\n<pre><code>syntax = \"proto2\";\n\npackage caffe;\n\nmessage Datum {\n  optional int32 channels = 1;\n  optional int32 height = 2;\n  optional int32 width = 3;\n  \/\/ the actual image data, in bytes\n  optional bytes data = 4;\n  optional int32 label = 5;\n  \/\/ Optionally, the datum could also hold float data.\n  repeated float float_data = 6;\n  \/\/ If true data contains an encoded image that need to be decoded\n  optional bool encoded = 7 [default = false];\n}\n<\/code><\/pre>\n\n<p>Then you may convert \u2018data.mv_data\u2019 into the feature vectors by<\/p>\n\n<pre><code>    while ((rc = mdb_cursor_get(cursor, &amp;key, &amp;data, MDB_NEXT)) == 0)\n    {\n        string str( (char*)data.mv_data, (int)data.mv_size );\n\n        datum.ParseFromString( str ); \n\n        if( datum.float_data_size()&gt;0 )\n        {\n            \/\/ datum.float_data_size() is the dimension of the feature vectors\n            for( int i = 0; i &lt; datum.float_data_size(); i++ )\n            {\n                float f = datum.float_data(i);\n                \/\/ do something\n            }\n        }\n    }\n<\/code><\/pre>\n\n<p>When I build the code above, there're lots of linking errors, such as undefined reference \u2026 and other stuffs related to Protobuf. If you encounter the same problem, a solution that I've found is to link the program against libprotobuf.a instead of simply -llmdb (static linkage instead of dynamic linkage).<\/p>\n\n<p>Another small problem is that the label I assigned to each example in the file processed by 'feature_extractor' is lost. I don't know why. So I've just put these labels into a separate file and process it along the LMDB file. For example, if you want to output LIBSVM file:<\/p>\n\n<pre><code>    int c = 0;\n    while ((rc = mdb_cursor_get(cursor, &amp;key, &amp;data, MDB_NEXT)) == 0)\n    {\n        string str( (char*)data.mv_data, (int)data.mv_size );\n        datum.ParseFromString( str ); \n        if( datum.float_data_size()&gt;0 )\n        {\n            cout&lt;&lt; label[c]&lt;&lt; \" \";\n            for( int i = 0; i &lt; datum.float_data_size(); i++ )\n            {\n                float f = datum.float_data(i);\n                cout&lt;&lt; (i+1)&lt;&lt; \":\"&lt;&lt; f&lt;&lt; \" \";\n            }\n            cout&lt;&lt; endl;\n            c++;\n        }\n    }\n<\/code><\/pre>\n\n<p>Good luck.<\/p>\n"}]}
{"QuestionId":32908025,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><caffe><conv-neural-network>","CreationDate":"2015-10-02T13:14:55.563","AcceptedAnswerId":null,"OwnerUserId":5321075.0,"Title":"Caffe: train, validation and test split","Body":"<p>I've been using caffe for a while, with some success, but I have noticed in examples given that there is only ever a two way split on the data set with  and  phases, where the  set seems to act as a validation set.<\/p>\n\n<p>Ideally I would like to have three sets, so that once the model is trained, I can save it and test it on a completely new test set - stored in a completed separate lmdb folder. Does anyone have any experience of this? Thanks.<\/p>\n","answers":[{"AnswerId":"32909946","CreationDate":"2015-10-02T14:51:12.887","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Differentiating between validation and testing is made to imply that hyperparameters may be tuned to the validation set while nothing is fitted to the test set in any way.\n<code>caffe<\/code> doesn't optimize anything but the weights, and since the test is only there for evaluation, it does exactly as expected.<\/p>\n\n<p>Assuming you're tuning hyper parameters between solver optimization runs. The lmdb passed to <code>caffe<\/code> for testing is really the validation set. If you're done with tuning your hyperparameters and do one more solver optimization with an lmdb for testing that holds data never used in previous runs. That last lmdb is your test set.<\/p>\n\n<p>Since caffe doesn't optimize hyperparameters, its test set is what it is, a test set. It's possible to come up with a some python code around the solver optimization calls that iterates through hyperparameter values. After it's done it can swap in a new lmdb with unseen data to tell you about how well the network generalizes with it.<\/p>\n\n<p>I don't recommend modifying caffe for an explicit val\/test distinction. You don't even have to do anything elaborate with setting up the prototxt file for the solver and network definition. You can do the val\/test swap at the end by simply moving the val lmdb somewhere else and moving the test lmdb in its place using <code>shutil.copy(src, dst)<\/code><\/p>\n"}]}
{"QuestionId":32909986,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><caffe><pycaffe>","CreationDate":"2015-10-02T14:53:24.643","AcceptedAnswerId":null,"OwnerUserId":1657827.0,"Title":"RMSprop, Adam, AdaDelta test accuracy does not improve using Caffe","Body":"<p>I am  using  on an image dataset on a . Using a , , , , , , the  decreases and  goes from  in  iterations which is quite good.<\/p>\n\n<p>When using other optimisers such as ,  and , the  remains almost the same even and no improvement in  after  iterations.<\/p>\n\n<p>For , I have changed the respective parameters as mentioned <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver_rmsprop.prototxt\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>For , I have changed the respective parameters as mentioned <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver_adam.prototxt\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>For , I have changed the respective parameters as mentioned <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_adadelta_solver.prototxt\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>Can someone please tell me what i am doing wrong?<\/p>\n","answers":[{"AnswerId":"36440449","CreationDate":"2016-04-06T02:31:59.547","ParentId":null,"OwnerUserId":"953512","Title":null,"Body":"<p>I saw similar results to pir: Adam would diverge when given the same base_lr that SGD used.  When I reduced base_lr to 1\/100 of its original value, Adam suddenly converged, and gave good results.<\/p>\n"}]}
{"QuestionId":32912149,"AnswerCount":1,"Tags":"<image-processing><neural-network><caffe><conv-neural-network>","CreationDate":"2015-10-02T16:55:47.210","AcceptedAnswerId":32927007.0,"OwnerUserId":5389199.0,"Title":"How to configure Caffe deploy.prototxt?","Body":"<p>I followed @ypx instructions on this <a href=\"https:\/\/stackoverflow.com\/questions\/32846518\/how-to-use-caffe-convnet-library-to-detect-facial-expressions?noredirect=1#comment53572955_32846518\">question<\/a>. Now I want to predict some pictures. So I'm using:<\/p>\n\n<pre><\/pre>\n\n<p>But I get this message:<\/p>\n\n<pre><\/pre>\n\n<p>I think that my problem is on my deploy.prototxt file. <a href=\"https:\/\/bpaste.net\/show\/42874fcfbbaa\" rel=\"nofollow noreferrer\">This is my deploy.prototxt<\/a> and <a href=\"https:\/\/bpaste.net\/show\/b5564e0f8c04\" rel=\"nofollow noreferrer\">This is my train.prototxt<\/a><\/p>\n\n<p>Can someone help me to configure my deploy file?<\/p>\n","answers":[{"AnswerId":"32927007","CreationDate":"2015-10-03T20:10:13.760","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You should remove the training input layer (lines 8--21) from your deploy net.<\/p>\n\n<p>That is, discard this:<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  transform_param {\n    scale: 0.00392156862745\n  }\n  data_param {\n    source: \"\/tmp\/db\"\n    batch_size: 64\n    backend: LMDB\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":32916255,"AnswerCount":2,"Tags":"<python><machine-learning><scikit-learn><theano>","CreationDate":"2015-10-02T21:40:40.753","AcceptedAnswerId":36193124.0,"OwnerUserId":342646.0,"Title":"sklearn: How to reset a Regressor or classifier object in sknn","Body":"<p>I have defined a regressor as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I am using this regressor in a k-fold cross-validation. In order for cross-validation to work properly and not learn from the previous folds, it's necessary that the regressor to be reset after each fold.<br>\nHow can I reset the regressor object?<\/p>\n","answers":[{"AnswerId":"34085940","CreationDate":"2015-12-04T10:16:40.717","ParentId":null,"OwnerUserId":"5638905","Title":null,"Body":"<p>The pattern that I use for cross validation instantiates a new classifier for each training\/test pair:<\/p>\n\n<pre><code>from sklearn.cross_validation import KFold\n\nkf = KFold(len(labels),n_folds=5, shuffle=True)\nfor train, test in kf:\n    clf = YourClassifierClass()\n    clf.fit(data[train],labels[train])\n    # Do evaluation with data[test] and labels[test]\n<\/code><\/pre>\n\n<p>You can save your current best classifier in a separate variable and access its parameters after cross validation (this is also useful if you want to try different parameters).<\/p>\n"},{"AnswerId":"36193124","CreationDate":"2016-03-24T04:07:46.900","ParentId":null,"OwnerUserId":"3014097","Title":null,"Body":"<p><a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.base.clone.html\" rel=\"noreferrer\">sklearn.base.clone<\/a> should achieve what you're looking to achieve<\/p>\n"}]}
{"QuestionId":32931280,"AnswerCount":1,"Tags":"<python><computer-vision><neural-network><caffe><conv-neural-network>","CreationDate":"2015-10-04T07:55:19.160","AcceptedAnswerId":32953308.0,"OwnerUserId":2300002.0,"Title":"Extract image features using Caffe for custom image classifier","Body":"<p>I would like to obtain the output of the 6th layer of a pre-built caffe model and train an SVM on top of it. My intention is to build a custom image classifier, where the user can create custom image classes, and classify input images among those classes, instead of the imagenet classes.Here is the pseudo code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32953308","CreationDate":"2015-10-05T16:23:42.207","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>You should use <code>Net<\/code> class, instead of <code>Classifier<\/code>. Thus, you just need to call <code>net.forward()<\/code>.<\/p>\n\n<p>Two things to pay attention to:<\/p>\n\n<ol>\n<li>Preprocess your input image. See <code>Transformer<\/code> class <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py\" rel=\"noreferrer\">here<\/a> for reference.<\/li>\n<li><p>If you extract the features by using just<\/p>\n\n<pre><code>features = net.blobs['fc6'].data\n<\/code><\/pre>\n\n<p>your array will be overwritten by the next <code>forward()<\/code> call. Be sure you are performing a <strong>deep copy<\/strong>, such as<\/p>\n\n<pre><code>features = net.blobs['fc6'].data.copy()\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":32936166,"AnswerCount":1,"Tags":"<python-2.7><gpu><gpgpu><theano>","CreationDate":"2015-10-04T17:18:53.907","AcceptedAnswerId":null,"OwnerUserId":620625.0,"Title":"How do you free up gpu memory?","Body":"<p>When running theano, I get an error: not enough memory. See below.\nWhat are some possible actions that can be taken to free up memory?\nI know I can close applications etc, but I just want see if anyone has other ideas. For example, is it possible to reserve memory?<\/p>\n\n<blockquote>\n  <p>THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python \n  conv_exp.py Using gpu device 0: GeForce GT 650M Trying to run under a\n  GPU.  If this is not desired, then modify network3.py to set the GPU\n  flag to False. <strong>Error allocating 156800000 bytes of device memory (out\n  of memory). Driver report 64192512 bytes free and 1073414144 bytes\n  total<\/strong>  Traceback (most recent call last):   File \"conv_exp.py\", line\n  25, in \n      training_data, validation_data, test_data = network3.load_data_shared()   File\n  \"\/Users\/xr\/courses\/deep_learning\/con_nn\/neural-networks-and-deep-learning\/src\/network3.py\",\n  line 78, in load_data_shared\n      return [shared(training_data), shared(validation_data), shared(test_data)]   File\n  \"\/Users\/xr\/courses\/deep_learning\/con_nn\/neural-networks-and-deep-learning\/src\/network3.py\",\n  line 74, in shared\n      np.asarray(data[0], dtype=theano.config.floatX), borrow=True)   File\n  \"\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/site-packages\/theano\/compile\/sharedvalue.py\",\n  line 208, in shared\n      allow_downcast=allow_downcast, **kwargs)   File \"\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/site-packages\/theano\/sandbox\/cuda\/var.py\",\n  line 203, in float32_shared_constructor\n      deviceval = type_support_filter(value, type.broadcastable, False, None) MemoryError: ('Error allocating 156800000 bytes of device memory\n  (out of memory).', \"you might consider using 'theano.shared(...,\n  borrow=True)'\")<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"33177976","CreationDate":"2015-10-16T19:11:28.687","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>If borrow is set to true garbage collection is on (default true: <code>config.allow_gc=True<\/code>) and the video card is not currently being used as a display device (doubtful, since you're using a mobile gpu), the only other options are to reduce the parameters of the network or possibly the batch size of the model.  The latter will be especially effective if the model uses dropout or noise-based masks (these will be equal to the number of examples in the batch x the number of parameters dropped out or noised).  <\/p>\n\n<p>Otherwise maybe you could boot to the command prompt to save a few mbs?\n:\/<\/p>\n"}]}
{"QuestionId":32939413,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-10-04T23:21:44.483","AcceptedAnswerId":32939467.0,"OwnerUserId":2002890.0,"Title":"Theano fails to use GPU when running Python as module (python -m)","Body":"<p>I built a Python package so that I can do relative imports like:<\/p>\n\n<pre><\/pre>\n\n<p>from<\/p>\n\n<pre><\/pre>\n\n<p>if my file structure is something like:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I run  I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>This error does not appear when I just run<\/p>\n\n<p><\/p>\n\n<p>and I'm able to use the GPU just fine in that case. I'm wondering if there's some Theano config setting that's begin altered by running python as a module?<\/p>\n","answers":[{"AnswerId":"32939467","CreationDate":"2015-10-04T23:31:47.407","ParentId":null,"OwnerUserId":"2334254","Title":null,"Body":"<p>How did you import the files from outside the directory of the top file? If you aren\u2019t using <\/p>\n\n<pre><code>import sys\nsys.path.insert(0, '.\/path\/to\/some\/place\/with\/files\/for\/importing1')\nsys.path.insert(0, '.\/path\/to\/some\/place\/with\/files\/for\/importing2')\nimport thing1\nimport thing2\n<\/code><\/pre>\n\n<p>I recommend that you do so.<\/p>\n"}]}
{"QuestionId":32939762,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><torch>","CreationDate":"2015-10-05T00:24:05.707","AcceptedAnswerId":null,"OwnerUserId":2173168.0,"Title":"How do Torch 7 load very large datasets that do not fit in memory?","Body":"<p>I have some very large datasets which I want to load into Torch. But I cannot load them fully to the memory before I run my learning algorithm. <\/p>\n\n<p>What should I do to maximize performance? <\/p>\n","answers":[{"AnswerId":"32943919","CreationDate":"2015-10-05T08:10:11.973","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Have a look at <a href=\"https:\/\/github.com\/soumith\/imagenet-multiGPU.torch\">imagenet-multiGPU.torch<\/a> full-stack sample code. It contains a data loader (<code>dataset.lua<\/code>) able to sample a batch of images at a time which prevents from pre-loading everything in memory:<\/p>\n\n<pre><code>local inputs, labels = trainLoader:sample(opt.batchSize)\n<\/code><\/pre>\n\n<p>(see <a href=\"https:\/\/github.com\/soumith\/imagenet-multiGPU.torch\/blob\/64872bf5d5c86ce0c7beb3fd7e148b6f19959437\/train.lua#L94-L105\"><code>train.lua<\/code><\/a> for more details)<\/p>\n"}]}
{"QuestionId":32952877,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><deep-learning>","CreationDate":"2015-10-05T16:00:53.760","AcceptedAnswerId":32964189.0,"OwnerUserId":492372.0,"Title":"Why does this method throw an error in terms of number of arguments?","Body":"<p>I have the following piece of code and when I run the <strong>theano_build<\/strong>() method, it throws as error saying<\/p>\n\n<pre><\/pre>\n\n<p>the following is the code in Theano. It is basically a two hidden layered recurrent neural network<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32964189","CreationDate":"2015-10-06T07:31:28.253","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Try changing<\/p>\n\n<pre><code>return [o_t[0], s_t1, s_t2]\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>return o_t[0], s_t1, s_t2\n<\/code><\/pre>\n\n<p>I think the former is causing the method to return something that is coerced by Theano into a single tensor while the latter is explicitly returningthree objects as indicated in <code>outputs_info<\/code>.<\/p>\n"}]}
{"QuestionId":32956598,"AnswerCount":2,"Tags":"<neural-network><caffe>","CreationDate":"2015-10-05T19:40:58.087","AcceptedAnswerId":32972012.0,"OwnerUserId":4975126.0,"Title":"Navigating hyper-parameters","Body":"<p>I was just wondering if someone could provide a good source for me to read on how I should approach choosing hyper-parameters of the solver based on the complexity of my problem.<\/p>\n\n<p>Basically, I understand that many feel that they are \"shooting around in the dark\" when it comes to setting and then modifying these parameters and a system or benchmark for choosing parameters based on specific problem\/data complexity has escaped me.<\/p>\n\n<p>If you care to explain your own methodology or simply provide commentary on your source, it would be much appreciated.<\/p>\n","answers":[{"AnswerId":"32972012","CreationDate":"2015-10-06T14:02:34.740","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>Since the hyperparameters we're talking about are related to <strong>backpropagation<\/strong>, which is a gradient-based approach, I believe the main reference is <a href=\"http:\/\/arxiv.org\/abs\/1206.5533\" rel=\"nofollow\">Y. Bengio<\/a>, along with the more classic <a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/lecun-98b.pdf\" rel=\"nofollow\">Lecun et al.<\/a>.<\/p>\n\n<p>There are three main approaches to find out the optimal value for an hyperparameter. The first two are well explained in the first paper I linked.<\/p>\n\n<ul>\n<li>Manual search. The researcher choose the optimal value through try-and-error.<\/li>\n<li>Automatic search. The researcher relies on an automated routine in order to speed up the search.<\/li>\n<li>Bayesian Optimization. You can find a video presenting it <a href=\"https:\/\/www.youtube.com\/watch?v=VG2uCpKJkSg\" rel=\"nofollow\">here<\/a>.<\/li>\n<\/ul>\n"},{"AnswerId":"32966048","CreationDate":"2015-10-06T09:13:06.323","ParentId":null,"OwnerUserId":"5299473","Title":null,"Body":"<p>I think this is the main reference:<\/p>\n\n<p><a href=\"http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow\">http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf<\/a><\/p>\n\n<p>Also take a look at Chapter 5 in:\n<a href=\"http:\/\/neuralnetworksanddeeplearning.com\/\" rel=\"nofollow\">http:\/\/neuralnetworksanddeeplearning.com\/<\/a><\/p>\n"}]}
{"QuestionId":32966517,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-10-06T09:34:21.277","AcceptedAnswerId":32967829.0,"OwnerUserId":2515628.0,"Title":"Multiple accuracy layers in Caffe","Body":"<p>I'm trying to classify a large set of images using nVidia DIGITS and Caffe. Everything works well when I use standard networks and networks I've constructed.<\/p>\n\n<p>However, when I run GoogleNet example, I can see several accuracy layers' results. How can there be multiple accuracy layers in a CNN? Having multiple loss layers is quite understandable, but what do multiple accuracy values mean? I get several accuracy graphs during learning. Similar to this picture: <a href=\"https:\/\/i.stack.imgur.com\/6jV7c.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6jV7c.png\" alt=\"learning process\"><\/a><\/p>\n\n<p>The  and  denote accuracy layers. I understand from the <em>prototxt<\/em> that these evaluate top 1 and top 5 accuracy values, but what are the lossX accuracy layers?<\/p>\n\n<p>Even though some of these graphs converge to around 98%, when I manually test the trained network on the , I get significantly lower value (ones corresponding to the lower three accuracy graphs).<\/p>\n\n<p>Can someone shed some light on this? How can there be multiple accuracy layers with different values?<\/p>\n","answers":[{"AnswerId":"32967829","CreationDate":"2015-10-06T10:36:35.510","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you look closely at the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/train_val.prototxt\" rel=\"nofollow\"><code>'train_val.prototxt'<\/code><\/a> you will notice that there are indeed several accuracy layers branching off the main \"path\" at different levels. <code>loss1<\/code> is evaluated after <code>inception 4a<\/code> layer, <code>loss2<\/code> is evaluated after <code>inception 4d<\/code> and <code>loss3<\/code> is the loss at the top of the net. Introducing loss (and accuracy layers) to intermediate representations of the deep net allows for faster gradient propagation during training. These intermediate accuracies measure how well the intermediate representations have converged.   <\/p>\n"}]}
{"QuestionId":32966627,"AnswerCount":1,"Tags":"<python><linear-algebra><matrix-multiplication><theano><blas>","CreationDate":"2015-10-06T09:38:41.010","AcceptedAnswerId":32972148.0,"OwnerUserId":1990516.0,"Title":"Matrix triple product with theano","Body":"<p>This is pretty much the same question as here <a href=\"https:\/\/stackoverflow.com\/questions\/30206293\/matrix-tensor-triple-product\">Matrix\/Tensor Triple Product?<\/a> , but for theano.<\/p>\n\n<p>So I have three matrices , ,  of sizes , , , and I want to compute the 3D tensor of shape  resulting from the triple (trilinear) product:<\/p>\n\n<pre><\/pre>\n\n<p>,  and  are shared variables:<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to write it with a single theano expression, is there a way to do so?\nIf there are many, which one is the fastest?<\/p>\n","answers":[{"AnswerId":"32972148","CreationDate":"2015-10-06T14:09:37.857","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<pre><code>np.einsum('nr,mr,lr-&gt;nml', A, B, C)\n<\/code><\/pre>\n\n<p>is equivalent to<\/p>\n\n<pre><code>np.dot(A[:, None, :] * B[None, :, :], C.T)\n<\/code><\/pre>\n\n<p>which can be implemented in Theano as<\/p>\n\n<pre><code>theano.dot(A[:, None, :] * B[None, :, :], C.T)\n<\/code><\/pre>\n"}]}
{"QuestionId":32966970,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><caffe>","CreationDate":"2015-10-06T09:54:44.823","AcceptedAnswerId":33013840.0,"OwnerUserId":1868436.0,"Title":"Deep learning Training dataset with Caffe","Body":"<p>I am a deep-learning newbie and working on creating a vehicle classifier for images using Caffe  and have a 3-part question:<\/p>\n\n<ol>\n<li><p>Are there any best practices in organizing classes for training a\nCNN? i.e. number of classes and number of samples for each class?\nFor example, would I be better off this way:<\/p>\n\n<ul>\n<li>(a) Vehicles - Car-Sedans\/Car-Hatchback\/Car-SUV\/Truck-18-wheeler\/....    (note this could mean several thousand classes), or  <\/li>\n<li>(b) have a    higher level\nmodel that classifies between car\/truck\/2-wheeler and so    on...\nand if car type then query the Car Model to get the car type<br>\n(sedan\/hatchback etc)<\/li>\n<\/ul><\/li>\n<li><p>How many training images per class is a typical best practice? I know there are several other variables that affect the accuracy of\nthe CNN, but what rough number is good to shoot for in each class?\nShould it be a function of the number of classes in the model? For\nexample, if I have many classes in my model, should I provide more\nsamples per class?<\/p><\/li>\n<li><p>How do we ensure we are not overfitting to class? Is there way to measure heterogeneity in training samples for a class?<\/p><\/li>\n<\/ol>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"33013840","CreationDate":"2015-10-08T10:57:18.863","ParentId":null,"OwnerUserId":"2832033","Title":null,"Body":"<ol>\n<li><p>Well, the first choice that you mentioned corresponds to a very challenging task in computer vision community: fine-grained image classification, where you want to classify the subordinates of a base class, say Car! To get more info on <a href=\"http:\/\/arxiv.org\/pdf\/1411.6447\" rel=\"nofollow\">this<\/a>, you may see this paper. \nAccording to the literature on image classification, classifying the high-level classes such as car\/trucks would be much simpler for CNNs to learn since there may exist more discriminative features. I suggest to follow the second approach, that is classifying all types of cars vs. truck and so on.<\/p><\/li>\n<li><p>Number of training samples is mainly proportional to the number of parameters, that is if you want to train a shallow model, much less samples are required. That also depends on your decision to fine-tune a pre-trained model or train a network from scratch. When sufficient samples are not available, you have to fine-tune a model on your task. <\/p><\/li>\n<li><p>Wrestling with over-fitting has been always a problematic issue in machine learning and even CNNs are not free of them. Within the literature, some practical suggestions have been introduced to reduce the occurrence of over-fitting such as dropout layers and data-augmentation procedures.<\/p><\/li>\n<\/ol>\n\n<p>May not included in your questions, but it seems that you should follow the fine-tuning procedure, that is initializing the network with pre-computed weights of a model on another task (say ILSVRC 201X) and adapt the weights according to your new task. This procedure is known as transfer learning (and sometimes domain adaptation) in community.<\/p>\n"}]}
{"QuestionId":32975809,"AnswerCount":1,"Tags":"<python><python-2.7><theano>","CreationDate":"2015-10-06T17:17:08.517","AcceptedAnswerId":32985080.0,"OwnerUserId":5415099.0,"Title":"Theano With Python2.7: SGD with multiple losses","Body":"<p>After Theano being praised to me I thought I'd make my first steps with a specific form of SGD. I've got a parameter vector theta which I want to optimise My loss function return a vector containing the column sums of the squared loss between a matrix A and B. Each of the elements is an independent loss for a specific dimension using the broadcasted theta. Theta should be updated so that the next iteration the loss for each dimension is lower. I chose this because the data (X,Y) is given in this way.<\/p>\n\n<p>Now the tutorial says T.grad() should be used to get the gradient for the updates. But T.grad doesn't allow me to calculate the gradient for non-scalars. The tutorial (<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html<\/a>) says 'Scalar costs only can be directly handled by grad. Arrays are handled through repeated applications.' So I tried (admitably an ugly attempt) to calculate the gradient for each of the losses. How do I calculate gradients for a multiple losses? And is there a clean, best-practice way? Is that even correct? Something else that I should consider?<\/p>\n\n<p>Martin<\/p>\n\n<pre><\/pre>\n\n<p>This is where it is getting weird:\nBecause T.grad(loss, theta) throws TypeError: cost must be a scalar. So I got this ugly attempt:<\/p>\n\n<pre><\/pre>\n\n<p>When I want to compile it I get this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32985080","CreationDate":"2015-10-07T06:20:27.277","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>As Mikael Rousson points out in a comment, for the purposes of the gradient you probably don't need to deal with separate losses; just sum up all the loss components into a scalar and then compute the partial derivatives with respect to the vector of parameters, yielding a vector of gradients.<\/p>\n\n<p>So add<\/p>\n\n<pre><code>loss = losses.sum()\n<\/code><\/pre>\n\n<p>or directly define a scalar loss<\/p>\n\n<pre><code>loss = T.sqr(theta * X - Y).sum()\n<\/code><\/pre>\n\n<p>then use<\/p>\n\n<pre><code>d_losses = T.grad(loss, theta)\nupdates = [(theta, theta - alpha * d_losses)]\n<\/code><\/pre>\n\n<p><code>d_losses[0]<\/code> equals the partial derivative of <code>loss<\/code> with respect to <code>theta[0]<\/code> but the only term in <code>loss<\/code> that involves <code>theta[0]<\/code> is the component of the sum over the first element of <code>losses<\/code> so it is also equal to the partial derivative of <code>losses[0]<\/code> with respect to <code>theta[0]<\/code> which is exactly what you want,  I think.<\/p>\n"}]}
{"QuestionId":32976868,"AnswerCount":2,"Tags":"<python><machine-learning><theano>","CreationDate":"2015-10-06T18:17:53.350","AcceptedAnswerId":null,"OwnerUserId":5016984.0,"Title":"how does theano.scan's updates work?","Body":"<p> return two variables: values variable and updates variable.  For example,<\/p>\n\n<pre><\/pre>\n\n<p>However, I notice that in most of the examples I work with, the updates variable is empty.  It seems only when we write the function in  is a certain way, we get the updates.  For example,<\/p>\n\n<pre><\/pre>\n\n<p>Can someone explain to me why in the first example the updates is empty, but in the second example, the updates variable is not empty?  and more generally, how does the updates variable in  work?  Thanks.<\/p>\n","answers":[{"AnswerId":"32984852","CreationDate":"2015-10-07T06:04:12.343","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Consider the following four variations (this code can be executed to observe the differences) and analysis below.<\/p>\n\n<pre><code>import theano\n\n\ndef v1a():\n    a = theano.shared(1)\n    outputs, updates = theano.scan(lambda x: x + 1, outputs_info=a, n_steps=10)\n    f = theano.function([], outputs=outputs)\n    print f(), a.get_value()\n\n\ndef v1b():\n    a = theano.shared(1)\n    outputs, updates = theano.scan(lambda x: x + 1, outputs_info=a, n_steps=10)\n    f = theano.function([], outputs=outputs, updates=updates)\n    print f(), a.get_value()\n\n\ndef v2a():\n    a = theano.shared(1)\n    outputs, updates = theano.scan(lambda: {a: a + 1}, n_steps=10)\n    f = theano.function([], outputs=outputs)\n    print f(), a.get_value()\n\n\ndef v2b():\n    a = theano.shared(1)\n    outputs, updates = theano.scan(lambda: {a: a + 1}, n_steps=10)\n    f = theano.function([], outputs=outputs, updates=updates)\n    print f(), a.get_value()\n\n\ndef main():\n    v1a()\n    v1b()\n    v2a()\n    v2b()\n\n\nmain()\n<\/code><\/pre>\n\n<p>The output of this code is<\/p>\n\n<pre><code>[ 2  3  4  5  6  7  8  9 10 11] 1\n[ 2  3  4  5  6  7  8  9 10 11] 1\n[] 1\n[] 11\n<\/code><\/pre>\n\n<p>The <code>v1x<\/code> variations use <code>lambda x: x + 1<\/code>. the result of the lambda function is a symbolic variable whose value is 1 greater than the input. The name of the lambda function's parameter has been changed to avoid shadowing the shared variable name. In these variations the shared variable is not used or manipulated in any way by the scan, other than using it as the initial value of the recurrent symbolic variable incremented by the scan step function.<\/p>\n\n<p>The <code>v2x<\/code> variations use <code>lambda {a: a + 1}<\/code>. The result of the lambda function is a dictionary that explains how to update the shared variable <code>a<\/code>.<\/p>\n\n<p>The <code>updates<\/code> from the <code>v1x<\/code> variations is empty because we have not returned a dictionary from the step function defining any shared variable updates. The <code>outputs<\/code> from the <code>v2x<\/code> variations is empty because we have not provided any symbolic output from the step function. <code>updates<\/code> only has use if the step function returns a shared variable update expression dictionary (as in <code>v2x<\/code>) and <code>outputs<\/code> only has use if the step function returns a symbolic variable output (as in <code>v1x<\/code>).<\/p>\n\n<p>When a dictionary is returned, it will have no effect if not provided to <code>theano.function<\/code>. Note that the shared variable has not been updated in <code>v2a<\/code> but it has been updated in <code>v2b<\/code>.<\/p>\n"},{"AnswerId":"37460329","CreationDate":"2016-05-26T12:02:52.667","ParentId":null,"OwnerUserId":"6220857","Title":null,"Body":"<p>To complement Daniel's answer, if you want to compute outputs and updates in theano scan at the same time, look at this example. <\/p>\n\n<p>This code loops over a sequence, computing the sum of its elements and updates a shared variable <code>t<\/code> (length of the sentence)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import theano\nimport numpy as np\n\nt = theano.shared(0)\ns = theano.tensor.vector('v')\n\ndef rec(s, first, t):\n    first = s + first\n    second = s\n    return (first, second), {t: t+1}\n\nfirst = np.float32(0)\n\n(firsts, seconds), updates = theano.scan(\n    fn=rec,\n    sequences=s,\n    outputs_info=[first, None],\n    non_sequences=t)\n\nf = theano.function([s], [firsts, seconds], updates=updates, allow_input_downcast=True)\n\nv = np.arange(10)\n\nprint f(v)\nprint t.get_value()\n<\/code><\/pre>\n\n<p>The output of this code is <\/p>\n\n<pre><code>[array([  0.,   1.,   3.,   6.,  10.,  15.,  21.,  28.,  36.,  45.], dtype=float32), \narray([ 0.,  1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9.], dtype=float32)]\n10\n<\/code><\/pre>\n\n<p><code>rec<\/code> function outputs a tuple and a dictionary. Scanning over a sequence will both compute the outputs and add the dictionary to the updates, allowing you to create a function updating <code>t<\/code>and computing <code>firsts<\/code> and <code>seconds<\/code> at the same time.<\/p>\n"}]}
{"QuestionId":32979683,"AnswerCount":1,"Tags":"<machine-learning><neural-network><convolution><deep-learning><caffe>","CreationDate":"2015-10-06T21:02:05.710","AcceptedAnswerId":32984370.0,"OwnerUserId":4882455.0,"Title":"How did they calculate the output volume for this convnet example in Caffe?","Body":"<p>In this <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">tutorial<\/a>, the output volumes are stated in output [25], and the receptive fields are specified in output [26]. <\/p>\n\n<p>Okay, the input volume  gets convolved with the region of size . <\/p>\n\n<p>Using this <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow\">formula<\/a> <strong><\/strong>, where:<br>\n<em> = the input volume size<\/em><br>\n<em> = the receptive field size<\/em><br>\n<em> = padding<\/em><br>\n<em> = stride<\/em><\/p>\n\n<p>...results with   i.e. <strong>[55*55*96]<\/strong>. So far so good :)<\/p>\n\n<p>For 'pool1' they used and  I think? The calculation checks out: .<\/p>\n\n<p>From this point I get a bit confused. The receptive field for the second convnet layer is , yet the output for 'conv2' is equal to . What calculation happened here?<\/p>\n\n<p>And then, the height and width of the output volumes of 'conv3' to 'conv4' are all the same ? What's going on? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"32984370","CreationDate":"2015-10-07T05:21:21.180","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you look closely at the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt#L55\" rel=\"nofollow\">parameters of <code>conv2<\/code> layer<\/a> you'll notice <\/p>\n\n<pre><code>   pad: 2\n<\/code><\/pre>\n\n<p>That is, the input blob is padded by 2 extra pixels all around, thus the formula now is <\/p>\n\n<pre><code>27 + 2 + 2 - ( 5 - 1 ) = 27\n<\/code><\/pre>\n\n<p>Padding a kernel size of <code>5<\/code> with <code>2<\/code> pixels from both sides yields the same output size.<\/p>\n"}]}
{"QuestionId":32989064,"AnswerCount":1,"Tags":"<3d><max><pool><theano><pooling>","CreationDate":"2015-10-07T09:48:36.333","AcceptedAnswerId":35720432.0,"OwnerUserId":5115006.0,"Title":"Theano max_pool_3d","Body":"<p>How do I extend theanos downsample.max_pool_2d_same_size in order to pool not only within a feature map, but also between those - in a efficient manner?<\/p>\n\n<p>Lets say i got 3 feature maps, each of size 10x10, that would be a 4D Tensor (1,3,10,10). First lets max pool ((2,2), no overlapping) each of the (10,10) feature map. The results are 3 sparse feature maps, still (10,10) but most values equal to zero: within a (2,2) window is at most one value greater than zero. This is what downsample.max_pool_2d_same_size does.<\/p>\n\n<p>Next, i want to compare every maximum of a certain (2,2) window to all other maxima of all other feature maps of the window at the same position.\nI want to keep only the maxima across all of the feature maps. The results are again 3 feature maps (10,10), with nearly all of the values being zero. <\/p>\n\n<p>Is there a fast way of doing so?\nI wouldn't mind other max_pooling functions, but i need the exact locations of the maxima for pooling\/unpooling purposes (but that's another topic).<\/p>\n","answers":[{"AnswerId":"35720432","CreationDate":"2016-03-01T10:28:08.007","ParentId":null,"OwnerUserId":"5115006","Title":null,"Body":"<p>I solved it using lasagne with cudnn. Here are some minimal examples of how to get the indices of a max pooling operation (2d and 3d). See <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/lasagne-users\/BhtKsRmFei4\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/lasagne-users\/BhtKsRmFei4<\/a><\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\nfrom theano.tensor.type import TensorType\nfrom theano.configparser import config\nimport lasagne\n\ndef tensor5(name=None, dtype=None):\n    if dtype is None:\n        dtype = config.floatX\n    type = TensorType(dtype, (False, False, False, False, False))\n    return type(name)\n\ndef max_pooling_2d():\n    input_var = T.tensor4('input')\n    input_layer = lasagne.layers.InputLayer(shape=(None, 2, 4, 4), input_var=input_var)\n    max_pool_layer = lasagne.layers.MaxPool2DLayer(input_layer, pool_size=(2, 2))\n\n    pool_in, pool_out = lasagne.layers.get_output([input_layer, max_pool_layer])\n    indices = T.grad(None, wrt=pool_in, known_grads={pool_out: T.ones_like(pool_out)})\n    get_indices_fn = theano.function([input_var], indices,allow_input_downcast=True)\n\n    data = np.random.randint(low=0, high=9, size=32).reshape((1,2,4,4))\n    indices = get_indices_fn(data)\n    print data, \"\\n\\n\", indices\n\ndef max_pooling_3d():\n    input_var = tensor5('input')\n    input_layer = lasagne.layers.InputLayer(shape=(1, 1, 2, 4, 4), input_var=input_var)\n    # 5 input dimensions: (batchsize, channels, 3 spatial dimensions)\n    max_pool_layer = lasagne.layers.dnn.MaxPool3DDNNLayer(input_layer, pool_size=(2, 2, 2))\n\n    pool_in, pool_out = lasagne.layers.get_output([input_layer, max_pool_layer])\n    indices = T.grad(None, wrt=pool_in, known_grads={pool_out: T.ones_like(pool_out)})\n    get_indices_fn = theano.function([input_var], indices,allow_input_downcast=True)\n\n    data = np.random.randint(low=0, high=9, size=32).reshape((1,1,2,4,4))\n    indices = get_indices_fn(data)\n    print data, \"\\n\\n\", indices\n<\/code><\/pre>\n"}]}
{"QuestionId":32990353,"AnswerCount":0,"Tags":"<python-3.x><scikit-learn><theano>","CreationDate":"2015-10-07T10:53:44.810","AcceptedAnswerId":null,"OwnerUserId":853462.0,"Title":"Training procedure","Body":"<p>I noticed that the  implements gradient descent for a <em>linear model<\/em>, therefore one could state that the class combines the fitting procedure (SGD) and the model (linear model) in one class.<\/p>\n\n<p>SGD is, however, not inherent to linear models, and linear_models can be trained using many other optimizations with particular pro's (memory usage, convergence speed, local optima avoidance*, ...). One could state that such optimization techniques implement how to iterate over the training data, whether to do it online or offline, in what feature-dimension to apply an update and when to stop (possibly based on an error function of a validation set).<\/p>\n\n<p>In particular, I implemented a model using theano and wrapped it in a fit\/predict interface. Theano is cool because it allows one to define a callable which applies gradient descent on one sample, or on a set of samples, as well as a callable which returns the error on a validation set. But this coolness is not inherent to theano, a lot more models can simply define an update and error-evaluation function, which can then be used by different iteration- and stopping policies for fitting.<\/p>\n\n<p>The theano examples often use minibatch, and the minibatch code is copy-pasted or reimplemented a lot with just minor adjustments which can easily be factored out. So I was hoping that sklearn implements something that you initialize with some parameters and an update\/error callable to fit 'any model'. Or possible there is some good practice on how to do this yourself (especially w.r.t. the interface of the fitter).<\/p>\n\n<p>Is there anything like this (in sklearn), that is Fitters which do not define the model?<\/p>\n\n<p>*In the particular case of linear models and a l2 cost function, local optima do not exist of course, but still.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Fair enough, this calls for a suggestion. I coded these two classes, which are not 100% clean, but they given an idea of what I mean:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Then for in the fit of the model one could support different training approaches and stopping criteria like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Obviosuly the hash-map part is hacky, and could be done more elegantly using a standardized interface for the two classes above (since the hashmaps are still O(N*M) in size for N fitters and M models).<\/p>\n","answers":[]}
{"QuestionId":32995102,"AnswerCount":3,"Tags":"<shell><lua><torch>","CreationDate":"2015-10-07T14:34:56.000","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch, how to execute a script with \"dofile\" and input parameters?","Body":"<p>I'm executing a Torch script from my Linux shell, with the  command. This Torch script takes two input parameters:<\/p>\n\n<p><\/p>\n\n<p>Now I'd like to run this script through the Torch shell. To do this, I have to use the  command. But in this case, I don't know how to pass the input parameters  and .<\/p>\n\n<p>In Torch, how to pass some input parameters to the  execution command?<\/p>\n\n<hr>\n\n<p><strong>EDIT<\/strong>: Here's the code that I'm trying to run. I'm not able to run it properly, maybe you can tell me why<\/p>\n\n<p>external_command.lua content:<\/p>\n\n<pre><\/pre>\n\n<p>On the shell:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>EDIT 2<\/strong>: I tried the solution posted by TonyHsu, but it does not work anyway.\nHere's what I'm doing.<\/p>\n\n<p>I define a function  in a script called  which contains the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Then I use the  script that I previously defined as  input parameter for this function:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, also in this case I get an error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"32995278","CreationDate":"2015-10-07T14:42:00.523","ParentId":null,"OwnerUserId":"1190388","Title":null,"Body":"<p>You use <code>loadfile<\/code>:<\/p>\n\n<pre><code>local TempFunc = loadfile \"torch_script.lua\"\nTempFunc(input_parameter1, input_parameter2)\n<\/code><\/pre>\n"},{"AnswerId":"33054761","CreationDate":"2015-10-10T13:58:00.200","ParentId":null,"OwnerUserId":"4677158","Title":null,"Body":"<p>Maybe you can try to define a function first :<\/p>\n\n<pre><code>function runfile(&lt;scriptName&gt;, ...)\nopt = nil\narg = {...}\ndofile(&lt;scriptName&gt;)\nend\n<\/code><\/pre>\n\n<p>then  run <\/p>\n\n<pre><code>runfile(&lt;scriptName&gt;,...)\n<\/code><\/pre>\n"},{"AnswerId":"42470928","CreationDate":"2017-02-26T16:16:06.330","ParentId":null,"OwnerUserId":"7596504","Title":null,"Body":"<p>I think the trick is to pass the parameters in global variable 'args'. Say, I have the following content in <code>external_command.lua<\/code>.<\/p>\n\n<pre><code>-- args has been set by the caller\nif not args or #args == 0 then\n    print('no input_parameter')\nelse\n    print('#args = ' .. #args, 'input_parameter: ' .. args[1])\nend\n<\/code><\/pre>\n\n<p>Then define runfile() as follows.<\/p>\n\n<pre><code>function runfile(f, ...)\n    local tmp = args  -- save the original global args\n    args = table.pack(...)\n    dofile(f)\n    args = tmp  -- restore args\nend\n<\/code><\/pre>\n\n<p>I've tested it in th. It looks like this.<\/p>\n\n<pre><code>th&gt; runfile('ext_command.lua', 10)\n#args = 1       input_parameter: 10\n                                                                      [0.0002s]\nth&gt; runfile('ext_command.lua', 'a', 'b', 'c')\n#args = 3       input_parameter: a\n                                                                      [0.0002s]\n<\/code><\/pre>\n"}]}
{"QuestionId":32995147,"AnswerCount":1,"Tags":"<python-2.7><theano>","CreationDate":"2015-10-07T14:36:44.353","AcceptedAnswerId":null,"OwnerUserId":648896.0,"Title":"Finding argmax indices along a dimension and negate the corresponding values if the index is 1","Body":"<p>Given a tensor in size DxCx2xHxW I try to find the argmax indices for the third dimension (between two values) and if the index is 0 I'll keep it or if it is 1 I'll negate the value and reduce the tensor to size DxCxHxW with the modified values and ignoring the non argmax indices.<\/p>\n\n<p>Could you suggest a way to do this correctly? Thanks in advance for any further comment...<\/p>\n\n<p>Here also my first stupid attempt that might give some insights about what am I trying to do.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33010156","CreationDate":"2015-10-08T08:11:26.807","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>For the potential benefit of other people, this question was answered on the theano-users mailing list: <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/6P5KxLph2I8\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/6P5KxLph2I8<\/a><\/p>\n\n<p>The recommended solution, by Pierre Luc Carrier, was:<\/p>\n\n<pre><code>mask = input_reshaped.argmax(axis=3).astype(\"float32\") * -2 + 1\nreturn input_reshaped.max(axis=3) * mask\n<\/code><\/pre>\n"}]}
{"QuestionId":32996281,"AnswerCount":5,"Tags":"<lua><torch>","CreationDate":"2015-10-07T15:25:08.147","AcceptedAnswerId":33008837.0,"OwnerUserId":1082019.0,"Title":"How to check if two Torch tensors or matrices are equal?","Body":"<p>I need a Torch command that checks if two tensors have the same content, and returns TRUE if they have the same content.<\/p>\n\n<p>For example:<\/p>\n\n<pre><\/pre>\n\n<p>What should I use in this script instead of  ?<\/p>\n\n<p>I tried simply with  but it does not work. <\/p>\n","answers":[{"AnswerId":"59059453","CreationDate":"2019-11-26T21:17:22.153","ParentId":null,"OwnerUserId":"3103059","Title":null,"Body":"<p>You can <strong>convert<\/strong> the two tensors to numpy arrays:<\/p>\n\n<pre><code>local tens_a = torch.Tensor((9,8,7,6));\nlocal tens_b = torch.Tensor((9,8,7,6));\n\na=tens_a.numpy()\nb=tens_b.numpy()\n<\/code><\/pre>\n\n<p>and then something like<\/p>\n\n<pre><code>np.sum(a==b)\n4\n<\/code><\/pre>\n\n<p>would give you a fairly good idea of how equals are they.<\/p>\n"},{"AnswerId":"33008837","CreationDate":"2015-10-08T07:04:24.700","ParentId":null,"OwnerUserId":"5118777","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torcheqa-b\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torcheqa-b<\/a><\/p>\n\n<pre><code>torch.eq(a, b)\n<\/code><\/pre>\n\n<p>Implements == operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.<\/p>\n\n<p>--UPDATE<\/p>\n\n<p>from @deltheil<\/p>\n\n<pre><code>torch.all(torch.eq(tens_a, tens_b))\n<\/code><\/pre>\n\n<p>or even simpler<\/p>\n\n<pre><code>torch.all(tens_a:eq(tens_b))\n<\/code><\/pre>\n"},{"AnswerId":"57772248","CreationDate":"2019-09-03T12:57:10.543","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>To compare tensors you can do element wise:<\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=eq#torch.eq\" rel=\"nofollow noreferrer\"><code>torch.eq<\/code><\/a> is element wise:<\/p>\n\n<pre><code>torch.eq(torch.tensor([[1., 2.], [3., 4.]]), torch.tensor([[1., 1.], [4., 4.]]))\ntensor([[True, False], [False, True]])\n<\/code><\/pre>\n\n<p>Or <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=eq#torch.equal\" rel=\"nofollow noreferrer\"><code>torch.equal<\/code><\/a> for the whole tensor exactly:<\/p>\n\n<pre><code>torch.equal(torch.tensor([[1., 2.], [3, 4.]]), torch.tensor([[1., 1.], [4., 4.]]))\n# False\ntorch.equal(torch.tensor([[1., 2.], [3., 4.]]), torch.tensor([[1., 2.], [3., 4.]]))\n# True\n<\/code><\/pre>\n\n<p>But then you may be lost because at some point there are small differences you would like to ignore. For instance floats <code>1.0<\/code> and <code>1.0000000001<\/code> are pretty close and you may consider these are equal. For that kind of comparison you have <a href=\"https:\/\/pytorch.org\/docs\/master\/torch.html?highlight=allclose#torch.allclose\" rel=\"nofollow noreferrer\"><code>torch.allclose<\/code><\/a>.<\/p>\n\n<pre><code>torch.allclose(torch.tensor([[1., 2.], [3., 4.]]), torch.tensor([[1., 2.000000001], [3., 4.]]))\n# True\n<\/code><\/pre>\n\n<p>At some point may be important to check element wise how many elements are equal, comparing to the full number of elements. If you have two tensors <code>dt1<\/code> and <code>dt2<\/code> you get number of elements of <code>dt1<\/code> as <code>dt1.nelement()<\/code><\/p>\n\n<p>And with this formula you get the percentage:<\/p>\n\n<pre><code>print(torch.sum(torch.eq(dt1, dt2)).item()\/dt1.nelement())\n<\/code><\/pre>\n"},{"AnswerId":"54187453","CreationDate":"2019-01-14T18:50:56.620","ParentId":null,"OwnerUserId":"10913498","Title":null,"Body":"<p>This below solution worked for me: <\/p>\n\n<pre><code>torch.equal(tensorA, tensorB)\n<\/code><\/pre>\n\n<p>From <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.equal\" rel=\"noreferrer\">the documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><code>True<\/code> if two tensors have the same size and elements, <code>False<\/code> otherwise.<\/p>\n<\/blockquote>\n"},{"AnswerId":"40111360","CreationDate":"2016-10-18T14:45:52.120","ParentId":null,"OwnerUserId":"3805131","Title":null,"Body":"<p>Try this if you want to ignore small precision differences which are common for floats<\/p>\n\n<pre><code>torch.all(torch.lt(torch.abs(torch.add(tens_a, -tens_b)), 1e-12))\n<\/code><\/pre>\n"}]}
{"QuestionId":33000271,"AnswerCount":2,"Tags":"<deep-learning><caffe>","CreationDate":"2015-10-07T18:56:31.887","AcceptedAnswerId":null,"OwnerUserId":3654968.0,"Title":"Caffe make runtest loops","Body":"<p>I am trying to work with caffe. make and make test succeed. When I run 'make runtest' it seems to go into an infinite loop.\n<a href=\"https:\/\/i.stack.imgur.com\/fiXCN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fiXCN.png\" alt=\"Screenshot of the make runtest looping\"><\/a> <\/p>\n","answers":[{"AnswerId":"45913543","CreationDate":"2017-08-28T07:11:42.233","ParentId":null,"OwnerUserId":"6106129","Title":null,"Body":"<pre><code>sudo make clean \nsudo make all\nsudo make pycaffe\nsudo make distribute\nsudo make test\nsudo make runtest\n<\/code><\/pre>\n"},{"AnswerId":"33423452","CreationDate":"2015-10-29T19:55:49.613","ParentId":null,"OwnerUserId":"2097312","Title":null,"Body":"<p>just run \"make runtest\". It should work.<\/p>\n"}]}
{"QuestionId":33014480,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2015-10-08T11:30:14.720","AcceptedAnswerId":33019209.0,"OwnerUserId":4968861.0,"Title":"keras example Type Error","Body":"<p>I installed WinPython and Keras 0.1.2 in my Win64 System, then I started to run the examples provided in Keras source code floder, but lots errors occured make me frustrated - -! <\/p>\n\n<p>Somebody help me, thanks for any tips.<\/p>\n\n<p>I try to run 'mnist_mlp.py', but an error occured in the line blow:<\/p>\n\n<pre><\/pre>\n\n<p>then I find the examples on <a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow\">http:\/\/keras.io\/examples\/<\/a>\nand when I run to the line below I still faced some error.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/ZhBK8.png\" rel=\"nofollow\">Keras Error<\/a><\/p>\n","answers":[{"AnswerId":"33019209","CreationDate":"2015-10-08T14:51:15.770","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>The examples <a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow\">http:\/\/keras.io\/examples\/<\/a> have not been updated after the recent API changes <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/iWVrWpR_eaQ\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/iWVrWpR_eaQ<\/a>. <\/p>\n\n<p>Make sure you have the latest version of Keras installed with:<\/p>\n\n<pre><code>sudo pip install git+git:\/\/github.com\/fchollet\/keras.git --upgrade\n<\/code><\/pre>\n\n<p>and use updated examples from the same repository <a href=\"https:\/\/github.com\/fchollet\/keras\/tree\/master\/examples\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/tree\/master\/examples<\/a><\/p>\n"},{"AnswerId":"33789501","CreationDate":"2015-11-18T20:14:50.397","ParentId":null,"OwnerUserId":"2775878","Title":null,"Body":"<p>As of 19th November, the latest version of Keras 0.2.0 isn't what you do <code>pip install<\/code>. You'll need to download it from <a href=\"https:\/\/pypi.python.org\/pypi\/Keras\" rel=\"nofollow\">https:\/\/pypi.python.org\/pypi\/Keras<\/a>. After extracting the files, run <code>python setup.py install<\/code><\/p>\n"}]}
{"QuestionId":33040854,"AnswerCount":2,"Tags":"<python><xcode><theano><pycuda>","CreationDate":"2015-10-09T14:30:13.973","AcceptedAnswerId":null,"OwnerUserId":1374969.0,"Title":"pycuda fail; Theano with Anaconda","Body":"<p>I'm using Anaconda to install Theano on MacOSX (Mavericks 10.9 ), just like this post explains: \"<a href=\"https:\/\/stackoverflow.com\/questions\/30398000\/how-to-make-theano-operate-on-mac-lion\">How to make Theano operate on Mac Lion?<\/a>\"<\/p>\n\n<blockquote>\n  <p>theano.test()      <\/p>\n<\/blockquote>\n\n<p>This command gives the same error as in the post above. It gives that error on an Ubuntu 14.1, System 76 as well. <\/p>\n\n<p><strong>I am able to import commands from Theano; but I still would like to understand why theano.test() fails. The packages CUDA and Boost were already installed before running...<\/strong> <\/p>\n\n<p>(Reference: See section: \"<em>Testing your Installation<\/em>\" <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install.html<\/a>)<\/p>\n\n<p>As the post suggests, I assumed the fix would come from installing the XCode command line, homebrew, and pycuda. The first two were installed just fine. But pycuda fails:<\/p>\n\n<blockquote>\n  <p>pip install pycuda<\/p>\n<\/blockquote>\n\n<p>....gives the following error:<\/p>\n\n<pre><\/pre>\n\n<p>It seems like the first error (gcc failed) is that the complier is not finding gcc. (Note again: I installed the MacOSX command line tools) <\/p>\n\n<p>I run <\/p>\n\n<pre><\/pre>\n\n<p>this gives usr\/bin\/gcc <\/p>\n\n<p>I also tried : <\/p>\n\n<pre><\/pre>\n\n<p>That didn't work as well. <\/p>\n\n<p>Has anyone else had this difficulty installing pycuda and can make a recommendation here? Thanks.<\/p>\n","answers":[{"AnswerId":"33076741","CreationDate":"2015-10-12T08:42:35.310","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The first error is not indicating that <code>gcc<\/code> failed to be found, rather it is saying that <code>gcc<\/code> reported an error and that error was that the file <code>cuda.h<\/code> could not be found.<\/p>\n\n<p>That error suggests you don't have the CUDA toolkit installed. You need to install this before proceeding to install PyCUDA. The CUDA toolkit can be <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow\">downloaded from NVIDIA's website<\/a>.<\/p>\n\n<p>The <a href=\"http:\/\/mathema.tician.de\/software\/pycuda\/\" rel=\"nofollow\">PyCUDA web page<\/a> indicates that it has the following prerequisites: BOOST, CUDA, and Numpy.<\/p>\n"},{"AnswerId":"33979362","CreationDate":"2015-11-29T04:16:51.390","ParentId":null,"OwnerUserId":"2695203","Title":null,"Body":"<p>I resolved this same issue with a combination of two things:<\/p>\n\n<ol>\n<li><p>Making sure <code>nvcc<\/code> is in your <code>PATH<\/code> seems resolve the compilation errors. If you then encounter the following linking error:<\/p>\n\n<p><code>ld: file not found: @rpath\/CUDA.framework\/Versions\/A\/CUDA for architecture x86_64\nclang: error: linker command failed with exit code 1 (use -v to see invocation)<\/code><\/p><\/li>\n<li><p>I fixed it changing the @rpath reference in libcuda.dylib to an absolute path:<br>\n<code>install_name_tool -change @rpath\/CUDA.framework\/Versions\/A\/CUDA \\     \/Library\/Frameworks\/CUDA.framework\/CUDA \\\n\/usr\/local\/cuda\/lib\/libcuda.dylib<\/code><\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":33047651,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-10-09T21:50:10.493","AcceptedAnswerId":33059235.0,"OwnerUserId":5429355.0,"Title":"In Lua Torch, the product of two zero matrices has nan entries","Body":"<p>I have encountered a strange behavior of the torch.mm function in Lua\/Torch. Here is a simple program that demonstrates the problem.<\/p>\n\n<pre><\/pre>\n\n<p>The program consists of one loop, in which the program multiplies two zero 2x2 matrices and tests if entry <em>ent<\/em> of the product matrix is equal to nan. It seems that the program should run forever since the product should always be equal to 0, and hence ent should be 0. However, the program prints:<\/p>\n\n<pre><\/pre>\n\n<p>Why is this happening?<\/p>\n\n<p><strong>Update:<\/strong> <\/p>\n\n<ol>\n<li>The problem disappears if I replace <em>prod = torch.mm(a,b)<\/em> with <em>torch.mm(prod,a,b)<\/em>, which suggests that something is wrong with the memory allocation. <\/li>\n<li>My version of Torch was compiled without BLAS &amp; LAPACK libraries. After I recompiled torch with OpenBLAS, the problem disappeared. However, I am still interested in its cause.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"33059235","CreationDate":"2015-10-10T21:40:22.543","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>The part of code that auto-generates the Lua wrapper for <code>torch.mm<\/code> can be found <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/28de026\/TensorMath.lua#L325-L341\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>When you write <code>prod = torch.mm(a,b)<\/code> within your loop it corresponds to the following C code behind the scenes (generated by this wrapper thanks to <a href=\"https:\/\/github.com\/torch\/cwrap\" rel=\"nofollow\">cwrap<\/a>):<\/p>\n\n<pre><code>\/* this is the tensor that will hold the results *\/\narg1 = THDoubleTensor_new(); \nTHDoubleTensor_resize2d(arg1, arg5-&gt;size[0], arg6-&gt;size[1]);\narg3 = arg1;\n\/* .... *\/\nluaT_pushudata(L, arg1, \"torch.DoubleTensor\");\n\/* effective matrix multiplication operation that will fill arg1 *\/\nTHDoubleTensor_addmm(arg1,arg2,arg3,arg4,arg5,arg6);\n<\/code><\/pre>\n\n<p>So:<\/p>\n\n<ul>\n<li>a new result tensor is created and resized with the proper dimensions,<\/li>\n<li>but <strong>this new tensor is NOT initialized<\/strong>, i.e. there is no <code>calloc<\/code> or explicit fill here so it points to junk memory and could contain NaN-s,<\/li>\n<li>this tensor is pushed on the stack so as to be available on the Lua side as the return value.<\/li>\n<\/ul>\n\n<p>The last point means that this returned tensor is different from the initial <code>prod<\/code> one (i.e. within the loop, <code>prod<\/code> shadows the initial value).<\/p>\n\n<p>On the other hand calling <code>torch.mm(prod,a,b)<\/code> <em>does<\/em> use your initial <code>prod<\/code> tensor to store the results (behind the scenes there is no need to create a dedicated tensor in that case). Since in your code snippet you do not initialize \/ fill it with given values it could also contain junk.<\/p>\n\n<p>In both cases the core operation is a <code>gemm<\/code> multiplication like C = beta * C + alpha * A * B, with beta=0 and alpha=1. The <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/28de02639e8e970532f2635d47ba3eabdf50e04f\/lib\/TH\/generic\/THBlas.c#L320-L336\" rel=\"nofollow\">naive implementation<\/a> looks like that:<\/p>\n\n<pre><code>  real *a_ = a;\n  for(i = 0; i &lt; m; i++)\n  {\n    real *b_ = b;\n    for(j = 0; j &lt; n; j++)\n    {\n      real sum = 0;\n      for(l = 0; l &lt; k; l++)\n        sum += a_[l*lda]*b_[l];\n      b_ += ldb;\n      \/*\n       * WARNING: beta*c[j*ldc+i] could give NaN even if beta=0\n       *          if the other operand c[j*ldc+i] is NaN!\n       *\/\n      c[j*ldc+i] = beta*c[j*ldc+i]+alpha*sum;\n    }\n    a_++;\n  }\n<\/code><\/pre>\n\n<p><em>Comments are mine.<\/em><\/p>\n\n<p>So:<\/p>\n\n<ol>\n<li>with <code>torch.mm(a,b)<\/code>: at each iteration, a new result tensor is created without being initialized (it could contain NaN-s). <strong>So every iteration presents a risk of returning NaN-s<\/strong> (see above warning),<\/li>\n<li>with <code>torch.mm(prod,a,b)<\/code>: there is the same risk since you do not initialized the <code>prod<\/code> tensor. BUT: this risk only exists at the first iteration of the repeat \/ until loop since right after <code>prod<\/code> is filled with 0-s and re-used for the subsequent iterations.<\/li>\n<\/ol>\n\n<p>So this is why you do not observe a problem here (it is less frequent).<\/p>\n\n<p>In case 1: this should be improved at the Torch level, i.e. make sure the wrapper initializes the output (e.g. with <code>THDoubleTensor_fill(arg1, 0);<\/code>).<\/p>\n\n<p>In case 2: you should initialize <code>prod<\/code> initially and use the <code>torch.mm(prod,a,b)<\/code> construct to avoid any NaN problem.<\/p>\n\n<p>--<\/p>\n\n<p><strong>EDIT<\/strong>: this problem is now fixed (see this <a href=\"https:\/\/github.com\/torch\/torch7\/pull\/401\" rel=\"nofollow\">pull request<\/a>).<\/p>\n"}]}
{"QuestionId":33055034,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-10-10T14:25:29.443","AcceptedAnswerId":null,"OwnerUserId":5431108.0,"Title":"Copies of cloned network do not train when placed in larger module","Body":"<p>I want to train a small network module by cloning it and placing it in a few different larger networks. But when I place it one such a context, it's parameters are not updated during training. Below is a small example that shows the problem.<\/p>\n\n<p>I make the the small primitive network like this:<\/p>\n\n<pre><\/pre>\n\n<p>Then I put it into a bigger network, called , like this:<\/p>\n\n<pre><\/pre>\n\n<p>Then I train the larger network, and print out it's parameters, and 's parameter's at iteration. What I see is that the larger  network's parameters change during training, while 's do not. Below is the training code. Is there a way to fix this?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35135201","CreationDate":"2016-02-01T16:16:26.057","ParentId":null,"OwnerUserId":"2104596","Title":null,"Body":"<p><code>getParameters()<\/code> changes the memory location of the network's parameters, so any sharing is lost. See <a href=\"https:\/\/github.com\/torch\/DEPRECEATED-torch7-distro\/issues\/33\" rel=\"nofollow\">this<\/a> for more details.<\/p>\n"}]}
{"QuestionId":33056492,"AnswerCount":1,"Tags":"<python><indexing><theano><deep-learning>","CreationDate":"2015-10-10T16:52:26.057","AcceptedAnswerId":33076627.0,"OwnerUserId":1861627.0,"Title":"Indexing a tensor in the 3rd dimension","Body":"<p>I have a batch of  sequences of integers of length  which is embedded into a  tensor. This sequence is auto-encoded by my network architecture. So, I have:  <\/p>\n\n<pre><\/pre>\n\n<p>as far as I could get, the indexing is in the first dimension of the tensor, so I had to use a . Is there a way to index the tensor directly?<\/p>\n","answers":[{"AnswerId":"33076627","CreationDate":"2015-10-12T08:36:28.513","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Sounds like you want a 3 dimensional version of <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#tensor.nnet.categorical_crossentropy\" rel=\"nofollow\"><code>theano.tensor.nnet.categorical_crossentropy<\/code><\/a>?<\/p>\n\n<p>If so, then I think you could simply flatten the matrix of true class label indexes into a vector and the 3D tensor of predicted class probabilities into a matrix and then use the built in function.<\/p>\n\n<pre><code>cost = T.nnet.categorical_crossentropy(\n    Y.reshape((Y.shape[0] * Y.shape[1], X.shape[2])),\n    X.flatten())\n<\/code><\/pre>\n\n<p>The order of entries in <code>Y<\/code> may need to be adjusted first (e.g. via a <code>dimshuffle<\/code>) to make sure the entries in the matrix and vector being compared correspond to each other.<\/p>\n\n<p>Here we assume, as the question suggests, that the sequences are not padded -- they are all exactly <code>L<\/code> elements in length. If the sequences are actually padded then you may need to do something much more complicated to avoid computing cost elements inside the padding regions.<\/p>\n"}]}
{"QuestionId":33068607,"AnswerCount":1,"Tags":"<lua><introspection><torch>","CreationDate":"2015-10-11T18:31:35.457","AcceptedAnswerId":33068608.0,"OwnerUserId":1804173.0,"Title":"How can I use the torch REPL for debugging?","Body":"<p>For quick debugging it is sometimes useful to start a REPL from a script at a certain break-point. I have found out that I can start the Torch REPL at any point by:<\/p>\n\n<pre><\/pre>\n\n<p>The only problem with this approach is that the REPL does not see any local variables from the calling chunk. Without being able to inspect locals, the REPL is not really useful as a debugger.<\/p>\n\n<p>Is it possible to start a REPL which has access to local variables?<\/p>\n\n<p><em>Disclaimer: I have found my own (newbie) solution to this problem, but I'm always open for alternatives\/suggestions.<\/em><\/p>\n","answers":[{"AnswerId":"33068608","CreationDate":"2015-10-11T18:31:35.457","ParentId":null,"OwnerUserId":"1804173","Title":null,"Body":"<p>One possible workaround is to use a wrapper, which copies the local variables of the calling scope to the global scope by using <code>debug.getlocal<\/code> before calling <code>repl()<\/code>:<\/p>\n\n<pre><code>require \"trepl\"\n\nfunction debugRepl(restoreGlobals)\n  restoreGlobals = restoreGlobals or false\n\n  -- optionally make a shallow copy of _G\n  local oldG = {}\n  if restoreGlobals then\n    for k, v in pairs(_G) do\n      oldG[k] = v\n    end\n  end\n\n  -- copy upvalues to _G\n  local i = 1\n  local func = debug.getinfo(2, \"f\").func\n  while true do\n    local k, v = debug.getupvalue(func, i)\n    if k ~= nil then\n      _G[k] = v\n    else\n      break\n    end\n    i = i + 1\n  end\n\n  -- copy locals to _G\n  local i = 1\n  while true do\n    local k, v = debug.getlocal(2, i)\n    if k ~= nil then\n      _G[k] = v\n    else\n      break\n    end\n    i = i + 1\n  end\n\n  repl()\n\n  if restoreGlobals then\n    _G = oldG\n  end\nend\n<\/code><\/pre>\n\n<p>Note (since it is not mentioned in the documentation and only visible from the <a href=\"https:\/\/github.com\/torch\/trepl\/blob\/f7960316998f24bd62aa2d1461a06f07e0196c55\/init.lua#L571\" rel=\"nofollow\">source<\/a>): Typing <code>break<\/code> in the REPL returns execution to the script, while <code>exit<\/code> (or <kbd>CTRL<\/kbd>+<kbd>D<\/kbd>) terminates the execution completely.<\/p>\n"}]}
{"QuestionId":33068711,"AnswerCount":1,"Tags":"<python><cmake><makefile><caffe>","CreationDate":"2015-10-11T18:41:16.643","AcceptedAnswerId":null,"OwnerUserId":3320135.0,"Title":"Compiling the Python Interface for Caffe using CMake","Body":"<p>On Ubuntu, I have downloaded Caffe to , and compiled it using cmake with the CMakeLists.txt file that it comes with. I now want to start using the python interface. According to the examples, I need to add  to the  variable in my  file. Then, I need to use use  in a python script. However, this gives me the error: .<\/p>\n\n<p>If I use the line  in my python script, this means that there must exist a file called  in the directories defined by . Is that correct? However, I cannot find any files called  on my system. But the  file is supposed to build the python files, and has the line .<\/p>\n\n<p>Does anybody have any experience with this? Had anybody got the python interface working by compiling with cmake?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"33094837","CreationDate":"2015-10-13T05:34:01.507","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You do not need to have <code>caffe.py<\/code> in <code>~\/caffe\/python<\/code>. In this case you should have a subfolder <code>~\/caffe\/python\/caffe<\/code> that should have <code>__init__.py<\/code> file in this subfolder. Make sure you have <code>~caffe\/python\/caffe\/_caffe.so<\/code> file - this shared object should be created during compilation.<br>\nIt may be the case that you added the right folder to your <code>.bashrc<\/code>, but have you loaded\/sourced the file after updating it? <\/p>\n"}]}
{"QuestionId":33081408,"AnswerCount":1,"Tags":"<python><matlab><neural-network><deep-learning><caffe>","CreationDate":"2015-10-12T12:40:16.280","AcceptedAnswerId":33144978.0,"OwnerUserId":1783952.0,"Title":"Python or Matlab for Caffe?","Body":"<p>I will be working on an implementation of DQNs and recent extensions from Google DeepMind in Caffe.<\/p>\n\n<p>For this I will write a simulator (in place of the Atari emulator) for creating training experiences for the agent.<\/p>\n\n<p>My question is: \n<strong>Which of the Matlab or Python interfaces to Caffe is the most mature and well functioning?<\/strong><br>\nAny other considerations for choosing between Python and Matlab for this task?<\/p>\n\n<p>It should be noted that I'm much more experienced with Python development than with Matlab development (apart from using it as a tool for math classes etc.), however a team I'm collaborating with seems to mostly use Matlab.<\/p>\n","answers":[{"AnswerId":"33144978","CreationDate":"2015-10-15T09:37:43.523","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>IMHO, it is better to use python to interface with caffe.<br>\nFrom my understanding caffe developers and community intend on developing the python interface more than the matlab one. One such direction is allowing for a net to be constructed purely in python (not using prototxt file at all), you can see such an example <a href=\"https:\/\/stackoverflow.com\/questions\/33140000\/how-to-feed-caffe-multi-label-data-in-hdf5-format?lq=1\">in this question<\/a>.<\/p>\n"}]}
{"QuestionId":33083852,"AnswerCount":1,"Tags":"<python><cuda><virtualenv><theano><lasagne>","CreationDate":"2015-10-12T14:37:58.193","AcceptedAnswerId":33127594.0,"OwnerUserId":3778384.0,"Title":"Use CUDA in virtualenv to serve Theano","Body":"<p>Lacking root permissions I installed Theano and Lasagne in a python3 virtualenv at Ubuntu 14.04. <\/p>\n\n<p>Running some code, I get an , which as far as I see results from missing CUDA header, although it is installed at .<\/p>\n\n<p>Following the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#using-the-gpu\" rel=\"nofollow\">theano documentation<\/a> I set environment variable , which yields an additional error: <\/p>\n\n<p><\/p>\n\n<p>Is there any way to include this CUDA library into the virtualenv? <\/p>\n","answers":[{"AnswerId":"33127594","CreationDate":"2015-10-14T14:06:18.483","ParentId":null,"OwnerUserId":"3778384","Title":null,"Body":"<p>Finally found the answer. In addition to environment variables I forgot to configure the <code>.theanorc<\/code> in my home directory like the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#using-the-gpu\" rel=\"nofollow\">theano documentation<\/a> suggested:<\/p>\n\n<pre><code>$ cat \/home\/user\/.theanorc\n[global]\nfloatX = float32\ndevice = gpu0\n\n[cuda]\nroot = \/usr\/local\/cuda-6.5\n<\/code><\/pre>\n\n<p>Thanks anyway! Learned the lesson to read docs carefully ...<\/p>\n"}]}
{"QuestionId":33092527,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2015-10-13T01:00:38.063","AcceptedAnswerId":33099416.0,"OwnerUserId":134397.0,"Title":"Theano scan for fast computations on an array","Body":"<p>I am trying to use Theano to speed up code that is already implemented in numpy that sums the elements in an array. In numpy, the function looks like below<\/p>\n\n<pre><\/pre>\n\n<p>with a sample call<\/p>\n\n<pre><\/pre>\n\n<p>The theano equivalent of the above function is <\/p>\n\n<pre><\/pre>\n\n<p>When called, this gives the correct output<\/p>\n\n<pre><\/pre>\n\n<p>However, when I benchmark the two, the numpy function surpasses theano in speed by a factor of three on my computer. <\/p>\n\n<pre><\/pre>\n\n<p>Since theano converts the outerloop into C, should it not be faster than Python? What can be done to speed up the theano code?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I am aware that the brute code in numpy can be optimized using summations, but the reason why I wanted to take the theano route was because I am interested in cases where the update to output can be any general function of  and , say <\/p>\n\n<pre><\/pre>\n\n<p> was just one specific example to make the point. Using mathematical notation what I am trying to implement is a fast summation , where  and  are integers, and  is a vector, and  can be any general function of  and  like the ones given above.<\/p>\n\n<pre><\/pre>\n\n<p>I was hoping that by using theano, I would be able to optimize the outter loop, and would get a faster solution than the brute numpy solution.<\/p>\n","answers":[{"AnswerId":"33099416","CreationDate":"2015-10-13T09:47:24.520","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Building on Divakar's answer...<\/p>\n\n<p>The circumstances where Theano can outperform numpy are quite specific. In general, Theano will only perform well in comparison to numpy when the computation involves vectorisable operations on large tensors.<\/p>\n\n<p>In this case the operation can be performed very efficiently in numpy. One need not use a loop at all by using the standard result for the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Arithmetic_progression#Sum\" rel=\"nofollow\">sum of an arithmetic sequence<\/a>. Here <code>n = kN - k0 + 1<\/code> is the number of items to sum.<\/p>\n\n<pre><code>numpy.arange(k0, kN + 1).sum() == (kN - k0 + 1) * (k0 + kN) \/ 2\n<\/code><\/pre>\n\n<p>If one needs to use Theano for some reason other than performance (e.g. to obtain gradients, or as part of some larger symbolic computation) then one can compute the same result without using sum or scan, just as in numpy.<\/p>\n\n<p>The following code implements the original numpy and Theano approaches, and compares them with Divakar's numpy approaches (and my Theano version of the arange sum approach) plus my numpy and Theano approaches that use the standard sum of an arithmetic sequence result.<\/p>\n\n<pre><code>import numpy\nimport timeit\nimport itertools\nimport theano\nimport theano.tensor as tt\n\n\ndef numpy1(k0, kN, x):\n    output = numpy.zeros_like(x)\n    for k in range(k0, kN + 1):\n        output += k * x\n    return output\n\n\ndef numpy2(k0, kN, x):\n    return numpy.arange(k0, kN + 1).sum() * x\n\n\ndef numpy3(k0, kN, x):\n    return numpy.einsum('i-&gt;', numpy.arange(k0, kN + 1)) * x\n\n\ndef theano1_step(k, s_tm1, x):\n    return s_tm1 + k * x\n\n\ndef compile_theano1():\n    k0 = tt.lscalar()\n    kN = tt.lscalar()\n    x = tt.vector()\n    outputs, _ = theano.scan(theano1_step, sequences=[tt.arange(k0, kN + 1)], outputs_info=[tt.zeros_like(x)],\n                             non_sequences=[x], strict=True)\n    return theano.function([k0, kN, x], outputs=outputs[-1])\n\n\ndef compile_theano2():\n    k0 = tt.lscalar()\n    kN = tt.lscalar()\n    x = tt.vector()\n    return theano.function([k0, kN, x], outputs=tt.arange(k0, kN + 1).sum() * x)\n\n\ndef numpy4(k0, kN, x):\n    return ((kN - k0 + 1) * (k0 + kN) \/ 2) * x\n\n\ndef compile_theano4():\n    k0 = tt.lscalar()\n    kN = tt.lscalar()\n    x = tt.vector()\n    return theano.function([k0, kN, x], outputs=((kN - k0 + 1) * (k0 + kN) \/ 2) * x)\n\n\ndef main():\n    iteration_count = 10\n    k0 = 10\n    kN = 10000\n    x = numpy.random.standard_normal(size=(20000,)).astype(theano.config.floatX)\n\n    functions = [numpy1, numpy2, numpy3, numpy4, compile_theano1(), compile_theano2(), compile_theano4()]\n    function_count = len(functions)\n    results = numpy.empty((iteration_count * function_count, x.shape[0]), dtype=theano.config.floatX)\n    times = numpy.empty((iteration_count * function_count,), dtype=theano.config.floatX)\n\n    for iteration in xrange(iteration_count):\n        for function_index, function in enumerate(functions):\n            start = timeit.default_timer()\n            results[iteration * function_count + function_index] = function(k0, kN, x)\n            times[iteration * function_count + function_index] = timeit.default_timer() - start\n\n    for result1, result2 in itertools.izip(results[0::2], results[1::2]):\n        assert numpy.allclose(result1, result2)\n\n    for function_name, function_index in itertools.izip(\n            ('numpy1', 'numpy2', 'numpy3', 'numpy4', 'theano1', 'theano2', 'theano4'),\n            xrange(function_count)):\n        time = times[function_index::function_count].mean()\n        print '%8s %.8f' % (function_name, float(time))\n\n\nmain()\n<\/code><\/pre>\n\n<p>On my crappy desktop computer, which uses a CPU (<em>not<\/em> a GPU) for the Theano computations, I get the following timings (in seconds, lower is better):<\/p>\n\n<pre><code>  numpy1 0.27894366\n  numpy2 0.00011240\n  numpy3 0.00008502\n  numpy4 0.00006357\n theano1 0.99175695\n theano2 0.00040656\n theano4 0.00017563\n<\/code><\/pre>\n\n<p>In this particular case, running the Theano code on a GPU is unlikely to be useful unless <code>x<\/code> is very large. But even then the cost of copying <code>x<\/code> into GPU memory could wash out any gain from the parallel element-wise multiplication.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>To address the new part in the edited version of the question...<\/p>\n\n<p>Theano is not good with explicit loops. If you can vectorize the function <code>f<\/code> then the computation can be performed much more efficiently (in time but perhaps not space) in both numpy and Theano by computing the sum along the <code>x<\/code> axis of the vectorized result.<\/p>\n\n<p>For example, if you want <code>output += exp(k*x)<\/code> then you can achieve this in numpy <em>without an explicit loop<\/em> like this:<\/p>\n\n<pre><code>k = numpy.arange(k0, kN + 1)\nresult = numpy.exp(numpy.outer(x, k)).sum(axis=0)\n<\/code><\/pre>\n\n<p>If <code>f<\/code> cannot be vectorized or a loop is necessary for some other reason then Theano may or may not offer better performance. You would have to try it out to find out. When an explicit loop is necessary, Theano is only likely to beat numpy if the computation occurring on the inside of the loop involves very large tensor operations.<\/p>\n"},{"AnswerId":"33095798","CreationDate":"2015-10-13T06:39:09.730","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>For the operation you are performing, you can simply sum all elements from <code>k0<\/code> to <code>kN<\/code> to get a scalar, which must be used to scale <code>x<\/code> for getting the desired output. This way you would have a vectorized method staying within the NumPy environment and working with <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/user\/whatisnumpy.html\" rel=\"nofollow\"><code>NumPy's strengths<\/code><\/a>.\nThe implementation with <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.sum.html\" rel=\"nofollow\"><code>np.sum()<\/code><\/a> would look like this -<\/p>\n\n<pre><code>np.arange(k0,kN+1).sum()*x\n<\/code><\/pre>\n\n<p>You can also perform the summation with <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.einsum.html\" rel=\"nofollow\"><code>np.einsum<\/code><\/a>, which might be slightly better on performance, like so -<\/p>\n\n<pre><code>np.einsum('i-&gt;',np.arange(k0,kN+1))*x\n<\/code><\/pre>\n\n<p>Runtime tests and output verification -<\/p>\n\n<pre><code>In [74]: k0 = 10; kN = 10000\n\nIn [75]: x = np.random.rand(20000)\n\nIn [76]: np.allclose(numpy_fn(k0,kN,x),np.arange(k0,kN+1).sum()*x)\nOut[76]: True\n\nIn [77]: np.allclose(numpy_fn(k0,kN,x),np.einsum('i-&gt;',np.arange(k0,kN+1))*x)\nOut[77]: True\n\nIn [78]: %timeit numpy_fn(k0,kN,x)\n1 loops, best of 3: 460 ms per loop\n\nIn [79]: %timeit np.arange(k0,kN+1).sum()*x\n10000 loops, best of 3: 54.9 \u00b5s per loop\n\nIn [80]: %timeit np.einsum('i-&gt;',np.arange(k0,kN+1))*x\n10000 loops, best of 3: 49.7 \u00b5s per loop\n<\/code><\/pre>\n"}]}
{"QuestionId":33098987,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2015-10-13T09:27:57.587","AcceptedAnswerId":33099405.0,"OwnerUserId":242871.0,"Title":"Calculating top-5 error rate in Caffe?","Body":"<p>I use synset for calculating sort top  predictions from softmax output. <\/p>\n\n<p>Which gives me the top 5 class names. But I want to know how can I calculate its percentage. \nI mean the top-5% error. <\/p>\n\n<p>Can anybody kindly guide me.\nthanks. <\/p>\n","answers":[{"AnswerId":"33099405","CreationDate":"2015-10-13T09:46:54.533","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>During training, you can compute the \"top k\" accuracy using \"top_k\" parameter to the accuracy layer. Here's an example modifying the existing accuracy layer in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/train_val.prototxt#L384\" rel=\"nofollow\">bvlc_reference_caffe's <code>train_val.prototxt<\/code><\/a>:<\/p>\n\n<pre><code>layer {\n  name: \"accuracy_top_5\"\n  type: \"Accuracy\"\n  bottom: \"fc8\"\n  bottom: \"label\"\n  top: \"accuracy_top_5\"\n  accuracy_param {\n    top_k: 5\n  }\n  include {\n    phase: TEST\n  }\n}     \n<\/code><\/pre>\n"}]}
{"QuestionId":33101145,"AnswerCount":1,"Tags":"<neural-network><computer-vision><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-10-13T11:11:21.037","AcceptedAnswerId":33101342.0,"OwnerUserId":1348187.0,"Title":"Trying to train the ImageNet model with Region CNN (R-CNN)","Body":"<p>After several month working with <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a>, I've been able to train my own models successfully. For example further than my own models, I've been able to train ImageNet with 1000 classes.<\/p>\n\n<p>In my project now, I'm trying to extract the region of my interest class. After that I've compiled and run the demo of <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\" rel=\"nofollow\"><strong>Fast R-CNN<\/strong><\/a> and it works ok, but the sample models contains only 20 classes and I'd like to have more classes, for example all of them.<\/p>\n\n<p>I've already downloaded the <a href=\"http:\/\/image-net.org\/download-bboxes\" rel=\"nofollow\"><strong>bounding boxes<\/strong><\/a> of ImageNet, with the real images.<\/p>\n\n<p>Now, I've gone blank, I can't figure out the next steps and there's not a documentation of how to do it. The only thing I've found is how to train the INRIA person model, and they provide dataset + annotations + python script.<\/p>\n\n<p>My questions are:<\/p>\n\n<ul>\n<li>Is there maybe any tutorial or guide that I've missed?<\/li>\n<li>Is there already a model trained with 1000 classes able to classify images and extract the bounding boxes?<\/li>\n<\/ul>\n\n<p>Thank you very much in advance.<\/p>\n\n<p>Regards.<\/p>\n\n<p>Rafael.<\/p>\n","answers":[{"AnswerId":"33101342","CreationDate":"2015-10-13T11:21:26.903","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><a href=\"http:\/\/www.cs.berkeley.edu\/~rbg\/\" rel=\"nofollow\">Dr Ross Girshik<\/a> has done a lot of work on object detection. You can learn a lot from his detailed git on <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\" rel=\"nofollow\">fast RCNN<\/a>: you should be able to find a caffe branch there, with a demo. I did not use it myself, but it seems very comprehensible.<\/p>\n\n<p>Another direction you might find interesting is <a href=\"http:\/\/lsda.berkeleyvision.org\/\" rel=\"nofollow\">LSDA<\/a>: using weak supervision to train object detection for many classes.<\/p>\n\n<p>BTW, have you looked into <a href=\"http:\/\/arxiv.org\/abs\/1506.01497\" rel=\"nofollow\">faster-rcnn<\/a>?<\/p>\n"}]}
{"QuestionId":33101410,"AnswerCount":1,"Tags":"<python><numpy><theano><deep-learning>","CreationDate":"2015-10-13T11:24:29.387","AcceptedAnswerId":33101585.0,"OwnerUserId":492372.0,"Title":"Why does this Theano code run successfully without any errors?","Body":"<p>I have the following piece of code borrowed from an online tutorial. I see that the below line written in the main method of the code<\/p>\n\n<pre><\/pre>\n\n<p>is adding the tensor 'a' of dimension (2,1,2,2) and tensor 'b' of dimension (2,2,2,2). How is it able to add correctly even when we have declared broadcastable as 'false' in the make_tensor method? Shouldn't we declare broadcastable as True so it can add with different dimensions? Shouldn't it throw an error saying dimensions don't match? Is my understanding of broadcastable wrong?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33101585","CreationDate":"2015-10-13T11:32:03.603","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The reason this works is that new dimensions added by <code>dimshuffle<\/code> via the <code>'x'<\/code> parameter value are always broadcastable.<\/p>\n\n<p>Notice that in <code>broadcasted_add<\/code> the only dimension that needs to be broadcasted is the dimension that was added to <code>a<\/code> via the <code>dimshuffle<\/code>. None of the other dimensions need to be broadcasted.<\/p>\n"}]}
{"QuestionId":33103111,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-10-13T12:46:46.730","AcceptedAnswerId":33103490.0,"OwnerUserId":492372.0,"Title":"Is there an error in the way the params are updated in the following Theano method?","Body":"<p>I was going through a tutorial online for momentum based learning and came across this method in Theano<\/p>\n\n<pre><\/pre>\n\n<p>Shouldn't the order of the following two lines be the other way round (interchanged) ?<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>I understand that after the train method is executed and the cost is calculated, only then the updates is run, correct? <\/p>\n\n<p>Doesn't that means we should use the current cost, and with the existing param_update value (that comes from the previous iteration), we should calculate the newer param_update and thus update the current param value?<\/p>\n\n<p>Why is it the other way round and why is that correct?<\/p>\n","answers":[{"AnswerId":"33103490","CreationDate":"2015-10-13T13:05:16.923","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The order of updates inside the updates list provided to <code>theano.function<\/code> is ignored. Updates are always computed using the the <em>old<\/em> values of shared variables.<\/p>\n\n<p>This snippet of code shows that the order of updates is ignored:<\/p>\n\n<pre><code>import theano\nimport theano.tensor\n\np = 0.5\nparam = theano.shared(1.)\nparam_update = theano.shared(2.)\ncost = 3 * param * param\nupdate_a = (param, param - param_update)\nupdate_b = (param_update, p * param_update + (1 - p) * theano.grad(cost, param))\nupdates1 = [update_a, update_b]\nupdates2 = [update_b, update_a]\nf1 = theano.function([], outputs=[param, param_update], updates=updates1)\nf2 = theano.function([], outputs=[param, param_update], updates=updates2)\nprint f1(), f1()\nparam.set_value(1)\nparam_update.set_value(2)\nprint f2(), f2()\n<\/code><\/pre>\n\n<p>If, logically, you want<\/p>\n\n<pre><code>new_a = old_a + a_update\nnew_b = new_a + b_update\n<\/code><\/pre>\n\n<p>Then you need to provide updates like this:<\/p>\n\n<pre><code>new_a = old_a + a_update\nnew_b = old_a + a_update + b_update\n<\/code><\/pre>\n"}]}
{"QuestionId":33112941,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><pycaffe><matcaffe>","CreationDate":"2015-10-13T21:27:25.743","AcceptedAnswerId":33142623.0,"OwnerUserId":4975126.0,"Title":"Multiple category classification in Caffe","Body":"<p>I thought we might be able to compile a Caffeinated description of some methods of performing <em>multiple category classification<\/em>.<\/p>\n\n<p><strong>By multi category classification I mean:<\/strong> The input data containing representations of multiple model output categories and\/or simply being classifiable under multiple model output categories.<\/p>\n\n<p>E.g. An image containing a cat &amp; dog would output (ideally) ~1 for both the cat &amp; dog prediction categories and ~0 for all others.<\/p>\n\n<ol>\n<li><p>Based on <a href=\"http:\/\/arxiv.org\/pdf\/1312.6082v4.pdf\" rel=\"noreferrer\">this paper<\/a>, <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/523\" rel=\"noreferrer\">this stale and closed PR<\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698\" rel=\"noreferrer\">this open PR<\/a>, it seems caffe is perfectly capable of accepting labels. Is this correct?<\/p><\/li>\n<li><p>Would the construction of such a network require the use of multiple neuron (inner product -> relu -> inner product) and softmax layers as in <a href=\"http:\/\/arxiv.org\/pdf\/1312.6082v4.pdf\" rel=\"noreferrer\">page 13 of this paper<\/a>; or does Caffe's ip &amp; softmax presently support multiple label dimensions?<\/p><\/li>\n<li><p>When I'm passing my labels to the network which example would illustrate the correct approach (if not both)?:<\/p>\n\n<p><strong>E.g. Cat eating apple<\/strong> Note: Python syntax, but I use the c++ source.<\/p>\n\n<p>Column 0 - Class is in input;\nColumn 1 - Class is not in input<\/p>\n\n<pre><\/pre>\n\n<p>or<\/p>\n\n<p>Column 0 - Class is in input<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p><strong>If anything lacks clarity please let me know and I will generate pictorial examples of the questions I'm trying to ask.<\/strong><\/p>\n","answers":[{"AnswerId":"33142623","CreationDate":"2015-10-15T07:38:56.010","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Nice question. I believe there is no single \"canonical\" answer here and you may find several different approaches to tackle this problem. I'll do my best to show one possible way. It is slightly different than the question you asked, so I'll re-state the problem and suggest a solution. <\/p>\n\n<p><strong>The problem:<\/strong> given an input image and a set of <code>C<\/code> classes, indicate <em>for each class<\/em> if it is depicted in the image or not. <\/p>\n\n<p><strong>Inputs:<\/strong> in training time, inputs are pairs of image and a <code>C<\/code>-dim <em>binary<\/em> vector indicating for each class of the <code>C<\/code> classes if it is present in the image or not.<\/p>\n\n<p><strong>Output:<\/strong> given an image, output a <code>C<\/code>-dim binary vector (same as the second form suggested in your question).<\/p>\n\n<p><strong>Making caffe do the job:<\/strong> In order to make this work we need to modify the top layers of the net using a different loss.<br>\nBut first, let's understand the usual way caffe is used and then look into the changes needed.<br>\n<em>The way things are now:<\/em> image is fed into the net, goes through conv\/pooling\/... layers and finally goes through an <code>\"InnerProduct\"<\/code> layer with <code>C<\/code> outputs. These <code>C<\/code> predictions goes into a <code>\"Softmax\"<\/code> layer that inhibits all but the most dominant class. Once a single class is highlighted <code>\"SoftmaxWithLoss\"<\/code> layer checks that the highlighted predicted class matches the ground truth class.<\/p>\n\n<p><em>What you need:<\/em> the problem with the existing approach is the <code>\"Softmax\"<\/code> layer that basically selects a single class. I suggest you <em>replace<\/em> it with a <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidLayer.html\" rel=\"nofollow noreferrer\"><code>\"Sigmoid\"<\/code><\/a> layer that maps <em>each<\/em> of the <code>C<\/code> outputs into an indicator whether this specific class is present in the image. For training, you should use <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidCrossEntropyLossLayer.html\" rel=\"nofollow noreferrer\"><code>\"SigmoidCrossEntropyLoss\"<\/code><\/a> instead of the <code>\"SoftmaxWithloss\"<\/code> layer.<\/p>\n"},{"AnswerId":"34993636","CreationDate":"2016-01-25T13:18:06.297","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>Since one image can have multiple labels. The most intuitive way is to think this problem as a C independt binary classification problem where C is the total number of different classes. So it is easy to understand what <a href=\"https:\/\/stackoverflow.com\/a\/33142623\/1714410\">@Shai have said<\/a>:<\/p>\n\n<blockquote>\n  <p>add a \"Sigmoid\" layer that maps each of the C outputs into an indicator whether this specific class is present in the image, and should use \"SigmoidCrossEntropyLoss\" instead of the \"SoftmaxWithloss\" layer.\n  The loss is the sum of these C SigmoidCrossEntropyLoss.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33117607,"AnswerCount":2,"Tags":"<python><caffe><lmdb>","CreationDate":"2015-10-14T05:53:29.987","AcceptedAnswerId":33123313.0,"OwnerUserId":639973.0,"Title":"Caffe: Reading LMDB from Python","Body":"<p>I've extracted features using caffe, which generates a .mdb file.\nThen I'm trying to read it using Python and display it as a readable number.<\/p>\n\n<pre><\/pre>\n\n<p>This prints out a very long line of unreadable, broken characters.<\/p>\n\n<p>Then I tried printing int(value), which returns the following:<\/p>\n\n<pre><\/pre>\n\n<p>float(value) gives the following:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a problem with the lmdb file itself, or does it have to do with conversion of data type?<\/p>\n","answers":[{"AnswerId":"37807822","CreationDate":"2016-06-14T09:18:47.940","ParentId":null,"OwnerUserId":"1545917","Title":null,"Body":"<p>If you have encoded images in <code>lmdb<\/code>, you'll probably see this error when using @ytrewq's code<\/p>\n\n<pre><code>ValueError: total size of new array must be unchanged\n<\/code><\/pre>\n\n<p>Use this function instead:<\/p>\n\n<pre><code>import caffe\nimport lmdb\nimport PIL.Image\nfrom StringIO import StringIO\nimport numpy as np\n\ndef read_lmdb(lmdb_file):\n    cursor = lmdb.open(lmdb_file, readonly=True).begin().cursor()\n    datum = caffe.proto.caffe_pb2.Datum()\n    for _, value in cursor:\n        datum.ParseFromString(value)\n        s = StringIO()\n        s.write(datum.data)\n        s.seek(0)\n\n        yield np.array(PIL.Image.open(s)), datum.label\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre><code>lmdb_dir = '\/save\/jobs\/20160613-125532-958f\/train_db\/'\nfor im, label in read_lmdb(lmdb_dir):\n    print label, im\n<\/code><\/pre>\n"},{"AnswerId":"33123313","CreationDate":"2015-10-14T10:47:24.600","ParentId":null,"OwnerUserId":"639973","Title":null,"Body":"<p>Here's the working code I figured out<\/p>\n\n<pre><code>import caffe\nimport lmdb\n\nlmdb_env = lmdb.open('directory_containing_mdb')\nlmdb_txn = lmdb_env.begin()\nlmdb_cursor = lmdb_txn.cursor()\ndatum = caffe.proto.caffe_pb2.Datum()\n\nfor key, value in lmdb_cursor:\n    datum.ParseFromString(value)\n    label = datum.label\n    data = caffe.io.datum_to_array(datum)\n    for l, d in zip(label, data):\n            print l, d\n<\/code><\/pre>\n"}]}
{"QuestionId":33120123,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-10-14T08:18:31.530","AcceptedAnswerId":null,"OwnerUserId":1000290.0,"Title":"Theano (GPU) memory footprint management \/ debug","Body":"<p>I'm using theano to do some computation involving a large (about 300,000 x 128) matrix.<\/p>\n\n<p>The theano function quit after outputting a , similar to <a href=\"https:\/\/stackoverflow.com\/questions\/32936166\/how-do-you-free-up-gpu-memory\">this question<\/a>.<\/p>\n\n<p>I think it's probably because the large matrix gets processed step by step, and each step leaves a large  (although the same shape as the first one) in the memory.<\/p>\n\n<p>So my question is:<\/p>\n\n<ol>\n<li><p>In the case of temporary(non-output) variables with the same shape and dtype, does theano make any effort to reuse such allocated memory? like, a pool for each array shape exist?<\/p><\/li>\n<li><p>If 1, can I inspect which node in the function graph reuses memory?<\/p><\/li>\n<\/ol>\n\n<p>Although I know there's <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/shared.html\" rel=\"nofollow\">shared<\/a> that can do explicit sharing, but I'm in doubt that using it would make the (already hard) computation code harder to understand.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>A simplified example of such situation:<\/p>\n\n<pre><\/pre>\n\n<p>If none of op1, op2, op3 can be optimized as in-place,\nwill theano try to reuse memory, e.g.  and  might \"share\"\nthe same address, to reduce memory footprint, since  is no\nlonger used at the time of ?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"33121010","CreationDate":"2015-10-14T09:02:43.563","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You need to use in-place operations as much as possible. For the most part this is not under user control (they are automatically used by the optimizer when circumstances allow) but there are a few things you can do to encourage their use.<\/p>\n\n<p>Take a look at the documentation on this issue:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/profiling.html\" rel=\"nofollow\">memory profiler<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/optimizations.html\" rel=\"nofollow\">Information on the optimizations that enable in-place operation<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/inplace.html\" rel=\"nofollow\">How to create custom operations that can work in-place<\/a><\/li>\n<\/ul>\n\n<p>Do not use the <code>inplace<\/code> parameter of the <code>inc_subtensor<\/code> operation, as indicated in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#indexing\" rel=\"nofollow\">documentation<\/a>. The <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#inplace\" rel=\"nofollow\">docs also indicate<\/a> that inplace operators are not supported for user specification (the optimizer will automatically apply them when possible).<\/p>\n\n<p>You could use Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/profiling.html\" rel=\"nofollow\">memory profiler<\/a> to help track down which operation(s) are using up the memory.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I'm not an expert in this area but I believe Theano uses a garbage collection mechanism to free up memory that is no longer needed. This is discussed in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/faq.html#faster-theano-function\" rel=\"nofollow\">documentation.<\/a> <\/p>\n"}]}
{"QuestionId":33124269,"AnswerCount":2,"Tags":"<machine-learning><neural-network><caffe><conv-neural-network><pycaffe>","CreationDate":"2015-10-14T11:36:29.313","AcceptedAnswerId":null,"OwnerUserId":5321075.0,"Title":"Visualising convolution kernels in caffe","Body":"<p>I have been following a Caffe example <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow noreferrer\">here<\/a> to plot the Convolution kernels from my ConvNet. I have attached an image below of my kernels, however it looks nothing like the kernels in the example. I have followed the example exactly, anyone know what the issue may be?<\/p>\n\n<p>My net is trained on a set of simulated images (with two classes) and the performance of the net is pretty good, around 80% test accuracy.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iwWNU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iwWNU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38546153","CreationDate":"2016-07-23T20:30:17.500","ParentId":null,"OwnerUserId":"2898233","Title":null,"Body":"<p>To get \"smoother\" filters you could try to add a small amount of L2 weight-decay (decay_mult) to the conv1 layer. <\/p>\n\n<p>See also <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n\n<pre><code>layer {\n  name: \"conv1\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"conv1\"\n  # learning rate and decay multipliers for the filters\n  param { lr_mult: 1 decay_mult: 1 }\n  # learning rate and decay multipliers for the biases\n  param { lr_mult: 2 decay_mult: 0 }\n  convolution_param {\n    num_output: 96     # learn 96 filters\n    kernel_size: 11    # each filter is 11x11\n    stride: 4          # step 4 pixels between each filter application\n    weight_filler {\n      type: \"gaussian\" # initialize the filters from a Gaussian\n      std: 0.01        # distribution with stdev 0.01 (default mean: 0)\n    }\n    bias_filler {\n      type: \"constant\" # initialize the biases to zero (0)\n      value: 0\n    }\n  }\n}\n<\/code><\/pre>\n"},{"AnswerId":"36656650","CreationDate":"2016-04-15T20:54:42.983","ParentId":null,"OwnerUserId":"369747","Title":null,"Body":"<p>Well, you might need to set the interpolation parameter to 'none' when you call imshow. Is that what you are referring to?<\/p>\n"}]}
{"QuestionId":33131265,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-10-14T16:54:25.667","AcceptedAnswerId":33158967.0,"OwnerUserId":1135950.0,"Title":"Multiply tensor lua","Body":"<p>I'm new to lua,<\/p>\n\n<p>Why the  below code does not work?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33158967","CreationDate":"2015-10-15T21:40:12.360","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p><code>data[5]<\/code> is a 1-D tensor and <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-t\" rel=\"nofollow\">transpose<\/a> only works for 2-D tensors. So you can use the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\">indexing<\/a> operator as follow to obtain an extra dimension:<\/p>\n\n<pre><code> -- matrix-matrix operation: result is a 1x1 tensor\n local tmp = w * data[{{5}}]:t()\n<\/code><\/pre>\n\n<p>Or alternatively <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-squeezedim\" rel=\"nofollow\">squeeze<\/a> the first singleton dimension of <code>w<\/code>:<\/p>\n\n<pre><code> -- dot prod between 1D tensors: result is a number\n local tmp = w:squeeze() * data[5]\n<\/code><\/pre>\n"}]}
{"QuestionId":33139114,"AnswerCount":2,"Tags":"<python><eclipse><cmd><theano>","CreationDate":"2015-10-15T03:01:23.947","AcceptedAnswerId":null,"OwnerUserId":1360133.0,"Title":"Theano + Windows + CUDA, work well using CMD, but encoutered error in eclipse (warning C4819)","Body":"<p>I am working with `\nWindows + CUDA 6.5 + VS2010 + Python2.7 + Theano. \nMy Python scripts works well when using command line, say:<\/p>\n\n<pre><\/pre>\n\n<p>and I get results as:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/uo5tY.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uo5tY.jpg\" alt=\"passed in CMD 1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OghuV.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OghuV.jpg\" alt=\"passed in CMD 2\"><\/a><\/p>\n\n<p>However, when I run the same  in , I get a lot of errors:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TQSQW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TQSQW.png\" alt=\"error in Eclipse 1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/myVTu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/myVTu.png\" alt=\"error in Eclipse 2\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/t7JYc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/t7JYc.png\" alt=\"error in Eclipse 3\"><\/a><\/p>\n\n<p>I think this is probably due to the difference of character Map of  and , because the  implies that a lot of  files are not parsed correctly. How to fix it?<\/p>\n","answers":[{"AnswerId":"37871970","CreationDate":"2016-06-17T01:58:57.327","ParentId":null,"OwnerUserId":"3148107","Title":null,"Body":"<p>I guess your system setting of 'system locale' is Chinese. Change it to English could possibly fix this problem.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/37871883\/3148107\">https:\/\/stackoverflow.com\/a\/37871883\/3148107<\/a><\/p>\n"},{"AnswerId":"33161063","CreationDate":"2015-10-16T01:21:43.960","ParentId":null,"OwnerUserId":"1360133","Title":null,"Body":"<p>I find the answer myself. This is because, I installed Microsoft C++ for Python first, and then installed VS2010. The environment variable of Eclipse can only be updated after restarting windows. <\/p>\n\n<p>However, after I restart windows, the problem changes, not  warning C4819 any more, it began to report errors like \"unresolved external ...\", which is mainly due to some DLLs or libs are not included correctly by Eclipse.<\/p>\n\n<p>Finally, I decide to change my IDE to PyCharm. Everything goes well so far.<\/p>\n"}]}
{"QuestionId":33140000,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><caffe>","CreationDate":"2015-10-15T04:38:41.843","AcceptedAnswerId":33166461.0,"OwnerUserId":290163.0,"Title":"How to feed caffe multi label data in HDF5 format?","Body":"<p>I want to use caffe with a vector label, not integer. I have checked some answers, and it seems HDF5 is a better way. But then I'm stucked with error like:<\/p>\n\n<blockquote>\n  <p>accuracy_layer.cpp:34] Check failed:  (50 vs. 200) Number of labels must match number of predictions; e.g., if label axis == 1 and prediction shape is (N, C, H, W), label count (number of labels) must be , with integer values in {0, 1, ..., C-1}.<\/p>\n<\/blockquote>\n\n<p>with HDF5 created as:<\/p>\n\n<pre><\/pre>\n\n<p>My network is generated by:<\/p>\n\n<pre><\/pre>\n\n<p>It seems I should increase label number and put things in integer rather than array, but if I do this, caffe complains number of data and label is not equal, then exists.<\/p>\n\n<p>So, what is the correct format to feed multi label data?<\/p>\n\n<p>Also, I'm so wondering why no one just simply write the data format how HDF5 maps to caffe blobs?<\/p>\n","answers":[{"AnswerId":"33141214","CreationDate":"2015-10-15T06:18:27.643","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your accuracy layer makes no sense. <\/p>\n\n<p><em>The way accuracy layer works<\/em>: in <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> accuracy layer expects two inputs<br>\n(i) a predicted probability <em>vector<\/em> and<br>\n(ii) ground-truth corresponding <em>scalar<\/em> integer label.<br>\nThe accuracy layer than checks if the probability of the predicted label is indeed the maximal (or within <code>top_k<\/code>).<br>\nTherefore if you have to classify <code>C<\/code> different classes, your inputs are going to be <code>N<\/code>-by-<code>C<\/code> (where <code>N<\/code> is batch size) input predicted probabilities for <code>N<\/code> samples belonging to each of the <code>C<\/code> classes, and <code>N<\/code> labels.<\/p>\n\n<p><em>The way it is defined in your net<\/em>: You input accuracy layer <code>N<\/code>-by-4 predictions and <code>N<\/code>-by-4 labels -- this makes no sense for caffe.<\/p>\n"},{"AnswerId":"33166461","CreationDate":"2015-10-16T09:05:52.870","ParentId":null,"OwnerUserId":"290163","Title":null,"Body":"<p>Answer to this question's title:<\/p>\n\n<p>The HDF5 file should have two dataset in root, named \"data\" and \"label\", respectively. The shape is (<code>data amount<\/code>, <code>dimension<\/code>). I'm using only one-dimension data, so I'm not sure what's the order of <code>channel<\/code>, <code>width<\/code>, and <code>height<\/code>. Maybe it does not matter. <code>dtype<\/code> should be float or double.<\/p>\n\n<p>A sample code creating train set with <code>h5py<\/code> is:<\/p>\n\n<pre>\nimport h5py, os\nimport numpy as np\n\nf = h5py.File('train.h5', 'w')\n# 1200 data, each is a 128-dim vector\nf.create_dataset('data', (1200, 128), dtype='f8')\n# Data's labels, each is a 4-dim vector\nf.create_dataset('label', (1200, 4), dtype='f4')\n\n# Fill in something with fixed pattern\n# Regularize values to between 0 and 1, or SigmoidCrossEntropyLoss will not work\nfor i in range(1200):\n    a = np.empty(128)\n    if i % 4 == 0:\n        for j in range(128):\n            a[j] = j \/ 128.0;\n        l = [1,0,0,0]\n    elif i % 4 == 1:\n        for j in range(128):\n            a[j] = (128 - j) \/ 128.0;\n        l = [1,0,1,0]\n    elif i % 4 == 2:\n        for j in range(128):\n            a[j] = (j % 6) \/ 128.0;\n        l = [0,1,1,0]\n    elif i % 4 == 3:\n        for j in range(128):\n            a[j] = (j % 4) * 4 \/ 128.0;\n        l = [1,0,1,1]\n    f['data'][i] = a\n    f['label'][i] = l\n\nf.close()\n<\/pre>\n\n<p>Also, the accuracy layer is not needed, simply removing it is fine. Next problem is the loss layer. Since <code>SoftmaxWithLoss<\/code> has only one output (index of the dimension with max value), it can't be used for multi-label problem. Thank to Adian and Shai, I find <code>SigmoidCrossEntropyLoss<\/code> is good in this case.<\/p>\n\n<p>Below is the full code, from data creation, training network, and getting test result:<\/p>\n\n<blockquote>\n  <p>main.py (modified from caffe lanet example)<\/p>\n<\/blockquote>\n\n<pre>\nimport os, sys\n\nPROJECT_HOME = '...\/project\/'\nCAFFE_HOME = '...\/caffe\/'\nos.chdir(PROJECT_HOME)\n\nsys.path.insert(0, CAFFE_HOME + 'caffe\/python')\nimport caffe, h5py\n\nfrom pylab import *\nfrom caffe import layers as L\n\ndef net(hdf5, batch_size):\n    n = caffe.NetSpec()\n    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)\n    n.ip1 = L.InnerProduct(n.data, num_output=50, weight_filler=dict(type='xavier'))\n    n.relu1 = L.ReLU(n.ip1, in_place=True)\n    n.ip2 = L.InnerProduct(n.relu1, num_output=50, weight_filler=dict(type='xavier'))\n    n.relu2 = L.ReLU(n.ip2, in_place=True)\n    n.ip3 = L.InnerProduct(n.relu2, num_output=4, weight_filler=dict(type='xavier'))\n    n.loss = L.SigmoidCrossEntropyLoss(n.ip3, n.label)\n    return n.to_proto()\n\nwith open(PROJECT_HOME + 'auto_train.prototxt', 'w') as f:\n    f.write(str(net(PROJECT_HOME + 'train.h5list', 50)))\nwith open(PROJECT_HOME + 'auto_test.prototxt', 'w') as f:\n    f.write(str(net(PROJECT_HOME + 'test.h5list', 20)))\n\ncaffe.set_device(0)\ncaffe.set_mode_gpu()\nsolver = caffe.SGDSolver(PROJECT_HOME + 'auto_solver.prototxt')\n\nsolver.net.forward()\nsolver.test_nets[0].forward()\nsolver.step(1)\n\nniter = 200\ntest_interval = 10\ntrain_loss = zeros(niter)\ntest_acc = zeros(int(np.ceil(niter * 1.0 \/ test_interval)))\nprint len(test_acc)\noutput = zeros((niter, 8, 4))\n\n# The main solver loop\nfor it in range(niter):\n    solver.step(1)  # SGD by Caffe\n    train_loss[it] = solver.net.blobs['loss'].data\n    solver.test_nets[0].forward(start='data')\n    output[it] = solver.test_nets[0].blobs['ip3'].data[:8]\n\n    if it % test_interval == 0:\n        print 'Iteration', it, 'testing...'\n        correct = 0\n        data = solver.test_nets[0].blobs['ip3'].data\n        label = solver.test_nets[0].blobs['label'].data\n        for test_it in range(100):\n            solver.test_nets[0].forward()\n            # Positive values map to label 1, while negative values map to label 0\n            for i in range(len(data)):\n                for j in range(len(data[i])):\n                    if data[i][j] > 0 and label[i][j] == 1:\n                        correct += 1\n                    elif data[i][j] %lt;= 0 and label[i][j] == 0:\n                        correct += 1\n        test_acc[int(it \/ test_interval)] = correct * 1.0 \/ (len(data) * len(data[0]) * 100)\n\n# Train and test done, outputing convege graph\n_, ax1 = subplots()\nax2 = ax1.twinx()\nax1.plot(arange(niter), train_loss)\nax2.plot(test_interval * arange(len(test_acc)), test_acc, 'r')\nax1.set_xlabel('iteration')\nax1.set_ylabel('train loss')\nax2.set_ylabel('test accuracy')\n_.savefig('converge.png')\n\n# Check the result of last batch\nprint solver.test_nets[0].blobs['ip3'].data\nprint solver.test_nets[0].blobs['label'].data\n<\/pre>\n\n<p>h5list files simply contain paths of h5 files in each line:<\/p>\n\n<blockquote>\n  <p>train.h5list<\/p>\n<\/blockquote>\n\n<pre><code>\/home\/foo\/bar\/project\/train.h5\n<\/code><\/pre>\n\n<blockquote>\n  <p>test.h5list<\/p>\n<\/blockquote>\n\n<pre><code>\/home\/foo\/bar\/project\/test.h5\n<\/code><\/pre>\n\n<p>and the solver:<\/p>\n\n<blockquote>\n  <p>auto_solver.prototxt<\/p>\n<\/blockquote>\n\n<pre>train_net: \"auto_train.prototxt\"\ntest_net: \"auto_test.prototxt\"\ntest_iter: 10\ntest_interval: 20\nbase_lr: 0.01\nmomentum: 0.9\nweight_decay: 0.0005\nlr_policy: \"inv\"\ngamma: 0.0001\npower: 0.75\ndisplay: 100\nmax_iter: 10000\nsnapshot: 5000\nsnapshot_prefix: \"sed\"\nsolver_mode: GPU\n<\/pre>\n\n<p>Converge graph:\n<a href=\"https:\/\/i.stack.imgur.com\/i0pHF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/i0pHF.png\" alt=\"Converge graph\"><\/a><\/p>\n\n<p>Last batch result:<\/p>\n\n<pre>\n[[ 35.91593933 -37.46276474 -6.2579031 -6.30313492]\n[ 42.69248581 -43.00864792 13.19664764 -3.35134125]\n[ -1.36403108 1.38531208 2.77786589 -0.34310576]\n[ 2.91686511 -2.88944006 4.34043217 0.32656598]\n...\n[ 35.91593933 -37.46276474 -6.2579031 -6.30313492]\n[ 42.69248581 -43.00864792 13.19664764 -3.35134125]\n[ -1.36403108 1.38531208 2.77786589 -0.34310576]\n[ 2.91686511 -2.88944006 4.34043217 0.32656598]]\n\n[[ 1. 0. 0. 0.]\n[ 1. 0. 1. 0.]\n[ 0. 1. 1. 0.]\n[ 1. 0. 1. 1.]\n...\n[ 1. 0. 0. 0.]\n[ 1. 0. 1. 0.]\n[ 0. 1. 1. 0.]\n[ 1. 0. 1. 1.]]\n<\/pre>\n\n<p>I think this code still has many things to improve. Any suggestion is appreciated.<\/p>\n"}]}
{"QuestionId":33149311,"AnswerCount":2,"Tags":"<python><cuda><gpu><nvidia><theano>","CreationDate":"2015-10-15T13:02:27.217","AcceptedAnswerId":null,"OwnerUserId":3338246.0,"Title":"Theano on a GPU: configuration on 64-bit Windows 7","Body":"<p>I'm running Python 2.7 (Anaconda distribution) on a 64-bit Windows 7 workstation, and I'm trying to get Theano (v0.7.0) to run on an NVIDIA Quadro 2000 GPU for a machine learning application. Here's where I'm up to:<\/p>\n\n<ol>\n<li><p>I've installed CUDA v6.5 and added all of the necessary paths to the Windows path environment variable, as far as I can tell. I've been through the CUDA installation guide for Windows (<a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-microsoft-windows\/index.html#axzz3oYtn9SKk\" rel=\"nofollow\">http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-microsoft-windows\/index.html#axzz3oYtn9SKk<\/a>) and passed each test successfully, suggesting that the CUDA toolkit is communicating with the GPU.<\/p><\/li>\n<li><p>I also have Visual Studio 12.0 installed, and paths set to the following directories:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Theano will run my code on the CPU. However, if I try and run some test code on the GPU, I get the following errors and warnings:<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>This seems to imply that the GPU and\/or CUDA are unavailable, or that the NVCC compiler can't be located in the PATH variable? If it's the latter, my path contains <\/p>\n\n<pre><\/pre>\n\n<p>but perhaps there are other directories that should be included?<\/p>\n\n<p>Any idea what could be going on?<\/p>\n","answers":[{"AnswerId":"35450840","CreationDate":"2016-02-17T07:56:06.273","ParentId":null,"OwnerUserId":"1762932","Title":null,"Body":"<p>I also encountered a similar problem about \"CUDA is installed, but device gpu is not available  (error: cuda unavilable)\" previously. <\/p>\n\n<p>The problem was solved right after following the instruction of one line on <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a>:<\/p>\n\n<hr>\n\n<p>download the <strong>stdint.h<\/strong> header from <a href=\"http:\/\/msinttypes.googlecode.com\/svn\/trunk\/stdint.h\" rel=\"nofollow\">here<\/a> and save it as <strong>C:\\Program Files (x86)\\Common Files\\Microsoft\\Visual C++ for Python\\9.0\\VC\\include\\stdint.h<\/strong><\/p>\n"},{"AnswerId":"37525926","CreationDate":"2016-05-30T12:40:16.743","ParentId":null,"OwnerUserId":"1903430","Title":null,"Body":"<p>It might happen because the compiling to use PyCUDA was failed.<\/p>\n\n<p>Install <code>MinGW<\/code> and make a path to its <code>bin<\/code> directory. I fixed the problem with it.<\/p>\n"}]}
{"QuestionId":33151755,"AnswerCount":1,"Tags":"<theano><deep-learning>","CreationDate":"2015-10-15T14:51:49.797","AcceptedAnswerId":33152587.0,"OwnerUserId":4200891.0,"Title":"'length not know' in theano.function","Body":"<p>I want to make some changes of the logistic_sgd.py in DeepLearning Tutorial. Details are below.<\/p>\n\n<p>The original codes:<\/p>\n\n<pre><\/pre>\n\n<p>My codes:<\/p>\n\n<pre><\/pre>\n\n<p>I want to use the index of train_set_x and train_set_y from a vector <strong>idx_list<\/strong>, not the one in original codes <strong>index<\/strong>,but I got the following errors:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33152587","CreationDate":"2015-10-15T15:29:33.150","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you're mixing Python with symbolic Theano code in an unsupported way.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>x: train_set_x[[i for i in idx_list]],\ny: train_set_y[[i for i in idx_list]]\n<\/code><\/pre>\n\n<p>you need<\/p>\n\n<pre><code>x: train_set_x[idx_list],\ny: train_set_y[idx_list]\n<\/code><\/pre>\n\n<p>Here's a full example that demonstrates the change in a bit more detail:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as T\n\n\ndef v1(all_x):\n    batch_size = 3\n    index = T.lscalar()\n    x_part = T.vector()\n    f = theano.function(\n        inputs=[index],\n        outputs=x_part,\n        givens={\n            x_part: all_x[index * batch_size: (index + 1) * batch_size]\n        }\n    )\n    print f(1)\n\n\ndef v2_broken(all_x):\n    idx_list = T.lvector()\n    x_part = T.vector()\n    f = theano.function(\n        inputs=[idx_list],\n        outputs=x_part,\n        givens={\n            x_part: all_x[[i for i in idx_list]]\n        }\n    )\n    print f([2, 4, 6, 8])\n\n\ndef v2_fixed(all_x):\n    idx_list = T.lvector()\n    x_part = T.vector()\n    f = theano.function(\n        inputs=[idx_list],\n        outputs=x_part,\n        givens={\n            x_part: all_x[idx_list]\n        }\n    )\n    print f([2, 4, 6, 8])\n\n\ndef main():\n    all_x = theano.shared(-numpy.arange(10, dtype=theano.config.floatX))\n    v1(all_x)\n    # v2_broken(all_x)  # raises ValueError: length not known\n    v2_fixed(all_x)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":33152078,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-10-15T15:05:50.737","AcceptedAnswerId":33152770.0,"OwnerUserId":853462.0,"Title":"Create a shared row in theano","Body":"<p>I noticed that in theano, when one creates a shared variable based on 1D numpy array, this becomes a vector, but not a row:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The same goes for a 1xN matrix, it becomes a matrix but not a row:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This is troublesome when I want to add a M x N matrix to a 1 X N row-vector, because the shared vector is not broadcastable in the first dimension. First of all, this will not work:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>With the error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Ok, that's clear, we can't assign vectors to rows. Unfortunately, this is also not fine:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>With the error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>So we can't just use a 1 x N matrix as a row as well (because the first dimension of a 1 x N matrix is not broadcastable).<\/p>\n\n<p>The question remains, what 'can' we do? How can I create a shared variable of type row, such that is i broadcastable using matrix-row addition?<\/p>\n","answers":[{"AnswerId":"40809096","CreationDate":"2016-11-25T16:22:24.683","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>I would use:<\/p>\n\n<pre><code>shared_row = theano.shared(numpy.zeros((1,10,)), broadcastable=(True, False))\nprint(shared_row.type)\n# TensorType(float64, row)\nprint(shared_row.broadcastable)\n(True, False)\n<\/code><\/pre>\n"},{"AnswerId":"33152770","CreationDate":"2015-10-15T15:37:46.770","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>An alternative to using <code>reshape(1, N)<\/code> is to use <code>dimshuffle('x', 0)<\/code> as <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.dimshuffle\" rel=\"nofollow\">described in the documentation<\/a>.<\/p>\n\n<p>Here's a demo of the two approaches:<\/p>\n\n<pre><code>import numpy\nimport theano\n\nx = theano.shared(numpy.arange(10))\nprint x\nprint x.dimshuffle('x', 0).type\nprint x.dimshuffle(0, 'x').type\nprint x.reshape((1, x.shape[0])).type\nprint x.reshape((x.shape[0], 1)).type\n\nf = theano.function([], outputs=[x, x.dimshuffle('x', 0), x.reshape((1, x.shape[0]))])\ntheano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>&lt;TensorType(int32, vector)&gt;\nTensorType(int32, row)\nTensorType(int32, col)\nTensorType(int32, row)\nTensorType(int32, col)\nDeepCopyOp [@A] ''   2\n |&lt;TensorType(int32, vector)&gt; [@B]\nDeepCopyOp [@C] ''   4\n |InplaceDimShuffle{x,0} [@D] ''   1\n   |&lt;TensorType(int32, vector)&gt; [@B]\nDeepCopyOp [@E] ''   6\n |Reshape{2} [@F] ''   5\n   |&lt;TensorType(int32, vector)&gt; [@B]\n   |MakeVector{dtype='int64'} [@G] ''   3\n     |TensorConstant{1} [@H]\n     |Shape_i{0} [@I] ''   0\n       |&lt;TensorType(int32, vector)&gt; [@B]\n<\/code><\/pre>\n\n<p>Demonstrating that the <code>dimshuffle<\/code> is probably preferable as it involves less work than the <code>reshape<\/code>.<\/p>\n"}]}
{"QuestionId":33153362,"AnswerCount":1,"Tags":"<lua><torch><luarocks>","CreationDate":"2015-10-15T16:04:52.223","AcceptedAnswerId":33155461.0,"OwnerUserId":249226.0,"Title":"How to correctly require lua packages using interactive lua or torch session?","Body":"<p>I'm having difficulty installing and requiring packages in general. For example, for the inspect.lua package, I first install via luarocks as instructed in the package (<a href=\"https:\/\/github.com\/kikito\/inspect.lua\" rel=\"nofollow\">https:\/\/github.com\/kikito\/inspect.lua<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>Then if I start either  or torch7 (), I will then require it via:<\/p>\n\n<pre><\/pre>\n\n<p>This inspect variable is always :<\/p>\n\n<pre><\/pre>\n\n<p>returns .<\/p>\n\n<p>Initially, I wasn't sure that it was returning , so when I would try, for example  I would get the error \"attempt to call global 'inspect' (a nil value)\".<\/p>\n\n<p>Using torch, it seems like I can use \"import 'inspect'\" successfully, although I'm not sure why this works and require does not.<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"33155461","CreationDate":"2015-10-15T18:03:24.647","ParentId":null,"OwnerUserId":"752976","Title":null,"Body":"<p>Credit for spotting that goes to <strong>@siffiejoe<\/strong>.<\/p>\n\n<p>The Lua interpreter works in blocks. Every block is treated as a separate execution set. Thus, if you write:<\/p>\n\n<pre><code>local a = 5\nlocal b = a\n<\/code><\/pre>\n\n<p>In a file, it will correctly set <code>b<\/code> to be equal to 5, because a lua file is treated as one big block. In the REPL, though, after the first line local variables are purged.<\/p>\n\n<p>That simply means you either should force your code into one block:<\/p>\n\n<pre><code>do local inspect = require 'inspect'; print(inspect) end\n<\/code><\/pre>\n\n<p>Or use a global variable that persists across block executions:<\/p>\n\n<pre><code>$ inspect = require 'inspect'\n$ print(inspect)\n<\/code><\/pre>\n"}]}
{"QuestionId":33160611,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-10-16T00:19:50.257","AcceptedAnswerId":null,"OwnerUserId":3768533.0,"Title":"Torch DoubleTensor CharTensor Incompatibility","Body":"<p>Essentially how does one multiply a CharTensor mask with a DoubleTensor? I making a module for torch, it is supposed to be similar to the ReLU module. Therefore self.mask is supposed to be a mask of 0s and 1s that I want to hit the DoubleTensor self.output with. Currently self.mask is a CharTensor and I am not able to multiply the two as seen in the log. It seems so trivial but I have been looking online forever now, unable to find a solution.<\/p>\n\n<p>Infinite thanks,<\/p>\n\n<p>-an anxious coder<\/p>\n\n<p>function ReQU:updateOutput(input)<\/p>\n\n<p>...<\/p>\n\n<p>self.mask = torch.gt(input, 0)<\/p>\n\n<p>self.output:cmul(self.mask)<\/p>\n\n<p>...<\/p>\n\n<p>Log:\ninvalid arguments: DoubleTensor ByteTensor \nexpected arguments: <em>DoubleTensor<\/em> [DoubleTensor] DoubleTensor\nstack traceback:\n    [C]: in function 'cmul'\n    .\/requ.lua:21: in function 'forward'<\/p>\n","answers":[{"AnswerId":"33175443","CreationDate":"2015-10-16T16:36:57.190","ParentId":null,"OwnerUserId":"1417179","Title":null,"Body":"<p>Torch exposes methods to do this without the need to multiply values by yourself. The simplest is probably to invert your mask and use the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\">index <code>[]<\/code> operator<\/a>:<\/p>\n\n<pre><code>self.mask = torch.le(input, 0)\nself.output[self.mask] = 0\n<\/code><\/pre>\n"}]}
{"QuestionId":33174142,"AnswerCount":1,"Tags":"<c++><opencv><caffe><lmdb>","CreationDate":"2015-10-16T15:25:57.567","AcceptedAnswerId":33180023.0,"OwnerUserId":1179925.0,"Title":"Caffe+Opencv without lmdb","Body":"<p>When using caffe, to create training dataset containing images we need to create database in special format like lmdb, but there is any option to pass to caffe batch of images as for example  ?<\/p>\n\n<p>To clarify I'm looking for solution that can handle large amount of images that can't fit into memory (but assume that one training batch (containing for example 50 images) can be stored in memory).<\/p>\n","answers":[{"AnswerId":"33180023","CreationDate":"2015-10-16T21:44:24.683","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Caffe can take many types of inputs, depending upon the input layer that we use. \nSome of the input methods that are available are:<\/p>\n\n<ol>\n<li>Data<\/li>\n<li>MemoryData<\/li>\n<li>HDF5Data <\/li>\n<li>ImageData etc.<\/li>\n<\/ol>\n\n<p>In the model file, the very first layer that you find will be <code>Layer type: Data<\/code>, which used <code>lmdb or leveldb<\/code> as input method. The conversion of a set of images to these databases are pretty easy as Caffe already provides the tools to convert the images.<\/p>\n\n<p>The <code>Layer type: MemoryData<\/code> reads data directly from memory, which will be extremely helpful while using camera inputs to be passed as Caffe input during Test phase. Using this layer for training is highly <strong>not<\/strong> recommended.<\/p>\n\n<p>The <code>Layer type: ImageData<\/code> takes a text file as input. The text file contains all the image names along with their complete path and the class number. Caffe uses OpenCV to read the images in this layer. It also takes care of all the transformations to the image. Thus instead of using OpenCV to read the image and then pass to MemoryData layer, use of ImageData is recommended.<\/p>\n\n<p>The format of the .txt from which ImageData layer reads the image must be:<\/p>\n\n<blockquote>\n  <p>\/path\/to\/the\/image\/imageName.jpg  classNumber<\/p>\n<\/blockquote>\n\n<p>Use of LMDB or LevelDB is highly recommended because, ImageData needn't work well if the image path or name contains spaces, or when any of the images are corrupt.<\/p>\n\n<p>Details of various layers can be found out <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Memory is allocated in GPU depending upon the model and batch size. If memory overflow occurs, you could try reducing the batch size. Caffe easily handled training the Imagenet database of 1.2million images. Thus with an optimal batch size, the algorithm should work without any issues.<\/p>\n"}]}
{"QuestionId":33199665,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-10-18T15:20:47.363","AcceptedAnswerId":null,"OwnerUserId":987599.0,"Title":"Caffe mean file creation without database","Body":"<p>I run caffe using an  and don't want to create an LMDB or LevelDB for the data, But The <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/compute_image_mean.cpp\" rel=\"nofollow\"><\/a> tool only works with LMDB\/LevelDB databases.<\/p>\n\n<p>Is there a simple solution for creating a mean file from a list of files (the same format that  is using)?<\/p>\n","answers":[{"AnswerId":"33207665","CreationDate":"2015-10-19T06:10:49.453","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You may notice that recent models (e.g., googlenet) do not use a mean file the same size as the input image, but rather a 3-vector representing a mean value per image <em>channel<\/em>. These values are quite \"immune\" to the specific dataset used (as long as it is large enough and contains \"natural images\").<br>\nSo, as long as you are working with natural images you may use the same values as e.g., GoogLenet is using: B=104, G=117, R=123.<\/p>\n"},{"AnswerId":"33201966","CreationDate":"2015-10-18T18:59:15.600","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>The simplest solution is to create a LMDB or LevelDB database of the image set. <\/p>\n\n<p>The complicated solution is to write a tool similar to compute_image_mean, which takes image inputs and do the transformations and find the mean!<\/p>\n"}]}
{"QuestionId":33208282,"AnswerCount":2,"Tags":"<computer-vision><neural-network><caffe>","CreationDate":"2015-10-19T06:51:52.443","AcceptedAnswerId":33216247.0,"OwnerUserId":639973.0,"Title":"Extracting Features from VGG","Body":"<p>I want to extract features from images in MS COCO dataset using a fine-tuned VGG-19 network.<\/p>\n\n<p>However, it takes about 6~7 seconds per image, roughly 2 hours per 1k images. (even longer for other fine-tuned models)<\/p>\n\n<p>There are 120k images in MS COCO dataset, so it'll take at least 10 days.<\/p>\n\n<p>Is there any way that I can speed up the feature extraction process?<\/p>\n","answers":[{"AnswerId":"33216247","CreationDate":"2015-10-19T13:51:17.103","ParentId":null,"OwnerUserId":"2832033","Title":null,"Body":"<p>Well, this is not just a command. First you must check whether your GPU is powerful enough to wrestle with deep CNNs. Knowing your GPU model can answer this question.<\/p>\n\n<p>Second, you have to compile and build Caffe framework with CUDA and GPU-enabled (CPU_Only disabled) in the Makefile.config (or CMakeLists.txt).<\/p>\n\n<p>Passing all required steps (installing Nvidia Driver, installing CUDA and etc.) you can build caffe for GPU-use. Then by passing the GPU_Device_ID in your command-line you can benefit from speed provided by them.<\/p>\n\n<p>Follow <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Install-Caffe-on-EC2-from-scratch-(Ubuntu,-CUDA-7,-cuDNN)\" rel=\"nofollow\">this<\/a> link for building Caffe using GPU.<\/p>\n\n<p>Hope it helps<\/p>\n"},{"AnswerId":"33229199","CreationDate":"2015-10-20T06:08:20.733","ParentId":null,"OwnerUserId":"325528","Title":null,"Body":"<p>This ipython notebook example explains the steps to extract features out of any caffe model really well: <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a><\/p>\n\n<p>In pycaffe, you can set gpu mode simply by using caffe.set_mode_gpu().<\/p>\n"}]}
{"QuestionId":33215122,"AnswerCount":1,"Tags":"<neural-network><hdf5><deep-learning><caffe><pycaffe>","CreationDate":"2015-10-19T13:00:50.847","AcceptedAnswerId":null,"OwnerUserId":824624.0,"Title":"how to feed the image data to HDF5 on caffe or existing examples?","Body":"<p>I had hard time working on caffe with HDF5 on the image classification and regression tasks, for some reason, the training on HDF5 will always fail at the first beginning that the test and train loss could very soon drop to close to zero. after trying all the tricks such reducing the learning rate, adding RELU, dropout, nothing started to work, so I started to doubt that the HDF5 data I am feeding to caffe  is wrong.   <\/p>\n\n<p>so currently I am working on the universal dataset (<a href=\"https:\/\/gist.github.com\/jimgoo\/0179e52305ca768a601f\" rel=\"nofollow\">Oxford 102 category flower dataset<\/a> and also it has <a href=\"https:\/\/github.com\/jimgoo\/caffe-oxford102\" rel=\"nofollow\">public code<\/a> ), firstly I started out by trying ImageData and LMDB layer for the classification, they all worked very well.  at last i used HDF5 data layer for the finetuning, the training_prototxt doesn't change unless on the data layer which uses HDF5 instead. and again, at the start of the learning, the loss drops from 5 to 0.14 at iteration 60, 0.00146 at iteration 100, that seems to prove that HDF5 data is incorrect.<\/p>\n\n<p>i have two image&amp;label to HDF5 snippet on the <a href=\"https:\/\/gist.github.com\/helxsz\/110280fe6e0f227a4ba1\" rel=\"nofollow\">github<\/a>, all of them seem to generate the HDF5 dataset, but for some reason these dataset doesn't seem to be not working with caffe<\/p>\n\n<p>I wonder anything wrong with this data, or anything that makes this example run in HDF5 or if you have some HDF5 examples for classification or regression, which can be helpful to me a lot.<\/p>\n\n<p>one snippet is shown as <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33774732","CreationDate":"2015-11-18T08:12:05.047","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>One input transformation that seems to happen in the original net and is missing from your HDF5 creation in <strong>mean subtraction<\/strong>.<br>\nYou should obtain mean file (looks like <a href=\"https:\/\/gist.github.com\/jimgoo\/0179e52305ca768a601f#file-train_val-prototxt-L15\" rel=\"nofollow\"><code>\"imagenet_mean.binaryproto\"<\/code><\/a> in your example), read it into python and subtract it from each image.<\/p>\n\n<p>BTW, the mean file can give you a clue as to the scale of the input image (if pixel values should be in [0..1] range or [0..255]).<\/p>\n\n<p>You might find <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L18\" rel=\"nofollow\"><code>caffe.io<\/code><\/a> useful converting binaryproto to numpy array.<\/p>\n"}]}
{"QuestionId":33219187,"AnswerCount":1,"Tags":"<python><function><theano>","CreationDate":"2015-10-19T16:12:22.917","AcceptedAnswerId":33229286.0,"OwnerUserId":5175919.0,"Title":"I want to understand theano function(given,updates)","Body":"<p>I would like to ask a few questions related to theano function.<\/p>\n\n<p>1.I saw the script in which input variable was not assigned. If so, how it is working?<\/p>\n\n<pre><\/pre>\n\n<p>As you can see above, input square bracket is empty. What is input in this case then?<\/p>\n\n<p>2.When it comes to 'given' parameter in function, it is a little bit difficult to understand. People say this is for boosting GPU process, but I would like to know what exactly variables should be assigned for 'given'.\nPlease look at the script below.<\/p>\n\n<pre><\/pre>\n\n<p>In the given, what x and y do? what colon(:) means? \nAs far as I know,(theano homepage says: givens (iterable over pairs (Var1, Var2) of Variables. List, tuple or dict. The Var1 and Var2 in each pair must have the same Type.) \u2013 specific substitutions to make in the computation graph (Var2 replaces Var1).) it needs 2 variables but it seems 4 variables in 1st example and 2nd one is pretty complicated to me. Could anyone tell me what given does with specific details? And explain what's going on the 2nd script in given variables.<\/p>\n\n<p>Furthermore, in the homepage, it says 'you can use the givens parameter of function which replaces a particular node in a graph for the purpose of one particular function.' I don't get what parpticular node in a graph it replaces. <\/p>\n\n<p>Please help me!!<\/p>\n","answers":[{"AnswerId":"33229286","CreationDate":"2015-10-20T06:14:29.770","ParentId":null,"OwnerUserId":"5417824","Title":null,"Body":"<p>Firstly, given means what is the real value of the T variable . IF in the givens there is (x,a) and a is a np.array then it will use a to replace x when it is computed, also in the given if there is x:a, it means the same thing. \nIn the first example, data_X is the input, just not given in the first parameter but in givens, which is the same.\nIn the second example, test_set_x is theano.shared which means its value is already exist, it's a matrix. But which part of it will you use, you will choose by index which is the input.<\/p>\n"}]}
{"QuestionId":33227747,"AnswerCount":1,"Tags":"<python-2.7><neural-network><artificial-intelligence><theano><lstm>","CreationDate":"2015-10-20T03:59:35.713","AcceptedAnswerId":33228279.0,"OwnerUserId":4075155.0,"Title":"cPickle.load() doesnt accept non-.gz files, what can I use for .pkl files?","Body":"<p>I am trying to run an example of a LSTM recurrent neural network that is presented in this git: <a href=\"https:\/\/github.com\/mesnilgr\/is13\" rel=\"nofollow\">https:\/\/github.com\/mesnilgr\/is13<\/a>. <\/p>\n\n<p>I've installed theano and everything and when I got to the point of running the code, I've noticed the data was not being downloaded, so I've opened an issue on the github (<a href=\"https:\/\/github.com\/mesnilgr\/is13\/issues\/12\" rel=\"nofollow\">https:\/\/github.com\/mesnilgr\/is13\/issues\/12<\/a>) and this guy came up with a solution that consisted in:<\/p>\n\n<p>1-get the data from the dropbox link he provides.\n2- change the code of the 'load.py' file to download, and read the data properly.<\/p>\n\n<p>The only issue is that the data in the dropbox folder(<a href=\"https:\/\/www.dropbox.com\/s\/3lxl9jsbw0j7h8a\/atis.pkl?dl=0\" rel=\"nofollow\">https:\/\/www.dropbox.com\/s\/3lxl9jsbw0j7h8a\/atis.pkl?dl=0<\/a>) is not a compacted .gz file as, I suppose, was the data from the original repository. So I dont have enough skill to change the code in order to do with the uncompressed data exaclty what it would do with the compressed one. Can someone help me? <\/p>\n\n<p>The modification suggested and the changes I've done are described on the issue I've opened on the git(<a href=\"https:\/\/github.com\/mesnilgr\/is13\/issues\/12\" rel=\"nofollow\">https:\/\/github.com\/mesnilgr\/is13\/issues\/12<\/a>).<\/p>\n","answers":[{"AnswerId":"33228279","CreationDate":"2015-10-20T04:54:59.407","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It looks like your code is using <\/p>\n\n<pre><code>gzip.open(...)\n<\/code><\/pre>\n\n<p>But if the file is not gzipped then you probably just need to remove the <code>gzip.<\/code> prefix and use<\/p>\n\n<pre><code>open(...)\n<\/code><\/pre>\n"}]}
{"QuestionId":33228775,"AnswerCount":1,"Tags":"<deep-learning><caffe><lmdb><pycaffe>","CreationDate":"2015-10-20T05:37:06.720","AcceptedAnswerId":null,"OwnerUserId":1864142.0,"Title":"Multi-label classification with Caffe","Body":"<p>I want to fine-tune GoogLeNet to do Multi-label classification with Caffe. I have already fine-tuned it to a single-label classification, but I can't make the transition to Multi-label yet. <\/p>\n\n<p>The main steps I am doing which are different:<\/p>\n\n<h1>Create LMDB for Data &amp; Ground Truth<\/h1>\n\n<p>I am modifying the code <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issuecomment-70211045\" rel=\"noreferrer\">here<\/a> and <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/caffe-users\/lmdb$20python\/caffe-users\/-vWuaM3bnro\/RB4xzpI8xeUJ\" rel=\"noreferrer\">here<\/a> to create one LMDB with the Data and the other with ground truth.<\/p>\n\n<h1>Replacing SoftmaxWithLoss with SigmoidCrossEntropyLoss<\/h1>\n\n<p>Updating the train_val.prototxt, I replace SoftmaxWithLoss layers to SigmoidCrossEntropyLoss, and set the data layers so that both DB's are loaded. I set the learning rate parameters as I have done with the single-label classification problem.<\/p>\n\n<p>This steps seems to be working. The data flow, and it is possible to perform solver.step(1). To verify the data and labels are loaded right, I have explicitly calculated the loss using the formula, and got the same result as Caffe.<\/p>\n\n<h1>Problem<\/h1>\n\n<p>The network does not converge. Running it several hundrads of iterations, each of the different classes averages around the class population. That is if class a has 0.35 1's and 0.65 0's in the population, the network will converge to ~0.35 classification probability for each observation, regardless of the true label. <\/p>\n\n<h1>Possible error 1<\/h1>\n\n<p>I suspect the problem is because I fail to load the images correctly into caffe in a way that GoogLeNet pretrained model can learn from them. My previous experience so far is convert_imageset which works perfectly. Right now I am using shelhamer code to save the images into the LMDB:<\/p>\n\n<pre><\/pre>\n\n<p>I normalize the mean in the data layer when loading the image.\nDoes that seem right? Is there another way to do that?<\/p>\n\n<h1>Possible error 2<\/h1>\n\n<p>It might also be that the train_val.prototxt has been defined wrong. Is there any thing else that need to be done than switching the SoftmaxWithLoss -> SigmoidCrossEntropyLoss?<\/p>\n\n<p>Any help will be greatly appreciated!\nThanks!<\/p>\n","answers":[{"AnswerId":"48751389","CreationDate":"2018-02-12T16:43:05.917","ParentId":null,"OwnerUserId":"5954475","Title":null,"Body":"<p>In GoogLeNet input data should be subtracted by mean:<\/p>\n\n<pre><code>...\nim = im.transpose((2,0,1))\nmean = np.array((104,117,123))\nim -= mean\nim_dat = caffe.io.array_to_datum(im)\n...\n<\/code><\/pre>\n"}]}
{"QuestionId":33229222,"AnswerCount":1,"Tags":"<theano><conv-neural-network>","CreationDate":"2015-10-20T06:09:59.143","AcceptedAnswerId":33229989.0,"OwnerUserId":3890207.0,"Title":"How to initialize weights when using RELU activation function","Body":"<p>I want to make a Conv network and I wish to use the RELU activation function. Can someone please give me a clue of the correct way to initialize weights (I'm using Theano)<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"33229989","CreationDate":"2015-10-20T06:58:56.443","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I'm not sure there is a hard and fast best way to initialize weights and bias for a ReLU layer.<\/p>\n\n<p>Some <a href=\"http:\/\/deepdish.io\/2015\/02\/24\/network-initialization\/\" rel=\"noreferrer\">claim<\/a> that (a slightly modified version of) Xavier initialization works well with ReLUs. Others that small Gaussian random weights plus bias=1 (ensuring the weighted sum of positive inputs will remain positive and thus not end up in the ReLUs zero region).<\/p>\n\n<p>In Theano, these can be achieved like this (assuming weights post-multiply the input):<\/p>\n\n<pre><code>w = theano.shared((numpy.random.randn((in_size, out_size)) * 0.1).astype(theano.config.floatX))\nb = theano.shared(numpy.ones(out_size))\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>w = theano.shared((numpy.random.randn((in_size, out_size)) * tt.sqrt(2 \/ (in_size + out_size))).astype(theano.config.floatX))\nb = theano.shared(numpy.zeros(out_size))\n<\/code><\/pre>\n"}]}
{"QuestionId":33229504,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-20T06:29:38.307","AcceptedAnswerId":33229569.0,"OwnerUserId":669482.0,"Title":"How to get a theano function to return the an array of the same length as another tensor variable","Body":"<p>I am really new to Theano, and I am just trying to figure out some basic functionality. I have a tensor variable x, and i would like the functio to return a tensor variable y of the same shape, but filled with value 0.2. I am not sure how to define y.<\/p>\n\n<p>For example if x = [1,2,3,4,5], then I would like y = [0,2, 0,2, 0,2, 0,2, 0.2]<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33229569","CreationDate":"2015-10-20T06:34:07.643","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There's probably a dozen different ways to do this and which is best will depend on the context: how this piece of code\/functionality fits into the wider program.<\/p>\n\n<p>Here's one approach:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\nx = tt.vector()\ny = tt.ones_like(x) * 0.2\nf = theano.function([x], outputs=y)\nprint f([1, 2, 3, 4, 5])\n<\/code><\/pre>\n"}]}
{"QuestionId":33229748,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-10-20T06:45:13.550","AcceptedAnswerId":33230107.0,"OwnerUserId":3214345.0,"Title":"Why do I get a Theano TypeError when trying to update a shared variable?","Body":"<p>I'm trying to run a very simple gradient descent on the function y=x^2.\nI've tried implementing it using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>But when I try compile the function \"fn\", I get the following error: <\/p>\n\n<pre><\/pre>\n\n<p>I thought this may be a problem with the learning_rate variable, since it might not be the same type as the x shared variable, but if I modify the code as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I still get the same error.<\/p>\n\n<p>I'm stuck :( Any ideas?<\/p>\n","answers":[{"AnswerId":"33230107","CreationDate":"2015-10-20T07:05:32.760","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that your shared variable <code>x<\/code> has no type specified so one is being inferred. Since the value you provide is a Python integer literal, the type is assumed to be <code>int32<\/code>. This is a problem because gradients don't work well with integers and so <code>dy_dx<\/code> is actually a <code>float64<\/code>. This in turn makes the update value a <code>float64<\/code> too. A shared variable can only be updated with a value of the same type (this is the error message) so you have a problem: the shared variable is an <code>int32<\/code> but the update is a <code>float64<\/code>.<\/p>\n\n<p>One solution is to make the shared variable a float as well. This can be achieved by simply adding a decimal point to the initial value of <code>x<\/code>.<\/p>\n\n<pre><code>x = theano.shared(2.)\n<\/code><\/pre>\n"}]}
{"QuestionId":33237703,"AnswerCount":1,"Tags":"<convolution><caffe>","CreationDate":"2015-10-20T13:16:11.127","AcceptedAnswerId":33543982.0,"OwnerUserId":1391376.0,"Title":"Deconvolution with caffe","Body":"<p>I was wondering if it is possible to perform a deconvolution of images in Caffe using a point spread function of objective at a given focal point. Something along the lines of <a href=\"http:\/\/yuzhikov.com\/articles\/BlurredImagesRestoration1.htm\" rel=\"nofollow\">this approach<\/a>.<\/p>\n\n<p>If yes, what would be the best way to proceed?<\/p>\n","answers":[{"AnswerId":"33543982","CreationDate":"2015-11-05T11:59:30.800","ParentId":null,"OwnerUserId":"5170044","Title":null,"Body":"<p>It is possible to deconvolve images using Caffe (and CNN in general), but the approach may not be as general as you hope it to be.<\/p>\n\n<p>CNNs can take blurry image as an input and output sharp image. As the networks are convolutional, the input can be of any size. This can be easily done in Caffe using Convolutional layers and Euclidean Loss layer. Optionally, you can experiment with adding some pooling and deconvolution layers.<\/p>\n\n<p>CNNs can be trained to deconvolve images for specific blur PSF as in your link. (see: <a href=\"http:\/\/papers.nips.cc\/paper\/5485-deep-convolutional-neural-network-for-image-deconvolution.pdf\" rel=\"nofollow\">[Xu et al.:Deep Convolutional Neural Network for Image Deconvolution. NIPS 2014]<\/a>). This works well but you have to re-train the CNN for each new PSF (which takes lot of time).<\/p>\n\n<p>I've tried to train CNNs to do blind deconvolution (PSF is not known) and it works very well for text documents. You can get trained nets and python-Caffe scripts at <a href=\"http:\/\/www.fit.vutbr.cz\/~ihradis\/CNN-Deblur\/\" rel=\"nofollow\">[Hradi\u0161 et al.: Convolutional Neural Networks for Direct Text Deblurring. BMVC 2015]<\/a>. This approach could work for other types of images, but it would not work for unrestricted photographs and diverse blurs. For general photos, I would guess It could work for small range of blurs. <\/p>\n\n<p>Another possibility is to do inverse filtration (e.g. using Wiener filter) and process the output using a CNN. The advantage of this is that you can compute the inverse filter for new PSF very fast and the CNN stays the same. [Schuler et al.: A machine learning approach for non-blind image deconvolution. CVPR 2013]<\/p>\n"}]}
{"QuestionId":33246553,"AnswerCount":0,"Tags":"<python-3.x><hdf5><deep-learning><caffe>","CreationDate":"2015-10-20T20:50:52.033","AcceptedAnswerId":null,"OwnerUserId":5331481.0,"Title":"libhdf5_hl.so.9 cannot open shared object file","Body":"<p>While running  file ,I am getting:<\/p>\n\n<p> <\/p>\n\n<p>It's very difficult for me to understand this error. I thought this error was just because of anaconda so I updated anaconda.. still getting the same error.. I don't know why this error is coming. please help me. Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":33246681,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><feature-extraction><deep-learning><caffe>","CreationDate":"2015-10-20T20:59:20.810","AcceptedAnswerId":null,"OwnerUserId":3230626.0,"Title":"Using Caffe to classify \"hand-crafted\" image features","Body":"<p>Does it make any sense to perform feature extraction on images using, e.g., OpenCV, then use Caffe for classification of those features?<\/p>\n\n<p>I am asking this as opposed to the traditional way of passing the images directly to Caffe, and letting Caffe do the extraction and classification procedures.<\/p>\n","answers":[{"AnswerId":"33252374","CreationDate":"2015-10-21T06:28:23.980","ParentId":null,"OwnerUserId":"997572","Title":null,"Body":"<p>Yes, it does make sense, but it may not be the first thing you want to try:\nIf you have already extracted hand-crafted features that are suitable for your domain, there is a good chance you'll get satisfactory results by using an easier-to-use machine learning tool (e.g. libsvm).<\/p>\n\n<p>Caffe can be used in many different ways with your features. If they are low-level features (e.g. Histogram of Gradients), then several convolutional layers may be able to extract the appropriate mid-level features for your problem. You may also use caffe as an alternative non-linear classifier (instead of SVM). You have the freedom to try (too) many things, but my advice is to first try a machine learning method with a smaller meta-parameter space, especially if you're new to neural nets and caffe.<\/p>\n"},{"AnswerId":"33251841","CreationDate":"2015-10-21T05:49:00.167","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe is a tool for training and evaluating deep neural networks. It is quite a versatile tool allowing for both deep convolutional nets as well as other architectures.<br>\nOf course it can be used to process pre-computed image features. <\/p>\n"}]}
{"QuestionId":33247682,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-10-20T22:09:19.407","AcceptedAnswerId":null,"OwnerUserId":1545917.0,"Title":"Update a slice of tensor in torch","Body":"<p>How to slice tensor's rows in  like we do in  with  ?<\/p>\n\n<p>Example of code that I want to run:<\/p>\n\n<pre><\/pre>\n\n<p>Expected value for <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33248011","CreationDate":"2015-10-20T22:36:55.643","ParentId":null,"OwnerUserId":"1545917","Title":null,"Body":"<p>I had to use brackets <code>m[]<\/code> instead of parenthesis<\/p>\n\n<p>Rest of the code is fine.<\/p>\n"}]}
{"QuestionId":33247800,"AnswerCount":1,"Tags":"<compilation><cuda><nvcc><luajit><torch>","CreationDate":"2015-10-20T22:20:10.860","AcceptedAnswerId":33249861.0,"OwnerUserId":3431929.0,"Title":"Unable to find Torch header files using nvcc","Body":"<p>I'm trying to call some CUDA code from luaJIT (Torch) but I'm running into compiling issues. nvcc seems unable to find my Torch header files. I have CUDA 6.5 and gcc 4.4.7.<\/p>\n\n<pre><\/pre>\n\n<p>im2col.cu includes the following<\/p>\n\n<pre><\/pre>\n\n<p>where \"utils.h\" is<\/p>\n\n<pre><\/pre>\n\n<p>This is relatively weird since the mentioned files are indeed in the include location I gave to the compiler..<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas about what I'm doing wrong?<\/p>\n\n<p>Thx in advance!<\/p>\n","answers":[{"AnswerId":"33249861","CreationDate":"2015-10-21T02:17:26.993","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>It seems that this compile command:<\/p>\n\n<pre><code>nvcc -o im2col -I\/deep\/u\/ibello\/torch\/include im2col.cu \n<\/code><\/pre>\n\n<p>did not give the necessary search paths for the compiler to find the header files like <code>THGeneral.h<\/code> that were located in <code>\/deep\/u\/ibello\/torch\/include\/TH<\/code><\/p>\n\n<p>The solution was to specify a compile command like this:<\/p>\n\n<pre><code>nvcc -o im2col -I\/deep\/u\/ibello\/torch\/include -I\/deep\/u\/ibello\/torch\/include\/TH im2col.cu \n<\/code><\/pre>\n"}]}
{"QuestionId":33248959,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-21T00:20:50.863","AcceptedAnswerId":33255235.0,"OwnerUserId":5205141.0,"Title":"Efficient way of loading minibatches < gpu memory","Body":"<p>I have the following scenario:<\/p>\n\n<ul>\n<li>My dataset >> gpu memory<\/li>\n<li>My minibatches &lt; gpu memory ... such that depending on size I can fit up to 10 in memory at once while still training no problem.<\/li>\n<\/ul>\n\n<p>The size of my dataset means I won't revisit datapoints, so I guess no point in making them shared?  Or is there?  I was thinking that maybe it would be beneficial to have up to 10 shared initialised variables of size=mini-batch, such that I can I swap 10 in at once instead of just one at a a time.  Also, is it possible to preload mini-batches in parallel?<\/p>\n","answers":[{"AnswerId":"33255235","CreationDate":"2015-10-21T09:10:31.437","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you're not revisiting datapoints then there probably isn't any value in using shared variables.<\/p>\n\n<p>The following code could be modified and used to evaluate the different methods of getting data into your specific computation.<\/p>\n\n<p>The \"input\" method is the one that will probably be best when you have no need to revisit data. The \"shared_all\" method may outperform everything else but only if you can fit the entire dataset in GPU memory. The \"shared_batched\" allows you to evaluate whether hierarchically batching your data could help.<\/p>\n\n<p>In the \"shared_batched\" method, the dataset is divided into many macro batches and each macro batch is divided into many micro batches. A single shared variable is used to hold a single macro batch. The code evaluates all the micro batches within the current macro batch. Once a complete macro batch has been processed the next macro batch is loaded into the shared variable and the code iterates over the micro batches within it again.<\/p>\n\n<p>In general, it might be expected that small numbers of large memory transfers will operate faster than larger numbers of smaller transfers (where the total transfered is the same for each). But this needs to be tested (e.g. with the code below) before it can be known for sure; YMMV.<\/p>\n\n<p>The use of the \"borrow\" parameter may also have a significant impact on the performance, but be aware of <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/aliasing.html\" rel=\"nofollow\">the implications<\/a> before using it.<\/p>\n\n<pre><code>import math\nimport timeit\nimport numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef test_input(data, batch_size):\n    assert data.shape[0] % batch_size == 0\n    batch_count = data.shape[0] \/ batch_size\n    x = tt.tensor4()\n    f = theano.function([x], outputs=x.sum())\n    total = 0.\n    start = timeit.default_timer()\n    for batch_index in xrange(batch_count):\n        total += f(data[batch_index * batch_size: (batch_index + 1) * batch_size])\n    print 'IN\\tNA\\t%s\\t%s\\t%s\\t%s' % (batch_size, batch_size, timeit.default_timer() - start, total)\n\n\ndef test_shared_all(data, batch_size):\n    batch_count = data.shape[0] \/ batch_size\n    for borrow in (True, False):\n        start = timeit.default_timer()\n        all = theano.shared(data, borrow=borrow)\n        load_time = timeit.default_timer() - start\n        x = tt.tensor4()\n        i = tt.lscalar()\n        f = theano.function([i], outputs=x.sum(), givens={x: all[i * batch_size:(i + 1) * batch_size]})\n        total = 0.\n        start = timeit.default_timer()\n        for batch_index in xrange(batch_count):\n            total += f(batch_index)\n        print 'SA\\t%s\\t%s\\t%s\\t%s\\t%s' % (\n            borrow, batch_size, batch_size, load_time + timeit.default_timer() - start, total)\n\n\ndef test_shared_batched(data, macro_batch_size, micro_batch_size):\n    assert data.shape[0] % macro_batch_size == 0\n    assert macro_batch_size % micro_batch_size == 0\n    macro_batch_count = data.shape[0] \/ macro_batch_size\n    micro_batch_count = macro_batch_size \/ micro_batch_size\n    macro_batch = theano.shared(numpy.empty((macro_batch_size,) + data.shape[1:], dtype=theano.config.floatX),\n                                borrow=True)\n    x = tt.tensor4()\n    i = tt.lscalar()\n    f = theano.function([i], outputs=x.sum(), givens={x: macro_batch[i * micro_batch_size:(i + 1) * micro_batch_size]})\n    for borrow in (True, False):\n        total = 0.\n        start = timeit.default_timer()\n        for macro_batch_index in xrange(macro_batch_count):\n            macro_batch.set_value(\n                data[macro_batch_index * macro_batch_size: (macro_batch_index + 1) * macro_batch_size], borrow=borrow)\n            for micro_batch_index in xrange(micro_batch_count):\n                total += f(micro_batch_index)\n        print 'SB\\t%s\\t%s\\t%s\\t%s\\t%s' % (\n            borrow, macro_batch_size, micro_batch_size, timeit.default_timer() - start, total)\n\n\ndef main():\n    numpy.random.seed(1)\n\n    shape = (20000, 3, 32, 32)\n\n    print 'Creating random data with shape', shape\n    data = numpy.random.standard_normal(size=shape).astype(theano.config.floatX)\n\n    print 'Running tests'\n    for macro_batch_size in (shape[0] \/ pow(10, i) for i in xrange(int(math.log(shape[0], 10)))):\n        test_shared_all(data, macro_batch_size)\n        test_input(data, macro_batch_size)\n        for micro_batch_size in (macro_batch_size \/ pow(10, i) for i in\n                                 xrange(int(math.log(macro_batch_size, 10)) + 1)):\n            test_shared_batched(data, macro_batch_size, micro_batch_size)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":33252617,"AnswerCount":1,"Tags":"<python><cuda><theano>","CreationDate":"2015-10-21T06:43:29.413","AcceptedAnswerId":33253026.0,"OwnerUserId":5412601.0,"Title":"Trying to install Theano but don't have an Nvidia card","Body":"<p>I am following:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a><\/p>\n\n<p>to install theano. I just want to play with the code, I don't need to use a GPU to improve my speed.<\/p>\n\n<p>I don't have an Nvidia card and when I try to install cuda, the installation fails. I watch as the installation tool deletes the files I need.<\/p>\n\n<p>I am using Anaconda python so I commented this line:<\/p>\n\n<pre><\/pre>\n\n<p>In the:<\/p>\n\n<blockquote>\n  <p>C:\\SciSoft\\env.bat<\/p>\n<\/blockquote>\n\n<p>file. I gave up and tried to install theano with easy_install.\nI try to import Theano from python, it fails with:<\/p>\n\n<blockquote>\n  <p><strong><em>ton of stuff<\/em><\/strong><br>\n  Problem occurred during compilation with the command line below: C:\\SciSoft\\TDM-GCC-64\\bin\\g++.exe -shared -g\n  -march=bdver2 -mmmx -mno-3dnow -mss<br>\n  <strong><em>more stuff<\/em><\/strong><br>\n  C:\\Users\\xxx\\Anaconda\\libs\/python27.lib: error adding symbols: File in wr ong format collect2.exe: error: ld returned\n  1 exit status<\/p>\n  \n  <p>--------------------------------------------------------------------------- Exception                                 Traceback (most recent call\n  last)  in ()\n  ----> 1 import theano<\/p>\n  \n  <p>C:\\Users\\xxx\\Anaconda\\lib\\site-packages\\theano-0.7.0-py2.7.egg\\theano__i\n  nit__.pyc in ()<\/p>\n  \n  <p><strong><em>Even more stuff<\/em><\/strong><br>\n  Exception: Compilation failed (return status=1): C:\\Users\\xxxx\\Anaconda\\li . collect2.exe: error: ld\n  returned 1 exit statusrong format<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"33253026","CreationDate":"2015-10-21T07:09:34.240","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you don't need to run on GPU then don't worry about installing Visual Studio or CUDA. I think you just need Anaconda, but maybe also TDM GCC.<\/p>\n\n<p>From a clean environment I install the latest version of Anaconda then run<\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n\n<p>I'd recommend <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">installing Theano from Github<\/a> (the bleeding edge version) since the \"stable\" release is not updated often and there are usually many significant improvements (especially for performance) in the bleeding edge version in comparison to the stable version.<\/p>\n\n<p>There is no need to perform all the steps in the \"Configuring the Environment\" section; just make sure your C++ compiler is in the PATH.<\/p>\n\n<p>If Theano fails to work after these minimal installation instructions, I'd recommend solving each problem on a case-by-case basis instead of trying to run the full installation instructions provided in the documentation (which may be out of date).<\/p>\n"}]}
{"QuestionId":33262803,"AnswerCount":2,"Tags":"<lua><torch>","CreationDate":"2015-10-21T15:03:16.220","AcceptedAnswerId":33263453.0,"OwnerUserId":1366749.0,"Title":"How do I use torch's class system to create a custom class","Body":"<p>I was attempting to follow the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/utility.md\" rel=\"nofollow\">torch documentation for utility functions<\/a>.<\/p>\n\n<p>And I did the following:<\/p>\n\n<pre><\/pre>\n\n<p>But I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I was expecting the  function to somehow work via the  metatable mechanism, but Blah doesn't even seem to have a metatable:<\/p>\n\n<pre><\/pre>\n\n<p>Maybe the documentation is out of date? But torch seems to be creating plenty of classes this way internally.<\/p>\n\n<p>I just updated to the latest torch, so I know it's not my torch version is too old...<\/p>\n\n<p>Thoughts?<\/p>\n","answers":[{"AnswerId":"33263453","CreationDate":"2015-10-21T15:34:30.693","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>do\n local Blah = torch.class('Blah')\n function Blah:__init() end\nend\n\nblah = Blah()\n<\/code><\/pre>\n"},{"AnswerId":"45123611","CreationDate":"2017-07-15T23:22:56.503","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p><strong>1.<\/strong> You need:<\/p>\n\n<p><code>local Blah = torch.class('Blah')<\/code> <\/p>\n\n<p><strong>2.<\/strong> You need use:<\/p>\n\n<pre><code>do \nend \n<\/code><\/pre>\n\n<p><a href=\"http:\/\/lua-users.org\/wiki\/ScopeTutorial\" rel=\"nofollow noreferrer\">lexical scoping<\/a>, if you want to call <code>class 'Blah'<\/code> from the <strong>same<\/strong> module.\nBut if you call it from an <strong>other<\/strong> module - as we mostly do with a class -, we don't need use <code>do-end<\/code> <a href=\"http:\/\/lua-users.org\/wiki\/ScopeTutorial\" rel=\"nofollow noreferrer\">lexical scoping<\/a>.<\/p>\n\n<p>So if your module's purpose is just declaring a torch type class and then using it several time from other modules, you just need declare as <code>local<\/code> as above in the section <strong>1.<\/strong> and you don't need (but you <strong>can<\/strong>) <strong>do-end<\/strong> <a href=\"http:\/\/lua-users.org\/wiki\/ScopeTutorial\" rel=\"nofollow noreferrer\">lexical scoping<\/a>.<\/p>\n\n<p>Actually <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/utility.md\" rel=\"nofollow noreferrer\">torch documentation<\/a> remark:<\/p>\n\n<pre><code>-- for naming convenience\n<\/code><\/pre>\n\n<p>is a bit misleading here, I guess.<\/p>\n"}]}
{"QuestionId":33262928,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-10-21T15:08:47.213","AcceptedAnswerId":33274241.0,"OwnerUserId":492372.0,"Title":"Why doesn this Theano.scan function not throw an error?","Body":"<p>I have this piece of code from the Theano documentation that calculates the polynomial<\/p>\n\n<pre><\/pre>\n\n<p>I think what it is doing is calculating 1*(3^0) + 0*(3^1) + 2*(3^2) which is 19.0 which is my output.<\/p>\n\n<p>My question is that full_range is an array of [0, 1, 2, ... ,998, 999] containing 1000 elements. This is being used as sequences in the Theano scan parameter. <\/p>\n\n<p>Why doesn't w.r.t the sequences parameter in the scan function. It has two sequences as input namely coefficients and full_range. This means that in every iteration of the scan function, it picks one value from each of these inputs and runs the code<\/p>\n\n<pre><\/pre>\n\n<p>where the free_var is 3. How does the code manage to run successfully when the number of elements in coeff is just 3 while the number of elements in power (which is the full_range input) has 1000 elements? I am missing something here. <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"33274241","CreationDate":"2015-10-22T05:54:45.357","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>When multiple sequences are passed to <code>scan<\/code>, it will iterate as many times as there are elements in the shortest sequence.<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#iterating-over-the-first-dimension-of-a-tensor-calculating-a-polynomial\" rel=\"nofollow\">The documentation<\/a> says:<\/p>\n\n<blockquote>\n  <p>given multiple sequences of uneven lengths, scan will truncate to the\n  shortest of them. This makes it safe to pass a very long arange, which\n  we need to do for generality, since arange must have its length\n  specified at creation time.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33266902,"AnswerCount":1,"Tags":"<numpy><theano><blas>","CreationDate":"2015-10-21T18:44:41.850","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"theano - speed test compared to numpy - BLAS libraries","Body":"<p>I just managed to install theano. I did the speed test with the code below- and theano is a lot slower than numpy on the CPU. I presume that I might inadvertently have linked theano to the wrong blas library? (presuming that they would show equal speed if they would use the same library)? is there an easy way to check and remedy this? please let me know any advice - Many thanks in advance, W<\/p>\n\n<p>NP time: 2.384000[s], theano time: 47.326000[s]<\/p>\n\n<pre><\/pre>\n\n<p>Upon the advice of ali_m, i checked theano.config.blas.ldflags (which showed CBLAS) and i checked np.<strong>config<\/strong>.show() - which shows the text on the bottom:<\/p>\n\n<p>following <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install.html<\/a> (under troubleshooting) - i can disable the Theano BLAS (which indeed speeds things up) - but i am wondering how i can either achieve option zero or option two - and get theano to use the same MKL library? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33270607","CreationDate":"2015-10-21T22:54:32.990","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>To make theano dynamically link against MKL rather than CBLAS you could try the following:<\/p>\n\n<ol>\n<li>Create the file <code>$HOME\/.theanorc.txt<\/code> if it doesn't already exist<\/li>\n<li><p>Add the following lines:<\/p>\n\n<pre><code>[blas]\nldflags = -L'C:\/Program Files (x86)\/Intel\/Composer XE\/mkl\/lib\/intel64' -lmkl_intel_thread -lmkl_core -lpthread -lm\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>I don't have a Windows machine at hand to test this, so you may have to play around with the paths a bit.<\/p>\n"}]}
{"QuestionId":33276539,"AnswerCount":1,"Tags":"<cuda><gpu-programming><caffe>","CreationDate":"2015-10-22T08:29:32.247","AcceptedAnswerId":33276793.0,"OwnerUserId":3442959.0,"Title":"CAFFE: Cuda Error \"(8 vs. 0) invalid device function\" when using GPU (GeForce GTX 970)?","Body":"<p>I'm trying to run the CNN Network \"CAFFE\" on a GTX 970. But I get the error mentioned in the title.<\/p>\n\n<p>Can someone help?<\/p>\n\n<p>I posted the issue with more details on the caffe group, but received no hints\/answers!<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/sVOfE0qhf_M\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/sVOfE0qhf_M<\/a><\/p>\n\n<p><strong>update 1<\/strong><\/p>\n\n<p>In my Makefile.config, I added <\/p>\n\n<pre><\/pre>\n\n<p>But when I try to make, it returns:<\/p>\n\n<pre><\/pre>\n\n<p><strong>update 2<\/strong><\/p>\n\n<p>The NCC version is:<\/p>\n\n<pre><\/pre>\n\n<p><strong>update 3<\/strong><\/p>\n\n<p>I'm using CUDA 6.5 with the 346.96 driver<\/p>\n\n<pre><\/pre>\n\n<p><strong>update 4<\/strong><\/p>\n\n<p>I found <a href=\"https:\/\/wiki.ubuntuusers.de\/Grafikkarten\/nvidia\/nvidia\" rel=\"nofollow\">here<\/a> (page is in German) that Ubuntu 15.04 only supports NVidia GPUs until the 800-series.\nWhat I'm doing right now is update Ubuntu to 15.10, where the NVidia driver version \"nvidia-352\" is available that is supposed to support a Geforce GTX 970.<\/p>\n\n<p>I'll post the results here.<\/p>\n\n<p><strong>update 5<\/strong><\/p>\n\n<p>It's updated to Ubuntu 15.10. This comes with GCC 5.2, but CAFFE only accepts versions below 4.9.\nI installed g++ 4.8 parallel to the existing one I've set the compiler to \"g++-4.8\" in cmake...<\/p>\n\n<p><strong>update 6<\/strong><\/p>\n\n<p>Now the error reads<\/p>\n\n<pre><\/pre>\n\n<p><strong>update 7<\/strong><\/p>\n\n<p>My next attempt:<\/p>\n\n<p>1) Install Ubuntu 15.04<\/p>\n\n<p>2) Install <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow\">CUDA 7.5 directly from nVidia<\/a><\/p>\n\n<p>3) Install <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow\">cuDNN 7.0<\/a><\/p>\n\n<p>4) install <a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow\">CAFFE<\/a><\/p>\n\n<p><strong>that finally worked!!<\/strong><\/p>\n","answers":[{"AnswerId":"33276793","CreationDate":"2015-10-22T08:42:56.543","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>Whenever the CUDA runtime API returns \"Invalid Device Function\", it means you are using code which wasn't built for the architecture you are trying to run it on (and doesn't have a JIT path). <\/p>\n\n<p>You probably need to check your CAFFE <code>Makefile.config<\/code> to make sure it is set for the correct architecture, by making sure that <code>CUDA_ARCH<\/code> includes <code>-gencode arch=compute_52,code=compute_52<\/code>.<\/p>\n\n<p>Because you are using a compute capability 52 device, you must use CUDA 7 or newer with your GPU. The CUDA 6.5 toolkit you are using cannot generate code for your GPU, and the CUDA 6.5 driver cannot JIT recompile lower compute capability code into something which will run on your GPU. This is a not negotiable requirement for your hardware.<\/p>\n\n<p><strong>solution is under update 7 in the post above<\/strong><\/p>\n"}]}
{"QuestionId":33279057,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2015-10-22T10:38:52.947","AcceptedAnswerId":null,"OwnerUserId":1783952.0,"Title":"Cannot import caffe - iPython doesn't find environment variables that 'ordinary' python finds","Body":"<p>I'm using Anaconda python on a Mac running OS X 10.11 El Capitan. <\/p>\n\n<p>I'm trying to use  in iPython, but importing it () fails with the error <\/p>\n\n<pre><\/pre>\n\n<p>While iPython can't import caffe, doing so in 'ordinary' python works.<\/p>\n\n<pre><\/pre>\n\n<p>I should also mention that I'm using CUDA 7.5, and although I can't see which version of caffe I have (downloaded zip) it was done from the master branch on September 17 2015.<\/p>\n\n<p>The actions I took for 'ordinary' python to be able to import caffe was to add the following to ~\/.bash_profile<\/p>\n\n<pre><\/pre>\n\n<p>iPython is reading the PYTHONPATH set in  since it will give  if the  line is commented out. But for some reason it doesn't find the , which 'ordinary' python does.<\/p>\n\n<p>Any ideas on what might be wrong?<\/p>\n","answers":[{"AnswerId":"33297699","CreationDate":"2015-10-23T08:13:10.113","ParentId":null,"OwnerUserId":"1783952","Title":null,"Body":"<p>Ok, so it turns out that the main issue was that due to the new feature called SIP (System Integrity Protection) in El Capitan, that prevents <code>DYLD_FALLBACK_LIBRARY_PATH<\/code> from being set. This is a feature that is supposed to provide extra protection against root attacks on Mac OSX.<\/p>\n\n<p>To disable SIP:<\/p>\n\n<ul>\n<li>Restart computer in recovery mode by holding in cmd-R when starting\nup.<\/li>\n<li>Open a terminal window (under Utilities in the top menu).<\/li>\n<li>Enter <code>csrutil disable<\/code> to turn SIP off (and <code>csrutil enable<\/code> to turn it back on).<\/li>\n<li>Restart computer from Apple symbol in the manu.<\/li>\n<\/ul>\n\n<p>This solved the issue for me. For more info: <a href=\"http:\/\/www.macworld.com\/article\/2986118\/security\/how-to-modify-system-integrity-protection-in-el-capitan.html\" rel=\"nofollow\">MacWorld article on SIP<\/a>, <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/waugt62RQMU\" rel=\"nofollow\">Caffe user mailing list<\/a>.<\/p>\n"}]}
{"QuestionId":33283233,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-22T14:16:11.577","AcceptedAnswerId":null,"OwnerUserId":4677531.0,"Title":"Is the use of 'givens' really necessary in the deeplearning tutorials?","Body":"<p>In the <a href=\"http:\/\/deeplearning.net\/tutorial\/contents.html\" rel=\"nofollow noreferrer\">deep learning tutorials<\/a>, all training data is stored in a  array and only an index into that array is passed to the training function to slice out a minibatch.\nI understand that this allows the data to be left in GPU memory, as opposed to passing small chunks of data as a parameter to the training function for each minibatch.\nIn some <a href=\"https:\/\/stackoverflow.com\/questions\/26879157\/purpose-of-given-variables-in\">previous questions<\/a>, this was given as an answer as to why the  mechanism is used in the tutorials.<\/p>\n\n<p>I don't yet see the connection between these two concepts, so I'm probably missing out on something essential.\nAs far as I understand, the givens mechanism swaps out a variable in the graph with a given symbolic expression (i.e., some given subgraph is inserted in place of that variable).\nThen why not define the computational graph the way we need it in the first place?<\/p>\n\n<p>Here is a minimal example. I define a shared variable  and an integer , and I either create a graph that already contains the slicing operation, or I create one where the slicing operation is inserted post-hoc via .\nBy all appearances, the two resulting functions  and  are identical (see the debugprints at the end).<\/p>\n\n<p>But then why do the tutorials use the  pattern?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"39686434","CreationDate":"2016-09-25T11:44:41.947","ParentId":null,"OwnerUserId":"3736362","Title":null,"Body":"<p>They use <code>givens<\/code> here only to decouple actual data which is passed to the graph from the input data variable. You could explicitly replace input variable with <code>X[index * batch_size: (index + 1) * batch_size]<\/code> but that is just a little more messy. <\/p>\n"}]}
{"QuestionId":33286455,"AnswerCount":1,"Tags":"<python-3.4><bayesian><theano><pymc3>","CreationDate":"2015-10-22T16:55:19.570","AcceptedAnswerId":null,"OwnerUserId":5206473.0,"Title":"Bayesian Lifetime estimates using pymc3\/theano","Body":"<p>I am trying to estimate the following model<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xKEst.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xKEst.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where I provide uniform priors <a href=\"https:\/\/i.stack.imgur.com\/pe9bO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pe9bO.png\" alt=\"enter image description here\"><\/a>  and I code the likelihood <a href=\"https:\/\/i.stack.imgur.com\/HTqQE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HTqQE.png\" alt=\"enter image description here\"><\/a>. The latter comes from this <a href=\"https:\/\/marketing.wharton.upenn.edu\/files\/?whdmsaction=public:main.file&amp;fileID=344\" rel=\"nofollow noreferrer\">paper<\/a> and goes as follows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1IncA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1IncA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In the theano\/pymc3 implementation I am computing the first and second term on the rhs in  and . Finally  sum over the entire sample.<\/p>\n\n<p>Theano, on his own is producing some output yet when I integrate it in a pymc3 model the following error appears:<\/p>\n\n<p><\/p>\n\n<p>I think the problem is how to supply pymc3 variables to theano.<\/p>\n\n<pre><\/pre>\n\n<p>My first guess was modifying  with  but then theano complains about some mysterious  missing from the graph. Any clue?<\/p>\n","answers":[{"AnswerId":"33367696","CreationDate":"2015-10-27T12:11:41.637","ParentId":null,"OwnerUserId":"5206473","Title":null,"Body":"<p>I figured it out by: i) simplifying things ii) avoiding using theano operators when coding the likelihood and iii) using the build-in wrapper Deterministic for deterministic transformations of variables (my lifetime). To speed up computation I vectorised the likelihood by writing the second term on the rhs as solution of a geometric series. Here is the code in case someone want to test it on his own lifetime application.<\/p>\n\n<pre><code>from pymc3 import Model, Uniform, Deterministic \nimport pymc3\nfrom scipy import optimize\nimport theano.tensor as T\n\nX = array([[ 5, 64,  8, 13],[ 4, 71, 23, 41],[ 7, 70,  4, 19]) \n#f, n, x, tx where f stands for the frequency of the triple (n, x, tx)\n\nclass CustomDist(pymc3.distributions.Discrete):\n    def __init__(self, p, theta, *args, **kwargs):\n        super(CustomDist, self).__init__(*args, **kwargs)\n        self.p = p\n        self.theta = theta\n\n    def logp(self, X):\n        p = self.theta\n        theta = self.theta\n        f, n, x, tx = X[0],(X[1] + 1),X[2],(X[3] + 1) #time indexed at 0, x &gt; n\n        ll =  f * T.log(  p ** x * (1 - p) ** (n - x) * (1 - theta) ** n + \n                [(1 - p) ** (tx - x) * (1 - theta) ** tx - (1 - p) ** (n - x) * (1 - theta) ** n] \/ (1 - (1 - p) * (1 - theta)) )\n        # eliminated -1 since would result in negatice ll\n        return(T.sum( ll ))\n\nwith Model() as bg_model:\n    p = Uniform('p', lower = 0, upper = 1)\n    theta = Uniform('theta', lower = 0, upper = 1)\n    like = CustomDist('like', p = p, theta = theta, observed=X.T) #likelihood\n\n    lt = pymc3.Deterministic('lt', p \/ theta)\n    # start = {'p':.5, 'theta':.1}\n    start = pymc3.find_MAP(fmin=optimize.fmin_powell)\n    step = pymc3.Slice([p, theta, lt])\n    trace = pymc3.sample(5000, start = start, step = [step])\n\npymc3.traceplot(trace[2000:])\n<\/code><\/pre>\n"}]}
{"QuestionId":33288006,"AnswerCount":1,"Tags":"<lua><parallel-processing><torch>","CreationDate":"2015-10-22T18:21:14.907","AcceptedAnswerId":33289716.0,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, how to save a trained neural network model to file?","Body":"<p>I working on a Torch \/ Lua project, in which I implemented an artificial neural network model.\nEverything works, but now I'd like to modify my code in the following way. Since my input dataset are very large, I'd like to divide it in N=20 spans. <\/p>\n\n<p>Then I want to train my neural network only on the 1st dataset span and then test on the other N-1=19 spans in parallel.<\/p>\n\n<p>To run all these parallel jobs, I need to save up my neural network model details to a file, and then load it for every 19 jobs.<\/p>\n\n<p><strong>Is there any way in torch to correctly \"write\" an artificial neural network model to file?<\/strong><\/p>\n","answers":[{"AnswerId":"33289716","CreationDate":"2015-10-22T20:04:16.777","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>-- save the model\ntorch.save(filename, model)\n<\/code><\/pre>\n\n<p>.<\/p>\n\n<pre><code>-- load the model\nmodel = torch.load(filename)\n<\/code><\/pre>\n"}]}
{"QuestionId":33292212,"AnswerCount":2,"Tags":"<python><lstm><keras>","CreationDate":"2015-10-22T23:04:01.037","AcceptedAnswerId":null,"OwnerUserId":4034203.0,"Title":"Simplest Lstm training with Keras io","Body":"<p>I would like to create the simplest LSTM there is using keras python library.<\/p>\n\n<p>I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>but does not seem to work...<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone explain me what exactly goes wrong ? <\/p>\n\n<p>I also tried to Transpose the matrices but that's not it.<\/p>\n","answers":[{"AnswerId":"33574500","CreationDate":"2015-11-06T19:44:48.883","ParentId":null,"OwnerUserId":"3298756","Title":null,"Body":"<p>First problem I see is using Pandas Dataframe. I think you should use numpy array here. The second problem is the X matrix. It should be a 3D array. For example if I try with<\/p>\n\n<pre><code>X_train = np.random.randn(6,2,2)\n<\/code><\/pre>\n\n<p>then it will work. <\/p>\n"},{"AnswerId":"45450108","CreationDate":"2017-08-02T01:37:52.163","ParentId":null,"OwnerUserId":"3873931","Title":null,"Body":"<p>The error it shows me when I run your code is<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected lstm_2_input to have 3 dimensions, but got array with shape (6, 2)<\/p>\n<\/blockquote>\n\n<p>You can fix this by inputting a 3D numpy array, as user108372 mentioned<\/p>\n\n<p>A good way to debug is this to print out model.summary(), which will give you the output shapes expected for each of the layers.  Additionally, you don't always need to specify the output and input shapes. Keras will take care of that for you :)  Now, a working example would be something like this:<\/p>\n\n<pre><code>X_train = np.random.randn(6,2,2)\ny_train = pd.DataFrame( np.array([1, 2, 3, 4, 3, 4]) ).values\nX_test = np.random.randn(2,2,2)\ny_test = pd.DataFrame( np.array([1, 2]) ).values\n\nmodel = Sequential()\nmodel.add(LSTM(32, \n               return_sequences=False, \n               input_dim=X_train.shape[1]))\n# The shape of the last Dense layer should always correspond to y_train.shape[1]\nmodel.add(Dense(y_train.shape[1])) \nmodel.add(Activation(\"linear\"))\nmodel.compile(loss=\"mean_squared_error\",\n              optimizer=\"rmsprop\")\n\nmodel.fit(X_train, y_train)\n<\/code><\/pre>\n\n<p>I recommend you print out shapes of the example above and align your shapes with the shapes described here.<\/p>\n"}]}
{"QuestionId":33298988,"AnswerCount":0,"Tags":"<torch>","CreationDate":"2015-10-23T09:27:15.893","AcceptedAnswerId":null,"OwnerUserId":988191.0,"Title":"Large Difference in Output: nnGraph based LSTM vs Sequencer LSTM (Torch)","Body":"<p>I have implemented a Sequence Labeler in Torch using rnn from <a href=\"https:\/\/github.com\/Element-Research\/rnn\" rel=\"nofollow\">Element Research<\/a> and also using nnGraph based LSTM code from <a href=\"https:\/\/github.com\/oxford-cs-ml-2015\/practical6\" rel=\"nofollow\">Oxford ML Group<\/a>. The training of the nnGraph based LSTM is done similar to the one given in Oxford ML Group.<\/p>\n\n<p>I have kep the hyperparameters same for both the modules. When i train both the modules on same dataset i get lower error(around 75 Fmeasure) on rnn from Element Research while the error is large(around 5 Fmeasure) using nnGraph based LSTM. <\/p>\n\n<p>For simplification i do Backpropagaton Through Time over the enitire sequence on both the models. For nnGraph based LSTM I clone it for maximum length of the sequence.<\/p>\n\n<p>Here is the snippet for training using  package:<\/p>\n\n<pre><\/pre>\n\n<p>and the snippet for training using nnGraph based LSTM is<\/p>\n\n<pre><\/pre>\n\n<p>I have shared the complete code snippet here : <a href=\"https:\/\/bitbucket.org\/snippets\/rmurthy\/rLopM\/sequence-labeling-using-lstm\" rel=\"nofollow\">Complete Code Snippet for both modules<\/a><\/p>\n\n<p>I know that i'm missing something in my nnGraph based LSTM implenetation but unable to figure out my error. Can someone please help me in finding where i am wrong?<\/p>\n","answers":[]}
{"QuestionId":33300575,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><lstm>","CreationDate":"2015-10-23T10:51:31.127","AcceptedAnswerId":33348718.0,"OwnerUserId":5103648.0,"Title":"How to vectorize LSTMs?","Body":"<p>In particular, I'm confused about what it means for an LSTM layer to have (say) 50 cells. Consider the following LSTM block from <a href=\"http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\" rel=\"nofollow noreferrer\">this awesome blog post<\/a>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/y91nY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/y91nY.png\" alt=\"LSTM Model from http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\"><\/a><\/p>\n\n<p>Say my input  is a  vector and the hidden layer  is a  vector. Given that the cell state  undergoes only point-wise operations (point-wise  and ) before becoming the new hidden state, I gather that . Now the forget gate looks at the input concatenated with the hidden layer, which would be a  vector, which means the forget gate must have a weight matrix of shape , such that . <\/p>\n\n<p>So my question at this point is that, am I looking at a LSTM block with 50 cells when ? Or am I misunderstanding what it means for a LSTM layer to have 50 cells? <\/p>\n","answers":[{"AnswerId":"33348718","CreationDate":"2015-10-26T14:50:36.097","ParentId":null,"OwnerUserId":"492372","Title":null,"Body":"<p>I understand what you are getting confused with. So basically, the black line connecting the two boxes at the top which represents the cell state is actually a set of very small 50 lines grouped together. These get multiplied point wise with the output of the forget gate which has an output consisting of 50 values. These 50 values multiply with the cell state point wise.<\/p>\n"}]}
{"QuestionId":33304859,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-10-23T14:29:28.167","AcceptedAnswerId":33359405.0,"OwnerUserId":1452257.0,"Title":"Shape mismatch when using combining layers in Caffe","Body":"<p>I'm using the Caffe library for training a convolutional neural network (CNN). However, I'm getting the following error when using the concat layer to combine the output from two convolutional layers before applying it to a inner_product layer. <\/p>\n\n<pre><\/pre>\n\n<p>As far as I know I am using the concat layer in the exact same way as in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/train_val.prototxt\" rel=\"nofollow\">BVLC_GoogLeNet<\/a>. The concat layer can be found in my train.prototxt at <a href=\"http:\/\/pastebin.com\/DbqkTFpd\" rel=\"nofollow\">pastebin<\/a> under the name . The dimensions of my input blob is , where the data format in Caffe is . I've tried training both using the pycaffe interface and the console. I get the same error. Below is code for training using the console.<\/p>\n\n<pre><\/pre>\n\n<p>What is the meaning of this error? And how can it be resolved?<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>As mentioned in the comments the log contains nothing else than the error. The stack trace for the error is the following:<\/p>\n\n<pre><\/pre>\n\n<p>It should also be noted that my solver and code works fine for training the exact same CNN with only 1 \"path\" along the network, i.e. without the CONCAT layer.<\/p>\n","answers":[{"AnswerId":"33359405","CreationDate":"2015-10-27T03:20:51.223","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>I believe the issue you're having is that your <strong>train<\/strong> net has been updated to have a concat layer while your <strong>test<\/strong> net hasn't.<\/p>\n\n<p>It would explain the 400x400 vs 400x800 issue you're having considering your concat merges two 400x400 layers. I can't know for certain without being able to see your test net.<\/p>\n"}]}
{"QuestionId":33306236,"AnswerCount":1,"Tags":"<caffe><roc><auc><pycaffe><matcaffe>","CreationDate":"2015-10-23T15:32:20.627","AcceptedAnswerId":null,"OwnerUserId":242871.0,"Title":"Calculating ROC and AUC in Caffe?","Body":"<p>I have trained imagenet in Caffe. Now i am trying to calculate ROC\/AUC for my model and the trained model provided by caffe. I have two questions:<\/p>\n\n<p>1) ROC\/AUC is mainly used for binary classes, but i also found that in some cases people used it for multi-classes. Is it possible for 1000 classes. And what will be its impact? As in  reviews people didn't give good answer for ROC\/AUC in multi-class problems. <\/p>\n\n<p>2) If possible, and comparing two models based on ROC\/AUC will be a good idea, Can anybody tell how to do it for these 1000 classes in Caffe? And do i have to retrain the models from scratch, or can i calculate only with final trained models?<\/p>\n\n<p>Regards<\/p>\n","answers":[{"AnswerId":"41681259","CreationDate":"2017-01-16T16:53:11.330","ParentId":null,"OwnerUserId":"2641587","Title":null,"Body":"<p><a href=\"https:\/\/stats.stackexchange.com\/questions\/2151\/how-to-plot-roc-curves-in-multiclass-classification\">This discussion<\/a> addresses multi-class ROC\/AUC analysis nicely. Answering your questions:<\/p>\n\n<ol>\n<li><p>You can do multiple one-vs-all classifications for each class, thus building multiple ROC curves. <\/p><\/li>\n<li><p>Having computed 1000 AUC values, you can come up with the mean AUC over all classes and use this metric to compare goodness of your models. No, you don't need to retrain your models.<\/p><\/li>\n<\/ol>\n\n<p>Also, pay an attention that ROC\/AUC metrics are quite specific and used mostly in detection\/biometry tasks like voice identification. <\/p>\n"}]}
{"QuestionId":33310424,"AnswerCount":0,"Tags":"<floating-point><neural-network><hdf5><deep-learning><caffe>","CreationDate":"2015-10-23T19:53:27.680","AcceptedAnswerId":null,"OwnerUserId":1646558.0,"Title":"Floating point exception in caffe","Body":"<p>I am trying learn caffe running some examples. I am using a dataset in format HDF5 and I am getting the following error when running caffe train:<\/p>\n\n<pre><\/pre>\n\n<p>My net:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone help me with the error?<\/p>\n","answers":[]}
{"QuestionId":33311167,"AnswerCount":1,"Tags":"<python><anaconda><theano>","CreationDate":"2015-10-23T20:49:19.030","AcceptedAnswerId":33323970.0,"OwnerUserId":1833640.0,"Title":"Error when importing theano using GPU","Body":"<p>I am trying to use theano for python 2.7 installed through anaconda on a CentOS 5.6. I am able to import theano if only using cpu, but when I try to run it using gpu I get hte error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on how to fix that? I have been trying everything, from reinstalling cuda, anaconda,python...<\/p>\n","answers":[{"AnswerId":"33323970","CreationDate":"2015-10-24T22:19:03.393","ParentId":null,"OwnerUserId":"1833640","Title":null,"Body":"<p>I have found the problem after reading this <a href=\"http:\/\/www.johnwittenauer.net\/configuring-theano-for-high-performance-deep-learning\/\" rel=\"nofollow\">link<\/a>. Apparently theano creates a cache folder .theano that was causing the issue. It was fixed after removing this folder. <\/p>\n"}]}
{"QuestionId":33327144,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-10-25T07:08:17.310","AcceptedAnswerId":33327702.0,"OwnerUserId":5485277.0,"Title":"MissingInputError with theano scan function","Body":"<p>I am experimenting with Theano and specifically with the function <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">scan<\/a>. <\/p>\n\n<p>I want to use it to apply a linear classifier to a set of feature vectors stored as columns of a matrix X (I am sure there are better ways to do it, this is just to get familiar with the function scan). <\/p>\n\n<p>This is my code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>When executing the snippet from iPython I get the following error (triggered by the last instruction):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33327702","CreationDate":"2015-10-25T08:35:37.973","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It's not entirely clear what you are trying to do here but my guess is that you're implementing two different versions of a linear classifier, one that does not use scan and another that does.<\/p>\n\n<p>The code below demonstrates my approach to doing this.<\/p>\n\n<p>To answer your specific question:<\/p>\n\n<p>The error message is appearing because your scan version uses <code>T_f<\/code> in the scan step function (this is strange and one reason it's not clear what you're trying to do; the step function is not using any of its input variables <code>x<\/code>, <code>W<\/code>, or <code>b<\/code> at all!) and <code>T_f<\/code> uses <code>T_x<\/code> but your scan version's function does not take <code>T_x<\/code> as input. Instead it takes <code>T_X<\/code> (note the case difference) which is not then used at all.<\/p>\n\n<p>Here's a few hints and explanations for the differences between your code and mine.<\/p>\n\n<ol>\n<li><p>It helps immensely to keep things separated into discrete methods. By splitting the code into the <code>v1<\/code> and <code>v2<\/code> methods, we ensure that the two different implementations don't interfere with one another.<\/p><\/li>\n<li><p>Using the <code>strict<\/code> parameter of <code>theano.scan<\/code> is recommended at all times. It ensures you don't accidentally introduce an error caused by naming conflicts in the step function's parameters. It isn't enabled by default because that could break older code, when strict didn't exist.<\/p><\/li>\n<li><p>Use a fully fledged function instead of a lambda for scan's step function. Like strict mode, this helps avoid accidental naming conflicts, and makes the step code easier to follow. The step function can also be tested in isolation.<\/p><\/li>\n<li><p>Use <code>compute_test_value<\/code> to ensure the computation works with simple sample data. In particular, this will identify shape mismatches (e.g. doing a <code>dot<\/code> with the parameters in the wrong order), and makes debugging easier by being able to print\/explore the intermediate values <em>while the computation graph is being constructing<\/em> instead of later when the computation is executed.<\/p><\/li>\n<li><p>This code has each input sample encoded as a row of <code>x<\/code> instead of as a column of <code>x<\/code>. This requires post-multiplying by <code>w<\/code> instead of pre-multiplying. It's possible to do it either way but pre-multiplying by <code>w<\/code> will make the addition with <code>b<\/code> a bit messier (a <code>dimshuffle<\/code> will need to be introduced).<\/p><\/li>\n<li><p>There's no need to use <code>theano.Param<\/code> unless you need to use non-standard behaviour with respect to default values, etc.<\/p><\/li>\n<li><p>Avoid naming things such that they differ only in case! In general, stick to the <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0008\/\" rel=\"nofollow\">Python style guide<\/a> (i.e. instance variables should be lower case with words separated with underscores).<\/p><\/li>\n<li><p>A dimshuffle and a selection of the first row are needed in the scan version's step function to ensure the dot product and subsequence addition of the bias are dimension compatible. This is not needed in the non-scan version because there we are doing a matrix-matrix dot-product.<\/p><\/li>\n<\/ol>\n\n<p>The code:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as T\n\n\ndef create_inputs(x_value, w_value, b_value):\n    x, w = T.matrices(2)\n    b = T.vector()\n    x.tag.test_value = x_value\n    w.tag.test_value = w_value\n    b.tag.test_value = b_value\n    return x, w, b\n\n\ndef v1(x_value, w_value, b_value):\n    x, w, b = create_inputs(x_value, w_value, b_value)\n    y = T.dot(x, w) + b\n    f = theano.function(inputs=[x, w, b], outputs=y)\n    print f(x_value, w_value, b_value)\n\n\ndef v2_step(x, w, b):\n    return (T.dot(x.dimshuffle('x', 0), w) + b)[0]\n\n\ndef v2(x_value, w_value, b_value):\n    x, w, b = create_inputs(x_value, w_value, b_value)\n    y, _ = theano.scan(v2_step, sequences=[x], non_sequences=[w, b], strict=True)\n    f = theano.function(inputs=[x, w, b], outputs=y)\n    print f(x_value, w_value, b_value)\n\n\ndef main():\n    batch_size = 2\n    input_size = 3\n    hidden_size = 4\n    theano.config.compute_test_value = 'raise'\n    numpy.random.seed(1)\n    x_value = numpy.random.standard_normal(size=(batch_size, input_size))\n    w_value = numpy.random.standard_normal(size=(input_size, hidden_size))\n    b_value = numpy.zeros((hidden_size,))\n    v1(x_value, w_value, b_value)\n    v2(x_value, w_value, b_value)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":33330779,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2015-10-25T14:25:35.330","AcceptedAnswerId":33349475.0,"OwnerUserId":115781.0,"Title":"What's the triplet loss back propagation gradient formula?","Body":"<p>I am trying to use caffe to implement triplet loss described in <a href=\"http:\/\/arxiv.org\/abs\/1503.03832\" rel=\"noreferrer\">Schroff, Kalenichenko and Philbin \"FaceNet: A Unified Embedding for Face Recognition and Clustering\", 2015<\/a>.<\/p>\n\n<p>I am new to this so how to calculate the gradient in back propagation?<\/p>\n","answers":[{"AnswerId":"33349475","CreationDate":"2015-10-26T15:26:45.047","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I assume you define the loss layer as<\/p>\n\n<pre><code>layer {\n  name: \"tripletLoss\"\n  type: \"TripletLoss\"\n  bottom: \"anchor\"\n  bottom: \"positive\"\n  bottom: \"negative\"\n  ...\n}\n<\/code><\/pre>\n\n<p>Now you need to compute a gradient w.r.t each of the \"bottom\"s.<\/p>\n\n<p>The loss is given by:<br>\n<a href=\"https:\/\/i.stack.imgur.com\/oEtp7.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oEtp7.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The gradient w.r.t the \"anchor\" input (<code>fa<\/code>):<br>\n<a href=\"https:\/\/i.stack.imgur.com\/jm7w8.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jm7w8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The gradient w.r.t the \"positive\" input (<code>fp<\/code>):<br>\n<a href=\"https:\/\/i.stack.imgur.com\/Za4bF.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Za4bF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The gradient w.r.t the \"negative\" input (<code>fn<\/code>):<br>\n<a href=\"https:\/\/i.stack.imgur.com\/lOmv0.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lOmv0.png\" alt=\"![enter image description here\"><\/a><\/p>\n\n<hr>\n\n<p>The original calculation (I leave here for sentimental reasons...) <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9QWXX.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9QWXX.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Please see <a href=\"https:\/\/stackoverflow.com\/questions\/33330779\/whats-the-triplet-loss-back-propagation-gradient-formula\/33349475?noredirect=1#comment54527446_33349475\">comment<\/a> correcting the last term.<\/p>\n"}]}
{"QuestionId":33340042,"AnswerCount":1,"Tags":"<amazon-ec2><neural-network><torch><lstm>","CreationDate":"2015-10-26T06:57:42.403","AcceptedAnswerId":null,"OwnerUserId":3612823.0,"Title":"Training LSTM in Torch","Body":"<p>I'm trying to implement an LSTM model in Torch on EC2.<\/p>\n\n<p>I don't have any experience with Torch but I found this code: <a href=\"http:\/\/apaszke.github.io\/lstm-explained.html\" rel=\"nofollow\">http:\/\/apaszke.github.io\/lstm-explained.html<\/a><\/p>\n\n<p>I have Torch running on EC2 with the LSTM.lua script loaded.<\/p>\n\n<p>I read through the guide but I'm not sure how to add my data and train the model. Do I just put a :forward and a :backward in a loop?<\/p>\n","answers":[{"AnswerId":"37237560","CreationDate":"2016-05-15T11:15:08.913","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<p>You can have a look at this web page.\n<a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/training.md#nn.traningneuralnet.dok\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/training.md#nn.traningneuralnet.dok<\/a><\/p>\n\n<p>input(data) and output(labels) can be Tensors so you can load your data in the Torch structure, then :forward, :backward and upgrade the model parameters.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":33341135,"AnswerCount":0,"Tags":"<macos><lua><gnuplot><torch>","CreationDate":"2015-10-26T08:11:37.563","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Gnu plot error with lua\/torch (OS X)","Body":"<p>I want to run some code in torch that requires the gnuplot lib however I get the following error;<\/p>\n\n<pre><\/pre>\n\n<p>I read here <a href=\"https:\/\/stackoverflow.com\/questions\/29882271\/lua-error-gnuplot-terminal-is-not-set\">Lua Error: &quot;Gnuplot terminal is not set&quot;<\/a> that I require the gnu plot executable installed. So I downloaded this and followed the website instructions here <a href=\"http:\/\/gnuplot.sourceforge.net\/ReleaseNotes_5_0.html\" rel=\"nofollow noreferrer\">http:\/\/gnuplot.sourceforge.net\/ReleaseNotes_5_0.html<\/a> and then I get this error;<\/p>\n\n<pre><\/pre>\n\n<p>Not too sure if I am going about this the correct way. Any help would be greatly appreciated! <\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":33354272,"AnswerCount":1,"Tags":"<python><cuda><celery><theano>","CreationDate":"2015-10-26T19:43:39.640","AcceptedAnswerId":null,"OwnerUserId":1062529.0,"Title":"RuntimeError when using Theano shared variable in a Celery celery worker","Body":"<p>I have a <a href=\"http:\/\/www.celeryproject.org\/\" rel=\"nofollow\">Celery task<\/a> called <strong>simple_theano_tasks<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>When THEANO is <em>configured<\/em> to use the CPU, everything works as expected (no error):<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>[INFO\/MainProcess] Received task: my_project.tasks.simple_theano_tasks[xxxx]<\/p>\n  \n  <p>[WARNING\/Worker-1] Done. Shared value is []<\/p>\n  \n  <p>[INFO\/MainProcess] Task my_project.tasks.simple_theano_tasks[xxxx] succeeded in 0.00407959899985s<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Now, when I do the exact same thing with GPU enabled, Theano (or CUDA) raise an error:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>...<\/p>\n  \n  <p>Using gpu device 0: GeForce GTX 670M (CNMeM is enabled)<\/p>\n  \n  <p>...<\/p>\n  \n  <p>[INFO\/MainProcess] Received task: my_project.tasks.simple_theano_tasks[xxx]<\/p>\n  \n  <p>[ERROR\/MainProcess] Task my_project.tasks.simple_theano_tasks[xxx] raised unexpected: RuntimeError(\"Cuda error 'initialization error' while copying %lli data element to device memory\",)<\/p>\n  \n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\/...\/local\/lib\/python2.7\/site-packages\/celery\/app\/trace.py\", line 240, in trace_task\n      R = retval = fun(*args, **kwargs)<\/p>\n  \n  <p>File \"\/...\/local\/lib\/python2.7\/site-packages\/celery\/app\/trace.py\", line 438, in <strong>protected_call<\/strong>\n      return self.run(*args, **kwargs)<\/p>\n  \n  <p>File \"\/...\/my_project\/tasks.py\", line 362, in simple_theano_tasks\n      shared = theano.shared(my_array, name='my_variable', borrow=True)<\/p>\n  \n  <p>File \"\/...\/local\/lib\/python2.7\/site-packages\/theano\/compile\/sharedvalue.py\", line 247, in shared\n      allow_downcast=allow_downcast, **kwargs)<\/p>\n  \n  <p>File \"\/...\/local\/lib\/python2.7\/site-packages\/theano\/sandbox\/cuda\/var.py\", line 229, in float32_shared_constructor\n      deviceval = type_support_filter(value, type.broadcastable, False, None)\n  RuntimeError: Cuda error 'initialization error' while copying %lli data element to device memory<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Finally, when I run the exact same code in a Python shell I have no error:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Does anyone has an idea of:<\/p>\n\n<ul>\n<li>Why does theano behaves differently inside a Celery worker?<\/li>\n<li>How to fix this?<\/li>\n<\/ul>\n\n<hr>\n\n<p><strong>Some additional context<\/strong>:<\/p>\n\n<ul>\n<li><p>I am using <a href=\"https:\/\/github.com\/Theano\/Theano\" rel=\"nofollow\">theano<\/a>@0.7.0 and <a href=\"http:\/\/www.celeryproject.org\/\" rel=\"nofollow\">Celery<\/a>@3.1.18<\/p><\/li>\n<li><p>\"~\/.theanorc\" file<\/p><\/li>\n<\/ul>\n\n<blockquote>\n  <p>[global]<\/p>\n  \n  <p>floatX=float32<\/p>\n  \n  <p>device=gpu<\/p>\n  \n  <p>[mode]=FAST_RUN<\/p>\n  \n  <p>[nvcc]<\/p>\n  \n  <p>fastmath=True<\/p>\n  \n  <p>[lib]<\/p>\n  \n  <p>cnmem=0.1<\/p>\n  \n  <p>[cuda]<\/p>\n  \n  <p>root=\/usr\/local\/cuda<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"33356142","CreationDate":"2015-10-26T21:41:30.790","ParentId":null,"OwnerUserId":"1062529","Title":null,"Body":"<p>A workaround is to:<\/p>\n\n<ol>\n<li>Specify the CPU as the target device (in \".theanorc\" or using \"THEANO_FLAGS=device=cpu\")<\/li>\n<li>Later, override the assigned device to a specified GPU<\/li>\n<\/ol>\n\n<hr>\n\n<p>Celery task is now:<\/p>\n\n<pre><code>@app.task(bind=True, queue='test')\ndef simple_theano_tasks(self):\n  # At this point, no theano import statements have been processed, and so the device is unbound\n  import theano, numpy as np\n  import theano.sandbox.cuda\n  theano.sandbox.cuda.use('gpu') # enable gpu\n  my_array = np.zeros((0,), dtype=theano.config.floatX)\n  shared = theano.shared(my_array, name='my_variable', borrow=True)\n  print 'Done. Shared value is {}'.format(shared.get_value())\n<\/code><\/pre>\n\n<hr>\n\n<p><em>Note<\/em>:  <a href=\"https:\/\/github.com\/Theano\/Theano\/wiki\/Using-Multiple-GPUs\" rel=\"nofollow\">I found the solution reading this article about using multiple GPU<\/a><\/p>\n"}]}
{"QuestionId":33358281,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-27T01:00:35.430","AcceptedAnswerId":33359365.0,"OwnerUserId":669482.0,"Title":"Calculate the normalized dot product of two matrices using Theano","Body":"<p>I would like to get the normalized dot product of 2 matrices using Theano. By normalized dot product of 2 matrices, I define the normalized inner product of 2 vectors as the following: \nTake v_a from matrix A, and vector v_b from matrix B. AB__dot_norm = v_a * v_b \/ |v_a| |v_b|. <\/p>\n\n<p>I can get the norm of v_a and v_b with the following code. I am not sure how to normalize the dot_product matrix with the normalized vectors.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33359365","CreationDate":"2015-10-27T03:16:32.800","ParentId":null,"OwnerUserId":"669482","Title":null,"Body":"<p>I have figured out a solution, I thought it might be useful to post it here.<\/p>\n\n<pre><code>in_norm = in_tensor \/ in_tensor.norm(2, axis =1 ).reshape((in_tensor.shape[0],1))\nw_norm = w_tensor \/ w_tensor.norm(2,axis = 0).reshape((1,  w_tensor.shape[1]))\noutput = theano.tensor.dot (in_norm, w_norm)\n<\/code><\/pre>\n"}]}
{"QuestionId":33361659,"AnswerCount":0,"Tags":"<vim><autocomplete><lua><torch>","CreationDate":"2015-10-27T06:58:10.170","AcceptedAnswerId":null,"OwnerUserId":4453332.0,"Title":"How to make Vim autosuggest Lua methods when typing?","Body":"<p>For auto-completion, I have installed the  and , but when I try to trigger the auto-completion, I get the error<\/p>\n\n<pre><\/pre>\n\n<p>I'm mostly working in torch7. I'd really appreciate a way of getting library methods suggested while typing.<\/p>\n","answers":[]}
{"QuestionId":33363582,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><lasagne>","CreationDate":"2015-10-27T09:01:45.980","AcceptedAnswerId":null,"OwnerUserId":5492709.0,"Title":"Lasagne mlp target out of bound","Body":"<p>Hi I am trying to modify the mnist example to match it to my dataset. I only try to use the mlp example and it gives a strange error.<\/p>\n\n<p>Tha dataset is a matrix with 2100 rows and 17 columns, and the output should be one of the 16 possible classes. The error seems happening in the secon phase of the training. The model is build correctly (log info confirmed).<\/p>\n\n<p>Here is the error log:<\/p>\n\n<blockquote>\n  <p>ValueError: y_i value out of bounds<\/p>\n  \n  <p>Apply node that caused the error:<\/p>\n  \n  <p>CrossentropySoftmaxArgmax1HotWithBias(Dot22.0, b, targets)<\/p>\n  \n  <p>Toposort index: 33<\/p>\n  \n  <p>Inputs types: [TensorType(float64, matrix), TensorType(float64, vector), >TensorType(int32, vector)]<\/p>\n  \n  <p>Inputs shapes: [(100, 17), (17,), (100,)]<\/p>\n  \n  <p>Inputs strides: [(136, 8), (8,), (4,)]<\/p>\n  \n  <p>Inputs values: ['not shown', 'not shown', 'not shown']<\/p>\n  \n  <p>Outputs clients: [[Sum{acc_dtype=float64}(CrossentropySoftmaxArgmax1HotWithBias.0)], [CrossentropySoftmax1HotWithBiasDx(Assert{msg=' and  do not have the same shape.'}.0, CrossentropySoftmaxArgmax1HotWithBias.1, targets)], []]<\/p>\n  \n  <p>HINT: Re-running with most Theano optimization disabled could give you a >back-trace of when this node was created. This can be done with by >setting the Theano flag 'optimizer=fast_compile'. If that does not work, >Theano optimizations can be disabled with 'optimizer=None'.\n  HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.<\/p>\n<\/blockquote>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33394520","CreationDate":"2015-10-28T15:00:49.660","ParentId":null,"OwnerUserId":"5492709","Title":null,"Body":"<p>I Figured out the problem:\nmy dataset does not have an output for every target, becouse it is too small! There are 17 target outputs but my dataset has only 16 different outputs, and it is missing examples of the 17th output.<\/p>\n\n<p>In order to resolve this problem, just change the softmax with rectify,<\/p>\n\n<p>from this:<\/p>\n\n<pre><code>l_out = lasagne.layers.DenseLayer(\n    l_hid2_drop, num_units=17,\n    nonlinearity=lasagne.nonlinearities.softmax)\n<\/code><\/pre>\n\n<p>to this:<\/p>\n\n<pre><code>l_out = lasagne.layers.DenseLayer(\n    l_hid2_drop, num_units=17,\n    nonlinearity=lasagne.nonlinearities.rectify)\n<\/code><\/pre>\n"}]}
{"QuestionId":33365425,"AnswerCount":1,"Tags":"<python><theano><numerical><deep-learning>","CreationDate":"2015-10-27T10:26:52.563","AcceptedAnswerId":null,"OwnerUserId":67405.0,"Title":"Theano example for solving linear programming","Body":"<p>I have the following problem. Let's say we have  it is an expression value of gene <em>j<\/em> in a sample <em>k<\/em>. It is the average of expression levels across the cell types <em>s_{ij}<\/em>, weighted by respective proportions <em>a_{ki}<\/em> (<em>i = 1 ... N<\/em>,N is the disease type):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iJfh6.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iJfh6.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Generally this can be expressed as matrix form<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7rVZU.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7rVZU.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What I want to do is to solve this equation<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/swFmK.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/swFmK.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>How can it be done using Theano?<\/p>\n","answers":[{"AnswerId":"33393404","CreationDate":"2015-10-28T14:13:29.790","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can do this in theano or not do it in theano. The only thing theano can help you with here is the gradient of the euclidean norm, which it can calculate for you, but which is also easy to write by hand. The algorithm to solve the problem needs to be implemented by yourself. Either you write the lagrangian and then solve the dual problem by gradient ascent and projection onto the constraint, or you solve the primal problem directly by gradient descent and projection onto the constraints. You need to program these optimization steps yourself, which is also the case for any other optimization you do in theano.<\/p>\n"}]}
{"QuestionId":33371954,"AnswerCount":1,"Tags":"<python><python-2.7><pycharm><caffe><pycaffe>","CreationDate":"2015-10-27T15:24:08.953","AcceptedAnswerId":33397590.0,"OwnerUserId":2527680.0,"Title":"PYCHARM import caffe error","Body":"<p>I have tried several ideas to make caffe work on pycharm.<\/p>\n\n<p>I have also added the path to the pycharm as well.\n<a href=\"https:\/\/i.stack.imgur.com\/giOek.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/giOek.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DpoDN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DpoDN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>import caffe doesnt work in pycharm but works in the terminal.<\/p>\n\n<p>the caffe folder is also in my pythonpath.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7Qpfd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7Qpfd.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"33397590","CreationDate":"2015-10-28T17:24:42.990","ParentId":null,"OwnerUserId":"2527680","Title":null,"Body":"<p>Fixed the issue.\nI just had to edit the <code>\/etc\/ld.so.conf<\/code> file and add the lib dir of the Cuda libraries. In my case I added:<\/p>\n\n<pre><code>\/usr\/local\/cuda-7.5\/lib64\n<\/code><\/pre>\n\n<p>Then run <code>sudo ldconfig<\/code> in the terminal and restart Pycharm. It should work.<\/p>\n"}]}
{"QuestionId":33376036,"AnswerCount":1,"Tags":"<cuda><caffe>","CreationDate":"2015-10-27T18:40:07.340","AcceptedAnswerId":33376633.0,"OwnerUserId":5454694.0,"Title":"\/usr\/bin\/ld: cannot find -<library>","Body":"<p>I am getting the error while running make-and-install.sh caffe . I have linked the \/usr\/local\/cuda\/libcudart.so to \/usr\/lib\/libcudart.so and other files also<\/p>\n\n<pre><\/pre>\n\n<p>any kind of help will be appreciated.<\/p>\n\n<p>Is there some problem with the libraries or the path .<\/p>\n\n<p><strong>Edit to add<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33376633","CreationDate":"2015-10-27T19:13:49.820","ParentId":null,"OwnerUserId":"54200","Title":null,"Body":"<p><strong>edit your post with the output of <code>ls -l \/usr\/lib\/cu*<\/code>. Additionally, do <code>file \/usr\/lib\/libcudart.so<\/code> (I'm guessing that your link (ln) is invalid.<\/strong><\/p>\n\n<p>^^^^^ Already done ^^^^^<\/p>\n\n<p>As the broken link is confirmed, go to the <code>\/usr\/lib\/<\/code> directory and link (<code>ln<\/code>) again.<\/p>\n\n<p><code>ln -s \/usr\/local\/cuda\/libcudart.so \/usr\/lib\/libcudart.so<\/code><\/p>\n\n<p>do this step for all desired libraries.\nAs mentioned by @RobertCrovella, it appears there is some library knowledge required too, so add <code>-L\/usr\/local\/cuda\/lib64<\/code>.<\/p>\n\n<p>Now, after typing all of that out, I would recommend not moving\/linking any files, just add <code>-L \/usr\/local\/cuda -L\/usr\/local\/cuda\/lib64<\/code> which will find all of your cuda libraries.<\/p>\n"}]}
{"QuestionId":33380117,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2015-10-27T23:11:19.657","AcceptedAnswerId":33384472.0,"OwnerUserId":388234.0,"Title":"What is a \"label\" in Caffe?","Body":"<p>In Caffe when you are defining your inputs for the NN in the protobuf file, you can input \"data\" and \"label\". I'm guessing label contains the expected output for training data (what it is normally considered the  values in Machine Learning literature).<\/p>\n\n<p>My problem is that in the caffe.proto file, label is defined as a scalar (int or long). At least with data, I can set it to an numpy array, because it takes String values. If I'm training for more than one prediction output, how could I pass it as an array?<\/p>\n\n<p>Or am I mistaken? What is ? What is it for? And how can I pass the  values to caffe?<\/p>\n","answers":[{"AnswerId":"33384472","CreationDate":"2015-10-28T06:39:16.523","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The basic use case of caffe used to be image classification: assigning a single integer label per input image. Thus, the \"datum\" data structure reserves space for a 4D float array (batches of 3 channels images) and an integer \"label\" per image in the batch.<\/p>\n\n<p>This restriction can be easily overcome using HDF5 input data layer.<br>\nSee e.g., <a href=\"https:\/\/stackoverflow.com\/questions\/33140000\/how-to-feed-caffe-multi-label-data-in-hdf5-format\">this answer<\/a>.<\/p>\n"}]}
{"QuestionId":33380897,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2015-10-28T00:30:19.573","AcceptedAnswerId":null,"OwnerUserId":4497662.0,"Title":"Keras + IndexError","Body":"<p>I am very new to keras. Trying to build a binary classifier for an NLP task. (My code is motivated from imdb example - <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_cnn.py\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_cnn.py<\/a>)<\/p>\n\n<p>Below is my code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>When I run model.fit(), I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Can you please help me resolve this ?<\/p>\n","answers":[{"AnswerId":"41652879","CreationDate":"2017-01-14T17:28:29.090","ParentId":null,"OwnerUserId":"2314651","Title":null,"Body":"<p>You need to Pad the imdb sequences you are using, add those lines:<\/p>\n\n<pre><code>from keras.preprocessing import sequence\nTrain_X = sequence.pad_sequences(Train_X, maxlen=maxlen)\nTest_X = sequence.pad_sequences(Test_X, maxlen=maxlen)\n<\/code><\/pre>\n\n<p>Before building the actual model.<\/p>\n"}]}
{"QuestionId":33384406,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe>","CreationDate":"2015-10-28T06:35:26.343","AcceptedAnswerId":33384632.0,"OwnerUserId":290163.0,"Title":"How to force caffe read all training data?","Body":"<p>I'm using <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> and STILL having trouble for input.<\/p>\n\n<p>Here is my solver.prototxt:<\/p>\n\n<pre>\ntrain_net: \"auto_train.prototxt\"\ntest_net: \"auto_test.prototxt\"\ntest_iter: 800\ntest_interval: 20\nbase_lr: 0.01\nmomentum: 0.9\nweight_decay: 0.0005\nlr_policy: \"inv\"\ngamma: 0.0001\npower: 0.75\ndisplay: 100\nmax_iter: 10000\nsnapshot: 5000\nsnapshot_prefix: \"sed\"\nsolver_mode: GPU\n<\/pre>\n\n<p>And this is the running python script:<\/p>\n\n<pre>\nimport os\n\nPROJECT_HOME = '\/home\/romulus\/code\/project\/'\nCAFFE_HOME = '\/home\/romulus\/code\/caffe\/'\nos.chdir(PROJECT_HOME)\n\nimport sys\nsys.path.insert(0, CAFFE_HOME + 'caffe\/python')\nimport caffe, h5py\nfrom pylab import *\nfrom caffe import layers as L, params as P\n\nOUTPUT_DIM = 8\n\ndef net(db, batch_size):\n    n = caffe.NetSpec()\n    n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LEVELDB, source=db,\n                             transform_param=dict(scale=1.\/255), ntop=2)\n    n.ip1 = L.InnerProduct(n.data, num_output=500, weight_filler=dict(type='xavier'))\n    n.relu1 = L.ReLU(n.ip1, in_place=True)\n    n.ip2 = L.InnerProduct(n.relu1, num_output=500, weight_filler=dict(type='xavier'))\n    n.relu2 = L.ReLU(n.ip2, in_place=True)\n    n.ip3 = L.InnerProduct(n.relu2, num_output=OUTPUT_DIM, weight_filler=dict(type='xavier'))\n    n.loss = L.SoftmaxWithLoss(n.ip3, n.label)\n\n    return n.to_proto()\n\n\nwith open('\/home\/romulus\/code\/project\/auto_train.prototxt', 'w') as f:\n    f.write(str(net('\/home\/romulus\/code\/project\/traindb', 64)))\nwith open('\/home\/romulus\/code\/project\/auto_test.prototxt', 'w') as f:\n    f.write(str(net('\/home\/romulus\/code\/project\/testdb', 100)))\n\ncaffe.set_device(0)\ncaffe.set_mode_gpu()\nsolver = caffe.SGDSolver(PROJECT_HOME + 'auto_solver.prototxt')\n\nsolver.net.forward()  # train net\nsolver.test_nets[0].forward()  # test net (there can be more than one)\n\nniter = 500\ntest_interval = 15\ntrain_loss = zeros(niter)\ntest_acc = zeros(int(np.ceil(niter * 1.0 \/ test_interval)))\noutput = zeros((niter, 8, OUTPUT_DIM))\n\nfor it in range(niter):\n    solver.step(1)  # SGD by Caffe\n\n    train_loss[it] = solver.net.blobs['loss'].data\n    solver.test_nets[0].forward(start='ip1')\n    output[it] = solver.test_nets[0].blobs['ip3'].data[:8]\n\n    if it % test_interval == 0:\n        print 'Iteration', it, 'testing...'\n        correct = 0\n\n        for test_it in range(1):\n            solver.test_nets[0].forward()\n            correct += sum(solver.test_nets[0].blobs['ip3'].data.argmax(1)\n                           == solver.test_nets[0].blobs['label'].data)\n        test_acc[it \/\/ test_interval] = correct * 1.0 \/ len(data)\n\n_, ax1 = subplots()\nax2 = ax1.twinx()\nax1.plot(arange(niter), train_loss)\nax2.plot(test_interval * arange(len(test_acc)), test_acc, 'r')\nax1.set_xlabel('iteration')\nax1.set_ylabel('train loss')\nax2.set_ylabel('test accuracy')\n_.savefig('converge.png')\n<\/pre>\n\n<p>Data is generated manually, each data is an 1x256 vector, with all same scale values is 8 * . That means, a datum with label 3 is [ 24, 24, 24, 24, 24...., 24, 24]. I have 8 labels and totally 80000 data.<\/p>\n\n<p>My problem is, if I put data into leveldb with label order like , caffe trains the network well. But if I put ordered by , caffe fails to learn . If I reduce the  in solver.prototxt to 100, caffe will always say output label is 0.<\/p>\n\n<p>It seems like caffe does not read all the training data, but only something in front. But I can't find anything describing it except by training batch. <\/p>\n\n<p>In fact if I increase the training batch size to 80000, caffe trains things well again. Though it is very slow and it's not so-called batch.<\/p>\n\n<p>Can anyone helps? Thank you!<\/p>\n","answers":[{"AnswerId":"33384632","CreationDate":"2015-10-28T06:49:12.060","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It is always a good practice to input the data in a randomized order: If your data is input is a \"sorted\" way, gradients will take very degenerate directions per batch yielding poor training results.<\/p>\n\n<p>The number of training examples caffe \"sees\" during training is <code>max_iter<\/code>*<code>batch_size<\/code>, so if you set these two parameters to exceed the number of training examples you have, you should cover all the data you have during training. <\/p>\n"}]}
{"QuestionId":33416303,"AnswerCount":1,"Tags":"<theano><logistic-regression>","CreationDate":"2015-10-29T13:59:36.893","AcceptedAnswerId":null,"OwnerUserId":5181464.0,"Title":"Logistic Regression in Theano","Body":"<p>I'm new to theano. I've learnt the basics and trying to implement simple models (Logistic Regression etc.).\nThe model is a very simple with 784 (28*28) input units and a 10 unit softmax non-linearity on top of it (Training on MNIST dataset). I'm using <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#tensor.nnet.binary_crossentropy\" rel=\"nofollow\">binary_crossentropy<\/a> as the loss function and using an L2 Regularizer from preventing overfitting.\nBut it seems like the model is still overfitting (by looking at the weights of the model; given below). I tried changing the regularization parameter (lambda) but nothing is working. Where did I go wrong ?\nThanks in advance.<\/p>\n\n<pre><\/pre>\n\n<p>The weights look like this\n<a href=\"http:\/\/i.stack.imgur.com\/4Tkmc.png\" rel=\"nofollow\">The weights of the model<\/a><\/p>\n","answers":[{"AnswerId":"33417534","CreationDate":"2015-10-29T14:52:30.637","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Not sure if this is caused the problem but shouldn't the loss function be a categorical cross-entropy, not a binary cross-entropy?<\/p>\n\n<p>The MNIST task is to classify each image into one class; an image cannot belong to many classes. Binary cross-entropy is the appropriate loss when an item can belong to many classes and categorical cross-entropy is the appropriate loss when an item can belong to only one class.<\/p>\n\n<p>I would also recommend trying this without any regularization initially (remove that component from the loss function altogether while testing), and make sure your learning rate is small enough (e.g. 0.001 should work).<\/p>\n"}]}
{"QuestionId":33416857,"AnswerCount":2,"Tags":"<matrix><machine-learning><neural-network><theano><conv-neural-network>","CreationDate":"2015-10-29T14:23:46.983","AcceptedAnswerId":null,"OwnerUserId":2542738.0,"Title":"TypeError: Cannot convert Type TensorType(float64, vector) (of Variable Subtensor{int64:int64:}.0) into Type TensorType(float64, matrix)","Body":"<p>I am using theano and their LeNet tutorial to train CNN with the dataset of Street View Google Images.<\/p>\n\n<p>I load the dataset:<\/p>\n\n<pre><\/pre>\n\n<p>I get the dimensions printed:<\/p>\n\n<pre><\/pre>\n\n<p>Then I do as in tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>Result printed:<\/p>\n\n<pre><\/pre>\n\n<p>I get the problem here (this is the first function defined, next is validation and traning, just so that the name does not confuse you, and you don't think that everything went OK with similar training and validation functions):<\/p>\n\n<h1>create a function to compute the mistakes that are made by the model<\/h1>\n\n<pre><\/pre>\n\n<p>Precisely, I have an error in line:<\/p>\n\n<pre><\/pre>\n\n<p>The error itself says (afaiu) that I am trying to convert smth to matrix.\nWhere is this happening, I wonder? <\/p>\n\n<pre><\/pre>\n\n<p>AFAIU, the dimensions of y and test_set_y agree. In my code, they are both 1D vectors.\n<strong>Why does slicing cause error? Does slicing create matrix? And more importantly, how do I solve this problem?<\/strong>\nThanks in advance!<\/p>\n\n<p>Thank you in advance for your help!<\/p>\n","answers":[{"AnswerId":"33417439","CreationDate":"2015-10-29T14:47:45.503","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Your input data looks strange and, specifically, the labels are presented as matrices, not vectors.<\/p>\n\n<pre><code>self.train_data_dims (70000, 1024, 1)\nself.train_labels_dims (70000, 1)\nself.valid_data_dims (3250, 1024, 1)\nself.valid_labels_dims (3250, 1)\nself.test_data_dims (26000, 1024, 1)\nself.test_labels_dims (26000, 1)\n<\/code><\/pre>\n\n<p>Why do all of these have an extra dimension of size one at the end?<\/p>\n\n<p>I think you need to change your data loading code so this prints<\/p>\n\n<pre><code>self.train_data_dims (70000, 1024, 1)\nself.train_labels_dims (70000, )\nself.valid_data_dims (3250, 1024, 1)\nself.valid_labels_dims (3250, )\nself.test_data_dims (26000, 1024, 1)\nself.test_labels_dims (26000, )\n<\/code><\/pre>\n\n<p>This ensure that, for example, <code>test_labels<\/code> is a vector instead of a matrix.<\/p>\n"},{"AnswerId":"37821508","CreationDate":"2016-06-14T20:28:18.393","ParentId":null,"OwnerUserId":"2498151","Title":null,"Body":"<p>Flatten your array<\/p>\n\n<pre><code> y_out = theano.shared(np.asarray(y, dtype=theano.config.floatX), borrow=True)\n y_out = y_out.flatten()\n y_out = T.cast(y_out, 'int32')\n<\/code><\/pre>\n"}]}
{"QuestionId":33421217,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-29T17:46:04.453","AcceptedAnswerId":null,"OwnerUserId":427785.0,"Title":"Buidling matrix using scan within theano","Body":"<p>I'm pretty certain this is trivial, but I haven't yet managed to quite get my head around . I want to iteratively build a matrix of values, , where<\/p>\n\n<pre><\/pre>\n\n<p>so you could think of it as a dynamic programming problem. However, I can't even generate the list [1..100] by iterating over the list [1..100] and updating the shared value as I go.<\/p>\n\n<pre><\/pre>\n\n<p>which returns<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33424037","CreationDate":"2015-10-29T20:31:20.543","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There's a few things here that point towards some misunderstandings. <code>scan<\/code> really can be a hard bit of Theano to wrap your head around!<\/p>\n\n<p>Here's some updated code that does what I think you're trying to do, but I wouldn't recommend using this code at all. The basic issue is that you seem to be using a shared variable inappropriately.<\/p>\n\n<pre><code>import numpy as np\nimport theano as T\nimport theano.tensor as TT\n\ndef test():\n    arr = T.shared(np.zeros(100))\n    def grid(idx, arr):\n        return {arr: TT.set_subtensor(arr[idx], idx)}\n\n    _, updates = T.scan(\n        grid,\n        sequences=TT.arange(100),\n        non_sequences=[arr])\n\n    return arr, updates\n\noutputs, updates = test()\nrun = T.function([], outputs=outputs, updates=updates)\nprint run()\nprint outputs.get_value()\n<\/code><\/pre>\n\n<p>This code is changed from the original in two ways:<\/p>\n\n<ol>\n<li><p>The updates from the scan have to be captured (originally discarded) and passed to the <code>theano.function<\/code>'s <code>updates<\/code> parameters. Without this the shared variable won't be updated at all.<\/p><\/li>\n<li><p>The contents of the shared variable need to be examined after the function is executed (see below).<\/p><\/li>\n<\/ol>\n\n<p>This code prints two sets of values. The first is the output of the Theano function from when it's executed. The second is the contents of the shared variable after the Theano function has executed. The Theano function returns the shared variable so you might think that these two sets of values should be the same, but you'd be wrong! No shared variables are updated until <em>after<\/em> all of the function's output values have been computed. So it's only after the function has been executed and we look at the contents of the shared variable that we see the values we expected to see originally.<\/p>\n\n<p>Here's an example of implementing a dynamic programming algorithm in Theano. The algorithm is a simplified version of dynamic time warping which has a lot of similarities to edit distance.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef inner_step(j, c_ijm1, i, c_im1, x, y):\n    insert_cost = tt.switch(tt.eq(j, 0), numpy.inf, c_ijm1)\n    delete_cost = tt.switch(tt.eq(i, 0), numpy.inf, c_im1[j])\n    match_cost = tt.switch(tt.eq(i, 0), numpy.inf, c_im1[j - 1])\n    in_top_left = tt.and_(tt.eq(i, 0), tt.eq(j, 0))\n    min_c = tt.min(tt.stack([insert_cost, delete_cost, match_cost]))\n    c_ij = tt.abs_(x[i] - y[j]) + tt.switch(in_top_left, 0., min_c)\n    return c_ij\n\n\ndef outer_step(i, c_im1, x, y):\n    outputs, _ = theano.scan(inner_step, sequences=[tt.arange(y.shape[0])],\n                             outputs_info=[tt.constant(0, dtype=theano.config.floatX)],\n                             non_sequences=[i, c_im1, x, y], strict=True)\n    return outputs\n\n\ndef main():\n    x = tt.vector()\n    y = tt.vector()\n    outputs, _ = theano.scan(outer_step, sequences=[tt.arange(x.shape[0])],\n                             outputs_info=[tt.zeros_like(y)],\n                             non_sequences=[x, y], strict=True)\n    f = theano.function([x, y], outputs=outputs)\n    a = numpy.array([1, 2, 4, 8], dtype=theano.config.floatX)\n    b = numpy.array([2, 3, 4, 7, 8, 9], dtype=theano.config.floatX)\n    print a\n    print b\n    print f(a, b)\n\n\nmain()\n<\/code><\/pre>\n\n<p>This is highly simplified and I wouldn't recommend using it for real. In general Theano is very bad at doing dynamic programming because <code>theano.scan<\/code> is so slow in comparison to native looping. If you need to propagate gradients through a dynamic program then you may not have any choice but if you don't need gradients you should probably avoid using Theano for dynamic programming.<\/p>\n\n<p>If you want a much more thorough implementation of DTW which gets over some of the performance hits Theano imposes by computing many comparisons in parallel (i.e. batching) then take a look here: <a href=\"https:\/\/github.com\/danielrenshaw\/TheanoBatchDTW\" rel=\"nofollow\">https:\/\/github.com\/danielrenshaw\/TheanoBatchDTW<\/a>.<\/p>\n"}]}
{"QuestionId":33423207,"AnswerCount":2,"Tags":"<python><python-3.x><protocol-buffers><caffe>","CreationDate":"2015-10-29T19:39:30.940","AcceptedAnswerId":null,"OwnerUserId":2097312.0,"Title":"NameError: name 'unicode' is not defined when compile with python3","Body":"<p>I am trying to compiled caffe with python3. After the compilation when I am trying to import caffe I am getting follwing error<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>The issue is Python 3 renamed the unicode type to str, the old str type has been replaced by bytes. <\/li>\n<\/ul>\n\n<p>I tried to modify the proto\/caffe_pb2 but it is not working. <\/p>\n\n<p>Anyone having similar issue with caffe. I have a working caffe with python3 support in my bare metal Now I am trying to dockerize it.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"49509828","CreationDate":"2018-03-27T09:58:15.880","ParentId":null,"OwnerUserId":"4711297","Title":null,"Body":"<p>In Python 3, Str is Unicode by default, \nso, in the line 977, please replace unicode function with str function<\/p>\n\n<pre><code>--&gt; 977       has_default_value=True, default_value=Str(\"constant\", \"utf-8\"),  ## Replace with Str\n<\/code><\/pre>\n\n<p>Hope this helps!!<\/p>\n"},{"AnswerId":"47658325","CreationDate":"2017-12-05T16:29:37.580","ParentId":null,"OwnerUserId":"4342118","Title":null,"Body":"<p>I had the same problem. <code>unicode<\/code> is for Python2. After <code>make pycaffe<\/code> command you can check <code>\/opt\/caffe-1.0\/python\/caffe\/proto\/caffe_pb2.py<\/code> file and if there is any <code>unicode<\/code> in the file the problem still exists. You can delete <code>caffe_pb2.py<\/code> and run <code>make pycaffe<\/code> again. <\/p>\n\n<p>The solution is installing appropriate versions of <code>protoc<\/code> and <code>protobuf<\/code>.<\/p>\n\n<p><strong>A) install protoc >= 2.6.1<\/strong><\/p>\n\n<p>(protoc == 2.5.0 not working)<\/p>\n\n<p><a href=\"https:\/\/cuongba.com\/protocol-buffers-and-rabbitmq-on-ubuntu-14-04-part-1\" rel=\"nofollow noreferrer\">https:\/\/cuongba.com\/protocol-buffers-and-rabbitmq-on-ubuntu-14-04-part-1<\/a><\/p>\n\n<pre><code>wget https:\/\/github.com\/google\/protobuf\/releases\/download\/v2.6.1\/protobuf-2.6.1.tar.gz\ntar -xzvf protobuf-2.6.1.tar.gz\ncd protobuf-2.6.1\nsudo .\/configure\nmake\nmake check\nsudo make install\nsudo ldconfig\nprotoc --version\n<\/code><\/pre>\n\n<p><strong>B) install protobuf>=3.4.0<\/strong><\/p>\n\n<pre><code>pip3 install protobuf&gt;=3.4.0\n<\/code><\/pre>\n"}]}
{"QuestionId":33425203,"AnswerCount":1,"Tags":"<python><matrix><theano>","CreationDate":"2015-10-29T21:47:57.320","AcceptedAnswerId":null,"OwnerUserId":4476617.0,"Title":"Theano ~ most efficient way to flatten then build back a matrix?","Body":"<p>I am trying to compute the hessian on a neural network using theano.<\/p>\n\n<p>However, I cannot compute the hessian on my loss function with respect to a weight matrix . The hessian needs to be computed with respect to a vector .<\/p>\n\n<p>So, a way around this is to flatten the weight matrix into a vector, then compute the hessian, etc.<\/p>\n\n<p>Once I am done updating my weight matrix in flattened form, I need to rebuild my weight matrix with the original dimensions, say <\/p>\n\n<p>What is the most efficient way to do this using theano syntax?<\/p>\n","answers":[{"AnswerId":"33425359","CreationDate":"2015-10-29T21:58:27.997","ParentId":null,"OwnerUserId":"5505084","Title":null,"Body":"<p>It sounds like your looking for \"reshape\", like @Pushkin mentioned.<\/p>\n\n<p>You can use reshape in Theano just like Numpy, but it's with their custom tensor object.<\/p>\n\n<ol>\n<li>Save weight matrix shape to variable. <\/li>\n<li>Flatten and compute.<\/li>\n<li>Call reshape with the previously saved shape variable.<\/li>\n<\/ol>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.reshape\" rel=\"nofollow\">\"Theano - Reshape Tensor\"<\/a><\/p>\n"}]}
{"QuestionId":33428064,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-10-30T03:14:44.770","AcceptedAnswerId":33474303.0,"OwnerUserId":3688217.0,"Title":"Theano: use ifelse to suppress evaluation of scan","Body":"<p>I am trying to implement this paper: <a href=\"http:\/\/arxiv.org\/abs\/1412.1632\" rel=\"nofollow\">Deep Learning for Answer Sentence Selection<\/a> and more accurately the Bigram model using theano's scan function like this:<\/p>\n\n<pre><\/pre>\n\n<p>My problem is that some answers are just one word long so <strong>s<\/strong> is a vector instead of a matrix. This causes problem with the scan function since the +1 tap of the sequence is missing. To handle this I want to use the ifelse statement.<\/p>\n\n<p>Now my first question: <\/p>\n\n<p>Is this even possible? Or will the scan always be evaluated and the ifelse only decides afterwards which value to use?<\/p>\n\n<hr>\n\n<p>second question: <\/p>\n\n<p>how can I make tensor.eq(a,b) return a scalar? Because I get the following error message:<\/p>\n\n<blockquote>\n  <p>TypeError: Condition given to the op has to be a scalar with 0\n  standing for False, anything else for True<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>EDIT some code that triggers an index error because ifelse will not be lazy-evaluated if theano computes the gradients<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33474303","CreationDate":"2015-11-02T09:22:31.613","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I made superficial changes to your code to make executable but it does work, and the <code>ifelse<\/code> <em>is<\/em> lazy (i.e. it only computes the scan if needed). To demonstrate, I added some symbolic print statements which show which value(s) are computed.<\/p>\n\n<pre><code>Tl = theano.shared(numpy.random.randn(2, 3).astype(theano.config.floatX))\nTr = theano.shared(numpy.random.randn(2, 3).astype(theano.config.floatX))\nb = theano.shared(numpy.zeros((2,), dtype=theano.config.floatX))\n\ns = tt.matrix('s')\n\nresults, updates = theano.scan(\n    lambda t_0, t_p1, prior_result, Tl, Tr, b: prior_result + tt.tanh(tt.dot(Tl, t_0) + tt.dot(Tr, t_p1) + b),\n    sequences=dict(input=s, taps=[0, 1]),\n    outputs_info=tt.zeros_like(b, dtype='float64'),\n    non_sequences=[Tl, Tr, b],\n    strict=True\n)\n\ntrue_value = theano.printing.Print('true')(s[0])\nfalse_value = theano.printing.Print('false')(results[-1])\nfinal_result = theano.ifelse.ifelse(tt.eq(s.shape[0], 1), true_value, false_value)\nf = theano.function(inputs=[s], outputs=final_result)\nprint 'a ', f([[1, 2, 3], [4, 5, 6]])\nprint 'b ', f([[1, 2, 3]])\n<\/code><\/pre>\n"}]}
{"QuestionId":33431799,"AnswerCount":3,"Tags":"<lua><torch><zerobrane>","CreationDate":"2015-10-30T08:43:39.740","AcceptedAnswerId":null,"OwnerUserId":5489845.0,"Title":"Lua version in ZeroBraneStudio vs Torch","Body":"<p>I am using ZeroBrane Studio as IDE to code deep learning. I have realized that the models I save when programming in the IDE (using Lua 5.1 as interpreter) do not load well when executing the same loading from Torch7. The same happens when learning from torch (.\/th code.lua) and then trying to load them inside the IDE. I get something like:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know how to check the lua version that torch is using? Any idea on how to workaround this?<\/p>\n\n<p>Thanks!<\/p>\n\n<p><strong>update:<\/strong> It seems that I am indeed using the same Lua version (5.1) in both Torch and ZeroBrane. I still get different behaviour (one successful and the other crashing) when passing through torch.load(). <\/p>\n","answers":[{"AnswerId":"33434838","CreationDate":"2015-10-30T11:23:03.300","ParentId":null,"OwnerUserId":"3979429","Title":null,"Body":"<p>To check the version of Lua that <em>anything<\/em> is running, you would usually print _VERSION. It's a global variable that stores the version of Lua (unless you overwrite it, of course).<\/p>\n\n<pre><code>print(_VERSION)\n<\/code><\/pre>\n\n<p>If this isn't available for some reason, they might state their version on their site (?)<\/p>\n"},{"AnswerId":"33437671","CreationDate":"2015-10-30T13:52:55.697","ParentId":null,"OwnerUserId":"3677376","Title":null,"Body":"<p>Most command line tools on Linux understand the <code>-v<\/code> command line switch (for \"version\"). So do Lua and LuaJIT.<\/p>\n\n<p>To figure out which interpreter is running a particular script, you can scan the <code>arg<\/code> table for the smallest (usually negative) index:<\/p>\n\n<pre><code>local exe, i = arg[ 0 ], -1\nwhile arg[ i ] do\n  exe, i = arg[ i ], i-1\nend\nprint( exe )\n<\/code><\/pre>\n\n<p>Or (on Linux) you can look into the <code>\/proc<\/code> file system while your script is running:<\/p>\n\n<pre><code>ls -l \/proc\/4425\/exe\n<\/code><\/pre>\n\n<p>(substitute <code>4425<\/code> with real process ID).<\/p>\n\n<p>Judging from the error message the interpreter used in ZeroBrane Studio seems to be <code>\/opt\/zbstudio\/bin\/linux\/x64\/lua<\/code> in your case.<\/p>\n"},{"AnswerId":"33495984","CreationDate":"2015-11-03T09:51:38.740","ParentId":null,"OwnerUserId":"5489845","Title":null,"Body":"<p>@siffiejoe: thanks for posing your question regarding versions, it gave me the correct directions to explore. <\/p>\n\n<pre><code>\/opt\/zbstudio\/bin\/linux\/x64\/lua version is LuaJIT 2.0.2\n\"lua\" command alone points to \/usr\/bin\/lua, and it is Lua 5.1.5\n~\/torch\/install\/share\/lua\/5.1 seemed to contain Lua 5.1\n~\/torch\/install\/bin\/luajit is 2.1.0-alpha\n<\/code><\/pre>\n\n<p>So after realizing that terminal \"th\" is using LuaJit 2.1.0 all I had to do is create a user.lua in ZeroBrane and add the line \"path.lua = \"~\/torch\/install\/bin\/luajit\". Now ZB is using the same luajit interpreter as th.<\/p>\n\n<p>Thanks all for your suggestions.<\/p>\n"}]}
{"QuestionId":33436403,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-10-30T12:49:53.063","AcceptedAnswerId":33461342.0,"OwnerUserId":133374.0,"Title":"dimshuffle equivalent function in Numpy","Body":"<p>I often use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.dimshuffle\"><\/a>. Is there an equivalent function for Numpy?<\/p>\n\n<p>I guess I could do the same via several  and  (for broadcast dimensions),  but is there some simpler or more direct way, just like ?<\/p>\n","answers":[{"AnswerId":"33461342","CreationDate":"2015-11-01T10:44:35.623","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The function <code>numpy.transpose<\/code> permits any permutation of the axes of an array.<\/p>\n\n<p>The variety <code>array.T<\/code> is a special case of this, corresponding to <code>array.transpose()<\/code> without arguments, which defaults to <code>array.transpose(range(array.ndim)[::-1])<\/code>.<\/p>\n\n<p><code>numpy.swapaxes<\/code> is <code>numpy.transpose<\/code> restricted to permutations of two axes.<\/p>\n\n<p><code>theano.tensor.dimshuffle<\/code> essentially corresponds to <code>numpy.transpose<\/code>, but in addition, it permits the creation of new axes of length 1 for broadcasting, by adding <code>'x'<\/code> wherever an axis should be created. In numpy, this can be achieved using a combination of <code>transpose<\/code> and <code>reshape<\/code>.<\/p>\n\n<p>Note that in numpy, care is taken to make <code>transpose<\/code> return a view on the data whenever possible. In theano this is probably the case, too, but may depend on how the code is optimized.<\/p>\n"}]}
{"QuestionId":33440453,"AnswerCount":2,"Tags":"<python><gcc><anaconda><theano>","CreationDate":"2015-10-30T16:16:18.200","AcceptedAnswerId":33473701.0,"OwnerUserId":1474615.0,"Title":"anaconda python error importing theano","Body":"<p>I'm quite new in python and of course I'm also new with Theano. I'm trying to use it under windows along with anaconda python. I have installed all the compulsory requirements (except CUDA since on this laptop I don't have a NVIDIA GPU).\nI installed the same GCC and set the path as suggested in the walkthrough page.\nStill I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any Suggestions?<\/p>\n\n<p>This \/ \\ thing in <\/p>\n\n<pre><\/pre>\n\n<p>sounds weird to me...<\/p>\n","answers":[{"AnswerId":"33440547","CreationDate":"2015-10-30T16:21:54.667","ParentId":null,"OwnerUserId":"5504265","Title":null,"Body":"<p>Have you read this link?<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2406\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/2406<\/a><\/p>\n\n<p>Hope you find solution from that. <\/p>\n"},{"AnswerId":"33473701","CreationDate":"2015-11-02T08:45:36.410","ParentId":null,"OwnerUserId":"1474615","Title":null,"Body":"<p>As suggested by Daniel in a comment, I did not run the command:<\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n\n<p>After that I was able to import Theano coerrectly. Solved! :)<\/p>\n"}]}
{"QuestionId":33443020,"AnswerCount":1,"Tags":"<deep-learning><caffe><lmdb>","CreationDate":"2015-10-30T18:52:57.803","AcceptedAnswerId":33447280.0,"OwnerUserId":2832033.0,"Title":"Merge two LMDB databases for feeding to the network (caffe)","Body":"<p>Here are two LMDB databases. Is there any way to merge these two databases and feed it to network using caffe? <\/p>\n","answers":[{"AnswerId":"33447280","CreationDate":"2015-10-31T01:52:41.810","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>Simply write a script using the python lmdb interface. Something like:<\/p>\n\n<pre><code>import lmdb\n\nenv = lmdb.open(\"path\/to\/lmdbFile\")\ntxn = env.begin(write=True)\n\ndatabase1 = txn.cursor(\"db1Name\")\ndatabase2 = txn.cursor(\"db2Name\")\n\nenv.open_db(key=\"newDBName\", txn=txn)\nnewDatabase = txt.cursor(\"newDBName\")\n\nfor (key, value) in database1:\n    newDatabase.put(key, value)\n\nfor (key, value) in database2:\n    newDatabase.put(key, value)\n<\/code><\/pre>\n\n<p>or you could just as simply add one to the other by:<\/p>\n\n<pre><code>for (key, value) in database2:\n    database1.put(key, value)\n<\/code><\/pre>\n"}]}
{"QuestionId":33446612,"AnswerCount":0,"Tags":"<matlab><deep-learning><caffe><multi-gpu><matcaffe>","CreationDate":"2015-10-31T00:07:58.190","AcceptedAnswerId":null,"OwnerUserId":3968758.0,"Title":"Caffe: Multi-GPU support with Matlab (Matcaffe)","Body":"<p>Caffe is a deep learning framework that also includes a Matlab interface (Matcaffe). While Caffe supports multiple GPUs it seems that Matcaffe does not at the moment. Does anyone know of a workaround?<\/p>\n","answers":[]}
{"QuestionId":33448650,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2015-10-31T06:14:26.243","AcceptedAnswerId":33566538.0,"OwnerUserId":2832033.0,"Title":"Is it necessary to apply mean-subtraction on FC7 features of an CNN before feeding into L-SVM?","Body":"<p>I want to use AlexNet off-the-self features to train an L-SVM on my own dataset. I found that it is recommended to zero-mean features before feeding them into SVM classifier. I just want to know whether it is necessary to do so when using CNNs features or not, since in the first data layer of an CNN, mean-subtraction is once applied on raw input images. <\/p>\n","answers":[{"AnswerId":"33566538","CreationDate":"2015-11-06T12:11:17.530","ParentId":null,"OwnerUserId":"5170044","Title":null,"Body":"<p>It is <strong>not necessary<\/strong> to subtract mean from FC7 features before using them in to train SVM, but it would not hurt. I'm not exactly sure what L-SVM is, but this answer should hold for any SVM if the data is <em>\"somewhere around zero\"<\/em>. <\/p>\n\n<p>On the other hand, it is a good idea to <strong>normalize<\/strong> the feature vectors (FC7 activations) to <strong>unit length<\/strong> (e.g. divide them by their L1 or L2 norm). Such normalization often improves classification results, and it makes tuning of SVM hyperparameters much easier.<\/p>\n\n<p>The decision how to normalize FC7 activations should not depend on the mean subtraction of input images in any way. The input mean subtraction possibly improves training of the network, but it should not affect what the network computes when fully trained. Keep in mind that the image mean is constant and computed on the training set - it affects all images in the same way.<\/p>\n"}]}
{"QuestionId":33459587,"AnswerCount":0,"Tags":"<torch><lstm>","CreationDate":"2015-11-01T05:45:27.820","AcceptedAnswerId":null,"OwnerUserId":3612823.0,"Title":"Training LSTM Model in Torch and Data Structure","Body":"<p>I am new to Torch and I am trying to implement LSTM.  Below is the code that I have.  Is this the correct way to implement an LSTM using the rnn library or is there more I need to do?  <\/p>\n\n<p>Also, what should my data look like for inputting into an LSTM? Just ordered with 2 inputs and 1 output?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33463217,"AnswerCount":1,"Tags":"<python><caffe><lmdb>","CreationDate":"2015-11-01T14:27:57.747","AcceptedAnswerId":null,"OwnerUserId":824624.0,"Title":"can't get access to python lmdb , ' object has no attribute 'Environment''","Body":"<p>I am using lmdb python to load the data<\/p>\n\n<pre><\/pre>\n\n<p>I got a error saying 'img_env = lmdb.Environment(img_db_fn, map_size=1099511627776) AttributeError: 'module' object has no attribute 'Environment''<\/p>\n","answers":[{"AnswerId":"34517039","CreationDate":"2015-12-29T18:40:53.307","ParentId":null,"OwnerUserId":"1057460","Title":null,"Body":"<p>I <strong>guess<\/strong> you might use <code>lmdb<\/code> as one of your personal module names and thus it conflicts with the standard <code>lmdb<\/code> module<\/p>\n"}]}
{"QuestionId":33464544,"AnswerCount":0,"Tags":"<ios><lua><torch>","CreationDate":"2015-11-01T16:42:12.850","AcceptedAnswerId":null,"OwnerUserId":5513021.0,"Title":"Access files in Torch iOS","Body":"<p>I have an image file and I want to access it in my Lua file. I already put the file in the project folder. However the call <\/p>\n\n<p>local path = system.pathForFile(\"xxx\", system.DocumentsDirectory) fails to work. <\/p>\n\n<p>How should I do it?<\/p>\n","answers":[]}
{"QuestionId":33473848,"AnswerCount":1,"Tags":"<python><windows><theano>","CreationDate":"2015-11-02T08:54:29.237","AcceptedAnswerId":null,"OwnerUserId":1934733.0,"Title":"Windows missing Python.h","Body":"<p>I'm trying to run a sample Theano code that uses GPU on windows.\nMy python (with python-dev and Theano and all required libraries) was installed from Anaconda.\nThis is the error I run into:<\/p>\n\n<p>Cannot open include file: 'Python.h': No such file or directory<\/p>\n\n<p>Python.h is actually in c:\/\/Anaconda\/include\/<\/p>\n\n<p>I'm guessing that I should add that directory to some environmental variable, but I don't know which.<\/p>\n","answers":[{"AnswerId":"46394599","CreationDate":"2017-09-24T20:35:13.613","ParentId":null,"OwnerUserId":"3657967","Title":null,"Body":"<p>If you are using Visual Studio in Windows, right-click on your project in the Solution Explorer and navigate as follows: Properties -> C\/C++ -> General -> Additional Include Directories -> Add C:\/Anaconda\/include\/ (or wherever your Anaconda install is located)<\/p>\n"}]}
{"QuestionId":33474424,"AnswerCount":1,"Tags":"<python><neural-network><theano><keras>","CreationDate":"2015-11-02T09:28:45.183","AcceptedAnswerId":33474623.0,"OwnerUserId":4842101.0,"Title":"Keras load weights of a neural network \/ error when predicting","Body":"<p>I'm using the Keras library to create a neural network. I have a iPython Notebook in order to load the training data, initializing the network and \"fit\" the weights of the neural network.\nFinally, I save the weights using the save_weights() method.\nCode is below :<\/p>\n\n<pre><\/pre>\n\n<p>Then, in another iPython Notebook, I just want to use these weights and predict some outputs values given inputs. I initialize the same neural network, and then load the weights.<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that apparently, the load_weights method is not enough to have a functional model. I'm getting an error :<\/p>\n\n<pre><\/pre>\n\n<p>Any idea?\nThanks a lot.<\/p>\n","answers":[{"AnswerId":"33474623","CreationDate":"2015-11-02T09:40:02.290","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You need to call <code>model.compile<\/code>. This can be done either before or after the <code>model.load_weights<\/code> call but must be after the model architecture is specified and before the <code>model.predict<\/code> call.<\/p>\n"}]}
{"QuestionId":33475428,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-11-02T10:26:36.437","AcceptedAnswerId":33475706.0,"OwnerUserId":4116272.0,"Title":"Troubles with simple accumulation with Theano scan function","Body":"<p>I am trying to accumulate matrix values while multiplying with the step: \nres = sum_i(i * a). My code looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This outputs:<\/p>\n\n<pre><\/pre>\n\n<p>While I would expect this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33475706","CreationDate":"2015-11-02T10:41:53.730","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The output is (perhaps unsurprising?) correct. The reason you get this output is as follows:<\/p>\n\n<p>In the first iteration,<\/p>\n\n<pre><code>res = 0\nstep = [[0, 0], [0, 0]]\nx = [[0, 0], [2, 2]]\n<\/code><\/pre>\n\n<p>and so<\/p>\n\n<pre><code>res + step * x = 0 + [[0, 0], [0, 0]] * [[0, 0], [2, 2]]\n               = 0 + [[0, 0], [0, 0]]\n               = [[0, 0], [0, 0]]\n<\/code><\/pre>\n\n<p>In the second iteration,<\/p>\n\n<pre><code>res = 1\nstep = [[0, 0], [0, 0]]\nx = [[0, 0], [2, 2]]\n<\/code><\/pre>\n\n<p>and so<\/p>\n\n<pre><code>res + step * x = 1 + [[0, 0], [0, 0]] * [[0, 0], [2, 2]]\n               = 1 + [[0, 0], [0, 0]]\n               = [[1, 1], [1, 1]]\n<\/code><\/pre>\n\n<p>Note that the 1 is broadcasted to same shape as the matrix that results from the elementwise multiplication of <code>step<\/code> and <code>x<\/code>.<\/p>\n"}]}
{"QuestionId":33478998,"AnswerCount":1,"Tags":"<python><scikit-learn><neural-network><theano><scikits>","CreationDate":"2015-11-02T13:42:13.483","AcceptedAnswerId":null,"OwnerUserId":3763498.0,"Title":"RuntimeError: NaN in hidden0_W, No handlers could be found for logger \"sknn\" scikit-neuralnetwork","Body":"<p>I am trying to implement the Neural Network Classifier in Python. Here is the snippet of the function<\/p>\n\n<pre><\/pre>\n\n<p>And after executing the code, I got the following error.<\/p>\n\n<pre><\/pre>\n\n<p>I have googled the error and there some suggestions to lower the learning rate by 10x. But even after that the problem persists. Plus no other solution seems o be working well. Neither I'm able to understand in what context is this error related to. Any help would really be helpful.<\/p>\n","answers":[{"AnswerId":"34539252","CreationDate":"2015-12-31T00:13:34.853","ParentId":null,"OwnerUserId":"5096153","Title":null,"Body":"<p>I also got the \"No handlers found .... Training diverged and returned NaN\" erroring out on the fit() command. You need to do two things. <\/p>\n\n<p>First, add the following to the top of your code:<\/p>\n\n<pre><code>import logging\nlogging.basicConfig()\n<\/code><\/pre>\n\n<p>Second, you need to normalize your data so that the input is [-1,1] or [0,1].<\/p>\n\n<p>I was training my neural network to classify images and once I normalize the image values using:<\/p>\n\n<pre><code>#normalize data\nimg *= (1.0\/img.max())\n<\/code><\/pre>\n\n<p>and it worked. Best wishes, hope this helps!<\/p>\n"}]}
{"QuestionId":33479546,"AnswerCount":1,"Tags":"<theano><conv-neural-network><training-data><large-data>","CreationDate":"2015-11-02T14:09:08.953","AcceptedAnswerId":null,"OwnerUserId":1317209.0,"Title":"Training Methodology of CNN in theano with large scale data","Body":"<p>I am training a CNN with 1M images with theano. Now I am puzzled on how to prepare the training data.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li><p>When the images resize to 64*64*3, the size of whole data is about 100G. Should I save the data into a single npy file or some smaller files? which one is efficient?<\/p><\/li>\n<li><p>How to decide the number of parameters of the CNN? How about 1M\/10 = 100K?<\/p><\/li>\n<li><p>Should I limit the memory cost of a training block and the CNN parameters less than GPU memory?<\/p><\/li>\n<\/ol>\n\n<p>My computer is with 16G memory and GPU Titian.<\/p>\n\n<p>Thank you very much.<\/p>\n","answers":[{"AnswerId":"33986786","CreationDate":"2015-11-29T19:05:31.547","ParentId":null,"OwnerUserId":"2805751","Title":null,"Body":"<p>If you're using a NN framework like pylearn2, lasagne, Keras, etc, check the docs to see if there are guidelines for iterating batches off disk from an hdf5 store or similar.  <\/p>\n\n<p>If there's nothing and you don't want to roll your own, the <a href=\"https:\/\/github.com\/mila-udem\/fuel\" rel=\"nofollow\">fuel package<\/a> provides lots of helpful data iteration schemes that can be adapted to models in theano (and probably most of the frameworks; there's a good tutorial in the fuel repository).<\/p>\n\n<p>As for the parameters, you'll have to cross validate to figure out the best parameters for your data.<\/p>\n\n<p>And yes, the model size + minibatch size + dropout mask for the batch has to be under the available vram.<\/p>\n"}]}
{"QuestionId":33495280,"AnswerCount":1,"Tags":"<lua><machine-learning><deep-learning><torch>","CreationDate":"2015-11-03T09:16:46.223","AcceptedAnswerId":null,"OwnerUserId":5519228.0,"Title":"overriding updateGradInput method in torch is not working for my self defined module","Body":"<p>I implemented a module called sharedModule for my own use. And I have overriden the updateGradInput method below:<\/p>\n\n<pre><\/pre>\n\n<p>However, when i call backward method on my module, the self.gradInput field is not updated, what's the problem, any one can help me out?<\/p>\n","answers":[{"AnswerId":"33501856","CreationDate":"2015-11-03T14:41:50.420","ParentId":null,"OwnerUserId":"5519228","Title":null,"Body":"<p>backward method will call two methods named updateGradInput and accGradParams, turns out that the error occurs in accGradParams because of typos <\/p>\n"}]}
{"QuestionId":33498836,"AnswerCount":1,"Tags":"<python-3.x><theano><deep-learning><openblas>","CreationDate":"2015-11-03T12:12:16.530","AcceptedAnswerId":null,"OwnerUserId":1935415.0,"Title":"installing theano \"blas error\"","Body":"<pre><\/pre>\n\n<p>When importing theano, python gives this warning.\nI setup theano according to this link. <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a><\/p>\n\n<p>Numpy and Scipy imports are OK. but i'm stuck with BLAS.?<\/p>\n\n<p>I clone to my computer OpenBLAS from github. \nhow i add blas to environment variables ( \"path\" ) ?<\/p>\n","answers":[{"AnswerId":"42451839","CreationDate":"2017-02-25T04:08:55.173","ParentId":null,"OwnerUserId":"6302808","Title":null,"Body":"<p>Apparently you have error from scipy.linalg.blas, try to install the package from\n, to install scipy please make sure you have install <\/p>\n\n<blockquote>\n  <p>numpy+mkl<\/p>\n<\/blockquote>\n\n<p>You can find it on <a href=\"http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#numpy\" rel=\"nofollow noreferrer\">http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#numpy<\/a><\/p>\n\n<hr>\n\n<p>The other problem you've mentioned is that g++ compiler isn't detected, theano requires g++ (Linux and Windows) in order to compile generated C code.\n(just google it deep learning theano setup might help). <\/p>\n\n<p>The tutorial part of installing MinGW:<\/p>\n\n<ol>\n<li>In C:\\ create file folder\u2019s name called Program <\/li>\n<li>Install MinGW under this file folder \u201dProgram\u201d <\/li>\n<li>Install msys under the path included bin folder.<\/li>\n<li>run <code>sh \/postinstall\/pi.sh<\/code> in mysy and the direction you've just install MinGW is supposed to be <code>C:\/Program\/mingw-w64\/mingw64<\/code>(depend on the install direction)<\/li>\n<li>follow the step on <a href=\"http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano\" rel=\"nofollow noreferrer\">http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano<\/a> and you might have your code works!<\/li>\n<\/ol>\n"}]}
{"QuestionId":33503025,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-03T15:32:27.877","AcceptedAnswerId":null,"OwnerUserId":5520753.0,"Title":"T.hessian gives NotImplementedError() in theano","Body":"<p>How it can be that it works <\/p>\n\n<pre><\/pre>\n\n<p>whereas this   <\/p>\n\n<pre><\/pre>\n\n<p>gives NotImplementedError()\nmay it be that the problem in such cost function: <\/p>\n\n<pre><\/pre>\n\n<p>Here y is the vector of class labels from 0 to n-1 and <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33503824","CreationDate":"2015-11-03T16:10:04.617","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I am unable to reproduce this problem with the limited code that has been provided. However, here is a fully working demo of <code>T.grad<\/code> and <code>T.hessian<\/code>.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as T\n\nx = T.matrix()\nw_flat = theano.shared(numpy.random.randn(3, 2).astype(theano.config.floatX).flatten())\nw = w_flat.reshape((3, 2))\ncost = T.pow(theano.dot(x, w), 2).sum()\ng_w = T.grad(cost=cost, wrt=[w])\nh_w = T.hessian(cost=cost, wrt=[w_flat])\nf = theano.function([x], outputs=g_w + h_w)\nfor output in f(numpy.random.randn(4, 3).astype(theano.config.floatX)):\n    print output.shape, '\\n', output\n<\/code><\/pre>\n\n<p>Note that the <code>wrt<\/code> value for <code>T.hessian<\/code> needs to be a vector.<\/p>\n"}]}
{"QuestionId":33510773,"AnswerCount":1,"Tags":"<python><arrays><numpy><lua><torch>","CreationDate":"2015-11-03T23:05:18.013","AcceptedAnswerId":null,"OwnerUserId":2006700.0,"Title":"How to quickly convert a returned Python-in-Lua numpy array into a Lua Torch Tensor?","Body":"<p>I have a Python function that returns a multi-dimensional numpy array. I want to call this Python function from Lua and get the data into a Lua Torch Tensor as quickly as possible. I have a solution that works quite slowly and am looking for a way that is significantly faster (order of 10fps or more). I'm not sure if this is possible. <\/p>\n\n<p>I believe this will be of use to others considering the growing popularity of Facebook backed Torch and the extensive easy-to-use image-processing tools available in Python of which Lua lacks. <\/p>\n\n<p>I am using the Bastibe fork of lunatic-python in order to call a Python function from Lua. With aid from this previous <a href=\"https:\/\/stackoverflow.com\/questions\/20400390\/how-to-call-a-python-function-from-lua\">question<\/a> and this <a href=\"https:\/\/labix.org\/lunatic-python#head-d0597456187b5ea4006f9427346633f765432975\" rel=\"nofollow noreferrer\">documentation<\/a>, I have come up with some code that works, but is far too slow. I am using Lua 5.1 and Python 2.7.6 and can update these if necessary.<\/p>\n\n<p>Lua Code: \"test_lua.lua\"<\/p>\n\n<pre><\/pre>\n\n<p>Python Code: \"test_python.py\"<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36896347","CreationDate":"2016-04-27T17:03:38.103","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>Try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a>, it has a lua engine in python and be able to share numpy memory with torch, so it's very fast, here is the code for your case:<\/p>\n\n<pre><code>import numpy\nimport Image\nimport lutorpy as lua\n\ngetImage = numpy.asarray(Image.open(\"image.jpg\"))\na = torch.fromNumpyArray(getImage)\n\n# now you can use your image as torch Tensor\n# for example: use SpatialConvolution from nn to process the image\nrequire(\"nn\")\nn = nn.SpatialConvolution(1,16,12,12)\nres = n._forward(a)\nprint(res._size())\n\n# convert back to numpy array\noutput = res.asNumpyArray()\n<\/code><\/pre>\n"}]}
{"QuestionId":33519182,"AnswerCount":3,"Tags":"<python><cuda><gpu><theano>","CreationDate":"2015-11-04T10:17:13.463","AcceptedAnswerId":33519932.0,"OwnerUserId":1935415.0,"Title":"How can I run theano on GPU","Body":"<p>If I run the following code with python 3.5<\/p>\n\n<pre><\/pre>\n\n<p>I get the output<\/p>\n\n<pre><\/pre>\n\n<p>it says if the times are close, it means that I am running on my CPU.<\/p>\n\n<p>How can I run this code on my GPU?<\/p>\n\n<p><strong>NOTE:<\/strong><\/p>\n\n<ul>\n<li>I have a workstation with Nvidia Quadro k4200.<\/li>\n<li>I have installed Cuda toolkit <\/li>\n<li>I have successfully worked an cuda vectorAdd sample project on VS2012.<\/li>\n<\/ul>\n","answers":[{"AnswerId":"35635753","CreationDate":"2016-02-25T18:48:19.163","ParentId":null,"OwnerUserId":"627042","Title":null,"Body":"<p>Alternatively, if you want to use the GPU programattically:<\/p>\n\n<pre><code>import theano.sandbox.cuda\ntheano.sandbox.cuda.use(\"gpu0\")\n<\/code><\/pre>\n\n<p>You should see a message like this:<\/p>\n\n<pre><code>Using gpu device 0: Tesla K80\n<\/code><\/pre>\n\n<p>Useful if the environment you are running in isn't easy to configure.<\/p>\n"},{"AnswerId":"34647854","CreationDate":"2016-01-07T05:18:38.977","ParentId":null,"OwnerUserId":"2004070","Title":null,"Body":"<p>If you are on Linux, create a .theanorc file in your home folder and add the following to set up theano to run on GPU.<\/p>\n\n<pre><code>[global]\ndevice = gpu\nfloatx = float32\n<\/code><\/pre>\n"},{"AnswerId":"33519932","CreationDate":"2015-11-04T10:50:58.883","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You configure Theano to use a GPU by specifying the <code>device=gpu<\/code> in Theano's config. There are two principle methods for setting the config: (1) in the <code>THEANO_FLAGS<\/code> environment variable, or (2) via the .theanorc file. Both methods, and all of Theano's configuration flags, are <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\">documented<\/a>.<\/p>\n\n<p>You will know that Theano is using the GPU if, after calling <code>import theano<\/code> you see a message that looks something like this<\/p>\n\n<pre><code>Using gpu device 0: GeForce GT 640 (CNMeM is disabled)\n<\/code><\/pre>\n\n<p>The details may vary for you but if no message appears at all then Theano is using the CPU only.<\/p>\n\n<p>Note also that even if you see the GPU message, your particular computation graph may not run on the GPU. To see which parts of your computation are running on the GPU print its compiled and optimized graph<\/p>\n\n<pre><code>f = theano.function(...)\ntheano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>Operations that start with the prefix 'Gpu' will run on the GPU. Operations that do not have that prefix to their name will run on the CPU.<\/p>\n"}]}
{"QuestionId":33523860,"AnswerCount":1,"Tags":"<python><python-2.7><theano>","CreationDate":"2015-11-04T14:03:51.650","AcceptedAnswerId":33524396.0,"OwnerUserId":1718152.0,"Title":"Minimize cosine distance theano","Body":"<p>I have a simple and working multi-layer perceptron in Theano, with 1 hidden layer and 1 regression layer with 2 outputs. In the regression layer a mean square error function is defined which is used as cost function. However, during learning I now want to minimize the cosine distance between two vectors, so I wanted to use the cosine distance as cost function instead. Below some relevant parts of my current implementation.<\/p>\n\n<pre><\/pre>\n\n<p>If I change the cost function from  to  I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't see why the cost (function) would not be scalar. Just for testing I tried:<\/p>\n\n<pre><\/pre>\n\n<p>The model builds then, but I get a dimension mismatch during training.<\/p>\n\n<pre><\/pre>\n\n<p>I think the problem is that I don't see how my cosine distance function is different from my mean square error function in Theano. What do I miss here?<\/p>\n","answers":[{"AnswerId":"33524396","CreationDate":"2015-11-04T14:27:26.543","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The difference is that in your <code>mse<\/code> function computes the <code>T.mean<\/code> without specifying an axis so gives the mean over all entries in the tensor, whatever shape that tensor might be. In comparison your first <code>cos<\/code> function does not aggregate at all so the return value will have the same shape as <code>T.dot(y,self.y_pred)<\/code>, i.e. not a scalar. You sum in the second version of the <code>cos<\/code> function which produces the required scalar but this may not be computing what you want it to compute, depending on the semantics of the shape of your inputs.<\/p>\n\n<p>The second error is probably due to the error in your <code>cos<\/code> function: you don't want to perform a dot product in numerator, i.e. <code>T.dot(y,self.y_pred)<\/code>, instead you want the element-wise multiplication, e.g. <code>y * self.y_pred<\/code><\/p>\n\n<p>Here's my code for doing various distances in Theano. Note that the <code>_magnitude<\/code> and <code>cosine<\/code> functions include tweaks that help avoid <code>NaN<\/code> or out of range values. These kinds of problems can occur in either the forward pass or backward pass (gradient).<\/p>\n\n<pre><code>import numpy\nimport theano.tensor as tt\n\n\ndef _squared_magnitude(x):\n    return tt.sqr(x).sum(axis=-1)\n\n\ndef _magnitude(x):\n    return tt.sqrt(tt.maximum(_squared_magnitude(x), numpy.finfo(x.dtype).tiny))\n\n\ndef cosine(x, y):\n    return tt.clip((1 - (x * y).sum(axis=-1) \/ (_magnitude(x) * _magnitude(y))) \/ 2, 0, 1)\n\n\ndef euclidean(x, y):\n    return _magnitude(x - y)\n\n\ndef squared_euclidean(x, y):\n    return _squared_magnitude(x - y)\n<\/code><\/pre>\n"}]}
{"QuestionId":33530261,"AnswerCount":3,"Tags":"<visual-studio-2013><caffe>","CreationDate":"2015-11-04T19:19:35.333","AcceptedAnswerId":33539031.0,"OwnerUserId":5431576.0,"Title":"How to build Caffe library on Windows (Visual Studio 2013)?","Body":"<p>have anyone successfully using Caffe library in Visual Studio development? If yes, can you make a short overview (step by step). <\/p>\n\n<p>What I've already done:\n - download Caffe from Github\n - building Boost\n - building Cuda Toolkit 7.5\n - building gflags\n - building glog<\/p>\n\n<p>Still have to do:\n- hdf5\n- protobuf\n- GLog \n....<\/p>\n\n<p>Have anyone any instructions?<\/p>\n","answers":[{"AnswerId":"33539031","CreationDate":"2015-11-05T07:42:44.777","ParentId":null,"OwnerUserId":"2832033","Title":null,"Body":"<p>It is not recommended to build and use Caffe on Windows,but if you're insisting on that you can find this perfect step-by-step guide so helpful. <\/p>\n\n<p><a href=\"https:\/\/initialneil.wordpress.com\/2015\/01\/11\/build-caffe-in-windows-with-visual-studio-2013-cuda-6-5-opencv-2-4-9\/\" rel=\"nofollow\">Build Caffe in Windows with Visual Studio 2013 + CUDA 6.5 + OpenCV 2.4.9<\/a><\/p>\n"},{"AnswerId":"35622708","CreationDate":"2016-02-25T09:06:08.423","ParentId":null,"OwnerUserId":"4382376","Title":null,"Body":"<p><strong>UPDATE:<\/strong>\nSince microsoft <a href=\"https:\/\/github.com\/Microsoft\/caffe\" rel=\"nofollow\">adopted the porting of CAFFE to windows<\/a> , there is now an even easier way to build CAFFE for windows with visual studio. The entire process is depicted in <a href=\"https:\/\/www.youtube.com\/watch?v=nrzAF2sxHHM\" rel=\"nofollow\">this 5 min video!<\/a><\/p>\n\n<p><s>\ni saw the answer by Saeed, and i think i have something easier to propose.<\/p>\n\n<p>Follow the built found <a href=\"https:\/\/github.com\/happynear\/caffe-windows\" rel=\"nofollow\">here<\/a> for a very easy build of Caffe on windows. However, you always have to check it for changes, since their code updates regularly, but their documentation stays behind. In my case (and many other co-workers) their code builts as-is, right out of the box!<\/s><\/p>\n"},{"AnswerId":"52569298","CreationDate":"2018-09-29T14:25:55.783","ParentId":null,"OwnerUserId":"3816017","Title":null,"Body":"<p>You can also use Collective Knowledge workflow framework to automatically install Caffe on Windows, Linux or MacOS from command line in just a few steps. See<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/5654#issuecomment-372123226\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/5654#issuecomment-372123226<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/dividiti\/ck-caffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dividiti\/ck-caffe<\/a><\/li>\n<\/ul>\n\n<p>Basically, you just need to do the following:<\/p>\n\n<pre><code>   pip install ck\n   ck pull repo --url=https:\/\/github.com\/dividiti\/ck-caffe\n   ck install package:lib-caffe-bvlc-master-cudnn-universal --env.CAFFE_BUILD_PYTHON=ON\n<\/code><\/pre>\n\n<p>CK will detect or rebuild all necessary dependencies on Windows (HDF5, Boost, OpenCV, etc). You can then use Caffe from virtual environment as follows:<\/p>\n\n<pre><code>   ck show env\n   ck virtual env --tags=lib,caffe\n\n   python\n   &gt; import caffe\n<\/code><\/pre>\n\n<p>Hope it's of any help!<\/p>\n"}]}
{"QuestionId":33536182,"AnswerCount":2,"Tags":"<python><sentiment-analysis><lstm><keras>","CreationDate":"2015-11-05T03:30:30.077","AcceptedAnswerId":34154972.0,"OwnerUserId":3200890.0,"Title":"Testing the Keras sentiment classification with model.predict","Body":"<p>I have trained the imdb_lstm.py on my PC.\nNow I want to test the trained network by inputting some text of my own. How do I do it?\nThank you!<\/p>\n","answers":[{"AnswerId":"34154972","CreationDate":"2015-12-08T11:40:59.090","ParentId":null,"OwnerUserId":"4583366","Title":null,"Body":"<p>So what you basically need to do is as follows:<\/p>\n\n<ol>\n<li>Tokenize sequnces: convert the string into words (features): For example: \"hello my name is georgio\" to [\"hello\", \"my\", \"name\", \"is\", \"georgio\"].<\/li>\n<li>Next, you want to remove stop words (check Google for what stop words are).<\/li>\n<li>This stage is optional, it may lead to faulty results but I think it worth a try. Stem your words (features), that way you'll reduce the number of features which will lead to a faster run. Again, that's optional and might lead to some failures, for example: if you stem the word 'parking' you get 'park' which has a different meaning.<\/li>\n<li>Next thing is to create a dictionary (check Google for that). Each word gets a unique number and from this point we will use this number only.<\/li>\n<li>Computers understand numbers only so we need to talk in their language. We'll take the dictionary from stage 4 and replace each word in our corpus with its matching number.<\/li>\n<li>Now we need to split our data set to two groups: training and testing sets. One (training) will train our NN model and the second (testing) will help us to figure out how good is our NN. You can use Keras' cross validation function.<\/li>\n<li>Next thing is defining whats the max number of features our NN can get as an input. Keras call this parameter - 'maxlen'. But you don't really have to do this manually, Keras can do that automatically just by searching for the longest sentence you have in your corpus. <\/li>\n<li>Next, let's say that Keras found out that the longest sentence in your corpus has 20 words (features) and one of your sentences is the example in the first stage, which its length is 5 (if we'll remove stop words it'll be shorter), in such case we'll need to add zeros, 15 zeros actually. This is called pad sequence, we do that so every input sequence will be in the same length.<\/li>\n<\/ol>\n"},{"AnswerId":"33559473","CreationDate":"2015-11-06T04:24:28.557","ParentId":null,"OwnerUserId":"4164727","Title":null,"Body":"<p>This might help.\n<a href=\"http:\/\/keras.io\/models\/\" rel=\"nofollow noreferrer\">http:\/\/keras.io\/models\/<\/a><\/p>\n\n<p>Here is an sample usage.\n<a href=\"https:\/\/stackoverflow.com\/questions\/31556268\/how-to-use-keras-for-xor\">How to use keras for XOR<\/a><\/p>\n\n<p>Probably you have to convert ur corpus into ndarray first and throw it to your model.predict<\/p>\n\n<p>From what it seem so far the model.predict input of the training model should be 100 words corpus which represent an index of each word in dictionary. So if you want to train it with ur corpus, you have to convert ur corpus according to those dictionary and see if the result is 0 or 1 <\/p>\n"}]}
{"QuestionId":33540653,"AnswerCount":1,"Tags":"<boost><visual-studio-2013><linker-errors><caffe><icc>","CreationDate":"2015-11-05T09:20:17.397","AcceptedAnswerId":33542100.0,"OwnerUserId":null,"Title":"link error for library which doesn't exist in Linker->Input","Body":"<p>I'm compiling Berkley's Caffe to a static library, MT, MSVC 12 (VS2013) and linking it to a dll.<br>\nIt works fine.<br>\nNow I'm trying to compile it with Intel C++ 2015 compiler - the static lib build fine, but the dependent dll fails with the following linker error:<\/p>\n\n<blockquote>\n  <p>fatal error LNK1104: cannot open file\n  'libboost_thread-iw-mt-s-1_58.lib'<\/p>\n<\/blockquote>\n\n<p>There isn't such library in my Boost build indeed, but I don't know where this dependency came from. Except for the compiler I didn't change any other setting, and I can't find that Boost lib in either Caffe of my dll's Linker->Input->Additional dependencies.<br>\nHow can I fix this?<br>\nThanks!<\/p>\n","answers":[{"AnswerId":"33542100","CreationDate":"2015-11-05T10:24:18.050","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Found the solution here: <a href=\"https:\/\/software.intel.com\/en-us\/articles\/intel-c-compiler-for-windows-fatal-link-error-lnk1104-when-using-intel-c-compiler-with-boost-libraries\" rel=\"nofollow\">https:\/\/software.intel.com\/en-us\/articles\/intel-c-compiler-for-windows-fatal-link-error-lnk1104-when-using-intel-c-compiler-with-boost-libraries<\/a><\/p>\n\n<p>Copied:<\/p>\n\n<blockquote>\n  <p>When building an application that uses the Boost libraries with the\n  Intel\u00ae C++ Compiler, you may get linker errors like the ones shown\n  below due to incorrect libraries being linked to the application:<\/p>\n  \n  <p>fatal error LNK1104: cannot open file\n  'libboost_thread-iw-mt-1_33_1.lib' fatal error LNK1104: cannot open\n  file 'libboost_thread-iw-1_33_1.lib' ... The root cause is missing\n  Boost libraries for the Intel\u00ae C++ Compiler.<\/p>\n  \n  <p>The preferred solution is to recompile all required Boost libraries\n  with the Intel\u00ae C++ Compiler (libraries with the infix \"iw\" are\n  created because of this). However, this is not mandatory. The\n  libraries provided for the different Microsoft Visual Studio* versions\n  are safe to use as well. Perform the following steps to use them\n  instead:<\/p>\n  \n  <p>Open the Boost configuration file \"auto_link.hpp\". Search for 1<\/p>\n  \n  <h1>elif defined(__ICL) 2   3   4<\/h1>\n\n<pre><code>                       \/\/ Intel C++, no version number: 5   6\n                    #  define BOOST_LIB_TOOLSET \"iw\" Change \"iw\" depending on which Microsoft Visual Studio version you're using:\n<\/code><\/pre>\n  \n  <p>\"vc71\": Microsoft Visual Studio .NET 2003 \"vc80\": Microsoft Visual\n  Studio 2005 \"vc90\": Microsoft Visual Studio 2008 \"vc100\": Microsoft\n  Visual Studio 2010 \"vc110\": Microsoft Visual Studio 2012 \"vc120\":\n  Microsoft Visual Studio 2013 \"vc140\": Microsoft Visual Studio 2015<\/p>\n  \n  <p>Rebuild your application to resolve the linker errors.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33545325,"AnswerCount":1,"Tags":"<lua><neural-network><torch><backpropagation>","CreationDate":"2015-11-05T13:04:19.810","AcceptedAnswerId":33554212.0,"OwnerUserId":2782619.0,"Title":"torch backward through gModule","Body":"<p>I have a graph as follows, where the input x has two paths to reach y. They are combined with a gModule that uses cMulTable. Now if I do gModule:backward(x,y), I get a table of two values. Do they correspond to the error derivative derived from the two paths? \n<a href=\"https:\/\/i.stack.imgur.com\/31ZoU.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/31ZoU.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>But since path2 contains other nn layers, I suppose I need to derive the derivates in this path in a stepwise fashion. But why did I get a table of two values for dy\/dx?<\/p>\n\n<p>To make things clearer, code to test this is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>So what is wrong with my steps?<\/p>\n\n<p>P.S, perhaps I have found out the reason because g:backward({x,x_p}, error) results in the same table. So I guess the two values stand for dy\/dx and dy\/dx_p respectively.<\/p>\n","answers":[{"AnswerId":"33554212","CreationDate":"2015-11-05T20:31:34.660","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>I think you simply made a mistake constructing your <code>gModule<\/code>. <code>gradInput<\/code> of every <code>nn.Module<\/code> has to have exactly the same structure as its <code>input<\/code> - that is the way backprop works.<\/p>\n\n<p>Here's an example how to create a module like yours using <code>nngraph<\/code>:<\/p>\n\n<pre><code>require 'torch'\nrequire 'nn'\nrequire 'nngraph'\n\nfunction CreateModule(input_size)\n    local input = nn.Identity()()   -- network input\n\n    local nn_module_1 = nn.Linear(input_size, 100)(input)\n    local nn_module_2 = nn.Linear(100, input_size)(nn_module_1)\n\n    local output = nn.CMulTable()({input, nn_module_2})\n\n    -- pack a graph into a convenient module with standard API (:forward(), :backward())\n    return nn.gModule({input}, {output})\nend\n\n\ninput = torch.rand(30)\n\nmy_module = CreateModule(input:size(1))\n\noutput = my_module:forward(input)\ncriterion_err = torch.rand(output:size())\n\ngradInput = my_module:backward(input, criterion_err)\nprint(gradInput)\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>As I said, <code>gradInput<\/code> of every <code>nn.Module<\/code> has to have exactly the same structure as its <code>input<\/code>. So, if you define your module as <code>nn.gModule({input1, input2}, {score})<\/code>, your <code>gradOutput<\/code> (the result of the backward pass) will be a table of gradients w.r.t. <code>input1<\/code> and <code>input2<\/code> which in your case are <code>x<\/code> and <code>x_p<\/code>. <\/p>\n\n<p>The only question remains: why on Earth don't you get an error when call:<\/p>\n\n<pre><code>gradient1 = g:backward(x, error) \ngradient2 = g:backward(x_p, error)\n<\/code><\/pre>\n\n<p>An exception must be raised because the first argument must be not a tensor but a table of two tensors. Well, most (perhaps all) of torch modules during calculating <code>:backward(input, gradOutput)<\/code> don't use <code>input<\/code> argument (they usually store a copy of <code>input<\/code> from the last <code>:forward(input)<\/code> call). In fact, this argument is so useless that modules don't even bother themselves to verify it.<\/p>\n"}]}
{"QuestionId":33547401,"AnswerCount":1,"Tags":"<python><theano><lasagne>","CreationDate":"2015-11-05T14:40:26.777","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"lasagne - unexpected keyword error","Body":"<p>I am trying to learn machine learning in Python - and would like to run the lasagne \/ nolearn packages. I have installed all packages - and am using the script per below (from <a href=\"http:\/\/semantive.com\/deep-learning-examples\/\" rel=\"nofollow\">http:\/\/semantive.com\/deep-learning-examples\/<\/a>) which gives the following error. Please let me know if anyone has an idea on how to resolve this error.<\/p>\n\n<p>the script gives an initial error in only one of the lasagne modules: <\/p>\n\n<pre><\/pre>\n\n<p>subsequently - there is an error around the pad argument:<\/p>\n\n<pre><\/pre>\n\n<p>the code <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33561453","CreationDate":"2015-11-06T07:19:33.080","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You appear to be using an incompatible version of Lasagne for your nolearn version.<\/p>\n\n<p>The <code>pad<\/code> feature was added to the <code>Conv2DLayer<\/code> class in Lasagne in July and August 2015 (see <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/commit\/7a6c9e59c547efaf1a34484ae587f96b03797993\" rel=\"nofollow\">here<\/a> and <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/commit\/6d18347bdb52b1c7f71ae1802ed6a6a418eb8ea1\" rel=\"nofollow\">here<\/a>). Your version of nolearn clearly expects to be using that version or later.<\/p>\n\n<p>There are two possibilities:<\/p>\n\n<ol>\n<li><p>You (presumably accidentally) have two versions of Lasagne on your system but its an older one that's found first by Python. If this is true, remove the older version and\/or make sure the newer version is found (first) by Python.<\/p><\/li>\n<li><p>You simply have an outdated version of Lasagne. Solution: update it! How you do that may depend on how you installed it in the first place. Ultimately you'll need to get the latest version from Github though.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":33554977,"AnswerCount":2,"Tags":"<python><matplotlib><theano><lasagne><nolearn>","CreationDate":"2015-11-05T21:18:57.250","AcceptedAnswerId":33612691.0,"OwnerUserId":3011570.0,"Title":"Installing nolearn (python machine learning) results in error","Body":"<p>I am attempting to install nolearn, a python machine learning library (based upon theano and lasagne), and am receiving an unusual error (note: the command is the first step in installing nolearn as install docs):<\/p>\n\n<p>COMMAND:<\/p>\n\n<pre><\/pre>\n\n<p>OUTPUT: <\/p>\n\n<pre><\/pre>\n\n<p>I have tried a number of things, including installing freetype as well as reinstalling matplotlib (both based upon another stackexchange posting) but unfortunately it still does not work (error above is shown).<\/p>\n\n<p>Any assistance is appreciated!<\/p>\n\n<p>My configuration is as follows:<\/p>\n\n<ul>\n<li>running on MAC OSX (Yosemite)<\/li>\n<li>running in ANACONDA environment<\/li>\n<li>using conda as environment manager <\/li>\n<\/ul>\n","answers":[{"AnswerId":"33612691","CreationDate":"2015-11-09T15:38:10.953","ParentId":null,"OwnerUserId":"3011570","Title":null,"Body":"<p>I ended up having difficulty reconciling the <code>nolearn<\/code> library and its requirements for recent\/current versions of <code>lasagne<\/code> and <code>theano<\/code> (etc) and the versions available \/ compatible with <code>anaconda<\/code>.  I may have been able to resolve these conflicts given more time, but being time constrained, I just stopped using <code>anaconda<\/code> (and <code>conda<\/code>) and just used the regular <code>python<\/code> distribution and <code>virtualenv<\/code> (and <code>virtualenvwrapper<\/code>) and got it all working very quickly.<\/p>\n"},{"AnswerId":"33561285","CreationDate":"2015-11-06T07:06:50.583","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>When using Anaconda it is generally better to install\/update using <code>conda<\/code> instead of using <code>pip<\/code>.<\/p>\n\n<p>You could try updating scipy using conda instead of pip: <code>condu update scipy<\/code>.<\/p>\n\n<p>Or update everything installed via conda: <code>conda update --all<\/code>.<\/p>\n\n<p>You may need to update Anaconda itself first: <code>conda update conda<\/code>.<\/p>\n\n<p>If you can get scipy installed\/updated that way, then try the same for the other entries in requirements.txt (running the install\/update manually\/individually).<\/p>\n"}]}
{"QuestionId":33558987,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-11-06T03:31:44.747","AcceptedAnswerId":33562131.0,"OwnerUserId":290163.0,"Title":"How to use Caffe snapshot to recognize in practice?","Body":"<p>I have a trained <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> snapshot, and now I want to use it to do some real recognition, i.e. input one or some files and get their NN outputs.<\/p>\n\n<p>I want to do this in a python script, but shell batch is also fine.<\/p>\n\n<p>Is there some documents or answers for this? Thanks!<\/p>\n","answers":[{"AnswerId":"33562131","CreationDate":"2015-11-06T08:08:06.867","ParentId":null,"OwnerUserId":"3509999","Title":null,"Body":"<p>For python script - take a look at this great ipython notebook:\n<a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a><\/p>\n\n<p>For simple bash you could also use the cpp_classification tool, see more here:\n<a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/cpp_classification.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/cpp_classification.html<\/a><\/p>\n"}]}
{"QuestionId":33564181,"AnswerCount":2,"Tags":"<python><neural-network><theano><keyerror><keras>","CreationDate":"2015-11-06T10:05:13.850","AcceptedAnswerId":null,"OwnerUserId":4842101.0,"Title":"Keras GRU NN KeyError when fitting : \"not in index\"","Body":"<p>I'm currently facing an issue while trying to fit my GRU model with my training data.\nAfter a quick look on StackOverflow, I found this post to be quite similar to my issue :<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/33292212\/simplest-lstm-training-with-keras-io\">Simplest Lstm training with Keras io<\/a><\/p>\n\n<p>My own model is as follow :<\/p>\n\n<pre><\/pre>\n\n<p>And the error is :<\/p>\n\n<pre><\/pre>\n\n<p>Any idea to solve this ?\nThanks<\/p>\n\n<p>EDIT : Some facts about the data :<\/p>\n\n<pre><\/pre>\n\n<p>X and Y are CSV file containing time series values (e.g. for each row, there are 37 consecutive values of the time series in the X file + 2 time values (considered as past) and 30 in the Y file (considered as the forecast to predict))<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37095407","CreationDate":"2016-05-08T01:49:25.097","ParentId":null,"OwnerUserId":"6305521","Title":null,"Body":"<p>I couldn't use Pandas DataFrames as inputs &amp; outputs to Keras model.fit, at least not Pandas 0.13.1, which is the standard package from Ubuntu.<\/p>\n\n<p>Instead, use np.array(X_train) and np.array(Y_train). That worked for me.<\/p>\n"},{"AnswerId":"44305348","CreationDate":"2017-06-01T10:40:59.547","ParentId":null,"OwnerUserId":"6751309","Title":null,"Body":"<p>I've experienced a similar issue. In my case the problem was in that you use Embeddings layer with predefined dimensions on input, so the sequences you pass to this layer should be padded or truncated to the input_size using keras.preprocessing.sequence.<\/p>\n"}]}
{"QuestionId":33574293,"AnswerCount":1,"Tags":"<neural-network><caffe><torch><conv-neural-network>","CreationDate":"2015-11-06T19:31:09.270","AcceptedAnswerId":33577040.0,"OwnerUserId":4975126.0,"Title":"Using Neural Networks for Data Manipulation","Body":"<p>Perhaps related to <a href=\"https:\/\/stackoverflow.com\/questions\/33112941\/multiple-category-classification-in-caffe\">this question<\/a>, but my goal is to have a network perform a manipulation on an input image and output the resulting image data.<\/p>\n\n<p>In the event this question lacks clarity, I would be glad to delve into deeper detail of my problem in the comments. However, I'll try to be as non-case-specific as possible, to make this question of use to others.<\/p>\n\n<h2>The Problem<\/h2>\n\n<p>I have a plethora of training data consisting of images before and after the proposed manipulation. My question relates to how I can train 1-to-1 each pixel using Caffe. My loss should take the form of something that computes the difference between the two images.<\/p>\n\n<p>If I have my last fully-connected\/inner-product layer outputting  and I have my label of the expected output image (same dimensions) what type of loss+accuracy structure should I use?<\/p>\n\n<h2>My Case<\/h2>\n\n<p>I've tried simply passing the inner-product data to a sigmoid cross-entropy loss with my label data, but it doesn't seem to be a supported method. <\/p>\n\n<p>My labels are non integer values, as they are pixel RGB data between 0 and 1 (<strong>Note:<\/strong> I could use integers in the form of 0 to 255) and Caffe seems to interpret labels as categories as opposed to simple values.<\/p>\n\n<p>I could have 255 categories per pixel channel, but that would result in 255 * 3 channels * 256 height * 256 width = 50,135,040 categories, which is absurdly over complicating what I'm trying to achieve.<\/p>\n\n<h2>My Questions<\/h2>\n\n<ul>\n<li>Does Caffe natively support what I'm trying to achieve?\n\n<ul>\n<li>If so, how should I change my structure to conform to these specification?<\/li>\n<li>If not, do any other neural network frameworks such as Torch support this?<\/li>\n<\/ul><\/li>\n<li>Is there a name for the type of problem I'm trying to solve with my network (certainly not categorical classification)?\n\n<ul>\n<li>What has been used in the past to solve this style of problem?<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<h2>Sources With Potential Value<\/h2>\n\n<ul>\n<li><a href=\"http:\/\/research.microsoft.com\/en-us\/um\/people\/kahe\/publications\/eccv14srcnn.pdf\" rel=\"nofollow noreferrer\">Training Super-Resolution Image Upscaling<\/a><\/li>\n<li><a href=\"https:\/\/youtu.be\/P78QYjWh5sM\" rel=\"nofollow noreferrer\">Karol Gregor on Variational Autoencoders and Image Generation<\/a> (less relevant as this touches on model reconstruction)<\/li>\n<\/ul>\n","answers":[{"AnswerId":"33577040","CreationDate":"2015-11-06T23:05:20.110","ParentId":null,"OwnerUserId":"5170044","Title":null,"Body":"<p>The loss layer you are looking for is <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1EuclideanLossLayer.html\" rel=\"nofollow\">Euclidean loss layer<\/a> (Mean Squared Error):<\/p>\n\n<pre><code>layers {\n  name: \"loss\"\n  type: EUCLIDEAN_LOSS\n  top: \"loss\"\n  bottom: \"CONVX_15\"\n  bottom: \"labels\"\n}\n<\/code><\/pre>\n\n<p>Your problem is multivariate regression and you have to use loss which is suitable for it. <strong>Sigmoid cross-entropy loss<\/strong> is for classification, where the target values (labels) have to be between 0 and 1 (e.g. probability that a pixel is on\/off). <\/p>\n\n<p>With euclidean loss, be carefull to manage your gradients. Keep your target values in range &lt;0,1> and use Xavier weight initialization. Still, you will probably need to set lower learning rate compared to classification problems to keep the SGD from exploding. <\/p>\n"}]}
{"QuestionId":33583743,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2015-11-07T14:17:35.217","AcceptedAnswerId":null,"OwnerUserId":2721716.0,"Title":"storing a dynamic dimension array list during the \"scan\" function in theano","Body":"<p>I want to make a dynamic array list during the theano  function like below.<\/p>\n\n<pre><\/pre>\n\n<p>What I want to do is in each iteration, append one vector of  to the end of the , then return the new matrix. However it seems that scan can not store the dynamic dimension matrix list in the runtime? <\/p>\n\n<p>I have found that the error happened on the line <\/p>\n\n<p>The theano report the error like <\/p>\n\n<blockquote>\n  <p>ValueError: could not broadcast input array from shape (2,3) into shape (1,3)\", other details about this error is omitted. <\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":33592778,"AnswerCount":2,"Tags":"<python><gpgpu><theano><deep-learning><lasagne>","CreationDate":"2015-11-08T10:03:56.150","AcceptedAnswerId":33593166.0,"OwnerUserId":5538861.0,"Title":"Lasagne dropoutlayer does not utilize GPU efficiently","Body":"<p>I am using theano and lasagne for a DNN speech enhancement project. I use a feed-forward network very similar to the mnist example in the lasagne documentation (\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py). This network uses several dropout layers. I train my network on an Nvidia Titan X GPU. However, when I do not use dropout my GPU utilization is approximately 60% and one epoch takes around 60s but when I use dropout my GPU utilization drops to 8% and each epoch takes approximately 600s. This is regardless of the dropout rate is set to 20% or 0.1%. <\/p>\n\n<p>Initially I thought it was due to the random number generator (RNG) used to generate the dropout mask, that did not run on the GPU. However, in the code (<a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/layers\/noise.py\" rel=\"nofollow\">https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/layers\/noise.py<\/a>) it seems like rng_mrg is used, which should be running on the GPU based on this link: <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#other-implementations\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#other-implementations<\/a> <\/p>\n\n<p>Running the theano profiler shows that \"theano.sandbox.rng_mrg.mrg_uniform\" takes up 86.7% of the execution time, which I do not understand. <\/p>\n\n<p>If anyone has an idea of what kills my GPU utilization I would appreciate it. <\/p>\n","answers":[{"AnswerId":"33593166","CreationDate":"2015-11-08T10:56:05.750","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>If you look at the <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/rng_mrg.py#L326\" rel=\"nofollow\">code<\/a> for <code>mrg_uniform<\/code>, you can see that it is a pure python CPU implementation of the random generator. You can also see that there is a <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/rng_mrg.py#L532\" rel=\"nofollow\">GPU version<\/a> of that same generator, but the code you are running apparently does not use it.<\/p>\n\n<p>So the answer isn't that your GPU utilisation is going down so much as your CPU utilisation is greatly increasing because you are using a pure Python random generator. The solution would obviously be to work out how to switch to a GPU accelerated random generator.<\/p>\n"},{"AnswerId":"33594665","CreationDate":"2015-11-08T13:48:50.420","ParentId":null,"OwnerUserId":"5538861","Title":null,"Body":"<p>As pointed out by talonmies the problem was that lasagne was using the CPU version of the RNG (mrg_uniform) and not the GPU version (GPU_mrg_uniform). \nI have not yet found an elegant solution but the following two hacks solves the problem. <\/p>\n\n<p>Either change line 93 <code>cuda_enabled = False<\/code> to <code>cuda_enabled = True<\/code> in <\/p>\n\n<pre><code>https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/__init__.py\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<p>Change line 57<br>\n<code>self._srng = RandomStreams(get_rng().randint(1, 2147462579))<\/code><br>\nto <br> \n<code>self._srng = \"RandomStreams(get_rng().randint(1, 2147462579),use_cuda = True)<\/code> in <code>https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/layers\/noise.py<\/code><\/p>\n\n<p>I also believe you should be able to do the same by simply typing <code>theano.sandbox.cuda.use(enable_cuda=True)<\/code> directly in the main script. However, this did not for some reasons not work for me.  <\/p>\n"}]}
{"QuestionId":33610685,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2015-11-09T13:52:52.940","AcceptedAnswerId":33610914.0,"OwnerUserId":16480.0,"Title":"In TensorFlow, what is the difference between Session.run() and Tensor.eval()?","Body":"<p>TensorFlow has two ways to evaluate part of graph:  on a list of variables and .  Is there a difference between these two?<\/p>\n","answers":[{"AnswerId":"46795905","CreationDate":"2017-10-17T17:31:19.243","ParentId":null,"OwnerUserId":"8752622","Title":null,"Body":"<p><code>eval()<\/code> can not handle the list object<\/p>\n\n<pre><code>tf.reset_default_graph()\n\na = tf.Variable(0.2, name=\"a\")\nb = tf.Variable(0.3, name=\"b\")\nz = tf.constant(0.0, name=\"z0\")\nfor i in range(100):\n    z = a * tf.cos(z + i) + z * tf.sin(b - i)\ngrad = tf.gradients(z, [a, b])\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    init.run()\n    print(\"z:\", z.eval())\n    print(\"grad\", grad.eval())\n<\/code><\/pre>\n\n<p>but <code>Session.run()<\/code> can<\/p>\n\n<pre><code>print(\"grad\", sess.run(grad))\n<\/code><\/pre>\n\n<p>correct me if I am wrong<\/p>\n"},{"AnswerId":"54296807","CreationDate":"2019-01-21T19:43:16.073","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>The most important thing to remember:<\/p>\n\n<blockquote>\n  <p><strong>The only way to get a constant, variable (any result) from TenorFlow is the session.<\/strong><\/p>\n<\/blockquote>\n\n<p>Knowing this everything else is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"nofollow noreferrer\">easy<\/a>:<\/p>\n\n<blockquote>\n  <p>Both <code>tf.Session.run()<\/code> and <code>tf.Tensor.eval()<\/code> get results from the session where <code>tf.Tensor.eval()<\/code> is a shortcut for calling <code>tf.get_default_session().run(t)<\/code><\/p>\n<\/blockquote>\n\n<hr>\n\n<p>I would also outline the method <code>tf.Operation.run()<\/code> as in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>After the graph has been launched in a session, an Operation can be executed by passing it to <code>tf.Session.run()<\/code>. <code>op.run()<\/code> is a shortcut for calling <code>tf.get_default_session().run(op)<\/code>.<\/p>\n<\/blockquote>\n"},{"AnswerId":"33610914","CreationDate":"2015-11-09T14:05:45.853","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you have a <code>Tensor<\/code> t, calling <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#eval\" rel=\"noreferrer\"><code>t.eval()<\/code><\/a> is equivalent to calling <code>tf.get_default_session().run(t)<\/code>.<\/p>\n\n<p>You can make a session the default as follows:<\/p>\n\n<pre><code>t = tf.constant(42.0)\nsess = tf.Session()\nwith sess.as_default():   # or `with sess:` to close on exit\n    assert sess is tf.get_default_session()\n    assert t.eval() == sess.run(t)\n<\/code><\/pre>\n\n<p>The most important difference is that you can use <code>sess.run()<\/code> to fetch the values of many tensors in the same step:<\/p>\n\n<pre><code>t = tf.constant(42.0)\nu = tf.constant(37.0)\ntu = tf.mul(t, u)\nut = tf.mul(u, t)\nwith sess.as_default():\n   tu.eval()  # runs one step\n   ut.eval()  # runs one step\n   sess.run([tu, ut])  # evaluates both tensors in a single step\n<\/code><\/pre>\n\n<p>Note that each call to <code>eval<\/code> and <code>run<\/code> will execute the whole graph from scratch. To cache the result of a computation, assign it to a <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a>.<\/p>\n"},{"AnswerId":"52096129","CreationDate":"2018-08-30T11:39:07.640","ParentId":null,"OwnerUserId":"2761007","Title":null,"Body":"<p>In tensorflow you create graphs and pass values to that graph. Graph does all the hardwork and generate the output based on the configuration that you have made in the graph.\nNow When you pass values to the graph then first you need to create a tensorflow session.<\/p>\n\n<pre><code>tf.Session()\n<\/code><\/pre>\n\n<p>Once session is initialized then you are supposed to use that session because all the variables and settings are now part of the session. So, there are two ways to pass external values to the graph so that graph accepts them. One is to call the .run() while you are using the session being executed.<\/p>\n\n<p>Other way which is basically a shortcut to this is to use .eval(). I said shortcut because the full form of .eval() is <\/p>\n\n<pre><code>tf.get_default_session().run(values)\n<\/code><\/pre>\n\n<p>You can check that yourself.\nAt the place of <code>values.eval()<\/code> run <code>tf.get_default_session().run(values)<\/code>. You must get the same behavior.<\/p>\n\n<p>what eval is doing is using the default session and then executing run().<\/p>\n"},{"AnswerId":"33666260","CreationDate":"2015-11-12T07:28:42.570","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The FAQ session on tensor flow has an <a href=\"https:\/\/www.tensorflow.org\/resources\/faq#running_a_tensorflow_computation\" rel=\"nofollow noreferrer\">answer to exactly the same question<\/a>. I will just go ahead and leave it here:<\/p>\n\n<hr>\n\n<p>If <code>t<\/code> is a <code>Tensor<\/code> object, <code>t.eval()<\/code> is shorthand for <code>sess.run(t)<\/code> (where <code>sess<\/code> is the current default session. The two following snippets of code are equivalent:<\/p>\n\n<pre><code>sess = tf.Session()\nc = tf.constant(5.0)\nprint sess.run(c)\n\nc = tf.constant(5.0)\nwith tf.Session():\n  print c.eval()\n<\/code><\/pre>\n\n<p>In the second example, the session acts as a context manager, which has the effect of installing it as the default session for the lifetime of the <code>with<\/code> block. The context manager approach can lead to more concise code for simple use cases (like unit tests); if your code deals with multiple graphs and sessions, it may be more straightforward to explicit calls to <code>Session.run()<\/code>.<\/p>\n\n<p>I'd recommend that you at least skim throughout the whole FAQ, as it might clarify a lot of things.<\/p>\n"}]}
{"QuestionId":33613945,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-09T16:43:21.977","AcceptedAnswerId":33615369.0,"OwnerUserId":88622.0,"Title":"How to deal with inputs outside 0-1 range in tensorflow?","Body":"<p>In the example provided at <a href=\"http:\/\/www.tensorflow.org\/get_started\" rel=\"nofollow\">http:\/\/www.tensorflow.org\/get_started<\/a> if I multiply the input by 2<\/p>\n\n<pre><\/pre>\n\n<p>I get non-sense output, while I expected to get the same solution.<\/p>\n\n<pre><\/pre>\n\n<p>How does tensorflow handle inputs that are not in the 0-1 range?<\/p>\n\n<p><em>EDIT<\/em>: Using  works without an issue.<\/p>\n","answers":[{"AnswerId":"33615369","CreationDate":"2015-11-09T18:05:43.650","ParentId":null,"OwnerUserId":"1625244","Title":null,"Body":"<p>The issue is that the example uses a very aggressive learning rate:<\/p>\n\n<pre><code>optimizer = tf.train.GradientDescentOptimizer(0.5)\n<\/code><\/pre>\n\n<p>This makes learning faster, but stops working if you change the problem a bit. A learning rate of <code>0.01<\/code> would be more typical:<\/p>\n\n<pre><code>optimizer = tf.train.GradientDescentOptimizer(0.01)\n<\/code><\/pre>\n\n<p>Now your modification works fine. :)<\/p>\n"}]}
{"QuestionId":33614221,"AnswerCount":0,"Tags":"<machine-learning><computer-vision><torch>","CreationDate":"2015-11-09T16:59:24.320","AcceptedAnswerId":null,"OwnerUserId":5543312.0,"Title":"itorch do not display any image","Body":"<p>I am doing iTorch but stucked<\/p>\n\n<p>I followed this installation step in <a href=\"https:\/\/github.com\/facebook\/iTorch(for\" rel=\"nofollow\">https:\/\/github.com\/facebook\/iTorch(for<\/a> itorch)<\/p>\n\n<p>I did just like github\/facebook\/iTorch said<\/p>\n\n<p>in itorch console mode, entered this code<\/p>\n\n<pre><\/pre>\n\n<p>But nothing happen\njust shows blank<\/p>\n\n<pre><\/pre>\n\n<p>And during plotting, I entered plot:save('out.html') then it makes out.html and that file shows correct image.<\/p>\n\n<p>version list is :\nUbuntu 12.04.5 LTS\nipython version 4.0.0\nPython 2.7.3\ntorch works well<\/p>\n\n<p>plz tell me how to solve it!!<\/p>\n","answers":[]}
{"QuestionId":33616094,"AnswerCount":6,"Tags":"<python><windows><tensorflow>","CreationDate":"2015-11-09T18:51:41.767","AcceptedAnswerId":33618580.0,"OwnerUserId":4389216.0,"Title":"Is Tensorflow compatible with a Windows workflow?","Body":"<p>I haven't seen anything about Windows compatibility -- is this on the way or currently available somewhere if I put forth some effort? (I have a Mac and an Ubuntu box but the Windows machine is the one with the discrete graphics card that I currently use with theano).<\/p>\n","answers":[{"AnswerId":"33635663","CreationDate":"2015-11-10T17:10:32.520","ParentId":null,"OwnerUserId":"278191","Title":null,"Body":"<p>As @mrry suggested, it is easier to set up TensorFlow with Docker. Here's how I managed to set it up as well as getting iPython Notebook up and running in my Docker environment (I find it really convenient to use iPython Notebook for all testing purposes as well as documenting my experiments).<\/p>\n\n<p>I assume that you have installed both docker and boot2docker for Windows here.<\/p>\n\n<p>First, run TensorFlow docker on daemon and set it up so Jupyter server (iPython Notebook) can be accessed from your main Windows system's browser:<\/p>\n\n<pre><code>docker run -dit -v \/c\/Users\/User\/:\/media\/disk -p 8888:8888 b.gcr.io\/tensorflow\/tensorflow:latest\n<\/code><\/pre>\n\n<p>Replace <code>\/c\/Users\/User\/<\/code> with a path in your host you wish to mount i.e. where you can keep your iPython files. <em>I don't know how to set it to other drives than C:, let me know if you do<\/em>. <code>\/media\/disk<\/code> is the location in your TensorFlow docker where your host path's mounted against.<\/p>\n\n<p><code>-p 8888:8888<\/code> basically means \"map port 8888 in docker to 8888 in host directory\". You can change the second part to other ports if you wish.<\/p>\n\n<p>When you got it running, you can access it by running the following code:<\/p>\n\n<pre><code>docker exec -ti [docker-id] bash\n<\/code><\/pre>\n\n<p>Where [docker-id] can be found by running:<\/p>\n\n<pre><code>docker ps\n<\/code><\/pre>\n\n<p>To start your ipython notebook server from within TensorFlow's docker, run the following command:<\/p>\n\n<pre><code>ipython notebook --ip='*'\n<\/code><\/pre>\n\n<p>To allow ipython server to listen to all ip so your app may be accessible from host machine.<\/p>\n\n<p>Instead of viewing your app in <code>http:\/\/localhost:8888<\/code>, you can only view it in <code>http:\/\/[boot2docker-ip]:8888<\/code>. To find <code>boot2docker-ip<\/code> run this in your terminal (not boot2docker terminal):<\/p>\n\n<pre><code>boot2docker ip\n<\/code><\/pre>\n"},{"AnswerId":"40875065","CreationDate":"2016-11-29T20:19:08.567","ParentId":null,"OwnerUserId":"4896087","Title":null,"Body":"<p>TensorFlow is now officially available on Windows!<\/p>\n\n<blockquote>\n  <p>TensorFlow now builds and runs on Microsoft Windows (tested on Windows\n  10, Windows 7, and Windows Server 2016). Supported languages include\n  Python (via a pip package) and C++. CUDA 8.0 and cuDNN 5.1 are\n  supported for GPU acceleration. Known limitations include: It is not\n  currently possible to load a custom op library. The GCS and HDFS file\n  systems are not currently supported. The following ops are not\n  currently implemented: DepthwiseConv2dNative,\n  DepthwiseConv2dNativeBackpropFilter,\n  DepthwiseConv2dNativeBackpropInput, Dequantize, Digamma, Erf, Erfc,\n  Igamma, Igammac, Lgamma, Polygamma, QuantizeAndDequantize,\n  QuantizedAvgPool, QuantizedBatchNomWithGlobalNormalization,\n  QuantizedBiasAdd, QuantizedConcat, QuantizedConv2D, QuantizedMatmul,\n  QuantizedMaxPool, QuantizeDownAndShrinkRange, QuantizedRelu,\n  QuantizedRelu6, QuantizedReshape, QuantizeV2, RequantizationRange, and\n  Requantize.<\/p>\n<\/blockquote>\n\n<ul>\n<li><p><a href=\"https:\/\/mspoweruser.com\/tensorflow-now-builds-runs-microsoft-windows\/\" rel=\"nofollow noreferrer\">TensorFlow now builds and runs on Microsoft Windows<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/tag\/0.12.0-rc0\" rel=\"nofollow noreferrer\">Release Note<\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"33618580","CreationDate":"2015-11-09T21:24:48.520","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>Updated 11\/28\/2016:<\/strong> Today we released the first release candidate of TensorFlow 0.12, which includes support for Windows. You can install the Python bindings using the following command in a Python shell:<\/p>\n\n<pre><code>C:\\&gt; pip install tensorflow\n<\/code><\/pre>\n\n<p>...or, if you want GPU support:<\/p>\n\n<pre><code>C:\\&gt; pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>You can also build TensorFlow yourself using Microsoft Visual C++ and NVCC (for the CUDA parts). The easiest way to build on Windows is currently to use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/cmake\">CMake build<\/a>, and we will soon provide support for <a href=\"https:\/\/www.bazel.io\/versions\/master\/docs\/windows.html\">Bazel on Windows<\/a>.<\/p>\n\n<hr>\n\n<p><strong>Previous answer:<\/strong> We haven't tried to build TensorFlow on Windows so far: the only supported platforms are Linux (Ubuntu) and Mac OS X, and we've only built binaries for those platforms.<\/p>\n\n<p>For now, on Windows, the easiest way to get started with TensorFlow would be to use Docker: <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation\">http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation<\/a><\/p>\n\n<p>It should become easier to add Windows support when Bazel (the build system we are using) adds support for building on Windows, which is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17#issuecomment-189599501\">on the roadmap for Bazel 0.3<\/a>.  You can see <a href=\"http:\/\/bazel.io\/roadmap.html\">the full Bazel roadmap here<\/a>.<\/p>\n\n<p>In the meantime, you can follow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\">issue 17 on the TensorFlow GitHub page<\/a>.<\/p>\n"},{"AnswerId":"33623888","CreationDate":"2015-11-10T06:02:17.573","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>Another way to run it on Windows is to install for example Vmware (a free version if you are not using it commercially), install Ubuntu Linux into that and then install TensorFlow using the Linux instructions. That is what I have been doing, it works well.<\/p>\n"},{"AnswerId":"37105739","CreationDate":"2016-05-08T23:06:42.253","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>TensorFlow is not supporting Windows at this point, but Windows has changed. Windows 10 Build 14432 includes bash.<\/p>\n\n<p>You can download the build from <a href=\"https:\/\/www.microsoft.com\/en-us\/software-download\/windowsinsiderpreviewadvanced\" rel=\"nofollow noreferrer\">https:\/\/www.microsoft.com\/en-us\/software-download\/windowsinsiderpreviewadvanced<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sLXLp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sLXLp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After installation, just install\/enable bash, and type <code>bash<\/code> in cmd. That's it.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LmTEH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LmTEH.png\" alt=\"enter image description here\"><\/a>\n(The image is from <a href=\"https:\/\/blogs.windows.com\/windowsexperience\/2016\/04\/06\/announcing-windows-10-insider-preview-build-14316\/\" rel=\"nofollow noreferrer\">https:\/\/blogs.windows.com\/windowsexperience\/2016\/04\/06\/announcing-windows-10-insider-preview-build-14316\/<\/a>)<\/p>\n\n<p>Then, run this (Python is already installed):<\/p>\n\n<pre><code>sudo apt-get install python-pip python-dev\nsudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.8.0-cp27-none-linux_x86_64.whl \n<\/code><\/pre>\n\n<p>Happy tensorflowing in Windows!<\/p>\n"},{"AnswerId":"40156608","CreationDate":"2016-10-20T13:59:09.360","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Initial support for building TensorFlow on Microsoft Windows was added on  2016-10-05 in commit <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/d0d975f8c3330b5402263b2356b038bc8af919a2\" rel=\"nofollow\">d0d975f8c3330b5402263b2356b038bc8af919a2<\/a>:<\/p>\n\n<blockquote>\n  <p>This PR contains an initial version of support for building TensorFlow\n  (CPU only) on Windows using CMake. It includes documentation for\n  building with CMake on Windows, platform-specific code for\n  implementing core functions on Windows, and CMake rules for building\n  the C++ example trainer program and a PIP package (Python 3.5 only).\n  The CMake rules support building TensorFlow with Visual Studio 2015.<\/p>\n  \n  <p>Windows support is a work in progress, and we welcome your feedback\n  and contributions.<\/p>\n  \n  <p>For full details of the features currently supported and instructions\n  for how to build TensorFlow on Windows, please see the file\n  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/cmake\/README.md\" rel=\"nofollow\"><code>tensorflow\/contrib\/cmake\/README.md<\/code><\/a>.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/twitter.com\/mrry\/status\/789167883158827008\" rel=\"nofollow\">Since 2016-10-20, GPU is supported<\/a>.<\/p>\n"}]}
{"QuestionId":33616593,"AnswerCount":3,"Tags":"<python><parallel-processing><deep-learning><tensorflow>","CreationDate":"2015-11-09T19:23:22.177","AcceptedAnswerId":33641967.0,"OwnerUserId":1133286.0,"Title":"How do I use distributed DNN training in TensorFlow?","Body":"<p>Google released TensorFlow today.<\/p>\n\n<p>I have been poking around in the code, and I don't see anything in the code or API about training across a cluster of GPU servers.<\/p>\n\n<p>Does it have distributed training functionality yet?<\/p>\n","answers":[{"AnswerId":"33622414","CreationDate":"2015-11-10T03:33:47.053","ParentId":null,"OwnerUserId":"4394807","Title":null,"Body":"<h3>Update<\/h3>\n\n<p>As you may have noticed. Tensorflow has already supported distributed DNN training for quite some time. Please refer to its offcial website for details.<\/p>\n\n<p>=========================================================================<\/p>\n\n<h3>Previous<\/h3>\n\n<p>No, it doesn't support distribute training yet, which is a little disappointing. But I don't think it is difficult to extend from single machine to multi-machine. Compared to other open source libraries, like Caffe, TF's data graph structure is more suitable for cross-machine tasks. <\/p>\n"},{"AnswerId":"33641967","CreationDate":"2015-11-11T00:00:48.197","ParentId":null,"OwnerUserId":"1695962","Title":null,"Body":"<p><strong>Updated:<\/strong> <\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/deploy\/distributed\" rel=\"nofollow noreferrer\">Distributed TensorFlow Documentation<\/a> <\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/distributed_runtime#distributed-tensorflow\" rel=\"nofollow noreferrer\">Distributed TensorFlow Source<\/a><\/p><\/li>\n<\/ul>\n\n<p>The release occurred on 2\/26\/2016 and was announced by coauthor Derek Murray in the original issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23#issuecomment-189285409\" rel=\"nofollow noreferrer\">here<\/a> and uses <a href=\"http:\/\/www.grpc.io\/\" rel=\"nofollow noreferrer\">gRPC<\/a> for inter-process communication.<\/p>\n\n<p><strong>Previous:<\/strong> <\/p>\n\n<p>Before the update above, a distributed implementation of <strong>TensorFlow<\/strong> had not been released yet. Support for a distributed implementation was the topic of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23\" rel=\"nofollow noreferrer\">this issue<\/a> where coauthor Vijay Vasudevan <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23#issuecomment-155151640\" rel=\"nofollow noreferrer\">wrote<\/a>:<\/p>\n\n<blockquote>\n  <p>we are working on making a distributed implementation available, it's\n  currently not in the initial release<\/p>\n<\/blockquote>\n\n<p>and Jeff Dean later provided <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23#issuecomment-155608002\" rel=\"nofollow noreferrer\">an update<\/a>:<\/p>\n\n<blockquote>\n  <p>Our current internal distributed extensions are somewhat entangled\n  with Google internal infrastructure, which is why we released the\n  single-machine version first. The code is not yet in GitHub, because\n  it has dependencies on other parts of the Google code base at the\n  moment, most of which have been trimmed, but there are some remaining\n  ones.<\/p>\n  \n  <p>We realize that distributed support is really important, and it's one\n  of the top features we're prioritizing at the moment.<\/p>\n<\/blockquote>\n"},{"AnswerId":"35653639","CreationDate":"2016-02-26T13:58:53.333","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It took us a few months, but today marks the release of the initial <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/distributed_runtime\">distributed TensorFlow runtime<\/a>. This includes support for multiple machines, each with multiple GPUs, with communication provided by <a href=\"http:\/\/grpc.io\">gRPC<\/a>.<\/p>\n\n<p>The current version includes the necessary backend components so that you can assemble a cluster manually and connect to it from a client program. More details are available in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/distributed_runtime\/README.md\">readme<\/a>.<\/p>\n"}]}
{"QuestionId":33616732,"AnswerCount":2,"Tags":"<python><macos><tensorflow>","CreationDate":"2015-11-09T19:32:04.603","AcceptedAnswerId":33616733.0,"OwnerUserId":3767229.0,"Title":"Where is the folder for Installing tensorflow with pip, Mac OSX?","Body":"<p>just installed tensorflow using pip with the command:<\/p>\n\n<p><\/p>\n\n<p>On the <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md\">\"Getting Started\" for Tensorflow<\/a> they have an example for convolutional neural networks<\/p>\n\n<p><\/p>\n\n<p>Where is that directory located when installing with pip?<\/p>\n","answers":[{"AnswerId":"33616733","CreationDate":"2015-11-09T19:32:04.603","ParentId":null,"OwnerUserId":"3767229","Title":null,"Body":"<p>Installing with pip, installs the packages to the directory \"site-packages\".<\/p>\n\n<p>The following code shows the location of tensorflow as well as where pip installs the packages:<\/p>\n\n<pre><code>$ pip show tensorflow\n<\/code><\/pre>\n\n<p>Which return:<\/p>\n\n<pre><code>Metadata-Version: 2.0\nName: tensorflow\nVersion: 0.5.0\nSummary: TensorFlow helps the tensors flow\nHome-page: http:\/\/tensorflow.com\/\nAuthor: Google Inc.\nAuthor-email: opensource@google.com\nLicense: Apache 2.0\nLocation: \/usr\/local\/lib\/python2.7\/site-packages\nRequires: six, numpy\n<\/code><\/pre>\n\n<p>here <code>Location:<\/code> shows where the package is installed with <\/p>\n\n<pre><code>$ cd \/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>As some people pointed out in the newer versions of tensorflow and depending on the <code>$ echo $TENSORFLOW<\/code> you need to look in either <\/p>\n\n<pre><code>$ cd \/usr\/local\/lib\/python{2,3}.X\/{site,dist}-packages\/tensorflow\n<\/code><\/pre>\n\n<p>Or<\/p>\n\n<pre><code>$ cd \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/include\/tensorflow\/core\/framework\n<\/code><\/pre>\n"},{"AnswerId":"40713607","CreationDate":"2016-11-21T05:53:23.583","ParentId":null,"OwnerUserId":"1954267","Title":null,"Body":"<p>It depends on where is $TENSORFLOW environment variable is set. Lets help it setup.<\/p>\n\n<p>First check,\n<code>$ echo $TENSORFLOW<\/code>\nif it return blank, you need to setup to access from any directory from your console.<\/p>\n\n<p>Case-1)\nIn case you have python from anaconda library\/environment (let say you have \nanaconda2)<\/p>\n\n<p>Usually installed location: \n~\/anaconda2\/lib\/python2.7\/site-package\/tensorflow<\/p>\n\n<p>case-2.)\nIn case of Python2.x or Python3.x, x = is subversion like 2.7 or 3.5\nUsually installed location: \n<code>\/usr\/local\/lib\/python2.x\/site-packages\/tensorflow<\/code><\/p>\n\n<p>Now you have identified python version.\nUse it as onetime path in bash or profile.<\/p>\n\n<p><code>$ vi ~\/.bashrc<\/code><\/p>\n\n<p>add this line in the bottom of the basrc file.<\/p>\n\n<p><code>$ export $TENSORFLOW=\"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow:$PATH\"<\/code><\/p>\n\n<p>check again in a new terminal<\/p>\n\n<p><code>$ echo $TENSORFLOW<\/code><\/p>\n\n<p>now you can use <\/p>\n\n<p><code>$ python tensorflow\/models\/image\/mnist\/convolutional.py<\/code><\/p>\n\n<p>directly by<\/p>\n\n<p><code>$ python -m tensorflow.models.image.mnist.convolutional.py<\/code> <\/p>\n\n<p>Now your program as per description in TensorFlow tutorial will search these path with period (.) instead of \"\/\" with -m argument.<\/p>\n"}]}
{"QuestionId":33617638,"AnswerCount":1,"Tags":"<macos><tensorflow>","CreationDate":"2015-11-09T20:27:08.087","AcceptedAnswerId":33618045.0,"OwnerUserId":904032.0,"Title":"TensorFlow (Mac OS X): can't determine number of CPU cores:","Body":"<p>There must be a simple setting for Mac OS X, to get rid of the following warning...something in .bash_profile?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33618045","CreationDate":"2015-11-09T20:52:32.577","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To provide explicit values for the relevant configuration options, you can do:<\/p>\n\n<pre><code>NUM_CORES = ...  # Choose how many cores to use.\nsess = tf.Session(\n    config=tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES,\n                   intra_op_parallelism_threads=NUM_CORES))\n<\/code><\/pre>\n\n<p>This issue is present in the initial binary release of TensorFlow for Mac OS X, but should be fixed in this commit: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/430a054d6134f00e5188906bc4080fb7c5035ad5\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/commit\/430a054d6134f00e5188906bc4080fb7c5035ad5<\/a><\/p>\n\n<p>The fix will be included in the next binary release. In the meantime, you can try building from source, by following the instructions here: <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"noreferrer\">http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources<\/a><\/p>\n"}]}
{"QuestionId":33619190,"AnswerCount":1,"Tags":"<python><installation><pip><tensorflow>","CreationDate":"2015-11-09T22:08:02.530","AcceptedAnswerId":33636869.0,"OwnerUserId":5244410.0,"Title":"Why I don't have permissions to remove six while installing a pip package?","Body":"<p>I've run this command to pip install TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm having trouble removing six (to reinstall it anyway?)<\/p>\n\n<p>Does anyone have any insight into this problem.  doesn't work either<\/p>\n\n<p>my terminal is spitting out:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33636869","CreationDate":"2015-11-10T18:20:29.617","ParentId":null,"OwnerUserId":"610569","Title":null,"Body":"<p>Ensure that your <code>pip<\/code> is <code>python2.x<\/code> for the <strong>location<\/strong> field:<\/p>\n\n<pre><code>$ pip show pip\n---\nMetadata-Version: 2.0\nName: pip\nVersion: 7.1.2\nSummary: The PyPA recommended tool for installing Python packages.\nHome-page: https:\/\/pip.pypa.io\/\nAuthor: The pip developers\nAuthor-email: python-virtualenv@groups.google.com\nLicense: MIT\nLocation: \/usr\/local\/lib\/python2.7\/dist-packages\nRequires: \n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>sudo pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":33620794,"AnswerCount":11,"Tags":"<c++><tensorflow>","CreationDate":"2015-11-10T00:18:33.113","AcceptedAnswerId":33622489.0,"OwnerUserId":3499258.0,"Title":"How to build and use Google TensorFlow C++ api","Body":"<p>I'm really eager to start using Google's new Tensorflow library in C++. The website and docs are just really unclear in terms of how to build the project's C++ API and I don't know where to start. <\/p>\n\n<p>Can someone with more experience help by discovering and sharing a guide to using tensorflow's C++ API?  <\/p>\n","answers":[{"AnswerId":"33833476","CreationDate":"2015-11-20T18:20:09.600","ParentId":null,"OwnerUserId":"307401","Title":null,"Body":"<p>To add to @mrry's post, I put together a tutorial that explains how to load a TensorFlow graph with the C++ API. It's very minimal and should help you understand how all of the pieces fit together. Here's the meat of it:<\/p>\n\n<p>Requirements:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/bazelbuild\/bazel\/releases\" rel=\"noreferrer\">Bazel<\/a> installed<\/li>\n<li>Clone TensorFlow repo<\/li>\n<\/ul>\n\n<p>Folder structure:<\/p>\n\n<ul>\n<li><code>tensorflow\/tensorflow\/|project name|\/<\/code><\/li>\n<li><code>tensorflow\/tensorflow\/|project name|\/|project name|.cc (e.g. https:\/\/gist.github.com\/jimfleming\/4202e529042c401b17b7)<\/code><\/li>\n<li><code>tensorflow\/tensorflow\/|project name|\/BUILD<\/code><\/li>\n<\/ul>\n\n<p>BUILD:<\/p>\n\n<pre><code>cc_binary(\n    name = \"&lt;project name&gt;\",\n    srcs = [\"&lt;project name&gt;.cc\"],\n    deps = [\n        \"\/\/tensorflow\/core:tensorflow\",\n    ]\n)\n<\/code><\/pre>\n\n<p>Two caveats for which there are probably workarounds:<\/p>\n\n<ul>\n<li>Right now, building things needs to happen <em>within<\/em> the TensorFlow repo.<\/li>\n<li>The compiled binary is huge (103MB).<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/medium.com\/@jimfleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f\" rel=\"noreferrer\">https:\/\/medium.com\/@jimfleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f<\/a><\/p>\n"},{"AnswerId":"33931217","CreationDate":"2015-11-26T05:31:20.413","ParentId":null,"OwnerUserId":"1234504","Title":null,"Body":"<p>You can use this ShellScript to install (most) of it's dependencies, clone, build, compile and get all the necessary files into <code>..\/src\/includes<\/code> folder:<\/p>\n\n<p><a href=\"https:\/\/github.com\/node-tensorflow\/node-tensorflow\/blob\/master\/tools\/install.sh\" rel=\"noreferrer\">https:\/\/github.com\/node-tensorflow\/node-tensorflow\/blob\/master\/tools\/install.sh<\/a><\/p>\n"},{"AnswerId":"54685575","CreationDate":"2019-02-14T07:57:51.513","ParentId":null,"OwnerUserId":"10871473","Title":null,"Body":"<p>Tensorflow itself only provides very basic examples about C++ APIs.<br\/>\nHere is a good resource which includes examples of datasets, rnn, lstm, cnn and more <br\/>\n<a href=\"https:\/\/github.com\/rockzhuang\/tensorflow\" rel=\"nofollow noreferrer\">tensorflow c++ examples<\/a><\/p>\n"},{"AnswerId":"43920376","CreationDate":"2017-05-11T15:55:48.127","ParentId":null,"OwnerUserId":"384094","Title":null,"Body":"<p>First, after installing <code>protobuf<\/code> and <code>eigen<\/code>, you'd like to build Tensorflow:<\/p>\n\n<pre><code>.\/configure\nbazel build \/\/tensorflow:libtensorflow_cc.so\n<\/code><\/pre>\n\n<p>Then Copy the following include headers and dynamic shared library to <code>\/usr\/local\/lib<\/code> and <code>\/usr\/local\/include<\/code>:<\/p>\n\n<pre><code>mkdir \/usr\/local\/include\/tf\ncp -r bazel-genfiles\/ \/usr\/local\/include\/tf\/\ncp -r tensorflow \/usr\/local\/include\/tf\/\ncp -r third_party \/usr\/local\/include\/tf\/\ncp -r bazel-bin\/libtensorflow_cc.so \/usr\/local\/lib\/\n<\/code><\/pre>\n\n<p>Lastly, compile using an example:<\/p>\n\n<pre><code>g++ -std=c++11 -o tf_example \\\n-I\/usr\/local\/include\/tf \\\n-I\/usr\/local\/include\/eigen3 \\\n-g -Wall -D_DEBUG -Wshadow -Wno-sign-compare -w  \\\n-L\/usr\/local\/lib\/libtensorflow_cc \\\n`pkg-config --cflags --libs protobuf` -ltensorflow_cc tf_example.cpp\n<\/code><\/pre>\n"},{"AnswerId":"55612771","CreationDate":"2019-04-10T12:37:59.993","ParentId":null,"OwnerUserId":"1285444","Title":null,"Body":"<p>answers above are good enough to show how to build the library, but how to collect the headers are still tricky. here I share the little script I use to copy the necessary headers. <\/p>\n\n<p><code>SOURCE<\/code> is the first param, which is the tensorflow source(build) direcoty;<br>\n<code>DST<\/code> is the second param, which is the <code>include directory<\/code> holds the collected headers. (eg. in cmake, <code>include_directories(.\/collected_headers_here)<\/code>).<\/p>\n\n<pre><code>#!\/bin\/bash\n\nSOURCE=$1\nDST=$2\necho \"-- target dir is $DST\"\necho \"-- source dir is $SOURCE\"\n\nif [[ -e $DST ]];then\n    echo \"clean $DST\"\n    rm -rf $DST\n    mkdir $DST\nfi\n\n\n# 1. copy the source code c++ api needs\nmkdir -p $DST\/tensorflow\ncp -r $SOURCE\/tensorflow\/core $DST\/tensorflow\ncp -r $SOURCE\/tensorflow\/cc $DST\/tensorflow\ncp -r $SOURCE\/tensorflow\/c $DST\/tensorflow\n\n# 2. copy the generated code, put them back to\n# the right directories along side the source code\nif [[ -e $SOURCE\/bazel-genfiles\/tensorflow ]];then\n    prefix=\"$SOURCE\/bazel-genfiles\/tensorflow\"\n    from=$(expr $(echo -n $prefix | wc -m) + 1)\n\n    # eg. compiled protobuf files\n    find $SOURCE\/bazel-genfiles\/tensorflow -type f | while read line;do\n        #echo \"procese file --&gt; $line\"\n        line_len=$(echo -n $line | wc -m)\n        filename=$(echo $line | rev | cut -d'\/' -f1 | rev )\n        filename_len=$(echo -n $filename | wc -m)\n        to=$(expr $line_len - $filename_len)\n\n        target_dir=$(echo $line | cut -c$from-$to)\n        #echo \"[$filename] copy $line $DST\/tensorflow\/$target_dir\"\n        cp $line $DST\/tensorflow\/$target_dir\n    done\nfi\n\n\n# 3. copy third party files. Why?\n# In the tf source code, you can see #include \"third_party\/...\", so you need it\ncp -r $SOURCE\/third_party $DST\n\n# 4. these headers are enough for me now.\n# if your compiler complains missing headers, maybe you can find it in bazel-tensorflow\/external\ncp -RLf $SOURCE\/bazel-tensorflow\/external\/eigen_archive\/Eigen $DST\ncp -RLf $SOURCE\/bazel-tensorflow\/external\/eigen_archive\/unsupported $DST\ncp -RLf $SOURCE\/bazel-tensorflow\/external\/protobuf_archive\/src\/google $DST\ncp -RLf $SOURCE\/bazel-tensorflow\/external\/com_google_absl\/absl $DST\n<\/code><\/pre>\n"},{"AnswerId":"50194367","CreationDate":"2018-05-05T21:56:54.510","ParentId":null,"OwnerUserId":"602173","Title":null,"Body":"<p>If you don't want to build Tensorflow yourself and your operating system is Debian or Ubuntu, you can download prebuilt packages with the Tensorflow C\/C++ libraries. This distribution can be used for C\/C++ inference with CPU, GPU support is not included:<\/p>\n\n<p><a href=\"https:\/\/github.com\/kecsap\/tensorflow_cpp_packaging\/releases\" rel=\"noreferrer\">https:\/\/github.com\/kecsap\/tensorflow_cpp_packaging\/releases<\/a><\/p>\n\n<p>There are instructions written how to freeze a checkpoint in Tensorflow (TFLearn) and load this model for inference with the C\/C++ API:<\/p>\n\n<p><a href=\"https:\/\/github.com\/kecsap\/tensorflow_cpp_packaging\/blob\/master\/README.md\" rel=\"noreferrer\">https:\/\/github.com\/kecsap\/tensorflow_cpp_packaging\/blob\/master\/README.md<\/a><\/p>\n\n<p>Beware: I am the developer of this Github project.<\/p>\n"},{"AnswerId":"44626090","CreationDate":"2017-06-19T08:53:21.027","ParentId":null,"OwnerUserId":"4932728","Title":null,"Body":"<p>If you don't mind using CMake, there is also <a href=\"https:\/\/github.com\/FloopCZ\/tensorflow_cc\" rel=\"noreferrer\">tensorflow_cc<\/a> project that builds and installs TF C++ API for you, along with convenient CMake targets you can link against. The project README contains an example and Dockerfiles you can easily follow.<\/p>\n"},{"AnswerId":"43629668","CreationDate":"2017-04-26T09:05:35.240","ParentId":null,"OwnerUserId":"1076564","Title":null,"Body":"<p>I use a hack\/workaround to avoid having to build the whole TF library myself (which saves both time (it's set up in 3 minutes), disk space, installing dev dependencies, and size of the resulting binary). It's officially unsupported, but works well if you just want to quickly jump in.<\/p>\n\n<p>Install TF through pip (<code>pip install tensorflow<\/code> or <code>pip install tensorflow-gpu<\/code>). Then find its library <code>_pywrap_tensorflow.so<\/code> (TF 0.* - 1.0) or <code>_pywrap_tensorflow_internal.so<\/code> (TF 1.1+). In my case (Ubuntu) it's located at <code>\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so<\/code>. Then create a symlink to this library called <code>lib_pywrap_tensorflow.so<\/code> somewhere where your build system finds it (e.g. <code>\/usr\/lib\/local<\/code>). The prefix <code>lib<\/code> is important! You can also give it another <code>lib*.so<\/code> name - if you call it <code>libtensorflow.so<\/code>, you may get better compatibility with other programs written to work with TF.<\/p>\n\n<p>Then create a C++ project as you are used to (CMake, Make, Bazel, whatever you like). <\/p>\n\n<p>And then you're ready to just link against this library to have TF available for your projects (and you also have to link against <code>python2.7<\/code> libraries)! In CMake, you e.g. just add <code>target_link_libraries(target _pywrap_tensorflow python2.7)<\/code>.<\/p>\n\n<p>The C++ header files are located around this library, e.g. in <code>\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/include\/<\/code>.<\/p>\n\n<p>Once again: this way is officially unsupported and you may run in various issues. The library seems to be statically linked against e.g. protobuf, so you may run in odd link-time or run-time issues. But I am able to load a stored graph, restore the weights and run inference, which is IMO the most wanted functionality in C++.<\/p>\n"},{"AnswerId":"33622489","CreationDate":"2015-11-10T03:41:23.363","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To get started, you should download the source code from Github, by <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\">following the instructions here<\/a> (you'll need <a href=\"http:\/\/bazel.io\">Bazel<\/a> and a recent version of GCC).<\/p>\n\n<p>The C++ API (and the backend of the system) is in <code>tensorflow\/core<\/code>. Right now, only the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/public\/session.h\">C++ Session interface<\/a>, and the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/public\/tensor_c_api.h\">C API<\/a> are being supported. You can use either of these to execute TensorFlow graphs that have been built using the Python API and serialized to a <code>GraphDef<\/code> protocol buffer. There is also an experimental feature for building graphs in C++, but this is currently not quite as full-featured as the Python API (e.g. no support for auto-differentiation at present). You can see an example program that <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/cc\/tutorials\/example_trainer.cc\">builds a small graph in C++ here<\/a>.<\/p>\n\n<p>The second part of the C++ API is the API for adding a new <code>OpKernel<\/code>, which is the class containing implementations of numerical kernels for CPU and GPU. There are numerous examples of how to build these in <code>tensorflow\/core\/kernels<\/code>, as well as a <a href=\"http:\/\/tensorflow.org\/how_tos\/adding_an_op\/index.md\">tutorial for adding a new op in C++<\/a>.<\/p>\n"},{"AnswerId":"38772539","CreationDate":"2016-08-04T16:15:16.637","ParentId":null,"OwnerUserId":"5079659","Title":null,"Body":"<p>If you wish to avoid both building your projects with Bazel and generating a large binary, I have assembled a repository instructing the usage of the TensorFlow C++ library with CMake.  You can find it <a href=\"https:\/\/github.com\/cjweeks\/tensorflow-cmake\" rel=\"noreferrer\">here<\/a>.  The general ideas are as follows:<\/p>\n\n<ul>\n<li>Clone the TensorFlow repository.<\/li>\n<li>Add a build rule to <code>tensorflow\/BUILD<\/code> (the provided ones do not include all of the C++ functionality).<\/li>\n<li>Build the TensorFlow shared library.<\/li>\n<li>Install specific versions of Eigen and Protobuf, or add them as external dependencies.<\/li>\n<li>Configure your CMake project to use the TensorFlow library.<\/li>\n<\/ul>\n"},{"AnswerId":"49764888","CreationDate":"2018-04-11T01:02:14.067","ParentId":null,"OwnerUserId":"5103762","Title":null,"Body":"<p>If you are thinking into using Tensorflow c++ api on a standalone package you probably will need tensorflow_cc.so ( There is also a c api version tensorflow.so ) to build the c++ version you can use:<\/p>\n\n<pre><code>bazel build -c opt \/\/tensorflow:libtensorflow_cc.so\n<\/code><\/pre>\n\n<p>Note1: If you want to add intrinsics support you can add this flags as: <code>--copt=-msse4.2 --copt=-mavx<\/code><\/p>\n\n<p>Note2: If you are thinking into using OpenCV on your project as well, there is an issue when using both libs together (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/14267\" rel=\"noreferrer\">tensorflow issue<\/a>) and you should use <code>--config=monolithic<\/code>.<\/p>\n\n<p>After building the library you need to add it to your project.\nTo do that you can include this paths:<\/p>\n\n<pre><code>tensorflow\ntensorflow\/bazel-tensorflow\/external\/eigen_archive\ntensorflow\/bazel-tensorflow\/external\/protobuf_archive\/src\ntensorflow\/bazel-genfiles\n<\/code><\/pre>\n\n<p>And link the library to your project:<\/p>\n\n<pre><code>tensorflow\/bazel-bin\/tensorflow\/libtensorflow_framework.so (unused if you build with --config=monolithic)\ntensorflow\/bazel-bin\/tensorflow\/libtensorflow_cc.so\n<\/code><\/pre>\n\n<p>And when you are building your project you should also specify to your compiler that you are going to use c++11 standards.<\/p>\n\n<p>Side Note: Paths relative to tensorflow version 1.5 (You may need to check if in your version anything changed).<\/p>\n\n<p>Also this link helped me a lot into finding all this infos: <a href=\"http:\/\/www.bitbionic.com\/2017\/08\/18\/run-your-keras-models-in-c-tensorflow\/\" rel=\"noreferrer\">link<\/a><\/p>\n"}]}
{"QuestionId":33621547,"AnswerCount":1,"Tags":"<python><docker><anaconda><tensorflow>","CreationDate":"2015-11-10T01:52:11.710","AcceptedAnswerId":33622640.0,"OwnerUserId":2073494.0,"Title":"\"No such file or directory\" when starting convolutional.py script on tensorflow docker image","Body":"<p>I don't have a Linux or Mac machine so in order to check out TensorFlow on Windows, installed docker and downloaded the image of tensorflow-full.\nWhen I run the following command:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error message:<\/p>\n\n<pre><\/pre>\n\n<p>Currently on Win 8.1 and have anaconda installed.<\/p>\n","answers":[{"AnswerId":"33622640","CreationDate":"2015-11-10T03:59:32.383","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like the error message is caused by trying to execute a script file (<code>...\/convolutional.py<\/code>) that is inside the container, using the Python interpreter outside the container.<\/p>\n\n<p>First of all, follow the steps here to ensure that Docker is configured successfully on your Windows machine:<\/p>\n\n<p><a href=\"http:\/\/docs.docker.com\/engine\/installation\/windows\/#using-docker-from-windows-command-prompt-cmd-exe\" rel=\"noreferrer\">http:\/\/docs.docker.com\/engine\/installation\/windows\/#using-docker-from-windows-command-prompt-cmd-exe<\/a><\/p>\n\n<p>Once you've successfully run the <code>hello-world<\/code> container, run the following command to start the TensorFlow container:<\/p>\n\n<pre><code>docker run -it b.gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p>(Note that, depending on your terminal, the previous step may or may not work. A common error is <code>cannot enable tty mode on non tty input<\/code>. In that case, run the following command to connect to the VM that is hosting the containers:<\/p>\n\n<pre><code>docker-machine ssh default\n<\/code><\/pre>\n\n<p>...then at the resulting prompt, the <code>docker run<\/code> command again.)<\/p>\n\n<p>At the resulting prompt, you should be able to run the script with the following command:<\/p>\n\n<pre><code>python \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n"}]}
{"QuestionId":33621643,"AnswerCount":3,"Tags":"<k-means><tensorflow>","CreationDate":"2015-11-10T02:03:39.033","AcceptedAnswerId":33637578.0,"OwnerUserId":2250985.0,"Title":"How would I implement k-means with TensorFlow?","Body":"<p>The intro tutorial, which uses the built-in gradient descent optimizer, makes a lot of sense. However, k-means isn't just something I can plug into gradient descent. It seems like I'd have to write my own sort of optimizer, but I'm not quite sure how to do that given the TensorFlow primitives.<\/p>\n\n<p>What approach should I take?<\/p>\n","answers":[{"AnswerId":"43930092","CreationDate":"2017-05-12T05:43:03.913","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Most of the answers I have seen so far focuses just on the 2d version (when you need to cluster points in 2 dimensions). Here is my implementation of the clustering in arbitrary dimensions.<\/p>\n\n<hr>\n\n<p>Basic idea of <a href=\"https:\/\/en.wikipedia.org\/wiki\/K-means_clustering\" rel=\"nofollow noreferrer\">k-means algorithm<\/a> in n dims:<\/p>\n\n<ul>\n<li>generate random k starting points<\/li>\n<li>do this till you exceed the patience or the cluster assignment does not change:\n\n<ul>\n<li>assign each point to the closest starting point<\/li>\n<li>recalculate the location of each starting point by taking the average among it's cluster<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<hr>\n\n<p>To be able to somehow validate the results I will attempt to cluster MNIST images. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nfrom random import randint\nfrom collections import Counter\nfrom tensorflow.examples.tutorials.mnist import input_data\n\nmnist = input_data.read_data_sets(\"MNIST_data\/\")\nX, y, k = mnist.test.images, mnist.test.labels, 10\n<\/code><\/pre>\n\n<p>So here <strong>X<\/strong> is my data to cluster <code>(10000, 784)<\/code>, <strong>y<\/strong> is the real number, and <strong>k<\/strong> is the number of cluster (which is the same as the number of digits. Now the actual algorithm:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># select random points as a starting position. You can do better by randomly selecting k points.\nstart_pos = tf.Variable(X[np.random.randint(X.shape[0], size=k),:], dtype=tf.float32)\ncentroids = tf.Variable(start_pos.initialized_value(), 'S', dtype=tf.float32)\n\n# populate points\npoints           = tf.Variable(X, 'X', dtype=tf.float32)\nones_like        = tf.ones((points.get_shape()[0], 1))\nprev_assignments = tf.Variable(tf.zeros((points.get_shape()[0], ), dtype=tf.int64))\n\n# find the distance between all points: http:\/\/stackoverflow.com\/a\/43839605\/1090562\np1 = tf.matmul(\n    tf.expand_dims(tf.reduce_sum(tf.square(points), 1), 1),\n    tf.ones(shape=(1, k))\n)\np2 = tf.transpose(tf.matmul(\n    tf.reshape(tf.reduce_sum(tf.square(centroids), 1), shape=[-1, 1]),\n    ones_like,\n    transpose_b=True\n))\ndistance = tf.sqrt(tf.add(p1, p2) - 2 * tf.matmul(points, centroids, transpose_b=True))\n\n# assign each point to a closest centroid\npoint_to_centroid_assignment = tf.argmin(distance, axis=1)\n\n# recalculate the centers\ntotal = tf.unsorted_segment_sum(points, point_to_centroid_assignment, k)\ncount = tf.unsorted_segment_sum(ones_like, point_to_centroid_assignment, k)\nmeans = total \/ count\n\n# continue if there is any difference between the current and previous assignment\nis_continue = tf.reduce_any(tf.not_equal(point_to_centroid_assignment, prev_assignments))\n\nwith tf.control_dependencies([is_continue]):\n    loop = tf.group(centroids.assign(means), prev_assignments.assign(point_to_centroid_assignment))\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\n# do many iterations. Hopefully you will stop because of has_changed is False\nhas_changed, cnt = True, 0\nwhile has_changed and cnt &lt; 300:\n    cnt += 1\n    has_changed, _ = sess.run([is_continue, loop])\n\n# see how the data is assigned\nres = sess.run(point_to_centroid_assignment)\n<\/code><\/pre>\n\n<p>Now it is time check how good are our clusters. To do this we will group all the real numbers that appeared in the cluster together. After that we will see the most popular choices in that cluster. In a case of the perfect clustering we will have the just one value in each group. In case of random cluster each value will be approximately equally represented in the group.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>nums_in_clusters = [[] for i in xrange(10)]\nfor cluster, real_num in zip(list(res), list(y)):\n    nums_in_clusters[cluster].append(real_num)\n\nfor i in xrange(10):\n    print Counter(nums_in_clusters[i]).most_common(3)\n<\/code><\/pre>\n\n<p>This gives me something like this:<\/p>\n\n<pre><code>[(0, 738), (6, 18), (2, 11)]\n[(1, 641), (3, 53), (2, 51)]\n[(1, 488), (2, 115), (7, 56)]\n[(4, 550), (9, 533), (7, 280)]\n[(7, 634), (9, 400), (4, 302)]\n[(6, 649), (4, 27), (0, 14)]\n[(5, 269), (6, 244), (0, 161)]\n[(8, 646), (5, 164), (3, 125)]\n[(2, 698), (3, 34), (7, 14)]\n[(3, 712), (5, 290), (8, 110)]\n<\/code><\/pre>\n\n<p>This is pretty good because majority of the counts is in the first group. You see that clustering confuses 7 and 9, 4 and 5. But 0 is clustered pretty nicely.<\/p>\n\n<p><strong>A few approaches how to improve this:<\/strong><\/p>\n\n<ul>\n<li>run the algorithm a few times and select the best one (based on the distance to clusters)<\/li>\n<li>handling cases when nothing is assigned to a cluster. In my case you will get Nan in <code>means<\/code> variable because <code>count<\/code> is 0.<\/li>\n<li>random points initialization.<\/li>\n<\/ul>\n"},{"AnswerId":"33637578","CreationDate":"2015-11-10T19:04:10.410","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p><strong>(note:  You can now get <a href=\"https:\/\/gist.github.com\/dave-andersen\/265e68a5e879b5540ebc\">a more polished version of this code as a gist on github<\/a>.)<\/strong><\/p>\n\n<p>you can definitely do it, but you need to define your own optimization criteria (for k-means, it's usually a max iteration count and when the assignment stabilizes).  Here's an example of how you might do it (there are probably more optimal ways to implement it, and definitely better ways to select the initial points).  It's basically like you'd do it in numpy if you were trying really hard to stay away from doing things iteratively in python:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nimport time\n\nN=10000\nK=4\nMAX_ITERS = 1000\n\nstart = time.time()\n\npoints = tf.Variable(tf.random_uniform([N,2]))\ncluster_assignments = tf.Variable(tf.zeros([N], dtype=tf.int64))\n\n# Silly initialization:  Use the first two points as the starting                \n# centroids.  In the real world, do this better.                                 \ncentroids = tf.Variable(tf.slice(points.initialized_value(), [0,0], [K,2]))\n\n# Replicate to N copies of each centroid and K copies of each                    \n# point, then subtract and compute the sum of squared distances.                 \nrep_centroids = tf.reshape(tf.tile(centroids, [N, 1]), [N, K, 2])\nrep_points = tf.reshape(tf.tile(points, [1, K]), [N, K, 2])\nsum_squares = tf.reduce_sum(tf.square(rep_points - rep_centroids),\n                            reduction_indices=2)\n\n# Use argmin to select the lowest-distance point                                 \nbest_centroids = tf.argmin(sum_squares, 1)\ndid_assignments_change = tf.reduce_any(tf.not_equal(best_centroids,\n                                                    cluster_assignments))\n\ndef bucket_mean(data, bucket_ids, num_buckets):\n    total = tf.unsorted_segment_sum(data, bucket_ids, num_buckets)\n    count = tf.unsorted_segment_sum(tf.ones_like(data), bucket_ids, num_buckets)\n    return total \/ count\n\nmeans = bucket_mean(points, best_centroids, K)\n\n# Do not write to the assigned clusters variable until after                     \n# computing whether the assignments have changed - hence with_dependencies\nwith tf.control_dependencies([did_assignments_change]):\n    do_updates = tf.group(\n        centroids.assign(means),\n        cluster_assignments.assign(best_centroids))\n\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\n\nchanged = True\niters = 0\n\nwhile changed and iters &lt; MAX_ITERS:\n    iters += 1\n    [changed, _] = sess.run([did_assignments_change, do_updates])\n\n[centers, assignments] = sess.run([centroids, cluster_assignments])\nend = time.time()\nprint (\"Found in %.2f seconds\" % (end-start)), iters, \"iterations\"\nprint \"Centroids:\"\nprint centers\nprint \"Cluster assignments:\", assignments\n<\/code><\/pre>\n\n<p>(Note that a real implementation would need to be more careful about initial cluster selection, avoiding problem cases with all points going to one cluster, etc.  This is just a quick demo.  I've updated my answer from earlier to make it a bit more clear and \"example-worthy\".)<\/p>\n"},{"AnswerId":"54630703","CreationDate":"2019-02-11T12:32:58.943","ParentId":null,"OwnerUserId":"6261650","Title":null,"Body":"<p>Nowadays you could directly use (or take inspiration from) the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/factorization\/KMeansClustering\" rel=\"nofollow noreferrer\">KMeansClustering Estimator<\/a>. You can take a look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/contrib\/factorization\/python\/ops\/clustering_ops.py\" rel=\"nofollow noreferrer\">its implementation on GitHub<\/a>.<\/p>\n"}]}
{"QuestionId":33622019,"AnswerCount":4,"Tags":"<python><pip><tensorflow>","CreationDate":"2015-11-10T02:45:28.997","AcceptedAnswerId":null,"OwnerUserId":5544872.0,"Title":"Tensorflow successfully installs on mac but gets ImportError on copyreg when used","Body":"<p>After successfully , importing the tensorflow library fails.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33636999","CreationDate":"2015-11-10T18:28:33.907","ParentId":null,"OwnerUserId":"610569","Title":null,"Body":"<p><code>copyreg<\/code> is a <code>python3<\/code> function that is available in the <code>six<\/code> module in <code>python2.x<\/code>, see <a href=\"https:\/\/docs.python.org\/2\/library\/copy_reg.html#module-copy_reg\" rel=\"nofollow\">https:\/\/docs.python.org\/2\/library\/copy_reg.html#module-copy_reg<\/a><\/p>\n\n<p>To get <code>copyreg<\/code>, you have to install <code>six<\/code>:<\/p>\n\n<pre><code>pip install -U six\n<\/code><\/pre>\n\n<p>(Note: In <code>python2<\/code>, you access can either access the function with (i) <code>six.copy_reg<\/code> or when a module is imported with <code>six.moves.*<\/code>, it keeps the <code>python3<\/code> syntax, i.e. <code>six.moves.copyreg<\/code>)<\/p>\n"},{"AnswerId":"33641112","CreationDate":"2015-11-10T22:46:51.767","ParentId":null,"OwnerUserId":"4237753","Title":null,"Body":"<p>Since non of the suggested fixes worked for me, I changed the line    <\/p>\n\n<pre><code>import six.moves.copyreg as copyreg\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>from six.moves import copyreg\n<\/code><\/pre>\n\n<p>and this seemed like this fixed this problem. However, I got another ImportError instead saying<\/p>\n\n<pre><code>dlopen(\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so, 2): no suitable image found.  Did find:\n\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so: mach-o, but wrong architecture\n<\/code><\/pre>\n\n<p>Just posting this as an answer if someone else maybe has more success in this than me.<\/p>\n"},{"AnswerId":"33622507","CreationDate":"2015-11-10T03:43:36.203","ParentId":null,"OwnerUserId":"4375459","Title":null,"Body":"<p>Solution: TensorFlow depends on protobuf which require six-1.10.0. Apple's default python environment has six-1.4.1 and may be difficult to upgrade. So we recommend either installing a separate copy of python via homebrew:<\/p>\n\n<p>brew install python<\/p>\n\n<p>or building \/ using TensorFlow within virtualenv as described above.<\/p>\n"},{"AnswerId":"33625805","CreationDate":"2015-11-10T08:37:35.777","ParentId":null,"OwnerUserId":"3929051","Title":null,"Body":"<p>You can upgrade to six-1.10.x using <\/p>\n\n<pre><code>easy_install -U six\n<\/code><\/pre>\n\n<p>This will upgrade the current version of six from 1.4 to 1.10.x, that is required by tensorflow. <\/p>\n"}]}
{"QuestionId":33622333,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><softmax>","CreationDate":"2015-11-10T03:24:42.220","AcceptedAnswerId":null,"OwnerUserId":5544931.0,"Title":"Caffe: Softmax with temperature","Body":"<p>I am working on implementing Hinton's Knowledge distillation <a href=\"http:\/\/arxiv.org\/abs\/1503.02531\" rel=\"nofollow\">paper<\/a>. The first step is to store the soft targets of a \"cumbersome model\" with a higher temperature (i.e. I don't need to train the network, just need to do forward pass per image and store the soft targets with a temperature ).<br>\nIs there a way I can get the output of Alexnet or googlenet soft targets but with a different temperature?<br>\nI need to modify the soft-max with .<br>\nNeed to divide the outputs of the final fully connected layer with a temperature . I only need this for the forward pass (not for training). <\/p>\n","answers":[{"AnswerId":"33624676","CreationDate":"2015-11-10T07:14:18.400","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I believe there are three options to solve this problem<\/p>\n\n<p><strong>1.<\/strong> Implement your own <code>Softmax<\/code> layer with a temperature parameter. It should be quite straight forward to modify the code of <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/softmax_layer.cpp\" rel=\"nofollow noreferrer\"><code>softmax_layer.cpp<\/code><\/a> to take into account a \"temperature\" <code>T<\/code>. You might need to tweak the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\" rel=\"nofollow noreferrer\"><code>caffe.proto<\/code><\/a> as well to allow for parsing <code>Softmax<\/code> layer with an extra parameter.<\/p>\n\n<p><strong>2.<\/strong> Implement the layer as a <a href=\"https:\/\/stackoverflow.com\/a\/41481539\/1714410\">python layer<\/a>.  <\/p>\n\n<p><strong>3.<\/strong> If you only need a forward pass, i.e. \"extracting features\", then you can simply output as features the \"top\" of the layer <strong>before<\/strong> the softmax layer and do the softmax with temperature outside caffe altogether.<\/p>\n\n<p><strong>4.<\/strong> You can add <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L1062-L1097\" rel=\"nofollow noreferrer\"><code>Scale<\/code><\/a> layer before the top <code>Softmax<\/code> layer:<\/p>\n\n<pre><code>layer {\n  type: \"Scale\"\n  name: \"temperature\"\n  bottom: \"zi\"\n  top: \"zi\/T\"\n  scale_param { \n    filler: { type: 'constant' value: 1\/T }  # replace \"1\/T\" with the actual value of 1\/T.\n  }\n  param { lr_mult: 0 decay_mult: 0 } # make sure temperature is fixed\n}\nlayer {\n  type: \"Softmax\"\n  name: \"prob\"\n  bottom: \"zi\/T\"\n  top: \"pi\"\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":33622613,"AnswerCount":11,"Tags":"<install><pip><python-wheel><tensorflow>","CreationDate":"2015-11-10T03:56:32.017","AcceptedAnswerId":33635645.0,"OwnerUserId":3775886.0,"Title":"Tensorflow installation error: not a supported wheel on this platform","Body":"<p>when I try to install tensorflow by cloning from git, I run into the error \"no module named copyreg,\" so I tried installing using a virtualenv. However, I then run into this error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't see this under the common problems section, so any help would be appreciated! Thank you. <\/p>\n","answers":[{"AnswerId":"46323955","CreationDate":"2017-09-20T13:51:40.737","ParentId":null,"OwnerUserId":"815612","Title":null,"Body":"<p>Make sure that the wheel is, well, supported by your platform. Pip uses the wheel's filename to determine compatibility. The format is:<\/p>\n\n<pre><code>tensorflow-{version}-{python version}-none-{your platform}.whl\n<\/code><\/pre>\n\n<p>I didn't realize that <code>x86_64<\/code> refers to x64, I thought it meant <em>either<\/em> x86 or x64, so I banged my head against this futilely for some time. Tensorflow is not available for 32 bit systems, unless you want to compile it yourself.<\/p>\n"},{"AnswerId":"33635645","CreationDate":"2015-11-10T17:09:59.547","ParentId":null,"OwnerUserId":"3057648","Title":null,"Body":"<p>I too got the same problem<br>\nI downloaded <code>get-pip.py<\/code> from <a href=\"https:\/\/bootstrap.pypa.io\/get-pip.py\" rel=\"noreferrer\">https:\/\/bootstrap.pypa.io\/get-pip.py<\/a><\/p>\n\n<p>and then ran <code>python2.7 get-pip.py<\/code> for installing <code>pip2.7<\/code><\/p>\n\n<p>and then ran the <code>pip install<\/code> command with <code>python2.7<\/code> as follows<\/p>\n\n<p><strong>For Ubuntu\/Linux:<\/strong><\/p>\n\n<pre><code>python2.7 -m pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p><strong>For Mac OS X:<\/strong><\/p>\n\n<pre><code>python2.7 -m pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>this should work just fine as it did for me :)<\/p>\n\n<p>I followed these instructions from <a href=\"https:\/\/askubuntu.com\/questions\/695981\/platform-not-supported-for-tensorflow-on-ubuntu-14-04-2\">here<\/a><\/p>\n"},{"AnswerId":"44042340","CreationDate":"2017-05-18T08:23:08.960","ParentId":null,"OwnerUserId":"4995051","Title":null,"Body":"<p>Seems that tensorflow only work on python 3.5 at the moment, try to run this command before running the <strong>pip install<\/strong><\/p>\n\n<pre><code>conda create --name tensorflow python=3.5\n<\/code><\/pre>\n\n<p>After this running the following lines :<\/p>\n\n<p>For <strong>cpu<\/strong> :<\/p>\n\n<pre><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>For <strong>gpu<\/strong> : <\/p>\n\n<pre><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.1.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>Should work like a charm,<\/p>\n\n<p>Cheers<\/p>\n"},{"AnswerId":"56306373","CreationDate":"2019-05-25T15:48:49.550","ParentId":null,"OwnerUserId":"10198965","Title":null,"Body":"<p>This may mean that you are installing the wrong pre-build binary<\/p>\n\n<p>since my CPU on Ubuntu 18.04 my download url was:\n<a href=\"https:\/\/github.com\/lakshayg\/tensorflow-build\/releases\/download\/tf1.12.0-ubuntu18.04-py2-py3\/tensorflow-1.12.0-cp36-cp36m-linux_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/github.com\/lakshayg\/tensorflow-build\/releases\/download\/tf1.12.0-ubuntu18.04-py2-py3\/tensorflow-1.12.0-cp36-cp36m-linux_x86_64.whl<\/a><\/p>\n\n<p>as it can be found on this github page:\n<a href=\"https:\/\/github.com\/lakshayg\/tensorflow-build\" rel=\"nofollow noreferrer\">https:\/\/github.com\/lakshayg\/tensorflow-build<\/a><\/p>\n\n<pre><code>pip install --ignore-installed --upgrade &lt;LOCAL PATH \/ BINARY-URL&gt;\n<\/code><\/pre>\n\n<p>resolved the issue for me.<\/p>\n"},{"AnswerId":"50667837","CreationDate":"2018-06-03T15:07:15.887","ParentId":null,"OwnerUserId":"4726173","Title":null,"Body":"<p>I was trying to install from source, and got that error. (Why would a wheel built on this machine not be compatible with it-?)<\/p>\n\n<p>For me, the tag --ignore-installed made all the difference.<\/p>\n\n<pre><code>pip install --ignore-installed \/tmp\/tensorflow_pkg\/tensorflow-1.8.0-cp36-cp36m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>worked, while<\/p>\n\n<pre><code>pip install \/tmp\/tensorflow_pkg\/tensorflow-1.8.0-cp36-cp36m-linux_x86_64.whl \n<\/code><\/pre>\n\n<p>threw the abovementioned error.<\/p>\n\n<p>Context: Conda environment; might have been a problem specific to this<\/p>\n"},{"AnswerId":"57317718","CreationDate":"2019-08-01T22:26:49.747","ParentId":null,"OwnerUserId":"2695227","Title":null,"Body":"<p>I faced the same issue and tried all the solutions that folks suggested here and other links (like <a href=\"https:\/\/askubuntu.com\/questions\/695981\/platform-not-supported-for-tensorflow-on-ubuntu-14-04-2\">https:\/\/askubuntu.com\/questions\/695981\/platform-not-supported-for-tensorflow-on-ubuntu-14-04-2<\/a>). <\/p>\n\n<p>It was so frustrating because using \n<code>print(wheel.pep425tags.get_supported())<\/code>\nI could see that my ubuntu supported ('cp37', 'cp37m', 'linux_x86_64') and that was exactly what I was trying to install (from <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.14.0-cp37-cp37m-linux_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.14.0-cp37-cp37m-linux_x86_64.whl<\/a>).<\/p>\n\n<p>What at the end fixed it was to simply download the package first and then <\/p>\n\n<pre><code>pip install tensorflow-1.14.0-cp37-cp37m-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"47582644","CreationDate":"2017-11-30T21:07:57.840","ParentId":null,"OwnerUserId":"7907591","Title":null,"Body":"<p>On Windows 10, with Python 3.6.X version I was facing same then after checking deliberately , I noticed I had Python-32 bit installation on my 64 bit machine. Remember TensorFlow is only compatible with 64bit installation of python. <strong>Not 32 bit of Python<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cABEM.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cABEM.jpg\" alt=\"installation requirements\"><\/a><\/p>\n\n<p>If we download Python from python.org , the default installation would be 32 bit. So we have to download 64 bit installer manually to install Python 64 bit. And then add <\/p>\n\n<ol>\n<li>C:\\Users\\\\AppData\\Local\\Programs\\Python\\Python36<\/li>\n<li>C:\\Users\\\\AppData\\Local\\Programs\\Python\\Python36\\Scripts<\/li>\n<\/ol>\n\n<p>Then run <code>gpupdate \/Force<\/code> on command prompt. If python command doesnt work for 64 bit restart your machine.<\/p>\n\n<p>Then run python on command prompt. It should show 64 bit<\/p>\n\n<pre><code>C:\\Users\\YOURNAME&gt;python\nPython 3.6.3 (v3.6.3:2c5fed8, Oct  3 2017, 18:11:49) [MSC v.1900 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n<\/code><\/pre>\n\n<p>Then run below command to install tensorflow CPU version(recommended) <\/p>\n\n<pre><code>pip3 install --upgrade tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"45029431","CreationDate":"2017-07-11T08:37:00.727","ParentId":null,"OwnerUserId":"4431354","Title":null,"Body":"<p>actually, you can use Python 3.5.<em>, I successfully solved this problem with Python 3.5.3.  Modify python version to 3.5.<\/em> in conda, see <a href=\"https:\/\/conda.io\/docs\/py2or3.html\" rel=\"nofollow noreferrer\">https:\/\/conda.io\/docs\/py2or3.html<\/a>.  then go to <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>, and repeat from \"Create a conda environment named tensorflow by invoking the following command\" BLA BLA.....<\/p>\n"},{"AnswerId":"35530708","CreationDate":"2016-02-21T00:14:54.473","ParentId":null,"OwnerUserId":"3637833","Title":null,"Body":"<p>After activating the virtualenv, be sure to upgrade pip to the latest version.<\/p>\n\n<pre><code>(your_virtual_env)$  pip install --upgrade pip\n<\/code><\/pre>\n\n<p>And now you'll be able to install tensor-flow correctly (for linux):<\/p>\n\n<pre><code>(your_virtual_env)$  pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.0-py2-none-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"40036679","CreationDate":"2016-10-14T06:49:19.407","ParentId":null,"OwnerUserId":"6929661","Title":null,"Body":"<p>The pip wheel contains the python version in its name (cp34-cp34m). If you download the whl file and rename it to say <strong>py3-none<\/strong> or instead, it should work. Can you try that?\n<\/p>\n\n<p>The installation won't work for anaconda users that choose python 3 support because the installation procedure is asking to create a python 3.5 environment and the file is currently called cp34-cp34m. So renaming it would do the job for now.<\/p>\n\n<pre><code>sudo pip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.7.0-cp34-cp34m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>This will produced the exact error message you got above. However, when you will downloaded the file yourself and rename it to \"tensorflow-0.7.0-py3-none-linux_x86_64.whl\", then execute the command again with changed filename, it should work fine.<\/p>\n"},{"AnswerId":"41408906","CreationDate":"2016-12-31T13:55:36.453","ParentId":null,"OwnerUserId":"94958","Title":null,"Body":"<p>I was trying to do the windows-based install and kept getting this error. <\/p>\n\n<p>Turns out you <strong>have to<\/strong> have python 3.5.2. Not 2.7, not 3.6.x-- nothing other than 3.5.2. <\/p>\n\n<p>After installing python 3.5.2 the <code>pip install<\/code> worked. <\/p>\n"}]}
{"QuestionId":33622842,"AnswerCount":5,"Tags":"<ubuntu><tensorflow>","CreationDate":"2015-11-10T04:24:02.350","AcceptedAnswerId":33623372.0,"OwnerUserId":46190.0,"Title":"Error in python after 'import tensorflow': TypeError: __init__() got an unexpected keyword argument 'syntax'","Body":"\n\n<p>I installed TensorFlow on my Ubuntu 15.10 machine as instructed for CPU only:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Then when I run the Python REPL and import tensorflow, I get:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>I have the Ubuntu protobuf-compiler package installed and it's version 2.6.1-1.2<\/p>\n","answers":[{"AnswerId":"51603748","CreationDate":"2018-07-30T23:58:37.743","ParentId":null,"OwnerUserId":"3160397","Title":null,"Body":"<p>I was having similar issue today. It was because I had <strong>python 3.7<\/strong> installed which is not compatible with tensorflow as of July 30 2018.\nDowngrading to <code>3.6<\/code> fixed the issue. Posted <a href=\"https:\/\/stackoverflow.com\/questions\/51600158\/issues-installing-tensorflow-on-mac#51603687\">solution<\/a> here if anyone is interested. <\/p>\n\n<p>Following github issue was helpful when resolving this issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022<\/a><\/p>\n"},{"AnswerId":"33623372","CreationDate":"2015-11-10T05:17:49.533","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Several users have reported issues that arise when an older version of protobuf is installed.<\/p>\n\n<p>TensorFlow requires (and uses a copy of) protobuf-3.0.0a4. However it seems to be conflicting with your installed version (perhaps due to how the PYTHONPATH environment variable is configured on your system?). Can you try uninstalling the <code>python-protobuf<\/code> package if it is installed, and then seeing if TensorFlow starts correctly?<\/p>\n\n<p>Some users have had success when using a virtualenv. For instructions on installing TensorFlow into a virtualenv, see <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#virtualenv-based_installation\">the getting started guide<\/a>.<\/p>\n"},{"AnswerId":"33638454","CreationDate":"2015-11-10T19:57:27.370","ParentId":null,"OwnerUserId":"3397171","Title":null,"Body":"<p>This <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11#issuecomment-155153983\" rel=\"noreferrer\">github comment<\/a> solved my problem.<\/p>\n\n<p>In short, you'll need <code>protobuf&gt;=3.0<\/code> installed to make tensorflow work.<\/p>\n\n<p>First uninstall whichever <code>protobuf<\/code> you have, then <code>pip install 'protobuf&gt;=3.0.0a3'<\/code>.<\/p>\n\n<p>It's not just an issue for Ubuntu, Mac OS also has a similar issue.<\/p>\n"},{"AnswerId":"34040439","CreationDate":"2015-12-02T10:31:12.703","ParentId":null,"OwnerUserId":"5629831","Title":null,"Body":"<p>It worked for me:<\/p>\n\n<pre><code>pip uninstall protobuf\npip uninstall tensorflow\nbrew uninstall protobuf\npip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n\nimport google.protobuf\n&gt;&gt;&gt; print google.protobuf.__version__\n3.0.0a4\n&gt;&gt;&gt; import tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"42552766","CreationDate":"2017-03-02T10:24:46.783","ParentId":null,"OwnerUserId":"7629885","Title":null,"Body":"<p>I got the same error while running on the python version 2.7. But everything went fine when i ran it using <b>python version 3<\/b>.<\/p>\n\n<p>Try the following commands on bash shell:<\/p>\n\n<pre><code>python3\nPython 3.5.2 (default, Nov 17 2016, 17:05:23)\n[GCC 5.4.0 20160609] on linux \nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; exit()\n<\/code><\/pre>\n\n<p>It worked for me.<\/p>\n"}]}
{"QuestionId":33623453,"AnswerCount":5,"Tags":"<python><tensorflow><pip><install>","CreationDate":"2015-11-10T05:25:15.040","AcceptedAnswerId":null,"OwnerUserId":2951094.0,"Title":"pip installation error \"No such file or directory: setup.py\"","Body":"\n\n<p>I'm getting installation error because pip couldn't find setup.py.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>I get this error message.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Can anyone tell me how to solve this problem?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"33861955","CreationDate":"2015-11-23T00:19:45.283","ParentId":null,"OwnerUserId":"4912028","Title":null,"Body":"<p>Tensorflow requires python -V 2.7. Since, lot of machine use the stable 2.6 version of python for their system build I would recommend parallel\/alternative install of python 2.7. Then you can install pip with 2.7 version python interpreter. Please see below:<\/p>\n\n<pre><code>#sudo wget https:\/\/bitbucket.org\/pypa\/setuptools\/raw\/bootstrap\/ez_setup.py\n#sudo python2.7 ez_setup.py\n#easy_install-2.7 pip\n#sudo pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"45384116","CreationDate":"2017-07-29T00:37:14.497","ParentId":null,"OwnerUserId":"768183","Title":null,"Body":"<p>I'm using ubuntu 12.04 64bits.\nI solve this issue by:<\/p>\n\n<pre><code>mkdir -p ~\/code\/download\/lib\/tensorflow\/\ncd ~\/code\/download\/lib\/tensorflow\/\nvirtualenv --system-site-packages tensorflow_for_ubuntu12_04_64bit_python2_7\nsource tensorflow_for_ubuntu12_04_64bit_python2_7\/bin\/activate\ncd tensorflow_for_ubuntu12_04_64bit_python2_7\npip install --upgrade tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"49989255","CreationDate":"2018-04-23T20:13:44.707","ParentId":null,"OwnerUserId":"4946302","Title":null,"Body":"<p>the filename you are running sdist with must be named <code>setup.py<\/code>\nrunning the file as say <code>setup_something.py<\/code> will fail<\/p>\n"},{"AnswerId":"33680205","CreationDate":"2015-11-12T19:56:37.123","ParentId":null,"OwnerUserId":"64253","Title":null,"Body":"<p>from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/56\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/56<\/a><\/p>\n\n<blockquote>\n  <p>The command to type is <strong>\"pip install --upgrade pip\"<\/strong>, and this should be\n  added to the instructions right after where they tell the user to\n  <strong>\"source bin\/activate\"<\/strong><\/p>\n<\/blockquote>\n"},{"AnswerId":"37322987","CreationDate":"2016-05-19T12:18:00.320","ParentId":null,"OwnerUserId":"5222127","Title":null,"Body":"<p>Try upgrade protobuf with pip install:<\/p>\n\n<pre><code>sudo pip install --upgrade protobuf \nsudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/protobuf-3.0.0b2.post2-cp27-none-any.whl\n<\/code><\/pre>\n\n<p>worked for me. The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#pip-installation\" rel=\"nofollow\">documentation<\/a> has provided many solutions for various installation problems.<\/p>\n"}]}
{"QuestionId":33624048,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-10T06:17:43.417","AcceptedAnswerId":33624832.0,"OwnerUserId":1230772.0,"Title":"Fail to run word embedding example in tensorflow tutorial with GPUs","Body":"<p>I am trying to run the word embedding example code at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/g3doc\/tutorials\/word2vec\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/g3doc\/tutorials\/word2vec<\/a> (installed with GPU version of tensorflow under Ubuntu 14.04), but it returns the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>When I run the code in CPU version tensorflow, it works just fine. But not for GPU version. I also tried to use tf.device('\/cpu:0') to force it using CUP instead of GPU, but it produces the same output.<\/p>\n\n<p>Is there any function in this example cannot be run in GPUs? And how do I switch to CPU without reinstalling CPU version of tensorflow since tf.device('\/cpu:0') not working?<\/p>\n","answers":[{"AnswerId":"33624715","CreationDate":"2015-11-10T07:17:55.540","ParentId":null,"OwnerUserId":"5545462","Title":null,"Body":"<ul>\n<li>scatter_sub is only supported on cpu in the current release.<\/li>\n<li>I expect that adding on Ln119: with tf.device(\"\/cpu:0\") should force everything to use cpus. How did you use tf.device?<\/li>\n<\/ul>\n"},{"AnswerId":"33624832","CreationDate":"2015-11-10T07:28:33.523","ParentId":null,"OwnerUserId":"1090721","Title":null,"Body":"<p>It seems a whole bunch of operations used in this example aren't supported on a GPU. A quick workaround is to restrict operations such that only matrix muls are ran on the GPU.<\/p>\n\n<p>There's an example in the docs: <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/framework.md\">http:\/\/tensorflow.org\/api_docs\/python\/framework.md<\/a><\/p>\n\n<p>See the section on tf.Graph.device(device_name_or_function)<\/p>\n\n<p>I was able to get it working with the following:<\/p>\n\n<pre><code>def device_for_node(n):\n  if n.type == \"MatMul\":\n    return \"\/gpu:0\"\n  else:\n    return \"\/cpu:0\"\n\nwith graph.as_default():\n  with graph.device(device_for_node):\n    ...\n<\/code><\/pre>\n"}]}
{"QuestionId":33625824,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2015-11-10T08:38:45.870","AcceptedAnswerId":33634788.0,"OwnerUserId":4630640.0,"Title":"ImportError: No module named core.framework.graph_pb2","Body":"\n\n<p>When I follow the Installation for linux from <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md\" rel=\"nofollow noreferrer\">tensorflow offical site<\/a>, everything is fine until the last step: <strong>Train your first TensorFlow neural net model<\/strong>. <\/p>\n\n<p>When I execute the command, , An ImportError promps.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>My OS is 14.04.1-Ubuntu.<\/p>\n","answers":[{"AnswerId":"57793949","CreationDate":"2019-09-04T18:25:18.507","ParentId":null,"OwnerUserId":"5692048","Title":null,"Body":"<p>For me, the solution was to run with python3 instead python<\/p>\n"},{"AnswerId":"33634788","CreationDate":"2015-11-10T16:26:24.770","ParentId":null,"OwnerUserId":"2779436","Title":null,"Body":"<p>I have met the exactly same issue.  once your have installed the tensorflow successfully, it's not about the library dependency anymore.<\/p>\n\n<p>if your executed the convolution.py 100% accurately as manual and get the exception like below<\/p>\n\n<pre><code>ImportError: No module named core.framework.graph_b2\n<\/code><\/pre>\n\n<p>this means you are executing the python script exactly under the cloned project root directory,let's say the root named \"src\". <\/p>\n\n<pre><code>src$python tensorflow\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n\n<p>please try to execute the script in the parent directory of the cloned root directory. for example, if your just clone the tensorflow under src dir, goto its parent dir like xxx and do it again.<\/p>\n\n<pre><code>xxx$python src\/tensorflow\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n\n<p>bingo, it works like a charm!<\/p>\n"},{"AnswerId":"33632845","CreationDate":"2015-11-10T14:53:46.697","ParentId":null,"OwnerUserId":"1763903","Title":null,"Body":"<p>I had the same issue and the solution was to <strong>uninstalled protocol buffer v2.6<\/strong>.<\/p>\n\n<p>TensorFlow requires protocol buffer v3.0, which should come with your TensorFlow installation\/source. Hope this helps.<\/p>\n"},{"AnswerId":"33637050","CreationDate":"2015-11-10T18:32:24.843","ParentId":null,"OwnerUserId":"4630640","Title":null,"Body":"<p>I solved it by using <code>virtualenv<\/code>, but there are many other factors which may cause the same error. <\/p>\n\n<p>More solutions are at Github issues <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/81\" rel=\"nofollow\">#81<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/61\" rel=\"nofollow\">#61<\/a>,\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/51\" rel=\"nofollow\">#51<\/a>.<\/p>\n"},{"AnswerId":"33625958","CreationDate":"2015-11-10T08:46:52.907","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Most probably you have either outdated version of:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/pypi.python.org\/pypi\/six\" rel=\"nofollow\">python's six module<\/a>. Install the newest one (<code>pip install six<\/code>) or upgrade your old one (<code>pip install --upgrade six<\/code>)<\/li>\n<li><a href=\"https:\/\/pypi.python.org\/pypi\/protobuf\" rel=\"nofollow\">protobuf<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":33627888,"AnswerCount":3,"Tags":"<python><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2015-11-10T10:29:30.830","AcceptedAnswerId":33628970.0,"OwnerUserId":937153.0,"Title":"Generating LMDB for Caffe","Body":"<p>I am trying to build a deep learning model for Saliency analysis using caffe (I am using the python wrapper). But I am unable to understand how to generate the lmdb data structure for this purpose. I have gone through the Imagenet and mnist examples and I understand that I should generate labels in the format <\/p>\n\n<pre><\/pre>\n\n<p>But in my case, I will be labeling each pixel with 0 or 1 indicating whether that pixel is salient or not. That won't be a single label for an image.  <\/p>\n\n<p>How to generate lmdb files for a per pixel based labeling ?<\/p>\n","answers":[{"AnswerId":"39272038","CreationDate":"2016-09-01T12:50:09.700","ParentId":null,"OwnerUserId":"5979961","Title":null,"Body":"<p>In caffe both lmdb and hdf5 supports multiple labels per image, matrices if you like, see this thread:<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issue-53768814\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issue-53768814<\/a><\/p>\n\n<p>See this tutorial on how to create a multi-label dataset (lmdb here) for caffe with python code:<\/p>\n\n<p><a href=\"http:\/\/www.kostyaev.me\/article\/Multilabel%20Dataset\/\" rel=\"nofollow\">http:\/\/www.kostyaev.me\/article\/Multilabel%20Dataset\/<\/a><\/p>\n\n<p>EDIT: For example for the labels it uses the caffe-python function which converts a 3-dimensional array to datum, found in caffe\/python\/caffe.io.py:\narray_to_datum(arr, label=None):<\/p>\n"},{"AnswerId":"33628970","CreationDate":"2015-11-10T11:29:53.983","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can approach this problem in two ways:<\/p>\n\n<p><strong>1.<\/strong> Using HDF5 data layer instead of LMDB. HDF5 is more flexible and can support labels the size of the image. You can see <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">this answer<\/a> for an example of constructing and using HDF5 input data layer.<\/p>\n\n<p><strong>2.<\/strong> You can have two LMDB input layers: one for the image and one for the label. Note that when you build the LMDB you must <strong>not<\/strong> use the <code>'shuffle'<\/code> option in order to have the images and their labels in sync.<\/p>\n\n<p><strong>Update:<\/strong> I recently gave a more detailed answer <a href=\"https:\/\/stackoverflow.com\/a\/45181897\/1714410\">here<\/a>.<\/p>\n"},{"AnswerId":"34038314","CreationDate":"2015-12-02T08:47:21.277","ParentId":null,"OwnerUserId":"5056575","Title":null,"Body":"<p>Check this one: <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow\">http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/<\/a><\/p>\n\n<p>Just load all images in <code>X<\/code> and corresponding labels in <code>Y<\/code>.<\/p>\n"}]}
{"QuestionId":33630840,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2015-11-10T13:12:54.487","AcceptedAnswerId":null,"OwnerUserId":5546694.0,"Title":"Convolutional neural networks and 3D images","Body":"<p>I am interested in applying CNNs to 3D images (i.e. medical data). \nDoes TensorFlow already incorporate this functionality?<\/p>\n","answers":[{"AnswerId":"45119786","CreationDate":"2017-07-15T15:21:45.183","ParentId":null,"OwnerUserId":"6440983","Title":null,"Body":"<p>The TensorFlow implementation for 3D Convolutional Neural Networks has been provided with the following open source projects:<\/p>\n\n<p><a href=\"https:\/\/github.com\/astorfi\/3D-convolutional-Audio-Visual\" rel=\"nofollow noreferrer\">Lip Reading - Cross Audio-Visual Recognition using 3D Convolutional Neural Networks<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/astorfi\/3D-convolutional-speaker-recognition\" rel=\"nofollow noreferrer\">Using 3D Convolutional Neural Networks for Speaker Verification<\/a><\/p>\n"},{"AnswerId":"33711518","CreationDate":"2015-11-14T17:57:48.930","ParentId":null,"OwnerUserId":"1468566","Title":null,"Body":"<p>No, the current implementations are made for 2D images (functions like nn.conv2d). They support multiple channels (eg RGB) and it is possible to express 3D images as a multichannel 2D image (each z-slice is a channel), but this isn't always ideal. Additionally for using these sort of approaches you need substantial amounts of image data which is typically difficult to come by in the medical area.<\/p>\n\n<p>Update: both TensorFlow and Theano (subsequently Keras, Lasagne, etc) now all support 3D operations as stated above. It is important to note that 3D operations are much more computationally and memory intensive than a similar 2D operation.<\/p>\n"},{"AnswerId":"35661317","CreationDate":"2016-02-26T20:40:32.787","ParentId":null,"OwnerUserId":"1599927","Title":null,"Body":"<p>If you want to use CNN with 3D images, a possible alternative is to use this <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2049\" rel=\"nofollow\">Caffe PR<\/a>.\nYou will need to convert your data to HDF5 format.<\/p>\n"},{"AnswerId":"37635583","CreationDate":"2016-06-04T22:21:08.337","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>TensorFlow now supports <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv3d\" rel=\"nofollow noreferrer\">3D convolution<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/max_pooling3d\" rel=\"nofollow noreferrer\">3D pooling<\/a> in the master branch.<\/p>\n\n<p>You can use them with 5D tensors as input with shape: <code>[batch_size, depth, height, width, channels]<\/code>.<\/p>\n"}]}
{"QuestionId":33633227,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-11-10T15:12:48.543","AcceptedAnswerId":null,"OwnerUserId":4248614.0,"Title":"Getting value of a tensor variable in Theano","Body":"<p>I am a noob at Theano - so please be patient.<\/p>\n\n<p>I'm trying to implement gradient clipping for a project. A code snippet is as below :<\/p>\n\n<pre><\/pre>\n\n<p>The error I receive is as follows : <\/p>\n\n<p>\nPlease advice.<\/p>\n","answers":[{"AnswerId":"33647769","CreationDate":"2015-11-11T09:37:33.253","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you're calling <code>eval<\/code> without providing the input values.<\/p>\n\n<p>Your code is incomplete so I'm guessing a bit but <code>cost<\/code> is probably a function of at least one input tensor (e.g. something like <code>x = T.matrix()<\/code>). That being the case, a value needs to be provided for <code>x<\/code> when the computation is executed. You execute the computation by calling <code>eval<\/code> and can provide a value for <code>x<\/code> like this:<\/p>\n\n<pre><code>print grad_norm.eval({x: numpy.random.randn(10, 20).astype(theano.config.floatX})\n<\/code><\/pre>\n\n<p>You need to make sure you provide a value for every symbolic input that, in this case, <code>grad_norm<\/code> is a function of (just add additional entries to the dictionary passed to <code>eval<\/code>). You also need to make sure the shapes of the values passed in the dictionary are appropriate for the computation (e.g. if there's a dot product with a shared variable, the multiplied axes must be of the same size).<\/p>\n\n<p>Also, be aware, the <code>eval<\/code> method is intended primarily for debug purposes. You shouldn't use this feature in 'real' code. Instead, you should compile the computation using <code>theano.function<\/code>.<\/p>\n"}]}
{"QuestionId":33633370,"AnswerCount":21,"Tags":"<python><tensorflow><tensor>","CreationDate":"2015-11-10T15:19:58.187","AcceptedAnswerId":33633839.0,"OwnerUserId":4993513.0,"Title":"How to print the value of a Tensor object in TensorFlow?","Body":"<p>I have been using the introductory example of matrix multiplication in TensorFlow.<\/p>\n\n<pre><\/pre>\n\n<p>When I print the product, it is displaying it as a  object:<\/p>\n\n<pre><\/pre>\n\n<p>But how do I know the value of ?<\/p>\n\n<p>The following doesn't help:<\/p>\n\n<pre><\/pre>\n\n<p>I know that graphs run on , but isn't there any way I can check the output of a  object without running the graph in a ?<\/p>\n","answers":[{"AnswerId":"55736854","CreationDate":"2019-04-17T22:35:10.353","ParentId":null,"OwnerUserId":"9936228","Title":null,"Body":"<p>In the recent Tensorflow 1.13.1<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n\n#print the product\nprint(product)         # tf.Tensor([[12.]], shape=(1, 1), dtype=float32)\nprint(product.numpy()) # [[12.]]\n<\/code><\/pre>\n\n<p>With Tensorflow 2.0, eager mode is enabled by default. so the following code works with TF2.0.<\/p>\n\n<pre><code>import tensorflow as tf\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n\n#print the product\nprint(product)         # tf.Tensor([[12.]], shape=(1, 1), dtype=float32)\nprint(product.numpy()) # [[12.]]\n<\/code><\/pre>\n"},{"AnswerId":"57433811","CreationDate":"2019-08-09T16:12:23.053","ParentId":null,"OwnerUserId":"1551308","Title":null,"Body":"<p>Using tips provided in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/print\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/print<\/a> I use the <code>log_d<\/code> function to print formatted strings.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef log_d(fmt, *args):\n    op = tf.py_func(func=lambda fmt_, *args_: print(fmt%(*args_,)),\n                    inp=[fmt]+[*args], Tout=[])\n    return tf.control_dependencies([op])\n\n\n# actual code starts now...\n\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n\nwith log_d('MAT1: %s, MAT2: %s', matrix1, matrix2): # this will print the log line\n    product = tf.matmul(matrix1, matrix2)\n\nwith tf.Session() as sess:\n    sess.run(product)\n<\/code><\/pre>\n"},{"AnswerId":"56796660","CreationDate":"2019-06-27T18:13:57.573","ParentId":null,"OwnerUserId":"4909087","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/eval\" rel=\"nofollow noreferrer\"><strong><code>tf.keras.backend.eval<\/code><\/strong><\/a> is useful for evaluating small expressions.<\/p>\n\n<pre><code>tf.keras.backend.eval(op)\n<\/code><\/pre>\n\n<p>TF 1.x and TF 2.0 compatible.<\/p>\n\n<hr>\n\n<p><strong>Minimal Verifiable Example<\/strong><\/p>\n\n<pre><code>from tensorflow.keras.backend import eval\n\nm1 = tf.constant([[3., 3.]])\nm2 = tf.constant([[2.],[2.]])\n\neval(tf.matmul(m1, m2))\n# array([[12.]], dtype=float32)\n<\/code><\/pre>\n\n<p>This is useful because you do not have to explicitly create a <code>Session<\/code> or <code>InteractiveSession<\/code>. <\/p>\n"},{"AnswerId":"33633839","CreationDate":"2015-11-10T15:41:24.600","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest<sup>[A]<\/sup> way to evaluate the actual value of a <code>Tensor<\/code> object is to pass it to the <code>Session.run()<\/code> method, or call <code>Tensor.eval()<\/code> when you have a default session (i.e. in a <code>with tf.Session():<\/code> block, or see below). In general<sup>[B]<\/sup>, you cannot print the value of a tensor without running some code in a session.<\/p>\n\n<p>If you are experimenting with the programming model, and want an easy way to evaluate tensors, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/InteractiveSession\" rel=\"nofollow noreferrer\"><code>tf.InteractiveSession<\/code><\/a> lets you open a session at the start of your program, and then use that session for all <code>Tensor.eval()<\/code> (and <code>Operation.run()<\/code>) calls. This can be easier in an interactive setting, such as the shell or an IPython notebook, when it's tedious to pass around a <code>Session<\/code> object everywhere. For example, the following works in a Jupyter notebook:<\/p>\n\n<pre><code>with tf.Session() as sess:  print(product.eval()) \n<\/code><\/pre>\n\n<p>This might seem silly for such a small expression, but one of the key ideas in Tensorflow is <em>deferred execution<\/em>: it's very cheap to build a large and complex expression, and when you want to evaluate it, the back-end (to which you connect with a <code>Session<\/code>) is able to schedule its execution more efficiently (e.g. executing independent parts in parallel and using GPUs).<\/p>\n\n<hr>\n\n<p>[A]: To print the value of a tensor without returning it to your Python program, you can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\"><code>tf.Print()<\/code><\/a> operator, as <a href=\"https:\/\/stackoverflow.com\/a\/36296783\/3574081\">Andrzej suggests in another answer<\/a>. Note that you still need to run part of the graph to see the output of this op, which is printed to standard output. If you're running distributed TensorFlow, <code>tf.Print()<\/code> will print its output to the standard output of the task where that op runs. This means that if you use <a href=\"https:\/\/colab.research.google.com\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com<\/a> for example, or any other Jupyter Notebook, then you will <em>not see<\/em> the output of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\"><code>tf.Print()<\/code><\/a> in the notebook; in that case refer to <a href=\"https:\/\/stackoverflow.com\/questions\/49969193\/no-result-of-tf-print-in-kerass-model-fit\/49969365#49969365\">this answer<\/a> on how to get it to print still.<\/p>\n\n<p>[B]: You <em>might<\/em> be able to use the experimental <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/util\/constant_value\" rel=\"nofollow noreferrer\"><code>tf.contrib.util.constant_value()<\/code><\/a> function to get the value of a constant tensor, but it isn't intended for general use, and it isn't defined for many operators.<\/p>\n"},{"AnswerId":"33640758","CreationDate":"2015-11-10T22:20:06.370","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>No, you can not see the content of the tensor without running the graph (doing <code>session.run()<\/code>). The only things you can see are:<\/p>\n\n<ul>\n<li>the dimensionality of the tensor (but I assume it is not hard to calculate it for the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/math_ops.md#contents\">list of the operations<\/a> that TF has)<\/li>\n<li>type of the operation that will be used to generate the tensor (<code>transpose_1:0<\/code>, <code>random_uniform:0<\/code>)<\/li>\n<li>type of elements in the tensor (<code>float32<\/code>)<\/li>\n<\/ul>\n\n<p>I have not found this in documentation, but I believe that the values of the variables (and some of the constants are not calculated at the time of assignment).<\/p>\n\n<hr>\n\n<p>Take a look at this example:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom datetime import datetime\ndim = 7000\n<\/code><\/pre>\n\n<p>The first example where I just initiate a constant Tensor of random numbers run approximately the same time irrespectibly of dim (<code>0:00:00.003261<\/code>)<\/p>\n\n<pre><code>startTime = datetime.now()\nm1 = tf.truncated_normal([dim, dim], mean=0.0, stddev=0.02, dtype=tf.float32, seed=1)\nprint datetime.now() - startTime\n<\/code><\/pre>\n\n<p>In the second case, where the constant is actually gets evaluated and the values are assigned, the time clearly depends on dim (<code>0:00:01.244642<\/code>)<\/p>\n\n<pre><code>startTime = datetime.now()\nm1 = tf.truncated_normal([dim, dim], mean=0.0, stddev=0.02, dtype=tf.float32, seed=1)\nsess = tf.Session()\nsess.run(m1)\nprint datetime.now() - startTime\n<\/code><\/pre>\n\n<p>And you can make it more clear by calculating something (<code>d = tf.matrix_determinant(m1)<\/code>, keeping in mind that the time will run in <code>O(dim^2.8)<\/code>)<\/p>\n\n<p>P.S. I found were it is explained in <a href=\"http:\/\/tensorflow.org\/resources\/faq.md#contents\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>A Tensor object is a symbolic handle to the result of an operation,\n  but does not actually hold the values of the operation's output.<\/p>\n<\/blockquote>\n"},{"AnswerId":"55735880","CreationDate":"2019-04-17T21:00:03.087","ParentId":null,"OwnerUserId":"3320697","Title":null,"Body":"<p>In Tensorflow 2.0+ (or in Eager mode environment) you can call <code>.numpy()<\/code> method:<\/p>\n\n<pre><code>import tensorflow as tf\n\nmatrix1 = tf.constant([[3., 3.0]])\nmatrix2 = tf.constant([[2.0],[2.0]])\nproduct = tf.matmul(matrix1, matrix2)\n\nprint(product.numpy()) \n<\/code><\/pre>\n"},{"AnswerId":"55733984","CreationDate":"2019-04-17T18:34:20.653","ParentId":null,"OwnerUserId":"4015879","Title":null,"Body":"<p>Question: How to print the value of a Tensor object in TensorFlow?<\/p>\n\n<p>Answer:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Variable\nx = tf.Variable([[1,2,3]])\n\n# initialize\ninit = (tf.global_variables_initializer(), tf.local_variables_initializer())\n\n# Create a session\nsess = tf.Session()\n\n# run the session\nsess.run(init)\n\n# print the value\nsess.run(x)\n<\/code><\/pre>\n"},{"AnswerId":"56066394","CreationDate":"2019-05-09T19:33:08.427","ParentId":null,"OwnerUserId":"10997543","Title":null,"Body":"<p><strong>Enable the eager execution which is introduced in tensorflow after version 1.10.\nIt's very easy to use.<\/strong><\/p>\n\n<pre><code># Initialize session\nimport tensorflow as tf\ntf.enable_eager_execution()\n\n\n# Some tensor we want to print the value of\na = tf.constant([1.0, 3.0])\n\nprint(a)\n<\/code><\/pre>\n"},{"AnswerId":"51629420","CreationDate":"2018-08-01T09:09:48.923","ParentId":null,"OwnerUserId":"5258060","Title":null,"Body":"<p>Basically, in tensorflow when you create a tensor of any sort they are created and stored inside which can be accessed only when you run a tensorflow session. Say you have created a constant tensor <br\/>\n<code>c = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])<\/code><br\/>\nWithout running a session, you can get<br\/>\n- <code>op<\/code>: An Operation. Operation that computes this tensor.<br\/>\n- <code>value_index<\/code>: An int. Index of the operation's endpoint that produces this tensor.<br\/>\n- <code>dtype<\/code>: A DType. Type of elements stored in this tensor.<br\/><\/p>\n\n<p>To get the values you can run a session with the tensor you require as:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    print(sess.run(c))\n    sess.close()\n<\/code><\/pre>\n\n<p>The output will be something like this:<\/p>\n\n<blockquote>\n  <p>array([[1., 2., 3.],\n         [4., 5., 6.]], dtype=float32) <br\/><\/p>\n<\/blockquote>\n"},{"AnswerId":"42188604","CreationDate":"2017-02-12T14:11:34.543","ParentId":null,"OwnerUserId":"1540476","Title":null,"Body":"<p>Try this simple code!  (it is self explanatory) <\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession() # see the answers above :)\nx = [[1.,2.,1.],[1.,1.,1.]]    # a 2D matrix as input to softmax\ny = tf.nn.softmax(x)           # this is the softmax function\n                               # you can have anything you like here\nu = y.eval()\nprint(u)\n<\/code><\/pre>\n"},{"AnswerId":"37543184","CreationDate":"2016-05-31T10:28:56.297","ParentId":null,"OwnerUserId":"214647","Title":null,"Body":"<p>Reiterating what others said, its not possible to check the values without running the graph.<\/p>\n\n<p>A simple snippet for anyone looking for an easy example to print values is as below. The code can be executed without any modification in ipython notebook<\/p>\n\n<pre><code>import tensorflow as tf\n\n#define a variable to hold normal random values \nnormal_rv = tf.Variable( tf.truncated_normal([2,3],stddev = 0.1))\n\n#initialize the variable\ninit_op = tf.initialize_all_variables()\n\n#run the graph\nwith tf.Session() as sess:\n    sess.run(init_op) #execute init_op\n    #print the random values that we sample\n    print (sess.run(normal_rv))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[-0.16702934  0.07173464 -0.04512421]\n [-0.02265321  0.06509651 -0.01419079]]\n<\/code><\/pre>\n"},{"AnswerId":"57233365","CreationDate":"2019-07-27T15:04:31.997","ParentId":null,"OwnerUserId":"2366125","Title":null,"Body":"<p>You can use Keras, one-line answer will be to use <code>eval<\/code> method like so:<\/p>\n\n<pre><code>import keras.backend as K\nprint(K.eval(your_tensor))\n<\/code><\/pre>\n"},{"AnswerId":"42380807","CreationDate":"2017-02-22T00:54:45.957","ParentId":null,"OwnerUserId":"7601910","Title":null,"Body":"<p>I think you need to get some fundamentals right. With the examples above you have created tensors (multi dimensional array). But for tensor flow to really work you have to initiate a \"<strong>session<\/strong>\" and run your \"<strong>operation<\/strong>\" in the session. Notice the word \"session\" and \"operation\".\nYou need to know 4 things to work with  tensorflow: <\/p>\n\n<ol>\n<li>tensors<\/li>\n<li>Operations<\/li>\n<li>Sessions<\/li>\n<li>Graphs<\/li>\n<\/ol>\n\n<p>Now from what you wrote out you have given the tensor, and the operation but you have no session running nor a graph. Tensor (edges of the graph) flow through graphs and are manipulated by operations (nodes of the graph). There is default graph but you can initiate yours in a session.<\/p>\n\n<p>When you say print , you only access the shape of the variable or constant you defined.<\/p>\n\n<p>So you can see what you are missing :<\/p>\n\n<pre><code> with tf.Session() as sess:     \n           print(sess.run(product))\n           print (product.eval())\n<\/code><\/pre>\n\n<p>Hope it helps!<\/p>\n"},{"AnswerId":"49664292","CreationDate":"2018-04-05T04:41:37.773","ParentId":null,"OwnerUserId":"2061991","Title":null,"Body":"<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\nx = [[1.,2.,1.],[1.,1.,1.]]    \ny = tf.nn.softmax(x)           \n\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n\nprint(product.eval())\ntf.reset_default_graph()\nsess.close()\n<\/code><\/pre>\n"},{"AnswerId":"41580745","CreationDate":"2017-01-11T00:18:38.700","ParentId":null,"OwnerUserId":"5725619","Title":null,"Body":"<p>Based on the answers above, with your particular code snippet you can print the product like this:<\/p>\n\n<pre><code>import tensorflow as tf\n#Initialize the session\nsess = tf.InteractiveSession()\n\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n\n#print the product\nprint(product.eval())\n\n#close the session to release resources\nsess.close()\n<\/code><\/pre>\n"},{"AnswerId":"49900935","CreationDate":"2018-04-18T13:25:24.623","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>I didn't find it easy to understand what is required even after reading all the answers until I executed this. TensofFlow is new to me too.<\/p>\n\n<pre><code>def printtest():\nx = tf.constant([1.0, 3.0])\nx = tf.Print(x,[x],message=\"Test\")\ninit = (tf.global_variables_initializer(), tf.local_variables_initializer())\nb = tf.add(x, x)\nwith tf.Session() as sess:\n    sess.run(init)\n    print(sess.run(b))\n    sess.close()\n<\/code><\/pre>\n\n<p>But still you may need the value returned by executing the session.<\/p>\n\n<pre><code>def printtest():\n    x = tf.constant([100.0])\n    x = tf.Print(x,[x],message=\"Test\")\n    init = (tf.global_variables_initializer(), tf.local_variables_initializer())\n    b = tf.add(x, x)\n    with tf.Session() as sess:\n        sess.run(init)\n        c = sess.run(b)\n        print(c)\n        sess.close()\n<\/code><\/pre>\n"},{"AnswerId":"36296783","CreationDate":"2016-03-29T23:13:22.653","ParentId":null,"OwnerUserId":"1576602","Title":null,"Body":"<p>While other answers are correct that you cannot print the value until you evaluate the graph, they do not talk about one easy way of actually printing a value inside the graph, once you evaluate it. <\/p>\n\n<p>The easiest way to see a value of a tensor whenever the graph is evaluated (using <code>run<\/code> or <code>eval<\/code>) is to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#Print\" rel=\"noreferrer\"><code>Print<\/code><\/a> operation as in this example:<\/p>\n\n<pre><code># Initialize session\nimport tensorflow as tf\nsess = tf.InteractiveSession()\n\n# Some tensor we want to print the value of\na = tf.constant([1.0, 3.0])\n\n# Add print operation\na = tf.Print(a, [a], message=\"This is a: \")\n\n# Add more elements of the graph using a\nb = tf.add(a, a)\n<\/code><\/pre>\n\n<p>Now, whenever we evaluate the whole graph, e.g. using <code>b.eval()<\/code>, we get:<\/p>\n\n<pre><code>I tensorflow\/core\/kernels\/logging_ops.cc:79] This is a: [1 3]\n<\/code><\/pre>\n"},{"AnswerId":"58859509","CreationDate":"2019-11-14T14:47:09.163","ParentId":null,"OwnerUserId":"835112","Title":null,"Body":"<p>tf.Print is now deprecated, here's how to use tf.print (lowercase p) instead.<\/p>\n\n<p>While running a session is a good option, it is not always the way to go. For instance, you may want to print some tensor in a particular session.<\/p>\n\n<p>The new print method returns a print operation which has no output tensors:<\/p>\n\n<pre><code>print_op = tf.print(tensor_to_print)\n<\/code><\/pre>\n\n<p>Since it has no outputs, you can't insert it in a graph the same way as you could with tf.Print. Instead, you can you can add it to control dependencies in your session in order to make it print.<\/p>\n\n<pre><code>sess = tf.compat.v1.Session()\nwith sess.as_default():\n  tensor_to_print = tf.range(10)\n  print_op = tf.print(tensor_to_print)\nwith tf.control_dependencies([print_op]):\n  tripled_tensor = tensor_to_print * 3\nsess.run(tripled_tensor)\n<\/code><\/pre>\n\n<p>Sometimes, in a larger graph, maybe created partly in subfunctions, it is cumbersome to propagate the print_op to the session call. Then, tf.tuple can be used to couple the print operation with another operation, which will then run with that operation whichever session executes the code. Here's how that is done: <\/p>\n\n<pre><code>print_op = tf.print(tensor_to_print)\nsome_tensor_list = tf.tuple([some_tensor], control_inputs=[print_op])\n# Use some_tensor_list[0] instead of any_tensor below.\n<\/code><\/pre>\n"},{"AnswerId":"44741406","CreationDate":"2017-06-24T22:12:28.033","ParentId":null,"OwnerUserId":"3532255","Title":null,"Body":"<p>Please note that <code>tf.Print()<\/code> will change the tensor name.\nIf the tensor you seek to print is a placeholder, feeding data to it will fail as the original name will not be found during feeding.\nFor example:<\/p>\n\n<pre><code>import tensorflow as tf\ntens = tf.placeholder(tf.float32,[None,2],name=\"placeholder\")\nprint(eval(\"tens\"))\ntens = tf.Print(tens,[tens, tf.shape(tens)],summarize=10,message=\"tens:\")\nprint(eval(\"tens\"))\nres = tens + tens\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nprint(sess.run(res))\n<\/code><\/pre>\n\n<p>Output is:<\/p>\n\n<pre><code>python test.py\nTensor(\"placeholder:0\", shape=(?, 2), dtype=float32)\nTensor(\"Print:0\", shape=(?, 2), dtype=float32)\nTraceback (most recent call last):\n[...]\nInvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'placeholder' with dtype float\n<\/code><\/pre>\n"},{"AnswerId":"44860574","CreationDate":"2017-07-01T11:49:59.460","ParentId":null,"OwnerUserId":"4500633","Title":null,"Body":"<p>You should think of TensorFlow Core programs as consisting of two discrete sections:<\/p>\n\n<ul>\n<li>Building the computational graph.<\/li>\n<li>Running the computational graph.<\/li>\n<\/ul>\n\n<p>So for the code below you just Build the computational graph.  <\/p>\n\n<pre><code>matrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nproduct = tf.matmul(matrix1, matrix2)\n<\/code><\/pre>\n\n<p>You need also To initialize all the variables in a TensorFlow program , you must explicitly call a special operation as follows:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>Now you build the graph and initialized all variables ,next step is to evaluate the nodes, you must run the computational graph within a session. A session encapsulates the control and state of the TensorFlow runtime.<\/p>\n\n<p>The following code creates a Session object and then invokes its run method to run enough of the computational graph to evaluate <code>product<\/code> :<\/p>\n\n<pre><code>sess = tf.Session()\n\/\/ run variables initializer\nsess.run(init)\n\nprint(sess.run([product]))\n<\/code><\/pre>\n"},{"AnswerId":"47058715","CreationDate":"2017-11-01T15:49:25.150","ParentId":null,"OwnerUserId":"5615276","Title":null,"Body":"<p>You can check the output of a TensorObject without running the graph in a session, by enabling <a href=\"https:\/\/research.googleblog.com\/2017\/10\/eager-execution-imperative-define-by.html\" rel=\"noreferrer\">eager execution<\/a>.<\/p>\n\n<p>Simply add the following two lines of code:\n<code>\nimport tensorflow.contrib.eager as tfe\ntfe.enable_eager_execution()\n<\/code><\/p>\n\n<p>right after you <code>import tensorflow<\/code>.<\/p>\n\n<p>The output of <code>print product<\/code> in your example will now be:\n<code>\ntf.Tensor([[ 12.]], shape=(1, 1), dtype=float32)\n<\/code><\/p>\n\n<p>Note that as of now (November 2017) you'll have to install a Tensorflow nightly build to enable eager execution. Pre-built wheels can be found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#installation\" rel=\"noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":33634008,"AnswerCount":12,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-10T15:48:57.080","AcceptedAnswerId":33634101.0,"OwnerUserId":363455.0,"Title":"How do I install TensorFlow's tensorboard?","Body":"<p>How do I install TensorFlow's tensorboard?<\/p>\n","answers":[{"AnswerId":"33634101","CreationDate":"2015-11-10T15:53:33.207","ParentId":null,"OwnerUserId":"1574374","Title":null,"Body":"<p>The steps to install Tensorflow are here: <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/<\/a><\/p>\n\n<p>For example, on Linux for CPU-only (no GPU), you would type this command:<\/p>\n\n<pre><code>pip install -U pip\npip install tensorflow\n<\/code><\/pre>\n\n<p>Since <a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow\" rel=\"noreferrer\">TensorFlow<\/a> depends on <a href=\"https:\/\/pypi.python.org\/pypi\/tensorboard\" rel=\"noreferrer\">TensorBoard<\/a>, running the following command should <strong>not<\/strong> be necessary:<\/p>\n\n<pre><code>pip install tensorboard\n<\/code><\/pre>\n"},{"AnswerId":"33641996","CreationDate":"2015-11-11T00:04:10.833","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>Try typing <code>which tensorboard<\/code> in your terminal. It should exist if you <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#pip_installation\" rel=\"noreferrer\">installed with <code>pip<\/code><\/a> as mentioned in the <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/README.md\" rel=\"noreferrer\">tensorboard README<\/a> (although the documentation doesn't tell you that you can now launch tensorboard without doing anything else).<\/p>\n\n<p>You need to give it a log directory. If you are in the directory where you saved your graph, you can launch it from your terminal with something like:<\/p>\n\n<pre><code>tensorboard --logdir .\n<\/code><\/pre>\n\n<p>or more generally:<\/p>\n\n<pre><code>tensorboard --logdir \/path\/to\/log\/directory\n<\/code><\/pre>\n\n<p>for any log directory.<\/p>\n\n<p>Then open your favorite web browser and type in <code>localhost:6006<\/code> to connect.<\/p>\n\n<p>That should get you started. As for logging anything useful in your training process, you need to use the <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/summary\" rel=\"noreferrer\">TensorFlow Summary API<\/a>. You can also use the <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"noreferrer\"><code>TensorBoard<\/code> callback in Keras<\/a>.<\/p>\n"},{"AnswerId":"33679854","CreationDate":"2015-11-12T19:38:04.747","ParentId":null,"OwnerUserId":"861809","Title":null,"Body":"<p>If your Tensorflow install is located here:<\/p>\n\n<pre><code>\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\n<\/code><\/pre>\n\n<p>then the python command to launch Tensorboard is:<\/p>\n\n<pre><code>$ python \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/tensorboard\/tensorboard.py --logdir=\/home\/user\/Documents\/...\/logdir\n<\/code><\/pre>\n\n<p>The installation from pip allows you to use:<\/p>\n\n<pre><code>$ tensorboard --logdir=\/home\/user\/Documents\/...\/logdir\n<\/code><\/pre>\n"},{"AnswerId":"51291107","CreationDate":"2018-07-11T17:19:56.380","ParentId":null,"OwnerUserId":"4939139","Title":null,"Body":"<p>It is better not to mix up the virtual environments or perform installation on the root directory. Steps I took for hassle free installation are as below. I used conda for installing all my dependencies instead of pip. I'm answering with extra details, because when I tried to install tensor board and tensor flow on my root env, it messed up. <\/p>\n\n<ul>\n<li><p>Create a virtual env <\/p>\n\n<p><code>conda create --name my_env python=3.6<\/code><\/p><\/li>\n<li><p>Activate virtual environment<\/p>\n\n<p><code>source activate my_env<\/code><\/p><\/li>\n<li><p>Install basic required modules<\/p>\n\n<p><code>conda install pandas<\/code><\/p>\n\n<p><code>conda install tensorflow<\/code><\/p><\/li>\n<li><p>Install tensor board <\/p>\n\n<p><code>conda install -c condo-forge tensor board<\/code><\/p><\/li>\n<\/ul>\n\n<p>Hope that helps<\/p>\n"},{"AnswerId":"51145326","CreationDate":"2018-07-03T00:56:57.867","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>If you're using the anaconda distribution of Python, then simply do:<\/p>\n\n<pre><code> $\u276f conda install -c conda-forge tensorboard \n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code> $\u276f conda install -c anaconda tensorboard \n<\/code><\/pre>\n\n<hr>\n\n<p>Also, you can have a look at various builds by search the packages repo by:<\/p>\n\n<pre><code>$\u276f anaconda search -t conda tensorboard\n<\/code><\/pre>\n\n<p>which would list the channels and the corresponding builds, the supported OS, Python versions etc.,<\/p>\n"},{"AnswerId":"33642184","CreationDate":"2015-11-11T00:21:56.743","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>TensorBoard isn't a separate component. TensorBoard comes packaged with TensorFlow.<\/p>\n"},{"AnswerId":"49708396","CreationDate":"2018-04-07T14:28:07.553","ParentId":null,"OwnerUserId":"860421","Title":null,"Body":"<p>Adding this just for the sake of completeness of this question (some questions may get closed as duplicate of this one).<\/p>\n\n<p>I usually use user mode for pip ie. <code>pip install --user<\/code> even if instructions assume root mode. That way, my tensorboard installation was in <code>~\/.local\/bin\/tensorboard<\/code>, and it was not in my path (which shouldn't be ideal either). So I was not able to access it.<\/p>\n\n<p>In this case, running<\/p>\n\n<pre><code>sudo ln -s ~\/.local\/bin\/tensorboard \/usr\/bin\n<\/code><\/pre>\n\n<p>should fix it.<\/p>\n"},{"AnswerId":"50032414","CreationDate":"2018-04-25T22:50:53.307","ParentId":null,"OwnerUserId":"9442951","Title":null,"Body":"<p>It may be helpful to make an alias for it.<\/p>\n\n<p>Install and find your tensorboard location:<\/p>\n\n<pre><code>pip install tensorboard\npip show tensorboard\n<\/code><\/pre>\n\n<p>Add the following alias in .bashrc:<\/p>\n\n<pre><code>alias tensorboard='python pathShownByPip\/tensorboard\/main.py'\n<\/code><\/pre>\n\n<p>Open another terminal or run exec bash.<\/p>\n\n<p>For Windows users, cd into <code>pathShownByPip\\tensorboard<\/code> and run <code>python main.py<\/code> from there.<\/p>\n\n<p>For Python 3.x, use <code>pip3<\/code> instead of <code>pip<\/code>, and don't forget to use <code>python3<\/code> in the alias.<\/p>\n"},{"AnswerId":"39481495","CreationDate":"2016-09-14T02:10:20.303","ParentId":null,"OwnerUserId":"6783862","Title":null,"Body":"<p>you may have installed tensorflow to virtualenv.\nactivate it and tensorboard command will become available.<\/p>\n"},{"AnswerId":"47943767","CreationDate":"2017-12-22T15:22:38.617","ParentId":null,"OwnerUserId":"2218905","Title":null,"Body":"<pre><code>pip install tensorflow.tensorboard  # install tensorboard\npip show tensorflow.tensorboard\n# Location: c:\\users\\&lt;name&gt;\\appdata\\roaming\\python\\python35\\site-packages\n# now just run tensorboard as:\npython c:\\users\\&lt;name&gt;\\appdata\\roaming\\python\\python35\\site-packages\\tensorboard\\main.py --logdir=&lt;logidr&gt;\n<\/code><\/pre>\n"},{"AnswerId":"44809595","CreationDate":"2017-06-28T18:01:42.023","ParentId":null,"OwnerUserId":"3749292","Title":null,"Body":"<p>The <code>pip<\/code> package you are looking for is <code>tensorflow-tensorboard<\/code> developed by Google.<\/p>\n"},{"AnswerId":"38396754","CreationDate":"2016-07-15T12:58:17.373","ParentId":null,"OwnerUserId":"5624248","Title":null,"Body":"<p>If you installed TensorFlow using <code>pip<\/code>, then the location of TensorBoard can be retrieved by issuing the command <code>which tensorboard<\/code> on the terminal. You can then edit the TensorBoard file, if necessary.<\/p>\n"}]}
{"QuestionId":33634126,"AnswerCount":1,"Tags":"<python><word2vec><tensorflow>","CreationDate":"2015-11-10T15:54:50.330","AcceptedAnswerId":33662406.0,"OwnerUserId":5231961.0,"Title":"Tensorflow - no module named 'embeddings' in tensorflow.models.embeddings","Body":"<p>I am trying to run the vector representations of words tutorial using Tensorflow found here:<\/p>\n\n<p><a href=\"http:\/\/www.tensorflow.org\/tutorials\/word2vec\/index.md\" rel=\"nofollow\">http:\/\/www.tensorflow.org\/tutorials\/word2vec\/index.md<\/a><\/p>\n\n<p>The first script, word2vec_basic.py, runs fine, but the second one (found here)<\/p>\n\n<p><a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/embedding\/word2vec.py\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/embedding\/word2vec.py<\/a><\/p>\n\n<p>gives me the error \"ImportError: No module named embedding\" at line 28. I installed tensorflow as per the instructions on the page, any reason why I wouldn't have gotten the embeddings module, containing word2vec?<\/p>\n","answers":[{"AnswerId":"33662406","CreationDate":"2015-11-12T00:27:04.707","ParentId":null,"OwnerUserId":"94688","Title":null,"Body":"<p>For now, the more full-featured versions of word2vec such as word2vec.py require building from source.<\/p>\n"}]}
{"QuestionId":33634525,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-10T16:14:05.877","AcceptedAnswerId":33635450.0,"OwnerUserId":5547331.0,"Title":"TensorFlow on 32-bit Linux?","Body":"<p>Is there a version of TensorFlow for 32-bit Linux?  I only see the 64-bit wheel available, and didn't find anything about it on the site.<\/p>\n","answers":[{"AnswerId":"41561561","CreationDate":"2017-01-10T05:13:29.583","ParentId":null,"OwnerUserId":"1143274","Title":null,"Body":"<p>I've built a CPU-only version of TensorFlow on 32-bit Ubuntu (16.04.1 Xubuntu). It went a lot more smoothly than anticipated, for such a complex library that doesn't support 32-bit architectures officially.<\/p>\n\n<p>It can be done by following a subset of the intersection of these two guides:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/cudamusing.blogspot.com.au\/2015\/11\/building-tensorflow-for-jetson-tk1.html\" rel=\"noreferrer\">November 2015 walkthrough about Jetson TK1<\/a>.<\/li>\n<li><a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/901148\/jetson-tx1\/tensorflow-on-jetson-tx1\/post\/5027786\/#5027786\" rel=\"noreferrer\">November 2016 walkthrough about Jetson TX1<\/a>.<\/li>\n<\/ul>\n\n<p>If I haven't forgotten anything, here are the steps I've taken:<\/p>\n\n<ol>\n<li><p>Install Oracle Java 8 JDK:<\/p>\n\n<pre><code>$ sudo apt-get remove icedtea-8-plugin  #This is just in case\n$ sudo add-apt-repository ppa:webupd8team\/java\n$ sudo apt-get update\n$ sudo apt-get install oracle-java8-installer\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>(This is all you need in a pristine Xubuntu install, but google the above keywords otherwise, to read about selecting a default JRE and <code>javac<\/code>.)<\/p>\n\n<ol start=\"2\">\n<li><p>Dependencies:<\/p>\n\n<pre><code>sudo apt-get update\nsudo apt-get install git zip unzip swig python-numpy python-dev python-pip python-wheel\npip install --upgrade pip\n<\/code><\/pre><\/li>\n<li><p>Following <a href=\"https:\/\/bazel.build\/versions\/master\/docs\/install.html#compiling-from-source\" rel=\"noreferrer\">the instructions that come with Bazel<\/a>, download a Bazel source zip (I got <code>bazel-0.4.3-dist.zip<\/code>), make a directory like <code>~\/tf\/bazel\/<\/code> and unzip it there.<\/p><\/li>\n<li><p>I was getting an <code>OutOfMemoryError<\/code> during the following build, but <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/1476#issuecomment-230719919\" rel=\"noreferrer\">this fix<\/a> took care of it (i.e. adding the <code>-J-Xmx512m<\/code> for the bootstrap build).<\/p><\/li>\n<li><p>Call <code>bash .\/compile.sh<\/code>, and wait for a long time (overnight for me, but see the remarks at the end).<\/p><\/li>\n<li><p><code>$ git clone -b r0.12 https:\/\/github.com\/tensorflow\/tensorflow<\/code><\/p><\/li>\n<li><p>This seems like the only change to the source code that was necessary!<\/p>\n\n<pre><code>$ cd tensorflow\n$ grep -Rl \"lib64\"| xargs sed -i 's\/lib64\/lib\/g'\n<\/code><\/pre><\/li>\n<li><p>Then <code>$ .\/configure<\/code> and say <em>no<\/em> to everything. (Accept defaults where relevant.)<\/p><\/li>\n<li><p>The following took quite a few hours with my setup:<\/p>\n\n<pre><code>$ bazel build -c opt --jobs 1 --local_resources 1024,0.5,1.0 --verbose_failures \/\/tensorflow\/tools\/pip_package:build_pip_package\n$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n$ pip install --user \/tmp\/tensorflow_pkg\/ten&lt;Press TAB here&gt;\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>To see that it's installed, see if it works on the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"noreferrer\">TensorFlow Beginners tutorial<\/a>. I use <code>jupyter qtconsole<\/code> (i.e. the new name of IPython). Run the code in the <code>mnist_softmax.py<\/code>. It should take little time even on very limited machines.<\/p>\n\n<p>For some reason, <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#installing_from_sources\" rel=\"noreferrer\">TensorFlow's guide to building from source<\/a> doesn't suggest <a href=\"https:\/\/stackoverflow.com\/q\/34204551\/1143274\">running the unit tests<\/a>:<\/p>\n\n<pre><code>$ bazel test \/\/tensorflow\/...\n<\/code><\/pre>\n\n<p>(Yes, type in the ellipses.)<\/p>\n\n<p>Though I couldn't run them \u2014 it spent 19 hours trying to link <code>libtensorflow_cc.so<\/code>, and then something killed the linker. This was with half a core and 1536 MB memory limit. Maybe someone else, with a larger machine, can report on how the unit tests go.<\/p>\n\n<p>Why didn't we need to do the other things mentioned in those two walkthroughs? Firstly, most of that work is about taking care of GPU interfacing. Secondly, both Bazel and TensorFlow have become more self-contained since the first of those walkthroughs was written.<\/p>\n\n<p>Note that the above settings provided to Bazel for the build are very conservative (1024 MB RAM, half a core, one job at a time), because I'm running this through VirtualBox using a single core of a $200 netbook of the type that Intel makes for disadvantaged kids in Venezuela, Pakistan and Nigeria. (By the way, if you do this, make sure the virtual HDD is 20 GB at the very least \u2014 trying to build the unit tests above took about 5 GB of space.) The build of the wheel took almost 20 hours and the modest deep CNN from <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"noreferrer\">the second tutorial<\/a>, which is quoted to take up to half an hour to run on modern desktop CPUs, takes about 80 hours under this setup. One might wonder why I don't get a desktop, but the truth is that actual training with TensorFlow only makes sense on a high-end GPU (or a bunch thereof), and when we can hire an AWS spot instance with such a GPU for about 10 cents an hour without commitment and on a workable ad-hoc basis, it doesn't make a lot of sense to be training elsewhere. The 480000% speed-up is really noticeable. On the other hand, the convenience of having a local installation is well worth going through a process such as above.<\/p>\n"},{"AnswerId":"34540752","CreationDate":"2015-12-31T03:51:28.760","ParentId":null,"OwnerUserId":"5660925","Title":null,"Body":"<p>It appears that Google does not yet support tensorflow on 32-bit machines.<\/p>\n\n<p>On a 32-bit machine running Centos 6.5,the following error is received after the \"import tensorflow as tf\" command:\n     ImportError: tensorflow\/python\/_pywrap_tensorflow.so: wrong ELF class: <strong>ELFCLASS64<\/strong><\/p>\n\n<p>Until Google distributes a 32-bit version of tensorflow, I also recommend building tensorflow from source as specified <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"33635450","CreationDate":"2015-11-10T17:00:09.400","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>We have only tested the TensorFlow distribution on 64-bit Linux and Mac OS X, and distribute binary packages for those platforms only. Try following the <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\">source installation instructions<\/a> to build a version for your platform.<\/p>\n\n<p><strong>EDIT:<\/strong> One user has published <a href=\"http:\/\/cudamusing.blogspot.com\/2015\/11\/building-tensorflow-for-jetson-tk1.html\">instructions for running TensorFlow on a 32-bit ARM processor<\/a>, which is promising for other 32-bit architectures. These instructions may have useful pointers for getting TensorFlow and Bazel to work in a 32-bit environment.<\/p>\n"}]}
{"QuestionId":33634711,"AnswerCount":2,"Tags":"<makefile><debian><caffe><fpic><glog>","CreationDate":"2015-11-10T16:23:02.223","AcceptedAnswerId":null,"OwnerUserId":2071581.0,"Title":"caffe recompile with -fPIC libglog.a error","Body":"<p>I'm trying to install Caffe on Debian, but when i do  i get this error:<\/p>\n\n<blockquote>\n  <p>\/usr\/bin\/ld: \/usr\/local\/lib\/libglog.a(logging.cc.o): relocation R_X86_64_32 against `.rodata' can not be used when making a shared object; recompile with -fPIC<\/p>\n  \n  <p>\/usr\/local\/lib\/libglog.a: could not read symbols: Bad value<\/p>\n  \n  <p>collect2: error: ld returned 1 exit status\n  make: *** [.build_release\/lib\/libcaffe.so] Error 1<\/p>\n<\/blockquote>\n\n<p>where should i add the  flag?\nI also read that it wasn't enough for .<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"53343056","CreationDate":"2018-11-16T17:54:21.783","ParentId":null,"OwnerUserId":"1726797","Title":null,"Body":"<p>In glog, adding <code>-fPIC<\/code> to <code>AM_CXXFLAGS<\/code> in <code>Makefile.am<\/code> works. Adding it to the <code>CXXFLAGS<\/code> environment variable had no effect for me.<\/p>\n"},{"AnswerId":"46071185","CreationDate":"2017-09-06T09:12:01.303","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>Seems like the only solution in to recompile the problematic 3rd parties with -fPIC. (Unless you can link to the .so instead of the .a)<br>\nThe standard build of static libraries in packages is to build without -fPIC.<br>\nIn glog, for example, you should add it to the C flags in compile.ac.<\/p>\n"}]}
{"QuestionId":33634763,"AnswerCount":1,"Tags":"<python><gpu><theano><deep-learning>","CreationDate":"2015-11-10T16:25:10.960","AcceptedAnswerId":null,"OwnerUserId":5547104.0,"Title":"Theano performance issue when copying data to GPU","Body":"<p>I had some performance issues when trying to train a deep convolutional neural net using theano and lasagne. I made some experiments to investigate where they come from. One thing I found is that it takes very long to load batches of images from main memory to the GPU. Here is a minimal example that illustrates the problem. It simply times how long it takes to evaluate a theano identity function on batches of images with batch sizes 1,2,4,8,16,... I am working with RGB images of size 448x448.<\/p>\n\n<pre><\/pre>\n\n<p>My results are the following:<\/p>\n\n<pre><\/pre>\n\n<p>Note the increase of factor 100 when increasing the batch size from 8 to 16. Is this normal or do I have some kind of technical problems? If so, do you have any idea where it might come from? Any help is appreciated. It would also help if you run the code snippet and report what you see.<\/p>\n\n<p>EDIT:\nDaniel Renshaw pointed out that it has probably nothing to do with host-GPU copying. Any other ideas where the problem might come from? Some more information:<\/p>\n\n<p>The theano debug print of the function reads<\/p>\n\n<pre><\/pre>\n\n<p>Output of the theano profiling:<\/p>\n\n<pre><\/pre>\n\n<p>THEANO_FLAGS:\n<\/p>\n","answers":[{"AnswerId":"33647587","CreationDate":"2015-11-11T09:26:52.877","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Your computation is almost certainly not running on the GPU! As long as you're using standard <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">configuration flags<\/a>, Theano's optimizer is clever enough to see that no operations are actually performed so it doesn't add any \"move data to GPU\" and \"move data back from GPU\" operations in the compiled computation. You can see this by adding the following line just after the <code>f = theano.function([var], var)<\/code> line.<\/p>\n\n<pre><code>theano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>If you want to understand the overhead of moving data to and from the GPU, you're probably better served by Theano's built in <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/profiling.html\" rel=\"nofollow\">profiling tools<\/a>. Turn profiling on then, in the output, look at how much time is spent in <code>GpuFromHost<\/code> and <code>HostFromGpu<\/code> operations. This, of course, must be done with a more meaningful computation, one where data does actually need to be moved around.<\/p>\n\n<p>However, it is curious that you get the results that you do. If the computation is indeed running on the CPU, I would still not expect to see such a step change as the batch size increases. This is probably of little interest to you if you don't continue seeing the same behaviour when the computation is actually running on the GPU though.<\/p>\n\n<p>By the way, running your code (which on my server actually ran on the CPU despite <code>device=gpu<\/code> in the config, as explained above) I didn't get the same huge step change; my time multipliers were 2.6, 1.9, 4.0, 3.9, 2.0 (i.e. the time increased by a multiple of 2.6 from batch size = 1 to batch size = 2, etc.)<\/p>\n"}]}
{"QuestionId":33634764,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-10T16:25:11.193","AcceptedAnswerId":null,"OwnerUserId":3067467.0,"Title":"TensorBoard doesn't find scalar statistics","Body":"<p>Has anyone gotten TensorBoard working? Doesn't seem to find my ScalarSummary files.\nI'm using:<\/p>\n\n<pre><\/pre>\n\n<p>and:  contains a file called <\/p>\n\n<p>Which I assume are the scalar summaries?<\/p>\n","answers":[{"AnswerId":"33652076","CreationDate":"2015-11-11T13:56:11.870","ParentId":null,"OwnerUserId":"4384710","Title":null,"Body":"<p>Try to call tensorboard with an absolute logdir path.\nFor me it helped at least to see the graph. Unfortunately summaries still are not shown.<\/p>\n"}]}
{"QuestionId":33636925,"AnswerCount":10,"Tags":"<docker><jupyter><tensorflow>","CreationDate":"2015-11-10T18:23:38.577","AcceptedAnswerId":33642903.0,"OwnerUserId":276310.0,"Title":"How do I start tensorflow docker jupyter notebook","Body":"<p>I've installed the tensorflow docker container on an ubuntu machine.  The tensorflow docker <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md\" rel=\"noreferrer\">setup instructions<\/a> specify:<\/p>\n\n<pre><\/pre>\n\n<p>This puts me into the docker container terminal, and I can run python and execute the Hello World example.  I can also manually run .\\run_jupyter.sh to start the jupyter notebook. However, I can't reach the notebook from host.<\/p>\n\n<p>How do I start the jupyter notebook such that I can use the notebook from the host machine? Ideally I would like to use docker to launch the container and start jupyter in a single command.<\/p>\n","answers":[{"AnswerId":"33638070","CreationDate":"2015-11-10T19:32:16.077","ParentId":null,"OwnerUserId":"276310","Title":null,"Body":"<p>After further reading of <a href=\"https:\/\/docs.docker.com\/v1.8\/reference\/commandline\/run\/\">docker documentation<\/a> I have a solution that works for me:<\/p>\n\n<pre><code>docker run -p 8888:8888 -p 6006:6006 b.gcr.io\/tensorflow\/tensorflow .\/run_jupyter.sh\n<\/code><\/pre>\n\n<p>The -p 8888:8888 and -p 6006:6006 expose the container ports to the host on the same port number.  If you just use -p 8888, a random port on the host will be assigned.<\/p>\n\n<p>The .\/run_jupyter.sh tells docker what to execute within the container.<\/p>\n\n<p>With this command, I can use a browser on the host machine to connect to <a href=\"http:\/\/localhost:8888\/\">http:\/\/localhost:8888\/<\/a> and access the jupyter notebook.<\/p>\n\n<p>UPDATE:\nAfter wrestling with docker on windows I switched back to a Ubuntu machine with docker.  My notebook was being erased between docker sessions which makes sense after reading more docker documentation.  Here is an updated command which also mounts a host directory within the container and starts jupyter pointing to that mounted directory.  Now my notebook is saved on the host and will be available next time start up tensorflow.<\/p>\n\n<pre><code>docker run -p 8888:8888 -p 6006:6006 -v \/home\/rob\/notebook:\/notebook b.gcr.io\/tensorflow\/tensorflow sh -c \"jupyter notebook \/notebook\"\n<\/code><\/pre>\n"},{"AnswerId":"36479554","CreationDate":"2016-04-07T14:41:23.773","ParentId":null,"OwnerUserId":"2704551","Title":null,"Body":"<p>To get this to run under hyper-v. Perform the following steps: <\/p>\n\n<p>1) Create a docker virtual machine using <a href=\"https:\/\/blogs.msdn.microsoft.com\/scicoria\/2014\/10\/09\/getting-docker-running-on-hyper-v-8-1-2012-r2\/\" rel=\"nofollow\">https:\/\/blogs.msdn.microsoft.com\/scicoria\/2014\/10\/09\/getting-docker-running-on-hyper-v-8-1-2012-r2\/<\/a> this will get you a working docker container. You can connect to it via the console or via ssh. I'd put at least 8gb of memory since I'm sure this will use a lot of memory. <\/p>\n\n<p>2) run \"ifconfig\" to determine the IP address of the Docker VM <\/p>\n\n<p>3) On the docker shell prompt type: <\/p>\n\n<p>docker run -p 8888:8888 -p 6006:6006 -it b.gcr.io\/tensorflow\/tensorflow <\/p>\n\n<p>4) Connect to the Jupyter Workbench using http:\/[ifconfig address]:8888\/<\/p>\n"},{"AnswerId":"43233221","CreationDate":"2017-04-05T13:52:45.293","ParentId":null,"OwnerUserId":"231589","Title":null,"Body":"<p>For some reason I ran into one additional problem that I needed to overcome beyond the examples provided, using the <code>--ip<\/code> flag:<\/p>\n\n<pre><code>nvidia-docker run --rm \\\n  -p 8888:8888 -p 6006:6006 \\\n  -v `pwd`:\/root \\\n  -it tensorflow\/tensorflow:latest-devel-gpu-py3 sh -c \"jupyter notebook --ip 0.0.0.0 .\"\n<\/code><\/pre>\n\n<p>And then I can access via <a href=\"http:\/\/localhost:8888\" rel=\"nofollow noreferrer\">http:\/\/localhost:8888<\/a> from my machine. In some ways this makes sense; within the container you bind to <code>0.0.0.0<\/code> which represents all available addresses. But whether I need to do this seems to vary (e.g I've started notebooks using <code>jupyter\/scipy-notebook<\/code> without having to do this). <\/p>\n\n<p>In any case, the above command works for me, might be of use to others.<\/p>\n"},{"AnswerId":"33642903","CreationDate":"2015-11-11T01:43:09.693","ParentId":null,"OwnerUserId":"1840190","Title":null,"Body":"<p>For a Linux host Robert Graves answer will work, but for Mac OS X or Windows there is more to be done because docker runs in a virtual machine. <\/p>\n\n<p>So to begin launch the docker shell (or any shell if you are using Linux) and run the following command to launch a new TensorFlow container:<\/p>\n\n<pre><code>docker run -p 8888:8888 -p 6006:6006 b.gcr.io\/tensorflow\/tensorflow .\/run_jupyter.sh\n<\/code><\/pre>\n\n<p>Then for Mac OS X and Windows you need to do the following only once:<\/p>\n\n<ol>\n<li>Open VirtualBox<\/li>\n<li>Click on the docker vm (mine was automatically named \"default\")<\/li>\n<li>Open the settings by clicking settings<\/li>\n<li>In the network settings open the port forwarding dialog<\/li>\n<li>Click the + symbol to add another port and connect a port from your mac to the VM by filling in the dialog as shown below. In this example I chose port 8810 because I run other notebooks using port 8888.\n<a href=\"https:\/\/i.stack.imgur.com\/2HB9P.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2HB9P.png\" alt=\"enter image description here\"><\/a><\/li>\n<li>then open a browser and connect to <a href=\"http:\/\/localhost:8810\" rel=\"noreferrer\">http:\/\/localhost:8810<\/a> (or whichever port you set in the host port section<\/li>\n<li>Make your fancy pants machine learning app!<\/li>\n<\/ol>\n"},{"AnswerId":"57930416","CreationDate":"2019-09-13T21:14:04.900","ParentId":null,"OwnerUserId":"4575433","Title":null,"Body":"<p>As an alternative to the official TensorFlow image, you can also use the <a href=\"https:\/\/github.com\/ml-tooling\/ml-workspace\" rel=\"nofollow noreferrer\"><strong>ML Workspace<\/strong><\/a> Docker image. The ML Workspace is an open-source web IDE that combines Jupyter, VS Code, TensorFlow, and many other tools &amp; libraries into one convenient Docker image. Deploying a single workspace instance is as simple as:<\/p>\n\n<pre><code>docker run -p 8080:8080 mltooling\/ml-workspace:latest\n<\/code><\/pre>\n\n<p>All tools are accessible from the same port and integrated into the Jupyter UI. You can find the documentation <a href=\"https:\/\/github.com\/ml-tooling\/ml-workspace\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"36505342","CreationDate":"2016-04-08T17:02:32.023","ParentId":null,"OwnerUserId":"2073494","Title":null,"Body":"<p>These steps worked for me if you are a total docker noob using a windows machine.<\/p>\n\n<p>Versions: Windows 8.1, docker 1.10.3, tensorflow r0.7<\/p>\n\n<ol>\n<li>Run Docker Quickstart Terminal<\/li>\n<li>After it is loaded, note the ip address. If you can't find it use this <code>docker-machine ip<\/code> and make a note. Lets call it 'ip address'. Will look something like this: 192.168.99.104 (I made up this ip address)<\/li>\n<li><p>Paste this command on the docker terminal:<\/p>\n\n<p><code>docker run -p 8888:8888 -p 6006:6006 b.gcr.io\/tensorflow\/tensorflow<\/code>. <\/p>\n\n<p>If you are running this for the first time, it will download and install the image on this light weight vm. Then it should say 'The Jupyter notebook is running at ....' -> This is a good sign!<\/p><\/li>\n<li>Open your browser at: <code>&lt;your ip address (see above)&gt;:8888<\/code>. Eg. 192.168.99.104:8888\/<\/li>\n<li>Hopefully you can see your ipython files.<\/li>\n<\/ol>\n"},{"AnswerId":"36797039","CreationDate":"2016-04-22T14:52:27.083","ParentId":null,"OwnerUserId":"3274693","Title":null,"Body":"<h2>My simple yet efficient workflow:<\/h2>\n\n<h3>TL;DR version:<\/h3>\n\n<ol>\n<li>Open <strong>Docker Quickstart Terminal<\/strong>. If it is already open, run <code>$ cd<\/code><\/li>\n<li>Run this <strong>once<\/strong>: <code>$ docker run -it -p 8888:8888 -p 6006:6006 -v \/$(pwd)\/tensorflow:\/notebooks --name tf b.gcr.io\/tensorflow\/tensorflow<\/code><\/li>\n<li>To start every time: <code>$ docker start -i tf<\/code><\/li>\n<\/ol>\n\n<blockquote>\n  <p>If you are not on windows, you should probably change <code>\/$(pwd)<\/code> to <code>$(pwd)<\/code><\/p>\n<\/blockquote>\n\n<p>You will get an empty folder named <code>tensorflow<\/code> in your home directory for use as a persistent storage of project files such as Ipython Notebooks and datasets.<\/p>\n\n<h3>Explanation:<\/h3>\n\n<ol>\n<li><code>cd<\/code> for making sure you are in your home directory.<\/li>\n<li>params:\n\n<ul>\n<li><code>-it<\/code> stands for interactive, so you can interact with the container in the terminal environment. <\/li>\n<li><code>-v host_folder:container_folder<\/code> enables sharing a folder between the host and the container. The host folder should be inside your home directory. <code>\/$(pwd)<\/code> translates to <code>\/\/c\/Users\/YOUR_USER_DIR<\/code> in Windows 10. This folder is seen as <code>notebooks<\/code> directory in the container which is used by Ipython\/Jupyter Notebook. <\/li>\n<li><code>--name tf<\/code> assigns the name <code>tf<\/code> to the container.<\/li>\n<li><code>-p 8888:8888 -p 6006:6006<\/code> mapping ports of container to host, first pair for Jupyter notebook, the second one for Tensorboard<\/li>\n<\/ul><\/li>\n<li><code>-i<\/code> stands for interactive<\/li>\n<\/ol>\n\n<h3><a href=\"https:\/\/stackoverflow.com\/a\/36927672\/3274693\">Running TensorFlow on the cloud<\/a><\/h3>\n"},{"AnswerId":"37807773","CreationDate":"2016-06-14T09:16:52.093","ParentId":null,"OwnerUserId":"2470546","Title":null,"Body":"<p>To tidy up the things a little bit, I want to give some additional explanations because I also suffered a lot setting up docker with tensorflow. For this I refer to <a href=\"https:\/\/youtu.be\/Wwnj8NayOcQ\" rel=\"nofollow\">this<\/a> video which is unfortunately not selfexplanatory in all cases.\nI assume you already installed docker. The really interesting general part of the video starts at minute 0:44 where he finally started docker. Until there he only downloads the tensorflow repo into the folder, that he then mounts into the container. You can of course put anything else into the container and access it later in the docker VM. <\/p>\n\n<hr>\n\n<ol>\n<li><p>First he runs the long docker command <code>docker run \u2013dit -v \/c\/Users\/Jay\/:\/media\/disk \u2013p 8000 \u2013p 8888 \u2013p 6006  b.gcr.io\/tensorflow\/tensorflow<\/code>. The \u201crun\u201d command starts containers. In this case it starts the container \u201cb.gcr.io\/tensorflow\/tensorflow\u201d, whose address is provided within the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/get_started\/os_setup.html#docker-installation\" rel=\"nofollow\">tensorflow docker installation tutorial<\/a>. The container will be downloaded by docker if not already locally available. \nThen he gives two additional kinds of arguments: He mounts a folder of the hostsystem at the given path to the container. DO NOT forget to give the partition in the beginning (eg. \"\/c\/\"). \nAdditionally he declares ports being available later from the host machine with the params -p. \nFrom all this command you get back the [CONTAINER_ID] of this container execution! \nYou can always see the currently running containers by running \u201cdocker ps\u201d in the docker console. Your container created above should appear in this list with the same id.<\/p>\n\n<hr><\/li>\n<li><p>Next Step: With your container running, you now want to execute something in it. In our case jupyter notebook or tensorflow or whatever: To do this you make docker execute the bash on the newly created container: <code>docker exec \u2013ti [CONTAINER_ID] bash<\/code>. This command now starts a bash shell on your container. You see this because the \u201c$\u201d now changed to root@[CONTAINER_ID]:. From here is no way back. If you want to get back to the docker terminal, you have to start another fresh docker console like he is doing in minute 1:10. Now with a bash shell running in the container you can do whatever you want and execute Jupiter or tensorflow or whatever. The folder of the host system, you gave in the run command, should be available now under  \u201c\/media\/disk\u201d. <\/p>\n\n<hr><\/li>\n<li><p>Last step accessing the VM output. It still did not want to work out for me and I could not access my notebook. You still have to find the correct IP and Port to access the launched notebook, tensorboard session or whatever. First find out the main IP by using <code>docker-machine \u2013ls<\/code>. In this list you get the URL. (If it is your only container it is called default.) You can leave away the port given here. Then from <code>docker ps<\/code> you get the list of forwarded ports. When there is written 0.0.0.32776->6006\/tcp in the list, you can access it from the hostmachine by using the port given in the first place (Awkyard). So in my case the executed tensorboard in the container said \u201claunched on port 6006\u201d. Then from my hostmachine I needed to enter <a href=\"http:\/\/192.168.99.100:32776\/\" rel=\"nofollow\">http:\/\/192.168.99.100:32776\/<\/a> to access it.<\/p><\/li>\n<\/ol>\n\n<p>-> And that\u2019s it! It ran for me like this!<\/p>\n"},{"AnswerId":"37982268","CreationDate":"2016-06-23T04:19:37.533","ParentId":null,"OwnerUserId":"6498984","Title":null,"Body":"<p>It gives you the terminal prompt:<\/p>\n\n<pre><code>FOR \/f \"tokens=*\" %i IN ('docker-machine env --shell cmd vdocker') DO %i\ndocker run -it tensorflow\/tensorflow:r0.9-devel\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>FOR \/f \"tokens=*\" %i IN ('docker-machine env --shell cmd vdocker') DO %i\ndocker run -it b.gcr.io\/tensorflow\/tensorflow:latest-devel\n<\/code><\/pre>\n\n<p>You should have 'vdocker' or change vdocker to 'default'.<\/p>\n"},{"AnswerId":"39577370","CreationDate":"2016-09-19T16:00:10.447","ParentId":null,"OwnerUserId":"881626","Title":null,"Body":"<p>Jupyter now has a ready to run <a href=\"https:\/\/github.com\/jupyter\/docker-stacks\/tree\/master\/tensorflow-notebook\" rel=\"nofollow noreferrer\">Docker image<\/a> for TensorFlow:<\/p>\n\n<p><code>docker run -d -v $(pwd):\/home\/jovyan\/work -p 8888:8888 jupyter\/tensorflow-notebook<\/code><\/p>\n"}]}
{"QuestionId":33637208,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-10T18:42:26.367","AcceptedAnswerId":33857405.0,"OwnerUserId":434408.0,"Title":"Getting: tensorflow is not a supported wheel on this platform","Body":"<p>I tried installing tensorflow on my system, but I received the following error:<\/p>\n\n<p>tensorflow-0.5.0-cp27-none-linux_x86_64.whl is not a supported wheel on this platform<\/p>\n\n<p><strong>Exception Information captured in pip.log file:-<\/strong><\/p>\n\n<hr>\n\n<p>\/home\/gansai\/tensorflow\/bin\/pip run on Wed Nov 11 00:19:05 2015<\/p>\n\n<p>tensorflow-0.5.0-cp27-none-linux_x86_64.whl is not a supported wheel on this platform.<\/p>\n\n<p>Exception information:\nTraceback (most recent call last):<\/p>\n\n<p>File \"\/home\/gansai\/tensorflow\/local\/lib\/python2.7\/site-packages\/pip\/basecommand.py\", line 122, in main\n    status = self.run(options, args)<\/p>\n\n<p>File \"\/home\/gansai\/tensorflow\/local\/lib\/python2.7\/site-packages\/pip\/commands\/install.py\", line 269, in run\n    InstallRequirement.from_line(name, None))<\/p>\n\n<p>File \"\/home\/gansai\/tensorflow\/local\/lib\/python2.7\/site-packages\/pip\/req.py\", line 168, in from_line\n    raise UnsupportedWheel(\"%s is not a supported wheel on this platform.\" % wheel.filename)<\/p>\n\n<p>UnsupportedWheel: tensorflow-0.5.0-cp27-none-linux_x86_64.whl is not a supported wheel on this platform.<\/p>\n\n<hr>\n\n<p>What could I do to install tensorflow and start experimenting it?<\/p>\n","answers":[{"AnswerId":"33857405","CreationDate":"2015-11-22T16:52:37.153","ParentId":null,"OwnerUserId":"434408","Title":null,"Body":"<p><em>tensorflow-0.5.0-cp27-none-linux_x86_64.whl is not a supported wheel on this platform<\/em><\/p>\n\n<p>The above error comes because of trying to install TensorFlow onto a 32 bit system. As you could observe, the wheel was linux_x86_64, which is intended to be installed on 64 bit.<\/p>\n\n<p>Steps to follow:-<\/p>\n\n<p>Direct Binary Installation<\/p>\n\n<ol>\n<li>Install Debian 64 bit OS. Download debian 64 bit ISO from torrent (<a href=\"http:\/\/cdimage.debian.org\/debian-cd\/current-live\/amd64\/bt-hybrid\/\" rel=\"noreferrer\">http:\/\/cdimage.debian.org\/debian-cd\/current-live\/amd64\/bt-hybrid\/<\/a>)<\/li>\n<li>Install python-dev. <code>apt-get install python-dev<\/code><\/li>\n<li>Install tensor-flow <code>python2.7 -m pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/code><\/li>\n<\/ol>\n\n<p>Docker Based Installation<\/p>\n\n<ol>\n<li>Install Debian 64 bit OS. Download debian 64 bit ISO from torrent (<a href=\"http:\/\/cdimage.debian.org\/debian-cd\/current-live\/amd64\/bt-hybrid\/\" rel=\"noreferrer\">http:\/\/cdimage.debian.org\/debian-cd\/current-live\/amd64\/bt-hybrid\/<\/a>)<\/li>\n<li>Install Docker on Debian. Follow steps mentioned in <a href=\"http:\/\/docs.docker.com\/engine\/installation\/debian\/#debian-jessie-80-64-bit\" rel=\"noreferrer\">http:\/\/docs.docker.com\/engine\/installation\/debian\/#debian-jessie-80-64-bit<\/a><\/li>\n<li>Run docker container for tensorflow. <code>docker run -it b.gcr.io\/tensorflow\/tensorflow<\/code>. Follow steps mentioned in <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation\" rel=\"noreferrer\">http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation<\/a><\/li>\n<\/ol>\n\n<p>The main point to be noted here is that, as of now, python wheel for tensorflow is supported for 64 bit system, as shared in <code>pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/code><\/p>\n"},{"AnswerId":"36236625","CreationDate":"2016-03-26T14:44:00.627","ParentId":null,"OwnerUserId":"828184","Title":null,"Body":"<p>It might be unrelated but I had the same error with tensorflow-0.7.1-cp34-none-linux_x86_64.whl on a fresh Ubuntu Linux 14.04 LTS (64bit) and this is what has helped:<\/p>\n\n<pre><code>sudo apt-get install python3-setuptools -y &amp;&amp; sudo easy_install3 pip -y &amp;&amp; sudo apt-get install python 3.5-dev -y &amp;&amp; sudo apt-get install python3.4-dev -y\nsudo pip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp34-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>You can call a tensorflow script like this:<\/p>\n\n<pre><code>#call it with python3.4\npython3.4 tensorflow_demo.py\n<\/code><\/pre>\n\n<p>While researching I often read about using wget and renaming the file but this became obsolete with the latest version, see here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1142#issuecomment-186740120\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1142#issuecomment-186740120<\/a><\/p>\n"},{"AnswerId":"33642934","CreationDate":"2015-11-11T01:46:55.853","ParentId":null,"OwnerUserId":"3057648","Title":null,"Body":"<p>I guess pip3 is being used for installation<br>\nit can be solved by using pip2.7<\/p>\n\n<p>I followed the steps in <a href=\"https:\/\/stackoverflow.com\/a\/33635645\/3057648\">here<\/a>  <\/p>\n\n<p>hope it helps you:)<\/p>\n"}]}
{"QuestionId":33639123,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-10T20:37:18.260","AcceptedAnswerId":33652509.0,"OwnerUserId":5520753.0,"Title":"Cannot find any scalar summaries in TensorBoard","Body":"<p>I am doing mnist tutorial, and fully_connected_feed.py works and saves events.out.tfevents.1447186888 file to ~..\\data\\\nwhen I trying to open TensorBoard like this <\/p>\n\n<pre><\/pre>\n\n<p>or like this <\/p>\n\n<pre><\/pre>\n\n<p>It opens, but then I see  \"No scalar summary tags were found.\" <\/p>\n","answers":[{"AnswerId":"33652509","CreationDate":"2015-11-11T14:19:57.557","ParentId":null,"OwnerUserId":"5550902","Title":null,"Body":"<p>Try to use<\/p>\n\n<p><code>tensorboard --logdir=home\/$USER\/tensorflow\/tensorflow\/g3doc\/tutorials\/mnist\/data<\/code><\/p>\n\n<p>or<\/p>\n\n<p><code>tensorboard --logdir=${PWD}<\/code>    in that directory<\/p>\n\n<p>Because tensorboard checks path existence by using os.path.exists()<\/p>\n\n<p>=<\/p>\n\n<p>Regarding that, I would like to set <code>alias tensorboard='tensorboard --logdir=${PWD}'<\/code>  for convenient<\/p>\n"}]}
{"QuestionId":33640581,"AnswerCount":9,"Tags":"<python><tensorflow>","CreationDate":"2015-11-10T22:07:54.417","AcceptedAnswerId":36784797.0,"OwnerUserId":230987.0,"Title":"How to do Xavier initialization on TensorFlow","Body":"<p>I'm porting my Caffe network over to TensorFlow but it doesn't seem to have xavier initialization. I'm using  but this seems to be making it a lot harder to train.<\/p>\n","answers":[{"AnswerId":"55334946","CreationDate":"2019-03-25T09:42:07.043","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>In <strong>Tensorflow 2.0<\/strong> and further both <code>tf.contrib.*<\/code> and <code>tf.get_variable()<\/code> are deprecated. In order to do Xavier initialization you now have to switch to:<\/p>\n\n<pre><code>init = tf.initializers.GlorotUniform()\nvar = tf.Variable(init(shape=shape))\n# or a oneliner with a little confusing brackets\nvar = tf.Variable(tf.initializers.GlorotUniform()(shape=shape))\n<\/code><\/pre>\n\n<p>Glorot uniform and Xavier uniform are two different names of the same initialization type. If you want to know more about how to use initializations in TF2.0 with or without Keras refer to <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/initializers\/GlorotUniform\" rel=\"noreferrer\">documentation<\/a>.<\/p>\n"},{"AnswerId":"34376010","CreationDate":"2015-12-19T22:25:34.570","ParentId":null,"OwnerUserId":"249341","Title":null,"Body":"<p>A nice wrapper around <code>tensorflow<\/code> called <a href=\"https:\/\/github.com\/google\/prettytensor\" rel=\"noreferrer\"><code>prettytensor<\/code><\/a> gives an implementation in the source code (copied directly from <a href=\"https:\/\/github.com\/google\/prettytensor\/blob\/a69f13998258165d6682a47a931108d974bab05e\/prettytensor\/layers.py\" rel=\"noreferrer\">here<\/a>):<\/p>\n\n<pre><code>def xavier_init(n_inputs, n_outputs, uniform=True):\n  \"\"\"Set the parameter initialization using the method described.\n  This method is designed to keep the scale of the gradients roughly the same\n  in all layers.\n  Xavier Glorot and Yoshua Bengio (2010):\n           Understanding the difficulty of training deep feedforward neural\n           networks. International conference on artificial intelligence and\n           statistics.\n  Args:\n    n_inputs: The number of input nodes into each output.\n    n_outputs: The number of output nodes for each input.\n    uniform: If true use a uniform distribution, otherwise use a normal.\n  Returns:\n    An initializer.\n  \"\"\"\n  if uniform:\n    # 6 was used in the paper.\n    init_range = math.sqrt(6.0 \/ (n_inputs + n_outputs))\n    return tf.random_uniform_initializer(-init_range, init_range)\n  else:\n    # 3 gives us approximately the same limits as above since this repicks\n    # values greater than 2 standard deviations from the mean.\n    stddev = math.sqrt(3.0 \/ (n_inputs + n_outputs))\n    return tf.truncated_normal_initializer(stddev=stddev)\n<\/code><\/pre>\n"},{"AnswerId":"43714083","CreationDate":"2017-05-01T04:00:37.810","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>TF-contrib has <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.layers\/initializers#xavier_initializer\" rel=\"noreferrer\"><code>xavier_initializer<\/code><\/a>. Here is an example how to use it:<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.get_variable(\"a\", shape=[4, 4], initializer=tf.contrib.layers.xavier_initializer())\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print sess.run(a)\n<\/code><\/pre>\n\n<p>In addition to this, tensorflow has other initializers:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.layers\/initializers#xavier_initializer_conv2d\" rel=\"noreferrer\">xavier_initializer_conv2d<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.layers\/initializers#variance_scaling_initializer\" rel=\"noreferrer\">variance_scaling_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant_initializer\" rel=\"noreferrer\">constant_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/zeros_initializer\" rel=\"noreferrer\">zeros_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/ones_initializer\" rel=\"noreferrer\">ones_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/uniform_unit_scaling_initializer\" rel=\"noreferrer\">uniform_unit_scaling_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/truncated_normal_initializer\" rel=\"noreferrer\">truncated_normal_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random_uniform_initializer\" rel=\"noreferrer\">random_uniform_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random_normal_initializer\" rel=\"noreferrer\">random_normal_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/orthogonal_initializer\" rel=\"noreferrer\">orthogonal_initializer<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/keras\/initializers\/he_uniform\" rel=\"noreferrer\">as well as a lot of initializers from keras<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"33679580","CreationDate":"2015-11-12T19:23:24.540","ParentId":null,"OwnerUserId":"2125302","Title":null,"Body":"<p>I looked and I couldn't find anything built in. However, according to this:<\/p>\n\n<p><a href=\"http:\/\/andyljones.tumblr.com\/post\/110998971763\/an-explanation-of-xavier-initialization\" rel=\"nofollow\">http:\/\/andyljones.tumblr.com\/post\/110998971763\/an-explanation-of-xavier-initialization<\/a><\/p>\n\n<p>Xavier initialization is just sampling a (usually Gaussian) distribution where the variance is a function of the number of neurons.  <code>tf.random_normal<\/code> can do that for you, you just need to compute the stddev (i.e. the number of neurons being represented by the weight matrix you're trying to initialize).<\/p>\n"},{"AnswerId":"50240889","CreationDate":"2018-05-08T19:19:54.367","ParentId":null,"OwnerUserId":"5520444","Title":null,"Body":"<p>Just in case you want to use one line as you do with:<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal((n_prev, n), stddev=0.1))\n<\/code><\/pre>\n\n<p>You can do:<\/p>\n\n<pre><code>W = tf.Variable(tf.contrib.layers.xavier_initializer()((n_prev, n)))\n<\/code><\/pre>\n"},{"AnswerId":"49798529","CreationDate":"2018-04-12T13:57:18.597","ParentId":null,"OwnerUserId":"315385","Title":null,"Body":"<p>Via the <code>kernel_initializer<\/code> parameter to <code>tf.layers.conv2d, tf.layers.conv2d_transpose, tf.layers.Dense<\/code>  etc<\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>layer = tf.layers.conv2d(\n     input, 128, 5, strides=2,padding='SAME',\n     kernel_initializer=tf.contrib.layers.xavier_initializer())\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d_transpose\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d_transpose<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/Dense\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/Dense<\/a><\/p>\n"},{"AnswerId":"45380994","CreationDate":"2017-07-28T19:24:43.197","ParentId":null,"OwnerUserId":"832621","Title":null,"Body":"<p>Just to add another example on how to define a <code>tf.Variable<\/code> initialized using <a href=\"http:\/\/proceedings.mlr.press\/v9\/glorot10a\/glorot10a.pdf\" rel=\"noreferrer\">Xavier and Yoshua<\/a>'s method:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    ...\n    initializer = tf.contrib.layers.xavier_initializer()\n    w1 = tf.Variable(initializer(w1_shape))\n    b1 = tf.Variable(initializer(b1_shape))\n    ...\n<\/code><\/pre>\n\n<p>This prevented me from having <code>nan<\/code> values on my loss function due to numerical instabilities when using multiple layers with RELUs.<\/p>\n"},{"AnswerId":"36784797","CreationDate":"2016-04-22T04:23:57.423","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Since version 0.8 there is a Xavier initializer, <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.layers.html#xavier_initializer\">see here for the docs<\/a>.<\/p>\n\n<p>You can use something like this:<\/p>\n\n<pre><code>W = tf.get_variable(\"W\", shape=[784, 256],\n           initializer=tf.contrib.layers.xavier_initializer())\n<\/code><\/pre>\n"},{"AnswerId":"33711310","CreationDate":"2015-11-14T17:37:42.593","ParentId":null,"OwnerUserId":"347257","Title":null,"Body":"<p>@Aleph7, Xavier\/Glorot initialization depends the number of incoming connections (fan_in), number outgoing connections (fan_out), and kind of activation function (sigmoid or tanh) of the neuron. See this: <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf\">http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf<\/a><\/p>\n\n<p>So now, to your question. This is how I would do it in TensorFlow:<\/p>\n\n<pre><code>(fan_in, fan_out) = ...\n    low = -4*np.sqrt(6.0\/(fan_in + fan_out)) # use 4 for sigmoid, 1 for tanh activation \n    high = 4*np.sqrt(6.0\/(fan_in + fan_out))\n    return tf.Variable(tf.random_uniform(shape, minval=low, maxval=high, dtype=tf.float32))\n<\/code><\/pre>\n\n<p>Note that we should be sampling from a uniform distribution, and not the normal distribution as suggested in the other answer.<\/p>\n\n<p>Incidentally, I wrote a <a href=\"http:\/\/deliprao.com\/archives\/100\">post yesterday<\/a> for something different using TensorFlow that happens to also use Xavier initialization. If you're interested, there's also a python notebook with an end-to-end example: <a href=\"https:\/\/github.com\/delip\/blog-stuff\/blob\/master\/tensorflow_ufp.ipynb\">https:\/\/github.com\/delip\/blog-stuff\/blob\/master\/tensorflow_ufp.ipynb<\/a> <\/p>\n"}]}
{"QuestionId":33641358,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-10T23:04:28.820","AcceptedAnswerId":null,"OwnerUserId":2948749.0,"Title":"Can someone help me with TensorFlow?","Body":"<p>Google just opened up TensorFlow as opened source.\nI read it a bit but looks like you can only train it with their given MNIST data.<\/p>\n\n<p>I am looking for example code where i can train with my own data, and output results for my test file.<\/p>\n\n<p>where I have .csv file (like a sample per line) as training data (with id,output,+72 more columns)<\/p>\n\n<p>and have another .csv file for test data where i'd to predict output(1 or 0).<\/p>\n\n<p>Anyone understand that TensorFlow enough to give me some sample code?<\/p>\n","answers":[{"AnswerId":"34123157","CreationDate":"2015-12-06T21:43:24.717","ParentId":null,"OwnerUserId":"334757","Title":null,"Body":"<p>The best solution I have found is:<\/p>\n\n<p><a href=\"https:\/\/github.com\/google\/skflow\">https:\/\/github.com\/google\/skflow<\/a><\/p>\n\n<p>Charles<\/p>\n"},{"AnswerId":"33708187","CreationDate":"2015-11-14T12:02:20.343","ParentId":null,"OwnerUserId":"5561751","Title":null,"Body":"<p>You can have a look at those examples (like linear regression): <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\" rel=\"nofollow\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples<\/a><\/p>\n\n<p>But, for examples using mnist, you just need to replace inputs (train and test mnist data by your own data arrays).<\/p>\n"},{"AnswerId":"33785517","CreationDate":"2015-11-18T16:37:41.297","ParentId":null,"OwnerUserId":"236247","Title":null,"Body":"<p>Ok, here is the code sample from the site for csv.  You need to use TextLineReader to handle the csv format if that is what you are interested in and it sounds like you are.  For all of your options to read files, the link is <a href=\"http:\/\/tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow\">here<\/a><\/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.concat(0, [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)\n<\/code><\/pre>\n"}]}
{"QuestionId":33641799,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-10T23:43:14.540","AcceptedAnswerId":33644778.0,"OwnerUserId":189456.0,"Title":"Why does TensorFlow example fail when increasing batch size?","Body":"<p>I was looking at the <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/beginners\/index.md\">Tensorflow MNIST example for beginners<\/a> and found that in this part:<\/p>\n\n<pre><\/pre>\n\n<p>changing the batch size from 100 to be above 204 causes the model to fail to converge. It works up to 204, but at 205 and any higher number I tried, the accuracy would end up &lt; 10%. Is this a bug, something about the algorithm, something else?<\/p>\n\n<p>This is running their binary installation for OS X, seems to be version 0.5.0.<\/p>\n","answers":[{"AnswerId":"33719722","CreationDate":"2015-11-15T12:35:54.803","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>@dga nicely explained you the reason of such behavior (the cross_entropy becomes too huge) and thus the algorithm will not be able to converge. There are a couple of ways to fix this. He already suggested to decrease the learning rate.<\/p>\n\n<p>Gradient descent is the most basic algorithm. Almost all other <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.md#AUTOGENERATED-optimizers\" rel=\"nofollow noreferrer\">optimizers<\/a> will be working properly:<\/p>\n\n<pre><code>train_step = tf.train.AdagradOptimizer(0.01).minimize(cross_entropy)\ntrain_step = tf.train.AdamOptimizer().minimize(cross_entropy)\ntrain_step = tf.train.FtrlOptimizer(0.01).minimize(cross_entropy)\ntrain_step = tf.train.RMSPropOptimizer(0.01, 0.1).minimize(cross_entropy)\n<\/code><\/pre>\n\n<p>Another approach is to 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> which handles numeric instabilities.<\/p>\n"},{"AnswerId":"33644778","CreationDate":"2015-11-11T05:30:05.850","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You're using the very basic linear model in the beginners example?<\/p>\n\n<p>Here's a trick to debug it - watch the cross-entropy as you increase the batch size (the first line is from the example, the second I just added):<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y))\ncross_entropy = tf.Print(cross_entropy, [cross_entropy], \"CrossE\")\n<\/code><\/pre>\n\n<p>At a batch size of 204, you'll see:<\/p>\n\n<pre><code>I tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[92.37558]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[90.107414]\n<\/code><\/pre>\n\n<p>But at 205, you'll see a sequence like this, from the start:<\/p>\n\n<pre><code>I tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[472.02966]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[475.11697]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[1418.6655]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[1546.3833]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[1684.2932]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[1420.02]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[1796.0872]\nI tensorflow\/core\/kernels\/logging_ops.cc:64] CrossE[nan]\n<\/code><\/pre>\n\n<p>Ack - NaN's showing up.  Basically, the large batch size is creating such a huge gradient that your model is spiraling out of control -- the updates it's applying are too large, and overshooting the direction it should go by a huge margin.<\/p>\n\n<p>In practice, there are a few ways to fix this.  You could reduce the learning rate from .01 to, say, .005, which results in a final accuracy of 0.92.  <\/p>\n\n<pre><code>train_step = tf.train.GradientDescentOptimizer(0.005).minimize(cross_entropy)\n<\/code><\/pre>\n\n<p>Or you could use a more sophisticated optimization algorithm (Adam, Momentum, etc.) that tries to do more to figure out the direction of the gradient.  Or you could use a more complex model that has more free parameters across which to disperse that big gradient.<\/p>\n"},{"AnswerId":"34364526","CreationDate":"2015-12-18T21:54:10.770","ParentId":null,"OwnerUserId":"5696306","Title":null,"Body":"<p><strong>Nan occurs when 0*log(0) occurs:<\/strong><\/p>\n\n<p>replace:<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y))\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y + 1e-10))\n<\/code><\/pre>\n"},{"AnswerId":"33645235","CreationDate":"2015-11-11T06:19:15.697","ParentId":null,"OwnerUserId":"1625244","Title":null,"Body":"<p>@dga gave a great answer, but I wanted to expand a little.<\/p>\n\n<p>When I wrote the beginners tutorial, I implemented the cost function like so:<\/p>\n\n<blockquote>\n  <p>cross_entropy = -tf.reduce_sum(y_*tf.log(y))<\/p>\n<\/blockquote>\n\n<p>I wrote it that way because that looks most similar to the mathematical definition of cross-entropy. But it might actually be better to do something like this:<\/p>\n\n<blockquote>\n  <p>cross_entropy = -tf.reduce_mean(y_*tf.log(y))<\/p>\n<\/blockquote>\n\n<p>Why might it be nicer to use a mean instead of a sum? Well, if we sum, then doubling the batch size doubles the cost, and also doubles the magnitude of the gradient. Unless we adjust our learning rate (or use an algorithm that adjusts it for us, like @dga suggested) our training will explode! But if we use a mean, then our learning rate becomes kind of independent of our batch size, which is nice.<\/p>\n\n<p>I'd encourage you to check out Adam (<code>tf.train.AdamOptimizer()<\/code>). It's often more tolerant to fiddling with things than SGD.<\/p>\n"}]}
{"QuestionId":33641922,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-10T23:56:03.110","AcceptedAnswerId":null,"OwnerUserId":3688217.0,"Title":"conditional graph in tensorflow and for loop that accesses tensor size","Body":"<p>First the broad questions: <\/p>\n\n<ul>\n<li>Is it possible to build a conditional graph with tensorflow? <\/li>\n<li>If yes, do the automated gradient-calcuation and the implemented optimizers work with it? <\/li>\n<li>Can I access the shape of a tensor and turn it into an integer for using it with an \"if\" condition and a \"for i in range()\" loop?<\/li>\n<\/ul>\n\n<p>My actual use case is that I want to do a 1D convolution with variable tensor length. For this I first need an if statement that only executes the convolution if the length is greater then one. Then I have a for loop that goes through the tensor for the convolution. The problem is that this code:<\/p>\n\n<pre><\/pre>\n\n<p>doesn't work because the range operator needs an integer. Can I turn this somehow into an integer?<\/p>\n\n<p>In the end I want to train this model with adagrad, either with the automatic differentiation or the already implemented optimiser<\/p>\n\n<hr>\n\n<p>Edit:<\/p>\n\n<p>this is the 1D convolution which will later be the first of two Layers in my Model. The type errors are behind each version of the for loop that triggers one<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33642096","CreationDate":"2015-11-11T00:13:51.273","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I recommend you to write each question differently. Otherwise it will be closed as too broad.<\/p>\n\n<p>I can answer only your 3-rd question. How to programmatically get a shape of the tensor. You are correctly using <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#shape\" rel=\"nofollow noreferrer\">shape to get the shape of the tensor<\/a>, but you still can not get the results before you run the graph (look at my <a href=\"https:\/\/stackoverflow.com\/a\/33640758\/1090562\">explanation here<\/a>).<\/p>\n\n<pre><code>a = tf.truncated_normal([2, 3], mean=0.0, stddev=0.02, dtype=tf.float32, seed=1)\nb = tf.shape(a)\nsess = tf.Session()\nprint sess.run(b) # will give you [2 3]\n<\/code><\/pre>\n\n<p>The ugly way that I have found to get the shape from constants, without running the graph is to do something like (do not really know why would you need it):<\/p>\n\n<pre><code>print a._shape._dims[0]._value\nprint a._shape._dims[1]._value\n<\/code><\/pre>\n\n<p>To get the shape from a variable, you can do this:<\/p>\n\n<pre><code>weights = tf.Variable(tf.random_normal([784, 200], stddev=0.35))\nprint weights.get_shape().as_list()\n<\/code><\/pre>\n\n<p>Another way to access a shape of a Tensor before the evaluation is: <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/framework.md#Tensor.get_shape\" rel=\"nofollow noreferrer\">tf.Tensor.get_shape()<\/a><\/p>\n"}]}
{"QuestionId":33643196,"AnswerCount":1,"Tags":"<virtualenv><tensorflow>","CreationDate":"2015-11-11T02:18:38.893","AcceptedAnswerId":null,"OwnerUserId":3057648.0,"Title":"tensorflow is being installed in anaconda\/lib\/python2.7\/site-packages","Body":"<p>I'm trying to install  in <\/p>\n\n<p>after installing I get the message that tensorflow has been installed in \n<\/p>\n\n<p>how can I get it installed in site-packages of virtual environment?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I'm using ubuntu 14.04.2 LTS<br>\nand following are the sequence of instructions I executed  <\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>I encountered   <\/p>\n\n<p>when trying to execute <\/p>\n\n<p> <\/p>\n\n<p>so I had to follow instuction from <a href=\"https:\/\/askubuntu.com\/questions\/695981\/platform-not-supported-for-tensorflow-on-ubuntu-14-04-2\">https:\/\/askubuntu.com\/questions\/695981\/platform-not-supported-for-tensorflow-on-ubuntu-14-04-2<\/a><\/p>\n\n<p>so after executing  <\/p>\n\n<p><\/p>\n\n<p>I got the output<\/p>\n\n<pre><\/pre>\n\n<p>so as I'm trying to install in virtual environment, I want to install  locally in virtual environment rather than in  <\/p>\n","answers":[{"AnswerId":"34766341","CreationDate":"2016-01-13T12:09:07.490","ParentId":null,"OwnerUserId":"588709","Title":null,"Body":"<p>It seems you already have anaconda installed. anaconda comes with an environment management system, already (see <a href=\"http:\/\/conda.pydata.org\/docs\/using\/envs.html\" rel=\"nofollow noreferrer\">here<\/a>). Having two systems (virtualenv and anaconda) seems to mess up your python paths.<\/p>\n\n<p>If you would like  to use conda environments, then you do not need to install virtualenv. Just create a new conda env for your tensorflow installation.<\/p>\n\n<p>If you still have troubles with your installation, you can have a look at my answer <a href=\"https:\/\/stackoverflow.com\/a\/34746395\/588709\">here<\/a>. It's for Mac OS X, but the steps described should work on any system.<\/p>\n"}]}
{"QuestionId":33643301,"AnswerCount":3,"Tags":"<deep-learning><caffe><conv-neural-network><nvidia-digits>","CreationDate":"2015-11-11T02:29:20.543","AcceptedAnswerId":33645134.0,"OwnerUserId":1868436.0,"Title":"CNN with 2000 classes?","Body":"<p>I need to classify images into one of 2000 classes. <\/p>\n\n<p>I am using the Nvidia DIGITS + caffe (GoogLeNet) and provided 10K samples per class (so a whopping 20 million images, ~1Tb data!). But the data prep (\"create db\") task itself is being estimated to be 102 days and I shudder to think what the actual training time will be if that estimate is correct.<\/p>\n\n<p>What is the best way to approach this challenge? should I break up dataset into 3-4 models? and use them separately? Use a smaller dataset and risk less accuracy? something else?<\/p>\n\n<p>Thanks for helping out a newbie.<\/p>\n","answers":[{"AnswerId":"34992866","CreationDate":"2016-01-25T12:40:02.767","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>I think you can use ImageDataLayer rather than LMDB or levelDB.<\/p>\n"},{"AnswerId":"37806092","CreationDate":"2016-06-14T07:52:13.797","ParentId":null,"OwnerUserId":"2469488","Title":null,"Body":"<p>First, you need to know what scenario you should use <a href=\"https:\/\/stackoverflow.com\/a\/35715139\/2469488\">LMDB\/LevelDB<\/a> and what benefit do you get(especially, training in <a href=\"http:\/\/image-net.org\/tutorials\/cvpr2015\/getting_started.pdf\" rel=\"nofollow noreferrer\">parallel<\/a>.) \nBut what things you will encounter is that creating LMDB is large and slow, and you also need to use <a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/3op2c3\/help_on_speeding_up_caffe_lmdb_creation_for\/\" rel=\"nofollow noreferrer\">SSD instead of HDD<\/a> to decrease time-consuming problem.<\/p>\n\n<p>Second, when you only need to train CNN with small datasets (e.g &lt; 2M images), you can just use <strong>ImageDataLayer<\/strong> same as @kli_nlpr said.<\/p>\n\n<p>Finally, just like @Shai said, the most challenge of all process always is <strong>preparing data<\/strong>, and <strong>check training whether the outcome is what you expected<\/strong>, if it is not you should check data, preparing data again or check training configuration. If you tried all possible solutions, however you still feel slow, and you can tried to change your hardware as <a href=\"http:\/\/on-demand.gputechconf.com\/supercomputing\/2014\/presentation\/SC424-deep-learning-gpu-clusters.pdf\" rel=\"nofollow noreferrer\">GPU cluster<\/a>.<\/p>\n"},{"AnswerId":"33645134","CreationDate":"2015-11-11T06:08:30.587","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You got the training data!! This is the most challenging part of all the process, don't give up on it!<\/p>\n\n<p>Try breaking the task of data creation into steps, and parrallelize them. Training will take long (depending on your GPU) but you should be patient. <\/p>\n"}]}
{"QuestionId":33643882,"AnswerCount":3,"Tags":"<python><reshape><tensorflow>","CreationDate":"2015-11-11T03:39:39.933","AcceptedAnswerId":33644208.0,"OwnerUserId":3067467.0,"Title":"Tensor Reshape No-op in RNN example","Body":"<p>What exactly is this code doing in the method  for the  in the file ? I'm not entirely sure what a shape of the form  means...<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ChSdt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ChSdt.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"33645053","CreationDate":"2015-11-11T06:00:44.640","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#reshape\" rel=\"nofollow\">Documentation vaguely suggests<\/a> the behavior for -1.<\/p>\n\n<blockquote>\n  <p>If shape is the special value [-1], then tensor is flattened and the\n  operation outputs a 1-D tensor with all elements of tensor.<\/p>\n<\/blockquote>\n\n<p>When the shape is an array, where one of the elements is -1, this value is just a convenient placeholder for some integer that might produce the same shape. Note that there might not be such an integer (for example if your starting matrix is <code>[5x5]<\/code> you can not reshape it with <code>[7, -1]<\/code>).<\/p>\n\n<p>Also as you see there can not be two -1, because this make the shape ambiguous. As it was noted, the behavior is similar to <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow\">numpy's reshape<\/a>.<\/p>\n"},{"AnswerId":"33644208","CreationDate":"2015-11-11T04:17:20.873","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The semantics of reshape are similar to the one from numpy:\n<a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow\">http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html<\/a><\/p>\n\n<p>It changes the tensor to have 2-dimensions and the second dimension should have self.state_size elements. E.g. if my tensor has 6 elements and I reshape it to [-1, 2], then the first dimension will have 6 \/ 2 = 3 elements.<\/p>\n"},{"AnswerId":"33644366","CreationDate":"2015-11-11T04:39:49.930","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Rafal's example is great.  The way I remember -1 is that it sets the size of that dimension to whatever's necessary to fit all of the data from the original tensor.  You can only have one -1 in a reshape.<\/p>\n\n<p>If the original tensor is of size a, b, c  (total elements = a<em>b<\/em>c)\nand you resize it to x, y, -1,<\/p>\n\n<p>then the effect will be that the -1 will end up being a<em>b<\/em>c\/(y*z).<\/p>\n\n<p>A 3,3,3 tensor (27 elements) reshaped to 9,3,-1 will actually have a size of 9,3,1 (27 elements)<\/p>\n"}]}
{"QuestionId":33644635,"AnswerCount":2,"Tags":"<python><macos><tensorflow>","CreationDate":"2015-11-11T05:12:37.307","AcceptedAnswerId":null,"OwnerUserId":5549326.0,"Title":"Python unable to import tensorflow to Mac OS X 10.8","Body":"<p>I am trying to install tensorflow on my mac air (specs: OSX 10.8.4, 4GB RAM). While tensorflow installs cleanly, as on the tensorflow.org page, and I am able to do $ source bin\/activate to get a tensorflow prompt, I am not able to import the package in python. This is what I get: <\/p>\n\n<pre><\/pre>\n\n<p>I looked at the common problems section on tensorflow, as well as searched for the ImportError above but have had no luck so far. I wonder if this is a protobuf issue as referenced in  <a href=\"https:\/\/stackoverflow.com\/questions\/33622842\/error-in-python-after-import-tensorflow-typeerror-init-got-an-unexpect\">Error in python after &#39;import tensorflow&#39;: TypeError: __init__() got an unexpected keyword argument &#39;syntax&#39;<\/a> .\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"33685174","CreationDate":"2015-11-13T02:53:05.733","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue appears to be that the binary distribution of TensorFlow is not compatible with Mac OS X 10.8. The relevant part of the error message is this one:<\/p>\n\n<pre><code>Symbol not found: ___sincos_stret\n<\/code><\/pre>\n\n<p>According to <a href=\"https:\/\/github.com\/pioneerspacesim\/pioneer\/issues\/2610\" rel=\"nofollow\">various<\/a> <a href=\"http:\/\/sourceforge.net\/p\/librecad\/bugs\/455\/\" rel=\"nofollow\">sources<\/a>, this is a known issue when using binaries compiled for a newer version of Mac OS X on 10.8 (Mavericks), because the symbol <code>___sincos_stret<\/code> is not defined in the standard libraries for that version of the OS.<\/p>\n\n<p>To address this, you will need to (i) upgrade to Mac OS X 10.9 or later, (ii) follow the instructions to <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"nofollow\">install from source<\/a>, or (iii) use <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation\" rel=\"nofollow\">Docker<\/a>.<\/p>\n"},{"AnswerId":"38107323","CreationDate":"2016-06-29T18:20:25.120","ParentId":null,"OwnerUserId":"6420569","Title":null,"Body":"<p>I had similar issue. I used docker toolbox and was able to run tensorflow. You can follow my <a href=\"https:\/\/stackoverflow.com\/questions\/38020709\/tensorflow-import-error-on-os-x-10-8-5-expected-in\">question<\/a> and my solution.<\/p>\n"}]}
{"QuestionId":33644882,"AnswerCount":0,"Tags":"<image-processing><neural-network><convolution><caffe><conv-neural-network>","CreationDate":"2015-11-11T05:42:10.813","AcceptedAnswerId":null,"OwnerUserId":5389199.0,"Title":"Caffe's configuration for Cohn-Kanade dataset","Body":"<p>I'm trying to make a facial expression recognizer using Caffe and the <a href=\"http:\/\/www.pitt.edu\/~emotion\/ck-spread.htm\" rel=\"nofollow\">Cohn-Kanade database<\/a>. <\/p>\n\n<p>This is my train prototxt configuration:<\/p>\n\n<pre><\/pre>\n\n<p>This is my trainning function, that I got it from lenet example:<\/p>\n\n<pre><\/pre>\n\n<p>I'm only using the Neutral and Surprise faces (if I solve my problem I'll use more emotions). But my net has only 28% of accuracy. I'd like to know the low accuracy has to with a problem in the network configuration, with the logic inside my train function or if my training db is too small?. This is my dataset discription:<\/p>\n\n<p>Train dataset: 56 images of Neutral faces. 60 images of Surprise faces.\nTest dataset: 15 images of Neutral faces. 15 images of Surprise faces.<\/p>\n\n<p>All images are 32x32 and in grayscale.\nAnd my batch_size is 32.<\/p>\n\n<p>Please, can someone help me to known what is my problem?<\/p>\n","answers":[]}
{"QuestionId":33646541,"AnswerCount":7,"Tags":"<python><anaconda><tensorflow>","CreationDate":"2015-11-11T08:15:16.530","AcceptedAnswerId":33698750.0,"OwnerUserId":3169198.0,"Title":"Tensorflow and Anaconda on Ubuntu?","Body":"<p>On my Ubuntu 14.04, I have installed tensorflow, using \"pip\", as specified in the <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md\" rel=\"noreferrer\">Tensorflow Installation instructions<\/a> and I made sure it was working by importing it in python and it did work.<\/p>\n\n<p>Then, I installed Anaconda and it changed my .bashrc file by adding the following line to it:<\/p>\n\n<pre><\/pre>\n\n<p>But because of this change, now it looks into the PATH above, which doesn't contain tensorflow. now I can't import tensorflow in my python code.<\/p>\n\n<p>What is the proper way to extend the $PATH environment variable so that it stays using everything from anaconda2 but it becomes able to import \"tensorflow\"?<\/p>\n","answers":[{"AnswerId":"33698750","CreationDate":"2015-11-13T17:40:22.600","ParentId":null,"OwnerUserId":"3169198","Title":null,"Body":"<p>I solved the problem but in a different way!\nI found a link where the tensorflow.whl files were converted to conda packages, so I went ahead and installed it using the command:<\/p>\n\n<pre><code>conda install -c https:\/\/conda.anaconda.org\/jjhelmus tensorflow\n<\/code><\/pre>\n\n<p>and it worked, since the $PATH points to anaconda packages, I can import it now!<\/p>\n\n<p><a href=\"https:\/\/anaconda.org\/jjhelmus\/tensorflow\">Source is here<\/a><\/p>\n"},{"AnswerId":"33646665","CreationDate":"2015-11-11T08:25:56.920","ParentId":null,"OwnerUserId":"80309","Title":null,"Body":"<p>I had the same problem and decided it was easiest to start over, install Anaconda first and then TensorFlow after that.<\/p>\n"},{"AnswerId":"33646711","CreationDate":"2015-11-11T08:29:31.573","ParentId":null,"OwnerUserId":"1158241","Title":null,"Body":"<p>I suspect that pip is giving you a TensorFlow installation in cpython, not anaconda.<\/p>\n\n<p>How about a virtualenv?<\/p>\n\n<pre><code># Create env\n$ virtualenv --python=\/path\/to\/anaconda \/path\/to\/your\/env\n\n# Activate env\n$ source \/path\/to\/your\/env\/bin\/activate\n\n# Install Tensorflow\n$ pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"49557440","CreationDate":"2018-03-29T13:31:50.217","ParentId":null,"OwnerUserId":"6524987","Title":null,"Body":"<p>Actually, the TensorFlow Official website made every detail of installing.\nThe Operation System <strong>Windows, Mac OS, Ubuntu<\/strong>; the environment with <strong>GPU<\/strong> or just <strong>CPU<\/strong>, every single detail of problems you may come up with.<\/p>\n\n<p>Check this out<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#installing_with_anaconda\" rel=\"nofollow noreferrer\">Installing TensorFlow on Ubuntu with Anaconda<\/a> <\/p>\n\n<p>you will not regret.<\/p>\n\n<p>Once you visit that you may also find like<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/install_windows#installing_with_anaconda\" rel=\"nofollow noreferrer\">Installing TensorFlow on Windows with Anaconda<\/a><\/p>\n"},{"AnswerId":"49400212","CreationDate":"2018-03-21T07:08:07.933","ParentId":null,"OwnerUserId":"8057450","Title":null,"Body":"<p>Install tensorflow from the following command. Conda will take care of the installation process.<\/p>\n\n<p>conda install -c conda-forge tensorflow<\/p>\n"},{"AnswerId":"39687760","CreationDate":"2016-09-25T14:10:14.933","ParentId":null,"OwnerUserId":"4537639","Title":null,"Body":"<p>Since v0.10.0, <code>tensorflow<\/code> is a community maintained conda package in the <code>conda-forge<\/code> channel. Hence, it can be installed directly with the following command:<\/p>\n\n<pre><code>conda install -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>The instructions on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/get_started\/os_setup.html#using-conda\" rel=\"noreferrer\">TensorFlow documentation<\/a> has also been updated.<\/p>\n\n<p>To facilitate future updates, it is probably a good idea to add <code>conda-forge<\/code> channel into your conda config:<\/p>\n\n<pre><code>conda config --add channels conda-forge\n<\/code><\/pre>\n\n<p>In fact, <code>tensorflow=0.10.0rc0<\/code> was recently added onto the <a href=\"https:\/\/anaconda.org\/anaconda\/tensorflow\" rel=\"noreferrer\">Anaconda default channel<\/a> and will be installed instead if the <code>conda-forge<\/code> channel is not specified:<\/p>\n\n<pre><code>conda install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"37168414","CreationDate":"2016-05-11T16:28:58.430","ParentId":null,"OwnerUserId":"4046157","Title":null,"Body":"<p>I solved the problem using this:<\/p>\n\n<pre><code>conda create --name=tensorenv python=3.4\nsource activate tensorenv\n<\/code><\/pre>\n"}]}
{"QuestionId":33648167,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-11T10:01:41.847","AcceptedAnswerId":33648339.0,"OwnerUserId":4165313.0,"Title":"Why do we name variables in Tensorflow?","Body":"<p>In some of the places, I saw the syntax, where variables are initialized with names, sometimes without names. For example:<\/p>\n\n<pre><\/pre>\n\n<p>What is the point of naming the variable  ?<\/p>\n","answers":[{"AnswerId":"51417079","CreationDate":"2018-07-19T07:46:09.757","ParentId":null,"OwnerUserId":"7309971","Title":null,"Body":"<p>In fact, from the aspect to distinguish different variables, we totally can use the python name (the left part of the assignment sign, and we call the name as <code>python name<\/code> to avoid confusion. such as <code>v<\/code> in the following example) to name the variables. However, in the programming process, we usually rebind the python name to other objects (i.e., the op in Tensorflow), for example, <\/p>\n\n<pre><code>v = tf.get_variable(\"v1\", [3], initializer = tf.zeros_initializer)\nv = tf.get_variable(\"v2\", [5], initializer = tf.zeros_initializer)\n<\/code><\/pre>\n\n<p>Firstly, the python name <code>v<\/code> bind the tensor form first line (<code>tf.get_variable(\"v1\", [3], initializer = tf.zeros_initializer)<\/code>). Then,the <code>v<\/code> rebind the tensor from the second line (<code>tf.get_variable(\"v2\", [5], initializer = tf.zeros_initializer<\/code>)) and didn't bind the first tensor anymore. If we didn't give the tensorflow attribute name <code>v1<\/code> and <code>v2<\/code>, how can we identify the tensor from the first line?<\/p>\n"},{"AnswerId":"46419671","CreationDate":"2017-09-26T06:56:16.443","ParentId":null,"OwnerUserId":"8089695","Title":null,"Body":"<p>Consider the following use case code and its output<\/p>\n\n<pre><code>def f():\n    a = tf.Variable(np.random.normal(), dtype = tf.float32, name = 'test123')\n\ndef run123():\n    f()\n    init = tf.global_variables_initializer()\n    with tf.Session() as sess123:\n        sess123.run(init)\n        print(sess123.run(fetches = ['test123:0']))\n        print(sess123.run(fetches = [a]))\n\nrun123()\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<blockquote>\n  <p>[0.10108799]<\/p>\n  \n  <p>NameError                                 Traceback (most recent call\n  last)  in ()\n       10         print(sess123.run(fetches = [a]))\n       11 \n  ---> 12 run123()<\/p>\n  \n  <p> in run123()\n        8         sess123.run(init)\n        9         print(sess123.run(fetches = ['test123:0']))\n  ---> 10         print(sess123.run(fetches = [a]))\n       11 \n       12 run123()<\/p>\n  \n  <p>NameError: name 'a' is not defined<\/p>\n<\/blockquote>\n\n<p>The 'a', as defined in the scope of f(), not available outside of its scope, i.e in run123(). But the default graph has to refer to them with something, so that the default graph can be referenced to, as needed, across various scopes and that is when its <strong>name<\/strong> comes handy.<\/p>\n"},{"AnswerId":"33648339","CreationDate":"2015-11-11T10:10:40.913","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The <code>name<\/code> parameter is optional (you can create variables and constants with or without it), and the variable you use in your program does not depend on it. Names can be helpful in a couple of places:<\/p>\n\n<p><strong>When you want to save or restore your variables<\/strong> (you can <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/state_ops.md#Saver\">save them to a binary file<\/a> after the computation). From <a href=\"http:\/\/tensorflow.org\/how_tos\/variables\/index.md\">docs<\/a>:<\/p>\n\n<blockquote>\n  <p>By default, it uses the value of the Variable.name property for each\n  variable<\/p>\n<\/blockquote>\n\n<pre><code>matrix_1 = tf.Variable([[1, 2], [2, 3]], name=\"v1\")\nmatrix_2 = tf.Variable([[3, 4], [5, 6]], name=\"v2\")\ninit = tf.initialize_all_variables()\n\nsaver = tf.train.Saver()\n\nsess = tf.Session()\nsess.run(init)\nsave_path = saver.save(sess, \"\/model.ckpt\")\nsess.close()\n<\/code><\/pre>\n\n<p>Nonetheless you have variables <code>matrix_1<\/code>, <code>matrix_2<\/code> they are saves as <code>v1<\/code>, <code>v2<\/code> in the file.<\/p>\n\n<p><strong>Also names are used in TensorBoard to nicely show names of edges<\/strong>. You can even <a href=\"http:\/\/tensorflow.org\/how_tos\/graph_viz\/index.md#name_scoping_and_nodes\">group them by using the same scope<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.name_scope('hidden') as scope:\n  a = tf.constant(5, name='alpha')\n  W = tf.Variable(tf.random_uniform([1, 2], -1.0, 1.0), name='weights')\n  b = tf.Variable(tf.zeros([1]), name='biases')\n<\/code><\/pre>\n"},{"AnswerId":"37084810","CreationDate":"2016-05-07T05:21:29.440","ParentId":null,"OwnerUserId":"4794308","Title":null,"Body":"<p>You can imagine Python namespace and TensorFlow namespace as two parallel universes. Names in TensorFlow space are actually the \"real\" attributes belonging to any TensorFlow variables, while names in Python space are just temporary pointers pointing to TensorFlow variables during this run of your script. That is the reason why when saving and restoring variables, only TensorFlow names are used, because the Python namespace no longer exists after script being terminated, but Tensorflow namespace is still there in your saved files.<\/p>\n"}]}
{"QuestionId":33648322,"AnswerCount":7,"Tags":"<python><tensorflow>","CreationDate":"2015-11-11T10:09:45.163","AcceptedAnswerId":33862534.0,"OwnerUserId":4952564.0,"Title":"Tensorflow image reading & display","Body":"<p>I've got a bunch of images in a format similar to Cifar10 (binary file,  bytes per image), one image after another (<a href=\"http:\/\/cs.stanford.edu\/~acoates\/stl10\/\" rel=\"noreferrer\">STL-10 dataset<\/a>). The file I'm opening has 138MB.<\/p>\n\n<p>I tried to read &amp; check the contents of the Tensors containing the images to be sure that the reading is done right, however I have two questions - <\/p>\n\n<ol>\n<li>Does the  load the whole file, however just provide inputs one at a time? Since reading the first  bytes should be relatively fast. However, the code takes about two minutes to run. <\/li>\n<li>How to get the actual image contents in a displayable format, or display them internally to validate that the images are read well? I did , however the result is empty.<\/li>\n<\/ol>\n\n<p>The code is below:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I'm running this on my CPU, if that adds anything.<\/p>\n\n<p>EDIT: I found the pure TensorFlow solution thanks to Rosa. Apparently, when using the , in order to see the results, you need to initialize the queue runners. \nThe only required thing to add to the code above is the second line from below:<\/p>\n\n<pre><\/pre>\n\n<p>Afterwards, the image in the  can be displayed with . I hope this helps someone. If you have any further questions, feel free to ask me or check the link in Rosa's answer.<\/p>\n","answers":[{"AnswerId":"52565820","CreationDate":"2018-09-29T06:39:18.587","ParentId":null,"OwnerUserId":"4988503","Title":null,"Body":"<p>You can use tf.keras API.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.keras.preprocessing.image import load_img, array_to_img\n\ntf.enable_eager_execution()\n\nimg = load_img(\"example.png\")\nimg = tf.convert_to_tensor(np.asarray(img))\nimage = tf.image.resize_images(img, (800, 800))\nto_img = array_to_img(image)\nto_img.show()\n<\/code><\/pre>\n"},{"AnswerId":"33649013","CreationDate":"2015-11-11T10:51:52.630","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>After speaking with you in the comments, I believe that you can just do this using numpy\/scipy. The ideas is to read the image in the <code>numpy<\/code> 3d-array and feed it into the variable.<\/p>\n\n<pre><code>from scipy import misc\nimport tensorflow as tf\n\nimg = misc.imread('01.png')\nprint img.shape    # (32, 32, 3)\n\nimg_tf = tf.Variable(img)\nprint img_tf.get_shape().as_list()  # [32, 32, 3]\n<\/code><\/pre>\n\n<p>Then you can run your graph:<\/p>\n\n<pre><code>init = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\nim = sess.run(img_tf)\n<\/code><\/pre>\n\n<p>and verify that it is the same:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nfig = plt.figure()\nfig.add_subplot(1,2,1)\nplt.imshow(im)\nfig.add_subplot(1,2,2)\nplt.imshow(img)\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/k5xkC.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/k5xkC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>P.S.<\/strong> you mentioned: <code>Since it's supposed to parallelize reading, it seems useful to know.<\/code>. To which I can say that rarely in data-analysis reading of the data is the bottleneck. Most of your time you will spend training your model.<\/p>\n"},{"AnswerId":"33766867","CreationDate":"2015-11-17T20:58:27.023","ParentId":null,"OwnerUserId":"1309552","Title":null,"Body":"<p>According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\" rel=\"nofollow noreferrer\">documentation<\/a> you can decode JPEG\/PNG images.<\/p>\n\n<p>It should be something like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nfilenames = ['\/image_dir\/img.jpg']\nfilename_queue = tf.train.string_input_producer(filenames)\n\nreader = tf.WholeFileReader()\nkey, value = reader.read(filename_queue)\n\nimages = tf.image.decode_jpeg(value, channels=3)\n<\/code><\/pre>\n\n<p>You can find a bit more info <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/ik7hSLIKEBM\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"33862534","CreationDate":"2015-11-23T01:39:14.943","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>Just to give a complete answer:<\/p>\n\n<pre><code>filename_queue = tf.train.string_input_producer(['\/Users\/HANEL\/Desktop\/tf.png']) #  list of files to read\n\nreader = tf.WholeFileReader()\nkey, value = reader.read(filename_queue)\n\nmy_img = tf.image.decode_png(value) # use png or jpg decoder based on your files.\n\ninit_op = tf.global_variables_initializer()\nwith tf.Session() as sess:\n  sess.run(init_op)\n\n  # Start populating the filename queue.\n\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(coord=coord)\n\n  for i in range(1): #length of your filename list\n    image = my_img.eval() #here is your image Tensor :) \n\n  print(image.shape)\n  Image.fromarray(np.asarray(image)).show()\n\n  coord.request_stop()\n  coord.join(threads)\n<\/code><\/pre>\n\n<p>Or if you have a directory of images you can add them all via <a href=\"https:\/\/github.com\/HamedMP\/ImageFlow\" rel=\"noreferrer\">this Github source file<\/a><\/p>\n\n<p>@mttk and @salvador-dali: I hope it is what you need<\/p>\n"},{"AnswerId":"42367882","CreationDate":"2017-02-21T12:53:05.333","ParentId":null,"OwnerUserId":"5015616","Title":null,"Body":"<p>Load names with tf.train.match_filenames_once get the number of files to iterate over with tf.size\nopen session and enjoy ;-)<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib;\nfrom PIL import Image\n\nmatplotlib.use('Agg')\nimport matplotlib.pyplot as plt\n\n\nfilenames = tf.train.match_filenames_once('.\/images\/*.jpg')\ncount_num_files = tf.size(filenames)\nfilename_queue = tf.train.string_input_producer(filenames)\n\nreader=tf.WholeFileReader()\nkey,value=reader.read(filename_queue)\nimg = tf.image.decode_jpeg(value)\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n    num_files = sess.run(count_num_files)\n    for i in range(num_files):\n        image=img.eval()\n        print(image.shape)\n        Image.fromarray(np.asarray(image)).save('te.jpeg')\n<\/code><\/pre>\n"},{"AnswerId":"44441850","CreationDate":"2017-06-08T17:10:23.687","ParentId":null,"OwnerUserId":"5595995","Title":null,"Body":"<p>I used CIFAR10 format instead of STL10 and code came out like<\/p>\n\n<pre><code>filename_queue = tf.train.string_input_producer(filenames)\nread_input = read_cifar10(filename_queue)\nwith tf.Session() as sess:       \n    tf.train.start_queue_runners(sess=sess)\n    result = sess.run(read_input.uint8image)        \nimg = Image.fromarray(result, \"RGB\")    \nimg.save('my.jpg')\n<\/code><\/pre>\n\n<p>The snippet is identical with mttk and Rosa Gronchi, but Somehow I wasn't able to show the image during run-time, so I saved as the JPG file.     <\/p>\n"},{"AnswerId":"37574145","CreationDate":"2016-06-01T16:28:45.967","ParentId":null,"OwnerUserId":"3817279","Title":null,"Body":"<p>(Can't comment, not enough reputation, but here is a modified version that worked for me)<\/p>\n\n<p>To @HamedMP error about the <code>No default session is registered<\/code> you can use <code>InteractiveSession<\/code> to get rid of this error:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/client.html#InteractiveSession\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/client.html#InteractiveSession<\/a><\/p>\n\n<p>And to @NumesSanguis issue with <code>Image.show<\/code>, you can use the regular PIL <code>.show()<\/code> method because <code>fromarray<\/code> returns an image object. <\/p>\n\n<p>I do both below (note I'm using JPEG instead of PNG):<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom PIL import Image\n\nfilename_queue = tf.train.string_input_producer(['my_img.jpg']) #  list of files to read\n\nreader = tf.WholeFileReader()\nkey, value = reader.read(filename_queue)\n\nmy_img = tf.image.decode_jpeg(value) # use png or jpg decoder based on your files.\n\ninit_op = tf.initialize_all_variables()\nsess = tf.InteractiveSession()\nwith sess.as_default():\n    sess.run(init_op)\n\n# Start populating the filename queue.\n\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(coord=coord)\n\nfor i in range(1): #length of your filename list\n  image = my_img.eval() #here is your image Tensor :) \n\nImage.fromarray(np.asarray(image)).show()\n\ncoord.request_stop()\ncoord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":33648737,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-11T10:35:07.100","AcceptedAnswerId":33948927.0,"OwnerUserId":2569501.0,"Title":"Is TensorFlow suitable for Recommendation Systems","Body":"<p>I have read blogpost about TensorFlow is being open sourced.<\/p>\n\n<p>In the tutorials and the examples on the TensorFlow website, I see that they are mostly classification problems. (e.g. given an image, classify the number written in it)<\/p>\n\n<p>I am curious about it the software also suitable for solving problems in recommendation system?\nFor example, is it good for solving problems on collaborative filtering \/ content-based filtering?<\/p>\n","answers":[{"AnswerId":"33948927","CreationDate":"2015-11-27T00:50:58.800","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>Tensorflow is great for deep learning, or training large neural nets. Although, it can be used for several other mathematical applications such as PDEs, various classifiers, recommendation systems etc, there doesn't seem to have a lot of support for them as yet.<\/p>\n\n<p>This <a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/3aeazf\/recommender_system_libraries_in_python\/\" rel=\"noreferrer\">reddit thread<\/a> might be a good place to start for searching libraries which are centred around recommendation systems. <\/p>\n"}]}
{"QuestionId":33648796,"AnswerCount":2,"Tags":"<torch>","CreationDate":"2015-11-11T10:38:41.870","AcceptedAnswerId":33660580.0,"OwnerUserId":1886138.0,"Title":"Add my custom loss function to torch","Body":"<p>I want to add a loss function to torch that calculates the edit distance between predicted and target values.\nIs there an easy way to implement this idea?\nOr do I have to write my own class with backward and forward functions?<\/p>\n","answers":[{"AnswerId":"44335476","CreationDate":"2017-06-02T18:40:09.900","ParentId":null,"OwnerUserId":"4633087","Title":null,"Body":"<p>Just to add to the accepted answer, you have to be careful that the loss function you define (edit distance in your case) is differentiable with respect to the network parameters.<\/p>\n"},{"AnswerId":"33660580","CreationDate":"2015-11-11T21:55:55.757","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>If your criterion can be represented as a composition of existing modules and criteria, it's a good idea to simply construct such composition using containers. The only problem is that standard containers are designed to work with modules only, not criteria. The difference is in <code>:forward<\/code> method signature:<\/p>\n\n<pre><code>module:forward(input)\ncriterion:forward(input, target)\n<\/code><\/pre>\n\n<p>Luckily, we are free to define our own container which is able work with criteria too. For example, sequential:<\/p>\n\n<pre><code>local GeneralizedSequential, _ = torch.class('nn.GeneralizedSequential', 'nn.Sequential')\n\nfunction GeneralizedSequential:forward(input, target)\n    return self:updateOutput(input, target)\nend\n\nfunction GeneralizedSequential:updateOutput(input, target)\n    local currentOutput = input\n    for i=1,#self.modules do\n        currentOutput = self.modules[i]:updateOutput(currentOutput, target)\n    end\n    self.output = currentOutput\n    return currentOutput\nend\n<\/code><\/pre>\n\n<p>Below is an illustration of how to implement <code>nn.CrossEntropyCriterion<\/code> having this generalized sequential container:<\/p>\n\n<pre><code>function MyCrossEntropyCriterion(weights)\n    criterion = nn.GeneralizedSequential()\n    criterion:add(nn.LogSoftMax())\n    criterion:add(nn.ClassNLLCriterion(weights))\n    return criterion\nend\n<\/code><\/pre>\n\n<p>Check whether everything is correct:<\/p>\n\n<pre><code>output = torch.rand(3,3)\ntarget = torch.Tensor({1, 2, 3})\n\nmycrit = MyCrossEntropyCriterion()\n-- print(mycrit)\nprint(mycrit:forward(output, target))\nprint(mycrit:backward(output, target))\n\ncrit = nn.CrossEntropyCriterion()\n-- print(crit)\nprint(crit:forward(output, target))\nprint(crit:backward(output, target))\n<\/code><\/pre>\n"}]}
{"QuestionId":33650653,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-11-11T12:34:02.173","AcceptedAnswerId":33676391.0,"OwnerUserId":639973.0,"Title":"extracting features from all bounding boxes of Fast R-CNN","Body":"<p>I just installed Fast RCNN and have run the demo,<\/p>\n\n<p>and I came to wonder if it's possible to extract features from all bounding boxes in the image (and do this for the entire dataset).<\/p>\n\n<p>For example, if Fast RCNN detects cat, dog, and a car from an image, <\/p>\n\n<p>I'd like to extract separate CNN features for each of cat, dog, and car.<\/p>\n\n<p>And do this for tens of thousands of images.<\/p>\n\n<p>The feature extraction example on Fast RCNN's Github (<a href=\"https:\/\/github.com\/rbgirshick\/caffe-fast-rcnn\/tree\/master\/examples\/feature_extraction\" rel=\"nofollow\">https:\/\/github.com\/rbgirshick\/caffe-fast-rcnn\/tree\/master\/examples\/feature_extraction<\/a>) seems to be the replica of feature extraction using caffe for the entire image, not each bounding box.<\/p>\n\n<p>Could anyone help me on this?<\/p>\n\n<p>UPDATED:<\/p>\n\n<p>Apparently, feature extraction for each bounding box is done in the following part of the code from <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\/blob\/master\/lib\/fast_rcnn\/test.py\" rel=\"nofollow\">https:\/\/github.com\/rbgirshick\/fast-rcnn\/blob\/master\/lib\/fast_rcnn\/test.py<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to figure out how to tweak this to save the features, as we do with Caffe for features of the entire images, which are saved to a mdb file.<\/p>\n","answers":[{"AnswerId":"33676391","CreationDate":"2015-11-12T16:26:40.190","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p><strong>UPDATE<\/strong><\/p>\n\n<p>During the process of determining the right bounding boxes, Fast-RCNN extracts CNN features from a high (~800-2000) number of image regions, called <em>object proposals<\/em>. These regions are obtained through different algorithms, typically <a href=\"http:\/\/koen.me\/research\/selectivesearch\/\" rel=\"nofollow\">selective search<\/a>. After this computation, it uses those features to recognize the \"right\" proposals and find out the \"right\" bounding box. This is called <em>bounding box regression<\/em>.<\/p>\n\n<p>Of course Fast-RCNN optimizes this process, but still has to extract CNN features features from many more regions than the ones related with the object of interest. <\/p>\n\n<p>Shortly, if you were to save the variable <code>blobs_out<\/code> in the code snap you pasted, you will save the features relative to <strong>all<\/strong> the object proposals, including the \"wrong\" proposals. But you can save all that and then try to prune and retrieve only the desired ones. To save the features, just use <code>pickle.dump()<\/code>.<\/p>\n\n<p>Look at the end of the <code>test_net<\/code> function, <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\/blob\/master\/lib\/fast_rcnn\/test.py#L324\" rel=\"nofollow\">here<\/a>. The <code>nms_dets<\/code> variable seems to store the final boxes. There may be a way to take the <code>blobs_out<\/code> you stored and throw the undesired features off, but it doesn't seem so straightforward.<\/p>\n\n<hr>\n\n<p>The simplest solution I'm able to think about is as follows.<\/p>\n\n<p>Let's Fast-RCNN compute the final bounding boxes. Then, extract the relative image patches, with something like the following (I'm assuming Python):<\/p>\n\n<pre><code>img = cv2.imread('\/path\/to\/image')\nfor bbox in bboxes_list:\n    x0, y0, x1, y1 = bbox\n    cut = img[y0:y1, x0:x1]\n    extract_cnn_features(cut)\n<\/code><\/pre>\n\n<p>The feature extraction is identical to the entire image case:<\/p>\n\n<pre><code> net = Caffe.NET('deploy.prototxt', 'caffemodel', caffe.TEST)\n # preprocess input\n net.blobs['data'].data[...] = net_input\n net.forward()\n feats = net.blobs['my_layer'].data.copy()\n<\/code><\/pre>\n\n<p>Of course this method is computationally expensive, since you are basically compute <em>twice<\/em> the CNN features. It depends on your requirements about speed and the size of the CNN models.<\/p>\n"}]}
{"QuestionId":33650678,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2015-11-11T12:35:54.793","AcceptedAnswerId":null,"OwnerUserId":2795395.0,"Title":"pymc3: Disaster example with deterministic switchpoint function","Body":"<p>I'm trying to reproduce coal mining example with deterministic function for switchpoint instead of using theano's switch function. Code:<\/p>\n\n<pre><\/pre>\n\n<p>But this code rise an error:<\/p>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- KeyError                                  Traceback (most recent call\n  last) c:\\program files\\git\\theano\\theano\\tensor\\type.py in\n  dtype_specs(self)\n      266                 'complex64': (complex, 'theano_complex64', 'NPY_COMPLEX64')\n  --> 267             }[self.dtype]\n      268         except KeyError:<\/p>\n  \n  <p>KeyError: 'object'<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>TypeError                                 Traceback (most recent call\n  last) c:\\program files\\git\\theano\\theano\\tensor\\basic.py in\n  constant_or_value(x, rtype, name, ndim, dtype)\n      407             rval = rtype(\n  --> 408                 TensorType(dtype=x_.dtype, broadcastable=bcastable),\n      409                 x_.copy(),<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\type.py in <strong>init<\/strong>(self,\n  dtype, broadcastable, name, sparse_grad)\n       49         self.broadcastable = tuple(bool(b) for b in broadcastable)\n  ---> 50         self.dtype_specs()  # error checking is done there\n       51         self.name = name<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\type.py in dtype_specs(self)\n      269             raise TypeError(\"Unsupported dtype for %s: %s\"\n  --> 270                             % (self.<strong>class<\/strong>.<strong>name<\/strong>, self.dtype))\n      271 <\/p>\n  \n  <p>TypeError: Unsupported dtype for TensorType: object<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>TypeError                                 Traceback (most recent call\n  last) c:\\program files\\git\\theano\\theano\\tensor\\basic.py in\n  as_tensor_variable(x, name, ndim)\n      201     try:\n  --> 202         return constant(x, name=name, ndim=ndim)\n      203     except TypeError:<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\basic.py in constant(x,\n  name, ndim, dtype)\n      421     ret = constant_or_value(x, rtype=TensorConstant, name=name, ndim=ndim,\n  --> 422                             dtype=dtype)\n      423 <\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\basic.py in\n  constant_or_value(x, rtype, name, ndim, dtype)\n      416     except Exception:\n  --> 417         raise TypeError(\"Could not convert %s to TensorType\" % x, type(x))\n      418 <\/p>\n  \n  <p>TypeError: ('Could not convert FromFunctionOp{rate1} to TensorType',\n  )<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>AsTensorError                             Traceback (most recent call\n  last)  in ()\n       14     mu2 = pymc3.Exponential('mu2',lam=1.)\n       15     #rate1 = pymc3.switch(switchpoint >= np.arange(len(data)), mu1,mu2)\n  ---> 16     disasters=pymc3.Poisson('disasters', mu=rate1, observed = data)<\/p>\n  \n  <p>C:\\Users\\User\\Anaconda3\\lib\\site-packages\\pymc3\\distributions\\distribution.py\n  in <strong>new<\/strong>(cls, name, *args, **kwargs)\n       19         if isinstance(name, str):\n       20             data = kwargs.pop('observed', None)\n  ---> 21             dist = cls.dist(*args, **kwargs)\n       22             return model.Var(name, dist, data)\n       23         elif name is None:<\/p>\n  \n  <p>C:\\Users\\User\\Anaconda3\\lib\\site-packages\\pymc3\\distributions\\distribution.py\n  in dist(cls, *args, **kwargs)\n       32     def dist(cls, *args, **kwargs):\n       33         dist = object.<strong>new<\/strong>(cls)\n  ---> 34         dist.<strong>init<\/strong>(*args, **kwargs)\n       35         return dist\n       36 <\/p>\n  \n  <p>C:\\Users\\User\\Anaconda3\\lib\\site-packages\\pymc3\\distributions\\discrete.py\n  in <strong>init<\/strong>(self, mu, *args, **kwargs)\n      185         super(Poisson, self).<strong>init<\/strong>(*args, **kwargs)\n      186         self.mu = mu\n  --> 187         self.mode = floor(mu).astype('int32')\n      188 \n      189     def random(self, point=None, size=None, repeat=None):<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\gof\\op.py in <strong>call<\/strong>(self,\n  *inputs, **kwargs)\n      598         \"\"\"\n      599         return_list = kwargs.pop('return_list', False)\n  --> 600         node = self.make_node(*inputs, **kwargs)\n      601 \n      602         if config.compute_test_value != 'off':<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\elemwise.py in\n  make_node(self, *inputs)\n      540         using DimShuffle.\n      541         \"\"\"\n  --> 542         inputs = list(map(as_tensor_variable, inputs))\n      543         shadow = self.scalar_op.make_node(\n      544             *[get_scalar_type(dtype=i.type.dtype).make_variable()<\/p>\n  \n  <p>c:\\program files\\git\\theano\\theano\\tensor\\basic.py in\n  as_tensor_variable(x, name, ndim)\n      206         except Exception:\n      207             str_x = repr(x)\n  --> 208         raise AsTensorError(\"Cannot convert %s to TensorType\" % str_x, type(x))\n      209 \n      210 # this has a different name, because _as_tensor_variable is the<\/p>\n  \n  <p>AsTensorError: ('Cannot convert FromFunctionOp{rate1} to TensorType',\n  )<\/p>\n<\/blockquote>\n\n<p>How i handle this?<\/p>\n\n<p>The second thing - when i'm using the pymc3.switch function like this:<\/p>\n\n<pre><\/pre>\n\n<p>And next try to sample:<\/p>\n\n<pre><\/pre>\n\n<p>I get an error:<\/p>\n\n<pre><\/pre>\n\n<p>Being simple analyst, should i learn all this stuff about theano to being able  to work with my statistical problems? Is a new mcmc sampler with gradient feature is only one thing that should motivates me to switch from pymc2 to pymc3? <\/p>\n","answers":[{"AnswerId":"33747503","CreationDate":"2015-11-17T01:09:54.197","ParentId":null,"OwnerUserId":"2708667","Title":null,"Body":"<p>For your first question, it looks like you're trying to pass a theano function as a variable. You need to call the function with the other variables as arguments, which will then return a theano variable. Try changing your line to<\/p>\n\n<pre><code>disasters=pymc3.Poisson('disasters', mu=rate1(switchpoint, mu1, mu2), observed = data)\n<\/code><\/pre>\n\n<p>I couldn't reproduce the error in your second part; the sampling worked just fine for me.<\/p>\n"}]}
{"QuestionId":33651732,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2015-11-11T13:37:31.440","AcceptedAnswerId":33656793.0,"OwnerUserId":4075155.0,"Title":"TypeError: rank mismatch between coding and true distributions","Body":"<p>I'm trying to adapt <a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/user\/tutorial.html#run-the-mnist-example\" rel=\"nofollow\">Lasagne tutorial<\/a> to an LSTM model. This is how my code is atm:<\/p>\n\n<pre><\/pre>\n\n<p>I can build my model without errors. But after inserting the code related to Loss evaluation I get this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33656793","CreationDate":"2015-11-11T17:56:26.297","ParentId":null,"OwnerUserId":"4075155","Title":null,"Body":"<p>My output layer should have <strong>batchsize * seqlen<\/strong> as parameter.<\/p>\n\n<p>Not <strong>batchsize, seqlen<\/strong>.<\/p>\n"}]}
{"QuestionId":33651810,"AnswerCount":3,"Tags":"<amazon-ec2><tensorflow>","CreationDate":"2015-11-11T13:41:25.023","AcceptedAnswerId":33686846.0,"OwnerUserId":2448716.0,"Title":"The minimum required Cuda capability is 3.5","Body":"<p>After installing TensorFlow and its dependencies on a g2.2xlarge EC2 instance I tried to run an MNIST example from the getting started page: <\/p>\n\n<pre><\/pre>\n\n<p>But I get the following warning:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a hard requirement? Any chance I could comment that check out in a fork of TensorFlow? It would be super nice to be able to train models in AWS.<\/p>\n","answers":[{"AnswerId":"33686846","CreationDate":"2015-11-13T06:04:44.437","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Currently only GPUs with compute capability >= 3.5 are officially supported. However, GitHub user @infojunkie has offered <a href=\"https:\/\/gist.github.com\/infojunkie\/cb6d1a4e8bf674c6e38e\">a patch<\/a> that makes it possible to use TensorFlow with a GPU with compute capability 3.0.<\/p>\n\n<p>The official fix is in development. Meanwhile, check out the discussion on the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25\">GitHub issue<\/a> for adding this support.<\/p>\n"},{"AnswerId":"40899749","CreationDate":"2016-11-30T22:53:20.053","ParentId":null,"OwnerUserId":"4501221","Title":null,"Body":"<p>There is a simple trick. You don't even have to build TF from sources.<\/p>\n\n<p>In the file <code>tensorflow\\python\\_pywrap_tensorflow.pyd<\/code> there are two occurences of regex <code>3\\.5.*5\\.2<\/code>. Just replace both <code>3.5<\/code> with <code>3.0<\/code>.<\/p>\n\n<p>Tested on Windows 10, Anaconda 4.2.13, Python 3.5.2, TensorFlow 0.12, CUDA 8, NVidia GTX 660m (CUDA cap. 3.0).<\/p>\n"},{"AnswerId":"35110686","CreationDate":"2016-01-31T05:32:16.690","ParentId":null,"OwnerUserId":"3903578","Title":null,"Body":"<p>There is a section in the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/get_started\/os_setup.html#install_cuda\">official installation page<\/a> that guides you to enable Cuda 3, but you need to build Tensorflow from source. <\/p>\n\n<pre><code>$ TF_UNOFFICIAL_SETTING=1 .\/configure\n\n# Same as the official settings above\n\nWARNING: You are configuring unofficial settings in TensorFlow. Because some\nexternal libraries are not backward compatible, these settings are largely\nuntested and unsupported.\n\nPlease specify a list of comma-separated Cuda compute capabilities you want to\nbuild with. You can find the compute capability of your device at:\nhttps:\/\/developer.nvidia.com\/cuda-gpus.\nPlease note that each additional compute capability significantly increases\nyour build time and binary size. [Default is: \"3.5,5.2\"]: 3.0\n\nSetting up Cuda include\nSetting up Cuda lib64\nSetting up Cuda bin\nSetting up Cuda nvvm\nConfiguration finished\n<\/code><\/pre>\n"}]}
{"QuestionId":33654754,"AnswerCount":4,"Tags":"<matrix><reshape><tensorflow>","CreationDate":"2015-11-11T16:08:49.320","AcceptedAnswerId":33655736.0,"OwnerUserId":2701950.0,"Title":"How can I change the shape of a variable in TensorFlow?","Body":"<p><a href=\"http:\/\/www.tensorflow.org\/how_tos\/variables\/index.md#creation\">TensorFlow tutorial<\/a> says that at creation time we need to specify the shape of tensors.  That shape automatically becomes the shape of the tensor.  It also says that TensorFlow provides advanced mechanisms to reshape variables. How can I do that? Any code example?<\/p>\n","answers":[{"AnswerId":"45284621","CreationDate":"2017-07-24T15:32:55.860","ParentId":null,"OwnerUserId":"2567512","Title":null,"Body":"<pre><code>tf.Variable(tf.placeholder(tf.float32))\n<\/code><\/pre>\n\n<p>is not valid at tensorflow 1.2.1<\/p>\n\n<p>in python shell:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.Variable(tf.placeholder(tf.float32))\n<\/code><\/pre>\n\n<p>You will get:<\/p>\n\n<pre><code>ValueError: initial_value must have a shape specified: Tensor(\"Placeholder:0\", dtype=float32)\n<\/code><\/pre>\n\n<p>Update: if you add <code>validate_shape=False<\/code>, there will be no error.<\/p>\n\n<pre><code>tf.Variable(tf.placeholder(tf.float32), validate_shape=False)\n<\/code><\/pre>\n\n<p>if <code>tf.py_func<\/code> matches your requirement:<\/p>\n\n<pre><code>def init():\n    return numpy.random.rand(2,3)\na = tf.pyfun(init, [], tf.float32)\n<\/code><\/pre>\n\n<p>You can create variable that has any shape by passing your own init function.<\/p>\n\n<p>Another way:<\/p>\n\n<pre><code>var = tf.get_varible('my-name', initializer=init, shape=(1,1))\n<\/code><\/pre>\n\n<p>You can pass <code>tf.constant<\/code> or any <code>init<\/code> function that returns numpy array. The shape provided will not be validated. The output shape is your real data shape.<\/p>\n"},{"AnswerId":"33662680","CreationDate":"2015-11-12T00:56:02.150","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/state_ops.md#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> class is the recommended way to create variables, but it restricts your ability to change the shape of the variable once it has been created.<\/p>\n\n<p>If you need to change the shape of a variable, you can do the following (e.g. for a 32-bit floating point tensor):<\/p>\n\n<pre><code>var = tf.Variable(tf.placeholder(tf.float32))\n# ...\nnew_value = ...  # Tensor or numpy array.\nchange_shape_op = tf.assign(var, new_value, validate_shape=False)\n# ...\nsess.run(change_shape_op)  # Changes the shape of `var` to new_value's shape.\n<\/code><\/pre>\n\n<p>Note that this feature is not in the documented public API, so it is subject to change. If you do find yourself needing to use this feature, let us know, and we can investigate a way to support it moving forward. <\/p>\n"},{"AnswerId":"33655736","CreationDate":"2015-11-11T16:59:16.700","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/array_ops#Shapes_and_Shaping\" rel=\"nofollow noreferrer\">shapes-and-shaping<\/a> from TensorFlow documentation. It describes different shape transformations available.<\/p>\n\n<p>The most common function is probably <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#reshape\" rel=\"nofollow noreferrer\">tf.reshape<\/a>, which is similar to its numpy equivalent. It allows you to specify any shape that you want as long as the number of elements stays the same. There are some examples available in the documentation.<\/p>\n"},{"AnswerId":"33659560","CreationDate":"2015-11-11T20:48:51.990","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#AUTOGENERATED-shapes-and-shaping\" rel=\"nofollow\">Documentation shows<\/a> methods for reshaping. They are:<\/p>\n\n<ul>\n<li>reshape<\/li>\n<li>squeeze (removes dimensions of size 1 from the shape of a tensor)<\/li>\n<li>expand_dims (adds dimensions of size 1)<\/li>\n<\/ul>\n\n<p>as well as bunch of methods to get <code>shape<\/code>, <code>size<\/code>, <code>rank<\/code> of your tensor. Probably the most used is <code>reshape<\/code> and here is a code example with a couple of edge cases (-1):<\/p>\n\n<pre><code>import tensorflow as tf\n\nv1 = tf.Variable([\n    [1, 2, 3, 4],\n    [5, 6, 7, 8],\n    [9, 10, 11, 12]\n])\nv2 = tf.reshape(v1, [2, 6])\nv3 = tf.reshape(v1, [2, 2, -1])\nv4 = tf.reshape(v1, [-1])\n# v5 = tf.reshape(v1, [2, 4, -1]) will fail, because you can not find such an integer for -1\nv6 = tf.reshape(v1, [1, 4, 1, 3, 1])\nv6_shape = tf.shape(v6)\nv6_squeezed = tf.squeeze(v6)\nv6_squeezed_shape = tf.shape(v6_squeezed)\n\ninit = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run(init)\na, b, c, d, e, f, g = sess.run([v2, v3, v4, v6, v6_shape, v6_squeezed, v6_squeezed_shape])\n# print all variables to see what is there\nprint e # shape of v6\nprint g # shape of v6_squeezed\n<\/code><\/pre>\n"}]}
{"QuestionId":33655731,"AnswerCount":14,"Tags":"<python><ubuntu><glibc><tensorflow>","CreationDate":"2015-11-11T16:58:54.000","AcceptedAnswerId":33699100.0,"OwnerUserId":5551505.0,"Title":"Error while importing Tensorflow in Python 2.7 in Ubuntu 12.04. 'GLIBC_2.17 not found'","Body":"<p>I have installed the Tensorflow bindings with python successfully. But when I try to import Tensorflow, I get the follwoing error.<\/p>\n\n<pre><\/pre>\n\n<p>I have tried to update GLIBC_2.15 to 2.17, but no luck.<\/p>\n","answers":[{"AnswerId":"34900471","CreationDate":"2016-01-20T12:45:41.107","ParentId":null,"OwnerUserId":"1990516","Title":null,"Body":"<p>Okay so here is the other solution I mentionned in my <a href=\"https:\/\/stackoverflow.com\/a\/34897674\/1990516\">previous answer<\/a>, it's more tricky, but should always work on systems with GLIBC>=2.12 and GLIBCXX>=3.4.13.\nIn my case it was on a CentOS 6.7, but it's also fine for Ubuntu 12.04.<\/p>\n\n<p>We're going to need a version of gcc that supports c++11, either on another machine or an isolated install; but not for the moment.<\/p>\n\n<p>What we're gonna do here is edit the _pywrap_tensorflow.so binary in order to 'weakify' its libc and libstdc++ dependencies, so that ld accepts to link the stubs we're gonna make. Then we'll make those stubs for the missing symbols, and finally we're gonna pre-load all of this when running python.<\/p>\n\n<p>First of all, I want to thank James for his great article ( <a href=\"http:\/\/www.lightofdawn.org\/wiki\/wiki.cgi\/NewAppsOnOldGlibc\" rel=\"nofollow noreferrer\">http:\/\/www.lightofdawn.org\/wiki\/wiki.cgi\/NewAppsOnOldGlibc<\/a> ) and precious advices, I couldn't have made it without him.<\/p>\n\n<p>So let's start by weakifying the dependencies, it's just about replacing the right bytes in _pywrap_tensorflow.so. Please note that this step only works for the current version of tensorflow (0.6.0). So if its not done already create and activate your <a href=\"http:\/\/docs.python-guide.org\/en\/latest\/dev\/virtualenvs\/\" rel=\"nofollow noreferrer\">virtualenv<\/a> if you have one (if you're not admin virtualenv is a solution, another is to add <code>--user<\/code> flag to pip command), and install tensorflow 0.6.0 (replace cpu by gpu in the url if you want the gpu version) :<\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.6.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>And let's weakify all the annoying dependencies, here is the command for the cpu version of tensorflow:<\/p>\n\n<pre><code>TENSORFLOW_DIR=`python -c \"import imp; print(imp.find_module('tensorflow')[1])\"`\nfor addr in 0xC6A93C 0xC6A99C 0xC6A9EC 0xC6AA0C 0xC6AA1C 0xC6AA3C; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so bs=1 seek=$((addr)) ; done\n<\/code><\/pre>\n\n<p>And here is the gpu one (run only the correct one or you'll corrupt the binary):<\/p>\n\n<pre><code>TENSORFLOW_DIR=`python -c \"import imp; print(imp.find_module('tensorflow')[1])\"`\nfor addr in 0xDC5EA4 0xDC5F04 0xDC5F54 0xDC5F74 0xDC5F84 0xDC5FA4; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so bs=1 seek=$((addr)) ; done\n<\/code><\/pre>\n\n<p>You can check it with:<\/p>\n\n<pre><code>readelf -V ${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so\n<\/code><\/pre>\n\n<p>Have a look at the article if you want to understand what's going on here.<\/p>\n\n<p>Now we're gonna make the stubs for the missing libc symbols:<\/p>\n\n<pre><code>mkdir ~\/my_stubs\ncd ~\/my_stubs\nMYSTUBS=~\/my_stubs\nprintf \"#include &lt;time.h&gt;\\n#include &lt;string.h&gt;\\nvoid* memcpy(void *dest, const void *src, size_t n) {\\nreturn memmove(dest, src, n);\\n}\\nint clock_gettime(clockid_t clk_id, struct timespec *tp) {\\nreturn clock_gettime(clk_id, tp);\\n}\" &gt; mylibc.c\ngcc -s -shared -o mylibc.so -fPIC -fno-builtin mylibc.c\n<\/code><\/pre>\n\n<p>You <strong>need<\/strong> to perform that step on the machine with the missing dependencies (or machine with similar versions of standard libraries (in a cluster for example)).<\/p>\n\n<p>Now we're gonna probably change of machine since we need a gcc that supports c++11, and it is probably not on the machine that lacks all the dependencies (or you can use an isolated install of a recent gcc). In the following I assume we're still in <code>~\/my_stubs<\/code> and somehow you share your home accross the machines, otherwise you'll just have to copy the .so files we're gonna generate when it's done.<\/p>\n\n<p>So, one stub that we can do for libstdc++, and for the remaining missing ones we're going to compile them from gcc source (it might take some time to clone the repository):<\/p>\n\n<pre><code>printf \"#include &lt;functional&gt;\\nvoid std::__throw_bad_function_call(void) {\\nexit(1);\\n}\" &gt; bad_function.cc\ngcc -std=c++11 -s -shared -o bad_function.so -fPIC -fno-builtin bad_function.cc\ngit clone https:\/\/github.com\/gcc-mirror\/gcc.git\ncd gcc\nmkdir my_include\nmkdir my_include\/ext\ncp libstdc++-v3\/include\/ext\/aligned_buffer.h my_include\/ext\ngcc -I$PWD\/my_include -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/hashtable.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/hashtable_c++0x.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/chrono.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/chrono.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/random.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/random.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/hash_bytes.so -fPIC -fno-builtin .\/libstdc++-v3\/libsupc++\/hash_bytes.cc\n<\/code><\/pre>\n\n<p>And that's it! You can now run a tensorflow python script by preloading all our shared libraries (and your local libstdc++):<\/p>\n\n<pre><code>LIBSTDCPP=`ldconfig -p | grep libstdc++.so.6 | grep 64 | cut -d' ' -f4` #For 64bit machines\nLD_PRELOAD=\"$MYSTUBS\/mylibc.so:$MYSTUBS\/random.so:$MYSTUBS\/hash_bytes.so:$MYSTUBS\/chrono.so:$MYSTUBS\/hashtable.so:$MYSTUBS\/bad_function.so:$LIBSTDCPP\" python ${TENSORFLOW_DIR}\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n\n<p>:)<\/p>\n"},{"AnswerId":"34897674","CreationDate":"2016-01-20T10:37:40.953","ParentId":null,"OwnerUserId":"1990516","Title":null,"Body":"<p>I tried <a href=\"https:\/\/stackoverflow.com\/a\/33699100\/1990516\">BR_User solution<\/a> and still had an annoying: <\/p>\n\n<pre><code>ImportError: \/usr\/lib64\/libstdc++.so.6: version `GLIBCXX_3.4.14' not found\n<\/code><\/pre>\n\n<p>I am on CentOS 6.7, it also lacks an updated c++ standard lib, so to build on BR_User solution I extracted the correct libstdc++ package, however I found no need for the virtual environment. <\/p>\n\n<p>Supposing you have already installed tensorflow, it gives:<\/p>\n\n<pre><code>mkdir ~\/my_libc_env\ncd ~\/my_libc_env\nwget http:\/\/launchpadlibrarian.net\/137699828\/libc6_2.17-0ubuntu5_amd64.deb\nwget http:\/\/launchpadlibrarian.net\/137699829\/libc6-dev_2.17-0ubuntu5_amd64.deb\nwget ftp.riken.jp\/Linux\/scientific\/7.0\/x86_64\/os\/Packages\/libstdc++-4.8.2-16.el7.x86_64.rpm\nar p libc6_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nar p libc6-dev_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nrpm2cpio libstdc++-4.8.2-7mgc30.x86_64.rpm| cpio -idmv\n<\/code><\/pre>\n\n<p>and then run python with:<\/p>\n\n<pre><code>LD_LIBRARY_PATH=\"$HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/:$HOME\/my_libc_env\/usr\/lib64\/\" $HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/ld-2.17.so `which python`\n<\/code><\/pre>\n\n<p>If it doesn't work, I have <a href=\"https:\/\/stackoverflow.com\/a\/34900471\/1990516\">another solution<\/a>, but you won't like it.<\/p>\n"},{"AnswerId":"34808203","CreationDate":"2016-01-15T09:42:32.703","ParentId":null,"OwnerUserId":"3246234","Title":null,"Body":"<p>That error mainly arises if your GNU C-Library is not up-to-date.\nYou can check what version you're running using a simple<\/p>\n\n<pre><code>$ ldd --version\n<\/code><\/pre>\n\n<p>The output should be like this:<\/p>\n\n<pre><code>ldd (Ubuntu EGLIBC 2.19-0ubuntu6.6) 2.19\nCopyright (C) 2014 Free Software Foundation, Inc.\nThis is free software; see the source for copying conditions.\nThere is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\nWritten by Roland McGrath and Ulrich Drepper.\n<\/code><\/pre>\n\n<p>The 2.19 is your GLIBC version. To upgrade you can visit the GNU-C Library project website and download the latest version.\nThe link to the latest glibc is here : <a href=\"https:\/\/www.gnu.org\/software\/libc\/index.html\" rel=\"nofollow\">GNU-C library Download<\/a>\nAt the time of writing this answer the <strong>latest stable version was 2.22<\/strong>. <\/p>\n\n<p>I tried running tensorflow on both Ubuntu 12.04 and 14.04. Ubuntu 14.04 doesn't throw this issue as it comes with glibc 2.19 installed by default.<\/p>\n\n<p>Hope it helps.  <\/p>\n"},{"AnswerId":"54628478","CreationDate":"2019-02-11T10:29:22.067","ParentId":null,"OwnerUserId":"974555","Title":null,"Body":"<p>I had a similar problem trying to install <code>numba<\/code> on an old HPC Linux system running \nSUSE Linux Enterprise Server 11, which has glibc 2.11.3 (according to <code>ldd --version<\/code>).  I have no superuser access, and any regular conda installation (either from the channels <code>defaults<\/code>, <code>conda-forge<\/code>, or <code>numba<\/code>) would fail with <code>OSError: \/lib64\/libpthread.so.0: version<\/code>GLIBC_2.12<code>not found (required by ...envs\/test1-py37\/lib\/python3.7\/site-packages\/llvmlite\/binding\/libllvmlite.so<\/code>).  I managed to install <code>numba<\/code> by the following steps:<\/p>\n\n<ol>\n<li>Create a conda environment<\/li>\n<li>Inside the conda environment, install <code>glibc<\/code> from the <code>nlsec<\/code> environment<\/li>\n<li>Then use <code>pip<\/code> to install <code>numba<\/code> and <code>llvmlite<\/code> built against the newly installed <code>glibc<\/code>:<\/li>\n<\/ol>\n\n<p>Thus:<\/p>\n\n<pre><code>conda create -n test -c nlesc glibc\nconda activate test\npip install numba\n<\/code><\/pre>\n\n<p>Then importing numba works.  However, one should subsequently be careful that any subsequent <code>conda install<\/code> does not \"upgrade\" <code>llvmlite<\/code> to a version not compatible with <code>numba<\/code>, or the old problem returns.  You may want <a href=\"https:\/\/stackoverflow.com\/a\/48733093\/974555\">to pin<\/a> either or both for this effect.  You'll have to pin the build as well as the version.  I created a file <code>pinned<\/code> containing:<\/p>\n\n<pre><code>llvmlite==0.27.0-pypi\nnumba==0.42.0-pypi\n<\/code><\/pre>\n\n<p>and placed it within my conda environment in a text file <code>conda-meta\/pinned<\/code>.<\/p>\n"},{"AnswerId":"57642801","CreationDate":"2019-08-25T02:25:50.027","ParentId":null,"OwnerUserId":"4231538","Title":null,"Body":"<p>I had the same problem and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/177\" rel=\"nofollow noreferrer\">found the following solution<\/a>:<\/p>\n\n<pre><code>conda install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"33699100","CreationDate":"2015-11-13T18:03:05.210","ParentId":null,"OwnerUserId":"935448","Title":null,"Body":"<p>I had the same problem, so googling I made these steps:<\/p>\n\n<pre><code>$ sudo pip install --upgrade virtualenv\n$ virtualenv --system-site-packages ~\/tensorflow\n$ cd ~\/tensorflow\n$ source bin\/activate\n$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n$ cd \/tmp\n$ wget http:\/\/launchpadlibrarian.net\/137699828\/libc6_2.17-0ubuntu5_amd64.deb\n$ wget http:\/\/launchpadlibrarian.net\/137699829\/libc6-dev_2.17-0ubuntu5_amd64.deb\n$ mkdir libc6_2.17\n$ cd libc6_2.17\n$ ar p ..\/libc6_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\n$ ar p ..\/libc6-dev_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\n$ cd -\n$ LD_LIBRARY_PATH=\/tmp\/libc6_2.17\/lib\/x86_64-linux-gnu\/ \/tmp\/libc6_2.17\/lib\/x86_64-linux-gnu\/ld-2.17.so bin\/python local\/lib\/python2.7\/site-packages\/tensorflow\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n\n<p>And to exit:<\/p>\n\n<pre><code>$ deactivate \n<\/code><\/pre>\n\n<p>That works for me.<\/p>\n"},{"AnswerId":"34993102","CreationDate":"2016-01-25T12:51:27.380","ParentId":null,"OwnerUserId":"5654065","Title":null,"Body":"<p>My solution is similar to that of Theo T; although fine tuned for Xubuntu 12.04 (CAELinux 2013)<\/p>\n\n<pre><code>sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.6.0-cp27-none-linux_x86_64.whl\nsudo su\n# prompt for password\nTENSORFLOW_DIR=`python -c \"import imp; print(imp.find_module('tensorflow')[1])\"`\nfor addr in 0xC6A93C 0xC6A99C 0xC6A9EC 0xC6AA0C 0xC6AA1C 0xC6AA3C; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so bs=1 seek=$((addr)) ; done\n\nreadelf -V ${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so\nexit\n\nmkdir ~\/.my_stubs\ncd ~\/.my_stubs\nMYSTUBS=~\/.my_stubs\n\n\nprintf \"#include &lt;time.h&gt;\\n#include &lt;string.h&gt;\\nvoid* memcpy(void *dest, const void *src, size_t n) {\\nreturn memmove(dest, src, n);\\n}\\nint clock_gettime(clockid_t clk_id, struct timespec *tp) {\\nreturn clock_gettime(clk_id, tp);\\n}\" &gt; mylibc.c\ngcc -s -shared -o mylibc.so -fPIC -fno-builtin mylibc.c\n\nprintf \"#include &lt;functional&gt;\\nvoid std::__throw_bad_function_call(void) {\\nexit(1);\\n}\" &gt; bad_function.cc\ngcc -std=c++11 -s -shared -o bad_function.so -fPIC -fno-builtin bad_function.cc\n\ngit clone https:\/\/github.com\/gcc-mirror\/gcc.git\n\ncd gcc\nmkdir my_include\nmkdir my_include\/ext\ncp libstdc++-v3\/include\/ext\/aligned_buffer.h my_include\/ext\ngcc -I$PWD\/my_include -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/hashtable.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/hashtable_c++0x.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/chrono.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/chrono.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/random.so -fPIC -fno-builtin libstdc++-v3\/src\/c++11\/random.cc\ngcc -std=c++11 -fpermissive -s -shared -o $MYSTUBS\/hash_bytes.so -fPIC -fno-builtin .\/libstdc++-v3\/libsupc++\/hash_bytes.cc\n<\/code><\/pre>\n\n<p>The following two steps would run the convolution.py script located in the tensorflow\/models\/image\/mnist directory: <\/p>\n\n<pre><code>LIBSTDCPP=`ldconfig -p | grep libstdc++.so.6 | grep 64 | cut -d' ' -f4` #For 64bit machines\nLD_PRELOAD=\"$MYSTUBS\/mylibc.so:$MYSTUBS\/random.so:$MYSTUBS\/hash_bytes.so:$MYSTUBS\/chrono.so:$MYSTUBS\/hashtable.so:$MYSTUBS\/bad_function.so:$LIBSTDCPP\" python ${TENSORFLOW_DIR}\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n\n<p>If you want python to be always loaded with these reduced dependencies, simply add this to the .bashrc file with the below step:<\/p>\n\n<pre><code>echo \"LIBSTDCPP=`ldconfig -p | grep libstdc++.so.6 | grep 64 | cut -d' ' -f4`\" &gt;&gt; ~\/.bashrc\necho alias python=\"'\"LD_PRELOAD='\"''$'MYSTUBS\/mylibc.so:'$'MYSTUBS\/random.so:'$'MYSTUBS\/hash_bytes.so:'$'MYSTUBS\/chrono.so:'$'MYSTUBS\/hashtable.so:'$'MYSTUBS\/bad_function.so:'$'LIBSTDCPP'\"' python\"'\" &gt;&gt; ~\/.bashrc\n<\/code><\/pre>\n\n<p>And if you have ipython:<\/p>\n\n<pre><code>echo alias ipython=\"'\"LD_PRELOAD='\"''$'MYSTUBS\/mylibc.so:'$'MYSTUBS\/random.so:'$'MYSTUBS\/hash_bytes.so:'$'MYSTUBS\/chrono.so:'$'MYSTUBS\/hashtable.so:'$'MYSTUBS\/bad_function.so:'$'LIBSTDCPP'\"' ipython\"'\" &gt;&gt; ~\/.bashrc\n<\/code><\/pre>\n\n<p>Basically, I updated it to the python installation being done in dist_packages instead of site-packages. Also, I clone from the <a href=\"https:\/\/github.com\/gcc-mirror\" rel=\"nofollow\">gcc-mirror<\/a> instead of the native git page for gcc. Rest all is the same.<\/p>\n\n<p>The last few steps ensure that every time you run python or ipython, it would be loaded with the reduced dependencies.<\/p>\n"},{"AnswerId":"41003023","CreationDate":"2016-12-06T19:10:59.920","ParentId":null,"OwnerUserId":"1032586","Title":null,"Body":"<p>I've just managed to install tensorflow 0.12rc0 on CentOS 6.5 with glibc 2.12, without having root privileges. Simply installing tensorflow binary via pip was giving me an error, related to GLIBC version as well.<\/p>\n\n<p>Basically, you have 4 options how to deal with this (each with some advantages and disadvantages):<\/p>\n\n<h2>Option 1 - Upgrade your system GLIBC globally.<\/h2>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/33658956\/1032586\">This is<\/a>, probably, the best option, <em>if<\/em> your system <a href=\"https:\/\/superuser.com\/a\/537694\/135260\">supports this<\/a>, you have root privileges, and you are confident that this upgrade won't break anything for some weird reason. Ultimately, this goes up to upgrading the whole Linux distribution. <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/#axzz4Rtai3pua\" rel=\"noreferrer\">Here<\/a>'s a nice short list of default GLIBC versions on popular distributions.<\/p>\n\n<h2>Option 2 - Add second GLIBC to your system<\/h2>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/33699100\/1032586\">Compile or download binary<\/a>. The most simple&amp;straightforward option. Especially if you only need to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110#issuecomment-220009120\" rel=\"noreferrer\">run few simple scripts<\/a>. <\/p>\n\n<ul>\n<li>It <a href=\"https:\/\/stackoverflow.com\/a\/851229\/1032586\">is possible<\/a> to have multiple versions of glibc on the same system, but one should do this with a great care. <\/li>\n<li>You won't destroy your system, if all your changes would be limited to a virtual environment. <\/li>\n<li>Many programs, installed\/compiled before might be relying on old GLIBC, <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\/34900471#comment65844745_33699100\">would just crash<\/a> in your new environment (e.g. your python IDE). Including most basic bash commands, like \"lc\", \"cd\", etc. <\/li>\n<li>Other side-effects like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1426#issuecomment-195049947\" rel=\"noreferrer\">significant memory leaks<\/a> are also possible.<\/li>\n<li>Thus, it's a very bad idea to <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\/34900471#comment62537898_34808203\">add new GLIBC to your normal environment<\/a>, e.g. via <code>.bashrc<\/code>. <\/li>\n<li>On the other hand, if you need some specific tool in your new virtual environment, you may <a href=\"https:\/\/stackoverflow.com\/a\/851229\/1032586\">recompile it, linking against new GLIBC<\/a>. So, that it would work OK in your new enviroment.<\/li>\n<li>However, personally, I quickly gave up recompiling everything I need in a new environment (without root and a package manager).<\/li>\n<li>A <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\/34897674#comment62546305_34808203\">slightly different approach<\/a> is officially offered by GLIBC developers, for testing new GLIBC builds.<\/li>\n<\/ul>\n\n<h2>Option 3 - Patch tensorflow<\/h2>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\/34900471#34900471\">This<\/a> may work for TF 0.6.0, but you would probably have to start again from scratch, when each new tensorflow version is released. E.g. <a href=\"http:\/\/www.it1me.com\/it-answers?id=33655731&amp;ttl=Error%20while%20importing%20Tensorflow%20in%20python2.7%20in%20Ubuntu%2012.04.%20%26%2339%3BGLIBC_2.17%20not%20found%26%2339%3B\" rel=\"noreferrer\">here<\/a>'s a fix for 0.9.0.<\/p>\n\n<h2>Option 4 - Compile tensorflow from source<\/h2>\n\n<p>If you re-compile it from source and link against your existing GLIBC, newer GLIBC would be no longer needed. Somehow, this option was not mentioned in any answer here yet. Imho, this is the best option, both \"<a href=\"https:\/\/superuser.com\/a\/537694\/135260\">in general<\/a>\", and \"specifically for tensorflow\".<\/p>\n\n<ul>\n<li>This works OK with r0.11 and would probably work for years, but theoretically, it might break in some newer tensorflow version, if they would decide to actually use some new GLIBC functionality, not present in older versions.<\/li>\n<li>To be honest, building tensorflow from source is not straightforward, especially on outdated systems. <\/li>\n<\/ul>\n\n<p><strong>A quick summary of \"building tensorflow on outdated system\":<\/strong><\/p>\n\n<p>Although the official guide provides a \"<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup.html#installing-from-sources\" rel=\"noreferrer\">installing from sources<\/a>\" section, there are few tricks you need to do to build it on an outdated system. Here I assume, that you do not have root privileges (if you do - you probably would be able to install the same pre-requestities with a package manager, rather them manually building them from source). <\/p>\n\n<p>I found two well-documented success stories: <a href=\"http:\/\/blog.abysm.org\/2016\/06\/building-tensorflow-centos-6\" rel=\"noreferrer\">#1<\/a>, <a href=\"https:\/\/gist.github.com\/taylorpaul\/3e4e405ffad1be8355bf742fa24b41f0\" rel=\"noreferrer\">#2<\/a> and a number of useful posts on the official github (mostly about a set of libraries to link inside the binary): <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110#issuecomment-201834137\" rel=\"noreferrer\">#1<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/53#issuecomment-164105922\" rel=\"noreferrer\">#2<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2291#issuecomment-244533549\" rel=\"noreferrer\">#3<\/a>, <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/649#issuecomment-166710509\" rel=\"noreferrer\">#4<\/a>. I had to combine tricks, described there to successfully compile TF in my case.<\/p>\n\n<ol>\n<li><p>First of all, check your <code>gcc --version<\/code>, and <a href=\"https:\/\/gcc.gnu.org\/projects\/cxx-status.html#cxx11\" rel=\"noreferrer\">verify<\/a> that it supports c++11. Mine was 4.4.7, so it won't work. I've <a href=\"https:\/\/gcc.gnu.org\/releases.html\" rel=\"noreferrer\">downloaded<\/a> gcc-4.9.4 source code, and <a href=\"https:\/\/gcc.gnu.org\/wiki\/InstallingGCC\" rel=\"noreferrer\">compiled it<\/a>. This step is pretty straightforward, but the compilation itself may take few hours. As a workaround for an issue in bazel, I've compiled gcc with <a href=\"https:\/\/stackoverflow.com\/a\/39181639\/1032586\">hardcoded paths<\/a> to <code>as<\/code>,<code>ld<\/code> and <code>nm<\/code>. However, you may try another workarounds: (<a href=\"https:\/\/stackoverflow.com\/questions\/39095577\/tensorflow-build-failure-with-gcc-error-as#comment69045706_39181639\">1<\/a>, <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/760#issuecomment-211037359\" rel=\"noreferrer\">2<\/a>).<\/p>\n\n<pre><code>#!\/bin\/sh\n\nunset LIBRARY_PATH CPATH C_INCLUDE_PATH \nunset PKG_CONFIG_PATH CPLUS_INCLUDE_PATH INCLUDE LD_LIBRARY_PATH\n\ncd gcc-4.9.4\n.\/contrib\/download_prerequisites\n\nmkdir objdir\ncd objdir\n\n\n# I've added --disable-multilib to fix the following error:\n# \/usr\/bin\/ld: crt1.o: No such file: No such file or directory\n# collect2: ld returned 1 exit status\n# configure: error: I suspect your system does not have 32-bit \n# developement libraries (libc and headers). If you have them,\n# rerun configure with --enable-multilib. If you do not have them, \n# and want to build a 64-bit-only compiler, rerun configure \n# with --disable-multilib.           \n\n..\/configure --prefix=$HOME\/opt\/gcc-4.9.4 \\\n             --disable-multilib \\\n             --disable-nls \\\n             --enable-languages=c,c++ \\\n             --with-ld=\/usr\/bin\/ld \\\n             --with-nm=\/usr\/bin\/nm \\\n             --with-as=\/usr\/bin\/as\n\nmake        \nmake install\n<\/code><\/pre><\/li>\n<li><p>Check your <code>java --version<\/code>. Bazel requires JDK 8, install it if necessary. (They still <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/releases\" rel=\"noreferrer\">provide<\/a> some jdk7 related downloads, for bazel-0.4.1 but it looks like they consider it deprecated)<\/p><\/li>\n<li><p>I've created a separate <code>use_gcc_4.9.4.sh<\/code> file, with necessary environment variables. I use <code>source .\/use_gcc_4.9.4.sh<\/code> when I need to so something related to this newer compiler.<\/p>\n\n<pre><code>#!\/bin\/sh\nthis=$HOME\/opt\/gcc-4.9.4\nexport PATH=$this\/bin:$PATH\nexport CPATH=$this\/include:$CPATH\nexport LIBRARY_PATH=$this\/lib:$LIBRARY_PATH\nexport LIBRARY_PATH=$this\/lib64:$LIBRARY_PATH\nexport LD_LIBRARY_PATH=$this\/lib:$LD_LIBRARY_PATH\nexport LD_LIBRARY_PATH=$this\/lib64:$LD_LIBRARY_PATH\n<\/code><\/pre><\/li>\n<li><p>The current bazel binary (0.4.1) <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/660\" rel=\"noreferrer\">requires GLIBC 2.14<\/a>, so we have to <a href=\"https:\/\/bazel.build\/versions\/master\/docs\/install.html#compiling-from-source\" rel=\"noreferrer\">compile bazel from source<\/a> as well (with our new gcc). Works OK, unless you are only <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/2177\" rel=\"noreferrer\">allowed to run a very limited number of threads<\/a> on the target machine. (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110#issuecomment-201834137\" rel=\"noreferrer\">This<\/a> post describes some additional workarounds, but in my case they were not needed, maybe due to recent updates in bazel code.)<\/p><\/li>\n<li><p>Obtain tensorflow source code <code>git clone https:\/\/github.com\/tensorflow\/tensorflow<\/code>, and install prerequisites you need (CUDA,cuDNN,python, etc). See <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup.html#installing-from-sources\" rel=\"noreferrer\">official guide<\/a>.<\/p><\/li>\n<li><p>If you're not using default system gcc (e.g. if you had to compile newer gcc, like discussed above), add <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/649#issuecomment-166710509\" rel=\"noreferrer\">the following linker flags<\/a> to <code>tensorflow\/third_party\/gpus\/crosstool\/CROSSTOOL.tpl<\/code>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v0.11.0\/third_party\/gpus\/crosstool\/CROSSTOOL.tpl#L59\" rel=\"noreferrer\">line 59<\/a>:<\/p>\n\n<pre><code>linker_flag: \"-L\/home\/username\/localinst\/opt\/gcc-4.9.4\/lib64\"\nlinker_flag: \"-Wl,-rpath,\/home\/username\/localinst\/opt\/gcc-4.9.4\/lib64\"\n<\/code><\/pre>\n\n<p>Without this step, you would likely run into error messages like <a href=\"https:\/\/github.com\/ContinuumIO\/anaconda-issues\/issues\/483\" rel=\"noreferrer\">this<\/a>:<\/p>\n\n<pre><code># ERROR: \/home\/username\/localdistr\/src\/tensorflow\/tensorflow\/tensorflow\/core\/debug\/BUILD:33:1: null failed: protoc failed: error executing command bazel-out\/host\/bin\/external\/protobuf\/protoc '--cpp_out=bazel-out\/local_linux-py3-opt\/genfiles\/' '--plugin=protoc-gen-grpc=bazel-out\/host\/bin\/external\/grpc\/grpc_cpp_plugin' ... (remaining 8 argument(s) skipped): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.\n# bazel-out\/host\/bin\/external\/protobuf\/protoc: \/usr\/lib64\/libstdc++.so.6: version `GLIBCXX_3.4.20' not found (required by bazel-out\/host\/bin\/external\/protobuf\/protoc)\n# bazel-out\/host\/bin\/external\/protobuf\/protoc: \/usr\/lib64\/libstdc++.so.6: version `CXXABI_1.3.8' not found (required by bazel-out\/host\/bin\/external\/protobuf\/protoc)\n# bazel-out\/host\/bin\/external\/protobuf\/protoc: \/usr\/lib64\/libstdc++.so.6: version `GLIBCXX_3.4.18' not found (required by bazel-out\/host\/bin\/external\/protobuf\/protoc)\n<\/code><\/pre><\/li>\n<li><p>Finally, to avoid GLIBC dependencies, we have to statically link some libraries, by adding the <code>-lrt<\/code> linker flag (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2291\" rel=\"noreferrer\">maybe<\/a> <code>-lm<\/code> as well). I found multiple posts, suggesting to add this in a different manner: <\/p>\n\n<ul>\n<li>via <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110#issuecomment-201834137\" rel=\"noreferrer\">bazel command line<\/a> (may sound reasonable, but not working for me on current tensorflow version, somehow),<\/li>\n<li>via <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2291#issuecomment-244533549\" rel=\"noreferrer\">\"bazel-tensorflow\/external\/protobuf\/BUILD\"<\/a>(not sure if it's working, but this does not look convenient - this file is only created during the build attempt itself) <\/li>\n<li><p>via <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2291#issuecomment-250025095\" rel=\"noreferrer\">\"third_party\/gpus\/crosstool\/CROSSTOOL.tpl\"<\/a> (the same file we've just edited in the previous step, just below the lines we've already added). <\/p>\n\n<pre><code>linker_flag: \"-lrt\"\nlinker_flag: \"-lm\"\n<\/code><\/pre><\/li>\n<li><p>via <a href=\"http:\/\/blog.abysm.org\/2016\/06\/building-tensorflow-centos-6\" rel=\"noreferrer\">\"tensorflow\/tensorflow.bzl\"<\/a> (works for me, but less convenient just because you have to edit one more file. I'm not sure it's 100% equivalent to the previous point)<\/p><\/li>\n<\/ul>\n\n<p>Without <code>-lrt<\/code> I ran into GLIBC-version-specific error again, trying to <code>import tensorflow<\/code>:<\/p>\n\n<pre><code># ImportError: \/lib64\/libc.so.6: version `GLIBC_2.14' not found (required by \/home\/username\/anaconda3\/envs\/myenvname\/lib\/python3.5\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so)\n<\/code><\/pre>\n\n<p>Without <code>-lm<\/code> you may run into <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2291\" rel=\"noreferrer\">this<\/a> (for me, it turned out to be not necessary).<\/p><\/li>\n<li><p>Run the build process.<\/p><\/li>\n<\/ol>\n\n<pre class=\"lang-bash prettyprint-override\"><code>    source .\/use_gcc_4.9.4.sh\n    .\/configure\n    bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n    bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n    pip install --upgrade \/tmp\/tensorflow_pkg\/tensorflow-0.12.0rc0-cp35-cp35m-linux_x86_64.whl\n<\/code><\/pre>\n\n<ol start=\"9\">\n<li><p>Try to run the following simple python script to test if the most basic stuff is functioning:<\/p>\n\n<pre><code>import tensorflow as tf\nhello = tf.constant('Hello, TensorFlow!')\nsess = tf.Session()\nprint(sess.run(hello))\n\na = tf.constant(10)\nb = tf.constant(32)\nprint(sess.run(a + b))\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"50021555","CreationDate":"2018-04-25T11:48:04.647","ParentId":null,"OwnerUserId":"9003405","Title":null,"Body":"<p>I got \"version `GLIBC_2.14' not found\" when importing tensorflow in python. Managed to import tensorflow in python3.5 on Centos6.6 using virtualenv:<\/p>\n\n<pre><code>mkdir ~\/tensorflow\ncd ~\/tensorflow\nvirtualenv --system-site-packages -p python3.5 ~\/tensorflow\nsource bin\/activate\nwget http:\/\/launchpadlibrarian.net\/137699828\/libc6_2.17-0ubuntu5_amd64.deb\nwget http:\/\/launchpadlibrarian.net\/137699829\/libc6-dev_2.17-0ubuntu5_amd64.deb\nwget ftp:\/\/195.220.108.108\/linux\/mageia\/distrib\/4\/x86_64\/media\/core\/updates\/libstdc++6-4.8.2-3.2.mga4.x86_64.rpm\nwget https:\/\/rpmfind.net\/linux\/centos\/7.4.1708\/updates\/x86_64\/Packages\/glibc-2.17-196.el7_4.2.i686.rpm\nar p libc6_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nar p libc6-dev_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nrpm2cpio libstdc++6-4.8.2-3.2.mga4.x86_64.rpm | cpio -idmv\nrpm2cpio glibc-2.17-196.el7_4.2.i686.rpm | cpio -idmv\npip3.5 install --upgrade tensorflow\nexport PYTHONPATH=\"$HOME\/tensorflow\/lib\/python3.5\/site-packages\/\"\nalias tfpython='LD_LIBRARY_PATH=\"$HOME\/tensorflow\/lib\/x86_64-linux-gnu\/:$HOME\/tensorflow\/usr\/lib64\/\" $HOME\/tensorflow\/lib\/x86_64-linux-gnu\/ld-2.17.so \/usr\/local\/bin\/python3.5'\ntfpython\n<\/code><\/pre>\n"},{"AnswerId":"45193695","CreationDate":"2017-07-19T14:30:46.903","ParentId":null,"OwnerUserId":"5138923","Title":null,"Body":"<p>Above answers are fine but I still faced some issues even after doing them. I followed as in @Theo answer on rhel and I was trying to export LD_LIBRARY_PATH which had thrown some errors, along with LD_LIBRARY_PATH also include LD_PRELOAD as <\/p>\n\n<pre><code>mkdir libstdc\ncd libstdc\nrpm2cpio ..\/libstdc++-4.8.3-9.el7.x86_64.rpm| cpio -idmv\nLD_LIBRARY_PATH=\"$HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/:$HOME\/my_libc_env\/lib64\/\" LD_PRELOAD=\"$HOME\/my_libc_env\/libstdc\/usr\/lib64\/libstdc++.so.6.0.19\" $HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/ld-2.17.so `which python`\n<\/code><\/pre>\n\n<p>you have to always do this while running python so make a script that loads these variables and calls python interpreter<\/p>\n\n<pre><code>vim tmp.sh\n\nLD_LIBRARY_PATH=\"$HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/:$HOME\/my_libc_env\/lib64\/\" LD_PRELOAD=\"$HOME\/my_libc_env\/libstdc\/usr\/lib64\/libstdc++.so.6.0.19\" $HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/ld-2.17.so `which python`\n<\/code><\/pre>\n"},{"AnswerId":"36348685","CreationDate":"2016-04-01T05:25:02.367","ParentId":null,"OwnerUserId":"6143296","Title":null,"Body":"<p>The link in @Th\u00e9o T's answer is no longer working, so I had to try several alternatives to finally find one that works. <\/p>\n\n<pre><code>mkdir ~\/my_libc_env\ncd ~\/my_libc_env\nwget http:\/\/launchpadlibrarian.net\/137699828\/libc6_2.17-0ubuntu5_amd64.deb\nwget http:\/\/launchpadlibrarian.net\/137699829\/libc6-dev_2.17-0ubuntu5_amd64.deb\nwget ftp:\/\/ftp.icm.edu.pl\/vol\/rzm5\/linux-slc\/centos\/7.0.1406\/cr\/x86_64\/Packages\/libstdc++-4.8.3-9.el7.x86_64.rpm\nar p libc6_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nar p libc6-dev_2.17-0ubuntu5_amd64.deb data.tar.gz | tar zx\nrpm2cpio libstdc++-4.8.3-9.el7.x86_64.rpm| cpio -idmv\n<\/code><\/pre>\n\n<p>Run python code with tensorflow<\/p>\n\n<pre><code>LD_LIBRARY_PATH=\"$HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/:$HOME\/my_libc_env\/usr\/lib64\/\" $HOME\/my_libc_env\/lib\/x86_64-linux-gnu\/ld-2.17.so `which python` your_program.py\n<\/code><\/pre>\n"},{"AnswerId":"36733420","CreationDate":"2016-04-20T03:42:54.917","ParentId":null,"OwnerUserId":"1637673","Title":null,"Body":"<p>Notice: you should download deb package by your platform:<\/p>\n\n<p>x86:<\/p>\n\n<p><a href=\"https:\/\/launchpad.net\/ubuntu\/raring\/i386\/libc6\/2.17-0ubuntu5.1\" rel=\"nofollow\">https:\/\/launchpad.net\/ubuntu\/raring\/i386\/libc6\/2.17-0ubuntu5.1<\/a><\/p>\n\n<pre><code>wget http:\/\/launchpadlibrarian.net\/151932048\/libc6_2.17-0ubuntu5.1_i386.deb\n\nsudo dpkg -i libc6_2.17-0ubuntu5.1_i386.deb\n<\/code><\/pre>\n\n<p>x64:<\/p>\n\n<p><a href=\"https:\/\/launchpad.net\/ubuntu\/raring\/amd64\/libc6\/2.17-0ubuntu5.1\" rel=\"nofollow\">https:\/\/launchpad.net\/ubuntu\/raring\/amd64\/libc6\/2.17-0ubuntu5.1<\/a><\/p>\n\n<pre><code>wget http:\/\/launchpadlibrarian.net\/151925896\/libc6_2.17-0ubuntu5.1_amd64.deb\n\nsudo dpkg -i libc6_2.17-0ubuntu5.1_amd64.deb\n<\/code><\/pre>\n\n<p>Test on my 12.04 Ubuntu x64.<\/p>\n"},{"AnswerId":"35767560","CreationDate":"2016-03-03T09:05:16.120","ParentId":null,"OwnerUserId":"5929065","Title":null,"Body":"<p>I have tried <a href=\"https:\/\/stackoverflow.com\/a\/34897674\/5929065\">https:\/\/stackoverflow.com\/a\/34897674\/5929065<\/a> solution, and its worked perfectly. But libstdc++-4.8.2-7 download link cannot be accessed any more, try to pick up a new one from this page: <a href=\"http:\/\/rpm.pbone.net\/index.php3\/stat\/4\/idpl\/31981489\/dir\/centos_7\/com\/libstdc++-4.8.5-4.el7.x86_64.rpm.html\" rel=\"nofollow noreferrer\">http:\/\/rpm.pbone.net\/index.php3\/stat\/4\/idpl\/31981489\/dir\/centos_7\/com\/libstdc++-4.8.5-4.el7.x86_64.rpm.html<\/a><\/p>\n\n<p>For me, I use this:<\/p>\n\n<p>wget <a href=\"ftp:\/\/mirror.switch.ch\/pool\/4\/mirror\/centos\/7.2.1511\/os\/x86_64\/Packages\/libstdc++-4.8.5-4.el7.x86_64.rpm\" rel=\"nofollow noreferrer\">ftp:\/\/mirror.switch.ch\/pool\/4\/mirror\/centos\/7.2.1511\/os\/x86_64\/Packages\/libstdc++-4.8.5-4.el7.x86_64.rpm<\/a><\/p>\n"},{"AnswerId":"38115363","CreationDate":"2016-06-30T06:23:09.110","ParentId":null,"OwnerUserId":"4702363","Title":null,"Body":"<p>Note to 'weakifying' solution of Theo Trouillon. It only applies for Tensorflow 0.6.0. If you want to apply it for Tensorflow 0.9.0, it gets trickier. \nMy case was CPU mode, Centos 6.7 where GLIBC 2.12 GLIBCXX 3.4.13.<\/p>\n\n<p>Installing tensorflow : <\/p>\n\n<pre><code>pip uninstall protobuf \npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.9.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>'weakifying' code should be CHANGED :<\/p>\n\n<pre><code>TENSORFLOW_DIR=`python -c \"import imp; print(imp.find_module('tensorflow')[1])\"`\nfor addr in 0x376e34 0x376e94 0x376ef4 0x376f14 0x376f24 0x376f44 0x376f54 ; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/python\/_pywrap_tensorflow.so bs=1 seek=$((addr)) ; done\nfor addr in 0x93aa4 0x93b04 ; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/contrib\/layers\/python\/ops\/_bucketization_op.so bs=1 seek=$((addr)) ; done\nfor addr in 0x95bec 0x95c4c ; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/contrib\/layers\/python\/ops\/_sparse_feature_cross_op.so bs=1 seek=$((addr)) ; done\nfor addr in 0x9ffec 0x9fffc 0xa005c ; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/contrib\/metrics\/python\/ops\/_set_ops.so bs=1 seek=$((addr)) ; done\nfor addr in 0x96aa4 0x96b04 0x96b24; do printf '\\x02' | dd conv=notrunc of=${TENSORFLOW_DIR}\/contrib\/linear_optimizer\/python\/ops\/_sdca_ops.so bs=1 seek=$((addr)) ; done\n<\/code><\/pre>\n\n<p>And following code should be ADDED to the stub generating code at newer machine.<\/p>\n\n<pre><code>mkdir my_include\/bits\ncp libstdc++-v3\/include\/bits\/shared_ptr_atomic.h my_include\/bits\/\ncp libstdc++-v3\/include\/std\/memory my_include\/\ngcc -I$PWD\/my_include -std=c++11 -fpermissive -s -shared -o ${MYSTUBS}\/shared_ptr.so -fPIC -fno-builtin .\/libstdc++-v3\/src\/c++11\/shared_ptr.cc\ngcc -I$PWD\/my_include -std=c++11 -fpermissive -s -shared -o ${MYSTUBS}\/list.so -fPIC -fno-builtin .\/libstdc++-v3\/src\/c++98\/list.cc\necho \"\n#include &lt;unistd.h&gt;\n#include &lt;stdlib.h&gt;\nchar * secure_getenv (const char *name) {\n          if ((getuid () == geteuid ()) &amp;&amp; (getgid () == getegid ())) return getenv (name); else  return NULL;\n}\" &gt; getenv.cc\ngcc -I$PWD\/my_include  -std=c++11 -fpermissive -s -shared -o    ${MYSTUBS}\/getenv.so -fPIC -fno-builtin getenv.cc\n<\/code><\/pre>\n\n<p>Of course the final running code should be changed to include additional stubs. <\/p>\n\n<pre><code>LD_PRELOAD=\"$MYSTUBS\/list.so:$MYSTUBS\/mylibc.so:$MYSTUBS\/shared_ptr.so:$MYSTUBS\/getenv.so:$MYSTUBS\/random.so:$MYSTUBS\/hash_bytes.so:$MYSTUBS\/chrono.so:$MYSTUBS\/hashtable.so:$MYSTUBS\/bad_function.so:$LIBSTDCPP\" python  ${TENSORFLOW_DIR}\/models\/image\/mnist\/convolutional.py\n<\/code><\/pre>\n"}]}
{"QuestionId":33656551,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-11T17:42:11.957","AcceptedAnswerId":33691154.0,"OwnerUserId":4487457.0,"Title":"Unable to import Tensorflow \"No module named copyreg\"","Body":"<p>El Capitan OS here. I've been trying to find a workaround with import Tensorflow into my ipython notebook, but so far no luck.<\/p>\n\n<p>Like many people in the forums, I've also had issues with install tensorflow because of the six package. I was able to install after some fidgeting with brew<\/p>\n\n<pre><\/pre>\n\n<p>I got a message that tensorflow was installed correctly. Even when I did  I get the message:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I'm on my ipython notebook and I did an  I get the message: <\/p>\n\n<p>I've dug further and found this error on the import as well:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33805088","CreationDate":"2015-11-19T13:19:20.937","ParentId":null,"OwnerUserId":"2470083","Title":null,"Body":"<p>Another simple way I tried to fix this is to edit the source code and <code>python_message.py<\/code> and change the line <code>import six.moves.copyreg as copyreg<\/code> into <code>from six.moves import copyreg<\/code>.<\/p>\n\n<p>Of course you might need to change the permission of this read-only file in order to modify it. <\/p>\n"},{"AnswerId":"33659659","CreationDate":"2015-11-11T20:56:01.467","ParentId":null,"OwnerUserId":"3718836","Title":null,"Body":"<p>Copyreg can be found in the six module in Python2.<\/p>\n\n<p>Short answer, try upgrading\/installing six:<\/p>\n\n<pre><code>pip install -U six\n<\/code><\/pre>\n"},{"AnswerId":"33663662","CreationDate":"2015-11-12T02:59:33.510","ParentId":null,"OwnerUserId":"4068846","Title":null,"Body":"<p>Maybe try reinstalling six. It looks like it is in a different location from your other libraries. <code>\/Library\/Python\/...<\/code> vs <code>\/usr\/local\/lib\/...<\/code> This probably means that six was install before homebrew python. <\/p>\n"},{"AnswerId":"33691154","CreationDate":"2015-11-13T10:56:51.880","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>As Jonah commented, it's solved by this:<\/p>\n\n<p>On MacOSX<\/p>\n\n<p>If you encounter:<\/p>\n\n<pre><code>import six.moves.copyreg as copyreg\n<\/code><\/pre>\n\n<pre class=\"lang-none prettyprint-override\"><code>ImportError: No module named copyreg\n<\/code><\/pre>\n\n<p>Solution: TensorFlow depends on protobuf which requires six-1.10.0. Apple's default python environment has six-1.4.1 and may be difficult to upgrade. So we recommend either installing a separate copy of python via homebrew:<\/p>\n\n<pre><code>brew install python\n<\/code><\/pre>\n\n<p>But I highly recommend using virtualenv for this purpose.<\/p>\n\n<pre><code># On Mac:\n$ sudo easy_install pip  # If pip is not already installed\n$ sudo pip install --upgrade virtualenv\n<\/code><\/pre>\n\n<p>Next, set up a new virtualenv environment. To set it up in the directory <code>~\/tensorflow<\/code>, run:<\/p>\n\n<pre><code>$ virtualenv --system-site-packages ~\/tensorflow\n$ cd ~\/tensorflow\n<\/code><\/pre>\n\n<p>Then activate the virtualenv:<\/p>\n\n<pre><code>$ source bin\/activate  # If using bash\n$ source bin\/activate.csh  # If using csh\n(tensorflow)$  # Your prompt should change\n<\/code><\/pre>\n\n<p>Inside the virtualenv, install TensorFlow:<\/p>\n\n<pre><code>(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>You can then run your TensorFlow program like:<\/p>\n\n<pre><code>(tensorflow)$ python tensorflow\/models\/image\/mnist\/convolutional.py\n\n# When you are done using TensorFlow:\n(tensorflow)$ deactivate  # Deactivate the virtualenv\n\n$  # Your prompt should change back\n<\/code><\/pre>\n"}]}
{"QuestionId":33656689,"AnswerCount":2,"Tags":"<macos><cuda><virtual-machine><gpu><tensorflow>","CreationDate":"2015-11-11T17:50:32.057","AcceptedAnswerId":33664156.0,"OwnerUserId":1185578.0,"Title":"Would TensorFlow utilize GPU on a Mac if installed on a VM?","Body":"<p>From TensorFlow's <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md\" rel=\"noreferrer\">\"Getting Started\"<\/a> page:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not super familiar with using GPU or CUDA libraries, but if I installed TensorFlow inside a Linux VM (say the <a href=\"http:\/\/files.vagrantup.com\/precise32.box\" rel=\"noreferrer\">precise32<\/a> available through Vagrant), then would TensorFlow utilize the GPU when running inside that VM?<\/p>\n","answers":[{"AnswerId":"33664156","CreationDate":"2015-11-12T03:52:22.303","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Probably not.  VirtualBox, for example, <a href=\"https:\/\/www.virtualbox.org\/manual\/ch14.html\">does not support PCI Passthrough on a MacOS host, only a Linux host<\/a> (and even then, I'd... uh, not get my hopes up).  MacOS ends up so tightly integrated with its GPU(s) that I'd be very dubious that any VM can do it at this point.<\/p>\n"},{"AnswerId":"36995253","CreationDate":"2016-05-03T03:46:50.837","ParentId":null,"OwnerUserId":"4494","Title":null,"Body":"<p>As an update: Tensorflow can now use GPUs on Mac OS X. The relevant PR is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/664\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/664<\/a> and after a <code>brew install coreutils<\/code> the Linux installation 'build from source' instructions should work. I see a 10x speedup compared to the CPU version with an NVIDIA gforce 960 and Intel i7-6700K.<\/p>\n"}]}
{"QuestionId":33657722,"AnswerCount":1,"Tags":"<deep-learning><tensorflow>","CreationDate":"2015-11-11T18:49:07.920","AcceptedAnswerId":33708257.0,"OwnerUserId":604964.0,"Title":"Can't understand shape(output) = (shape(value) - ksize + 1) \/ strides in TensorFlow docs","Body":"<p>In the following excerpt from <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/nn.md#pooling\" rel=\"nofollow\">http:\/\/tensorflow.org\/api_docs\/python\/nn.md#pooling<\/a><\/p>\n\n<p>shape(output) = (shape(value) - ksize + 1) \/ strides<\/p>\n\n<p>where the rounding direction depends on padding:<\/p>\n\n<ul>\n<li><p>padding = 'SAME': Round down (only full size windows are considered).<\/p><\/li>\n<li><p>padding = 'VALID': Round up (partial windows are included).<\/p><\/li>\n<\/ul>\n\n<p>I can't understand the formula above. I am familiar with following formula though: <\/p>\n\n<p>shape(out) = (shape(value) - ksize + 2*pad)\/strides+1.<\/p>\n\n<p>Are the two formulas equivalent?<\/p>\n\n<p>For example, let's say shape(value) = 9, ksize = 3, strides = 2, and padding = 'SAME'.<\/p>\n\n<p>In the first formula the shape(output) will be (9-3+1)\/2 = 7\/2 = 3.5 and rounding down results in 3.<\/p>\n\n<p>In the second formula the shape(output) will be (9-3+2*1)\/2 + 1 = 5<\/p>\n\n<p>It doesn't seem to be the same formula. Even if I round up the first one, the result will be 4.<\/p>\n\n<p>In addition to that, the padding definition seems to be inverted. Isn't the 'SAME' padding that includes partial windows?<\/p>\n","answers":[{"AnswerId":"33708257","CreationDate":"2015-11-14T12:11:44.590","ParentId":null,"OwnerUserId":"604964","Title":null,"Body":"<p>I think I figured out the problem. The two formulas become equivalent if you assume that the appropriate padding is already included in shape(value). But I still think the definition of the padding types have been swapped in the documentation. I created an issue to report that: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/196\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/196<\/a><\/p>\n"}]}
{"QuestionId":33659047,"AnswerCount":4,"Tags":"<python><gpu><nvidia><tensorflow>","CreationDate":"2015-11-11T20:15:14.900","AcceptedAnswerId":36089757.0,"OwnerUserId":4802797.0,"Title":"where is the .\/configure of TensorFlow and how to enable the GPU support?","Body":"<p>When installing TensorFlow on my Ubuntu, I would like to use GPU with CUDA. <\/p>\n\n<p>But I am stopped at this step in the <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md\" rel=\"nofollow noreferrer\">Official Tutorial<\/a> :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dmwOC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dmwOC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Where exactly is this  ? Or where is my root of source tree.<\/p>\n\n<p>My TensorFlow is located here .   But I still did not find .  <\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I have found the  according to <a href=\"https:\/\/stackoverflow.com\/a\/33659119\/4802797\">Salvador Dali's answer<\/a>. But when doing the example code, I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The cuda device cannot be found. <\/p>\n\n<p><strong>Answer<\/strong><\/p>\n\n<p>See the answer about how did I enable GPU support <a href=\"https:\/\/stackoverflow.com\/a\/36089757\/4802797\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"36089757","CreationDate":"2016-03-18T16:39:04.560","ParentId":null,"OwnerUserId":"4802797","Title":null,"Body":"<ul>\n<li><p>Answer to first question: <code>.\/configure<\/code> has already been found according to the answer <a href=\"https:\/\/stackoverflow.com\/a\/33659119\/4802797\">here<\/a>. It is under the source folder of <code>tensorflow<\/code> as shown <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/configure\" rel=\"nofollow noreferrer\">here<\/a>. <\/p><\/li>\n<li><p>Answer to second question: <\/p><\/li>\n<\/ul>\n\n<p>Actually, I have the GPU <code>NVIDIA Corporation GK208GLM [Quadro K610M]<\/code>. I also have <code>CUDA<\/code> + <code>cuDNN<\/code> installed. (Therefore, the following answer is based on you have already installed <code>CUDA 7.0+<\/code> + <code>cuDNN<\/code> correctly with the correct versions.) However the problem is: I have driver installed but the GPU is just not working. I made it working in the following steps:<\/p>\n\n<p>At first, I did this <code>lspci<\/code> and got: <\/p>\n\n<pre><code>01:00.0 VGA compatible controller: NVIDIA Corporation GK208GLM [Quadro K610M] (rev ff)\n<\/code><\/pre>\n\n<p>The status here is <strong>rev ff<\/strong>. Then, I did <code>sudo update-pciids<\/code>, and check <code>lspci<\/code> again, and got:<\/p>\n\n<pre><code>01:00.0 VGA compatible controller: NVIDIA Corporation GK208GLM [Quadro K610M] (rev a1)\n<\/code><\/pre>\n\n<p>Now, the status of Nvidia GPU is correct as <strong>rev a1<\/strong>. But now, the <code>tensorflow<\/code> is not supporting GPU yet. The next steps are (the Nvidia driver I installed is version <code>nvidia-352<\/code>): <\/p>\n\n<pre><code>sudo modprobe nvidia_352\nsudo modprobe nvidia_352_uvm\n<\/code><\/pre>\n\n<p>in order to add the driver into correct mode. Check again: <\/p>\n\n<pre><code>cliu@cliu-ubuntu:~$ lspci -vnn | grep -i VGA -A 12\n01:00.0 VGA compatible controller [0300]: NVIDIA Corporation GK208GLM [Quadro K610M] [10de:12b9] (rev a1) (prog-if 00 [VGA controller])\n    Subsystem: Hewlett-Packard Company Device [103c:1909]\n    Flags: bus master, fast devsel, latency 0, IRQ 16\n    Memory at cb000000 (32-bit, non-prefetchable) [size=16M]\n    Memory at 50000000 (64-bit, prefetchable) [size=256M]\n    Memory at 60000000 (64-bit, prefetchable) [size=32M]\n    I\/O ports at 5000 [size=128]\n    Expansion ROM at cc000000 [disabled] [size=512K]\n    Capabilities: &lt;access denied&gt;\n    Kernel driver in use: nvidia\ncliu@cliu-ubuntu:~$ lsmod | grep nvidia\nnvidia_uvm             77824  0 \nnvidia               8646656  1 nvidia_uvm\ndrm                   348160  7 i915,drm_kms_helper,nvidia\n<\/code><\/pre>\n\n<p>We can find that the <code>Kernel driver in use: nvidia<\/code> is shown and <code>nvidia<\/code> is in correct mode. <\/p>\n\n<p>Now, use the example <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/using_gpu\/index.html#logging-device-placement\" rel=\"nofollow noreferrer\">here<\/a> for testing the GPU: <\/p>\n\n<pre><code>cliu@cliu-ubuntu:~$ python\nPython 2.7.9 (default, Apr  2 2015, 15:33:21) \n[GCC 4.9.2] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n&gt;&gt;&gt; b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n&gt;&gt;&gt; c = tf.matmul(a, b)\n&gt;&gt;&gt; sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\nI tensorflow\/core\/common_runtime\/local_device.cc:25] Local device intra op parallelism threads: 8\nI tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:888] 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_init.cc:88] Found device 0 with properties: \nname: Quadro K610M\nmajor: 3 minor: 5 memoryClockRate (GHz) 0.954\npciBusID 0000:01:00.0\nTotal memory: 1023.81MiB\nFree memory: 1007.66MiB\nI tensorflow\/core\/common_runtime\/gpu\/gpu_init.cc:112] DMA: 0 \nI tensorflow\/core\/common_runtime\/gpu\/gpu_init.cc:122] 0:   Y \nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:643] Creating TensorFlow device (\/gpu:0) -&gt; (device: 0, name: Quadro K610M, pci bus id: 0000:01:00.0)\nI tensorflow\/core\/common_runtime\/gpu\/gpu_region_allocator.cc:47] Setting region size to 846897152\nI tensorflow\/core\/common_runtime\/local_session.cc:45] Local session inter op parallelism threads: 8\nDevice mapping:\n\/job:localhost\/replica:0\/task:0\/gpu:0 -&gt; device: 0, name: Quadro K610M, pci bus id: 0000:01:00.0\nI tensorflow\/core\/common_runtime\/local_session.cc:107] Device mapping:\n\/job:localhost\/replica:0\/task:0\/gpu:0 -&gt; device: 0, name: Quadro K610M, pci bus id: 0000:01:00.0\n\n&gt;&gt;&gt; print sess.run(c)\nb: \/job:localhost\/replica:0\/task:0\/gpu:0\nI tensorflow\/core\/common_runtime\/simple_placer.cc:289] b: \/job:localhost\/replica:0\/task:0\/gpu:0\na: \/job:localhost\/replica:0\/task:0\/gpu:0\nI tensorflow\/core\/common_runtime\/simple_placer.cc:289] a: \/job:localhost\/replica:0\/task:0\/gpu:0\nMatMul: \/job:localhost\/replica:0\/task:0\/gpu:0\nI tensorflow\/core\/common_runtime\/simple_placer.cc:289] MatMul: \/job:localhost\/replica:0\/task:0\/gpu:0\n[[ 22.  28.]\n [ 49.  64.]]\n<\/code><\/pre>\n\n<p>As you can see, the GPU is utilized. <\/p>\n"},{"AnswerId":"33662595","CreationDate":"2015-11-12T00:45:13.093","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>For your second question:  Do you have a compatible GPU (NVIDIA compute capability 3.5 or higher) installed, and do you have CUDA 7.0 + cuDNN installed as per the instructions?  That's the most likely reason you're seeing a failure.  It could be a cuda install problem if the answer is yes.  Do you see your GPU listed when you run nvidia-smi?  If not, you need to get that fixed first.  This may require getting a newer driver and\/or re-running nvidia-xconfig, etc.<\/p>\n"},{"AnswerId":"33724855","CreationDate":"2015-11-15T20:47:54.747","ParentId":null,"OwnerUserId":"5564902","Title":null,"Body":"<p>you can rebuild a GPU version from the source only if you have the 7.0 cuda libraries and the 6.5 cudnn libraries.\nthis need to be updated by google, I think<\/p>\n"},{"AnswerId":"33659119","CreationDate":"2015-11-11T20:21:23.277","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This is a bash script which suppose to be in <\/p>\n\n<blockquote>\n  <p>the root of your source tree<\/p>\n<\/blockquote>\n\n<p>when you <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\">cloned the repo<\/a>. Here it is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/configure\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/configure<\/a><\/p>\n"}]}
{"QuestionId":33659424,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-11T20:40:17.040","AcceptedAnswerId":33662396.0,"OwnerUserId":5552083.0,"Title":"TensorFlow MNIST example not running with fully_connected_feed.py","Body":"<p>I am able to run the  fine, but when running , I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am new to Python so could also just be a general setup problem.<\/p>\n","answers":[{"AnswerId":"33662396","CreationDate":"2015-11-12T00:26:11.227","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is a Python path issue. Assuming that the directory <code>tensorflow\/g3doc\/tutorials\/mnist<\/code> is your current working directory (or in your Python path), the easiest way to resolve it is to change the following lines in fully_connected_feed.py from:<\/p>\n\n<pre><code>from tensorflow.g3doc.tutorials.mnist import input_data\nfrom tensorflow.g3doc.tutorials.mnist import mnist\n<\/code><\/pre>\n\n<p>...to:<\/p>\n\n<pre><code>import input_data\nimport mnist\n<\/code><\/pre>\n"},{"AnswerId":"33682776","CreationDate":"2015-11-12T22:39:14.970","ParentId":null,"OwnerUserId":"4275865","Title":null,"Body":"<p>Another alternative is to link the 'g3doc' directory from the github repo into the tensorflow python wheel folder. That way you don't need to change the code.<\/p>\n"}]}
{"QuestionId":33662648,"AnswerCount":1,"Tags":"<python><logging><tensorflow>","CreationDate":"2015-11-12T00:51:25.543","AcceptedAnswerId":33664610.0,"OwnerUserId":2381279.0,"Title":"Tensorflow causes logging messages to double","Body":"<p>So I was playing around with Google's <a href=\"http:\/\/www.tensorflow.org\/\" rel=\"noreferrer\">Tensorflow<\/a> library they published yesterday and encountered an annoying bug that keeps biting me.<\/p>\n\n<p>What I did was setup the python logging functions as I usually do, and the result was that, if I import the tensorflow library, all messages in the console started doubling. Interestingly, this does <em>not<\/em> happen if you just use the  function.<\/p>\n\n<p>An example of a code that does <em>not<\/em> double the messages:<\/p>\n\n<pre><\/pre>\n\n<p>An example of a code that <em>does<\/em> double all messages:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I'm a simple man. I like the functionality of , so I use it. The setup with the  object and the adding of a  is something I picked up looking at how other people did this, but it looks like it fits with how the thing was meant to be used. However, I do not have in-depth knowledge of the logging library, as it always just kind of worked.<\/p>\n\n<p>So, any help explaining why the doubling of the messages occurs will be most helpful.<\/p>\n\n<p>I am using Ubuntu 14.04.3 LTS with Python 2.7.6, but the error happens in all Python 2.7 versions I tried.<\/p>\n","answers":[{"AnswerId":"33664610","CreationDate":"2015-11-12T04:45:29.343","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I get this output:<\/p>\n\n<pre><code>test\nWARNING:TEST:test\n<\/code><\/pre>\n\n<p>Tensorflow is <em>also<\/em> using the logging framework and has set up its own handlers, so when you log, by default, it propagates up to the parent logging handlers inside tensorflow.  You can change this behavior by setting:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logger.propagate = False\n<\/code><\/pre>\n\n<p>See also <a href=\"https:\/\/stackoverflow.com\/questions\/19561058\/duplicate-output-in-simple-python-logging-configuration\">duplicate output in simple python logging configuration<\/a><\/p>\n\n<p>Followup: This was an unintended side-effect of the way tensorflow was using the logging package. I've changed it at HEAD to scope its internal loggers under the name \"tensorflow\" to avoid this pollution. Should be in the github head within a day or so. In the meantime, the logger.propagate solution will work and won't break once that fix is in, so you should be safe to go. Thanks again for spotting this!<\/p>\n\n<p>Followup-Followup:\nStarting with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_logger\" rel=\"nofollow noreferrer\">TensorFlow 1.14<\/a> exposes the <code>logger<\/code> directly:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nlogger = tf.get_logger()\n<\/code><\/pre>\n"}]}
{"QuestionId":33663081,"AnswerCount":7,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-11-12T01:43:23.693","AcceptedAnswerId":null,"OwnerUserId":2629530.0,"Title":"How to create a Tensorflow Tensorboard Empty Graph","Body":"<p>launch tensorboard with <\/p>\n\n<p>at tensorboard:6006 > graph, it says No graph definition files were found.<\/p>\n\n<p>To store a graph, create a tf.python.training.summary_io.SummaryWriter and pass the graph either via the constructor, or by calling its add_graph() method.<\/p>\n\n<pre><\/pre>\n\n<p>However the page is still empty, how can I start playing with tensorboard?<\/p>\n\n<h1>current tensorboard<\/h1>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UB8dX.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UB8dX.png\" alt=\"Current Tensorboard\"><\/a><\/p>\n\n<h1>result wanted<\/h1>\n\n<p>An empty graph that can add nodes, editable.<\/p>\n\n<h1>update<\/h1>\n\n<p>Seems like tensorboard is unable to create a graph to add nodes, drag and edit etc ( I am confused by the official video ).<\/p>\n\n<p>running <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/mnist\/fully_connected_feed.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/mnist\/fully_connected_feed.py<\/a> and then  is able to view the graph<\/p>\n\n<p>However seems like tensorflow only provide ability to view summary, nothing much different to make it standout<\/p>\n","answers":[{"AnswerId":"42452126","CreationDate":"2017-02-25T04:56:21.030","ParentId":null,"OwnerUserId":"7473616","Title":null,"Body":"<p><strong>i solved by on windows:<\/strong><\/p>\n\n<pre><code>       file_writer = tf.summary.FileWriter(\"output\", sess.graph)\n<\/code><\/pre>\n\n<p>for that directory \"output\". I opened command on windows. <\/p>\n\n<p>typed<\/p>\n\n<pre><code>tensorboard --logdir=\"C:\\Users\\kiran\\machine Learning\\output\"\n<\/code><\/pre>\n\n<p>my mistake was on that line..<\/p>\n"},{"AnswerId":"42193017","CreationDate":"2017-02-12T20:44:39.307","ParentId":null,"OwnerUserId":"6183298","Title":null,"Body":"<blockquote>\n  <p><strong>result wanted<\/strong><\/p>\n  \n  <p>An empty graph that can add nodes, editable.<\/p>\n<\/blockquote>\n\n<p>I think you will find the <a href=\"http:\/\/orange.biolab.si\/\" rel=\"nofollow noreferrer\" title=\"Orange\">Orange<\/a> tool useful. It allows you to drag and drop various nodes and implement algorithms via GUI.<\/p>\n"},{"AnswerId":"35302468","CreationDate":"2016-02-09T21:24:50.693","ParentId":null,"OwnerUserId":"4586307","Title":null,"Body":"<p>The graphs in TensorBoard do not show up if you are using Firefox. You have to install Chrome. <\/p>\n"},{"AnswerId":"33702335","CreationDate":"2015-11-13T21:49:43.470","ParentId":null,"OwnerUserId":"2065691","Title":null,"Body":"<p>In this page, there is a very simple code that you can use to test your installation: <a href=\"http:\/\/tensorflow.org\/get_started\" rel=\"nofollow noreferrer\">http:\/\/tensorflow.org\/get_started<\/a><\/p>\n\n<p>I included this line <\/p>\n\n<pre><code>tf.train.write_graph(sess.graph_def, '\/home\/daniel\/Documents\/Projetos\/Prorum\/ProgramasEmPython\/TestingTensorFlow\/fileGraph', 'graph.pbtxt')\n<\/code><\/pre>\n\n<p>After this \"sess.run(init)\" <\/p>\n\n<p>This will generate a file that you have to upload to the \"TensorBoard\".<\/p>\n\n<p>In order to open the TensorBoard, supposing that it is installed in your computer (it must be if you use pip to install), I used the terminal of Ubuntu and wrote:<\/p>\n\n<p>\"tensorboard --logdir nameOfDirectory\"<\/p>\n\n<p>Then, you should open your browser in Port 6006:<\/p>\n\n<pre><code>http:\/\/localhost:6006\/\n<\/code><\/pre>\n\n<p>This will open the TensorBoard. I went to the \"Graph Menu\" and uploaded the file. It generated this figure below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wB7tP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wB7tP.png\" alt=\"http:\/\/www.prorum.com\"><\/a><\/p>\n\n<p>So, what I have done is to transfer the model I created in Python to TensorBoard. I believe that it is possible to create an empty one, if no model is created (only the session is initiated). However, I am not sure if you are able to change this directly in the TensorBoard.<\/p>\n\n<p>I have answered before this question here in Portuguese with more details for Brazilian users. Maybe it can be useful for other people: <a href=\"http:\/\/prorum.com\/index.php\/1843\/recentemente-plataforma-aprendizagem-primeira-impressao\" rel=\"nofollow noreferrer\">http:\/\/prorum.com\/index.php\/1843\/recentemente-plataforma-aprendizagem-primeira-impressao<\/a><\/p>\n"},{"AnswerId":"33685349","CreationDate":"2015-11-13T03:15:26.087","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorBoard is a tool for <a href=\"http:\/\/tensorflow.org\/how_tos\/graph_viz\/index.md\">visualizing the TensorFlow graph<\/a> and analyzing recorded metrics during training and inference. The graph is created using the Python API, then written out using the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.md#SummaryWriter.add_graph\"><code>tf.train.SummaryWriter.add_graph()<\/code><\/a> method. When you load the file written by the SummaryWriter into TensorBoard, you can see the graph that was saved, and interactively explore it.<\/p>\n\n<p>However, TensorBoard is not a tool for <em>building<\/em> the graph itself. It does not have any support for adding nodes to the graph.<\/p>\n"},{"AnswerId":"34461172","CreationDate":"2015-12-25T08:14:31.970","ParentId":null,"OwnerUserId":"1092746","Title":null,"Body":"<p>Starting from the following <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/index.html\" rel=\"noreferrer\">Code Example<\/a>, I can add one line as shown below:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nsess = tf.InteractiveSession()  #define a session\n# Create 100 phony x, y data points in NumPy, y = x * 0.1 + 0.3\nx_data = np.random.rand(100).astype(\"float32\")\ny_data = x_data * 0.1 + 0.3\n\n# Try to find values for W and b that compute y_data = W * x_data + b\n# (We know that W should be 0.1 and b 0.3, but Tensorflow will\n# figure that out for us.)\nW = tf.Variable(tf.random_uniform([1], -1.0, 1.0))\nb = tf.Variable(tf.zeros([1]))\ny = W * x_data + b\n\n# Minimize the mean squared errors.\nloss = tf.reduce_mean(tf.square(y - y_data))\noptimizer = tf.train.GradientDescentOptimizer(0.5)\ntrain = optimizer.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\n#### ----&gt; ADD THIS LINE &lt;---- ####\nwriter = tf.train.SummaryWriter(\"\/tmp\/test\", sess.graph)\n\n# Fit the line.\nfor step in xrange(201):\n    sess.run(train)\n    if step % 20 == 0:\n        print(step, sess.run(W), sess.run(b))\n\n# Learns best fit is W: [0.1], b: [0.3]\n<\/code><\/pre>\n\n<p>And then run tensorboard from the command line, pointing to the appropriate directory.  This shows a complete call for the SummaryWriter.  It is important to note the following things:<\/p>\n\n<ol>\n<li>SummaryWriter is passed a Session, and so must happen after the Session (or InteractiveSession) is created<\/li>\n<li>That Session may be created early in the program, but when the Session is passed to the SummaryWriter, the graph <em>as it exists at that point<\/em> is written to the file that the TensorBoard will use.  <\/li>\n<\/ol>\n"},{"AnswerId":"44608710","CreationDate":"2017-06-17T19:39:55.973","ParentId":null,"OwnerUserId":"5595995","Title":null,"Body":"<p>I had to use <\/p>\n\n<pre><code>python -m tensorflow.tensorboard --logdir=\"C:\\tmp\\tensorflow\\..\"\n<\/code><\/pre>\n\n<p>somehow <code>tensorboard --logdir<\/code> didn't work. <\/p>\n\n<p>My environment <\/p>\n\n<p>OS: Windows 7, Python 3.5, and Tensorflow 1.1.0<\/p>\n"}]}
{"QuestionId":33663980,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-12T03:31:43.627","AcceptedAnswerId":33664544.0,"OwnerUserId":4068846.0,"Title":"How to run tensor flow seq2seq demo","Body":"\n\n<p>I tensor flow installed and successfully went through the MNIST demo. Now, I am trying to run the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/seq2seq\/index.md#\" rel=\"nofollow noreferrer\">seq2seq demo<\/a>, but this is not working for me. <\/p>\n\n<p>I cloned a version of their github repo and attempted to run some of the listed commands from the repo root.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>No surprise here, as it doesn't really make sense to have bazel execute a python script. <\/p>\n\n<p>Later in the tutorial,<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>If I remove the parameters from the invocation above, it will attempt (and fail) to build the entire tensor flow project before it executes . This is not what I want as I have already successfully installed tensor flow with pip. <\/p>\n\n<p>The last thing I tried running,<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Environment info: OS X 10.11.1, Python 2.7.10 (anaconda)<\/p>\n","answers":[{"AnswerId":"33664544","CreationDate":"2015-11-12T04:37:28.377","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>There are two ways to run the script:<\/p>\n\n<p>1) separate the script arguments with -- as part of bazel run<\/p>\n\n<pre><code>bazel run -c opt \/\/tensorflow\/models\/rnn\/translate:translate -- \\\n--data_dir .\/data_dir --train_dir .\/checkpoints_directory \\\n--en_vocab_size=40000 --fr_vocab_size=40000\n<\/code><\/pre>\n\n<p>2) build and then run from <code>.\/bazel-bin\/<\/code>:<\/p>\n\n<pre><code>bazel build -c opt \/\/tensorflow\/models\/rnn\/translate:translate\n\n.\/bazel-bin\/tensorflow\/models\/rnn\/translate\/translate \\\n--data_dir .\/data_dir --train_dir .\/checkpoints_directory \\\n--en_vocab_size=40000 --fr_vocab_size=40000\n<\/code><\/pre>\n"}]}
{"QuestionId":33664651,"AnswerCount":9,"Tags":"<python><import><machine-learning><tensorflow><mnist>","CreationDate":"2015-11-12T04:51:26.380","AcceptedAnswerId":33664749.0,"OwnerUserId":678572.0,"Title":"import input_data MNIST tensorflow not working","Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/33659424\/tensorflow-mnist-example-not-running\">TensorFlow MNIST example not running with fully_connected_feed.py<\/a><\/p>\n\n<p>I checked this out and realized that  was not built-in.  So I downloaded the whole folder from <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py\" rel=\"noreferrer\">here<\/a>. How can I start the tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>I'm using iPython (Jupyter) so do I need to change my working directory to this folder I downloaded? or can I add this to my  directory? If so, where do I add the files? I installed  with  (on my OSX) and the current location is <\/p>\n\n<p>Are these files meant to be accessed directly through  like  datasets? or am I just supposed to cd into the directory and work from there? The example is not clear. <\/p>\n","answers":[{"AnswerId":"43170104","CreationDate":"2017-04-02T15:29:52.657","ParentId":null,"OwnerUserId":"6400639","Title":null,"Body":"<p>MNIST input_data was built-in, it's just not a individual module, it's inside Tensorflow module, try <\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\n<\/code><\/pre>\n"},{"AnswerId":"37540230","CreationDate":"2016-05-31T08:08:40.663","ParentId":null,"OwnerUserId":"4049614","Title":null,"Body":"<p>The old tutorial said, to import the MNIST data, use:<\/p>\n\n<pre><code>import input_data\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n<\/code><\/pre>\n\n<p>This will cause the error.\nThe new tutorial uses the following code to do so:<\/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\n<p>And this works well.<\/p>\n"},{"AnswerId":"35192289","CreationDate":"2016-02-04T04:04:55.257","ParentId":null,"OwnerUserId":"1332841","Title":null,"Body":"<blockquote>\n  <p>How can I start the tutorial <\/p>\n<\/blockquote>\n\n<p>I didn't download the folder you did but I installed tensorflow by pip and then I had similar problem.<\/p>\n\n<p>My workaround was to replace <\/p>\n\n<p><code>import tensorflow.examples.tutorials.mnist.input_data<\/code><\/p>\n\n<p>with<\/p>\n\n<p><code>import tensorflow.examples.tutorials.mnist.input_data as input_data<\/code><\/p>\n"},{"AnswerId":"41886668","CreationDate":"2017-01-27T03:43:56.493","ParentId":null,"OwnerUserId":"2875928","Title":null,"Body":"<p>I might be kinda late, but for tensorflow version 0.12.1, you might wanna use input_data.read_data_sets instead.<\/p>\n\n<p>Basically using this function to load the data from your local drive that you had downloaded from <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a>.<\/p>\n\n<p><code>from tensorflow.examples.tutorials.mnist import input_data<\/code>\n<code>mnist = input_data.read_data_sets('data_set\/')<\/code><\/p>\n"},{"AnswerId":"42639770","CreationDate":"2017-03-07T03:41:41.670","ParentId":null,"OwnerUserId":"243072","Title":null,"Body":"<pre><code>cd your_mnist_dir &amp;&amp;\\\nwget https:\/\/github.com\/HIPS\/hypergrad\/raw\/master\/data\/mnist\/mnist_data.pkl &amp;&amp;\\\nwget https:\/\/github.com\/HIPS\/hypergrad\/raw\/master\/data\/mnist\/t10k-images-idx3-ubyte.gz &amp;&amp;\\\nwget https:\/\/github.com\/HIPS\/hypergrad\/raw\/master\/data\/mnist\/t10k-labels-idx1-ubyte.gz &amp;&amp;\\\nwget https:\/\/github.com\/HIPS\/hypergrad\/raw\/master\/data\/mnist\/train-images-idx3-ubyte.gz &amp;&amp;\\\nwget https:\/\/github.com\/HIPS\/hypergrad\/raw\/master\/data\/mnist\/train-labels-idx1-ubyte.gz\n<\/code><\/pre>\n"},{"AnswerId":"49475195","CreationDate":"2018-03-25T11:16:32.270","ParentId":null,"OwnerUserId":"6201599","Title":null,"Body":"<p>As TensorFlow official website shown, All MNIST data is hosted on <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GGjPF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GGjPF.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"34112569","CreationDate":"2015-12-06T00:27:54.103","ParentId":null,"OwnerUserId":"5645276","Title":null,"Body":"<p>I am using different version - following Install on Windows with Docker <a href=\"http:\/\/www.netinstructions.com\/how-to-install-and-run-tensorflow-on-a-windows-pc\/\" rel=\"nofollow\">here<\/a> - and had similar problem.<\/p>\n\n<p>An easy workaround I've found was:<\/p>\n\n<p>1.Into the Linux command line, figure out where is the input_data.py on my Docker image (in your case you mentionned that you had to download it manually. In my case, it was already here). I used the follwing linux command:<\/p>\n\n<pre><code>$ sudo find . -print | grep -i '.*[.]py'\n<\/code><\/pre>\n\n<p>I've got the files &amp; path<\/p>\n\n<pre><code>.\/tensorflow\/g3doc\/tutorials\/mnist\/mnist.py\n.\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py\n<\/code><\/pre>\n\n<p>2.launch Python and type the following command using SYS:<\/p>\n\n<pre><code>&gt;&gt; import sys\n&gt;&gt; print(sys.path)\n<\/code><\/pre>\n\n<p>you will get the existing paths.<\/p>\n\n<pre><code>['', '\/usr\/lib\/python2.7', '\/usr\/lib\/python2.7\/plat-x86_64-linux-gnu', '\/usr\/lib\/python2.7\/lib-tk', '\/usr\/lib\/python2.7\/lib-old', '\/usr\/lib\/python2.7\/lib-dynload', '\/usr\/local\/lib\/python2.7\/dist-packages', '\/usr\/lib\/python2.7\/dist-packages', '\/usr\/lib\/python2.7\/dist-packages\/PILcompat']\n<\/code><\/pre>\n\n<p>4.add the path of inputa_data.py:<\/p>\n\n<pre><code>&gt;&gt; sys.path.insert(1,'\/tensorflow\/tensorflow\/g3doc\/tutorials\/mnist')\n<\/code><\/pre>\n\n<p>Hope that it can help. If you found better option, let me know. :)<\/p>\n"},{"AnswerId":"33664749","CreationDate":"2015-11-12T05:01:10.703","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>So let's assume that you are in the directory: <code>\/somePath\/tensorflow\/tutorial<\/code> (and this is your working directory).<\/p>\n\n<p>All you need to do is to download the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/input_data.py\">input_data.py<\/a> and put it this. Let the file name where you invoke:<\/p>\n\n<pre><code>import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n...\n<\/code><\/pre>\n\n<p>is <code>main.py<\/code> and it is also in this directory.<\/p>\n\n<p>Whenever this is done, you can just start running <code>main.py<\/code> which will start downloading the files and will put them in the MNIST_data folder (once they are there the script will not be downloading them next time).<\/p>\n"},{"AnswerId":"46536917","CreationDate":"2017-10-03T03:47:54.127","ParentId":null,"OwnerUserId":"5106574","Title":null,"Body":"<p>MNIST data set included as a part of tensorflow examples tutorial, If we want to use this :<\/p>\n\n<h1>Import MNIST data to identify handwritten digites<\/h1>\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":33665749,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2015-11-12T06:48:37.310","AcceptedAnswerId":null,"OwnerUserId":2015954.0,"Title":"Error when running tensorflow","Body":"<p>I am trying to run the first example (\"MNIST For ML Beginners\") from tensorflow tutorial with Python 2.7.10 and getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>If I rebuild Python with <\/p>\n\n<blockquote>\n  <p>--enable-unicode=ucs4<\/p>\n<\/blockquote>\n\n<p>I am getting another problem:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know how I can workaround this?<\/p>\n","answers":[]}
{"QuestionId":33669719,"AnswerCount":0,"Tags":"<python><profiling><theano>","CreationDate":"2015-11-12T10:48:59.270","AcceptedAnswerId":null,"OwnerUserId":988125.0,"Title":"Theano profiling summary","Body":"<p>I am trying to learn how to profile in Theano. Below is some very basic code.<\/p>\n\n<p>I set  to 1 as advised. Also, I am not sure if it is necessary but I set  as well. This prints some output when the code completes running. But how do I capture \/ print \/ log this so I can actually analyse the results?<\/p>\n\n<p>I can see from the web documentation that in the deprecated profiling approach with <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/profilemode.html\" rel=\"nofollow\">profilemode<\/a>, there is a handy  function. Unfortunately the current profiling is not as well documented as the deprecated one.<\/p>\n\n<p>How can I analyse the profiling results for these two functions?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33671372,"AnswerCount":1,"Tags":"<python><linux><cuda><tensorflow>","CreationDate":"2015-11-12T12:26:02.237","AcceptedAnswerId":33675465.0,"OwnerUserId":4696856.0,"Title":"ImportError: libcudart.so.7.0: cannot open shared object file: No such file or directory","Body":"<pre><\/pre>\n\n<p>When I execute python command, a \"ImportError\" error occurs.Some solved it by adding \"export LD_LIBRARY_PATH=\/usr\/local\/cuda-5.5\/lib:\/usr\/local\/cuda-5.5\/lib64\" in the \/etc\/profile.I try it but do no effect.I find the \"libcudart.so.7.0\" in \/usr\/share\/man\/man7\/libcudart.so.7 by executing the whereis command and have no idea what should I do next to solve it.<\/p>\n","answers":[{"AnswerId":"33675465","CreationDate":"2015-11-12T15:44:46.813","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This error is being raised because the loader cannot find version 7.0 of the CUDA runtime on your system. <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#install_cuda\" rel=\"noreferrer\">TensorFlow requires CUDA 7.0.<\/a><\/p>\n\n<p>From the path in your question (<code>\/usr\/local\/cuda-5.5\/...<\/code>) it looks like you have CUDA 5.5 installed. The solution is to upgrade your CUDA runtime to version 7.0, which is available from NVIDIA <a href=\"https:\/\/developer.nvidia.com\/cuda-toolkit-70\" rel=\"noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":33671453,"AnswerCount":2,"Tags":"<nlp><theano><deep-learning><keras>","CreationDate":"2015-11-12T12:31:21.007","AcceptedAnswerId":null,"OwnerUserId":4622951.0,"Title":"\"g++ not detected\" while data set goes larger, is there any limit to matrix size in GPU?","Body":"<p>I got this message in using Keras to train an RNN for language model with a big 3D tensor (generated from a text, one hot encoded, and results a shape of (165717, 25, 7631)):<\/p>\n\n<pre><\/pre>\n\n<p>But everything goes well while I limit the size of data set into small. Thus I wonder that does Theano or CUDA limit the size of matrix? <\/p>\n\n<p>Besides, do I have a better way to do one hot representation? I mean, in the large 3D tensor, most elements are 0 due to the one-hot representation. However, I didn't found a layer which accepts index representation of words.<\/p>\n","answers":[{"AnswerId":"33706911","CreationDate":"2015-11-14T09:19:01.710","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Your Theano installation is not complete.<\/p>\n\n<p>There are two problems mentioned in the question's pasted result:<\/p>\n\n<blockquote>\n  <p>WARNING (theano.configdefaults): g++ not detected ! Theano will be\n  unable to  execute optimized C-implementations (for both CPU and GPU)\n  and will default to  Python implementations. Performance will be\n  severely degraded. To remove this  warning, set Theano flags cxx to an\n  empty string.<\/p>\n<\/blockquote>\n\n<p>I suspect you're seeing this one even with small data sizes but it's a warning so things continue running successfully (using the pure Python implementation automatically).<\/p>\n\n<blockquote>\n  <p>ERROR (theano.sandbox.cuda): nvcc compiler not found on $PATH. Check\n  your nvcc  installation and try again.<\/p>\n<\/blockquote>\n\n<p>This is the one that happens when the data size increases because now the GPU is trying to be used.<\/p>\n\n<p>Both messages indicate an incomplete Theano installation. The first indicates that you've not set up your C++ compiler properly. The second indicates that you've not set up CUDA properly. You need to follow the appropriate sections of the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html\" rel=\"nofollow\">installation documentation<\/a> to fix these problems. Note that simply doing <code>pip install Theano<\/code> is not sufficient when you want to use anything other than the pure Python implementations.<\/p>\n"},{"AnswerId":"39813530","CreationDate":"2016-10-02T03:18:03.570","ParentId":null,"OwnerUserId":"3598832","Title":null,"Body":"<pre><code>conda install mingw libpython\n<\/code><\/pre>\n\n<p>Make sure this is installed. Get this answer from another post, <a href=\"https:\/\/stackoverflow.com\/a\/31109547\/3598832\">https:\/\/stackoverflow.com\/a\/31109547\/3598832<\/a>, which indicated from the manual.<\/p>\n"}]}
{"QuestionId":33673495,"AnswerCount":4,"Tags":"<deep-learning><tensorflow>","CreationDate":"2015-11-12T14:14:40.643","AcceptedAnswerId":null,"OwnerUserId":5555047.0,"Title":"Are AlexNet weights available for Tensorflow?","Body":"<p>The benchmark for AlexNet is part of the Tensorflow repository when I clone it from git. The benchmark implements the layers, but it looks to me that the actual weights for AlexNet are not loaded at any point.<\/p>\n\n<p>I would like to play around with Tensorflow, but my applications (in caffe) are using a pre-trained AlexNet. <\/p>\n\n<p>Do you think they will release the weights as well?<\/p>\n","answers":[{"AnswerId":"33681059","CreationDate":"2015-11-12T20:49:28.103","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I don't know of an existing one, but someone wrote a <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"noreferrer\">converter to import Caffe models into tensorflow<\/a>, and you can find <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet\" rel=\"noreferrer\">pre-trained Alexnet models for Caffe<\/a> (also see <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo\" rel=\"noreferrer\">the BVLC Model-Zoo<\/a>).  I can't promise it will work, but you can quite possibly glue those two together to get what you want.<\/p>\n"},{"AnswerId":"41631833","CreationDate":"2017-01-13T09:58:28.050","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p><a href=\"http:\/\/www.cs.toronto.edu\/~guerzhoy\/tf_alexnet\/\" rel=\"nofollow noreferrer\">http:\/\/www.cs.toronto.edu\/~guerzhoy\/tf_alexnet\/<\/a> has weights as numpy arrays of trained weights on AlexNet. You can use those in TensorFlow.<\/p>\n"},{"AnswerId":"40080468","CreationDate":"2016-10-17T07:15:13.733","ParentId":null,"OwnerUserId":"7029635","Title":null,"Body":"<p>No,the benchmark file is just implementation and you have to train and evaluate<\/p>\n"},{"AnswerId":"48624547","CreationDate":"2018-02-05T14:16:01.047","ParentId":null,"OwnerUserId":"7039825","Title":null,"Body":"<p><strong>Yes, there is AlexNet pretrained weights available for Tensorflow, you can download it <a href=\"http:\/\/www.cs.toronto.edu\/%7Eguerzhoy\/tf_alexnet\/\" rel=\"nofollow noreferrer\">here<\/a><\/strong><\/p>\n\n<p>To load them to your project, you can use the following code (adapted from <a href=\"https:\/\/github.com\/kratzert\/finetune_alexnet_with_tensorflow\/blob\/master\/alexnet.py\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<pre><code># Load the weights into memory\nweights_dict = np.load('.\/weight-path\/bvlc_alexnet.npy', encoding='bytes').item()\n\n# Loop over all layer names stored in the weights dict\nfor op_name in weights_dict:\n\n    # Check if layer should be trained from scratch\n    if op_name not in self.SKIP_LAYER:\n\n        with tf.variable_scope(op_name, reuse=True):\n\n            # Assign weights\/biases to their corresponding tf variable\n            for data in weights_dict[op_name]:\n\n                # Biases\n                if len(data.shape) == 1:\n                    var = tf.get_variable('biases', trainable=False)\n                    session.run(var.assign(data))\n\n                # Weights\n                else:\n                    var = tf.get_variable('weights', trainable=False)\n                    session.run(var.assign(data))\n<\/code><\/pre>\n"}]}
{"QuestionId":33673583,"AnswerCount":3,"Tags":"<bash><installation><tensorflow>","CreationDate":"2015-11-12T14:18:59.410","AcceptedAnswerId":33674247.0,"OwnerUserId":4611375.0,"Title":"Installing TensorFlow in a Virtualenv on OSX","Body":"<p>I am following the <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#virtualenv-based_installation\" rel=\"nofollow\">instructions for installing TensorFlow in a virtual environment on Mac<\/a>, however I am not sure what this instruction means:<\/p>\n\n<pre><\/pre>\n\n<p>Specifically the  bit.<\/p>\n\n<p>Installation has progressed fine so far. I tried using that exact command but it printed an error:<\/p>\n\n<pre><\/pre>\n\n<p>I assume that everything between the &lt;> is a placeholder and refers to some location, not sure what though.<\/p>\n","answers":[{"AnswerId":"33674247","CreationDate":"2015-11-12T14:51:09.920","ParentId":null,"OwnerUserId":"4611375","Title":null,"Body":"<p>Solved.<\/p>\n\n<p>It's now installed and up and running.<\/p>\n\n<p><code>&lt;$url_to_binary.whl&gt;<\/code> should be replaced by <code>https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl<\/code><\/p>\n\n<p>I assume that was the file they meant.<\/p>\n"},{"AnswerId":"33680441","CreationDate":"2015-11-12T20:12:06.607","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>They meant the binary they told about previously in the <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#virtualenv-based_installation\" rel=\"nofollow\">binary installation section for MacOS<\/a><\/p>\n\n<blockquote>\n  <p>Only CPU-version is available at the moment. <\/p>\n  \n  <p><code>$ pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl<\/code><\/p>\n<\/blockquote>\n\n<p>So the path to binary is <code>https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl<\/code><\/p>\n"},{"AnswerId":"49995449","CreationDate":"2018-04-24T07:11:09.103","ParentId":null,"OwnerUserId":"9689594","Title":null,"Body":"<p>In my opinion,you may try like this:<\/p>\n\n<p><code>(tensorflow)$ pip install --upgrade<\/code> <\/p>\n\n<p><a href=\"https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.8.0-cp27-none-linux_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.8.0-cp27-none-linux_x86_64.whl<\/a><\/p>\n"}]}
{"QuestionId":33676656,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2015-11-12T16:39:02.337","AcceptedAnswerId":null,"OwnerUserId":3502360.0,"Title":"Caffe HDF5 not learning","Body":"<p>I'm fine-tuning the GoogleNet network with Caffe to my own dataset. If I use IMAGE_DATA layers as input learning takes place. However, I need to switch to an HDF5 layer for further extensions that I require. When I use HDF5 layers no learning takes place.<\/p>\n\n<p>I am using the exact same input images, and the labels match also. I have also checked to ensure that the data in .h5 files can be loaded correctly. It does, and Caffe is also able to find the number of examples I feed it as well as the correct number of classes (2).<\/p>\n\n<p>This leads me to think that the issue lies in the transformations I am performing manually (since HDF5 layers do not perform any built-in transformations). The code for these is below. I do the following:<\/p>\n\n<ul>\n<li>Convert image from RGB to BGR<\/li>\n<li>Resize it to 256x256 so I can subtract the mean file from ImageNet (<a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/python\/caffe\/imagenet\" rel=\"nofollow noreferrer\">included in the Caffe library<\/a>)<\/li>\n<li>Since the original GoogleNet prototxt does not divide by 255, I also do not (<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/train_val.prototxt#L10\" rel=\"nofollow noreferrer\">see here<\/a>)<\/li>\n<li>I resize the image down to 224x224, which is the crop size required by GoogleNet<\/li>\n<li>I transpose the image as needed to satisfy CxHxW, as required by Caffe<\/li>\n<li>At the moment I am not performing data augmentation, which could be turned on if I let oversample=True.<\/li>\n<\/ul>\n\n<p>Can anyone see anything wrong with this approach? Is data augmentation so critical that no learning would take place without it?<\/p>\n\n<h1>The HDF5 conversion code<\/h1>\n\n<pre><\/pre>\n\n<h1>Relevant differences between IMAGE_DATA and HDF5 prototxt files<\/h1>\n\n<pre><\/pre>\n\n<h1>Update<\/h1>\n\n<p>When I say no learning is taking place I mean that my training loss is not going down consistently when using HDF5 data compared to the IMG_Data. In the images below, the first plot is plot the change in the training loss for the IMG_DATA network, and the other is the HDF5 data network. <\/p>\n\n<p>One possibility that I am considering is that the network is overfitting to each of the .h5 that I am feeding it. At the moment I am using data augmentation, but all of the augmented examples are stored into a single .h5 file, along with other examples. However, because all of the augmented versions of a single input image are all contained within the same .h5 file, I think this could cause the network to overfit to that specific .h5 file. However, I am not sure whether this is what the second plot suggests.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8PRPz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8PRPz.png\" alt=\"Training loss with IMG_DATA input\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/0BYLs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0BYLs.png\" alt=\"Training loss with HDF5 input\"><\/a><\/p>\n","answers":[{"AnswerId":"36628066","CreationDate":"2016-04-14T15:50:37.400","ParentId":null,"OwnerUserId":"6205009","Title":null,"Body":"<p>I faced the same problem and found out that for some reason doing the transformation manually as you are doing in your code causes the images to be all black (all zeros). try to debug your code and see if that is happening.\nthe solution is to use the same methodology explained in the Caffe tutorial here\n<a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a>\nthe part where you see<\/p>\n\n<pre><code># create transformer for the input called 'data'\ntransformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})\ntransformer.set_transpose('data', (2,0,1))  # move image channels to outermost dimension\ntransformer.set_mean('data', mu)            # subtract the dataset-mean value in each channel\ntransformer.set_raw_scale('data', 255)      # rescale from [0, 1] to [0, 255]\ntransformer.set_channel_swap('data', (2,1,0))  # swap channels from RGB to BGR\n<\/code><\/pre>\n\n<p>then few lines down<\/p>\n\n<pre><code>image = caffe.io.load_image(caffe_root + 'examples\/images\/cat.jpg')\ntransformed_image = transformer.preprocess('data', image)\n<\/code><\/pre>\n"}]}
{"QuestionId":33678180,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-12T17:59:20.247","AcceptedAnswerId":33685318.0,"OwnerUserId":1319312.0,"Title":"What is the format for device filters in TensorFlow?","Body":"<p>So the Session config proto has a device_filters option, with the comment:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have concrete explanation of the format? For example, I want to exclude \/gpu:0 as an option because I use it for running other models.<\/p>\n\n<p>I've tried<\/p>\n\n<pre><\/pre>\n\n<p>But I'm still getting ops allocated to gpu 0. I don't want to override the devices on a per-op basis.<\/p>\n","answers":[{"AnswerId":"33685318","CreationDate":"2015-11-13T03:10:39.990","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>ConfigProto.device_filters<\/code> field is currently ignored by TensorFlow, although it is intended to support your use case in future. If you want to achieve the same end of running ops on <code>\/gpu:1<\/code> and <code>\/cpu:0<\/code>, you can do that as follows, using \"soft placement\":<\/p>\n\n<pre><code>with tf.device(\"\/gpu:1\"):\n  # Build your model in this with context. All nodes will get the\n  # device \"\/gpu:1\".\n\nwith tf.Session(config=tf.ConfigProto(allow_soft_placement=True)):\n  # Execute your mode in this with context.\n  # Soft placement will use \/gpu:1 for GPU-compatible ops, and \/cpu:0\n  # for CPU-only ops.\n<\/code><\/pre>\n"}]}
{"QuestionId":33679382,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-12T19:11:52.290","AcceptedAnswerId":33680246.0,"OwnerUserId":3731804.0,"Title":"How do I get the current value of a Variable?","Body":"<p>Suppose we have a variable:<\/p>\n\n<pre><\/pre>\n\n<p>This variable can be updated during the training process using the  method.<\/p>\n\n<p><em>What is the best way to get the current value of a variable?<\/em><\/p>\n\n<p>I know we could use this:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm afraid this would trigger a whole chain of operations.<\/p>\n\n<p>In Theano, you could just do<\/p>\n\n<pre><\/pre>\n\n<p>I'm looking for the equivalent thing in TensorFlow.<\/p>\n","answers":[{"AnswerId":"40304259","CreationDate":"2016-10-28T11:36:32.803","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p><code>tf.Print<\/code> can simplify your life!<\/p>\n\n<p><code>tf.Print<\/code> will print the value of the tensor(s) you tell it to print at the moment where the <code>tf.Print<\/code> line is called in your code when your code is evaluated.<\/p>\n\n<p>So for example:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.Variable([1.0, 2.0])\nx = tf.Print(x,[x])\nx = 2* x\n\ntf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run()\n<\/code><\/pre>\n\n<blockquote>\n  <p>[1.0 2.0 ]<\/p>\n<\/blockquote>\n\n<p>because it prints the value of <code>x<\/code> at the moment when the <code>tf.Print<\/code> line is. If instead you do<\/p>\n\n<pre><code>v = x.eval()\nprint(v)\n<\/code><\/pre>\n\n<p>you will get:<\/p>\n\n<blockquote>\n  <p>[2.0 4.0 ]<\/p>\n<\/blockquote>\n\n<p>because it will give you the final value of x. <\/p>\n"},{"AnswerId":"56948436","CreationDate":"2019-07-09T08:32:08.843","ParentId":null,"OwnerUserId":"11758515","Title":null,"Body":"<p>As they cancelled <code>tf.Variable()<\/code> in <code>tensorflow 2.0.0<\/code>,<\/p>\n\n<p>If you want to extract values from a <code>tensor(ie \"net\")<\/code>, you can use this,<\/p>\n\n<pre><code>net.[tf.newaxis,:,:].numpy().\n<\/code><\/pre>\n"},{"AnswerId":"33680246","CreationDate":"2015-11-12T19:59:01.920","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>In general, <code>session.run(x)<\/code> will evaluate only the nodes that are necessary to compute <code>x<\/code> and nothing else, so it should be relatively cheap if you want to inspect the value of the variable.<\/p>\n\n<p>Take a look at this great answer <a href=\"https:\/\/stackoverflow.com\/a\/33610914\/5543198\">https:\/\/stackoverflow.com\/a\/33610914\/5543198<\/a> for more context.<\/p>\n"},{"AnswerId":"33680834","CreationDate":"2015-11-12T20:35:15.960","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The only way to get the value of the variable is by running it in a <code>session<\/code>. In the <a href=\"http:\/\/tensorflow.org\/resources\/faq.md#contents\" rel=\"noreferrer\">FAQ it is written<\/a> that:<\/p>\n\n<blockquote>\n  <p>A Tensor object is a symbolic handle to the result of an operation,\n  but does not actually hold the values of the operation's output.<\/p>\n<\/blockquote>\n\n<p>So TF equivalent would be:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.Variable([1.0, 2.0])\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    v = sess.run(x)\n    print(v)  # will show you your variable.\n<\/code><\/pre>\n\n<p>The part with <code>init = global_variables_initializer()<\/code> is important and should be done in order to initialize variables.<\/p>\n\n<p>Also, take a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/InteractiveSession\" rel=\"noreferrer\">InteractiveSession<\/a> if you work in IPython.<\/p>\n"}]}
{"QuestionId":33680397,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-11-12T20:08:25.327","AcceptedAnswerId":33682724.0,"OwnerUserId":861809.0,"Title":"Tensorboard doesn't display the graph (HTML error)","Body":"<p>I created Tensorboard logs with a graph, using the  script (I also have the same issue on other scripts).<\/p>\n\n<pre><\/pre>\n\n<p>The Tensorboard works well for Scalars, Images and Histograms, but not for graphs. I get a white page, with only the header.<\/p>\n\n<p>If I open the browser console (<em>Firefox 42, Ubuntu 14.04<\/em>), I get those errors when I switch to the GRAPH tab:<\/p>\n\n<pre><\/pre>\n\n<p>How to fix this ?<\/p>\n","answers":[{"AnswerId":"33682724","CreationDate":"2015-11-12T22:34:37.973","ParentId":null,"OwnerUserId":"4275865","Title":null,"Body":"<p>Tensorboard is only working in Google Chrome as far as I can tell. I tried Safari, Firefox and Chrome on my Mac and it only showed the graph properly in Chrome.<\/p>\n"}]}
{"QuestionId":33681517,"AnswerCount":14,"Tags":"<python><machine-learning><neural-network><tensorflow>","CreationDate":"2015-11-12T21:16:01.233","AcceptedAnswerId":33682213.0,"OwnerUserId":276310.0,"Title":"Tensorflow One Hot Encoder?","Body":"<p>Does tensorflow have something similar to scikit learn's <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.OneHotEncoder.html\">one hot encoder<\/a> for processing categorical data?  Would using a placeholder of tf.string behave as categorical data?<\/p>\n\n<p>I realize I can manually pre-process the data before sending it to tensorflow, but having it built in is very convenient.<\/p>\n","answers":[{"AnswerId":"40942677","CreationDate":"2016-12-02T23:55:55.103","ParentId":null,"OwnerUserId":"1055001","Title":null,"Body":"<p>There are  a couple ways to do it.<\/p>\n\n<pre><code>ans = tf.constant([[5, 6, 0, 0], [5, 6, 7, 0]]) #batch_size*max_seq_len\nlabels = tf.reduce_sum(tf.nn.embedding_lookup(np.identity(10), ans), 1)\n\n&gt;&gt;&gt; [[ 0.  0.  0.  0.  0.  1.  1.  0.  0.  0.]\n&gt;&gt;&gt; [ 0.  0.  0.  0.  0.  1.  1.  1.  0.  0.]]\n<\/code><\/pre>\n\n<p>The other way to do it is.<\/p>\n\n<pre><code>labels2 = tf.reduce_sum(tf.one_hot(ans, depth=10, on_value=1, off_value=0, axis=1), 2)\n\n &gt;&gt;&gt; [[0 0 0 0 0 1 1 0 0 0]\n &gt;&gt;&gt; [0 0 0 0 0 1 1 1 0 0]]\n<\/code><\/pre>\n"},{"AnswerId":"38982556","CreationDate":"2016-08-16T19:00:34.220","ParentId":null,"OwnerUserId":"2452184","Title":null,"Body":"<p>A simple and short way to one-hot encode any integer or list of intergers:<\/p>\n\n<pre><code>a = 5 \nb = [1, 2, 3]\n# one hot an integer\none_hot_a = tf.nn.embedding_lookup(np.identity(10), a)\n# one hot a list of integers\none_hot_b = tf.nn.embedding_lookup(np.identity(max(b)+1), b)\n<\/code><\/pre>\n"},{"AnswerId":"41776894","CreationDate":"2017-01-21T07:38:24.717","ParentId":null,"OwnerUserId":"1616677","Title":null,"Body":"<p>My version of @CFB and @dga example, shortened a bit to ease understanding.<\/p>\n\n<pre><code>num_labels = 10\nlabels_batch = [2, 3, 5, 9]\n\nsparse_labels = tf.reshape(labels_batch, [-1, 1])\nderived_size = len(labels_batch)\nindices = tf.reshape(tf.range(0, derived_size, 1), [-1, 1])\nconcated = tf.concat(1, [indices, sparse_labels]) \nlabels = tf.sparse_to_dense(concated, [derived_size, num_labels], 1.0, 0.0)\n<\/code><\/pre>\n"},{"AnswerId":"33810489","CreationDate":"2015-11-19T17:26:19.417","ParentId":null,"OwnerUserId":"3728184","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn.html#embedding_lookup\" rel=\"nofollow\">tf.nn.embedding_lookup<\/a>. It maps from categorical IDs to their embeddings. <\/p>\n\n<p>For an example of how it's used for input data, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L95\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"33681732","CreationDate":"2015-11-12T21:28:56.683","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/one_hot\" rel=\"noreferrer\"><code>tf.one_hot()<\/code><\/a> is available in TF and easy to use. <\/p>\n\n<p>Lets assume you have 4 possible categories (cat, dog, bird, human) and 2 instances (cat, human). So your <code>depth=4<\/code> and your <code>indices=[0, 3]<\/code><\/p>\n\n<pre><code>import tensorflow as tf\nres = tf.one_hot(indices=[0, 3], depth=4)\nwith tf.Session() as sess:\n    print sess.run(res)\n<\/code><\/pre>\n\n<p>Keep in mind that if you provide index=-1 you will get all zeros in your one-hot vector.<\/p>\n\n<p><strong>Old answer, when this function was not available.<\/strong><\/p>\n\n<p>After looking though the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/index.md\" rel=\"noreferrer\">python documentation<\/a>, I have not found anything similar. One thing that strengthen my belief that it does not exist is that in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1d76583411038767f673a0c96174c80eaf9ff42f\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py\" rel=\"noreferrer\">their own example<\/a> they write <code>one_hot<\/code> manually.<\/p>\n\n<pre><code>def dense_to_one_hot(labels_dense, num_classes=10):\n  \"\"\"Convert class labels from scalars to one-hot vectors.\"\"\"\n  num_labels = labels_dense.shape[0]\n  index_offset = numpy.arange(num_labels) * num_classes\n  labels_one_hot = numpy.zeros((num_labels, num_classes))\n  labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1\n  return labels_one_hot\n<\/code><\/pre>\n\n<p>You can also do this in <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.OneHotEncoder.html\" rel=\"noreferrer\">scikitlearn<\/a>.<\/p>\n"},{"AnswerId":"35902365","CreationDate":"2016-03-09T20:53:13.020","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>Recent versions of TensorFlow (nightlies and maybe even 0.7.1) have an op called tf.one_hot that does what you want.  Check it out!<\/p>\n\n<p>On the other hand if you have a dense matrix and you want to look up and aggregate values in it, you would want to use the embedding_lookup function.<\/p>\n"},{"AnswerId":"33682213","CreationDate":"2015-11-12T21:57:38.037","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>As of TensorFlow 0.8, there is now a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/one_hot\" rel=\"noreferrer\">native one-hot op, <code>tf.one_hot<\/code><\/a> that can convert a set of sparse labels to a dense one-hot representation.  This is in addition to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code><\/a>, which can in some cases let you compute the cross entropy directly on the sparse labels instead of converting them to one-hot.<\/p>\n\n<p><strong>Previous answer, in case you want to do it the old way:<\/strong>\n@Salvador's answer is correct - there (used to be) no native op to do it.  Instead of doing it in numpy, though, you can do it natively in tensorflow using the sparse-to-dense operators:<\/p>\n\n<pre><code>num_labels = 10\n\n# label_batch is a tensor of numeric labels to process\n# 0 &lt;= label &lt; num_labels\n\nsparse_labels = tf.reshape(label_batch, [-1, 1])\nderived_size = tf.shape(label_batch)[0]\nindices = tf.reshape(tf.range(0, derived_size, 1), [-1, 1])\nconcated = tf.concat(1, [indices, sparse_labels])\noutshape = tf.pack([derived_size, num_labels])\nlabels = tf.sparse_to_dense(concated, outshape, 1.0, 0.0)\n<\/code><\/pre>\n\n<p>The output, labels, is a one-hot matrix of batch_size x num_labels.<\/p>\n\n<p>Note also that as of 2016-02-12 (which I assume will eventually be part of a 0.7 release), TensorFlow also has the <code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code> op, which in some cases can let you do training without needing to convert to a one-hot encoding.<\/p>\n\n<p>Edited to add:  At the end, you may need to explicitly set the shape of labels.  The shape inference doesn't recognize the size of the num_labels component.  If you don't need a dynamic batch size with derived_size, this can be simplified.<\/p>\n\n<p>Edited 2016-02-12 to change the assignment of outshape per comment below.<\/p>\n"},{"AnswerId":"44911049","CreationDate":"2017-07-04T16:58:45.667","ParentId":null,"OwnerUserId":"7403220","Title":null,"Body":"<p>As mentioned above by @dga, Tensorflow has <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/one_hot\" rel=\"nofollow noreferrer\">tf.one_hot<\/a> now:<\/p>\n\n<pre><code>labels = tf.constant([5,3,2,4,1])\nhighest_label = tf.reduce_max(labels)\nlabels_one_hot = tf.one_hot(labels, highest_label + 1)\n\narray([[ 0.,  0.,  0.,  0.,  0.,  1.],\n       [ 0.,  0.,  0.,  1.,  0.,  0.],\n       [ 0.,  0.,  1.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  1.,  0.],\n       [ 0.,  1.,  0.,  0.,  0.,  0.]], dtype=float32)\n<\/code><\/pre>\n\n<p>You need to specify depth, otherwise you'll get a pruned one-hot tensor.<\/p>\n\n<p>If you like to do it manually:<\/p>\n\n<pre><code>labels = tf.constant([5,3,2,4,1])\nsize = tf.shape(labels)[0]\nhighest_label = tf.reduce_max(labels)\nlabels_t = tf.reshape(labels, [-1, 1])\nindices = tf.reshape(tf.range(size), [-1, 1])\nidx_with_labels = tf.concat([indices, labels_t], 1)\nlabels_one_hot = tf.sparse_to_dense(idx_with_labels, [size, highest_label + 1], 1.0)\n\narray([[ 0.,  0.,  0.,  0.,  0.,  1.],\n       [ 0.,  0.,  0.,  1.,  0.,  0.],\n       [ 0.,  0.,  1.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  1.,  0.],\n       [ 0.,  1.,  0.,  0.,  0.,  0.]], dtype=float32)\n<\/code><\/pre>\n\n<p>Note arguments order in tf.concat()<\/p>\n"},{"AnswerId":"37640473","CreationDate":"2016-06-05T09:41:53.467","ParentId":null,"OwnerUserId":"366553","Title":null,"Body":"<p>Current versions of tensorflow implement the following function for creating one-hot tensors:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#one_hot\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#one_hot<\/a><\/p>\n"},{"AnswerId":"40067488","CreationDate":"2016-10-16T06:22:42.743","ParentId":null,"OwnerUserId":"4955995","Title":null,"Body":"<p><code>numpy<\/code> does it!<\/p>\n\n<pre><code>import numpy as np\nnp.eye(n_labels)[target_vector]\n<\/code><\/pre>\n"},{"AnswerId":"35447639","CreationDate":"2016-02-17T03:57:23.163","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>There's <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/ops\/embeddings_ops.py\" rel=\"nofollow\">embedding_ops<\/a> in Scikit Flow and examples that deal with categorical variables, etc. <\/p>\n\n<p>If you just begin to learn TensorFlow, I would suggest you trying out <a href=\"https:\/\/github.com\/tensorflow\/skflow\/tree\/master\/examples\" rel=\"nofollow\">examples<\/a> in <a href=\"https:\/\/github.com\/tensorflow\/skflow\" rel=\"nofollow\">TensorFlow\/skflow<\/a> first and then once you are more familiar with TensorFlow it would be fairly easy for you to insert TensorFlow code to build a custom model you want (there are also examples for this). <\/p>\n\n<p>Hope those examples for images and text understanding could get you started and let us know if you encounter any issues! (post issues or tag skflow in SO). <\/p>\n"},{"AnswerId":"46441298","CreationDate":"2017-09-27T06:53:20.747","ParentId":null,"OwnerUserId":"3907250","Title":null,"Body":"<pre><code>In [7]: one_hot = tf.nn.embedding_lookup(np.eye(5), [1,2])\n\nIn [8]: one_hot.eval()\nOut[8]: \narray([[ 0.,  1.,  0.,  0.,  0.],\n       [ 0.,  0.,  1.,  0.,  0.]])\n<\/code><\/pre>\n\n<p>works on TF version 1.3.0. As of Sep 2017. <\/p>\n"},{"AnswerId":"35236597","CreationDate":"2016-02-06T02:18:56.697","ParentId":null,"OwnerUserId":"5890756","Title":null,"Body":"<p>Maybe it's due to changes to Tensorflow since Nov 2015, but @dga's answer produced errors. I did get it to work with the following modifications:<\/p>\n\n<pre><code>sparse_labels = tf.reshape(label_batch, [-1, 1])\nderived_size = tf.shape(sparse_labels)[0]\nindices = tf.reshape(tf.range(0, derived_size, 1), [-1, 1])\nconcated = tf.concat(1, [indices, sparse_labels])\noutshape = tf.concat(0, [tf.reshape(derived_size, [1]), tf.reshape(num_labels, [1])])\nlabels = tf.sparse_to_dense(concated, outshape, 1.0, 0.0)\n<\/code><\/pre>\n"},{"AnswerId":"35236673","CreationDate":"2016-02-06T02:30:49.510","ParentId":null,"OwnerUserId":"5530615","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/sparse_ops.html#sparse_to_dense\" rel=\"nofollow\">tf.sparse_to_dense<\/a>:<\/p>\n\n<p>The sparse_indices argument indicates where the ones should go, output_shape should be set to the number of possible outputs (e.g. the number of labels), and sparse_values should be 1 with the desired type (it will determine the type of the output from the type of sparse_values).<\/p>\n"}]}
{"QuestionId":33684441,"AnswerCount":4,"Tags":"<numpy><tensorflow>","CreationDate":"2015-11-13T01:22:21.970","AcceptedAnswerId":33749112.0,"OwnerUserId":697065.0,"Title":"writing a custom cost function in tensorflow","Body":"<p>I'm trying to write my own cost function in tensor flow, however apparently I cannot 'slice' the tensor object?<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:\n<\/p>\n","answers":[{"AnswerId":"33684528","CreationDate":"2015-11-13T01:32:03.897","ParentId":null,"OwnerUserId":"697065","Title":null,"Body":"<p>just realized that the following failed:<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*np.log(y))\n<\/code><\/pre>\n\n<p>you cant use numpy functions on tf objects, and the indexing my be different too.<\/p>\n"},{"AnswerId":"33684551","CreationDate":"2015-11-13T01:35:02.723","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This is because you have not initialized your variable and because of this it does not have your Tensor there right now (can read more <a href=\"https:\/\/stackoverflow.com\/a\/33640758\/1090562\">in my answer here<\/a>)<\/p>\n\n<p>Just do something like this:<\/p>\n\n<pre><code>def normVec(y):\n    print y\n    return np.cross(y[:,[0,2,4]],y[:,[1,3,5]])\n\nt1 = normVec(y_)\n# and comment everything after it.\n<\/code><\/pre>\n\n<p>To see that you do not have a Tensor now and only <code>Tensor(\"Placeholder_1:0\", shape=TensorShape([Dimension(None), Dimension(6)]), dtype=float32)<\/code>.<\/p>\n\n<p>Try initializing your variables<\/p>\n\n<pre><code>init = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n<\/code><\/pre>\n\n<p>and evaluate your variable <code>sess.run(y)<\/code>. P.S. you have not fed your placeholders up till now.<\/p>\n"},{"AnswerId":"33749112","CreationDate":"2015-11-17T04:22:18.790","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>At present, tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"noreferrer\">can't gather on axes other than the first - it's requested<\/a>.<\/p>\n\n<p>But for what you want to do in this specific situation, you can transpose, then gather 0,2,4, and then transpose back.  It won't be crazy fast, but it works:<\/p>\n\n<pre><code>tf.transpose(tf.gather(tf.transpose(y), [0,2,4]))\n<\/code><\/pre>\n\n<p>This is a useful workaround for some of the limitations in the current implementation of gather.<\/p>\n\n<p>(But it is also correct that you can't use a numpy slice on a tensorflow node - you can run it and slice the output, and also that you need to initialize those variables before you run. :).  You're mixing tf and np in a way that doesn't work.<\/p>\n\n<pre><code>x = tf.Something(...)\n<\/code><\/pre>\n\n<p>is a tensorflow graph object.  Numpy has no idea how to cope with such objects.<\/p>\n\n<pre><code>foo = tf.run(x)\n<\/code><\/pre>\n\n<p>is back to an object python can handle.<\/p>\n\n<p>You typically want to keep your loss calculation in pure tensorflow, so do the cross and other functions in tf.  You'll probably have to do the arccos the long way, as tf doesn't have a function for it.<\/p>\n"},{"AnswerId":"39404333","CreationDate":"2016-09-09T05:21:02.367","ParentId":null,"OwnerUserId":"6472650","Title":null,"Body":"<p>I think you can use \"Wraps Python function\" method in tensorflow. Here's the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/script_ops.html\" rel=\"nofollow\">link<\/a> to the documentation. <\/p>\n\n<p>And as for the people who answered \"Why don't you just use tensorflow's built in function to construct it?\" - sometimes the cost function people are looking for cannot be expressed in tf's functions or extremely difficult.<\/p>\n"}]}
{"QuestionId":33684648,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><conv-neural-network><caffe>","CreationDate":"2015-11-13T01:46:25.970","AcceptedAnswerId":33717093.0,"OwnerUserId":5389199.0,"Title":"What is batch size in Caffe or convnets","Body":"<p>I thought that batch size is only for performance. The bigger the batch, more images are computed at the same time to train my net. But I realized, if I change my batch size, my net accuracy gets better. So I did not understand what batch size is. Can someone explain me what is batch size?  <\/p>\n","answers":[{"AnswerId":"33717093","CreationDate":"2015-11-15T06:32:19.720","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe is trained using <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent\" rel=\"nofollow noreferrer\">Stochastic-Gradient-Descend (SGD)<\/a>: that is, at each iteration it computes the (stochastic) gradient of the parameters w.r.t the training data and makes a move (=change the parameters) in the direction of the gradient.<br>\nNow, if you write the equations of the gradient w.r.t. training data you'll notice that in order to compute the gradient <em>exactly<\/em> you need to evaluate <em>all<\/em> your training data <em>at each iteration<\/em>: this is prohibitively time consuming, especially when the training data gets bigger and bigger.<br>\nIn order to overcome this, SGD approximates the exact gradient, in a stochastic manner, by <em>sampling<\/em> only a small portion of the training data at each iteration. This small portion is the batch.<br>\nThus, the larger the batch size the more accurate the gradient estimate at each iteration.<\/p>\n\n<p><strong>TL;DR<\/strong>: batch size affect the accuracy of the estimated gradient at each iteration, changing the batch size therefore affect the \"path\" the optimization takes and may change the results of the training process.<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong><br>\nIn ICLR 2018 conference an interesting work was presented:<br>\n<em>Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le<\/em> <a href=\"https:\/\/openreview.net\/forum?id=B1Yy1BxCZ\" rel=\"nofollow noreferrer\"><strong>Don't Decay the Learning Rate, Increase the Batch Size<\/strong><\/a>.<br>\nThis work basically relates the effect of changing batch size and learning rate.<\/p>\n"}]}
{"QuestionId":33684657,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-11-13T01:47:05.187","AcceptedAnswerId":33685256.0,"OwnerUserId":5557105.0,"Title":"Issue feeding a list into feed_dict in TensorFlow","Body":"<p>I'm trying to pass a list into , however I'm having trouble doing so. Say I have:<\/p>\n\n<pre><\/pre>\n\n<p>where inputs is fed into some function  that I want to compute. So to run this in tensorflow, I created a session and ran the following:<\/p>\n\n<pre><\/pre>\n\n<p>but I get an error, \nHowever, I'm able to pass the data element-wise like so:<\/p>\n\n<pre><\/pre>\n\n<p>So I'm wondering if there's a way I can solve this issue. I've also tried to construct a dictionary(using a  loop), however this results in a dictionary with a single element, where they key is: \n<\/p>\n","answers":[{"AnswerId":"33685256","CreationDate":"2015-11-13T03:04:05.103","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are two issues that are causing problems here:<\/p>\n\n<p>The first issue is that the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/client.md#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a> call only accepts a small number of types as the keys of the <code>feed_dict<\/code>. In particular, lists of tensors are <strong><em>not<\/em><\/strong> supported as keys, so you have to put each tensor as a separate key.<sup>*<\/sup> One convenient way to do this is using a dictionary comprehension:<\/p>\n\n<pre><code>inputs = [tf.placeholder(...), ...]\ndata = [np.array(...), ...]\nsess.run(y, feed_dict={i: d for i, d in zip(inputs, data)})\n<\/code><\/pre>\n\n<p>The second issue is that the <code>10 * [tf.placeholder(...)]<\/code> syntax in Python creates a list with ten elements, where each element is <strong><em>the same tensor object<\/em><\/strong> (i.e. has the same <code>name<\/code> property, the same <code>id<\/code> property, and is reference-identical if you compare two elements from the list using <code>inputs[i] is inputs[j]<\/code>). This explains why, when you tried to create a dictionary using the list elements as keys, you ended up with a dictionary with a single element - because all of the list elements were identical.<\/p>\n\n<p>To create 10 different placeholder tensors, as you intended, you should instead do the following:<\/p>\n\n<pre><code>inputs = [tf.placeholder(tf.float32, shape=(batch_size, input_size))\n          for _ in xrange(10)]\n<\/code><\/pre>\n\n<p>If you print the elements of this list, you'll see that each element is a tensor with a different name.<\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong> <sup>*<\/sup> You can now pass <strong>tuples<\/strong> as the keys of a <code>feed_dict<\/code>, because these may be used as dictionary keys.<\/p>\n"},{"AnswerId":"54104492","CreationDate":"2019-01-09T06:42:14.537","ParentId":null,"OwnerUserId":"5246875","Title":null,"Body":"<p>feed_dict can be provided by preparing a dictionary beforehand as follows<\/p>\n\n<pre><code>n = 10\ninput_1 = [tf.placeholder(...) for _ in range(n)]\ninput_2 = tf.placeholder(...)\ndata_1 = [np.array(...) for _ in range(n)]\ndata_2 = np.array(...)\n\n\nfeed_dictionary = {}\nfor i in range(n):\n    feed_dictionary[input_1[i]] = data_1[i]\nfeed_dictionary[input_2] = data_2\nsess.run(y, feed_dict=feed_dictionary)\n<\/code><\/pre>\n"},{"AnswerId":"33684847","CreationDate":"2015-11-13T02:11:21.420","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Here is a correct example:<\/p>\n\n<pre><code>batch_size, input_size, n = 2, 3, 2\n# in your case n = 10\nx = tf.placeholder(tf.types.float32, shape=(n, batch_size, input_size))\ny = tf.add(x, x)\n\ndata = np.random.rand(n, batch_size, input_size)\n\nsess = tf.Session()\nprint sess.run(y, feed_dict={x: data})\n<\/code><\/pre>\n\n<p>And here is a strange things I see in your approach. For some reason you use <code>10 * [tf.placeholder(...)]<\/code>, which creates 10 tensors of size <code>(batch_size, input_size)<\/code>. No idea why do you do this, if you can just create on Tensor of rank 3 (where the first dimension is 10).<\/p>\n\n<p>Because you have a list of tensors (and not a tensor), you can not feed your data to this list (but in my case I can feed to my tensor). <\/p>\n"}]}
{"QuestionId":33685864,"AnswerCount":2,"Tags":"<c++><math><machine-learning><neural-network><theano>","CreationDate":"2015-11-13T04:17:44.137","AcceptedAnswerId":33706836.0,"OwnerUserId":5557434.0,"Title":"How do I get the gradient expression when implementing a specific neural network?","Body":"<p>I mean if I want to implement a normal neural network, it's easy to find the gradient or derivative expression anywhere, but when I implement a specific neural network, its's not that easy. So, normally, when you do that you get gradient expression by hand or you use some tool? Thanks a lot!<\/p>\n","answers":[{"AnswerId":"33706836","CreationDate":"2015-11-14T09:07:28.967","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano can tell you the symbolic expression for the gradient. The only problem is that it can take a little effort to interpret the output.<\/p>\n\n<p>For example, let <code>x<\/code>, <code>w<\/code>, and <code>b<\/code> be a scalars, and<\/p>\n\n<pre><code>y = x * w + b\n<\/code><\/pre>\n\n<p>In Theano this can be implemented as<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\nx = tt.scalar('x')\nw = tt.scalar('w')\nb = tt.scalar('b')\ny = x * w + b\ng = tt.grad(y, w)\ntheano.printing.debugprint(g)\nf = theano.function([x, w, b], outputs=g)\ntheano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>When executed (using the current bleeding edge version of Theano) this script will print<\/p>\n\n<pre><code>Elemwise{mul} [id A] ''   \n |Elemwise{second,no_inplace} [id B] ''   \n | |Elemwise{add,no_inplace} [id C] ''   \n | | |Elemwise{mul,no_inplace} [id D] ''   \n | | | |x [id E]\n | | | |w [id F]\n | | |b [id G]\n | |TensorConstant{1.0} [id H]\n |x [id E]\n\nDeepCopyOp [id A] 'x'   0\n |x [id B]\n<\/code><\/pre>\n\n<p>The first output is the unoptimized gradient expression. The second is the optimized gradient expression.<\/p>\n\n<p>As we know from our differential calculus class, for the given expression of <code>y<\/code>,<\/p>\n\n<pre><code>dy\/dw = x\n<\/code><\/pre>\n\n<p>And this is exactly what we see from Theano.<\/p>\n\n<p>So you can build your neural network in Theano and then print the automatically derived symbolic gradient to see what it should be, then implement that expression in your alternate implementation.<\/p>\n\n<p>For more complex networks interpreting the symbolic gradient from Theano can be very tedious and it's probably easier to learn the necessary maths to derive the gradients yourself, especially if using reasonably standard neural network constructs. The Theano automatic differentiation could then be used to verify your manual derivation.<\/p>\n"},{"AnswerId":"33686758","CreationDate":"2015-11-13T05:56:05.643","ParentId":null,"OwnerUserId":"4253485","Title":null,"Body":"<p>Using and neural network architecture, the formulas for the gradient decent learning should be computed. If you're using a neural network package like caffe or torch, you'll just have to specify if you'll be working on the negative log likelihood or the mean square error or what (this is what changes the learning rule). The formulas would be done automatically. However, if you're writing your own neural network platform, you'll need to do some math or just get the formulas from a paper or something. There is no tool that works as a formula finder for the neural networks learning.<\/p>\n"}]}
{"QuestionId":33686464,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-13T05:27:40.277","AcceptedAnswerId":33686790.0,"OwnerUserId":4418769.0,"Title":"What am I missing from this csv reader for TensorFlow?","Body":"<p>It is mostly a copy paste from the tutorial, on the website. I am getting an error:<\/p>\n\n<blockquote>\n  <p>Invalid argument: ConcatOp : Expected concatenating dimensions in the\n  range [0, 0), but got 0    [[Node: concat = Concat[N=4, T=DT_INT32,\n  _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](concat\/concat_dim, DecodeCSV, DecodeCSV:1, DecodeCSV:2, DecodeCSV:3)]]<\/p>\n<\/blockquote>\n\n<p>the contents of my csv file is:<\/p>\n\n<blockquote>\n  <p>3,4,1,8,4<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33686652","CreationDate":"2015-11-13T05:44:46.413","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I am also stuck with this <a href=\"http:\/\/tensorflow.org\/how_tos\/reading_data\/index.md\" rel=\"nofollow\">tutorial<\/a>. I was able to exchange one problem for another, when I changed your <code>with tf.Session()<\/code> for:<\/p>\n\n<pre><code>sess = tf.Session()\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(coord=coord)\n\nfor i in range(2):\n    #print i\n    example, label = sess.run([features, col5])\n\ncoord.request_stop()\ncoord.join(threads)\n\nsess.close()\n<\/code><\/pre>\n\n<p>The error disappeared, TF started to run, but it looks like it is stuck. If you uncomment <code># print<\/code> you will see that only one iteration runs. Most probably this is not really helpful (because I trade an error for infinite execution).<\/p>\n"},{"AnswerId":"33686790","CreationDate":"2015-11-13T05:59:09.947","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue arises due to the shape of the tensors in your program. <strong><em>TL;DR<\/em><\/strong> Instead of <code>tf.concat()<\/code> you should use <code>tf.pack()<\/code>, which will transform the four scalar <code>col<\/code> tensors into a 1-D tensor of length 4. <\/p>\n\n<p>Before we start, note that you can use the <code>get_shape()<\/code> method on any <code>Tensor<\/code> object to get static shape information about that tensor. For example, the commented-out line in your code could be:<\/p>\n\n<pre><code>print col1.get_shape()\n# ==&gt; 'TensorShape([])' - i.e. `col1` is a scalar.\n<\/code><\/pre>\n\n<p>The <code>value<\/code> tensor returned by <code>reader.read()<\/code> is a scalar string. <code>tf.decode_csv(value, record_defaults=[...])<\/code> produces, for each element of <code>record_defaults<\/code>, a tensor of the same shape as <code>value<\/code>, i.e. a scalar in this case. A scalar is a 0-dimensional tensor with a single element. <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#concat\" rel=\"noreferrer\"><code>tf.concat(i, xs)<\/code><\/a> is not defined on scalars: it concatenates a list of N-dimensional tensors (<code>xs<\/code>) into a new N-dimensional tensor, along dimension <code>i<\/code>, where <code>0 &lt;= i &lt; N<\/code>, and there is no valid <code>i<\/code> if <code>N = 0<\/code>.<\/p>\n\n<p>The <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#pack\" rel=\"noreferrer\"><code>tf.pack(xs)<\/code><\/a> operator is designed to solve this problem simply. It takes a list of <code>k<\/code> N-dimensional tensors (with the same shape) and packs them into an N+1-dimensional tensor with size <code>k<\/code> in the 0th dimension. If you replace the <code>tf.concat()<\/code> with <code>tf.pack()<\/code>, your program will work:<\/p>\n\n<pre><code># features = tf.concat(0, [col1, col2, col3, col4])\nfeatures = tf.pack([col1, col2, col3, col4])\n\nwith tf.Session() as sess:\n  # Start populating the filename queue.\n  # ...\n<\/code><\/pre>\n"}]}
{"QuestionId":33686555,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-13T05:35:48.953","AcceptedAnswerId":33686627.0,"OwnerUserId":4593600.0,"Title":"How do we deploy a trained tensorflow model on a mobile device?","Body":"<p>One of the highlights of tensorflow appears to be \"true portability\" - seamless deployment of trained models across different platforms - especially running a trained model on a mobile device? Do you have an example or some tutorial that walks through how a trained tensorflow model can be packaged and executed within a mobile app?<\/p>\n","answers":[{"AnswerId":"33686627","CreationDate":"2015-11-13T05:42:34.117","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow repository includes an example Android application that uses the mobile device camera as a data source, and the Inception image classification model for inference. The source can be found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\">here<\/a>, and the repository includes both the full source code and a link to download a <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\">trained model<\/a>.<\/p>\n\n<p>The model is the Inception model that won Imagenet\u2019s Large Scale Visual Recognition Challenge in 2014.<\/p>\n"}]}
{"QuestionId":33686607,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-13T05:40:45.297","AcceptedAnswerId":null,"OwnerUserId":5557582.0,"Title":"not able to import tensorflow inside anaconda python 2.7","Body":"<p>First of all,  I am an inexperienced with most things related to software engineer, so bear with me if the question is too simple.  <\/p>\n\n<p>On my windows desktop,  I installed VMware.  Inside the vmware,  I installed Anaconda.  then using the pip command I was able to install tensorflow. <\/p>\n\n<p>However when I try to import it, the following error shows up <\/p>\n\n<pre><\/pre>\n\n<p>I can't find the GLIBC_2.17 anywhere, how can I fix this?<\/p>\n","answers":[{"AnswerId":"33686917","CreationDate":"2015-11-13T06:13:02.613","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can either ugprade your libc (see <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\">Error while importing Tensorflow in python2.7 in Ubuntu 12.04. &#39;GLIBC_2.17 not found&#39;<\/a> for example)<\/p>\n\n<p>Or you could just build from source so that it uses your local version of libc.  The latter is probably the easiest way to go.  See <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"nofollow noreferrer\">http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources<\/a><\/p>\n\n<p>(When you configure, say no to using GPU to keep things simple)<\/p>\n"},{"AnswerId":"33686929","CreationDate":"2015-11-13T06:14:18.587","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow binary packages require that the installed version of <code>glibc<\/code> (the GNU C Library) is at least 2.17. It looks like your VM has an old version of <code>glibc<\/code>, which is causing this error when you try to load TensorFlow.<\/p>\n\n<p>Since you are using VMWare, can you try creating a VM with a Ubuntu 14.04 image? We have tested with this operating system, and it has the necessary libraries to run TensorFlow.<\/p>\n"}]}
{"QuestionId":33686730,"AnswerCount":1,"Tags":"<neural-network><convolution><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-11-13T05:53:31.617","AcceptedAnswerId":33695652.0,"OwnerUserId":3929051.0,"Title":"Labels in Caffe as Images","Body":"<p>I'm new to Caffe. I am trying to implement a Fully Convolution Neural Network (<a href=\"https:\/\/gist.github.com\/shelhamer\/91eece041c19ff8968ee\" rel=\"nofollow\">FCN-8s<\/a>) for semantic segmentation. I have image data and label data, which are both images. This is for pixel-wise predictions. <\/p>\n\n<p>I tried using ImageData as the data type, but it asks for an integer label, which is not applicable to this scenario. Kindly advise as how to I can give Caffe a 2D label. Should I prefer LMDB instead of ImageData? If so, how do I proceed? I could not find any good tutorial\/documentation for a situation like this. <\/p>\n","answers":[{"AnswerId":"33695652","CreationDate":"2015-11-13T15:01:28.947","ParentId":null,"OwnerUserId":"5293046","Title":null,"Body":"<p>Since you need to achieve <strong>pixel-wise<\/strong> predictions, you can't use a single label as ground-truth. Instead, you should use a ground-truth matrix of labels.<\/p>\n\n<p>One of the Caffe guys wrote a code snippet for creating an LMDB with image data, see <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1698#issuecomment-70211045\" rel=\"noreferrer\">here<\/a>:<\/p>\n\n<pre><code>import caffe\nimport lmdb\nfrom PIL import Image\n\nin_db = lmdb.open('image-lmdb', map_size=int(1e12))\nwith in_db.begin(write=True) as in_txn:\n    for in_idx, in_ in enumerate(inputs):\n        # load image:\n        # - as np.uint8 {0, ..., 255}\n        # - in BGR (switch from RGB)\n        # - in Channel x Height x Width order (switch from H x W x C)\n        im = np.array(Image.open(in_)) # or load whatever ndarray you need\n        im = im[:,:,::-1]\n        im = im.transpose((2,0,1))\n        im_dat = caffe.io.array_to_datum(im)\n        in_txn.put('{:0&gt;10d}'.format(in_idx), im_dat.SerializeToString())\nin_db.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":33686902,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-13T06:11:04.007","AcceptedAnswerId":33687086.0,"OwnerUserId":4601596.0,"Title":"Conditional execution in TensorFlow","Body":"<p>How can I choose to execute a portion of the graph based on a condition?<\/p>\n\n<p>I have a part of my network which is to be executed only if a placeholder value is provided in . An alternate path is taken if the value is not provided. How do I go about implementing this using tensorflow?<\/p>\n\n<p>Here are the relevant portions of my code:<\/p>\n\n<pre><\/pre>\n\n<p>In the above code, I'm looking for something to use in the place of <\/p>\n","answers":[{"AnswerId":"33687086","CreationDate":"2015-11-13T06:29:39.190","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Hrm.  It's possible that what you want is tf.control_flow_ops.cond()\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py#L597\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py#L597<\/a><\/p>\n\n<p>But that's not exported into the tf namespace, and I'm answering without checking how guaranteed-stable this interface is, but it's used in released models, so go for it. :)<\/p>\n\n<p>However:  Because you actually know in advance what path you want when you construct the feed_dict, you could also take a different approach of invoking a separate path through your model.  The standard way to do this is to, e.g., set up code like:<\/p>\n\n<pre><code>def model(input, n_greater_than):\n  ... cleverness ...\n  if n_greater_than:\n     ... other cleverness...\n  return tf.reduce_mean(input)\n\n\nout1 = model(input, True)\nout2 = model(input, False)\n<\/code><\/pre>\n\n<p>And then pull the out1 or out2 nodes depending upon what you know when you're about to run your computation and set the feed_dict.  Remember that by default, if the model references the same variables (create them <em>outside<\/em> the model() func), then you'll basically have two separate paths through.<\/p>\n\n<p>You can see an example of this in the convolutional mnist example:  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/mnist\/convolutional.py#L165\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/mnist\/convolutional.py#L165<\/a><\/p>\n\n<p>I'm a fan of doing it this way without introducing control flow dependencies if you can.<\/p>\n"},{"AnswerId":"33687046","CreationDate":"2015-11-13T06:25:26.100","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Here is a simple example, that can get you started. It executes different parts of the graph based on the shape of the tensor:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.Variable([[3.0, 3.0], [3.0, 3.0]])\nb = tf.Variable([[1.0, 1.0], [2.0, 2.0]])\nl = tf.shape(a)\n\nadd_op, sub_op = tf.add(a, b), tf.sub(a, b)\n\nsess = tf.Session()\ninit = tf.initialize_all_variables()\nsess.run(init)\nt = sess.run(l)\n\nprint sess.run(sub_op if t[0] == 3 else add_op)\n\nsess.close()\n<\/code><\/pre>\n\n<p>Change 3 to 2 to see how tensor will be subtracted. As you see I initiated the nodes for <code>add<\/code> and <code>sub<\/code> and <code>shape<\/code>, then in the graph I check for the shape and go run the specific part.<\/p>\n"}]}
{"QuestionId":33687103,"AnswerCount":3,"Tags":"<python-2.7><windows-7><anaconda><theano>","CreationDate":"2015-11-13T06:31:19.810","AcceptedAnswerId":33706634.0,"OwnerUserId":395857.0,"Title":"How to install Theano on Anaconda Python 2.7 x64 on Windows?","Body":"<p>I wonder how to install Theano on Anaconda Python 2.7 x64 on Windows 7 x64. The Theano website provides some <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\">instructions<\/a> but is not clear as to what is specific to Anaconda.<\/p>\n","answers":[{"AnswerId":"33706634","CreationDate":"2015-11-14T08:39:12.857","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I'm not 100% certain but this may be a minimal set of instructions, but only if you don't want to use a GPU. Getting Theano to use a GPU on Windows is quite a bit more difficult.<\/p>\n\n<ol>\n<li>Install <a href=\"http:\/\/tdm-gcc.tdragon.net\/\" rel=\"noreferrer\">TDM GCC<\/a> x64.<\/li>\n<li>Install Anaconda x64.<\/li>\n<li>run <code>conda update conda<\/code>.<\/li>\n<li>run <code>conda update --all<\/code>.<\/li>\n<li>run <code>conda install mingw libpython<\/code>.<\/li>\n<li>Install Theano (how you do this depends on whether you want to interact with the Theano source code or not, and whether you want the \"bleeding edge\" version, or are happy with the last, but out-of-date major release).\n\n<ol>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#basic-user-install-instructions\" rel=\"noreferrer\">Older version<\/a>: e.g. <code>pip install Theano<\/code>.<\/li>\n<li><a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"noreferrer\">Bleeding edge version<\/a>: e.g. <code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git<\/code> (see linked documentation for more options)<\/li>\n<\/ol><\/li>\n<\/ol>\n\n<p>If you want multithreading support via OpenMP then things get more complicated.<\/p>\n\n<p>If you want GPU support things get <em>much<\/em> more complicated.<\/p>\n\n<p>The Windows installation instructions in the Theano documentation are fragmented at best, and terribly out of date at worst. If you need more than the basics working on Windows, you'll need to pick your way through to find an approach that works for you.<\/p>\n"},{"AnswerId":"39162453","CreationDate":"2016-08-26T09:04:14.113","ParentId":null,"OwnerUserId":"3747801","Title":null,"Body":"<p>Adding GPU support is not that much more complicated (although not intuitive)<\/p>\n\n<ol>\n<li>Install theano as in Daniel Renshaw's answer<\/li>\n<li>Go to the <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/wiki\/From-Zero-to-Lasagne-on-Windows-7-%2864-bit%29#nvidia-gpu-support-cuda\" rel=\"nofollow\">\"From Zero to Lasagne\"<\/a> tutorial, and follow it from the <code>Nvidia GPU support (CUDA)<\/code> section with the following changes:\n\n<ul>\n<li>I installed <code>visual studio 2013 community<\/code> instead of the Windows SDK<\/li>\n<li>And <code>.theanorc<\/code> should be placed in <code>C:\\Users\\USERNAME<\/code> <\/li>\n<\/ul><\/li>\n<\/ol>\n"},{"AnswerId":"42212049","CreationDate":"2017-02-13T19:30:20.577","ParentId":null,"OwnerUserId":"400589","Title":null,"Body":"<p>The trick is that <strong>YOU NEED TO CREATE AN ENVIRONMENT\/WORKSPACE FOR PYTHON<\/strong>. This solution should work for Python 2.7 but at the time of writing keras can run on python 3.5, especially if you have the latest anaconda installed (this took me awhile to figure out so I'll outline the steps I took to install KERAS in python 3.5):<\/p>\n\n<p>-- <em>CREATE ENVIRONMENT\/WORKSPACE FOR PYTHON 3.5:<\/em><\/p>\n\n<ol>\n<li><code>C:\\conda create --name neuralnets python=3.5<\/code><\/li>\n<li><code>C:\\activate neuralnets<\/code><\/li>\n<\/ol>\n\n<p>-- <em>INSTALL EVERYTHING<\/em> (notice the neuralnets workspace in parenthesis on each line). <strong>ACCEPT ANY DEPENDENCIES EACH OF THOSE STEPS WANTS TO INSTALL<\/strong>:<\/p>\n\n<ol>\n<li><code>(neuralnets) C:\\conda install theano<\/code><\/li>\n<li><code>(neuralnets) C:\\conda install mingw libpython<\/code><\/li>\n<li><code>(neuralnets) C:\\pip install tensorflow<\/code><\/li>\n<li><code>(neuralnets) C:\\pip install keras<\/code><\/li>\n<\/ol>\n\n<p>-- <em>TEST IT OUT:<\/em><\/p>\n\n<pre><code>(neuralnets) C:\\python -c \"from keras import backend; print(backend._BACKEND)\"\n<\/code><\/pre>\n\n<hr>\n\n<p>Just remember, if you want to work in the workspace you always have to do:<\/p>\n\n<pre><code>C:\\activate neuralnets\n<\/code><\/pre>\n\n<p>so you can launch Jypiter for example (assuming you also have jypiter installed in this environment\/workspace) as:<\/p>\n\n<pre><code>C:\\activate neuralnets\n(neuralnets) jypiter notebook\n<\/code><\/pre>\n\n<blockquote>\n  <p>You can read more about managing and creating conda environments\/workspaces \n  at the follwing URL: <a href=\"https:\/\/conda.io\/docs\/using\/envs.html\" rel=\"noreferrer\">https:\/\/conda.io\/docs\/using\/envs.html<\/a><\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33688713,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-13T08:41:47.760","AcceptedAnswerId":null,"OwnerUserId":5558025.0,"Title":"Raindrops Differential Equations Demo","Body":"<p>When I write all the code <a href=\"http:\/\/tensorflow.org\/tutorials\/pdes\/index.md\" rel=\"nofollow\">here<\/a> to a  file, and then run it, I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>However, it doesn't actually display an image. How can I see the actual images?<\/p>\n","answers":[{"AnswerId":"33699073","CreationDate":"2015-11-13T18:01:43.910","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This code in this tutorial was designed for pasting into IPython. If you run it in a notebook the <code>&lt;IPython.core.display.Image object&gt;<\/code> strings will be replaced with an image of the current state of the simulation.<\/p>\n\n<p>An alternative would be to generate image files that contain the same data. For example, you could replace the following code:<\/p>\n\n<pre><code>f = StringIO()\nPIL.Image.fromarray(a).save(f, fmt)\ndisplay(Image(data=f.getvalue()))\n<\/code><\/pre>\n\n<p>...with some code that writes the image to the file:<\/p>\n\n<pre><code>with open(\"\/tmp\/image.jpg\", \"w\") as f:\n  PIL.Image.fromarray(a).save(f, \"jpeg\")\n<\/code><\/pre>\n\n<p>...and then open the file <code>\/tmp\/image.jpg<\/code> in your favorite image viewer.<\/p>\n"}]}
{"QuestionId":33689598,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-13T09:32:54.480","AcceptedAnswerId":null,"OwnerUserId":2904237.0,"Title":"How to Pause \/ Resume Training in Tensorflow","Body":"<p>This question was made before the documentation for save and restore was available.\nFor now I would consider this question deprecated and say people to rely on the official documentation on <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">Save and Restore<\/a><\/p>\n\n<p>Gist of old question: <\/p>\n\n<blockquote>\n  <p>I got <strong>TF<\/strong> working fine for the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/index.md#launching_and_training_the_model\" rel=\"nofollow noreferrer\">CIFAR Tutorial<\/a>.  I've changed\n  the code to save the  (directory with checkpoint and\n  models) to a know location.<\/p>\n  \n  <p>Which brings me to my <strong>question<\/strong> :How can I pause and resume some\n  training with <strong>TF<\/strong> ?<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"56471432","CreationDate":"2019-06-06T05:36:28.403","ParentId":null,"OwnerUserId":"7240363","Title":null,"Body":"<p>As described by Hamed, the right way to do it on tensorflow is<\/p>\n\n<pre><code>    saver=tf.train.Saver()\n    save_path='checkpoints\/'\n    -----&gt; while training you can store using\n    saver.save(sess=session,save_path=save_path)\n    -----&gt; and restore\n    saver.restore(sess=session,save_path=save_path)\n<\/code><\/pre>\n\n<p>this will load the model where you last saved it and will the training(if you want) from there only.<\/p>\n"},{"AnswerId":"54677468","CreationDate":"2019-02-13T18:55:47.510","ParentId":null,"OwnerUserId":"8250471","Title":null,"Body":"<h3>Using tf.train.MonitoredTrainingSession() helped me to resume my training when my machine restarted.<\/h3>\n\n<h3>Things to keep in Mind:<\/h3>\n\n<ol>\n<li>Make sure you are saving your checkpoints. In tf.train.saver() you can specify max_checkpoints to keep.<\/li>\n<li>Specify the directory of the checkpoints in the tf.train.MonitoredTrainingSession(<em>checkpoint='dir_path',save_checkpoint_secs=<\/em>).\nBased on the <em>save_checkpoint_secs<\/em> argument, the above session would keep saving and updating the checkpoints.<\/li>\n<li>When you constantly keep saving the checkpoints, above function, looks for the latest checkpoint and resumes training from there. <\/li>\n<\/ol>\n"},{"AnswerId":"33690809","CreationDate":"2015-11-13T10:37:26.100","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>TensorFlow uses Graph-like computation, Nodes(Ops) and Edges(Variables aka states) and it provide a <code>Saver<\/code> for it's Vars. <\/p>\n\n<p>So as it's distributed computation you can run part of a graph in one machine\/processor and the rest in the other, meanwhile you can save the state(Vars) and feed it next time to continue your work. <\/p>\n\n<pre><code>saver.save(sess, 'my-model', global_step=0) ==&gt; filename: 'my-model-0'\n...\nsaver.save(sess, 'my-model', global_step=1000) ==&gt; filename: 'my-model-1000'\n<\/code><\/pre>\n\n<p>which later you can use <\/p>\n\n<pre><code>tf.train.Saver.restore(sess, save_path)\n<\/code><\/pre>\n\n<p>to restore your saved Vars.<\/p>\n\n<p><a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/state_ops.html#saving_and_restoring_variables\">Saver Usage<\/a><\/p>\n"}]}
{"QuestionId":33694368,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2015-11-13T13:56:20.417","AcceptedAnswerId":null,"OwnerUserId":3731804.0,"Title":"What is the best way to implement weight constraints in TensorFlow?","Body":"<p>Suppose we have weights<\/p>\n\n<pre><\/pre>\n\n<p>And we use the GD optimizer:<\/p>\n\n<pre><\/pre>\n\n<p>How can we implement for example non-negativity on weights?<\/p>\n\n<p>I tried clipping them:<\/p>\n\n<pre><\/pre>\n\n<p>But this slows down my training by a factor of 50.<\/p>\n\n<p>Does anybody know a good way to implement such constraints on the weights in TensorFlow?<\/p>\n\n<p>P.S.: in the equivalent Theano algorithm, I had<\/p>\n\n<pre><\/pre>\n\n<p>and it ran smoothly.<\/p>\n","answers":[{"AnswerId":"50052742","CreationDate":"2018-04-26T22:30:26.360","ParentId":null,"OwnerUserId":"9672508","Title":null,"Body":"<p>I recently had this problem as well. I discovered that you can import keras which has nice weight constraint functions as use them directly in the kernen constraint in tensorflow. Here is an example of my code. You can do similar things with kernel regularizer <\/p>\n\n<pre><code>from keras.constraints import non_neg\n\nconv1 = tf.layers.conv2d(\n    inputs=features['x'],\n    filters=32,\n    kernel_size=[5,5],\n    strides = 2,\n    padding='valid',\n    activation=tf.nn.relu,\n    kernel_regularizer=None,\n    kernel_constraint=non_neg(),\n    use_bias=False)\n<\/code><\/pre>\n"},{"AnswerId":"50026224","CreationDate":"2018-04-25T15:35:26.760","ParentId":null,"OwnerUserId":"39123","Title":null,"Body":"<p>There is a practical solution:  Your <strong>cost function<\/strong> can be written by you, to put high cost onto negative weights. I did this in a matrix factorization model in TensorFlow with python, and it worked well enough.  Right? I mean it's obvious. But nobody else mentioned it so here you go. EDIT: I just saw that Mark Borderding also gave another loss and cost-based solution implementation before I did.<\/p>\n\n<p>And if \"the best way\" is wanted, as the OP asked, what then?  Well \"best\" might actually be application-specific, in which case you'd need to try a few different ways with <em>your<\/em> dataset and consider <em>your<\/em> application requirements.<\/p>\n\n<p>Here is working code for increasing the cost for unwanted negative solution variables:<\/p>\n\n<pre><code>cost = tf.reduce_sum(keep_loss) + Lambda * reg # Cost = sum of losses for training set, except missing data.        \nif prefer_nonneg: # Optionally increase cost for negative values in rhat, if you want that.\n    negs_indices = tf.where(rhat &lt; tf.constant(0.0))\n    neg_vals = tf.gather_nd(rhat, negs_indices)\n    cost += 2. * tf.reduce_sum(tf.abs(neg_vals))  # 2 is a magic number (empirical parameter)         \n<\/code><\/pre>\n\n<p>You are free to use my code but please give me some credit if you choose to use it. Give a link to this answer on stackoverflow.com please.<\/p>\n\n<p>This design would be considered a soft constraint, because you can still get negative weights, if you let it, depending on your cost definition.<\/p>\n\n<p>It seems that constraint= is also available in TF v1.4+ as a parameter to tf.get_variable(), where you can pass a function like tf.clip_by_value. This seems like another soft constraint, not hard constraint, in my opinion, because it depends on your function to work well or not.  It also might be slow, as the other answerer tried the same function and reported it was slow to converge, although they didn't use the constraint= parameter to do this.  I don't see any reason why one would be any faster than the other since they both use the same clipping approach. So if you use the constraint= parameter then you should expect slow convergence in the context of the original poster's application.<\/p>\n\n<p>It would be nicer if also TF provided true hard constraints to the API, and let TF figure out how to both implement that as well as make it efficient on the back end. I mean, I have seen this done in linear programming solvers already for a long time. The application declares a constraint, and the back end makes it happen.<\/p>\n"},{"AnswerId":"37426800","CreationDate":"2016-05-25T02:23:58.303","ParentId":null,"OwnerUserId":"3343","Title":null,"Body":"<p>You can take the Lagrangian approach and simply add a penalty for features of the variable you don't want.  <\/p>\n\n<p>e.g.  To <em>encourage<\/em> <code>theta<\/code> to be non-negative, you could add the following to the optimizer's objective function.<\/p>\n\n<pre><code>    added_loss = -tf.minimum( tf.reduce_min(theta),0)\n<\/code><\/pre>\n\n<p>If any <code>theta<\/code> are negative, then add2loss will be positive, otherwise zero.  Scaling that to a meaningful value is left as an exercise to the reader.  Scaling too little will not exert enough pressure. Too much may make things unstable.<\/p>\n"},{"AnswerId":"43133477","CreationDate":"2017-03-31T06:02:46.463","ParentId":null,"OwnerUserId":"7700915","Title":null,"Body":"<p>By running<\/p>\n\n<pre><code>sess.run(tf.assign(x, tf.clip_by_value(x, 0, np.infty)))\n<\/code><\/pre>\n\n<p>you are consistently adding nodes to the graph and making it slower and slower.<\/p>\n\n<p>Actually you may just define a <code>clip_op<\/code> when building the graph and run it each time after updating the weights:<\/p>\n\n<pre><code># build the graph\nx = tf.Variable(np.random.random((5,10)))\nloss = ...\ntrain_op = tf.train.GradientDescentOptimizer(lr).minimize(loss)\nclip_op = tf.assign(x, tf.clip(x, 0, np.infty))\n\n# train\nsess.run(train_op)\nsess.run(clip_op)\n<\/code><\/pre>\n"},{"AnswerId":"47189751","CreationDate":"2017-11-08T21:10:07.653","ParentId":null,"OwnerUserId":"5096199","Title":null,"Body":"<p>As of TensorFlow 1.4, there is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/92838685241ca22ea2797e937e380bdbe8325784\/tensorflow\/python\/ops\/variable_scope.py#L295\" rel=\"noreferrer\">new argument<\/a> to <code>tf.get_variable<\/code> that allows to pass a constraint function that is applied after the update of the optimizer. Here is an example that enforces a non-negativity constraint:<\/p>\n\n<pre><code>with tf.variable_scope(\"MyScope\"):\n  v1 = tf.get_variable(\"v1\", \u2026, constraint=lambda x: tf.clip_by_value(x, 0, np.infty))\n<\/code><\/pre>\n\n<blockquote>\n  <p>constraint: An optional projection function to be applied to the\n  variable\n          after being updated by an <code>Optimizer<\/code> (e.g. used to implement norm\n          constraints or value constraints for layer weights). The function must\n          take as input the unprojected Tensor representing the value of the\n          variable and return the Tensor for the projected value\n          (which must have the same shape). Constraints are not safe to\n          use when doing asynchronous distributed training.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33694988,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-13T14:28:25.223","AcceptedAnswerId":33697275.0,"OwnerUserId":5559110.0,"Title":"tensorflow MNIST fully_connected_feed.py fails: range() takes at least 2 arguments (1 given)","Body":"<p>I'm having trouble running the example in one of the <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/tf\/index.md\" rel=\"nofollow noreferrer\">tensor flow tutorials<\/a>. The tutorial says to run I just need to type . When I do this it gets through fetching the input data, but then fails, like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I <em>think<\/em> this error is caused because there is some problem with session setup and\/or tensor evaluation. This is the function in mnist.py causing the problem:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>If I put all the code in the  function inside a  block, it gets past this error. However, I get other errors later about uninitialised variables, so I'm guessing something major is going wrong with session setup or initialisation, or something. Being new to tensor flow I'm a little at a loss. Any ideas?  <\/p>\n\n<p>[NB: I havent edited the code at all, just downloaded from the tensorflow tutorials and tried to run as instructed, with ]   <\/p>\n","answers":[{"AnswerId":"33697275","CreationDate":"2015-11-13T16:21:23.983","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This issue arises because in the latest version of the TensorFlow source on GitHub, <code>tf.range()<\/code> has been <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/6b12d081d54b89869e26d9d99828f13de381761e#diff-f0675f2568ff9470bcfc1f2bc79c5386L539\">updated<\/a> to be more permissive with its arguments (previously it required two arguments; now it has the same semantics as Python's <code>range()<\/code> built-in function), and the <code>fully_connected_feed.py<\/code> example has been updated to exploit this.<\/p>\n\n<p>However, if you try to run this version against the binary distribution of TensorFlow, you will get this error because the change to <code>tf.range()<\/code> has not been incorporated into the binary package.<\/p>\n\n<p>The easiest solution is to download <a href=\"https:\/\/raw.githubusercontent.com\/tensorflow\/tensorflow\/1d76583411038767f673a0c96174c80eaf9ff42f\/tensorflow\/g3doc\/tutorials\/mnist\/mnist.py\">the old version of <code>mnist.py<\/code><\/a>. Alternatively, you could <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\">build from source<\/a> to use the latest version of the tutorial.<\/p>\n"},{"AnswerId":"33741923","CreationDate":"2015-11-16T18:13:12.157","ParentId":null,"OwnerUserId":"5568791","Title":null,"Body":"<p>you can right result fix mnist code like this : <\/p>\n\n<pre><code>indices = tf.expand_dims(tf.range(0,batch_size),1)\n<\/code><\/pre>\n"},{"AnswerId":"33696551","CreationDate":"2015-11-13T15:46:32.487","ParentId":null,"OwnerUserId":"1953800","Title":null,"Body":"<p><code>TypeError: range() takes at least 2 arguments (1 given)<\/code><\/p>\n\n<p>That's the error.<\/p>\n\n<p>Looking at the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/constant_op.md#range\" rel=\"nofollow\">tensorflow docs for range<\/a>, we can see that <code>range<\/code> has a function signature of <code>start, limit, delta=1, name='range'<\/code>. This means that at least two arguments are required for function invocation. Your example only shows one argument provided.<\/p>\n\n<p>An example can be found in the docs:<\/p>\n\n<pre><code># 'start' is 3\n# 'limit' is 18\n# 'delta' is 3\ntf.range(start, limit, delta) ==&gt; [3, 6, 9, 12, 15]\n<\/code><\/pre>\n"}]}
{"QuestionId":33695113,"AnswerCount":1,"Tags":"<lua><neural-network><torch><conv-neural-network>","CreationDate":"2015-11-13T14:35:21.437","AcceptedAnswerId":33727957.0,"OwnerUserId":1114272.0,"Title":"Visualize Neural Net Layers in Torch, without itorch","Body":"<p>I have a neural net with several convolution layers and I'd like to visualize the feature maps I'm creating.<\/p>\n\n<p>I've seen this post:\n<a href=\"https:\/\/stackoverflow.com\/questions\/31509912\/visualize-images-in-intermediate-layers-in-torch-lua\">Visualize images in intermediate layers in torch (lua)<\/a><\/p>\n\n<p>which suggests using itorch, but it requires running my code in an itorch notebook which I would like to avoid. <\/p>\n\n<p>Are there any other Torch packages which can be used to visualize convolution layers?<\/p>\n\n<p>EDIT (with detailed solution):<\/p>\n\n<p>Since I found so few resources online about how to go about doing this I documented my full solution and slapped it on Github. Anyone who wants to visualize neural nets in torch can just go here to get started!<\/p>\n\n<p><a href=\"https:\/\/github.com\/egaebel\/torch-neural-network-visualization\" rel=\"nofollow\">https:\/\/github.com\/egaebel\/torch-neural-network-visualization<\/a><\/p>\n\n<p>Much thanks again to YuTse for the gnuplot tip!<\/p>\n","answers":[{"AnswerId":"33727957","CreationDate":"2015-11-16T03:13:22.443","ParentId":null,"OwnerUserId":"5118777","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/torch\/gnuplot\" rel=\"nofollow\">gnuplot<\/a><\/p>\n\n<p>in itorch console mode(<code>itorch<\/code>) \/ torch mode(<code>th<\/code>)<\/p>\n\n<pre><code>require 'image';\n\na = image.lena();\n\nrequire 'gnuplot';\n\ngnuplot.figure(1);\n\ngnuplot.imagesc(a[1])\n<\/code><\/pre>\n"}]}
{"QuestionId":33696775,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-11-13T15:56:35.783","AcceptedAnswerId":33706797.0,"OwnerUserId":1447786.0,"Title":"Why does this minimal RNN code throw a type error for a type never used?","Body":"<p>I'm trying to implement a minimal recurrent neural network example in . I expected the following python script to print a 10-by-20 matrix representing a hidden state sequence.<\/p>\n\n\n\n<pre><\/pre>\n\n<p>Instead, it returns the error:\n    <\/p>\n\n<p>This is especially confusing since, as far as I can tell, none of my variables are 3-tensors!<\/p>\n","answers":[{"AnswerId":"33706797","CreationDate":"2015-11-14T08:59:56.617","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There is a syntax error in the code the question's code: the <code>reshape((1,3))<\/code> at the end of the <code>fn =<\/code> is invalid and appears to have been added in error. The code runs when this element of that line is simple deleted. The remainder of this answer assumes that edit is as intended by the question author.<\/p>\n\n<p>I have not reproduced the stated error. This may be because I am using the bleeding edge version of Theano. It's quite possible the error message for this situation has been changed in the most recent versions of the code. However, the syntax error in the code above suggests another possibility: the question code is not actually the code that produced the error pasted into the question.<\/p>\n\n<p>Using the edited code, and the latest version of Theano, I get the error<\/p>\n\n<blockquote>\n  <p>TypeError: ('The following error happened while compiling the node', forall_inplace,cpu,scan_fn}(Shape_i{0}.0, Subtensor{int64:int64:int8}.0, IncSubtensor{InplaceSet;:int64:}.0, , , ), '\\n', \"Inconsistency in the inner graph of scan 'scan_fn' : an input and an output are associated with the same recurrent state and should have the same type but have type 'TensorType(float64, row)' and 'TensorType(float64, matrix)' respectively.\")<\/p>\n<\/blockquote>\n\n<p>This is similar in spirit to the question's error, but refers to a mismatch between a matrix and row vector; there is no reference to a 3-tensor.<\/p>\n\n<p>The simplest change to avoid this error is to change the shape of the <code>b<\/code> shared variable.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>b = shared(np.random.uniform(-1.0, 1.0, (1, 20)))\n<\/code><\/pre>\n\n<p>use<\/p>\n\n<pre><code>b = shared(np.random.uniform(-1.0, 1.0, (20,)))\n<\/code><\/pre>\n\n<p>I would also recommend doing the same for the initial value of <code>h_tm1<\/code>.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>outputs_info = np.zeros((1, 20))\n<\/code><\/pre>\n\n<p>use<\/p>\n\n<pre><code>outputs_info = np.zeros((20,))\n<\/code><\/pre>\n"},{"AnswerId":"39955247","CreationDate":"2016-10-10T09:30:17.867","ParentId":null,"OwnerUserId":"6945117","Title":null,"Body":"<p>The problem is because theano wrap the output_info into unsuitable variables. For instance <code>outputs_info = np.zeros((3, 20))<\/code> will be wrapped into <code>TensorType(float64, matrix)<\/code>. But for the dim equals 1, theano autonatically braodcasts the dim, so <code>outputs_info = np.zeros((1, 20))<\/code> will be <code>TensorType(float64, row)<\/code> since the first dim is infered broadcastable. <\/p>\n\n<p>A solution is <code>outputs_info = T.unbroadcast(T.zeros((1,20)), 0))<\/code>, this makes sure it wrapped as matrix.<\/p>\n\n<pre><code># import packages\/functions\nfrom theano import shared, scan, function, config, tensor as T\nimport numpy as np\n\n# declare variables\nX = T.tensor3(\"X\")\nWx = shared(np.asarray(np.random.uniform(-1.0, 1.0, (10, 20)), dtype=config.floatX))\nWh = shared(np.asarray(np.random.uniform(-1.0, 1.0, (20, 20)), dtype=config.floatX))\nb = shared(np.asarray(np.random.uniform(-1.0, 1.0, (1, 20)), dtype=config.floatX))\n\n# define recurrence function\ndef recurrence(x_t, h_tm1):\n    print h_tm1.type, x_t.type\n    return T.nnet.sigmoid(T.dot(h_tm1, Wh) + T.dot(x_t, Wx) + b)\n\n# compute hidden state sequence with scan\nht, _ = scan(fn = recurrence, sequences = X,\n             outputs_info = T.unbroadcast(T.zeros((1,20)), 0))\n\n# define function producing hidden state sequence\nfn = function([X], ht)\n\n# test function\nprint fn(np.eye(10,10,dtype=config.floatX).reshape(10,1,10))\n<\/code><\/pre>\n"}]}
{"QuestionId":33698510,"AnswerCount":1,"Tags":"<python><matrix><machine-learning><scikit-learn><tensorflow>","CreationDate":"2015-11-13T17:27:34.790","AcceptedAnswerId":33712950.0,"OwnerUserId":678572.0,"Title":"Use attribute and target matrices for TensorFlow Linear Regression Python","Body":"<p>I'm trying to follow <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.md\" rel=\"noreferrer\">this tutorial<\/a>.<\/p>\n\n<p>TensorFlow just came out and I'm really trying to understand it.  I'm familiar with <em>penalized linear regression<\/em> like Lasso, Ridge, and ElasticNet and its usage in .  <\/p>\n\n<p>For  Lasso regression, all I need to input into the regression algorithm is  [an M x N dimensional attribute matrix (pd.DataFrame)] and  [an M dimensional target vector (pd.Series)].  The  structure in TensorFlow is a bit new to me and I'm not sure how to structure my input data into what it wants. <\/p>\n\n<p>It seems as if softmax regression is for classification.  <strong>How can I restructure my  (M x N attribute matrix) and  (M dimensional target vector) to input into  for linear regression?<\/strong> <\/p>\n\n<p>My current method for doing a Linear Regression uses pandas, numpy, and sklearn and it's shown below. I think this question will be really helpful for people getting familiar with TensorFlow:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33712950","CreationDate":"2015-11-14T20:24:28.933","ParentId":null,"OwnerUserId":"5520753","Title":null,"Body":"<p>Softmax is an only addition function (in logistic regression for example), it is not a model like <\/p>\n\n<pre><code>model = LassoCV()\nmodel.fit(DF_X,SR_y)\n<\/code><\/pre>\n\n<p>Therefore you can't simply give it data with fit method. However, you can simply create your model with the help of TensorFlow functions.<\/p>\n\n<p>First of all, you have to create a computational graph, for example for linear  regression you will create tensors with the size of your data. They are only tensors and you will give them your array in another part of the program. <\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.placeholder(\"float\", [4, 3])      \ny_ = tf.placeholder(\"float\",[4])\n<\/code><\/pre>\n\n<p>When you create two variables, that will contain initial weights of our model<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([3,1]))\nb = tf.Variable(tf.zeros([1]))\n<\/code><\/pre>\n\n<p>And now you can create the model (you want to create regression, not classification therefore you don't need to use tf.nn.softmax ) <\/p>\n\n<pre><code>y=tf.matmul(x,W) + b\n<\/code><\/pre>\n\n<p>As you have regression and linear model you will use <\/p>\n\n<pre><code>loss=tf.reduce_sum(tf.square(y_ - y))\n<\/code><\/pre>\n\n<p>Then we will train our model with the same step as in the tutorial <\/p>\n\n<pre><code>train_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\n<\/code><\/pre>\n\n<p>Now that you created the computational graph you have to write one more part of the program, where you will use this graph to work with your data. <\/p>\n\n<pre><code>init = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)       \nsess.run(train_step, feed_dict={x:np.asarray(DF_X),y_:np.asarray(SR_y)})\n<\/code><\/pre>\n\n<p>Here you give your data to this computational graph with the help of feed_dict. In TensorFlow you provide information in numpy arrays.\nIf you want to see your mistake you can write <\/p>\n\n<pre><code>sess.run(loss,feed_dict={x:np.asarray(DF_X),y_:np.asarray(SR_y)}) \n<\/code><\/pre>\n"}]}
{"QuestionId":33699174,"AnswerCount":4,"Tags":"<gradient-descent><tensorflow>","CreationDate":"2015-11-13T18:07:10.640","AcceptedAnswerId":34858399.0,"OwnerUserId":1111929.0,"Title":"TensorFlow's ReluGrad claims input is not finite","Body":"<p>I'm trying out TensorFlow and I'm running into a strange error. I edited the deep MNIST example to use another set of images, and the algorithm converges nicely again, until around iteration 8000 (accuracy 91% at that point) when it crashes with the following error.<\/p>\n\n<pre><\/pre>\n\n<p>At first I thought maybe some coefficients were reaching the limit for a float, but adding l2 regularization on all weights &amp; biases didn't resolve the issue. It's always the first relu application that comes out of the stacktrace: <\/p>\n\n<pre><\/pre>\n\n<p>I'm working on CPU only for now. Any idea what could cause this and how to work around it?<\/p>\n\n<p>Edit: I traced it down to this problem <a href=\"https:\/\/stackoverflow.com\/questions\/33712178\/tensorflow-nan-bug\">Tensorflow NaN bug?<\/a>, the solution there works.<\/p>\n","answers":[{"AnswerId":"33702283","CreationDate":"2015-11-13T21:46:13.850","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I have experienced this error: <code>input is not finite<\/code> earlier (not with <code>tf.nn.relu<\/code>). In my case the problem was that the elements in my tensor variable reached very big number (which marked them as infinite and hence the message <code>input is not finite<\/code>).<\/p>\n\n<p>I would suggest to add a bunch of debugging output to <code>tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)<\/code> at every n-th iteration to track when exactly it reached infinity.<\/p>\n\n<p>This looks consistent with your comment:  <\/p>\n\n<blockquote>\n  <p>I modify the value to 1e-3, the crash occurs significantly earlier.\n  However, changing it to 1e-5 prevents the algorithm from converging<\/p>\n<\/blockquote>\n"},{"AnswerId":"34362080","CreationDate":"2015-12-18T18:44:54.047","ParentId":null,"OwnerUserId":"5696306","Title":null,"Body":"<p><strong>Error is due to 0log(0)<\/strong><\/p>\n\n<p>This can be avoided by:<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y*tf.log(yconv+ 1e-9))\n<\/code><\/pre>\n"},{"AnswerId":"34858399","CreationDate":"2016-01-18T15:31:32.280","ParentId":null,"OwnerUserId":"1111929","Title":null,"Body":"<p>Since I had another topic on this issue [ <a href=\"https:\/\/stackoverflow.com\/questions\/33712178\/tensorflow-nan-bug\">Tensorflow NaN bug?<\/a> ] I didn't keep this one updated, but the solution has been there for a while and has since been echoed by posters here. The problem is indeed 0*log(0) resulting in an NaN. <\/p>\n\n<p>One option is to use the line Muaaz suggests here or the one I wrote in the linked topic. But in the end TensorFlow has this routine embedded: <code>tf.nn.softmax_cross_entropy_with_logits<\/code> and this is more efficient and should hence be preferred when possible. This should be used where possible instead of the things me and Muaaz suggested earlier, as pointed out by a commentor on said link.<\/p>\n"},{"AnswerId":"34829494","CreationDate":"2016-01-16T16:48:19.807","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>Can't comment because of reputation, but Muaaz has the answer. The error can be replicated by training a system which has 0 error - resulting in log(0). His solution works to prevent this. Alternatively catch the error and move on.<\/p>\n\n<pre><code>...your other code...\ntry :\n  for i in range(10000):\n    train_accuracy = accuracy.eval(feed_dict={\n            x:batch_xs, y_: batch_ys, keep_prob: 1.0})\n\nexcept : print(\"training interupted. Hopefully deliberately\")\n<\/code><\/pre>\n"}]}
{"QuestionId":33699798,"AnswerCount":1,"Tags":"<gpu><theano><wubi>","CreationDate":"2015-11-13T18:47:31.993","AcceptedAnswerId":33706480.0,"OwnerUserId":395857.0,"Title":"Can Theano use the GPU on a Linux installed with Wubi?","Body":"<p>I have a computer running Microsoft Windows 7 with an Nvidia card. If I install Ubuntu with Wubi, can Theano use the GPU, and if so would it be just as efficiently as I was using Ubuntu without Wubi?<\/p>\n","answers":[{"AnswerId":"33706480","CreationDate":"2015-11-14T08:18:50.660","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Wubi is just an easy way to install Ubuntu so it dual-boots with Windows. Once Ubuntu is installed neither Windows nor Ubuntu operate in any fundamentally different way compared to the same end result via an alternate installation method.<\/p>\n\n<p>If you think that installing Ubuntu via Wubi will in some way change Windows so Theano on Windows can\/must use the GPU differently\/better then you're mistaken. You'll need to install Theano in each operating system you want to run it in. And the way in which Theano uses the GPU will be depend on each installation independent of any others. So, if you have a Windows\/Ubuntu dual-boot system (whether that is achieved via Wubi or not), an installation of Theano in Windows will have no impact on an installation of Theano in Ubuntu, and vice versa.<\/p>\n"}]}
{"QuestionId":33700049,"AnswerCount":6,"Tags":"<python><tensorflow>","CreationDate":"2015-11-13T19:04:25.530","AcceptedAnswerId":36744555.0,"OwnerUserId":5560069.0,"Title":"Get the diagonal of a matrix in TensorFlow","Body":"<p>Is there a way to extract the diagonal of a square matrix in TensorFlow? That is, for a matrix like this:<\/p>\n\n<pre><\/pre>\n\n<p>I want to fetch the elements: <\/p>\n\n<p>In numpy, this is pretty straight-forward via <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.diag.html\" rel=\"nofollow\">np.diag<\/a>:<\/p>\n\n<p>In TensorFlow, there is a <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/math_ops.md#diag\" rel=\"nofollow\">diag function<\/a>, but it only forms a new matrix with the elements specified in the argument on the diagonal, which is not what I want.<\/p>\n\n<p>I could imagine how this could be done via striding... but I don't see striding for tensors in TensorFlow.<\/p>\n","answers":[{"AnswerId":"33704034","CreationDate":"2015-11-14T01:03:21.160","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>This is probably is a workaround, but works.<\/p>\n\n<pre><code>&gt;&gt; sess = tensorflow.InteractiveSession()\n&gt;&gt; x = tensorflow.Variable([[1,2,3],[4,5,6],[7,8,9]])\n&gt;&gt; x.initializer.run()\n&gt;&gt; z = tensorflow.pack([x[i,i] for i in range(3)])\n&gt;&gt; z.eval()\narray([1, 5, 9], dtype=int32)\n<\/code><\/pre>\n"},{"AnswerId":"36744555","CreationDate":"2016-04-20T12:57:54.423","ParentId":null,"OwnerUserId":"863543","Title":null,"Body":"<p>with tensorflow 0.8 its possible to extract the diagonal elements with <code>tf.diag_part()<\/code> (see <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/math_ops.html#diag_part\" rel=\"nofollow noreferrer\">documentation<\/a>)<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>for tensorflow >= r1.12 its <code>tf.linalg.tensor_diag_part<\/code> (see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/linalg\/tensor_diag_part\" rel=\"nofollow noreferrer\">documentation<\/a>)<\/p>\n"},{"AnswerId":"41372642","CreationDate":"2016-12-29T03:11:32.043","ParentId":null,"OwnerUserId":"6472284","Title":null,"Body":"<p>Use the tf.diag_part()<\/p>\n\n<pre><code>with tf.Session() as sess:\n    x = tf.ones(shape=[3, 3])\n    x_diag = tf.diag_part(x)\n    print(sess.run(x_diag ))\n<\/code><\/pre>\n"},{"AnswerId":"34831682","CreationDate":"2016-01-16T20:17:48.563","ParentId":null,"OwnerUserId":"152244","Title":null,"Body":"<p>Depending on the context, a mask can be a nice way to `cancel' off diagonal elements of the matrix, especially if you plan in reducing it anyway:<\/p>\n\n<pre><code>mask = tf.diag(tf.ones([n]))\ny = tf.mul(mask,y)\ncost = -tf.reduce_sum(y)\n<\/code><\/pre>\n"},{"AnswerId":"34604943","CreationDate":"2016-01-05T05:44:41.157","ParentId":null,"OwnerUserId":"3148000","Title":null,"Body":"<p>Use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#gather\" rel=\"nofollow\"><code>gather<\/code><\/a> operation.<\/p>\n\n<pre><code>x = tensorflow.Variable([[1,2,3],[4,5,6],[7,8,9]])\nx_flat = tf.reshape(x, [-1])  # flatten the matrix\nx_diag = tf.gather(x, [0, 3, 6])\n<\/code><\/pre>\n"},{"AnswerId":"33704565","CreationDate":"2015-11-14T02:36:31.043","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Currently it is possible to extract diagonal elements with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/matrix_math_functions#diag_part\" rel=\"nofollow noreferrer\">tf.diag_part<\/a>. Here is their example:<\/p>\n\n<pre><code>\"\"\"\n'input' is [[1, 0, 0, 0],\n            [0, 2, 0, 0],\n            [0, 0, 3, 0],\n            [0, 0, 0, 4]]\n\"\"\"\n\ntf.diag_part(input) ==&gt; [1, 2, 3, 4]\n<\/code><\/pre>\n\n<hr>\n\n<p>Old answer (when diag_part) was not available (still relevant if you want to achieve something that is not available now):<\/p>\n\n<p>After looking though the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/math_ops.md\" rel=\"nofollow noreferrer\">math operations<\/a> and <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md\" rel=\"nofollow noreferrer\">tensor transformations<\/a>, it does not look like such operation exists. Even if you can extract this data with matrix multiplications it would not be efficient (get diagonal is <code>O(n)<\/code>).<\/p>\n\n<p>You have three approaches, starting with easy to hard.<\/p>\n\n<ol>\n<li>Evaluate the tensor, extract diagonal with numpy, build a variable with TF<\/li>\n<li>Use <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.md#pack\" rel=\"nofollow noreferrer\">tf.pack<\/a> in a way Anurag suggested (also extract the value 3 using <code>tf.shape<\/code><\/li>\n<li>Write your own <a href=\"http:\/\/tensorflow.org\/how_tos\/adding_an_op\/index.md\" rel=\"nofollow noreferrer\">op in C++<\/a>, rebuild TF and use it natively.<\/li>\n<\/ol>\n"}]}
{"QuestionId":33702064,"AnswerCount":1,"Tags":"<macos><docker><tensorflow>","CreationDate":"2015-11-13T21:28:29.173","AcceptedAnswerId":null,"OwnerUserId":274502.0,"Title":"TensorFlow docker dev workflow on mac","Body":"<p>There is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/docker\" rel=\"nofollow noreferrer\">an official guide<\/a> on how to install it that doesn't say much about actually developing in it.<\/p>\n\n<p>From what I understand, there is a <a href=\"https:\/\/stackoverflow.com\/questions\/30090007\/whats-the-right-way-to-setup-a-development-environment-on-os-x-with-docker\">quite big challenge in developing with Docker in general<\/a>. Not to mention there could be <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/203#issuecomment-156507498\" rel=\"nofollow noreferrer\">deeper technical complications<\/a> about going with it for TensorFlow, maybe mostly <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/149\" rel=\"nofollow noreferrer\">thanks to GPUs<\/a>. So there is a lot of stuff to after pulling the docker image...<\/p>\n\n<p>Does anyone have a step by step guide on how to get development going here?<\/p>\n","answers":[{"AnswerId":"34444753","CreationDate":"2015-12-23T22:33:29.630","ParentId":null,"OwnerUserId":"2138200","Title":null,"Body":"<p>You could mount a local directory to the docker container so that you can still use your preferred editor in osx. Here's a command to start the container with a mounted directory and run a command:<\/p>\n\n<p><code>docker run --name tensorflow --rm -v \/Users\/me\/Code\/web\/tensorflow_dev:\/tensorflow_dev b.gcr.io\/tensorflow\/tensorflow \/bin\/sh -c 'cd \/tensorflow_dev &amp;&amp; python mnist.py'<\/code><\/p>\n\n<p><code>-v<\/code> will mount the local directory and the <code>-c<\/code> will run the specified command. So your flow might look like:<\/p>\n\n<ol>\n<li>Edit python script in your favorite editor<\/li>\n<li>Run the above command to excute your script<\/li>\n<\/ol>\n\n<p>However, I actually use <a href=\"https:\/\/www.jetbrains.com\/pycharm\/\" rel=\"nofollow\">pycharm<\/a> so that I can place breakpoints and run the python script interactively within the editor. <\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":33702251,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-13T21:42:58.673","AcceptedAnswerId":33702428.0,"OwnerUserId":1936768.0,"Title":"tensorflow loss minimization type error","Body":"<p>I have a loss function implemented in TensorFlow that computes mean squared error. All tensors being used to compute the objective are of type float64 and therefore the loss function itself is of dtype float64. In particular,<\/p>\n\n<pre><\/pre>\n\n<p>However, when I attempt to minimize I obtain a value error with respect to type of the tensor:<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand why the expected dtype of the tensor is a single precision float when all variables leading up to the computation are of type float64. I have confirmed that when I coerce all variables to be float32 the computation executes correctly.<\/p>\n\n<p>Does anyone have any insight as to why this could be happening? My computer is a 64bit machine.<\/p>\n\n<p>Here is an example that reproduces the behavior<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33702428","CreationDate":"2015-11-13T21:57:06.123","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Currently the <code>tf.train.GradientDescentOptimizer<\/code> class only <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/optimizer.py#L325\" rel=\"nofollow\">supports<\/a> training on 32-bit floating-point variables and loss values. <\/p>\n\n<p>However, it looks like the kernel is implemented for double-precision values, so it should be possible to train in your scenario. <\/p>\n\n<p>A quick workaround would be to define a subclass that supports <code>tf.float64<\/code> values as well:<\/p>\n\n<pre><code>class DoubleGDOptimizer(tf.train.GradientDescentOptimizer):\n  def _valid_dtypes(self):\n    return set([tf.float32, tf.float64])\n<\/code><\/pre>\n\n<p>...and then use DoubleGDOptimizer in place of <code>tf.train.GradientDescentOptimizer<\/code>.<\/p>\n\n<p><strong>EDIT:<\/strong> You'll need to pass in the learning rate as <code>tf.constant(learning_rate, tf.float64)<\/code> to make this work.<\/p>\n\n<p>(<strong>N.B.<\/strong> This isn't a supported interface and it may be subject to change in future, but the team is aware of the desire for optimizing double-precision floats, and intends to provide a built-in solution.)<\/p>\n"}]}
{"QuestionId":33702307,"AnswerCount":1,"Tags":"<speech-recognition><speech-to-text><tensorflow>","CreationDate":"2015-11-13T21:47:45.233","AcceptedAnswerId":null,"OwnerUserId":5560490.0,"Title":"Speech to text using TensorFlow","Body":"<p>I'm new to TensorFlow and I am looking for help on a speech to text recognition project. Is there an example that showcases how to use TensorFlow for speech to text? I hear that it was used within Google to improve accuracy by 25%<\/p>\n","answers":[{"AnswerId":"33703819","CreationDate":"2015-11-14T00:31:11.083","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>Although there is no speech to text demo in tensorflow. There are demos involving <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md\" rel=\"noreferrer\">sequence-to-sequence modelling<\/a> which can be easily used in the context of speech to text.<\/p>\n"}]}
{"QuestionId":33702668,"AnswerCount":2,"Tags":"<windows><theano>","CreationDate":"2015-11-13T22:16:21.973","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Using Theano on Windows: missing lazylinker_ext.pyd","Body":"<p>When running a Python script that contain , I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>It looks like  is caused by the absence of the file . How am I supposed to get this file?<\/p>\n\n<hr>\n\n<p>I have installed Theano in Anaconda Python 2.7.10 X64 as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Then I changed the g++ compiler namely <a href=\"http:\/\/sourceforge.net\/projects\/tdm-gcc\/?source=typ_redirect\" rel=\"nofollow\">TDM-GCC MinGW Compiler<\/a> so that it compiles to x64 (otherwise I get ).<\/p>\n","answers":[{"AnswerId":"51176183","CreationDate":"2018-07-04T14:40:00.617","ParentId":null,"OwnerUserId":"10032785","Title":null,"Body":"<p>I've spent a couple of days on this.  Ultimately I had to update theano AFTER changing my G++ compiler to 64 bit (I used <a href=\"http:\/\/mingw-w64.org\/doku.php\" rel=\"nofollow noreferrer\">http:\/\/mingw-w64.org\/doku.php<\/a>). Windows \"where\" command makes it easy to see where the g++ is installed: <strong>where g++<\/strong> gave me several locations including the 32bit version.<\/p>\n"},{"AnswerId":"33706413","CreationDate":"2015-11-14T08:09:24.357","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>From <a href=\"https:\/\/stackoverflow.com\/questions\/3848357\/createprocess-no-such-file-or-directory\">here<\/a>.<\/p>\n\n<p>Have you closed your Command window and opened a new one? Changes to your environment variables (via the standard Windows dialogue window) will not apply to open Command windows (you could set the variables separately and manually in an open Command window though).<\/p>\n\n<p>If setting environment variables doesn't help, the other answers to the <a href=\"https:\/\/stackoverflow.com\/questions\/3848357\/createprocess-no-such-file-or-directory\">linked question<\/a> may help.<\/p>\n"}]}
{"QuestionId":33703624,"AnswerCount":5,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2015-11-14T00:03:32.687","AcceptedAnswerId":null,"OwnerUserId":2669077.0,"Title":"How does tf.app.run() work?","Body":"<p>How does  work in Tensorflow translate demo? <\/p>\n\n<p>In , there is a call to . How is it being handled?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43691911","CreationDate":"2017-04-29T04:52:55.777","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>There is nothing special in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/app\" rel=\"noreferrer\"><code>tf.app<\/code><\/a>. This is just a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/platform\/app.py\" rel=\"noreferrer\">generic entry point script<\/a>, which <\/p>\n\n<blockquote>\n  <p>Runs the program with an optional 'main' function and 'argv' list.<\/p>\n<\/blockquote>\n\n<p>It has nothing to do with neural networks and it just calls the main function, passing through any arguments to it.<\/p>\n"},{"AnswerId":"53585086","CreationDate":"2018-12-02T22:11:19.477","ParentId":null,"OwnerUserId":"690529","Title":null,"Body":"<p>Google code depends on a lot on global flags being accessing in libraries\/binaries\/python scripts and so tf.app.run() parses out those flags to create a global state in FLAGs(or something similar) variable and then calls python main() as it should. <\/p>\n\n<p>If they didn't have this call to tf.app.run(), then users might forget to do FLAGs parsing, leading to these libraries\/binaries\/scripts not having access to FLAGs they need. <\/p>\n"},{"AnswerId":"40766480","CreationDate":"2016-11-23T13:59:49.353","ParentId":null,"OwnerUserId":"6808091","Title":null,"Body":"<pre><code>if __name__ == \"__main__\":\n<\/code><\/pre>\n\n<p>means current file is executed under a shell instead of imported as a module.<\/p>\n\n<pre><code>tf.app.run()\n<\/code><\/pre>\n\n<p>As you can see through the file <code>app.py<\/code><\/p>\n\n<pre><code>def run(main=None, argv=None):\n  \"\"\"Runs the program with an optional 'main' function and 'argv' list.\"\"\"\n  f = flags.FLAGS\n\n  # Extract the args from the optional `argv` list.\n  args = argv[1:] if argv else None\n\n  # Parse the known flags from that list, or from the command\n  # line otherwise.\n  # pylint: disable=protected-access\n  flags_passthrough = f._parse_flags(args=args)\n  # pylint: enable=protected-access\n\n  main = main or sys.modules['__main__'].main\n\n  # Call the main function, passing through any arguments\n  # to the final program.\n  sys.exit(main(sys.argv[:1] + flags_passthrough))\n<\/code><\/pre>\n\n<p>Let's break line by line:<\/p>\n\n<pre><code>flags_passthrough = f._parse_flags(args=args)\n<\/code><\/pre>\n\n<p>This ensures that the argument you pass through command line is valid,e.g. \n<code>python my_model.py --data_dir='...' --max_iteration=10000<\/code> Actually, this feature is implemented based on python standard <code>argparse<\/code> module.<\/p>\n\n<pre><code>main = main or sys.modules['__main__'].main\n<\/code><\/pre>\n\n<p>The first <code>main<\/code> in right side of <code>=<\/code> is the first argument of current function <code>run(main=None, argv=None)<\/code>\n. While <code>sys.modules['__main__']<\/code> means current running file(e.g. <code>my_model.py<\/code>).<\/p>\n\n<p>So there are two cases:<\/p>\n\n<ol>\n<li><p>You don't have a <code>main<\/code> function in <code>my_model.py<\/code> Then you have to\ncall <code>tf.app.run(my_main_running_function)<\/code><\/p><\/li>\n<li><p>you have a <code>main<\/code> function in <code>my_model.py<\/code>. (This is mostly the case.)<\/p><\/li>\n<\/ol>\n\n<p>Last line:<\/p>\n\n<pre><code>sys.exit(main(sys.argv[:1] + flags_passthrough))\n<\/code><\/pre>\n\n<p>ensures your <code>main(argv)<\/code> or <code>my_main_running_function(argv)<\/code> function is called with parsed arguments properly.<\/p>\n"},{"AnswerId":"46613037","CreationDate":"2017-10-06T19:53:01.247","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>In simple terms, the job of <strong><code>tf.app.run()<\/code><\/strong> is to <em>first<\/em> set the global flags for later usage like:<\/p>\n\n<pre><code>from tensorflow.python.platform import flags\nf = flags.FLAGS\n<\/code><\/pre>\n\n<p>and then run your <em>custom main<\/em> function with a set of arguments.<\/p>\n\n<p>For e.g. in <a href=\"https:\/\/github.com\/tensorflow\/nmt\/tree\/master\/nmt\" rel=\"nofollow noreferrer\">TensorFlow NMT<\/a> codebase, the very first entry point for the program execution for training\/inference starts at this point (see below code)<\/p>\n\n<pre><code>if __name__ == \"__main__\":\n  nmt_parser = argparse.ArgumentParser()\n  add_arguments(nmt_parser)\n  FLAGS, unparsed = nmt_parser.parse_known_args()\n  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)\n<\/code><\/pre>\n\n<p>After parsing the arguments using <code>argparse<\/code>, with <code>tf.app.run()<\/code> you run the function \"main\" which is defined like:<\/p>\n\n<pre><code>def main(unused_argv):\n  default_hparams = create_hparams(FLAGS)\n  train_fn = train.train\n  inference_fn = inference.inference\n  run_main(FLAGS, default_hparams, train_fn, inference_fn)\n<\/code><\/pre>\n\n<p>So, after setting the flags for global use, <strong><code>tf.app.run()<\/code><\/strong> simply runs that <code>main<\/code> function that you pass to it with <code>argv<\/code> as its parameters.<\/p>\n\n<p>P.S.: As <a href=\"https:\/\/stackoverflow.com\/a\/43691911\/2956066\">Salvador Dali's answer<\/a> says, it's just a good software engineering practice, I guess, although I'm not sure whether TensorFlow performs any optimized run of the <code>main<\/code> function than that was run using normal CPython.<\/p>\n"},{"AnswerId":"33704794","CreationDate":"2015-11-14T03:29:18.993","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>It's just a very quick wrapper that handles flag parsing and then dispatches to your own main. See the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9dc6c17797c065796603d9259b2aa57b3c07ff71\/tensorflow\/python\/platform\/app.py#L31-L48\" rel=\"noreferrer\">code<\/a>.<\/p>\n"}]}
{"QuestionId":33705284,"AnswerCount":2,"Tags":"<text-classification><tensorflow>","CreationDate":"2015-11-14T05:00:56.467","AcceptedAnswerId":null,"OwnerUserId":4040618.0,"Title":"TensorFlow - Text Classification using Neural Networks","Body":"<p>Is there any example on how can TensorFlow be used for text classification using neural networks?<\/p>\n","answers":[{"AnswerId":"33928416","CreationDate":"2015-11-25T23:49:46.810","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>I've started putting together a set of examples for text classification on DBPedia dataset (predicting class of object from its description) as part of examples for <strong>Scikit Flow<\/strong>:\n<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>Going to expand this example and write a blog post when will have enough different models showcased. Feel free to suggest other datasets and models you would be interested to see.<\/p>\n"},{"AnswerId":"37337441","CreationDate":"2016-05-20T03:26:10.517","ParentId":null,"OwnerUserId":"3896984","Title":null,"Body":"<p>Denny Britz has some great tutorials for Deep learning on his blog at <a href=\"http:\/\/www.wildml.com\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>and has an example on github <a href=\"https:\/\/github.com\/dennybritz\/cnn-text-classification-tf\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>Some of his examples don't use tensor flow but the one I linked to github uses CNN for text classification using tensor flow<\/p>\n"}]}
{"QuestionId":33705404,"AnswerCount":2,"Tags":"<python><reshape><tensorflow>","CreationDate":"2015-11-14T05:22:20.347","AcceptedAnswerId":null,"OwnerUserId":2938232.0,"Title":"Reshaping a Tensor in TensorFlow with scalar tensors","Body":"<p>I have a tensor,  of shape .<\/p>\n\n<p>If I knew the shape beforehand, it is easy to reshape,  where  are ints, not tensors.  But is there a way to do this if I don't know the shapes.  It seems that getting the shapes like  and reshaping doesn't work.  That is,<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas?  In my application,  and  are known before runtime but  is only known at runtime. <\/p>\n","answers":[{"AnswerId":"33705884","CreationDate":"2015-11-14T06:43:31.990","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Take a look at this:<\/p>\n\n<pre><code>import tensorflow as tf\n\na, b, c = 2, 3, 4\nx = tf.Variable(tf.random_normal([a, b, c], mean=0.0, stddev=1.0, dtype=tf.float32))\ns = tf.shape(x)\n\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\nv1, v2, v3 = sess.run(s)\ny = tf.reshape(x, [v1 * v2, v3])\nshape = tf.shape(y)\n\nprint sess.run(y)\nprint sess.run(shape)\n<\/code><\/pre>\n\n<p>I am getting the shape of the variable after it's initialization and then use it later. Also take a look at <a href=\"https:\/\/stackoverflow.com\/a\/33642096\/1090562\">this answer<\/a>, as it deals with a similar thing.<\/p>\n"},{"AnswerId":"33717309","CreationDate":"2015-11-15T07:09:23.303","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Now that you told that you use placeholders to populate data, it started to make sense. Here is an example of how can you reshape your data in this case:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\ndata = np.random.rand(2, 3, 4)\n\nx = tf.placeholder(\"float\", None)\ns = tf.shape(x)\n\nsess = tf.Session()\nshape_original = sess.run(s, feed_dict={x: data})\n\nx_ = tf.reshape(x, [shape_original[0] * shape_original[1], shape_original[2]])\ns_ = tf.shape(x_)\n\nshape_now = sess.run(s_, feed_dict={x: data})\nprint 'Original\\t', shape_original\nprint 'Now\\t\\t\\t', shape_now\n\nsess.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":33706301,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-14T07:47:17.927","AcceptedAnswerId":null,"OwnerUserId":5558025.0,"Title":"Running seq2seq model error","Body":"<p>I am trying to run the code in this <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md\" rel=\"nofollow\">tutorial<\/a>. <\/p>\n\n<p>When I try to run this command:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas how to resolve?<\/p>\n","answers":[{"AnswerId":"33744123","CreationDate":"2015-11-16T20:36:06.217","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>I ran it by going to the directory and running:<\/p>\n\n<pre><code>python translate.py\n<\/code><\/pre>\n"},{"AnswerId":"33714209","CreationDate":"2015-11-14T22:36:08.913","ParentId":null,"OwnerUserId":"4383691","Title":null,"Body":"<p>It seems there are a lot of mistakes in the Tensorflow tutorial..\nI was able to run it by removing the .py, and adding an extra -- before the options like:<\/p>\n\n<p>bazel run -c opt tensorflow\/models\/rnn\/translate\/translate -- --data_dir \/home\/minsoo\/tensorflowrnn\/data<\/p>\n\n<p>the directory part should be changed according to your system.<\/p>\n"}]}
{"QuestionId":33707263,"AnswerCount":2,"Tags":"<python><theano><lasagne>","CreationDate":"2015-11-14T10:09:40.853","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"Lasagne - error","Body":"<p>i am trying to learn neural nets \/ lasagne by using the code in Robert Layton's Learning Data Mining with Python. I think i am following the code to the letter, but i get the following error message. any hints or intuition what i am doing wrong much appreciated;<\/p>\n\n<pre><\/pre>\n\n<p>code below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33707484","CreationDate":"2015-11-14T10:36:27.523","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you're not computing the loss with respect to the real input variable. <code>net_input=T.matrix('net_input')<\/code> is your own symbolic input to the network but Lasagne had already created one for you when you created the <code>InputLayer<\/code>. You also don't need to get the output with respect to a particular input, just get the output with respect to the input layer.<\/p>\n\n<p>So, replace the two lines<\/p>\n\n<pre><code>net_input=T.matrix('net_input')\nnet_output=output_layer.get_output_for(net_input)\n<\/code><\/pre>\n\n<p>with the single line<\/p>\n\n<pre><code>net_output=lasagne.layers.get_output(output_layer)\n<\/code><\/pre>\n\n<p>Anticipating the next issue you'll have, you can get the input variable Lasagne created for you via <code>input_layer.input_var<\/code> so you can compile your training function like this:<\/p>\n\n<pre><code>import theano\nf = theano.function([input_layer.input_var, true_output], outputs=loss, updates=updates)\n<\/code><\/pre>\n"},{"AnswerId":"42688313","CreationDate":"2017-03-09T06:32:44.040","ParentId":null,"OwnerUserId":"7387482","Title":null,"Body":"<p><code>input_layer=lasagne.layers.InputLayer(shape=(10,X.shape[1]),input_var=input)<\/code><\/p>\n\n<p>while the input is the tensor you defined before<\/p>\n"}]}
{"QuestionId":33709598,"AnswerCount":2,"Tags":"<scipy><tensorflow>","CreationDate":"2015-11-14T14:43:30.030","AcceptedAnswerId":null,"OwnerUserId":1936768.0,"Title":"TensorFlow cholesky decomposition","Body":"<p>From reading the TensorFlow documentation I see that there is a method for computing the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/math_ops.md#cholesky\" rel=\"noreferrer\">Cholesky decomposition of a square matrix<\/a>. However, usually when I want to use Cholesky decomposition, I do it for the purposes of solving a linear system where direct matrix inversion might be unstable. <\/p>\n\n<p>Therefore, I am looking for a method similar to the one implemented in <a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.16.0\/reference\/generated\/scipy.linalg.cho_solve.html\" rel=\"noreferrer\">Scipy<\/a>. Does anyone know if this exists in TensorFlow or if there is a way it could be incorporated?<\/p>\n","answers":[{"AnswerId":"34111976","CreationDate":"2015-12-05T23:04:24.410","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p><strong>Update (2017\/04\/23)<\/strong><\/p>\n\n<p>TensorFlow now has many <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/math_ops#Matrix_Math_Functions\" rel=\"nofollow noreferrer\">linear algebra operations<\/a>. For instance, checkout <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cholesky_solve\" rel=\"nofollow noreferrer\">tf.cholesky_solve<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matrix_solve_ls\" rel=\"nofollow noreferrer\">tf.matrix_solve_ls<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matrix_solve\" rel=\"nofollow noreferrer\">tf.matrix_solve<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/qr\" rel=\"nofollow noreferrer\">tf.qr<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/svd\" rel=\"nofollow noreferrer\">tf.svd<\/a>, etc. Of course, the original answer below may be helpful as well.<\/p>\n\n<p><strong>Original<\/strong>\nDoes <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matrix_inverse\" rel=\"nofollow noreferrer\">matrix_inverse<\/a> do what you need? It uses Cholesky or LU Decomposition, depending on the input. For example,<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; x = [[1.,1.],[-2.,3.],[1.,-1.]]\n&gt;&gt;&gt; y = [[-1.],[-8.],[3.]]\n&gt;&gt;&gt; a = tf.matrix_inverse(tf.matmul(x, x, transpose_a=True))\n&gt;&gt;&gt; b = tf.matmul(tf.matmul(a, x, transpose_b=True), y)\n&gt;&gt;&gt; with tf.Session():\n...   print b.eval()\n... \n[[ 1.]\n [-2.]]\n<\/code><\/pre>\n"},{"AnswerId":"33711820","CreationDate":"2015-11-14T18:29:11.287","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>user19..8:  The way to do this for now if you want to keep things \"mostly\" in tensorflow would be to do what you and Berci were discussing in the comments:  Run the tensorflow graph until the point where you need to solve the linear system, and then feed the results back in with a feed_dict.  In pseudocode:<\/p>\n\n<pre><code>saved_tensor1 = tf.Variable(...)\nsaved_tensor2 = tf.Variable(...)\n\nstart_of_model...\ntensor1, tensor2 = various stuff...\ndo_save_tensor1 = saved_tensor1.assign(tensor1)\ndo_save_tensor2 = saved_tensor2.assign(tensor2)\nyour_cholesky = tf.cholesky(your_other_tensor, ...)\n\n## THIS IS THE SPLIT POINT\n# Second half of your model starts here\nsolved_system = tf.placeholder(...)  # You'll feed this in with feed_dict\nfinal_answer = do_something_with(saved_tensor1, saved_tensor2, solved_system)\n<\/code><\/pre>\n\n<p>Then to run the whole thing, do:<\/p>\n\n<pre><code>_, _, cho = tf.run([do_save_tensor1, do_save_tensor2, your_cholesky])\nsolution = ... solve your linear system with scipy ...\nfeed_dict = {solved_system: solution}\nanswer = tf.run(final_answer, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>The key here is stashing your intermediate results in tf.Variables so that you can resume the computation afterwards.<\/p>\n\n<p>(I'm not promising that what you get out of tf.cholesky is in the right format to feed directly to scipy, or that you shouldn't just pull out the matrix in an earlier step and feed it to scipy, but this overall workflow should work for you).<\/p>\n\n<p>Note that this will create a performance bottleneck if you're doing heavily multicore or GPU operations and then have to serialize on spitting the matrix out to scipy, but it might also be just fine - depends a lot on your setting.<\/p>\n"}]}
{"QuestionId":33710593,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-11-14T16:26:50.053","AcceptedAnswerId":null,"OwnerUserId":5531398.0,"Title":"Theano MLP with 2 hidden layers throws Shape Mismatch error","Body":"<p>I'm approaching to neural networks implementations, trying to build a working MLP using Theano. Following the tutorial, I tried to enhance the net by adding a layer, for a total of two hidden layers each with the same amount of units (250). The problem is that when I run the script I meet \"Shape mismatch\" ValueError. My code is a modified version of the tutorial code that can be found here <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/mlp.html<\/a>.<\/p>\n\n<p>The part I modified is the snippet-2, namely the MLP object, as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I also removed some comments for readability. The output error I get is:<\/p>\n\n<blockquote>\n  <p>ValueError: Shape mismatch: x has 250 cols (and 20 rows) but y has 784\n  rows (and 250 cols) Apply node that caused the error:\n  Dot22(Elemwise{Composite{tanh((i0 + i1))}}[(0, 0)].0, W) Inputs types:\n  [TensorType(float64, matrix), TensorType(float64, matrix)] Inputs\n  shapes: [(20, 250), (784, 250)] Inputs strides: [(2000, 8), (2000, 8)]\n  Inputs values: ['not shown', 'not shown']<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"33710987","CreationDate":"2015-11-14T17:03:45.083","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The size of the input to layer 2 needs to be the same size as the output from layer 1.<\/p>\n\n<p><code>hiddenLayer2<\/code> takes <code>hiddenLayer1<\/code> as input and <code>hiddenLayer1.n_out == n_hidden<\/code> but 'hiddenLayer2.n_in == n_in'. In this case <code>n_hidden == 250<\/code> and <code>n_in == 784<\/code>. They should match but don't hence the error.<\/p>\n\n<p>The solution is to make <code>hiddenLayer2.n_in == hiddenLayer1.n_out<\/code>.<\/p>\n"}]}
{"QuestionId":33711427,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-14T17:48:40.237","AcceptedAnswerId":33711582.0,"OwnerUserId":1936768.0,"Title":"TensorFlow initializing Tensor of ones","Body":"<p>So suppose I have a tensor <\/p>\n\n<pre><\/pre>\n\n<p>So that I know the number of columns but not the number of rows. I need to initialize a vector of ones of dimension <\/p>\n\n<p>Now the following block of code does not work,<\/p>\n\n<pre><\/pre>\n\n<p>Nor does,<\/p>\n\n<pre><\/pre>\n\n<p>Now, I have found that one way to get around this is to actually make my vector of ones a placeholder,<\/p>\n\n<pre><\/pre>\n\n<p>And to pass in a numpy array of ones of appropriate size in my . But this solution strikes me as inelegant and not the intended use of a placeholder. I could be wrong here, but surely there's a better way.<\/p>\n","answers":[{"AnswerId":"49270430","CreationDate":"2018-03-14T05:41:21.180","ParentId":null,"OwnerUserId":"4268078","Title":null,"Body":"<p>Another not so elegant way is to <\/p>\n\n<pre><code>0 * tf.identity(X) + 1\n<\/code><\/pre>\n"},{"AnswerId":"33711582","CreationDate":"2015-11-14T18:03:19.130","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The way to solve your problem is to use tf.pack operation:<\/p>\n\n<p><code>o = tf.ones(shape=tf.pack([tf.shape(X)[0], 1]))<\/code><\/p>\n\n<p>The reason you had errors is that TensorFlow shape is expected to be a list of integers or a tensor <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/constant_op.md#ones\" rel=\"noreferrer\">link<\/a>. tf.pack makes it easy to convert a list of integers and\/or TensorFlow scalars into a Tensor object.<\/p>\n"}]}
{"QuestionId":33711519,"AnswerCount":2,"Tags":"<neural-network><tensorflow><mnist>","CreationDate":"2015-11-14T17:57:52.513","AcceptedAnswerId":null,"OwnerUserId":1760543.0,"Title":"Tensorflow - Does data adjancency matter ? - MNIST example","Body":"<p>I looked at the <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"nofollow\">MNIST example<\/a> and noticed that when the array of the image is flattened into a 728 array, would it matter if that array was randomized? I mean does the NN take into account the adjacency of the data, or is there one input node put input number (therefore 728 nodes). <\/p>\n\n<p>What I am asking is, will I get the same network if I train with the images flattened as in the example, as I would if I randomised the 728 data array ?<\/p>\n","answers":[{"AnswerId":"33714420","CreationDate":"2015-11-14T23:04:34.720","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You are looking at the very basic tutorial, whose whole point is to familiarize you with TF and a couple of important ML concepts which would be used in many harder models. They are not trying to do anything hard (in fact the accuracy is not higher than what you would achieve by running SVM out of the box). If you read the tutorial carefully they say:<\/p>\n\n<blockquote>\n  <p>It doesn't matter how we flatten the array, as long as we're\n  consistent between images.<\/p>\n<\/blockquote>\n\n<p>and<\/p>\n\n<blockquote>\n  <p>Flattening the data throws away information about the 2D structure of\n  the image. Isn't that bad? Well, the best computer vision methods do\n  exploit this structure, and we will in later tutorials. But the simple\n  method we will be using here, a softmax regression, won't.<\/p>\n<\/blockquote>\n\n<p>Your question about the network: the topology of the network would be the same, the weights\/biases would be different.<\/p>\n\n<p>Also an example with <a href=\"http:\/\/tensorflow.org\/tutorials\/deep_cnn\/index.md\" rel=\"nofollow\">convolutional neural network<\/a> takes into the account the adjacency of the data<\/p>\n"},{"AnswerId":"33711909","CreationDate":"2015-11-14T18:38:13.550","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Depends on which mnist example you're looking at.  convolutional.py runs a 5x5 spatial convolutional window across the image, which does take into account spatial correlation.<\/p>\n\n<p>The MNIST for beginners example that uses a simple weight matrix:<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([784,10]))\nb = tf.Variable(tf.zeros([10]))\n<\/code><\/pre>\n\n<p>does not.  You could permute the order of entries in the points and not change anything, <em>as long as you permute all inputs the same way<\/em>.<\/p>\n\n<p>(There's a reason that convolutional approaches are winning for most image recognition applications -- spatial locality <em>is<\/em> useful. :)<\/p>\n"}]}
{"QuestionId":33711556,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-14T18:00:57.327","AcceptedAnswerId":33713962.0,"OwnerUserId":247866.0,"Title":"Making predictions with a TensorFlow model","Body":"<p>I followed the given mnist tutorials and was able to train a model and evaluate its accuracy. However, the tutorials don't show how to make predictions given a model. I'm not interested in accuracy, I just want to use the model to predict a new example and in the output see all the results (labels), each with its assigned score (sorted or not).<\/p>\n","answers":[{"AnswerId":"33717109","CreationDate":"2015-11-15T06:34:14.423","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>As @dga suggested, you need to run your new instance of the data though your already predicted model.<\/p>\n\n<p>Here is an example:<\/p>\n\n<p>Assume you went though the first tutorial and calculated the accuracy of your model (the model is this: <code>y = tf.nn.softmax(tf.matmul(x, W) + b)<\/code>). Now you grab your model and apply the new data point to it. In the following code I calculate the vector, getting the position of the maximum value. Show the image and print that maximum position.<\/p>\n\n<pre><code>from matplotlib import pyplot as plt\nfrom random import randint\nnum = randint(0, mnist.test.images.shape[0])\nimg = mnist.test.images[num]\n\nclassification = sess.run(tf.argmax(y, 1), feed_dict={x: [img]})\nplt.imshow(img.reshape(28, 28), cmap=plt.cm.binary)\nplt.show()\nprint 'NN predicted', classification[0]\n<\/code><\/pre>\n"},{"AnswerId":"33713962","CreationDate":"2015-11-14T22:06:53.443","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>In the \"<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"noreferrer\">Deep MNIST for Experts<\/a>\" example, see this line:<\/p>\n\n<blockquote>\n  <p>We can now implement our regression model. It only takes one line! We\n  multiply the vectorized input images x by the weight matrix W, add the\n  bias b, and compute the softmax probabilities that are assigned to\n  each class.<\/p>\n\n<pre><code>y = tf.nn.softmax(tf.matmul(x,W) + b)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Just pull on node y and you'll have what you want.<\/p>\n\n<pre><code>feed_dict = {x: [your_image]}\nclassification = tf.run(y, feed_dict)\nprint classification\n<\/code><\/pre>\n\n<p>This applies to just about any model you create - you'll have computed the prediction probabilities as one of the last steps before computing the loss.<\/p>\n"},{"AnswerId":"51002167","CreationDate":"2018-06-23T14:33:30.077","ParentId":null,"OwnerUserId":"1637798","Title":null,"Body":"<p>The question is specifically about the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/layers\" rel=\"nofollow noreferrer\">Google MNIST tutorial<\/a>, which defines a predictor but doesn't apply it.   Using guidance from <a href=\"https:\/\/jhui.github.io\/2017\/03\/14\/TensorFlow-Estimator\/\" rel=\"nofollow noreferrer\">Jonathan Hui's TensorFlow Estimator blog post<\/a>, here is code which exactly fits the Google tutorial and does predictions:<\/p>\n\n<pre><code>from matplotlib import pyplot as plt\n\nimages = mnist.test.images[0:10]\n\npredict_input_fn = tf.estimator.inputs.numpy_input_fn(\n      x={\"x\":images},\n      num_epochs=1,\n      shuffle=False)\n\nmnist_classifier.predict(input_fn=predict_input_fn)\n\nfor image,p in zip(images,mnist_classifier.predict(input_fn=predict_input_fn)):\n    print(np.argmax(p['probabilities']))\n    plt.imshow(image.reshape(28, 28), cmap=plt.cm.binary)\n    plt.show()\n<\/code><\/pre>\n"}]}
{"QuestionId":33712178,"AnswerCount":12,"Tags":"<nan><tensorflow>","CreationDate":"2015-11-14T19:01:58.087","AcceptedAnswerId":33713196.0,"OwnerUserId":1111929.0,"Title":"Tensorflow NaN bug?","Body":"<p>I'm using TensorFlow and I modified the <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/pros\/index.md\">tutorial<\/a> example to take my RGB images.<\/p>\n\n<p>The algorithm works flawlessly out of the box on the new image set, until suddenly (still converging, it's around 92% accuracy usually), it crashes with the error that ReluGrad received non-finite values. Debugging shows that nothing unusual happens with the numbers until very suddenly, for unknown reason, the error is thrown. Adding<\/p>\n\n<pre><\/pre>\n\n<p>as debug code to each loop, yields the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Since none of my values is very high, the only way a NaN can happen is by a badly handled 0\/0, but since this tutorial code doesn't do any divisions or similar operations, I see no other explanation than that this comes from the internal TF code. <\/p>\n\n<p>I'm clueless on what to do with this. Any suggestions? The algorithm is converging nicely, its accuracy on my validation set was steadily climbing and just reached 92.5% at iteration 8600.<\/p>\n","answers":[{"AnswerId":"38673185","CreationDate":"2016-07-30T11:04:25.643","ParentId":null,"OwnerUserId":"6657378","Title":null,"Body":"<p>Actually, clipping is not a good idea as it will stop the gradient from propagating backwards when the threshold is reached. Instead we can add a little bit of constant to the softmax output.<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv + 1e-10))\n<\/code><\/pre>\n"},{"AnswerId":"43997715","CreationDate":"2017-05-16T09:37:11.230","ParentId":null,"OwnerUserId":"852592","Title":null,"Body":"<p>Here is the implementation of the binary (sigmoid) and categorical (softmax) cross-entropy losses in TensorFlow 1.1:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/nn_impl.py#L159\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/nn_impl.py#L159<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/nn_ops.py#L1609\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/nn_ops.py#L1609<\/a><\/li>\n<\/ul>\n\n<p>As one can see in the binary case they consider some special cases to achieve numerical stability:<\/p>\n\n<pre><code># The logistic loss formula from above is\n#   x - x * z + log(1 + exp(-x))\n# For x &lt; 0, a more numerically stable formula is\n#   -x * z + log(1 + exp(x))\n# Note that these two expressions can be combined into the following:\n#   max(x, 0) - x * z + log(1 + exp(-abs(x)))\n# To allow computing gradients at zero, we define custom versions of max and\n# abs functions.\nzeros = array_ops.zeros_like(logits, dtype=logits.dtype)\ncond = (logits &gt;= zeros)\nrelu_logits = array_ops.where(cond, logits, zeros)\nneg_abs_logits = array_ops.where(cond, -logits, logits)\nreturn math_ops.add(relu_logits - logits * labels,\n                    math_ops.log1p(math_ops.exp(neg_abs_logits)),\n                    name=name)\n<\/code><\/pre>\n"},{"AnswerId":"51138819","CreationDate":"2018-07-02T14:49:04.913","ParentId":null,"OwnerUserId":"3457417","Title":null,"Body":"<p>Besides all the great answers above, I will add mine. It's a scenario less common to run into, but does cause NaN: <strong>divide by zero<\/strong>.<\/p>\n\n<p>In my network for a NLP task, there is a layer that does <em>average pooling<\/em>. Namely, each data is a sequence of tokens. My layer does some token embedding and then calculates the average of the embedded vector. <\/p>\n\n<p>The average calculation is coded as <\/p>\n\n<pre><code>tf.reduce_sum(embedded)\/tf.reduce_sum(tf.not_equal(input, pad)) \n<\/code><\/pre>\n\n<p>Here <code>pad<\/code> is some dummy token I use in batch processing. <\/p>\n\n<p>Now if some data contains empty token list (for whatever reason), its length (the denominator in the code snippet above) would be 0. Then it causes a <strong>divide by zero<\/strong> issue and the NaN will remain in all the following layers\/ optimization steps.<\/p>\n\n<p>In case anyone ran into this issue, I used <code>tf.where<\/code> to smooth those length: <\/p>\n\n<pre><code>sum_embedding = tf.reduce_sum(embedded, 1)\nembedding_length = tf.reduce_sum(tf.cast(tf.not_equal(input, pad), dtype=tf.float32), axis=1, keep_dims=True)\nembedding_length_smoothed = tf.where(tf.greater(embedding_length, 0.0), embedding_length, tf.ones(tf.shape(embedding_length)))\navg_embedding = sum_embedding \/ embedding_length_smoothed\n<\/code><\/pre>\n\n<p>Essentially this treats all those data with 0-length token list to be of length 1, and avoids the NaN issue.<\/p>\n"},{"AnswerId":"33713196","CreationDate":"2015-11-14T20:49:12.377","ParentId":null,"OwnerUserId":"1111929","Title":null,"Body":"<p>Actually, it turned out to be something stupid. I'm posting this in case anyone else would run into a similar error.<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))\n<\/code><\/pre>\n\n<p>is actually a horrible way of computing the cross-entropy. In some samples, certain classes could be excluded with certainty after a while, resulting in y_conv=0 for that sample. That's normally not a problem since you're not interested in those, but in the way cross_entropy is written there, it yields 0*log(0) for that particular sample\/class. Hence the NaN.<\/p>\n\n<p>Replacing it with <\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y_conv,1e-10,1.0)))\n<\/code><\/pre>\n\n<p>solved all my problems.<\/p>\n"},{"AnswerId":"53018297","CreationDate":"2018-10-27T02:44:50.537","ParentId":null,"OwnerUserId":"6763456","Title":null,"Body":"<p>Sometimes you use <code>tf.sqrt()<\/code> function without adding a small constant <code>1e-10<\/code> in it, inducing this <code>nan<\/code> problem.<\/p>\n"},{"AnswerId":"47682117","CreationDate":"2017-12-06T19:33:17.393","ParentId":null,"OwnerUserId":"5841473","Title":null,"Body":"<p>I used LSTM for long sequences and got nan gradients. None of these answers helped me. But I came up with three own solutions. I hope they will be useful for some other people who came here from google search.<\/p>\n\n<ol>\n<li><p>Gradient clipping didn't help me because gradients turned nan in one batch update. In this case, you can replace nans with zeros with such lines:<\/p>\n\n<pre><code>opt = tf.train.AdamOptimizer(args.lr)\ngrads = opt.compute_gradients(loss)\ngrads2 = [(tf.where(tf.is_nan(grad), tf.zeros(grad.shape), grad), var) for grad, var in grads]\nopt_op = opt.apply_gradients(grads2)\n<\/code><\/pre>\n\n<p>If you want to track if nans appeared you can use this code:<\/p>\n\n<pre><code>was_nan = tf.reduce_any(tf.convert_to_tensor([tf.reduce_any(tf.is_nan(g)) for g in grads]))\n<\/code><\/pre><\/li>\n<li><p>Replace LSTMCell with LayerNormBasicLSTMCell - an LSTM cell with layer norm - something similar to batch norm between timesteps.<\/p><\/li>\n<li><p>If you use regular recurrent state dropout you can replace it with \"Recurrent Dropout without Memory Loss\". Code:<\/p>\n\n<pre><code>LayerNormBasicLSTMCell(neurons, dropout_keep_prob=0.8)\n<\/code><\/pre>\n\n<p>Note that you can also turn on the dropout feature alone without layer normalization:<\/p>\n\n<pre><code>LayerNormBasicLSTMCell(neurons, layer_norm=False, dropout_keep_prob=0.8)\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"42497444","CreationDate":"2017-02-27T23:08:35.483","ParentId":null,"OwnerUserId":"433287","Title":null,"Body":"<h1>A bias free alternative.<\/h1>\n\n<p>Many of the other solutions use clipping to avoid an undefined gradient. Depending on your problem, clipping introduces bias and may not be acceptable in all cases. As the following code demonstrates, we need only handle the point of discontinuity--not the region near it.<\/p>\n\n<h1>Specific Answer<\/h1>\n\n<pre><code>def cross_entropy(x, y, axis=-1):\n  safe_y = tf.where(tf.equal(x, 0.), tf.ones_like(y), y)\n  return -tf.reduce_sum(x * tf.log(safe_y), axis)\n\ndef entropy(x, axis=-1):\n  return cross_entropy(x, x, axis)\n<\/code><\/pre>\n\n<p><em>But did it work?<\/em><\/p>\n\n<pre><code>x = tf.constant([0.1, 0.2, 0., 0.7])\ne = entropy(x)\n# ==&gt; 0.80181855\ng = tf.gradients(e, x)[0]\n# ==&gt; array([1.30258512,  0.60943794, 0., -0.64332503], dtype=float32)  Yay! No NaN.\n<\/code><\/pre>\n\n<p>(Note: deleted <a href=\"https:\/\/stackoverflow.com\/questions\/33699174\/tensorflows-relugrad-claims-input-is-not-finite\">dup cross-post<\/a>.) <\/p>\n\n<h1>General Recipe<\/h1>\n\n<p>Use an inner <code>tf.where<\/code> to ensure the function has no asymptote.\nThat is, alter the input to the inf generating function such that no inf can be created.\nThen use a second <code>tf.where<\/code> to always select the valid code-path.\nThat is, implement the mathematical condition as you would \"normally\", i.e., the \"naive\" implementation.<\/p>\n\n<p>In Python code, the recipe is:<\/p>\n\n<p>Instead of this:<\/p>\n\n<pre><code>tf.where(x_ok, f(x), safe_f(x))\n<\/code><\/pre>\n\n<p>Do this:<\/p>\n\n<pre><code>safe_x = tf.where(x_ok, x, safe_x)\ntf.where(x_ok, f(safe_x), safe_f(x))\n<\/code><\/pre>\n\n<h2>Example<\/h2>\n\n<p>Suppose you wish to compute:<\/p>\n\n<pre><code>f(x) = { 1\/x, x!=0\n       { 0,   x=0\n<\/code><\/pre>\n\n<p>A naive implementation results in NaNs in the gradient, i.e.,<\/p>\n\n<pre><code>def f(x):\n  x_ok = tf.not_equal(x, 0.)\n  f = lambda x: 1. \/ x\n  safe_f = tf.zeros_like\n  return tf.where(x_ok, f(x), safe_f(x))\n<\/code><\/pre>\n\n<p><em>Does it work?<\/em><\/p>\n\n<pre><code>x = tf.constant([-1., 0, 1])\ntf.gradients(f(x), x)[0].eval()\n# ==&gt; array([ -1.,  nan,  -1.], dtype=float32)\n#  ...bah! We have a NaN at the asymptote despite not having\n# an asymptote in the non-differentiated result.\n<\/code><\/pre>\n\n<p>The basic pattern for avoiding NaN gradients when using <code>tf.where<\/code> is to call <code>tf.where<\/code> twice.  The innermost <code>tf.where<\/code> ensures that the result <code>f(x)<\/code> is always finite. The outermost <code>tf.where<\/code> ensures the correct result is chosen.  For the running example, the trick plays out like this:<\/p>\n\n<pre><code>def safe_f(x):\n  x_ok = tf.not_equal(x, 0.)\n  f = lambda x: 1. \/ x\n  safe_f = tf.zeros_like\n  safe_x = tf.where(x_ok, x, tf.ones_like(x))\n  return tf.where(x_ok, f(safe_x), safe_f(x))\n<\/code><\/pre>\n\n<p><em>But did it work?<\/em><\/p>\n\n<pre><code>x = tf.constant([-1., 0, 1])\ntf.gradients(safe_f(x), x)[0].eval()\n# ==&gt; array([-1.,  0., -1.], dtype=float32)\n# ...yay! double-where trick worked. Notice that the gradient\n# is now a constant at the asymptote (as opposed to being NaN).\n<\/code><\/pre>\n"},{"AnswerId":"38489657","CreationDate":"2016-07-20T19:52:28.363","ParentId":null,"OwnerUserId":"6529650","Title":null,"Body":"<p>If <code>y_conv<\/code> is the result of a softmax, say, <code>y_conv = tf.nn.softmax(x)<\/code>, then an even better solution is to replace it with <code>log_softmax<\/code>:<\/p>\n\n<pre><code>y = tf.nn.log_softmax(x)\ncross_entropy = -tf.reduce_sum(y_*y)\n<\/code><\/pre>\n"},{"AnswerId":"43692142","CreationDate":"2017-04-29T05:32:47.787","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You are trying to calculate <a href=\"https:\/\/en.wikipedia.org\/wiki\/Cross_entropy\" rel=\"nofollow noreferrer\">cross-entropy<\/a> using the standard formula. Not only the value is undefinined when <code>x=0<\/code>, it is also numerically unstable.<\/p>\n\n<p>It is better to 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> or if you really want to use hand-crafted formula, to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/clip_by_value\" rel=\"nofollow noreferrer\">tf.clip_by_value<\/a> zeros to very small number in the log. <\/p>\n"},{"AnswerId":"48994990","CreationDate":"2018-02-26T18:39:45.757","ParentId":null,"OwnerUserId":"9415029","Title":null,"Body":"<p>I was getting nans sometimes and not other times while working on a standard feed-forward network. I have previously used similar TensorFlow code and it worked fine.<\/p>\n\n<p>It turns out that I imported the variable names by accident. So, as soon as the first row (the variable names) was selected in a batch, the nan losses started. Maybe keep an eye out for that?<\/p>\n"},{"AnswerId":"56999360","CreationDate":"2019-07-12T02:06:57.967","ParentId":null,"OwnerUserId":"774907","Title":null,"Body":"<p>I've been using Tensorflow Estimator, which I believe account for those division by zero and other numerical stability issues, and occasionally get this error (<code>ERROR:tensorflow:Model diverged with loss = NaN during training<\/code>). Most of the time when I get this is because my inputs include <code>nan<\/code>s. So: be sure that your input dataframes (or whatever you use) don't have NaN values hidden somewhere in them.<\/p>\n"},{"AnswerId":"56848302","CreationDate":"2019-07-02T08:27:43.600","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>I will add here one of my previous problems with NaNs. I was using the sigmoid function as the activation of the last layer of my network. However, the sigmoid activation function uses the exponential function to be computed and I got some really big numbers entering the sigmoid.<\/p>\n\n<p>It resulted in infinite gradients and some NaNs started to appear.<\/p>\n"}]}
{"QuestionId":33714814,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-14T23:55:51.260","AcceptedAnswerId":null,"OwnerUserId":5558025.0,"Title":"Using Tensorboard to graph from a log directory","Body":"<p>Can anyone give me a quick and dirty tutorial about how to modify the code in the <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/beginners\/index.md\/\" rel=\"nofollow\">this MNIST tutorial<\/a> and <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md\" rel=\"nofollow\">this seq2seq tutorial<\/a> to log things to a log directory which can then be used in TensorBoard? I didn't really understand the ones on the official site. <\/p>\n","answers":[{"AnswerId":"33715745","CreationDate":"2015-11-15T02:28:51.187","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>You can create a <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.md#SummaryWriter\" rel=\"nofollow\">SummaryWriter<\/a> object, passing it the log directory, and call <code>add_summary<\/code> to log summaries and events to files in that directory. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/embedding\/word2vec.py#L383\" rel=\"nofollow\">word2vec.py<\/a> has an example. You can simply point TensorBoard at the log directory by passing it through <code>--logdir<\/code> and visualize the summaries.<\/p>\n"}]}
{"QuestionId":33714832,"AnswerCount":3,"Tags":"<python><tensorflow><determinants>","CreationDate":"2015-11-14T23:58:22.047","AcceptedAnswerId":33716423.0,"OwnerUserId":1936768.0,"Title":"matrix determinant differentiation in tensorflow","Body":"<p>I am interested in computing the derivative of a matrix determinant using TensorFlow. I can see from experimentation that TensorFlow has not implemented a method of differentiating through a determinant:<\/p>\n\n<pre><\/pre>\n\n<p>A little further investigation revealed that it is actually possible to compute the derivative; see for example <a href=\"https:\/\/en.wikipedia.org\/wiki\/Jacobi%27s_formula\" rel=\"noreferrer\">Jacobi's formula<\/a>. I determined that in order to implement this means of differentiating through a determinant that I need to use the function decorator,<\/p>\n\n<pre><\/pre>\n\n<p>However, I am not familiar enough with tensor flow to understand how this can be accomplished. Does anyone have any insight on this matter?<\/p>\n\n<p>Here's an example where I run into this issue:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33716423","CreationDate":"2015-11-15T04:30:13.827","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Please check \"Implement Gradient in Python\" section <a href=\"http:\/\/tensorflow.org\/how_tos\/adding_an_op\/index.html#implement-the-gradient-in-python\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>In particular, you can implement it as follows<\/p>\n\n<pre><code>@ops.RegisterGradient(\"MatrixDeterminant\")\ndef _MatrixDeterminantGrad(op, grad):\n  \"\"\"Gradient for MatrixDeterminant. Use formula from 2.2.4 from\n  An extended collection of matrix derivative results for forward and reverse\n  mode algorithmic differentiation by Mike Giles\n  -- http:\/\/eprints.maths.ox.ac.uk\/1079\/1\/NA-08-01.pdf\n\"\"\"\n  A = op.inputs[0]\n  C = op.outputs[0]\n  Ainv = tf.matrix_inverse(A)\n  return grad*C*tf.transpose(Ainv)\n<\/code><\/pre>\n\n<p>Then a simple training loop to check that it works:<\/p>\n\n<pre><code>a0 = np.array([[1,2],[3,4]]).astype(np.float32)\na = tf.Variable(a0)\nb = tf.square(tf.matrix_determinant(a))\ninit_op = tf.initialize_all_variables()\nsess = tf.InteractiveSession()\ninit_op.run()\n\nminimization_steps = 50\nlearning_rate = 0.001\noptimizer = tf.train.GradientDescentOptimizer(learning_rate)\ntrain_op = optimizer.minimize(b)\n\nlosses = []\nfor i in range(minimization_steps):\n  train_op.run()\n  losses.append(b.eval())\n<\/code><\/pre>\n\n<p>Then you can visualize your loss over time<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\n\nplt.ylabel(\"Determinant Squared\")\nplt.xlabel(\"Iterations\")\nplt.plot(losses)\n<\/code><\/pre>\n\n<p>Should see something like this\n<a href=\"https:\/\/i.stack.imgur.com\/1DA5J.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1DA5J.png\" alt=\"Loss plot\"><\/a><\/p>\n"},{"AnswerId":"33715984","CreationDate":"2015-11-15T03:10:32.200","ParentId":null,"OwnerUserId":"1936768","Title":null,"Body":"<p>For those who are interested, I discovered the solution that works on my problems:<\/p>\n\n<pre><code>@tf.RegisterGradient(\"MatrixDeterminant\")\ndef _MatrixDeterminant(op, grad):\n    \"\"\"Gradient for MatrixDeterminant.\"\"\"\n    return op.outputs[0] * tf.transpose(tf.matrix_inverse(op.inputs[0]))\n<\/code><\/pre>\n"},{"AnswerId":"33715264","CreationDate":"2015-11-15T01:04:55.510","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I think you are confused with what is a <strong>derivative of a matrix determinant<\/strong>.<\/p>\n\n<p>Matrix determinant is a function which is calculated over the elements of the matrix by some formula. So if all the elements of the matrix are numbers, you the determinant will you you just one number and the derivative will be <code>0<\/code>. When some of the elements are variables, you will get an expression of these variables. For example:<\/p>\n\n<pre><code>x, x^2\n1, sin(x)\n<\/code><\/pre>\n\n<p>The determinant will be <code>x*sin(x) - x^2<\/code> and the derivative is <code>2x + sin(x) + x*cos(x)<\/code>. The Jacobi formula just connects the determinant with adjunct matrix.<\/p>\n\n<hr>\n\n<p>In your example your matrix <code>A<\/code> consists of only numbers and therefore the determinant is just a number and the <code>loss<\/code> is just a number as well. <code>GradientDescentOptimizer<\/code> needs to have some free variables to minimize and does not have any because your <code>loss<\/code> is just a number.<\/p>\n"}]}
{"QuestionId":33715728,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-15T02:25:31.173","AcceptedAnswerId":33797408.0,"OwnerUserId":2857219.0,"Title":"Using Queues to uniformly sample from multiple input files","Body":"<p>I have one serialized file for each class in my dataset. I would like to use queues to load up each of these files and then place them in a RandomShuffleQueue that will pull them off so I get a random mix of examples from each class. I thought this code would work.<\/p>\n\n<p>In this example each file has 10 examples.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This works fine for 10 calls, but on the 11th it says that the queue is empty. <\/p>\n\n<p>I believe this is due to a misunderstanding on my part of what these queues operate on. I add 10 variables to the , but each of those variables is itself pulling from a queue, so I assumed the queue would not be emptied until each of the file queues was empty. <\/p>\n\n<p>What am I doing wrong here?<\/p>\n","answers":[{"AnswerId":"33797408","CreationDate":"2015-11-19T07:14:17.753","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The correct answer to this question will depend on how many files you have, how large they are, and how their sizes are distributed.<\/p>\n\n<p>The immediate problem with your example is that <code>rq<\/code> only gets one element for each <code>filename in filenames<\/code>, then the queue is closed. I'm presuming that there are 10 <code>filenames<\/code>, since <code>rq.dequeue()<\/code> will consume one element of <code>rq<\/code> each time you call <code>sess.run([label, image])<\/code>. Since the queue is closed, no more elements can be added, and the 11th activation of the <code>rq.dequeue()<\/code> operation fails.<\/p>\n\n<p>The general solution is that you have to create additional threads to keep running <code>rq.enqueue([out_string])<\/code> in a loop. TensorFlow includes a <code>QueueRunner<\/code> class that is designed to simplify this, and some other functions that handle common cases. The <a href=\"http:\/\/tensorflow.org\/how_tos\/threading_and_queues\/index.html#threading-and-queues\" rel=\"nofollow\">documentation for threading and queues<\/a> explains how they are used, and there is also some good information on <a href=\"http:\/\/tensorflow.org\/how_tos\/reading_data\/index.html#reading-from-files\" rel=\"nofollow\">using queues to read from files<\/a>.<\/p>\n\n<p>As to your particular problem, one way you could handle this would be to create <code>N<\/code> readers (for each of <code>N<\/code> files). You could then <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.html#pack\" rel=\"nofollow\"><code>tf.pack()<\/code><\/a> <code>N<\/code> elements (one from each reader) into a batch, and use <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/io_ops.html#QueueBase.enqueue_many\" rel=\"nofollow\"><code>enqueue_many<\/code><\/a> to add a batch at a time into a <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/io_ops.html#RandomShuffleQueue\" rel=\"nofollow\"><code>tf.RandomShuffleQueue<\/code><\/a> with a sufficiently large capacity and <code>min_after_dequeue<\/code> to ensure that there is sufficient mixing between the classes. Calling <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\" rel=\"nofollow\"><code>dequeue_many(k)<\/code><\/a> on the <code>RandomShuffleQueue<\/code> would give you a batch of <code>k<\/code> elements sampled from each file with equal probability.<\/p>\n"}]}
{"QuestionId":33715746,"AnswerCount":0,"Tags":"<python><ubuntu><cuda><theano>","CreationDate":"2015-11-15T02:29:11.997","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Theano: \"CUDA is installed, but device gpu is not available\"","Body":"<p>I am trying to run some Theano script using , I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I run Kubuntu 15.10 x64 with a Nvidia GTX 690.<\/p>\n\n<p>CUDA seems to be correctly installed:<\/p>\n\n<pre><\/pre>\n\n<p>I used the instructions from the official guide <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html\" rel=\"nofollow\">Easy Installation of an Optimized Theano on Current Ubuntu<\/a>, on a fresh installation of Kubuntu, namely:<\/p>\n\n<pre><\/pre>\n\n<p>Then I created a  file in that contains in my home folder (namely ):<\/p>\n\n<pre><\/pre>\n\n<p>and I rebooted.<\/p>\n\n<p>Did I miss something?<\/p>\n\n<hr>\n\n<p>I tried the following commands but get the same error:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33717772,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-15T08:26:47.360","AcceptedAnswerId":null,"OwnerUserId":4165313.0,"Title":"How can I copy a variable in tensorflow","Body":"<p>In numpy I can create a copy of the variable with <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.copy.html\">numpy.copy<\/a>. Is there a similar method, that I can use to create a copy of a Tensor in TensorFlow?<\/p>\n","answers":[{"AnswerId":"33717784","CreationDate":"2015-11-15T08:28:57.153","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You can do this in a couple of ways.<\/p>\n\n<ul>\n<li>this will create you a copy: <code>v2 = tf.Variable(v1)<\/code><\/li>\n<li>you can also use <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/control_flow_ops.html#identity\" rel=\"noreferrer\">identity op<\/a>: <code>v2 = tf.identity(v1)<\/code> (which I think is a proper way of doing it.<\/li>\n<\/ul>\n\n<p>Here is a code example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nv1 = tf.Variable([[1, 2], [3, 4]])\nv_copy1 = tf.Variable(v1)\nv_copy2 = tf.identity(v1)\n\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\na, b = sess.run([v_copy1, v_copy2])\nsess.close()\n\nprint a\nprint b\n<\/code><\/pre>\n\n<p>Both of them would print the same tensors.<\/p>\n"},{"AnswerId":"33792057","CreationDate":"2015-11-18T22:54:28.357","ParentId":null,"OwnerUserId":"858862","Title":null,"Body":"<p>This performs a deep copy<\/p>\n\n<pre><code>copied_variable = tf.Variable(source_variable.initialized_value())\n<\/code><\/pre>\n\n<p>It also handles intialization properly, i.e. <\/p>\n\n<pre><code>tf.intialize_all_variables()\n<\/code><\/pre>\n\n<p>will properly initialize source_variable first and then copy that value to copied_variable<\/p>\n"},{"AnswerId":"33794922","CreationDate":"2015-11-19T03:48:39.457","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You asked how to copy a variable in the title, but how to copy a tensor in the question.  Let's look at the different possible answers.<\/p>\n\n<p>(1)  You want to create a <em>tensor<\/em> that has the same value that is currently stored in a <em>variable<\/em> that we'll call <code>var<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensor = tf.identity(var)\n<\/code><\/pre>\n\n<p>But remember, 'tensor' is a graph node that will have that value when evaluated, and any time you evaluate it, it will grab the <em>current<\/em> value of <code>var<\/code>.  You can play around with <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/control_flow_ops#Control_Flow_Operations\" rel=\"noreferrer\">control flow ops such as <code>with_dependencies()<\/code><\/a> to see the ordering of updates to the variable and the timing of the identity.<\/p>\n\n<p>(2)  You want to create another variable and set its value to the value currently stored in a variable:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nvar = tf.Variable(0.9)\nvar2 = tf.Variable(0.0)\ncopy_first_variable = var2.assign(var)\ninit = tf.initialize_all_variables()\nsess = tf.Session()\n\nsess.run(init)\n\nprint sess.run(var2)\nsess.run(copy_first_variable)\nprint sess.run(var2)\n<\/code><\/pre>\n\n<p>(3)  You want to define a variable and set its starting value to the same thing you already initialized a variable to (this is what nivwu.. above answered):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>var2 = tf.Variable(var.initialized_value())\n<\/code><\/pre>\n\n<p><code>var2<\/code> will get initialized when you call <code>tf.initialize_all_variables<\/code>.  You can't use this to copy var after you've already initialized the graph and started running things.<\/p>\n"}]}
{"QuestionId":33717906,"AnswerCount":3,"Tags":"<linker><caffe>","CreationDate":"2015-11-15T08:49:14.003","AcceptedAnswerId":null,"OwnerUserId":4570493.0,"Title":"Caffe Installation error","Body":"<p>I am installing Caffe on Ubunutu 15.04. I installed all the prerequisites mentioned in the installation page of Caffe from the official page: <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/installation.html<\/a><\/p>\n\n<p>I am getting error during make. I am posting the error log (could not post whole log, too big a file. I cropped messages from starting till it fits the block). Please inform me if this is not useful<\/p>\n\n<pre><\/pre>\n\n<p>How should I proceed with this ?<\/p>\n\n<p>Edit: I am also posting my Makeconfig.config file (only the uncommented part)<\/p>\n","answers":[{"AnswerId":"37148844","CreationDate":"2016-05-10T20:57:13.103","ParentId":null,"OwnerUserId":"3096839","Title":null,"Body":"<p>Make sure OpenCv is installed and the Makefile is edited for the installed version<\/p>\n"},{"AnswerId":"41252270","CreationDate":"2016-12-20T22:32:17.407","ParentId":null,"OwnerUserId":"7322058","Title":null,"Body":"<p>The installation of caffe can be a problem sometimes, that is why authors have published a Docker image. I would suggest to use it, it works great.<\/p>\n\n<p>Assuming you have docker installed (<a href=\"https:\/\/docs.docker.com\/engine\/installation\/\" rel=\"nofollow noreferrer\">https:\/\/docs.docker.com\/engine\/installation\/<\/a>):<\/p>\n\n<pre><code>cd caffe\/docker\ndocker build -t caffe:cpu standalone\/cpu\n<\/code><\/pre>\n\n<p>Then, you can access via:<\/p>\n\n<pre><code>docker run -ti -v \/path\/to\/folder\/conf_and_images_for_caffe\/:\/datas caffe:cpu bash\n<\/code><\/pre>\n\n<p>(the -v option let you mount a volume on the docker container to work with files from your computer)<\/p>\n"},{"AnswerId":"33884146","CreationDate":"2015-11-24T01:57:38.190","ParentId":null,"OwnerUserId":"5522934","Title":null,"Body":"<p>The error show that your file .\/tools\/finetune_net.cpp may be not exist or be with wrong path.I advise you check that file first.<\/p>\n"}]}
{"QuestionId":33720331,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-15T13:39:45.747","AcceptedAnswerId":33720380.0,"OwnerUserId":904032.0,"Title":"MNIST For ML Beginners - Why is \"one-hot\" vector length 11?","Body":"<p>If you carefully count the digits displayed in the tutorial, also shown below, you'll see that it has length 11.  The indexes 0 through 9 stand for each digit, a total of 10. However, what's the significance of the 11th position shown below? <\/p>\n\n<p>Also, the code is defined with length 10, [60000,10].  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NFFsr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NFFsr.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"33720380","CreationDate":"2015-11-15T13:44:28.973","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This is just a typo, do not worry about it (actually this is not the only typo there). If you will check the shape of the data <code>mnist.train.labels.shape<\/code> you will get <code>(55000, 10)<\/code> which is not the same as they claim <code>(60000, 10)<\/code>.<\/p>\n\n<p>Also this shape of the data shows you that the length of one-hot vector is 10.<\/p>\n"}]}
{"QuestionId":33720645,"AnswerCount":2,"Tags":"<python><matlab><neural-network><tensorflow>","CreationDate":"2015-11-15T14:12:06.850","AcceptedAnswerId":33723404.0,"OwnerUserId":1715157.0,"Title":"Why is this TensorFlow implementation vastly less successful than Matlab's NN?","Body":"<p>As a toy example I'm trying to fit a function  from 100 no-noise data points. The matlab default implementation is phenomenally successful with mean square difference ~10^-10, and interpolates perfectly.<\/p>\n\n<p>I implement a neural network with one hidden layer of 10 sigmoid neurons. I'm a beginner at neural networks so be on your guard against dumb code.<\/p>\n\n<pre><\/pre>\n\n<p>Mean square difference ends at ~2*10^-3, so about 7 orders of magnitude worse than matlab. Visualising with<\/p>\n\n<pre><\/pre>\n\n<p>we can see the fit is systematically imperfect:\n<a href=\"https:\/\/i.stack.imgur.com\/Blxq9.png\"><img src=\"https:\/\/i.stack.imgur.com\/Blxq9.png\" alt=\"enter image description here\"><\/a>\nwhile the matlab one looks perfect to the naked eye with the differences uniformly &lt; 10^-5:\n<a href=\"https:\/\/i.stack.imgur.com\/kC8aJ.jpg\"><img src=\"https:\/\/i.stack.imgur.com\/kC8aJ.jpg\" alt=\"enter image description here\"><\/a>\nI have tried to replicate with TensorFlow the diagram of the Matlab network:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ORLXL.png\"><img src=\"https:\/\/i.stack.imgur.com\/ORLXL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Incidentally, the diagram seems to imply a tanh rather than sigmoid activation function. I cannot find it anywhere in documentation to be sure. However, when I try to use a tanh neuron in TensorFlow the fitting quickly fails with  for variables. I do not know why.<\/p>\n\n<p>Matlab uses Levenberg\u2013Marquardt training algorithm. Bayesian regularization is even more successful with mean squares at 10^-12 (we are probably in the area of vapours of float arithmetic).<\/p>\n\n<p>Why is TensorFlow implementation so much worse, and what can I do to make it better?<\/p>\n","answers":[{"AnswerId":"33728310","CreationDate":"2015-11-16T04:00:59.783","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>btw, here's a slightly cleaned up version of the above that cleans up some of the shape issues and unnecessary bouncing between tf and np.  It achieves 3e-08 after 40k steps, or about 1.5e-5 after 4000:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef weight_variable(shape):\n  initial = tf.truncated_normal(shape, stddev=0.1)\n  return tf.Variable(initial)\n\ndef bias_variable(shape):\n  initial = tf.constant(0.1, shape=shape)\n  return tf.Variable(initial)\n\nxTrain = np.linspace(0.2, 0.8, 101).reshape([1, -1])\nyTrain = (1\/xTrain)\n\nx = tf.placeholder(tf.float32, [1,None])\nhiddenDim = 10\n\nb = bias_variable([hiddenDim,1])\nW = weight_variable([hiddenDim, 1])\n\nb2 = bias_variable([1])\nW2 = weight_variable([1, hiddenDim])\n\nhidden = tf.nn.sigmoid(tf.matmul(W, x) + b)\ny = tf.matmul(W2, hidden) + b2\n\n# Minimize the squared errors.                                                                \nloss = tf.reduce_mean(tf.square(y - yTrain))\nstep = tf.Variable(0, trainable=False)\nrate = tf.train.exponential_decay(0.15, step, 1, 0.9999)\noptimizer = tf.train.AdamOptimizer(rate)\ntrain = optimizer.minimize(loss, global_step=step)\ninit = tf.initialize_all_variables()\n\n# Launch the graph                                                                            \nsess = tf.Session()\nsess.run(init)\n\nfor step in xrange(0, 40001):\n    train.run({x: xTrain}, sess)\n    if step % 500 == 0:\n        print loss.eval({x: xTrain}, sess)\n<\/code><\/pre>\n\n<p>All that said, it's probably not too surprising that LMA is doing better than a more general DNN-style optimizer for fitting a 2D curve.  Adam and the rest are targeting very high dimensionality problems, and <a href=\"http:\/\/www.eng.auburn.edu\/~wilambm\/pap\/2011\/K10149_C012.pdf\" rel=\"noreferrer\">LMA starts to get glacially slow for very large networks<\/a> (see 12-15).<\/p>\n"},{"AnswerId":"33723404","CreationDate":"2015-11-15T18:34:47.387","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I tried training for 50000 iterations it got to 0.00012 error. It takes about 180 seconds on Tesla K40.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cH2hN.png\"><img src=\"https:\/\/i.stack.imgur.com\/cH2hN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It seems that for this kind of problem, first order gradient descent is not a good fit (pun intended), and you need Levenberg\u2013Marquardt or l-BFGS. I don't think anyone implemented them in TensorFlow yet.<\/p>\n\n<p><strong>Edit<\/strong>\nUse <code>tf.train.AdamOptimizer(0.1)<\/code> for this problem. It gets to <code>3.13729e-05<\/code> after 4000 iterations. Also, GPU with default strategy also seems like a bad idea for this problem. There are many small operations and the overhead causes GPU version to run 3x slower than CPU on my machine.<\/p>\n"}]}
{"QuestionId":33721270,"AnswerCount":1,"Tags":"<plot><lua><gnuplot><torch>","CreationDate":"2015-11-15T15:14:35.703","AcceptedAnswerId":null,"OwnerUserId":5102245.0,"Title":"Combining plots produced produced using optim logger","Body":"<p>I have built and MNIST classifier using torch, the results of this are plotted by using optim logger.<\/p>\n\n<p>When plotting I call , where  is . The same is repeated for the test set. So this gives me two separate plots, is there a way to combine these two logs in one plot?<\/p>\n","answers":[{"AnswerId":"40917470","CreationDate":"2016-12-01T18:05:17.123","ParentId":null,"OwnerUserId":"7237677","Title":null,"Body":"<pre><code>gnuplot\nset title \"Mean accuracy\"\nset xlabel \"Epochs\"\nset ylabel \"Accuracy (%)\"\nset yrange [0:100]\nplot 'test.log' title \"Testing\" with linespoints ls 1 linetype 6,'train.log' title \"Training\" with linespoints ls 1 linetype 7  \n<\/code><\/pre>\n"}]}
{"QuestionId":33722235,"AnswerCount":3,"Tags":"<machine-learning><neural-network><tensorflow><conv-neural-network>","CreationDate":"2015-11-15T16:53:06.763","AcceptedAnswerId":33722335.0,"OwnerUserId":3849143.0,"Title":"Tensorflow model zoo?","Body":"<p>One of the main advantages of caffe for me was the possibility of doing transfer learning on freely distributed pretrained models.<\/p>\n\n<p>Is there a place to get trained models from papers\/competitions in tensorflow format? <\/p>\n\n<p>If not, is there a possibility to convert existing caffe(or any other) models into tensorflow models?<\/p>\n","answers":[{"AnswerId":"45207767","CreationDate":"2017-07-20T07:12:25.067","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>As of now, official TF github page has a <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">list with the group of models<\/a>. Also you can use <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">keras.applications<\/a> which has pretrained models.<\/p>\n"},{"AnswerId":"51858161","CreationDate":"2018-08-15T11:51:25.973","ParentId":null,"OwnerUserId":"5838671","Title":null,"Body":"<p>There are a few ways to do transfer learning with Tensorflow some of which combine TF Slim and custom codes, but there is a very nice collection of pretrained Tensorflow models <a href=\"https:\/\/github.com\/taehoonlee\/tensornets\" rel=\"nofollow noreferrer\">Tensornets<\/a> which contain almost all popular models and their pretrained weights<\/p>\n"},{"AnswerId":"33722335","CreationDate":"2015-11-15T17:02:04.857","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can likely use the <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\">caffe to tensorflow model converter<\/a> to convert <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo\">model zoo<\/a> models.  If you try it and report back, it would be great to know.  There's a potential issue with <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\/issues\/3\">converting maxpooling and padding<\/a>, but it seems to work for many models.<\/p>\n"}]}
{"QuestionId":33722904,"AnswerCount":2,"Tags":"<cuda><gpu><tensorflow>","CreationDate":"2015-11-15T17:49:04.170","AcceptedAnswerId":null,"OwnerUserId":5564902.0,"Title":"nvidia cuda 7.5 driver in tensorFlow are not properly handled (ubuntu 14.04)","Body":"<p>After a an install without pb, I am trying the tutorial about GPUs : \nI type :<\/p>\n\n<pre><\/pre>\n\n<p>I got :<\/p>\n\n<blockquote>\n  <p>Tensor(\"MatMul_1:0\", shape=TensorShape([Dimension(2), Dimension(2)]), dtype=float32, device=\/gpu:0)<\/p>\n<\/blockquote>\n\n<p>.<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"\", line 1, in\n    File \"\/home\/olivier\/anaconda\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 345, in run\n      results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)\n    File \"\/home\/olivier\/anaconda\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 419, in _do_run\n      e.code)\n  tensorflow.python.framework.errors.InvalidArgumentError: Cannot assign a device to node 'b_1': Could not satisfy explicit device specification '\/gpu:0'\n       [[Node: b_1 = Constdtype=DT_FLOAT, value=Tensor, _device=\"\/gpu:0\"]]\n  Caused by op u'b_1', defined at:\n    File \"\", line 3, in\n    File \"\/home\/olivier\/anaconda\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/constant_op.py\", line 147, in constant\n      attrs={\"value\": tensor_value, \"dtype\": dtype_value}, name=name).outputs[0]\n    File \"\/home\/olivier\/anaconda\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1710, in create_op\n      original_op=self._default_original_op, op_def=op_def)\n    File \"\/home\/olivier\/anaconda\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 988, in <strong>init<\/strong>\n      self._traceback = _extract_stack()<\/p>\n<\/blockquote>\n\n<p>In Torch7, my GPU works normally<\/p>\n","answers":[{"AnswerId":"33744428","CreationDate":"2015-11-16T20:54:18.237","ParentId":null,"OwnerUserId":"5564902","Title":null,"Body":"<p>The binaries published by google need to find <code>libcudart.so.7.0<\/code> in the path library , you just need to add it to  <code>LD_LIBRARY_PATH<\/code> by something like <\/p>\n\n<p><code>export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:\/home\/olivier\/digits-2.0\/lib\/cuda\"<\/code><\/p>\n\n<p>that you put in your <code>.bashrc<\/code><\/p>\n"},{"AnswerId":"35972364","CreationDate":"2016-03-13T15:52:48.387","ParentId":null,"OwnerUserId":"601314","Title":null,"Body":"<p>On an optimus laptop (running Manjaro Linux) it's possible to run TensorFlow with cuda acceleration by starting a python console with optirun:<\/p>\n\n<pre><code>$optirun python\n<\/code><\/pre>\n\n<p>I detailled the way to do it <a href=\"http:\/\/bit.ly\/1RHApU5\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":33724225,"AnswerCount":2,"Tags":"<python><python-2.7><lua><torch>","CreationDate":"2015-11-15T19:50:18.147","AcceptedAnswerId":36896678.0,"OwnerUserId":3106889.0,"Title":"Call torch7 (Lua) function from python?","Body":"<p>I have a program is written in python and I have model ConvNet trained using Toch7. I would like to call forward and backpro to the model from python program as I find difficult and hard to write it again in lua.<\/p>\n\n<p>Any idea please?<\/p>\n","answers":[{"AnswerId":"36896678","CreationDate":"2016-04-27T17:20:26.877","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>I think you now have a much better solution, which is <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow noreferrer\">lutorpy<\/a>.\nDifferent from pytorch, you have a lua engine in python, so it's more flexible to import <strong>any lua module<\/strong> and code in python, and it's easy to use and flexible. For pytorch you only have very few ported module which you can directly use in python. <\/p>\n\n<p>With lutorpy, you can convert between numpy and torch tensor easily and very fast. <\/p>\n\n<p>For you case, you can write your code in python like this:<\/p>\n\n<pre><code>import numpy as np\nimport lutorpy as lua\n\nmodel = torch.load('PATH TO YOUR MODEL FILE')\n\n# generate your input data with numpy\narr = np.random.randn(100)\n\n# convert your numpy array into torch tensor\nx = torch.fromNumpyArray(arr)\n\n# apply model forward method with \"._\" syntax(which is equivalent to \":\" in lua)\ny = model._forward(x)\n<\/code><\/pre>\n\n<p>A brief comparison between different library:\n<a href=\"https:\/\/stackoverflow.com\/questions\/36446899\/how-can-i-load-and-use-torch-deep-learning-models-from-python\">How can I load and use torch deep learning models from python?<\/a><\/p>\n"},{"AnswerId":"33739697","CreationDate":"2015-11-16T16:07:24.957","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>As was suggested by one of the Torch authors on <a href=\"https:\/\/groups.google.com\/d\/msg\/torch7\/xmRXKHU5ihA\/jGULM_X3CgAJ\" rel=\"nofollow\">torch7 maillist<\/a>, you can try <a href=\"https:\/\/github.com\/hughperkins\/pytorch\" rel=\"nofollow\">pytorch<\/a>.<\/p>\n"}]}
{"QuestionId":33724374,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-15T20:03:00.273","AcceptedAnswerId":null,"OwnerUserId":5558025.0,"Title":"Run training verbosely to check status if command","Body":"<p>I have been running this command: <\/p>\n\n<p>for about 10 hours, and still no output. Is there a way to run the command in a verbose mode?<\/p>\n","answers":[{"AnswerId":"33726315","CreationDate":"2015-11-15T23:22:21.653","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You can check all the flags for this bazel command with <code>bazel help<\/code>, which will give you something <a href=\"http:\/\/bazel.io\/docs\/bazel-user-manual.html\" rel=\"nofollow\">like this<\/a>. I have not run the command, but looking through that page I see the following section <strong>Verbosity options: options that control what Bazel prints<\/strong>, which has what you want:<\/p>\n\n<ul>\n<li><code>--explain logfile<\/code><\/li>\n<li><code>--verbose_explanations<\/code><\/li>\n<li><code>--profile file<\/code><\/li>\n<li><code>--[no]show_progress<\/code><\/li>\n<li><code>--show_result n<\/code><\/li>\n<\/ul>\n\n<p>All sounds close to what you want. Read docs for details.<\/p>\n"},{"AnswerId":"33729406","CreationDate":"2015-11-16T06:10:00.887","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The RNN translation example does not have a specific \"verbose\" flag, but all of its modes produce output on stdout. The default mode is to train a model, which runs indefinitely, producing output (and a model checkpoint) every <code>--steps_per_checkpoint=N<\/code> steps.<\/p>\n\n<p>It is possible that Bazel is buffering standard output from the process, so you aren't able to see progress being made. Try running the built binary instead:<\/p>\n\n<pre><code>$ bazel build tensorflow\/models\/rnn\/translate:translate -c opt\n$ bazel-bin\/tensorflow\/models\/rnn\/translate\/translate --data_dir \/Users\/Username\/data\/ --train_dir \/Users\/User\/train\/  --en_vocab_size=40000 --fr_vocab_size=40000 --size 256 --num_layers 2 --steps_per_checkpoint=50\n<\/code><\/pre>\n"}]}
{"QuestionId":33725159,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-15T21:17:28.877","AcceptedAnswerId":null,"OwnerUserId":5557105.0,"Title":"Understanding Variable scope example in Tensorflow","Body":"<p>I was looking at the mechanics section for Tensorflow, specifically on <a href=\"https:\/\/chromium.googlesource.com\/external\/github.com\/tensorflow\/tensorflow\/+\/r0.7\/tensorflow\/g3doc\/how_tos\/variable_scope\/index.md\" rel=\"nofollow noreferrer\">shared variables<\/a>. In the section \"The problem\", they are dealing with a convolutional neural net, and provide the following code (which runs an image through the model):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>If the model was implemented in such a way, would it then be impossible to learn\/update the parameters because there's a new set of parameters for each image in my training set?<\/p>\n\n<p>Edit:\nI've also tried \"the problem\" approach on a simple linear regression <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/1_linear_regression.py\" rel=\"nofollow noreferrer\">example<\/a>, and there do not appear to be any issues with this method of implementation. Training seems to work as well as can be shown by the last line of the code. So I'm wondering if there is a subtle discrepancy in the tensorflow documentation and what I'm doing. :<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34547289","CreationDate":"2015-12-31T13:39:25.910","ParentId":null,"OwnerUserId":"1716733","Title":null,"Body":"<p>One has to create the variable set only once per whole training (and testing) set. The goal of <em>variable scopes<\/em> is to allow for modularization of <em>subsets of parameters<\/em>, such as those belonging to layers (e.g. when architecture of a layer is repeated, the same names can be used within each layer scope).<\/p>\n\n<p>In your example you create parameters only in the <code>model<\/code> function. You can print out your variable names to see that it is assigned to the specified scope:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from __future__ import print_function\n\nX = tf.placeholder(\"float\") # create symbolic variables\nY = tf.placeholder(\"float\")\nprint(\"X:\", X.name)\nprint(\"Y:\", Y.name)\n\ndef model(X):\n    with tf.variable_scope(\"param\"):\n        w = tf.Variable(0.0, name=\"weights\") # create a shared variable (like theano.shared) for the weight matrix\n    print(\"w:\", w.name)\n    return tf.mul(X, w) \n<\/code><\/pre>\n\n<p>The call to <code>sess.run(train_op, feed_dict={X: x, Y: y})<\/code> only evaluates the value of <code>train_op<\/code> given the provided values of <code>X<\/code> and <code>Y<\/code>. No new  variables (incl. parameters) are created there; therefore, it has no effect. You can make sure the variable names stay the same by again printing them out:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.variable_scope(\"train\"):\n    print(\"X:\", X.name)\n    print(\"Y:\", Y.name)\n    for i in range(100):\n        for (x, y) in zip(trX, trY):\n            sess.run(train_op, feed_dict={X: x, Y: y})\n<\/code><\/pre>\n\n<p>You will see that variable names stay the same, as they are already initialized.<\/p>\n\n<p>If you'd like to <em>retrieve a variable<\/em> using its scope, you need to use <code>get_variable<\/code> within a <code>tf.variable_scope<\/code> enclosure:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.variable_scope(\"param\"):\n    w = tf.get_variable(\"weights\", [1])\nprint(\"w:\", w.name)\n<\/code><\/pre>\n"}]}
{"QuestionId":33726333,"AnswerCount":3,"Tags":"<parallel-processing><neural-network><gpu><deep-learning><caffe>","CreationDate":"2015-11-15T23:23:54.010","AcceptedAnswerId":33729979.0,"OwnerUserId":4086273.0,"Title":"Caffe\/pyCaffe: set all GPUs","Body":"<p>is possible to set all GPUs for Caffe (especially pyCaffe)?<\/p>\n\n<p>Something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33729979","CreationDate":"2015-11-16T07:03:02.987","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>AFAIK Caffe is not supporting multi gpu training at the moment. It is planned for future release. See a discussion <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/876\" rel=\"noreferrer\">here<\/a>.<br>\nIt seems like <a href=\"https:\/\/github.com\/NVIDIA\/caffe\" rel=\"noreferrer\">NVIDIA's branch of caffe<\/a> has this functionality. See the issue <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/issues\/92\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"33784028","CreationDate":"2015-11-18T15:28:21.653","ParentId":null,"OwnerUserId":"2570842","Title":null,"Body":"<p>It seems that Caffe now starts supporting training on multi-GPUs for C++ interface.\nSee the docs on <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html\" rel=\"nofollow\">CommandLine interface<\/a>. <\/p>\n\n<pre><code># train on GPUs 0 &amp; 1 (doubling the batch size)\ncaffe train -solver examples\/mnist\/lenet_solver.prototxt -gpu 0,1\n# train on all GPUs (multiplying batch size by number of devices)\ncaffe train -solver examples\/mnist\/lenet_solver.prototxt -gpu all\n<\/code><\/pre>\n"},{"AnswerId":"34953508","CreationDate":"2016-01-22T18:25:58.010","ParentId":null,"OwnerUserId":"2404152","Title":null,"Body":"<p>Both forks have supported multi-GPU for a while now.<\/p>\n\n<ul>\n<li>BVLC\/caffe got support for multi-GPU on 08\/13\/2015 (see <a href=\"https:\/\/github.com\/BVLC\/caffe\/commit\/bb0a90e8aabb0e1ffd5732439591814d10d7fd45\">commit<\/a>, <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2903\">issue<\/a>).<\/li>\n<li>NVIDIA\/caffe got support for multi-GPU on 06\/19\/2015 (see <a href=\"https:\/\/github.com\/NVIDIA\/caffe\/releases\/tag\/v0.12.0\">release note<\/a>).<\/li>\n<\/ul>\n\n<p>You may be interested to know that there is a pretty serious outstanding issue with multi-GPU and PythonLayers (see <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2936\">issue<\/a>, <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3032\">temporary fix<\/a>).<\/p>\n"}]}
{"QuestionId":33727116,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-16T01:11:58.033","AcceptedAnswerId":null,"OwnerUserId":3728184.0,"Title":"How to specify number of GPUs in Python interface?","Body":"<p>When I start a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow\">tensorflow<\/a>  it immediately allocates 100% of the memory on both GPUs in my dekstop (devices 0 and 2).<\/p>\n\n<p>How can I use the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/index.md\" rel=\"nofollow\">Python API<\/a> to restrict it to just 1 GPU?<\/p>\n","answers":[{"AnswerId":"33728220","CreationDate":"2015-11-16T03:50:07.220","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>If you want some operations to be done only with a particular device, you can do the following:<\/p>\n\n<pre><code>with tf.device('\/gpu:0'):\n   ...\n   # all your operations\n<\/code><\/pre>\n\n<p>Where the devices are specified in the following way: <code>\"\/cpu:0\"<\/code>, <code>\"\/gpu:0\"<\/code>, <code>\"\/gpu:1\"<\/code>. I suggest you to read more details in the topic <a href=\"http:\/\/tensorflow.org\/how_tos\/using_gpu\/index.md\" rel=\"nofollow\">using GPU devices<\/a>.<\/p>\n\n<p>If you want to see some tutorials, take a look at <a href=\"http:\/\/tensorflow.org\/tutorials\/deep_cnn\/index.md\" rel=\"nofollow\">cifar10 tutorial.<\/a><\/p>\n"},{"AnswerId":"33730425","CreationDate":"2015-11-16T07:37:25.223","ParentId":null,"OwnerUserId":"3728184","Title":null,"Body":"<p>A crude way to restrict it to use just GPU #0, for example, is to define this:<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0\n<\/code><\/pre>\n"}]}
{"QuestionId":33727540,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-16T02:13:49.903","AcceptedAnswerId":33729233.0,"OwnerUserId":1187026.0,"Title":"Does the output of tf.nn.bias_add(value, bias) ever have a different shape than shape of value","Body":"<p>So in the convolutional neural network cifar10 example in tensorflow, in the  method of cifar10.py I see several instances of this:<\/p>\n\n<pre><\/pre>\n\n<p>It seems like the reshape is making sure the output of  has the shape of value<\/p>\n\n<p>My question is, is the  necessary? Is there a situation where  will not return a tensor with the same shape as value?<\/p>\n","answers":[{"AnswerId":"33729233","CreationDate":"2015-11-16T05:50:17.040","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/common_shapes.py#L87\" rel=\"nofollow noreferrer\">shape of the result<\/a> of <code>tf.nn.bias_add(value, bias)<\/code> is always the same as the shape of value, so these calls to <code>tf.reshape()<\/code> are unnecessary.<\/p>\n\n<p>Occasionally, calls to <code>tf.reshape()<\/code> are used to add explicit information about the shape, but the <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/faq#what_is_the_difference_between_xset_shape_and_x_tfreshapex\" rel=\"nofollow noreferrer\">recommended way to do this, per the FAQ<\/a>, is to use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#set_shape\" rel=\"nofollow noreferrer\"><code>Tensor.set_shape()<\/code><\/a> method to add shape information without adding a redundant operation to the graph.<\/p>\n"}]}
{"QuestionId":33727649,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-16T02:29:51.770","AcceptedAnswerId":33739256.0,"OwnerUserId":2829987.0,"Title":"Error downloading Tensorflow MNIST data","Body":"<p>I am not able to use input_data.py to download the MNIST dataset. When I try to call the method, I see some html and the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Not sure what I am doing wrong. The code I have is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33739256","CreationDate":"2015-11-16T15:45:29.613","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The syntax error is raised because <code>input_data.py<\/code> is in HTML format \u2014 notice the HTML tags in the error message \u2014 presumably because you downloaded the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py\" rel=\"nofollow\">webpage<\/a> that displays the content of the file from GitHub.<\/p>\n\n<p>Instead, download the raw contents of <code>input_data.py<\/code> from this link to get a valid Python file: <a href=\"https:\/\/raw.githubusercontent.com\/tensorflow\/tensorflow\/master\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py\" rel=\"nofollow\">https:\/\/raw.githubusercontent.com\/tensorflow\/tensorflow\/master\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py<\/a><\/p>\n"}]}
{"QuestionId":33727726,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-16T02:40:30.963","AcceptedAnswerId":33737060.0,"OwnerUserId":5565899.0,"Title":"Tensor flow install OSX","Body":"<p>I'm trying to install TensorFlow on El 10.11.1, and I'm having issues creating and installing the pip package that I do not understand at all.<\/p>\n\n<p>I successfully installed Bazel, SWIG and NumPy <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md\">per instructions<\/a> (at least they all seem to work at this time).<\/p>\n\n<p>Creating the pip package\/install I get the following error when I run the command:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>\/Users\/christophermichael\/tensorflow\/tensorflow\/tensorflow\/python\/BUILD:730:1: C++ compilation of rule '\/\/tensorflow\/python:_pywrap_tensorflow.so' failed: osx_gcc_wrapper.sh failed: error executing command <\/p>\n<\/blockquote>\n\n<p>Running with , I get:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n","answers":[{"AnswerId":"33737060","CreationDate":"2015-11-16T13:54:11.447","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>Why do you want to install from source? <\/p>\n\n<p>Create a virtualenv and install it there, I have the same OS version and it works super fine from this method:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/33691154\/1973820\">https:\/\/stackoverflow.com\/a\/33691154\/1973820<\/a><\/p>\n"},{"AnswerId":"34073102","CreationDate":"2015-12-03T18:03:47.687","ParentId":null,"OwnerUserId":"5636237","Title":null,"Body":"<p>Yeah, as Hamed said above, it's better directly install the binaries.<\/p>\n\n<p>However, if you do need to build from the source, then these steps should work:<\/p>\n\n<ul>\n<li><code>brew install python<\/code><\/li>\n<li><code>virtualenv ~\/tf<\/code><\/li>\n<li><code>source ~\/tf\/bin\/activate<\/code><\/li>\n<li>Install bazel for Mac<\/li>\n<li>Install PCRE (not PCRE2) for Mac<\/li>\n<li>Install SWIG for Mac<\/li>\n<li><code>git clone --recurse-submodules https:\/\/github.com\/tensorflow\/tensorflow<\/code><\/li>\n<li><code>cd tensorflow<\/code><\/li>\n<li><code>.\/configure<\/code><\/li>\n<li><code>bazel build -c opt \/\/tensorflow\/tools\/pip_package:build_pip_package<\/code><\/li>\n<li><code>bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg<\/code><\/li>\n<li><code>pip install \/tmp\/tensorflow_pkg\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/code><\/li>\n<li><code>bazel build -c opt \/\/tensorflow\/cc:tutorials_example_trainer<\/code><\/li>\n<\/ul>\n"},{"AnswerId":"34746395","CreationDate":"2016-01-12T14:32:46.890","ParentId":null,"OwnerUserId":"588709","Title":null,"Body":"<p>Another option is to install anaconda and use its environment management. The advantage of using environments either with anaconda or virtualenv is that you separate sources and dependencies for different use cases. This means, if you need to make changes to your python libraries for something else, it won't affect your tensorflow environmnet.<\/p>\n\n<p>Here is how to install tensorflow with anaconda.<\/p>\n\n<ol>\n<li>Install anaconda <a href=\"http:\/\/docs.continuum.io\/anaconda\/install\" rel=\"nofollow\">http:\/\/docs.continuum.io\/anaconda\/install<\/a> <\/li>\n<li>Create a new conda workspace <\/li>\n<li>Download the specific protobuf version that tensorflow needs: <a href=\"https:\/\/pypi.python.org\/pypi\/protobuf\/3.0.0a3\" rel=\"nofollow\">https:\/\/pypi.python.org\/pypi\/protobuf\/3.0.0a3<\/a><\/li>\n<li>Install it via <code>sudo easy_install ~\/Downloads\/protobuf-3.0.0a3-py2.7.egg<\/code> <\/li>\n<li>Install a numpy version greater than 1.08.x via <code>conda install numpy<\/code> <\/li>\n<li><p>Download the 0.6.0 version of tensorflow:\n<a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl<\/a><\/p><\/li>\n<li><p>Install via <code>pip install ~\/Downloads\/tensorflow-0.6.0-py2-none-any.whl<\/code> <\/p><\/li>\n<\/ol>\n\n<p>When you install\n    tensorflow from the whl file directly, it should tell you when\n    dependencies are not there. It seems not to be able to resolve these\n    conflicts independently. My setup had issues with protobuf and\n    numpy. After installing them manually everything worked fine.<\/p>\n\n<p>I hope this helps!<\/p>\n"}]}
{"QuestionId":33727935,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-16T03:10:01.573","AcceptedAnswerId":null,"OwnerUserId":5565980.0,"Title":"How to use stop_gradient in Tensorflow","Body":"<p>I'm wondering how to use  in tensorflow, and the documentation is not clear to me. <\/p>\n\n<p>I'm currently using  to produce the gradient of the loss function w.r.t. the word embeddings in a CBOW word2vec model. I want to just get the value, and not do backpropagation (as I'm generating adversarial examples). <\/p>\n\n<p>Currently, I'm using the code:<\/p>\n\n<pre><\/pre>\n\n<p><s>But when I run this, it does the backpropogation anyway!<\/s> What am I doing wrong, and just as importantly, how can I fix this? <\/p>\n\n<p>CLARIFICATION: To clarify by \"backpropagation\" I mean \"calculating values and updating model parameters\".<\/p>\n\n<h3>UPDATE<\/h3>\n\n<p>If I run the two lines above after the first training step, the I get a different loss after 100 training steps than when I don't run those two lines. I might be fundamentally misunderstanding something about Tensorflow.<\/p>\n\n<p>I've tried setting using  both in the beginning of the graph declaration and before each training step. The total loss is consistent between multiple runs, but not between including\/excluding those two lines. So if it's not the RNG causing the disparity, and it's not unanticipated updating of the model parameters between training steps, do you have any idea what would cause this behavior? <\/p>\n\n<h3>SOLUTION<\/h3>\n\n<p>Welp, it's a bit late but here's how I solved it. I only wanted to optimize over some, but not all, variables. I thought that the way to prevent optimizing some variables would be to use  - but I never found a way to make that work. Maybe there is a way, but what worked for me was to adjust my  to only optimize over a list of variables. So instead of:<\/p>\n\n<pre><\/pre>\n\n<p>I used:<\/p>\n\n<pre><\/pre>\n\n<p>This prevented  from updating the variables not in . Hopefully it works for you, too!<\/p>\n","answers":[{"AnswerId":"39356229","CreationDate":"2016-09-06T19:10:39.937","ParentId":null,"OwnerUserId":"4001737","Title":null,"Body":"<p>tf.stop_gradient provides a way to not compute gradient with respect to some variables during back-propagation.  <\/p>\n\n<p>For example, in the code below, we have three variables, w1, w2, w3 and input x. The loss is square((x1.dot(w1) - x.dot(w2 * w3))). We want to minimize this loss wrt to w1 but want to keep w2 and w3 fixed. To achieve this we can just put tf.stop_gradient(tf.matmul(x, w2*w3)).  <\/p>\n\n<p>In the figure below, I plotted how w1, w2, and w3 from their initial values as the function of training iterations. It can be seen that w2 and w3 remains fixed while w1 changes until it becomes equal to w2 * w3.<\/p>\n\n<p>An image showing that w1 only learns but not w2 and w3:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5qEBW.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5qEBW.png\" alt=\"An image showing that w1 only learns but not w2 and w3\"><\/a><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nw1 = tf.get_variable(\"w1\", shape=[5, 1], initializer=tf.truncated_normal_initializer())\nw2 = tf.get_variable(\"w2\", shape=[5, 1], initializer=tf.truncated_normal_initializer())\nw3 = tf.get_variable(\"w3\", shape=[5, 1], initializer=tf.truncated_normal_initializer())\nx = tf.placeholder(tf.float32, shape=[None, 5], name=\"x\")\n\n\na1 = tf.matmul(x, w1)\na2 = tf.matmul(x, w2*w3)\na2 = tf.stop_gradient(a2)\nloss = tf.reduce_mean(tf.square(a1 - a2))\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)\ngradients = optimizer.compute_gradients(loss)\ntrain_op = optimizer.apply_gradients(gradients)\n<\/code><\/pre>\n"},{"AnswerId":"33729320","CreationDate":"2015-11-16T05:59:59.310","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gradients\" rel=\"noreferrer\"><code>tf.gradients(loss, embed)<\/code><\/a> computes the partial derivative of the tensor <code>loss<\/code> with respect to the tensor <code>embed<\/code>. TensorFlow computes this partial derivative by backpropagation, so it is expected behavior that evaluating the result of <code>tf.gradients(...)<\/code> performs backpropagation. However, evaluating that tensor does not perform any variable updates, because the expression does not include any <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/state_ops#Variable_helper_functions\" rel=\"noreferrer\">assignment operations<\/a>.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stop_gradient\" rel=\"noreferrer\"><code>tf.stop_gradient()<\/code><\/a> is an operation that acts as the identity function in the forward direction, but stops the accumulated gradient from flowing through that operator in the backward direction. It does not prevent backpropagation altogether, but instead prevents an individual tensor from contributing to the gradients that are computed for an expression. The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stop_gradient\" rel=\"noreferrer\">documentation for the operation<\/a> has more details about the operation, and when to use it. <\/p>\n"}]}
{"QuestionId":33730259,"AnswerCount":2,"Tags":"<python><macos><tensorflow>","CreationDate":"2015-11-16T07:24:37.130","AcceptedAnswerId":33856549.0,"OwnerUserId":5518689.0,"Title":"TensorFlow installation error on Mac","Body":"<p>I want to use TensorFlow on my Mac with PyCharm, but when I use:<\/p>\n\n<pre><\/pre>\n\n<p>it gives an error:<\/p>\n\n<pre><\/pre>\n\n<p>How do I fix it?<\/p>\n","answers":[{"AnswerId":"33856549","CreationDate":"2015-11-22T15:32:07.163","ParentId":null,"OwnerUserId":"2976939","Title":null,"Body":"<p>I got it to install by running the following for a virtualenv setup recommended by the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#user-content-virtualenv-based-installation-virtualenv_install\" rel=\"nofollow\">documentation<\/a><\/p>\n\n<h3>Install python<\/h3>\n\n<pre><code>brew install python\n<\/code><\/pre>\n\n<h3>Install tools<\/h3>\n\n<pre><code># On Mac:\n$ sudo easy_install pip  # If pip is not already installed\n$ sudo pip install --upgrade virtualenv\n<\/code><\/pre>\n\n<h3>Setup virtualenv in ~\/tensorflow directory<\/h3>\n\n<pre><code>$ virtualenv --system-site-packages ~\/tensorflow\n$ cd ~\/tensorflow\n<\/code><\/pre>\n\n<h3>Activate virtualenv<\/h3>\n\n<pre><code>$ source bin\/activate  # If using bash\n$ source bin\/activate.csh  # If using csh\n(tensorflow)$  # Your prompt should change\n<\/code><\/pre>\n\n<h3>Install tensorflow (this is where I got the SSL error)<\/h3>\n\n<p>Download tensorflow from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#user-content-mac-os-x\" rel=\"nofollow\">google<\/a> getting the latest version using curl<\/p>\n\n<pre><code>curl -O https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>Install the package you downloaded<\/p>\n\n<pre><code>pip install tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>You should be able to type <code>python<\/code> and try out tensorflow in a python terminal<\/p>\n"},{"AnswerId":"34122049","CreationDate":"2015-12-06T20:04:34.463","ParentId":null,"OwnerUserId":"2917986","Title":null,"Body":"<p>Another common issue in Mac while installing Tensor flow is Protobuf version.<\/p>\n\n<p>Please make sure you have protobuf 3.0.0a3 or above.<\/p>\n\n<p>Thanks to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11#issuecomment-155174905\" rel=\"nofollow\">Github Issue on Tensorflow<\/a> To install using homebrew<\/p>\n\n<pre><code>brew install --devel protobuf\n<\/code><\/pre>\n\n<p>To verify, please run <\/p>\n\n<pre><code>pip show numpy protobuf\n<\/code><\/pre>\n"}]}
{"QuestionId":33731366,"AnswerCount":1,"Tags":"<python><python-2.7><redhat><glibc><tensorflow>","CreationDate":"2015-11-16T08:50:30.413","AcceptedAnswerId":null,"OwnerUserId":2353472.0,"Title":"Error while importing Tensorflow in python2.7 in Red Hat release 6.6. 'GLIBC_2.17 not found'","Body":"<p>This is essentially a repeat of question asked <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\">here<\/a>. However, I am using Red Hat Version 6.6, which has glibc 2.12 (glibc 2.17, I think was introduced with RHEL ver 7). Is it possible to install tensorflow locally, without upgrading OS. (I don't have admin privileges). This is the error I am getting<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33747710","CreationDate":"2015-11-17T01:36:01.640","ParentId":null,"OwnerUserId":"50617","Title":null,"Body":"<blockquote>\n  <p>This is essentially a repeat of question asked here.<\/p>\n<\/blockquote>\n\n<p>And the same <a href=\"https:\/\/stackoverflow.com\/a\/33658956\/50617\">answer<\/a> should work. What stops you from downloading glibc-2.17 packages, unpacking them, and using the <code>glibc-2.17\/ld.so<\/code> with correct <code>--library-path<\/code>?<\/p>\n\n<blockquote>\n  <p>Is it possible to install tensorflow locally, without upgrading OS.<\/p>\n<\/blockquote>\n\n<p>Yes, using above answer.<\/p>\n"}]}
{"QuestionId":33731875,"AnswerCount":3,"Tags":"<python-2.7><tensorflow>","CreationDate":"2015-11-16T09:18:42.780","AcceptedAnswerId":42618440.0,"OwnerUserId":5566910.0,"Title":"TensorFlow IOError: [Errno socket error] [Errno 104] Connection reset by peer","Body":"<p>I have installed TensorFlow. When I download MNISTdataset, there exists one error. Who can tell me what's wrong? Thanks very much! The error detail as follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47495810","CreationDate":"2017-11-26T11:47:59.533","ParentId":null,"OwnerUserId":"5634636","Title":null,"Body":"\n\n<p>As @user3851234 's answer, you can download the dataset by changing the <code>SOURCE_URL<\/code>, like: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>SOURCE_URL = 'https:\/\/s3.amazonaws.com\/lasagne\/recipes\/datasets\/mnist\/'\n<\/code><\/pre>\n\n<p>There're few options for you:<\/p>\n\n<ul>\n<li>Offical source: <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a><\/li>\n<li>Google mirror: <a href=\"https:\/\/storage.googleapis.com\/cvdf-datasets\/mnist\/\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/cvdf-datasets\/mnist\/<\/a><\/li>\n<li>Amazon mirror: <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<hr>\n\n<p><strong>Special note for programmers in China<\/strong><\/p>\n\n<p>The mirror provided by Google and Amazon often failed in Mainland China, you should try <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">Yann LeCun's page<\/a> first.<\/p>\n\n<hr>\n\n<p><strong>Make your own MNIST mirror<\/strong><\/p>\n\n<p>Preparation:<\/p>\n\n<p>4 MNIST data files wherever you got:<\/p>\n\n<ul>\n<li>train-images-idx3-ubyte.gz:  training set images (9912422 bytes) <\/li>\n<li>train-labels-idx1-ubyte.gz:  training set labels (28881 bytes) <\/li>\n<li>t10k-images-idx3-ubyte.gz:   test set images (1648877 bytes) <\/li>\n<li>t10k-labels-idx1-ubyte.gz:   test set labels (4542 bytes)<\/li>\n<\/ul>\n\n<p>and put them under your domain, like:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Amy\/mnistrepo\/mnist\/train-images-idx3-ubyte.gz\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Amy\/mnistrepo\/mnist\/train-images-idx3-ubyte.gz<\/a><\/p>\n\n<p>and change <code>SOURCE_URL<\/code> to your domain:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>SOURCE_URL = 'https:\/\/github.com\/Amy\/mnistrepo\/mnist\n<\/code><\/pre>\n"},{"AnswerId":"42618440","CreationDate":"2017-03-06T05:51:55.063","ParentId":null,"OwnerUserId":"3851234","Title":null,"Body":"<p>If the site is still not accessible even after hours, you can modify the \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py and change the source link to '<a href=\"https:\/\/s3.amazonaws.com\/lasagne\/recipes\/datasets\/mnist\/\" rel=\"nofollow noreferrer\">https:\/\/s3.amazonaws.com\/lasagne\/recipes\/datasets\/mnist\/<\/a>' and run the example again.<\/p>\n"},{"AnswerId":"33741079","CreationDate":"2015-11-16T17:22:09.157","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Python was not able to download the MNIST dataset from lecun.com.  First, check to make sure you can browse <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"noreferrer\">Yann LeCun's MNIST page<\/a> from that computer.  If you can't, it may be a firewall or Internet connectivity issue.  If you can, try running the download again in a few minutes - I've seen this error transiently a few times since the release of Tensorflow and it's always gone away within 5 minutes.<\/p>\n"}]}
{"QuestionId":33732477,"AnswerCount":1,"Tags":"<python><python-2.7><deep-learning><protocol-buffers><caffe>","CreationDate":"2015-11-16T09:52:56.897","AcceptedAnswerId":null,"OwnerUserId":4415918.0,"Title":"Google DeepDream: Message type \"caffe.NetParameter\" has no field named \"input_shape\"","Body":"<p>I want to start the Google DeepDream IPython Notebook from <a href=\"https:\/\/github.com\/google\/deepdream\" rel=\"nofollow\">https:\/\/github.com\/google\/deepdream<\/a>. I installed the Dependencies listed in the notebook (caffe and protobuf) and started the notebook with the anaconda distribution of python.<\/p>\n\n<pre><\/pre>\n\n<p>When running this code cell an error is thrown:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the complete Stack Trace:<\/p>\n\n<pre><\/pre>\n\n<p>Don't really know how to fix this error. I Hope you guys can help me.<\/p>\n","answers":[{"AnswerId":"46071314","CreationDate":"2017-09-06T09:17:16.217","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>In your prototxt, change:<\/p>\n\n<pre><code>input_shape: {\ndim: 1\ndim: 3\ndim: 224\ndim: 224\n}\n<\/code><\/pre>\n\n<p>form to:<\/p>\n\n<pre><code>input_dim: 1\ninput_dim: 3\ninput_dim: 224\ninput_dim: 224\n<\/code><\/pre>\n\n<p>The proto API has changed somewhere.<\/p>\n"}]}
{"QuestionId":33735372,"AnswerCount":0,"Tags":"<theano><pylearn>","CreationDate":"2015-11-16T12:26:23.253","AcceptedAnswerId":null,"OwnerUserId":641072.0,"Title":"Adding L2 regularization to SGD trainer in MLP from pylearn2\/Theano","Body":"<p>I am new to Pylearn -so this question might be pretty basic. I want to add L2 regularization to the SGD trainer for MLP, but I am not sure how to do so. Here's my code (that fails):<\/p>\n\n<pre><\/pre>\n\n<p>NOTE: I have skipped the data reading part - assumed it doesn't affect the question. \nThis is how I run the mlp:<\/p>\n\n<pre><\/pre>\n\n<p>This is the error I see:<\/p>\n\n<pre><\/pre>\n\n<p>I am not sure how to include a L2 cost into the regularization? Can someone please help?<\/p>\n","answers":[]}
{"QuestionId":33735833,"AnswerCount":1,"Tags":"<install><tensorflow><six>","CreationDate":"2015-11-16T12:49:57.163","AcceptedAnswerId":33741143.0,"OwnerUserId":5567602.0,"Title":"Tensor Flow installation on RHEL 7","Body":"<p>I am new to Tensor Flow and trying to install it on my RHEL 7 machine using the below command. I cannot connect the machine directly to internet due to some restrictions so using a whl file downloaded from Tensor flow site.<\/p>\n\n<pre><\/pre>\n\n<p>But getting error as below :<\/p>\n\n<blockquote>\n  <p>Could not find a version that satisfies the requirement six>=1.10.0\n  (from tensorflow==0.5.0) (from versions: ) No matching distribution\n  found for six>=1.10.0 (from tensorflow==0.5.0)<\/p>\n<\/blockquote>\n\n<p>I am using pip installed on my machine through anaconda and have python 2.7 installed on the machine as well <\/p>\n\n<p>Can you please help me to decode the error?<\/p>\n","answers":[{"AnswerId":"33741143","CreationDate":"2015-11-16T17:25:38.060","ParentId":null,"OwnerUserId":"2556980","Title":null,"Body":"<p>I had the same issue.  Your six package needs to be updated.  Try:<\/p>\n\n<pre><code>easy_install --upgrade six\n<\/code><\/pre>\n\n<p>Note: I am doing this from within a virtual environment and I do have internet.    <\/p>\n"}]}
{"QuestionId":33736795,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-11-16T13:41:10.517","AcceptedAnswerId":null,"OwnerUserId":3731804.0,"Title":"TensorFlow - numpy-like tensor indexing","Body":"<p>In numpy, we can do this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I try something equivalent in TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>The last line gives a \"Bad slice index tensor\" exception. It seems TensorFlow doesn't support indexing like numpy or Theano.<\/p>\n\n<p>Does anybody know if there is a TensorFlow way of doing this (indexing a tensor by arbitrary values). I've seen the tf.nn.embedding part, but I'm not sure they can be used for this and even if they can, it's a huge workaround for something this straightforward.<\/p>\n\n<p>(Right now, I'm feeding the data from  as an input and doing the indexing in numpy but I hoped to put  inside TensorFlow to get higher efficiency)<\/p>\n","answers":[{"AnswerId":"33741008","CreationDate":"2015-11-16T17:18:11.293","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>LDGN's comment is correct.  This is not possible at the moment, and is a requested feature.  If you follow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\">issue#206 on github<\/a> you'll get updated if\/when this is available.  Many people would like this feature.<\/p>\n"},{"AnswerId":"41893288","CreationDate":"2017-01-27T11:47:41.793","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can actually do that now with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"noreferrer\"><code>tf.gather_nd<\/code><\/a>. Let's say you have a matrix <code>m<\/code> like the following:<\/p>\n\n<pre><code>| 1 2 3 4 |\n| 5 6 7 8 |\n<\/code><\/pre>\n\n<p>And you want to build a matrix <code>r<\/code> of size, let's say, 3x2, built from elements of <code>m<\/code>, like this:<\/p>\n\n<pre><code>| 3 6 |\n| 2 7 |\n| 5 3 |\n| 1 1 |\n<\/code><\/pre>\n\n<p>Each element of <code>r<\/code> corresponds to a row and column of <code>m<\/code>, and you can have matrices <code>rows<\/code> and <code>cols<\/code> with these indices (zero-based, since we are programming, not doing math!):<\/p>\n\n<pre><code>       | 0 1 |         | 2 1 |\nrows = | 0 1 |  cols = | 1 2 |\n       | 1 0 |         | 0 2 |\n       | 0 0 |         | 0 0 |\n<\/code><\/pre>\n\n<p>Which you can stack into a 3-dimensional tensor like this:<\/p>\n\n<pre><code>| | 0 2 | | 1 1 | |\n| | 0 1 | | 1 2 | |\n| | 1 0 | | 2 0 | |\n| | 0 0 | | 0 0 | |\n<\/code><\/pre>\n\n<p>This way, you can get from <code>m<\/code> to <code>r<\/code> through <code>rows<\/code> and <code>cols<\/code> as follows:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nm = np.array([[1, 2, 3, 4], [5, 6, 7, 8]])\nrows = np.array([[0, 1], [0, 1], [1, 0], [0, 0]])\ncols = np.array([[2, 1], [1, 2], [0, 2], [0, 0]])\n\nx = tf.placeholder('float32', (None, None))\nidx1 = tf.placeholder('int32', (None, None))\nidx2 = tf.placeholder('int32', (None, None))\nresult = tf.gather_nd(x, tf.stack((idx1, idx2), -1))\n\nwith tf.Session() as sess:\n    r = sess.run(result, feed_dict={\n        x: m,\n        idx1: rows,\n        idx2: cols,\n    })\nprint(r)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[ 3.  6.]\n [ 2.  7.]\n [ 5.  3.]\n [ 1.  1.]]\n<\/code><\/pre>\n"},{"AnswerId":"41053686","CreationDate":"2016-12-09T05:30:09.987","ParentId":null,"OwnerUserId":"4172685","Title":null,"Body":"<p>For <code>Tensorflow 0.11<\/code>, basic indexing has been implemented. More advanced indexing (like boolean indexing) is still missing but apparently is planned for future versions.<\/p>\n\n<p>Advanced indexing can be tracked with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4638\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4638<\/a><\/p>\n"}]}
{"QuestionId":33737611,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-11-16T14:21:59.393","AcceptedAnswerId":null,"OwnerUserId":1846250.0,"Title":"Using tensorflow for sequence tagging : Synced sequence input and output","Body":"<p>I would like to use Tensorflow for sequence tagging namely Part of Speech tagging. I tried to use the same model outlined here: <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md\" rel=\"nofollow\">http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md<\/a> (which outlines a model to translate English to French). <\/p>\n\n<p>Since in tagging, the input sequence and output sequence have exactly the same length, I configured the buckets so that input and output sequences have same length and tried to learn a POS tagger using this model on ConLL 2000. <\/p>\n\n<p>However it seems that the decoder sometimes outputs a taggedsequence shorter than the input sequence (it seems to feel that the EOS tag appears prematurely)<\/p>\n\n<p>For example:\n<strong>He reckons the current account deficit will narrow to only # 1.8 billion in September .<\/strong><\/p>\n\n<p>The above sentence is tokenized to have 18 tokens which gets padded to 20 (due to bucketing).<\/p>\n\n<p>When asked to decode the above, the decoder spits out the following:<\/p>\n\n<p><strong>PRP VBD DT JJ JJ NN MD VB TO VB DT NN IN NN . _EOS . _EOS CD CD<\/strong><\/p>\n\n<p>So here it ends the sequence (EOS) after 15 tokens not 18. <\/p>\n\n<p>How can I force the sequence to learn that the decoded sequence should be the same length as the encoded one in my scenario.<\/p>\n","answers":[{"AnswerId":"33772126","CreationDate":"2015-11-18T04:56:58.247","ParentId":null,"OwnerUserId":"26094","Title":null,"Body":"<p>If your input and output sequences are the same length you probably want something simpler than a seq2seq model (since handling <em>different<\/em> sequence lengths is one of it's strengths)<\/p>\n\n<p>Have you tried just training (word -> tag) ?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6pdIT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6pdIT.png\" alt=\"tokens to tags\"><\/a><\/p>\n\n<p>note: that for something like pos tagging where there is clear signal from tokens on either side you'll definitely get a benefit from a bidirectional net.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/C8ntb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/C8ntb.png\" alt=\"bidir net\"><\/a><\/p>\n\n<p>If you want to go all crazy there would be some fun character level variants too where you only emit the tag at the token boundary (the rationale being that pos tagging benefits from character level features; e.g. things like out of vocab names). So many variants to try! :D<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/91Bwr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/91Bwr.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"38962548","CreationDate":"2016-08-15T20:28:30.447","ParentId":null,"OwnerUserId":"3413482","Title":null,"Body":"<p>I came to the same problem. At the end I found ptb_word_lm.py example in tensorflow's examples is exactly what we need for tokenization, NER and POS tagging.<\/p>\n\n<p>If you look into details of the language model example, you can find out that it treats the input character sequence as X and right shift X for 1 space as Y. It is exactly what fixed length sequence labeling needs. <\/p>\n"},{"AnswerId":"33770456","CreationDate":"2015-11-18T01:51:12.203","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>There are various ways of specifying an end of sequence parameter. The translate demo uses a flag <code>&lt;EOS&gt;<\/code> to determine the end of sequence. However, you can also specify end of sequence by counting the number of expected words in the output. In the lines 225-227 of the <code>translate.py<\/code>:<\/p>\n\n<pre><code># If there is an EOS symbol in outputs, cut them at that point.\nif data_utils.EOS_ID in outputs:\n    outputs = outputs[:outputs.index(data_utils.EOS_ID)]\n<\/code><\/pre>\n\n<p>You can see that outputs are being cut off whenever <code>&lt;EOS&gt;<\/code> is encountered. You can easily tweak it to constrain the number of output words. You might also consider getting rid of <code>&lt;EOS&gt;<\/code> flag altogether while training, considering your application.<\/p>\n"}]}
{"QuestionId":33738440,"AnswerCount":2,"Tags":"<python><filter><convolution><tensorflow>","CreationDate":"2015-11-16T15:04:26.910","AcceptedAnswerId":33749505.0,"OwnerUserId":1111929.0,"Title":"How to interpret TensorFlow's convolution filter and striding parameters?","Body":"<p>I'm trying to understand TensorFlow's <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/nn.md#convolution\" rel=\"nofollow\">convolution<\/a>, in particular the formula <\/p>\n\n<pre><\/pre>\n\n<p>I would have expected the formula to be <\/p>\n\n<pre><\/pre>\n\n<p>instead. Starting from a 32x32 image, and applying a 5x5 filter with strides [1,3,3,1], then in my understanding this should yield a 10x10 output, whose values are the convolutions of the areas <\/p>\n\n<pre><\/pre>\n\n<p>so both dimensions should be floor((32-5)\/3)+1=10 and not floor((32-5+1)\/3)=9. What am I missing here? Have I misunderstood the way convolution is done here and\/or what the parameters mean? If so, what parameters should I use in order to obtain the above selection?<\/p>\n","answers":[{"AnswerId":"33747152","CreationDate":"2015-11-17T00:30:10.587","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You're correct - it should be:<\/p>\n\n<p>ceil(float(in_height - filter_height + 1) \/ float(strides[1]))<\/p>\n\n<p>For 32, 5, stride=3, this becomes:  ceil(9.33) = 10.<\/p>\n\n<p>Fixed and will be pushed into github soon.  Thanks for catching this!  For more info, see the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/196\" rel=\"nofollow\">github bug discussion, issue #196<\/a><\/p>\n"},{"AnswerId":"33749505","CreationDate":"2015-11-17T05:01:27.447","ParentId":null,"OwnerUserId":"4454870","Title":null,"Body":"<p>According to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/196\" rel=\"nofollow\">issue #196<\/a>, this part of the documentation is apparently wrong; and I think there is still problem in dga's answer. <\/p>\n\n<p>It should be: <\/p>\n\n<p>floor((in_height+y_padding-filter_height)\/y_stride) + 1, <\/p>\n\n<ul>\n<li>When padding=VALID, y_padding=0.<\/li>\n<li>When padding=SAME, in general y_padding should be adjusted to make (in_height+y_padding-filter_height)\/y_stride an integer so that 'floor' becomes unnecessary.<\/li>\n<\/ul>\n"}]}
{"QuestionId":33739964,"AnswerCount":0,"Tags":"<python><hdf5><caffe><h5py><hdf>","CreationDate":"2015-11-16T16:21:01.050","AcceptedAnswerId":null,"OwnerUserId":5568402.0,"Title":"Relatively large HDF5 files using h5py - sanity check","Body":"<p>I'm doing deep learning with caffe and generating my own dataset in HDF5 format. I have 131 976 images all 224x224 which come to about 480MB, and each image has a 1x6 array as a label. I've found that when I generate the .h5 files, they come to 5GB each, 125GB in total. I just want to make sure this is expected. I've checked the contents, but i don't understand how the memory requirement is 250 times bigger. All I'm doing is filling numpy arrays X and Y and creating the datasets (25 in total).<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33740516,"AnswerCount":1,"Tags":"<deep-learning><caffe><conv-neural-network>","CreationDate":"2015-11-16T16:51:03.517","AcceptedAnswerId":33743906.0,"OwnerUserId":1868436.0,"Title":"Generating \"artificial\" images for training CNN","Body":"<p>I am trying to identify different types of vehicles and logos etc. As with most efforts in this domain the challenge has been a paucity of training images.<\/p>\n\n<p>Has anyone tried generating artificial images? By \"drawing\" the logo atop (say) a vehicle hood, vehicle rear and performing distortions (e.g. change color, shape etc.) to give it a more diverse training set.<\/p>\n\n<p>Are there guidelines and\/or best practices in this approach?<\/p>\n\n<p>Edit: downvoters, at least please comment why you downvoted so I can learn. I am not trying to troll this forum and sincerely asked a question. If you disagree please share your thoughts.<\/p>\n","answers":[{"AnswerId":"33743906","CreationDate":"2015-11-16T20:19:55.787","ParentId":null,"OwnerUserId":"4253485","Title":null,"Body":"<p>Data augmentation is a well known technique to address the problem of over fitting. It is advisable that you do so by performing those distortions on the original training images. This could be done by adding noise, performing skew, rotation, cropping random windows and training on those windows. Those are just some examples.<\/p>\n\n<p>Doing this to generated images would not be the best option to try because generated images would not be as good as the available training images.<\/p>\n\n<p>Start with this <a href=\"http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow\">paper<\/a> It explains the AlexNet and mentions some of the tricks they used which includes some nice data augmentation.<\/p>\n"}]}
{"QuestionId":33743346,"AnswerCount":1,"Tags":"<ubuntu><theano>","CreationDate":"2015-11-16T19:44:42.927","AcceptedAnswerId":33743347.0,"OwnerUserId":395857.0,"Title":"How to install Theano on Ubuntu 14.04 x64, and configure it so that it uses the GPU?","Body":"<p>I tried to follow the instructions on <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html\" rel=\"nofollow\">Easy Installation of an Optimized Theano on Current Ubuntu<\/a> but it doesn't work: whenever I run a Theano script using GPU, it gives me the error message:<\/p>\n\n<blockquote>\n  <p>CUDA is installed, but device gpu is not available (error: Unable to get the number of gpus available: no CUDA-capable device is detected) <\/p>\n<\/blockquote>\n\n<hr>\n\n<p>More specifically, following the instructions in the linked webpage, I executed the following steps:<\/p>\n\n<pre><\/pre>\n\n<p>Then I rebooted and tried running:\u00a0<\/p>\n\n<pre><\/pre>\n\n<p>But I get:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33743347","CreationDate":"2015-11-16T19:44:42.927","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>(I tested the following on Ubuntu 14.04.4 LTS x64 and Kubuntu 14.04.4 LTS x64, I guess it should work on most Ubuntu variants)<\/p>\n\n<h2>Installing Theano and configuring the GPU (CUDA)<\/h2>\n\n<p>The instructions on the official website are outdated. Instead you can use the following instructions (assume a freshly installed Kubuntu 14.04 LTS x64):<\/p>\n\n<pre><code># Install Theano\nsudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git\nsudo pip install Theano\n\n# Install Nvidia drivers, CUDA and CUDA toolkit, following some instructions from http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/index.html\nwget http:\/\/developer.download.nvidia.com\/compute\/cuda\/7.5\/Prod\/local_installers\/cuda-repo-ubuntu1404-7-5-local_7.5-18_amd64.deb # Got the link at https:\/\/developer.nvidia.com\/cuda-downloads\nsudo dpkg -i cuda-repo-ubuntu1404-7-5-local_7.5-18_amd64.deb\nsudo apt-get update\nsudo apt-get install cuda\n\nsudo reboot\n<\/code><\/pre>\n\n<p>At that point, running <code>nvidia-smi<\/code> should work, but running <code>nvcc<\/code> won't work.<\/p>\n\n<pre><code># Execute in console, or (add in ~\/.bash_profile then run \"source ~\/.bash_profile\"):\nexport PATH=\/usr\/local\/cuda-7.5\/bin:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-7.5\/lib64:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>At that point, both <code>nvidia-smi<\/code> and <code>nvcc<\/code> should work.<\/p>\n\n<p>To test whether Theano is able to use the GPU:<\/p>\n\n<p>Copy-paste the following in <code>gpu_test.py<\/code>:<\/p>\n\n<pre><code># Start gpu_test.py\n# From http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#using-gpu\nfrom theano import function, config, shared, sandbox\nimport theano.tensor as T\nimport numpy\nimport time\n\nvlen = 10 * 30 * 768  # 10 x #cores x # threads per core\niters = 1000\n\nrng = numpy.random.RandomState(22)\nx = shared(numpy.asarray(rng.rand(vlen), config.floatX))\nf = function([], T.exp(x))\nprint(f.maker.fgraph.toposort())\nt0 = time.time()\nfor i in xrange(iters):\n    r = f()\nt1 = time.time()\nprint(\"Looping %d times took %f seconds\" % (iters, t1 - t0))\nprint(\"Result is %s\" % (r,))\nif numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):\n    print('Used the cpu')\nelse:\n    print('Used the gpu')\n# End gpu_test.py\n<\/code><\/pre>\n\n<p>and run it:<\/p>\n\n<pre><code>THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32' python gpu_test.py\n<\/code><\/pre>\n\n<p>which should return:<\/p>\n\n<pre><code>f@f-Aurora-R4:~$ THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32' python gpu_test.py\nUsing gpu device 0: GeForce GTX 690\n[GpuElemwise{exp,no_inplace}(&lt;CudaNdarrayType(float32, vector)&gt;), HostFromGpu(GpuElemwise{exp,no_inplace}.0)]\nLooping 1000 times took 0.658292 seconds\nResult is [ 1.23178029  1.61879349  1.52278066 ...,  2.20771813  2.29967761\n  1.62323296]\nUsed the gpu\n<\/code><\/pre>\n\n<p>To know your CUDA version:<\/p>\n\n<pre><code>\u200bnvcc -V\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre><code>username@server:~$ nvcc -V\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2015 NVIDIA Corporation\nBuilt on Tue_Aug_11_14:27:32_CDT_2015\nCuda compilation tools, release 7.5, V7.5.17\n<\/code><\/pre>\n\n<hr>\n\n<h2>Adding cuDNN<\/h2>\n\n<p>To add cuDNN (instructions from <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/cuda\/dnn.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/cuda\/dnn.html<\/a>):<\/p>\n\n<ol>\n<li>Download cuDNN from <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/rdp\/cudnn-download<\/a> (need registration, which is free)<\/li>\n<li><code>tar -xvf cudnn-7.0-linux-x64-v3.0-prod.tgz<\/code><\/li>\n<li>Do one of the following<\/li>\n<\/ol>\n\n<p>Option 1: Copy the <code>*.h<\/code> files to <code>CUDA_ROOT\/include<\/code> and the <code>*.so*<\/code> files to <code>CUDA_ROOT\/lib64<\/code> (by default, <code>CUDA_ROOT<\/code> is <code>\/usr\/local\/cuda<\/code> on Linux).<\/p>\n\n<pre><code>sudo cp cuda\/lib64\/* \/usr\/local\/cuda\/lib64\/\nsudo cp cuda\/include\/cudnn.h \/usr\/local\/cuda\/include\/\n<\/code><\/pre>\n\n<p>Option 2:<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\/home\/user\/path_to_CUDNN_folder\/lib64:$LD_LIBRARY_PATH\nexport CPATH=\/home\/user\/path_to_CUDNN_folder\/include:$CPATH\nexport LIBRARY_PATH=\/home\/user\/path_to_CUDNN_folder\/lib64:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>By default, Theano will detect if it can use cuDNN. If so, it will use it. If not, Theano optimizations will not introduce cuDNN ops. So Theano will still work if the user did not introduce them manually.<\/p>\n\n<p>To get an error if Theano can not use cuDNN, use this Theano flag: <code>optimizer_including=cudnn<\/code>.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32,optimizer_including=cudnn' python gpu_test.py\n<\/code><\/pre>\n\n<p>To know your cuDNN version:<\/p>\n\n<pre><code>cat \/usr\/local\/cuda\/include\/cudnn.h | grep CUDNN_MAJOR -A 2\n<\/code><\/pre>\n\n<hr>\n\n<h2>Adding CNMeM<\/h2>\n\n<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<pre><code># Build CNMeM without the unit tests\ngit clone https:\/\/github.com\/NVIDIA\/cnmem.git cnmem\ncd cnmem\nmkdir build\ncd build\nsudo apt-get install -y cmake\ncmake ..\nmake\n\n# Copy files to proper location\nsudo cp ..\/include\/cnmem.h \/usr\/local\/cuda\/include\nsudo cp *.so \/usr\/local\/cuda\/lib64\/\ncd ..\/..\n<\/code><\/pre>\n\n<p>To use with Theano, you need to add the <code>lib.cnmem<\/code> flag. Example:<\/p>\n\n<pre><code>THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32,lib.cnmem=0.8,optimizer_including=cudnn' python gpu_test.py\n<\/code><\/pre>\n\n<p>The first output of the script should be:<\/p>\n\n<pre><code>Using gpu device 0: GeForce GTX TITAN X (CNMeM is enabled with initial size: 80.0% of memory, cuDNN 5005)\n<\/code><\/pre>\n\n<p><code>lib.cnmem=0.8<\/code> means that it can use up to 80% of the GPU.<\/p>\n\n<p>CNMeM has been reported to give some interesting speed improvements, and is supported by Theano, Torch, and Caffee.<\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/987\" rel=\"nofollow noreferrer\">Theano - source 1<\/a>:<\/p>\n\n<blockquote>\n  <p>The speed up depend of many factor, like the shapes and the model itself. The speed up go from 0 to 2x faster.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/groups.google.com\/d\/msg\/theano-users\/nvfrmlJ96TA\/Y9NhxP7k6qgJ\" rel=\"nofollow noreferrer\">Theano - source 2<\/a>:<\/p>\n\n<blockquote>\n  <p>If you don't change the Theano flag allow_gc, you can expect 20% speed up on the GPU. In some case (small models), we saw a 50% speed up.<\/p>\n<\/blockquote>\n\n<hr>\n\n<h2>Running Theano on multiple CPU cores<\/h2>\n\n<p>As a side note, you can run Theano on multiple CPU cores with the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html\" rel=\"nofollow noreferrer\"><code>OMP_NUM_THREADS=[number_of_cpu_cores]<\/code> flag<\/a>. Example:<\/p>\n\n<pre><code>OMP_NUM_THREADS=4 python gpu_test.py \n<\/code><\/pre>\n\n<p>The script <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/misc\/check_blas.py\" rel=\"nofollow noreferrer\"><code>theano\/misc\/check_blas.py<\/code><\/a> outputs information regarding which BLAS is used:<\/p>\n\n<pre><code>cd [theano_git_directory]\nOMP_NUM_THREADS=4 python theano\/misc\/check_blas.py\n<\/code><\/pre>\n\n<hr>\n\n<p>To run Theano's test suite:<\/p>\n\n<pre><code>nosetests theano\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>sudo pip install nose-parameterized\nimport theano\ntheano.test()\n<\/code><\/pre>\n\n<p>Common issues:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/q\/38446771\/395857\">Importing theano: AttributeError: 'module' object has no attribute 'find_graphviz'<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":33743396,"AnswerCount":1,"Tags":"<python><anaconda><theano><openblas>","CreationDate":"2015-11-16T19:48:08.477","AcceptedAnswerId":33750544.0,"OwnerUserId":2147347.0,"Title":"Why OpenBLAS Lapack not Available in Theano Windows?","Body":"<p>I'm using Windows 10 64bit, and already installed theano. I've downloaded OpenBLAS from: <a href=\"http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.15\/OpenBLAS-v0.2.15-Win64-int32.zip\/download\" rel=\"nofollow noreferrer\">http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.15\/OpenBLAS-v0.2.15-Win64-int32.zip\/download<\/a> and put libopenblas.dll in  file from  as well<\/p>\n\n<p>my  file:<\/p>\n\n<pre><\/pre>\n\n<p>when I try to run <\/p>\n\n<p>I got something like this:<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand, but there is written:<\/p>\n\n<pre><\/pre>\n\n<p>Is it correct? or something is error?<\/p>\n\n<p>I've try to print ldflags using command<\/p>\n\n<pre><\/pre>\n\n<p>And I think I got the correct output<\/p>\n\n<pre><\/pre>\n\n<p>So, why when I test using check_blas.py my openblas status is not available?<\/p>\n\n<p>I'm new in Theano so I hope anybody here can help me,\nThank You<\/p>\n","answers":[{"AnswerId":"33750544","CreationDate":"2015-11-17T06:32:34.880","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I don't think Theano uses OpenBLAS directly, I think it only uses it via numpy. So you need to configure numpy to use OpenBLAS. This does not appear to be an easy task on Windows.<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/openblas-users\/EuGTc6bWhZw\" rel=\"nofollow\">See this thread for example<\/a>.<\/p>\n\n<p>Your numpy configuration is being printed by Theano's <code>check_blas.py<\/code> script. That configuration appears to be all set up for using MKL. If you have MKL installed (and appropriately licensed) then you don't need OpenBLAS.<\/p>\n\n<p>If you don't have MKL installed, despite the numpy config, an alternative to OpenBLAS is to use Anaconda's MKL extensions. This is a commercial product but a license is free to students and non-commercial researchers.<\/p>\n"}]}
{"QuestionId":33743944,"AnswerCount":1,"Tags":"<python><machine-learning><theano><deep-learning><pylearn>","CreationDate":"2015-11-16T20:22:14.207","AcceptedAnswerId":33749749.0,"OwnerUserId":2637212.0,"Title":"Pre-training ANN with RBM in pylearn2","Body":"<p>I'm trying to train a multi-layered ANN with , using pre-training with RBM.  I've slightly modified the script called <strong>run_deep_trainer<\/strong> that is contained in <em>pylearn2\\pylearn2\\scripts\\tutorials\\deep_trainer<\/em>.  I want a 4-layered net, where the first 3 are made with 500  and the last one is a  layer.<\/p>\n\n<p>Here's the script I've created:<\/p>\n\n<pre><\/pre>\n\n<p>It correctly does the unsupervised pre-training part, but there's an error in the supervised training part:<\/p>\n\n<pre><\/pre>\n\n<p>If I use the  (as model) in the last layer, which means substituting the functions  and  with  and , everything works fine.<\/p>\n\n<p>It seems that the model  doesn't return the the parameters () through the function .<\/p>\n\n<p>Does anyone know how to fix this?<\/p>\n","answers":[{"AnswerId":"33749749","CreationDate":"2015-11-17T05:23:49.360","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>The issue is because <code>SoftmaxRegressor<\/code> is a model, but <code>Softmax<\/code> is a layer for an <code>MLP<\/code>. A way to fix it would be something along the lines of<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_mlp_softmax(structure):\n    n_input, n_output = structure\n\n    layer = MLP(nvis=500, layers=[Softmax(n_classes=n_output, irange=0.02, layer_name='y')])\n\n    return layer\n<\/code><\/pre>\n\n<p>Where <code>MLP<\/code> is <code>mlp.MLP<\/code><\/p>\n"}]}
{"QuestionId":33744852,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-11-16T21:19:39.493","AcceptedAnswerId":34169463.0,"OwnerUserId":2397318.0,"Title":"require a package that has a different file name in lua","Body":"<p>I am very new to Lua, so pardon for the retarded question.<\/p>\n\n<p>Following <a href=\"https:\/\/github.com\/soumith\/cvpr2015\/blob\/master\/DQN%20Training%20iTorch.ipynb\" rel=\"nofollow\">this tutorial<\/a> I am trying to import DeepMind  into an iTorch notebook.<\/p>\n\n<p>I have cloned <a href=\"https:\/\/github.com\/soumith\/deepmind-atari\" rel=\"nofollow\">this repo<\/a> that contains a folder called  where the source code lives.<\/p>\n\n<p>I have added the path to the  folder <\/p>\n\n<pre><\/pre>\n\n<p>When I try to do <\/p>\n\n<pre><\/pre>\n\n<p>I obtain an (expected) error, since there is no file called  in the folder. The source code of such module is, in fact, contained in the file . <\/p>\n\n<p>I have seen the <a href=\"http:\/\/www.lua.org\/pil\/15.3.html\" rel=\"nofollow\">documentation<\/a> that hints at the problem between file and package name.<\/p>\n\n<p>So my question is: how can I import this module correctly? How can I let Lua know that  it should actually look for  (without hardcoding into the path I guess)? <\/p>\n","answers":[{"AnswerId":"34169463","CreationDate":"2015-12-09T01:54:38.283","ParentId":null,"OwnerUserId":"2397318","Title":null,"Body":"<p>The trick is in:<\/p>\n\n<pre><code>if not dqn then\n    require 'initenv'\nend\n<\/code><\/pre>\n\n<p>put in every file. The <code>initenv<\/code> file will then define what <code>dqn<\/code> is supposed to mean. Once the path to <code>initenv<\/code> is <code>package.path<\/code>, this seems a nice trick to overcome the limitation.<\/p>\n"}]}
{"QuestionId":33745345,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2015-11-16T21:54:18.747","AcceptedAnswerId":33747913.0,"OwnerUserId":1715157.0,"Title":"Recovering probability distribution from binary observations - what are the reasons for the defects of this implementation?","Body":"<p>I am trying to recover a probability distribution (not a probability density, any function with range in [0,1] with f(x) encoding probability of success for a observation at x). I use a hidden layer with 10 neurons and softmax. Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here are two typical results:\n<a href=\"https:\/\/i.stack.imgur.com\/rS0oV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rS0oV.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/VYaOF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VYaOF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Questions:<\/p>\n\n<p>What is the difference between doing  and applying softmax manually with minimizing cross entropy?<\/p>\n\n<p>It is typical for the model not to snap to the last period of the distribution. I've had it do so successfully only once. Perhaps it will be fixed by doing more training runs, but it doesn't look like it as the results often stabilise for the last ~20k runs. Would it most likely be improved by better selection of the optimising algorithm, by more hidden layers, or by more dimensions of the hidden layer? (partially answered by Edit)<\/p>\n\n<p>The aberrations close to x=0 are typical. What causes them?<\/p>\n\n<p><strong>Edit<\/strong>: The fit has improved a lot by doing<\/p>\n\n<pre><\/pre>\n\n<p>and changing the activations to tanh from sigmoids.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GLjzY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GLjzY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Further questions:<\/p>\n\n<p>Is it typical that a higher hidden dimension makes braking out of local minima easier?<\/p>\n\n<p>What is the approximate typical relation between the optimal dimension of hidden layers and dimension of inputs ? Linear, weaker than linear or stronger than linear?<\/p>\n","answers":[{"AnswerId":"33747913","CreationDate":"2015-11-17T02:00:06.147","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>It's over-fitting on the left and under-fitting on the right.<\/p>\n\n<p>Because of the small random biases your hidden units all get near zero activation near <code>x=0<\/code>, and because of the asymetry and large range of the x values, most of the hidden units are saturated out around <code>x = 10<\/code>.<\/p>\n\n<p>The gradients can't flow through saturated units, so they all get used up to overfit the values they can feel, near zero.<\/p>\n\n<p>I think centering the data on <code>x=0<\/code> will help.\nTry reducing the weight-initialization-variance, and\/or increasing the bias-initialization-variance (or equivalently, reducing the range of the data to a smaller region, like <code>[-1,1]<\/code>).<\/p>\n\n<p>You would get the same problem if you used RBF's and initializad them all near zero. with the linear-sigmoid units the second layer is using pairs of linear-sigmoids to make RBF's.<\/p>\n"}]}
{"QuestionId":33746071,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2015-11-16T22:45:38.550","AcceptedAnswerId":null,"OwnerUserId":5568317.0,"Title":"TensorFlow core debug; missing debug symbols","Body":"<p>I'm trying to learn TensorFlow's internals by stepping from its CIFAR-10 model training's python code into its core C++ code.\nUsing Eclipse+PyDev for step by step debugging of the python code works great, but I can't find how to step into the C++ code of the TensorFlow core. \nI tried using Eclipse CDT to build the C++ code in a separate project, and attach the debugger to the python process running  as described <a href=\"https:\/\/stackoverflow.com\/questions\/25507911\/debug-c-library-from-python-ctypes\">here<\/a>, but the symbols are never loaded and (obviously) deferred breakpoints are never hit. <\/p>\n\n<p><strong>Background and setup:<\/strong>  <\/p>\n\n<p>I'm running on Ubuntu 14.04 LTS, installed the TensorFlow code from sources as described <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"nofollow noreferrer\">here<\/a> and my CDT project uses a Makefile containing  <\/p>\n\n<p>.<\/p>\n","answers":[{"AnswerId":"33777606","CreationDate":"2015-11-18T10:30:24.047","ParentId":null,"OwnerUserId":"5568317","Title":null,"Body":"<p>TensorFlow loads a library called <code>_pywrap_tensorflow.so<\/code> that includes its C API (as defined in <code>tensorflow\/tensorflow\/core\/client\/tensor_c_api.cc<\/code> ).  <\/p>\n\n<p>In my case the library loaded during runtime was located in<br>\n<code>~\/tensorflow\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so<\/code><br>\nbut the library that was built from the local source code was located in <code>~\/.cache\/bazel\/_bazel_&lt;username&gt;\/dbb3c677efbf9967e464a5c6a1e69337\/tensorflow\/bazel-out\/local_linux-dbg\/bin\/tensorflow\/python\/_pywrap_tensorflow.so<\/code>.  <\/p>\n\n<p>Copying the locally built library over the loaded library, and attaching to the python process as defined in the question solved the problem.<\/p>\n"}]}
{"QuestionId":33746557,"AnswerCount":0,"Tags":"<opencv><ocr><tesseract><image-recognition><torch>","CreationDate":"2015-11-16T23:28:54.127","AcceptedAnswerId":null,"OwnerUserId":4618519.0,"Title":"OCR recognition improvement","Body":"<p>I\u2019m working with tesseract on some recognition on some video streams. i need assistance improving and also maybe looking at other image recognition libraries. I have a number of streams that have different elements in them, so each have to be designed for recognition differently.<\/p>\n\n<p>So the current streams are over twitch. 1 of the issues is that sometimes twitch streams in different quality - I have 720p, 480p, 360p. what i need to know is the winning team and the score. <\/p>\n\n<p>main issue atm is that tesseract cannot recognise characters or font on a image from a 360p stream. here is the sample image...<\/p>\n\n<p><a href=\"http:\/\/tinypic.com\/view.php?pic=9hm4vp&amp;s=8#.Vi95xxDhCSM\" rel=\"nofollow\">http:\/\/tinypic.com\/view.php?pic=9hm4vp&amp;s=8#.Vi95xxDhCSM<\/a><\/p>\n\n<p>And here's some more 360p quality images in a drive...<\/p>\n\n<p><a href=\"http:\/\/1drv.ms\/1M1O75J\" rel=\"nofollow\">http:\/\/1drv.ms\/1M1O75J<\/a><\/p>\n\n<p>So yeah thats my issue, mainly how to recognise the text well on a 360p image :) i have no idea of best method or libraries so any help would be great :)<\/p>\n","answers":[]}
{"QuestionId":33747423,"AnswerCount":0,"Tags":"<python><theano><recurrent-neural-network>","CreationDate":"2015-11-17T00:59:50.820","AcceptedAnswerId":null,"OwnerUserId":1474903.0,"Title":"Theano strange error on basic RNN","Body":"<p>I'm a beginner in Theano\/RNN programming and I made a first simple implementation based on the code <a href=\"https:\/\/11350770138305416713.googlegroups.com\/attach\/2628c6c56beb34bf\/rnn.py?part=0.1&amp;view=1&amp;vt=ANaJVrGnDrd7muHLLF-3RV2_0O2ZIl1FqKjvkiMLgr_gdX91v-uw-qJB9axMqCxmqq2uVVmKB9TlyVbHcA8njBSfVW-US24rv6HSdP5EuYoiYI0es6KniIo\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>in my code I create as a dataset a simple autoregressive process in which every term of the output series is the sum of the previous 4 values of the input series (normally distributed values)<\/p>\n\n<p>The full code is<\/p>\n\n<pre><\/pre>\n\n<p>there are very few modifications over the basic behavior defined in the linked page. I get a strange error which prevents the program for running further instructions after the line <\/p>\n\n<pre><\/pre>\n\n<p>the error that I get is<\/p>\n\n<blockquote>\n  <p>\/Library\/Python\/2.7\/site-packages\/theano\/scan_module\/scan_perform_ext.py:133: RuntimeWarning: numpy.ndarray size changed, may indicate binary incompatibility\n    from scan_perform.scan_perform import *\n  BLAS error: Parameter incX passed to cblas_dgemv was 0, which is invalid.<\/p>\n<\/blockquote>\n\n<p>I really don't understand how to fix this and have found nothing about this error<\/p>\n","answers":[]}
{"QuestionId":33747526,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-17T01:12:14.190","AcceptedAnswerId":33751773.0,"OwnerUserId":5545794.0,"Title":"How can I make power operation on tensor list","Body":"<p>How to make power operation on tensor list in Theano?<\/p>\n\n<pre><\/pre>\n\n<p>Where w is a list contains several TensorType (float64,4D), above codes make the error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix it?<\/p>\n","answers":[{"AnswerId":"33751773","CreationDate":"2015-11-17T07:57:31.207","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If, for example, <code>w<\/code> is a Python list of Theano symbolic matrices,<\/p>\n\n<pre><code>w = [theano.tensor.matrix() for _ in xrange(10)]\n<\/code><\/pre>\n\n<p>then one can get a Python list of symbolic expressions for the sum of the square of each matrix using a regular Python list comprehension:<\/p>\n\n<pre><code>a = [T.sum(w_i ** 2) for w_i in w]\n<\/code><\/pre>\n\n<p>If you then want the symbolic sum of the entries in <code>a<\/code>, just use<\/p>\n\n<pre><code>a_sum = T.sum(a)\n<\/code><\/pre>\n\n<p>Note that this all assumes the number of entries in <code>w<\/code> is known in advance, i.e. the size of <code>w<\/code> doesn't need to be symbolic. Theano doesn't yet fully support symbolic lists but partial support is available via <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/typed_list.html\" rel=\"nofollow\"><code>typed_list<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":33747596,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2015-11-17T01:22:50.713","AcceptedAnswerId":33750395.0,"OwnerUserId":4418769.0,"Title":"Problems implementing an XOR gate with Neural Nets in Tensorflow","Body":"<p>I want to make a trivial neural network, it should just implement the XOR gate. I am using the TensorFlow library, in python. \nFor an XOR gate, the only data I train with, is the complete truth table, that should be enough right? Over optimization is what I will expect to happen very quickly. Problem with the code is that the <em>weights<\/em> and <em>biases<\/em> do not update. Somehow it still gives me 100% accuracy with zero for the biases and weights.<\/p>\n\n<pre><\/pre>\n\n<p>I am following the tutorial from <a href=\"http:\/\/tensorflow.org\/tutorials\/mnist\/beginners\/index.md#softmax_regressions\" rel=\"noreferrer\">http:\/\/tensorflow.org\/tutorials\/mnist\/beginners\/index.md#softmax_regressions<\/a>\nand in the final for-loop I am printing the results form the matrix(as described in the link).<\/p>\n\n<p>Can anybody point out my error and what I should do to fix it?<\/p>\n","answers":[{"AnswerId":"33750395","CreationDate":"2015-11-17T06:21:18.187","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are a few issues with your program. <\/p>\n\n<p>The first issue is that the function you're learning isn't XOR - it's NOR. The lines:<\/p>\n\n<pre><code>xTrain = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])\nyTrain = np.array([[1], [0], [0], [0]])\n<\/code><\/pre>\n\n<p>...should be:<\/p>\n\n<pre><code>xTrain = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])\nyTrain = np.array([[0], [1], [1], [0]])\n<\/code><\/pre>\n\n<p>The next big issue is that the network you've designed <a href=\"https:\/\/en.wikipedia.org\/wiki\/Perceptrons_(book)#The_XOR_affair\" rel=\"noreferrer\">isn't capable of learning XOR<\/a>. You'll need to use a non-linear function (such as <code>tf.nn.relu()<\/code> and define at least one more layer to learn the XOR function. For example:<\/p>\n\n<pre><code>x = tf.placeholder(\"float\", [None, 2])\nW_hidden = tf.Variable(...)\nb_hidden = tf.Variable(...)\nhidden = tf.nn.relu(tf.matmul(x, W_hidden) + b_hidden)\n\nW_logits = tf.Variable(...)\nb_logits = tf.Variable(...)\nlogits = tf.matmul(hidden, W_logits) + b_logits\n<\/code><\/pre>\n\n<p>A further issue is that initializing the weights to zero will <a href=\"https:\/\/stats.stackexchange.com\/questions\/27112\/danger-of-setting-all-initial-weights-to-zero-in-backpropagation\">prevent your network from training<\/a>. Typically, you should initialize your weights randomly, and your biases to zero. Here's one popular way to do it:<\/p>\n\n<pre><code>HIDDEN_NODES = 2\n\nW_hidden = tf.Variable(tf.truncated_normal([2, HIDDEN_NODES], stddev=1.\/math.sqrt(2)))\nb_hidden = tf.Variable(tf.zeros([HIDDEN_NODES]))\n\nW_logits = tf.Variable(tf.truncated_normal([HIDDEN_NODES, 2], stddev=1.\/math.sqrt(HIDDEN_NODES)))\nb_logits = tf.Variable(tf.zeros([2]))\n<\/code><\/pre>\n\n<p>Putting it all together, and using TensorFlow routines for cross-entropy (with a one-hot encoding of <code>yTrain<\/code> for convenience), here's a program that learns XOR:<\/p>\n\n<pre><code>import math\nimport tensorflow as tf\nimport numpy as np\n\nHIDDEN_NODES = 10\n\nx = tf.placeholder(tf.float32, [None, 2])\nW_hidden = tf.Variable(tf.truncated_normal([2, HIDDEN_NODES], stddev=1.\/math.sqrt(2)))\nb_hidden = tf.Variable(tf.zeros([HIDDEN_NODES]))\nhidden = tf.nn.relu(tf.matmul(x, W_hidden) + b_hidden)\n\nW_logits = tf.Variable(tf.truncated_normal([HIDDEN_NODES, 2], stddev=1.\/math.sqrt(HIDDEN_NODES)))\nb_logits = tf.Variable(tf.zeros([2]))\nlogits = tf.matmul(hidden, W_logits) + b_logits\n\ny = tf.nn.softmax(logits)\n\ny_input = tf.placeholder(tf.float32, [None, 2])\n\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, y_input)\nloss = tf.reduce_mean(cross_entropy)\n\ntrain_op = tf.train.GradientDescentOptimizer(0.2).minimize(loss)\n\ninit_op = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run(init_op)\n\nxTrain = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])\nyTrain = np.array([[1, 0], [0, 1], [0, 1], [1, 0]])\n\nfor i in xrange(500):\n  _, loss_val = sess.run([train_op, loss], feed_dict={x: xTrain, y_input: yTrain})\n\n  if i % 10 == 0:\n    print \"Step:\", i, \"Current loss:\", loss_val\n    for x_input in [[0, 0], [0, 1], [1, 0], [1, 1]]:\n      print x_input, sess.run(y, feed_dict={x: [x_input]})\n<\/code><\/pre>\n\n<p>Note that this is probably not the most efficient neural network for computing XOR, so suggestions for tweaking the parameters are welcome!<\/p>\n"}]}
{"QuestionId":33747978,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-17T02:09:11.540","AcceptedAnswerId":null,"OwnerUserId":2669077.0,"Title":"Running translate on ipython kills the process","Body":"\n\n<p>I am running the translate demo on ipython using:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>The process gets killed automatically for some reason. Here is the output of the run.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>I am running it on a vagrant box running  Ubuntu 14.04 and no gpu. What might be happening here?<\/p>\n","answers":[{"AnswerId":"33748712","CreationDate":"2015-11-17T03:36:41.067","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>That's not ipython\/tensorflow specific.<\/p>\n\n<p>Kills like that can come from the linux kernel if it decides, basically, that a process is using too much memory.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/726690\/who-killed-my-process-and-why\">Who &quot;Killed&quot; my process and why?<\/a><\/p>\n"}]}
{"QuestionId":33748552,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-17T03:17:24.803","AcceptedAnswerId":33770771.0,"OwnerUserId":997378.0,"Title":"Tensorflow: How to replace a node in a calculation graph?","Body":"<p>If you have two disjoint graphs, and want to link them, turning this:<\/p>\n\n<pre><\/pre>\n\n<p>into this:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to do that? It seems like it could make construction easier in some cases. <\/p>\n\n<p>For example if you have a graph that has the input image as a , and want to optimize the input image, deep-dream style, is there a way to just replace the placeholder with a  node? Or do you have to think of that before building the graph?<\/p>\n","answers":[{"AnswerId":"52613079","CreationDate":"2018-10-02T17:03:38.003","ParentId":null,"OwnerUserId":"7050212","Title":null,"Body":"<p>Practical example: <\/p>\n\n<pre><code>import tensorflow as tf\ng1 = tf.Graph()\nwith g1.as_default():\n    # set variables\/placeholders\n    tf.placeholder(tf.int32, [], name='g1_a')\n    tf.placeholder(tf.int32, [], name='g1_b')\n\n    # example on exacting tensor by name\n    a = g1.get_tensor_by_name('g1_a:0')\n    b = g1.get_tensor_by_name('g1_b:0')\n\n    # operation ==&gt;&gt;     c = 2 * 3 = 6\n    mul_op = tf.multiply(a, b, name='g1_mul')\n    sess = tf.Session()\n    g1_mul_results = sess.run(mul_op, feed_dict={'g1_a:0': 2, 'g1_b:0': 3})\n    print('graph1 mul = ', g1_mul_results)  # output = 6\n\n    print('\\ngraph01 operations\/variables:')\n    for op in g1.get_operations():\n        print(op.name)\n\ng2 = tf.Graph()\nwith g2.as_default():\n    # set variables\/placeholders\n    tf.import_graph_def(g1.as_graph_def())\n    g2_c = tf.placeholder(tf.int32, [], name='g2_c')\n\n    # example on exacting tensor by name\n    g1_b = g2.get_tensor_by_name('import\/g1_b:0')\n    g1_mul = g2.get_tensor_by_name('import\/g1_mul:0')\n\n    # operation ==&gt;&gt;\n    b = tf.multiply(g1_b, g2_c, name='g2_var_times_g1_a')\n    f = tf.multiply(g1_mul, g1_b, name='g1_mul_times_g1_b')\n\n    print('\\ngraph01 operations\/variables:')\n    for op in g2.get_operations():\n        print(op.name)\n    sess = tf.Session()\n\n    # graph1 variable 'a' times graph2 variable 'c'(graph2)\n    ans = sess.run('g2_var_times_g1_a:0', feed_dict={'g2_c:0': 4, 'import\/g1_b:0': 5})\n    print('\\ngraph2 g2_var_times_g1_a = ', ans)  # output = 20\n\n    # graph1 mul_op (a*b) times graph1 variable 'b'\n    ans = sess.run('g1_a_times_g1_b:0',\n                   feed_dict={'import\/g1_a:0': 6, 'import\/g1_b:0': 7})\n    print('\\ngraph2 g1_mul_times_g1_b:0 = ', ans)  # output = (6*7)*7 = 294\n\n''' output\ngraph1 mul =  6\n\ngraph01 operations\/variables:\ng1_a\ng1_b\ng1_mul\n\ngraph01 operations\/variables:\nimport\/g1_a\nimport\/g1_b\nimport\/g1_mul\ng2_c\ng2_var_times_g1_a\ng1_a_times_g1_b\n\ngraph2 g2_var_times_g1_a =  20\n\ngraph2 g1_a_times_g1_b:0 =  294\n'''\n<\/code><\/pre>\n\n<p>reference <a href=\"http:\/\/www.guidetomlandai.com\/tutorials\/tensorflow\/inherit_tensors_from_another_graph\/\" rel=\"nofollow noreferrer\">LINK<\/a> <\/p>\n"},{"AnswerId":"43565044","CreationDate":"2017-04-22T21:51:52.763","ParentId":null,"OwnerUserId":"651945","Title":null,"Body":"<p>It turns out that <code>tf.train.import_meta_graph<\/code> passes all additional arguments to the underlying <code>import_scoped_meta_graph<\/code> which has the <code>input_map<\/code> argument and utilizes it when it gets to it's own (internal) invocation of <code>import_graph_def<\/code>. <\/p>\n\n<p>It is not documented, and took me waaaay toooo much time to find it, but it works!<\/p>\n"},{"AnswerId":"33770771","CreationDate":"2015-11-18T02:24:34.747","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TL;DR: If you can define the two computations as Python functions, you should do that. If you can't, there's more advanced functionality in TensorFlow to serialize and import graphs, which allows you to compose graphs from different sources.<\/p>\n\n<p>One way to do this in TensorFlow is to build the disjoint computations as separate <code>tf.Graph<\/code> objects, then convert them to serialized protocol buffers using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#as_graph_def\" rel=\"noreferrer\"><code>Graph.as_graph_def()<\/code><\/a>:<\/p>\n\n<pre><code>with tf.Graph().as_default() as g_1:\n  input = tf.placeholder(tf.float32, name=\"input\")\n  y = f(input)\n  # NOTE: using identity to get a known name for the output tensor.\n  output = tf.identity(y, name=\"output\")\n\ngdef_1 = g_1.as_graph_def()\n\nwith tf.Graph().as_default() as g_2:  # NOTE: g_2 not g_1       \n  input = tf.placeholder(tf.float32, name=\"input\")\n  z = g(input)\n  output = tf.identity(y, name=\"output\")\n\ngdef_2 = g_2.as_graph_def()\n<\/code><\/pre>\n\n<p>Then you could compose <code>gdef_1<\/code> and <code>gdef_2<\/code> into a third graph, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/import_graph_def\" rel=\"noreferrer\"><code>tf.import_graph_def()<\/code><\/a>:<\/p>\n\n<pre><code>with tf.Graph().as_default() as g_combined:\n  x = tf.placeholder(tf.float32, name=\"\")\n\n  # Import gdef_1, which performs f(x).\n  # \"input:0\" and \"output:0\" are the names of tensors in gdef_1.\n  y, = tf.import_graph_def(gdef_1, input_map={\"input:0\": x},\n                           return_elements=[\"output:0\"])\n\n  # Import gdef_2, which performs g(y)\n  z, = tf.import_graph_def(gdef_2, input_map={\"input:0\": y},\n                           return_elements=[\"output:0\"]\n<\/code><\/pre>\n"},{"AnswerId":"39838757","CreationDate":"2016-10-03T19:10:41.163","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>If you want to combine trained models (for example to reuse parts of a pretrained model in a new model), you can use a <code>Saver<\/code> to save a checkpoint of the first model, then restore that model (entirely or partially) into another model.<\/p>\n\n<p>For example, say you want to reuse model 1's weights <code>w<\/code> in model 2, and also convert <code>x<\/code> from a placeholder to a variable:<\/p>\n\n<pre><code>with tf.Graph().as_default() as g1:\n    x = tf.placeholder('float')\n    w = tf.Variable(1., name=\"w\")\n    y = x * w\n    saver = tf.train.Saver()\n\nwith tf.Session(graph=g1) as sess:\n    w.initializer.run()\n    # train...\n    saver.save(sess, \"my_model1.ckpt\")\n\nwith tf.Graph().as_default() as g2:\n    x = tf.Variable(2., name=\"v\")\n    w = tf.Variable(0., name=\"w\")\n    z = x + w\n    restorer = tf.train.Saver([w]) # only restore w\n\nwith tf.Session(graph=g2) as sess:\n    x.initializer.run()  # x now needs to be initialized\n    restorer.restore(sess, \"my_model1.ckpt\") # restores w=1\n    print(z.eval())  # prints 3.\n<\/code><\/pre>\n"}]}
{"QuestionId":33752772,"AnswerCount":3,"Tags":"<python-2.7><raspberry-pi2><tensorflow>","CreationDate":"2015-11-17T08:59:24.443","AcceptedAnswerId":33757698.0,"OwnerUserId":871742.0,"Title":"Tensorflow on Raspberry Pi","Body":"<p>I'm trying to install  on . The OS is ,  and  version is <\/p>\n\n<p>When I run this<\/p>\n\n<pre><\/pre>\n\n<p>it tells that <\/p>\n\n<blockquote>\n  <p>\"tensorflow-0.5.0... is not a supported wheel on this platform.\"<\/p>\n<\/blockquote>\n\n<p>Is it possible to install  on ?\nAnd someone knows what I'm missing?<\/p>\n","answers":[{"AnswerId":"36024132","CreationDate":"2016-03-15T23:23:49.453","ParentId":null,"OwnerUserId":"278046","Title":null,"Body":"<p>It is now possible to install Tensorflow on Raspberry Pi. There are detailed build instructions at <a href=\"https:\/\/github.com\/samjabrahams\/tensorflow-on-raspberry-pi\">https:\/\/github.com\/samjabrahams\/tensorflow-on-raspberry-pi<\/a> as well as an unofficial downloadable build.<\/p>\n"},{"AnswerId":"33754980","CreationDate":"2015-11-17T10:47:23.977","ParentId":null,"OwnerUserId":"5571565","Title":null,"Body":"<p>x86 is for intel-like cpu and 64 is for 64 bits that's not Raspberry pi's hardware...<\/p>\n"},{"AnswerId":"33757698","CreationDate":"2015-11-17T13:04:43.623","ParentId":null,"OwnerUserId":"4418769","Title":null,"Body":"<p>Raspberry Pi is a 32-bit architecture, and the TensorFlow program was made for a 64-bit. \nI am not sure how well A Raspberry Pi can handle it, but you could install a virtual machine and run the TensorFlow inside there.<\/p>\n\n<p>Maybe the newer generations of the Rasberry Pi could handle this a bit better, but the original ones, will probably struggle to run this setup. <\/p>\n\n<p>You could do the training on a more powerful machine and just run the learned machine on the Rasberry Pi, that might be feasible in terms of computing power.<\/p>\n"}]}
{"QuestionId":33753251,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2015-11-17T09:25:35.270","AcceptedAnswerId":33756072.0,"OwnerUserId":867505.0,"Title":"Using squared difference of two images as loss function in tensorflow","Body":"\n\n<p>I'm trying to use the SSD between two images as loss function for my network.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Problem is, that the weights then diverge and I get the error<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Why's that? I did try some other stuff like normalizing the ssd by the image size (did not work) or cropping the output values to 1 (does not crash anymore, but I still need to evaluate this):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Are my observations to be expected?<\/p>\n\n<p>Edit:\n@mdaoust suggestions proved to be correct. The main point was normalizing by batch size. This can be done independent of batch size by using this code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>With this change, only a slight decrease of the learning rate (to 0.0001) was necessary.<\/p>\n","answers":[{"AnswerId":"33756072","CreationDate":"2015-11-17T11:44:07.147","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>There are a lot of ways you can end up with non-finite results.<\/p>\n\n<p>But optimizers, especially simple ones like gradient descent, can diverge if the learning rate is 'too high'.<\/p>\n\n<p>Have you tried simply dividing your learning rate by 10\/100\/1000? Or normalizing by <code>pixels*batch_size<\/code> to get the average error per pixel?<\/p>\n\n<p>Or one of the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.md#AUTOGENERATED-processing-gradients-before-applying-them.\" rel=\"nofollow\">more advanced optimizers<\/a>? For example <code>tf.train.AdamOptimizer()<\/code> with default options.<\/p>\n"}]}
{"QuestionId":33758669,"AnswerCount":3,"Tags":"<python><parallel-processing><python-multiprocessing><tensorflow>","CreationDate":"2015-11-17T13:52:07.407","AcceptedAnswerId":null,"OwnerUserId":1936768.0,"Title":"Running multiple tensorflow sessions concurrently","Body":"<p>I am trying to run several sessions of TensorFlow concurrently on a CentOS 7 machine with 64 CPUs. My colleague reports that he can use the following two blocks of code to produce a parallel speedup on his machine using 4 cores:<\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>In particular, he says that he observes <\/p>\n\n<pre><\/pre>\n\n<p>However, when I run the code:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, I get a significant slowdown by using multiprocessing whereas my colleague does not. Does anyone have any insight as to why this could be occurring and what can be done to fix it?<\/p>\n\n<p>EDIT<\/p>\n\n<p>Here is some example output. Notice that loading the data seems to occur in parallel, but training the individual models has a very sequential look in the output (and which can be verified by looking at CPU usage in  as the program executes)<\/p>\n\n<pre><\/pre>\n\n<p>Another EDIT<\/p>\n\n<p>Suppose we wish to confirm that the issue is seemingly with TensorFlow and not with multiprocessing. I replaced the contents of  with a big loop as follows:<\/p>\n\n<pre><\/pre>\n\n<p>For output:<\/p>\n\n<pre><\/pre>\n\n<p>Hence I think the problem here is not with multiprocessing itself.<\/p>\n","answers":[{"AnswerId":"34827660","CreationDate":"2016-01-16T13:43:28.420","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>From comment by OP (<a href=\"https:\/\/stackoverflow.com\/users\/1936768\/user1936768\">user1936768<\/a>):<\/p>\n\n<p>I have good news: It turns out, on my system at least, my trial programs didn't execute long enough for the other instances of TF to start up. When I put a longer running example program in main, I do indeed see concurrent computations <\/p>\n"},{"AnswerId":"33764538","CreationDate":"2015-11-17T18:37:38.263","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><s>One possibility is that your sessions are trying to use 64 cores each and stomping on each other\nPerhaps try setting <code>NUM_CORES<\/code> to a lower value for each session\n<\/s><\/p>\n\n<pre><code>sess = tf.Session(\n    tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES,\n                   intra_op_parallelism_threads=NUM_CORES))\n<\/code><\/pre>\n"},{"AnswerId":"54797096","CreationDate":"2019-02-20T23:54:11.500","ParentId":null,"OwnerUserId":"7858504","Title":null,"Body":"<p>This can be done elegantly with <a href=\"https:\/\/github.com\/ray-project\/ray\" rel=\"nofollow noreferrer\">Ray<\/a>, which is a library for parallel and distributed Python, which lets you train your models in parallel from a single Python script.<\/p>\n\n<p>This has the advantage of letting you parallelize \"classes\" by turning them in to \"actors\", which can be hard to do with regular Python multiprocessing. This is important because the expensive part of often initializing the TensorFlow graph. If you create an actor and then call the <code>train<\/code> method multiple times, the cost of initializing the graph is amortized.<\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow.examples.tutorials.mnist import input_data\nfrom PIL import Image\nimport ray\nimport tensorflow as tf\nimport time\n\n\n@ray.remote\nclass TrainingActor(object):\n    def __init__(self, seed):\n        print('Set new seed:', seed)\n        np.random.seed(seed)\n        tf.set_random_seed(seed)\n        self.mnist = input_data.read_data_sets('MNIST_data\/', one_hot=True)\n\n        # Setting up the softmax architecture.\n        self.x = tf.placeholder('float', [None, 784])\n        W = tf.Variable(tf.zeros([784, 10]))\n        b = tf.Variable(tf.zeros([10]))\n        self.y = tf.nn.softmax(tf.matmul(self.x, W) + b)\n\n        # Setting up the cost function.\n        self.y_ = tf.placeholder('float', [None, 10])\n        cross_entropy = -tf.reduce_sum(self.y_*tf.log(self.y))\n        self.train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)\n\n        # Initialization\n        self.init = tf.initialize_all_variables()\n        self.sess = tf.Session(\n            config=tf.ConfigProto(\n                inter_op_parallelism_threads=1,\n                intra_op_parallelism_threads=1\n            )\n        )\n\n    def train(self):\n        self.sess.run(self.init)\n\n        for i in range(1000):\n            batch_xs, batch_ys = self.mnist.train.next_batch(100)\n            self.sess.run(self.train_step, feed_dict={self.x: batch_xs, self.y_: batch_ys})\n\n        correct_prediction = tf.equal(tf.argmax(self.y, 1), tf.argmax(self.y_, 1))\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))\n\n        return self.sess.run(accuracy, feed_dict={self.x: self.mnist.test.images,\n                                                  self.y_: self.mnist.test.labels})\n\n\nif __name__ == '__main__':\n    # Start Ray.\n    ray.init()\n\n    # Create 3 actors.\n    training_actors = [TrainingActor.remote(seed) for seed in range(3)]\n\n    # Make them all train in parallel.\n    accuracy_ids = [actor.train.remote() for actor in training_actors]\n    print(ray.get(accuracy_ids))\n\n    # Start new training runs in parallel.\n    accuracy_ids = [actor.train.remote() for actor in training_actors]\n    print(ray.get(accuracy_ids))\n<\/code><\/pre>\n\n<p>If you only want to create one copy of the dataset instead of having each actor read the dataset, you can rewrite things as follows. Under the hood, this uses the <a href=\"https:\/\/arrow.apache.org\/docs\/python\/plasma.html\" rel=\"nofollow noreferrer\">Plasma shared memory object store<\/a> and the <a href=\"https:\/\/arrow.apache.org\/docs\/index.html\" rel=\"nofollow noreferrer\">Apache Arrow data format<\/a>.<\/p>\n\n<pre><code>@ray.remote\nclass TrainingActor(object):\n    def __init__(self, mnist, seed):\n        self.mnist = mnist\n        ...\n\n    ...\n\nif __name__ == \"__main__\":\n    ray.init()\n\n    # Read the mnist dataset and put it into shared memory once\n    # so that workers don't create their own copies.\n    mnist = input_data.read_data_sets('MNIST_data\/', one_hot=True)\n    mnist_id = ray.put(mnist)\n\n    training_actors = [TrainingActor.remote(mnist_id, seed) for seed in range(3)]\n<\/code><\/pre>\n\n<p>You can see more in the <a href=\"https:\/\/ray.readthedocs.io\/en\/latest\/?badge=latest\" rel=\"nofollow noreferrer\">Ray documentation<\/a>. Note I'm one of the Ray developers.<\/p>\n"}]}
{"QuestionId":33759623,"AnswerCount":25,"Tags":"<python><tensorflow><machine-learning><model>","CreationDate":"2015-11-17T14:37:26.367","AcceptedAnswerId":50852627.0,"OwnerUserId":1978504.0,"Title":"Tensorflow: how to save\/restore a model?","Body":"<p>After you train a model in Tensorflow: <\/p>\n\n<ol>\n<li>How do you save the trained model?<\/li>\n<li>How do you later restore this saved model?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"43784657","CreationDate":"2017-05-04T13:45:31.493","ParentId":null,"OwnerUserId":"5243083","Title":null,"Body":"<p>I am improving my answer to add more details for saving and restoring models. <\/p>\n\n<p>In(and after) <strong>Tensorflow version 0.11<\/strong>: <\/p>\n\n<p><strong>Save the model:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\n\n#Prepare to feed input, i.e. feed_dict and placeholders\nw1 = tf.placeholder(\"float\", name=\"w1\")\nw2 = tf.placeholder(\"float\", name=\"w2\")\nb1= tf.Variable(2.0,name=\"bias\")\nfeed_dict ={w1:4,w2:8}\n\n#Define a test operation that we will restore\nw3 = tf.add(w1,w2)\nw4 = tf.multiply(w3,b1,name=\"op_to_restore\")\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\n#Create a saver object which will save all the variables\nsaver = tf.train.Saver()\n\n#Run the operation by feeding input\nprint sess.run(w4,feed_dict)\n#Prints 24 which is sum of (w1+w2)*b1 \n\n#Now, save the graph\nsaver.save(sess, 'my_test_model',global_step=1000)\n<\/code><\/pre>\n\n<p><strong>Restore the model:<\/strong> <\/p>\n\n<pre><code>import tensorflow as tf\n\nsess=tf.Session()    \n#First let's load meta graph and restore weights\nsaver = tf.train.import_meta_graph('my_test_model-1000.meta')\nsaver.restore(sess,tf.train.latest_checkpoint('.\/'))\n\n\n# Access saved Variables directly\nprint(sess.run('bias:0'))\n# This will print 2, which is the value of bias that we saved\n\n\n# Now, let's access and create placeholders variables and\n# create feed-dict to feed new data\n\ngraph = tf.get_default_graph()\nw1 = graph.get_tensor_by_name(\"w1:0\")\nw2 = graph.get_tensor_by_name(\"w2:0\")\nfeed_dict ={w1:13.0,w2:17.0}\n\n#Now, access the op that you want to run. \nop_to_restore = graph.get_tensor_by_name(\"op_to_restore:0\")\n\nprint sess.run(op_to_restore,feed_dict)\n#This will print 60 which is calculated \n<\/code><\/pre>\n\n<p>This and some more advanced use-cases have been explained very well here.<\/p>\n\n<p><a href=\"http:\/\/cv-tricks.com\/tensorflow-tutorial\/save-restore-tensorflow-models-quick-complete-tutorial\/\" rel=\"noreferrer\">A quick complete tutorial to save and restore Tensorflow models<\/a><\/p>\n"},{"AnswerId":"44087746","CreationDate":"2017-05-20T15:35:09.650","ParentId":null,"OwnerUserId":"523503","Title":null,"Body":"<p>If you use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/monitored_session.py\" rel=\"noreferrer\">tf.train.MonitoredTrainingSession<\/a> as the default session, you don't need to add extra code to do save\/restore things. Just pass a checkpoint dir name to MonitoredTrainingSession's constructor, it will use session hooks to handle these. <\/p>\n"},{"AnswerId":"33762168","CreationDate":"2015-11-17T16:30:03.613","ParentId":null,"OwnerUserId":"998293","Title":null,"Body":"<p>For TensorFlow version &lt; 0.11.0RC1:<\/p>\n\n<p>The checkpoints that are saved contain values for the <code>Variable<\/code>s in your model, not the model\/graph itself, which means that the graph should be the same when you restore the checkpoint.<\/p>\n\n<p>Here's an example for a linear regression where there's a training loop that saves variable checkpoints and an evaluation section that will restore variables saved in a prior run and compute predictions. Of course, you can also restore variables and continue training if you'd like.<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32)\ny = tf.placeholder(tf.float32)\n\nw = tf.Variable(tf.zeros([1, 1], dtype=tf.float32))\nb = tf.Variable(tf.ones([1, 1], dtype=tf.float32))\ny_hat = tf.add(b, tf.matmul(x, w))\n\n...more setup for optimization and what not...\n\nsaver = tf.train.Saver()  # defaults to saving all variables - in this case w and b\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    if FLAGS.train:\n        for i in xrange(FLAGS.training_steps):\n            ...training loop...\n            if (i + 1) % FLAGS.checkpoint_steps == 0:\n                saver.save(sess, FLAGS.checkpoint_dir + 'model.ckpt',\n                           global_step=i+1)\n    else:\n        # Here's where you're restoring the variables w and b.\n        # Note that the graph is exactly as it was when the variables were\n        # saved in a prior training run.\n        ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)\n        if ckpt and ckpt.model_checkpoint_path:\n            saver.restore(sess, ckpt.model_checkpoint_path)\n        else:\n            ...no checkpoint found...\n\n        # Now you can run the model to get predictions\n        batch_x = ...load some data...\n        predictions = sess.run(y_hat, feed_dict={x: batch_x})\n<\/code><\/pre>\n\n<p>Here are the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"noreferrer\">docs<\/a> for <code>Variable<\/code>s, which cover saving and restoring. And here are the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"noreferrer\">docs<\/a> for the <code>Saver<\/code>.<\/p>\n"},{"AnswerId":"53827972","CreationDate":"2018-12-18T07:11:30.810","ParentId":null,"OwnerUserId":"7750274","Title":null,"Body":"<p>Wherever you want to save the model,<\/p>\n\n<pre><code>self.saver = tf.train.Saver()\nwith tf.Session() as sess:\n            sess.run(tf.global_variables_initializer())\n            ...\n            self.saver.save(sess, filename)\n<\/code><\/pre>\n\n<p>Make sure, all your <code>tf.Variable<\/code> have names, because you may want to restore them later using their names.\nAnd where you want to predict,<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph(filename)\nname = 'name given when you saved the file' \nwith tf.Session() as sess:\n      saver.restore(sess, name)\n      print(sess.run('W1:0')) #example to retrieve by variable name\n<\/code><\/pre>\n\n<p>Make sure that saver runs inside the corresponding session. \nRemember that, if you use the <code>tf.train.latest_checkpoint('.\/')<\/code>, then only the latest check point will be used. <\/p>\n"},{"AnswerId":"50852627","CreationDate":"2018-06-14T08:10:10.717","ParentId":null,"OwnerUserId":"3867406","Title":null,"Body":"<h1>Docs<\/h1>\n\n<p>They built an exhaustive and useful tutorial -> <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/guide\/saved_model<\/a><\/p>\n\n<p>From the docs:<\/p>\n\n<h3>Save<\/h3>\n\n<pre><code># Create some variables.\nv1 = tf.get_variable(\"v1\", shape=[3], initializer = tf.zeros_initializer)\nv2 = tf.get_variable(\"v2\", shape=[5], initializer = tf.zeros_initializer)\n\ninc_v1 = v1.assign(v1+1)\ndec_v2 = v2.assign(v2-1)\n\n# Add an op to initialize the variables.\ninit_op = tf.global_variables_initializer()\n\n# Add ops to save and restore all the variables.\nsaver = tf.train.Saver()\n\n# Later, launch the model, initialize the variables, do some work, and save the\n# variables to disk.\nwith tf.Session() as sess:\n  sess.run(init_op)\n  # Do some work with the model.\n  inc_v1.op.run()\n  dec_v2.op.run()\n  # Save the variables to disk.\n  save_path = saver.save(sess, \"\/tmp\/model.ckpt\")\n  print(\"Model saved in path: %s\" % save_path)\n<\/code><\/pre>\n\n<h3>Restore<\/h3>\n\n<pre><code>tf.reset_default_graph()\n\n# Create some variables.\nv1 = tf.get_variable(\"v1\", shape=[3])\nv2 = tf.get_variable(\"v2\", shape=[5])\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  # Check the values of the variables\n  print(\"v1 : %s\" % v1.eval())\n  print(\"v2 : %s\" % v2.eval())\n<\/code><\/pre>\n\n<h1>Tensorflow 2<\/h1>\n\n<p>This is still beta so I'd advise against for now. If you still want to go down that road here is the <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/saved_model\" rel=\"noreferrer\"><code>tf.saved_model<\/code> usage guide<\/a><\/p>\n\n<h1>Tensorflow &lt; 2<\/h1>\n\n<h2><code>simple_save<\/code><\/h2>\n\n<p>Many good answer, for completeness I'll add my 2 cents: <strong><a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/saved_model\" rel=\"noreferrer\">simple_save<\/a><\/strong>. Also a standalone code example using the <code>tf.data.Dataset<\/code> API.<\/p>\n\n<p>Python 3 ; Tensorflow <strong>1.14<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.saved_model import tag_constants\n\nwith tf.Graph().as_default():\n    with tf.Session() as sess:\n        ...\n\n        # Saving\n        inputs = {\n            \"batch_size_placeholder\": batch_size_placeholder,\n            \"features_placeholder\": features_placeholder,\n            \"labels_placeholder\": labels_placeholder,\n        }\n        outputs = {\"prediction\": model_output}\n        tf.saved_model.simple_save(\n            sess, 'path\/to\/your\/location\/', inputs, outputs\n        )\n<\/code><\/pre>\n\n<p>Restoring:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith restored_graph.as_default():\n    with tf.Session() as sess:\n        tf.saved_model.loader.load(\n            sess,\n            [tag_constants.SERVING],\n            'path\/to\/your\/location\/',\n        )\n        batch_size_placeholder = graph.get_tensor_by_name('batch_size_placeholder:0')\n        features_placeholder = graph.get_tensor_by_name('features_placeholder:0')\n        labels_placeholder = graph.get_tensor_by_name('labels_placeholder:0')\n        prediction = restored_graph.get_tensor_by_name('dense\/BiasAdd:0')\n\n        sess.run(prediction, feed_dict={\n            batch_size_placeholder: some_value,\n            features_placeholder: some_other_value,\n            labels_placeholder: another_value\n        })\n<\/code><\/pre>\n\n<h1>Standalone example<\/h1>\n\n<p><strong><a href=\"http:\/\/vict0rsch.github.io\/2018\/05\/17\/restore-tf-model-dataset\/\" rel=\"noreferrer\">Original blog post<\/a><\/strong><\/p>\n\n<p>The following code generates random data for the sake of the demonstration.<\/p>\n\n<ol>\n<li>We start by creating the placeholders. They will hold the data at runtime. From them, we create the <code>Dataset<\/code> and then its <code>Iterator<\/code>. We get the iterator's generated tensor, called <code>input_tensor<\/code> which will serve as input to our model.<\/li>\n<li>The model itself is built from <code>input_tensor<\/code>: a GRU-based bidirectional RNN followed by a dense classifier. Because why not.<\/li>\n<li>The loss is a <code>softmax_cross_entropy_with_logits<\/code>, optimized with <code>Adam<\/code>. After 2 epochs (of 2 batches each), we save the \"trained\" model with <code>tf.saved_model.simple_save<\/code>. If you run the code as is, then the model will be saved in a folder called <code>simple\/<\/code> in your current working directory.<\/li>\n<li>In a new graph, we then restore the saved model with <code>tf.saved_model.loader.load<\/code>. We grab the placeholders and logits with <code>graph.get_tensor_by_name<\/code> and the <code>Iterator<\/code> initializing operation with <code>graph.get_operation_by_name<\/code>.<\/li>\n<li>Lastly we run an inference for both batches in the dataset, and check that the saved and restored model both yield the same values. They do!<\/li>\n<\/ol>\n\n<p>Code:<\/p>\n\n<pre><code>import os\nimport shutil\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.python.saved_model import tag_constants\n\n\ndef model(graph, input_tensor):\n    \"\"\"Create the model which consists of\n    a bidirectional rnn (GRU(10)) followed by a dense classifier\n\n    Args:\n        graph (tf.Graph): Tensors' graph\n        input_tensor (tf.Tensor): Tensor fed as input to the model\n\n    Returns:\n        tf.Tensor: the model's output layer Tensor\n    \"\"\"\n    cell = tf.nn.rnn_cell.GRUCell(10)\n    with graph.as_default():\n        ((fw_outputs, bw_outputs), (fw_state, bw_state)) = tf.nn.bidirectional_dynamic_rnn(\n            cell_fw=cell,\n            cell_bw=cell,\n            inputs=input_tensor,\n            sequence_length=[10] * 32,\n            dtype=tf.float32,\n            swap_memory=True,\n            scope=None)\n        outputs = tf.concat((fw_outputs, bw_outputs), 2)\n        mean = tf.reduce_mean(outputs, axis=1)\n        dense = tf.layers.dense(mean, 5, activation=None)\n\n        return dense\n\n\ndef get_opt_op(graph, logits, labels_tensor):\n    \"\"\"Create optimization operation from model's logits and labels\n\n    Args:\n        graph (tf.Graph): Tensors' graph\n        logits (tf.Tensor): The model's output without activation\n        labels_tensor (tf.Tensor): Target labels\n\n    Returns:\n        tf.Operation: the operation performing a stem of Adam optimizer\n    \"\"\"\n    with graph.as_default():\n        with tf.variable_scope('loss'):\n            loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(\n                    logits=logits, labels=labels_tensor, name='xent'),\n                    name=\"mean-xent\"\n                    )\n        with tf.variable_scope('optimizer'):\n            opt_op = tf.train.AdamOptimizer(1e-2).minimize(loss)\n        return opt_op\n\n\nif __name__ == '__main__':\n    # Set random seed for reproducibility\n    # and create synthetic data\n    np.random.seed(0)\n    features = np.random.randn(64, 10, 30)\n    labels = np.eye(5)[np.random.randint(0, 5, (64,))]\n\n    graph1 = tf.Graph()\n    with graph1.as_default():\n        # Random seed for reproducibility\n        tf.set_random_seed(0)\n        # Placeholders\n        batch_size_ph = tf.placeholder(tf.int64, name='batch_size_ph')\n        features_data_ph = tf.placeholder(tf.float32, [None, None, 30], 'features_data_ph')\n        labels_data_ph = tf.placeholder(tf.int32, [None, 5], 'labels_data_ph')\n        # Dataset\n        dataset = tf.data.Dataset.from_tensor_slices((features_data_ph, labels_data_ph))\n        dataset = dataset.batch(batch_size_ph)\n        iterator = tf.data.Iterator.from_structure(dataset.output_types, dataset.output_shapes)\n        dataset_init_op = iterator.make_initializer(dataset, name='dataset_init')\n        input_tensor, labels_tensor = iterator.get_next()\n\n        # Model\n        logits = model(graph1, input_tensor)\n        # Optimization\n        opt_op = get_opt_op(graph1, logits, labels_tensor)\n\n        with tf.Session(graph=graph1) as sess:\n            # Initialize variables\n            tf.global_variables_initializer().run(session=sess)\n            for epoch in range(3):\n                batch = 0\n                # Initialize dataset (could feed epochs in Dataset.repeat(epochs))\n                sess.run(\n                    dataset_init_op,\n                    feed_dict={\n                        features_data_ph: features,\n                        labels_data_ph: labels,\n                        batch_size_ph: 32\n                    })\n                values = []\n                while True:\n                    try:\n                        if epoch &lt; 2:\n                            # Training\n                            _, value = sess.run([opt_op, logits])\n                            print('Epoch {}, batch {} | Sample value: {}'.format(epoch, batch, value[0]))\n                            batch += 1\n                        else:\n                            # Final inference\n                            values.append(sess.run(logits))\n                            print('Epoch {}, batch {} | Final inference | Sample value: {}'.format(epoch, batch, values[-1][0]))\n                            batch += 1\n                    except tf.errors.OutOfRangeError:\n                        break\n            # Save model state\n            print('\\nSaving...')\n            cwd = os.getcwd()\n            path = os.path.join(cwd, 'simple')\n            shutil.rmtree(path, ignore_errors=True)\n            inputs_dict = {\n                \"batch_size_ph\": batch_size_ph,\n                \"features_data_ph\": features_data_ph,\n                \"labels_data_ph\": labels_data_ph\n            }\n            outputs_dict = {\n                \"logits\": logits\n            }\n            tf.saved_model.simple_save(\n                sess, path, inputs_dict, outputs_dict\n            )\n            print('Ok')\n    # Restoring\n    graph2 = tf.Graph()\n    with graph2.as_default():\n        with tf.Session(graph=graph2) as sess:\n            # Restore saved values\n            print('\\nRestoring...')\n            tf.saved_model.loader.load(\n                sess,\n                [tag_constants.SERVING],\n                path\n            )\n            print('Ok')\n            # Get restored placeholders\n            labels_data_ph = graph2.get_tensor_by_name('labels_data_ph:0')\n            features_data_ph = graph2.get_tensor_by_name('features_data_ph:0')\n            batch_size_ph = graph2.get_tensor_by_name('batch_size_ph:0')\n            # Get restored model output\n            restored_logits = graph2.get_tensor_by_name('dense\/BiasAdd:0')\n            # Get dataset initializing operation\n            dataset_init_op = graph2.get_operation_by_name('dataset_init')\n\n            # Initialize restored dataset\n            sess.run(\n                dataset_init_op,\n                feed_dict={\n                    features_data_ph: features,\n                    labels_data_ph: labels,\n                    batch_size_ph: 32\n                }\n\n            )\n            # Compute inference for both batches in dataset\n            restored_values = []\n            for i in range(2):\n                restored_values.append(sess.run(restored_logits))\n                print('Restored values: ', restored_values[i][0])\n\n    # Check if original inference and restored inference are equal\n    valid = all((v == rv).all() for v, rv in zip(values, restored_values))\n    print('\\nInferences match: ', valid)\n<\/code><\/pre>\n\n<p>This will print:<\/p>\n\n<pre><code>$ python3 save_and_restore.py\n\nEpoch 0, batch 0 | Sample value: [-0.13851789 -0.3087595   0.12804556  0.20013677 -0.08229901]\nEpoch 0, batch 1 | Sample value: [-0.00555491 -0.04339041 -0.05111827 -0.2480045  -0.00107776]\nEpoch 1, batch 0 | Sample value: [-0.19321944 -0.2104792  -0.00602257  0.07465433  0.11674127]\nEpoch 1, batch 1 | Sample value: [-0.05275984  0.05981954 -0.15913513 -0.3244143   0.10673307]\nEpoch 2, batch 0 | Final inference | Sample value: [-0.26331693 -0.13013336 -0.12553    -0.04276478  0.2933622 ]\nEpoch 2, batch 1 | Final inference | Sample value: [-0.07730117  0.11119192 -0.20817074 -0.35660955  0.16990358]\n\nSaving...\nINFO:tensorflow:Assets added to graph.\nINFO:tensorflow:No assets to write.\nINFO:tensorflow:SavedModel written to: b'\/some\/path\/simple\/saved_model.pb'\nOk\n\nRestoring...\nINFO:tensorflow:Restoring parameters from b'\/some\/path\/simple\/variables\/variables'\nOk\nRestored values:  [-0.26331693 -0.13013336 -0.12553    -0.04276478  0.2933622 ]\nRestored values:  [-0.07730117  0.11119192 -0.20817074 -0.35660955  0.16990358]\n\nInferences match:  True\n<\/code><\/pre>\n"},{"AnswerId":"53183223","CreationDate":"2018-11-07T03:31:18.600","ParentId":null,"OwnerUserId":"7745839","Title":null,"Body":"<p>Use tf.train.Saver to save a model, remerber, you need to specify the var_list, if you want to reduce the model size. The val_list can be tf.trainable_variables or tf.global_variables.<\/p>\n"},{"AnswerId":"56116641","CreationDate":"2019-05-13T16:45:14.330","ParentId":null,"OwnerUserId":"10035385","Title":null,"Body":"<p>I'm on Version:<\/p>\n\n<pre><code>tensorflow (1.13.1)\ntensorflow-gpu (1.13.1)\n<\/code><\/pre>\n\n<p>Simple way is<\/p>\n\n<p>Save:<\/p>\n\n<pre><code>model.save(\"model.h5\")\n<\/code><\/pre>\n\n<p>Restore:<\/p>\n\n<pre><code>model = tf.keras.models.load_model(\"model.h5\")\n<\/code><\/pre>\n"},{"AnswerId":"43526252","CreationDate":"2017-04-20T17:34:18.510","ParentId":null,"OwnerUserId":"1076564","Title":null,"Body":"<p>Here's my simple solution for the two basic cases differing on whether you want to load the graph from file or build it during runtime.<\/p>\n\n<p>This answer holds for Tensorflow 0.12+ (including 1.0).<\/p>\n\n<h2>Rebuilding the graph in code<\/h2>\n\n<h3>Saving<\/h3>\n\n<pre><code>graph = ... # build the graph\nsaver = tf.train.Saver()  # create the saver after the graph\nwith ... as sess:  # your session object\n    saver.save(sess, 'my-model')\n<\/code><\/pre>\n\n<h3>Loading<\/h3>\n\n<pre><code>graph = ... # build the graph\nsaver = tf.train.Saver()  # create the saver after the graph\nwith ... as sess:  # your session object\n    saver.restore(sess, tf.train.latest_checkpoint('.\/'))\n    # now you can use the graph, continue training or whatever\n<\/code><\/pre>\n\n<h2>Loading also the graph from a file<\/h2>\n\n<p><strong>When using this technique, make sure all your layers\/variables have explicitly set unique names.<\/strong> Otherwise Tensorflow will make the names unique itself and they'll be thus different from the names stored in the file. It's not a problem in the previous technique, because the names are \"mangled\" the same way in both loading and saving.<\/p>\n\n<h3>Saving<\/h3>\n\n<pre><code>graph = ... # build the graph\n\nfor op in [ ... ]:  # operators you want to use after restoring the model\n    tf.add_to_collection('ops_to_restore', op)\n\nsaver = tf.train.Saver()  # create the saver after the graph\nwith ... as sess:  # your session object\n    saver.save(sess, 'my-model')\n<\/code><\/pre>\n\n<h3>Loading<\/h3>\n\n<pre><code>with ... as sess:  # your session object\n    saver = tf.train.import_meta_graph('my-model.meta')\n    saver.restore(sess, tf.train.latest_checkpoint('.\/'))\n    ops = tf.get_collection('ops_to_restore')  # here are your operators in the same order in which you saved them to the collection\n<\/code><\/pre>\n"},{"AnswerId":"43333803","CreationDate":"2017-04-10T22:22:47.870","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>In most cases, saving and restoring from disk using a <code>tf.train.Saver<\/code> is your best option:<\/p>\n\n<pre><code>... # build your model\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess:\n    ... # train the model\n    saver.save(sess, \"\/tmp\/my_great_model\")\n\nwith tf.Session() as sess:\n    saver.restore(sess, \"\/tmp\/my_great_model\")\n    ... # use the model\n<\/code><\/pre>\n\n<p>You can also save\/restore the graph structure itself (see the <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"noreferrer\">MetaGraph documentation<\/a> for details). By default, the <code>Saver<\/code> saves the graph structure into a <code>.meta<\/code> file. You can call <code>import_meta_graph()<\/code> to restore it. It restores the graph structure and returns a <code>Saver<\/code> that you can use to restore the model's state:<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph(\"\/tmp\/my_great_model.meta\")\n\nwith tf.Session() as sess:\n    saver.restore(sess, \"\/tmp\/my_great_model\")\n    ... # use the model\n<\/code><\/pre>\n\n<p>However, there are cases where you need something much faster. For example, if you implement early stopping, you want to save checkpoints every time the model improves during training (as measured on the validation set), then if there is no progress for some time, you want to roll back to the best model.  If you save the model to disk every time it improves, it will tremendously slow down training. The trick is to save the variable states to <em>memory<\/em>, then just restore them later:<\/p>\n\n<pre><code>... # build your model\n\n# get a handle on the graph nodes we need to save\/restore the model\ngraph = tf.get_default_graph()\ngvars = graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)\nassign_ops = [graph.get_operation_by_name(v.op.name + \"\/Assign\") for v in gvars]\ninit_values = [assign_op.inputs[1] for assign_op in assign_ops]\n\nwith tf.Session() as sess:\n    ... # train the model\n\n    # when needed, save the model state to memory\n    gvars_state = sess.run(gvars)\n\n    # when needed, restore the model state\n    feed_dict = {init_value: val\n                 for init_value, val in zip(init_values, gvars_state)}\n    sess.run(assign_ops, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>A quick explanation: when you create a variable <code>X<\/code>, TensorFlow automatically creates an assignment operation <code>X\/Assign<\/code> to set the variable's initial value. Instead of creating placeholders and extra assignment ops (which would just make the graph messy), we just use these existing assignment ops. The first input of each assignment op is a reference to the variable it is supposed to initialize, and the second input (<code>assign_op.inputs[1]<\/code>) is the initial value. So in order to set any value we want (instead of the initial value), we need to use a <code>feed_dict<\/code> and replace the initial value. Yes, TensorFlow lets you feed a value for any op, not just for placeholders, so this works fine.<\/p>\n"},{"AnswerId":"54306505","CreationDate":"2019-01-22T10:44:50.017","ParentId":null,"OwnerUserId":"8079690","Title":null,"Body":"<p>You can <strong>save the variables in the network<\/strong> using<\/p>\n\n<pre><code>saver = tf.train.Saver() \nsaver.save(sess, 'path of save\/fileName.ckpt')\n<\/code><\/pre>\n\n<p>To <strong>restore the network<\/strong> for reuse later or in another script, use:<\/p>\n\n<pre><code>saver = tf.train.Saver()\nsaver.restore(sess, tf.train.latest_checkpoint('path of save\/')\nsess.run(....) \n<\/code><\/pre>\n\n<p>Important points:<\/p>\n\n<ol>\n<li><code>sess<\/code> must be same between first and later runs (coherent structure). <\/li>\n<li><code>saver.restore<\/code> needs the path of the folder of the saved files, not an individual file path. <\/li>\n<\/ol>\n"},{"AnswerId":"53577920","CreationDate":"2018-12-02T06:18:57.453","ParentId":null,"OwnerUserId":"1462770","Title":null,"Body":"<p>According to the new Tensorflow version, <code>tf.train.Checkpoint<\/code> is the preferable way of saving and restoring a model:<\/p>\n\n<blockquote>\n  <p><code>Checkpoint.save<\/code> and <code>Checkpoint.restore<\/code> write and read object-based\n  checkpoints, in contrast to tf.train.Saver which writes and reads\n  variable.name based checkpoints. Object-based checkpointing saves a\n  graph of dependencies between Python objects (Layers, Optimizers,\n  Variables, etc.) with named edges, and this graph is used to match\n  variables when restoring a checkpoint. It can be more robust to\n  changes in the Python program, and helps to support restore-on-create\n  for variables when executing eagerly. <strong>Prefer <code>tf.train.Checkpoint<\/code> over\n  <code>tf.train.Saver<\/code> for new code<\/strong>.<\/p>\n<\/blockquote>\n\n<p>Here is an example:<\/p>\n\n<pre><code>import tensorflow as tf\nimport os\n\ntf.enable_eager_execution()\n\ncheckpoint_directory = \"\/tmp\/training_checkpoints\"\ncheckpoint_prefix = os.path.join(checkpoint_directory, \"ckpt\")\n\ncheckpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)\nstatus = checkpoint.restore(tf.train.latest_checkpoint(checkpoint_directory))\nfor _ in range(num_training_steps):\n  optimizer.minimize( ... )  # Variables will be restored on creation.\nstatus.assert_consumed()  # Optional sanity checks.\ncheckpoint.save(file_prefix=checkpoint_prefix)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Checkpoint\" rel=\"noreferrer\">More information and example here.<\/a> <\/p>\n"},{"AnswerId":"33763208","CreationDate":"2015-11-17T17:22:19.267","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>There are two parts to the model, the model definition, saved by <code>Supervisor<\/code> as <code>graph.pbtxt<\/code> in the model directory and the numerical values of tensors, saved into checkpoint files like <code>model.ckpt-1003418<\/code>.<\/p>\n\n<p>The model definition can be restored using <code>tf.import_graph_def<\/code>, and the weights are restored using <code>Saver<\/code>.<\/p>\n\n<p>However, <code>Saver<\/code> uses special collection holding list of variables that's attached to the model Graph, and this collection is not initialized using import_graph_def, so you can't use the two together at the moment (it's on our roadmap to fix). For now, you have to use approach of Ryan Sepassi -- manually construct a graph with identical node names, and use <code>Saver<\/code> to load the weights into it.<\/p>\n\n<p>(Alternatively you could hack it by using by using <code>import_graph_def<\/code>, creating variables manually, and using <code>tf.add_to_collection(tf.GraphKeys.VARIABLES, variable)<\/code> for each variable, then using <code>Saver<\/code>)<\/p>\n"},{"AnswerId":"55770331","CreationDate":"2019-04-20T04:45:07.200","ParentId":null,"OwnerUserId":"1587329","Title":null,"Body":"<p>For <strong>tensorflow 2.0<\/strong>, it is <a href=\"https:\/\/www.tensorflow.org\/alpha\/guide\/keras\/saving_and_serializing\" rel=\"noreferrer\">as simple as<\/a> <\/p>\n\n<blockquote>\n<pre><code># Save the model\nmodel.save('path_to_my_model.h5')\n<\/code><\/pre>\n<\/blockquote>\n\n<p>To restore:<\/p>\n\n<pre><code>new_model = tensorflow.keras.models.load_model('path_to_my_model.h5')\n<\/code><\/pre>\n"},{"AnswerId":"58978357","CreationDate":"2019-11-21T15:27:51.337","ParentId":null,"OwnerUserId":"7032887","Title":null,"Body":"<p>For tensorflow-2.0<\/p>\n\n<p>it's very simple.<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<h3>SAVE<\/h3>\n\n<pre><code>model.save(\"model_name\")\n<\/code><\/pre>\n\n<h3>RESTORE<\/h3>\n\n<pre><code>model = tf.keras.models.load_model('model_name')\n<\/code><\/pre>\n"},{"AnswerId":"34609566","CreationDate":"2016-01-05T10:35:14.897","ParentId":null,"OwnerUserId":"521933","Title":null,"Body":"<p>As Yaroslav said, you can hack restoring from a graph_def and checkpoint by importing the graph, manually creating variables, and then using a Saver.<\/p>\n\n<p>I implemented this for my personal use, so I though I'd share the code here.<\/p>\n\n<p>Link: <a href=\"https:\/\/gist.github.com\/nikitakit\/6ef3b72be67b86cb7868\">https:\/\/gist.github.com\/nikitakit\/6ef3b72be67b86cb7868<\/a><\/p>\n\n<p>(This is, of course, a hack, and there is no guarantee that models saved this way will remain readable in future versions of TensorFlow.)<\/p>\n"},{"AnswerId":"41999075","CreationDate":"2017-02-02T09:57:52.050","ParentId":null,"OwnerUserId":"7505159","Title":null,"Body":"<p>As described in issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6255\" rel=\"noreferrer\">6255<\/a>:<\/p>\n\n<pre><code>use '**.\/**model_name.ckpt'\nsaver.restore(sess,'.\/my_model_final.ckpt')\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>saver.restore('my_model_final.ckpt')\n<\/code><\/pre>\n"},{"AnswerId":"42076582","CreationDate":"2017-02-06T20:21:06.847","ParentId":null,"OwnerUserId":"4420961","Title":null,"Body":"<p>You can also take this easier way.<\/p>\n\n<h3>Step 1: initialize all your variables<\/h3>\n\n<pre><code>W1 = tf.Variable(tf.truncated_normal([6, 6, 1, K], stddev=0.1), name=\"W1\")\nB1 = tf.Variable(tf.constant(0.1, tf.float32, [K]), name=\"B1\")\n\nSimilarly, W2, B2, W3, .....\n<\/code><\/pre>\n\n<h3>Step 2: save the session inside model <code>Saver<\/code> and save it<\/h3>\n\n<pre><code>model_saver = tf.train.Saver()\n\n# Train the model and save it in the end\nmodel_saver.save(session, \"saved_models\/CNN_New.ckpt\")\n<\/code><\/pre>\n\n<h3>Step 3: restore the model<\/h3>\n\n<pre><code>with tf.Session(graph=graph_cnn) as session:\n    model_saver.restore(session, \"saved_models\/CNN_New.ckpt\")\n    print(\"Model restored.\") \n    print('Initialized')\n<\/code><\/pre>\n\n<h3>Step 4: check your variable<\/h3>\n\n<pre><code>W1 = session.run(W1)\nprint(W1)\n<\/code><\/pre>\n\n<hr>\n\n<p>While running in different python instance, use<\/p>\n\n<pre><code>with tf.Session() as sess:\n    # Restore latest checkpoint\n    saver.restore(sess, tf.train.latest_checkpoint('saved_model\/.'))\n\n    # Initalize the variables\n    sess.run(tf.global_variables_initializer())\n\n    # Get default graph (supply your custom graph if you have one)\n    graph = tf.get_default_graph()\n\n    # It will give tensor object\n    W1 = graph.get_tensor_by_name('W1:0')\n\n    # To get the value (numpy array)\n    W1_value = session.run(W1)\n<\/code><\/pre>\n"},{"AnswerId":"38179139","CreationDate":"2016-07-04T07:32:49.983","ParentId":null,"OwnerUserId":"586275","Title":null,"Body":"<p>If it is an internally saved model, you just specify a restorer for all variables as<\/p>\n\n<pre><code>restorer = tf.train.Saver(tf.all_variables())\n<\/code><\/pre>\n\n<p>and use it to restore variables in a current session:<\/p>\n\n<pre><code>restorer.restore(self._sess, model_file)\n<\/code><\/pre>\n\n<p>For the external model you need to specify the mapping from the its variable names to your variable names. You can view the model variable names using the command <\/p>\n\n<pre><code>python \/path\/to\/tensorflow\/tensorflow\/python\/tools\/inspect_checkpoint.py --file_name=\/path\/to\/pretrained_model\/model.ckpt\n<\/code><\/pre>\n\n<p>The inspect_checkpoint.py script can be found in '.\/tensorflow\/python\/tools' folder of the Tensorflow source.<\/p>\n\n<p>To specify the mapping, you can use my <a href=\"https:\/\/github.com\/sdemyanov\/tensorflow-worklab\/\">Tensorflow-Worklab<\/a>, which contains a set of classes and scripts to train and retrain different models. It includes an example of retraining ResNet models, located <a href=\"https:\/\/raw.githubusercontent.com\/ry\/tensorflow-resnet\/master\/data\/tensorflow-resnet-pretrained-20160509.tar.gz.torrent\">here<\/a> <\/p>\n"},{"AnswerId":"58073719","CreationDate":"2019-09-24T05:41:06.607","ParentId":null,"OwnerUserId":"9936228","Title":null,"Body":"<h1>tf.keras Model saving with <code>TF2.0<\/code><\/h1>\n\n<p>I see great answers for saving models using TF1.x. I want to provide couple of more pointers in saving <code>tensorflow.keras<\/code> models which is a little complicated as there are many ways to save a model. <\/p>\n\n<p>Here I am providing an example of saving a <code>tensorflow.keras<\/code> model to <code>model_path<\/code> folder under current directory. This works well with most recent tensorflow (TF2.0). I will update this description if there is any change in near future.<\/p>\n\n<h1>Saving and loading entire model<\/h1>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nmnist = tf.keras.datasets.mnist\n\n#import data\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\n# create a model\ndef create_model():\n  model = tf.keras.models.Sequential([\n    tf.keras.layers.Flatten(input_shape=(28, 28)),\n    tf.keras.layers.Dense(512, activation=tf.nn.relu),\n    tf.keras.layers.Dropout(0.2),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n    ])\n# compile the model\n  model.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n  return model\n\n# Create a basic model instance\nmodel=create_model()\n\nmodel.fit(x_train, y_train, epochs=1)\nloss, acc = model.evaluate(x_test, y_test,verbose=1)\nprint(\"Original model, accuracy: {:5.2f}%\".format(100*acc))\n\n# Save entire model to a HDF5 file\nmodel.save('.\/model_path\/my_model.h5')\n\n# Recreate the exact same model, including weights and optimizer.\nnew_model = keras.models.load_model('.\/model_path\/my_model.h5')\nloss, acc = new_model.evaluate(x_test, y_test)\nprint(\"Restored model, accuracy: {:5.2f}%\".format(100*acc))\n<\/code><\/pre>\n\n<h1>Saving and loading model Weights only<\/h1>\n\n<p>If you are interested in saving model weights only and then load weights to restore the model, then<\/p>\n\n<pre><code>model.fit(x_train, y_train, epochs=5)\nloss, acc = model.evaluate(x_test, y_test,verbose=1)\nprint(\"Original model, accuracy: {:5.2f}%\".format(100*acc))\n\n# Save the weights\nmodel.save_weights('.\/checkpoints\/my_checkpoint')\n\n# Restore the weights\nmodel = create_model()\nmodel.load_weights('.\/checkpoints\/my_checkpoint')\n\nloss,acc = model.evaluate(x_test, y_test)\nprint(\"Restored model, accuracy: {:5.2f}%\".format(100*acc))\n<\/code><\/pre>\n\n<h1>Saving and restoring using keras checkpoint callback<\/h1>\n\n<pre><code># include the epoch in the file name. (uses `str.format`)\ncheckpoint_path = \"training_2\/cp-{epoch:04d}.ckpt\"\ncheckpoint_dir = os.path.dirname(checkpoint_path)\n\ncp_callback = tf.keras.callbacks.ModelCheckpoint(\n    checkpoint_path, verbose=1, save_weights_only=True,\n    # Save weights, every 5-epochs.\n    period=5)\n\nmodel = create_model()\nmodel.save_weights(checkpoint_path.format(epoch=0))\nmodel.fit(train_images, train_labels,\n          epochs = 50, callbacks = [cp_callback],\n          validation_data = (test_images,test_labels),\n          verbose=0)\n\nlatest = tf.train.latest_checkpoint(checkpoint_dir)\n\nnew_model = create_model()\nnew_model.load_weights(latest)\nloss, acc = new_model.evaluate(test_images, test_labels)\nprint(\"Restored model, accuracy: {:5.2f}%\".format(100*acc))\n<\/code><\/pre>\n\n<h1>saving model with custom metrics<\/h1>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\n# Custom Loss1 (for example) \n@tf.function() \ndef customLoss1(yTrue,yPred):\n  return tf.reduce_mean(yTrue-yPred) \n\n# Custom Loss2 (for example) \n@tf.function() \ndef customLoss2(yTrue, yPred):\n  return tf.reduce_mean(tf.square(tf.subtract(yTrue,yPred))) \n\ndef create_model():\n  model = tf.keras.models.Sequential([\n    tf.keras.layers.Flatten(input_shape=(28, 28)),\n    tf.keras.layers.Dense(512, activation=tf.nn.relu),  \n    tf.keras.layers.Dropout(0.2),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n    ])\n  model.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy', customLoss1, customLoss2])\n  return model\n\n# Create a basic model instance\nmodel=create_model()\n\n# Fit and evaluate model \nmodel.fit(x_train, y_train, epochs=1)\nloss, acc,loss1, loss2 = model.evaluate(x_test, y_test,verbose=1)\nprint(\"Original model, accuracy: {:5.2f}%\".format(100*acc))\n\nmodel.save(\".\/model.h5\")\n\nnew_model=tf.keras.models.load_model(\".\/model.h5\",custom_objects={'customLoss1':customLoss1,'customLoss2':customLoss2})\n<\/code><\/pre>\n\n<h1>Saving keras model with custom ops<\/h1>\n\n<p>When we have custom ops as in the following case (<code>tf.tile<\/code>), we need to create a function and wrap with a Lambda layer. Otherwise, model cannot be saved. <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras.layers import Input, Lambda\nfrom tensorflow.keras import Model\n\ndef my_fun(a):\n  out = tf.tile(a, (1, tf.shape(a)[0]))\n  return out\n\na = Input(shape=(10,))\n#out = tf.tile(a, (1, tf.shape(a)[0]))\nout = Lambda(lambda x : my_fun(x))(a)\nmodel = Model(a, out)\n\nx = np.zeros((50,10), dtype=np.float32)\nprint(model(x).numpy())\n\nmodel.save('my_model.h5')\n\n#load the model\nnew_model=tf.keras.models.load_model(\"my_model.h5\")\n<\/code><\/pre>\n\n<p>I think I have covered a few of the many ways of saving tf.keras model. However, there are many other ways. Please comment below if you see your use case is not covered above. Thanks!<\/p>\n"},{"AnswerId":"47235448","CreationDate":"2017-11-11T07:19:47.207","ParentId":null,"OwnerUserId":"7093436","Title":null,"Body":"<p>My environment: Python 3.6, Tensorflow 1.3.0<\/p>\n\n<p>Though there have been many solutions, most of them is based on <code>tf.train.Saver<\/code>. When we load a <code>.ckpt<\/code> saved by <code>Saver<\/code>, we have to either redefine the tensorflow network or use some weird and hard-remembered name, e.g. <code>'placehold_0:0'<\/code>,<code>'dense\/Adam\/Weight:0'<\/code>. Here I recommend to use <code>tf.saved_model<\/code>, one simplest example given below, your can learn more from <a href=\"https:\/\/www.tensorflow.org\/serving\/serving_basic\" rel=\"noreferrer\">Serving a TensorFlow Model<\/a>:<\/p>\n\n<p><strong>Save the model:<\/strong><\/p>\n\n\n\n<pre><code>import tensorflow as tf\n\n# define the tensorflow network and do some trains\nx = tf.placeholder(\"float\", name=\"x\")\nw = tf.Variable(2.0, name=\"w\")\nb = tf.Variable(0.0, name=\"bias\")\n\nh = tf.multiply(x, w)\ny = tf.add(h, b, name=\"y\")\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\n# save the model\nexport_path =  '.\/savedmodel'\nbuilder = tf.saved_model.builder.SavedModelBuilder(export_path)\n\ntensor_info_x = tf.saved_model.utils.build_tensor_info(x)\ntensor_info_y = tf.saved_model.utils.build_tensor_info(y)\n\nprediction_signature = (\n  tf.saved_model.signature_def_utils.build_signature_def(\n      inputs={'x_input': tensor_info_x},\n      outputs={'y_output': tensor_info_y},\n      method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))\n\nbuilder.add_meta_graph_and_variables(\n  sess, [tf.saved_model.tag_constants.SERVING],\n  signature_def_map={\n      tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:\n          prediction_signature \n  },\n  )\nbuilder.save()\n<\/code><\/pre>\n\n<p><strong>Load the model:<\/strong><\/p>\n\n\n\n<pre><code>import tensorflow as tf\nsess=tf.Session() \nsignature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY\ninput_key = 'x_input'\noutput_key = 'y_output'\n\nexport_path =  '.\/savedmodel'\nmeta_graph_def = tf.saved_model.loader.load(\n           sess,\n          [tf.saved_model.tag_constants.SERVING],\n          export_path)\nsignature = meta_graph_def.signature_def\n\nx_tensor_name = signature[signature_key].inputs[input_key].name\ny_tensor_name = signature[signature_key].outputs[output_key].name\n\nx = sess.graph.get_tensor_by_name(x_tensor_name)\ny = sess.graph.get_tensor_by_name(y_tensor_name)\n\ny_out = sess.run(y, {x: 3.0})\n<\/code><\/pre>\n"},{"AnswerId":"44769893","CreationDate":"2017-06-26T23:03:01.793","ParentId":null,"OwnerUserId":"7806129","Title":null,"Body":"<p>All the answers here are great, but I want to add two things.<\/p>\n\n<p>First, to elaborate on @user7505159's answer, the \".\/\" can be important to add to the beginning of the file name that you are restoring.<\/p>\n\n<p>For example, you can save a graph with no \".\/\" in the file name like so:<\/p>\n\n<pre><code># Some graph defined up here with specific names\n\nsaver = tf.train.Saver()\nsave_file = 'model.ckpt'\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    saver.save(sess, save_file)\n<\/code><\/pre>\n\n<p>But in order to restore the graph, you may need to prepend a \".\/\" to the file_name:<\/p>\n\n<pre><code># Same graph defined up here\n\nsaver = tf.train.Saver()\nsave_file = '.\/' + 'model.ckpt' # String addition used for emphasis\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    saver.restore(sess, save_file)\n<\/code><\/pre>\n\n<p>You will not always need the \".\/\", but it can cause problems depending on your environment and version of TensorFlow.<\/p>\n\n<p>It also want to mention that the <code>sess.run(tf.global_variables_initializer())<\/code> can be important before restoring the session.<\/p>\n\n<p>If you are receiving an error regarding uninitialized variables when trying to restore a saved session, make sure you include <code>sess.run(tf.global_variables_initializer())<\/code> before the <code>saver.restore(sess, save_file)<\/code> line. It can save you a headache.<\/p>\n"},{"AnswerId":"35447265","CreationDate":"2016-02-17T03:21:03.720","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>You can also check out <a href=\"https:\/\/github.com\/tensorflow\/skflow\/tree\/master\/examples\" rel=\"noreferrer\">examples<\/a> in <a href=\"https:\/\/github.com\/tensorflow\/skflow\" rel=\"noreferrer\">TensorFlow\/skflow<\/a>, which offers <code>save<\/code> and <code>restore<\/code> methods that can help you easily manage your models. It has parameters that you can also control how frequently you want to back up your model. <\/p>\n"},{"AnswerId":"58057293","CreationDate":"2019-09-23T07:10:04.453","ParentId":null,"OwnerUserId":"9857120","Title":null,"Body":"<p>In the new version of tensorflow 2.0, the process of saving\/loading a model is a lot easier. Because of the Implementation of the Keras API, a high-level API for TensorFlow.<\/p>\n\n<p>To save a model:\nCheck the documentation for reference:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/models\/save_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/models\/save_model<\/a><\/p>\n\n<pre><code>tf.keras.models.save_model(model_name, filepath, save_format)\n<\/code><\/pre>\n\n<p>To load a model:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/models\/load_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/models\/load_model<\/a><\/p>\n\n<pre><code>model = tf.keras.models.load_model(filepath)\n<\/code><\/pre>\n"},{"AnswerId":"40765759","CreationDate":"2016-11-23T13:24:26.667","ParentId":null,"OwnerUserId":"6808091","Title":null,"Body":"<p>In (and after) TensorFlow version 0.11.0RC1, you can save and restore your model directly by calling <code>tf.train.export_meta_graph<\/code> and <code>tf.train.import_meta_graph<\/code> according to <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph<\/a>.<\/p>\n\n<h3>Save the model<\/h3>\n\n<pre><code>w1 = tf.Variable(tf.truncated_normal(shape=[10]), name='w1')\nw2 = tf.Variable(tf.truncated_normal(shape=[20]), name='w2')\ntf.add_to_collection('vars', w1)\ntf.add_to_collection('vars', w2)\nsaver = tf.train.Saver()\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsaver.save(sess, 'my-model')\n# `save` method will call `export_meta_graph` implicitly.\n# you will get saved graph files:my-model.meta\n<\/code><\/pre>\n\n<h3>Restore the model<\/h3>\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('.\/'))\nall_vars = tf.get_collection('vars')\nfor v in all_vars:\n    v_ = sess.run(v)\n    print(v_)\n<\/code><\/pre>\n"},{"AnswerId":"58326742","CreationDate":"2019-10-10T15:49:26.533","ParentId":null,"OwnerUserId":"9957560","Title":null,"Body":"<p>Following @Vishnuvardhan Janapati 's answer, here is another way to save and reload model with <strong>custom layer\/metric\/loss<\/strong> under <strong>TensorFlow 2.0.0<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Layer\nfrom tensorflow.keras.utils.generic_utils import get_custom_objects\n\n# custom loss (for example)  \ndef custom_loss(y_true,y_pred):\n  return tf.reduce_mean(y_true - y_pred)\nget_custom_objects().update({'custom_loss': custom_loss}) \n\n# custom loss (for example) \nclass CustomLayer(Layer):\n  def __init__(self, ...):\n      ...\n  # define custom layer and all necessary custom operations inside custom layer\n\nget_custom_objects().update({'CustomLayer': CustomLayer})  \n<\/code><\/pre>\n\n<p>In this way, once you have executed such codes, and saved your model with <code>tf.keras.models.save_model<\/code> or <code>model.save<\/code> or <code>ModelCheckpoint<\/code> callback, you can re-load your model without the need of precise custom objects, as simple as<\/p>\n\n<pre><code>new_model = tf.keras.models.load_model(\".\/model.h5\"})\n<\/code><\/pre>\n"}]}
{"QuestionId":33761643,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2015-11-17T16:07:17.930","AcceptedAnswerId":33763851.0,"OwnerUserId":5389199.0,"Title":"Caffe accuracy bigger than 100%","Body":"<p>I'm building one but, and when I use the custom train function provided on <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow\">lenet example<\/a> with a batch size bigger than 110 my accuracy gets bigger than 1 (100%).<\/p>\n\n<p>If I use batch size 32, I get 30 percent of accuracy. Batch size equal 64 my net accuracy is 64. And batch size equal to 128, the accuracy is 1.2. <\/p>\n\n<p>My images are 32x32.\nTrain dataset: 56 images of Neutral faces. 60 images of Surprise faces. Test dataset: 15 images of Neutral faces. 15 images of Surprise faces. <\/p>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>So, what is wrong with my code?<\/p>\n","answers":[{"AnswerId":"33763851","CreationDate":"2015-11-17T17:58:00.560","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>In your testing code you run 100 iterations (<code>for test_it in range(100)<\/code>), on each iteration you compute <code>correct<\/code> as number of examples in a <strong>batch<\/strong> that are correct. You then divide that number by 1e4.<\/p>\n\n<p>Let's assume your model is very good and has almost 100% prediction rate. Then with batch size of 32 on each of 100 iterations you will add 32 to <code>correct<\/code>, yielding 3200. You then divide it by 1e4, ending up with 0.32, which is almost consistent with what you see (your number is slightly less because sometimes your model does mispredict the target).<\/p>\n\n<p>To fix it, you can replace<\/p>\n\n<pre><code>test_acc[it \/\/ test_interval] = correct \/ 1e4\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>test_acc[it \/\/ test_interval] = correct \/ (100.0 * batch_size)\n<\/code><\/pre>\n"}]}
{"QuestionId":33762831,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><deep-learning><tensorflow>","CreationDate":"2015-11-17T17:01:44.717","AcceptedAnswerId":33764034.0,"OwnerUserId":1782011.0,"Title":"Error when building seq2seq model with tensorflow","Body":"<p>I'm trying to understand the seq2seq models defined in seq2seq.py in tensorflow. I use bits of code I copy from the translate.py example that comes with tensorflow. I keep getting the same error and really do not understand where it comes from.<\/p>\n\n<p>A minimal code example to reproduce the error:<\/p>\n\n<pre><\/pre>\n\n<p>The error I get when evaluating the last line (I evaluated it interactively in the python interpreter):<\/p>\n\n<pre><\/pre>\n\n<p>I suspect the error comes from my side :)\nOn a sidenote. The documentation and the tutorials are really great but the example code for the sequence to sequence model (the english to french translation example) is quite dense. You also have to jump a lot between files to understand what's going on. Me at least got lost several times in the code.<\/p>\n\n<p>A minimal example (perhaps on some toy data) of constructing and training a basic seq2seq model would really be helpful here. Somebody know if this already exist somewhere?<\/p>\n\n<p><strong>EDIT<\/strong>\nI have fixed the code above according @Ishamael suggestions (meaning, no errors returns) (see below), but there are still some things not clear in this fixed version. My input is a sequence of vectors of length 2 of real valued values. And my output is a sequence of binary vectors of length 22. Should my tf.placeholder code not look like the following? <strong>(EDIT yes)<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I also had to change tf.int32 to tf.float32 above. Since my output is binary. Should this not be tf.int32 for the tf.placeholder of my decoder? But tensorflow complains again if I do this. I'm not sure what the reasoning is behind this.<\/p>\n\n<p>The size of my hidden layer is 512 here.<\/p>\n\n<p>the complete fixed code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33763850","CreationDate":"2015-11-17T17:58:00.467","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>There is a self-test method in the translate module that shows its minimal usage.<a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/translate\/translate.py#123\" rel=\"nofollow\">[here]<\/a><\/p>\n\n<p>I just ran the self-test method using.<\/p>\n\n<pre><code>python translate.py --self_test 1\n<\/code><\/pre>\n"},{"AnswerId":"33764034","CreationDate":"2015-11-17T18:08:26.750","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Most of the models (seq2seq is not an exception) expect their input to be in batches, so if the shape of your logical input is <code>[n]<\/code>, then a shape of a tensor you will be using as an input to your model should be <code>[batch_size x n]<\/code>. In practice the first dimension of the shape is usually left out as <code>None<\/code> and inferred to be the batch size at runtime.<\/p>\n\n<p>Since the logical input to seq2seq is a vector of numbers, the actual tensor shape should be <code>[None, input_sequence_length]<\/code>. So fixed code would look along the lines of:<\/p>\n\n\n\n<pre><code>input_sequence_length = 2; # the length of one vector in your input sequence\n\nfor i in xrange(350):  \n    encoder_inputs.append(tf.placeholder(tf.int32, shape=[None, input_sequence_length],\n                                              name=\"encoder{0}\".format(i)))\n<\/code><\/pre>\n\n<p>(and then the same for the decoder)<\/p>\n"}]}
{"QuestionId":33765029,"AnswerCount":2,"Tags":"<python><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2015-11-17T19:05:35.140","AcceptedAnswerId":33765506.0,"OwnerUserId":3009514.0,"Title":"Caffe, how to predict from a pretrained net","Body":"<p>I'm using this code to load my net:<\/p>\n\n<pre><\/pre>\n\n<p>I have doubts on three lines.<\/p>\n\n<p>1- <\/p>\n\n<p>What is mean? Should I use this mean value or another? And if yes, where can I get custom mean value? I'm using a custom dataset.<\/p>\n\n<p>2- <\/p>\n\n<p>What channel_swap means? And again, should I use this value or an custom value?<\/p>\n\n<p>And the last<\/p>\n\n<p>3-   <\/p>\n\n<p>What is raw_scale? And what value should I use?<\/p>\n\n<p>I'm using <a href=\"http:\/\/www.pitt.edu\/~emotion\/ck-spread.htm\" rel=\"nofollow\">Cohn Kanade<\/a> dataset. All images are 64x64 and in grayscale.  <\/p>\n","answers":[{"AnswerId":"42569401","CreationDate":"2017-03-03T02:10:51.573","ParentId":null,"OwnerUserId":"6264122","Title":null,"Body":"<p>I agree on comments of @Ishamael on channel_swap and mean. I just wanted to add further clarification on raw_scale. Assuming that images are loaded with <code>caffe.io.load_image<\/code>, values are always in the range of 0 to 1 [1]. Just to note that: <\/p>\n\n<blockquote>\n  <p>While Python represents images in [0, 1], certain Caffe models\n  CaffeNet and AlexNet represent images in [0, 255] so the raw_scale\n  these models must be 255.<\/p>\n<\/blockquote>\n\n<p>And I think it's wise to check the input image values prior to feeding to the data layer of network in order to choose appropriate raw_scale.<\/p>\n\n<p>Thank you.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L224\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L224<\/a><\/p>\n"},{"AnswerId":"33765506","CreationDate":"2015-11-17T19:34:04.407","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>The <code>channel_swap<\/code> is to reverse <code>RGB<\/code> into <code>BGR<\/code>, which is apparently necessary if you use a reference image net model, based on a comment in [1]. In your case the images are greyscale, so you probably do not have three channels. You might need to set it to (0, 0, 0), but even that might not help (I am unsure on the exact implementation of <code>channel_swap<\/code>). If that does not help, the simplest solution might be to preprocess you data by splitting every pixel into three values (RGB) with equal values. After that you might drop <code>channel_swap<\/code> altogether, because your channels have the same value, and swapping them is a no-op.<\/p>\n\n<p>Mean is what will be subtracted from your input data to center it. (Remember that neural networks need the data to have zero mean, while the input images usually have positive mean, hence the need of the subtraction). The mean you subtract should be the same that was used for training, so using mean from the file associated with the model is correct. I am not sure, however, on whether you should call <code>.mean(1)<\/code> on it -- did you get that line from some example? If yes, then it is most likely the correct thing to do.<\/p>\n\n<p><code>raw_scale<\/code> is a scale of your input data. The model expects pixels to be normalized, so if your input data has values between 0 and 255, then <code>raw_scale<\/code> set to 255 is correct. If your data has values between 0 and 1, then <code>raw_scale<\/code> should be set to 1.<\/p>\n\n<p>Finally, based on my understanding of the comment in [2] you do not need to provide <code>image_dims<\/code><\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L204\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L204<\/a><\/p>\n\n<p>[2] <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L18\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L18<\/a><\/p>\n"}]}
{"QuestionId":33765336,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-17T19:24:07.520","AcceptedAnswerId":33768179.0,"OwnerUserId":2615676.0,"Title":"Remove nodes from graph or reset entire default graph","Body":"<p>When working with the default global graph, is it possible to remove nodes after they've been added, or alternatively to reset the default graph to empty? When working with TF interactively in IPython, I find myself having to restart the kernel repeatedly. I would like to be able to experiment with graphs more easily if possible.<\/p>\n","answers":[{"AnswerId":"35905948","CreationDate":"2016-03-10T01:27:02.447","ParentId":null,"OwnerUserId":"2642845","Title":null,"Body":"<p>By default, a session is constructed around the default graph.\nTo avoid leaving dead nodes in the session, you need to either control the default graph or use an explicit graph.<\/p>\n\n<ul>\n<li><p>To clear the default graph, you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#reset_default_graph\" rel=\"noreferrer\">tf.reset_default_graph<\/a> function.<\/p>\n\n<pre><code>tf.reset_default_graph()\nsess = tf.InteractiveSession()\n<\/code><\/pre><\/li>\n<li><p>You can also construct explicitly a graph and avoid using the default one. If you use a normal <code>Session<\/code>, you will need to fully create the graph before constructing the session. For <code>InteractiveSession<\/code>, you can just declare the graph and use it as a context to declare further changes:<\/p>\n\n<pre><code>g = tf.Graph()\nsess = tf.InteractiveSession(graph=g)\nwith g.asdefault():\n    # Put variable declaration and other tf operation\n    # in the graph context\n    ....\n    b = tf.matmul(A, x)\n    ....\n\n sess.run([b], ...)\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>EDIT: For recent versions of <code>tensorflow<\/code> (1.0+), the correct function is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#as_default\" rel=\"noreferrer\"><code>g.as_default<\/code><\/a>.<\/p>\n"},{"AnswerId":"33768179","CreationDate":"2015-11-17T22:21:44.920","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><strong>Update 11\/2\/2016<\/strong><\/p>\n\n<p><code>tf.reset_default_graph()<\/code><\/p>\n\n<p><strong>Old stuff<\/strong><\/p>\n\n<p>There's <code>reset_default_graph<\/code>, but not part of public API (I think it should be, does someone wants to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\">file an issue<\/a> on GitHub?)<\/p>\n\n<p>My work-around to reset things is this:<\/p>\n\n<pre><code>from tensorflow.python.framework import ops\nops.reset_default_graph()\nsess = tf.InteractiveSession()\n<\/code><\/pre>\n"},{"AnswerId":"57177680","CreationDate":"2019-07-24T07:41:01.267","ParentId":null,"OwnerUserId":"8441105","Title":null,"Body":"<p>Not sure if I faced the very same problem, but<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.keras.backend.clear_session()\n<\/code><\/pre>\n\n<p>at the beginning of the cell in which the model (Keras, in my case) was constructed and trained helped to \"cut the clutter\" so only the current graph remains in the TensorBoard visualization after repeated runs of the same cell.<\/p>\n\n<p>Environment: TensorFlow 2.0 (<code>tensorflow-gpu==2.0.0b1<\/code>) in Colab with built-in TensorBoard (using the <code>%load_ext tensorboard<\/code> trick).<\/p>\n"},{"AnswerId":"46823233","CreationDate":"2017-10-19T05:17:56.973","ParentId":null,"OwnerUserId":"556876","Title":null,"Body":"<p>IPython \/ Jupyter notebook cells keep state between runs of a cell.<\/p>\n\n<p>Create a custom graph:<\/p>\n\n<pre><code>def main():\n    # Define your model\n    data = tf.placeholder(...)\n    model = ...\n\nwith tf.Graph().as_default():\n    main()\n<\/code><\/pre>\n\n<p>Once ran, the graph is cleaned up.<\/p>\n"}]}
{"QuestionId":33766689,"AnswerCount":1,"Tags":"<neural-network><deep-learning><hdf5><caffe><lmdb>","CreationDate":"2015-11-17T20:47:45.930","AcceptedAnswerId":33773553.0,"OwnerUserId":937153.0,"Title":"Caffe HDF5 Pre-processing","Body":"<p>I am getting started with Caffe and Deep learning and I am not able to understand what are the required pre-processing steps to train a model using Caffe on HDF5 data. Specifically, <\/p>\n\n<ol>\n<li>Is it required to convert the image into [0-1] range. The notebook example (00-classification.ipynb) states that the model operates in [0-255] range while some of the references show that it should be [0-1]. How do I decide this? <\/li>\n<li>As per the documentation, the conventional blob dimensions for batches of image data is N x channel K x height H x width W. There are no conflicts on this<\/li>\n<li>Channel swap step for RGB to BGR conversion is mandatory ? <\/li>\n<li>How to perform image mean computation for HDF5 data? For , the backend is lmdb. This is only for improving performance?<\/li>\n<\/ol>\n\n<p>As for the use of LMDB, questions 1-3 still hold. Any clarification on this will be highly appreciated. <\/p>\n","answers":[{"AnswerId":"33773553","CreationDate":"2015-11-18T06:51:38.000","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Welcome to caffe.<\/p>\n\n<p><strong>1.<\/strong> Scaling the input data to the range of [0..1] or [0..255] is entirely up to you. Some models works in [0..1] range, others in [0..255] and it is completely unrelated to the choice of input method (LMDB\/HDF5).<br>\n<em>The most important thing here is being consistent<\/em>. If you decided to work in the range [0..1] you must make sure both training and validation sets are prepared in the same manner and that new examples during \"deploy\" phase are scaled to the same range.<\/p>\n\n<p><strong>2.<\/strong> Caffe blobs are always 4-D batch-channel-width-height as you already observed.<\/p>\n\n<p><strong>3.<\/strong> RGB to BGR is again not mandatory but very common since BGR is the way opencv reads images. Again, the most important thing here is consistency throughout the life cycle of your net.<\/p>\n\n<p><strong>4.<\/strong> Recently, models subtract mean per channel, rather than mean per pixel. It is more convenient, especially if you change the input size of your net. When processing the HDF5 data you can compute the mean image and save it into a binaryproto. See <a href=\"https:\/\/stackoverflow.com\/a\/27645934\/1714410\">an example here<\/a>.<\/p>\n"}]}
{"QuestionId":33768110,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-11-17T22:17:08.940","AcceptedAnswerId":34086218.0,"OwnerUserId":2452761.0,"Title":"How to delete class variable in Torch?","Body":"<p>I have problems in understanding how a class variable in Torch works.<\/p>\n\n<p>I did the following:<\/p>\n\n<pre><\/pre>\n\n<p>I checked the User variable by typing   and it shows:<\/p>\n\n<pre><\/pre>\n\n<p>I  first of all tried to delete  by<\/p>\n\n<pre><\/pre>\n\n<p>it works.  is now freed and can be reinitialized to any values. But when I tried to delete the variable  by typing<\/p>\n\n<pre><\/pre>\n\n<p>It seems it's not working even though the variable  was not list in  anymore. When I try:<\/p>\n\n<pre><\/pre>\n\n<p>the error pops out:<\/p>\n\n<pre><\/pre>\n\n<p>Could anyone tell me the reason behind it?<\/p>\n","answers":[{"AnswerId":"34086218","CreationDate":"2015-12-04T10:30:04.010","ParentId":null,"OwnerUserId":"387870","Title":null,"Body":"<p>torch.class() stores the class metatable in the lua-registry, see <a href=\"http:\/\/www.lua.org\/pil\/27.3.1.html\" rel=\"nofollow\">http:\/\/www.lua.org\/pil\/27.3.1.html<\/a> and the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/4606e88758ea815e956ca7e3234fbedd02671a01\/lib\/luaT\/luaT.c#L641\" rel=\"nofollow\">luaT_lua_newmetatable()<\/a> function in the torch C-backend.<\/p>\n\n<p>In order to unregister an existing class it is necessary to remove the entry from the lua-registry. You can access the registry from lua with help of the <a href=\"http:\/\/www.lua.org\/manual\/5.2\/manual.html#pdf-debug.getregistry\" rel=\"nofollow\">debug.getregistry()<\/a> function.<\/p>\n\n<p>With removeal from registry your example works:<\/p>\n\n<pre><code>mydata = torch.class('something')\nmydata = nil\nsoemthing = nil\n\n-- remove the global registration:\ndebug.getregistry()['something'] = nil    \n\n-- now it is possible to register the class again\nmydata2 = torch.class('something')\n<\/code><\/pre>\n"}]}
{"QuestionId":33769041,"AnswerCount":5,"Tags":"<python><indexing><tensorflow>","CreationDate":"2015-11-17T23:27:39.567","AcceptedAnswerId":null,"OwnerUserId":189456.0,"Title":"Tensorflow indexing with boolean tensor","Body":"<p>In numpy, with two arrays of the same shape,  and , it is possible to do slices like this . How do you achieve the same result in tensorflow?  doesn't work and  doesn't support anything like this either. Is there a way to index with a boolean tensor right now or is that currently unsupported?<\/p>\n","answers":[{"AnswerId":"43011597","CreationDate":"2017-03-25T01:40:34.660","ParentId":null,"OwnerUserId":"281545","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>ones = tf.ones_like(x) # create a tensor all ones\nmask = tf.greater(x, ones) # boolean tensor, mask[i] = True iff x[i] &gt; 1\nslice_y_greater_than_one = tf.boolean_mask(y, mask)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/boolean_mask\" rel=\"noreferrer\">tf.boolean_mask<\/a><\/p>\n\n<p><strong>EDIT<\/strong>: another (better ?) way to do it:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.constant([1, 2, 0, 4])\ny = tf.Variable([1, 2, 0, 4])\nmask = x &gt; 1\nslice_y_greater_than_one = tf.boolean_mask(y, mask)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print (sess.run(slice_y_greater_than_one)) # [2 4]\n<\/code><\/pre>\n"},{"AnswerId":"53510122","CreationDate":"2018-11-28T00:04:36.260","ParentId":null,"OwnerUserId":"7696162","Title":null,"Body":"<p>Was looking for similar capability to reduce a TensorFlow.js tensor by a defined criteria, but TensorFlow.js does not have the boolean_mask function.  After much hair pulling and teeth gnashing, cooked up the following, which essentially sums up the total number of true criteria, and then simply selects the topk values to create the subset tensor.<\/p>\n\n<pre><code>const a = tf.tensor1d([1, 2, 0, 4]);\nconst b = a.greater(1).sum().get();\nconst {values, indices} = tf.topk(a, b);\nvalues.print();   # 4,2\nindices.print();  # 3,1\n<\/code><\/pre>\n\n<p>And to create a subset tensor of values less than or equal to 1, it's a matter of using tf.neg on the tensor as there is no bottomk function, and then after obtaining the subset tensor via topk, applying tf.neg again to restore the original values.<\/p>\n"},{"AnswerId":"33770206","CreationDate":"2015-11-18T01:23:23.113","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This is not implemented at this moment, here's GitHub issue tracking the progress -- <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206<\/a><\/p>\n"},{"AnswerId":"34934873","CreationDate":"2016-01-21T21:41:36.393","ParentId":null,"OwnerUserId":"5350761","Title":null,"Body":"<p>I would not say it is completely not implemented. How's that for a double negative?  <\/p>\n\n<p>Tensorflow actually supports quite a lot of slicing and dicing, although the syntax may be slightly less pretty.  For example, if you want to create a new array which is equal to <code>y<\/code> when <code>x&gt;1<\/code> but equal to 0 otherwise, you can definitely do that.  Check out <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/control_flow_ops#Comparison_Operators\" rel=\"noreferrer\">comparison operators<\/a> e.g.<\/p>\n\n<pre><code>masked = tf.greater(x,1)\nzeros = tf.zeros_like(x)\nnew_tensor = tf.where(masked, y, zeros)\n<\/code><\/pre>\n\n<p>If, on the other hand, you want to make a new array which contains only the guys where <code>x&gt;1<\/code> you can do that by combining <code>where<\/code> with the <code>gather<\/code> function.  Details for <code>gather<\/code> can be found at<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops\/slicing_and_joining\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops\/slicing_and_joining<\/a><\/p>\n\n<p>PS. Of course, <code>x&gt;1<\/code> is not differentiable with respect to <code>x<\/code>... tf may be great, but it doesn't work magic :).    <\/p>\n"},{"AnswerId":"48320139","CreationDate":"2018-01-18T11:22:50.127","ParentId":null,"OwnerUserId":"3350724","Title":null,"Body":"<p><code>tf.boolean_mask<\/code> does the job, but on some platforms like Raspberry Pi or OSX, the operation is not supported in Tensorflow wheel distributions (Check this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3543\" rel=\"nofollow noreferrer\">tf.boolean_mask not supported on OSX<\/a>. So an alternative is to use <code>where<\/code> and <code>gather<\/code> as @Jackson Loper suggested. For example:<\/p>\n\n<pre><code>x = tf.Variable([1, 2, 0, 4])\nix = tf.where(x &gt; 1)\ny = tf.gather(x, ix)\n\nwith tf.Session() as sess:\n sess.run(tf.global_variables_initializer())\n print(sess.run(y))\n<\/code><\/pre>\n"}]}
{"QuestionId":33770190,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><conv-neural-network><caffe>","CreationDate":"2015-11-18T01:21:28.897","AcceptedAnswerId":33773152.0,"OwnerUserId":3009514.0,"Title":"How to create caffe.deploy from train.prototxt","Body":"<p>This is my <a href=\"https:\/\/bpaste.net\/show\/ffb4600d4071\" rel=\"noreferrer\">train.prototxt<\/a>. And this is my <a href=\"https:\/\/bpaste.net\/show\/39e1064bc386\" rel=\"noreferrer\">deploy.prototxt<\/a>.<\/p>\n\n<p>When I want to load my deploy file I get this error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>So, I removed the data layer:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Than, I removed  from conv1 layer.<\/p>\n\n<p>After it, I got this error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I removed  from loss layer. And I got this error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>What should I do to fix it and create my deploy file?<\/p>\n","answers":[{"AnswerId":"38629622","CreationDate":"2016-07-28T07:22:13.690","ParentId":null,"OwnerUserId":"4696698","Title":null,"Body":"<p>Besides advices from @Shai, you may also want to disable the dropout layers.<\/p>\n\n<p>Although Jia Yangqing, author of Caffe once said that dropout layers have negligible impact on the testing results (<a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/lojWhG_V57k\/sGZO-id9wqEJ\" rel=\"nofollow\">google group conversation, 2014<\/a>), other Deeplearning tools suggest to disable dropout in the deploy phase (for example, <a href=\"http:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/noise.html\" rel=\"nofollow\">lasange<\/a>).<\/p>\n"},{"AnswerId":"33773152","CreationDate":"2015-11-18T06:25:42.267","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>There are two main differences between a \"train\" prototxt and a \"deploy\" one:<\/p>\n\n<p><strong>1. Inputs:<\/strong> While for training data is fixed to a pre-processed training dataset (lmdb\/HDF5 etc.), deploying the net require it to process other inputs in a more \"random\" fashion.<br>\nTherefore, the first change is to remove the input layers (layers that push \"data\" and \"labels\" during TRAIN and TEST phases). To replace the input layers you need to add the following declaration:<\/p>\n\n<pre><code>input: \"data\"\ninput_shape: { dim:1 dim:3 dim:224 dim:224 }\n<\/code><\/pre>\n\n<p>This declaration does not provide the actual data for the net, but it tells the net what shape to expect, allowing caffe to pre-allocate necessary resources.<\/p>\n\n<p><strong>2. Loss:<\/strong> the top most layers in a training prototxt define the loss function for the training. This usually involve the ground truth labels. When deploying the net, you no longer have access to these labels. Thus loss layers should be converted to \"prediction\" outputs. For example, a \"SoftmaxWithLoss\" layer should be converted to a simple \"Softmax\" layer that outputs class probability instead of log-likelihood loss. Some other loss layers already have predictions as inputs, thus it is sufficient just to remove them.<\/p>\n\n<p><strong>Update<\/strong>: see <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Using-a-Trained-Network:-Deploy\" rel=\"noreferrer\">this tutorial<\/a> for more information.<\/p>\n"}]}
{"QuestionId":33771458,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2015-11-18T03:44:39.050","AcceptedAnswerId":null,"OwnerUserId":5557434.0,"Title":"How to get the error line of code with theano?","Body":"<p>In the FAQ page <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html<\/a>,\nI saw with this example<\/p>\n\n<pre><\/pre>\n\n<p>If I use optimizer=fast_compile or optimizer=None and  exception_verbosity=high I should see<\/p>\n\n<pre><\/pre>\n\n<p>But instead I just saw<\/p>\n\n<pre><\/pre>\n\n<p>which the backtrace didn't tell me the line. What is the problem?<\/p>\n","answers":[]}
{"QuestionId":33772819,"AnswerCount":2,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2015-11-18T05:59:46.583","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"tensorflow error on running the seq2seq model","Body":"<p>On running the RNN tutorial <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.md#run_it\" rel=\"nofollow\">example<\/a> , I get the following error after a reading data line statements:<\/p>\n\n<pre><\/pre>\n\n<p>\/default\/_app.py\", line 15, in run\n    sys.exit(main(sys.argv))\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/models\/rnn\/translate\/translate.py\", line 261, in main\n    train()\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/models\/rnn\/translate\/translate.py\", line 130, in train\n    model = create_model(sess, False)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/models\/rnn\/translate\/translate.py\", line 109, in create_model\n    forward_only=forward_only)<br>\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/models\/rnn\/translate\/seq2seq_model.py\", line 153, in <strong>init<\/strong>\n    self.saver = tf.train.Saver(tf.all_variables())\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/training\/saver.py\", line 693, in <strong>init<\/strong>\n    restore_sequentially=restore_sequentially)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/training\/saver.py\", line 411, in build<br>\n    save_tensor = self._AddSaveOps(filename_tensor, vars_to_save)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/training\/saver.py\", line 114, in _AddSaveOps\n    save = self.save_op(filename_tensor, vars_to_save)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/training\/saver.py\", line 68, in save_op<br>\n    tensor_slices=[vs.slice_spec for vs in vars_to_save])\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/ops\/io_ops.py\", line 149, in _save\n    tensors, name=name)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/ops\/gen_io_ops.py\", line 343, in _save_slices\n    name=name)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/ops\/op_def_library.py\", line 646, in apply_op\n    op_def=op_def)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/framework\/ops.py\", line 1767, in create_op\n    original_op=self._default_original_op, op_def=op_def)\n  File \"\/home\/temp_user\/.cache\/bazel\/_bazel_temp_user\/7cf40d683d56020fae2d5abbde7f9f05\/tensorflow\/bazel-out\/local_linux-opt\/bin\/tensorflow\/models\/rnn\/translate\/translate.runfiles\/tensorflow\/python\/framework\/ops.py\", line 1008, in <strong>init<\/strong>\n    self._traceback = _extract_stack()<\/p>\n\n<p>ERROR: Non-zero return code '1' from command: Process exited with status 1.<\/p>\n\n<p>So what is the cause of this problem as the other language model example is working and the library has also been built. As per comments I created the checkpoint directory , still throwing the same error:\ntensorflow\/core\/common_runtime\/executor.cc:1052] 0x400d2bbe0 Compute status: Not found: .\/checkpoints_directory\/translate.ckpt-200.tempstate9246663217899500702<\/p>\n","answers":[{"AnswerId":"33794161","CreationDate":"2015-11-19T02:22:49.370","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>I think this is one of the problems that arises when the previous checkpoint was not saved properly. You can correct it in the following steps.<\/p>\n\n<p>1.You can delete the all checkpoint files and restart the training.:<\/p>\n\n<pre><code>rm checkpoint\nrm translate-ckpt-*\n<\/code><\/pre>\n\n<p>Now, restart your training again.<\/p>\n\n<p>Alternatively, you can remove the most latest checkpoint and start it from the previous checkpoint.<\/p>\n\n<p>1.Go to the directory and delete the most latest checkpoint, in this case it is:<\/p>\n\n<pre><code>rm translate-ckpt-200\n<\/code><\/pre>\n\n<p>2.Now edit the checkpoint file. You might see something like<\/p>\n\n<pre><code>model_checkpoint_path: \"data\/translate.ckpt-200\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-170\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-180\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-190\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-200\"\n<\/code><\/pre>\n\n<p>3.Remove the last line and set the checkpoint to a previous stage.<\/p>\n\n<pre><code>model_checkpoint_path: \"data\/translate.ckpt-190\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-170\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-180\"\nall_model_checkpoint_paths: \"data\/translate.ckpt-190\"\n<\/code><\/pre>\n\n<p>4.Restart your training. <\/p>\n"},{"AnswerId":"41002102","CreationDate":"2016-12-06T18:12:56.947","ParentId":null,"OwnerUserId":"5660353","Title":null,"Body":"<p>I got the same issue running Sequence-to-Sequence Models. \nAnd creating the [checkpoint directory] before running the code solves the problem !<\/p>\n"}]}
{"QuestionId":33772833,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-18T06:01:16.867","AcceptedAnswerId":33775312.0,"OwnerUserId":2829987.0,"Title":"error while merging summaries for tensorboard","Body":"\n\n<p>I am trying to generate the graph for MNIST beginner tutorial but is getting the following error. For some reason, merged_summary_op object is None.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I think I am missing a step here. I launched the session first and then running the statement:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"33775312","CreationDate":"2015-11-18T08:43:55.163","ParentId":null,"OwnerUserId":"5575330","Title":null,"Body":"\n\n<p>I had the same error.<\/p>\n\n<p>In my case, adding at least one <code>tf.scalar_summary()<\/code> before calling <code>tf.merge_all_summaries()<\/code> solved the problem.<\/p>\n\n<p>For example,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>cross_entropy = -tf.reduce_sum(y_*tf.log(y))\ntf.scalar_summary(\"cross_entropy\", cross_entropy)\nmerged_summary_op = tf.merge_all_summaries()\n<\/code><\/pre>\n\n<p>I hope this snippet helps you.<\/p>\n"}]}
{"QuestionId":33773271,"AnswerCount":1,"Tags":"<neural-network><deep-learning><tensorflow><tensorboard>","CreationDate":"2015-11-18T06:33:46.863","AcceptedAnswerId":34909941.0,"OwnerUserId":1782011.0,"Title":"Graph visualisaton is not showing in tensorboard for seq2seq model","Body":"\n\n<p>I build a seq2seq model using the seq2seq.py library provided with tensorflow.\nBefore training anything I wanted to visualize the  graph network of my untrained model in tensorboard, but it does not want to display this.<\/p>\n\n<p>Below a minimal example to reproduce my problem.\nAnybody an idea why this does not work? Can you only visualize a grap of a model after it has been trained?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34909941","CreationDate":"2016-01-20T20:33:05.567","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>It looks like this might be related to a bug where the graph visualization does not work in the firefox browser. Try using chrome or safari if possible.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/650\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/650<\/a><\/p>\n"}]}
{"QuestionId":33773272,"AnswerCount":0,"Tags":"<python><cuda><theano>","CreationDate":"2015-11-18T06:33:48.920","AcceptedAnswerId":null,"OwnerUserId":5575191.0,"Title":"ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu","Body":"<p><strong>When I run import theano the error is as follows:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>The contents of my .theanorc file are as follows:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>And the Environment variable is as follows\uff1a<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>If I run import theano without the GPU option on, it runs without a problem. Also CUDA samples run without a problem.I have confirmed that the computer GPU has been opened.\nA few weeks ago\uff0ci had run it without a problem. Because of a software everything was unwork. Now the problem confuse me...<\/p>\n","answers":[]}
{"QuestionId":33773661,"AnswerCount":3,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2015-11-18T06:59:02.053","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"Predicting next word using the language model tensorflow example","Body":"<p>The <a href=\"http:\/\/tensorflow.org\/tutorials\/recurrent\/index.md#language_modeling\" rel=\"noreferrer\">tensorflow tutorial<\/a> on language model allows  to compute the probability of sentences :<\/p>\n\n<pre><\/pre>\n\n<p>in the comments below it also specifies a way of predicting the next word instead of probabilities but does not specify how this can be done. So how to output a word instead of probability using this example?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33781091","CreationDate":"2015-11-18T13:15:00.407","ParentId":null,"OwnerUserId":"4418769","Title":null,"Body":"<p>It is actually an advantage that the function returns a probability instead of the word itself. Since it is using a list of words, with the associated probabilities, you can do further processing, and increase the accuracy of your result.<\/p>\n\n<p>To answer your question:\nYou can take the list of words, iterate though it , and make the program display the word with the highest probability. <\/p>\n"},{"AnswerId":"33839742","CreationDate":"2015-11-21T04:32:51.633","ParentId":null,"OwnerUserId":"2476920","Title":null,"Body":"<p>Your output is a TensorFlow list and it is possible to get its max argument (the predicted most probable class) with a TensorFlow function. This is normally the list that contains the next word's probabilities. <\/p>\n\n<p>At \"Evaluate the Model\" from this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/mnist\/pros\/index.md#evaluate-the-model\" rel=\"nofollow noreferrer\">page<\/a>, your output list is <code>y<\/code> in the following example:<\/p>\n\n<blockquote>\n  <p>First we'll figure out where we predicted the correct label. <code>tf.argmax<\/code>\n  is an extremely useful function which gives you the index of the\n  highest entry in a tensor along some axis. For example, <code>tf.argmax(y,1)<\/code>\n  is the label our model thinks is most likely for each input, while\n  <code>tf.argmax(y_,1)<\/code> is the true label. We can use <code>tf.equal<\/code> to check if our\n  prediction matches the truth.\n  <code>correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))<\/code><\/p>\n<\/blockquote>\n\n<p>Another approach that is different is to have pre-vectorized (embedded\/encoded) words. You could vectorize your words (therefore embed them) with Word2vec to accelerate learning, you might want to take a look at this. Each word could be represented as a point in a 300 dimensions space of meaning, and you could find automatically the \"N words\" closest to the predicted point in space at the output of the network. In that case, the <code>argmax<\/code> way to proceed does not work anymore and you could probably compare on cosine similarity with the words you truly wanted to compare to, but for that I am not sure actually how does this could cause numerical instabilities. In that case <code>y<\/code> will not represent words as features, but word embeddings over a dimensionality of, let's say, 100 to 2000 in size according to different models. You could Google something like this for more info: \"man woman queen word addition word2vec\" to understand the subject of embeddings more. <\/p>\n\n<p>Note: when I talk about word2vec here, it is about using an external pre-trained word2vec model to help your training to only have pre-embedded inputs and create embedding outputs. Those outputs' corresponding words can be re-figured out by word2vec to find the corresponding similar top predicted words. <\/p>\n\n<p>Notice that the approach I suggest is not exact since it would be only useful to know if we predict EXACTLY the word that we wanted to predict. For a more soft approach, it would be possible to use ROUGE or BLEU metrics for evaluating your model in case you use sentences or something longer than a word. <\/p>\n"},{"AnswerId":"39282697","CreationDate":"2016-09-02T00:30:05.890","ParentId":null,"OwnerUserId":"456185","Title":null,"Body":"<p>You need to find the argmax of the probabilities, and translate the index back to a word by reversing the word_to_id map. To get this to work, you must save the probabilities in the model and then fetch them from the run_epoch function (you could also save just the argmax itself). Here's a snippet: <\/p>\n\n<pre><code>inverseDictionary = dict(zip(word_to_id.values(), word_to_id.keys()))\n\ndef run_epoch(...):\n  decodedWordId = int(np.argmax(logits))\n  print (\" \".join([inverseDictionary[int(x1)] for x1 in np.nditer(x)])  \n    + \" got\" + inverseDictionary[decodedWordId] + \n    + \" expected:\" + inverseDictionary[int(y)])\n<\/code><\/pre>\n\n<p>See full implementation here: <a href=\"https:\/\/github.com\/nelken\/tf\" rel=\"nofollow\">https:\/\/github.com\/nelken\/tf<\/a><\/p>\n"}]}
{"QuestionId":33776077,"AnswerCount":1,"Tags":"<theano><caffe><lasagne>","CreationDate":"2015-11-18T09:24:02.487","AcceptedAnswerId":null,"OwnerUserId":4570493.0,"Title":"Parse '.caffemodel.h5' for Lasagne","Body":"<p>I have trained a model from caffe in \".caffemodel.h5\" format. I want to parse it to extract parameters and feed it to a lasagne model. How can I do it?<\/p>\n","answers":[{"AnswerId":"33776230","CreationDate":"2015-11-18T09:31:17.137","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You will have to replicate the architecture of your caffemodel in terms of Lasagne layers first, by hand, or use a script that does this (a script that takes a caffe protobuf and translates it to lasagne exists AFAIK, or should be made if not. In any case it exists in sklearn-theano.)<\/p>\n\n<p>By hand, you need to open the hf5 file using e.g. <code>pytables<\/code>, and then replicate the architecture in Lasagne.<\/p>\n"}]}
{"QuestionId":33778020,"AnswerCount":1,"Tags":"<machine-learning><theano>","CreationDate":"2015-11-18T10:49:13.940","AcceptedAnswerId":33781412.0,"OwnerUserId":5557434.0,"Title":"How to fix the NAN or INF when we use cross entropy with theano?","Body":"<p>We'll have to compute: <\/p>\n\n<p><\/p>\n\n<p>so when we get  <strong>1<\/strong>. or <strong>0<\/strong>.\uff0cthis problem would show up. What should I do to avoid it?<\/p>\n","answers":[{"AnswerId":"33781412","CreationDate":"2015-11-18T13:30:37.213","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Your expression <code>y_compute<\/code> maybe contains an exponential, e.g. coming from <code>theano.tensor.nnet.sigmoid<\/code>? In that case, it should usually never reach exact 0 or 1. In those cases you can then just use your expression or <code>theano.tensor.nnet.crossentropy_categorical_1hot<\/code> directly.<\/p>\n\n<p>If for whatever reason you have exact 0 and 1, another way is to clip the input to the crossentropy. Try e.g. replacing <code>y_compute<\/code> with <code>theano.tensor.clip(y_compute, 0.001, 0.999)<\/code>, knowing that this will restrict the range of the logarithm.<\/p>\n"}]}
{"QuestionId":33778225,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2015-11-18T10:57:44.920","AcceptedAnswerId":33797142.0,"OwnerUserId":2584518.0,"Title":"Building custom Caffe layer in python","Body":"<p>After parsing many links regarding building Caffe layers in Python i still have difficulties in understanding few concepts. Can please someone clarify them?  <\/p>\n\n<ul>\n<li>Blobs and weights python structure for network is explained here: <a href=\"https:\/\/stackoverflow.com\/questions\/31324739\/finding-gradient-of-a-caffe-conv-filter-with-regards-to-input\">Finding gradient of a Caffe conv-filter with regards to input<\/a>.  <\/li>\n<li>Network and Solver structure is explained here: <a href=\"https:\/\/stackoverflow.com\/questions\/32379878\/cheat-sheet-for-caffe-pycaffe\">Cheat sheet for caffe \/ pycaffe?<\/a>.  <\/li>\n<li>Example of defining python layer is here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py\" rel=\"noreferrer\">pyloss.py on git<\/a>.  <\/li>\n<li>Layer tests here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_python_layer_with_param_str.py\" rel=\"noreferrer\">test layer on  git<\/a>.  <\/li>\n<li>Development of new layers for C++ is described here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Development\" rel=\"noreferrer\">git wiki<\/a>.<\/li>\n<\/ul>\n\n<p>What I am still missing is:<\/p>\n\n<ol>\n<li> method: what I should do here? Why in example I should compare the lenght of 'bottom' param with '2'? Why it should be 2? It seems not a batch size because its arbitrary? And bottom as I understand is blob, and then the first dimension is batch size?<\/li>\n<li> method: as I understand 'bottom' input param is blob of below layer, and 'top' param is blob of upper layer, and I need to reshape top layer according to output shape of my calculations with forward pass. But why do I need to do this every forward pass if these shapes do not change from pass to pass, only weights change?<\/li>\n<li> and  methods have 0 indexes for 'top' input param used. Why would I need to use  or  instead of  and ? Whats this index about? If we do not use other part of this top list, why it is exposed in this way? I can suspect its or C++ backbone coincidence, but it would be good to know exactly.<\/li>\n<li><p> method, line with:<\/p>\n\n<pre><\/pre>\n\n<p>what I do here? why its dimension is 2 again? And what I am counting here? Why both part of blobs (0 and 1) should be equal in amount of some members ()?<\/p><\/li>\n<li><p> method, what I define by this line:<\/p>\n\n<pre><\/pre>\n\n<p>When it is used after forward path if I define it? Can we just use<\/p>\n\n<pre><\/pre>\n\n<p>instead to count loss later in this method, without assigning to , or its done with some purpose?<\/p><\/li>\n<li><p> method: what's this about: ? Why again range is 2?<\/p><\/li>\n<li> method,  parameter: why it is defined? I mean if its True, gradient should be assigned to  as I see, but why would someone call method which would do nothing with , if it just do nothing and still cycling inside?<\/li>\n<\/ol>\n\n<p>I'm sorry if those questions are too obvious, I just wasn't able to find a good guide to understand them and asking for help here.<\/p>\n","answers":[{"AnswerId":"36369650","CreationDate":"2016-04-02T05:42:12.537","ParentId":null,"OwnerUserId":"3350028","Title":null,"Body":"<p><strong>Why it should be 2?<\/strong><\/p>\n\n<p>That specific gist is talking about the Euclidian loss layer. Euclidian loss is the mean square error between <strong>2<\/strong> vectors. Hence there must be 2 vectors in the input blob to this layer. The length of each vector must be same because it is element-wise difference. You can see this check in the <strong>reshape<\/strong> method.<\/p>\n\n<p>Thanks.<\/p>\n"},{"AnswerId":"33797142","CreationDate":"2015-11-19T06:56:26.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You asked a lot of questions here, I'll give you some highlights and pointers that I hope will clarify matters for you. I will not explicitly answer all your questions.<\/p>\n\n<p>It seems like you are most confused about the the difference between a blob and a layer's input\/output. Indeed most of the layers has a <em>single<\/em> blob as input and a <em>single<\/em> blob as output, but it is not always the case. Consider a loss layer: it has <em>two<\/em> inputs: predictions and ground truth labels. So, in this case <code>bottom<\/code> is a vector of length <strong>2<\/strong>(!) with <code>bottom[0]<\/code> being a (4-D) blob representing predictions, while <code>bottom[1]<\/code> is another blob with the labels. Thus, when constructing such a layer you must ascertain that you have exactly (hard coded) 2 input blobs (see e.g., <code>ExactNumBottomBlobs()<\/code> in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/loss_layers.hpp#L40\" rel=\"nofollow noreferrer\"><code>AccuracyLayer<\/code><\/a> definition).  <\/p>\n\n<p>The same goes for <code>top<\/code> blobs as well: indeed in most cases there is a single <code>top<\/code> for each layer, but it's not always the case (see e.g., <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/loss_layers.hpp#L40\" rel=\"nofollow noreferrer\"><code>AccuracyLayer<\/code><\/a>). Therefore, <code>top<\/code> is also a <em>vector<\/em> of 4-D blobs, one for each <code>top<\/code> of the layer. Most of the time there would be a single element in that vector, but sometimes you might find more than one.<\/p>\n\n<p>I believe this covers your questions 1,3,4 and 6.<\/p>\n\n<p>As of <code>reshape()<\/code> (Q.2) this function is not called every forward pass, it is called only when net is setup to allocate space for inputs\/outputs and params.<br>\nOccasionally, you might want to change input size for your net (e.g., for detection nets) then you need to call <code>reshape()<\/code> for all layers of the net to accommodate the new input size.<\/p>\n\n<p>As for <code>propagate_down<\/code> parameter (Q.7): since a layer may have more than one <code>bottom<\/code> you would need, in principle, to pass the gradient to <em>all<\/em> <code>bottom<\/code>s during backprop. However, what is the meaning of a gradient to the <code>label<\/code> bottom of a loss layer? There are cases when you do not want to propagate to <em>all<\/em> <code>bottom<\/code>s: this is what this flag is for. (here's an <a href=\"https:\/\/stackoverflow.com\/a\/33349475\/1714410\">example<\/a> with a loss layer with three <code>bottom<\/code>s that expect gradient to all of them).<\/p>\n\n<p>For more information, see <a href=\"https:\/\/stackoverflow.com\/a\/41481539\/1714410\">this <code>\"Python\"<\/code> layer tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":33779095,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2015-11-18T11:39:48.030","AcceptedAnswerId":null,"OwnerUserId":4063989.0,"Title":"Docker download Google's TensorFlow issue","Body":"<p>Recently Google introduces the TensorFlow (machine learning library) which is not distributed for Windows I searched and find that it is possible to download it via Docker, \nI installed the docker on windows 8.1 machine and followed <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#docker-based_installation\" rel=\"nofollow noreferrer\">this<\/a> and <a href=\"http:\/\/caffinc.github.io\/2015\/11\/tensorflow-windows\/\" rel=\"nofollow noreferrer\">this<\/a> \nto get things work but since I am using a local proxy then I tried to export HTTP_PROXY, HTTPS_PROXY to \/var\/lib\/boot2docker\/profile file and \nnow when I run below command <\/p>\n\n<blockquote>\n  <p>docker pull b.gcr.io\/tensorflow\/tensorflow<\/p>\n<\/blockquote>\n\n<p>I get <\/p>\n\n<blockquote>\n  <p>Using default tag: latest Error response from daemon: unable to ping\n  registry endpoint <a href=\"https:\/\/b.gcr.io\/v0\/\" rel=\"nofollow noreferrer\">https:\/\/b.gcr.io\/v0\/<\/a> v2 ping attempt failed with\n  error: Get <a href=\"https:\/\/b.gcr.io\/v2\/\" rel=\"nofollow noreferrer\">https:\/\/b.gcr.io\/v2\/<\/a>: http: error connecting to proxy\n  <a href=\"https:\/\/localhost:8580\" rel=\"nofollow noreferrer\">https:\/\/localhost:8580<\/a>: dial tcp 127.0.0.1:8580: connection refused \n  v1 ping attempt failed with error: Get <a href=\"https:\/\/b.gcr.io\/v1\/_ping\" rel=\"nofollow noreferrer\">https:\/\/b.gcr.io\/v1\/_ping<\/a>:\n  http: error connecting to proxy <a href=\"https:\/\/localhost:8580\" rel=\"nofollow noreferrer\">https:\/\/localhost:8580<\/a>: dial tcp\n  127.0.0.1:8580: connection refused<\/p>\n<\/blockquote>\n\n<p>Can someone tell me how to fix this ?<\/p>\n\n<p>my docker-machine's profile<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lAQFO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lAQFO.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"33785208","CreationDate":"2015-11-18T16:23:42.060","ParentId":null,"OwnerUserId":"6309","Title":null,"Body":"<p>If you are using docker-machine, you shouldn't have to tinker directly with the docker daemon profile.<\/p>\n\n<p>Use the <code>--engine-env<\/code> option when creating your VM instance for docker.<br>\nSee <a href=\"https:\/\/docs.docker.com\/machine\/reference\/create\/#specifying-configuration-options-for-the-created-docker-engine\" rel=\"nofollow\"><strong><code>docker-machine create<\/code><\/strong><\/a>.<\/p>\n\n<p>Simply define <code>%HOME%\/.bashrc<\/code> (which will be read when you open your bash session, <em>before<\/em> doing an ssh to your VM)<\/p>\n\n<pre><code>alias dm=docker-machine\nexport http_proxy=$HTTP_PROXY\nexport https_proxy=$HTTPS_PROXY\nexport NO_PROXY=$NO_PROXY\nexport no_proxy=$NO_PROXY\n\nalias dmcv='docker-machine create -d virtualbox --engine-env HTTP_PROXY=$http_proxy --engine-env HTTPS_PROXY=$https_proxy --engine-env http_proxy=$http_proxy --engine-env https_proxy=$https_proxy --engine-env NO_PROXY=$no_proxy --engine-env no_proxy=$no_proxy'\n\nalias d=docker\nalias dpsa='docker ps -a'\ndenv() { eval $(docker-machine env \"$@\"); }\nvbmctr() { eval $(VBoxManage controlvm $1 natpf1 \"$1-$2-tcp,tcp,,$2,,$2\"); eval $(VBoxManage controlvm $1 natpf1 \"$1-$2-udp,udp,,$2,,$2\"); }\n<\/code><\/pre>\n\n<p>Make sure your htt(s)_proxy are defined with:<\/p>\n\n<pre><code>http:\/\/username:password@proxy-server.com:port\n<\/code><\/pre>\n\n<p>(note that it always starts with <code>http:\/\/<\/code> even for <code>https_proxy<\/code>)<\/p>\n\n<p>Also make sure to define <code>no_proxy<\/code>:<\/p>\n\n<pre><code>NO_PROXY=.company,.sock,localhost,127.0.0.1,::1,192.168.99.100,192.168.99.101,192.168.99.102,192.168.99.103,192.168.99.104\n<\/code><\/pre>\n\n<p>(replace <code>.company<\/code> by your company extension)<\/p>\n\n<p>From there, you can do a:<\/p>\n\n<pre><code>dmcv default\ndenv default\ndm ssh default\n<\/code><\/pre>\n\n<p>The key here is the <code>dmcv<\/code> alias: it will create the VM with a <code>\/var\/lib\/boot2docker\/profile<\/code> <em>already<\/em> modified for you  with proxy.<\/p>\n\n<p>Note that I always use the upercase and lowercase versions of those proxy variables, in order to be interpreted by different unix commands (like curl, wget, ...) which rely sometime on lowercase, other times on upercase variable names.<\/p>\n"}]}
{"QuestionId":33779965,"AnswerCount":0,"Tags":"<deep-learning><caffe><conv-neural-network>","CreationDate":"2015-11-18T12:22:40.220","AcceptedAnswerId":null,"OwnerUserId":3929051.0,"Title":"Caffe training iteration loss is -nan","Body":"<p>I'm trying to implement <a href=\"https:\/\/gist.github.com\/shelhamer\/91eece041c19ff8968ee\" rel=\"nofollow\">FCN-8s<\/a> using my own custom data. While training, from scratch, I see that my loss = -nan.<\/p>\n\n<p>Could someone suggest what's going wrong and how can I correct this? My solver.prototxt is as follows:- <\/p>\n\n<p>The train_val.prototxt is the same as given in the <a href=\"https:\/\/gist.github.com\/shelhamer\/91eece041c19ff8968ee\" rel=\"nofollow\">link above<\/a>. My custom images are of size 3x512x640 and labels of 1x512x640.There are 11 different types of labels.  <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33780779,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-11-18T13:01:21.773","AcceptedAnswerId":33786620.0,"OwnerUserId":3929051.0,"Title":"Caffe | solver.prototxt values setting strategy","Body":"<p>On Caffe, I am trying to implement a Fully Convolution Network for semantic segmentation. I was wondering is there a specific strategy to set up your  values for the following hyper-parameters:<\/p>\n\n<ul>\n<li>test_iter<\/li>\n<li>test_interval<\/li>\n<li>iter_size<\/li>\n<li>max_iter<\/li>\n<\/ul>\n\n<p>Does it depend on the number of images you have for your training set? If so, how? <\/p>\n","answers":[{"AnswerId":"33786620","CreationDate":"2015-11-18T17:33:18.100","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In order to set these values in a meaningful manner, you need to have a few more bits of information regarding your data:<\/p>\n\n<p><strong>1. Training set size<\/strong> the total number of training examples you have, let's call this quantity <code>T<\/code>.<br>\n<strong>2. Training batch size<\/strong> the number of training examples processed together in a single batch, this is usually set by the input data layer in the <code>'train_val.prototxt'<\/code>. For example, in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/train_val.prototxt#L25\" rel=\"nofollow noreferrer\">this file<\/a> the train batch size is set to 256. Let's denote this quantity by <code>tb<\/code>.<br>\n<strong>3. Validation set size<\/strong> the total number of examples you set aside for validating your model, let's denote this by <code>V<\/code>.<br>\n<strong>4. Validation batch size<\/strong> value set in <code>batch_size<\/code> for the TEST phase. In <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/train_val.prototxt#L52\" rel=\"nofollow noreferrer\">this example<\/a> it is set to 50. Let's call this <code>vb<\/code>.<\/p>\n\n<p>Now, during training, you would like to get an un-biased estimate of the performance of your net every once in a while. To do so you run your net on the validation set for <code>test_iter<\/code> iterations. To cover the entire validation set you need to have <code>test_iter = V\/vb<\/code>.<br>\nHow often would you like to get this estimation? It's really up to you. If you have a very large validation set and a slow net, validating too often will make the training process too long. On the other hand, not validating often enough may prevent you from noting if and when your training process failed to converge. <code>test_interval<\/code> determines how often you validate: usually for large nets you set <code>test_interval<\/code> in the order of 5K, for smaller and faster nets you may choose lower values. Again, all up to you. <\/p>\n\n<p>In order to cover the entire training set (completing an \"epoch\") you need to run <code>T\/tb<\/code> iterations. Usually one trains for several epochs, thus <code>max_iter=#epochs*T\/tb<\/code>.<\/p>\n\n<p>Regarding <code>iter_size<\/code>: this allows to average gradients over several training mini batches, see <a href=\"https:\/\/stackoverflow.com\/a\/36532014\/1714410\">this thread<\/a> fro more information.<\/p>\n"}]}
{"QuestionId":33783672,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-11-18T15:13:38.313","AcceptedAnswerId":null,"OwnerUserId":5563431.0,"Title":"How can I visualize the weights(variables) in cnn in Tensorflow?","Body":"<p>After training the cnn model, I want to visualize the weight or print out the weights, what can I do?\nI cannot even print out the variables after training.\nThank you!<\/p>\n","answers":[{"AnswerId":"35858862","CreationDate":"2016-03-08T04:17:57.137","ParentId":null,"OwnerUserId":"1760830","Title":null,"Body":"<p>Like @mrry said, you can use <code>tf.image_summary<\/code>. For example, for <code>cifar10_train.py<\/code>, you can put this code somewhere under <code>def train()<\/code>. Note how you access a var under scope 'conv1'<\/p>\n\n<pre><code># Visualize conv1 features\nwith tf.variable_scope('conv1') as scope_conv:\n  weights = tf.get_variable('weights')\n\n  # scale weights to [0 255] and convert to uint8 (maybe change scaling?)\n  x_min = tf.reduce_min(weights)\n  x_max = tf.reduce_max(weights)\n  weights_0_to_1 = (weights - x_min) \/ (x_max - x_min)\n  weights_0_to_255_uint8 = tf.image.convert_image_dtype (weights_0_to_1, dtype=tf.uint8)\n\n  # to tf.image_summary format [batch_size, height, width, channels]\n  weights_transposed = tf.transpose (weights_0_to_255_uint8, [3, 0, 1, 2])\n\n  # this will display random 3 filters from the 64 in conv1\n  tf.image_summary('conv1\/filters', weights_transposed, max_images=3)\n<\/code><\/pre>\n\n<p>If you want to visualize all your <code>conv1<\/code> filters in one nice grid, you would have to organize them into a grid yourself. I did that today, so now I'd like to share a <a href=\"https:\/\/gist.github.com\/kukuruza\/03731dc494603ceab0c5\" rel=\"noreferrer\">gist for visualizing conv1 as a grid<\/a><\/p>\n"},{"AnswerId":"33794463","CreationDate":"2015-11-19T02:55:47.200","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To visualize the weights, you can use a <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.html#image_summary\" rel=\"noreferrer\"><code>tf.image_summary()<\/code><\/a> op to transform a convolutional filter (or a slice of a filter) into a summary proto, write them to a log using a <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/train.html#SummaryWriter\" rel=\"noreferrer\"><code>tf.train.SummaryWriter<\/code><\/a>, and visualize the log using <a href=\"http:\/\/tensorflow.org\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"noreferrer\">TensorBoard<\/a>.<\/p>\n\n<p>Let's say you have the following (simplified) program:<\/p>\n\n<pre><code>filter = tf.Variable(tf.truncated_normal([8, 8, 3]))\nimages = tf.placeholder(tf.float32, shape=[None, 28, 28])\n\nconv = tf.nn.conv2d(images, filter, strides=[1, 1, 1, 1], padding=\"SAME\")\n\n# More ops...\nloss = ...\noptimizer = tf.GradientDescentOptimizer(0.01)\ntrain_op = optimizer.minimize(loss)\n\nfilter_summary = tf.image_summary(filter)\n\nsess = tf.Session()\nsummary_writer = tf.train.SummaryWriter('\/tmp\/logs', sess.graph_def)\nfor i in range(10000):\n  sess.run(train_op)\n  if i % 10 == 0:\n    # Log a summary every 10 steps.\n    summary_writer.add_summary(filter_summary, i)\n<\/code><\/pre>\n\n<p>After doing this, you can start TensorBoard to visualize the logs in <code>\/tmp\/logs<\/code>, and you will be able to see a visualization of the filter.<\/p>\n\n<p>Note that this trick visualizes depth-3 filters as RGB images (to match the channels of the input image). If you have deeper filters, or they don't make sense to interpret as color channels, you can use the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/array_ops.html#split\" rel=\"noreferrer\"><code>tf.split()<\/code><\/a> op to split the filter on the depth dimension, and generate one image summary per depth.<\/p>\n"},{"AnswerId":"40652881","CreationDate":"2016-11-17T10:49:12.823","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>You can extract the values as numpy arrays the following way:<\/p>\n\n<pre><code>with tf.variable_scope('conv1', reuse=True) as scope_conv:\n    W_conv1 = tf.get_variable('weights', shape=[5, 5, 1, 32])\n    weights = W_conv1.eval()\n    with open(\"conv1.weights.npz\", \"w\") as outfile:\n        np.save(outfile, weights)\n<\/code><\/pre>\n\n<p>Note that you have to adjust the scope (<code>'conv1'<\/code> in my case) and the variable name (<code>'weights'<\/code> in my case).<\/p>\n\n<p>Then it boils down on visualizing numpy arrays. One example how to visualize numpy arrays is<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n\n\"\"\"Visualize numpy arrays.\"\"\"\n\nimport numpy as np\nimport scipy.misc\n\narr = np.load('conv1.weights.npb')\n\n# Get each 5x5 filter from the 5x5x1x32 array\nfor filter_ in range(arr.shape[3]):\n    # Get the 5x5x1 filter:\n    extracted_filter = arr[:, :, :, filter_]\n\n    # Get rid of the last dimension (hence get 5x5):\n    extracted_filter = np.squeeze(extracted_filter)\n\n    # display the filter (might be very small - you can resize the window)\n    scipy.misc.imshow(extracted_filter)\n<\/code><\/pre>\n"}]}
{"QuestionId":33784214,"AnswerCount":3,"Tags":"<python><testing><machine-learning><tensorflow>","CreationDate":"2015-11-18T15:36:18.373","AcceptedAnswerId":34054866.0,"OwnerUserId":5577321.0,"Title":"How to test tensorflow cifar10 cnn tutorial model","Body":"<p>I am relatively new to machine-learning and currently have almost no experiencing in developing it.<\/p>\n\n<p>So my <strong>Question<\/strong> is: after training and evaluating the cifar10 dataset from the tensorflow <a href=\"http:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/index.html\" rel=\"noreferrer\">tutorial<\/a> I was wondering how could one test it with sample images?<\/p>\n\n<p>I could train and evaluate the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"noreferrer\">Imagenet tutorial from the caffe machine-learning framework<\/a> and it was relatively easy to use the trained model on custom applications using the python API.<\/p>\n\n<p>Any help would be very appreciated!<\/p>\n","answers":[{"AnswerId":"33785193","CreationDate":"2015-11-18T16:23:03.710","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>I recommend taking a look at the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">basic MNIST tutorial<\/a> on the TensorFlow website. It looks like you define some function that generates the type of output that you want, and then run your session, passing it this evaluation function (<code>correct_prediction<\/code> below), and a dictionary containing whatever arguments you require (<code>x<\/code> and <code>y_<\/code> below).<\/p>\n\n<p>If you have defined and trained some network that takes an input <code>x<\/code>, generates a response <code>y<\/code> based on your inputs, and you know your expected responses for your testing set <code>y_<\/code>, you may be able to print out every response to your testing set with something like:<\/p>\n\n<pre><code>correct_prediction = tf.equal(y, y_)  % Check whether your prediction is correct\nprint(sess.run(correct_prediction, feed_dict={x: test_images, y_: test_labels}))\n<\/code><\/pre>\n\n<p>This is just a modification of what is done in the tutorial, where instead of trying to print each response, they determine the percent of correct responses. Also note that the tutorial uses <em>one-hot vectors<\/em> for the prediction <code>y<\/code> and actual value <code>y_<\/code>, so in order to return the associated numeral, they have to find which index of these vectors are equal to one with <code>tf.argmax(y, 1)<\/code>.<\/p>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>In general, if you define something in your graph, you can output it later when you run your graph. Say you define something that determines the result of the softmax function on your output logits as:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n  ...\n  prediction = tf.nn.softmax(logits)\n  ...\n<\/code><\/pre>\n\n<p>then you can output this at run time with:<\/p>\n\n<pre><code>with tf.Session(graph=graph) as sess:\n  ...\n  feed_dict = { ... }  # define your feed dictionary\n  pred = sess.run([prediction], feed_dict=feed_dict)\n  # do stuff with your prediction vector\n<\/code><\/pre>\n"},{"AnswerId":"34054866","CreationDate":"2015-12-02T23:10:10.797","ParentId":null,"OwnerUserId":"5577321","Title":null,"Body":"<p>This isn't 100% the answer to the question, but it's a similar way of solving it, based on a MNIST NN training example suggested in the comments to the question.<\/p>\n\n<p>Based on the TensorFlow begginer MNIST tutorial, and thanks to <a href=\"http:\/\/opensourc.es\/blog\/tensorflow-mnist\" rel=\"noreferrer\">this tutorial<\/a>, this is a way of training and using your Neural Network with custom data.<\/p>\n\n<p>Please note that similar should be done for tutorials such as the CIFAR10, as @Yaroslav Bulatov mentioned in the comments.<\/p>\n\n<pre><code>import input_data\nimport datetime\nimport numpy as np\nimport tensorflow as tf\nimport cv2\nfrom matplotlib import pyplot as plt\nimport matplotlib.image as mpimg\nfrom random import randint\n\n\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n\nx = tf.placeholder(\"float\", [None, 784])\n\nW = tf.Variable(tf.zeros([784,10]))\nb = tf.Variable(tf.zeros([10]))\n\ny = tf.nn.softmax(tf.matmul(x,W) + b)\ny_ = tf.placeholder(\"float\", [None,10])\n\ncross_entropy = -tf.reduce_sum(y_*tf.log(y))\n\ntrain_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)\n\ninit = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run(init)\n\n#Train our model\niter = 1000\nfor i in range(iter):\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#Evaluationg our model:\ncorrect_prediction=tf.equal(tf.argmax(y,1), tf.argmax(y_,1))\n\naccuracy=tf.reduce_mean(tf.cast(correct_prediction,\"float\"))\nprint \"Accuracy: \", sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels})\n\n#1: Using our model to classify a random MNIST image from the original test set:\nnum = randint(0, mnist.test.images.shape[0])\nimg = mnist.test.images[num]\n\nclassification = sess.run(tf.argmax(y, 1), feed_dict={x: [img]})\n'''\n#Uncomment this part if you want to plot the classified image.\nplt.imshow(img.reshape(28, 28), cmap=plt.cm.binary)\nplt.show()\n'''\nprint 'Neural Network predicted', classification[0]\nprint 'Real label is:', np.argmax(mnist.test.labels[num])\n\n\n#2: Using our model to classify MNIST digit from a custom image:\n\n# create an an array where we can store 1 picture\nimages = np.zeros((1,784))\n# and the correct values\ncorrect_vals = np.zeros((1,10))\n\n# read the image\ngray = cv2.imread(\"my_digit.png\", 0 ) #0=cv2.CV_LOAD_IMAGE_GRAYSCALE #must be .png!\n\n# rescale it\ngray = cv2.resize(255-gray, (28, 28))\n\n# save the processed images\ncv2.imwrite(\"my_grayscale_digit.png\", gray)\n\"\"\"\nall images in the training set have an range from 0-1\nand not from 0-255 so we divide our flatten images\n(a one dimensional vector with our 784 pixels)\nto use the same 0-1 based range\n\"\"\"\nflatten = gray.flatten() \/ 255.0\n\"\"\"\nwe need to store the flatten image and generate\nthe correct_vals array\ncorrect_val for a digit (9) would be\n[0,0,0,0,0,0,0,0,0,1]\n\"\"\"\nimages[0] = flatten\n\n\nmy_classification = sess.run(tf.argmax(y, 1), feed_dict={x: [images[0]]})\n\n\"\"\"\nwe want to run the prediction and the accuracy function\nusing our generated arrays (images and correct_vals)\n\"\"\"\nprint 'Neural Network predicted', my_classification[0], \"for your digit\"\n<\/code><\/pre>\n\n<p>For further image conditioning (digits should be completely dark in a white background) and better NN training (accuracy>91%) please check the Advanced MNIST tutorial from TensorFlow or the 2nd tutorial i've mentioned.<\/p>\n"},{"AnswerId":"36364242","CreationDate":"2016-04-01T19:18:16.890","ParentId":null,"OwnerUserId":"6146626","Title":null,"Body":"<p>The below example is not for the mnist tutorial, but a simple XOR example. Note the <code>train()<\/code> and <code>test()<\/code> methods. All that we declare &amp; keep globally are the weights, biases, and session. In the test method we redefine the shape of the input and reuse the same weights &amp; biases (and session) that we refined in training. <\/p>\n\n<pre><code>import tensorflow as tf\n\n#parameters for the net\nw1 = tf.Variable(tf.random_uniform(shape=[2,2], minval=-1, maxval=1, name='weights1'))\nw2 = tf.Variable(tf.random_uniform(shape=[2,1], minval=-1, maxval=1, name='weights2'))\n\n#biases\nb1 = tf.Variable(tf.zeros([2]), name='bias1')\nb2 = tf.Variable(tf.zeros([1]), name='bias2')\n\n#tensorflow session\nsess = tf.Session()\n\n\ndef train():\n\n    #placeholders for the traning inputs (4 inputs with 2 features each) and outputs (4 outputs which have a value of 0 or 1)\n    x = tf.placeholder(tf.float32, [4, 2], name='x-inputs')\n    y = tf.placeholder(tf.float32, [4, 1], name='y-inputs')\n\n    #set up the model calculations\n    temp = tf.sigmoid(tf.matmul(x, w1) + b1)\n    output = tf.sigmoid(tf.matmul(temp, w2) + b2)\n\n    #cost function is avg error over training samples\n    cost = tf.reduce_mean(((y * tf.log(output)) + ((1 - y) * tf.log(1.0 - output))) * -1)\n\n    #training step is gradient descent\n    train_step = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(cost)\n\n    #declare training data\n    training_x = [[0,1], [0,0], [1,0], [1,1]]\n    training_y = [[1], [0], [1], [0]]\n\n    #init session\n    init = tf.initialize_all_variables()\n    sess.run(init)\n\n    #training\n    for i in range(100000):\n        sess.run(train_step, feed_dict={x:training_x, y:training_y})\n\n        if i % 1000 == 0:\n            print (i, sess.run(cost, feed_dict={x:training_x, y:training_y}))\n\n    print '\\ntraining done\\n'\n\n\ndef test(inputs):\n    #redefine the shape of the input to a single unit with 2 features\n    xtest = tf.placeholder(tf.float32, [1, 2], name='x-inputs')\n\n    #redefine the model in terms of that new input shape\n    temp = tf.sigmoid(tf.matmul(xtest, w1) + b1)\n    output = tf.sigmoid(tf.matmul(temp, w2) + b2)\n\n    print (inputs, sess.run(output, feed_dict={xtest:[inputs]})[0, 0] &gt;= 0.5)\n\n\ntrain()\n\ntest([0,1])\ntest([0,0])\ntest([1,1])\ntest([1,0])\n<\/code><\/pre>\n"}]}
{"QuestionId":33785203,"AnswerCount":1,"Tags":"<video><tensorflow>","CreationDate":"2015-11-18T16:23:30.007","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"Tensorflow - classification of videos","Body":"<p>Is the open-source version of TensorFlow suitable for classification of videos - akin the examples stated in <a href=\"http:\/\/vision.stanford.edu\/pdf\/karpathy14.pdf\">http:\/\/vision.stanford.edu\/pdf\/karpathy14.pdf<\/a> ? If not - are there any other packages in the public domain that could be used for this purpose? <\/p>\n","answers":[{"AnswerId":"33791476","CreationDate":"2015-11-18T22:13:31.237","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It would be possible to modify the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/models\/image\">image models<\/a> shipped with TensorFlow to train a neural network on such video data.<\/p>\n\n<p>The main challenges would be:<\/p>\n\n<ol>\n<li><p>Creating a <a href=\"http:\/\/tensorflow.org\/how_tos\/new_data_formats\/index.html#custom-data-readers\">custom reader<\/a> that handled the appropriate video file format (or converting your video data to one of the supported format).<\/p><\/li>\n<li><p>Using the <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/image.html#images\">image ops<\/a> to generate the multi-resolution images required in the Karpathy et al. model.<\/p><\/li>\n<li><p>(Possibly) <a href=\"http:\/\/tensorflow.org\/how_tos\/threading_and_queues\/index.html\">using queues<\/a> to perform pipelined video processing and ensure that I\/O doesn't become a bottleneck.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":33785936,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-18T16:57:38.137","AcceptedAnswerId":33786141.0,"OwnerUserId":3849791.0,"Title":"TensorFlow Error found in Tutorial","Body":"<p>Dare I even ask? This is such a new technology at this point that I can't find a way to solve this seemingly simple error. The tutorial I'm going over can be found here- <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/index.html#deep-mnist-for-experts\">http:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/index.html#deep-mnist-for-experts<\/a><\/p>\n\n<p>I literally copied and pasted all of the code into IPython Notebook and at the very last chunk of code I get an error.<\/p>\n\n<pre><\/pre>\n\n<p>After running this code, I receive this error.<\/p>\n\n<pre><\/pre>\n\n<p>I thought that I may need to install or reinstall TensorFlow via conda install <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\">https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/a> but conda doesn't even know how to install it.<\/p>\n\n<p>Does anyone have any idea of how to work around this error?<\/p>\n","answers":[{"AnswerId":"33786141","CreationDate":"2015-11-18T17:09:07.890","ParentId":null,"OwnerUserId":"3849791","Title":null,"Body":"<p>I figured it out. As you see in the value error, it says <code>No default session is registered. Use 'with DefaultSession(sess)' or pass an explicit session to eval(session=sess)<\/code> so the answer I came up with is to pass an explicit session to eval, just like it says. Here is where I made the changes.<\/p>\n\n<pre><code>if i%100 == 0:\n        train_accuracy = accuracy.eval(session=sess, feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})\n<\/code><\/pre>\n\n<p>And<\/p>\n\n<pre><code>train_step.run(session=sess, feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})\n<\/code><\/pre>\n\n<p>Now the code is working fine.<\/p>\n"},{"AnswerId":"48031624","CreationDate":"2017-12-30T06:33:50.333","ParentId":null,"OwnerUserId":"5503531","Title":null,"Body":"<p>I encountered a similar error when I tried a simple tensorflow example.<\/p>\n\n<pre><code>import tensorflow as tf\nv = tf.Variable(10, name=\"v\")\nsess = tf.Session()\nsess.run(v.initializer)\nprint(v.eval())\n<\/code><\/pre>\n\n<p>My solution is to use sess.as_default(). For example, I changed my code to the following and it worked:<\/p>\n\n<pre><code>import tensorflow as tf\nv = tf.Variable(10, name=\"v\")\nwith tf.Session().as_default() as sess:\n  sess.run(v.initializer)      \n  print(v.eval())\n<\/code><\/pre>\n\n<p>Another solution can be use InteractiveSession. The difference between InteractiveSession and Session is that an InteractiveSession makes itself the default session so you can run() or eval() without explicitly call the session.<\/p>\n\n<pre><code>v = tf.Variable(10, name=\"v\")\nsess = tf.InteractiveSession()\nsess.run(v.initializer)\nprint(v.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":33786879,"AnswerCount":0,"Tags":"<computer-vision><image-segmentation><deep-learning><caffe><vision>","CreationDate":"2015-11-18T17:48:17.363","AcceptedAnswerId":null,"OwnerUserId":5577837.0,"Title":"Crude image clustering using a Caffe model?","Body":"<p>Let's say I have a Caffe Model that can recognize textures. Is there a straightforward way to get a crude segmentation of an image using the classification results for a sliding window? <\/p>\n","answers":[]}
{"QuestionId":33788989,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2015-11-18T19:45:42.470","AcceptedAnswerId":33846659.0,"OwnerUserId":189456.0,"Title":"Tensorflow: Using Adam optimizer","Body":"<p>I am experimenting with some simple models in tensorflow, including one that looks very similar to the first <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.md\">MNIST for ML Beginners example<\/a>, but with a somewhat larger dimensionality. I am able to use the gradient descent optimizer with no problems, getting good enough convergence. When I try to use the ADAM optimizer, I get errors like this:<\/p>\n\n<pre><\/pre>\n\n<p>where the specific variable that complains about being uninitialized changes depending on the run. What does this error mean? And what does it suggest is wrong? It seems to occur regardless of the learning rate I use.<\/p>\n","answers":[{"AnswerId":"43773554","CreationDate":"2017-05-04T03:47:51.413","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/q\/34001922\/1090562\">FailedPreconditionError: Attempting to use uninitialized value<\/a> is one of the most frequent errors related to tensorflow. From official documentation, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/errors\/FailedPreconditionError\" rel=\"nofollow noreferrer\">FailedPreconditionError<\/a> <\/p>\n\n<blockquote>\n  <p>This exception is most commonly raised when running an operation that\n  reads a tf.Variable before it has been initialized.<\/p>\n<\/blockquote>\n\n<p>In your case the error even explains what variable was not initialized: <code>Attempting to use uninitialized value Variable_1<\/code>. One of the TF tutorials explains a lot about variables, their <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables\" rel=\"nofollow noreferrer\">creation\/initialization\/saving\/loading<\/a><\/p>\n\n<p>Basically to initialize the variable you have 3 options:<\/p>\n\n<ul>\n<li>initialize all global variables with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/global_variables_initializer\" rel=\"nofollow noreferrer\"><code>tf.global_variables_initializer()<\/code><\/a><\/li>\n<li>initialize variables you care about with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variables_initializer\" rel=\"nofollow noreferrer\"><code>tf.variables_initializer(list_of_vars)<\/code><\/a>. Notice that you can use this function to mimic global_variable_initializer: <code>tf.variable_initializers(tf.global_variables())<\/code><\/li>\n<li>initialize only one variable with <code>var_name.initializer<\/code><\/li>\n<\/ul>\n\n<p>I almost always use the first approach. Remember you should put it inside a session run. So you will get something like this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n\n<p>If your are curious about more information about variables, read <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/state_ops\" rel=\"nofollow noreferrer\">this documentation<\/a> to know how to <code>report_uninitialized_variables<\/code> and check <code>is_variable_initialized<\/code>.<\/p>\n"},{"AnswerId":"47542608","CreationDate":"2017-11-28T23:23:44.217","ParentId":null,"OwnerUserId":"4876765","Title":null,"Body":"<p>You need to call <code>tf.global_variables_initializer()<\/code> on you session, like<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\nsess.run(init)\n<\/code><\/pre>\n\n<p>Full example is available in this great tutorial\n<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/mechanics\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/mechanics<\/a><\/p>\n"},{"AnswerId":"46275909","CreationDate":"2017-09-18T09:27:46.110","ParentId":null,"OwnerUserId":"7299740","Title":null,"Body":"<p>run init after AdamOptimizer\uff0cand without define init before or run init<\/p>\n\n<p>sess.run(tf.initialize_all_variables()) <\/p>\n\n<p>or <\/p>\n\n<p>sess.run(tf.global_variables_initializer())<\/p>\n"},{"AnswerId":"43648306","CreationDate":"2017-04-27T03:44:56.740","ParentId":null,"OwnerUserId":"5412161","Title":null,"Body":"<p>I was having a similar problem. (No problems training with GradientDescent optimizer, but error raised when using to Adam Optimizer, or any other optimizer with its own variables)<\/p>\n\n<p>Changing to an interactive session solved this problem for me.<\/p>\n\n<pre><code>sess = tf.Session()\n<\/code><\/pre>\n\n<p>into<\/p>\n\n<pre><code>sess = tf.InteractiveSession()\n<\/code><\/pre>\n"},{"AnswerId":"33846659","CreationDate":"2015-11-21T17:53:30.513","ParentId":null,"OwnerUserId":"5535547","Title":null,"Body":"<p>The AdamOptimizer class creates additional variables, called \"slots\", to hold values for the \"m\" and \"v\" accumulators.<\/p>\n\n<p>See the source here if you're curious, it's actually quite readable:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/adam.py#L39\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/adam.py#L39<\/a> . Other optimizers, such as Momentum and Adagrad use slots too.<\/p>\n\n<p>These variables must be initialized before you can train a model.<\/p>\n\n<p>The normal way to initialize variables is to call <code>tf.initialize_all_variables()<\/code> which adds ops to initialize the variables present in the graph <em>when it is called<\/em>.<\/p>\n\n<p>(Aside: unlike its name suggests, initialize_all_variables() does not initialize anything, it only add ops that will initialize the variables when run.)<\/p>\n\n<p>What you must do is call initialize_all_variables() <em>after<\/em> you have added the optimizer:<\/p>\n\n<pre><code>...build your model...\n# Add the optimizer\ntrain_op = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)\n# Add the ops to initialize variables.  These will include \n# the optimizer slots added by AdamOptimizer().\ninit_op = tf.initialize_all_variables()\n\n# launch the graph in a session\nsess = tf.Session()\n# Actually intialize the variables\nsess.run(init_op)\n# now train your model\nfor ...:\n  sess.run(train_op)\n<\/code><\/pre>\n"}]}
{"QuestionId":33789485,"AnswerCount":2,"Tags":"<python><csv><numpy><tensorflow>","CreationDate":"2015-11-18T20:14:10.353","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"Training TensorFlow for Predicting a Column in a csv file","Body":"<p>I have data that is structured in a csv file. I want to be able to predict whether column 1 is going to be a 1 or a 0 given all other columns. How do I go about training the program (preferably using Neural Networks) to use all of the given data in order to make that prediction. Is there code that someone can show me? I've tried feeding it ,  (sorry if I spelt that wrong), and a ; nothing has worked yet. Here is the code I am running until I get the error-<\/p>\n\n<pre><\/pre>\n\n<p>At which point I run into this error-<\/p>\n\n<pre><\/pre>\n\n<p>Any help is greatly appreciated. All I need to do is predict whether column 1 is going to be a 1 or a 0. Even if all you do is get me past this one error, I should be able to take it from there.<\/p>\n\n<p>EDIT: This is what the csv looks like when I print it out.<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to predict the first column.<\/p>\n","answers":[{"AnswerId":"33832376","CreationDate":"2015-11-20T17:12:16.410","ParentId":null,"OwnerUserId":"904032","Title":null,"Body":"<p>The following reads from a CSV file and builds a tensorflow program. The example uses the Iris data set, since that maybe a more meaningful example. However, it should probably work for your data as well.<\/p>\n\n<p>Please note, the first column will be [0,1 or 2], since there are 3 species of iris. <\/p>\n\n<pre><code>#!\/usr\/bin\/env python\nimport tensorflow as tf\nimport numpy as np\nfrom numpy import genfromtxt\n\n# Build Example Data is CSV format, but use Iris data\nfrom sklearn import datasets\nfrom sklearn.cross_validation import train_test_split\nimport sklearn\ndef buildDataFromIris():\n    iris = datasets.load_iris()\n    X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.33, random_state=42)\n    f=open('cs-training.csv','w')\n    for i,j in enumerate(X_train):\n        k=np.append(np.array(y_train[i]),j   )\n        f.write(\",\".join([str(s) for s in k]) + '\\n')\n    f.close()\n    f=open('cs-testing.csv','w')\n    for i,j in enumerate(X_test):\n        k=np.append(np.array(y_test[i]),j   )\n        f.write(\",\".join([str(s) for s in k]) + '\\n')\n    f.close()\n\n\n# Convert to one hot\ndef convertOneHot(data):\n    y=np.array([int(i[0]) for i in data])\n    y_onehot=[0]*len(y)\n    for i,j in enumerate(y):\n        y_onehot[i]=[0]*(y.max() + 1)\n        y_onehot[i][j]=1\n    return (y,y_onehot)\n\n\nbuildDataFromIris()\n\n\ndata = genfromtxt('cs-training.csv',delimiter=',')  # Training data\ntest_data = genfromtxt('cs-testing.csv',delimiter=',')  # Test data\n\nx_train=np.array([ i[1::] for i in data])\ny_train,y_train_onehot = convertOneHot(data)\n\nx_test=np.array([ i[1::] for i in test_data])\ny_test,y_test_onehot = convertOneHot(test_data)\n\n\n#  A number of features, 4 in this example\n#  B = 3 species of Iris (setosa, virginica and versicolor)\nA=data.shape[1]-1 # Number of features, Note first is y\nB=len(y_train_onehot[0])\ntf_in = tf.placeholder(\"float\", [None, A]) # Features\ntf_weight = tf.Variable(tf.zeros([A,B]))\ntf_bias = tf.Variable(tf.zeros([B]))\ntf_softmax = tf.nn.softmax(tf.matmul(tf_in,tf_weight) + tf_bias)\n\n# Training via backpropagation\ntf_softmax_correct = tf.placeholder(\"float\", [None,B])\ntf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax))\n\n# Train using tf.train.GradientDescentOptimizer\ntf_train_step = tf.train.GradientDescentOptimizer(0.01).minimize(tf_cross_entropy)\n\n# Add accuracy checking nodes\ntf_correct_prediction = tf.equal(tf.argmax(tf_softmax,1), tf.argmax(tf_softmax_correct,1))\ntf_accuracy = tf.reduce_mean(tf.cast(tf_correct_prediction, \"float\"))\n\n# Initialize and run\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\nprint(\"...\")\n# Run the training\nfor i in range(30):\n    sess.run(tf_train_step, feed_dict={tf_in: x_train, tf_softmax_correct: y_train_onehot})\n\n# Print accuracy\n    result = sess.run(tf_accuracy, feed_dict={tf_in: x_test, tf_softmax_correct: y_test_onehot})\n    print \"Run {},{}\".format(i,result)\n\n\n\"\"\"\nBelow is the ouput\n  ...\n  Run 0,0.319999992847\n  Run 1,0.300000011921\n  Run 2,0.379999995232\n  Run 3,0.319999992847\n  Run 4,0.300000011921\n  Run 5,0.699999988079\n  Run 6,0.680000007153\n  Run 7,0.699999988079\n  Run 8,0.680000007153\n  Run 9,0.699999988079\n  Run 10,0.680000007153\n  Run 11,0.680000007153\n  Run 12,0.540000021458\n  Run 13,0.419999986887\n  Run 14,0.680000007153\n  Run 15,0.699999988079\n  Run 16,0.680000007153\n  Run 17,0.699999988079\n  Run 18,0.680000007153\n  Run 19,0.699999988079\n  Run 20,0.699999988079\n  Run 21,0.699999988079\n  Run 22,0.699999988079\n  Run 23,0.699999988079\n  Run 24,0.680000007153\n  Run 25,0.699999988079\n  Run 26,1.0\n  Run 27,0.819999992847\n  ...\n\n Ref:\n https:\/\/gist.github.com\/mchirico\/bcc376fb336b73f24b29#file-tensorflowiriscsv-py\n\"\"\"\n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"},{"AnswerId":"33790863","CreationDate":"2015-11-18T21:35:16.193","ParentId":null,"OwnerUserId":"538524","Title":null,"Body":"<p>You just need to provide an input matching your x,y_ shape.<\/p>\n\n<pre><code>x = tf.placeholder(\"float\", [None, 11])\ny_ = tf.placeholder(\"float\", [None,2])\n<\/code><\/pre>\n\n<p>So instead of data.train.next_batch(100) create and use a function \"my_csv_batch(count)\" which returns an array of shape [[count,11], [count,2]]  The first set of arrays is your x and the next is your y_s labels<\/p>\n\n<p>my_csv_batch will return a batch (perhaps stochastically generated if you have a lot of data)  from your csv file.<\/p>\n\n<p>Btw, you'll also need something similar for doing  your eval as well.  You'll have to generate a batch of data and labels similarly.<\/p>\n"}]}
{"QuestionId":33789559,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe>","CreationDate":"2015-11-18T20:17:26.543","AcceptedAnswerId":33799546.0,"OwnerUserId":937153.0,"Title":"Caffe HDF5 pixel-wise classification","Body":"<p>I am trying to implement a pixel-wise binary classification for images using caffe. For each image having dimension  3x256x256, I have a 256x256 label array in which each entry is marked as either 0 or 1. Also, when I read my HDF5 file using the below code, <\/p>\n\n<pre><\/pre>\n\n<p>I get the data and label dimension as <\/p>\n\n<pre><\/pre>\n\n<p>I am trying to feed this data into the given hdf5 classification network and while training, I have the following output(using the default solver, but in GPU mode). <\/p>\n\n<pre><\/pre>\n\n<p>gives<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Basically the error is <\/p>\n\n<pre><\/pre>\n\n<p>I am not able to understand why the number of labels expected is just same as my batch size. How exactly should I tackle this problem ? Is this a problem with my labeling method ? <\/p>\n","answers":[{"AnswerId":"33799546","CreationDate":"2015-11-19T09:14:51.923","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your problem is that <code>\"SoftmaxWithLoss\"<\/code> layer tries to compare a prediction vector of 2 elements per input image to a label of size 256-by-256 per image.<br>\nThis makes no sense.<\/p>\n\n<p><strong>Root cause of the error:<\/strong> I guess what you tired to do is to have a binary classifier applied to <em>each pixel<\/em> in the image. To that end you defined \"fc1\" as an <code>\"InnerProduct\"<\/code> layer  with <code>num_output=2<\/code>. However, the way caffe sees this is that you have a single binary classifier applied to the <em>entire<\/em> image. Thus caffe gives you a <em>single<\/em> binary prediction to the entire image. <\/p>\n\n<p><strong>How to solve:<\/strong> when working on pixel-wise predictions you no longer need to use <code>\"InnerProduct\"<\/code> layers and you have a \"fully convolutional net\". If you <strong>replace<\/strong> \"fc1\" with a conv layer (for instance a kernel that examine the 5-by-5 environment of each pixel and makes a decision according to this patch):<\/p>\n\n<pre><code>layer {\n  name: \"bin_class\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"bin_class\"\n  convolution_param {\n    num_output: 2 # binary class output\n    kernel_size: 5 # 5-by-5 patch for prediciton\n    pad: 2 # make sure spatial output size equals size of label \n  }\n}\n<\/code><\/pre>\n\n<p>Now applying <code>\"SoftmaxWithLoss\"<\/code> to <code>bottom: bin_class<\/code> and <code>bottom: label<\/code> should work.<\/p>\n"}]}
{"QuestionId":33790135,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-18T20:53:35.003","AcceptedAnswerId":null,"OwnerUserId":5568428.0,"Title":"Compiling tensorflow\/models\/rnn\/translate:translate with local numpy","Body":"\n\n<p>I am trying to run the neural machine translation demo on a gpu. The GPU example in tensorflow's getting started page works.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>produces the expected output.<\/p>\n\n<p>But when I try to compile the <a href=\"http:\/\/tensorflow.org\/tutorials\/seq2seq\/index.html\" rel=\"nofollow noreferrer\">translation demo<\/a>:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>it fails:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Probably because of<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>This is probably because the global numpy installation is old and does not know PyArray_SHAPE. I don't have admin rights to update it globally, but I have installed an updated numpy in $HOME\/.local\/lib\/python2.7\/site-packages\/ using . If I add the path to the corresponding rule in tensorflow\/tensorflow\/python\/BUILD like this:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>it complains:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>How can I tell tensorflow to use the local numpy version?<\/p>\n\n<p>(gcc 4.9.3, bleeding edge tensorflow + bazel, local numpy 1.10.1, ubuntu 12.04)<\/p>\n\n<p>EDIT:<\/p>\n\n<p>When I follow the instructions from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/109\" rel=\"nofollow noreferrer\">here<\/a> as suggested by syncd I get<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Various attempts to add them to hdrs in the tf_cuda_library rule do not help:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"33842934","CreationDate":"2015-11-21T11:45:06.860","ParentId":null,"OwnerUserId":"5568428","Title":null,"Body":"<p>Here is a <strong>very<\/strong> nasty workaround if you get the \"undeclared inclusion(s) in rule\" errors:<\/p>\n\n<p>1.) Pick a path in your gccs build-in include directories (i.e. one of the cxx_builtin_include_directory paths in bazel-workspace\/tools\/cpp\/CROSSTOOL - should be equal to those given by g++ -v bla.cc)<\/p>\n\n<p>2.) Lets say you picked directory <em>dir<\/em> .Create a directory called \"bla\" within <em>dir<\/em><\/p>\n\n<p>3.) In <em>dir<\/em>\/bla, create a symlink to your numpy include directory (ln -s ...\/core\/include\/numpy .)<\/p>\n\n<p>4.) Add \"<em>dir<\/em>\/bla\" to  tensorflow\/python\/BUILD and tensorflow\/tensorflow.bzl as described in the link.<\/p>\n\n<p>5.) Feel guilty but happy that it compiles<\/p>\n"},{"AnswerId":"33802841","CreationDate":"2015-11-19T11:37:20.347","ParentId":null,"OwnerUserId":"5580978","Title":null,"Body":"<p>One possible solution is presented <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/109\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>Create a link to <code>$HOME\/.local\/lib\/python2.7\/site-packages\/numpy\/core\/include\/numpy<\/code> in the <code>tensorflow\/third_party<\/code> dir and edit <code>-Ithird_party<\/code> to <code>tensorflow\/python\/build<\/code> and <code>tensorflow\/tensorflow.bzl<\/code><\/p>\n"}]}
{"QuestionId":33790366,"AnswerCount":3,"Tags":"<amazon-web-services><neural-network><deep-learning><caffe>","CreationDate":"2015-11-18T21:06:03.360","AcceptedAnswerId":33796339.0,"OwnerUserId":3929051.0,"Title":"Caffe | Check failed: error == cudaSuccess (2 vs. 0) out of memory","Body":"<p>I am trying to train a network on Caffe. I have image size of 512x640. Batch size is 1. I'm trying to implement <a href=\"https:\/\/gist.github.com\/shelhamer\/91eece041c19ff8968ee\" rel=\"noreferrer\">FCN-8s<\/a>. <\/p>\n\n<p>I am currently running this on a Amazon EC2 instance (g2.2xlarge) with 4GB of GPU memory. But when I run the solver, it immediately throws out an error<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Can someone help me proceed from here?<\/p>\n","answers":[{"AnswerId":"33796339","CreationDate":"2015-11-19T06:00:20.990","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The error you get is indeed out of memory, but it's not the RAM, but rather GPU memory (note that the error comes from CUDA).<br>\nUsually, when caffe is out of memory - the first thing to do is reduce the batch size (at the cost of <a href=\"https:\/\/stackoverflow.com\/a\/33717093\/1714410\">gradient accuracy<\/a>), but since you are already at batch size = 1...<br>\nAre you sure batch size is 1 for <strong>both<\/strong> TRAIN and TEST phases?<\/p>\n"},{"AnswerId":"57689098","CreationDate":"2019-08-28T09:24:52.430","ParentId":null,"OwnerUserId":"11945652","Title":null,"Body":"<p>I was facing a similar issue when running Deeplab v2 on a PC with following configuration: <\/p>\n\n<pre><code>----------\nOS: Ubuntu 18.04.3 LTS (64-bit)\n----------\nProcessor: Intel Core i7-6700k CPU @ 4.00 GHz x 8\n----------\nGPU: GeForce GTX 780 (3022 MiB)\n----------\nRAM : 31.3 GiB\n----------\n<\/code><\/pre>\n\n<p>Changing both the test and training batch sizes to 1 didn't help me. But, changing the dimensions of the output image sure did! <\/p>\n"},{"AnswerId":"41191485","CreationDate":"2016-12-16T19:45:56.263","ParentId":null,"OwnerUserId":"470433","Title":null,"Body":"<p>Caffe can use multiple GPU's. This is only supported in the C++ interface, not in the python one. You could also enable cuDNN for a lower memory footprint.<\/p>\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"}]}
{"QuestionId":33794211,"AnswerCount":2,"Tags":"<docker><tensorflow>","CreationDate":"2015-11-19T02:29:17.257","AcceptedAnswerId":null,"OwnerUserId":4627745.0,"Title":"How to run Recurrent Neural Networks sample programs is tensorflow?","Body":"\n\n<p>I installed tensorflow in ubuntu using the following below commands<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>After installing all the basic programs whatever given in tensorflow website is working fine. But I am able to run Recurrent Neural Networks sample programs (whatever pseudocode they had given) becoz i cam to know if we install it using the pip  <strong>models\/rnn\/ptb<\/strong> folder and also some others folder are missing. How to solve this?\n          Later i came to know if we install it using the docker it installs all its dependencies. but having problems in docker installation also. whenevr i check (docker version command)it is showing, the following information<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>How to resolve this?(I installed docker using the below commands<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>How to solve this also?<\/p>\n","answers":[{"AnswerId":"33796704","CreationDate":"2015-11-19T06:27:54.523","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>To answer the question about the ptb data, you have to download and install the data as described on this page (because we cannot distribute it with our install):<\/p>\n\n<p><a href=\"http:\/\/www.tensorflow.org\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">http:\/\/www.tensorflow.org\/tutorials\/recurrent\/index.html<\/a><\/p>\n"},{"AnswerId":"33815715","CreationDate":"2015-11-19T22:42:39.250","ParentId":null,"OwnerUserId":"2282170","Title":null,"Body":"<p>I am not sure if your question is specific to Docker but I can tell you how I did it:<\/p>\n\n<ul>\n<li>cloned the git repository and compiled the tensorflow code from source<\/li>\n<li>generated installer using bazel and pip installed tensorflow<\/li>\n<li>build ptb_word_lm application, downloaded the data and ran the app as outlined in the link highlighted by @LudiMagister<\/li>\n<\/ul>\n\n<p>seems to work and I get about 2k wps in terms of speed with GPU support. No idea if that's any good but it gives you a benchmark.<\/p>\n\n<p>hope this helps.<\/p>\n"}]}
{"QuestionId":33794909,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2015-11-19T03:46:52.570","AcceptedAnswerId":33794927.0,"OwnerUserId":395857.0,"Title":"Installing Torch7 on Kubuntu 14.04 LTS x64: \"Error: Build error: Failed compiling object readline.o\"","Body":"<p>I followed the instructions from <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">http:\/\/torch.ch\/docs\/getting-started.html<\/a> to install Torch on Torch7 on Kubuntu 14.04 LTS x64.<\/p>\n\n<pre><\/pre>\n\n<p>When executing  it failed with this error message:<\/p>\n\n<pre><\/pre>\n\n<p>What did I miss?<\/p>\n","answers":[{"AnswerId":"33794927","CreationDate":"2015-11-19T03:49:04.747","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>To fix this issue you need to run:<\/p>\n\n<pre><code>sudo apt-get install -y libreadline-dev\n<\/code><\/pre>\n\n<p>then remove the build folder (otherwise it will give another error: <code>collect2: error: ld returned 1 exit status<\/code>) and run the installation commands again.<\/p>\n"}]}
{"QuestionId":33795982,"AnswerCount":1,"Tags":"<python><heroku><tensorflow>","CreationDate":"2015-11-19T05:34:13.677","AcceptedAnswerId":33796439.0,"OwnerUserId":2056772.0,"Title":"undefined symbol: PyUnicodeUCS4_FromStringAndSize with tensorflow on heroku","Body":"<p>trying to install tensorflow on heroku, in the end <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/24\" rel=\"nofollow\">it fails<\/a> with:<\/p>\n\n<pre><\/pre>\n\n<p>requirements.txt:<\/p>\n\n<pre><\/pre>\n\n<p>install:<\/p>\n\n<pre><\/pre>\n\n<p>looks good ! unfortunately, when i try:<\/p>\n\n<pre><\/pre>\n\n<p>anything i can do about it ?<\/p>\n","answers":[{"AnswerId":"33796439","CreationDate":"2015-11-19T06:08:08.667","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This issue is caused by your Python binary having an incompatible Unicode definition (UCS2) from the one assumed by the TensorFlow binary (UCS4).<\/p>\n\n<p>At present, the best solution is to build TensorFlow <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.html#source\" rel=\"noreferrer\">from source<\/a>. The latest version does not use Unicode strings in the native extension, so this problem should not arise. When updated binaries are available, they will include this fix.<\/p>\n\n<p>If you can't build from source on the target machine (for example when using Heroku), one option is to <a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.html#create-pip\" rel=\"noreferrer\">build your own PIP package<\/a> on a different machine (with the same architecture), and add that to your <code>requirements.txt<\/code> instead.<\/p>\n"}]}
{"QuestionId":33801624,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2015-11-19T10:44:02.220","AcceptedAnswerId":null,"OwnerUserId":3054356.0,"Title":"Graph Model evaluation index error","Body":"<p>I am trying to learn Keras GraphNN using <a href=\"http:\/\/keras.io\/models\/\" rel=\"nofollow noreferrer\">simple examples<\/a>. I have a simple example dataset with 784 features and I want to run this example:<\/p>\n\n<pre><\/pre>\n\n<p>The documentation mentions that :<\/p>\n\n<blockquote>\n  <p><strong>evaluate<\/strong>(data, batch_size=128, verbose=1): Show performance of the model over some validation data.\n  <strong>Return<\/strong>: The loss score over the data.\n  <strong>Arguments<\/strong>: Same meaning as fit method above. verbose is used as a binary flag (progress bar or nothing).<\/p>\n<\/blockquote>\n\n<p>And from the definition of the  we know that:<\/p>\n\n<blockquote>\n  <p><strong>Arguments<\/strong>:\n  <strong>data<\/strong>:dictionary mapping input names out outputs names to appropriate numpy arrays. All arrays should contain the same number of samples.<\/p>\n<\/blockquote>\n\n<p>Although my fit method runs perfect, I get  on <\/p>\n\n<p>My input shapes are:<\/p>\n\n<pre><\/pre>\n\n<p>What am I missing here ?<\/p>\n","answers":[]}
{"QuestionId":33802336,"AnswerCount":5,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2015-11-19T11:14:26.003","AcceptedAnswerId":33816991.0,"OwnerUserId":867505.0,"Title":"Visualizing output of convolutional layer in tensorflow","Body":"\n\n<p>I'm trying to visualize the output of a convolutional layer in tensorflow using the function . I'm already using it successfully in other instances (e. g. visualizing the input image), but have some difficulties reshaping the output here correctly. I have the following conv layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>So the output of  would have the shape . Just using  Doesn't account for the 32 different kernels, so I'm basically slicing through different feature maps here.<\/p>\n\n<p>How can I reshape them correctly? Or is there another helper function I could use for including this output in the summary?<\/p>\n","answers":[{"AnswerId":"35247876","CreationDate":"2016-02-06T23:36:20.960","ParentId":null,"OwnerUserId":"5766785","Title":null,"Body":"<p>In case someone would like to \"jump\" to numpy and visualize \"there\" here is an example how to display both <code>Weights<\/code> and <code>processing result<\/code>. All transformations are based on prev answer by <code>mdaoust<\/code>.   <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># to visualize 1st conv layer Weights\nvv1 = sess.run(W_conv1)\n\n# to visualize 1st conv layer output\nvv2 = sess.run(h_conv1,feed_dict = {img_ph:x, keep_prob: 1.0})\nvv2 = vv2[0,:,:,:]   # in case of bunch out - slice first img\n\n\ndef vis_conv(v,ix,iy,ch,cy,cx, p = 0) :\n    v = np.reshape(v,(iy,ix,ch))\n    ix += 2\n    iy += 2\n    npad = ((1,1), (1,1), (0,0))\n    v = np.pad(v, pad_width=npad, mode='constant', constant_values=p)\n    v = np.reshape(v,(iy,ix,cy,cx)) \n    v = np.transpose(v,(2,0,3,1)) #cy,iy,cx,ix\n    v = np.reshape(v,(cy*iy,cx*ix))\n    return v\n\n# W_conv1 - weights\nix = 5  # data size\niy = 5\nch = 32   \ncy = 4   # grid from channels:  32 = 4x8\ncx = 8\nv  = vis_conv(vv1,ix,iy,ch,cy,cx)\nplt.figure(figsize = (8,8))\nplt.imshow(v,cmap=\"Greys_r\",interpolation='nearest')\n\n#  h_conv1 - processed image\nix = 30  # data size\niy = 30\nv  = vis_conv(vv2,ix,iy,ch,cy,cx)\nplt.figure(figsize = (8,8))\nplt.imshow(v,cmap=\"Greys_r\",interpolation='nearest')\n<\/code><\/pre>\n"},{"AnswerId":"39543059","CreationDate":"2016-09-17T05:11:09.737","ParentId":null,"OwnerUserId":"6783862","Title":null,"Body":"<p>you may try to get convolution layer activation image this way:<\/p>\n\n<pre><code>    h_conv1_features = tf.unpack(h_conv1, axis=3)\n    h_conv1_imgs = tf.expand_dims(tf.concat(1, h_conv1_features_padded), -1)\n<\/code><\/pre>\n\n<p>this gets one vertical stripe with all images concatenated vertically.<\/p>\n\n<p>if you want them padded (in my case of relu activations to pad with white line):<\/p>\n\n<pre><code>    h_conv1_features = tf.unpack(h_conv1, axis=3)\n    h_conv1_max = tf.reduce_max(h_conv1)\n    h_conv1_features_padded = map(lambda t: tf.pad(t-h_conv1_max, [[0,0],[0,1],[0,0]])+h_conv1_max, h_conv1_features)\n    h_conv1_imgs = tf.expand_dims(tf.concat(1, h_conv1_features_padded), -1)\n<\/code><\/pre>\n"},{"AnswerId":"57834168","CreationDate":"2019-09-07T13:34:49.473","ParentId":null,"OwnerUserId":"712452","Title":null,"Body":"<p>Another way, which I think very easy, is using the <code>get_operation_by_name<\/code> function. I had hard time visualizing the layers with other methods but this helped me.<\/p>\n\n<pre><code>#first, find out the operations, many of those are micro-operations such as add etc.\ngraph = tf.get_default_graph()\ngraph.get_operations()\n\n#choose relevant operations\nop_name = '...' \nop = graph.get_operation_by_name(op_name)\nout = sess.run([op.outputs[0]], feed_dict={x: img_batch, is_training: False}) \n#img_batch is a single image whose dimensions are (1,n,n,1). \n\n# out is the output of the layer, do whatever you want with the output\n#in my case, I wanted to see the output of a convolution layer\nout2 = np.array(out)\nprint(out2.shape)\n\n# determine, row, col, and fig size etc.\nfor each_depth in range(out2.shape[4]):\n    fig.add_subplot(rows, cols, each_depth+1)\n    plt.imshow(out2[0,0,:,:,each_depth], cmap='gray')\n<\/code><\/pre>\n\n<p>For example below is the input(colored cat) and output of the second conv layer in my model.<a href=\"https:\/\/i.stack.imgur.com\/F7kYo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/F7kYo.png\" alt=\"below\"><\/a> <\/p>\n\n<p>Note that I am aware this question is old and there are easier methods with Keras but for people who use an old model from other people (such as me), this may be useful.<\/p>\n"},{"AnswerId":"33816991","CreationDate":"2015-11-20T00:33:53.763","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>I don't know of a helper function but if you want to see all the filters you can pack them into one image with some fancy uses of <code>tf.transpose<\/code>.<\/p>\n\n<p>So if you have a tensor that's <code>images<\/code> x <code>ix<\/code> x <code>iy<\/code> x <code>channels<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; V = tf.Variable()\n&gt;&gt;&gt; print V.get_shape()\n\nTensorShape([Dimension(-1), Dimension(256), Dimension(256), Dimension(32)])\n<\/code><\/pre>\n\n<p>So in this example <code>ix = 256<\/code>, <code>iy=256<\/code>, <code>channels=32<\/code><\/p>\n\n<p>first slice off 1 image, and remove the <code>image<\/code> dimension<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>V = tf.slice(V,(0,0,0,0),(1,-1,-1,-1)) #V[0,...]\nV = tf.reshape(V,(iy,ix,channels))\n<\/code><\/pre>\n\n<p>Next add a couple of pixels of zero padding around the image<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ix += 4\niy += 4\nV = tf.image.resize_image_with_crop_or_pad(image, iy, ix)\n<\/code><\/pre>\n\n<p>Then reshape so that instead of 32 channels you have 4x8 channels, lets call them <code>cy=4<\/code> and <code>cx=8<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>V = tf.reshape(V,(iy,ix,cy,cx)) \n<\/code><\/pre>\n\n<p>Now the tricky part. <code>tf<\/code> seems to return results in C-order, numpy's default. <\/p>\n\n<p>The current order, if flattened, would list all the channels for the first pixel (iterating over <code>cx<\/code> and <code>cy<\/code>), before listing the channels of the second pixel (incrementing <code>ix<\/code>). Going across the rows of pixels (<code>ix<\/code>) before incrementing to the next row (<code>iy<\/code>).<\/p>\n\n<p>We want the order that would lay out the images in a grid.\nSo you go across a row of an image (<code>ix<\/code>), before stepping along the row of channels (<code>cx<\/code>), when you hit the end of the row of channels you step to the next row in the image (<code>iy<\/code>) and when you run out or rows in the image you increment to the next row of channels (<code>cy<\/code>). so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>V = tf.transpose(V,(2,0,3,1)) #cy,iy,cx,ix\n<\/code><\/pre>\n\n<p>Personally I prefer <code>np.einsum<\/code> for fancy transposes, for readability, but it's not in <code>tf<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/175\" rel=\"noreferrer\">yet<\/a>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>newtensor = np.einsum('yxYX-&gt;YyXx',oldtensor)\n<\/code><\/pre>\n\n<p>anyway, now that the pixels are in the right order, we can safely flatten it into a 2d tensor:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># image_summary needs 4d input\nV = tf.reshape(V,(1,cy*iy,cx*ix,1))\n<\/code><\/pre>\n\n<p>try <code>tf.image_summary<\/code> on that, you should get a grid of little images. <\/p>\n\n<p>Below is an image of  what one gets after following all the steps here.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7qYQ3.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7qYQ3.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"46487955","CreationDate":"2017-09-29T11:34:38.510","ParentId":null,"OwnerUserId":"4462831","Title":null,"Body":"<p>I personally try to tile every 2d-filter in a single image.<\/p>\n\n<p>For doing this <sub>-if i'm not terribly mistaken since I'm quite new to DL-<\/sub> I found out that it could be helpful to exploit the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/depth_to_space\" rel=\"nofollow noreferrer\">depth_to_space<\/a> function, since it takes a 4d tensor <\/p>\n\n<p><code>[batch, height, width, depth]<\/code><\/p>\n\n<p>and produces an output of shape<\/p>\n\n<p><code>[batch, height*block_size, width*block_size, depth\/(block_size*block_size)]<\/code> <\/p>\n\n<p>Where block_size is the number of \"tiles\" in the output image. The only limitation to this is that the depth should be the square of block_size, which is an integer, otherwise it cannot \"fill\" the resulting image correctly. \nA possible solution could be of padding the depth of the input tensor up to a depth that is accepted by the method, but I sill havn't tried this.<\/p>\n"}]}
{"QuestionId":33807398,"AnswerCount":2,"Tags":"<luajit><torch>","CreationDate":"2015-11-19T15:03:31.487","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch Cuda - Generates two processes on both GPU cores","Body":"<p>When I run;<\/p>\n\n<pre><\/pre>\n\n<p>in lua it automatically allocates two processes to two of the cores in my GPU. For example I get the following output in ;<\/p>\n\n<pre><\/pre>\n\n<p>I would like to be able to control which GPU the process goes on. I have tried;<\/p>\n\n<pre><\/pre>\n\n<p>but this just creates more processes on the GPU.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"33808446","CreationDate":"2015-11-19T15:49:03.583","ParentId":null,"OwnerUserId":"1417179","Title":null,"Body":"<p>You can control which GPU your process will run on <em>before<\/em> launching it using the <code>CUDA_VISIBLE_DEVICES<\/code> environment variable, <em>e.g.<\/em> to run only on CPU 0:<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0\nluajit your-script.lua\n<\/code><\/pre>\n"},{"AnswerId":"41614437","CreationDate":"2017-01-12T13:27:18.280","ParentId":null,"OwnerUserId":"5979961","Title":null,"Body":"<p>As the previous answer said, choosing GPU can be done using <code>CUDA_VISIBLE_DEVICES<\/code> environment variable on the command line before calling torch or torch-lua-script. This is a general way with CUDA and can be used with any application, not only torch. The number here can though clash with the number set in <code>cutorch.setDevice()<\/code> (also which instead is 1-based). One can select multiple specific GPU:s by a comma-separated list, for example:<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES=1,2\n<\/code><\/pre>\n\n<p>This will result in torch running only on GPU 1 and 2. More information can be found here:<\/p>\n\n<p><a href=\"https:\/\/devblogs.nvidia.com\/parallelforall\/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices\/\" rel=\"nofollow noreferrer\">https:\/\/devblogs.nvidia.com\/parallelforall\/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices\/<\/a><\/p>\n"}]}
{"QuestionId":33808368,"AnswerCount":2,"Tags":"<python><csv><tensorflow>","CreationDate":"2015-11-19T15:45:29.213","AcceptedAnswerId":33810040.0,"OwnerUserId":3849791.0,"Title":"How do I change the dtype in TensorFlow for a csv file?","Body":"<p>Here is the code that I am trying to run-<\/p>\n\n<pre><\/pre>\n\n<p>This code return the error below.<\/p>\n\n<pre><\/pre>\n\n<p>It the has a lot of information following it which I think is irrelevant to the problem. Obviously the problem is that a lot of the data that I am feeding to the program is not of the dtype int32. It's mostly float numbers. I've tried a few things to change the dtype like explicitly setting the  argument in  as well as the . Neither worked. It's an invalid argument. To top it all off, I don't know if this code is going to actually make a prediction on the data. I want it to predict whether col1 is going to be a 1 or a 0 and I don't see anything in the code that would hint that it's going to actually make that prediction. Maybe I'll save that question for a different thread. Any help is greatly appreciated!<\/p>\n","answers":[{"AnswerId":"33810040","CreationDate":"2015-11-19T17:02:29.007","ParentId":null,"OwnerUserId":"3849791","Title":null,"Body":"<p>The answer to changing the dtype is to just change the defaults like so-<\/p>\n\n<pre><code>record_defaults = [[1.], [1.], [1.], [1.], [1.], [1.], [1.], [1.], [1.], [1.], [1.]]\n<\/code><\/pre>\n\n<p>After you do that, if you print out col1, you'll receive this message.<\/p>\n\n<pre><code>Tensor(\"DecodeCSV_43:0\", shape=TensorShape([]), dtype=float32)\n<\/code><\/pre>\n\n<p>But there is another error that you will run into, <a href=\"https:\/\/stackoverflow.com\/questions\/33686464\/what-am-i-missing-from-this-csv-reader-for-tensorflow\/33686652#33686652\">which has been answered here.<\/a> To recap the answer, the workaround is to change <code>tf.concat<\/code> to <code>tf.pack<\/code> like so.<\/p>\n\n<pre><code>features = tf.pack([col2, col3, col4, col5, col6, col7, col8, col9, col10, col11])\n<\/code><\/pre>\n"},{"AnswerId":"33810130","CreationDate":"2015-11-19T17:06:38.837","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The interface to <code>tf.decode_csv()<\/code> is a little tricky. The <code>dtype<\/code> of each column is determined by the corresponding element of the <code>record_defaults<\/code> argument. The value for <code>record_defaults<\/code> in your code is interpreted as each column having <code>tf.int32<\/code> as its type, which leads to an error when it encounters floating-point data.<\/p>\n\n<p>Let's say you have the following CSV data, containing three integer columns, followed by a floating point column:<\/p>\n\n<pre><code>4, 8, 9, 4.5\n2, 5, 1, 3.7\n2, 2, 2, 0.1\n<\/code><\/pre>\n\n<p>Assuming all of the columns are <em>required<\/em>, you would build <code>record_defaults<\/code> as follows:<\/p>\n\n<pre><code>value = ...\n\nrecord_defaults = [tf.constant([], dtype=tf.int32),    # Column 0\n                   tf.constant([], dtype=tf.int32),    # Column 1\n                   tf.constant([], dtype=tf.int32),    # Column 2\n                   tf.constant([], dtype=tf.float32)]  # Column 3\n\ncol0, col1, col2, col3 = tf.decode_csv(value, record_defaults=record_defauts)\n\nassert col0.dtype == tf.int32\nassert col1.dtype == tf.int32\nassert col2.dtype == tf.int32\nassert col3.dtype == tf.float32\n<\/code><\/pre>\n\n<p>An empty value in <code>record_defaults<\/code> signifies that the value is required. Alternatively, if (e.g.) column 2 is allowed to have missing values, you would define <code>record_defaults<\/code> as follows:<\/p>\n\n<pre><code>record_defaults = [tf.constant([], dtype=tf.int32),     # Column 0\n                   tf.constant([], dtype=tf.int32),     # Column 1\n                   tf.constant([0], dtype=tf.int32),    # Column 2\n                   tf.constant([], dtype=tf.float32)]   # Column 3\n<\/code><\/pre>\n\n<p>The second part of your question concerns how to build and train a model that predicts the value of one of the columns from the input data. Currently, the program doesn't: it simply concatenates the columns into a single tensor, called <code>features<\/code>. You will need to define and train a model, that interprets that data. One of the simplest such approaches is linear regression, and you might find this tutorial on <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/1_linear_regression.py\">linear regression in TensorFlow<\/a> adaptable to your problem.<\/p>\n"}]}
{"QuestionId":33809869,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-19T16:54:17.860","AcceptedAnswerId":null,"OwnerUserId":3441991.0,"Title":"Tensorflow: data source can be dynamic?","Body":"<p>In all the examples I've seen in the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/index.html\" rel=\"nofollow\">Tensorflow documentation<\/a>, the input is parsed from a static source -- typically a file.\nIs it possible to construct a graph in which the data come from a different source?  In my use case, data comes from a message string from a server: I want to cast these messages, as they come, in a tensor and use these tensors as inputs for the computation part of the graph.<\/p>\n","answers":[{"AnswerId":"33810248","CreationDate":"2015-11-19T17:12:48.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to provide data from a dynamic source is to put one or more <a href=\"http:\/\/tensorflow.org\/api_docs\/python\/io_ops.html#placeholders\" rel=\"nofollow\"><code>tf.placeholder()<\/code><\/a> ops in your graph to represent the input tensors, and use the <a href=\"http:\/\/tensorflow.org\/how_tos\/reading_data\/index.html#Feeding\" rel=\"nofollow\">feed mechanism<\/a> to supply different values for those tensors each time you call <code>run()<\/code>. If you do this, it is possible to write arbitrary Python code to generate the input data, which could involve invoking a remote server - or even on the server side handling an incoming request.<\/p>\n\n<p>If you're doing this in C++, the <a href=\"http:\/\/tensorflow.org\/api_docs\/cc\/ClassSession.html\" rel=\"nofollow\"><code>tensorflow::Session<\/code><\/a> class offers the same ability to feed placeholder values, using the <a href=\"http:\/\/tensorflow.org\/api_docs\/cc\/ClassSession.html#virtual_Status_tensorflow_Session_Run\" rel=\"nofollow\"><code>Session::Run()<\/code><\/a> method. The fed values must be <a href=\"http:\/\/tensorflow.org\/api_docs\/cc\/ClassTensor.html\" rel=\"nofollow\"><code>tensorflow::Tensor<\/code><\/a> objects, which you can create by specifying a datatype and shape (list of dimensions). The <code>Tensor<\/code> objects have methods that allow you to access them as multi-dimensional arrays (like <code>Tensor::scalar&lt;T&gt;()<\/code>, <code>Tensor::matrix&lt;T&gt;()<\/code>, and <code>Tensor::tensor&lt;T, NDIMS&gt;()<\/code>) which allows you to fill in their values as follows:<\/p>\n\n<pre class=\"lang-c++ prettyprint-override\"><code>tensorflow::Tensor t(DT_FLOAT, tensorflow::TensorShape({2, 2}));\nt.matrix()(0, 0) = 1.0;\nt.matrix()(0, 1) = 0.0;\nt.matrix()(1, 0) = 0.0;\nt.matrix()(1, 1) = 1.0;\n<\/code><\/pre>\n\n<p>You can also use all the methods of the <a href=\"http:\/\/eigen.tuxfamily.org\/dox-devel\/unsupported\/group__CXX11__Tensor__Module.html\" rel=\"nofollow\">Eigen Tensor<\/a> library to build these values.<\/p>\n"}]}
{"QuestionId":33809920,"AnswerCount":2,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2015-11-19T16:57:00.333","AcceptedAnswerId":null,"OwnerUserId":2711403.0,"Title":"Trouble with Caffe network initialization","Body":"<p>Attached towards the end is the output of Caffe Training initialization. <\/p>\n\n<p>My network consists of 4 convolutional layers followed by 3 fully connected layers. <\/p>\n\n<p>I computed my filter and feature map sizes, and my calculations show them to be consistent. <\/p>\n\n<ol>\n<li>Input is an image of dimensions (195 X 65).  <\/li>\n<li>First conv1 layer filters are (32 X 13 X 5), where 32 is number of filters. Stride is 1 along both axes.  <\/li>\n<li>Each output map of conv1 should be of size (183 X 61).  <\/li>\n<li>A MAX Pooling layer of size (2 X 2) with stride 2.  <\/li>\n<li>Output of pooling layer should be of size (92 X 31).  <\/li>\n<li>This is followed by conv2 layer of size (48 X 7 X 3).  <\/li>\n<li>Each feature map of conv2 layer should be of size ( 86 X 29)  <\/li>\n<li>This is followed by an average pooling with window size (2 X 2).  <\/li>\n<li>The output of pooling layer is of size ( 43 X 15)  <\/li>\n<li>This is followed by conv3 layer of size (64 X 7 X 3), where 64 is number of filters.  <\/li>\n<li>The output of conv3 should, each be of size ( 37 X 13)  <\/li>\n<li>This is immediately followed by conv4 layer of size ( 64 X 5 X 3),  <\/li>\n<li>Each feature map of output of conv4 should be of size (33 X 11)  <\/li>\n<li>Conv4 is followed by an average pooling of window size (2 X 2)  <\/li>\n<li>The output of this is of size of ( 17 X 6)  <\/li>\n<li>This is followed by 3 fully connected layers, each with 300 output neurons.<\/li>\n<\/ol>\n\n<p>No Padding is done at any layer. Below is the output of caffe training initialization. It shows some errors at conv2, conv3 and then at fc6. \nI am not able to understand the source of those errors. Could someone please clarify ?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33884598","CreationDate":"2015-11-24T02:51:26.217","ParentId":null,"OwnerUserId":"5522934","Title":null,"Body":"<p>you can read the tutorial in the following page,there is the layer's explanation to show how inner_product layer be computed. You can check your dimensions of the FC6 or debug the code to trace the relevant variables.<a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html<\/a><\/p>\n"},{"AnswerId":"33887770","CreationDate":"2015-11-24T07:31:59.347","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Look at the stack trace of the error you got (I put it together from the different places it appeared in your log):<\/p>\n\n<blockquote>\n<pre><code>OpenCV Error: Assertion failed (0 &lt;= roi.x &amp;&amp; 0 &lt;= roi.width &amp;&amp; roi.x + roi.width &lt;= m.cols &amp;&amp; 0 &lt;= roi.y &amp;&amp; 0 &lt;= roi.height &amp;&amp; roi.y + roi.height &lt;= m.rows) in Mat, file \/home\/uujjwal\/libraries\/opencv-3.0.0\/modules\/core\/src\/matrix.cpp, line 495\nterminate called after throwing an instance of 'cv::Exception'\nwhat():  \/home\/uujjwal\/libraries\/opencv-3.0.0\/modules\/core\/src\/matrix.cpp:495: error: (-215) 0 &lt;= roi.x &amp;&amp; 0 &lt;= roi.width &amp;&amp; roi.x + roi.width &lt;= m.cols &amp;&amp; 0 &lt;= roi.y &amp;&amp; 0 &lt;= roi.height &amp;&amp; roi.y + roi.height &lt;= m.rows in function Mat\n*** Aborted at 1447952145 (unix time) try \"date -d @1447952145\" if you are using GNU date ***\nPC: @     0x7f9446ec75d7 __GI_raise\n*** SIGABRT (@0x9f71700003803) received by PID 14339 (TID 0x7f9430a6f700) from PID 14339; stack trace: ***\n@     0x7f945ba27130 (unknown)\n@     0x7f9446ec75d7 __GI_raise\n@     0x7f9446ec8cc8 __GI_abort\n@     0x7f94477cb9b5 (unknown)\n@     0x7f94477c9926 (unknown)\n@     0x7f94477c9953 (unknown)\n@     0x7f94477c9b73 (unknown)\n@     0x7f945269b58d cv::error()\n@     0x7f945269b6f0 cv::error()\n@     0x7f945253219a cv::Mat::Mat()\n@     0x7f945c14a3ea caffe::DataTransformer&lt;&gt;::Transform()\n@     0x7f945c1cb478 caffe::DataLayer&lt;&gt;::load_batch()\n@     0x7f945c1b0f98 caffe::BasePrefetchingDataLayer&lt;&gt;::InternalThreadEntry()\n@     0x7f945c144d64 caffe::InternalThread::entry()\n@     0x7f945bc4124a (unknown)\n@     0x7f945ba1fdf5 start_thread\n@     0x7f9446f881ad __clone\n<\/code><\/pre>\n<\/blockquote>\n\n<p>You can clearly see that:<\/p>\n\n<ol>\n<li>This is a <strong>single<\/strong> error, and not multiple ones.  <\/li>\n<li>It occurred in a <strong>thread<\/strong> that is not the main thread (and this is why its error message and trace are printed between the output lines of the main thread building the model).  <\/li>\n<li>This error happens when the pre-fetching thread in the data layer tries to transform one of the input images.<\/li>\n<\/ol>\n\n<p>What can you do?<\/p>\n\n<ul>\n<li>You can try and change the <code>prefetch: 4<\/code> parameter in your input data layer.  <\/li>\n<li>You can try to test with a different input lmdb constructed a new where you can verify that all images in the new dataset are indeed valid.  <\/li>\n<li>Reconstruct the input lmdb using <a href=\"https:\/\/stackoverflow.com\/questions\/31427094\/guide-to-use-convert-imageset-cpp\/31431716#\"><code>convert_imageset<\/code><\/a> utility, not trusting DIGITS to do the job for you.<\/li>\n<\/ul>\n"}]}
{"QuestionId":33810016,"AnswerCount":3,"Tags":"<windows><gpu><tensorflow>","CreationDate":"2015-11-19T17:01:21.437","AcceptedAnswerId":40848593.0,"OwnerUserId":1243762.0,"Title":"TensorFlow dependencies needed. How to run TensorFlow on Windows","Body":"\n\n<p>I am interested in getting TensorFlow to run on Windows, however at present I realize that this is not possible due to some of the dependencies not being usable with Windows, e.g. Bazel.<\/p>\n\n<p>The need arises because as I currently understand it the only way to access the GPU from TensorFlow is via a non-virtual install of Linux. I do realize I can dual boot into a Linux install, but would prefer to avoid that route.<\/p>\n\n<p>To resolve the problem I am in need of the entire dependency chain to build TensorFlow as was wondering if this already existed. <\/p>\n\n<p>I also realize that I can capture the build output when building from source as a solid start, but would like to avoid that work if it is already known.<\/p>\n\n<p>There is a beta of Bazel that runs on Windows - <a href=\"https:\/\/github.com\/dslomov\/bazel-windows\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dslomov\/bazel-windows<\/a><\/p>\n\n<p>See related GitHub Issue to run TensorFlow on Windows. - <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17<\/a><\/p>\n\n<p>Another reason to run on Windows is the possibility to port to Xbox One.<\/p>\n\n<p>I found a possible answer, still need to check it. This will generate a dependency graph as a dot file.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42473142","CreationDate":"2017-02-26T19:28:55.420","ParentId":null,"OwnerUserId":"448357","Title":null,"Body":"<p>Since the begin of 2017, Tensorflow is now <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">officially supported on Windows and can be installed via pip<\/a>:<\/p>\n\n<pre><code>pip install --upgrade tensorflow\npip install --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p>or by fetching packages directly (pick the one that matches your needs, e.g. x64\/gpu)<\/p>\n\n<pre><code># x86 \/ CPU\npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_x86_64.whl\n\n# x64 \/ CPU\npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\n# x64 \/ GPU\npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n"},{"AnswerId":"40848593","CreationDate":"2016-11-28T16:03:01.723","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are now three main options for building and\/or running TensorFlow on Windows:<\/p>\n\n<ul>\n<li>You can install a GPU-enabled PIP package of TensorFlow 0.12rc0 from PyPI: <code>pip install tensorflow-gpu<\/code><\/li>\n<li>You can build the GPU-enabled PIP package yourself using the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/d6b25985ac219a6e58d186a2beb74d5e8d9e4533\/tensorflow\/contrib\/cmake\" rel=\"nofollow noreferrer\">experimental CMake build<\/a>. This also gives you the ability to work with TensorFlow in Visual Studio. The documentation for this build can be found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d6b25985ac219a6e58d186a2beb74d5e8d9e4533\/tensorflow\/contrib\/cmake\/README.md\" rel=\"nofollow noreferrer\">here<\/a>.<\/li>\n<li>There is preliminary support for building TensorFlow using <a href=\"https:\/\/www.bazel.io\/versions\/master\/docs\/windows.html\" rel=\"nofollow noreferrer\">Bazel for Windows<\/a>. However, we are still ironing out some bugs with this build.<\/li>\n<\/ul>\n"},{"AnswerId":"33834062","CreationDate":"2015-11-20T18:55:04.830","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>This may not be exactly what you want one way to run TensorFlow under Windows is to install a virtual machine (VMWare player v12 is free to use for non-commercial) and then install Ubuntu in that and finally TensorFlow in Ubuntu. Works well for me.<\/p>\n"}]}
{"QuestionId":33810990,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-19T17:52:29.890","AcceptedAnswerId":33812296.0,"OwnerUserId":826983.0,"Title":"How to feed a placeholder?","Body":"<p>I am trying to implement a simple feed forward network. However, I can't figure out how to feed a . This example:<\/p>\n\n<pre><\/pre>\n\n<p>Gives me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried<\/p>\n\n<pre><\/pre>\n\n<p>but that does not work apparently.<\/p>\n","answers":[{"AnswerId":"33812296","CreationDate":"2015-11-19T19:03:15.523","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To feed a placeholder, you use the <code>feed_dict<\/code> argument to <code>Session.run()<\/code> (or <code>Tensor.eval()<\/code>). Let's say you have the following graph, with a placeholder:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=[2, 2])\ny = tf.constant([[1.0, 1.0], [0.0, 1.0]])\nz = tf.matmul(x, y)\n<\/code><\/pre>\n\n<p>If you want to evaluate <code>z<\/code>, you must feed a value for <code>x<\/code>. You can do this as follows:<\/p>\n\n<pre><code>sess = tf.Session()\nprint sess.run(z, feed_dict={x: [[3.0, 4.0], [5.0, 6.0]]})\n<\/code><\/pre>\n\n<p>For more information, see the <a href=\"http:\/\/tensorflow.org\/how_tos\/reading_data\/index.html#Feeding\" rel=\"noreferrer\">documentation on feeding<\/a>.<\/p>\n"}]}
{"QuestionId":33811892,"AnswerCount":1,"Tags":"<image><theano><predict>","CreationDate":"2015-11-19T18:40:52.960","AcceptedAnswerId":null,"OwnerUserId":5531398.0,"Title":"Feeding images to Theano throws ShapeMismatch error","Body":"<p>I'm trying a really basic example of Theano Logistic Regression model, and after training the net I wanted to test some images to see how they're classified. The code of the training and testing can be found at <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py<\/a>. In fact the only part I tried to modify is the predict() function as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I followed some hints found here <a href=\"http:\/\/blog.eduardovalle.com\/2015\/08\/25\/input-images-theano\/\" rel=\"nofollow\">http:\/\/blog.eduardovalle.com\/2015\/08\/25\/input-images-theano\/<\/a> but apparently I'm having problems since what I get is:<\/p>\n\n<blockquote>\n  <p>ValueError: Shape mismatch: x has 28 cols (and 28 rows) but y has 784\n  rows (and 10 cols) Apply node that caused the error: Dot22(x, W)\n  Inputs types: [TensorType(float64, matrix), TensorType(float64,\n  matrix)] Inputs shapes: [(28, 28), (784, 10)] Inputs strides: [(224,\n  8), (80, 8)] Inputs values: ['not shown', 'not shown']<\/p>\n<\/blockquote>\n\n<p>What is then the correct way to feed images (in my case 28x28) to Theano for prediction?<\/p>\n","answers":[{"AnswerId":"33821282","CreationDate":"2015-11-20T07:37:01.693","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The model expects <em>batches<\/em> of <em>flattened<\/em> images as input.<\/p>\n\n<p>The input needs to be a matrix with a row per image. There are 784 columns which is a flattened (or reshaped) version of a 28x28 pixel image (note that 28*28=784).<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>img = Image.open('index.png','r')\nshared_x = theano.shared(numpy.asarray(img,dtype=theano.config.floatX))\npredicted_values = predict_model(shared_x.get_value())\n<\/code><\/pre>\n\n<p>Use<\/p>\n\n<pre><code>img = Image.open('index.png','r')\nimg = numpy.asarray(img, dtype=theano.config.floatX)\nimg = img.flatten()\nx_batch = numpy.expand_dims(img, 0)\npredicted_values = predict_model(x_batch)\n<\/code><\/pre>\n\n<p>There's no need to use a shared variable since you're just passing the numpy array it contains to the Theano function.<\/p>\n"}]}
{"QuestionId":33821449,"AnswerCount":1,"Tags":"<docker><google-cloud-platform><tensorflow>","CreationDate":"2015-11-20T07:46:34.430","AcceptedAnswerId":33833246.0,"OwnerUserId":594353.0,"Title":"Starting TensorFlow on Docker on Google Cloud","Body":"<p>I followed the directions to install TensorFlow on Docker on Google Cloud here :<\/p>\n\n<p><a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.html#docker-based-installation\" rel=\"nofollow\">http:\/\/tensorflow.org\/get_started\/os_setup.html#docker-based-installation<\/a><\/p>\n\n<p>The first time, it did work and showed the tensorflow prompt.\nNow that I have logged out and back in, I get this:<\/p>\n\n<p><\/p>\n\n<p>I also tried this:<\/p>\n\n<p><\/p>\n\n<p>Is there a different way to start TensorFlow on Docker after it is installed?<\/p>\n","answers":[{"AnswerId":"33833246","CreationDate":"2015-11-20T18:04:53.743","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>docker run -it<\/code> command brings up a bash shell in a container where TensorFlow is installed. Once you are at the <code>root@2e87064f0743:\/#<\/code> prompt you can start an interactive TensorFlow session by starting <code>ipython<\/code> as the following example shows:<\/p>\n\n<pre><code>$ docker run -it b.gcr.io\/tensorflow\/tensorflow\n\nroot@2e87064f0743:\/# ipython\n\nPython 2.7.6 ...\nIn [1]: import tensorflow as tf\nIn [2]: c = tf.constant(5.0)\nIn [3]: sess = tf.InteractiveSession()\nI tensorflow\/core\/...\n\nIn [4]: c.eval()\nOut[4]: 5.0\n<\/code><\/pre>\n"}]}
{"QuestionId":33828464,"AnswerCount":1,"Tags":"<neural-network><deep-learning><conv-neural-network><caffe><vgg-net>","CreationDate":"2015-11-20T13:55:29.850","AcceptedAnswerId":null,"OwnerUserId":375986.0,"Title":"Fine-tune VGG or AlexNet for non-square inputs","Body":"<p>VGG and AlexNet, amongst others, require a fixed image input of square dimensions ( == ). How can one fine-tune or otherwise perform net surgery such that non-square inputs can be provided? <\/p>\n\n<p>For your reference, I'm using Caffe and intend to extract  features for non-square image inputs.<\/p>\n","answers":[{"AnswerId":"33852528","CreationDate":"2015-11-22T07:15:24.847","ParentId":null,"OwnerUserId":"2469775","Title":null,"Body":"<p>For the convolutional part of the net - the input size does not really matter: the shape of the output will change as you change the input size.<br>\nHowever, when it comes to <code>\"InnerProduct\"<\/code> layers - the shape of the weights is fixed and it is determined by input size.  <\/p>\n\n<p>You can perform <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb#Casting-a-Classifier-into-a-Fully-Convolutional-Network\" rel=\"nofollow\">\"net surgery\"<\/a> converting your <code>\"InnerProduct\"<\/code> layers into <code>\"Convolution\"<\/code> layers: This way your net can process inputs at any size they come. However, your outputs will also vary in shape. <\/p>\n\n<p>Another option is to define your net according to a new <strong>fixed<\/strong> input size, re-use all the learned weights of the covolutions and only fine-tune the weights of the fully connected layers.<\/p>\n"}]}
{"QuestionId":33828582,"AnswerCount":3,"Tags":"<python><matlab><deep-learning><caffe><vgg-net>","CreationDate":"2015-11-20T14:01:04.293","AcceptedAnswerId":33858058.0,"OwnerUserId":3962073.0,"Title":"VGG Face Descriptor in python with caffe","Body":"<p>I want implement <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/vgg_face\/\" rel=\"noreferrer\">VGG Face Descriptor<\/a> in python. But I keep getting an error:<\/p>\n\n<blockquote>\n  <p>TypeError: can only concatenate list (not \"numpy.ndarray\") to list<\/p>\n<\/blockquote>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n\n<p>Can you help me ?<\/p>\n\n<p><strong>UPDATE 1<\/strong><\/p>\n\n<p>This working code is example in matlab<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33858058","CreationDate":"2015-11-22T17:50:12.723","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>To use python interface you need to transform the input image before feeding it to the net<\/p>\n\n<pre><code>img = caffe.io.load_image( \"ak.png\" )\nimg = img[:,:,::-1]*255.0 # convert RGB-&gt;BGR\navg = np.array([93.5940, 104.7624, 129.1863])  # BGR mean values\nimg = img - avg # subtract mean (numpy takes care of dimensions :)\n<\/code><\/pre>\n\n<p>Now <code>img<\/code> is <code>H<\/code>-by-<code>W<\/code>-by-3 numpy array.<br>\nCaffe expects its inputs as 4D: batch_index <code>x<\/code> channel <code>x<\/code> width <code>x<\/code> height.<br>\nTherefore you need to <code>transpose<\/code> the input and add a singleton dimension to represent the \"batch_index\" leading dimension<\/p>\n\n<pre><code>img = img.transpose((2,0,1)) \nimg = img[None,:] # add singleton dimension\n<\/code><\/pre>\n\n<p>Now you can run the forward pass<\/p>\n\n<pre><code>out = net.forward_all( data = img )\n<\/code><\/pre>\n"},{"AnswerId":"34038441","CreationDate":"2015-12-02T08:54:51.530","ParentId":null,"OwnerUserId":"2584518","Title":null,"Body":"<p>OpenCV reads in BGR and scaled to 255 format by default, so:<\/p>\n\n<pre><code>img = cv2.imread('ak.png')\navg = np.array([93.5940,104.7624,129.1863]) # BGR mean from VGG\nimg -= avg # subtract mean\nimg = img.transpose((2,0,1)) # to match image input dimension: 3x224x224\nimg = img[None,:] # add singleton dimension to match batch dimension\nout = net.forward_all(data = img)\n<\/code><\/pre>\n"},{"AnswerId":"33834735","CreationDate":"2015-11-20T19:36:40.947","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Try passing a single element list to the method.<\/p>\n\n<pre><code>net.forward ([img])\n<\/code><\/pre>\n"}]}
{"QuestionId":33828923,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2015-11-20T14:17:46.407","AcceptedAnswerId":33834800.0,"OwnerUserId":3623648.0,"Title":"Seq2Seq for prediction of complex states","Body":"<p><strong>My problem<\/strong>:\nI have a sequence of complex states and I want to predict the future states.<\/p>\n\n<p><strong>Input<\/strong>:\nI have a sequence of states. Each sequence can be of variable length. Each state is a moment in time and is described by several attributes: [att1, att2, ...]. Where each attribute is a number between an interval [[0..5], [1..3651], ...]<\/p>\n\n<p>The example (and paper) of Seq2Seq is based on that each state (word) is taken from their dictionary. So each state has around 80.000 possibilities. But how would you represent each state when it is taken from a set of vectors and the set is just each possible combination of the attributes.<\/p>\n\n<p>Is there any method to work with more complex states with TensorFlow? Also, what is a good method do decide the boundaries of your buckets when the relation between input length and output length is unclear?<\/p>\n","answers":[{"AnswerId":"33834800","CreationDate":"2015-11-20T19:40:44.527","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>May I suggest a rephrasing and splitting of your question into two parts?  The first is really a general machine learning\/LSTM question that's independent of tensorflow:  How to use an LSTM to predict when the sequence elements are general vectors, and the second is how to represent this in tensorflow.  For the former - there's nothing really magical to do there.  <\/p>\n\n<p>But a very quick answer:  You've really just skipped the embedding lookup part of seq2seq.  You can feed dense tensors in to a suitably modified version of it -- your state is just a dense vector representation of the state.  That's the same thing that comes out of an embedding lookup.<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/word2vec\/index.md\" rel=\"nofollow\">vector representation tutorial<\/a> discusses the preprocessing that turns, e.g., words into embeddings for use in later parts of the learning pipeline.<\/p>\n\n<p>If you look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/seq2seq.py#L139\" rel=\"nofollow\">line 139 of seq2seq.py<\/a> you'll see that the embedding_rnn_decoder takes in a 1D batch of things to decide (the dimension is elements in the batch), but then uses the embedding lookup to turn it into a batch_size * cell.input_size tensor.  You want to directly input a batch_size * cell.input_size tensor into the RNN, skipping the embedding step.<\/p>\n"}]}
{"QuestionId":33828959,"AnswerCount":1,"Tags":"<neural-network><theano><deep-learning><caffe><torch>","CreationDate":"2015-11-20T14:19:18.563","AcceptedAnswerId":33847461.0,"OwnerUserId":1010393.0,"Title":"Is there a common format for neural networks","Body":"<p>Different teams use different libraries to train and run neural networks (caffe, torch, theano...). This makes sharing difficult: each library has its own format to store networks and you have to install a new library each time you want to test other teams' work.<\/p>\n\n<p>I am looking for solutions to make this less tedious:\n- Is there a preferred (shared?) format to store neural networks?\n- Is there a service or library that can help handle different types of networks \/ or transform one type into another?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"33847461","CreationDate":"2015-11-21T19:09:43.483","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>Is there a preferred (shared?) format to store neural networks?<\/p>\n<\/blockquote>\n\n<p>Each library \/ framework has its own serialization, e.g. Caffe uses <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html#model-format\" rel=\"noreferrer\">Protocol Buffers<\/a>, Torch has a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/840e7312e51e3dd6e28583d4020bfe134eadb9cf\/doc\/serialization.md\" rel=\"noreferrer\">built-in serialization scheme<\/a> and Theano objects can be <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loading_and_saving.html\" rel=\"noreferrer\">serialized with pickle<\/a>.<\/p>\n\n<p>In some cases like <a href=\"https:\/\/github.com\/sermanet\/OverFeat\" rel=\"noreferrer\">OverFeat<\/a> or <a href=\"https:\/\/github.com\/pjreddie\/darknet\" rel=\"noreferrer\">darknet<\/a> the weights and biases are stored on-disk in binary format via plain <code>fwrite<\/code>-s of the corresponding <code>float<\/code>(or <code>double<\/code>) contiguous arrays (see <a href=\"https:\/\/stackoverflow.com\/a\/30058467\/1688185\">this answer<\/a> for more details). Note that this does not cover the architecture of the network \/ model which has to be known or represented separately (like <a href=\"https:\/\/github.com\/jhjin\/overfeat-torch\/blob\/17e308dcd093b16fa39dedc452cac72ce85a6e8b\/run.lua#L37-L56\" rel=\"noreferrer\">declared explicitly at load time<\/a>).<\/p>\n\n<p>Also: a library like <a href=\"http:\/\/libccv.org\/\" rel=\"noreferrer\">libccv<\/a> stores the structure and the weights in a <a href=\"https:\/\/github.com\/deltheil\/loadccv\/blob\/1c965183163b95a100113491c8b305b55df267cb\/init.lua#L103-L128\" rel=\"noreferrer\">SQLite database<\/a>.<\/p>\n\n<blockquote>\n  <p>Is there a service or library that can help handle different types of networks \/ or transform one type into another?<\/p>\n<\/blockquote>\n\n<p>I don't think there is a single (meta) library that claims to do so. But it exists distinct projects that provide convenient converters.<\/p>\n\n<p>Some examples (non exhaustive):<\/p>\n\n<ul>\n<li>Caffe -> Torch: <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\" rel=\"noreferrer\">https:\/\/github.com\/szagoruyko\/loadcaffe<\/a><\/li>\n<li>Torch -> Caffe: <a href=\"https:\/\/github.com\/facebook\/fb-caffe-exts\" rel=\"noreferrer\">https:\/\/github.com\/facebook\/fb-caffe-exts<\/a><\/li>\n<li>Caffe -> TensorFlow: <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"noreferrer\">https:\/\/github.com\/ethereon\/caffe-tensorflow<\/a><\/li>\n<\/ul>\n\n<p>--<\/p>\n\n<p><strong>UPDATE<\/strong> (2017-09): two noticeable initiatives are:<\/p>\n\n<p>(1) the <a href=\"https:\/\/github.com\/onnx\" rel=\"noreferrer\">ONNX<\/a> format (a.k.a. Open Neural Network Exchange):<\/p>\n\n<blockquote>\n  <p>[...] a standard for representing deep learning models that enables models to be transferred between frameworks<\/p>\n<\/blockquote>\n\n<p><em>See these <a href=\"https:\/\/research.fb.com\/facebook-and-microsoft-introduce-new-open-ecosystem-for-interchangeable-ai-frameworks\/\" rel=\"noreferrer\">blog<\/a> <a href=\"https:\/\/www.microsoft.com\/en-us\/cognitive-toolkit\/blog\/2017\/09\/microsoft-facebook-create-open-ecosystem-ai-model-interoperability\/\" rel=\"noreferrer\">posts<\/a>.<\/em><\/p>\n\n<p>(2) the <a href=\"https:\/\/apple.github.io\/coremltools\/coremlspecification\/\" rel=\"noreferrer\">CoreML<\/a> format introduced by Apple:<\/p>\n\n<blockquote>\n  <p>[...] a public file format (<code>.mlmodel<\/code>) for a broad set of ML methods [...] Models in this format can be directly integrated into apps through Xcode.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":33833100,"AnswerCount":1,"Tags":"<cross-validation><caffe>","CreationDate":"2015-11-20T17:55:55.943","AcceptedAnswerId":null,"OwnerUserId":3929051.0,"Title":"4 fold cross validation | Caffe","Body":"<p>So I trying to perform a 4-fold cross validation on my training set. I have divided my training data into four quarters. I use three quarters for training and one quarter for validation. I repeat this three more times till all the quarters are given a chance to be the validation set, atleast once. <\/p>\n\n<p>Now after training I have four caffemodels. I test the models on my validation sets. I am getting different accuracy in each case. How should I proceed from here? Should I just choose the model with the highest accuracy? <\/p>\n","answers":[{"AnswerId":"38915630","CreationDate":"2016-08-12T10:22:32.140","ParentId":null,"OwnerUserId":"1195402","Title":null,"Body":"<p>Maybe it is a late reply, but in any case...\nThe short answer is that, if the performances of the four models are similar and good enough, then you re-train the model on all the data available, because you don't want to waste any of them. <\/p>\n\n<p>The n-fold cross validation is a practical technique to get some insights on the learning and generalization properties of the model you are trying to train, when you don't have a lot of data to start with. You can find details everywhere on the web, but I suggest the open-source book <a href=\"http:\/\/www-bcf.usc.edu\/~gareth\/ISL\/\" rel=\"nofollow noreferrer\">Introduction to Statistical Learning<\/a>, Chapter 5.<\/p>\n\n<p>The general rule says that after you trained your n models, you average the prediction error (MSE, accuracy, or whatever) to get a general idea of the performance of that particular model (in your case maybe the network architecture and learning strategy) on that dataset.\nThe main idea is to assess the models learned on the training splits checking if they have an acceptable performance on the validation set. If they do not, then your models probably overfitted tha training data. If both the errors on training and validation splits are high, then the models should be reconsidered, since they don't have predictive capacity. <\/p>\n\n<p>In any case, I would also consider the <a href=\"https:\/\/www.quora.com\/Is-cross-validation-heavily-used-in-Deep-Learning-or-is-it-too-expensive-to-be-used\/answer\/Yoshua-Bengio?srid=ipgD\" rel=\"nofollow noreferrer\">advice of Yoshua Bengio<\/a> who says that for the kind of problem deep learning is meant for, you usually have enough data to simply go with a training\/test split. In this case <a href=\"https:\/\/stackoverflow.com\/questions\/13610074\/is-there-a-rule-of-thumb-for-how-to-divide-a-dataset-into-training-and-validatio\">this answer on Stackoverflow <\/a>could be useful to you. <\/p>\n"}]}
{"QuestionId":33833818,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-20T18:41:15.570","AcceptedAnswerId":33834561.0,"OwnerUserId":3320135.0,"Title":"Printing the loss during TensorFlow training","Body":"<p>I am looking at the TensorFlow \"<a href=\"https:\/\/www.tensorflow.org\/versions\/r1.1\/get_started\/mnist\/beginners#evaluating_our_model\" rel=\"noreferrer\">MNIST For ML Beginners<\/a>\" tutorial, and I want to print out the training loss after every training step.<\/p>\n\n<p>My training loop currently looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now,  is defined as:<\/p>\n\n<pre><\/pre>\n\n<p>Where  is the loss which I want to print out:<\/p>\n\n<pre><\/pre>\n\n<p>One way to print this would be to explicitly compute  in the training loop:<\/p>\n\n<pre><\/pre>\n\n<p>I now have two questions regarding this:<\/p>\n\n<ol>\n<li><p>Given that  is already computed during , it seems inefficient to compute it twice, requiring twice the number of forward passes of all the training data. Is there a way to access the value of  when it was computed during ?<\/p><\/li>\n<li><p>How do I even print a ? Using  gives me an error...<\/p><\/li>\n<\/ol>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"33834539","CreationDate":"2015-11-20T19:25:19.810","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Instead of just running the training_step, run also the cross_entropy node so that its value is returned to you.  Remember that:<\/p>\n\n<pre><code>var_as_a_python_value = sess.run(tensorflow_variable)\n<\/code><\/pre>\n\n<p>will give you what you want, so you can do this:<\/p>\n\n<pre><code>[_, cross_entropy_py] = sess.run([train_step, cross_entropy],\n                                 feed_dict={x: batch_xs, y_: batch_ys})\n<\/code><\/pre>\n\n<p>to both run the training and pull out the value of the cross entropy as it was computed during the iteration.  Note that I turned both the arguments to sess.run and the return values into a list so that both happen.<\/p>\n"},{"AnswerId":"33834561","CreationDate":"2015-11-20T19:26:32.137","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can fetch the value of <code>cross_entropy<\/code> by adding it to the list of arguments to <code>sess.run(...)<\/code>. For example, your <code>for<\/code>-loop could be rewritten as follows:<\/p>\n\n<pre><code>for i in range(100):\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))\n    _, loss_val = sess.run([train_step, cross_entropy],\n                           feed_dict={x: batch_xs, y_: batch_ys})\n    print 'loss = ' + loss_val\n<\/code><\/pre>\n\n<p>The same approach can be used to print the current value of a variable. Let's say, in addition to the value of <code>cross_entropy<\/code>, you wanted to print the value of a <code>tf.Variable<\/code> called <code>W<\/code>, you could do the following:<\/p>\n\n<pre><code>for i in range(100):\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))\n    _, loss_val, W_val = sess.run([train_step, cross_entropy, W],\n                                  feed_dict={x: batch_xs, y_: batch_ys})\n    print 'loss = %s' % loss_val\n    print 'W = %s' % W_val\n<\/code><\/pre>\n"}]}
{"QuestionId":33835552,"AnswerCount":2,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-11-20T20:30:09.487","AcceptedAnswerId":33858039.0,"OwnerUserId":4415923.0,"Title":"In torch \/ nn, how to use nn.View(-1):setNumInputDims(2) with a minibatch 4D tensor?","Body":"<p>I am writing a neural network with torch nn.<\/p>\n\n<p>As part of it I have to transform a 3D tensor with dimensions  into a 2D tensor with dimensions .<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to modify my network so that it can handle minibatches as input. Therefore, I want to transform a 4D tensor with dimensions  into a 3D tensor with dimensions  where  is the number of elements in my minibatch.  is known in advance, but the other dimensions aren't.<\/p>\n\n<p>When I feed a 4D tensor into the network above, I get a 2D tensor  out of it. How do I have to modify the network so that it creates a 3D tensor instead as desired?<\/p>\n\n<p>I have tried different combinations like ,  and  but none of them created a tensor of the format I wanted.<\/p>\n","answers":[{"AnswerId":"33858039","CreationDate":"2015-11-22T17:49:01.420","ParentId":null,"OwnerUserId":"4415923","Title":null,"Body":"<p><code>nn<\/code> seems to be very limited in this case. The best I could find to solve my problem is to create a parallel-like network via. <code>nn.Concat<\/code> and <code>nn.Select<\/code>:<\/p>\n\n<pre><code>local net = nn.Concat(1)\nfor i=1, d do\n  local subNet = nn.Sequential()\n  subNet:add(nn.Select(1, i))\n  subNet:add(nn.View(-1):setNumInputDims(2))\n  subNet:add(nn.Replicate(1, 1))\n  net:add(subNet)\nend\n<\/code><\/pre>\n\n<p>Note: You really need <code>nn.Replicate<\/code> in this case, otherwise it will create a <code>d*a x b*c<\/code> tensor as well.<\/p>\n\n<p>P.S. If someone can provide a better solution (preferably without splitting), I will give him or her the accepted answer marking.<\/p>\n"},{"AnswerId":"37039393","CreationDate":"2016-05-04T22:50:03.490","ParentId":null,"OwnerUserId":"6293129","Title":null,"Body":"<p>My idea is to transpose 2nd dim into 1st dim, and reshape has an optional boolean argument to take care of the unknown dimension. Hacky but seems working.<br>\n    nn.Transpose({1, 2})<br>\n    nn.Reshape(d, -1, true)<br>\n    nn.Transpose({1, 2})  <\/p>\n"}]}
{"QuestionId":33836728,"AnswerCount":3,"Tags":"<docker><google-cloud-platform><tensorflow><tensorboard>","CreationDate":"2015-11-20T21:56:57.150","AcceptedAnswerId":33838211.0,"OwnerUserId":594353.0,"Title":"View Tensorboard on Docker on Google Cloud","Body":"<p>I am trying to display TensorBoard from TensorFlow on Docker on Google Cloud.<\/p>\n\n<p><a href=\"http:\/\/tensorflow.org\/how_tos\/summaries_and_tensorboard\/index.md\" rel=\"noreferrer\">http:\/\/tensorflow.org\/how_tos\/summaries_and_tensorboard\/index.md<\/a><\/p>\n\n<p><\/p>\n\n<p>I have Apache running on Google Cloud (it may be in my first container \"ai-unicorn\" Docker made its own container \"docker-playground\").\nI can see the default page from Google Cloud at <a href=\"http:\/\/104.197.119.57\/\" rel=\"noreferrer\">http:\/\/104.197.119.57\/<\/a> .<\/p>\n\n<p>I start TensorBoard on Google Cloud like this:<\/p>\n\n<p><\/p>\n\n<p>I tried the Google Cloud SSH option called \"Open in browser window on custom port\" using port 6006.<\/p>\n\n<p>It displays:\n\"We are unable to connect to the VM on port 6006.\"<\/p>\n\n<p>What is the correct way to view TensorBoard from Google Cloud?<\/p>\n","answers":[{"AnswerId":"42049234","CreationDate":"2017-02-05T06:35:46.837","ParentId":null,"OwnerUserId":"1165814","Title":null,"Body":"<p>You do not have to use Docker to display the TensorBoard. But if you do want to use Docker, just run the TensorBoard inside of your Docker image.<\/p>\n\n<p>The trick is to allow external access to the default TensorBoard tcp port 6006.<\/p>\n\n<p>I tried the following working solution to display TensorBoard in my Google Cloud VM.<\/p>\n\n<ol>\n<li><p>ensure you pass the gcloud authentication:<\/p>\n\n<p>gcloud auth login<\/p><\/li>\n<li><p>allow public access to the tcp port 6006<\/p>\n\n<p>gcloud compute firewall-rules create tensorboard-port --allow tcp:6006<\/p><\/li>\n<li><p>Run TensorBoard on your VM<\/p>\n\n<p>tensorboard --logdir=workspace\/train\/<\/p><\/li>\n<li><p>Use external IP address to access the TensorBoard outside your VM:<\/p>\n\n<p>Open address <a href=\"http:\/\/your_vm_external\" rel=\"noreferrer\">http:\/\/your_vm_external<\/a> IP:6006\/,<\/p>\n\n<p>e.g. <a href=\"http:\/\/104.196.140.145:6006\/\" rel=\"noreferrer\">http:\/\/104.196.140.145:6006\/<\/a>, where 104.196.140.145 is the external IP address of my VM.<\/p><\/li>\n<\/ol>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Aaxjb.jpg\" alt=\"\"><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/vAbeg.jpg\" alt=\"Enjoy TensorBoard\"><\/p>\n"},{"AnswerId":"48575637","CreationDate":"2018-02-02T04:17:01.740","ParentId":null,"OwnerUserId":"264410","Title":null,"Body":"<p>another option is to use <code>ngrok<\/code> to tunnel. see: <a href=\"https:\/\/stackoverflow.com\/questions\/47818822\/can-i-use-tensorboard-with-google-colab\">Can I use Tensorboard with Google Colab?<\/a><\/p>\n\n<pre><code>$ from jupyter notebook\nps = !ps -ax\nis_tensorboard_running = len([f for f in ps if \"tensorboard\" in f ]) &gt; 0\n\nis_ngrok_running = len([f for f in ps if \"ngrok\" in f ]) &gt; 0\nprint(\"tensorbord={}, ngrok={}\".format(is_tensorboard_running, is_ngrok_running))\n\nif not is_ngrok_running:  \n#    grok should be installed in \/content\/ngrok\n  get_ipython().system_raw('\/content\/ngrok http 6006 &amp;')\n  is_ngrok_running = True\n\n# get public url for tensorboard\ntensorboard_url = !curl -s http:\/\/localhost:4040\/api\/tunnels | python3 -c \\\n    \"import sys, json; print(json.load(sys.stdin)['tunnels'][0]['public_url'])\"    \nprint(\"tensorboard url=\", tensorboard_url)\n<\/code><\/pre>\n"},{"AnswerId":"33838211","CreationDate":"2015-11-21T00:22:09.663","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>By <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tensorboard\/tensorboard.py#L60\">default<\/a>, TensorBoard serves requests on <code>127.0.0.1<\/code>, which is only accessible to processes running on the same machine. If you start TensorBoard with <code>--host 0.0.0.0<\/code>, it will also serve requests on the remote interfaces, so you should be able to connect to it remotely:<\/p>\n\n<pre><code>$ tensorboard --logdir .\/ --host 0.0.0.0\n<\/code><\/pre>\n\n<p>Note that the \"Open in browser window on custom port\" will not connect you to the TensorBoard server - this option is used to connect to an <em>SSH<\/em> server on a non-standard port. The <a href=\"https:\/\/cloud.google.com\/compute\/docs\/networking#externalhttpconnection\">Google Cloud Platform docs<\/a> have information on how to expose ports from your VM. You will need to allow connections on TCP port 6006 for remote access to your VM. You may also need to expose port 6006 from your Docker container, by following the <a href=\"https:\/\/github.com\/wsargent\/docker-cheat-sheet#exposing-ports\">instructions here<\/a>.<\/p>\n\n<p><strong>EDIT:<\/strong> Added some step-by-step instructions to help with your Docker configuration. There are several issues here, and it's not possible to tell which one is failing.<\/p>\n\n<ol>\n<li><p>Configure port forwarding when you start your Docker container:<\/p>\n\n<pre><code>(vm)$ docker run -p 0.0.0.0:7007:6006 -it b.gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p>This forwards connections from port 7007 on your VM to 6006 in your Docker container. (Other values are possible.)<\/p><\/li>\n<li><p>Ensure that you can connect to TensorBoard from within the Docker container:<\/p>\n\n<pre><code>(container)$ tensorboard --logdir .\/ --host 0.0.0.0 --port 6006 &amp;\n(container)$ curl http:\/\/localhost:6006\/\n<\/code><\/pre>\n\n<p>The second command should print some HTML to the console.<\/p><\/li>\n<li><p>In a shell on the VM, ensure that you can connect to the TensorBoard instance running in the container:<\/p>\n\n<pre><code>(vm)$ curl http:\/\/localhost:7007\/\n<\/code><\/pre>\n\n<p>The command should print the same HTML to the console.<\/p><\/li>\n<li><p>Configure the Google Cloud firewall to allow your local client to connect to port 7007 on your VM.<\/p>\n\n<pre><code>(client)$ gcloud compute firewall-rules create tensorboard --allow tcp:7007\n<\/code><\/pre>\n\n<p>You should now be able to connect to TensorBoard in a web browser on your client.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":33837423,"AnswerCount":3,"Tags":"<python><linux><virtualenv><virtualenvwrapper><tensorflow>","CreationDate":"2015-11-20T22:56:06.630","AcceptedAnswerId":null,"OwnerUserId":4415759.0,"Title":"not able to install tensorflow on server","Body":"<p>I am trying to use tensorflow on my college server, as some of the things that I have written are too heavy for my laptop.\nSo I don't have sudo privileges. Below is what I unsuccessfully tried.<\/p>\n\n<p>I am able to install tensorflow by \n. But when I import tensorflow I get the error . <\/p>\n\n<p>I found <a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\">this<\/a> link which solved the exact same problem but when I run <\/p>\n\n<p> <\/p>\n\n<p>I get this error: <\/p>\n\n<p><\/p>\n\n<p>For resolving this I tried:<\/p>\n\n<p><\/p>\n\n<p>And then I get following error:<\/p>\n\n<p><\/p>\n\n<p>And I don't have sudo access so I can't use sudo with above command. <\/p>\n\n<p>For installing virtualenv I used:<\/p>\n\n<p><\/p>\n\n<p>I think that the problem is with virtualenv setup but I tried removing it and installing virtualenv by: <\/p>\n\n<p><\/p>\n\n<p>But this fails and I get:<\/p>\n\n<p><\/p>\n\n<p>I have looked at many (<a href=\"https:\/\/stackoverflow.com\/questions\/29134512\/insecureplatformwarning-a-true-sslcontext-object-is-not-available-this-prevent\">1<\/a>,<a href=\"https:\/\/stackoverflow.com\/questions\/19331782\/cant-install-virtualenv-on-ubuntu-12-04\">2<\/a>) SO answers dealing with this problem but none work for me. I get stuck at some other error.<\/p>\n\n<p>I am now stuck in a loop of errors without sudo privilege. <\/p>\n\n<p>So, my question is that do I necessarily have to install virtualenv or can my problem of installing a working tensorflow be solved much simply. <\/p>\n\n<p>It is a linux server and default python version is 2.6. So, I had to install 2.7 separately for my use.<\/p>\n","answers":[{"AnswerId":"33838096","CreationDate":"2015-11-21T00:06:57.053","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Try <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md#installing_from_sources\" rel=\"nofollow\">building from source<\/a> instead of using the precompiled binary version.  That way you won't have to worry about the glibc incompatibility with the installed system one.<\/p>\n"},{"AnswerId":"33841129","CreationDate":"2015-11-21T08:04:06.383","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p>@Pukki you may be able to have a professor request the system administrator to install it. Then everyone on the system could use it.<\/p>\n"},{"AnswerId":"38258359","CreationDate":"2016-07-08T03:08:32.837","ParentId":null,"OwnerUserId":"196917","Title":null,"Body":"<p>I followed the instructions from <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#virtualenv-installation\" rel=\"nofollow\">this page<\/a> and it resolved my issue. You need the <code>--upgrade URL<\/code>  parameter<\/p>\n"}]}
{"QuestionId":33837903,"AnswerCount":1,"Tags":"<neural-network><ubuntu-14.04><deep-learning><caffe><io-redirection>","CreationDate":"2015-11-20T23:44:39.033","AcceptedAnswerId":33848103.0,"OwnerUserId":4561745.0,"Title":"Terminal output redirection not working for Caffe","Body":"<p>I want to redirect Caffe's output from the terminal to a file (say ). I'm using the command  <\/p>\n\n<pre><\/pre>\n\n<p>However, the  operator doesn't seem to be working and Caffe spits out all the output on to the terminal. I'm using Ubuntu 14.04. <\/p>\n\n<p>Can't seem to figure out why  is not working with Caffe. Any help is much appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"33848103","CreationDate":"2015-11-21T20:12:53.397","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to <a href=\"http:\/\/www.tldp.org\/LDP\/abs\/html\/io-redirection.html\" rel=\"noreferrer\">redierect stderr<\/a> as well<\/p>\n\n<pre><code>caffe train ... &gt; output.txt 2&gt;&amp;1\n<\/code><\/pre>\n\n<p>The redirection operator <code>&gt;<\/code> redirects only stdout, caffe is using sterr as well. You might want to set <code>GLOG_logtosterr=1<\/code> as well.<\/p>\n"}]}
{"QuestionId":33840612,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-11-21T06:54:18.553","AcceptedAnswerId":33840723.0,"OwnerUserId":3688217.0,"Title":"Making a matrix from several vectors","Body":"<p>Is it possible to make a Matrix out of several vectors in theano?<\/p>\n\n<p>like:<\/p>\n\n<pre><\/pre>\n\n<p>similar to the numpy operation:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33840723","CreationDate":"2015-11-21T07:09:17.423","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.stack\" rel=\"nofollow\">theano.tensor.stack<\/a>.<\/p>\n\n<p>Here's a working example:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\nvector1, vector2, vector3 = tt.vectors(3)\nmatrix = tt.stack(vector1, vector2, vector3)\nf = theano.function([vector1, vector2, vector3], matrix)\nprint f([1, 2, 3], [4, 5, 6], [7, 8, 9])\n<\/code><\/pre>\n\n<p>where prints<\/p>\n\n<pre><code>[[ 1.  2.  3.]\n [ 4.  5.  6.]\n [ 7.  8.  9.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":33844777,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-21T14:57:43.900","AcceptedAnswerId":33845177.0,"OwnerUserId":5285204.0,"Title":"How to broadcast across 3d tensor in theano?","Body":"<p>If I have a 3d tensor block  and I would like to set some of its \"faces\" to 0 with probability 0.5. Here axis 1 are rows, axis 2 are columns, and axis 3 are the \"faces\". I have tried<\/p>\n\n<pre><\/pre>\n\n<p>But this isn't working, the shapes aren't lining up and I get an error.<br>\nFor example, I would like <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33845177","CreationDate":"2015-11-21T15:33:52.647","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use <code>dimshuffle<\/code> to add the dimensions necessary to enable broadcasting.<\/p>\n\n<p>Here's a working example:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\nx = tt.tensor3()\ny = tt.bvector()\nz = x * y.dimshuffle(0, 'x', 'x')\nf = theano.function([x, y], z)\nx_value = numpy.array([[[2, 2, 2], [3, 3, 3], [4, 4, 4]],\n                       [[6, 6, 6], [7, 7, 7], [8, 8, 8]]], dtype=theano.config.floatX)\ny_value = numpy.array([1, 0], dtype=numpy.int8)\nprint f(x_value, y_value)\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>[[[ 2.  2.  2.]\n  [ 3.  3.  3.]\n  [ 4.  4.  4.]]\n\n [[ 0.  0.  0.]\n  [ 0.  0.  0.]\n  [ 0.  0.  0.]]]\n<\/code><\/pre>\n"}]}
{"QuestionId":33846069,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-11-21T16:58:43.547","AcceptedAnswerId":37511638.0,"OwnerUserId":3027854.0,"Title":"how to set rmse cost function in tensorflow","Body":"<p>I have cost function in tensorflow.<\/p>\n\n<pre><\/pre>\n\n<p>I am trying out <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/1_linear_regression.py\" rel=\"noreferrer\">this example<\/a>. How can I change it to rmse cost function?<\/p>\n","answers":[{"AnswerId":"43830200","CreationDate":"2017-05-07T10:07:02.167","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The formula for <a href=\"https:\/\/en.wikipedia.org\/wiki\/Root-mean-square_deviation\" rel=\"noreferrer\">root mean square error<\/a> is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/oeNfI.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oeNfI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The way to implement it in TF is <code>tf.sqrt(tf.reduce_mean(tf.squared_difference(Y1, Y2)))<\/code>.<\/p>\n\n<hr>\n\n<p>The important thing to remember is that there is no need to minimize RMSE loss with the optimizer. With the same result you can minimize just <code>tf.reduce_mean(tf.squared_difference(Y1, Y2))<\/code> or even <code>tf.reduce_sum(tf.squared_difference(Y1, Y2))<\/code> but because they have a smaller graph of operations, they will be optimized faster.<\/p>\n\n<p>But you can use this function if you just want to tract the value of RMSE.<\/p>\n"},{"AnswerId":"37511638","CreationDate":"2016-05-29T15:23:36.653","ParentId":null,"OwnerUserId":"2452184","Title":null,"Body":"<pre><code>tf.sqrt(tf.reduce_mean(tf.square(tf.subtract(targets, outputs))))\n<\/code><\/pre>\n\n<p>And slightly simplified (TensorFlow overloads the most important operators):<\/p>\n\n<pre><code>tf.sqrt(tf.reduce_mean((targets - outputs)**2))\n<\/code><\/pre>\n"},{"AnswerId":"49912380","CreationDate":"2018-04-19T03:45:44.137","ParentId":null,"OwnerUserId":"5608649","Title":null,"Body":"<p>Now we have <code>tf.losses.mean_squared_error<\/code><\/p>\n\n<p>Therefore, <\/p>\n\n<pre><code>RMSE = tf.sqrt(tf.losses.mean_squared_error(prediction, label))\n<\/code><\/pre>\n"},{"AnswerId":"33849210","CreationDate":"2015-11-21T22:14:32.450","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>(1)  Are you sure you need this?  Minimizing the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/l2_loss\" rel=\"nofollow noreferrer\">l2 loss<\/a> will give you the same result as minimizing the RMSE error.  (Walk through the math:  You don't need to take the square root, because minimizing x^2 still minimizes x for x>0, and you know that the sum of a bunch of squares is positive. Minimizing x*n minimizes x for constant n).<\/p>\n\n<p>(2)  If you need to know the numerical value of the RMSE error, then  implement it directly from the <a href=\"http:\/\/statweb.stanford.edu\/~susan\/courses\/s60\/split\/node60.html\" rel=\"nofollow noreferrer\">definition of RMSE<\/a>:<\/p>\n\n<pre><code>tf.sqrt(tf.reduce_sum(...)\/n)\n<\/code><\/pre>\n\n<p>(You need to know or calculate n - the number of elements in the sum, and set the reduction axis appropriately in the call to reduce_sum).<\/p>\n"}]}
{"QuestionId":33846101,"AnswerCount":1,"Tags":"<python><neural-network><theano><gradient-descent><recurrent-neural-network>","CreationDate":"2015-11-21T17:01:17.417","AcceptedAnswerId":null,"OwnerUserId":5589501.0,"Title":"theano GRU rnn adam optimizer","Body":"<p>Technical information:<\/p>\n\n<p>OS: Mac OS X 10.9.5<\/p>\n\n<p>IDE: Eclipse Mars.1 Release (4.5.1), with PyDev and Anaconda interpreter (grammar version 3.4)<\/p>\n\n<p>GPU: NVIDIA GeForce GT 650M<\/p>\n\n<p>Libs: numpy, aeosa, Sphinx-1.3.1, Theano 0.7, nltk-3.1<\/p>\n\n<p>My background: I am very new to theano and numpy and haven't taken a formal course in machine learning or discrete math.<\/p>\n\n<p>The recurrent neural network for natural language processing I currently use is taken from here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/dennybritz\/rnn-tutorial-gru-lstm\/blob\/master\/gru_theano.py\" rel=\"nofollow\">https:\/\/github.com\/dennybritz\/rnn-tutorial-gru-lstm\/blob\/master\/gru_theano.py<\/a><\/p>\n\n<p>The only change made to this file is replacing references to  with the string .<\/p>\n\n<p>I also use the utils.py and train.py modules included in the repository, with only minor changes.<\/p>\n\n<p>The adam optimizer I plan to incorporate in place of the sgd\/rms code implemented in the example repository is found here: <a href=\"https:\/\/gist.github.com\/skaae\/ae7225263ca8806868cb\" rel=\"nofollow\">https:\/\/gist.github.com\/skaae\/ae7225263ca8806868cb<\/a><\/p>\n\n<p>Reproduced here (again with references to the  replaced with the hard-coded ):<\/p>\n\n<p>( as ,  as ,  as ,  as )<\/p>\n\n<pre><\/pre>\n\n<p><strong>My question<\/strong> is this:<\/p>\n\n<p>How would you modify the GRUTheano module to use the Adam method above in place of the builtin sgd\/rmsprop function?<\/p>\n\n<p>It looks like the key changes would be to lines 99-126 of GRUTheano:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34644608","CreationDate":"2016-01-06T23:11:08.657","ParentId":null,"OwnerUserId":"557478","Title":null,"Body":"<p>I haven't tested this code, but the only thing you need to change is to tell <strong><em>updates<\/em><\/strong> to use adam(..) instead of the updates already provided here, so something like this should work (complete code looks like this (we need to get rid of rmsprop stuff)):<\/p>\n\n<pre><code>import numpy as np\nimport theano as theano\nimport theano.tensor as T\nfrom theano.gradient import grad_clip\nimport time\nimport operator\n\nclass GRUTheano(object):\n    def __init__(self, word_dim, hidden_dim=128, bptt_truncate=-1):\n        # Assign instance variables\n        self.word_dim = word_dim\n        self.hidden_dim = hidden_dim\n        self.bptt_truncate = bptt_truncate\n        # Initialize the network parameters\n        E = np.random.uniform(-np.sqrt(1.\/word_dim), np.sqrt(1.\/word_dim), (hidden_dim, word_dim))\n        U = np.random.uniform(-np.sqrt(1.\/hidden_dim), np.sqrt(1.\/hidden_dim), (6, hidden_dim, hidden_dim))\n        W = np.random.uniform(-np.sqrt(1.\/hidden_dim), np.sqrt(1.\/hidden_dim), (6, hidden_dim, hidden_dim))\n        V = np.random.uniform(-np.sqrt(1.\/hidden_dim), np.sqrt(1.\/hidden_dim), (word_dim, hidden_dim))\n        b = np.zeros((6, hidden_dim))\n        c = np.zeros(word_dim)\n        # Theano: Created shared variables\n        self.E = theano.shared(name='E', value=E.astype(theano.config.floatX))\n        self.U = theano.shared(name='U', value=U.astype(theano.config.floatX))\n        self.W = theano.shared(name='W', value=W.astype(theano.config.floatX))\n        self.V = theano.shared(name='V', value=V.astype(theano.config.floatX))\n        self.b = theano.shared(name='b', value=b.astype(theano.config.floatX))\n        self.c = theano.shared(name='c', value=c.astype(theano.config.floatX))\n        # We store the Theano graph here\n        self.theano = {}\n        self.__theano_build__()\n\n    def __theano_build__(self):\n        E, V, U, W, b, c = self.E, self.V, self.U, self.W, self.b, self.c\n\n        x = T.ivector('x')\n        y = T.ivector('y')\n\n        def forward_prop_step(x_t, s_t1_prev, s_t2_prev):\n            # This is how we calculated the hidden state in a simple RNN. No longer!\n            # s_t = T.tanh(U[:,x_t] + W.dot(s_t1_prev))\n\n            # Word embedding layer\n            x_e = E[:,x_t]\n\n            # GRU Layer 1\n            z_t1 = T.nnet.hard_sigmoid(U[0].dot(x_e) + W[0].dot(s_t1_prev) + b[0])\n            r_t1 = T.nnet.hard_sigmoid(U[1].dot(x_e) + W[1].dot(s_t1_prev) + b[1])\n            c_t1 = T.tanh(U[2].dot(x_e) + W[2].dot(s_t1_prev * r_t1) + b[2])\n            s_t1 = (T.ones_like(z_t1) - z_t1) * c_t1 + z_t1 * s_t1_prev\n\n            # GRU Layer 2\n            z_t2 = T.nnet.hard_sigmoid(U[3].dot(s_t1) + W[3].dot(s_t2_prev) + b[3])\n            r_t2 = T.nnet.hard_sigmoid(U[4].dot(s_t1) + W[4].dot(s_t2_prev) + b[4])\n            c_t2 = T.tanh(U[5].dot(s_t1) + W[5].dot(s_t2_prev * r_t2) + b[5])\n            s_t2 = (T.ones_like(z_t2) - z_t2) * c_t2 + z_t2 * s_t2_prev\n\n            # Final output calculation\n            # Theano's softmax returns a matrix with one row, we only need the row\n            o_t = T.nnet.softmax(V.dot(s_t2) + c)[0]\n\n            return [o_t, s_t1, s_t2]\n\n    [o, s, s2], updates = theano.scan(\n        forward_prop_step,\n        sequences=x,\n        truncate_gradient=self.bptt_truncate,\n        outputs_info=[None,\n                      dict(initial=T.zeros(self.hidden_dim)),\n                      dict(initial=T.zeros(self.hidden_dim))])\n\n    prediction = T.argmax(o, axis=1)\n    o_error = T.sum(T.nnet.categorical_crossentropy(o, y))\n\n    # Total cost (could add regularization here)\n    cost = o_error\n\n    # Gradients\n    dE = T.grad(cost, E)\n    dU = T.grad(cost, U)\n    dW = T.grad(cost, W)\n    db = T.grad(cost, b)\n    dV = T.grad(cost, V)\n    dc = T.grad(cost, c)\n\n    # Assign functions\n    self.predict = theano.function([x], o)\n    self.predict_class = theano.function([x], prediction)\n    self.ce_error = theano.function([x, y], cost)\n    self.bptt = theano.function([x, y], [dE, dU, dW, db, dV, dc])\n\n    self.params = [self.E, self.U, self.W, self.V, self.b, self.c]\n\n    updates=adam(cost, self.params)\n    self.sgd_step = theano.function(\n        inputs=[x, y],\n        outputs=[],\n        updates=updates\n    )\n\n\ndef calculate_total_loss(self, X, Y):\n    return np.sum([self.ce_error(x,y) for x,y in zip(X,Y)])\n\n\ndef calculate_loss(self, X, Y):\n    # Divide calculate_loss by the number of words\n    num_words = np.sum([len(y) for y in Y])\n    return self.calculate_total_loss(X,Y)\/float(num_words)\n\n\ndef adam(loss, all_params, learning_rate=0.001, b1=0.9, b2=0.999, e=1e-8,\n     gamma=1-1e-8):\n    \"\"\"\n    ADAM update rules\n    Default values are taken from [Kingma2014]\n\n    References:\n    [Kingma2014] Kingma, Diederik, and Jimmy Ba.\n    \"Adam: A Method for Stochastic Optimization.\"\n    arXiv preprint arXiv:1412.6980 (2014).\n    http:\/\/arxiv.org\/pdf\/1412.6980v4.pdf\n\n    \"\"\"\n    updates = []\n    all_grads = theano.grad(loss, all_params)\n    alpha = learning_rate\n    t = theano.shared(np.float32(1))\n    b1_t = b1*gamma**(t-1)   #(Decay the first moment running average coefficient)\n\n    for theta_previous, g in zip(all_params, all_grads):\n        m_previous =   theano.shared(np.zeros(theta_previous.get_value().shape,\n                                        dtype=theano.config.floatX))\n        v_previous = theano.shared(np.zeros(theta_previous.get_value().shape,\n                                        dtype=theano.config.floatX))\n\n        m = b1_t*m_previous + (1 - b1_t)*g                             # (Update biased first moment estimate)\n        v = b2*v_previous + (1 - b2)*g**2                              # (Update biased second raw moment estimate)\n        m_hat = m \/ (1-b1**t)                                          # (Compute bias-corrected first moment estimate)\n        v_hat = v \/ (1-b2**t)                                          # (Compute bias-corrected second raw moment estimate)\n        theta = theta_previous - (alpha * m_hat) \/ (T.sqrt(v_hat) + e) #(Update parameters)\n\n        updates.append((m_previous, m))\n        updates.append((v_previous, v))\n        updates.append((theta_previous, theta) )\n    updates.append((t, t + 1.))\n    return updates\n<\/code><\/pre>\n"}]}
{"QuestionId":33847168,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-21T18:42:35.637","AcceptedAnswerId":33851308.0,"OwnerUserId":488517.0,"Title":"How to retrieve the execution ordering of fetches?","Body":"<p>Given an array of fetches, how do I retrieve the (possibly non-unique) ordering of fetches that will be executed in a single call to ?<\/p>\n","answers":[{"AnswerId":"33851308","CreationDate":"2015-11-22T03:31:38.560","ParentId":null,"OwnerUserId":"488517","Title":null,"Body":"<p>A reasonable solution is to do re-compute the topological sort in python. Looks like the C++ implementation isn't exposed in the python API. Please let me know if there are situations where this wouldn't work.<\/p>\n\n<p>Here's an example:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom toposort import toposort\n\n\nsess = tf.InteractiveSession()\n\nmatrix1=tf.constant([[3., 3.]])\nmatrix2=tf.constant([[2.], [2.]])\n\nsum = tf.add(matrix1, matrix2)\nproduct = tf.matmul(matrix1, matrix2)\nfinal = tf.mul(sum, product)\n\ng = sess.graph\ndeps = {}\n\nfor op in g.get_operations():\n    # op node\n    op_inputs = set()\n    op_inputs.update([t.name for t in op.inputs])\n    deps[op.name] = op_inputs\n\n    # tensor output node\n    for t in op.outputs:\n        deps[t.name]={op.name}\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>deps\n{u'Add': {u'Const:0', u'Const_1:0'},\n u'Add:0': {u'Add'},\n u'Const': set(),\n u'Const:0': {u'Const'},\n u'Const_1': set(),\n u'Const_1:0': {u'Const_1'},\n u'MatMul': {u'Const:0', u'Const_1:0'},\n u'MatMul:0': {u'MatMul'},\n u'Mul': {u'Add:0', u'MatMul:0'},\n u'Mul:0': {u'Mul'}}\n\nlist(toposort(deps))\n[{u'Const', u'Const_1'},\n {u'Const:0', u'Const_1:0'},\n {u'Add', u'MatMul'},\n {u'Add:0', u'MatMul:0'},\n {u'Mul'},\n {u'Mul:0'}]\n<\/code><\/pre>\n\n<p>Subsequently, we can manually step through evaluation of each node in the graph - subsequent calls to <code>Session.run()<\/code> involve passing in a <code>feed_dict<\/code> that accumulates the results of all previous inputs. This is pretty slow because of the constant shuffling between C++ and numpy data, and memory-intensive because we are caching output values of everything.<\/p>\n"}]}
{"QuestionId":33849617,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2015-11-21T22:57:47.263","AcceptedAnswerId":33862568.0,"OwnerUserId":5590385.0,"Title":"How do I convert a directory of jpeg images to TFRecords file in tensorflow?","Body":"<p>I have training data that is a directory of jpeg images and a corresponding text file containing the file name and the associated category label.  I am trying to convert this training data into a tfrecords file as described in the tensorflow documentation. I have spent quite some time trying to get this to work but there are no examples in tensorflow that demonstrate how to use any of the readers to read in jpeg files and add them to a tfrecord using tfrecordwriter<\/p>\n","answers":[{"AnswerId":"51318067","CreationDate":"2018-07-13T05:15:07.363","ParentId":null,"OwnerUserId":"4494","Title":null,"Body":"<p>Note that images will be saved in TFRecord as uncompressed tensors, possibly increasing the size by a factor of about 5. That's wasting storage space, and likely to be rather slow because of the amount of data that needs to be read.<\/p>\n\n<p>It's far better to just save the filename in the TFRecord, and read the file on demand. The new <code>Dataset<\/code> API works well, and <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#decoding_image_data_and_resizing_it\" rel=\"nofollow noreferrer\">the documentation<\/a> has this example:<\/p>\n\n<pre><code># Reads an image from a file, decodes it into a dense tensor, and resizes it\n# to a fixed shape.\ndef _parse_function(filename, label):\n  image_string = tf.read_file(filename)\n  image_decoded = tf.image.decode_jpeg(image_string)\n  image_resized = tf.image.resize_images(image_decoded, [28, 28])\n  return image_resized, label\n\n# A vector of filenames.\nfilenames = tf.constant([\"\/var\/data\/image1.jpg\", \"\/var\/data\/image2.jpg\", ...])\n\n# `labels[i]` is the label for the image in `filenames[i].\nlabels = tf.constant([0, 37, ...])\n\ndataset = tf.data.Dataset.from_tensor_slices((filenames, labels))\ndataset = dataset.map(_parse_function)\n<\/code><\/pre>\n"},{"AnswerId":"33862568","CreationDate":"2015-11-23T01:43:54.813","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>I hope this helps:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>filename_queue = tf.train.string_input_producer(['\/Users\/HANEL\/Desktop\/tf.png']) #  list of files to read\n\nreader = tf.WholeFileReader()\nkey, value = reader.read(filename_queue)\n\nmy_img = tf.image.decode_png(value) # use decode_png or decode_jpeg decoder based on your files.\n\ninit_op = tf.initialize_all_variables()\nwith tf.Session() as sess:\n  sess.run(init_op)\n\n# Start populating the filename queue.\n\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(coord=coord)\n\nfor i in range(1): #length of your filename list\n  image = my_img.eval() #here is your image Tensor :) \n\nprint(image.shape)\nImage.show(Image.fromarray(np.asarray(image)))\n\ncoord.request_stop()\ncoord.join(threads)\n<\/code><\/pre>\n\n<p>For getting all images as an array of tensors use the following code example. <\/p>\n\n<p><a href=\"https:\/\/github.com\/HamedMP\/ImageFlow\" rel=\"noreferrer\">Github repo of ImageFlow<\/a><\/p>\n\n<hr>\n\n<p>Update:<\/p>\n\n<p>In the previous answer I just told how to read an image in TF format, but not saving it in TFRecords. For that you should use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _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# images and labels array as input\ndef convert_to(images, labels, name):\n  num_examples = labels.shape[0]\n  if images.shape[0] != num_examples:\n    raise ValueError(\"Images size %d does not match label size %d.\" %\n                     (images.shape[0], num_examples))\n  rows = images.shape[1]\n  cols = images.shape[2]\n  depth = images.shape[3]\n\n  filename = os.path.join(FLAGS.directory, name + '.tfrecords')\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<\/code><\/pre>\n\n<p>More info <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>And you read the data like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Remember to generate a file name queue of you 'train.TFRecord' file path\ndef read_and_decode(filename_queue):\n  reader = tf.TFRecordReader()\n  _, serialized_example = reader.read(filename_queue)\n  features = tf.parse_single_example(\n    serialized_example,\n    dense_keys=['image_raw', 'label'],\n    # Defaults are not specified since both keys are required.\n    dense_types=[tf.string, tf.int64])\n\n  # Convert from a scalar string tensor (whose single string has\n  image = tf.decode_raw(features['image_raw'], tf.uint8)\n\n  image = tf.reshape(image, [my_cifar.n_input])\n  image.set_shape([my_cifar.n_input])\n\n  # OPTIONAL: Could reshape into a 28x28 image and apply distortions\n  # here.  Since we are not applying any distortions in this\n  # example, and the next step expects the image to be flattened\n  # into a vector, we don't bother.\n\n  # Convert from [0, 255] -&gt; [-0.5, 0.5] floats.\n  image = tf.cast(image, tf.float32)\n  image = tf.cast(image, tf.float32) * (1. \/ 255) - 0.5\n\n  # Convert label from a scalar uint8 tensor to an int32 scalar.\n  label = tf.cast(features['label'], tf.int32)\n\n  return image, label\n<\/code><\/pre>\n"},{"AnswerId":"42201224","CreationDate":"2017-02-13T09:59:09.143","ParentId":null,"OwnerUserId":"1115215","Title":null,"Body":"<p>I have same problem, too.<\/p>\n\n<p>So here is how i get the tfrecords files of my own jpeg files<\/p>\n\n<p><strong>Edit:<\/strong> add sol 1 - a better &amp; faster way<\/p>\n\n<h2><strong>(Recommended) Solution 1:<\/strong><\/h2>\n\n<p>From <em>tensorflow official github: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/inception#how-to-construct-a-new-dataset-for-retraining\" rel=\"nofollow noreferrer\">How to Construct a New Dataset for Retraining<\/a><\/em>, use official python script <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/data\/build_image_data.py\" rel=\"nofollow noreferrer\">build_image_data.py<\/a> directly and <a href=\"https:\/\/bazel.build\/\" rel=\"nofollow noreferrer\">bazel<\/a> is a better idea.<\/p>\n\n<p>Here is the instruction:<\/p>\n\n<blockquote>\n  <p>To run <code>build_image_data.py<\/code>, you can run the following command line:<\/p>\n\n<pre><code># location to where to save the TFRecord data.        \nOUTPUT_DIRECTORY=$HOME\/my-custom-data\/\n\n# build the preprocessing script.\nbazel build inception\/build_image_data\n\n# convert the data.\nbazel-bin\/inception\/build_image_data \\\n  --train_directory=\"${TRAIN_DIR}\" \\\n  --validation_directory=\"${VALIDATION_DIR}\" \\\n  --output_directory=\"${OUTPUT_DIRECTORY}\" \\\n  --labels_file=\"${LABELS_FILE}\" \\\n  --train_shards=128 \\\n  --validation_shards=24 \\\n  --num_threads=8\n<\/code><\/pre>\n  \n  <p>where the <code>$OUTPUT_DIRECTORY<\/code> is the location of the sharded\n  <code>TFRecords<\/code>. The <code>$LABELS_FILE<\/code> will be a text file that is read by\n  the script that provides a list of all of the labels.<\/p>\n<\/blockquote>\n\n<p>then, it should do the trick.<\/p>\n\n<p>ps. bazel, which is made by Google, turn code into makefile.<\/p>\n\n<h2>Solution 2:<\/h2>\n\n<p>First, i reference the instruction by @capitalistpug and check the shell script file <\/p>\n\n<p>(shell script file providing by Google: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data\/download_and_preprocess_flowers.sh\" rel=\"nofollow noreferrer\">download_and_preprocess_flowers.sh<\/a>)<\/p>\n\n<p>Second, i also find out a mini inception-v3 training tutorial by NVIDIA<\/p>\n\n<p>(NVIDIA official <a href=\"http:\/\/www.nvidia.com\/object\/gpu-accelerated-applications-tensorflow-running-jobs.html\" rel=\"nofollow noreferrer\">SPEED UP TRAINING WITH GPU-ACCELERATED TENSORFLOW<\/a>)<\/p>\n\n<p><strong>Be careful<\/strong>, the following steps need to be executed in the Bazel WORKSAPCE enviroment<\/p>\n\n<p>so Bazel build file can run successfully<\/p>\n\n<hr>\n\n<p><em>First step<\/em>, I <strong>comment out<\/strong> the part of downloading the imagenet data set that i already downloaded <\/p>\n\n<p>and the rest of the part that i don't need of download_and_preprocess_flowers.sh <\/p>\n\n<p><em>Second step<\/em>, change directory to tensorflow\/models\/inception<\/p>\n\n<p>where it is the Bazel environment and <strong>it is build by Bazel before<\/strong><\/p>\n\n<pre><code>$ cd tensorflow\/models\/inception \n<\/code><\/pre>\n\n<p>Optional : If it is not builded before, type in the following code in cmd<\/p>\n\n<pre><code>$ bazel build inception\/download_and_preprocess_flowers \n<\/code><\/pre>\n\n<p>You need to figure out the content in the following image<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CzQDs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CzQDs.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><em>And last step<\/em>, type in the following code:<\/p>\n\n<pre><code>$ bazel-bin\/inception\/download_and_preprocess_flowers $Your\/own\/image\/data\/path\n<\/code><\/pre>\n\n<p>Then, it will start calling <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/data\/build_image_data.py\" rel=\"nofollow noreferrer\">build_image_data.py<\/a> and creating tfrecords file<\/p>\n"},{"AnswerId":"37735786","CreationDate":"2016-06-09T20:42:14.250","ParentId":null,"OwnerUserId":"690430","Title":null,"Body":"<p>Tensorflow's inception model has a file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/f87a58cd96d45de73c9a8330a06b2ab56749a7fa\/research\/inception\/inception\/data\/build_image_data.py\" rel=\"noreferrer\">build_image_data.py<\/a> that can accomplish the same thing with the assumption that each subdirectory represents a label.<\/p>\n"}]}
{"QuestionId":33850090,"AnswerCount":1,"Tags":"<python><windows><64-bit><anaconda><theano>","CreationDate":"2015-11-21T23:57:41.353","AcceptedAnswerId":33991220.0,"OwnerUserId":4380945.0,"Title":"compilation issue when running theano","Body":"<p>I installed theano on windows 8, 64 bit. I am using anaconda implementation, python 3.4. Trying to install theano, I diligently followed all steps on this link (which helped on another computer with similar configuration):<\/p>\n\n<p><a href=\"http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano\" rel=\"nofollow\">http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano<\/a> (English and Korean)<\/p>\n\n<p>whenever I type 'import theano' on my IDE (pycharm) I get a long error message, but I believe the most meaningful portion is:<\/p>\n\n<pre><\/pre>\n\n<p>Would sincerely appreciate some input\nbtw, I installed mignw64, msys, and all other dependencies.<\/p>\n","answers":[{"AnswerId":"33991220","CreationDate":"2015-11-30T03:56:52.537","ParentId":null,"OwnerUserId":"4380945","Title":null,"Body":"<p>I found the problem. the path environment variable was not pointing to the correct mingw-w64 (which is needed to run theano on 64 bit windows) directory.<\/p>\n\n<p>The environment variable should always point to the bin subdirectory of the mingw-w64 directory, as in:<\/p>\n\n<pre><code>C:\\Users\\xxxxx\\Anaconda3\\mingw-w64\\mingw64\\bin\n<\/code><\/pre>\n\n<p>Again, the following post is VERY USEFUL to install theano on 64-bit windows (in English and Korean):<\/p>\n\n<p><a href=\"http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano\" rel=\"nofollow\">http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano<\/a><\/p>\n"}]}
{"QuestionId":33851950,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-22T05:31:59.743","AcceptedAnswerId":null,"OwnerUserId":5590875.0,"Title":"initialising Seq2seq embedding with pretrained word2vec","Body":"<p>I am interested in initialising tensorflow seq2seq implementation with pretrained word2vec.<\/p>\n\n<p>I have seen the code. It seems embedding is initialized<\/p>\n\n<pre><\/pre>\n\n<p>how do I change this to initialise with pretrained word2vec??<\/p>\n","answers":[{"AnswerId":"33948852","CreationDate":"2015-11-27T00:40:25.003","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>You can change the tokanizer present in <code>tensorflow\/models\/rnn\/translate\/data_utils.py<\/code> to use a pre-trained word2vec model for tokenizing. The lines <code>187-190<\/code> of <code>data_utils.py<\/code>:<\/p>\n\n<pre><code>if tokenizer:\n    words = tokenizer(sentence)\nelse:\n    words = basic_tokenizer(sentence)\n<\/code><\/pre>\n\n<p>use <code>basic_tokenizer<\/code>. You can write a <code>tokenizer<\/code> method that uses a pre-trained word2vec model for tokenizing the sentences.<\/p>\n"},{"AnswerId":"33978872","CreationDate":"2015-11-29T02:47:05.637","ParentId":null,"OwnerUserId":"2802468","Title":null,"Body":"<p>I think you've gotten your answer in the mailing list but I am putting it here for posterity.<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/bH6S98NpIJE\" rel=\"noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/bH6S98NpIJE<\/a><\/p>\n\n<blockquote>\n  <p>You can initialize it randomly and afterwards do:\n  session.run(embedding.assign(my_word2vec_matrix))<\/p>\n  \n  <p>This will override the init values.<\/p>\n<\/blockquote>\n\n<p>This seems to work for me. I believe <code>trainable=False<\/code> is needed to keep the values fixed?<\/p>\n\n<pre><code># load word2vec model (say from gensim)\nmodel = load_model(FILENAME, binary=True)\n\n# embedding matrix\nX = model.syn0\nprint(type(X)) # numpy.ndarray\nprint(X.shape) # (vocab_size, embedding_dim)\n\n# start interactive session\nsess = tf.InteractiveSession()\n\n# set embeddings\nembeddings = tf.Variable(tf.random_uniform(X.shape, minval=-0.1, maxval=0.1), trainable=False)\n\n# initialize\nsess.run(tf.initialize_all_variables())\n\n# override inits\nsess.run(embeddings.assign(X))\n<\/code><\/pre>\n"}]}
{"QuestionId":33853032,"AnswerCount":1,"Tags":"<deep-learning><torch>","CreationDate":"2015-11-22T08:43:33.490","AcceptedAnswerId":null,"OwnerUserId":2711403.0,"Title":"Using Torch installation in one Linux system on another linux system","Body":"<p>I have a very amusing and troubling problem. On my home system, I can install Torch7 and use it. <\/p>\n\n<p>I do not have administrative privileges on my office system. Unfortunately, many dependencies of Torch such as readline are hence not satisfied. Is it possible that I can use the torch installation of my system, by copying it to my office system ? What would I have to do for that if that were possible ?<\/p>\n\n<p>To detail my problem, here is what happens when I install torch7 on my office system.<\/p>\n\n<p>I installed ncurses and readline libraries in my home folder. When I install the readline library, (after installing ncurses), it gives me undefined reference problem at the time of make install like this :<\/p>\n\n<p>libreadline.so: undefined reference to tgoto'\nlibreadline.so: undefined reference to UP'\nlibreadline.so: undefined reference to tgetnum'\nlibreadline.so: undefined reference to tgetstr'\nlibreadline.so: undefined reference to `BC'<\/p>\n\n<p>These errors just dont seem to get away. I dont have administrative privileges and i cannot just do sudo yum install.\nIt will take a lot of channels to get them install any new packages,so is it possible to address this issue somehow ?<\/p>\n","answers":[{"AnswerId":"35793920","CreationDate":"2016-03-04T10:40:55.893","ParentId":null,"OwnerUserId":"6017797","Title":null,"Body":"<p>I was facing the exact same problem. Found a solution provided by <a href=\"https:\/\/github.com\/alexbw\/conda-lua-recipes#to-install-packages-warning-this-is-in-flux-quite-a-bit-might-not-be-current\" rel=\"nofollow\">Alex Wiltschko<\/a>. In my case Anaconda was installed, so I could simply run the following command to install Lua and Torch:<\/p>\n\n<pre><code>conda install lua=5.2 lua-science -c alexbw\n<\/code><\/pre>\n\n<p>I have now posted the complete procedure on my blog. You can view the post <a href=\"https:\/\/milindpadalkar.wordpress.com\/2016\/03\/04\/installing-torch-without-root-privileges\/\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>Regards,\nMilind<\/p>\n"}]}
{"QuestionId":33853713,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2015-11-22T10:20:03.887","AcceptedAnswerId":null,"OwnerUserId":1795377.0,"Title":"TensorFlow installation on Ubuntu 14.04 LTS","Body":"<p>When I try to install TensorFlow Google's Machine Learning library on Ubunto using command :<\/p>\n\n<pre><\/pre>\n\n<p>i keep getting this error<\/p>\n\n<pre><\/pre>\n\n<p>I am using pip installed on my machine and have <strong>python 2.7<\/strong> installed on the machine as well<\/p>\n\n<p><strong>vmware image info :<\/strong> <\/p>\n\n<p>No LSB modules are available.<br>\nDistributor ID: Ubuntu <br>\nDescription:    Ubuntu 14.04 LTS<br>\nRelease:    14.04<br>\nCodename:   trusty<br><\/p>\n\n<p>Can anyone please help me to solve this error?<\/p>\n","answers":[{"AnswerId":"33857680","CreationDate":"2015-11-22T17:15:42.493","ParentId":null,"OwnerUserId":"2900034","Title":null,"Body":"<p>Use super user for installation.<\/p>\n\n<p><code>$ sudo pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/code><\/p>\n\n<p>This worked for me.<\/p>\n"},{"AnswerId":"33886255","CreationDate":"2015-11-24T05:43:42.100","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>Tensorflow works fine for me with VMware &amp; Ubuntu. Please try:<\/p>\n\n<pre><code>sudo apt-get install python-pip python-dev\nsudo pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":33855488,"AnswerCount":0,"Tags":"<speech-recognition><theano><lstm>","CreationDate":"2015-11-22T13:49:53.787","AcceptedAnswerId":null,"OwnerUserId":3609599.0,"Title":"Theano based libraries for RNN based speech recognition","Body":"<p>Um trying to do speech recognition using RNN. It would be really helpful if i could get some suggestions on best theano based libraries that I can use for RNN based speech recognition. (I am hoping to use CTC Connectionist Temporal Classification as well)<\/p>\n","answers":[]}
{"QuestionId":33857041,"AnswerCount":0,"Tags":"<theano><deep-learning><caffe>","CreationDate":"2015-11-22T16:17:45.903","AcceptedAnswerId":null,"OwnerUserId":4299527.0,"Title":"Can i turn the CIFAR-10 dataset to grayscale images and convert it to same dimension as MNIST dataset. Will the model be invalid or fail to learn?","Body":"<p>I'm new in the field of Deep Neural Network. There are various deep learning frameworks nearby. Notably , , , and recently open sourced . I have tried out a couple of tutorials with  provided on their site. Specifically the . I guess this is the  of every deep learning framework out there. I also viewed tutorials from <a href=\"http:\/\/cs231n.stanford.edu\" rel=\"nofollow\">here<\/a>. This one was explained in detail, but they do not provide hands on experience with any deep learning frameworks. So which framework should be better for beginners? I looked up similar questions asked on Quora. Some said that theano is tougher to learn but it gives more control, Caffe is easier, but it gives less control over the network. And nothing on Tensorflow, as it is new, but from what i've seen the documentation is not That well written, also it seems tougher to understand. So as a newbie what should i choose to learn?<\/p>\n\n<p>Another question, As I said, MNIST is the  of every deep learning framework, and many neural networks can be found for recognizing  dataset. So, if I use the same network to detect other dataset, say  dataset, will it work?? Let's just say that i turn the  dataset to grayscale images and convert it to same dimension as MNIST dataset. Will the model be invalid or fail to learn? or have bad accuracy or what?<\/p>\n","answers":[]}
{"QuestionId":33858021,"AnswerCount":6,"Tags":"<tensorflow>","CreationDate":"2015-11-22T17:47:05.993","AcceptedAnswerId":33858244.0,"OwnerUserId":4680447.0,"Title":"Outer product in tensorflow","Body":"<p>In tensorflow, there are nice functions for entrywise and matrix multiplication, but after looking through the docs, I cannot find any internal function for taking an <em>outer product<\/em> of two tensors, i.e., making a bigger tensor by all possible products of elements of smaller tensors (like numpy.outer): <\/p>\n\n<pre><\/pre>\n\n<p>or<\/p>\n\n<pre><\/pre>\n\n<p>Does tensorflow have such a function? <\/p>\n","answers":[{"AnswerId":"52202822","CreationDate":"2018-09-06T11:14:02.743","ParentId":null,"OwnerUserId":"10325146","Title":null,"Body":"<p>I would have commented to MasDra, but SO wouldn't let me as a new registered user.<\/p>\n\n<p>The general outer product of multiple vectors arranged in a list U of length order can be obtained via <\/p>\n\n<pre><code>tf.einsum(','.join(string.ascii_lowercase[0:order])+'-&gt;'+string.ascii_lowercase[0:order], *U)\n<\/code><\/pre>\n"},{"AnswerId":"51706531","CreationDate":"2018-08-06T11:25:41.903","ParentId":null,"OwnerUserId":"1079110","Title":null,"Body":"<p>As pointed out in the other answers, the outer product can be done using broadcasting:<\/p>\n\n<pre><code>a = tf.range(10)\nb = tf.range(5)\nouter = a[..., None] * b[None, ...]\n\ntf.InteractiveSession().run(outer)\n# array([[ 0,  0,  0,  0,  0],\n#        [ 0,  1,  2,  3,  4],\n#        [ 0,  2,  4,  6,  8],\n#        [ 0,  3,  6,  9, 12],\n#        [ 0,  4,  8, 12, 16],\n#        [ 0,  5, 10, 15, 20],\n#        [ 0,  6, 12, 18, 24],\n#        [ 0,  7, 14, 21, 28],\n#        [ 0,  8, 16, 24, 32],\n#        [ 0,  9, 18, 27, 36]], dtype=int32)\n<\/code><\/pre>\n\n<p>Explanation:<\/p>\n\n<ul>\n<li>The <code>a[..., None]<\/code> inserts a new dimension of length 1 after the last axis.<\/li>\n<li>Similarly, <code>b[None, ...]<\/code> inserts a new dimension of length 1 before the first axis.<\/li>\n<li>The element-wide multiplication then broadcasts the tensors from shapes <code>(10, 1) * (1, 5)<\/code> to <code>(10,  5) * (10, 5)<\/code>, computing the outer product.<\/li>\n<\/ul>\n\n<p>Where you insert the additional dimensions determines for which dimensions the outer product is computed. For example, if both tensors have a batch size, you can skip that using <code>:<\/code> which gives <code>a[:, ..., None] * b[:, None, ...]<\/code>. This can be further abbreviated as <code>a[..., None] * b[:, None]<\/code>. To perform the outer product over the last dimension and thus supporting any number of batch dimensions, use <code>a[..., None] * b[..., None, :]<\/code>.<\/p>\n"},{"AnswerId":"47750909","CreationDate":"2017-12-11T10:31:33.900","ParentId":null,"OwnerUserId":"6820320","Title":null,"Body":"<p>In case someone else stumbles upon this, according to the tensorflow docs you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/tf\/einsum\" rel=\"noreferrer\"><code>tf.einsum()<\/code><\/a> function to compute the outer product of two tensors <code>a<\/code> and <code>b<\/code>:<\/p>\n\n<pre><code># Outer product\n&gt;&gt;&gt; einsum('i,j-&gt;ij', u, v)  # output[i,j] = u[i]*v[j]\n<\/code><\/pre>\n"},{"AnswerId":"51323886","CreationDate":"2018-07-13T11:16:09.943","ParentId":null,"OwnerUserId":"3424521","Title":null,"Body":"<p>It is somewhat surprising that until recently there was no easy and \"natural\" way of doing an outer product between arbitrary tensors (also known as \"tensor product\") in tensorflow, especially given the name of the library...<\/p>\n\n<p>With <code>tensorflow&gt;=1.6<\/code> you can do now finally get what you want with a simple:<\/p>\n\n<pre><code>M = tf.tensordot(A, B, axes=0)\n<\/code><\/pre>\n\n<p>In earlier versions of tensorflow, <code>axes=0<\/code> raises a <code>ValueError: 'axes' must be at least 1.<\/code>. Somehow <code>tf.tensordot()<\/code> used to need at least one dimension to actually sum over. The easy way out is to simply add a \"fake\" dimension with <code>tf.expand_dims()<\/code>.<\/p>\n\n<p>On <code>tensorflow&lt;=1.5<\/code> you can thus get the same result as above by doing:<\/p>\n\n<pre><code>M = tf.tensordot(tf.expand_dims(A, 0), tf.expand_dims(B, 0), axes=[[0],[0]])\n<\/code><\/pre>\n\n<p>This adds a new index of dimension 1 in location 0 for both tensors and then lets <code>tf.tensordot()<\/code> sum over those indices.<\/p>\n"},{"AnswerId":"51641382","CreationDate":"2018-08-01T19:57:45.890","ParentId":null,"OwnerUserId":"10006823","Title":null,"Body":"<p>tf.multiply (and its '*' shortcut) result in an outer product, whether or not a batch is used. In particular, if the two input tensors have a 3D shapes of [batch, n, 1] and [batch, 1, n] then this op will calculate the outer product for [n,1],[1,n] <em>per each sample in the batch<\/em>. If there is no batch, so that the two input tensors are 2D, this op will calculate the outer product just the same.\nOn the other hand, while tf.tensordot yields the outer product for 2D matrices, it did not broadcast similarly when a batch was added.<\/p>\n\n<p><strong>Without a batch:<\/strong><\/p>\n\n<pre><code>a_np = np.array([[1, 2, 3]]) # shape: (1,3) [a row vector], 2D Tensor\nb_np = np.array([[4], [5], [6]]) # shape: (3,1) [a column vector], 2D Tensor\na = tf.placeholder(dtype='float32', shape=[1, 3])\nb = tf.placeholder(dtype='float32', shape=[3, 1])\nc = a*b # Result: an outer-product of a,b\nd = tf.multiply(a,b) # Result: an outer-product of a,b\ne = tf.tensordot(a,b, axes=[0,1]) # Result: an outer-product of a,b\n<\/code><\/pre>\n\n<p><strong>With a batch:<\/strong><\/p>\n\n<pre><code>a_np = np.array([[[1, 2, 3]], [[4, 5, 6]]]) # shape: (2,1,3) [a batch of two row vectors], 3D Tensor\nb_np = np.array([[[7], [8], [9]], [[10], [11], [12]]]) # shape: (2,3,1) [a batch of two column vectors], 3D Tensor\na = tf.placeholder(dtype='float32', shape=[None, 1, 3])\nb = tf.placeholder(dtype='float32', shape=[None, 3, 1])\nc = a*b # Result: an outer-product per batch\nd = tf.multiply(a,b) # Result: an outer-product per batch\ne = tf.tensordot(a,b, axes=[1,2]) # Does NOT result with an outer-product per batch\n<\/code><\/pre>\n\n<p><strong>Running any of these two graphs:<\/strong><\/p>\n\n<pre><code>sess = tf.Session()\nresult_astrix = sess.run(c, feed_dict={a:a_np, b: b_np})\nresult_multiply = sess.run(d, feed_dict={a:a_np, b: b_np})\nresult_tensordot = sess.run(e, feed_dict={a:a_np, b: b_np})\nprint('a*b:')\nprint(result_astrix)\nprint('tf.multiply(a,b):')\nprint(result_multiply)\nprint('tf.tensordot(a,b, axes=[1,2]:')\nprint(result_tensordot)\n<\/code><\/pre>\n"},{"AnswerId":"33858244","CreationDate":"2015-11-22T18:08:08.790","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Yes, you can do this by taking advantage of the broadcast semantics of tensorflow. Size the first out to size 1xN of itself, and the second to size Mx1 of itself, and you'll get a broadcast to MxN of all of the results when you multiply them.<\/p>\n\n<p>(You can play around with the same thing in numpy to see how it behaves in a simpler context, btw:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = np.array([1, 2, 3, 4, 5]).reshape([5,1])\nb = np.array([6, 7, 8, 9, 10]).reshape([1,5])\na*b\n<\/code><\/pre>\n\n<p>How exactly you do it in tensorflow depends a bit on which axes you want to use and what semantics you want for the resulting multiply, but the general idea applies.<\/p>\n"}]}
{"QuestionId":33858785,"AnswerCount":2,"Tags":"<python><numpy><constants><linear-regression><tensorflow>","CreationDate":"2015-11-22T18:54:40.320","AcceptedAnswerId":null,"OwnerUserId":1936768.0,"Title":"Caching a computed value as a constant in TensorFlow","Body":"<p>Suppose I want to compute the least squares coefficients in TensorFlow using the closed form solution. Normally, I would do this like so,<\/p>\n\n<pre><\/pre>\n\n<p>Where  and  are TensorFlow placeholders corresponding to the covariates and target variable, respectively.<\/p>\n\n<p>If I then wanted to perform prediction, I would do something like,<\/p>\n\n<pre><\/pre>\n\n<p>If I were to execute,<\/p>\n\n<pre><\/pre>\n\n<p>I would of course get an error that I did not provide a necessary value for the placeholder . I would like to have the flexibility to treat  as constant after I have computed it (so that I would not need to define a new placeholder for the new covariate matrix for prediction). One way to accomplish this is,<\/p>\n\n<pre><\/pre>\n\n<p>I was wondering if there were a more elegant way to treat the tensor as constant so that I neither need to execute the session and obtain a constant nor create a separate placeholder for incoming data to be used for prediction.<\/p>\n\n<p>Here is some sample code that demonstrates the circumstance I'm describing.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40848729","CreationDate":"2016-11-28T16:09:31.277","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Perhaps counter-intuitively, the simplest way to re-use <code>beta_hat<\/code> as a constant in subsequent steps would be to assign it to a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow noreferrer\"><code>tf.Variable<\/code><\/a>:<\/p>\n\n<pre><code>n, k = 100, 5\nX = tf.placeholder(dtype=tf.float32, shape=[None, k])\ny = tf.placeholder(dtype=tf.float32, shape=[None, 1])\n\nbeta = np.random.normal(size=(k, ))\ndata_X = np.random.normal(size=(n, k))\n\ndata_y = data_X.dot(beta)\ndata_y += np.random.normal(size=data_y.shape) \/ 3.0\ndata_y = np.atleast_2d(data_y).T\n\n# Convert to 32-bit precision.\ndata_X, data_y = np.float32(data_X), np.float32(data_y)\n\n# Compute the least squares solution.\nbeta_hat = tf.matmul(\n    tf.matmul(tf.matrix_inverse(tf.matmul(tf.transpose(X), X)),\n              tf.transpose(X)), y\n)\n\nbeta_hat_cached = tf.Variable(beta_hat)\n\n# Launch the graph\nsess = tf.Session()\n\nprint \"True beta: {}\".format(beta)\n# Run the initializer, which computes `beta_hat` once:\nsess.run(beta_hat_cached.initializer, feed_dict={X: data_X, y: data_y})\n# To access the value of `beta_hat`, \"run\" the variable to read its contents.\nprint \"Est. beta: {}\".format(beta_hat_cached.ravel())\n\n# Use the cached version to compute predictions.\ny_pred = tf.matmul(X, beta_hat_cached)\nprint \"Predictions:\"\nprint sess.run(y_pred, feed_dict={X: data_X})\n<\/code><\/pre>\n"},{"AnswerId":"48623239","CreationDate":"2018-02-05T13:02:45.200","ParentId":null,"OwnerUserId":"3309610","Title":null,"Body":"<p>mrry has indeed presented an elegant solution. You should consider marking his answer as correct if that is indeed what you want. <\/p>\n\n<p>However, I think this is a good place to clear up what I perceive to be a source of confusion regarding placeholders... This isn't necessarily directed at the person who asked the question, but I believe it will be relevant for a lot of beginners who stumble on this question... <\/p>\n\n<hr>\n\n<p>Placeholders should be thought of like function inputs. So first, let's review how that works in Python, and then I will show the equivalent form in Tensorflow... <\/p>\n\n<p>If I want to have a function that computes the output given various inputs <code>x<\/code> and <code>y<\/code>, then I could do it like this... <\/p>\n\n<pre><code>def f(x,y):\n    # For example... \n    return x * y \n<\/code><\/pre>\n\n<p>Specifically, I can call this function with various values for <code>x<\/code> and <code>y<\/code>:<\/p>\n\n<pre><code>f(1,3) = 3\nf(1,4) = 4\nf(2,3) = 6\nf(2,4) = 8\n<\/code><\/pre>\n\n<p>However, in my particular case, I may have a fixed value of <code>y<\/code>. So in my case, it doesn't make sense to pass <code>y<\/code> as an argument. Instead, I want to bake my value of <code>y<\/code> into the function, and just vary <code>x<\/code>. To do that, I can simply capture the outer value of <code>y<\/code>: <\/p>\n\n<pre><code>y = 3\ndef g(x):\n    return x * y\n<\/code><\/pre>\n\n<p>Now whenever I call <code>g<\/code>, <code>y<\/code> will have the fixed value of 3:<\/p>\n\n<pre><code>g(1) = 3\ng(2) = 6\n<\/code><\/pre>\n\n<p>Similarly, if I also know that <code>x<\/code> is fixed, I could capture the outer value of <code>x<\/code>:<\/p>\n\n<pre><code>x = 2\ndef h():\n   return g(x)\n<\/code><\/pre>\n\n<p>Now, when I call <code>h<\/code>, I am implicitly calling <code>h()=g(2)=f(2,3)<\/code>.<\/p>\n\n<p>That's great, but the problem is that everytime I call <code>h<\/code>, it will redo the multiplication because it is equivalent to calling <code>f(2,3)<\/code>. Thus to improve performance, I can evaluate the expression, and then have a function that just returns this precomputed value: <\/p>\n\n<pre><code>val = h()\ndef h23():\n    return val\n<\/code><\/pre>\n\n<p>No matter how many times I call <code>h23<\/code>, the multiplication is only performed once (on the line <code>val = h()<\/code>). <\/p>\n\n<p>Tensorflow has analogous concepts. <\/p>\n\n<p>If you want to have a function where you can vary both inputs, then you should make placeholder objects for both instances, and pass the values to the function in a feed dictionary when running in a session:<\/p>\n\n<pre><code>dtype = tf.float64\nshape = ()\nx = tf.placeholder( dtype, shape )\ny = tf.placeholder( dtype, shape )\nfxy = f(x,y)\nwith tf.Session() as sess: \n    print( sess.run( fxy, {x:1,y:3} ) )\n    print( sess.run( fxy, {x:1,y:4} ) )\n    print( sess.run( fxy, {x:2,y:3} ) )\n    print( sess.run( fxy, {x:2,y:4} ) )\n<\/code><\/pre>\n\n<p>However, if one of my values does not change, then I can directly initialize it as a constant and create a new function with this value \"baked into it\":<\/p>\n\n<pre><code>y = tf.constant( 3 )\ngx = f( x, y )\nwith tf.Session() as sess: \n    print( sess.run( gx, {x:1} ) )\n    print( sess.run( gx, {x:2} ) )\n<\/code><\/pre>\n\n<p>The key point is that now I do not need to pass a value for <code>y<\/code> in my feed dictionary. It is constant and captured in the expression <code>gx<\/code>.<br>\nSimilarly, if <code>x<\/code> is also a constant, then I should declare it so:<\/p>\n\n<pre><code>x = tf.constant(2)\nh = f(x,y)\nwith tf.Session() as sess: \n    print( sess.run( h ) )\n<\/code><\/pre>\n\n<p>As you can see, since all of my variables are constant, I don't need a feed dictionary at all. This is the Tensorflow equivalent of calling a function without arguments, like <code>h()<\/code>. <\/p>\n\n<p>However, just as before, when I call <code>h<\/code>, it may need to reevaluate the graph every time. So I have two options. <\/p>\n\n<ol>\n<li>I can compute the result in numpy, then wrap that value with a tensorflow constant. <\/li>\n<li>I can compute the output in tensorflow, run it in a session to get the numpy value, and then wrap it in a constant. <\/li>\n<\/ol>\n\n<p>In the first option, I would do something like this<\/p>\n\n<pre><code>fxy = tf.constant( f(2,3) ) \n<\/code><\/pre>\n\n<p>Now I have precomputed the value of the function outside of Tensorflow, and then wrapped that value as a constant so that I can use it in other tensorflow functions. <\/p>\n\n<p>Conversely, you would only consider option 2 if your function uses some complicated Tensorflow intrinsics, or if your function takes really long to run and you think it will be faster to computer in Tensorflow:<\/p>\n\n<pre><code>with tf.Session() as sess: \n    fxy = tf.constant( sess.run( h ) )\n<\/code><\/pre>\n\n<p>To understand what is going on here, recall that <\/p>\n\n<pre><code>h = f( tf.constant(1), tf.constant(3) )\n<\/code><\/pre>\n\n<p>So I don't need to pass a feed dict. The snippet <code>sess.run( h )<\/code> runs that multiplication inside of tensorflow and returns it as a Numpy array. Then finally, I wrap that value with a tf.constant so that I can use it in other Tensorflow functions. <\/p>\n"}]}
{"QuestionId":33858990,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-22T19:09:57.320","AcceptedAnswerId":null,"OwnerUserId":5285204.0,"Title":"How to reuse computation graph for different inputs?","Body":"<p>I have my main flow of computation set up that I can train using <\/p>\n\n<pre><\/pre>\n\n<p>Similarly, I have a function for predictions<\/p>\n\n<pre><\/pre>\n\n<p>Both of these functions accept the input  and send it through the same computation graph. <\/p>\n\n<p>I would now like to modify things so that when training, I can train using a noisy input, so I have something like<\/p>\n\n<pre><\/pre>\n\n<p>at the beginning of the computations.<\/p>\n\n<p>But this will also affect my  function, since its input will get corrupted as well. How can I reuse the same code for  and , but only provide the former with the noisy input?<\/p>\n","answers":[{"AnswerId":"33866260","CreationDate":"2015-11-23T08:13:43.983","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can organise your code like this:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\nimport theano.tensor.shared_randomstreams\n\n\ndef get_cost(x, y):\n    return tt.mean(tt.sum(tt.sqr(x - y), axis=1))\n\n\ndef get_output(x, w, b_h, b_y):\n    h = tt.tanh(tt.dot(x, w) + b_h)\n    y = tt.dot(h, w.T) + b_y\n    return y\n\n\ndef corrupt_input(x, corruption_level):\n    rng = tt.shared_randomstreams.RandomStreams()\n    return rng.binomial(size=x.shape, n=1, p=1 - corruption_level,\n                        dtype=theano.config.floatX) * x\n\n\ndef compile(input_size, hidden_size, corruption_level, learning_rate):\n    x = tt.matrix()\n    w = theano.shared(numpy.random.randn(input_size,\n                      hidden_size).astype(theano.config.floatX))\n    b_h = theano.shared(numpy.zeros(hidden_size, dtype=theano.config.floatX))\n    b_y = theano.shared(numpy.zeros(input_size, dtype=theano.config.floatX))\n    cost = get_cost(x, get_output(corrupt_input(x, corruption_level), w, b_h, b_y))\n    updates = [(p, p - learning_rate * tt.grad(cost, p)) for p in (w, b_h, b_y)]\n    train = theano.function(inputs=[x], outputs=cost, updates=updates)\n    predict = theano.function(inputs=[x], outputs=get_output(x, w, b_h, b_y))\n    return train, predict\n\n\ndef main():\n    train, predict = compile(input_size=3, hidden_size=2,\n                             corruption_level=0.2, learning_rate=0.01)\n\n\nmain()\n<\/code><\/pre>\n\n<p>Note that <code>get_output<\/code> is called twice. For the <code>train<\/code> function it is provided with the corrupted input but for the <code>predict<\/code> function it is provided with the clean input. <code>get_output<\/code> needs to contain \"the same computation graph\" that you talk of. I've just put a tiny autoencoder in there but you can put whatever you want in there.<\/p>\n\n<p>Assuming the corrupted input has the same shape as the input, the <code>get_output<\/code> function won't care whether its input is <code>x<\/code> or the corrupted version of <code>x<\/code>. So <code>get_output<\/code> can be shared but need not contain the corruption code.<\/p>\n"}]}
{"QuestionId":33859864,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2015-11-22T20:35:39.867","AcceptedAnswerId":null,"OwnerUserId":4495248.0,"Title":"How to create custom objective function in Keras?","Body":"<p>There are many objective functions in Keras <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/objectives.py\" rel=\"noreferrer\">here.<\/a><\/p>\n\n<p>But how can you create your own objective function, I tried to create a very basic objective function but it gives an error and I there is no way to know the size of the parameters passed to the function at run time.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting 2 contradicting errors, <\/p>\n\n<pre><\/pre>\n\n<p>It says cost or loss returned in the function must be a scalar but if I change the line 2 from\nloss = T.vector('float64')<br>\nto<br>\nloss = T.scalar('float64')<\/p>\n\n<p>it shows this error <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39789380","CreationDate":"2016-09-30T10:29:12.833","ParentId":null,"OwnerUserId":"6901690","Title":null,"Body":"<p>(Answer Fixed) A simple way to do it is calling Keras backend:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef custom_loss(y_true,y_pred):\n    return K.mean((y_true - y_pred)**2)\n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>model.compile(loss=custom_loss, optimizer=sgd,metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>that equals<\/p>\n\n<pre><code>model.compile(loss='mean_squared_error', optimizer=sgd,metrics = ['accuracy'])\n<\/code><\/pre>\n"},{"AnswerId":"40622302","CreationDate":"2016-11-16T00:51:41.277","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Here is my small snippet to write new loss functions and test them before using:<\/p>\n\n<pre><code>import numpy as np\n\nfrom keras import backend as K\n\n_EPSILON = K.epsilon()\n\ndef _loss_tensor(y_true, y_pred):\n    y_pred = K.clip(y_pred, _EPSILON, 1.0-_EPSILON)\n    out = -(y_true * K.log(y_pred) + (1.0 - y_true) * K.log(1.0 - y_pred))\n    return K.mean(out, axis=-1)\n\ndef _loss_np(y_true, y_pred):\n    y_pred = np.clip(y_pred, _EPSILON, 1.0-_EPSILON)\n    out = -(y_true * np.log(y_pred) + (1.0 - y_true) * np.log(1.0 - y_pred))\n    return np.mean(out, axis=-1)\n\ndef check_loss(_shape):\n    if _shape == '2d':\n        shape = (6, 7)\n    elif _shape == '3d':\n        shape = (5, 6, 7)\n    elif _shape == '4d':\n        shape = (8, 5, 6, 7)\n    elif _shape == '5d':\n        shape = (9, 8, 5, 6, 7)\n\n    y_a = np.random.random(shape)\n    y_b = np.random.random(shape)\n\n    out1 = K.eval(_loss_tensor(K.variable(y_a), K.variable(y_b)))\n    out2 = _loss_np(y_a, y_b)\n\n    assert out1.shape == out2.shape\n    assert out1.shape == shape[:-1]\n    print np.linalg.norm(out1)\n    print np.linalg.norm(out2)\n    print np.linalg.norm(out1-out2)\n\n\ndef test_loss():\n    shape_list = ['2d', '3d', '4d', '5d']\n    for _shape in shape_list:\n        check_loss(_shape)\n        print '======================'\n\nif __name__ == '__main__':\n    test_loss()\n<\/code><\/pre>\n\n<p>Here as you can see I am testing the binary_crossentropy loss, and have 2 separate losses defined, one numpy version (_loss_np) another tensor version (_loss_tensor) [Note: if you just use the keras functions then it will work with both Theano and Tensorflow... but if you are depending on one of them you can also reference them by K.theano.tensor.function or K.tf.function]<\/p>\n\n<p>Later I am comparing the output shapes and the L2 norm of the outputs (which should be almost equal) and the L2 norm of the difference (which should be towards 0) <\/p>\n\n<p>Once you are satisfied that your loss function is working properly you can use it as:<\/p>\n\n<pre><code>model.compile(loss=_loss_tensor, optimizer=sgd)\n<\/code><\/pre>\n"}]}
{"QuestionId":33862317,"AnswerCount":0,"Tags":"<python><python-2.7><theano><lasagne><nolearn>","CreationDate":"2015-11-23T01:12:12.337","AcceptedAnswerId":null,"OwnerUserId":741099.0,"Title":"Very High Errors in Deep Belief Network with Nolearn in Python","Body":"<p>I'm new to Nolearn and Theano in general. When I tried the code in <a href=\"http:\/\/www.pyimagesearch.com\/2014\/09\/22\/getting-started-deep-learning-python\/\" rel=\"nofollow noreferrer\">a Nolearn tutorial<\/a>, I get extremely high error rates of 0.9!<\/p>\n\n<p>Why am I getting such high errors while the tutorial is getting errors of 0.005? Is anyone else able to reproduce this problem?<\/p>\n\n<p>Using Theano 0.7.0, Lasagne v0.1, nolearn v0.5 on OS X Yosemite.<\/p>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Code<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Classification report<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0NncR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0NncR.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
{"QuestionId":33862983,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-23T02:42:40.627","AcceptedAnswerId":null,"OwnerUserId":5593400.0,"Title":"Two feedforward propagation needed for a single weight update in TensorFlow?","Body":"<p>In TensorFlow, it seems that we have to propagate the input to the top layer once to compute the current training error, and propagate another time to do the parameter update.<\/p>\n\n<p>At the bottom of <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/index.md\" rel=\"nofollow\">the Tensorflow MNIST example<\/a>, the line:<\/p>\n\n<pre><\/pre>\n\n<p>was used to compute accuracy, which required a feedforward process. <\/p>\n\n<p>Next, the line<\/p>\n\n<pre><\/pre>\n\n<p>was used to update the weight parameters. I think this process also requires a feedforward propagation first to get the network output so that the mean squared error can be calculated for back-propagation algorithm. <\/p>\n\n<p>Do we really need to go through the whole network twice to get current training error and update parameters??? <\/p>\n","answers":[{"AnswerId":"33865102","CreationDate":"2015-11-23T06:45:30.113","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>No need to go through twice.  Check out the other examples, such as <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/mnist\/convolutional.py\" rel=\"nofollow\">mnist\/convolutional.py<\/a>:<\/p>\n\n<pre><code>_, l, lr, predictions = s.run(\n          [optimizer, loss, learning_rate, train_prediction],\n          feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>You pull both nodes at the same time in run, and get both the training done and the train prediction at the same time.  This is the standard way of training.<\/p>\n\n<p>In general, I'd suggest checking out the examples in models\/ first.  The \"red pill\" and \"blue pill\" examples are meant to be a very gentle introduction to tensorflow, but the examples in models are a bit more real.  They're not production, but they're closer to what you'd want to do.<\/p>\n"}]}
{"QuestionId":33863081,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-23T02:55:47.333","AcceptedAnswerId":null,"OwnerUserId":803253.0,"Title":"TensorFlow Tutorial on Sequence-to-Sequence Models: Running translate.py return CRC check failure","Body":"<p>I am trying to run tutorial for sequence-to-sequence model, but running the following command<\/p>\n\n<pre><\/pre>\n\n<p>ends with the following error<\/p>\n\n<pre><\/pre>\n\n<p>Did I do something wrong, or is there (another) error in the tutorial?<\/p>\n","answers":[{"AnswerId":"33906475","CreationDate":"2015-11-25T00:28:51.090","ParentId":null,"OwnerUserId":"5602038","Title":null,"Body":"<p>Try to remove files in dir \/tmp:<\/p>\n\n<ul>\n<li>training-giga-fren.tar<\/li>\n<li>giga-fren.release2.en.gz<\/li>\n<li>giga-fren.release2.fr.gz<\/li>\n<\/ul>\n\n<p>And then rerun \nbazel run -c opt &lt;...>\/models\/rnn\/translate\/translate.py --data_dir [your_data_directory]<\/p>\n"}]}
{"QuestionId":33868378,"AnswerCount":1,"Tags":"<python><correlation><theano>","CreationDate":"2015-11-23T10:11:20.303","AcceptedAnswerId":33949789.0,"OwnerUserId":232371.0,"Title":"Circular Correlation in Theano","Body":"<p>I'm trying to compute <a href=\"https:\/\/en.wikipedia.org\/wiki\/Discrete_Fourier_transform#Circular_convolution_theorem_and_cross-correlation_theorem\" rel=\"nofollow\">circular cross-correlation<\/a> of two signals with Theano to use it in further calculation of loss that I would optimize over. But I'm not quite sure how to do that.<\/p>\n\n<p>It is defined as following:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>\"periodic\" summation or like \"for each k-th component\".<\/li>\n<\/ul>\n\n<p>I could do an ordinary correlation and then perform periodic summation, but it's not quite clear how to do that (periodic summation) symbolically (using scan, probably?)<\/p>\n\n<pre><\/pre>\n\n<p>or use circular cross-correlation theorem and compute as a iFFT(FFT(x)*FFT(y)):<\/p>\n\n<pre><\/pre>\n\n<p>but in this case I can't actually compute a gradient because gradient for ifft (and all functions that has something to do with complex numbers in general, afaik) is not implemented yet, I guess (aborts with an error: )<\/p>\n","answers":[{"AnswerId":"33949789","CreationDate":"2015-11-27T03:18:40.983","ParentId":null,"OwnerUserId":"232371","Title":null,"Body":"<p>Here's a working solution I came up with (definitely not optimal as soon as FFT is not used):<\/p>\n\n<pre><code>def circular_crosscorelation(X, y):\n    \"\"\" \n    Input:\n        symbols for X [n, m]\n        and y[m,]\n\n    Returns: \n        symbol for circular cross corelation of each of row in X with \n        cc[n, m]\n    \"\"\"\n    n, m = X.shape\n    corr_expr = T.signal.conv.conv2d(X, y[::-1].reshape((1, -1)), image_shape=(1, m), border_mode='full')\n    corr_len = corr_expr.shape[1]\n    pad = m - corr_len%m\n    v_padded = T.concatenate([corr_expr, T.zeros((n, pad))], axis=1)\n    circ_corr_exp = T.sum(v_padded.reshape((n, v_padded.shape[1] \/ m, m)), axis=1)\n    return circ_corr_exp[:, ::-1]\n\nX = T.dmatrix()\ny = T.dmatrix()\ncc = theano.function([X, y], circular_crosscorelation(X, y))\nprint cc( np.array([[2, 3, 5], [4, 5, 6]]), np.array([[7, 11, 13]]) )\n<\/code><\/pre>\n\n<p>returns<\/p>\n\n<pre><code>[[  94.  108.  108.]\n [ 149.  157.  159.]]\n<\/code><\/pre>\n\n<p>as expected.<\/p>\n\n<p>And can be analytically differentiated:<\/p>\n\n<pre><code>score = T.sum(circ_corr_exp**2)\ngrad = T.grad(score, x)\ng = theano.function([x, y], outputs=grad)\nprint g( np.array([[2, 3, 5], [4, 5, 6]]), np.array([[7, 11, 13]]) )\n\n&gt;&gt; [[ 6332.  6388.  6500.]\n&gt;&gt;  [ 9554.  9610.  9666.]]\n<\/code><\/pre>\n\n<hr>\n\n<p>here's also few more options (through direct circulant calculation) and time-comparation:<\/p>\n\n<pre><code>def circulant_np(v):\n    row = np.arange(len(v))\n    col = -np.arange(len(v))\n    idx = (row[:, np.newaxis] + col)%len(v)\n    return v[idx]\n\nprint circulant_np(np.array([1, 2, 3, 5]))\n\ndef c_corr_np(a, b):\n    return circulant_np(a).dot(b[::-1])\n\ndef circulant_t(v):\n    row = T.arange(v.shape[0])\n    col = -T.arange(v.shape[0])\n    idx = (row.reshape((-1, 1)) + col)%v.shape[0]\n    return v[idx]\n\ndef c_corr_t_f(a, b):\n    \"\"\" 1d correlation using circulant matrix \"\"\"\n    return circulant_t(a).dot(b[::-1])\n\na = T.dvector('a')\nb = T.dvector('b')\nc_corr_t = theano.function([a, b], c_corr_t_f(a, b))\n\nprint c_corr_np(np.array([2, 3, 5]), np.array([7, 11, 13]))\nprint c_corr_t(np.array([2, 3, 5]), np.array([7, 11, 13]))\nprint c_corr( np.array([[2, 3, 5]]), np.array([[7, 11, 13]]) )\n\n%timeit c_corr_np(np.array([2, 3, 5]), np.array([7, 11, 13]))\n%timeit c_corr_t(np.array([2, 3, 5]), np.array([7, 11, 13]))\n%timeit c_corr( np.array([[2, 3, 5]]), np.array([[7, 11, 13]]) ) # = circular_crosscorelation\n<\/code><\/pre>\n\n<p>which gives <\/p>\n\n<pre><code>10000 loops, best of 3: 30.6 \u00b5s per loop\n10000 loops, best of 3: 132 \u00b5s per loop\n10000 loops, best of 3: 149 \u00b5s per loop\n<\/code><\/pre>\n\n<p>and inverse cross-corr:<\/p>\n\n<pre><code>def inverse_circular_crosscorelation(y):\n    \"\"\" \n    Input:\n        symbol for y[1, m]\n\n    Returns: \n        symbol for y_inv s.t. \n        cc( y, y_inv ) = (1, 0 ... 0)\n    \"\"\"\n\n    A = circulant_t(y.reshape((-1, )))\n    b = T.concatenate([T.zeros((y.shape[1] - 1, )), T.ones((1, ))]).reshape((-1, 1))\n    return T.nlinalg.matrix_inverse(A).dot(b).reshape((1, -1))[:, ::-1]\n<\/code><\/pre>\n"}]}
{"QuestionId":33868433,"AnswerCount":1,"Tags":"<python><ubuntu><cuda><tensorflow>","CreationDate":"2015-11-23T10:14:44.377","AcceptedAnswerId":33873885.0,"OwnerUserId":4186233.0,"Title":"python - tensorflow execution on Ubuntu server with CUDA GeForce9600GT","Body":"<p>I just went through the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/index.md\" rel=\"nofollow\">cifar10 tensorflow tutorial<\/a> on my mac. This tutorial aroused my enthusiasm for NN so I set a Ubuntu Server up in order to train in large scale.<\/p>\n\n<p>However I cant run the cifar10_train.py on my server. I always run into this error:<\/p>\n\n<p>Path of the tensorflow source code:<\/p>\n\n<pre><\/pre>\n\n<p>Path of the tensorflow virtualenv installation:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Cmd:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I installed tensorflow just like on my mac in a virtualenv and activated it correctly before executing the script. \nAs by many in other threads suggested I upgraded already six, however I still got the same error.<\/p>\n\n<p><strong>Update 1<\/strong>\nAfter going the issue threads on github\/tensorflow through I noticed this is a bug involving cuda. I added these to my path environment:<\/p>\n\n<pre><\/pre>\n\n<p>I still run into a error, however it has shortened down to the following:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Update 2<\/strong> As someone suggested, I install protobuf via pip. The Error has once again change for some weird reason:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33873885","CreationDate":"2015-11-23T14:48:52.237","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>TensorFlow currently requires CUDA toolkit 7.0 and cuDNN.<\/p>\n\n<p>cuDNN requires a cc 3.0 GPU, and CUDA toolkit 7.0 requires a cc 2.0 GPU.<\/p>\n\n<p>Your 9600GT does not satisfy these <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md\" rel=\"nofollow\">requirements<\/a>:<\/p>\n\n<blockquote>\n  <p>In order to build or run TensorFlow with GPU support, both Cuda Toolkit 7.0 and CUDNN 6.5 V2 from NVIDIA need to be installed.<\/p>\n  \n  <p>TensorFlow GPU support requires having a GPU card with NVidia Compute Capability >= 3.5. <\/p>\n<\/blockquote>\n\n<p>So if you want to use TensorFlow with GPU support, you will need a cc3.5 or higher GPU, and follow the steps to install the needed support software correctly.  Alternatively, you could install TensorFlow without GPU support.<\/p>\n"}]}
{"QuestionId":33870404,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-11-23T11:49:54.287","AcceptedAnswerId":null,"OwnerUserId":1682139.0,"Title":"Problems installing TensorFlow on Mac","Body":"<p>I am trying to follow to the installation guide on <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.md\" rel=\"nofollow\">tensorflow.org<\/a> and have installed Python version 2 again for that reason using Homebrew. <\/p>\n\n<p>When I run the installation as described<\/p>\n\n<pre><\/pre>\n\n<p>I get this error message:<\/p>\n\n<pre><\/pre>\n\n<p>I am obviously doing something wrong, but have no idea. Any clues?<\/p>\n","answers":[{"AnswerId":"33875367","CreationDate":"2015-11-23T16:03:31.710","ParentId":null,"OwnerUserId":"904032","Title":null,"Body":"<p>On the Mac, I didn't have any problem installing tensorflow with the anaconda version of python: <a href=\"https:\/\/www.continuum.io\/downloads\" rel=\"nofollow\">https:\/\/www.continuum.io\/downloads<\/a> <\/p>\n\n<p>The anaconda version also provides science, math, engineering, and data analysis packages.  A lot of people on <a href=\"https:\/\/www.kaggle.com\/\" rel=\"nofollow\">https:\/\/www.kaggle.com\/<\/a> seem to use this...just a thought.<\/p>\n"},{"AnswerId":"34746189","CreationDate":"2016-01-12T14:23:56.003","ParentId":null,"OwnerUserId":"588709","Title":null,"Body":"<p>I do not want to use virtualenv, since anaconda already comes with its own environment management <code>conda<\/code>. When installing the newest version 0.6.0 directly with pip install, I had a similar error. It seemed to not resolve the dependencies correctly.<\/p>\n\n<p>Here is what you can try:<\/p>\n\n<ol>\n<li>Install anaconda<\/li>\n<li>Create a new conda workspace<\/li>\n<li>Download the specific protobuf version that tensorflow needs: <a href=\"https:\/\/pypi.python.org\/pypi\/protobuf\/3.0.0a3\" rel=\"nofollow\">https:\/\/pypi.python.org\/pypi\/protobuf\/3.0.0a3<\/a><\/li>\n<li>Install it via <code>sudo easy_install ~\/Downloads\/protobuf-3.0.0a3-py2.7.egg<\/code><\/li>\n<li>Install a numpy version greater than 1.08.x via <code>conda install numpy<\/code><\/li>\n<li>Download the 0.6.0 version of tensorflow: <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl<\/a><\/li>\n<li>Install via <code>pip install ~\/Downloads\/tensorflow-0.6.0-py2-none-any.whl<\/code><\/li>\n<\/ol>\n\n<p>When you install tensorflow from the whl file directly, it should tell you when dependencies are not there. It seems not to be able to resolve these conflicts independently. My setup had issues with protobuf and numpy. After installing them manually everything worked fine.<\/p>\n\n<p>I hope this helps!<\/p>\n"},{"AnswerId":"33870888","CreationDate":"2015-11-23T12:15:59.540","ParentId":null,"OwnerUserId":"4186233","Title":null,"Body":"<p>It seems to be a common issue. Try to install it in the virtualenv. Its a much better solution, as you can always easily set up a new version of tensorflow without conflicts.<\/p>\n\n<p>VirutalEnv Tutorial:\n<a href=\"http:\/\/tensorflow.org\/get_started\/os_setup.md#virtualenv-based_installation\" rel=\"nofollow\">http:\/\/tensorflow.org\/get_started\/os_setup.md#virtualenv-based_installation<\/a> <\/p>\n"}]}
{"QuestionId":33870427,"AnswerCount":0,"Tags":"<ios><torch>","CreationDate":"2015-11-23T11:50:58.040","AcceptedAnswerId":null,"OwnerUserId":4918620.0,"Title":"iOS torchActive always return false","Body":"<p>I want to test the torch, so I write some code  <\/p>\n\n<pre><\/pre>\n\n<p>}<\/p>\n\n<p>and the   which is a member variable. Like this <\/p>\n\n<p>The question is I only can make the torchMode = .On, because the torchActive always return flase!!<\/p>\n","answers":[]}
{"QuestionId":33872495,"AnswerCount":1,"Tags":"<lua><lua-table><torch>","CreationDate":"2015-11-23T13:38:44.593","AcceptedAnswerId":null,"OwnerUserId":4364020.0,"Title":"torch.Tensor manipulation - call methods on sub vectors","Body":"<p>I've got a tensor which is a vector with 9 cells, I would like to pull cells 3 to 6 with out looping, like in Matlab.<\/p>\n\n<p>More Info:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33872791","CreationDate":"2015-11-23T13:54:37.903","ParentId":null,"OwnerUserId":"1190388","Title":null,"Body":"<p>I have no experience with the torch library, but the documentation for <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md\" rel=\"nofollow\"><code>torch.Tensor<\/code><\/a> shows that there exist a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-subdim1s-dim1e---dim4s--dim4e\" rel=\"nofollow\"><code>:sub<\/code><\/a> and a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-selectdim-index\" rel=\"nofollow\"><code>:select<\/code><\/a> meta methods as well as a really great <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\"><code>__index<\/code><\/a> method, which supports slicing of a tensor.<\/p>\n\n<p>You can use one of those to get the tensor for index 3 to 6 and apply <code>torch.mean<\/code> on it.<\/p>\n"}]}
{"QuestionId":33873211,"AnswerCount":3,"Tags":"<c++><linux><blas><caffe><openblas>","CreationDate":"2015-11-23T14:17:37.690","AcceptedAnswerId":33875083.0,"OwnerUserId":1179925.0,"Title":"Is it possible to switch between BLAS libraries without recompiling program?","Body":"<p>For example can I have <a href=\"http:\/\/math-atlas.sourceforge.net\/\">Atlas<\/a>, <a href=\"http:\/\/www.openblas.net\/\">OpenBlas<\/a>, <a href=\"https:\/\/software.intel.com\/en-us\/intel-mkl\">MKL<\/a> installed on my Ubuntu 14.04 at the same time and switch between them without recompiling <a href=\"http:\/\/caffe.berkeleyvision.org\/\">Caffe<\/a>?<\/p>\n","answers":[{"AnswerId":"33888750","CreationDate":"2015-11-24T08:31:01.860","ParentId":null,"OwnerUserId":"4128833","Title":null,"Body":"<p>You could also do it without changing system wide settings, for example by adding the library you want to use to the <code>LD_PRELOAD<\/code> or <code>LD_LIBRARY_PATH<\/code> environment variables. The first library on this path will be the one used to resolve the symbols against.<\/p>\n\n<p>For example, you could run with<\/p>\n\n<pre><code>LD_PRELOAD=\/path\/to\/blas\/lib.so .\/my_executable_using_caffe\n<\/code><\/pre>\n\n<p>You can see that this approach would be extremely useful as part of a benchmarking script for different implementations since it doesn't affect the benchmarking environment itself. For example (in bash):<\/p>\n\n<pre><code>my_libraries=\/path\/to\/blas1.so \/path\/to\/blas2.so\n\nfor lib in $my_libraries\ndo\n  LD_PRELOAD=${lib} .\/my_executable_using_caffe\ndone\n<\/code><\/pre>\n\n<p>This approach, based on dynamic linking, applies to any other <em>shared library<\/em> your program is compiled against.<\/p>\n"},{"AnswerId":"33875083","CreationDate":"2015-11-23T15:48:57.467","ParentId":null,"OwnerUserId":"4044696","Title":null,"Body":"<p>Sure, you have to have them installed, and on Ubuntu\/Debian issue command<\/p>\n\n<pre><code>update-alternatives --config libblas.so.3\n<\/code><\/pre>\n\n<p>You'll get numbered list of alternatives and could switch between them easily<\/p>\n\n<p>Link: <a href=\"https:\/\/wiki.debian.org\/DebianScience\/LinearAlgebraLibraries\" rel=\"noreferrer\">https:\/\/wiki.debian.org\/DebianScience\/LinearAlgebraLibraries<\/a><\/p>\n"},{"AnswerId":"33874874","CreationDate":"2015-11-23T15:38:17.287","ParentId":null,"OwnerUserId":"2707697","Title":null,"Body":"<p>Yes, you can. You just need to use dynamic library linking.\nThe <code>libblas.so.3<\/code> is a soft linking to <code>\/etc\/alternatives\/libblas.so.3<\/code>\nWhich in turn points to the <code>BLAS<\/code> library implementation<\/p>\n\n<p>For example if you installing <code>Atlas<\/code>  (<code>sudo apt-get install libatlas3-base<\/code>):<\/p>\n\n<pre><code>\/usr\/lib\/libblas.so.3 -&gt; \/etc\/alternatives\/libblas.so.3\n\/etc\/alternatives\/libblas.so.3 -&gt; \/usr\/lib\/atlas-base\/atlas\/libblas.so.3\n<\/code><\/pre>\n\n<p>while after installing <code>Openblas<\/code> (<code>sudo apt-get install libopenblas-base<\/code>):<\/p>\n\n<pre><code>\/usr\/lib\/libblas.so.3 -&gt; \/etc\/alternatives\/libblas.so.3\n\/etc\/alternatives\/libblas.so.3 -&gt; \/usr\/lib\/openblas-base\/libblas.so.3\n<\/code><\/pre>\n\n<p>Of course you can use your own soft link to your library builds.<\/p>\n"}]}
{"QuestionId":33875262,"AnswerCount":1,"Tags":"<neural-network><torch><lstm><recurrent-neural-network>","CreationDate":"2015-11-23T15:58:30.763","AcceptedAnswerId":null,"OwnerUserId":3612823.0,"Title":"LSTM Model in Torch not Learning","Body":"<p>I have a toy dataset with one input and one output, the inputs are random 0s or 1s and the output equals the input from 3 samples previous<\/p>\n\n<pre><\/pre>\n\n<p>Model 1:<\/p>\n\n<pre><\/pre>\n\n<p>Model 2:<\/p>\n\n<pre><\/pre>\n\n<p>Criterion:<\/p>\n\n<pre><\/pre>\n\n<p>The first model doesn't learn and the second doesn't work (possible because my data is the wrong kind for a Sequencer?)<\/p>\n\n<p>edit:<\/p>\n\n<p>Training:<\/p>\n\n<pre><\/pre>\n\n<p>I now get an error when running Model 2 through training.  The error <\/p>\n\n<blockquote>\n  <p>AbstractRecurrent.lua:61: Missing gradInput<\/p>\n<\/blockquote>\n\n<p>occurs on line <\/p>\n\n<blockquote>\n  <p>lstm.modules[1]:backwardThroughTime()<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"33931282","CreationDate":"2015-11-26T05:36:55.783","ParentId":null,"OwnerUserId":"988191","Title":null,"Body":"<p>look at the toy example for using LSTM with Sequencer. The model doesn't do anything except try to predict random numbers.<\/p>\n\n<pre><code>require 'nn'\nrequire 'torch'\nrequire 'rnn'\n\nlstm = nn.Sequencer(\n   nn.Sequential()\n      :add(nn.LSTM(1,100))\n      :add(nn.Linear(100,1))\n      :add(nn.Sigmoid())\n   )\n\n-- Encapsulate thr criterion using a Sequencer\n-- Just provide the entire sequence as input and the corresponding\n-- target sequence as expected output\ncriterion = nn.SequencerCriterion(nn.BCECriterion())\n\ndata = torch.zeros(10,2)\nfor i=1,data:size(1) do\n  data[i][1] = torch.uniform()\n  data[i][2] = torch.uniform()\nend\n\nlocal inputs, targets = {},{}\n\nfor epoch=1,5 do\n  lstm:training()\n\n  for i=1,data:size(1) do\n    --useful for minibatch\n    inputs[1] = torch.zeros(1)\n    inputs[1][1] = data[i][1]\n\n    targets[1] = torch.zeros(1)\n    targets[1][1] = data[i][2]\n\n    local output = lstm:forward(inputs)\n    local err = criterion:forward(output, targets)\n\n    local gradOutputs = criterion:backward(output, targets)\n    -- Sequencer handles the backwardThroughTime internally\n    lstm:backward(inputs, gradOutputs)\n    lstm:updateParameters(0.01)\n    lstm:zeroGradParameters()\n\n    inputs = {}\n    targets = {}\n  end\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":33875887,"AnswerCount":1,"Tags":"<installation><machine-learning><tensorflow>","CreationDate":"2015-11-23T16:29:41.027","AcceptedAnswerId":null,"OwnerUserId":1795377.0,"Title":"TensorFlow installation on Ubuntu","Body":"<p>When I try to install TensorFlow Machine Learning library on Ubunto (vmware image) using command :<\/p>\n\n<pre><\/pre>\n\n<p>after downloading the package I got this error :<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n        File \"\", line 14, in \n      IOError: [Errno 2] No such file or directory: '\/tmp\/pip-GgS7fR-build\/setup.py'\n      Complete output from command python setup.py egg_info:\n      Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 14, in <\/p>\n  \n  <p>IOError: [Errno 2] No such file or directory:\n  '\/tmp\/pip-GgS7fR-build\/setup.py'<\/p>\n<\/blockquote>\n\n<p>I am using pip, python 2.7 and Ubuntu 12.04 LTS vmware image<\/p>\n\n<p>Can anyone please help me to solve this error?<\/p>\n\n<p>full  file error :<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"33877605","CreationDate":"2015-11-23T18:06:06.163","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p>Try installing Ubuntu 14.04 in VMware and use the same command.<\/p>\n"}]}
{"QuestionId":33878321,"AnswerCount":1,"Tags":"<python><nlp><neural-network><theano><keras>","CreationDate":"2015-11-23T18:46:47.203","AcceptedAnswerId":null,"OwnerUserId":1337499.0,"Title":"Keras\/Theano - how to multiply by vector in Lambda layer","Body":"<p>I'm trying to implement a lambda layer, where the function performed by the lambda layer is performing , where  is an array with  rows and each entry of  is .<\/p>\n\n<p>The issue I'm running into is that I can't figure out how to create a static array with each entry of  set to . I know I can get  using the  function defined for theano tensors, but I'm not sure how to then create the vector.<\/p>\n","answers":[{"AnswerId":"39533985","CreationDate":"2016-09-16T14:28:09.727","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>I think you are over-complicating things. If you want to multiply with a matrix filled with constants, you can simply divide by a scalar which is then broadcasted over your input matrix, e.g.<\/p>\n\n<pre><code>Lambda(lambda x: x \/ M)\n<\/code><\/pre>\n\n<p>where <code>M<\/code> can be defined as<\/p>\n\n<pre><code>from keras import backend as K\nM = K.shape(x)[0]\n<\/code><\/pre>\n\n<p>giving<\/p>\n\n<pre><code>Lambda(lambda x: x \/ K.shape(M)[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":33879060,"AnswerCount":1,"Tags":"<vector><lua><torch>","CreationDate":"2015-11-23T19:29:44.893","AcceptedAnswerId":33889675.0,"OwnerUserId":4364020.0,"Title":"torch.Tensor manipulation - Comparing two vectors","Body":"<p>I've got two tensor objects repenting vectors:\n    (0110010),\n    (0111011)\nI would like to compare between the two and create a new tensor vector:\n    (0110010)\nIterating over them in a loop is very slow, I know there is a solution for this in Matlab so I assume there is one for tensors as well.<\/p>\n","answers":[{"AnswerId":"33889675","CreationDate":"2015-11-24T09:19:13.953","ParentId":null,"OwnerUserId":"387870","Title":null,"Body":"<p>To do a logical and operation for tensors containing only 1 and 0 elements you could use the :cmul() member function (element-wise multiplication).<\/p>\n\n<pre><code>th&gt; torch.Tensor({0,1,1,0,0,1,0}):cmul(torch.Tensor({0,1,1,1,0,1,1}))\n 0\n 1\n 1\n 0\n 0\n 1\n 0\n<\/code><\/pre>\n\n<p>To compare two tensors element-wise you can use :eq():<\/p>\n\n<pre><code>th&gt; torch.Tensor({0,1,1,0,0,1,0}):eq(torch.Tensor({0,1,1,1,0,1,1}))\n 1\n 1\n 1\n 0\n 1\n 1\n 0 \n<\/code><\/pre>\n"}]}
{"QuestionId":33879281,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><tensorflow>","CreationDate":"2015-11-23T19:41:37.347","AcceptedAnswerId":37682834.0,"OwnerUserId":310378.0,"Title":"How can I implement a custom RNN (specifically an ESN) in Tensorflow?","Body":"<p>I am trying to define my own RNNCell (Echo State Network) in Tensorflow, according to below definition.<\/p>\n\n<p>x(t + 1) = tanh(Win*u(t) + W*x(t) + Wfb*y(t))<\/p>\n\n<p>y(t) = Wout*z(t)<\/p>\n\n<p>z(t) = [x(t), u(t)]<\/p>\n\n<p>x is state, u is input, y is output. Win, W, and Wfb are not trainable. All weights are randomly initialized, but W is modified like this: \"Set a certain percentage of elements of W to 0, scale W to keep its spectral radius below 1.0<\/p>\n\n<p>I have this code to generate the equation.<\/p>\n\n<pre><\/pre>\n\n<p>My problem is two-fold. First I don't know how to implement this as a superclass of RNNCell. Second I don't know how to generate a W tensor according to above specification.<\/p>\n\n<p>Any help about any of these question is greatly appreciated. Maybe I can figure out a way to prepare W, but I sure as hell don't understand how to implement my own RNN as a superclass of RNNCell.<\/p>\n","answers":[{"AnswerId":"37682834","CreationDate":"2016-06-07T14:58:33.757","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>To give a quick summary:<\/p>\n\n<p>Look in the TensorFlow source code under <code>python\/ops\/rnn_cell.py<\/code> too see how to <strong>subclass<\/strong> RNNCell.  It's usually like this:<\/p>\n\n<pre><code>class MyRNNCell(RNNCell):\n  def __init__(...):\n\n  @property\n  def output_size(self):\n  ...\n\n  @property\n  def state_size(self):\n  ...\n\n  def __call__(self, input_, state, name=None):\n     ... your per-step iteration here ...\n<\/code><\/pre>\n"}]}
{"QuestionId":33879736,"AnswerCount":1,"Tags":"<neural-network><theano><backpropagation><lasagne>","CreationDate":"2015-11-23T20:08:34.023","AcceptedAnswerId":33889861.0,"OwnerUserId":279858.0,"Title":"Can I (selectively) invert Theano gradients during backpropagation?","Body":"<p>I'm keen to make use of the architecture proposed in the recent paper \"<a href=\"http:\/\/arxiv.org\/pdf\/1409.7495.pdf\" rel=\"noreferrer\">Unsupervised Domain Adaptation by Backpropagation<\/a>\" in the Lasagne\/Theano framework.<\/p>\n\n<p>The thing about this paper that makes it a bit unusual is that it incorporates a 'gradient reversal layer', which <em>inverts<\/em> the gradient during backpropagation:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/akoSm.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/akoSm.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>(The arrows along the bottom of the image are the backpropagations which have their gradient inverted).<\/p>\n\n<p>In the paper the authors claim that the approach \"can be implemented using any deep learning package\", and indeed they provide a <a href=\"https:\/\/github.com\/ddtm\/caffe\/tree\/grl\" rel=\"noreferrer\">version made in caffe<\/a>.<\/p>\n\n<p>However, I'm using the Lasagne\/Theano framework, for various reasons. <\/p>\n\n<p>Is it possible to create such a gradient reversal layer in Lasagne\/Theano? I haven't seen any examples of where one can apply custom scalar transforms to gradients like this. If so, can I do it by creating a custom layer in Lasagne? <\/p>\n","answers":[{"AnswerId":"33889861","CreationDate":"2015-11-24T09:28:20.320","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's a sketch implementation using plain Theano. This can be integrated into Lasagne easily enough.<\/p>\n\n<p>You need to create a custom operation which acts as an identity operation in the forward pass but reverses the gradient in the backward pass.<\/p>\n\n<p>Here's a suggestion for how that could be implemented. It is not tested and I'm not 100% certain I've understood everything correctly, but you may be able to verify and fix as required.<\/p>\n\n<pre><code>class ReverseGradient(theano.gof.Op):\n    view_map = {0: [0]}\n\n    __props__ = ('hp_lambda',)\n\n    def __init__(self, hp_lambda):\n        super(ReverseGradient, self).__init__()\n        self.hp_lambda = hp_lambda\n\n    def make_node(self, x):\n        return theano.gof.graph.Apply(self, [x], [x.type.make_variable()])\n\n    def perform(self, node, inputs, output_storage):\n        xin, = inputs\n        xout, = output_storage\n        xout[0] = xin\n\n    def grad(self, input, output_gradients):\n        return [-self.hp_lambda * output_gradients[0]]\n<\/code><\/pre>\n\n<p>Using the paper notation and naming conventions, here's a simple Theano implementation of the complete general model they propose.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef g_f(z, theta_f):\n    for w_f, b_f in theta_f:\n        z = tt.tanh(theano.dot(z, w_f) + b_f)\n    return z\n\n\ndef g_y(z, theta_y):\n    for w_y, b_y in theta_y[:-1]:\n        z = tt.tanh(theano.dot(z, w_y) + b_y)\n    w_y, b_y = theta_y[-1]\n    z = tt.nnet.softmax(theano.dot(z, w_y) + b_y)\n    return z\n\n\ndef g_d(z, theta_d):\n    for w_d, b_d in theta_d[:-1]:\n        z = tt.tanh(theano.dot(z, w_d) + b_d)\n    w_d, b_d = theta_d[-1]\n    z = tt.nnet.sigmoid(theano.dot(z, w_d) + b_d)\n    return z\n\n\ndef l_y(z, y):\n    return tt.nnet.categorical_crossentropy(z, y).mean()\n\n\ndef l_d(z, d):\n    return tt.nnet.binary_crossentropy(z, d).mean()\n\n\ndef mlp_parameters(input_size, layer_sizes):\n    parameters = []\n    previous_size = input_size\n    for layer_size in layer_sizes:\n        parameters.append((theano.shared(numpy.random.randn(previous_size, layer_size).astype(theano.config.floatX)),\n                           theano.shared(numpy.zeros(layer_size, dtype=theano.config.floatX))))\n        previous_size = layer_size\n    return parameters, previous_size\n\n\ndef compile(input_size, f_layer_sizes, y_layer_sizes, d_layer_sizes, hp_lambda, hp_mu):\n    r = ReverseGradient(hp_lambda)\n\n    theta_f, f_size = mlp_parameters(input_size, f_layer_sizes)\n    theta_y, _ = mlp_parameters(f_size, y_layer_sizes)\n    theta_d, _ = mlp_parameters(f_size, d_layer_sizes)\n\n    xs = tt.matrix('xs')\n    xs.tag.test_value = numpy.random.randn(9, input_size).astype(theano.config.floatX)\n    xt = tt.matrix('xt')\n    xt.tag.test_value = numpy.random.randn(10, input_size).astype(theano.config.floatX)\n    ys = tt.ivector('ys')\n    ys.tag.test_value = numpy.random.randint(y_layer_sizes[-1], size=9).astype(numpy.int32)\n\n    fs = g_f(xs, theta_f)\n    e = l_y(g_y(fs, theta_y), ys) + l_d(g_d(r(fs), theta_d), 0) + l_d(g_d(r(g_f(xt, theta_f)), theta_d), 1)\n\n    updates = [(p, p - hp_mu * theano.grad(e, p)) for theta in theta_f + theta_y + theta_d for p in theta]\n    train = theano.function([xs, xt, ys], outputs=e, updates=updates)\n\n    return train\n\n\ndef main():\n    theano.config.compute_test_value = 'raise'\n    numpy.random.seed(1)\n    compile(input_size=2, f_layer_sizes=[3, 4], y_layer_sizes=[7, 8], d_layer_sizes=[5, 6], hp_lambda=.5, hp_mu=.01)\n\n\nmain()\n<\/code><\/pre>\n\n<p>This is untested but the following may allow this custom op to be used as a Lasagne layer:<\/p>\n\n<pre><code>class ReverseGradientLayer(lasagne.layers.Layer):\n    def __init__(self, incoming, hp_lambda, **kwargs):\n        super(ReverseGradientLayer, self).__init__(incoming, **kwargs)\n        self.op = ReverseGradient(hp_lambda)\n\n    def get_output_for(self, input, **kwargs):\n        return self.op(input)\n<\/code><\/pre>\n"}]}
{"QuestionId":33879775,"AnswerCount":1,"Tags":"<python><matrix><theano>","CreationDate":"2015-11-23T20:11:10.447","AcceptedAnswerId":33887835.0,"OwnerUserId":4476617.0,"Title":"theano ~ use an index matrix and embeddings matrix to produce a 3D tensor?","Body":"<p>Let's say I have an index matrix:<\/p>\n\n<pre><\/pre>\n\n<p>And an embeddings matrix:<\/p>\n\n<pre><\/pre>\n\n<p>Each element in the  corresponds to a particualr row in the . For example,  in the  corresponds to  in the .<\/p>\n\n<p>Using theano syntax only, how can I build a 3D tensor, where each index in the  is replaced with a row or vector from the ?<\/p>\n\n<p>Edit: I was able to solve a similar problem: that of using an  and the  above to generate a 2D tensor. For that, I needed the following theano syntax:<\/p>\n\n<pre><\/pre>\n\n<p>However, I am a bit stuck on the 3D case.<\/p>\n","answers":[{"AnswerId":"33887835","CreationDate":"2015-11-24T07:35:52.320","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can simply index into the embedding matrix with the index matrix:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\nembeddings_matrix = theano.shared(numpy.array([[1, 2, 3], [2, 3, 4], [4, 5, 6]], dtype=theano.config.floatX))\nindex_matrix = tt.imatrix()\ny = embeddings_matrix[index_matrix]\nf = theano.function([index_matrix], y)\noutput = f(numpy.array([[0, 1], [2, 1]], dtype=numpy.int32))\nprint output.shape, '\\n', output\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>(2L, 2L, 3L) \n[[[ 1.  2.  3.]\n  [ 2.  3.  4.]]\n\n [[ 4.  5.  6.]\n  [ 2.  3.  4.]]]\n<\/code><\/pre>\n\n<p>The output is 3D with each element in the index matrix replaced with the corresponding embedding vector as required.<\/p>\n"}]}
{"QuestionId":33883036,"AnswerCount":1,"Tags":"<operation><tensorflow>","CreationDate":"2015-11-23T23:54:40.980","AcceptedAnswerId":null,"OwnerUserId":236247.0,"Title":"Tensorflow: Cannot import user_ops from my python file","Body":"<p>I have run the tutorials and created my own neural network implementation in tensorflow successfully.  I then decided to go one bit further an add my own op because I needed to do some of my own preprocessing on the data.  I followed the tutorial on the tensorflow site to add an op.  I successfully built tensorflow after writing my own c++ file.  Then, when I try to use it from my code, I get<\/p>\n\n<pre><\/pre>\n\n<p>My code does get reflected in bazel-genfiles\/tensorflow\/python\/ops\/gen_user_ops.py so the wrapper does get generated for it correctly.  It just looks like I can't see the    tensorflow\/python\/user_ops\/user_ops.py which is what imports that file.<\/p>\n\n<p>Now when I when I go through the testing of this module, I get the following odd behavior.  It should not pass because the expected vector I give it does not match what the result should be.  But maybe the test never gets executed despite saying passed?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33904715","CreationDate":"2015-11-24T22:02:34.760","ParentId":null,"OwnerUserId":"236247","Title":null,"Body":"<p>Hmmm.  Well, I uninstalled tensorflow and then I reinstalled from what I just built and what I wrote was suddenly recognized.  I have seen this behavior twice in a row now where an uninstall is necessary.  So to sum, the steps after adding my own op are:<\/p>\n\n<pre><code>$ pip uninstall tensorflow\n$ bazel build -c opt \/\/tensorflow\/tools\/pip_package:build_pip_package\n\n# To build with GPU support:\n$ bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n\n$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n\n# The name of the .whl file will depend on your platform.\n$ pip install \/tmp\/tensorflow_pkg\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":33883382,"AnswerCount":4,"Tags":"<python><machine-learning><computer-vision><tensorflow>","CreationDate":"2015-11-24T00:29:24.050","AcceptedAnswerId":33900275.0,"OwnerUserId":5507905.0,"Title":"TensorFlow - why doesn't this sofmax regression learn anything?","Body":"<p>I am aiming to do big things with TensorFlow, but I'm trying to start small. <\/p>\n\n<p>I have small greyscale squares (with a little noise) and I want to classify them according to their colour (e.g. 3 categories: black, grey, white). I wrote a little Python class to generate squares, and 1-hot vectors, and modified their basic MNIST example to feed them in.<\/p>\n\n<p>But it won't learn anything - e.g. for 3 categories it always guesses \u224833% correct.<\/p>\n\n<pre><\/pre>\n\n<p>My data generator looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33899251","CreationDate":"2015-11-24T16:44:14.030","ParentId":null,"OwnerUserId":"5580978","Title":null,"Body":"<p>You issue is that the your gradients are increasing\/decreasing without bounds, causing the loss function to become nan. <\/p>\n\n<p>Take a look at this question: <a href=\"https:\/\/stackoverflow.com\/questions\/33641799\/why-does-tensorflow-example-fail-when-increasing-batch-size?rq=1\">Why does TensorFlow example fail when increasing batch size?<\/a><\/p>\n\n<p>Furthermore, make sure that you run the model for a sufficient number of steps. You are only running it once through your train dataset (100 times * 100 examples), and this is not enough for it to converge. Increase it to something like 2000 at a minimum (running 20 times through your dataset).<\/p>\n\n<p>Edit (can't comment, so i'll add my thoughts here):\nThe point of the post i linked is that you can use <code>GradientDescentOptimizer<\/code>, as long as you make the learning rate something like 0.001. That's the issue, your learning rate was too high for the loss function you were using.<\/p>\n\n<p>Alternatively, use a different loss function, that doesn't increase\/decrease the gradients as much. Use <code>tf.reduce_mean<\/code> instead of <code>tf.reduce_sum<\/code> in the definition of <code>crossEntropy<\/code>.<\/p>\n"},{"AnswerId":"33883639","CreationDate":"2015-11-24T00:57:03.950","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>For starters, try initializing your W matrix with random values, not zeros - you're not giving the optimizer anything to work with when the output is all zeros for all inputs.<\/p>\n\n<p>Instead of:<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([28*28, num_categories]))\n<\/code><\/pre>\n\n<p>Try:<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([28*28, num_categories],\n                                    stddev=0.1))\n<\/code><\/pre>\n"},{"AnswerId":"33900275","CreationDate":"2015-11-24T17:38:02.260","ParentId":null,"OwnerUserId":"5507905","Title":null,"Body":"<p>While dga and syncd's responses were helpful, I tried using non-zero weight initialization and larger datasets but to no avail. The thing that finally worked was using a different optimization algorithm.<\/p>\n\n<p>I replaced:<\/p>\n\n<p><code>train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)<\/code><\/p>\n\n<p>with<\/p>\n\n<p><code>train_step = tf.train.AdamOptimizer(0.0005).minimize(cross_entropy)<\/code><\/p>\n\n<p>I also embedded the training for loop in another for loop to train for several epochs, resulting in convergence like this:<\/p>\n\n<pre><code> ===# EPOCH 0 #===\nError: 0.370000004768\n ===# EPOCH 1 #===\nError: 0.333999991417\n ===# EPOCH 2 #===\nError: 0.282000005245\n ===# EPOCH 3 #===\nError: 0.222000002861\n ===# EPOCH 4 #===\nError: 0.152000010014\n ===# EPOCH 5 #===\nError: 0.111999988556\n ===# EPOCH 6 #===\nError: 0.0680000185966\n ===# EPOCH 7 #===\nError: 0.0239999890327\n ===# EPOCH 8 #===\nError: 0.00999999046326\n ===# EPOCH 9 #===\nError: 0.00400000810623\n<\/code><\/pre>\n\n<p>EDIT - WHY IT WORKS: I suppose the problem was that I didn't manually choose a good learning rate schedule, and Adam was able to generate a better one automatically.<\/p>\n"},{"AnswerId":"39187146","CreationDate":"2016-08-28T01:05:56.787","ParentId":null,"OwnerUserId":"4992374","Title":null,"Body":"<p>Found this question when I was having a similar issue..I fixed mine by scaling the features. <\/p>\n\n<p>A little background: I was following the tensorflow tutorial, however I wanted to use the data from Kaggle(<a href=\"https:\/\/www.kaggle.com\/c\/digit-recognizer\/data\" rel=\"nofollow\">see data here<\/a>) to do the modeling, but in the beginning I kepted having the same issue: the model just doesn't learn..after rounds of trouble-shooting, I realized that the Kaggle data was on a completely different scale. Therefore, I scaled the data so that it shares the same scale(0,1) as the tensorflow's MNIST dataset.<\/p>\n\n<p>Just figured that I would add my two cents here..in case some beginners who are trying to follow the tutorial's settings get stuck like I did =)<\/p>\n"}]}
{"QuestionId":33884286,"AnswerCount":0,"Tags":"<debugging><pthreads><caffe><data-layer>","CreationDate":"2015-11-24T02:15:38.917","AcceptedAnswerId":null,"OwnerUserId":5522934.0,"Title":"Debug errors: pthread_join.c can't find","Body":"<p>Everybody,when I training the net,there comes an critical problem, then I debug the code and it shows:<\/p>\n\n<blockquote>\n  <p>0x00007ffff23c866b in pthread_join (threadid=140736858203904, \n      thread_return=0x0) at pthread_join.c:92\n      92  pthread_join.c: No such file or directory.<\/p>\n<\/blockquote>\n\n<p>This error cause the data_layer can't working normally. But there is an declaration in data_lay.cpp. Does anybody know how to resolve this problem?<\/p>\n","answers":[]}
{"QuestionId":33884978,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-24T03:32:25.727","AcceptedAnswerId":null,"OwnerUserId":5597685.0,"Title":"Build a graph that works with variable batch size using Tensorflow","Body":"<p>I use tf.placeholders() ops for feeding variable batch size of input which are 2D tensors and use the feed mechanism to supply different values for those tensors when I call run().\nI got <\/p>\n\n<blockquote>\n  <p>TypeError: 'Tensor' object is not iterable.<\/p>\n<\/blockquote>\n\n<p>Following is my code:<\/p>\n\n<pre><\/pre>\n\n<p>As I can not see the content of the tensor \"train_index_input\" without running the graph,  so the error of \"'Tensor' object is not iterable\" raises for the code:<\/p>\n\n<pre><\/pre>\n\n<p>What I want to obtain is an embedding matrix \"embedding_input\" that is of shape [batch_size, embedding_dimension] where  is not fixed. Do I have to define a new operation in Tensorflow for embedding lookup for 2D tensors? Or any other ways to do it?Thanks <\/p>\n","answers":[{"AnswerId":"33907688","CreationDate":"2015-11-25T02:47:34.867","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You're trying to do a python-level list comprehension (<code>for x in train_index_input<\/code>) over a Tensorflow placeholder.  That won't work - Python has no idea what's inside a tf object.<\/p>\n\n<p>To accomplish a batch embedding lookup, what you can do is just flatten your batch:<\/p>\n\n<pre><code>train_indexes_flat = tf.reshape(train_index_input, [-1])\n<\/code><\/pre>\n\n<p>run it through embedding lookup:<\/p>\n\n<pre><code>looked_up_embeddings = tf.nn.embedding_lookup(train_embeddings, train_indexes_flat)\n<\/code><\/pre>\n\n<p>and then reshape it back into the right groups:<\/p>\n\n<pre><code>embedding_input = tf.reshape(looked_up_embeddings, [-1, window_size])\n<\/code><\/pre>\n"}]}
{"QuestionId":33885276,"AnswerCount":2,"Tags":"<build><swig><tensorflow><bazel>","CreationDate":"2015-11-24T04:07:41.277","AcceptedAnswerId":null,"OwnerUserId":150851.0,"Title":"Bazel: Use swig from non-standard location","Body":"<p>Because I don't have root access on my machine, I have built and installed swig from source in a non-standard directory () and I want bazel to use it. So when I try to build tensorflow, I get the error that swig cannot be found:<\/p>\n\n<pre><\/pre>\n\n<p>Checking the  script mentioned in the error it's just a script that does:<\/p>\n\n<pre><\/pre>\n\n<p>Printing the  from this script shows it's:<\/p>\n\n<pre><\/pre>\n\n<p>I can't append the $PATH variable in the swig script because it gets remade on build. So how do I tell bazel to use my non-standard swig location?<\/p>\n","answers":[{"AnswerId":"33885655","CreationDate":"2015-11-24T04:46:55.813","ParentId":null,"OwnerUserId":"148870","Title":null,"Body":"<p>That isn't Bazel-specific, it's part of the tensorflow code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/swig\/swig.sh\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/swig\/swig.sh<\/a><\/p>\n\n<p>You can just modify that file before you build.<\/p>\n\n<p>The specific build rule that uses that file is also defined by tensorflow, since it's not a rule included with the core Bazel distribution:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tensorflow.bzl#L275\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tensorflow.bzl#L275<\/a><\/p>\n"},{"AnswerId":"34352707","CreationDate":"2015-12-18T09:50:37.217","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>This problem looks like it is due to <a href=\"http:\/\/bazel.io\/docs\/bazel-user-manual.html#sandboxing\" rel=\"nofollow\">Bazel Sandboxing<\/a>.<\/p>\n\n<p>If so, then using the following Bazel options should workaround the problem, as it turns sandboxing off:<\/p>\n\n<pre><code>--spawn_strategy=standalone --genrule_strategy=standalone\n<\/code><\/pre>\n\n<p>I have not tested this, though.<\/p>\n\n<p>This <a href=\"http:\/\/bazel.io\/blog\/2015\/09\/11\/sandboxing.html\" rel=\"nofollow\">blog post<\/a> explains:<\/p>\n\n<blockquote>\n  <p>Sandboxing is the technique of restricting the access rights of a\n  process. In the context of Bazel, we're mostly concerned with\n  restricting file system access. More specifically, Bazel's file system\n  sandbox contains only known inputs, such that compilers and other\n  tools can't even see files they should not access.<\/p>\n<\/blockquote>\n\n<p>How it works on Linux:<\/p>\n\n<blockquote>\n  <p>On Linux, we're using user namespaces, which are available in Linux\n  3.8 and later. Specifically, we create a new mount namespace. We create a temporary directory into which we mount all the files that\n  the subprocess is allowed to see. We then use pivot_root to make the\n  temporary directory appear as the root directory for all subprocesses.<\/p>\n  \n  <p>We also mount \/proc, \/dev\/null, \/dev\/zero, and a temporary filesystem\n  (tmpfs) on \/tmp. We mount \/dev\/random and \/dev\/urandom, but recommend\n  against their usage, as it can lead to non-reproducible builds.<\/p>\n  \n  <p>We currently also mount \/bin, \/etc, \/usr (except \/usr\/local), and\n  every directory starting with \/lib, to allow running local tools. In\n  the future, we are planning to provide a shell with a set of Linux\n  utilities, and to require that all other tools are specified as\n  inputs.<\/p>\n<\/blockquote>\n\n<p><a href=\"http:\/\/bazel.io\/roadmap.html\" rel=\"nofollow\">Bazel's roadmap<\/a> says that sandboxing on MacOSX is scheduled for April 2016.<\/p>\n\n<p>Sandboxing is supposed to be a good thing, since it guarantees that build rules are consistent (as any dependency that is not made explicit will bomb).  So perhaps instead of turning Sandboxing off, better solutions would be:<\/p>\n\n<ul>\n<li>fixing TensorFlow's build rules to make the dependency to swig explicit.<\/li>\n<li>or making swig available in one of the directories above (eg. <code>\/bin<\/code>).<\/li>\n<li>or explicitly configuring Bazel to mount the directory that swig lives in (if may be possible with <code>-M<\/code> and <code>-m<\/code>, I have not tested though).<\/li>\n<\/ul>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":33886036,"AnswerCount":0,"Tags":"<neural-network><deep-learning><tensorflow><recurrent-neural-network>","CreationDate":"2015-11-24T05:25:12.873","AcceptedAnswerId":null,"OwnerUserId":1782011.0,"Title":"Use output from last or both layer in a 2 Layer GRU RNN classifier?","Body":"<p>I'm experimenting with tensorflow. I'm trying to build a classification system based on a sequence as input. I made a network existing of 2 layers of regular RNN and 1 layer of GRU cells (MultiRNNCell in tensorflow) Then I intended to couple a softmax output to the hidden state of my last GRU cell. Now I noticed that when I take the last state from the MultiRNNCell output. It's 3x the size as the hidden layer I specified for my GRU cell. I'm guessing that this is the state of all 3 RNN layers concatenated. Is this correct? If so, does it make sense to couple the state of all the layers to the softmax output? Or should I only do it on the state of the last GRU cell.<\/p>\n\n<p>I tested both cases on some small toy dataset and the algorithm seems to learn in both cases but I'm not sure what makes most sense in a real situation.<\/p>\n\n<p>Some code how I specified the RNN model<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33886558,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-11-24T06:08:41.950","AcceptedAnswerId":null,"OwnerUserId":5307788.0,"Title":"Numpy installation error. (Mingw32CCompiler instance has no attribute 'compile_options')","Body":"<p>I use python with windows8 \/ anaconda spyder (2.7)\nI'm trying to update Theano up-to-date. When I installing the theano by <\/p>\n\n<pre><\/pre>\n\n<p>The error happens at numpy installation part. <\/p>\n\n<p>Though numpy is available to use in anaconda GUI, But is not shown in anaconda terminal. (pip show numpy outputs nothing)<\/p>\n\n<p>The error is like below (when I try to install numpy, same error happens)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34696614","CreationDate":"2016-01-09T17:26:09.593","ParentId":null,"OwnerUserId":"614241","Title":null,"Body":"<p>This is most likely a numpy issue. Unfortunately I could not upgrade theano even after the fix I will suggest, but following a step by step guide like <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">this one<\/a> might work.<\/p>\n\n<p>This specific problem occurs because <code>numpy.distutils.misc_util.mingw32()<\/code> returns False while it shouldn't. See also <a href=\"https:\/\/github.com\/numpy\/numpy\/pull\/6978\" rel=\"nofollow\">this discussion<\/a> and <a href=\"https:\/\/github.com\/numpy\/numpy\/blob\/master\/numpy\/distutils\/misc_util.py#L381\" rel=\"nofollow\">the source\n<\/a>. I decided to set manually the system variable <code>MSYSTEM<\/code> to <code>MINGW32<\/code> which was not defined in my system and which I find unlikely to break any other behavior. <\/p>\n\n<p>As I said this does not guarantee that the upgrade will work nor that there will be no side effects, but it will solve the issue mentioned here. <\/p>\n"}]}
{"QuestionId":33889366,"AnswerCount":1,"Tags":"<python><neural-network><convolution><tensorflow><conv-neural-network>","CreationDate":"2015-11-24T09:04:15.990","AcceptedAnswerId":33901141.0,"OwnerUserId":2849678.0,"Title":"A reusable Tensorflow convolutional Network","Body":"<p>I want to reuse code from the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/index.md\" rel=\"nofollow\">Tensorflow \"MNIST for Pros\" CNN example<\/a>.\nMy images are 388px X 191px, with only 2 output classes. The original code can be <a href=\"http:\/\/martin-thoma.com\/tensor-flow-quick\/\" rel=\"nofollow\">found here<\/a>.\nI tried to reuse this code by changing the <strong><em>input &amp; output layers ONLY<\/em><\/strong>, as shown below:<\/p>\n\n<h1>input layer<\/h1>\n\n<pre><\/pre>\n\n<h1>output layer<\/h1>\n\n<pre><\/pre>\n\n<p><strong>Running the modified code gives a vague stacktrace:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33901141","CreationDate":"2015-11-24T18:27:51.763","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<pre><code>tensorflow.python.framework.errors.InvalidArgumentError: Input has 14005248 values, which isn't divisible by 3136\n [[Node: Reshape_4 = Reshape[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](MaxPool_5, Reshape_4\/shape)]]\n<\/code><\/pre>\n\n<p>But the way you executed it prevents you from seeing the actual line causing the problem.  Save it to a file and python <code>&lt;file&gt;<\/code> it.<\/p>\n\n<pre><code>  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n<\/code><\/pre>\n\n<p>But the answer is that you haven't changed the size of your convolutional and pooling layers, so when you used to run 28x28 images through, they eventually shrunk down to a 7x7x(convolutional_depth) layer.  Now you're running huge images through, so after the first convolutional layer and the 2x2 maxpool, you've got a VERY BIG thing you're trying to feed in, but you're reshaping to:<\/p>\n\n<pre><code>h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)\nh_pool2 = max_pool_2x2(h_conv2)\n\nW_fc1 = weight_variable([7 * 7 * 64, 1024])\nb_fc1 = bias_variable([1024])\n<\/code><\/pre>\n\n<p>The output of h_pool2 is much larger with your larger images.  You need to shrink them down more - likely with more convolutional and maxpooling layers.  You could also try increasing the size of W_fc1 to match the input size that's getting there.  It's running through two 2x2 maxpools - each shrinks the size by 2 in the x and y dimensions.  28x28x1 --> 14x14x32 --> 7x7x64.  So YOUR images are going from 388 x 191 --> 194 x 95 --> 97 x 47<\/p>\n\n<p>As a warning, a fully connected layer with 97*47 = 4559 inputs is going to be glacially slow.<\/p>\n"}]}
{"QuestionId":33900842,"AnswerCount":2,"Tags":"<c++><tensorflow><bazel>","CreationDate":"2015-11-24T18:10:24.847","AcceptedAnswerId":null,"OwnerUserId":3441991.0,"Title":"tensorflow and bazel \"undefined reference to\" compilation error","Body":"<p>I am writing a C++ code in tensorflow framework and I want to use a dynamic library written using makefile.\nIn source code I put the path to header file:<\/p>\n\n<pre><\/pre>\n\n<p>to use a function called fix_client(int argc, char **argv)\nand in the BUILD file I put the path to the dynamic library, called libtrading.so:<\/p>\n\n<pre><\/pre>\n\n<p>bazel-build retrieves me an error of:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33914359","CreationDate":"2015-11-25T10:41:36.063","ParentId":null,"OwnerUserId":"3441991","Title":null,"Body":"<p>I found the problem.\nThe linking of library in bazel is correct: the problem was that libtrading is a C library and tensorflow is built in a c++ environment: all the functions linked need to be included in the following guards:<\/p>\n\n<pre><code>#ifdef __cplusplus\nextern \"C\" {\n#endif\n...\n#ifdef __cplusplus\n}\n#endif\n<\/code><\/pre>\n"},{"AnswerId":"33912478","CreationDate":"2015-11-25T09:13:43.047","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p>Bazel requires all dependencies to be declared, so the TensorFlow library should be in your deps attribute. It looks like it is not the case in your target (especially the flag for tensorflow includes is out of place).<\/p>\n\n<p>After a quick look at TensorFlow build file I would say it needs the following deps attribute:<\/p>\n\n<pre><code> deps = [\n     \"\/\/tensorflow\/cc:cc_ops\",\n     \"\/\/tensorflow\/core:kernels\",\n     \"\/\/tensorflow\/core:tensorflow\",\n ],\n<\/code><\/pre>\n\n<p>But I am really unfamiliar with TensorFlow itself.<\/p>\n\n<p>What's the deps attribute of your cc_binary?<\/p>\n"}]}
{"QuestionId":33901226,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2015-11-24T18:32:34.300","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"How do I use TensorFlow to add Predicted Value to an empty column in a CSV file?","Body":"<p>So I have this great bit of code that comes out with approximately a 93% accuracy rate on its predictions. What I'm wondering how to do now is to take the trained program, make it look at actual test data without the answer on it, and make it fill in the answer regardless of the accuracy. Here's the code that I have that predicts with a ~93% accuracy rate.<\/p>\n\n<pre><\/pre>\n\n<p>Now I have the actual test set  where the first column is entirely empty and I need to fill it in with a predicted 1 or 0. How do I go about doing that?<\/p>\n","answers":[{"AnswerId":"33903190","CreationDate":"2015-11-24T20:28:39.250","ParentId":null,"OwnerUserId":"904032","Title":null,"Body":"<p>The program above doesn't appear to be saving the trained session. I think you want to do this in two steps.<\/p>\n\n<ol>\n<li>Train and save the session<\/li>\n<li>Restore the save session, and run test data through it.<\/li>\n<\/ol>\n\n<p><strong>Step 1:<\/strong> <\/p>\n\n<pre><code> #!\/usr\/bin\/env python\n\n import tensorflow as tf\n import numpy as np\n from numpy import genfromtxt\n import sklearn\n\n # Convert to one hot\n def convertOneHot(data):\n     y=np.array([int(i[0]) for i in data])\n     y_onehot=[0]*len(y)\n     for i,j in enumerate(y):\n         y_onehot[i]=[0]*(y.max() + 1)\n         y_onehot[i][j]=1\n     return (y,y_onehot)\n\n # Build Example Data is CSV format, but use Iris data\n from sklearn import datasets\n from sklearn.model_selection import train_test_split\n def buildDataFromIris():\n     iris = datasets.load_iris()\n     X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.33, random_state=42)\n     f=open('cs-training.csv','w')\n     for i,j in enumerate(X_train):\n         k=np.append(np.array(y_train[i]),j   )\n         f.write(\",\".join([str(s) for s in k]) + '\\n')\n     f.close()\n     f=open('cs-test.csv','w')\n     for i,j in enumerate(X_test):\n         k=np.append(np.array(y_test[i]),j   )\n         f.write(\",\".join([str(s) for s in k]) + '\\n')\n     f.close()\n\n # Recreate logging and save dir\n # Seems the tensorflow won't always overwrite\n import shutil, os, sys\n TMPDir='.\/tensorTMP'\n try:\n  shutil.rmtree(TMPDir)\n except:\n  print \"Tmp Dir did not exist...that's okay\"\n os.mkdir(TMPDir, 0755 )\n\n\n\n # Populate the data\n buildDataFromIris()\n\n data = genfromtxt('cs-training.csv',delimiter=',')  # Training data\n test_data = genfromtxt('cs-test.csv',delimiter=',')  # Test data\n\n x_train=np.array([ i[1::] for i in data])\n y_train,y_train_onehot = convertOneHot(data)\n\n x_test=np.array([ i[1::] for i in test_data])\n y_test,y_test_onehot = convertOneHot(test_data)\n\n A=data.shape[1]-1 # Number of features, Note first is y\n B=len(y_train_onehot[0])\n tf_in = tf.placeholder(\"float\", [None, A]) # Features\n tf_weight = tf.Variable(tf.zeros([A,B]))\n tf_bias = tf.Variable(tf.zeros([B]))\n tf_softmax = tf.nn.softmax(tf.matmul(tf_in,tf_weight) + tf_bias)\n\n # Training via backpropagation\n tf_softmax_correct = tf.placeholder(\"float\", [None,B])\n tf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax))\n\n # Train using tf.train.GradientDescentOptimizer\n tf_train_step = tf.train.GradientDescentOptimizer(0.01).minimize(tf_cross_entropy)\n\n # Add accuracy checking nodes\n tf_correct_prediction = tf.equal(tf.argmax(tf_softmax,1), tf.argmax(tf_softmax_correct,1))\n tf_accuracy = tf.reduce_mean(tf.cast(tf_correct_prediction, \"float\"))\n\n saver = tf.train.Saver([tf_weight,tf_bias])\n\n # Initialize and run\n init = tf.initialize_all_variables()\n sess = tf.Session()\n sess.run(init)\n\n THRESHOLD = 0.98\n saved = False\n print(\"...\")\n # Run the training\n for i in range(100):\n     sess.run(tf_train_step, feed_dict={tf_in: x_train, tf_softmax_correct: y_train_onehot})\n     result = sess.run(tf_accuracy, feed_dict={tf_in: x_test, tf_softmax_correct: y_test_onehot})\n     # If it's well trained on this iteration, save it. We just need one save.\n     if result &gt; THRESHOLD  and saved == False:\n         saved = True\n         print \"saving result {}\".format(result)\n         saver.save(sess,TMPDir +\"\/savedSess\")\n<\/code><\/pre>\n\n<p>The only modifications made were generating sample data using Iris, establishing a THRESHOLD or confidence interval for the session. If it's over that THRESHOLD, then, save the session.  After running step one, the model should be trained and saved.<\/p>\n\n<p><strong>Step 2:<\/strong> <\/p>\n\n<p>Restore the saved session, and run the training data through it.<\/p>\n\n<pre><code> #!\/usr\/bin\/env python\n\n import tensorflow as tf\n import numpy as np\n from numpy import genfromtxt\n import sklearn\n\n # Convert to one hot\n def convertOneHot(data):\n     y=np.array([int(i[0]) for i in data])\n     y_onehot=[0]*len(y)\n     for i,j in enumerate(y):\n         y_onehot[i]=[0]*(y.max() + 1)\n         y_onehot[i][j]=1\n     return (y,y_onehot)\n\n\n data = genfromtxt('cs-training.csv',delimiter=',')  # Training data\n test_data = genfromtxt('cs-test.csv',delimiter=',')  # Test data\n\n x_train=np.array([ i[1::] for i in data])\n y_train,y_train_onehot = convertOneHot(data)\n\n x_test=np.array([ i[1::] for i in test_data])\n y_test,y_test_onehot = convertOneHot(test_data)\n\n A=data.shape[1]-1 # Number of features, Note first is y\n B=len(y_train_onehot[0])\n tf_in = tf.placeholder(\"float\", [None, A]) # Features\n tf_weight = tf.Variable(tf.zeros([A,B]))\n tf_bias = tf.Variable(tf.zeros([B]))\n tf_softmax = tf.nn.softmax(tf.matmul(tf_in,tf_weight) + tf_bias)\n\n # Training via backpropagation\n tf_softmax_correct = tf.placeholder(\"float\", [None,B])\n tf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax))\n\n # Train using tf.train.GradientDescentOptimizer\n tf_train_step = tf.train.GradientDescentOptimizer(0.01).minimize(tf_cross_entropy)\n\n # Add accuracy checking nodes\n tf_correct_prediction = tf.equal(tf.argmax(tf_softmax,1), tf.argmax(tf_softmax_correct,1))\n tf_accuracy = tf.reduce_mean(tf.cast(tf_correct_prediction, \"float\"))\n\n saver = tf.train.Saver([tf_weight,tf_bias])\n\n # Initialize and run\n init = tf.initialize_all_variables()\n sess = tf.Session()\n sess.run(init)\n\n TMPDir='.\/tensorTMP'\n saver.restore(sess, TMPDir + '\/savedSess')\n ans = sess.run(tf_softmax, feed_dict={tf_in: x_test, tf_softmax_correct: y_test_onehot})\n\n print ans\n<\/code><\/pre>\n\n<p>Note, your output will look like the following...<\/p>\n\n<pre><code>[[  6.17585704e-02   8.63590300e-01   7.46511072e-02]\n[  9.98804331e-01   1.19561062e-03   3.25832108e-13]\n[  1.52018686e-07   4.49650863e-04   9.99550164e-01]\n<\/code><\/pre>\n"}]}
{"QuestionId":33901908,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-24T19:11:29.977","AcceptedAnswerId":33903865.0,"OwnerUserId":395857.0,"Title":"How to get name of a shared variable in Theano?","Body":"<p>How to get name of a shared variable in Theano? I see that printing it or using  works, but is the correct way to do so? Is there any function similar to  (e.g. )?<\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33903865","CreationDate":"2015-11-24T21:10:14.767","ParentId":null,"OwnerUserId":"1828879","Title":null,"Body":"<p>You get the name of a shared variable <code>x<\/code> by getting <code>x.name<\/code>. Unfortunately this is <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/shared.html\" rel=\"nofollow\">not documented<\/a>.<\/p>\n"}]}
{"QuestionId":33903569,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-24T20:52:23.870","AcceptedAnswerId":33904534.0,"OwnerUserId":1259390.0,"Title":"How to filter tensor from queue based on some predicate in tensorflow?","Body":"<p>How can I filter data stored in a queue using a predicate function? For example, let's say we have a queue that stores tensors of features and labels and we just need those that meet the predicate. I tried the following implementation without success:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33904534","CreationDate":"2015-11-24T21:50:17.813","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The most straightforward way to do this is to dequeue a batch, run them through the predicate test, use <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops.html#where\" rel=\"nofollow noreferrer\"><code>tf.where<\/code><\/a> to produce a dense vector of the ones that match the predicate, and use <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/array_ops.html#gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a> to collect the results, and enqueue that batch.  If you want that to happen automatically, you can start a queue runner on the second queue - the easiest way to do that is to use <code>tf.train.batch<\/code>:<\/p>\n\n<p>Example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\na = tf.constant(np.array([5, 1, 9, 4, 7, 0], dtype=np.int32))\n\nq = tf.FIFOQueue(6, dtypes=[tf.int32], shapes=[])\nenqueue = q.enqueue_many([a])\ndequeue = q.dequeue_many(6)\npredmatch = tf.less(dequeue, [5])\nselected_items = tf.reshape(tf.where(predmatch), [-1])\nfound = tf.gather(dequeue, selected_items)\n\nsecondqueue = tf.FIFOQueue(6, dtypes=[tf.int32], shapes=[])\nenqueue2 = secondqueue.enqueue_many([found])\ndequeue2 = secondqueue.dequeue_many(3) # XXX, hardcoded\n\nwith tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  sess.run(enqueue)  # Fill the first queue\n  sess.run(enqueue2) # Filter, push into queue 2\n  print sess.run(dequeue2) # Pop items off of queue2\n<\/code><\/pre>\n\n<p>The predicate produces a boolean vector;  the <code>tf.where<\/code> produces a dense vector of the indexes of the true values, and the <code>tf.gather<\/code> collects items from your original tensor based upon those indexes.<\/p>\n\n<p>A lot of things are hardcoded in this example that you'd need to make not-hardcoded in reality, of course, but hopefully it shows the structure of what you're trying to do (create a filtering pipeline).  In practice, you'd want QueueRunners on there to keep things churning automatically.  Using <code>tf.train.batch<\/code> is very useful to handle that automatically -- see <a href=\"http:\/\/www.tensorflow.org\/how_tos\/threading_and_queues\/index.html\" rel=\"nofollow noreferrer\">Threading and Queues<\/a> for more detail.<\/p>\n"}]}
{"QuestionId":33904232,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-11-24T21:33:35.580","AcceptedAnswerId":33910396.0,"OwnerUserId":2869180.0,"Title":"Theano Installation On windows 64","Body":"<p>Im new in Python and Theano library. I want to install Theano on windows 7-64. I have a display adapters :\nIntel(R) HD Graphics 3000 which is not compatible with NVIDA.<\/p>\n\n<p>My QUESTIONS: <\/p>\n\n<p>1-Is obligatory to install CUDA to i can use Theano?\n2- Even if i have an Ubuntu opearting system, with the same display adapters, CUDA still mandatory?<\/p>\n\n<p>Any help!<\/p>\n\n<p>Thanks <\/p>\n","answers":[{"AnswerId":"33910396","CreationDate":"2015-11-25T07:05:43.690","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You do not need CUDA to run Theano.<\/p>\n\n<p>Theano can run on either CPU or GPU. If you want to run on GPU you must (currently) use CUDA which means you must be using a NVIDIA display adapter. Without CUDA\/NVIDIA you must run on CPU.<\/p>\n\n<p>There is no disadvantage to running on CPU other than speed -- Theano can be much faster on GPU but everything that runs on a GPU will also run on a CPU as long as it has been coded generically (the default and standard method for Theano code).<\/p>\n"}]}
{"QuestionId":33905326,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-11-24T22:45:13.067","AcceptedAnswerId":34387104.0,"OwnerUserId":2386836.0,"Title":"Caffe training without testing","Body":"<p>I am using <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"noreferrer\">Caffe<\/a> to train AlexNet on a known image database. I am benchmarking and want to exclude a testing phase.<\/p>\n\n<p>Here is the  for AlexNet:<\/p>\n\n<pre><\/pre>\n\n<p>While I have never found a definitive doc that detailed all of the prototxt options, comments within Caffe tutorials indicate this \"\" represents the number of iterations after which we test the trained network. <\/p>\n\n<p>I figured that I might set it to zero to turn off testing. Nope.<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>So I set the  to one million, but still of course, Caffe tests the network at iteration zero.<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>How do I turn testing off while training?<\/p>\n","answers":[{"AnswerId":"33909593","CreationDate":"2015-11-25T06:12:12.227","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You have a flag for that too. Add<\/p>\n\n<pre><code>test_initialization: false\n<\/code><\/pre>\n\n<p>To your <code>'solver.prototxt'<\/code> and you are done ;)<\/p>\n"},{"AnswerId":"34387104","CreationDate":"2015-12-20T23:32:23.883","ParentId":null,"OwnerUserId":"2386836","Title":null,"Body":"<p>Caffe's documentation is somewhat scant on details. What I was finally told is this counterintuitive solution:<\/p>\n\n<p>In your solver.prototxt, take the lines for <code>test_iter<\/code> and <code>test_interval<\/code><\/p>\n\n<pre><code>test_iter: 1000\ntest_interval: 1000\n<\/code><\/pre>\n\n<p>and simply omit them. If you'd like to prevent the test at the beginning, you would add a line as <a href=\"https:\/\/stackoverflow.com\/a\/33909593\/1714410\">@shai suggested<\/a>:<\/p>\n\n<pre><code>test_initialization: false\n<\/code><\/pre>\n"}]}
{"QuestionId":33910638,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-25T07:21:51.040","AcceptedAnswerId":33948945.0,"OwnerUserId":4383691.0,"Title":"Tensorflow translate.py import error: No module named translate","Body":"<p>I'm trying to run Tensorflow's translate.py from a python console rather than through bazel -build, but I get an error at these two lines:<\/p>\n\n<pre><\/pre>\n\n<p>I've checked the folder to see that the \"<strong>init<\/strong>.py\" file is there, but python seems to think there is no such module as translate. \nHow can i fix this?<\/p>\n","answers":[{"AnswerId":"33948945","CreationDate":"2015-11-27T00:54:49.230","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>The best way to do this is to navigate to folder containing the translate module and running it. You can also download the translate module to any other place and run it. However, don't forget to change the above lines to:<\/p>\n\n<pre><code>from translate import data_utils\nfrom translate import seq2seq_model\n<\/code><\/pre>\n"},{"AnswerId":"39994992","CreationDate":"2016-10-12T09:21:37.913","ParentId":null,"OwnerUserId":"4129739","Title":null,"Body":"<p>I resolved this issue by removing all the <code>from tensorflow.models.rnn.translate<\/code> statements, leaving just <\/p>\n\n<pre><code>import data_utils\nimport seq2seq_model\n<\/code><\/pre>\n\n<p>in <code>translate.py<\/code> and <\/p>\n\n<pre><code>import data_utils\n<\/code><\/pre>\n\n<p>in <code>seq2seq_model.py<\/code>.<\/p>\n"}]}
{"QuestionId":33912342,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-25T09:08:22.503","AcceptedAnswerId":null,"OwnerUserId":5602977.0,"Title":"TensorFlow InvalidArgumentError when reading record for training on own image set","Body":"<p>i'am trying to train my own images (682x1024x3= 2095104 Pixels)  with TensorFlow. Therefore i combined several published scripts to 1) create a .tfrcords file using the TFRecord writer 2) to read this record and 3) to train them.\nTo read the images i took the function <em>read_images<\/em>() from here: <a href=\"https:\/\/github.com\/HamedMP\/ImageFlow\/blob\/master\/ImageFlow.py\" rel=\"nofollow\">https:\/\/github.com\/HamedMP\/ImageFlow\/blob\/master\/ImageFlow.py<\/a>\nand change it into these code<\/p>\n\n<pre><\/pre>\n\n<p>the result of <em>read_images<\/em>() i pass  with list of labels and filename to the <em>convert_to<\/em>() function published here: <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/convert_to_records.py\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/convert_to_records.py<\/a><\/p>\n\n<p>my changes in script where<\/p>\n\n<pre><\/pre>\n\n<p>and the for loop i am cutting the labelId out of the string <\/p>\n\n<pre><\/pre>\n\n<p>i did no change inside the <em>_int64_feature<\/em>() an <em>_bytes_feature<\/em>()\nDebuger say at that poin that my values of variables are: <\/p>\n\n<ul>\n<li>rows=682  cols=1024  depth=3\nand <em>len(image_raw)<\/em> gives me 2095104 this equal to pixels<\/li>\n<\/ul>\n\n<p>so thats the way i write the record.\nto read the record i'am using this script:<\/p>\n\n<p>\/master\/tensorflow\/g3doc\/tutorials\/mnist\/mnist.py<\/p>\n\n<p>and this one:<\/p>\n\n<p>\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/fully_connected_reader.py\n(sry just 2 links allowed)<\/p>\n\n<p>i changed the vlaue 1000 into capacity=10+ 3 * batchsize and min_after_deque=10 inside <em>inputs<\/em>() function (second last line):<\/p>\n\n<pre><\/pre>\n\n<p>Now i don't (see)have any issue one creating a TFRecord but when i run the whole script i'am getting the foliwing error:<\/p>\n\n<pre><\/pre>\n\n<p>which say i Expected <strong>[2095104]=682x1024x3<\/strong>, got <strong>[2101248]=684x1024x3<\/strong><\/p>\n\n<p>so i don't have no idea where thes 2 extra Pixels come from. can somebody help?<\/p>\n\n<p>the error occures in the <em>inputs<\/em>() function on line (117 published code)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47367578","CreationDate":"2017-11-18T15:14:35.537","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Moving Alexander's comment-answer to a real answer:<\/p>\n\n<p>1.) resize image into 1024x768 <\/p>\n\n<p>2.) resize it again into IMAGE_HIGHT = 192 IMAGE_WIDTH = 256. (how you see it's just divided by 4.)<\/p>\n\n<p>OP used this code snippet. <\/p>\n\n<pre><code>im = Image.open(source_folders[i]).resize((1024,768)) \nim = im.resize((256,192),Image.ANTIALIAS)\n<\/code><\/pre>\n"}]}
{"QuestionId":33916981,"AnswerCount":1,"Tags":"<python><tensorflow><sparse-matrix><softmax>","CreationDate":"2015-11-25T12:50:23.873","AcceptedAnswerId":null,"OwnerUserId":5604212.0,"Title":"Using Sparse Tensors to feed a placeholder for a softmax layer in TensorFlow","Body":"<p>Has anyone tried using Sparse Tensors for Text Analysis with TensorFlow with success? Everything is ready and I manage to feed  in  for a Softmax layer with numpy arrays, but I am unable to feed the dictionary with SparseTensorValues. <\/p>\n\n<p>I have not found either documentation about using sparse matrices to train a model ( softmax for example ) with Tensor Flow, which is strange, as classes  and  or  methods are ready for it, but there is no documentation about how to feed the  dictionary of values in the  method. <\/p>\n\n<p>Thanks a lot,<\/p>\n","answers":[{"AnswerId":"43197063","CreationDate":"2017-04-04T01:56:36.113","ParentId":null,"OwnerUserId":"2467383","Title":null,"Body":"<p>I have found a way of putting sparse images into tensorflow including batch processing if that is of any help. <\/p>\n\n<p>I create a 4-d sparse matrix in a dictionary where the dimensions are batchSize, xLen, ylen, zLen (where zLen is 3 for colour for example). The following pseudo code is for a batch of 50 32x96 pixel 3-color images. Values are the intensity of each pixel. In the snippet below I show the first 2 pixels of the first batch being initialised...<\/p>\n\n<pre><code>shape = [50, 32, 96, 3]\nindices = [[0, 20, 31, 0],[0, 22, 33, 1], etc...]\nvalues = [12, 24, etc...]\nbatch = {\"indices\": indices, \"values\": values, \"shape\": shape}\n<\/code><\/pre>\n\n<p>When setting up the computational graph I create a sparse-placeholder of the correct dimensions<\/p>\n\n<pre><code>images = tf.sparse_placeholder(tf.float32, shape=[None, 32, 96, 3])\n<\/code><\/pre>\n\n<p>'None' is used so I can vary the batch size.<\/p>\n\n<p>When I first want to use the images, e.g. to feed into a batch convolution, I convert them back to a dense tensor:<\/p>\n\n<pre><code>images = tf.sparse_tensor_to_dense(batch) \n<\/code><\/pre>\n\n<p>Then when I am ready to run a session, e.g. for training, I pass the 3 components of the batch into the dictionary so that they will be picked up by the sparse_placeholder:<\/p>\n\n<pre><code>train_dict = {images: (batch['indices'], batch['values'], batch['shape']), etc...}\nsess.run(train_step, feed_dict=train_dict)                \n<\/code><\/pre>\n\n<p>If you are not needing to do batch processing just leave off the first dimension and remove 'none' from the placeholder shape.<\/p>\n\n<p>I couldn't find any way of passing the images across in batch as an array of sparse matrices. It only worked if I created the 4th dimension. I'd be interested to know of alternatives.<\/p>\n\n<p>Whilst this doesn't give an exact answer to your question I hope it is of use as I have been struggling with similar issues.<\/p>\n"}]}
{"QuestionId":33919948,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2015-11-25T15:08:58.027","AcceptedAnswerId":33922859.0,"OwnerUserId":826983.0,"Title":"How to set adaptive learning rate for GradientDescentOptimizer?","Body":"<p>I am using TensorFlow to train a neural network. This is how I am initializing the :<\/p>\n\n<pre><\/pre>\n\n<p>The thing here is that I don't know how to set an update rule for the learning rate or a decay value for that. <\/p>\n\n<p>How can I use an adaptive learning rate here?<\/p>\n","answers":[{"AnswerId":"42921909","CreationDate":"2017-03-21T08:35:14.100","ParentId":null,"OwnerUserId":"5861256","Title":null,"Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/exponential_decay\" rel=\"noreferrer\">tensorflow<\/a> official docs<\/p>\n\n<pre><code>global_step = tf.Variable(0, trainable=False)\nstarter_learning_rate = 0.1\nlearning_rate = tf.train.exponential_decay(starter_learning_rate, global_step,\n                                       100000, 0.96, staircase=True)\n\n# Passing global_step to minimize() will increment it at each step.\nlearning_step = (\ntf.train.GradientDescentOptimizer(learning_rate)\n.minimize(...my loss..., global_step=global_step))\n<\/code><\/pre>\n"},{"AnswerId":"56465453","CreationDate":"2019-06-05T17:37:20.913","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>If you want to set specific learning rates for intervals of epochs like  <code>0 &lt; a &lt; b &lt; c &lt; ...<\/code>. Then you can define your learning rate as a conditional tensor, conditional on the global step, and feed this as normal to the optimiser. <\/p>\n\n<p>You could achieve this with a bunch of nested <code>tf.cond<\/code> statements, but its easier to build the tensor recursively:<\/p>\n\n<pre><code>def make_learning_rate_tensor(reduction_steps, learning_rates, global_step):\n    assert len(reduction_steps) + 1 == len(learning_rates)\n    if len(reduction_steps) == 1:\n        return tf.cond(\n            global_step &lt; reduction_steps[0],\n            lambda: learning_rates[0],\n            lambda: learning_rates[1]\n        )\n    else:\n        return tf.cond(\n            global_step &lt; reduction_steps[0],\n            lambda: learning_rates[0],\n            lambda: make_learning_rate_tensor(\n                reduction_steps[1:],\n                learning_rates[1:],\n                global_step,)\n            )\n<\/code><\/pre>\n\n<p>Then to use it you need to know how many training steps there are in a single epoch, so that we can use the global step to switch at the right time, and finally define the epochs and learning rates you want. So if I want the learning rates <code>[0.1, 0.01, 0.001, 0.0001]<\/code> during the epoch intervals of <code>[0, 19], [20, 59], [60, 99], [100, \\infty]<\/code> respectively, I would do:<\/p>\n\n<pre><code>global_step = tf.train.get_or_create_global_step()\nlearning_rates = [0.1, 0.01, 0.001, 0.0001]\nsteps_per_epoch = 225\nepochs_to_switch_at = [20, 60, 100]\nepochs_to_switch_at = [x*steps_per_epoch for x in epochs_to_switch_at ]\nlearning_rate = make_learning_rate_tensor(epochs_to_switch_at , learning_rates, global_step)\n<\/code><\/pre>\n"},{"AnswerId":"43930924","CreationDate":"2017-05-12T06:40:52.000","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Gradient descent algorithm uses the constant learning rate which you can provide in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/GradientDescentOptimizer#__init__\" rel=\"noreferrer\">during the initialization<\/a>. You can pass various learning rates in a way showed by Mrry.<\/p>\n\n<p>But instead of it you can also use <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/train#Optimizers\" rel=\"noreferrer\">more advanced optimizers<\/a> which have faster convergence rate and adapts to the situation.<\/p>\n\n<p>Here is a brief explanation based on my understanding:<\/p>\n\n<ul>\n<li><strong>momentum<\/strong> <a href=\"https:\/\/www.youtube.com\/watch?v=8yg2mRJx-z4\" rel=\"noreferrer\">helps<\/a> SGD to navigate along the relevant directions and softens the oscillations in the irrelevant. It simply adds a fraction of the direction of the previous step to a current step. This achieves amplification of speed in the correct dirrection and softens oscillation in wrong directions. This fraction is usually in the (0, 1) range. It also makes sense to use adaptive momentum. In the beginning of learning a big momentum will only hinder your progress, so it makse sense to use something like 0.01 and once all the high gradients disappeared you can use a bigger momentom. There is one problem with momentum: when we are very close to the goal, our momentum in most of the cases is very high and it does not know that it should slow down. This can cause it to miss or oscillate around the minima<\/li>\n<li><strong>nesterov accelerated gradient<\/strong> overcomes this problem by starting to slow down early. In momentum we first compute gradient and then make a jump in that direction amplified by whatever momentum we had previously. NAG does the same thing but in another order: at first we make a big jump based on our stored information, and then we calculate the gradient and make a small correction. This seemingly irrelevant change gives significant practical speedups.<\/li>\n<li><strong>AdaGrad<\/strong> or adaptive gradient allows the learning rate to adapt based on parameters. It performs larger updates for infrequent parameters and smaller updates for frequent one. Because of this it is well suited for sparse data (NLP or image recognition). Another advantage is that it basically illiminates the need to tune the learning rate. Each parameter has its own learning rate and due to the peculiarities of the algorithm the learning rate is monotonically decreasing. This causes the biggest problem: at some point of time the learning rate is so small that the system stops learning<\/li>\n<li><strong>AdaDelta<\/strong> resolves the problem of monotonically decreasing learning rate in AdaGrad. In AdaGrad the learning rate was calculated approximately as one divided by the sum of square roots. At each stage you add another square root to the sum, which causes denominator to constantly decrease. In AdaDelta instead of summing all past square roots it uses sliding window which allows the sum to decrease. <strong>RMSprop<\/strong> is very similar to AdaDelta<\/li>\n<li><p><strong>Adam<\/strong> or adaptive momentum is an algorithm similar to AdaDelta. But in addition to storing learning rates for each of the parameters it also stores momentum changes for each of them separately<\/p>\n\n<p>A <a href=\"http:\/\/sebastianruder.com\/optimizing-gradient-descent\/\" rel=\"noreferrer\">few visualizations<\/a>:\n<a href=\"https:\/\/i.stack.imgur.com\/qAx2i.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qAx2i.gif\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/1obtV.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1obtV.gif\" alt=\"enter image description here\"><\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"33922859","CreationDate":"2015-11-25T17:28:00.047","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>First of all, <code>tf.train.GradientDescentOptimizer<\/code> is designed to use a constant learning rate for all variables in all steps. TensorFlow also provides out-of-the-box adaptive optimizers including the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#AdagradOptimizer\"><code>tf.train.AdagradOptimizer<\/code><\/a> and the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#AdamOptimizer\"><code>tf.train.AdamOptimizer<\/code><\/a>, and these can be used as drop-in replacements.<\/p>\n\n<p>However, if you want to control the learning rate with otherwise-vanilla gradient descent, you can take advantage of the fact that the <code>learning_rate<\/code> argument to the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#GradientDescentOptimizer.__init__\"><code>tf.train.GradientDescentOptimizer<\/code> constructor<\/a> can be a <code>Tensor<\/code> object. This allows you to compute a different value for the learning rate in each step, for example:<\/p>\n\n<pre><code>learning_rate = tf.placeholder(tf.float32, shape=[])\n# ...\ntrain_step = tf.train.GradientDescentOptimizer(\n    learning_rate=learning_rate).minimize(mse)\n\nsess = tf.Session()\n\n# Feed different values for learning rate to each training step.\nsess.run(train_step, feed_dict={learning_rate: 0.1})\nsess.run(train_step, feed_dict={learning_rate: 0.1})\nsess.run(train_step, feed_dict={learning_rate: 0.01})\nsess.run(train_step, feed_dict={learning_rate: 0.01})\n<\/code><\/pre>\n\n<p>Alternatively, you could create a scalar <code>tf.Variable<\/code> that holds the learning rate, and assign it each time you want to change the learning rate.<\/p>\n"},{"AnswerId":"33931754","CreationDate":"2015-11-26T06:14:54.030","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Tensorflow provides an op to automatically apply an exponential decay to a learning rate tensor: <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#exponential_decay\" rel=\"noreferrer\"><code>tf.train.exponential_decay<\/code><\/a>.  For an example of it in use, see <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/mnist\/convolutional.py#L243\" rel=\"noreferrer\">this line in the MNIST convolutional model example<\/a>.  Then use @mrry's suggestion above to supply this variable as the learning_rate parameter to your optimizer of choice.<\/p>\n\n<p>The key excerpt to look at is:<\/p>\n\n<pre><code># Optimizer: set up a variable that's incremented once per batch and\n# controls the learning rate decay.\nbatch = tf.Variable(0)\n\nlearning_rate = tf.train.exponential_decay(\n  0.01,                # Base learning rate.\n  batch * BATCH_SIZE,  # Current index into the dataset.\n  train_size,          # Decay step.\n  0.95,                # Decay rate.\n  staircase=True)\n# Use simple momentum for the optimization.\noptimizer = tf.train.MomentumOptimizer(learning_rate,\n                                     0.9).minimize(loss,\n                                                   global_step=batch)\n<\/code><\/pre>\n\n<p>Note the <code>global_step=batch<\/code> parameter to minimize.  That tells the optimizer to helpfully increment the 'batch' parameter for you every time it trains.<\/p>\n"}]}
{"QuestionId":33920777,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-25T15:46:47.747","AcceptedAnswerId":null,"OwnerUserId":3511304.0,"Title":"Replacing a variable for an expression in a theano graph","Body":"<p>Given the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to replace <em>y<\/em> by <em>z+A<\/em> in the expression <em>f<\/em>, so it becomes equivalent to <em>x+z+A+B<\/em>?. It could be done manually, by searching the graph for occurrences of <em>y<\/em> and replacing them by <em>z+A<\/em>. However, it seems logical to have a simpler high-level way to do it in the API.<\/p>\n","answers":[{"AnswerId":"33933617","CreationDate":"2015-11-26T08:20:57.763","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This has been <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/xjjwOXGioh4\" rel=\"nofollow\">discussed on the theano-users mailing list<\/a>.<\/p>\n\n<p>You can use either the <code>givens<\/code> mechanism of <code>theano.function<\/code> or use <code>theano.clone<\/code>.<\/p>\n\n<p>Here's some example code:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\nx, y, z = T.dmatrices('x', 'y', 'z')\nA = theano.shared(np.random.rand(3, 4), borrow=True, name='A')\nB = theano.shared(np.random.rand(3, 4), borrow=True, name='B')\n\nh1 = x + y + B\nh2 = theano.clone(h1, {y: z + A})\n\nf1 = theano.function([x, y], h1)\nf2 = theano.function([x, z], h2)\nf3 = theano.function([x, z], h1, givens={y: z + A})\n\na = np.random.randn(3, 4)\nb = np.random.randn(3, 4)\nprint f1(a, b)\nprint f2(a, b)\nprint f3(a, b)\n<\/code><\/pre>\n\n<p>Note that the inputs to the theano functions have to be adjusted to ensure only the unspecified tensors are accepted as inputs (i.e. <code>y<\/code> is no longer an input once it is replaced by a function of <code>z<\/code>).<\/p>\n"}]}
{"QuestionId":33921398,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-11-25T16:18:00.607","AcceptedAnswerId":33922069.0,"OwnerUserId":395857.0,"Title":"Updating part of a shared variable in Theano","Body":"<p>How to update part of a shared variable in Theano?<\/p>\n\n<p>E.g. instead of doing:<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to only update part of , e.g. only the first column:<\/p>\n\n<pre><\/pre>\n\n<p>But doing so gives the error :<\/p>\n\n<pre><\/pre>\n\n<p>What is the typical way to do so in Theano?<\/p>\n\n<p> is a typical weight matrix:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33922069","CreationDate":"2015-11-25T16:48:46.093","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can use e.g. <code>theano.tensor.set_subtensor<\/code>.<\/p>\n\n<p>Taking the updates line you mention above, this becomes:<\/p>\n\n<pre><code>updates.append((W, T.set_subtensor(W[:, 0], W[:, 0]-learning_rate*gradient_W)))\n<\/code><\/pre>\n\n<p>where <code>T = theano.tensor<\/code>.<\/p>\n\n<p>Another possibility, a little more concise in your case, is <code>T.inc_subtensor<\/code>, where you specify only the increment instead of the new value:<\/p>\n\n<pre><code>updates.append((W, T.inc_subtensor(W[:, 0], -learning_rate*gradient_W)))\n<\/code><\/pre>\n"}]}
{"QuestionId":33922659,"AnswerCount":2,"Tags":"<python><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2015-11-25T17:18:02.040","AcceptedAnswerId":33923064.0,"OwnerUserId":3009514.0,"Title":"What output on this code means?","Body":"<p>I'm using <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow\">this tutorial<\/a> as example to build my caffe custom training function. On section 15 there is the following code:<\/p>\n\n<pre><\/pre>\n\n<p>On line 8 there is an  (output), what is the meaning of this code and it demensions. What is the meaning of . Why , why 8 and why 10? Should I change this array according to my own data set? If yes, what dimension should I use? Can someone explain me it?<\/p>\n","answers":[{"AnswerId":"33923064","CreationDate":"2015-11-25T17:39:05.857","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you read closely the tutorial you'll see that it deals with digits classification, hence the <strong>10<\/strong> classes. Moreover, they use a trick to tile 8 examples together (Section 11, near the <code>In [11]:<\/code>):<\/p>\n\n<blockquote>\n  <p># we use a little trick to tile the first eight images<\/p>\n<\/blockquote>\n\n<p>Hence the <strong>8<\/strong> dimension.<\/p>\n\n<p>Section 15 shows an example of tracking the progress of the network. It saves the output prediction probability per iteration. There are <strong>10<\/strong> classes times <strong>8<\/strong> examples per iteration, and there are <strong><code>niter<\/code><\/strong> iterations to track. All this information is stored in the 3D <code>output<\/code> array.<\/p>\n"},{"AnswerId":"33922817","CreationDate":"2015-11-25T17:26:10.657","ParentId":null,"OwnerUserId":"33258","Title":null,"Body":"<p>It looks like a call to <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/reference\/generated\/numpy.zeros.html\" rel=\"nofollow\"><code>numpy.zeros<\/code><\/a> where <code>shape = (niter, 8, 10)<\/code> which creates a 200 * 8 * 10 array of float 0.<\/p>\n"}]}
{"QuestionId":33922937,"AnswerCount":2,"Tags":"<python><csv><tensorflow>","CreationDate":"2015-11-25T17:32:07.653","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"Why does TensorFlow return [[nan nan]] instead of probabilities from a CSV file?","Body":"<p>Here is the code that I am using. I'm trying to get a 1, 0, or hopefully a probability in result to a real test set. When I just split up the training set and run it on the training set I get a ~93% accuracy rate, but when I train the program and run it on the actual test set (the one without the 1's and 0's filling in column 1) it returns nothing but nan's.<\/p>\n\n<pre><\/pre>\n\n<p>When I print  I get the following.<\/p>\n\n<pre><\/pre>\n\n<p>I don't know what I'm doing wrong here. All I want is for  to yield a 1, 0, or especially an array of probabilities where every unit inside the array has a length of 2.<\/p>\n\n<p>I don't expect that many people are going to be able to answer this question for me, but please try at the very least. I'm stuck here waiting for a stroke of genius moment which hasn't come in 2 days now so I figured that I would ask. Thank you!<\/p>\n\n<p>The  comes out looking like this-<\/p>\n\n<pre><\/pre>\n\n<p>And the only reason that the first unit in the data is equal to 1 is because I got rid of the nan's that filled that position in order to avoid errors. Note that everything after the first column is a feature. The first column is what I'm trying to be able to predict.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I changed the code to the following-<\/p>\n\n<pre><\/pre>\n\n<p>After the print out I see that one of the objects is Boolean. I don't know if that is the issue but take a look at the following and see if there is any way that you can help.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33929658","CreationDate":"2015-11-26T02:28:00.693","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I don't know the direct answer, but I know how I'd approach debugging it:  <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops.html#Print\" rel=\"noreferrer\"><code>tf.Print<\/code><\/a>.  It's an op that prints the value as tensorflow is executing, and returns the tensor for further computation, so you can just sprinkle them inline in your model.<\/p>\n\n<p>Try throwing in a few of these.  Instead of this line:<\/p>\n\n<pre><code>tf_softmax = tf.nn.softmax(tf.matmul(tf_in,tf_weight) + tf_bias)\n<\/code><\/pre>\n\n<p>Try:<\/p>\n\n<pre><code>tf_bias = tf.Print(tf_bias, [tf_bias], \"Bias: \")\ntf_weight = tf.Print(tf_weight, [tf_weight], \"Weight: \")\ntf_in = tf.Print(tf_in, [tf_in], \"TF_in: \")\nmatmul_result = tf.matmul(tf_in, tf_weight)\nmatmul_result = tf.Print(matmul_result, [matmul_result], \"Matmul: \")\ntf_softmax = tf.nn.softmax(matmul_result + tf_bias)\n<\/code><\/pre>\n\n<p>to see what Tensorflow thinks the intermediate values are.  If the NaNs are showing up earlier in the pipeline, it should give you a better idea of where the problem lies.  Good luck!  If you get some data out of this, feel free to follow up and we'll see if we can get you further.<\/p>\n\n<p>Updated to add:  Here's a stripped-down debugging version to try, where I got rid of the input functions and just generated some random data:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef dense_to_one_hot(labels_dense, num_classes=10):\n  \"\"\"Convert class labels from scalars to one-hot vectors.\"\"\"\n  num_labels = labels_dense.shape[0]\n  index_offset = np.arange(num_labels) * num_classes\n  labels_one_hot = np.zeros((num_labels, num_classes))\n  labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1\n  return labels_one_hot\n\nx_train=np.random.normal(0, 1, [50,10])\ny_train=np.random.randint(0, 10, [50])\ny_train_onehot = dense_to_one_hot(y_train, 10)\n\nx_test=np.random.normal(0, 1, [50,10])\ny_test=np.random.randint(0, 10, [50])\ny_test_onehot = dense_to_one_hot(y_test, 10)\n\n#  A number of features, 4 in this example\n#  B = 3 species of Iris (setosa, virginica and versicolor)\n\nA=10\nB=10\ntf_in = tf.placeholder(\"float\", [None, A]) # Features\ntf_weight = tf.Variable(tf.zeros([A,B]))\ntf_bias = tf.Variable(tf.zeros([B]))\ntf_softmax = tf.nn.softmax(tf.matmul(tf_in,tf_weight) + tf_bias)\n\ntf_bias = tf.Print(tf_bias, [tf_bias], \"Bias: \")\ntf_weight = tf.Print(tf_weight, [tf_weight], \"Weight: \")\ntf_in = tf.Print(tf_in, [tf_in], \"TF_in: \")\nmatmul_result = tf.matmul(tf_in, tf_weight)\nmatmul_result = tf.Print(matmul_result, [matmul_result], \"Matmul: \")\ntf_softmax = tf.nn.softmax(matmul_result + tf_bias)\n\n# Training via backpropagation\ntf_softmax_correct = tf.placeholder(\"float\", [None,B])\ntf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax))\n\n# Train using tf.train.GradientDescentOptimizer\ntf_train_step = tf.train.GradientDescentOptimizer(0.01).minimize(tf_cross_entropy)\n\n# Add accuracy checking nodes\ntf_correct_prediction = tf.equal(tf.argmax(tf_softmax,1), tf.argmax(tf_softmax_correct,1))\ntf_accuracy = tf.reduce_mean(tf.cast(tf_correct_prediction, \"float\"))\n\nprint tf_correct_prediction\nprint tf_accuracy\n\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\nfor i in range(1):\n    print \"Running the training step\"\n    sess.run(tf_train_step, feed_dict={tf_in: x_train, tf_softmax_correct: y_train_onehot})\n    #print y_train_onehot\n    #saver.save(sess, 'trained_csv_model')\n\n    print \"Running the eval step\"\n    ans = sess.run(tf_softmax, feed_dict={tf_in: x_test})\n    print ans\n<\/code><\/pre>\n\n<p>You should see lines starting with \"Bias: \", etc.  <\/p>\n"},{"AnswerId":"34764585","CreationDate":"2016-01-13T10:44:50.087","ParentId":null,"OwnerUserId":"461421","Title":null,"Body":"<p><code>tf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax))\n<\/code><\/p>\n\n<p>This was my problem on a project I was testing on. Specificaly it ended up being 0*log(0) which produces nan.<\/p>\n\n<p>If you replace this with:<\/p>\n\n<p><code>tf_cross_entropy = -tf.reduce_sum(tf_softmax_correct*tf.log(tf_softmax  + 1e-50))\n<\/code>\nIt should avoid the problem.<\/p>\n\n<p>Ive also used reduce_mean rather than reduce_sum. If you double the batch size and use reduce_sum it will double the cost (and the magnitude of the gradient). In addition to that when using tf.print (which prints to the console tensorfow was started from) it makes it a bit more comparable when varying batch size.<\/p>\n\n<p>Specifically this is what I'm using now when debugging:<\/p>\n\n<p><code>cross_entropy = -tf.reduce_sum(y*tf.log(model + 1e-50)) ## avoid nan due to 0*log(0)\ncross_entropy = tf.Print(cross_entropy, [cross_entropy], \"cost\") #print to the console tensorflow was started from\n<\/code><\/p>\n"}]}
{"QuestionId":33923288,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2015-11-25T17:52:05.803","AcceptedAnswerId":33923873.0,"OwnerUserId":4304769.0,"Title":"How to work with Tensorflow on Android platform?","Body":"<p>Google has made the TENSORFLOW open source for developers..<\/p>\n\n<p>Is there any way out for using this on android?<\/p>\n\n<p>The link is here <a href=\"http:\/\/tensorflow.org\" rel=\"nofollow\">TensorFlow<\/a>.<\/p>\n\n<p>I would love to have some directions to work on this API.<\/p>\n","answers":[{"AnswerId":"33923873","CreationDate":"2015-11-25T18:25:19.223","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow source repository includes an Android example application, with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/README.md\">some documentation<\/a>.<\/p>\n\n<p>The Android example includes a pre-trained model for image classification, and uses this to classify images captured by the camera. Typically you would build and train a model using the Python API; generate a serialised version of the model as a GraphDef protocol buffer (and possibly a checkpoint of the model parameters); and then load that and run inference steps using the C++ API.<\/p>\n"}]}
{"QuestionId":33925274,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-11-25T19:51:36.297","AcceptedAnswerId":33933424.0,"OwnerUserId":395857.0,"Title":"Theano: this shared variable already has an update expression","Body":"<p>Is there any way to define more than one update on the same share variable when calling ?<\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n\n<p>will throw the error :<\/p>\n\n<pre><\/pre>\n\n<p>I was interested in doing so as I have weight matrix that I need to update in such a way that different parts of it has a different update (which I do it using <a href=\"https:\/\/stackoverflow.com\/q\/33921398\/395857\"><\/a>).<\/p>\n","answers":[{"AnswerId":"33933424","CreationDate":"2015-11-26T08:10:07.527","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>As explained in <a href=\"https:\/\/groups.google.com\/forum\/m\/#!topic\/theano-users\/L6Efhjj4YOE\" rel=\"nofollow\">this thread<\/a> on the theano-users mailing list...<\/p>\n\n<p>You need to chain the <code>set_subtensor<\/code>s so that you only end up with a single entry for the whole tensor in the updates list.<\/p>\n\n<p>For example (code not like original because original does not actually include set_subtensor anywhere):<\/p>\n\n<pre><code>new_w = self.W\nnew_W = theano.tensor.set_subtensor(new_W[0], -1)\nnew_W = theano.tensor.set_subtensor(new_W[1], -2)\nupdates = [(self.W, new_W)]\ntrain = theano.function(inputs=[index], outputs=[cost], updates=updates)\n<\/code><\/pre>\n"}]}
{"QuestionId":33925595,"AnswerCount":1,"Tags":"<python><macos><tensorflow>","CreationDate":"2015-11-25T20:11:53.570","AcceptedAnswerId":34643711.0,"OwnerUserId":2713740.0,"Title":"Installing tensor flow on mac","Body":"<p>I am trying to install tensor flow on my mac with<\/p>\n\n<pre><\/pre>\n\n<p>And I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>I am on OSX 10.11.1 and all the development tools, pip and python2.7 are installed.<\/p>\n","answers":[{"AnswerId":"34643711","CreationDate":"2016-01-06T22:06:08.940","ParentId":null,"OwnerUserId":"1504540","Title":null,"Body":"<p>This may not be exactly the solution you're looking for, but I personally spent the last two hours diagnosing a similar issue where TensorFlow wouldn't work after correctly installing and having its packages satisfied. <\/p>\n\n<p>It was throwing one of the strange _IO module import errors that seemed to be saying python was working correctly, and I ended up tracing it back all the way to the System Integrity Protection issue introduced in El Cap. <\/p>\n\n<p>See <a href=\"https:\/\/apple.stackexchange.com\/questions\/208478\/how-do-i-disable-system-integrity-protection-sip-aka-rootless-on-os-x-10-11\">here<\/a> for easy fix that will let pip work correctly again.<\/p>\n\n<blockquote>\n  <p>For those that are curious, Apple introduced SIP to make sure no user accidentally hurts system files, but in the process it makes it trickier for users to run commands as root <\/p>\n<\/blockquote>\n\n<p>Also, remember to run <\/p>\n\n<pre><code>sudo easy_install --upgrade six\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>sudo pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>After disabling SIP to make sure you patch up any requirements.  You may need to uninstall some pip packages if this fails for some reason. <\/p>\n\n<p>As a final note, you should be off the ground if, in the python shell, you can <\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>and it doesn't throw any errors. <\/p>\n\n<p><a href=\"\/questions\/tagged\/python\" class=\"post-tag\" title=\"show questions tagged &#39;python&#39;\" rel=\"tag\">python<\/a><a href=\"\/questions\/tagged\/tensorflow\" class=\"post-tag\" title=\"show questions tagged &#39;tensorflow&#39;\" rel=\"tag\">tensorflow<\/a><a href=\"\/questions\/tagged\/elcapitan\" class=\"post-tag\" title=\"show questions tagged &#39;elcapitan&#39;\" rel=\"tag\">elcapitan<\/a><\/p>\n"}]}
{"QuestionId":33929368,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-26T01:49:00.083","AcceptedAnswerId":33933305.0,"OwnerUserId":395857.0,"Title":"How to perform a range on a Theano's TensorVariable?","Body":"<p>How to perform a range on a Theano's TensorVariable?<\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n\n<p>returns the error:<\/p>\n\n<pre><\/pre>\n\n<p>What is the correct way to use a range on a Theano's TensorVariable?<\/p>\n","answers":[{"AnswerId":"33933305","CreationDate":"2015-11-26T08:02:00.663","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It's not clear what this code is attempting to do because even if the loop worked, it wouldn't compute anything useful: <code>result<\/code> will always equal the sum of <code>start<\/code> and <code>constant<\/code> irrespective of the number of iterations.<\/p>\n\n<p>I'll assume that the intention was to compute <code>result<\/code> like this:<\/p>\n\n<pre><code>for iter in range(n_iters):\n    result = result + constant\n<\/code><\/pre>\n\n<p>The problem is that you're mixing symbolic, delayed execution, Theano operations with non-symbolic, immediately executed, Python operations.<\/p>\n\n<p><code>range<\/code> is a Python function that expects a Python integer parameter but you're providing a Theano symbolic value (<code>n_iters<\/code>, which has a double type instead of an integer type but let's assume it's actually an <code>iscalar<\/code> instead of a <code>dscalar<\/code>). As far as Python is concerned all Theano symbolic tensors are just objects: instances of a class type somewhere within the Theano library; they are most assuredly not integers. Even if you squint and try to pretend a Theano <code>iscalar<\/code> looks like a Python integer, it still doesn't work because Python operations execute immediately which means <code>n_iters<\/code> needs to have a value immediately. Theano on the other hand doesn't have a value for any <code>iscalar<\/code> until one is provided by calling a compiled Theano function (or via <code>eval<\/code>).<\/p>\n\n<p>To create a symbolic range, you can use <code>theano.tensor.arange<\/code> which operates just like NumPy's <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.arange.html\" rel=\"noreferrer\"><code>arange<\/code><\/a>, but symbolically.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import theano.tensor as T\nfrom theano import function\n\nmy_range_max = T.iscalar('my_range_max')\nmy_range = T.arange(my_range_max)\n\nf = function([my_range_max], my_range)\nprint('f(10): {0}'.format(f(10)))\n<\/code><\/pre>\n\n<p>outputs:<\/p>\n\n<pre><code>f(10): [0 1 2 3 4 5 6 7 8 9]\n<\/code><\/pre>\n\n<p>By making <code>n_iters<\/code> a symbolic variable you are implicitly saying \"I don't know how many iterations there need to be in this loop until a value is provided for <code>n_iters<\/code> later\". That being the case, you must use a symbolic loop instead of a Python <code>for<\/code> loop. In the latter case you must tell Python how many times to iterate <em>now<\/em>, you can't delay that decision until a value is provided for <code>n_iters<\/code> later. To solve this you need to switch to a symbolic loop, which is provided by Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"noreferrer\"><code>scan<\/code><\/a> operator.<\/p>\n\n<p>Here's the code changed to use <code>scan<\/code> (as well as the other assumed changes).<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nfrom theano import function\n\nconstant = T.dscalar('constant')\nn_iters = T.iscalar('n_iters')\nstart = T.dscalar('start')\n\nresults, _ = theano.scan(lambda result, constant: result + constant,\n                         outputs_info=[start], non_sequences=[constant], n_steps=n_iters)\n\nf = function([start, constant, n_iters], results[-1])\nprint('f(0,2,5): {0}'.format(f(0, 2, 5)))\n<\/code><\/pre>\n"}]}
{"QuestionId":33932901,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2015-11-26T07:34:17.340","AcceptedAnswerId":null,"OwnerUserId":5607347.0,"Title":"What's the purpose of tf.app.flags in TensorFlow?","Body":"<p>I am reading some example codes in Tensorflow, I found following code <\/p>\n\n<pre><\/pre>\n\n<p>in <\/p>\n\n<p>But I can't find any docs about this usage of . <\/p>\n\n<p>And I found the implementation of this flags is in the \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/platform\/default\/_flags.py\"><\/a><\/p>\n\n<p>Obviously, this  is somehow used to configure a network, so why  is it not in the API docs? Can anyone explain what is going on here? <\/p>\n","answers":[{"AnswerId":"52420982","CreationDate":"2018-09-20T08:44:49.070","ParentId":null,"OwnerUserId":"5106177","Title":null,"Body":"<p>The <code>tf.app.flags<\/code> module is a functionality provided by Tensorflow to implement command line flags for your Tensorflow program. As an example, the code you came across would do the following:<\/p>\n\n<pre><code>flags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate.')\n<\/code><\/pre>\n\n<p>The first parameter defines the name of the flag while the second defines the default value in case the flag is not specified while executing the file.<\/p>\n\n<p>So if you run the following:<\/p>\n\n<pre><code>$ python fully_connected_feed.py --learning_rate 1.00\n<\/code><\/pre>\n\n<p>then the learning rate is set to 1.00 and will remain 0.01 if the flag is not specified.<\/p>\n\n<p>As mentioned <a href=\"https:\/\/medium.com\/@zzzuzum\/command-line-flags-python-and-tensorflow-85ab217dbd5\" rel=\"noreferrer\" title=\"Command line flags: Python and Tensorflow\">in this article<\/a>, the docs are probably not present because this might be something that Google requires internally for its developers to use.<\/p>\n\n<p>Also, as mentioned in the post, there are several advantages of using Tensorflow flags over flag functionality provided by other Python packages such as <code>argparse<\/code> especially when dealing with Tensorflow models, the most important being that you can supply Tensorflow specific information to the code such as information about which GPU to use.<\/p>\n"},{"AnswerId":"52421144","CreationDate":"2018-09-20T08:52:30.510","ParentId":null,"OwnerUserId":"2135035","Title":null,"Body":"<p>At Google, they use flag systems to set default values for arguments. It's similar to argparse. They use their own flag system instead of argparse or sys.argv.<\/p>\n\n<p>Source: I worked there before.<\/p>\n"},{"AnswerId":"51821086","CreationDate":"2018-08-13T11:16:04.943","ParentId":null,"OwnerUserId":"6435600","Title":null,"Body":"<p>After trying many times I found this to print all FLAGS key as well as actual value -<\/p>\n\n<pre><code>for key in tf.app.flags.FLAGS.flag_values_dict():\n\n  print(key, FLAGS[key].value)\n<\/code><\/pre>\n"},{"AnswerId":"33938519","CreationDate":"2015-11-26T12:17:25.243","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.app.flags<\/code> module is presently a thin wrapper around <strike>python-gflags, so the <a href=\"https:\/\/github.com\/gflags\/python-gflags\">documentation for that project<\/a> is the best resource for how to use it<\/strike> <a href=\"https:\/\/docs.python.org\/2.7\/library\/argparse.html\"><code>argparse<\/code><\/a>, which implements a subset of the functionality in <a href=\"https:\/\/github.com\/gflags\/python-gflags\"><code>python-gflags<\/code><\/a>.<\/p>\n\n<p>Note that this module is currently packaged as a convenience for writing demo apps, and is not technically part of the public API, so it may change in future.<\/p>\n\n<p>We recommend that you implement your own flag parsing using <code>argparse<\/code> or whatever library you prefer.<\/p>\n\n<p><strong>EDIT:<\/strong> The <code>tf.app.flags<\/code> module is not in fact implemented using <code>python-gflags<\/code>, but it uses a similar API. <\/p>\n"},{"AnswerId":"50462209","CreationDate":"2018-05-22T07:35:56.017","ParentId":null,"OwnerUserId":"6400639","Title":null,"Body":"<p>When you use <code>tf.app.run()<\/code>, you can transfer the variable very conveniently between threads using <code>tf.app.flags<\/code>. See <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10\" rel=\"noreferrer\">this<\/a> for further usage of <code>tf.app.flags<\/code>.<\/p>\n"}]}
{"QuestionId":33933846,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-26T08:35:19.873","AcceptedAnswerId":34763748.0,"OwnerUserId":5602977.0,"Title":"TensorFlow Indices are not valid (out of bounds)","Body":"<p>Hi i'am currently trying to run TensorFlow with own image data. \nBut it crashes when i'am trying to run these function: its from mnist.py<\/p>\n\n<pre><\/pre>\n\n<p>with this error:<\/p>\n\n<pre><\/pre>\n\n<p>the number 100 is my batch_size and 447 is my number of classes.<\/p>\n\n<p>i also try to solve that issue like here\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/194\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/194<\/a>\nchanging the indeces line into this line:<\/p>\n\n<pre><\/pre>\n\n<p>didn't solve my problem.\nDoes anybody have an idea?<\/p>\n","answers":[{"AnswerId":"34763748","CreationDate":"2016-01-13T10:07:09.743","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>I had this error as well. And I realized my mistake. If you have 10 classes, your label values should range between 0-9, inclusive. The error was reproduced on the TensorFlow CIFAR10 example, used with <a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"nofollow noreferrer\">SVHN<\/a> dataset. Refer to question and answer below.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/34759227\/tensorflow-cifar10-example?lq=1\">TensorFlow CIFAR10 Example<\/a><\/p>\n"}]}
{"QuestionId":33943992,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-26T17:11:08.113","AcceptedAnswerId":null,"OwnerUserId":5609344.0,"Title":"In Tensorflow, how do I generate a scalar summary?","Body":"<p>Does anyone have a minimal example of using a SummaryWriter with a scalar_summary in order to see (say) a cross entropy result during a training run?<\/p>\n\n<p>The example <a href=\"http:\/\/www.tensorflow.org\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">given in the documentation<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>Returns an error: TypeError: Fetch argument None of None has invalid type , must be a string or Tensor. (Can not convert a NoneType into a Tensor or Operation.)\nWhen I run it.        <\/p>\n\n<p>If I add a:<\/p>\n\n<pre><\/pre>\n\n<p>operation after my cross entropy calculation, then instead I get the error:<\/p>\n\n<p>InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_2' with dtype float<\/p>\n\n<p>Which suggests that I need to add a feed_dict to the <\/p>\n\n<pre><\/pre>\n\n<p>call, but I am not clear what that feed_dict should contain....<\/p>\n","answers":[{"AnswerId":"33948801","CreationDate":"2015-11-27T00:32:15.977","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The error is probably coming from:<\/p>\n\n<pre><code>session.run(training_op)\n<\/code><\/pre>\n\n<p>Did you paste the example code into a version of the mnist code that requires a feed_dict for feeding in training examples?  Check the backtrace it gave you (and include it above if that doesn't solve the problem).<\/p>\n"},{"AnswerId":"34191229","CreationDate":"2015-12-09T23:48:44.180","ParentId":null,"OwnerUserId":"1576602","Title":null,"Body":"<p>The <code>feed_dict<\/code> should contain the same values that you use for running the <code>training_op<\/code>. It basically specifies the input values to your network for which you want to calculate the summaries. <\/p>\n"}]}
{"QuestionId":33944683,"AnswerCount":3,"Tags":"<python><functional-programming><tensorflow>","CreationDate":"2015-11-26T17:58:43.797","AcceptedAnswerId":33946482.0,"OwnerUserId":3849143.0,"Title":"Tensorflow \"map operation\" for tensor?","Body":"<p>I am  adapting the <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow noreferrer\">cifar10 convolution example<\/a> to my problem.  I'd like to change the data input from a design that reads images one-at-a-time from a file to a design that operates on an already-in-memory set of images.  The original  function looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>In the original version,  is a tensor containing one image.<\/p>\n\n<p>I keep all my images in RAM, so instead of using , I have one huge , where  is (something, 32, 32, 3).<\/p>\n\n<p>My question is very-very basic:  what is the best way to apply some function ( in my case) to all elements of  ? <\/p>\n\n<p>Iterating is problematic in tensorflow, with limited slices(<a href=\"https:\/\/stackoverflow.com\/questions\/33736795\/tensorflow-numpy-like-tensor-indexing\">TensorFlow - numpy-like tensor indexing<\/a>).  Is there a solution to achieving this using just one command?<\/p>\n","answers":[{"AnswerId":"33946482","CreationDate":"2015-11-26T20:25:34.383","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>There are a few answers - none quite as elegant as a map function.  Which is best depends a bit on your desire for memory efficiency.<\/p>\n\n<p>(a)  You can use <code>enqueue_many<\/code> to throw them into a <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/io_ops.html#FIFOQueue\"><code>tf.FIFOQueue<\/code><\/a> and then dequeue and <code>tf.image.resize_image_with_crop_or_pad<\/code> an image at a time, and concat it all back into one big smoosh.  This is probably slow.  Requires N calls to run for N images.<\/p>\n\n<p>(b)  You could use a single placeholder feed and run to resize and crop them on their way in from your original datasource.  This is possibly the best option from a memory perspective, because you never have to store the unresized data in memory.<\/p>\n\n<p>(c)  You could use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py#L1077\"><code>tf.control_flow_ops.While<\/code><\/a> op to iterate through the full batch and build up the result in a <code>tf.Variable<\/code>.  Particularly if you take advantage of the parallel execution permitted by while, this is likely to be the fastest approach.<\/p>\n\n<p>I'd probably go for option (c) unless you want to minimize memory use, in which case filtering it on the way in (option b) would be a better choice.<\/p>\n"},{"AnswerId":"45451927","CreationDate":"2017-08-02T05:10:12.597","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Tensorflow provides a couple of <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/functional_ops\" rel=\"nofollow noreferrer\">higher-order functions<\/a> and one of them is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\"><code>tf.map_fn<\/code><\/a>. The usage is very easy: you define your mappping and apply it to the tensor:<\/p>\n\n<pre><code>variable = tf.Variable(...)\nmapping = lambda x: f(x)\nres = tf.map_fn(mapping, variable)\n<\/code><\/pre>\n"},{"AnswerId":"38008313","CreationDate":"2016-06-24T07:51:33.050","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>As of version 0.8 there is <code>map_fn<\/code>. From the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>map_fn(fn, elems, dtype=None, parallel_iterations=10, back_prop=True,\n  swap_memory=False, name=None)<\/p>\n  \n  <p>map on the list of tensors unpacked from <code>elems<\/code> on dimension 0.<\/p>\n  \n  <p>This map operator repeatedly applies the callable <code>fn<\/code> to a sequence of elements from first to last. The elements are made of the tensors unpacked from <code>elems<\/code>. <code>dtype<\/code> is the data type of the return value of <code>fn<\/code>. Users must provide <code>dtype<\/code> if it is different from the data type of <code>elems<\/code>.<\/p>\n  \n  <p>Suppose that <code>elems<\/code> is unpacked into <code>values<\/code>, a list of tensors. The shape of the result tensor is <code>[len(values)] + fn(values[0]).shape<\/code>.<\/p>\n  \n  <p>Args:<\/p>\n  \n  <p>fn: The callable to be performed.<\/p>\n  \n  <p>elems: A tensor to be unpacked to apply <code>fn<\/code>.<\/p>\n  \n  <p>dtype: (optional) The output type of <code>fn<\/code>.<\/p>\n  \n  <p>parallel_iterations: (optional) The number of iterations allowed to run\n                             in parallel.\n        back_prop: (optional) True enables back propagation.\n        swap_memory: (optional) True enables GPU-CPU memory swapping.\n        name: (optional) Name prefix for the returned tensors.<\/p>\n  \n  <p>Returns:<\/p>\n  \n  <p>A tensor that packs the results of applying <code>fn<\/code> to the list of tensors\n        unpacked from <code>elems<\/code>, from first to last.<\/p>\n  \n  <p>Raises:<\/p>\n  \n  <p>TypeError: if <code>fn<\/code> is not callable.<\/p>\n  \n  <p>Example:<\/p>\n<\/blockquote>\n\n<pre><code>  elems = [1, 2, 3, 4, 5, 6]\n  squares = map_fn(lambda x: x * x, elems)\n  # squares == [1, 4, 9, 16, 25, 36]\n  ```\n<\/code><\/pre>\n"}]}
{"QuestionId":33946090,"AnswerCount":1,"Tags":"<python><scikit-learn><theano><lasagne>","CreationDate":"2015-11-26T19:52:28.350","AcceptedAnswerId":33955332.0,"OwnerUserId":1492106.0,"Title":"Incorrect shape for arguments to lasagne function","Body":"<p>I'm trying to build a neural network regressor using the <a href=\"https:\/\/github.com\/aigamedev\/scikit-neuralnetwork\" rel=\"nofollow\">scikit-neuralnetwork<\/a> library.<\/p>\n\n<p>As I understand it, ny NN seems to be being built fine, but I keep running into the following error at the  call:<\/p>\n\n<pre><\/pre>\n\n<p>My code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Might someone here know what's going wrong here? Any help would be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"33955332","CreationDate":"2015-11-27T10:26:51.427","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that the model expects its input to be a matrix but you're providing a vector.<\/p>\n\n<p>In the line<\/p>\n\n<pre><code>print nn.predict(X_train[0])\n<\/code><\/pre>\n\n<p>why do you only pass the first row of <code>X_train<\/code>?<\/p>\n\n<p>I expect if you passed the whole matrix, i.e.<\/p>\n\n<pre><code>print nn.predict(X_train)\n<\/code><\/pre>\n\n<p>or stacked the first row so it's passed as a matrix with only one row:<\/p>\n\n<pre><code>print nn.predict(np.expand_dims(X_train[0], 0))\n<\/code><\/pre>\n\n<p>then it may work as expected.<\/p>\n"}]}
{"QuestionId":33947726,"AnswerCount":1,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2015-11-26T22:15:12.733","AcceptedAnswerId":33971421.0,"OwnerUserId":4476617.0,"Title":"Indexing tensor with index matrix in theano?","Body":"<p>I have a theano tensor A such that A.shape = (40, 20, 5) and a theano matrix B such that B.shape = (40, 20). Is there a one-line operation I can perform to get a matrix C, where C.shape = (40, 20) and C(i,j) = A[i, j, B[i,j]] with theano syntax?<\/p>\n\n<p>Essentially, I want to use B as an indexing matrix; what is the most efficient\/elegant to do this using theano?<\/p>\n","answers":[{"AnswerId":"33971421","CreationDate":"2015-11-28T12:42:38.607","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can do the following in numpy:<\/p>\n\n<pre><code>import numpy as np\n\nA = np.arange(4 * 2 * 5).reshape(4, 2, 5)\nB = np.arange(4 * 2).reshape(4, 2) % 5\n\nC = A[np.arange(A.shape[0])[:, np.newaxis], np.arange(A.shape[1]), B]\n<\/code><\/pre>\n\n<p>So you can do the same thing in theano:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nAA = T.tensor3()\nBB = T.imatrix()\n\nCC = AA[T.arange(AA.shape[0]).reshape((-1, 1)), T.arange(AA.shape[1]), BB]\n\nf = theano.function([AA, BB], CC)\n\nf(A.astype(theano.config.floatX), B)\n<\/code><\/pre>\n"}]}
{"QuestionId":33949786,"AnswerCount":8,"Tags":"<python><tensorflow>","CreationDate":"2015-11-27T03:17:52.330","AcceptedAnswerId":33950177.0,"OwnerUserId":3090897.0,"Title":"How could I use batch normalization in TensorFlow?","Body":"<p>I would like to use <em>batch normalization<\/em> in TensorFlow. I found the related C++ source code in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/ops\/nn_ops.cc\" rel=\"noreferrer\"><\/a>. However, I did not find it documented on tensorflow.org.<\/p>\n\n<p>BN has different semantics in MLP and CNN, so I am not sure what exactly this BN does.<\/p>\n\n<p>I did not find a method called  either.<\/p>\n","answers":[{"AnswerId":"34634291","CreationDate":"2016-01-06T13:26:41.827","ParentId":null,"OwnerUserId":"3632556","Title":null,"Body":"<p>The following works fine for me, it does not require invoking EMA-apply outside.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.python import control_flow_ops\n\ndef batch_norm(x, n_out, phase_train, scope='bn'):\n    \"\"\"\n    Batch normalization on convolutional maps.\n    Args:\n        x:           Tensor, 4D BHWD input maps\n        n_out:       integer, depth of input maps\n        phase_train: boolean tf.Varialbe, true indicates training phase\n        scope:       string, variable scope\n    Return:\n        normed:      batch-normalized maps\n    \"\"\"\n    with tf.variable_scope(scope):\n        beta = tf.Variable(tf.constant(0.0, shape=[n_out]),\n                                     name='beta', trainable=True)\n        gamma = tf.Variable(tf.constant(1.0, shape=[n_out]),\n                                      name='gamma', trainable=True)\n        batch_mean, batch_var = tf.nn.moments(x, [0,1,2], name='moments')\n        ema = tf.train.ExponentialMovingAverage(decay=0.5)\n\n        def mean_var_with_update():\n            ema_apply_op = ema.apply([batch_mean, batch_var])\n            with tf.control_dependencies([ema_apply_op]):\n                return tf.identity(batch_mean), tf.identity(batch_var)\n\n        mean, var = tf.cond(phase_train,\n                            mean_var_with_update,\n                            lambda: (ema.average(batch_mean), ema.average(batch_var)))\n        normed = tf.nn.batch_normalization(x, mean, var, beta, gamma, 1e-3)\n    return normed\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import math\n\nn_in, n_out = 3, 16\nksize = 3\nstride = 1\nphase_train = tf.placeholder(tf.bool, name='phase_train')\ninput_image = tf.placeholder(tf.float32, name='input_image')\nkernel = tf.Variable(tf.truncated_normal([ksize, ksize, n_in, n_out],\n                                   stddev=math.sqrt(2.0\/(ksize*ksize*n_out))),\n                                   name='kernel')\nconv = tf.nn.conv2d(input_image, kernel, [1,stride,stride,1], padding='SAME')\nconv_bn = batch_norm(conv, n_out, phase_train)\nrelu = tf.nn.relu(conv_bn)\n\nwith tf.Session() as session:\n    session.run(tf.initialize_all_variables())\n    for i in range(20):\n        test_image = np.random.rand(4,32,32,3)\n        sess_outputs = session.run([relu],\n          {input_image.name: test_image, phase_train.name: True})\n<\/code><\/pre>\n"},{"AnswerId":"43285333","CreationDate":"2017-04-07T19:00:17.217","ParentId":null,"OwnerUserId":"7832197","Title":null,"Body":"<p>As of TensorFlow 1.0 (February 2017) there's also the high-level <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/batch_normalization\" rel=\"noreferrer\" title=\"tf.layers.batch_normalization\"><code>tf.layers.batch_normalization<\/code><\/a> API included in TensorFlow itself.<\/p>\n\n<p>It's super simple to use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Set this to True for training and False for testing\ntraining = tf.placeholder(tf.bool)\n\nx = tf.layers.dense(input_x, units=100)\nx = tf.layers.batch_normalization(x, training=training)\nx = tf.nn.relu(x)\n<\/code><\/pre>\n\n<p>...except that it adds extra ops to the graph (for updating its mean and variance variables) in such a way that they won't be dependencies of your training op. You can either just run the ops separately:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nsess.run([train_op, extra_update_ops], ...)\n<\/code><\/pre>\n\n<p>or add the update ops as dependencies of your training op manually, then just run your training op as normal:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nwith tf.control_dependencies(extra_update_ops):\n    train_op = optimizer.minimize(loss)\n...\nsess.run([train_op], ...)\n<\/code><\/pre>\n"},{"AnswerId":"38320613","CreationDate":"2016-07-12T05:23:56.177","ParentId":null,"OwnerUserId":"3167448","Title":null,"Body":"<p>There is also an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b826b79718e3e93148c3545e7aa3f90891744cc0\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L100\" rel=\"nofollow noreferrer\">\"official\" batch normalization layer<\/a> coded by the developers. They don't have very good docs on how to use it but here is how to use it (according to me):<\/p>\n\n<pre><code>from tensorflow.contrib.layers.python.layers import batch_norm as batch_norm\n\ndef batch_norm_layer(x,train_phase,scope_bn):\n    bn_train = batch_norm(x, decay=0.999, center=True, scale=True,\n    updates_collections=None,\n    is_training=True,\n    reuse=None, # is this right?\n    trainable=True,\n    scope=scope_bn)\n    bn_inference = batch_norm(x, decay=0.999, center=True, scale=True,\n    updates_collections=None,\n    is_training=False,\n    reuse=True, # is this right?\n    trainable=True,\n    scope=scope_bn)\n    z = tf.cond(train_phase, lambda: bn_train, lambda: bn_inference)\n    return z\n<\/code><\/pre>\n\n<p>to actually use it you need to create a placeholder for <code>train_phase<\/code> that indicates if you are in training or inference phase (as in <code>train_phase = tf.placeholder(tf.bool, name='phase_train')<\/code>). Its value can be filled during inference or training with a <code>tf.session<\/code> as in:<\/p>\n\n<pre><code>test_error = sess.run(fetches=cross_entropy, feed_dict={x: batch_xtest, y_:batch_ytest, train_phase: False})\n<\/code><\/pre>\n\n<p>or during training:<\/p>\n\n<pre><code>sess.run(fetches=train_step, feed_dict={x: batch_xs, y_:batch_ys, train_phase: True})\n<\/code><\/pre>\n\n<p>I'm pretty sure this is correct according to the discussion in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1122#issuecomment-231917736\" rel=\"nofollow noreferrer\">github<\/a>.<\/p>\n\n<hr>\n\n<p>Seems there is another useful link:<\/p>\n\n<p><a href=\"http:\/\/r2rt.com\/implementing-batch-normalization-in-tensorflow.html\" rel=\"nofollow noreferrer\">http:\/\/r2rt.com\/implementing-batch-normalization-in-tensorflow.html<\/a><\/p>\n"},{"AnswerId":"33950177","CreationDate":"2015-11-27T04:16:11.250","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p><strong>Update July 2016<\/strong>  The easiest way to use batch normalization in TensorFlow is through the higher-level interfaces provided in either <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\" rel=\"noreferrer\">contrib\/layers<\/a>, <a href=\"http:\/\/tflearn.org\/layers\/normalization\/\" rel=\"noreferrer\">tflearn<\/a>, or <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/slim\/ops.py\" rel=\"noreferrer\">slim<\/a>.<\/p>\n\n<p><strong>Previous answer if you want to DIY<\/strong>:\nThe documentation string for this has improved since the release - see the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/ops\/nn_ops.cc#L65\" rel=\"noreferrer\">docs comment in the master branch<\/a> instead of the one you found.  It clarifies, in particular, that it's the output from <code>tf.nn.moments<\/code>.<\/p>\n\n<p>You can see a very simple example of its use in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3972c791b9f4d9a61b9ad6399b481df396f359ff\/tensorflow\/python\/ops\/nn_test.py#L518\" rel=\"noreferrer\">batch_norm test code<\/a>.  For a more real-world use example, I've included below the helper class and use notes that I scribbled up for my own use (no warranty provided!):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\"\"\"A helper class for managing batch normalization state.                   \n\nThis class is designed to simplify adding batch normalization               \n(http:\/\/arxiv.org\/pdf\/1502.03167v3.pdf) to your model by                    \nmanaging the state variables associated with it.                            \n\nImportant use note:  The function get_assigner() returns                    \nan op that must be executed to save the updated state.                      \nA suggested way to do this is to make execution of the                      \nmodel optimizer force it, e.g., by:                                         \n\n  update_assignments = tf.group(bn1.get_assigner(),                         \n                                bn2.get_assigner())                         \n  with tf.control_dependencies([optimizer]):                                \n    optimizer = tf.group(update_assignments)                                \n\n\"\"\"\n\nimport tensorflow as tf\n\n\nclass ConvolutionalBatchNormalizer(object):\n  \"\"\"Helper class that groups the normalization logic and variables.        \n\n  Use:                                                                      \n      ewma = tf.train.ExponentialMovingAverage(decay=0.99)                  \n      bn = ConvolutionalBatchNormalizer(depth, 0.001, ewma, True)           \n      update_assignments = bn.get_assigner()                                \n      x = bn.normalize(y, train=training?)                                  \n      (the output x will be batch-normalized).                              \n  \"\"\"\n\n  def __init__(self, depth, epsilon, ewma_trainer, scale_after_norm):\n    self.mean = tf.Variable(tf.constant(0.0, shape=[depth]),\n                            trainable=False)\n    self.variance = tf.Variable(tf.constant(1.0, shape=[depth]),\n                                trainable=False)\n    self.beta = tf.Variable(tf.constant(0.0, shape=[depth]))\n    self.gamma = tf.Variable(tf.constant(1.0, shape=[depth]))\n    self.ewma_trainer = ewma_trainer\n    self.epsilon = epsilon\n    self.scale_after_norm = scale_after_norm\n\n  def get_assigner(self):\n    \"\"\"Returns an EWMA apply op that must be invoked after optimization.\"\"\"\n    return self.ewma_trainer.apply([self.mean, self.variance])\n\n  def normalize(self, x, train=True):\n    \"\"\"Returns a batch-normalized version of x.\"\"\"\n    if train:\n      mean, variance = tf.nn.moments(x, [0, 1, 2])\n      assign_mean = self.mean.assign(mean)\n      assign_variance = self.variance.assign(variance)\n      with tf.control_dependencies([assign_mean, assign_variance]):\n        return tf.nn.batch_norm_with_global_normalization(\n            x, mean, variance, self.beta, self.gamma,\n            self.epsilon, self.scale_after_norm)\n    else:\n      mean = self.ewma_trainer.average(self.mean)\n      variance = self.ewma_trainer.average(self.variance)\n      local_beta = tf.identity(self.beta)\n      local_gamma = tf.identity(self.gamma)\n      return tf.nn.batch_norm_with_global_normalization(\n          x, mean, variance, local_beta, local_gamma,\n          self.epsilon, self.scale_after_norm)\n<\/code><\/pre>\n\n<p>Note that I called it a <code>ConvolutionalBatchNormalizer<\/code> because it pins the use of <code>tf.nn.moments<\/code> to sum across axes 0, 1, and 2, whereas for non-convolutional use you might only want axis 0.<\/p>\n\n<p>Feedback appreciated if you use it.<\/p>\n"},{"AnswerId":"39128711","CreationDate":"2016-08-24T16:35:12.897","ParentId":null,"OwnerUserId":"6753555","Title":null,"Body":"<p>You can simply use the build-in batch_norm layer:<\/p>\n\n<pre><code>batch_norm = tf.cond(is_train, \n    lambda: tf.contrib.layers.batch_norm(prev, activation_fn=tf.nn.relu, is_training=True, reuse=None),\n    lambda: tf.contrib.layers.batch_norm(prev, activation_fn =tf.nn.relu, is_training=False, reuse=True))\n<\/code><\/pre>\n\n<p>where prev is the output of your previous layer (can be both fully-connected or a convolutional layer) and is_train is a boolean placeholder. Just use batch_norm as the input to the next layer, then.<\/p>\n"},{"AnswerId":"36511416","CreationDate":"2016-04-09T00:36:01.687","ParentId":null,"OwnerUserId":"2008401","Title":null,"Body":"<p>Since someone recently edited this, I'd like to clarify that this is no longer an issue.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/34634291\/3924118\">This answer<\/a> does not seem correct  When <code>phase_train<\/code> is set to false, it still updates the ema mean and variance. This can be verified with the following code snippet.<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, 20, 20, 10], name='input')\nphase_train = tf.placeholder(tf.bool, name='phase_train')\n\n# generate random noise to pass into batch norm\nx_gen = tf.random_normal([50,20,20,10])\npt_false = tf.Variable(tf.constant(True))\n\n#generate a constant variable to pass into batch norm\ny = x_gen.eval()\n\n[bn, bn_vars] = batch_norm(x, 10, phase_train)\n\ntf.initialize_all_variables().run()\ntrain_step = lambda: bn.eval({x:x_gen.eval(), phase_train:True})\ntest_step = lambda: bn.eval({x:y, phase_train:False})\ntest_step_c = lambda: bn.eval({x:y, phase_train:True})\n\n# Verify that this is different as expected, two different x's have different norms\nprint(train_step()[0][0][0])\nprint(train_step()[0][0][0])\n\n# Verify that this is same as expected, same x's (y) have same norm\nprint(train_step_c()[0][0][0])\nprint(train_step_c()[0][0][0])\n\n# THIS IS DIFFERENT but should be they same, should only be reading from the ema.\nprint(test_step()[0][0][0])\nprint(test_step()[0][0][0])\n<\/code><\/pre>\n"},{"AnswerId":"36525884","CreationDate":"2016-04-10T04:43:47.547","ParentId":null,"OwnerUserId":"6104317","Title":null,"Body":"<p>So a simple example of the use of this batchnorm class:<\/p>\n\n<pre><code>from bn_class import *\n\nwith tf.name_scope('Batch_norm_conv1') as scope:\n    ewma = tf.train.ExponentialMovingAverage(decay=0.99)                  \n    bn_conv1 = ConvolutionalBatchNormalizer(num_filt_1, 0.001, ewma, True)           \n    update_assignments = bn_conv1.get_assigner() \n    a_conv1 = bn_conv1.normalize(a_conv1, train=bn_train) \n    h_conv1 = tf.nn.relu(a_conv1)\n<\/code><\/pre>\n"},{"AnswerId":"38325288","CreationDate":"2016-07-12T09:45:06.733","ParentId":null,"OwnerUserId":"3633250","Title":null,"Body":"<p>Using TensorFlow built-in batch_norm layer, below is the code to load data, build a network with one hidden ReLU layer and L2 normalization and introduce batch normalization for both hidden and out layer. This runs fine and trains fine. Just FYI this example is mostly built upon the data and code from Udacity DeepLearning course.\nP.S. Yes, parts of it were discussed one way or another in answers earlier but I decided to gather in one code snippet everything so that you have example of whole network training process with Batch Normalization and its evaluation<\/p>\n\n<pre><code># These are all the modules we'll be using later. Make sure you can import them\n# before proceeding further.\nfrom __future__ import print_function\nimport numpy as np\nimport tensorflow as tf\nfrom six.moves import cPickle as pickle\n\npickle_file = '\/home\/maxkhk\/Documents\/Udacity\/DeepLearningCourse\/SourceCode\/tensorflow\/examples\/udacity\/notMNIST.pickle'\n\nwith open(pickle_file, 'rb') as f:\n  save = pickle.load(f)\n  train_dataset = save['train_dataset']\n  train_labels = save['train_labels']\n  valid_dataset = save['valid_dataset']\n  valid_labels = save['valid_labels']\n  test_dataset = save['test_dataset']\n  test_labels = save['test_labels']\n  del save  # hint to help gc free up memory\n  print('Training set', train_dataset.shape, train_labels.shape)\n  print('Validation set', valid_dataset.shape, valid_labels.shape)\n  print('Test set', test_dataset.shape, test_labels.shape)\n\nimage_size = 28\nnum_labels = 10\n\ndef reformat(dataset, labels):\n  dataset = dataset.reshape((-1, image_size * image_size)).astype(np.float32)\n  # Map 2 to [0.0, 1.0, 0.0 ...], 3 to [0.0, 0.0, 1.0 ...]\n  labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)\n  return dataset, labels\ntrain_dataset, train_labels = reformat(train_dataset, train_labels)\nvalid_dataset, valid_labels = reformat(valid_dataset, valid_labels)\ntest_dataset, test_labels = reformat(test_dataset, test_labels)\nprint('Training set', train_dataset.shape, train_labels.shape)\nprint('Validation set', valid_dataset.shape, valid_labels.shape)\nprint('Test set', test_dataset.shape, test_labels.shape)\n\n\ndef accuracy(predictions, labels):\n  return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))\n          \/ predictions.shape[0])\n\n\n#for NeuralNetwork model code is below\n#We will use SGD for training to save our time. Code is from Assignment 2\n#beta is the new parameter - controls level of regularization.\n#Feel free to play with it - the best one I found is 0.001\n#notice, we introduce L2 for both biases and weights of all layers\n\nbatch_size = 128\nbeta = 0.001\n\n#building tensorflow graph\ngraph = tf.Graph()\nwith graph.as_default():\n      # Input data. For the training data, we use a placeholder that will be fed\n  # at run time with a training minibatch.\n  tf_train_dataset = tf.placeholder(tf.float32,\n                                    shape=(batch_size, image_size * image_size))\n  tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))\n  tf_valid_dataset = tf.constant(valid_dataset)\n  tf_test_dataset = tf.constant(test_dataset)\n\n  #introduce batchnorm\n  tf_train_dataset_bn = tf.contrib.layers.batch_norm(tf_train_dataset)\n\n\n  #now let's build our new hidden layer\n  #that's how many hidden neurons we want\n  num_hidden_neurons = 1024\n  #its weights\n  hidden_weights = tf.Variable(\n    tf.truncated_normal([image_size * image_size, num_hidden_neurons]))\n  hidden_biases = tf.Variable(tf.zeros([num_hidden_neurons]))\n\n  #now the layer itself. It multiplies data by weights, adds biases\n  #and takes ReLU over result\n  hidden_layer = tf.nn.relu(tf.matmul(tf_train_dataset_bn, hidden_weights) + hidden_biases)\n\n  #adding the batch normalization layerhi()\n  hidden_layer_bn = tf.contrib.layers.batch_norm(hidden_layer)\n\n  #time to go for output linear layer\n  #out weights connect hidden neurons to output labels\n  #biases are added to output labels  \n  out_weights = tf.Variable(\n    tf.truncated_normal([num_hidden_neurons, num_labels]))  \n\n  out_biases = tf.Variable(tf.zeros([num_labels]))  \n\n  #compute output  \n  out_layer = tf.matmul(hidden_layer_bn,out_weights) + out_biases\n  #our real output is a softmax of prior result\n  #and we also compute its cross-entropy to get our loss\n  #Notice - we introduce our L2 here\n  loss = (tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(\n    out_layer, tf_train_labels) +\n    beta*tf.nn.l2_loss(hidden_weights) +\n    beta*tf.nn.l2_loss(hidden_biases) +\n    beta*tf.nn.l2_loss(out_weights) +\n    beta*tf.nn.l2_loss(out_biases)))\n\n  #now we just minimize this loss to actually train the network\n  optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)\n\n  #nice, now let's calculate the predictions on each dataset for evaluating the\n  #performance so far\n  # Predictions for the training, validation, and test data.\n  train_prediction = tf.nn.softmax(out_layer)\n  valid_relu = tf.nn.relu(  tf.matmul(tf_valid_dataset, hidden_weights) + hidden_biases)\n  valid_prediction = tf.nn.softmax( tf.matmul(valid_relu, out_weights) + out_biases) \n\n  test_relu = tf.nn.relu( tf.matmul( tf_test_dataset, hidden_weights) + hidden_biases)\n  test_prediction = tf.nn.softmax(tf.matmul(test_relu, out_weights) + out_biases)\n\n\n\n#now is the actual training on the ANN we built\n#we will run it for some number of steps and evaluate the progress after \n#every 500 steps\n\n#number of steps we will train our ANN\nnum_steps = 3001\n\n#actual training\nwith tf.Session(graph=graph) as session:\n  tf.initialize_all_variables().run()\n  print(\"Initialized\")\n  for step in range(num_steps):\n    # Pick an offset within the training data, which has been randomized.\n    # Note: we could use better randomization across epochs.\n    offset = (step * batch_size) % (train_labels.shape[0] - batch_size)\n    # Generate a minibatch.\n    batch_data = train_dataset[offset:(offset + batch_size), :]\n    batch_labels = train_labels[offset:(offset + batch_size), :]\n    # Prepare a dictionary telling the session where to feed the minibatch.\n    # The key of the dictionary is the placeholder node of the graph to be fed,\n    # and the value is the numpy array to feed to it.\n    feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}\n    _, l, predictions = session.run(\n      [optimizer, loss, train_prediction], feed_dict=feed_dict)\n    if (step % 500 == 0):\n      print(\"Minibatch loss at step %d: %f\" % (step, l))\n      print(\"Minibatch accuracy: %.1f%%\" % accuracy(predictions, batch_labels))\n      print(\"Validation accuracy: %.1f%%\" % accuracy(\n        valid_prediction.eval(), valid_labels))\n      print(\"Test accuracy: %.1f%%\" % accuracy(test_prediction.eval(), test_labels))\n<\/code><\/pre>\n"}]}
{"QuestionId":33949968,"AnswerCount":1,"Tags":"<python><python-2.7><numpy><tensorflow><kaggle>","CreationDate":"2015-11-27T03:47:38.953","AcceptedAnswerId":null,"OwnerUserId":4941992.0,"Title":"Tensorflow multi-class ML model issues","Body":"<p>I have been trying to get tensor flow working on a multi-class kaggle problem. Basically the data consists of 6 features which I have converted to all numeric observations. The goal is to use these 6 features to predict a trip type, where there are 38 different trip types. I have been trying to use tensorflow to predict these trip type classes. The following code is what I have thus far, including what I had used to format the csv file. The code will run, but the output starts off ok for run 1, and then is very poor with the same output for the remainder of the runs. The following is the example of the output when it is running:<\/p>\n\n<pre><\/pre>\n\n<p>And the code:<\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts regarding what might be going wrong here as to why the runs would degenerate like this would be greatly appreciated. Thanks!<\/p>\n","answers":[{"AnswerId":"35448000","CreationDate":"2016-02-17T04:32:34.697","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>If you just want something up and running quickly for Kaggle competition, I would suggest you trying out <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/learn\/\" rel=\"nofollow noreferrer\">examples<\/a> in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/learn\/python\/learn\" rel=\"nofollow noreferrer\">TFLearn<\/a> first. There's embedding_ops for one-hot, examples for early stopping, custom decay, and more importantly, the multi-class classification\/regression you are encountering. Once you are more familiar with TensorFlow it would be fairly easy for you to insert TensorFlow code to build a custom model you want (there are also examples for this).  <\/p>\n"}]}
{"QuestionId":33953011,"AnswerCount":2,"Tags":"<torch>","CreationDate":"2015-11-27T08:20:38.223","AcceptedAnswerId":null,"OwnerUserId":5607520.0,"Title":"How to install Torch7 on windows7?","Body":"<p>Is there a way to install torch7 on Windows?\nI query a lot of information, but for transplant Torch on WINDOWS no idea,\n@Paul Kulchenko \nthank you very much<\/p>\n","answers":[{"AnswerId":"33958257","CreationDate":"2015-11-27T13:15:47.203","ParentId":null,"OwnerUserId":"2979342","Title":null,"Body":"<p>You can find many guides on google search <code>Cygwin<\/code>.<\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=TjxEH_tr7e0\" rel=\"nofollow\">https:\/\/www.youtube.com\/watch?v=TjxEH_tr7e0<\/a><br>\n<a href=\"https:\/\/www.youtube.com\/watch?v=68PEBV9NpL8\" rel=\"nofollow\">https:\/\/www.youtube.com\/watch?v=68PEBV9NpL8<\/a><br>\n<a href=\"https:\/\/www.physionet.org\/physiotools\/cygwin\/\" rel=\"nofollow\">https:\/\/www.physionet.org\/physiotools\/cygwin\/<\/a><\/p>\n"},{"AnswerId":"41924518","CreationDate":"2017-01-29T18:36:26.283","ParentId":null,"OwnerUserId":"5831205","Title":null,"Body":"<p>Better way is to run a virtual Linux environment and install torch on it.\nCygwin may not perform as well<\/p>\n"}]}
{"QuestionId":33955277,"AnswerCount":1,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2015-11-27T10:24:22.157","AcceptedAnswerId":33962786.0,"OwnerUserId":2802468.0,"Title":"How do I define my own operators in TensorFlow","Body":"<p>In TensorFlow, we can use  for doing L2 regularization. Let's say I want to define my own regularization operator for L1 regularization (call it ). How would I go about it? I am having a hard time locating operator definitions in the TensorFlow source code.<\/p>\n","answers":[{"AnswerId":"33962786","CreationDate":"2015-11-27T18:06:30.483","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As the comment suggested, there is a how-to guide for <a href=\"http:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/index.html\" rel=\"noreferrer\">adding an op to TensorFlow<\/a>. This guide covers adding a new op that is implemented in C++. In general, you should do this in the following situations:<\/p>\n\n<ul>\n<li>The op cannot be implemented using existing TensorFlow ops (for example, <code>l1_loss<\/code> <em>could<\/em> be implemented using the existing <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/math_ops.html#arithmetic-operators\" rel=\"noreferrer\">element-wise<\/a> and <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/math_ops.html#reduction\" rel=\"noreferrer\">reduction<\/a> operators as a Python function).<\/li>\n<li>A C++ implementation is necessary for performance (or memory consumption) reasons.<\/li>\n<li>The op could be implemented as a composition of ops, <strong>but<\/strong> it has a gradient that can be computed more efficiently (or with better numerical stability) than computing the gradients op-by-op. (This is why <code>tf.nn.l2_loss<\/code> is implemented as a fused op in C++.)<\/li>\n<\/ul>\n"}]}
{"QuestionId":33955368,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-27T10:28:48.473","AcceptedAnswerId":33961914.0,"OwnerUserId":867505.0,"Title":"Device placement unknown in Tensorboard","Body":"\n\n<p>I'd like to investigate device placement in the tensorboard using the following code for generating the graph in the summary<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This works for displaying the graph and summaries defined in the graph. But when selecting 'device placement' in the tensorboard, all nodes are assigned to 'unknown device'. Do I need to dump the device placement in some other way?<\/p>\n","answers":[{"AnswerId":"33961914","CreationDate":"2015-11-27T16:58:22.190","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>The TensorBoard graph visualizer only sees the explicit device assignments that you have made in your program (i.e. those made using <code>with tf.Device(\"...\"):<\/code> blocks).<\/p>\n\n<p>The reason for this is that the nodes in a TensorFlow graph are assigned to devices in multiple stages. The first stage, in the client (e.g. your Python program) allows you to explicitly\u2014and optionally\u2014assign devices to each node, and it is the output of this stage that is written to the TensorBoard logs. A later placement stage runs inside the TensorFlow backend, and assigns every node to a device.<\/p>\n\n<p>I suspect you want to analyze the results of the later placement stage. Currently there is no support for this in TensorBoard, but you can extract some information by creating the <code>tf.Session<\/code> as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess = tf.Session(config=tf.ConfigProto(\n    log_device_placement=True))\n<\/code><\/pre>\n\n<p>\u2026and then the device placement decisions will be logged to stderr.<\/p>\n"}]}
{"QuestionId":33955662,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-27T10:45:36.097","AcceptedAnswerId":33956190.0,"OwnerUserId":5611802.0,"Title":"change theano shared variable do not effect the value of internal","Body":"<p>i have a little problem with shared variable.  codes as below:<\/p>\n\n<pre><\/pre>\n\n<p><strong>i can understand the above.  but reverse can't work,as below:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>however,the below work:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>i cannot understand the logic of shared variable, hope to get help<\/p>\n","answers":[{"AnswerId":"33956190","CreationDate":"2015-11-27T11:16:21.333","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Answered on <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/HpBQcLgsmbE\" rel=\"nofollow\">theano-users mailing list<\/a>.<\/p>\n\n<p>A shared variable wraps a piece of memory but there is no guarantee that only the piece of memory used to initialize the shared variable will be used throughout the computation.<\/p>\n\n<p><code>borrow=True<\/code> should be thought of as a hint rather than an explicit instruction to be followed in all cases.<\/p>\n\n<p>In the first example the shared variable wraps the initial numpy array because <code>borrow=True<\/code>. Changing the numpy array thus changes the contents of the shared variable. But this should not be relied upon and is poor practice for Theano programming.<\/p>\n\n<p>In the second example <code>s_true+=1.0<\/code> is a symbolic expression. It does not actually change any state in memory other than making <code>s_true<\/code> point at an object representing the expression rather than the shared variable. <code>print s_true.eval()<\/code> shows the result of executing the symbolic computation but that does not change the shared variable. The only \"approved\" ways to change the contents of a shared variable are via <code>shared_var.set_value(...)<\/code> or via the <code>updates=...<\/code> mechanism of <code>theano.function(...)<\/code>. Changing the backing store, as in the first example, will sometimes work but not always and should generally be avoided.<\/p>\n\n<p>The third example is just a more roundabout way of doing the first example so is again bad practice.<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/aliasing.html\" rel=\"nofollow\">This page in the documentation<\/a> may help.<\/p>\n"}]}
{"QuestionId":33956625,"AnswerCount":0,"Tags":"<machine-learning><neural-network><theano><keras>","CreationDate":"2015-11-27T11:42:30.180","AcceptedAnswerId":null,"OwnerUserId":4497662.0,"Title":"Continuously adapting neural nets with production data?","Body":"<p>I have created a simple neural net (using keras) for my NLP application which classify data point into 1 of the 2 classes. \nMy model gave > 95% accuracy on Test data. I have deployed the model in production. Its running good. The prediction of the model is sent as an email to users. <\/p>\n\n<p>*These end users can send back a feedback if the prediction was right or not. I want to incorporate this feedback into the model so as continuously learn from this feedback (both positive and negative feedback) and improve * <\/p>\n\n<p><strong>How can I achieve the above ?<\/strong> I have read about Reinforcement Learning, Online Learning and Active Learning. \nAny pointers\/tutorials (possibly with code) will be of great help.<\/p>\n","answers":[]}
{"QuestionId":33956956,"AnswerCount":1,"Tags":"<python><cuda><tensorflow>","CreationDate":"2015-11-27T11:59:31.930","AcceptedAnswerId":34015393.0,"OwnerUserId":4652515.0,"Title":"Counter on GPU insanely slow compared to default counter?","Body":"<p>Edit -- see edit at the bottom, tensorflow on gpu is insanely faster for incrementing large vectors of counters.<\/p>\n\n<p>I'm trying to see if using the GPU is giving me any speed benefit, and this following program that simply counts 200,000 times, once using tensor-flow and the GPU and another time using plain-ol-python. The tensor-flow loop takes over 14 seconds to run while plain ol python takes only .013 seconds? What am I doing wrong? Code below :<\/p>\n\n<pre><\/pre>\n\n<p>which outputs this<\/p>\n\n<pre><\/pre>\n\n<p>EDIT -- after the suggestion to change to vectors of counters, tensorflow on gpu is unbelievably faster. <\/p>\n\n<p>With 10,000 counters per vector :<\/p>\n\n<pre><\/pre>\n\n<p>output :<\/p>\n\n<pre><\/pre>\n\n<p>with 100,000 counters, the output was :<\/p>\n\n<pre><\/pre>\n\n<p>the plain ol python took over a minute until I gave up<\/p>\n","answers":[{"AnswerId":"34015393","CreationDate":"2015-12-01T08:03:14.963","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In some sense, both programs are \"surprisingly\" slow, compared to the number of instructions that must be executed. The single-element counter is performing 200,000 increment instructions, using 200,000 calls to <code>sess.run()<\/code>, in 14.4 seconds. The vector counter is performing 100,000,000 increment instructions, using 10,000 calls to <code>sess.run()<\/code>, in 0.99 seconds. If you wrote these programs in C, you would expect to find that each counter increment takes a few nanoseconds at most, so where is the time being spent?<\/p>\n\n<p>TensorFlow imposes some per-step overhead, on the order of a few microseconds per call to <code>Session.run()<\/code>. This is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/120\" rel=\"nofollow\">known issue<\/a>, and it is something the team is trying to reduce, but it is rarely a concern for most of the neural network algorithms that one would typically run in a single step. The overhead can be broken down as follows:<\/p>\n\n<ul>\n<li><strong>Per-step dispatch overhead:<\/strong> The TensorFlow session API is string based, so some string manipulation and hashing has to be done to identify the correct subgraph to run in each step. This involves some Python and some C++ code.<\/li>\n<li><strong>Per-op dispatch overhead:<\/strong> This is implemented in C++, and involves setting up a context and dispatching a TensorFlow kernel. There are three ops in your counter benchmark (<code>VariableOp<\/code>, <code>Add<\/code>, and <code>Assign<\/code>).<\/li>\n<li><strong>GPU kernel dispatch overhead:<\/strong> Dispatching a kernel to the GPU involves a kernel entry to invoke the GPU driver.<\/li>\n<li><strong>GPU copy overhead:<\/strong> Perhaps surprisingly, <code>sess.run(update)<\/code> will copy the result back from the GPU, because <code>update<\/code> is a <code>Tensor<\/code> object (corresponding to the result of the assignment), and its value will be returned from the call.<\/li>\n<\/ul>\n\n<p>There are a few things you could try that should speed up both versions of the code.<\/p>\n\n<ul>\n<li><p>Using <code>state.assign_add(one)<\/code> instead of separate <code>tf.add<\/code> and <code>tf.assign<\/code> ops will cut down on the per-op dispatch overhead (and also do a more efficient in-place addition).<\/p><\/li>\n<li><p>Calling <code>sess.run(update.op)<\/code> will avoid the copy back to the client on each step.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":33957945,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2015-11-27T12:57:25.813","AcceptedAnswerId":33960296.0,"OwnerUserId":2191652.0,"Title":"Negative Padding in Convolution (Caffe)","Body":"<p>I have a vector which consists of a concatenation the features of a video sequence across 7 frames.<br>\nI would like to apply 1D convolution to this vector such that only a part of a frame is processed.<br>\nSay a feature vector for one frame has the length 10 <\/p>\n\n<ul>\n<li>my input feature vector would be of total length 7 x 10 = 70 <\/li>\n<\/ul>\n\n<p>Now I want two convolutions to work on different parts of that vector<\/p>\n\n<ul>\n<li>conv1 should treat the features 1:5   <\/li>\n<li>conv2 should treat 6:10  <\/li>\n<li>stride for both would be 10  <\/li>\n<li>so the convolution filters apply to the same features only in different frames<\/li>\n<\/ul>\n\n<p>Basically I would need to specify an offset for the second conv filter. Is that possible?\nOn the Caffe website they speak only about a zero padding, but for an offset I would need a negative padding.<\/p>\n\n<p>Is something like this possible?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"33960296","CreationDate":"2015-11-27T15:14:12.820","ParentId":null,"OwnerUserId":"2515628","Title":null,"Body":"<p>I think you can achieve this by using a <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html#slicing\" rel=\"nofollow\">slicing layer<\/a> instead of negative padding.<\/p>\n"}]}
{"QuestionId":33962226,"AnswerCount":4,"Tags":"<machine-learning><neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2015-11-27T17:23:30.443","AcceptedAnswerId":33980220.0,"OwnerUserId":4975126.0,"Title":"Common causes of nans during training","Body":"<p>I've noticed that a frequent occurrence during training is s being introduced.<\/p>\n\n<p>Often times it seems to be introduced by weights in inner-product\/fully-connected or convolution layers blowing up.<\/p>\n\n<p>Is this occurring because the gradient computation is blowing up? Or is it because of weight initialization (if so, why does weight initialization have this effect)? Or is it likely caused by the nature of the input data?<\/p>\n\n<p>The overarching question here is simply: <strong>What is the most common reason for NANs to occurring during training?<\/strong> And secondly, what are some methods for combatting this (and why do they work)?<\/p>\n","answers":[{"AnswerId":"33980220","CreationDate":"2015-11-29T06:50:29.820","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Good question.<br>\nI came across this phenomenon several times. Here are my observations:<\/p>\n\n<hr>\n\n<h3>Gradient blow up<\/h3>\n\n<p><strong>Reason:<\/strong> large gradients throw the learning process off-track.<\/p>\n\n<p><strong>What you should expect:<\/strong> Looking at the runtime log, you should look at the loss values per-iteration. You'll notice that the loss starts to grow <em>significantly<\/em> from iteration to iteration, eventually the loss will be too large to be represented by a floating point variable and it will become <code>nan<\/code>.<\/p>\n\n<p><strong>What can you do:<\/strong> Decrease the <code>base_lr<\/code> (in the solver.prototxt) by an order of magnitude (at least). If you have several loss layers, you should inspect the log to see which layer is responsible for the gradient blow up and decrease the <code>loss_weight<\/code> (in train_val.prototxt) for that specific layer, instead of the general <code>base_lr<\/code>.<\/p>\n\n<hr>\n\n<h3>Bad learning rate policy and params<\/h3>\n\n<p><strong>Reason:<\/strong> caffe fails to compute a valid learning rate and gets <code>'inf'<\/code> or <code>'nan'<\/code> instead, this invalid rate multiplies all updates and thus invalidating all parameters.<\/p>\n\n<p><strong>What you should expect:<\/strong> Looking at the runtime log, you should see that the learning rate itself becomes <code>'nan'<\/code>, for example:<\/p>\n\n<blockquote>\n<pre><code>... sgd_solver.cpp:106] Iteration 0, lr = -nan\n<\/code><\/pre>\n<\/blockquote>\n\n<p><strong>What can you do:<\/strong> fix all parameters affecting the learning rate in your <code>'solver.prototxt'<\/code> file.<br>\nFor instance, if you use <code>lr_policy: \"poly\"<\/code> and you forget to define <code>max_iter<\/code> parameter, you'll end up with  <code>lr = nan<\/code>...<br>\nFor more information about learning rate in caffe, see <a href=\"https:\/\/stackoverflow.com\/q\/30033096\/1714410\">this thread<\/a>.<\/p>\n\n<hr>\n\n<h3>Faulty Loss function<\/h3>\n\n<p><strong>Reason:<\/strong> Sometimes the computations of the loss in the loss layers causes <code>nan<\/code>s to appear. For example, Feeding <a href=\"https:\/\/stackoverflow.com\/a\/27645934\/1714410\"><code>InfogainLoss<\/code> layer with non-normalized values<\/a>, using custom loss layer with bugs, etc.<\/p>\n\n<p><strong>What you should expect:<\/strong> Looking at the runtime log you probably won't notice anything unusual: loss is decreasing gradually, and all of a sudden a <code>nan<\/code> appears. <\/p>\n\n<p><strong>What can you do:<\/strong> See if you can reproduce the error, add printout to the loss layer and debug the error.<\/p>\n\n<p>For example: Once I used a loss that normalized the penalty by the frequency of label occurrence in a batch. It just so happened that if one of the training labels did not appear in the batch at all - the loss computed produced <code>nan<\/code>s. In that case, working with large enough batches (with respect to the number of labels in the set) was enough to avoid this error.<\/p>\n\n<hr>\n\n<h3>Faulty input<\/h3>\n\n<p><strong>Reason:<\/strong> you have an input with <code>nan<\/code> in it!<\/p>\n\n<p><strong>What you should expect:<\/strong> once the learning process \"hits\" this faulty input - output becomes <code>nan<\/code>. Looking at the runtime log you probably won't notice anything unusual: loss is decreasing gradually, and all of a sudden a <code>nan<\/code> appears.<\/p>\n\n<p><strong>What can you do:<\/strong> re-build your input datasets (lmdb\/leveldn\/hdf5...) make sure you do not have bad image files in your training\/validation set. For debug you can build a simple net that read the input layer, has a dummy loss on top of it and runs through all the inputs: if one of them is faulty, this dummy net should also produce <code>nan<\/code>.<\/p>\n\n<hr>\n\n<h3>stride larger than kernel size in <code>\"Pooling\"<\/code> layer<\/h3>\n\n<p>For some reason, choosing <code>stride<\/code> > <code>kernel_size<\/code> for pooling may results with <code>nan<\/code>s. For example:<\/p>\n\n<pre><code>layer {\n  name: \"faulty_pooling\"\n  type: \"Pooling\"\n  bottom: \"x\"\n  top: \"y\"\n  pooling_param {\n    pool: AVE\n    stride: 5\n    kernel: 3\n  }\n}\n<\/code><\/pre>\n\n<p>results with <code>nan<\/code>s in <code>y<\/code>.<\/p>\n\n<hr>\n\n<h3>Instabilities in <code>\"BatchNorm\"<\/code><\/h3>\n\n<p>It was reported that under some settings <code>\"BatchNorm\"<\/code> layer may output <code>nan<\/code>s due to numerical instabilities.<br>\nThis <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/5136\" rel=\"noreferrer\">issue<\/a> was raised in bvlc\/caffe and <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/5136\" rel=\"noreferrer\">PR #5136<\/a> is attempting to fix it.<\/p>\n\n<hr>\n\n<p>Recently, I became aware of <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L86-L88\" rel=\"noreferrer\"><code>debug_info<\/code><\/a> flag: setting <code>debug_info: true<\/code> in <code>'solver.prototxt'<\/code> will make caffe print to log more debug information (including gradient magnitudes and activation values) during training: This information can <a href=\"https:\/\/stackoverflow.com\/q\/40510706\/1714410\">help in spotting gradient blowups and other problems in the training process<\/a>.<\/p>\n"},{"AnswerId":"34626903","CreationDate":"2016-01-06T06:25:16.310","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>This answer is not about a cause for <code>nan<\/code>s, but rather proposes a way to help debug it. \nYou can have this python layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class checkFiniteLayer(caffe.Layer):\n  def setup(self, bottom, top):\n    self.prefix = self.param_str\n  def reshape(self, bottom, top):\n    pass\n  def forward(self, bottom, top):\n    for i in xrange(len(bottom)):\n      isbad = np.sum(1-np.isfinite(bottom[i].data[...]))\n      if isbad&gt;0:\n        raise Exception(\"checkFiniteLayer: %s forward pass bottom %d has %.2f%% non-finite elements\" %\n                        (self.prefix,i,100*float(isbad)\/bottom[i].count))\n  def backward(self, top, propagate_down, bottom):\n    for i in xrange(len(top)):\n      if not propagate_down[i]:\n        continue\n      isf = np.sum(1-np.isfinite(top[i].diff[...]))\n        if isf&gt;0:\n          raise Exception(\"checkFiniteLayer: %s backward pass top %d has %.2f%% non-finite elements\" %\n                          (self.prefix,i,100*float(isf)\/top[i].count))\n<\/code><\/pre>\n\n<p>Adding this layer into your <code>train_val.prototxt<\/code> at certain points you suspect may cause trouble:<\/p>\n\n<pre><code>layer {\n  type: \"Python\"\n  name: \"check_loss\"\n  bottom: \"fc2\"\n  top: \"fc2\"  # \"in-place\" layer\n  python_param {\n    module: \"\/path\/to\/python\/file\/check_finite_layer.py\" # must be in $PYTHONPATH\n    layer: \"checkFiniteLayer\"\n    param_str: \"prefix-check_loss\" # string for printouts\n  }\n}\n<\/code><\/pre>\n"},{"AnswerId":"38477639","CreationDate":"2016-07-20T09:39:14.787","ParentId":null,"OwnerUserId":"2197678","Title":null,"Body":"<p>I was trying to build a sparse autoencoder and had several layers in it to induce sparsity. While running my net, I encountered the NaN's. On removing some of the layers (in my case, I actually had to remove 1), I found that the NaN's disappeared. So, I guess too much sparsity may lead to NaN's as well (some 0\/0 computations may have been invoked!?)<\/p>\n"},{"AnswerId":"45043871","CreationDate":"2017-07-11T20:24:44.937","ParentId":null,"OwnerUserId":"8291948","Title":null,"Body":"<p>In my case, not setting the bias in the convolution\/deconvolution layers was the cause. <\/p>\n\n<p><strong>Solution:<\/strong> add the following to the convolution layer parameters.<\/p>\n\n<p>bias_filler {\n      type: \"constant\"\n      value: 0\n    }<\/p>\n"}]}
{"QuestionId":33962959,"AnswerCount":1,"Tags":"<python><linear-algebra><tensorflow>","CreationDate":"2015-11-27T18:22:16.897","AcceptedAnswerId":33963330.0,"OwnerUserId":5613112.0,"Title":"Cholesky factor differentiation in TensorFlow","Body":"<p>I would like to get the gradient of  with respect to its input. As of the moment, the  does not have a registered gradient:\n<\/p>\n\n<pre><\/pre>\n\n<p>The code used to generate this error is:\n<\/p>\n\n<pre><\/pre>\n\n<p>While it is possible for me to compute the gradient myself and register it, the only existing means by which I've seen the Cholesky gradient computed involved <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/slinalg.py\" rel=\"nofollow\">the use of for loops and needs the shape of the input matrix.<\/a> However, to the best of my knowledge, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/208\" rel=\"nofollow\">symbolic loops aren't currently available to TensorFlow.<\/a><\/p>\n\n<p>One possible workaround to getting the shape of the input matrix  would probably be to use: \n<\/p>\n\n<pre><\/pre>\n\n<p>But this approach doesn't work if the dimensions of  is dependent on a TensorFlow placeholder object with shape . <\/p>\n\n<p>If anyone has any idea for how to compute and register a gradient of , I would very much appreciate knowing about it.<\/p>\n","answers":[{"AnswerId":"33963330","CreationDate":"2015-11-27T18:51:55.660","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>We discussed this a bit in the answers and comments to this question:  <a href=\"https:\/\/stackoverflow.com\/questions\/33709598\/tensorflow-cholesky-decomposition\">TensorFlow cholesky decomposition<\/a>.\nIt might (?) be possible to port the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/slinalg.html#theano.tensor.slinalg.CholeskyGrad\" rel=\"nofollow noreferrer\">Theano implementation of CholeskyGrad<\/a>, provided its semantics are actually what you want.  Theano's is based upon <a href=\"http:\/\/www.jstor.org\/stable\/1390762\" rel=\"nofollow noreferrer\">Smith's \"Differentiation of the Cholesky Algorithm\"<\/a>.<\/p>\n\n<p>If you implement it as a C++ operation that the Python just calls into, you have unrestricted access to all the looping constructs you could desire, and anything Eigen provides.  If you wanted to do it in pure tensorflow, you could use the control flow ops, such as <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3972c791b9f4d9a61b9ad6399b481df396f359ff\/tensorflow\/python\/ops\/control_flow_ops.py#L1077\" rel=\"nofollow noreferrer\"><code>tf.control_flow_ops.While<\/code><\/a> to loop.<\/p>\n\n<p>Once you know the actual formula you want to apply, the answer here:  <a href=\"https:\/\/stackoverflow.com\/questions\/33714832\/matrix-determinant-differentiation-in-tensorflow\">matrix determinant differentiation in tensorflow<\/a>\nshows how to implement and register a gradient for an op in tensorflow.<\/p>\n\n<p>You could also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">create an issue on github<\/a> to request this feature, though, of course, you'll probably get it faster if you implement it yourself and then send in a pull request. :)<\/p>\n"}]}
{"QuestionId":33966549,"AnswerCount":0,"Tags":"<tensorflow><bazel>","CreationDate":"2015-11-28T00:55:01.560","AcceptedAnswerId":null,"OwnerUserId":988717.0,"Title":"Building Tensorflow with Bazel: Executing genrule @png_archive\/\/:configure failed","Body":"\n\n<p>I've been trying to understand why I can't build Tensorflow with Bazel.\nI've been following the install instructions from the Tensorflow webpage <a href=\"http:\/\/www.tensorflow.org\/get_started\/os_setup.html#binary-installation\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Here's my output when I build with the command <\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>It's a bit cryptic.  Bash just fails with the following message which doesn't really provide a whole lot of insight.  <\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>Has anyone else come across this error when compiling with bazel?<\/p>\n\n<h2>EDIT: extra info<\/h2>\n\n<p>OS is Fedora23 (kernel 4.2.6-300.fc23)<\/p>\n\n<p>gcc version is 5.1.1 20150618 (Red Hat 5.1.1-4) (GCC) <\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>ldd version is: (GNU libc) 2.22<\/p>\n","answers":[]}
{"QuestionId":33967634,"AnswerCount":1,"Tags":"<neural-network><tensorflow>","CreationDate":"2015-11-28T04:18:04.220","AcceptedAnswerId":33968250.0,"OwnerUserId":2131871.0,"Title":"Why does this variant of the XOR function not always converge?","Body":"<p>I'm trying to implement a simple XOR gate in TensorFlow. My problem is that my function doesn't always converge.<\/p>\n\n<p>If I'm not wrong, XOR space don't have local minimum, so I don't understand why this would happen.<\/p>\n\n<p>--<\/p>\n\n<p>I saw this answer: <a href=\"https:\/\/stackoverflow.com\/a\/33750395\/2131871\">https:\/\/stackoverflow.com\/a\/33750395\/2131871<\/a>, and it always converge.\nI have taken the code from the answer by @mrry and slightly modified it, so that instead of having two output nodes, it only have one, I used tanh activation function instead of relu &amp; softmax and adapted the cross_entropy function.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Can anybody explain me why my solution sometimes fail to converge? Thanks.<\/p>\n","answers":[{"AnswerId":"33968250","CreationDate":"2015-11-28T06:04:10.277","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The way you're computing your error is letting your network fall into local minima too easily.  I suspect it's because the l1 norm of the xor function has too many equal-weight poor solutions when moving from an existing solution.  (But I'm not positive - an ML expert can give you a more precise answer here.  I'm just a systems schmoo.)<\/p>\n\n<p>Easy fix:  replace these lines:<\/p>\n\n<pre><code>cross_entropy = tf.abs(tf.sub(y_input, y))\nloss = tf.reduce_mean(cross_entropy)\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>loss = tf.nn.l2_loss(y_input - y)\n<\/code><\/pre>\n"}]}
{"QuestionId":33967757,"AnswerCount":1,"Tags":"<python-2.7><build><tensorflow><bazel>","CreationDate":"2015-11-28T04:39:07.837","AcceptedAnswerId":null,"OwnerUserId":2423354.0,"Title":"Building TensorFlow: libpython2.7.so.1.0: cannot open shared object file","Body":"<p>When building the GPU enabled version of TensorFlow from source using the command <\/p>\n\n<pre><\/pre>\n\n<p>I am getting the error: <\/p>\n\n<pre><\/pre>\n\n<p>The .so file is in a directory pointed to in my LD_LIBRARY_PATH:<\/p>\n\n<pre><\/pre>\n\n<p>To provide a little more info, here's the entire error message that I get:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42989176","CreationDate":"2017-03-24T00:07:39.390","ParentId":null,"OwnerUserId":"6088060","Title":null,"Body":"<p>Would you be able to build the non-Cuda version of the TensorFlow? It would tell us whether it is your python environment that is somehow broken for TensorFlow.<\/p>\n"}]}
{"QuestionId":33968233,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-28T06:01:07.617","AcceptedAnswerId":33969338.0,"OwnerUserId":2715268.0,"Title":"How to perform conditional updates on shared variables in Theano?","Body":"<p>Is there a way to conditionally update the shared variables depending on the  result of the current function.\nEg.<\/p>\n\n<pre><\/pre>\n\n<p>In this model, I need to update the weight parameter only if the cost is greater than 0. There is a no_default_updates parameter in the function but it doesn't apply to 'updates' parameter.<\/p>\n","answers":[{"AnswerId":"33969338","CreationDate":"2015-11-28T08:35:30.543","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use a symbolic conditional operation. Theano has two: <code>switch<\/code> and <code>ifelse<\/code>. <code>switch<\/code> is performed element-wise while <code>ifelse<\/code> works more like a conventional conditional. See the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html\" rel=\"nofollow\">documentation<\/a> for more.<\/p>\n\n<p>Here's an example that updates parameters only when the cost is positive.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile(input_size, hidden_size, output_size, learning_rate):\n    w_h = theano.shared(numpy.random.standard_normal((input_size, hidden_size))\n                        .astype(theano.config.floatX), name='w_h')\n    b_h = theano.shared(numpy.random.standard_normal((hidden_size,))\n                        .astype(theano.config.floatX), name='b_h')\n    w_y = theano.shared(numpy.random.standard_normal((hidden_size, output_size))\n                        .astype(theano.config.floatX), name='w_y')\n    b_y = theano.shared(numpy.random.standard_normal((output_size,))\n                        .astype(theano.config.floatX), name='b_y')\n    x = tt.matrix()\n    z = tt.vector()\n    h = tt.tanh(theano.dot(x, w_h) + b_h)\n    y = theano.dot(h, w_y) + b_y\n    c = tt.sum(y - z)\n    updates = [(p, p - tt.switch(tt.gt(c, 0), learning_rate * tt.grad(cost=c, wrt=p), 0))\n               for p in (w_h, b_h, w_y, b_y)]\n    return theano.function([x, z], outputs=c, updates=updates)\n\n\ndef main():\n    f = compile(input_size=3, hidden_size=2, output_size=4, learning_rate=0.01)\n\n\nmain()\n<\/code><\/pre>\n\n<p>In this case either <code>switch<\/code> or <code>ifelse<\/code> could be used but <code>switch<\/code> is generally preferable in such cases because <code>ifelse<\/code> does not appear to be as well supported throughout the Theano framework and requires a special import.<\/p>\n"}]}
{"QuestionId":33969059,"AnswerCount":2,"Tags":"<python><neural-network><theano><keras>","CreationDate":"2015-11-28T07:57:29.503","AcceptedAnswerId":null,"OwnerUserId":4652515.0,"Title":"Trying to get simple Keras neural net example to work","Body":"<p>I've been fooling-around trying to get simple examples that I create working, because I find the examples given with large complicated datasets to be hard to grasp intuitively. The program below takes a list of weights  and uses them to create a random scattering of points on a plane with some random noise added in. I then train the simple neural nets on this data and check the results.<\/p>\n\n<p>When I do this with the Graph models everything works perfectly, the loss score goes down to zero predictably as the model converges on the weights given. However, when I try to use a sequential model, nothing happens. Code below<\/p>\n\n<p>If you want I can post my other script that uses the Graph instead of sequential and show that it finds the input weights perfectly.<\/p>\n\n<pre><\/pre>\n\n<p>The output is as follows<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36970749","CreationDate":"2016-05-01T19:05:09.890","ParentId":null,"OwnerUserId":"1397061","Title":null,"Body":"<p>The reason for the huge error is that your labels are not binary and very large, but the output of softmax is binary.  For instance, if the label is 10000 but you can only predict something between 0 and 1, there will be huge error regardless of what you predict.  Did you mean <code>activation='linear'<\/code> in the last layer, to do a regression?  Or did you want to put your labels through a softmax at the end of <code>get_data()<\/code>?<\/p>\n"},{"AnswerId":"41529700","CreationDate":"2017-01-08T05:10:37.460","ParentId":null,"OwnerUserId":"7388421","Title":null,"Body":"<p>Because your y_train is consist of 5 element, your output model should be 5 element too, <\/p>\n\n<pre><code>[ 0.06698217  0.70033048  0.4317502   0.78504855  0.26173543] --&gt; [ 1.]\n[ 0.28940025  0.21746189  0.93097653  0.94885535  0.56790348] --&gt; [ 1.]\n[ 0.69430499  0.1622601   0.22802859  0.75709315  0.88948355] --&gt; [ 1.]\n<\/code><\/pre>\n\n<p>for example try this network<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(INDIM, input_dim=INDIM, init='uniform', activation='tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10, init='uniform', activation='tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(5, init='uniform', activation='softmax'))\n<\/code><\/pre>\n"}]}
{"QuestionId":33970755,"AnswerCount":3,"Tags":"<gpu><tensorflow>","CreationDate":"2015-11-28T11:26:01.743","AcceptedAnswerId":null,"OwnerUserId":3298756.0,"Title":"Tensorflow not using GPU","Body":"<p>I'm trying to run the example seq2seq by Tensorflow, but it won't use the GPU. Here are the step I took to install Tensorflow on a Linux system with Tesla K20x<\/p>\n\n<pre><\/pre>\n\n<p>After all of this step, I have tensorflow installed. I then try to run the seq2seq example, by<\/p>\n\n<pre><\/pre>\n\n<p>but it will not use the GPU. I then try the example <\/p>\n\n<pre><\/pre>\n\n<p>and it gives an error     <\/p>\n\n<pre><\/pre>\n\n<p>Was wondering if anyone know what might cause the program to not use the GPU? Any help is appreciated.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"34341643","CreationDate":"2015-12-17T18:23:11.083","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue looks to be that when you <code>bazel run<\/code> the translation example, it rebuilds without GPU support. Try adding <code>--config=cuda<\/code> to the <code>bazel run<\/code> command, as follows:<\/p>\n\n<pre><code>$ bazel run -c opt --config=cuda \/\/tensorflow\/models\/rnn\/translate:translate\n<\/code><\/pre>\n\n<p>Without this option, Bazel will recompile the entire TensorFlow runtime without GPU support, and use this version when it runs the example application.<\/p>\n"},{"AnswerId":"38684892","CreationDate":"2016-07-31T14:17:12.277","ParentId":null,"OwnerUserId":"5035995","Title":null,"Body":"<p>it occurs since the cuda is not properly linked. Enter the following command in the terminal<\/p>\n\n<p><code>sudo ldconfig \/usr\/local\/cuda\/lib64<\/code><\/p>\n"},{"AnswerId":"46989968","CreationDate":"2017-10-28T12:55:46.703","ParentId":null,"OwnerUserId":"7124519","Title":null,"Body":"<p>I\u2019m guessing you should install it with a GPU version <\/p>\n\n<pre><code>pip install tensorflow-gpu\n<\/code><\/pre>\n"}]}
{"QuestionId":33974231,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2015-11-28T17:27:16.693","AcceptedAnswerId":33974372.0,"OwnerUserId":1076336.0,"Title":"Tensorflow error using my own data","Body":"<p>I've been playing with the Tensorflow library doing the tutorials. Now I wanted to play with my own data, but I fail horribly. This is perhaps a noob question but I can't figure it out.<\/p>\n\n<p>I'm using this example: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/convolutional_network.py\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/convolutional_network.py<\/a><\/p>\n\n<p>I want to use my own images, for converting my images to use with tensorflow i'm using this: <a href=\"https:\/\/github.com\/HamedMP\/ImageFlow\/blob\/master\/ImageFlow.py\">https:\/\/github.com\/HamedMP\/ImageFlow\/blob\/master\/ImageFlow.py<\/a><\/p>\n\n<p>Now I change the parameters in the example from this:<\/p>\n\n<pre><\/pre>\n\n<p>to this:<\/p>\n\n<pre><\/pre>\n\n<p>I did that because my images are 96 * 96 and there are only 2 classes of my images<\/p>\n\n<p>I also change the weights and biases to the numbers I need.<\/p>\n\n<p>I read the data like this:<\/p>\n\n<pre><\/pre>\n\n<p>imgReader being the ImageFlow file<\/p>\n\n<p>but when I try to run it I gives me an error:<\/p>\n\n<pre><\/pre>\n\n<p>I feel like i'm overlooking something small but I don't see it.<\/p>\n","answers":[{"AnswerId":"33974372","CreationDate":"2015-11-28T17:40:05.677","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This error arises because the shape of the data that you're trying to feed (104 x 96 x 96 x 1) does not match the shape of the input placeholder (<code>batch_size<\/code> x 9216, where <code>batch_size<\/code> may be variable).<\/p>\n\n<p>To make it work, add the following line before running a training step:<\/p>\n\n<pre><code>batch_xs = np.reshape(batch_xs, (-1, 9216))\n<\/code><\/pre>\n\n<p>This uses numpy to reshape the images read in, which are 4-D arrays of <code>batch_size<\/code> x h x w x channels, into a <code>batch_size<\/code> x 9216 element matrix as expected by the placeholder.<\/p>\n"},{"AnswerId":"40895995","CreationDate":"2016-11-30T18:51:10.130","ParentId":null,"OwnerUserId":"3291563","Title":null,"Body":"<p>I solved this problem by upgrading tensorflow through pip.<\/p>\n"}]}
{"QuestionId":33977273,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-11-28T22:41:29.733","AcceptedAnswerId":null,"OwnerUserId":5616368.0,"Title":"No module named tensorflow.python.platform","Body":"<p>I am running . I get:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"56139057","CreationDate":"2019-05-14T21:46:06.310","ParentId":null,"OwnerUserId":"9612929","Title":null,"Body":"<p>Does <code>import tensorflow<\/code> work in a Python shell?  <\/p>\n\n<p>If not, try reinstalling with <code>pip install --upgrade --force-reinstall &lt;package&gt;<\/code>.  <\/p>\n\n<p>For example, for TensorFlow version 1.12 with GPU support: <code>pip install --upgrade --force-reinstall tensorflow-gpu==1.12.0<\/code><\/p>\n"},{"AnswerId":"42198267","CreationDate":"2017-02-13T06:59:08.693","ParentId":null,"OwnerUserId":"7229161","Title":null,"Body":"<p>Try this, it worked for me:<\/p>\n\n<pre><code>sudo pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":33978421,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><xor><tensorflow>","CreationDate":"2015-11-29T01:27:32.840","AcceptedAnswerId":33979537.0,"OwnerUserId":3314143.0,"Title":"Get a simple MLP in TensorFlow to model XOR","Body":"<p>I tried to build a simple MLP with an input layer (2 neurons), a hidden layer (5 neurons) and an output layer (1 neuron). I planned to train and feed it with  for getting the desired output of  (elementwise).<\/p>\n\n<p>Unfortunately my code refuses to run. I keep getting dimensionality errors no matter what I'm trying. Quite frustrating :\/ I think I'm missing something but I can not figure out what is wrong.<\/p>\n\n<p>For better readability I also uploaded the code to a pastebin: <a href=\"http:\/\/pastebin.com\/GpH7WEEi\" rel=\"nofollow\">code<\/a><\/p>\n\n<p>Any ideas?<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong> I am still getting errors :\/<\/p>\n\n<pre><\/pre>\n\n<p>outputs . Altering the line to:<\/p>\n\n<pre><\/pre>\n\n<p>seems to be working, but then the error appears in:<\/p>\n\n<pre><\/pre>\n\n<p>telling me: <\/p>\n\n<p>Turning the statement to:<\/p>\n\n<pre><\/pre>\n\n<p>also throws an exception: .<\/p>\n\n<p><strong>EDIT2:<\/strong> I do not really understand this. Shouldn't  be , since in dimensions this would be ? If I transpose  hidden is still working (I even don't get the point why it is working without a transpose at all). However,  keeps throwing errors but this operation in dimensions should be ?!<\/p>\n","answers":[{"AnswerId":"33979537","CreationDate":"2015-11-29T04:51:32.790","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>(0)  It's helpful to include the error output - it's also a useful thing to look at,  because it does identify exactly where you were having shape problems.<\/p>\n\n<p>(1)  The shape errors arose because you have the arguments to matmul backwards in both of your matmuls, and have the tf.Variable backwards.  The general rule is that the weights for layer that has <code>input_size, output_size<\/code> should be <code>[input_size, output_size]<\/code>, and the matmul should be <code>tf.matmul(input_to_layer, weights_for_layer)<\/code> (and then add the biases, which are of shape <code>[output_size]<\/code>).<\/p>\n\n<p>So with your code,<\/p>\n\n<pre><code>W_hidden = tf.Variable(tf.random_uniform([hidden_nodes, 2], -1.0, 1.0))\n<\/code><\/pre>\n\n<p>should be:<\/p>\n\n<pre><code>W_hidden = tf.Variable(tf.random_uniform([2, hidden_nodes], -1.0, 1.0))\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>hidden = tf.sigmoid(tf.matmul(W_hidden, n_input) + b_hidden)\n<\/code><\/pre>\n\n<p>should be <code>tf.matmul(n_input, W_hidden)<\/code>; and<\/p>\n\n<pre><code>output = tf.sigmoid(tf.matmul(W_output, hidden))\n<\/code><\/pre>\n\n<p>should be <code>tf.matmul(hidden, W_output)<\/code><\/p>\n\n<p>(2)  Once you've fixed those bugs, your run needs to be fed a feed_dict:<\/p>\n\n<pre><code>sess.run(train)\n<\/code><\/pre>\n\n<p>should be:<\/p>\n\n<pre><code>sess.run(train, feed_dict={n_input: input_data}) \n<\/code><\/pre>\n\n<p>At least, I presume that this is what you're trying to achieve.<\/p>\n"}]}
{"QuestionId":33979259,"AnswerCount":1,"Tags":"<docker><tensorflow><tensorboard>","CreationDate":"2015-11-29T03:58:50.397","AcceptedAnswerId":null,"OwnerUserId":5616704.0,"Title":"View Tensorboard via Chorme on Windows","Body":"<p>I used Docker-based installation.<\/p>\n\n<pre><\/pre>\n\n<p>I started Tensorboard like this:<\/p>\n\n<pre><\/pre>\n\n<p>Then, I opened Chrome. I got TensorBoard on the title, but it was empty.<\/p>\n\n<p>Here is some pictures: <a href=\"http:\/\/imgur.com\/a\/ZwdzD\" rel=\"nofollow\">http:\/\/imgur.com\/a\/ZwdzD<\/a><\/p>\n\n<p>How to view tensorboard on Docker-based installation correctly?<\/p>\n","answers":[{"AnswerId":"33996336","CreationDate":"2015-11-30T10:26:09.610","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p>I am having a similar problem. \n<a href=\"https:\/\/stackoverflow.com\/questions\/33836728\/view-tensorboard-on-docker-on-google-cloud\">View Tensorboard on Docker on Google Cloud<\/a><\/p>\n\n<p>I used:<\/p>\n\n<p><code>docker run -p 127.0.0.1:$HOSTPORT:$6006 --name ai-unicorn -t XXXXXXXXXXXX<\/code><\/p>\n\n<p>(The image number is replaced by X... here)<\/p>\n"}]}
{"QuestionId":33979360,"AnswerCount":1,"Tags":"<google-app-engine><lua><google-compute-engine><torch>","CreationDate":"2015-11-29T04:15:57.420","AcceptedAnswerId":null,"OwnerUserId":1373421.0,"Title":"Access compute engine from app engine","Body":"<p>I have an instance running on the Compute Engine which uses Torch to predict objects in images. I wanted to make a simple web interface using which a user can upload an image, the image is sent to the server(compute engine), the objects are predicted and the list is returned back to the user. \nIn my compute engine (Ubuntu 14.04) this line of code is used to predict objects in images. (All the other setup has been already done in the compute engine.)<\/p>\n\n<pre><\/pre>\n\n<p>I want to call this line from the web app and pass the image to the image folder and get back the list of objects. How do I go about it?<\/p>\n","answers":[{"AnswerId":"33981964","CreationDate":"2015-11-29T10:54:39.183","ParentId":null,"OwnerUserId":"1037753","Title":null,"Body":"<p>In past projects I have discussed and used different approaches to communicate between Google App Engine and Google Compute Engine. Generally speaking the two usual suspects are:<\/p>\n\n<ul>\n<li><strong>Orchestration from App Engine<\/strong>: In this approach the App Engine application is the active part and sends requests to a service on the compute instance. This is what <a href=\"https:\/\/stackoverflow.com\/users\/222467\/\">Igor Artamonov<\/a> already described in his comment. We used a tomcat instance on the compute instances which ran a full rest api to invoke commands on the instance. Possible helpers:\n\n<ul>\n<li>When using the <a href=\"https:\/\/cloud.google.com\/compute\/docs\/reference\/latest\/\" rel=\"nofollow noreferrer\">Google Compute API<\/a> from App Engine you can get the external IP address of compute instances. So you know where your requests will have to go.<\/li>\n<\/ul><\/li>\n<li><strong>Polling from the compute instance<\/strong>: Since you know the app id of your App Engine application you could code a simple loop on your compute instances that request new jobs from the app engine application. I have used this approach in combination with an orchestration that will send a shutdown command to instances that are no longer required, therefor reducing the polling load on app engine. If new jobs were created I would start a new compute instance which would then poll until it receives a shutdown command again.<\/li>\n<\/ul>\n\n<p>Both approaches work well. If you use the Compute API and know the IPs of your compute instances you can restrict your polling endpoints and command invoke requests to these IPs for basic security.<\/p>\n\n<p>I would try to avoid too much polling though since, well let me give you a quote:<\/p>\n\n<blockquote>\n  <p>Actively polling is the poor man's solution to kicking off a workflow process. (<a href=\"http:\/\/www.javaworld.com\/article\/2071865\/web-app-frameworks\/use-spring-to-create-a-simple-workflow-engine.html?page=2\" rel=\"nofollow noreferrer\">javaworld.com<\/a>)<\/p>\n<\/blockquote>\n\n<p>But if you shut down your compute instances when they are finished with their workload i don't see a good reason why you shouldn't use polling. If you don't and you increase the number of compute instances to a couple instances you will have load on your App Engine application without achieving anything but cost.<\/p>\n"}]}
{"QuestionId":33979474,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-11-29T04:41:16.433","AcceptedAnswerId":34253082.0,"OwnerUserId":4012405.0,"Title":"Recurrent Mixture Density Networks for Tensorflow?","Body":"<p>Is there an implementation of Schuster's <a href=\"http:\/\/homepages.inf.ed.ac.uk\/amos\/pmrg\/schuster.pdf\" rel=\"nofollow\">Recurrent Mixture Density Networks<\/a> for Tensorflow yet?  They were used by Alex Graves to great effect in his \"<a href=\"http:\/\/arxiv.org\/abs\/1308.0850\" rel=\"nofollow\">Generating Sequences With Recurrent Neural Networks<\/a>\" paper and <a href=\"http:\/\/www.cs.toronto.edu\/~graves\/handwriting.html\" rel=\"nofollow\">handwriting generation demo<\/a>.<\/p>\n","answers":[{"AnswerId":"34057514","CreationDate":"2015-12-03T03:46:10.880","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>Not yet, but good point. Recurrent neural nets are in TensorFlow, if you combine those with a loss that maximizes likelihood of a Gaussian distribution then you can set this up. We haven't tried this yet but theoretically it should be possible with what is there now -- however, it may be possible to simplify it by having a special loss for this case (which we currently don't have).<\/p>\n"},{"AnswerId":"34253082","CreationDate":"2015-12-13T16:00:51.947","ParentId":null,"OwnerUserId":"4012405","Title":null,"Body":"<p>I looks like someone <a href=\"http:\/\/blog.otoro.net\/2015\/12\/12\/handwriting-generation-demo-in-tensorflow\/\" rel=\"nofollow\">made a version of the handwriting generation code in Tensorflow<\/a>.  It includes <a href=\"https:\/\/github.com\/hardmaru\/write-rnn-tensorflow\" rel=\"nofollow\">code<\/a> for a 2D Recurrent Mixture Density Network.<\/p>\n"}]}
{"QuestionId":33980109,"AnswerCount":1,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2015-11-29T06:34:00.550","AcceptedAnswerId":33993325.0,"OwnerUserId":5616984.0,"Title":"undefined symbol: clock_gettime with tensorflow on ubuntu14.04","Body":"<p>I have installed tensorflow from source with bazel. But when I import tensorflow in python,the error happens.<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix it now?<\/p>\n","answers":[{"AnswerId":"33993325","CreationDate":"2015-11-30T07:20:25.413","ParentId":null,"OwnerUserId":"3728184","Title":null,"Body":"<p>Try adding the <code>rt<\/code> library during the linking step, like in this diff:<\/p>\n\n<pre><code>--- a\/tensorflow\/tensorflow.bzl\n+++ b\/tensorflow\/tensorflow.bzl\n@@ -284,7 +284,7 @@ _py_wrap_cc = rule(attrs={\n\n def tf_extension_linkopts():\n-  return []  # No extension link opts\n+  return [\"-lrt\"]\n<\/code><\/pre>\n\n<p>See also this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/121\" rel=\"noreferrer\">github issue<\/a>.<\/p>\n"}]}
{"QuestionId":33980496,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-11-29T07:35:39.663","AcceptedAnswerId":null,"OwnerUserId":3035447.0,"Title":"tensorflow rnn model path","Body":"<p>I have trained the language model using Tensorflow as given in thie <a href=\"http:\/\/www.tensorflow.org\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">tutorial<\/a><\/p>\n\n<p>For training I used the following command.<\/p>\n\n<pre><\/pre>\n\n<p>The training was successful with the following o\/p at the end.<\/p>\n\n<pre><\/pre>\n\n<p>But I am still confused with, where the training model is stored and how to use it.<\/p>\n","answers":[{"AnswerId":"33981028","CreationDate":"2015-11-29T08:56:03.343","ParentId":null,"OwnerUserId":"1667256","Title":null,"Body":"<p>The demo code probably did not include the ability to save a model; you might want to explicitly use <code>tf.train.Saver<\/code> to save and restore variables to and from checkpoints. <\/p>\n\n<p>See <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow\">doc<\/a> and <a href=\"http:\/\/www.tensorflow.org\/how_tos\/variables\/index.html#saving-and-restoring\" rel=\"nofollow\">examples<\/a>.<\/p>\n\n<p>It's pretty straightforward according to the doc. In below example, I saved\nall the variables in the model. Instead, you can choose which variable(s) to save by following the <a href=\"http:\/\/www.tensorflow.org\/how_tos\/variables\/index.html#saving-and-restoring\" rel=\"nofollow\">examples<\/a>. <\/p>\n\n<pre><code># ... \ntf.initialize_all_variables().run()\n####################################################\n# Add ops to save and restore all the variables.\n####################################################\nsaver = tf.train.Saver()\n\nfor i in range(config.max_max_epoch):\n  lr_decay = config.lr_decay ** max(i - config.max_epoch, 0.0)\n  m.assign_lr(session, config.learning_rate * lr_decay)\n\n  print(\"Epoch: %d Learning rate: %.3f\" % (i + 1, session.run(m.lr)))\n  train_perplexity = run_epoch(session, m, train_data, m.train_op,\n                               verbose=True)\n  print(\"Epoch: %d Train Perplexity: %.3f\" % (i + 1, train_perplexity))\n  valid_perplexity = run_epoch(session, mvalid, valid_data, tf.no_op())\n  print(\"Epoch: %d Valid Perplexity: %.3f\" % (i + 1, valid_perplexity))\n\n  ####################################################\n  # Save the variables to disk.\n  ####################################################\n  save_path = saver.save(session, \"\/tmp\/model.epoch.%03d.ckpt\" % (i + 1))\n  print(\"Model saved in file: %s\" % save_path)\n  # ....\n<\/code><\/pre>\n\n<p>In my case, each checkpoint file has a disk size of 18.61M (<code>--model small<\/code>). <\/p>\n\n<p>Regarding how to use the model, just follow the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow\">doc<\/a> to restore the checkpoints from saved files. Then it's at your will how to use it. <\/p>\n"}]}
{"QuestionId":33981877,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2015-11-29T10:45:42.807","AcceptedAnswerId":36491430.0,"OwnerUserId":1667256.0,"Title":"See the contents of Checkpoint files?","Body":"<p>Per <a href=\"http:\/\/www.tensorflow.org\/how_tos\/variables\/index.html#saving-and-restoring\" rel=\"nofollow\">documentation<\/a>, variables in a session can be saved and restored to\/from a binary file with  object. <\/p>\n\n<p>But is there any way to see the content of the binary file?<\/p>\n","answers":[{"AnswerId":"36491430","CreationDate":"2016-04-08T04:16:06.467","ParentId":null,"OwnerUserId":"1667256","Title":null,"Body":"<p>A checkpoint file is an sstable. The value for each record is a serialized\n<code>SavedTensorSlices<\/code> message. (resource <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9c3043ff3bf31a6a81810b4ce9e87ef936f1f529\/tensorflow\/core\/util\/saved_tensor_slice.proto#L4\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<p>To see the content of the serialized <code>SavedTensorSlices<\/code> message, we just unerialize the content into a <code>SavedTensorSlices<\/code> object. Something like below:<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>SavedTensorSlices message;\nmessage.ParseFromString(value);\ncout &lt;&lt; message.DebugString();\n<\/code><\/pre>\n"},{"AnswerId":"34003567","CreationDate":"2015-11-30T16:42:41.370","ParentId":null,"OwnerUserId":"998293","Title":null,"Body":"<p>The files are read\/written using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9c3043ff3bf31a6a81810b4ce9e87ef936f1f529\/tensorflow\/core\/util\/tensor_slice_reader.cc\" rel=\"nofollow\">TensorSliceReader<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9c3043ff3bf31a6a81810b4ce9e87ef936f1f529\/tensorflow\/core\/util\/tensor_slice_writer.h\" rel=\"nofollow\">TensorSliceWriter<\/a> in C++ using what seems to be a special format for tensor data.<\/p>\n\n<p>The files contain the values of the saved tensors. The simplest way to inspect those values would be to restore the tensors from the checkpoint file and inspect the tensors directly. <\/p>\n"}]}
{"QuestionId":33982697,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2015-11-29T12:21:10.383","AcceptedAnswerId":null,"OwnerUserId":5617669.0,"Title":"Features protobuf performance","Body":"<p>I'm playing with MNIST examples and noticed that protobuf serialized features are horribly slow both serialize and deserialize.<\/p>\n\n<p>My simple test reads CSV with 42000 images and writes it into binary file using TFRecordWriter. Benchmark results are quite surprising:<\/p>\n\n<ul>\n<li>pickle of image as string: 19.292 seconds, size=130MB<\/li>\n<li>Example\/Features encoding, byte array: 108.510 seconds, size=100MB<\/li>\n<li>Example\/Features encoding, int64 array: 145.014 seconds, size=39MB<\/li>\n<\/ul>\n\n<p>With quite plausible size results, looks like protobuf feature encoding is very slow. I see the same slow results on reading, Example.FromString() is ~5-7 times slower than pickle.<\/p>\n\n<p>Are there tricks\/suggestions how to overcome this? <\/p>\n\n<p>My encoding code snippet is below.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":33983026,"AnswerCount":0,"Tags":"<python><visual-studio><deep-learning><caffe><pycaffe>","CreationDate":"2015-11-29T12:58:26.440","AcceptedAnswerId":null,"OwnerUserId":5389418.0,"Title":"In caffe installation, where to set python path in vs studio","Body":"<p>Hi Im trying to install caffe deep learning tool box in windows. <\/p>\n\n<p>I have to use it in python. I have so far completed first 2 steps mentioned in <\/p>\n\n<p>For python wrapper, it says\n  .\nI dont know where to set python include and library path in vs studio 2015.<\/p>\n\n<p>I have also run  in  folder. It gives the error . I'm trying to install caffe for python in windows for several hours. Please help. A step by step simple procedure to install caffe on windows and use it in python will be highly appreciated.<\/p>\n","answers":[]}
{"QuestionId":33983528,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2015-11-29T13:50:05.397","AcceptedAnswerId":null,"OwnerUserId":2476373.0,"Title":"How to change the learning rate of specific layer from the solver prototxt (CAFFE)","Body":"<p>Anybody knows how to change the learning rate  of a specific layer in CAFFE from the solver prototxt? I know there's , however I would like to target the rate of a specific layer, and doing it from the solver instead of the network prototxt.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"33996888","CreationDate":"2015-11-30T10:51:47.020","ParentId":null,"OwnerUserId":"3509999","Title":null,"Body":"<p>Every layer that requiers learning (i.e convultional, fully-connected, etc.) has a specific <code>lr_mult<\/code> parameter that can be controlled specifically for that layer. <code>lr_mult<\/code> is a \"multiplier on the global learning rate for this parameter.\"<\/p>\n\n<p>Simply define or change the <code>lr_mult<\/code> for your layer in <code>train_val.prototxt<\/code>.<\/p>\n\n<p>This is useful for fine-tuning, where you might want to have increased learning rate only for the new layer.<\/p>\n\n<p>For more info check the caffe <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow\">fine-tuning<\/a> tutorial. (Note: it is a bit outdated and the deprecated term <code>blobs_lr<\/code> is used there instead of <code>lr_mult<\/code>)<\/p>\n\n<p><strong>EDIT<\/strong>: To my best knowledge it is not possible to define a layer-specific learning rate from the <code>solver.prototxt<\/code>. Hence, assuming the <code>solver.prototxt<\/code> limitation is not strict, I suggest a different method to achieve the same result.<\/p>\n"}]}
{"QuestionId":33983627,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2015-11-29T13:59:32.823","AcceptedAnswerId":33986024.0,"OwnerUserId":2476373.0,"Title":"How to reuse same network twice within a new network in CAFFE","Body":"<p>I have a pretrained network (let's call it ) I would like to use twice within a new network. Anybody knows how to duplicate it? Then I would like to assign a different learning rate to each copy.<\/p>\n\n<p>For example ( is the 1st copy of ,  is the 2nd copy of ), the new network might look like:<\/p>\n\n<pre><\/pre>\n\n<p>I know how to reuse  with a single copy, however, since  and  will have different (finetune) learning rates, I don't know how can I make 2 copies of  and assign different learning rate for each.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"33986024","CreationDate":"2015-11-29T17:55:33.907","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Using the same net twice is something called <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html\" rel=\"nofollow\">\"Siamese network\"<\/a>. The way it is implemented in caffe is by explicitly duplicating the network, but using <code>\"name\"<\/code> param for each parameters blob to create a <em>single<\/em> copy of the underlying parameters. See <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/siamese\/mnist_siamese_train_test.prototxt\" rel=\"nofollow\">this prototxt for example<\/a>.<br>\nOnce you explicitly defeine the net twice, you can assign different <code>\"lr_mult\"<\/code> params for each copy.<\/p>\n\n<p>So suppose your reference network <code>N<\/code> has an input layer (which I'll skip in this example) and an inner product layer named <code>\"ip1\"<\/code>. Then <\/p>\n\n<pre><code> layer {\n   name: \"ip1_a\"\n   bottom: \"data_a\"\n   top: \"ip1_a\"\n   type: \"InnerProduct\"\n   inner_product_param {\n     num_output: 10\n   }\n   param {\n     name: \"ip1_w\"  # NOTE THIS NAME!\n     lr_mult: 1\n   }\n   param {\n     name: \"ip1_b\"\n     lr_mult: 2\n   }\n }\n layer {\n   name: \"ip1_b\"\n   bottom: \"data_b\"\n   top: \"ip1_b\"\n   type: \"InnerProduct\"\n   inner_product_param {\n     num_output: 10\n   }\n   param {\n     name: \"ip1_w\"  # NOTE THIS NAME: it's the same!\n     lr_mult: 10 # different LR for this branch\n   }\n   param {\n     name: \"ip1_b\"\n     lr_mult: 20\n   }\n }\n # one layer to combine them     \n layer {\n   type: \"Concat\"\n   bottom: \"ip1_a\"\n   bottom: \"ip1_b\"\n   top: \"ip1_combine\"\n   name: \"concat\"\n }\n layer {\n   name: \"joint_ip\"\n   type: \"InnerProduct\"\n   bottom: \"ip1_combine\"\n   top: \"joint_ip\"\n   inner_product_param {\n     num_output: 30\n   }\n } \n<\/code><\/pre>\n\n<p>If you finetune, you might need to do some net-surgery in order of the original wieghts to be saved in the <code>.caffemodel<\/code> file with the names <code>\"ip1_w\"<\/code> and <code>\"ip1_b\"<\/code>.<\/p>\n"}]}
{"QuestionId":33983698,"AnswerCount":3,"Tags":"<python><numpy><lua><torch>","CreationDate":"2015-11-29T14:08:27.040","AcceptedAnswerId":33987157.0,"OwnerUserId":2267681.0,"Title":"Torch7 Tensor Non-Contiguos Index (Similar to Numpy)","Body":"<p>I am new to torch7, and I can't find a way to get the some non contiguous indices of a tensor based on another tensor. In numpy, what I do is the following:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Is there a way to do something similar in torch7? Something like:<\/p>\n\n<pre class=\"lang-lua prettyprint-override\"><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"33987157","CreationDate":"2015-11-29T19:38:43.337","ParentId":null,"OwnerUserId":"2267681","Title":null,"Body":"<p>Ok, looking arround, I couldn't find an exact solution, but I found an approximation of what I wanted to do, I share it in case someone else finds it useful:<\/p>\n\n<pre class=\"lang-lua prettyprint-override\"><code>array = torch.zeros(5) -- array = [0 0 0 0 0]\nindices = torch.LongTensor({1, 3, 5}) -- Is important that this is a LongTensor\narray:indexAdd(1, indices, torch.Tensor({1, 2, 3})) -- array = [1 0 2 0 3]\n<\/code><\/pre>\n"},{"AnswerId":"34166405","CreationDate":"2015-12-08T21:31:29.947","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-indexcopydim-index-tensor\" rel=\"nofollow\"><code>torch.IndexCopy<\/code><\/a> does exactly what you need:<\/p>\n\n<pre><code>array:indexCopy(1, indices, torch.Tensor({1, 2, 3}))\n<\/code><\/pre>\n"},{"AnswerId":"36897020","CreationDate":"2016-04-27T17:36:32.890","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>If you are a python user, maybe you could also try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">https:\/\/github.com\/imodpasteur\/lutorpy<\/a> . For example, you can process your array in python and then convert it to torch tensor. If you want to convert back to numpy array, the conversion is instant because it only pass the pointer, and two object in python and lua are sharing the same memory.<\/p>\n\n<pre><code>array = np.zeros(5) # array = [0 0 0 0 0]\nindices = np.array([0, 2, 4])\narray[indices] = np.array([1, 2, 3]) # array = [1 0 2 0 3]\n\nrequire(\"torch\")\n# convert numpy array to torch tensor\ntensor = torch.fromNumpyArray(array) \n\n# put your torch code here\n\n#convert back to numpy array\narray = tensor.asNumpyArray()\n# now array is sharing memory with tensor \n<\/code><\/pre>\n"}]}
{"QuestionId":33983919,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-11-29T14:31:27.507","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Why some example in tensor flow tutorial need to be compiled?","Body":"<p>In some tensorflow tutorial (e.g RNN) the code need to be compiled with bazel. Why we have to compile these examples even they are written in python ? <\/p>\n","answers":[{"AnswerId":"34033720","CreationDate":"2015-12-02T02:10:15.227","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Because Python calls into C code using SWIG and that needs to be compiled<\/p>\n"}]}
{"QuestionId":33985305,"AnswerCount":2,"Tags":"<python><scipy><theano><lapack><keras>","CreationDate":"2015-11-29T16:49:59.470","AcceptedAnswerId":null,"OwnerUserId":4332827.0,"Title":"Scipy output error :undefined symbol: sgegv_","Body":"<p>When I tried running keras examples, the program output an error,<\/p>\n\n<pre><\/pre>\n\n<p>Anyone knows how to solve it?<\/p>\n","answers":[{"AnswerId":"34219523","CreationDate":"2015-12-11T08:57:24.383","ParentId":null,"OwnerUserId":"5668028","Title":null,"Body":"<p>In lapack 3.6.0 some deprecated functions which are used in scipy were dropped.\nYou have to use lapack 3.5.0 or lower in order to get things to work back.<\/p>\n"},{"AnswerId":"34429976","CreationDate":"2015-12-23T06:37:22.247","ParentId":null,"OwnerUserId":"4629487","Title":null,"Body":"<p>No need to downgrade your lapack. It is better to upgrade your scipy <\/p>\n\n<pre><code>pip install git+https:\/\/github.com\/scipy\/scipy.git\n<\/code><\/pre>\n"}]}
{"QuestionId":33988334,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-29T21:25:01.857","AcceptedAnswerId":33992733.0,"OwnerUserId":1364327.0,"Title":"Theano config directly in script","Body":"<p>I'm new to Theano and I wonder how to configure the default setting directly from script (without setting envir. variables). E.g. this is a working solution (<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"noreferrer\">source<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>I intend to come up with the identical solution that is executed by only:<\/p>\n\n<pre><\/pre>\n\n<p>and the additional parameters are set directly in the script itself. E.g. somehow like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is it even possible? I read the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"noreferrer\">config page<\/a> which provides the information that allows me to read the already set values (but not to set them by myself).<\/p>\n","answers":[{"AnswerId":"33992733","CreationDate":"2015-11-30T06:34:42.773","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>When you do this:<\/p>\n\n<pre><code>$ THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python check1.py\n<\/code><\/pre>\n\n<p>All you're actually doing is setting an environment variable before running the Python script.<\/p>\n\n<p>You can set environment variables in Python too. For example, the <code>THEANO_FLAGS<\/code> environment variable can be set inside Python like this:<\/p>\n\n<pre><code>import os\nos.environ[\"THEANO_FLAGS\"] = \"mode=FAST_RUN,device=gpu,floatX=float32\"\n<\/code><\/pre>\n\n<p>Note that some Theano config variables cannot be changed after importing Theano so this is fine:<\/p>\n\n<pre><code>import os\nos.environ[\"THEANO_FLAGS\"] = \"mode=FAST_RUN,device=gpu,floatX=float32\"\nimport theano\n<\/code><\/pre>\n\n<p>But this will not work as expected:<\/p>\n\n<pre><code>import theano\nimport os\nos.environ[\"THEANO_FLAGS\"] = \"mode=FAST_RUN,device=gpu,floatX=float32\"\n<\/code><\/pre>\n"}]}
{"QuestionId":33988422,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2015-11-29T21:35:15.710","AcceptedAnswerId":34007510.0,"OwnerUserId":688549.0,"Title":"Theano: Discrepancy vs Numpy when taking matrix dot with CPU vs GPU","Body":"<p>I recently got Theano working on Windows 10 with CUDA v7.5, CUDNN v3, and Visual Studio 2013 Community Edition. In order to verify it was working correctly, I tested the following code from the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\" title=\"Theano Windows install page\">Theano Windows install page<\/a> using both CPU and GPU:<\/p>\n\n<pre><\/pre>\n\n<p>I got the following results:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, there is a fairly significant difference of 0.022 when doing the calculation on GPU. Just wondering whether this is to be expected or I am doing something wrong.<\/p>\n\n<p>Here is my .theanorc:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34007510","CreationDate":"2015-11-30T20:35:05.827","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The GPU doesn't do the addition and multiplication in the same order. As floats are not exact, it is normal to see some differences.<\/p>\n\n<p>An absolute difference of that size can be normal if the relative difference is small.<\/p>\n\n<p>To compare them more \"correctly\" use <code>theano.tensor.basic._allclose(result1, result2)<\/code><\/p>\n"}]}
{"QuestionId":33991566,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-11-30T04:41:26.220","AcceptedAnswerId":null,"OwnerUserId":3090897.0,"Title":"Is it possible to plot more than two scalars in the same plot in tensorboard?","Body":"<p>Is it possible to plot more than two scalars in the same plot? More specifically. it could allow:<\/p>\n\n<ul>\n<li>training loss and test loss being plotted together<\/li>\n<li>regularization loss and top level loss to be plotted together.<\/li>\n<li>maybe adding accuracy on the right vertical axis<\/li>\n<\/ul>\n\n<p>Just name a few. This would help a lot.<\/p>\n","answers":[{"AnswerId":"42768433","CreationDate":"2017-03-13T16:14:46.867","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tensorboard\/README.md\" rel=\"nofollow noreferrer\">github documentation for TensorBoard<\/a> says that it only overlaps plots from different runs with the same tag name, so this won't work. If you'd like this feature please file a bug.<\/p>\n"}]}
{"QuestionId":33992029,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-11-30T05:36:01.270","AcceptedAnswerId":34006997.0,"OwnerUserId":2857219.0,"Title":"Implementing batch normalization with tensorflow","Body":"<p>I am trying to implement a batch normalization layer in . I am having no problem running the train step of this using  to get the <strong>mean<\/strong> and <strong>variance<\/strong>. <\/p>\n\n<p>For test time, I'd like to set up an exponential moving average to track the mean and variance. I am trying to do it like this:<\/p>\n\n<pre><\/pre>\n\n<p>Where _variable_on_cpu is defined as:<\/p>\n\n<pre><\/pre>\n\n<p>I believe that I am setting<\/p>\n\n<pre><\/pre>\n\n<p>Incorrectly, but I am not sure how. When I use tensorboard to track these mean and variance variables, they are just flat that their initialized values.<\/p>\n","answers":[{"AnswerId":"34006997","CreationDate":"2015-11-30T20:05:06.207","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Rafal's comment gets at the core of the problem:  You're not running the assign nodes.  You might try using the batchnorm helper I posted in another answer - <a href=\"https:\/\/stackoverflow.com\/questions\/33949786\/how-could-i-use-batch-normalization-in-tensorflow\/33950177?noredirect=1#comment55758348_33950177\">How could I use Batch Normalization in TensorFlow?<\/a> - or you can force the assign to happen by adding with_dependencies, as he suggests.<\/p>\n\n<p>The general principle is that you should only count on a node being run if data or control dependencies flow \"through\" it.  <code>with_dependencies<\/code> ensures that before the output op is used, the specified dependencies will have completed.<\/p>\n"}]}
{"QuestionId":33993604,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-11-30T07:40:01.323","AcceptedAnswerId":34002684.0,"OwnerUserId":1646131.0,"Title":"Using TensorFlow to classify Handwritten Digits without validation labels","Body":"<p>I have been going through the TensorFlow Tutorial and in general reading up on Machine Learning. <\/p>\n\n<p>Its my understanding that one of the major benefits to using a neural network is their ability to quickly classify the presented input after training. <\/p>\n\n<p>To start off I began by stepping through the example code and looking at how the training data is structured and I was able to successfully use the basic example (91% accuracy) to recognize images (only digits) that I have created using the following code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>Example of an image that is produced from the above code:\n<a href=\"http:\/\/i.stack.imgur.com\/CQDiC.png\" rel=\"nofollow\">Bitmap extracted from test data used.<\/a><\/p>\n\n<p><strong>Note:<\/strong> <em>This image is built from a list of points and then scaled down to fit a 28<\/em> * <em>28 array. The colors are inverted as the image is simply converted straight from a numpy array to a bitmap. Each point in the list is set to 0xFF as per the MNIST file format where 0 is white and 255 black<\/em>.<\/p>\n\n<p>This snippet above outputs  (sometimes  depending on the accuracy from training) and it correctly classifies the inputs to the labels.<\/p>\n\n<p>So my question is what are the necessary steps involved in using neural network built using TensorFlow to simply classify what the input is so that for example if the input is a '7' the output would for example be:<\/p>\n\n<pre><\/pre>\n\n<p>I have had a look through the TensorFlow documentation and I have not been able to come up with an solution. I suspect it is probably down to missing something in the tutorial.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34002684","CreationDate":"2015-11-30T15:57:35.647","ParentId":null,"OwnerUserId":"998293","Title":null,"Body":"<p>Assuming you've followed the <a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">MNIST for ML Beginners Tutorial<\/a>, to get a simple prediction, add an argmax node like so:<\/p>\n\n<pre><code>prediction = tf.argmax(y, 1)\n<\/code><\/pre>\n\n<p>And then run it, feeding in the data you'd like to classify:<\/p>\n\n<pre><code>prediction_val = sess.run(prediction, feed_dict={x: dataset.images})\n<\/code><\/pre>\n\n<p><code>prediction_val<\/code> will be of shape <code>(batch_size,)<\/code> and contains the likeliest label for each image in the batch.<\/p>\n\n<p>Note that the <code>feed_dict<\/code> only includes <code>x<\/code> and not <code>y_<\/code> because <code>prediction<\/code> does not depend on <code>y_<\/code>. <\/p>\n"}]}
{"QuestionId":33994241,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-11-30T08:26:15.590","AcceptedAnswerId":33994670.0,"OwnerUserId":5118777.0,"Title":"how to extract matrix from big matrix in torch\/lua","Body":"<p>In torch7 \/ lua<\/p>\n\n<p>There is a matrix:<\/p>\n\n<pre><\/pre>\n\n<p>how to extract this:<\/p>\n\n<pre><\/pre>\n\n<p>and how to overwrite it by matrix operation<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"33994670","CreationDate":"2015-11-30T08:55:25.500","ParentId":null,"OwnerUserId":"5118777","Title":null,"Body":"<pre><code>matrix\n\nmatrix:sub(2,3,2,3)\n\nz = torch.Tensor({{78,66},{45,21}})\n\nmatrix:sub(2,3,2,3):copy(z)\n<\/code><\/pre>\n"}]}
{"QuestionId":33997823,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><tensorflow><supervised-learning>","CreationDate":"2015-11-30T11:42:25.543","AcceptedAnswerId":34022791.0,"OwnerUserId":3314143.0,"Title":"TensorFlow MLP not training XOR","Body":"<p>I've built an MLP with Google's <a href=\"http:\/\/www.tensorflow.org\/\" rel=\"noreferrer\">TensorFlow<\/a> library. The network is working but somehow it refuses to learn properly. It always converges to an output of nearly 1.0 no matter what the input actually is.<\/p>\n\n<p>The <strong>complete code<\/strong> can be seen <a href=\"http:\/\/pastebin.com\/xKX7XpYx\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Any ideas?<\/p>\n\n<hr>\n\n<p>The <strong>input and output<\/strong> (batch size 4) is as follows:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Hidden layer configuration<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Output layer configuration<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>My <strong>learning<\/strong> methods look like this:<\/p>\n\n<pre><\/pre>\n\n<p>I tried both setups for <strong>cross entropy<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>where  is the original output as described in  and  the predicted\/calculated value by my network.<\/p>\n\n<hr>\n\n<p>The <strong>training inside the for-loop<\/strong> (for n epochs) goes like this:<\/p>\n\n<pre><\/pre>\n\n<p>I am saving the outcome to cvalues for debug printig of , , ...<\/p>\n\n<hr>\n\n<p>No matter what I've tried, when I test my network, trying to validate the output, it always produces something like this:<\/p>\n\n<pre><\/pre>\n\n<p>So it <strong>is not<\/strong> learning properly but always converging to nearly 1.0 no matter which input is fed.<\/p>\n","answers":[{"AnswerId":"34009074","CreationDate":"2015-11-30T22:19:21.857","ParentId":null,"OwnerUserId":"2879934","Title":null,"Body":"<p>I can't comment because I don't have enough reputation but I have some questions on that answer mrry. The $L_2$ loss function makes sense because it is basically the MSE function, but why wouldn't cross-entropy work? Certainly works for other NN libs. Second of all why in the world would translating your input space from $[0,1] -> [-1,1]$ have any affect <em>especially<\/em> since you added bias vectors.<\/p>\n\n<p><strong>EDIT<\/strong> This is a solution using cross entropy and one-hot compiled from multiple sources \n<strong>EDIT^2<\/strong> changed the code to use cross-entropy without any extra encoding or any weird target value shifting <\/p>\n\n<pre><code>import math\nimport tensorflow as tf\nimport numpy as np\n\nHIDDEN_NODES = 10\n\nx = tf.placeholder(tf.float32, [None, 2])\nW_hidden = tf.Variable(tf.truncated_normal([2, HIDDEN_NODES]))\nb_hidden = tf.Variable(tf.zeros([HIDDEN_NODES]))\nhidden = tf.nn.relu(tf.matmul(x, W_hidden) + b_hidden)\n\nW_logits = tf.Variable(tf.truncated_normal([HIDDEN_NODES, 1]))\nb_logits = tf.Variable(tf.zeros([1]))\nlogits = tf.add(tf.matmul(hidden, W_logits),b_logits)\n\n\ny = tf.nn.sigmoid(logits)\n\n\ny_input = tf.placeholder(tf.float32, [None, 1])\n\n\n\nloss = -(y_input * tf.log(y) + (1 - y_input) * tf.log(1 - y))\n\ntrain_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\n\ninit_op = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run(init_op)\n\nxTrain = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])\n\n\nyTrain = np.array([[0], [1], [1], [0]])\n\n\nfor i in xrange(2000):\n  _, loss_val,logitsval = sess.run([train_op, loss,logits], feed_dict={x: xTrain, y_input: yTrain})\n\n  if i % 10 == 0:\n    print \"Step:\", i, \"Current loss:\", loss_val,\"logits\",logitsval\n\nprint \"---------\"\nprint sess.run(y,feed_dict={x: xTrain})\n<\/code><\/pre>\n"},{"AnswerId":"34022791","CreationDate":"2015-12-01T14:28:14.343","ParentId":null,"OwnerUserId":"3314143","Title":null,"Body":"<p>In the meanwhile with the help of a colleague I were able to fix my solution and wanted to post it for completeness. My solution works <strong>with cross entropy<\/strong> and <strong>without altering the training data<\/strong>. Additionally it has the desired <strong>input shape of (1, 2)<\/strong> and <strong>ouput is scalar<\/strong>.<\/p>\n\n<p>It makes use of an <code>AdamOptimizer<\/code> which decreases the error <strong>much<\/strong> faster than a <code>GradientDescentOptimizer<\/code>. See <a href=\"https:\/\/stats.stackexchange.com\/questions\/184448\/difference-between-gradiendescentoptimizer-and-adamoptimizer-tensorflow?noredirect=1#comment350240_184448\">this post<\/a> for more information (&amp; questions^^) about the optimizer.<\/p>\n\n<p>In fact, my network produces reasonably good results in only 400-800 learning steps.<\/p>\n\n<p>After 2000 learning steps the output is nearly \"perfect\":<\/p>\n\n<pre><code>step: 2000\nloss: 0.00103311243281\n\ninput: [0.0, 0.0] | output: [[ 0.00019799]]\ninput: [0.0, 1.0] | output: [[ 0.99979786]]\ninput: [1.0, 0.0] | output: [[ 0.99996307]]\ninput: [1.0, 1.0] | output: [[ 0.00033751]]\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>import tensorflow as tf    \n\n#####################\n# preparation stuff #\n#####################\n\n# define input and output data\ninput_data = [[0., 0.], [0., 1.], [1., 0.], [1., 1.]]  # XOR input\noutput_data = [[0.], [1.], [1.], [0.]]  # XOR output\n\n# create a placeholder for the input\n# None indicates a variable batch size for the input\n# one input's dimension is [1, 2] and output's [1, 1]\nn_input = tf.placeholder(tf.float32, shape=[None, 2], name=\"n_input\")\nn_output = tf.placeholder(tf.float32, shape=[None, 1], name=\"n_output\")\n\n# number of neurons in the hidden layer\nhidden_nodes = 5\n\n\n################\n# hidden layer #\n################\n\n# hidden layer's bias neuron\nb_hidden = tf.Variable(tf.random_normal([hidden_nodes]), name=\"hidden_bias\")\n\n# hidden layer's weight matrix initialized with a uniform distribution\nW_hidden = tf.Variable(tf.random_normal([2, hidden_nodes]), name=\"hidden_weights\")\n\n# calc hidden layer's activation\nhidden = tf.sigmoid(tf.matmul(n_input, W_hidden) + b_hidden)\n\n\n################\n# output layer #\n################\n\nW_output = tf.Variable(tf.random_normal([hidden_nodes, 1]), name=\"output_weights\")  # output layer's weight matrix\noutput = tf.sigmoid(tf.matmul(hidden, W_output))  # calc output layer's activation\n\n\n############\n# learning #\n############\ncross_entropy = -(n_output * tf.log(output) + (1 - n_output) * tf.log(1 - output))\n# cross_entropy = tf.square(n_output - output)  # simpler, but also works\n\nloss = tf.reduce_mean(cross_entropy)  # mean the cross_entropy\noptimizer = tf.train.AdamOptimizer(0.01)  # take a gradient descent for optimizing with a \"stepsize\" of 0.1\ntrain = optimizer.minimize(loss)  # let the optimizer train\n\n\n####################\n# initialize graph #\n####################\ninit = tf.initialize_all_variables()\n\nsess = tf.Session()  # create the session and therefore the graph\nsess.run(init)  # initialize all variables  \n\n#####################\n# train the network #\n#####################\nfor epoch in xrange(0, 2001):\n    # run the training operation\n    cvalues = sess.run([train, loss, W_hidden, b_hidden, W_output],\n                       feed_dict={n_input: input_data, n_output: output_data})\n\n    # print some debug stuff\n    if epoch % 200 == 0:\n        print(\"\")\n        print(\"step: {:&gt;3}\".format(epoch))\n        print(\"loss: {}\".format(cvalues[1]))\n        # print(\"b_hidden: {}\".format(cvalues[3]))\n        # print(\"W_hidden: {}\".format(cvalues[2]))\n        # print(\"W_output: {}\".format(cvalues[4]))\n\n\nprint(\"\")\nprint(\"input: {} | output: {}\".format(input_data[0], sess.run(output, feed_dict={n_input: [input_data[0]]})))\nprint(\"input: {} | output: {}\".format(input_data[1], sess.run(output, feed_dict={n_input: [input_data[1]]})))\nprint(\"input: {} | output: {}\".format(input_data[2], sess.run(output, feed_dict={n_input: [input_data[2]]})))\nprint(\"input: {} | output: {}\".format(input_data[3], sess.run(output, feed_dict={n_input: [input_data[3]]})))\n<\/code><\/pre>\n"}]}
{"QuestionId":33998777,"AnswerCount":3,"Tags":"<python><gcc><theano><pymc3>","CreationDate":"2015-11-30T12:32:28.910","AcceptedAnswerId":null,"OwnerUserId":5475994.0,"Title":"Theano step gives File in wrong format error","Body":"<p>I am trying to replicate the example \n<a href=\"https:\/\/pymc-devs.github.io\/pymc3\/getting_started\/\" rel=\"nofollow\">Getting Started PyMC3<\/a>.<\/p>\n\n<p>on Windows 2012 R2.<\/p>\n\n<p>The step\n<br>\ngives<\/p>\n\n<pre><\/pre>\n\n<p>I have for gcc<\/p>\n\n<pre><\/pre>\n\n<p>what am i missing?<\/p>\n\n<p>thanks<\/p>\n","answers":[{"AnswerId":"34028462","CreationDate":"2015-12-01T19:25:45.283","ParentId":null,"OwnerUserId":"5475994","Title":null,"Body":"<p>Whatever the issue was creating a python 3.4 environment together with <\/p>\n\n<pre><code>conda install mingw libpython\n<\/code><\/pre>\n\n<p>solved it.<\/p>\n"},{"AnswerId":"37026340","CreationDate":"2016-05-04T11:20:30.873","ParentId":null,"OwnerUserId":"4225948","Title":null,"Body":"<p>What solved my issue is installing Python 2.7 version of anaconda<\/p>\n"},{"AnswerId":"41539057","CreationDate":"2017-01-08T23:28:14.850","ParentId":null,"OwnerUserId":"297990","Title":null,"Body":"<p>I got the same error using ipython. Using just python I imported theano without problem. Subsequent uses with ipython worked fine after the first compilation with just python.<\/p>\n"}]}
{"QuestionId":34001922,"AnswerCount":9,"Tags":"<python><pandas><classification><tensorflow>","CreationDate":"2015-11-30T14:41:20.107","AcceptedAnswerId":null,"OwnerUserId":3654387.0,"Title":"FailedPreconditionError: Attempting to use uninitialized in Tensorflow","Body":"<p>I am working through the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/index.html\" rel=\"noreferrer\">TensorFlow tutorial<\/a>, which uses a \"weird\" format to upload the data. I would like to use the NumPy or pandas format for the data, so that I can compare it with scikit-learn results.<\/p>\n\n<p>I get the digit recognition data from Kaggle: <a href=\"https:\/\/www.kaggle.com\/c\/digit-recognizer\/data\" rel=\"noreferrer\">https:\/\/www.kaggle.com\/c\/digit-recognizer\/data<\/a>.<\/p>\n\n<p>Here the code from the TensorFlow tutorial (which works fine):<\/p>\n\n<pre><\/pre>\n\n<p>Here I read the data, strip out the target variables and split the data into testing and training datasets (this all works fine):<\/p>\n\n<pre><\/pre>\n\n<p>Now, when I try to run the training step, I get a :<\/p>\n\n<pre><\/pre>\n\n<p>Here's the full error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas as to how I can fix this?<\/p>\n","answers":[{"AnswerId":"43583960","CreationDate":"2017-04-24T09:08:30.467","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>From official documentation, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/errors\/FailedPreconditionError\" rel=\"noreferrer\">FailedPreconditionError<\/a> <\/p>\n\n<blockquote>\n  <p>This exception is most commonly raised when running an operation that\n  reads a tf.Variable before it has been initialized.<\/p>\n<\/blockquote>\n\n<p>In your case the error even explains what variable was not initialized: <code>Attempting to use uninitialized value Variable_1<\/code>. One of the TF tutorials explains a lot about variables, their <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables\" rel=\"noreferrer\">creation\/initialization\/saving\/loading<\/a><\/p>\n\n<p>Basically to initialize the variable you have 3 options:<\/p>\n\n<ul>\n<li>initialize all global variables with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/global_variables_initializer\" rel=\"noreferrer\"><code>tf.global_variables_initializer()<\/code><\/a><\/li>\n<li>initialize variables you care about with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variables_initializer\" rel=\"noreferrer\"><code>tf.variables_initializer(list_of_vars)<\/code><\/a>. Notice that you can use this function to mimic global_variable_initializer: <code>tf.variable_initializers(tf.global_variables())<\/code><\/li>\n<li>initialize only one variable with <code>var_name.initializer<\/code><\/li>\n<\/ul>\n\n<p>I almost always use the first approach. Remember you should put it inside a session run. So you will get something like this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n\n<p>If your are curious about more information about variables, read <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/state_ops\" rel=\"noreferrer\">this documentation<\/a> to know how to <code>report_uninitialized_variables<\/code> and check <code>is_variable_initialized<\/code>.<\/p>\n"},{"AnswerId":"50719652","CreationDate":"2018-06-06T11:46:51.770","ParentId":null,"OwnerUserId":"374858","Title":null,"Body":"<p>When I had this issue with <code>tf.train.string_input_producer()<\/code> and <code>tf.train.batch()<\/code> initializing the local variables before I started the Coordinator solved the problem. I had been getting the error when I initialized the local variables after starting the Coordinator.<\/p>\n"},{"AnswerId":"34013098","CreationDate":"2015-12-01T05:15:19.060","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>FailedPreconditionError<\/code> arises because the program is attempting to read a variable (named <code>\"Variable_1\"<\/code>) before it has been initialized. In TensorFlow, all variables must be explicitly initialized, by running their \"initializer\" operations. For convenience, you can run all of the variable initializers in the current session by executing the following statement before your training loop:<\/p>\n\n<pre><code>tf.initialize_all_variables().run()\n<\/code><\/pre>\n\n<p>Note that this answer assumes that, as in the question, you are using <code>tf.InteractiveSession<\/code>, which allows you to run operations without specifying a session. For non-interactive uses, it is more common to use <code>tf.Session<\/code>, and initialize as follows:<\/p>\n\n<pre><code>init_op = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run(init_op)\n<\/code><\/pre>\n"},{"AnswerId":"53851885","CreationDate":"2018-12-19T13:04:41.813","ParentId":null,"OwnerUserId":"8180401","Title":null,"Body":"<p>The <strong>FailedPreconditionError<\/strong> comes because the session is trying to read a variable that hasn\"t been initialized.<\/p>\n\n<p>As of <em>Tensorflow<\/em> version 1.11.0, you need to take this :<\/p>\n\n<pre><code>init_op = tf.global_variables_initializer()\n\nsess = tf.Session()\nsess.run(init_op)\n<\/code><\/pre>\n"},{"AnswerId":"54401883","CreationDate":"2019-01-28T12:18:11.467","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>You have to initialize variables before using them.\u200b<\/p>\n\n<p>If you try to evaluate the variables before initializing them you'll run into:\n<code>FailedPreconditionError: Attempting to use uninitialized value tensor.<\/code><\/p>\n\n<p>The easiest way is initializing all variables at once using: <code>tf.global_variables_initializer()<\/code><\/p>\n\n<pre><code>init = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n<\/code><\/pre>\n\n<p>You use <code>sess.run(init)<\/code> to run the initializer, without fetching any value.<\/p>\n\n<p>To initialize only a subset of variables, you use <code>tf.variables_initializer()<\/code> listing the variables:<\/p>\n\n<pre><code>var_ab = tf.variables_initializer([a, b], name=\"a_and_b\")\nwith tf.Session() as sess:\n    sess.run(var_ab)\n<\/code><\/pre>\n\n<p>You can also initialize each variable separately using <code>tf.Variable.initializer<\/code><\/p>\n\n<pre><code># create variable W as 784 x 10 tensor, filled with zeros\nW = tf.Variable(tf.zeros([784,10])) with tf.Session() as sess:\n    sess.run(W.initializer)\n<\/code><\/pre>\n"},{"AnswerId":"54817353","CreationDate":"2019-02-21T22:38:41.323","ParentId":null,"OwnerUserId":"4549682","Title":null,"Body":"<p>Possibly something has changed in recent TensorFlow builds, because for me, running<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.local_variables_initializer())\n<\/code><\/pre>\n\n<p>before fitting any models seems to do the trick.  Most older examples and comments seem to suggest <code>tf.global_variables_initializer()<\/code>.<\/p>\n"},{"AnswerId":"41607207","CreationDate":"2017-01-12T07:16:36.377","ParentId":null,"OwnerUserId":"3144836","Title":null,"Body":"<p><code>tf.initialize_all_variables()<\/code> is deprecated. Instead initialize tensorflow variables with: <\/p>\n\n<pre><code>tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>A common example usage is:<\/p>\n\n<pre><code>with tf.Session() as sess:\n     sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n"},{"AnswerId":"50601988","CreationDate":"2018-05-30T10:10:21.093","ParentId":null,"OwnerUserId":"2426324","Title":null,"Body":"<p>Different use case, but set your session as the default session did the trick for me:<\/p>\n\n<pre><code>with sess.as_default():\n    result = compute_fn([seed_input,1])\n<\/code><\/pre>\n\n<p>This is one of these mistakes that is so obvious, once you have solved it. <\/p>\n\n<p>My use-case is the following: <br>\n1) store keras VGG16 as tensorflow graph <br>\n2) load kers VGG16 as a graph <br>\n3) run tf function on the graph and get: <\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value block1_conv2\/bias\n     [[Node: block1_conv2\/bias\/read = Identity[T=DT_FLOAT, _class=[\"loc:@block1_conv2\/bias\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](block1_conv2\/bias)]]\n     [[Node: predictions\/Softmax\/_7 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_168_predictions\/Softmax\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n"},{"AnswerId":"45347595","CreationDate":"2017-07-27T10:06:15.487","ParentId":null,"OwnerUserId":"5082576","Title":null,"Body":"<p>I got this error message from a completely different case. It seemed that the exception handler in tensorflow raised it. You can check each row in the Traceback. In my case, it happened in <code>tensorflow\/python\/lib\/io\/file_io.py<\/code>, because this file contained a different bug, where <code>self.__mode<\/code> and <code>self.__name<\/code> weren't initialized, and it needed to call <code>self._FileIO__mode<\/code>, and <code>self_FileIO__name<\/code> instead. <\/p>\n"}]}
{"QuestionId":34002591,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-11-30T15:53:28.130","AcceptedAnswerId":34002962.0,"OwnerUserId":5612876.0,"Title":"Tensorflow slicing based on variable","Body":"<p>I've found that indexing still is an open issue in tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\">(#206)<\/a>, so I'm wondering what I could use as a workaround at the moment. I want to index\/slice a row\/column of a matrix based on a variable that changes for every training example.<\/p>\n\n<p>What I've tried so far:<\/p>\n\n<ol>\n<li>Slicing based on placeholder (doesn't work)<\/li>\n<\/ol>\n\n<p>The following (working) code slices based on a fixed number.<\/p>\n\n<pre><\/pre>\n\n<p>However, it seems that I can't simply replace one of these fixed numbers with a tf.placeholder. The following code gives me the error <em>\"TypeError: List of Tensors when single Tensor expected.\"<\/em><\/p>\n\n<pre><\/pre>\n\n<p>This sounds like the brackets around <em>[i]<\/em> are too much, but removing them  doesn't help either. How to use a placeholder\/variable as index?<\/p>\n\n<ol start=\"2\">\n<li>Slicing based on python variable (doesn't backprop\/update properly)<\/li>\n<\/ol>\n\n<p>I've also tried using a normal python variable as index. This does not lead to an error, but the network doesn't learn anything while training. I suppose because the changing variable is not properly registered, the graph is malformed and updates don't work?<\/p>\n\n<ol start=\"3\">\n<li>Slicing via one-hot vector + multiplication (works, but is slow)<\/li>\n<\/ol>\n\n<p>One workaround I found is using a one-hot vector. Making a one-hot vector in numpy, passing this using a placeholder, then doing the slicing via matrix multiplication. This works, but is quite slow.<\/p>\n\n<p>Any ideas how to efficiently slice\/index based on a variable?<\/p>\n","answers":[{"AnswerId":"58411226","CreationDate":"2019-10-16T10:31:49.943","ParentId":null,"OwnerUserId":"3293408","Title":null,"Body":"<p>If you have an extra dimension, this works.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef reorder0(e, i, length):\n    '''\n    e: a two dimensional tensor\n    i: a one dimensional int32 tensor, of shape (e.shape[0])\n    returns: a tensor of the same shape as e, where the jth entry is entry i[j] from e\n    '''\n    return tf.concat(\n        [ tf.expand_dims( e[i[j],:], axis=0)  for j in range(length) ],\n        axis=0\n    )\n\ne = tf.placeholder(tf.float32, shape=(2,3,5), name='e' )  # sentences, words, embedding\ni = tf.placeholder(tf.int32, shape=(2,3), name='i' ) # for each word, index of parent\np = tf.concat(\n    [ tf.expand_dims(reorder0(e[k,:,:], i[k,:], 3), axis=0)  for k in range(2) ],\n    axis=0,\n    name='p'\n)\n\n#initialize\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\n#run\nresult = sess.run(p, feed_dict={\n    e: [ \n        ( (1.0,1.1,1.2,1.3,1.4),(2.0,2.1,2.2,2.3,2.4),(3.0,3.1,3.2,3.3,3.4) ), \n        ( (21.0,21.1,21.2,21.3,21.4),(22.0,22.1,22.2,22.3,22.4),(23.0,23.1,23.2,23.3,23.4) ),  \n    ], \n    i: [ (1,1,1), (2,0,2)]\n})\nprint(result)\n<\/code><\/pre>\n"},{"AnswerId":"58412607","CreationDate":"2019-10-16T11:50:35.500","ParentId":null,"OwnerUserId":"3293408","Title":null,"Body":"<p>If the sizes are not known when building the model, use TensorArray.<\/p>\n\n<pre><code>e = tf.placeholder(tf.float32, shape=(3,5) )  # words, embedding\ni = tf.placeholder(tf.int32, shape=(3) ) # for each word, index of parent\n#p = reorder0(e, i, 3)\na = tf.TensorArray(\n    tf.float32, \n    size=e.get_shape()[0],\n    dynamic_size=True,\n    infer_shape= True,\n    element_shape=e.get_shape()[1],\n    clear_after_read = False\n)\n\n\n#initialize\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\n#run\nresult = sess.run(\n    a.unstack(e).gather(i), \n    feed_dict={\n        e: ( (1.0,1.1,1.2,1.3,1.4),(2.0,2.1,2.2,2.3,2.4),(3.0,3.1,3.2,3.3,3.4) ),\n            #( (21.0,21.1,21.2,21.3,21.4),(22.0,22.1,22.2,22.3,22.4),(23.0,23.1,23.2,23.3,23.4) ),  \n        i: (2,0,2)\n    }\n)\nprint(result)\n<\/code><\/pre>\n"},{"AnswerId":"34002962","CreationDate":"2015-11-30T16:11:55.663","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Slicing based on a placeholder should work just fine. It looks like you are running into a type error, due to some subtle issues of shapes and types. Where you have the following:<\/p>\n\n<pre><code>x = tf.placeholder(\"float\")\ni = tf.placeholder(\"int32\")\ny = tf.slice(x,[i],[1])\n<\/code><\/pre>\n\n<p>...you should instead have:<\/p>\n\n<pre><code>x = tf.placeholder(\"float\")\ni = tf.placeholder(\"int32\")\ny = tf.slice(x,i,[1])\n<\/code><\/pre>\n\n<p>...and then you should feed <code>i<\/code> as <code>[0]<\/code> in the call to <code>sess.run()<\/code>.<\/p>\n\n<p>To make this a little clearer, I would recommend rewriting the code as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.placeholder(tf.float32, shape=[None])  # 1-D tensor\ni = tf.placeholder(tf.int32, shape=[1])\ny = tf.slice(x, i, [1])\n\n#initialize\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\n#run\nresult = sess.run(y, feed_dict={x: [1, 2, 3, 4, 5], i: [0]})\nprint(result)\n<\/code><\/pre>\n\n<p>The additional <code>shape<\/code> arguments to the <code>tf.placeholder<\/code> op help to ensure that the values you feed have the appropriate shapes, and also that TensorFlow will raise an error if the shapes are not correct.<\/p>\n"}]}
{"QuestionId":34004474,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><gated-recurrent-unit>","CreationDate":"2015-11-30T17:30:23.047","AcceptedAnswerId":34011303.0,"OwnerUserId":395857.0,"Title":"GRU implementation in Theano","Body":"<p>Based on the LSTM code provided in the official Theano tutorial (<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/lstm.py<\/a>), I changed the LSTM layer code (i.e. the functions  and ) to perform a GRU instead. <\/p>\n\n<p>The provided LSTM code trains well, but not the GRU I coded: the accuracy on the training set with the LSTM goes up to 1 (train cost = 0), while with the GRU it stagnates at 0.7 (train cost = 0.3).<\/p>\n\n<p>Below is the code I use for the GRU. I kept the same function names as in tutorial, so that one can copy paste the code directly in it. What could explain the poor performance of the GRU?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34011303","CreationDate":"2015-12-01T01:53:34.280","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>The issue comes from the last line, <code>return rval[0]<\/code>: it should instead be <code>return rval<\/code>.<\/p>\n\n<p>The LSTM code provided in the official Theano tutorial (<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/code\/lstm.py<\/a>) uses <code>return rval[0]<\/code> because <code>outputs_info<\/code> contains 2 elements:<\/p>\n\n<pre><code>rval, updates = theano.scan(_step,\n                            sequences=[mask, state_below],\n                            outputs_info=[tensor.alloc(numpy_floatX(0.),\n                                                       n_samples,\n                                                       dim_proj),\n                                          tensor.alloc(numpy_floatX(0.),\n                                                       n_samples,\n                                                       dim_proj)],\n                            name=_p(prefix, '_layers'),\n                            n_steps=nsteps)\nreturn rval[0]\n<\/code><\/pre>\n\n<p>In the GRU, <code>outputs_info<\/code> contains just one element:<\/p>\n\n<pre><code>outputs_info=[tensor.alloc(numpy_floatX(0.),\n                           n_samples,\n                           dim_proj)],\n<\/code><\/pre>\n\n<p>and despite the brackets, it won't return a list of a list of Theano variables representing the outputs of scan, but directly a Theano variable.<\/p>\n\n<p>The <code>rval<\/code> is then fed to a pooling layer (in this case, a mean pooling layer):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vEQW8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vEQW8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>By taking only <code>rval[0]<\/code> in the GRU, since in the GRU code <code>rval<\/code> is a Theano variable and not a list of a Theano variables, you removed the part in the red rectangle:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/q7WMZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/q7WMZ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>which means you tried to perform the sentence classification just using the first word.<\/p>\n\n<hr>\n\n<p>Another GRU implementation that can be plugged in the LSTM tutorial:<\/p>\n\n<pre><code># weight initializer, normal by default\ndef norm_weight(nin, nout=None, scale=0.01, ortho=True):\n    if nout is None:\n        nout = nin\n    if nout == nin and ortho:\n        W = ortho_weight(nin)\n    else:\n        W = scale * numpy.random.randn(nin, nout)\n    return W.astype('float32')\n\ndef param_init_lstm(options, params, prefix='lstm'):\n    \"\"\"\n    GRU. Source: https:\/\/github.com\/kyunghyuncho\/dl4mt-material\/blob\/master\/session0\/lm.py\n    \"\"\"\n    nin = options['dim_proj']\n    dim = options['dim_proj']\n    # embedding to gates transformation weights, biases\n    W = numpy.concatenate([norm_weight(nin, dim),\n                           norm_weight(nin, dim)], axis=1)\n    params[_p(prefix, 'W')] = W\n    params[_p(prefix, 'b')] = numpy.zeros((2 * dim,)).astype('float32')\n\n    # recurrent transformation weights for gates\n    U = numpy.concatenate([ortho_weight(dim),\n                           ortho_weight(dim)], axis=1)\n    params[_p(prefix, 'U')] = U\n\n    # embedding to hidden state proposal weights, biases\n    Wx = norm_weight(nin, dim)\n    params[_p(prefix, 'Wx')] = Wx\n    params[_p(prefix, 'bx')] = numpy.zeros((dim,)).astype('float32')\n\n    # recurrent transformation weights for hidden state proposal\n    Ux = ortho_weight(dim)\n    params[_p(prefix, 'Ux')] = Ux\n    return params\n\n\ndef lstm_layer(tparams, state_below, options, prefix='lstm', mask=None):\n\n    nsteps = state_below.shape[0]\n\n    if state_below.ndim == 3:\n        n_samples = state_below.shape[1]\n    else:\n        n_samples = state_below.shape[0]\n\n    dim = tparams[_p(prefix, 'Ux')].shape[1]\n\n    if mask is None:\n        mask = tensor.alloc(1., state_below.shape[0], 1)\n\n    # utility function to slice a tensor\n    def _slice(_x, n, dim):\n        if _x.ndim == 3:\n            return _x[:, :, n*dim:(n+1)*dim]\n        return _x[:, n*dim:(n+1)*dim]\n\n    # state_below is the input word embeddings\n    # input to the gates, concatenated\n    state_below_ = tensor.dot(state_below, tparams[_p(prefix, 'W')]) + \\\n        tparams[_p(prefix, 'b')]\n    # input to compute the hidden state proposal\n    state_belowx = tensor.dot(state_below, tparams[_p(prefix, 'Wx')]) + \\\n        tparams[_p(prefix, 'bx')]\n\n    # step function to be used by scan\n    # arguments    | sequences |outputs-info| non-seqs\n    def _step_slice(m_, x_, xx_,  h_,          U, Ux):\n        preact = tensor.dot(h_, U)\n        preact += x_\n\n        # reset and update gates\n        r = tensor.nnet.sigmoid(_slice(preact, 0, dim))\n        u = tensor.nnet.sigmoid(_slice(preact, 1, dim))\n\n        # compute the hidden state proposal\n        preactx = tensor.dot(h_, Ux)\n        preactx = preactx * r\n        preactx = preactx + xx_\n\n        # hidden state proposal\n        h = tensor.tanh(preactx)\n\n        # leaky integrate and obtain next hidden state\n        h = u * h_ + (1. - u) * h\n        h = m_[:, None] * h + (1. - m_)[:, None] * h_\n\n        return h\n\n    # prepare scan arguments\n    seqs = [mask, state_below_, state_belowx]\n    _step = _step_slice\n    shared_vars = [tparams[_p(prefix, 'U')],\n                   tparams[_p(prefix, 'Ux')]]\n\n    init_state = tensor.unbroadcast(tensor.alloc(0., n_samples, dim), 0)\n\n    rval, updates = theano.scan(_step,\n                                    sequences=seqs,\n                                    outputs_info=[init_state],\n                                    non_sequences=shared_vars,\n                                    name=_p(prefix, '_layers'),\n                                    n_steps=nsteps,\n                                    strict=True)\n    return rval\n<\/code><\/pre>\n\n<hr>\n\n<p>As a side note, Keras fixed this issue as <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/backend\/theano_backend.py#L433\" rel=\"nofollow noreferrer\">follows<\/a>:<\/p>\n\n<pre><code>results, _ = theano.scan(\n    _step,\n    sequences=inputs,\n    outputs_info=[None] + initial_states,\n    go_backwards=go_backwards)\n\n# deal with Theano API inconsistency\nif type(results) is list:\n    outputs = results[0]\n    states = results[1:]\nelse:\n    outputs = results\n    states = []\n<\/code><\/pre>\n"}]}
{"QuestionId":34005271,"AnswerCount":1,"Tags":"<python><numpy><linear-algebra><theano><numpy-einsum>","CreationDate":"2015-11-30T18:16:22.690","AcceptedAnswerId":34009857.0,"OwnerUserId":699814.0,"Title":"Theano version of a numpy einsum for two 3dim matrices","Body":"<p>I have two 3dim numpy matrices and I want to do a dot product according to one axis without using a loop in theano. a numpy solution with sample data would be like:<\/p>\n\n<pre><\/pre>\n\n<p>where \"a\", has the shape of (7, 4, 15) and \"b\", has the shape of (7, 4, 5). \"c\", is defined as dot product of \"a\" and \"b\":<\/p>\n\n<pre><\/pre>\n\n<p>I am looking for a theano implementation of this example to calculate \"c\".<\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"34009857","CreationDate":"2015-11-30T23:21:53.897","ParentId":null,"OwnerUserId":"975477","Title":null,"Body":"<p>To finish out this question:<\/p>\n\n<pre><code>import theano as th\nimport then.Tensor as T\n\nta = T.tensor3('a')\ntb = T.tensor3('b')\n\ntc = T.batched_tensordot(ta, tb, axes=[[1],[1]])\n<\/code><\/pre>\n\n<p>......<\/p>\n"}]}
{"QuestionId":34005460,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-11-30T18:28:23.533","AcceptedAnswerId":34012696.0,"OwnerUserId":3928385.0,"Title":"Theano -- Unresolved symbol when multiplying two matrices. All works for vectors and tensor3","Body":"<p>When I try to run the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I get the following error and stack trace:<\/p>\n\n<pre><\/pre>\n\n<p>If I replace  with  or  (and supply properly shaped tensors to the function call), it compiles and runs. For  none of , ,  work, but addition and subtraction work.<\/p>\n\n<p>I ran , and the errors stayed.  reports my current version to be . Since it only fails for matrices, could it be because of some library  depends on that I somehow happen to have misconfigured?<\/p>\n","answers":[{"AnswerId":"34012696","CreationDate":"2015-12-01T04:36:46.350","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>What I did wrong was <code>pip install --upgrade theano<\/code>. Turned out I already had theano 0.7, which is the latest, so <code>pip install --upgrade<\/code> did nothing (but it was not apparent because it did upgrade the dependencies).<\/p>\n\n<p><code>pip uninstall theano &amp;&amp; pip install theano<\/code> fixed the problem<\/p>\n"}]}
{"QuestionId":34007395,"AnswerCount":2,"Tags":"<eigenvalue><tensorflow>","CreationDate":"2015-11-30T20:27:22.540","AcceptedAnswerId":null,"OwnerUserId":920253.0,"Title":"Eigendecomposition in tensorflow from python API","Body":"<p>What would be the best way to achieve an iterative eigendecomposition using TensorFlow? (something like matlabs eigs).<\/p>\n\n<p>I know that Eigen has the ability to perform an eigendecomposition, but it is unclear to me how to access this from the tensorflow python API. If tensorflow is just acting as an API to Eigen, then would there be any advantage over simply using scipy.sparse.linalg.eigs?<\/p>\n","answers":[{"AnswerId":"42821701","CreationDate":"2017-03-15T22:14:12.073","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow does not currently expose a generic eigendecomposition operator. There are two main options:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">Add a new C++ operator<\/a> that invokes Eigen's eigendecomposition functions. Since this may be generally useful, you might consider contributing it to the project in a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pulls\" rel=\"nofollow noreferrer\">pull request<\/a>.<\/p><\/li>\n<li><p>Use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\"><code>tf.py_func()<\/code><\/a> op to run a Python function in your graph, and have this function call <a href=\"https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.sparse.linalg.eigs.html\" rel=\"nofollow noreferrer\"><code>scipy.sparse.linalg.eigs()<\/code><\/a>.<\/p><\/li>\n<\/ul>\n\n<p>The main benefit of the C++ operator is that it would avoid serialization overhead (copying the arrays to and from Python), and it would be portable to other platforms (e.g. Android, iOS) that do not always have a Python interpreter available.<\/p>\n"},{"AnswerId":"34425410","CreationDate":"2015-12-22T22:19:35.927","ParentId":null,"OwnerUserId":"278046","Title":null,"Body":"<p>tf.self_adjoint_eig and tf.batch_self_adjoint_eig are all I see in the API, \n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html<\/a><\/p>\n"}]}
{"QuestionId":34010859,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><mnist>","CreationDate":"2015-12-01T01:02:10.287","AcceptedAnswerId":null,"OwnerUserId":5623540.0,"Title":"tensorflow beginner tutorial - read_data_sets fails","Body":"<p>I'm trying to run the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">tutorial<\/a> on Linux. I installed , , , . <\/p>\n\n<p>I can import the data but there seems to be some kind of problem unpacking it. <\/p>\n\n<p>Can anyone help? <\/p>\n\n<hr>\n\n<pre><\/pre>\n\n<hr>\n","answers":[{"AnswerId":"34060136","CreationDate":"2015-12-03T07:29:56.777","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This appears to be an issue with the latest version of Numpy. A <a href=\"https:\/\/github.com\/numpy\/numpy\/commit\/c87c7a612ee7f90281797e447255c402b099c73a#diff-ed156d214ca580fa436c7859cd48d21cL1028\" rel=\"noreferrer\">recent change<\/a> made it an error to treat a single-element array as a scalar for the purposes of indexing.<\/p>\n\n<p>I have made the relevant change to the upstream TensorFlow code, but in the meantime you can edit <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/bb7a7a8858dc18ba733ed64e0733e27a4224ece8\/tensorflow\/g3doc\/tutorials\/mnist\/input_data.py#L45\" rel=\"noreferrer\">this line in <code>input_data.py<\/code><\/a> (L45) to be the following (adding <code>[0]<\/code> at the end of the line):<\/p>\n\n<pre><code>return numpy.frombuffer(bytestream.read(4), dtype=dt)[0]\n<\/code><\/pre>\n"}]}
{"QuestionId":34010935,"AnswerCount":1,"Tags":"<neural-network><tensorflow>","CreationDate":"2015-12-01T01:10:21.997","AcceptedAnswerId":34012242.0,"OwnerUserId":2879934.0,"Title":"TensorFlow XOR code works fine with two dimensional target but not without?","Body":"<p>Trying to implement a very basic XOR FFNN in TensorFlow. I may just be misunderstanding the code but can anyone see an obvious reason why this won't work-- blows up to NaNs and starts with loss of $0$.\nToggles are on works\/ doesn't work if you want to mess around with it.\nThanks!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34012242","CreationDate":"2015-12-01T03:47:15.530","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> For this to work, you should use<\/p>\n\n<pre><code>loss = tf.nn.l2_loss(logits - y_input)\n<\/code><\/pre>\n\n<p>...instead of <code>tf.nn.softmax_cross_entropy_with_logits<\/code>.<\/p>\n\n<p>The <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"nofollow\"><code>tf.nn.softmax_cross_entropy_with_logits<\/code><\/a> operator expects the logits and labels inputs to be a matrix of size <code>batch_size<\/code> by <code>num_classes<\/code>. Each row of logits is an unscaled probability distribution across the classes; and each row of labels is a one-hot encoding of the true class for each example in the batch. If the inputs do not match these assumptions, the training process may diverge.<\/p>\n\n<p>In this code, the logits are <code>batch_size<\/code> by 1, which means that there is only a single class, and the softmax outputs a prediction of class 0 for all of the examples; the labels are not one-hot. If you look at the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3972c791b9f4d9a61b9ad6399b481df396f359ff\/tensorflow\/core\/kernels\/xent_op.h#L107\" rel=\"nofollow\">implementation of the operator<\/a>, the backprop value for <code>tf.nn.softmax_cross_entropy_with_logits<\/code> is:<\/p>\n\n<pre><code>\/\/ backprop: prob - labels, where\n\/\/   prob = exp(logits - max_logits) \/ sum(exp(logits - max_logits))\n<\/code><\/pre>\n\n<p>This will be <code>[[1], [1], [1], [1]] - [[0], [1], [1], [0]]<\/code> in every step, which clearly does not converge.<\/p>\n"}]}
{"QuestionId":34010962,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><operand>","CreationDate":"2015-12-01T01:13:42.023","AcceptedAnswerId":null,"OwnerUserId":4625730.0,"Title":"Tesnor Flow unsupported opperand","Body":"<p>I am going to use tensor flow package to run models\/rnn\/ptb\/\nptb_worl_lm.py.\nHowever, I have got this error in seq2Seq.py,<\/p>\n\n<blockquote>\n  <p>line 653, in sequence_loss_by_example<\/p>\n  \n  <p>log_perps \/= total_size<\/p>\n  \n  <p>TypeError: unsupported operand type(s) for \/=: 'Tensor' and 'Tensor'<\/p>\n<\/blockquote>\n\n<p>I am using Ubuntu, and python 2.7.<\/p>\n","answers":[{"AnswerId":"34013929","CreationDate":"2015-12-01T06:25:56.737","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Are you running with the released version of tensorflow and using a post-release model, by chance?  This sounds a lot like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/293\" rel=\"nofollow\">Github issue 293<\/a>.  My suggestion would be to either:  (a)  Update your install;  (b)  Try removing the <code>from __future__ import division<\/code> from the top of the file; or (c) changing the line to invoke the underlying <code>log_perps = tf.div(log_perps, total_size)<\/code> function directly.<\/p>\n\n<p>(b) or (c) is the fastest fix, but in the long run, I'd go with (a).<\/p>\n"}]}
{"QuestionId":34010987,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-01T01:16:08.100","AcceptedAnswerId":34012564.0,"OwnerUserId":3444294.0,"Title":"Does Tensorflow rerun for each eval() call?","Body":"<p>In Tensorflow and Python, I am doing the following sort of thing to understand Tensorflow and debug code by observing multiple variables at the end of a computation. <\/p>\n\n<pre><\/pre>\n\n<p>Does Tensorflow rerun the entire graph computation for each  call? Seems inefficient to rerun the entire graph just to print out one variable (tensor). If that is what is happening, is there a way to run the graph\/process\/eval once and then print out the values of each variable, without rerunning the entire graph?<\/p>\n","answers":[{"AnswerId":"34012564","CreationDate":"2015-12-01T04:22:11.790","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>When you call <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/framework.html#Tensor.eval\" rel=\"nofollow\"><code>Tensor.eval()<\/code><\/a>, TensorFlow (i) works out what subgraph of the whole graph it needs to run to produce the value of that tensor, then (ii) runs that entire graph.<\/p>\n\n<p>It is often more efficient to use <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow\"><code>Session.run()<\/code><\/a> to fetch the values of multiple tensors at once. For example, you could rewrite your code as follows to run the graph once:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    val1, val2 = sess.run([var1, var2], {x:myInputs, y:myOutputs})\n    print \"var1 =\", val1\n    print \"var2 =\", val2\n<\/code><\/pre>\n"}]}
{"QuestionId":34013520,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-12-01T05:50:20.110","AcceptedAnswerId":34014103.0,"OwnerUserId":773226.0,"Title":"Interpreting Caffe models","Body":"<p>I am trying to interpret and understand the models that are written in Caffe .proto.<\/p>\n\n<p>Yesterday I came across a sample  by <a href=\"https:\/\/stackoverflow.com\/users\/1714410\/shai\">Shai<\/a> in <a href=\"https:\/\/stackoverflow.com\/a\/33986024\/773226\">here<\/a>,  as quoted below:<\/p>\n\n<pre><\/pre>\n\n<p>I understand this model definition as:<\/p>\n\n<pre><\/pre>\n\n<p>blob  is trained by layer , with weights initialized with (lr:1) and bias initialized with (lr:2).\nblob  is actually the new learned weights which was initialized with . The learned bias doesn't have a name.<\/p>\n\n<p>In some models, we can find some layers have:<\/p>\n\n<pre><\/pre>\n\n<p>Where the first instance of  always correspond to weights and the next instance for bias.<\/p>\n\n<p>Are my above understandings correct?<\/p>\n","answers":[{"AnswerId":"34014103","CreationDate":"2015-12-01T06:38:30.140","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are mixing two data types: the input (training) data and the net's parameters.<br>\nDuring training the input data is fixed to a known training\/validation set and only the net parameters are changed. In contrast, when deploying the net, the data changes to new images while the net parameters are fixed. See <a href=\"https:\/\/stackoverflow.com\/a\/31847179\/1714410\">this answer<\/a> for some in-depth description of the way caffe stores these two types of data.<\/p>\n\n<p>In the example you showed, there are two input training data paths: <code>data_a<\/code> and <code>data_b<\/code> that might be <em>different<\/em> images each time. The input blobs pass through an InnerProduct layer to become <code>ip1_a<\/code> and <code>ip1_b<\/code> blobs respectively. Then they are concatenated into a single blob <code>ip1_combined<\/code> which in turn is fed into the final InnerProduct layer.<\/p>\n\n<p>On the other hand, the model have a set of parameters: <code>ip1_w<\/code> and <code>ip1_b<\/code> (the weights and bias) of the first inner product layer. In this particular examples the parameters of the layer were explicitly named to indicate the fact that they are <em>shared<\/em> between <code>ip1_a<\/code> and <code>ip1_b<\/code> layers.<\/p>\n\n<p>As for the two <code>lr_mult<\/code>: then yes, the first is the LR multiplier of the weights, and the second one is for the bias term.<\/p>\n"}]}
{"QuestionId":34015467,"AnswerCount":1,"Tags":"<machine-learning><string-matching><tensorflow><recurrent-neural-network>","CreationDate":"2015-12-01T08:07:53.330","AcceptedAnswerId":null,"OwnerUserId":5546193.0,"Title":"String Matching Using Recurrent Neural Networks","Body":"<p>I have recently started exploring Recurrent Neural Networks. So far I have trained character level language model on tensorFlow using Andrej Karpathy's <a href=\"http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"noreferrer\">blog<\/a>. It works great.<\/p>\n\n<p>I couldnt however find any study on using RNNs for string matching or keyword spotting. For one of my project I require OCR of scanned documents and then parsing the converted text for key data points. Most string matching techniques fail to incorporate the OCR conversion mistakes and that leads to significant error.<\/p>\n\n<p>Is it possible to train the RNN on the variations of converted text I receive and use it for finding keywords.<\/p>\n","answers":[{"AnswerId":"51373784","CreationDate":"2018-07-17T05:27:38.827","ParentId":null,"OwnerUserId":"1851492","Title":null,"Body":"<p>This paper may the thing you are looking for: <\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1608.02214\" rel=\"nofollow noreferrer\">[1608.02214] Robsut Wrod Reocginiton via semi-Character Recurrent Neural Network<\/a><\/p>\n\n<p>A Brief introduction:<\/p>\n\n<p>The author of this paper demonstrated a method to recognize jumbled words which like Cmabrigde Uinervtisy(Cambridge  University). Training the neural network with correct begin, end characters and the encoded internal characters which doesn't contain it's position information, the neural network can learn to recognize and correct it.<\/p>\n\n<p>You can easily modify the network structure to adapt your own need, the OCR, as you had mentioned.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bCMyK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bCMyK.png\" alt=\"\"><\/a><br>\n<sub>(source: <a href=\"https:\/\/screenshotscdn.firefoxusercontent.com\/images\/24464385-3bef-4555-80ce-669d252fe5fe.png\" rel=\"nofollow noreferrer\">firefoxusercontent.com<\/a>)<\/sub>  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ydmKA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ydmKA.png\" alt=\"\"><\/a><br>\n<sub>(source: <a href=\"https:\/\/screenshotscdn.firefoxusercontent.com\/images\/313b52fb-f590-4904-9144-f1044160a0df.png\" rel=\"nofollow noreferrer\">firefoxusercontent.com<\/a>)<\/sub>  <\/p>\n"}]}
{"QuestionId":34015547,"AnswerCount":0,"Tags":"<python><neural-network><theano><deep-learning>","CreationDate":"2015-12-01T08:13:40.413","AcceptedAnswerId":null,"OwnerUserId":3694557.0,"Title":"How to use a trained model to make a prediction from a given input?","Body":"<p>I am following a <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html\" rel=\"nofollow\">tutorial<\/a> with a public <a href=\"https:\/\/github.com\/mnielsen\/neural-networks-and-deep-learning\/blob\/master\/src\/network3.py\" rel=\"nofollow\">code repo<\/a> available. I have been banging my head trying to figure out how to apply the model after it has been trained. Since the tutorial only covers the training, testing and validation, I'm not sure how to use it on new unclassified data. Here's what I've tried:<\/p>\n\n<p>(1) Under the main class \"Network\" in <a href=\"https:\/\/github.com\/mnielsen\/neural-networks-and-deep-learning\/blob\/master\/src\/network3.py\" rel=\"nofollow\">network3.py<\/a> used to construct and train networks, I added the following method on :<\/p>\n\n<pre><\/pre>\n\n<p>(2) From the Python interactive console (iPython notebook), I imported the data and trained the model as usual:<\/p>\n\n<pre><\/pre>\n\n<p>(3) I then issued the command:<\/p>\n\n<pre><\/pre>\n\n<p>which caused the following error<\/p>\n\n<blockquote>\n  <p>ValueError: ('Bad input argument to theano function with name\n  \"network3.py:181\"  at index 0(0-based)', 'setting an array element\n  with a sequence.')<\/p>\n<\/blockquote>\n\n<p>It looks like my inputs are not defined or interpreted as I expected. The data I used for training, testing, validating and predicting are all the same format, i.e. col 0 is the classification and cols 1-59 are numerical inputs. In the code above, I just used the training data to do the prediction. I would greatly appreciate any help or advice to steer me in the right direction. Thanks!<\/p>\n","answers":[]}
{"QuestionId":34017379,"AnswerCount":0,"Tags":"<command-line><c-preprocessor><caffe>","CreationDate":"2015-12-01T09:52:44.957","AcceptedAnswerId":null,"OwnerUserId":4680426.0,"Title":"where to find the previous definition of #define value","Body":"<p>I have made a mistake to change the .h file for some reason. And I do the  in the file, but I forgot its name.<\/p>\n\n<p>Now I come across this problem when I make caffe test in terminal<\/p>\n\n<pre><\/pre>\n\n<p>How can I find the .h file I had changed, and I have searched in finder on MAC OS yet.<\/p>\n","answers":[]}
{"QuestionId":34020375,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-12-01T12:23:50.540","AcceptedAnswerId":34036054.0,"OwnerUserId":5625511.0,"Title":"caffe multi-label training with lmdb to classifiy facial regions","Body":"<p>I'm using two lmdb inputs for identifying eyes, nosetip and mouth regions of a face. The data lmdb is of dimension x3xx while the label lmdb is of dimension x1x\/4x\/4. The label image is created by masking regions using numbers 1-4 on an opencv Mat that was initialized to be all 0s (so in total there are 5 labels with 0 being the background label). I scaled down the label image to be 1\/4 in width and height of the corresponding image because I have 2 pooling layers in my net. This downscaling ensures the label image dimension will match the output of the last convolution layer.<\/p>\n\n<p>My train_val.prototxt:<\/p>\n\n<pre><\/pre>\n\n<p>In the last convolution layer, I set the output size to be 5 because I have 5 label classes. The training is converged with final loss at about 0.3 and accuracy 0.9 (although some sources suggest this accuracy is not correctly measured for multi-lablels). When using the trained model, the output layer correctly produces a blob of dimension 1x5x\/4x\/4 which I managed to visualze as 5 separate single channel images. However, while the first image correctly hightlighted the background pixels, the remaining 4 images looks almost the same with all 4 regions highlighted.<\/p>\n\n<p>visualization of 5 output channels(intensity increases from blue to red):<br>\n<img src=\"https:\/\/i.stack.imgur.com\/gpktM.jpg\" alt=\"click to view picture\"><\/p>\n\n<p>original image(the concentric circles labels highest intensity from each channel. some are bigger just to distinguish from others. as you can see other than the background, the rest channels have highest activations almost on the same mouth region which should not be the case. )<br>\n<img src=\"https:\/\/i.stack.imgur.com\/qVNtJ.jpg\" alt=\"click to view picture\"><\/p>\n\n<p>Could someone help me spot the mistake I made?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"34036054","CreationDate":"2015-12-02T06:14:49.187","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you are facing <a href=\"http:\/\/www.chioka.in\/class-imbalance-problem\/\" rel=\"nofollow noreferrer\"><em>class imbalance<\/em><\/a>: most of your labeled pixels are labeled 0 (Background), hence, during training the net learns to predict background almost regardless of what it \"sees\". Since predicting background is correct most of the time, the training loss decreases and the accuracy increases up to a certain point.<br>\nHowever, when you actually try to visualize the output prediction it is mostly background with little information regarding the other scarce labels. <\/p>\n\n<p>One way of tackling class imbalance, in caffe, is to use <a href=\"https:\/\/stackoverflow.com\/a\/30497907\/1714410\">\"InfogainLoss\"<\/a> layer with weights tuned to counter-effect the imbalance of the labels.<\/p>\n"}]}
{"QuestionId":34023056,"AnswerCount":2,"Tags":"<python><numpy><caffe><tensorflow>","CreationDate":"2015-12-01T14:41:46.260","AcceptedAnswerId":34024383.0,"OwnerUserId":3531608.0,"Title":"TensorFlow, how to look inside 'blob', the response in through CNN","Body":"<p>I am currently using VGG-face-discriptor from <a href=\"https:\/\/github.com\/AKSHAYUBHAT\/TensorFace\" rel=\"nofollow\">https:\/\/github.com\/AKSHAYUBHAT\/TensorFace<\/a>. <\/p>\n\n<p>VGG-face-discriptor uses VGG16 and output vector 2622, a number of class of celebrities. \n What I really need is response of second last fully connected layer, which size is 4096.\n Using the code provided from the repository I mentioned above<\/p>\n\n<pre><\/pre>\n\n<p>Works very fine, giving me the closest celebrity face result.<\/p>\n\n<p>result:<\/p>\n\n<pre><\/pre>\n\n<p>Looking at the  varaible, I see 2622 numpy ndarry.\nBut I actually want the 2nd last feature vector.. How can I achieve this?<\/p>\n\n<p>I've looked around all the TensorFlow tutorial codes, but cannot find something like this.\nWith Caffe, I just <\/p>\n\n<pre><\/pre>\n\n<p>Simple was that. How can I see through 'blob' in TensorFlow?\nWith numpy array feature vector?<\/p>\n","answers":[{"AnswerId":"34024038","CreationDate":"2015-12-01T15:31:37.767","ParentId":null,"OwnerUserId":"3531608","Title":null,"Body":"<p>Oh.. I think I found solution.<\/p>\n\n<p>Just removing some lines in TensorFace\/vggface\/<strong>init<\/strong>.py<\/p>\n\n<pre><code># (37): nn.ReLU\n# (38): nn.Dropout(0.500000)\n# (39): nn.Linear(4096 -&gt; 2622)\n#self.layers.append(('linear','39',2622,False))\n# (40): nn.SoftMax\n#self.layers.append(('softmax'))\n<\/code><\/pre>\n\n<p>looks working<\/p>\n"},{"AnswerId":"34024383","CreationDate":"2015-12-01T15:48:54.387","ParentId":null,"OwnerUserId":"5612876","Title":null,"Body":"<p>You can use session.run to get the current values of elements in your computation graph. <\/p>\n\n<pre><code>layer7_values = session.run(layer7_tf, feed_dict={&lt;your inputs&gt;})\n<\/code><\/pre>\n\n<p>In this example <code>session<\/code> is a <code>tf.Session()<\/code> object. <code>layer7_tf<\/code> is a reference to the <code>Tensor<\/code> output of a layer in the TensorFlow model, and <code>layer7_values<\/code> will contain the layer values for the given input as a numpy array.<\/p>\n\n<p>To get a handle to <code>layer7_tf<\/code>, you have a couple of options. You can either modify <code>TensorFace\/vggface\/init.py<\/code> to return a reference to the appropriate layer; or you can explore the <code>session.graph_def<\/code> structure to find the <code>name<\/code> of the node that corresponds to that tensor, and pass the string name of the tensor (e.g. <code>layer7_tf\/foo\/bar:0<\/code>, where the <code>:0<\/code> corresponds to the 0th output of the op called <code>layer7_tf\/foo\/bar<\/code>) to <code>session.run()<\/code> instead.<\/p>\n"}]}
{"QuestionId":34023805,"AnswerCount":1,"Tags":"<python><theano><elementwise-operations>","CreationDate":"2015-12-01T15:20:01.903","AcceptedAnswerId":34032627.0,"OwnerUserId":5626182.0,"Title":"theano (python): elementwise gradient","Body":"<p>I'm trying to perform elementwise gradient with<\/p>\n\n<p>e.g.,<\/p>\n\n<p>output-f(x): 5 by 1 vector,<\/p>\n\n<p>with respect to \ninput-X: 5 by 1 vector<\/p>\n\n<p>I can do this like,<\/p>\n\n<pre><\/pre>\n\n<p>and the result is<\/p>\n\n<pre><\/pre>\n\n<p>but since it's not efficient, i want to get 5 by 1 vector as a result<\/p>\n\n<p>by doing something like..<\/p>\n\n<pre><\/pre>\n\n<p>which doesn't work.<\/p>\n\n<p>Is there any way of do this?\n(sorry for bad format..I'm new here and learning)<\/p>\n\n<hr>\n\n<p>(I added more clear example):<\/p>\n\n<p>given input vector: x[1], x[2], ..., x[n]<\/p>\n\n<p>and output vector: y[1], y[2], .., y[n],<\/p>\n\n<p>where y[i] = f(x[i]).<\/p>\n\n<p>I want the result of <\/p>\n\n<p>df(x[i])\/dx[i] only<\/p>\n\n<p>and not the<\/p>\n\n<p>df(x[i])\/dx[j] for (i&lt;>j)<\/p>\n\n<p>, for computational efficiency (n is number of data > 10000)<\/p>\n","answers":[{"AnswerId":"34032627","CreationDate":"2015-12-02T00:06:07.217","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You are looking for <code>theano.tensor.jacobian<\/code>.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nx = T.fvector()\np = T.as_tensor_variable([(x ** i).sum() for i in range(5)])\n\nj = T.jacobian(p, x)\n\nf = theano.function([x], [p, j])\n<\/code><\/pre>\n\n<p>Now evaluating yields<\/p>\n\n<pre><code>In [31]: f([1., 2., 3.])\nOut[31]: \n[array([  3.,   6.,  14.,  36.,  98.], dtype=float32),\n array([[   0.,    0.,    0.],\n        [   1.,    1.,    1.],\n        [   2.,    4.,    6.],\n        [   3.,   12.,   27.],\n        [   4.,   32.,  108.]], dtype=float32)]\n<\/code><\/pre>\n\n<p>If you are interested in only one, or a few partial derivatives, you can obtain only them also. It would be necessary to take a close look at the theano optimization rules to be able to see how much more efficient this gets (a benchmark is a first test). (It is possible that already indexing into the gradient makes it clear to theano that it does not need to calculate the rest).<\/p>\n\n<pre><code>x = T.fscalar()\ny = T.fvector()\nz = T.concatenate([x.reshape((1,)), y.reshape((-1,))])\n\ne = (z ** 2).sum()\ng = T.grad(e, wrt=x)\n\nff = theano.function([x, y], [e, g])\n<\/code><\/pre>\n"}]}
{"QuestionId":34030140,"AnswerCount":5,"Tags":"<sparse-matrix><tensorflow>","CreationDate":"2015-12-01T21:06:57.537","AcceptedAnswerId":34032563.0,"OwnerUserId":411038.0,"Title":"Is sparse tensor multiplication implemented in TensorFlow?","Body":"\n\n<p>Multiplication of sparse tensors with themselves or with dense tensors does not seem to work in TensorFlow.  The following example<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>fails with the error message<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Both tensors have values of type float32 as seen by evaluating them  without the multiplication op.  Multiplication of y with itself returns a similar error message.  Multipication of x with itself works fine.<\/p>\n","answers":[{"AnswerId":"57097632","CreationDate":"2019-07-18T15:08:07.817","ParentId":null,"OwnerUserId":"7548610","Title":null,"Body":"<p>tf.sparse_matmul is for multiplying two dense tensor not sparse type of data structure. That function is just an optimized version of tensor multiplication if the given matrix (or both two matrixes) have many zero value. Again it does not accept sparse tensor data type. It accepts dense tensor data type. It might fasten your calculations if values are mostly zero.<\/p>\n\n<p>as far as I know there is no implementation of two sparse type tensor muliplication. but just one sparse one dense which is tf.sparse_tensor_dense_matmul(x, y) !<\/p>\n"},{"AnswerId":"38100857","CreationDate":"2016-06-29T13:21:30.073","ParentId":null,"OwnerUserId":"198401","Title":null,"Body":"<p>Recently, <code>tf.sparse_tensor_dense_matmul(...)<\/code> was added that allows multiplying a sparse matrix by a dense matrix.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/sparse_ops.html#sparse_tensor_dense_matmul\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/sparse_ops.html#sparse_tensor_dense_matmul<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1241\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1241<\/a><\/p>\n"},{"AnswerId":"44473041","CreationDate":"2017-06-10T11:56:06.490","ParentId":null,"OwnerUserId":"2238038","Title":null,"Body":"<p>To make the answer more complete:<\/p>\n\n<pre><code>tf.sparse_matmul(\n    a,\n    b,\n    transpose_a=None,\n    transpose_b=None,\n    a_is_sparse=None,\n    b_is_sparse=None,\n    name=None\n)\n<\/code><\/pre>\n\n<p>exists as well:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_matmul\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_matmul<\/a><\/p>\n"},{"AnswerId":"44897976","CreationDate":"2017-07-04T06:00:06.193","ParentId":null,"OwnerUserId":"2567512","Title":null,"Body":"<p>It seems that<\/p>\n\n<pre><code>tf.sparse_matmul(\n    a,\n    b,\n    transpose_a=None,\n    transpose_b=None,\n    a_is_sparse=None,\n    b_is_sparse=None,\n    name=None\n)\n<\/code><\/pre>\n\n<p>is not for multiplication of two <code>SparseTensors<\/code>.<\/p>\n\n<p><code>a<\/code> and <code>b<\/code> are <code>Tensors<\/code> not <code>SparseTensors<\/code>. And I have tried that, it is not working with <code>SparseTensors<\/code>.<\/p>\n"},{"AnswerId":"34032563","CreationDate":"2015-12-01T23:59:01.930","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>General-purpose multiplication for <code>tf.SparseTensor<\/code> is not currently implemented in TensorFlow. However, there are three partial solutions, and the right one to choose will depend on the characteristics of your data:<\/p>\n\n<ul>\n<li><p>If you have a <code>tf.SparseTensor<\/code> and a <code>tf.Tensor<\/code>, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/sparse_ops\/math_operations#sparse_tensor_dense_matmul\" rel=\"noreferrer\"><code>tf.sparse_tensor_dense_matmul()<\/code><\/a> to multiply them. This is more efficient than the next approach if one of the tensors is too large to fit in memory when densified: the documentation has more guidance about how to decide between these two methods. Note that it accepts a <code>tf.SparseTensor<\/code> as the <strong>first<\/strong> argument, so to solve your exact problem you will need to use the <code>adjoint_a<\/code> and <code>adjoint_b<\/code> arguments, and transpose the result.<\/p><\/li>\n<li><p>If you have two sparse tensors and need to multiply them, the simplest (if not the most performant) way is to convert them to dense and use <code>tf.matmul<\/code>:<\/p>\n\n<pre><code>a = tf.SparseTensor(...)\nb = tf.SparseTensor(...)\n\nc = tf.matmul(tf.sparse_tensor_to_dense(a, 0.0),\n              tf.sparse_tensor_to_dense(b, 0.0),\n              a_is_sparse=True, b_is_sparse=True)\n<\/code><\/pre>\n\n<p>Note that the optional <code>a_is_sparse<\/code> and <code>b_is_sparse<\/code> arguments mean that \"<code>a<\/code> (or <code>b<\/code>) has a dense representation but a large number of its entries are zero\", which triggers the use of a different multiplication algorithm.<\/p><\/li>\n<li><p>For the special case of sparse <em>vector<\/em> by (potentially large and sharded) dense matrix multiplication, and the values in the vector are 0 or 1, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/embeddings#embedding_lookup\" rel=\"noreferrer\"><code>tf.nn.embedding_lookup<\/code><\/a> operator may be more appropriate. <a href=\"http:\/\/www.tensorflow.org\/tutorials\/word2vec\/index.html#building_the_graph\" rel=\"noreferrer\">This tutorial<\/a> discusses when you might use embeddings and how to invoke the operator in more detail.<\/p><\/li>\n<li><p>For the special case of sparse <em>matrix<\/em> by (potentially large and sharded) dense matrix, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/embeddings#embedding_lookup_sparse\" rel=\"noreferrer\"><code>tf.nn.embedding_lookup_sparse()<\/code><\/a> may be appropriate. This function accepts one or two <code>tf.SparseTensor<\/code> objects, with <code>sp_ids<\/code> representing the non-zero values, and the optional <code>sp_weights<\/code> representing their values (which otherwise default to one).<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":34032158,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-12-01T23:24:47.473","AcceptedAnswerId":null,"OwnerUserId":824624.0,"Title":"RuntimeError: module compiled against API version a but this version of numpy is 9 in ubuntu","Body":"<p>I just installed theano, but <\/p>\n\n<pre><\/pre>\n\n<p>here is the output<\/p>\n\n<pre><\/pre>\n\n<p>how to solve it?<\/p>\n","answers":[{"AnswerId":"37343218","CreationDate":"2016-05-20T09:38:47.693","ParentId":null,"OwnerUserId":"6220857","Title":null,"Body":"<p>Package management in python is sometimes tricky. To avoid these issues, I recommend using Anaconda which is a good python package and environment manager (<a href=\"https:\/\/www.continuum.io\/downloads\" rel=\"nofollow\">download here<\/a>, and <a href=\"http:\/\/conda.pydata.org\/docs\/test-drive.html\" rel=\"nofollow\">30 minutes guide<\/a>).<\/p>\n\n<p>Follow the official instructions to install Anaconda and chose the default values that will add Anaconda to your $PATH and make Anaconda your default python interpreter. Then open your terminal and use the following commands :<\/p>\n\n<ul>\n<li><code>conda install pip six nose numpy scipy<\/code> <\/li>\n<li><code>pip install theano<\/code><\/li>\n<\/ul>\n\n<p>Then try running your script using anaconda.<\/p>\n"}]}
{"QuestionId":34035632,"AnswerCount":1,"Tags":"<python><gpgpu><theano>","CreationDate":"2015-12-02T05:36:40.353","AcceptedAnswerId":null,"OwnerUserId":4127806.0,"Title":"theano gives \"...Waiting for existing lock by unknown process...\"","Body":"<p>My code was working fine. However, now I am getting an error that says:<\/p>\n\n<pre><\/pre>\n\n<p>I delete the folder lock_dir, but I still get the same message. <\/p>\n","answers":[{"AnswerId":"35458973","CreationDate":"2016-02-17T14:09:32.737","ParentId":null,"OwnerUserId":"5940875","Title":null,"Body":"<p>I solved this problem by adding the following code to my  ~\/.theanorc :<\/p>\n\n<p>[global]<br>\n base_compiledir=\/tmp\/%(user)s\/theano.NOBACKUP<\/p>\n"}]}
{"QuestionId":34036689,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2015-12-02T06:58:38.730","AcceptedAnswerId":null,"OwnerUserId":66455.0,"Title":"Example for Deploying a Tensorflow Model via a RESTful API","Body":"<p>Is there any example code for deploying a Tensorflow Model via a RESTful API? I see examples for a command line program and for a mobile app. Is there a framework for this or people just load the model and expose the predict method via a web framework (like Flask)to take input (say via JSON) and return the response? By framework I mean scaling for large number of predict requests. Of course since the models are immutable we can launch multiple instances of our prediction server and put it behind a load balancer (like HAProxy). My question is, are people using some framework for this or doing this from scratch, or, maybe this is already available in Tensorflow and I have not noticed it.<\/p>\n","answers":[{"AnswerId":"36819551","CreationDate":"2016-04-24T05:18:18.450","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/sugyan\/tensorflow-mnist\" rel=\"noreferrer\">https:\/\/github.com\/sugyan\/tensorflow-mnist<\/a> shows a simple restAPI example by using Flask and loading pre-trained mode (restore).<\/p>\n\n<pre><code>@app.route('\/api\/mnist', methods=['POST'])\ndef mnist():\n    input = ((255 - np.array(request.json, dtype=np.uint8)) \/ 255.0).reshape(1, 784)\n    output1 = simple(input)\n    output2 = convolutional(input)\n    return jsonify(results=[output1, output2])\n<\/code><\/pre>\n\n<p>Also, see the online demo at <a href=\"https:\/\/tensorflow-mnist.herokuapp.com\/\" rel=\"noreferrer\">https:\/\/tensorflow-mnist.herokuapp.com\/<\/a>. It seems the API is fast enough.<\/p>\n"},{"AnswerId":"35440410","CreationDate":"2016-02-16T18:31:33.163","ParentId":null,"OwnerUserId":"1397927","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"noreferrer\">TensorFlow Serving<\/a> is a high performance, open source serving system for machine learning models, designed for production environments and optimized for TensorFlow. The initial release contains examples built with <a href=\"http:\/\/www.grpc.io\/\" rel=\"noreferrer\">gRPC<\/a>, but you can easily replace the front-end (denoted as \"client\" in the diagram below) with a RESTful API to suit your needs. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YKWTf.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YKWTf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To get started quickly, check out the <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"noreferrer\">tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":34045096,"AnswerCount":1,"Tags":"<python><theano><mnist>","CreationDate":"2015-12-02T14:22:30.807","AcceptedAnswerId":null,"OwnerUserId":5404771.0,"Title":"Theano CUDA exception","Body":"<p>I'm relatively new to theano and I want to run the mnist example on my\nGPU but I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>It keeps outputting DEBUG messages without any output of mnist.\nI have a working version of nvcc:<\/p>\n\n<pre><\/pre>\n\n<p>And my  file:<\/p>\n\n<pre><\/pre>\n\n<p>How can I solve this?<\/p>\n","answers":[{"AnswerId":"36527095","CreationDate":"2016-04-10T07:30:16.570","ParentId":null,"OwnerUserId":"1105231","Title":null,"Body":"<p>I have a similar problem. Search google and get the code. <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/nvcc_compiler.py\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/sandbox\/cuda\/nvcc_compiler.py<\/a><\/p>\n\n<pre><code>        p = subprocess.Popen(\n                cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)\n        nvcc_stdout_raw, nvcc_stderr_raw = p.communicate()[:2]\n        console_encoding = getpreferredencoding()\n        nvcc_stdout = decode_with(nvcc_stdout_raw, console_encoding)\n        nvcc_stderr = decode_with(nvcc_stderr_raw, console_encoding)\n\n    if nvcc_stdout:\n        # this doesn't happen to my knowledge\n        print(\"DEBUG: nvcc STDOUT\", nvcc_stdout, file=sys.stderr)\n<\/code><\/pre>\n\n<p>It seems to nvcc have error output.<\/p>\n\n<p>But in my occasion, its output looks like<\/p>\n\n<blockquote>\n  <p>DEBUG: nvcc STDOUT mod.cu<\/p>\n  \n  <p>DEBUG: nvcc STDOUT mod.cu<\/p>\n<\/blockquote>\n\n<p>Sometimes the program work fine after this, sometimes doesn't work. It's very strange.\nSorry I can not make comment, so I just post the answer.<\/p>\n"}]}
{"QuestionId":34045295,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-02T14:31:56.183","AcceptedAnswerId":null,"OwnerUserId":5630018.0,"Title":"Multivariate time-series RNN using Tensorflow. Is this possible with an LSTM cell or similar?","Body":"<p>I am looking for examples of how to build a multivariate time-series RNN using Tensorflow. Is this possible with an LSTM cell or similar? <\/p>\n\n<p>e.g. the data might look something like this:<\/p>\n\n<p>Time,A,B,C,...<\/p>\n\n<p>0,3.5,4.5,7.7,...<\/p>\n\n<p>1,2.1,6.4,8.2,...<\/p>\n\n<p>...<\/p>\n\n<p>Any help much appreciated.  Thanks, John<\/p>\n","answers":[{"AnswerId":"34081427","CreationDate":"2015-12-04T05:22:16.183","ParentId":null,"OwnerUserId":"5545285","Title":null,"Body":"<p>It depends on exactly what you mean, but yes, it should be possible. If you write more specifically how exactly your input and target data looks like, somebody may be able to help. You can generally have sequential continuous or categorical input data and sequential continuous or categorical output data or a mix of those. I would suggest you look at the tutorials and try out a few things, then ask again here.<\/p>\n"},{"AnswerId":"34090025","CreationDate":"2015-12-04T13:58:38.477","ParentId":null,"OwnerUserId":"5630018","Title":null,"Body":"<p>Thanks.  I have figured it out now.  I misunderstood the docs 'inputs: A length T list of inputs, each a vector with shape [batch_size].'  <\/p>\n\n<p>The following link was useful:<\/p>\n\n<p><a href=\"https:\/\/m.reddit.com\/r\/MachineLearning\/comments\/3sok8k\/tensorflow_basic_rnn_example_with_variable_length\/\" rel=\"nofollow\">https:\/\/m.reddit.com\/r\/MachineLearning\/comments\/3sok8k\/tensorflow_basic_rnn_example_with_variable_length\/<\/a><\/p>\n"}]}
{"QuestionId":34046048,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-02T15:07:03.407","AcceptedAnswerId":34049684.0,"OwnerUserId":2615676.0,"Title":"Debugging nans in the backward pass","Body":"<p>I'm trying to debug a somewhat complicated and non-canonical NN architecture. Computing the forward pass is fine and is giving me the expected results, but when I try to optimize using Adam or any of the standard optimizers, even after one iteration with a very small learning rate I get nans everywhere. I'm trying to localize them and was wondering if there's a way to catch the first occurrence of a nan and detect in which op it arose? I tried  but it doesn't appear to be doing anything, or perhaps I'm using it incorrectly.<\/p>\n","answers":[{"AnswerId":"34049684","CreationDate":"2015-12-02T17:55:31.693","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>Debugging NaNs can be tricky, especially if you have a large network. <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops.html#add_check_numerics_ops\" rel=\"noreferrer\"><code>tf.add_check_numerics_ops()<\/code><\/a> adds ops to the graph that assert that each floating point tensor in the graph does not contain any NaN values, but does not run these checks by default. Instead it returns an op that you can run periodically, or on every step, as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_op = ...\ncheck_op = tf.add_check_numerics_ops()\n\nsess = tf.Session()\nsess.run([train_op, check_op])  # Runs training and checks for NaNs\n<\/code><\/pre>\n"},{"AnswerId":"34057106","CreationDate":"2015-12-03T03:00:52.677","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"\n\n<p>Maybe you could add Print ops to suspect ops print values, something like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print_ops = []\nfor op in ops:\n  print_ops.append(tf.Print(op, [op],\n                   message='%s :' % op.name, summarize=10))\nprint_op = tf.group(*print_ops)\nsess.run([train_op, print_op])\n<\/code><\/pre>\n\n<p>To add to all ops, you could do a loop along the lines of <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/python\/ops\/numerics.py#63\" rel=\"nofollow noreferrer\"><code>add_check_numerics_ops<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":34047152,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-02T15:55:49.780","AcceptedAnswerId":34073808.0,"OwnerUserId":2286138.0,"Title":"tutorials_example_trainer fails in debug mode (-c dbg)","Body":"<p>The build for tutorials_example_trainer works fine in release mode (-c opt), but fails in debug mode (-c dbg).<\/p>\n\n<p>Did anyone encounter this? It seems to be a bug.<\/p>\n\n<p>The command I run:<\/p>\n\n<pre><\/pre>\n\n<p>The build fails with the following message:<\/p>\n\n<blockquote>\n  <p>\/usr\/include\/c++\/4.8\/mutex(125) (col. 5): error: calling a <strong>host<\/strong>\n  function(\"std::<strong>mutex_base::__mutex_base [subobject]\") from a\n  __device<\/strong> function(\"std::mutex::mutex\") is not allowed<\/p>\n<\/blockquote>\n\n<p>&lt; some warnings><\/p>\n\n<blockquote>\n  <p>1 error detected in the compilation of\n  \"\/tmp\/tmpxft_00005e78_00000000-10_cwise_op_gpu_log.cu.compute_52.cpp1.ii\".\n  ERROR:\n  \/home\/uriv\/git\/tensorflow\/tensorflow\/tensorflow\/core\/BUILD:248:1:\n  output\n  'tensorflow\/core\/_objs\/gpu_kernels\/tensorflow\/core\/kernels\/cwise_op_gpu_log.cu.pic.o'\n  was not created. ERROR:\n  \/home\/uriv\/git\/tensorflow\/tensorflow\/tensorflow\/core\/BUILD:248:1: not\n  all outputs were created.<\/p>\n<\/blockquote>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"34073808","CreationDate":"2015-12-03T18:46:32.237","ParentId":null,"OwnerUserId":"3309044","Title":null,"Body":"<p>You can workaround the problem by editing  <\/p>\n\n<p><code>tensorflow\/third_party\/eigen3\/unsupported\/Eigen\/CXX11\/src\/Tensor\/TensorDeviceType.h<\/code>  <\/p>\n\n<p>and commenting out the following 2 lines of code:<\/p>\n\n<p><code>static tensorflow::mutex m_devicePropInitMutex(tensorflow::LINKER_INITIALIZED);<\/code><\/p>\n\n<p>and<\/p>\n\n<p><code>tensorflow::mutex_lock l(m_devicePropInitMutex);<\/code><\/p>\n\n<p>I'll push a proper fix to the tensorflow repository shortly.<\/p>\n"}]}
{"QuestionId":34048833,"AnswerCount":1,"Tags":"<python><theano><sampling><convolution>","CreationDate":"2015-12-02T17:11:32.183","AcceptedAnswerId":34250856.0,"OwnerUserId":5631237.0,"Title":"Drawing samples from multinomial using Theano","Body":"<p>I'm currently implementing a convolutional RBM and I'm using Theano for that.<\/p>\n\n<p>My current implementation seems to be pretty slow and profiling showed that it's mainly due to the Gibbs sampling steps.\nIn fact, I am using Theano's <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/shared_randomstreams.html\" rel=\"nofollow\">shared randomstreams<\/a> to generate multinomial samples.<\/p>\n\n<p>However, I found an improved version of Theano's random stream <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/sandbox\/rng_mrg.html\" rel=\"nofollow\">here<\/a> which meets all the performance demands I have.<\/p>\n\n<p>Unfortunately, this experimental random generator only supports two-dimensional matrices and I have to use it for a tensor4-object (4D-Matrix) as this is the result from Theano's nnet conv2d operation.<\/p>\n\n<p>Do you know if there's an efficient way to draw samples from a 4D-matrix that has the following structure:<\/p>\n\n<p>Samplesize x 1 x N x M<\/p>\n\n<p>where I would like to draw from one of columns (third dimension), giving me a code like this:<\/p>\n\n<pre><\/pre>\n\n<p>But this code would be really slow and I'd like to do this efficiently and on the GPU.<\/p>\n\n<p>So any help would be much appreciated.<\/p>\n","answers":[{"AnswerId":"34250856","CreationDate":"2015-12-13T12:02:09.840","ParentId":null,"OwnerUserId":"5631237","Title":null,"Body":"<p>So I found the solution that worked out for me with a rather simple dimshuffle\/reshape combination that is undone after the sampling.\n<\/p>\n\n<pre><code>def sampleVisibleLayer (self, V):\n    reshaped = V.dimshuffle(0, 1, 3, 2).reshape((V.shape[0]*V.shape[3], V.shape[2]))\n    S_reshaped = self.theano_rng.multinomial(n=1,pvals=reshaped)\n    S = S_reshaped.reshape((V.shape[0], 1, V.shape[3], V.shape[2])).dimshuffle(0, 1, 3, 2)\n<\/code><\/pre>\n\n<p>This solution worked well for me even though the batch size has some limits following this approach.\nSince the reshaped matrix can become very large, we can run into situations were the random generator raises an error message even though it's not mentioned in the documentation.<\/p>\n\n<p>The solution is also quite fast since dimshuffle and reshape are performed in O(1).<\/p>\n"}]}
{"QuestionId":34050071,"AnswerCount":6,"Tags":"<python><tensorflow>","CreationDate":"2015-12-02T18:15:36.273","AcceptedAnswerId":null,"OwnerUserId":1973820.0,"Title":"TensorFlow random_shuffle_queue is closed and has insufficient elements","Body":"<p>I'm reading batch of images by getting idea <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/fully_connected_reader.py\">here<\/a> from tfrecords(converted by <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/convert_to_records.py\">this<\/a>) <\/p>\n\n<p>My images are cifar images, [32, 32, 3] and as you can see while reading and passing images the shapes are normal ()<\/p>\n\n<p>the 2 most notable problems stated in the log, as far as I know is<\/p>\n\n<ol>\n<li>Shape of 12228, which I don't know from where I get this. All my tensors are either in shape [32, 32, 3] or [None, 3072]<\/li>\n<li>Running out of sample<\/li>\n<\/ol>\n\n<p><\/p>\n\n<p>How can I solve this?<\/p>\n\n<p>Logs:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43370673","CreationDate":"2017-04-12T13:16:46.173","ParentId":null,"OwnerUserId":"1861627","Title":null,"Body":"<p>I had a similar problem. Digging around the web, it turned out that if you use some <code>num_epochs<\/code> argument, you have to initialize all the <code>local<\/code> variables, so your code should end up looking like:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.local_variables_initializer())\n    sess.run(tf.global_variables_initializer())\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    # do your stuff here\n\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>If you post some more code, maybe I could take a deeper look into it. In the meantime, HTH.<\/p>\n"},{"AnswerId":"46992278","CreationDate":"2017-10-28T17:01:35.813","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p>This could also be caused by a wrong tf record file name that doesn't exist at all. Make sure you have correct file paths specified before you do other checks.<\/p>\n"},{"AnswerId":"42563378","CreationDate":"2017-03-02T18:40:56.837","ParentId":null,"OwnerUserId":"4115902","Title":null,"Body":"<p>I had the exactly same issue today and later I found it was the input data file I downloaded from \"famous data set\" (such as <a href=\"https:\/\/archive.ics.uci.edu\/ml\/machine-learning-databases\/iris\/iris.data\" rel=\"nofollow noreferrer\">https:\/\/archive.ics.uci.edu\/ml\/machine-learning-databases\/iris\/iris.data<\/a>) that caused the error: It has some empty lines at the end of the file. Remove the empty lines, the error was gone!<\/p>\n"},{"AnswerId":"38357092","CreationDate":"2016-07-13T16:28:04.040","ParentId":null,"OwnerUserId":"4392784","Title":null,"Body":"<p>To summarize the comments, the <\/p>\n\n<pre><code>Compute status: Out of range: RandomSuffleQueue '_2_input\/shuffle_batch\/random_shuffle_queue' is closed and has insufficient elements (requested 100, current size 0)\n<\/code><\/pre>\n\n<p>was caused by the queue running out of data. This is often caused by thinking you have enough data for N iterations when really you only have enough for M iterations where M &lt; N. <\/p>\n\n<p>One suggestion for figuring out how much data you actually have is to count how many times you can read data before an OutOfRangeError exception is thrown by the queue.<\/p>\n"},{"AnswerId":"46432131","CreationDate":"2017-09-26T17:04:13.320","ParentId":null,"OwnerUserId":"8678019","Title":null,"Body":"<p>I had this same problem and none of the previous answers seemed to solve it so I will also chime in.<\/p>\n\n<p>For me the problem ended up being the features list I was passing to parse_single_example. For whatever reason (since I am using a float_list ?) in my tfrecords file I needed to specify the length of the array in my features list or use tf.VarLenFeature ie:<\/p>\n\n<pre><code>feature_structure = {'features': tf.FixedLenFeature([FEATURE_SIZE], tf.float32),\n           'outputs': tf.FixedLenFeature([OUTPUT_SIZE], tf.float32)}\nd_features = tf.parse_single_example(serialized_example, features=feature_structure)\n<\/code><\/pre>\n\n<p>Without this I kept getting the \"random_shuffle_queue is closed and has insufficient elements\" error which I am guessing is because my parsed example had no data in it.<\/p>\n"},{"AnswerId":"44438695","CreationDate":"2017-06-08T14:37:38.733","ParentId":null,"OwnerUserId":"3821732","Title":null,"Body":"<p>You're likely processing the parsed TFRecord example wrong. E.g. trying to reshape a tensor to an incompatible size. You can debug using a tf_record_iterator to confirm the data you're reading is stored the way you think it is:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntfrecords_filename = '\/path\/to\/some.tfrecord'\nrecord_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)\n\nfor string_record in record_iterator:\n    # Parse the next example\n    example = tf.train.Example()\n    example.ParseFromString(string_record)\n\n    # Get the features you stored (change to match your tfrecord writing code)\n    height = int(example.features.feature['height']\n                                 .int64_list\n                                 .value[0])\n\n    width = int(example.features.feature['width']\n                                .int64_list\n                                .value[0])\n\n    img_string = (example.features.feature['image_raw']\n                                  .bytes_list\n                                  .value[0])\n    # Convert to a numpy array (change dtype to the datatype you stored)\n    img_1d = np.fromstring(img_string, dtype=np.float32)\n    # Print the image shape; does it match your expectations?\n    print(img_1d.shape)\n<\/code><\/pre>\n"}]}
{"QuestionId":34051205,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2015-12-02T19:21:41.697","AcceptedAnswerId":34055986.0,"OwnerUserId":2002890.0,"Title":"Accessing filename from file queue in Tensor Flow","Body":"<p>I have a directory of images, and a separate file matching image filenames to labels. So the directory of images has files like 'train\/001.jpg' and the labeling file looks like:<\/p>\n\n<pre><\/pre>\n\n<p>I can easily load images from the image directory in Tensor Flow by creating a filequeue from the filenames:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm at a loss for how to couple these files with the labels from the labeling file. It seems I need access to the filenames inside the queue at each step. Is there a way to get them? Furthermore, once I have the filename, I need to be able to look up the label keyed by the filename. It seems like a standard Python dictionary wouldn't work because these computations need to happen at each step in the graph.<\/p>\n","answers":[{"AnswerId":"44870199","CreationDate":"2017-07-02T11:15:07.343","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Another suggestion is to save your data in <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/python_io#tfrecords_format_details\" rel=\"nofollow noreferrer\">TFRecord<\/a> format. In this case you would be able to save all images and all labels in the same file. For a big number of files it gives a lot of advantages:<\/p>\n\n<ul>\n<li>can store data and labels at the same place<\/li>\n<li>data is allocated at one place (no need to remember various directories)<\/li>\n<li>if there are many files (images), opening\/closing a file is time consuming. Seeking the location of the file from ssd\/hdd also takes time<\/li>\n<\/ul>\n"},{"AnswerId":"43048260","CreationDate":"2017-03-27T13:48:39.420","ParentId":null,"OwnerUserId":"4820422","Title":null,"Body":"<p>There is <code>tf.py_func()<\/code> you could utilize to implement a mapping from file path to label.<\/p>\n\n<pre><code>files = gfile.Glob(data_pattern)\nfilename_queue = tf.train.string_input_producer(\nfiles, num_epochs=num_epochs, shuffle=True) #  list of files to read\n\ndef extract_label(s):\n    # path to label logic for cat&amp;dog dataset\n    return 0 if os.path.basename(str(s)).startswith('cat') else 1\n\ndef read(filename_queue):\n  key, value = reader.read(filename_queue)\n  image = tf.image.decode_jpeg(value, channels=3)\n  image = tf.cast(image, tf.float32)\n  image = tf.image.resize_image_with_crop_or_pad(image, width, height)\n  label = tf.cast(tf.py_func(extract_label, [key], tf.int64), tf.int32)\n  label = tf.reshape(label, [])\n\ntraining_data = [read(filename_queue) for _ in range(num_readers)]\n\n...\n\ntf.train.shuffle_batch_join(training_data, ...)\n<\/code><\/pre>\n"},{"AnswerId":"43105240","CreationDate":"2017-03-29T22:21:59.683","ParentId":null,"OwnerUserId":"2498151","Title":null,"Body":"<p>I used this:<\/p>\n\n<pre><code> filename = filename.strip().decode('ascii')\n<\/code><\/pre>\n"},{"AnswerId":"34055986","CreationDate":"2015-12-03T00:54:29.577","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Given that your data is not too large for you to supply the list of filenames as a python array, I'd suggest just doing the preprocessing in Python.  Create two lists (same order) of the filenames and the labels, and insert those into either a randomshufflequeue or a queue, and dequeue from that.  If you want the \"loops infinitely\" behavior of the string_input_producer, you could re-run the 'enqueue' at the start of every epoch.<\/p>\n\n<p>A very toy example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nf = [\"f1\", \"f2\", \"f3\", \"f4\", \"f5\", \"f6\", \"f7\", \"f8\"]\nl = [\"l1\", \"l2\", \"l3\", \"l4\", \"l5\", \"l6\", \"l7\", \"l8\"]\n\nfv = tf.constant(f)\nlv = tf.constant(l)\n\nrsq = tf.RandomShuffleQueue(10, 0, [tf.string, tf.string], shapes=[[],[]])\ndo_enqueues = rsq.enqueue_many([fv, lv])\n\ngotf, gotl = rsq.dequeue()\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    tf.train.start_queue_runners(sess=sess)\n    sess.run(do_enqueues)\n    for i in xrange(2):\n        one_f, one_l = sess.run([gotf, gotl])\n        print \"F: \", one_f, \"L: \", one_l\n<\/code><\/pre>\n\n<p>The key is that you're effectively enqueueing pairs of filenames\/labels when you do the <code>enqueue<\/code>, and those pairs are returned by the <code>dequeue<\/code>.<\/p>\n"},{"AnswerId":"34071619","CreationDate":"2015-12-03T16:47:44.450","ParentId":null,"OwnerUserId":"2002890","Title":null,"Body":"<p>Here's what I was able to do.<\/p>\n\n<p>I first shuffled the filenames and matched the labels to them in Python:<\/p>\n\n<pre><code>np.random.shuffle(filenames)\nlabels = [label_dict[f] for f in filenames]\n<\/code><\/pre>\n\n<p>Then created a string_input_producer for the filenames with shuffle off, and a FIFO for labels:<\/p>\n\n<pre><code>lv = tf.constant(labels)\nlabel_fifo = tf.FIFOQueue(len(filenames),tf.int32, shapes=[[]])\nfile_fifo = tf.train.string_input_producer(filenames, shuffle=False, capacity=len(filenames))\nlabel_enqueue = label_fifo.enqueue_many([lv])\n<\/code><\/pre>\n\n<p>Then to read the image I could use a WholeFileReader and to get the label I could dequeue the fifo:<\/p>\n\n<pre><code>reader = tf.WholeFileReader()\nimage = tf.image.decode_jpeg(value, channels=3)\nimage.set_shape([128,128,3])\nresult.uint8image = image\nresult.label = label_fifo.dequeue()\n<\/code><\/pre>\n\n<p>And generate the batches as follows:<\/p>\n\n<pre><code>min_fraction_of_examples_in_queue = 0.4\nmin_queue_examples = int(num_examples_per_epoch *\n                         min_fraction_of_examples_in_queue)\nnum_preprocess_threads = 16\nimages, label_batch = tf.train.shuffle_batch(\n  [result.uint8image, result.label],\n  batch_size=FLAGS.batch_size,\n  num_threads=num_preprocess_threads,\n  capacity=min_queue_examples + 3 * FLAGS.batch_size,\n  min_after_dequeue=min_queue_examples)\n<\/code><\/pre>\n"}]}
{"QuestionId":34060332,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><tensorflow><data-science>","CreationDate":"2015-12-03T07:42:20.507","AcceptedAnswerId":34092314.0,"OwnerUserId":2663210.0,"Title":"How to get predicted class labels in TensorFlow's MNIST example?","Body":"<p>I am new to Neural Networks and went through the MNIST example for beginners.<\/p>\n\n<p>I am currently trying to use this example on another dataset from Kaggle that does not have test labels.<\/p>\n\n<p>If I run the model on the test data set without corresponding labels and therefore unable to compute the accuracy like in the MNIST example, I would like to be able to see the predictions. Is it possible to access observations and their predicted labels somehow and print them out nicely?<\/p>\n","answers":[{"AnswerId":"47607324","CreationDate":"2017-12-02T11:17:13.940","ParentId":null,"OwnerUserId":"7686695","Title":null,"Body":"<pre><code>prediction=tf.argmax(y,1)\nprint prediction.eval(feed_dict={x: mnist.test.images}).\n<\/code><\/pre>\n\n<p>For more details, check out <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/97\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/97<\/a><\/p>\n"},{"AnswerId":"34092314","CreationDate":"2015-12-04T15:58:24.693","ParentId":null,"OwnerUserId":"5486165","Title":null,"Body":"<p>I think you just need to evaluate your output-tensor as stated in the tutorial:<\/p>\n\n<pre><code>accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\nprint(sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels}))\n<\/code><\/pre>\n\n<p>To get the output of a tensor see the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework.html#Tensor\" rel=\"nofollow\">docs<\/a>:<\/p>\n\n<blockquote>\n  <p>After the graph has been launched in a session, the value of the Tensor can be computed by passing it to Session.run(). t.eval() is a shortcut for calling tf.get_default_session().run(t).<\/p>\n<\/blockquote>\n\n<p>If you want to get predictions rather than accuracy, you need to evaluate your ouput tensor <code>y<\/code> in the same way:<\/p>\n\n<pre><code>print(sess.run(y, feed_dict={x: mnist.test.images}))\n<\/code><\/pre>\n"}]}
{"QuestionId":34062718,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-12-03T09:52:28.010","AcceptedAnswerId":34063604.0,"OwnerUserId":5431231.0,"Title":"Neural Networks: Understanding theano Library","Body":"<p>Can anyone explain me the output of following python code:<\/p>\n\n<pre><\/pre>\n\n<h3>Testing the function<\/h3>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34063604","CreationDate":"2015-12-03T10:33:47.130","ParentId":null,"OwnerUserId":"1149632","Title":null,"Body":"<p>You are actually telling Theano to calculate three different functions, where each succeeding function depends on the output of the previously executed function.<\/p>\n\n<p>In your example, you are using two input parameters: matrix A and matrix B.<\/p>\n\n<pre><code>A = [[ 1, 1 ],\n     [ 1, 1 ]]\n\nB = [[ 0, 1 ],\n     [ 2, 3 ]] \n<\/code><\/pre>\n\n<p>The first output line: <code>[[ 1.,  0.], [-1., -2.]]<\/code> is calculated by subtracting your A and B matrices:<\/p>\n\n<pre><code>[[1, 1],   -   [[0, 1],    = [[ 1, 0 ],\n [1, 1]]        [2, 3]]       [-1, -2]\n<\/code><\/pre>\n\n<p>The second output line <code>[[ 1.,  0.], [ 1.,  x2.]]<\/code> is merely the absolute value of the difference we just calculated:<\/p>\n\n<pre><code>abs [[ 1, 0 ],   =   [[ 1, 0],\n     [-1, -2]]        [ 1, 2]]\n<\/code><\/pre>\n\n<p>The third and final line calculates the squared values, element-wise.<\/p>\n\n<h3>Theano magic<\/h3>\n\n<p>Theano actually interprets your Python code, and infer which variables (or mathematical operations) a given variable depend upon. Thus, if you are only interested in the <code>diff_squared<\/code> output, you don't need to include the calls to <code>diff<\/code> and <code>abs_diff<\/code> too.<\/p>\n\n<pre><code>f = function([a, b], [diff_squared])\n<\/code><\/pre>\n"}]}
{"QuestionId":34063588,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><neural-network><caffe><object-recognition>","CreationDate":"2015-12-03T10:32:59.487","AcceptedAnswerId":null,"OwnerUserId":1167349.0,"Title":"Training DeepBelief Network to recognize multiple categories?","Body":"<p>The learning example of the DeepBelief framework demonstrates how to train a neural network to recognize one object category. The method used for training  does not have a category label parameter.<\/p>\n\n<p>However, in the DeepBelief simple example, the given neural network can categorize multiple object categories. Is there a way to do that kind of training through DeepBelief? Or should I look in to Caffe and use that instead as DeepBelief is based on Caffe?<\/p>\n","answers":[{"AnswerId":"34076072","CreationDate":"2015-12-03T21:03:24.560","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Based on their documentation, in particular on a docs for functions <a href=\"https:\/\/github.com\/jetpacapp\/DeepBeliefSDK#jpcnn_train\" rel=\"nofollow\"><code>jpcnn_train<\/code><\/a> and <a href=\"https:\/\/github.com\/jetpacapp\/DeepBeliefSDK#jpcnn_predict\" rel=\"nofollow\"><code>jpcnn_predict<\/code><\/a>, it does not appear to support multiclass classification for custom labels out of the box. It does seem to support multiclass classification for ImageNet labels.<\/p>\n\n<p>However, you can train multiple predictors (<a href=\"https:\/\/github.com\/jetpacapp\/DeepBeliefSDK\/wiki\/How-to-recognize-custom-objects\" rel=\"nofollow\">here's how to train one<\/a>), one per your custom class, and then choose the class for which the corresponding predictor outputs the highest value.<\/p>\n"}]}
{"QuestionId":34065897,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-03T12:18:08.167","AcceptedAnswerId":null,"OwnerUserId":1187026.0,"Title":"Initializing a shared variable with unknown shape","Body":"<p>The  initializer provides a  parameter so that you can change a  shape after initialization.<\/p>\n\n<p>However, it doesn't look like  has that parameter and instead uses the default value of  when creating a new . How can you create a shared  whose shape you can change after initialization? <\/p>\n","answers":[{"AnswerId":"44461073","CreationDate":"2017-06-09T15:03:22.750","ParentId":null,"OwnerUserId":"5806606","Title":null,"Body":"<p>It seems based on the docs <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_variable\" rel=\"nofollow noreferrer\">here<\/a>\nthat the <code>validate_shape<\/code> parameter is now supported as desired, in case anyone else stumbles upon this question.<\/p>\n"}]}
{"QuestionId":34068922,"AnswerCount":2,"Tags":"<pickle><theano><prediction>","CreationDate":"2015-12-03T14:46:19.743","AcceptedAnswerId":null,"OwnerUserId":5531398.0,"Title":"Save Theano model doens't work for MLP net","Body":"<p>I'm trying to save a model obtained during training of a Multi layer Perceptron network built with Theano according to <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/mlp.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/mlp.py<\/a> using the code shown in the Logistic regressor at <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py<\/a>, in particular<\/p>\n\n<pre><\/pre>\n\n<p>but what I get is<\/p>\n\n<blockquote>\n  <p>... loading data ... building the model ... training epoch 1,\n  minibatch 74\/74, validation error 38.333333 %\n       epoch 1, minibatch 74\/74, test error of best model 41.666667 % Traceback (most recent call last):   File \"mlp.py\", line 423, in\n  \n      test_mlp()   File \"mlp.py\", line 406, in test_mlp\n      cPickle.dump(classifier, f, protocol=cPickle.HIGHEST_PROTOCOL) cPickle.PicklingError: Can't pickle : attribute\n  lookup <strong>builtin<\/strong>.instancemethod failed<\/p>\n<\/blockquote>\n\n<p>Since I met this problem also with the convolutional network, my question is: there is a general way to store a model in Theano, to be reused for prediction?<\/p>\n\n<p><strong>EDIT<\/strong>\nAs suggested in the comments I'm using now <\/p>\n\n<pre><\/pre>\n\n<p>for saving and <\/p>\n\n<pre><\/pre>\n\n<p>for setting the weights (for instance) of the hiddenLayer in the classifier defined as<\/p>\n\n<pre><\/pre>\n\n<p>but when I'm calling the very function<\/p>\n\n<pre><\/pre>\n\n<p>I always have [0] as a prediction, even with a well trained net.\nI'm still doing wrong in setting or saving parameters?<\/p>\n","answers":[{"AnswerId":"34128672","CreationDate":"2015-12-07T07:47:59.197","ParentId":null,"OwnerUserId":"5648928","Title":null,"Body":"<p>I've been facing the same problem today. I don't know exactly the reason why this works with <code>'logistic_sgd.py<\/code>' and not with <code>'mlp.py'<\/code> but what you can do is to just store <code>'classifier.params'<\/code><\/p>\n\n<p><code>'classifier.params'<\/code> are the only elements you need once you're done with the training. It should not be too hard to predict your classes from this parameters (this is already in your code).<\/p>\n"},{"AnswerId":"37580720","CreationDate":"2016-06-02T00:10:37.370","ParentId":null,"OwnerUserId":"6012532","Title":null,"Body":"<p>I also ran into this problem and found this solution.\nEven if only <code>'classifier.params'<\/code> should be necessary, <code>y_pred<\/code> and <code>input<\/code> also need to be initialized. Simple way is to store them via pickle and reload them.<\/p>\n\n<p>Saving:<\/p>\n\n<pre><code>with open('best_model.pkl', 'wb') as f:\n    cPickle.dump((classifier.params, classifier.logRegressionLayer.y_pred, \n                 classifier.input), f)\n<\/code><\/pre>\n\n<p>Predicting function:<\/p>\n\n<pre><code>def predict(dataset, n_hidden, n_in, n_out):\n    datasets = load_data(dataset)\n    test_set_x, test_set_y = datasets[2]\n    test_set_x = test_set_x.get_value()\n    test_set_y = test_set_y.eval()\n\n    rng = numpy.random.RandomState(1234)\n    x = T.matrix('x')\n\n    # Declare MLP classifier\n    classifier = MLP(\n        rng=rng,\n        input=x,\n        n_in=n_in,\n        n_hidden=n_hidden,\n        n_out=n_out\n    )\n\n    # load the saved model\n    classifier.params, classifier.logRegressionLayer.y_pred,\n        classifier.input = cPickle.load(open('best_model.pkl'))\n\n    predict_model = theano.function(\n        inputs=[classifier.input],*emphasized text*\n        outputs=classifier.logRegressionLayer.y_pred)\n\n    print(\"Expected values: \", test_set_y[:10])\n    predicted_values = predict_model(test_set_x[:10])\n    print(\"Predicted values:\", predicted_values)\n<\/code><\/pre>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":34071340,"AnswerCount":1,"Tags":"<python><matrix><linear-algebra><division><theano>","CreationDate":"2015-12-03T16:34:22.947","AcceptedAnswerId":34084295.0,"OwnerUserId":699814.0,"Title":"matrix divided by rows of another matrix, without a loop in theano","Body":"<p>What is the equivalent theano implementation of the code below without using a loop?<\/p>\n\n<pre><\/pre>\n\n<p>the loop in numpy also can be replaced with:<\/p>\n\n<pre><\/pre>\n\n<p>What I need to do is to divide columns of \"a\" by each row of \"b\". At the end, there will 4 matrices of size 3*2 (size of \"a\") where each are \"a\" divided by one of the rows of \"b\".<\/p>\n\n<p>-Regards<\/p>\n","answers":[{"AnswerId":"34084295","CreationDate":"2015-12-04T08:55:26.193","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This works in Theano just the same as it does in numpy. Here's a script comparing the three approaches:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as tt\n\n\ndef numpy_v1(a, b):\n    ainvb = np.zeros((3, 2, 4))\n    for i in range(4):\n        ainvb[:, :, i] = a \/ b[:, i].T\n    return ainvb\n\n\ndef numpy_v2(a, b):\n    return a[:, :, None] \/ b\n\n\ndef compile_theano_v1():\n    a, b = tt.matrices('a', 'b')\n    return theano.function([a, b], a[:, :, None] \/ b)\n\n\ndef main():\n    dt = np.dtype(np.float32)\n\n    a = [[12, 3],\n         [2, 4],\n         [2, 4], ]\n\n    b = [[12, 3, 2, 3],\n         [2, 4, 4, 5]]\n\n    a = np.asarray(a, dtype=dt)\n    b = np.asarray(b, dtype=dt)\n    print(a.shape)\n    print(b.shape)\n\n    theano_v1 = compile_theano_v1()\n\n    numpy_v1_ainvb = numpy_v1(a, b)\n    numpy_v2_ainvb = numpy_v2(a, b)\n    theano_v1_ainvb = theano_v1(a, b)\n\n    assert np.allclose(numpy_v1_ainvb, numpy_v2_ainvb)\n    assert np.allclose(numpy_v2_ainvb, theano_v1_ainvb)\n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":34072132,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-03T17:11:42.853","AcceptedAnswerId":34254042.0,"OwnerUserId":3444294.0,"Title":"Validation and Test Evaluations in TensorFlow Demos","Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"nofollow\">TensorFlow beginner demo<\/a> code includes files  and . My understanding is that the beginner demo is based on running  while the expert demo is based on running .<\/p>\n\n<p>Questions:<\/p>\n\n<ol>\n<li>Is my understanding above correct?<\/li>\n<li>If the above is correct, then is it also true that the\nbeginner code () does <em>not<\/em> use validation data\nto stop training, and instead simply runs until the max number of\nepochs (1000 in the demo code)?  evaluates against validation data\nbut I have not found where the beginner code () does.<\/li>\n<li>The beginner demo should be run by executing  and  is meant for use only in the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/tf\/index.html\" rel=\"nofollow\">TensorFlow Mechanics 101 tutorial<\/a>?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"34254042","CreationDate":"2015-12-13T17:28:21.710","ParentId":null,"OwnerUserId":"3444294","Title":null,"Body":"<p>My own anwer: Yes, the code runs to a max number of steps, and <code>fully_connected_feed.py<\/code> is for the later demos.<\/p>\n"}]}
{"QuestionId":34075754,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-gpu>","CreationDate":"2015-12-03T20:42:14.663","AcceptedAnswerId":34075904.0,"OwnerUserId":2002890.0,"Title":"Tensor Flow Explicit Device Requirement Error","Body":"<p>I'm trying to run the CIFAR10 tutorial with the training code on one gpu and the eval code on the other. I know for sure I have two gpus on my computer, and I can test this by running the simple examples here: <a href=\"https:\/\/www.tensorflow.org\/how_tos\/using_gpu\/index.html\">https:\/\/www.tensorflow.org\/how_tos\/using_gpu\/index.html<\/a><\/p>\n\n<p>However, using a  does not work for most variables in the CIFAR example. I tried a whole lot of combinations of different variables on gpu vs. cpu, or all the variables on one or the other. Always the same error for some variable, something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is this possibly a bug in Tensor Flow or am I missing something?<\/p>\n","answers":[{"AnswerId":"34075904","CreationDate":"2015-12-03T20:52:53.080","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p><code>Could not satisfy explicit device specification<\/code> means you do not have the corresponding device. Do you actually have a CUDA-enabled GPU on your machine?<\/p>\n\n<p><strong>UPDATE:<\/strong> As it turned out in the discussion below, this error is also raised if the particular operation (in this case, <code>RandomShuffleQueue<\/code>) cannot be executed on the GPU, because it only has a CPU implementation.<\/p>\n\n<p>If you are fine with TensorFlow choosing a device for you (particularly, falling back to CPU when no GPU implementation is available), consider setting <code>allow_soft_placement<\/code> in your configuration, as per <a href=\"http:\/\/www.tensorflow.org\/how_tos\/using_gpu\/index.html\">this article<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess = tf.Session(config=tf.ConfigProto(\n    allow_soft_placement=True, log_device_placement=True))\n<\/code><\/pre>\n"}]}
{"QuestionId":34075853,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-03T20:49:25.767","AcceptedAnswerId":34077445.0,"OwnerUserId":4511070.0,"Title":"Initialize string producer from CSV","Body":"<p>I have a CSV file that contains pairs of filenames of images and a label. I would like to load those images.<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, I read in the CSV file line-by-line and then try to initialize a second input producer using the filename I get from the CSV decoder.<\/p>\n\n<p>I get an error in the last line, though:<\/p>\n\n<pre><\/pre>\n\n<p>Why is that?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"34077445","CreationDate":"2015-12-03T22:28:10.940","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is an error due to the incompatible <em>shapes<\/em> of <code>img1_path<\/code> (a 0-D, or scalar, Tensor) and the expected <code>string_tensor<\/code> argument to <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"nofollow\"><code>tf.train.string_input_producer<\/code><\/a> (a 1-D, or vector, Tensor).<\/p>\n\n<p>Fortunately, the solution is simple, using <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/array_ops.html#expand_dims\" rel=\"nofollow\"><code>tf.expand_dims<\/code><\/a> to convert <code>img1_path<\/code> to a (one-element) vector:<\/p>\n\n<pre><code># Convert img1_path from a scalar to a 1-element vector.\nimg1_path = tf.expand_dims(img1_path, 0)\n<\/code><\/pre>\n\n<p>Keeping track of the shapes of tensors can be tricky, so TensorFlow provides the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/framework.html#Tensor.get_shape\" rel=\"nofollow\"><code>Tensor.get_shape()<\/code><\/a> method, which you can call to get information about the shape of any tensor object. For example:<\/p>\n\n<pre><code>print img1_path.get_shape()\n# ==&gt; \"TensorShape([])\" (a 0-dimensional tensor, or scalar)\nimg1_path = tf.expand_dims(img1_path, 0)\nprint img1_path.get_shape()\n# ==&gt; \"TensorShape([Dimension(1)])\" (a 1-dimensional tensor with length 1)\n<\/code><\/pre>\n"}]}
{"QuestionId":34076405,"AnswerCount":2,"Tags":"<lua><torch><conv-neural-network>","CreationDate":"2015-12-03T21:23:30.333","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch 7 - parameters questions","Body":"<p>I am confused when I see this line in torch 7 demos;<\/p>\n\n<pre><\/pre>\n\n<p>for example in this documentation at line 147;<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/4_train.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/4_train.lua<\/a><\/p>\n\n<p>Does anyone know what this is trying to do in the context of the training process? Thanks.<\/p>\n","answers":[{"AnswerId":"34077392","CreationDate":"2015-12-03T22:24:54.193","ParentId":null,"OwnerUserId":"5622901","Title":null,"Body":"<p>if table <em>parameters<\/em> has not the same address as table <em>x<\/em>, then do parameters:copy(x)<\/p>\n"},{"AnswerId":"38774529","CreationDate":"2016-08-04T18:11:29.053","ParentId":null,"OwnerUserId":"213123","Title":null,"Body":"<p>The x is the parameter of feval(x), and feval is called in optimMethod (e.g. optim.sgd). This means that optimMethod will call the feval function with updated parameters (x). This if statement you mention, will update the parameters of the model with the parameters sent from optimMethod. The statement parameters:copy(x) means copy from x.<\/p>\n"}]}
{"QuestionId":34077105,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-12-03T22:04:51.100","AcceptedAnswerId":null,"OwnerUserId":5637012.0,"Title":"Tensorflow input format for Object Detection","Body":"<p>Is there any way to read in data for object detection in TensorFlow (e.g. similar to caffe's <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1WindowDataLayer.html\" rel=\"nofollow\"><\/a>)? I've tried looking around for examples that do this, but haven't found any. <\/p>\n","answers":[{"AnswerId":"34077803","CreationDate":"2015-12-03T22:54:32.367","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The standard data format that TensorFlow uses is the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto\" rel=\"nofollow\"><code>Example<\/code> protocol buffer<\/a>, which has a generic notion of \"Feature\" that should support Caffe-style <code>WindowData<\/code>. The <a href=\"http:\/\/www.tensorflow.org\/how_tos\/reading_data\/index.html#standard_tensorflow_format\" rel=\"nofollow\">documentation<\/a> has some information on this format, and the source code includes an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/convert_to_records.py\" rel=\"nofollow\">example application<\/a> for converting image data (the simple MNIST format) to this format, for use with the standard input pipeline.<\/p>\n\n<p>If you follow these steps, you would most likely store the image as a \"bytes\" feature, and add dense integer features, corresponding to the coordinates of the windows and the labels.<\/p>\n"},{"AnswerId":"44681729","CreationDate":"2017-06-21T16:38:19.500","ParentId":null,"OwnerUserId":"2701950","Title":null,"Body":"<p>Take a look at the following <a href=\"https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF\/blob\/master\/lib\/datasets\/kitti.py\" rel=\"nofollow noreferrer\">file<\/a> in <a href=\"https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF\" rel=\"nofollow noreferrer\">smallcorgi's github repository<\/a> that implements Faster RCNN architecture in tensorflow github repository.  That file is an example on how to read PASCAL VOC formatted xml files with bounding box annotations.<\/p>\n"},{"AnswerId":"35592124","CreationDate":"2016-02-24T02:24:52.363","ParentId":null,"OwnerUserId":"2237635","Title":null,"Body":"<p>I have been using TensorFlow for object detection over the last few weeks and have released some of my code as <a href=\"https:\/\/github.com\/Russell91\/tensorbox\" rel=\"nofollow\">TensorBox<\/a>. The input format is a text file in IDL format (see <a href=\"http:\/\/russellsstewart.com\/s\/tensorbox\/brainwash_test.txt\" rel=\"nofollow\">here<\/a>, for example) with a list of image names and a list of the bounding boxes in each image. You can switch out these input files to train and test on your own images.<\/p>\n"}]}
{"QuestionId":34078048,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-03T23:12:44.127","AcceptedAnswerId":34084580.0,"OwnerUserId":5637190.0,"Title":"How to update a segment in the matrix in Torch\/Lua?","Body":"<p>I'm trying to update a segment in the matrix in Torch, but I have not got a simple function to do it.<\/p>\n\n<p>For example: <\/p>\n\n<p>and I tried to update a segment in the matrix like:<\/p>\n\n<p> and<\/p>\n\n<p><\/p>\n\n<p>Neither of them works. <\/p>\n\n<p>Should I just do it with a loop or does there exist a simple function?    <\/p>\n","answers":[{"AnswerId":"34084580","CreationDate":"2015-12-04T09:11:13.077","ParentId":null,"OwnerUserId":"387870","Title":null,"Body":"<p>Simply create a tensor instead of trying to directly assign a table:<\/p>\n\n<pre><code>x[{2, {2,4}}] = torch.Tensor({1, 2, 3})\n<\/code><\/pre>\n"}]}
{"QuestionId":34078063,"AnswerCount":2,"Tags":"<theano><keras>","CreationDate":"2015-12-03T23:13:21.940","AcceptedAnswerId":null,"OwnerUserId":1628393.0,"Title":"Keras Convolution Neural Network","Body":"<p>Right now I am trying to construct a basic convolutional neural network to do simple classification with mnist dataset using keras. Eventually I want to put my own images in I just want to build a simple network first to make sure I have the structure working. so I downloaded the mnist data as mnint.pkl.gz unpacked it and loaded it into tuples and eventually bumpy arrays. Here is my code: <\/p>\n\n<pre><\/pre>\n\n<p>I get an error as such:<\/p>\n\n<pre><\/pre>\n\n<p>I think this means I need to put it into a theano 4d tensor such that\nis has (samples,channels,rows,columns) but I have no idea how to do that. Furthermore when I specifically want to solve the problem I am after I will we loading '.png' files in, I was then going to put them into numpy matrices to feed in but it looks like that is not going to work. Can anybody tell me how I can get Images into theano4d tensors to use in this code? Thanks<\/p>\n","answers":[{"AnswerId":"34083866","CreationDate":"2015-12-04T08:27:54.880","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You are correct that the code is expecting a <code>tensor4<\/code>. The conventional structure is <code>(batch, channel, width, height)<\/code>. In this case the images are monochrome so <code>channel=1<\/code> It looks like you're using a batch size of 10 and the MNIST images are 28 pixels in width and 28 pixels in height.<\/p>\n\n<p>You can simply reshape the data into the format required. If <code>x<\/code> is of shape (10, 784) then <code>x.reshape(10, 1, 28, 28)<\/code> will have the required format.<\/p>\n"},{"AnswerId":"39685276","CreationDate":"2016-09-25T09:21:47.050","ParentId":null,"OwnerUserId":"2686319","Title":null,"Body":"<p>The code is expecting a 4-dimensional <strong>numpy<\/strong> array, not a Theano tensor (keras does all the Theano tensor manipulation under the hood).<\/p>\n\n<p>Your inputs, X_train and X_test need to be reshaped as follows:<\/p>\n\n<pre><code>X_train = X_train.reshape(-1, 1, 28, 28)\nX_test = X_test.reshape(-1, 1, 28, 28)\n<\/code><\/pre>\n"}]}
{"QuestionId":34079398,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2015-12-04T01:26:34.923","AcceptedAnswerId":34081101.0,"OwnerUserId":2476373.0,"Title":"How to concat the output of a pooling layer and inner-product layer in CAFFE?","Body":"<p>While trying to concat the output of a pooling layer (pooling a convolution layer) with the outputs of an inner product layer, I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to concat both and feed them into another (fully-connected) inner-product layer.<\/p>\n\n<p>Can anyone suggests how to solve this?<\/p>\n\n<p>my concat layer definition is:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"34081101","CreationDate":"2015-12-04T04:48:55.117","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>Solved my issue by flattening the conv. layer output.<\/p>\n\n<p>e.g. by adding a Flatten layer between the pool and concat layers:<\/p>\n\n<pre><code>layer {\n  type: \"Flatten\"\n  bottom: \"pool3\"\n  top: \"pool3flat\"\n  name: \"p3flat\"\n}\n\nlayer {\n  type: \"Concat\"\n  bottom: \"pool3flat\"\n  bottom: \"ip1_prior\"\n  top: \"ip1_combine\"\n  name: \"concat\"\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":34079543,"AnswerCount":2,"Tags":"<computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2015-12-04T01:43:14.340","AcceptedAnswerId":null,"OwnerUserId":639973.0,"Title":"Fast-RCNN final bounding box","Body":"<p>I've been playing around with Fast-RCNN for a while, but still can't get some of the core mechanisms.<\/p>\n\n<p>In the tutorial slides (page 28 of <a href=\"http:\/\/tutorial.caffe.berkeleyvision.org\/caffe-cvpr15-detection.pdf\" rel=\"nofollow\">http:\/\/tutorial.caffe.berkeleyvision.org\/caffe-cvpr15-detection.pdf<\/a>), they have an example output with only one bounding box per object:<\/p>\n\n<p><a href=\"http:\/\/s22.postimg.org\/7rbu05xbl\/Screen_Shot_2015_12_04_at_2_19_57_PM.png\" rel=\"nofollow\">http:\/\/s22.postimg.org\/7rbu05xbl\/Screen_Shot_2015_12_04_at_2_19_57_PM.png<\/a><\/p>\n\n<p>Specifically, non-maximum suppression is performed on all region proposals(<a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\/blob\/master\/lib\/fast_rcnn\/test.py#L324\" rel=\"nofollow\">https:\/\/github.com\/rbgirshick\/fast-rcnn\/blob\/master\/lib\/fast_rcnn\/test.py#L324<\/a>), but in my case it still contains tens of regions for each object in the image. <\/p>\n\n<p>My bounding boxes look like the following with threshold of 0.99:<\/p>\n\n<p><a href=\"http:\/\/s29.postimg.org\/oc33ujgrb\/foo.jpg\" rel=\"nofollow\">http:\/\/s29.postimg.org\/oc33ujgrb\/foo.jpg<\/a><\/p>\n\n<p>How and where are the bounding boxes for overlapping region finalized into one?<\/p>\n","answers":[{"AnswerId":"48704394","CreationDate":"2018-02-09T11:01:47.340","ParentId":null,"OwnerUserId":"8096367","Title":null,"Body":"<p>Non-maximum suppression rejects overlap regions larger than the threshold. As your threshold is 0.99, few overlap regions will be rejected.<\/p>\n"},{"AnswerId":"39469990","CreationDate":"2016-09-13T12:22:30.020","ParentId":null,"OwnerUserId":"3716285","Title":null,"Body":"<p>Non-maximum supression should definitely filter out the overlapping bounding boxes in you example image.\nCheck again that you use it properly, and do it after fine-tuning the initial bounding box using the network output.<\/p>\n"}]}
{"QuestionId":34079625,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-04T01:52:32.880","AcceptedAnswerId":34083807.0,"OwnerUserId":5611802.0,"Title":"theano-scan: how outputs_info with placeholder feed as input","Body":"<p>i am confused about outputs_info of scan in theano. below is an example of scan which computes A**k:<\/p>\n\n<pre><\/pre>\n\n<p>i make a little change as below:<\/p>\n\n<pre><\/pre>\n\n<p>it still get the same result as the original.  and when i change outpus_info=[None, T.ones_like(A), None] or ouputs_info=[T.ones_like(A), None, None], it still get the same result. \nT.ones_like(A) will serve as the initial of prior_result in inner_fct, so can i assume that theano will find an not None one as the initial of prior_result? <\/p>\n\n<p>how exactly theano choose initials in outputs_info as input? <\/p>\n\n<pre><\/pre>\n\n<p>the second param in outputs_info is None, which means the second output of inner_fct is not recurrent, but  shows that  are same as .<\/p>\n","answers":[{"AnswerId":"34083807","CreationDate":"2015-12-04T08:23:17.277","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p><code>outputs_info<\/code> works just as <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#theano.scan\" rel=\"nofollow\">the documentation<\/a> says it works.<\/p>\n\n<p>In particular,<\/p>\n\n<blockquote>\n  <p>outputs_info is the list of Theano variables or dictionaries\n  describing the initial state of the outputs computed recurrently.<\/p>\n<\/blockquote>\n\n<p>and,<\/p>\n\n<blockquote>\n  <p>If you provide a None instead of a variable or a empty dictionary scan\n  assumes that you will not use any taps for this output (like for\n  example in case of a map)<\/p>\n<\/blockquote>\n\n<p>The number of elements in the <code>outputs_info<\/code> list should match the number of values returned by the <code>scan<\/code> step function. In your first example, the step function returns a single value so the <code>outputs_info<\/code> should (and does) contain a single element. In your second example the step function returns three values so the <code>outputs_info<\/code> should (and does) contain three elements.<\/p>\n\n<p>For each value in <code>outputs_info<\/code>, if it is <code>None<\/code> then you are saying that the value is not recurrent (outputs in this position are not made available to later iterations). A non-<code>None<\/code> <code>outputs_info<\/code> value indicates that the step output in that position <em>is<\/em> recurrent and the value provided is the initial value to be be passed to the step function in the first step; subsequent steps receive the output from the previous step.<\/p>\n\n<p>So <code>scan<\/code> is not doing some clever magic to automatically figure out which element contains the initial value. You've simply told Theano to do something specific in the second example, the step function returns two non-recurrent values (hence the two <code>None<\/code>s) followed by one recurrent value (hence the initial value in the third position).<\/p>\n\n<p>In the third example (added in a question edit), there is still only one recurrent output from the step function (the third). It is the <code>outputs_info<\/code> values that determines which outputs are recurrent and, in the third example, there is only one non-<code>None<\/code> <code>outputs_info<\/code> value. The <code>B2<\/code> function output and the <code>result<\/code> function output are the same because they have identical computation expressions: <code>prior_result*B<\/code>. It doesn't matter which of these two values is iterated, they remain identical throughout.<\/p>\n"}]}
{"QuestionId":34079781,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-12-04T02:11:04.480","AcceptedAnswerId":34385838.0,"OwnerUserId":2002890.0,"Title":"Tensor Flow shuffle_batch() blocks at end of epoch","Body":"<p>I'm using tf.train.shuffle_batch() to create batches of input images. It includes a min_after_dequeue parameter that makes sure there's a specified number of elements inside the internal queue, and blocks everything else if there isn't. <\/p>\n\n<pre><\/pre>\n\n<p>At the end of an epoch, when I'm doing evaluation (I'm sure this is also true in training but I haven't tested it), everything blocks. I figured out it's at the same moment the internal shuffle batch queue would be left with less than min_after_dequeue elements. At this time in the program I would ideally like to just dequeue the remaining elements but I'm not sure how.<\/p>\n\n<p>Apparently this type of blocking inside TF queues can be shut off when you know there's no more elements to enqueue with the .close() method. However, since the underlying queue is hidden inside the function, how do I call that method?<\/p>\n","answers":[{"AnswerId":"34385838","CreationDate":"2015-12-20T21:05:18.310","ParentId":null,"OwnerUserId":"2002890","Title":null,"Body":"<p>Here's the code that I eventually got to work, although with a bunch of warnings that elements I enqueued were cancelled.<\/p>\n\n<pre><code>lv = tf.constant(label_list)\n\nlabel_fifo = tf.FIFOQueue(len(filenames),tf.int32,shapes=[[]])\n# if eval_data:\n    # num_epochs = 1\n# else:\n    # num_epochs = None\nfile_fifo = tf.train.string_input_producer(filenames, shuffle=False, capacity=len(filenames))\nlabel_enqueue = label_fifo.enqueue_many([lv])\n\n\nreader = tf.WholeFileReader()\nresult.key, value = reader.read(file_fifo)\nimage = tf.image.decode_jpeg(value, channels=3)\nimage.set_shape([128,128,3])\nresult.uint8image = image\nresult.label = label_fifo.dequeue()\n\nimages, label_batch = tf.train.shuffle_batch(\n  [result.uint8image, result.label],\n  batch_size=FLAGS.batch_size,\n  num_threads=num_preprocess_threads,\n  capacity=FLAGS.min_queue_size + 3 * FLAGS.batch_size,\n  min_after_dequeue=FLAGS.min_queue_size)\n\n#in eval file:\nlabel_enqueue, images, labels = load_input.inputs()\n#restore from checkpoint in between\ncoord = tf.train.Coordinator()\ntry:\n  threads = []\n  for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):\n    threads.extend(qr.create_threads(sess, coord=coord, daemon=True,\n                                     start=True))\n\n  num_iter = int(math.ceil(FLAGS.num_examples \/ FLAGS.batch_size))\n  true_count = 0  # Counts the number of correct predictions.\n  total_sample_count = num_iter * FLAGS.batch_size\n\n  sess.run(label_enqueue)\n  step = 0\n  while step &lt; num_iter and not coord.should_stop():\n    end_epoch = False\n    if step &gt; 0:\n        for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):\n            #check if not enough elements in queue\n            size = qr._queue.size().eval()\n            if size - FLAGS.batch_size &lt; FLAGS.min_queue_size:\n                end_epoch = True\n    if end_epoch:\n        #enqueue more so that we can finish\n        sess.run(label_enqueue)\n    #actually run step\n    predictions = sess.run([top_k_op])\n<\/code><\/pre>\n"},{"AnswerId":"43748660","CreationDate":"2017-05-02T23:15:02.217","ParentId":null,"OwnerUserId":"2448616","Title":null,"Body":"<p>There is  an optional argument allow_smaller_final_batch<\/p>\n\n<p>\"allow_smaller_final_batch: (Optional) Boolean. If True, allow the final batch to be smaller if there are insufficient items left in the queue.\"<\/p>\n"},{"AnswerId":"34082495","CreationDate":"2015-12-04T06:50:20.450","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You are correct that running the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/io_ops.html#QueueBase.close\" rel=\"nofollow\"><code>RandomShuffleQueue.close()<\/code><\/a> operation will stop the dequeuing threads from blocking when there are fewer than <code>min_after_dequeue<\/code> elements in the queue.<\/p>\n\n<p>The <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/io_ops.html#shuffle_batch\" rel=\"nofollow\"><code>tf.train.shuffle_batch()<\/code><\/a> function creates a <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#QueueRunner\" rel=\"nofollow\"><code>tf.train.QueueRunner<\/code><\/a> that performs operations on the queue in a background thread. If you start it as follows, passing a <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/train.html#Coordinator\" rel=\"nofollow\"><code>tf.train.Coordinator<\/code><\/a>, you will be able to close the queue cleanly (based on the example <a href=\"http:\/\/www.tensorflow.org\/how_tos\/threading_and_queues\/index.html#queuerunner\" rel=\"nofollow\">here<\/a>):<\/p>\n\n<pre><code>sess = tf.Session()\ncoord = tf.train.Coordinator()\ntf.train.start_queue_runners(sess, coord=coord)\n\nwhile not coord.should_stop():\n  sess.run(train_op)\n# When done, ask the threads to stop.\ncoord.request_stop()\n# And wait for them to actually do it.\ncoord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":34079787,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-12-04T02:11:33.500","AcceptedAnswerId":34082273.0,"OwnerUserId":348412.0,"Title":"Tensor with unspecified dimension in tensorflow","Body":"<p>I'm playing around with tensorflow and ran into a problem with the following code:<\/p>\n\n<pre><\/pre>\n\n<p>(I'm using  as suggested in <a href=\"https:\/\/stackoverflow.com\/questions\/33711427\/tensorflow-initializing-tensor-of-ones\">this question<\/a>, since I was getting the same error)<\/p>\n\n<p>When I run it (from a larger script that invokes this one), I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to replicate the process in the interactive shell. Indeed, the dimensions of  are unspecified, although the supplied values do exist:<\/p>\n\n<pre><\/pre>\n\n<p>This is weird. Tensorflow generates the vector but can't say its shape. Am I doing something wrong?<\/p>\n","answers":[{"AnswerId":"43960046","CreationDate":"2017-05-14T03:17:47.407","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Similar question is nicely explained in <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/faq#tensor_shapes\" rel=\"nofollow noreferrer\">TF FAQ<\/a>:<\/p>\n\n<blockquote>\n  <p>In TensorFlow, a tensor has both a static (inferred) shape and a\n  dynamic (true) shape. The static shape can be read using the\n  <code>tf.Tensor.get_shape<\/code> method: this shape is inferred from the operations\n  that were used to create the tensor, and may be partially complete. If\n  the static shape is not fully defined, the dynamic shape of a Tensor t\n  can be determined by evaluating <code>tf.shape(t)<\/code>.<\/p>\n<\/blockquote>\n\n<p>So <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/shape\" rel=\"nofollow noreferrer\"><code>tf.shape()<\/code><\/a> returns you a tensor, will always have a size of <code>shape=(N,)<\/code>, and can be calculated in a session:<\/p>\n\n<pre><code>a = tf.Variable(tf.zeros(shape=(2, 3, 4)))\nwith tf.Session() as sess:\n    print sess.run(tf.shape(a))\n<\/code><\/pre>\n\n<p>On the other hand you can extract the static shape by using <code>x.get_shape().as_list()<\/code> and this can be calculated anywhere.<\/p>\n"},{"AnswerId":"34081045","CreationDate":"2015-12-04T04:42:46.803","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>The variable can have a dynamic shape. <code>get_shape()<\/code> returns the static shape.<\/p>\n\n<p>In your case you have a tensor that has a dynamic shape, and currently happens to hold value that is 4x3 (but at some other time it can hold a value with a different shape -- because the shape is dynamic). To set the static shape, use <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/framework.html#returns-22\" rel=\"nofollow\"><code>set_shape(w_shape)<\/code><\/a>. After that the shape you set will be enforced, and the tensor will be a valid <code>initial_value<\/code>.<\/p>\n"},{"AnswerId":"34082273","CreationDate":"2015-12-04T06:33:15.400","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As Ishamael says, all tensors have a static shape, which is known at graph construction time and accessible using <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/framework.html#Tensor.get_shape\" rel=\"noreferrer\"><code>Tensor.get_shape()<\/code><\/a>; and a dynamic shape, which is only known at runtime and is accessible by fetching the value of the tensor, or passing it to an operator like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/shape\" rel=\"noreferrer\"><code>tf.shape<\/code><\/a>. In many cases, the static and dynamic shapes are the same, but they can be different - the static shape can be <em>partially defined<\/em> - in order allow the dynamic shape to vary from one step to the next.<\/p>\n\n<p>In your code <code>normal_dist<\/code> has a partially-defined static shape, because <code>w_shape<\/code> is a computed value. (TensorFlow sometimes attempts to evaluate\nthese computed values at graph construction time, but it gets stuck at <code>tf.pack<\/code>.) It infers the shape <code>TensorShape([Dimension(None), Dimension(None)])<\/code>, which means \"a matrix with an unknown number of rows and columns,\" because it knowns that <code>w_shape<\/code> is a vector of length 2, so the resulting <code>normal_dist<\/code> must be 2-dimensional.<\/p>\n\n<p>You have two options to deal with this. You can set the static shape as Ishamael suggests, but this requires you to know the shape at graph construction time. For example, the following may work:<\/p>\n\n<pre><code>normal_dist.set_shape([input_data.get_shape()[1], labels.get_shape()[1]])\n<\/code><\/pre>\n\n<p>Alternatively, you can pass <code>validate_shape=False<\/code> to the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/state_ops.html#Variable.__init__\" rel=\"noreferrer\"><code>tf.Variable<\/code> constructor<\/a>. This allows you to create a variable with a partially-defined shape, but it limits the amount of static shape information that can be inferred later on in the graph.<\/p>\n"}]}
{"QuestionId":34079813,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><tensorflow>","CreationDate":"2015-12-04T02:15:02.717","AcceptedAnswerId":null,"OwnerUserId":2663210.0,"Title":"Retrieving data from TensorFlow object - list of booleans from correct_prediction","Body":"<p>I am going over MNIST beginner tutorial (<a href=\"http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">http:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/index.html<\/a>) and trying to get a boolean list of accurately predicted values from the correct_prediction tensor object. I am finding this confusing.<\/p>\n\n<p>According to the tutorial\n suppost to give us a list of booleans:<\/p>\n\n<blockquote>\n  <p>That gives us a list of booleans. To determine what fraction are\n  correct, we cast to floating point numbers and then take the mean. For\n  example, [True, False, True, True] would become [1,0,1,1] which would\n  become 0.75.<\/p>\n<\/blockquote>\n\n<p>However, Trying  gives us .\n gives us  which is not a list. Calling  to see methods and then  gives us .<\/p>\n\n<p>How do I access the list of predicted booleans and for that matter values of y, W and b? Should they be accessed somehow from tf.Session?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"34080116","CreationDate":"2015-12-04T02:51:58.600","ParentId":null,"OwnerUserId":"348412","Title":null,"Body":"<p>The tensor variables actually describe the computations that must be performed in order to obtain the values you are interested in. <\/p>\n\n<p>In other words, the tensor defined with <code>correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))<\/code> doesn't contain the list of booleans, it contains the instructions for computing it in a tensorflow graph. In order to get the actual values, you need to tell tensorflow to compute it in a graph. <\/p>\n\n<p>First, you need a <code>tf.Session<\/code> variable. A simple way to get it for testing in a interactive shell is <code>sess = tf.InteractiveSession()<\/code>, followed by variable initialization: <code>sess.run(tf.initialize_all_variables())<\/code>. <\/p>\n\n<p>Then, you can call <code>sess.run(tensor_variable)<\/code> to compute the value of a given tensor (or a list of them). If your tensors include placeholders in their computations (which they usually do), you'll also have to provide a feed dictionary. This is exemplified in the tutorial.<\/p>\n\n<p>Instead of <code>session.run()<\/code>, you can also call the <code>.eval()<\/code> method from tensors. This also needs that a default session exists.<\/p>\n"}]}
{"QuestionId":34083144,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-12-04T07:34:21.093","AcceptedAnswerId":34083531.0,"OwnerUserId":3425224.0,"Title":"When I send a numpy array to a givens parameter in a theano function, why do I get this Theano TypeError","Body":"<p>I am new to theano, and am still struggling with the\"pseudo code\" style of theano on the one hand and strict type checking on the other. I am more of a C programmer and a python programmer. Can someone please point out where I am going wrong in this example code which uses mean square error between predicted y points and training y points for x values, to get the optimal slope and intercept of a linear fit?<\/p>\n\n<p>The code is below:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"~\/EclipseWorkspace\/MemoryNetworkQA.Theano\/linear_regression.py\", line\n  70, in \n      lin.fit(x,y,learning_rate=0.01,num_iter=100)   File \"~\/EclipseWorkspace\/MemoryNetworkQA.Theano\/linear_regression.py\", line\n  54, in fit\n      self.x: trainset_x,   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/compile\/function.py\",\n  line 266, in function\n      profile=profile)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/compile\/pfunc.py\", line\n  489, in pfunc\n      no_default_updates=no_default_updates)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/compile\/pfunc.py\", line\n  217, in rebuild_collect_shared\n      raise TypeError(err_msg, err_sug)<\/p>\n  \n  <p>TypeError: ('An update must have the same type as the original shared\n  variable (shared_var=b, shared_var.type=TensorType(int64, scalar),\n  update_val=Elemwise{sub,no_inplace}.0,\n  update_val.type=TensorType(float64, scalar)).', 'If the difference is\n  related to the broadcast pattern, you can call the\n  tensor.unbroadcast(var, axis_to_unbroadcast[, ...]) function to remove\n  broadcastable dimensions.')<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34083531","CreationDate":"2015-12-04T08:03:35.933","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This code won't execute until the following problems have been addressed.<\/p>\n\n<ol>\n<li><p>The error reported in the question is due to the type of <code>self.b<\/code> not matching the type of the update for <code>self.b<\/code>. <code>self.b<\/code> has no type specified so one has been inferred. The initial value is a Python integer so the inferred type is an <code>int64<\/code>. The update is a <code>floatX<\/code> because the learning rate is a <code>floatX<\/code>. You can't update an <code>int64<\/code> with a <code>floatX<\/code>. The solution is to make the initial value a Python float resulting in an inferred <code>floatX<\/code> type. Change <code>self.b = theano.shared(value=1, name='b')<\/code> to <code>self.b = theano.shared(value=1., name='b')<\/code> (note the decimal point after the <code>1<\/code>).<\/p><\/li>\n<li><p>The next problem is that <code>self.x<\/code> is defined as a matrix but the value passed in the function call in the last line is a vector. A solution is to reshape <code>x<\/code> into a matrix, e.g. change <code>x = np.arange(10)<\/code> to <code>x = np.arange(10).reshape(1,10)<\/code>.<\/p><\/li>\n<li><p>The trainset shared variables have type <code>float32<\/code> but this conflicts with other areas of the code that is working with <code>floatX<\/code>. If your <code>floatX=float32<\/code> then there should be no problem but it would be safer to simply use <code>floatX<\/code> to maintain the same float type throughout. Change <code>trainset_x = theano.tensor._shared(x.astype(np.dtype(np.float32)),borrow=True)<\/code> to <code>trainset_x = theano.tensor._shared(x.astype(theano.config.floatX),borrow=True)<\/code> and similarly for <code>trainset_y<\/code>.<\/p><\/li>\n<li><p>The number of epochs is not currently having any effect because <code>epoch<\/code> is not being incremented. Change <code>while epoch &lt; num_iter:<\/code> to <code>for epoch in xrange(num_iter):<\/code> and remove <code>epoch = 0<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>In addition,<\/p>\n\n<ul>\n<li><p>The parameters look like they're not updating but this is a false view. The iterations pass quickly and never stop because of problem 4 above and the learning rate is large enough to make the model converge very quickly. Try changing the learning rate to something much smaller, e.g. 0.0001, and look at the output for only the first 100 epochs.<\/p><\/li>\n<li><p>I'd recommend avoiding the use of <code>theano.tensor._shared<\/code> unless you really do need to force the shared variable to be allocated on the CPU when <code>device=gpu<\/code>. The preferred method is <code>theano.shared<\/code>.<\/p><\/li>\n<li><p>The <code>n_train<\/code> variable isn't used anywhere.<\/p><\/li>\n<li><p>You're using <code>givens<\/code> inconsistently. I'd recommend using it for both <code>x<\/code> and <code>y<\/code>, or for neither. Take a look at the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">logistic regression tutorial<\/a> for more pointers on this.<\/p><\/li>\n<li><p>The Theano function is being recompiled on every call to <code>fit<\/code> but you'd be better off compiling it only once and reusing it on each <code>fit<\/code>.<\/p><\/li>\n<li><p>This model can be implemented without using <code>scan<\/code>. In general, <code>scan<\/code> is often only needed when the output of a step is a function of the output from an earlier step. <code>scan<\/code> is also generally much slower than alternatives and should be avoided when possible. You can remove <code>scan<\/code> by using <code>self.y = self.m * self.x + self.b<\/code> instead.<\/p><\/li>\n<li><p>If you do use scan, it's good practice to enable strict mode, via <code>strict=True<\/code> in the <code>scan<\/code> call.<\/p><\/li>\n<li><p>It's good practice to explicitly provide types for all shared variables. You do this for <code>trainset_x<\/code> and <code>trainset_y<\/code> but not for <code>self.m<\/code> and <code>self.b<\/code>.<\/p><\/li>\n<\/ul>\n"},{"AnswerId":"34083542","CreationDate":"2015-12-04T08:04:16.930","ParentId":null,"OwnerUserId":"3425224","Title":null,"Body":"<p>Okay, I found that the problem was really in the self.b. After initializing it with an explicit float, the type error goes away.<\/p>\n\n<p>But the slope and intercepts (self.m and self.b) which are still theano shared variables and are being passed in through updates, are not really getting updated. If anyone can tell me why, it will be a great help. Thanks.<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\nfrom collections import OrderedDict\n\nclass LinearModel:\n    def __init__(self,num_points):\n        self.m = theano.shared(value=0.1,name='m')\n        self.b = theano.shared(value=1.0, name='b')\n        self.params = [self.m, self.b]\n\n        def step(x_t):\n            y_t = self.m * x_t + self.b\n            return y_t\n\n        #self.x = T.matrix('x',dtype=theano.config.floatX)\n        #self.x = T.dmatrix('x')\n        self.x = T.vector('x',dtype=theano.config.floatX)\n        self.y, _ = theano.scan(\n                        fn=step,\n                        sequences=self.x,\n                    ) \n\n        self.loss = lambda y_train: self.mse(y_train)\n\n    def mse(self, y_train):\n        return T.mean((self.y - y_train) ** 2)\n\n    def fit(self,x, y, learning_rate=0.01, num_iter=100):\n        trainset_x = theano.tensor._shared(x.astype(np.dtype(np.float32)),borrow=True)\n        trainset_y = theano.tensor._shared(y.astype(np.dtype(np.float32)),borrow=True)\n        n_train = trainset_x.get_value(borrow=True).shape[0]\n\n        cost = self.loss(trainset_y)\n        gparams = T.grad(cost,self.params)\n\n        l_r = T.scalar('l_r', dtype=theano.config.floatX)\n\n        updates = OrderedDict()\n        for param,gparam in zip(self.params,gparams):\n            updates[param] = param - l_r * gparam\n\n        self.train_model = theano.function(  inputs=[l_r],\n                                        outputs=[cost,self.y],\n                                        updates=updates,\n                                        givens={\n                                              self.x: trainset_x,\n                                            }\n                                        )\n\n\n        epoch = 0\n        while epoch &lt; num_iter:\n            cost, _ = self.train_model(learning_rate)\n            m = self.m.get_value()\n            b = self.b.get_value()\n            print \"epoch: \",epoch,\" cost: \",cost,\" m: \",m,\" b: \",b\n            epoch += 1\n\n\nif __name__ == '__main__':\n    lin = LinearModel(10)\n    x = np.array([1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0,10.0])\n    y = np.random.rand(10)\n    lin.fit(x,y,learning_rate=0.01,num_iter=100)\n<\/code><\/pre>\n"}]}
{"QuestionId":34084050,"AnswerCount":2,"Tags":"<python><labels><tensorflow>","CreationDate":"2015-12-04T08:40:34.540","AcceptedAnswerId":34095427.0,"OwnerUserId":5602977.0,"Title":"Using string labels in Tensorflow","Body":"<p>i'am still trying to run Tensorflow with own image data. \nI was able to create a .tfrecords file with the conevert_to() function from this example <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/convert_to_records.py\" rel=\"noreferrer\">link<\/a><\/p>\n\n<p>Now i i'd like to train the network with code from that example <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"noreferrer\">link<\/a>. <\/p>\n\n<p>But it fails in the read_and_decode() function. My changes in that function are:<\/p>\n\n<pre><\/pre>\n\n<p>The Error is:<\/p>\n\n<pre><\/pre>\n\n<p>So how to 1) read and 2) use string labels for training in tensorflow.<\/p>\n","answers":[{"AnswerId":"34095427","CreationDate":"2015-12-04T18:49:41.090","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/90121d582dbad4bd13dd2a9750c3a908e89469dd\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py\" rel=\"nofollow noreferrer\"><code>convert_to_records.py<\/code><\/a> script creates a <code>.tfrecords<\/code> file in which each record is an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto\" rel=\"nofollow noreferrer\"><code>Example<\/code><\/a> protocol buffer. That protocol buffer supports string features using the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/feature.proto\" rel=\"nofollow noreferrer\"><code>bytes_list<\/code> kind<\/a>.<\/p>\n\n<p>The <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/tf\/decode_raw\" rel=\"nofollow noreferrer\"><code>tf.decode_raw<\/code><\/a> op is used to parse binary strings into image data; it is not designed to parse string (textual) labels. Assuming that <code>features['label']<\/code> is a <code>tf.string<\/code> tensor, you can use the <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/tf\/string_to_number\" rel=\"nofollow noreferrer\"><code>tf.string_to_number<\/code><\/a> op to convert it to a number. There is limited other support for string processing inside your TensorFlow program, so if you need to perform some more complicated function to convert the string label to an integer, you should perform this conversion in Python in the modified version of <code>convert_to_tensor.py<\/code>.<\/p>\n"},{"AnswerId":"46706708","CreationDate":"2017-10-12T09:59:28.767","ParentId":null,"OwnerUserId":"281545","Title":null,"Body":"<p>To add to @mrry 's answer, supposing your string is <code>ascii<\/code>, you can:<\/p>\n\n<pre><code>def _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef write_proto(cls, filepath, ..., item_id): # itemid is an ascii encodable string\n    # ...\n    with tf.python_io.TFRecordWriter(filepath) as writer:\n        example = tf.train.Example(features=tf.train.Features(feature={\n             # write it as a bytes array, supposing your string is `ascii`\n            'item_id': _bytes_feature(bytes(item_id, encoding='ascii')), # python 3\n            # ...\n        }))\n        writer.write(example.SerializeToString())\n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>def parse_single_example(cls, example_proto, graph=None):\n    features_dict = tf.parse_single_example(example_proto,\n        features={'item_id': tf.FixedLenFeature([], tf.string),\n        # ...\n        })\n    # decode as uint8 aka bytes\n    instance.item_id = tf.decode_raw(features_dict['item_id'], tf.uint8)\n<\/code><\/pre>\n\n<p>and then when you get it back in your session, transform back to string:<\/p>\n\n<pre><code>item_id, ... = session.run(your_tfrecords_iterator.get_next())\nprint(str(item_id.flatten(), 'ascii')) # python 3\n<\/code><\/pre>\n\n<p>I took the <code>uint8<\/code> trick from this <a href=\"https:\/\/stackoverflow.com\/a\/38902908\/281545\">related so answer<\/a>. Works for me but comments\/improvements welcome.<\/p>\n"}]}
{"QuestionId":34089823,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2015-12-04T13:47:23.157","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"Image plotting - after processing","Body":"<p>I am studying the Keras package for deep learning, and found a nice code example on <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py<\/a> that nicely integrates image pre-processing (e.g. rotations and shifts). I was wondering - is there an easy to plot the training images after pre-processing to observe the impact of these rotations and shifts? <\/p>\n","answers":[{"AnswerId":"34110280","CreationDate":"2015-12-05T20:08:27.380","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>You can save the generated images to the disk by giving <code>save_to_dir='path_to_dir'<\/code> to the <code>flow()<\/code> function of the data generator.<\/p>\n"},{"AnswerId":"40087186","CreationDate":"2016-10-17T13:04:16.557","ParentId":null,"OwnerUserId":"4031654","Title":null,"Body":"<p>Yes it is possible to plot images. For example in case of MNIST dataset:<\/p>\n\n<pre><code>from keras.datasets import mnist\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom matplotlib import pyplot\n\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\n\nX_train = X_train.reshape(X_train.shape[0], 1, 28, 28)\nX_test = X_test.reshape(X_test.shape[0], 1, 28, 28)\n\nX_train = X_train.astype('float32')\nX_test = X_test.astype('float32')\n\ndatagen = ImageDataGenerator(horizontal_flip=True, vertical_flip=True)\n\ndatagen.fit(X_train)\n\nfor X_batch, y_batch in datagen.flow(X_train, y_train, batch_size=9):\n# grid of 3x3 images\n    for i in range(0, 9):\n        pyplot.subplot(330 + 1 + i)\n        pyplot.imshow(X_batch[i].reshape(28, 28), cmap=pyplot.get_cmap('gray'))\n\n    pyplot.show()\n    break\n<\/code><\/pre>\n\n<p>For more details refer <a href=\"http:\/\/machinelearningmastery.com\/image-augmentation-deep-learning-keras\/\" rel=\"nofollow\">this<\/a> link.<\/p>\n"}]}
{"QuestionId":34090213,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-12-04T14:08:08.580","AcceptedAnswerId":34095831.0,"OwnerUserId":3106820.0,"Title":"tf.matmul doesn't works as expected","Body":"<p>i try to write and (logic operation) in tensor flow , there are two inputs and two weights multiply them to get one number and add this number to bias, my problem in matmul a send X (input) and W (weight) to method in shape .\n[[1],\n[1]] \nfor X (vertical), and [0.49900547 , 0.49900547] for W (horizontal) to get one number as result but it's give me two numbers , how i can make is multiply right ??\nthis is my code >><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34091342","CreationDate":"2015-12-04T15:08:18.247","ParentId":null,"OwnerUserId":"5486165","Title":null,"Body":"<p><code>matmul<\/code> operates directly on the tensors which in your case have 2 rows and 1 column.<\/p>\n\n<p>There is an argument in <code>matmul<\/code> to transpose either entry like:<\/p>\n\n<pre><code>matmul(X, W, transpose_a=True)\n<\/code><\/pre>\n\n<p>You can check out the docs here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops.html#matmul\" rel=\"nofollow\">docs<\/a><\/p>\n"},{"AnswerId":"34095831","CreationDate":"2015-12-04T19:15:55.700","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As with standard matrix multiplication, if <code>A<\/code> has shape <code>[m, k]<\/code>, and <code>B<\/code> has shape <code>[k, n]<\/code>, then <code>tf.matmul(A, B)<\/code> has shape <code>[m, n]<\/code> (<code>m<\/code> rows, <code>n<\/code> columns in the order TensorFlow uses).<\/p>\n\n<p>In your program, you are computing <code>tf.matmul(X, W)<\/code>. <code>X<\/code> is defined to be a placeholder with shape <code>[2, 1]<\/code>; <code>W<\/code> is defined to be variable initialized to a <code>[1, 2]<\/code> matrix of zeros. As a result, <code>mulres = tf.matmul(X, W)<\/code> will have shape <code>[2, 2]<\/code>, which is what is printed (<code>result: ...<\/code>) when I run your code locally.<\/p>\n\n<p>If you want to define a hidden layer with a single output, the change is simple:<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([1,2]), name=\"weight\")\n<\/code><\/pre>\n\n<p>...should be replaced with:<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([2, 1]), name=\"weight\")\n<\/code><\/pre>\n\n<p>(Indeed, initializing your weights to <code>tf.zeros<\/code> will prevent it from training, because all of the input elements will get the same gradient in backpropagation. Instead, you should initialize them randomly, for example using:<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([2, 1], stddev=0.5), name=\"weight\")\n<\/code><\/pre>\n\n<p>This will enable the network to learn different values for each component of the weight.)<\/p>\n"}]}
{"QuestionId":34090846,"AnswerCount":1,"Tags":"<python><linear-algebra><theano><matrix-inverse>","CreationDate":"2015-12-04T14:42:26.423","AcceptedAnswerId":34103695.0,"OwnerUserId":699814.0,"Title":"broadcasting linalg.pinv on a 3D theano tensor","Body":"<p>in the example below, there is a 3d numpy matrix of size (4, 3, 3)+ a solution about how to calculate pinv of each of 4 of those 3*3 matrices in numpy. I also tried to use the same function worked in numpy, in theano hoping that it is implemented the same, but it failed. Any idea how to do it in theano?<\/p>\n\n<pre><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34103695","CreationDate":"2015-12-05T10:18:38.260","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The error message is<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/Dropbox\/source\/intro_theano\/pinv.py\", line 32, in &lt;module&gt;\n    apinvt = map(lambda n: T.nlinalg.pinv(n), at)\n  File \"d:\\dropbox\\source\\theano\\theano\\tensor\\var.py\", line 549, in __iter__\n    raise TypeError(('TensorType does not support iteration. '\nTypeError: TensorType does not support iteration. Maybe you are using builtin.sum instead of theano.tensor.sum? (Maybe .max?)\n<\/code><\/pre>\n\n<p>This is occurring because, as the error message indicates, the symbolic variable <code>at<\/code> is not iterable.<\/p>\n\n<p>The fundamental problem here is that you're incorrectly mixing immediately executed Python code with delayed execution Theano symbolic code.<\/p>\n\n<p>You need to use a symbolic loop, not a Python loop. The correct solution is to use Theano's <code>scan<\/code> operator:<\/p>\n\n<pre><code>at=T.tensor3('a')\napinvt, _ = theano.scan(lambda n: T.nlinalg.pinv(n), at, strict=True)\nf = theano.function([at], apinvt)\nprint np.allclose(f(a), apinv)\n<\/code><\/pre>\n"}]}
{"QuestionId":34092606,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><lmdb><pycaffe>","CreationDate":"2015-12-04T16:12:18.060","AcceptedAnswerId":34117558.0,"OwnerUserId":1735003.0,"Title":"How to get the dataset size of a Caffe net in python?","Body":"<p>I look at the python <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow\">example for Lenet<\/a> and see that the number of iterations needed to run over the entire MNIST test dataset is hard-coded. However, can this value be not hard-coded at all? How to get the number of samples of the dataset pointed by a network in python?<\/p>\n","answers":[{"AnswerId":"34168545","CreationDate":"2015-12-09T00:12:39.517","ParentId":null,"OwnerUserId":"2915032","Title":null,"Body":"<p>It seems that you mixed <strong>iterations<\/strong> and <strong>amount of samples<\/strong> in one question. In the provided <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow\">example<\/a> we can see only number of <strong>iterations<\/strong>, i. e. how many times training phase will be repeated. The is no any direct relationship between amount of <strong>iterations<\/strong> (network training parameters) and amount of <strong>samples<\/strong> in dataset (network input).<\/p>\n\n<p>Some more detailed explanation:<\/p>\n\n<p><strong>EDIT<\/strong>: <strong>Caffe<\/strong> will totally load (<strong>batch size<\/strong> x <strong>iterations<\/strong>) samples for training or testing, but there is no relation with <strong>amount of loaded samples<\/strong> and <strong>actual database size<\/strong>: it will start reading from the beginning after reaching database last record - it other words, database in <strong>caffe<\/strong> acts like a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Circular_buffer\" rel=\"nofollow\">circular buffer<\/a>.<\/p>\n\n<p>Mentioned example points to <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">this configuration<\/a>. We can see that it expects <strong>lmdb<\/strong> input, and sets <em>batch size<\/em> to <strong>64<\/strong> (some more info about <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\" rel=\"nofollow\">batches and BLOBs<\/a>) for <strong>training<\/strong> phase and <strong>100<\/strong> for <strong>testing<\/strong> phase. Really we don't make any assumption about input <strong>dataset size<\/strong>, i. e. <strong>number of samples in dataset<\/strong>: <em>batch size<\/em> is only processing chunk size, <em>iterations<\/em> is how many batches <strong>caffe<\/strong> will take. It won't stop after reaching database end.<\/p>\n\n<p>In other words, network itself (i. e. <em>protobuf<\/em> config files) doesn't point to any number of samples <strong>in database<\/strong> - only to dataset name and format and <em>desired<\/em> amount of samples. There is no way to determine <strong>database size<\/strong> with <strong>caffe<\/strong> at the current moment, as I know.<\/p>\n\n<p>Thus if you want to load entire dataset for testing, you have only option to firstly determine <strong>amount of samples<\/strong> in <em>mnist_test_lmdb<\/em> or <em>mnist_train_lmdb<\/em> manually, and then specify corresponding values for <strong>batch size<\/strong> and <strong>iterations<\/strong>.<\/p>\n\n<p>You have some options for this:<\/p>\n\n<ol>\n<li>Look at <code>.\/examples\/mnist\/create_mnist.sh<\/code> console output - it prints amount of samples while converting from initial format (I believe that you followed <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">this tutorial<\/a>);<\/li>\n<li>follow @Shai's advice (read lmdb file directly).<\/li>\n<\/ol>\n"},{"AnswerId":"34117558","CreationDate":"2015-12-06T12:43:44.167","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/lmdb.readthedocs.org\/en\/release\" rel=\"nofollow noreferrer\"><code>lmdb<\/code><\/a> library to access the lmdb directly<\/p>\n\n<pre><code>import lmdb\ndb = lmdb.open('\/path\/to\/lmdb_folder')       \/\/Needs lmdb - method\nnum_examples = int( db.stat()['entries'] )\n<\/code><\/pre>\n\n<p>Should do the trick for you.<\/p>\n"}]}
{"QuestionId":34092850,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-12-04T16:24:23.480","AcceptedAnswerId":34186154.0,"OwnerUserId":3849791.0,"Title":"How do I fix a dimension error in TensorFlow?","Body":"<p>I'm trying to apply the expert portion of the tutorial to my own data but I keep running into dimension errors. Here's the code leading up to the error.<\/p>\n\n<pre><\/pre>\n\n<p>And then when I try to run this command:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Just for some background information, the data that I'm dealing with is a CSV file where each row contains 10 features and 1 empty column that can be a 1 or a 0. What I'm trying to get is a probability in the empty column that the column will equal a 1.<\/p>\n","answers":[{"AnswerId":"34186154","CreationDate":"2015-12-09T18:18:03.953","ParentId":null,"OwnerUserId":"3849791","Title":null,"Body":"<p>You have to shape the input so it is compatible with both the training tensor and the output. If you input is length 1, your output should be length 1 (length is substituted for dimension).<\/p>\n\n<p>When you're dealing with-<\/p>\n\n<pre><code>def conv2d(x, W):\n    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')\n\ndef max_pool_2x2(x):\n    return tf.nn.max_pool(x, ksize=[1, 1, 1, 1],\n                    strides=[1, 1, 1, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>Notice how I changed the strides and the ksize to <code>[1, 1, 1, 1]<\/code>. This will match an output to a 1 dimensional input and prevent errors down the road.<\/p>\n\n<p>When you're defining your weight variable (see code below)-<\/p>\n\n<pre><code>def weight_variable(shape):\n    initial = tf.truncated_normal(shape, stddev=0.1)\n    return tf.Variable(initial)\n\ndef bias_variable(shape):\n    initial = tf.constant(0.1, shape=shape)\n    return tf.Variable(initial)\n<\/code><\/pre>\n\n<p>you're going to have to make the first 2 numbers conform to the feature tensor that you are using to train your model, the last two numbers will be the dimension of the predicted output (same as the dimension of the input).<\/p>\n\n<pre><code>W_conv1 = weight_variable([1, 10, 1, 1])\nb_conv1 = bias_variable([1])\n<\/code><\/pre>\n\n<p>Notice the <code>[1, 10,<\/code> in the beginning which signifies that the feature tensor is going to be a 1x10 feature tensor; the last two numbers <code>1, 1]<\/code> correspond to the dimensions of the input and output tensors\/predictors.<\/p>\n\n<p>When you reshape your x_foo tensor (I call it x_ [x prime]), you, for whatever reason, have to define it like so-<\/p>\n\n<pre><code>x_ = tf.reshape(x, [-1,1,10,1])\n<\/code><\/pre>\n\n<p>Notice the 1 and 10 in the middle- <code>...1,10,...<\/code>. Once again, these numbers correspond to the dimension of your feature tensor.<\/p>\n\n<p>For every bias variable, you choose the final number of the previously defined variable. For example, if <code>W_conv1 = weight_variable([1, 10, 1, 1])<\/code> appears like so, you take the final number and put that into your bias variable so it can match the dimensions of the input. This is done like so- <code>b_conv1 = bias_variable([1])<\/code>.<\/p>\n\n<p>If you need any more explanation please comment below.<\/p>\n"},{"AnswerId":"34094552","CreationDate":"2015-12-04T17:59:23.803","ParentId":null,"OwnerUserId":"348412","Title":null,"Body":"<p>The dimensions you are using for the filter are not matching the output of the hidden layer.<\/p>\n\n<p>Let me see if I understood you: your input is composed of 8 features, and you want to reshape it into a 2x4 matrix, right?<\/p>\n\n<p>The weights you created with <code>weight_variable([1, 8, 1, 4])<\/code> expect a 1x8 input, in one channel, and produce a 1x8 output in 4 channels (or hidden units). The filter you are using sweeps the input in 2x2 squares. However, since the result of the weights is 1x8, they won't match.<\/p>\n\n<p>You should reshape the input as<\/p>\n\n<pre><code>x_image = tf.reshape(tf_in, [-1,2,4,1])\n<\/code><\/pre>\n\n<p>Now, your input is actually 2x4 instead of 1x8. Then you need to change the weight shape to <code>(2, 4, 1, hidden_units)<\/code> to deal with a 2x4 output. It will also produce a 2x4 output, and the 2x2 filter now can be applied.<\/p>\n\n<p>After that, the filter will match the output of the weights. Also note that you will have to change the shape of your second weight matrix to <code>weight_variable([2, 4, hidden_units, hidden2_units])<\/code><\/p>\n"}]}
{"QuestionId":34095136,"AnswerCount":0,"Tags":"<python><deep-learning><pycaffe>","CreationDate":"2015-12-04T18:32:30.880","AcceptedAnswerId":null,"OwnerUserId":4058368.0,"Title":"Training py_faster_rcnn on VOC dataset with one class","Body":"<p>I am using py_faster_rcnn to train the system for one class ('person'). Originally, it gave me an assertion error similar to this post\n<a href=\"https:\/\/stackoverflow.com\/questions\/31005463\/how-to-train-new-fast-rcnn-imageset\">How to train new fast-rcnn imageset<\/a><\/p>\n\n<p>So I made the following changes to my imdb.py file: <\/p>\n\n<pre><\/pre>\n\n<p>After the above changes, I get this new error. Has anyone come across this error or what may I be doing wrong?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34097281,"AnswerCount":7,"Tags":"<python><numpy><tensorflow>","CreationDate":"2015-12-04T20:55:54.633","AcceptedAnswerId":37912938.0,"OwnerUserId":1978504.0,"Title":"How can I convert a tensor into a numpy array in TensorFlow?","Body":"<p>How to convert a tensor into a numpy array when using Tensorflow with Python bindings?<\/p>\n","answers":[{"AnswerId":"37912938","CreationDate":"2016-06-19T23:37:24.980","ParentId":null,"OwnerUserId":"852592","Title":null,"Body":"<p>Any tensor returned by <code>Session.run<\/code> or <code>eval<\/code> is a NumPy array.<\/p>\n\n<pre><code>&gt;&gt;&gt; print(type(tf.Session().run(tf.constant([1,2,3]))))\n&lt;class 'numpy.ndarray'&gt;\n<\/code><\/pre>\n\n<p>Or:<\/p>\n\n<pre><code>&gt;&gt;&gt; sess = tf.InteractiveSession()\n&gt;&gt;&gt; print(type(tf.constant([1,2,3]).eval()))\n&lt;class 'numpy.ndarray'&gt;\n<\/code><\/pre>\n\n<p>Or, equivalently:<\/p>\n\n<pre><code>&gt;&gt;&gt; sess = tf.Session()\n&gt;&gt;&gt; with sess.as_default():\n&gt;&gt;&gt;    print(type(tf.constant([1,2,3]).eval()))\n&lt;class 'numpy.ndarray'&gt;\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong> Not <em>any<\/em> tensor returned by <code>Session.run<\/code> or <code>eval()<\/code> is a NumPy array. Sparse Tensors for example are returned as SparseTensorValue:<\/p>\n\n<pre><code>&gt;&gt;&gt; print(type(tf.Session().run(tf.SparseTensor([[0, 0]],[1],[1,2]))))\n&lt;class 'tensorflow.python.framework.sparse_tensor.SparseTensorValue'&gt;\n<\/code><\/pre>\n"},{"AnswerId":"42925913","CreationDate":"2017-03-21T11:32:42.433","ParentId":null,"OwnerUserId":"6358845","Title":null,"Body":"<p>Maybe you can try\uff0cthis method:<\/p>\n\n<pre><code>import tensorflow as tf\nW1 = tf.Variable(tf.random_uniform([1], -1.0, 1.0))\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\narray = W1.eval(sess)\nprint (array)\n<\/code><\/pre>\n"},{"AnswerId":"36678157","CreationDate":"2016-04-17T14:59:12.743","ParentId":null,"OwnerUserId":"2204511","Title":null,"Body":"<p>You need to: <\/p>\n\n<ol>\n<li>encode the image tensor in some format (jpeg, png) to binary tensor <\/li>\n<li>evaluate (run) the binary tensor in a session <\/li>\n<li>turn the binary to stream  <\/li>\n<li>feed to PIL image  <\/li>\n<li>(optional) displaythe image with matplotlib<\/li>\n<\/ol>\n\n<p>Code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport PIL\n\n...\n\nimage_tensor = &lt;your decoded image tensor&gt;\njpeg_bin_tensor = tf.image.encode_jpeg(image_tensor)\n\nwith tf.Session() as sess:\n    # display encoded back to image data\n    jpeg_bin = sess.run(jpeg_bin_tensor)\n    jpeg_str = StringIO.StringIO(jpeg_bin)\n    jpeg_image = PIL.Image.open(jpeg_str)\n    plt.imshow(jpeg_image)\n<\/code><\/pre>\n\n<p>This worked for me. You can try it in a ipython notebook. Just don't forget to add the following line:<\/p>\n\n<pre><code>%matplotlib inline\n<\/code><\/pre>\n"},{"AnswerId":"55205916","CreationDate":"2019-03-17T10:07:15.620","ParentId":null,"OwnerUserId":"7240363","Title":null,"Body":"<p>A simple example could be,<\/p>\n\n<pre><code>    import tensorflow as tf\n    import numpy as np\n    a=tf.random_normal([2,3],0.0,1.0,dtype=tf.float32)  #sampling from a std normal\n    print(type(a))\n    #&lt;class 'tensorflow.python.framework.ops.Tensor'&gt;\n    tf.InteractiveSession()  # run an interactive session in Tf.\n<\/code><\/pre>\n\n<p>n\nnow if we want this tensor a to be converted into a numpy array<\/p>\n\n<pre><code>    a_np=a.eval()\n    print(type(a_np))\n    #&lt;class 'numpy.ndarray'&gt;\n<\/code><\/pre>\n\n<p>As simple as that!<\/p>\n"},{"AnswerId":"52454578","CreationDate":"2018-09-22T07:41:00.450","ParentId":null,"OwnerUserId":"1052899","Title":null,"Body":"<p>I have faced and solved the <em>tensor->ndarray<\/em> conversion in the specific case of tensors representing (adversarial) images, obtained with <em>cleverhans<\/em> library\/tutorials.<\/p>\n\n<p>I think that my question\/answer (<a href=\"https:\/\/stackoverflow.com\/questions\/52443777\/how-to-convert-tensor-to-ndarray-tensor-with-adversarial-images-inside\/52445055#52445055\">here<\/a>) may be an helpful example also for other cases.<\/p>\n\n<p>I'm new with TensorFlow, mine is an empirical conclusion:<\/p>\n\n<p>It seems that tensor.eval() method may need, in order to succeed, also the value for input <em>placeholders<\/em>. \nTensor may work like a function that needs its input values (provided into <code>feed_dict<\/code>) in order to return an output value, e.g.<\/p>\n\n<pre><code>array_out = tensor.eval(session=sess, feed_dict={x: x_input})\n<\/code><\/pre>\n\n<p>Please note that the placeholder name is <strong>x<\/strong> in my case, but I suppose you should find out the right name for the input <em>placeholder<\/em>. \n<code>x_input<\/code> is a scalar value or array containing input data.<\/p>\n\n<p>In my case also providing <code>sess<\/code> was mandatory.<\/p>\n\n<p>My example also covers the <em>matplotlib<\/em> image visualization part, but this is OT.<\/p>\n"},{"AnswerId":"56554855","CreationDate":"2019-06-12T04:50:23.750","ParentId":null,"OwnerUserId":"4909087","Title":null,"Body":"<h1><strong>TensorFlow 2.0<\/strong><\/h1>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"noreferrer\">Eager Execution<\/a> is enabled by default, so just call <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L752-L770\" rel=\"noreferrer\"><strong><code>.numpy()<\/code><\/strong><\/a> on the Tensor object.<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant([[1, 2], [3, 4]])                 \nb = tf.add(a, 1)\n\na.<b>numpy()<\/b>\n# array([[1, 2],\n#        [3, 4]], dtype=int32)\n\nb.<b>numpy()<\/b>\n# array([[2, 3],\n#        [4, 5]], dtype=int32)\n\ntf.multiply(a, b).<b>numpy()<\/b>\n# array([[ 2,  6],\n#        [12, 20]], dtype=int32)\n<\/code><\/pre>\n\n<p>It is worth noting (from the docs),<\/p>\n\n<blockquote>\n  <p>Numpy array may share memory with the Tensor object. <strong>Any changes to one may be reflected in the other.<\/strong><\/p>\n<\/blockquote>\n\n<p>Bold emphasis mine. A copy may or may not be returned, and this is an implementation detail.<\/p>\n\n<hr>\n\n<p>If Eager Execution is disabled, you can build a graph and then run it through <code>tf.compat.v1.Session<\/code>:<\/p>\n\n<pre><code>a = tf.constant([[1, 2], [3, 4]])                 \nb = tf.add(a, 1)\nout = tf.multiply(a, b)\n\nout.eval(session=<b>tf.compat.v1.Session()<\/b>)    \n# array([[ 2,  6],\n#        [12, 20]], dtype=int32)<\/code><\/pre>\n\n<p>See also <a href=\"https:\/\/docs.google.com\/spreadsheets\/d\/1FLFJLzg7WNP6JHODX5q8BDgptKafq_slHpnHVbJIteQ\/edit#gid=0\" rel=\"noreferrer\">TF 2.0 Symbols Map<\/a> for a mapping of the old API to the new one.<\/p>\n"},{"AnswerId":"34097344","CreationDate":"2015-12-04T20:59:39.473","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>To convert back from tensor to numpy array you can simply run <code>.eval()<\/code> on the transformed tensor.<\/p>\n"}]}
{"QuestionId":34097457,"AnswerCount":2,"Tags":"<neural-network><tensorflow>","CreationDate":"2015-12-04T21:07:24.997","AcceptedAnswerId":34097753.0,"OwnerUserId":5231637.0,"Title":"TensorFlow Training","Body":"\n\n<p>Assuming I have a very simple neural network, like multilayer perceptron. For each layer the activation function is sigmoid and the network are fully connected.<\/p>\n\n<p>In TensorFlow this might be defined like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The design target for this model is to map a  points fft spectrogram to another  target spectrogram. Let's assume both the training data and target data are of size . They are stored in variables  and .<\/p>\n\n<p>Now here comes the question. For TensorFlow is there any difference between these two trainings?<\/p>\n\n<p>Training 1:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Training 2:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Thank you very much!<\/p>\n","answers":[{"AnswerId":"43215495","CreationDate":"2017-04-04T18:58:24.680","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>Yes there's a difference. I think the second way loss function can be bit messy. It's more like online training. For each data point in the whole batch you update all of your parameters. But in the first way it's called the batch gradient where you take one batch at a time and take the average loss then update the parameters. <\/p>\n\n<p>Please refer this link \n<strong><a href=\"https:\/\/stats.stackexchange.com\/questions\/49528\/batch-gradient-descent-versus-stochastic-gradient-descent\">https:\/\/stats.stackexchange.com\/questions\/49528\/batch-gradient-descent-versus-stochastic-gradient-descent<\/a><\/strong>\nFirst answer is really good in this link<\/p>\n"},{"AnswerId":"34097753","CreationDate":"2015-12-04T21:26:36.330","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In the first training version, you are training the entire batch of training data at once, which means that the first and the 3000th element of <code>spec_train<\/code> will be processed using the same model parameters in a single step. This is known as <strong>(Batch) Gradient Descent<\/strong>.<\/p>\n\n<p>In the second training version, you are training a single example from the training data at once, which means that the 3000th element of <code>spec_train<\/code> will be processed using model parameters that have been updated 2999 times since the first element was most recently processed. This is known as <strong>Stochastic Gradient Descent<\/strong> (or it would be if the element was selected at random).<\/p>\n\n<p>In general, TensorFlow is used with datasets that are too large to process in one batch, so mini-batch SGD (where a subset of the examples are processed in one step) is favored. Processing a single element at a time is theoretically desirable, but is inherently sequential and has high fixed costs because the matrix multiplications and other operations are not as computationally dense. Therefore, processing a small batch (e.g. 32 or 128) of examples at once is the usual approach, with multiple replicas training on different batches in parallel.<\/p>\n\n<p>See this <a href=\"https:\/\/stats.stackexchange.com\/questions\/49528\/batch-gradient-descent-versus-stochastic-gradient-descent\">Stats StackExchange question<\/a> for a more theoretical discussion of when you should use one approach versus the other.<\/p>\n"}]}
{"QuestionId":34097884,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-04T21:37:47.820","AcceptedAnswerId":null,"OwnerUserId":5641162.0,"Title":"TensorFlow very slow on second run (Ubuntu)","Body":"\n\n<p>I'm having a problem with TensorFlow (CPU) on Ubuntu 14.04 (VM, droplet), where running a script is fast the first time, but when running the same (or another) script directly after completion of the first run, things become very slow. \nI'm talking minutes instead of seconds. Even simple test scripts (like those provided in the tutorial) take forever, with no visible CPU load. <\/p>\n\n<p>For comparison: first run of the test script from the tutorial gives:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Second run of the same script, directly after completion of the first run gives:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Eventually, things seem to clear up and performance is back (only for one run though).<\/p>\n\n<p>I narrowed the problem down to this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>takes very long. Apparently resources used by a previous Session are not properly released [?]. My scripts use the Context manager, like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>My latest hypothesis is that this has to do with system properties (virtual machine settings, hypervisor issues interacting with the context manager of TF). Using the docker container of TF makes no difference. Rebooting didn't help either. The same scripts run OK on OS X.<\/p>\n","answers":[{"AnswerId":"34234680","CreationDate":"2015-12-12T00:18:07.433","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>To make sure it's obvious what happened and that this question is answered:  This occurred because tensorflow was reading from \/dev\/random instead of \/dev\/urandom.  On some systems, \/dev\/random can exhaust its supply of randomness and block until more is available, causing the slowdown.  This has now been <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/cd53f3c3302c9312c1840389a9988a879b8b9dd5\" rel=\"nofollow\">fixed in github<\/a>.  The fixes are included in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/tag\/0.6.0\" rel=\"nofollow\">release 0.6.0<\/a> and later.<\/p>\n"}]}
{"QuestionId":34097988,"AnswerCount":8,"Tags":"<python-2.7><python-3.x><anaconda><theano><keras>","CreationDate":"2015-12-04T21:44:47.167","AcceptedAnswerId":34975902.0,"OwnerUserId":5641188.0,"Title":"How do I install Keras and Theano in Anaconda Python on Windows?","Body":"<p>I am trying to work on neural networks in Python using the following Keras packages:<\/p>\n\n<pre><\/pre>\n\n<p>But, I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Installing installed . Later I tried to use , but it did not work. I Tried to install using , but I am getting this error:  So I installed Git and I set the environment variables.<\/p>\n\n<p>So, is there any procedure to install these packages?<\/p>\n","answers":[{"AnswerId":"45730499","CreationDate":"2017-08-17T08:47:15.670","ParentId":null,"OwnerUserId":"4866415","Title":null,"Body":"<p>I use macOS and used to have the same problem.<br>\nRunning the following command in the terminal saved me:<\/p>\n\n<pre><code>conda install -c conda-forge keras tensorflow\n<\/code><\/pre>\n\n<p>Hope it helps.<\/p>\n"},{"AnswerId":"39435319","CreationDate":"2016-09-11T10:54:34.610","ParentId":null,"OwnerUserId":"4132112","Title":null,"Body":"<p>In case you want to train CNN's with the theano backend like the Keras mnist_cnn.py example:<\/p>\n\n<p>You better use theano bleeding edge version. Otherwise there may occur assertion errors.<\/p>\n\n<ul>\n<li>Run Theano bleeding edge<br>\n<code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git<\/code><\/li>\n<li>Run Keras (like 1.0.8 works fine)<br>\n<code>pip install git+git:\/\/github.com\/fchollet\/keras.git<\/code><\/li>\n<\/ul>\n"},{"AnswerId":"42172336","CreationDate":"2017-02-11T04:45:44.283","ParentId":null,"OwnerUserId":"400589","Title":null,"Body":"<p>The trick is that <strong>you need to create an environment\/workspace for Python<\/strong>. This solution should work for Python 2.7 but at the time of writing keras can run on python 3.5, especially if you have the latest anaconda installed (this took me awhile to figure out so I'll outline the steps I took to install KERAS in python 3.5):<\/p>\n\n<p>Create environment\/workspace for Python 3.5<\/p>\n\n<ol>\n<li><code>C:\\conda create --name neuralnets python=3.5<\/code><\/li>\n<li><code>C:\\activate neuralnets<\/code><\/li>\n<\/ol>\n\n<p><em>Install everything<\/em> (notice the neuralnets workspace in parenthesis on each line). Accept any dependencies each of those steps wants to install:<\/p>\n\n<ol>\n<li><code>(neuralnets) C:\\conda install theano<\/code><\/li>\n<li><code>(neuralnets) C:\\conda install mingw libpython<\/code><\/li>\n<li><code>(neuralnets) C:\\pip install tensorflow<\/code><\/li>\n<li><code>(neuralnets) C:\\pip install keras<\/code><\/li>\n<\/ol>\n\n<p>Test it out:<\/p>\n\n<pre><code>(neuralnets) C:\\python -c \"from keras import backend; print(backend._BACKEND)\"\n<\/code><\/pre>\n\n<hr>\n\n<p>Just remember, if you want to work in the workspace you always have to do:<\/p>\n\n<pre><code>C:\\activate neuralnets\n<\/code><\/pre>\n\n<p>so you can launch Jupyter for example (assuming you also have Jupyter installed in this environment\/workspace) as:<\/p>\n\n<pre><code>C:\\activate neuralnets\n(neuralnets) jupyter notebook\n<\/code><\/pre>\n\n<blockquote>\n  <p>You can read more about managing and creating conda environments\/workspaces \n  at the follwing URL: <a href=\"https:\/\/conda.io\/docs\/using\/envs.html\" rel=\"nofollow noreferrer\">https:\/\/conda.io\/docs\/using\/envs.html<\/a><\/p>\n<\/blockquote>\n"},{"AnswerId":"34975902","CreationDate":"2016-01-24T12:41:08.747","ParentId":null,"OwnerUserId":"3667840","Title":null,"Body":"<p>It is my solution for the same problem<\/p>\n\n<ul>\n<li>Install <a href=\"http:\/\/tdm-gcc.tdragon.net\/\">TDM GCC<\/a> x64.<\/li>\n<li>Install <a href=\"https:\/\/www.continuum.io\/downloads\">Anaconda<\/a> x64.<\/li>\n<li>Open the Anaconda prompt<\/li>\n<li>Run <code>conda update conda<\/code><\/li>\n<li>Run <code>conda update --all<\/code><\/li>\n<li>Run <code>conda install mingw libpython<\/code><\/li>\n<li>Install the latest version of Theano,\n<code>pip install git+git:\/\/github.com\/Theano\/Theano.git<\/code><\/li>\n<li>Run <code>pip install git+git:\/\/github.com\/fchollet\/keras.git<\/code><\/li>\n<\/ul>\n"},{"AnswerId":"41099371","CreationDate":"2016-12-12T11:01:19.590","ParentId":null,"OwnerUserId":"3662213","Title":null,"Body":"<p>In windows with anaconda, just go on conda prompt and use this command <\/p>\n\n<pre><code>conda install --channel https:\/\/conda.anaconda.org\/conda-forge keras\n<\/code><\/pre>\n"},{"AnswerId":"55334030","CreationDate":"2019-03-25T08:46:39.387","ParentId":null,"OwnerUserId":"7916730","Title":null,"Body":"<p>In windows environment with Anconda. Go to anconda prompt from start. Then if you are behind proxy then .copndarc file needs to eb updated with the proxy details.<\/p>\n\n<p>ssl_verify: false\nchannels:\n  - defaults\nproxy_servers:\n  http: <a href=\"http:\/\/xx.xx.xx.xx:xxxx\" rel=\"nofollow noreferrer\">http:\/\/xx.xx.xx.xx:xxxx<\/a>\n  https: <a href=\"https:\/\/xx.xx.xx.xx:xxxx\" rel=\"nofollow noreferrer\">https:\/\/xx.xx.xx.xx:xxxx<\/a><\/p>\n\n<p>I had ssl_verify initially marked as 'True' then I was getting ssl error. So i turned it to false as above and then ran the below commands<\/p>\n\n<p>conda update conda\nconda update --all\nconda install --channel <a href=\"https:\/\/conda.anaconda.org\/conda-forge\" rel=\"nofollow noreferrer\">https:\/\/conda.anaconda.org\/conda-forge<\/a> keras\nconda install --channel <a href=\"https:\/\/conda.anaconda.org\/conda-forge\" rel=\"nofollow noreferrer\">https:\/\/conda.anaconda.org\/conda-forge<\/a> tensorflow<\/p>\n\n<p>My python version is 3.6.7<\/p>\n"},{"AnswerId":"52201372","CreationDate":"2018-09-06T09:52:24.037","ParentId":null,"OwnerUserId":"8667660","Title":null,"Body":"<p>install by this command given below\n<strong>conda install -c conda-forge keras<\/strong><\/p>\n\n<p>this is error \"CondaError: Cannot link a source that does not exist\" ive get in win 10.\nfor your error put this command in your command line.<\/p>\n\n<p><strong>conda update conda<\/strong><\/p>\n\n<p>this work for me .<\/p>\n"},{"AnswerId":"55759418","CreationDate":"2019-04-19T09:13:44.913","ParentId":null,"OwnerUserId":"4510954","Title":null,"Body":"<p><strong>Anaconda with Windows<\/strong><\/p>\n\n<ul>\n<li>Run anaconda prompt with administrator privilages<\/li>\n<li>conda update conda<\/li>\n<li>conda update --all<\/li>\n<li>conda install mingw libpython<\/li>\n<li>conda install theano<\/li>\n<\/ul>\n\n<p>After conda commands it's required to accept process - <strong>Proceed ([y]\/n)?<\/strong><\/p>\n"}]}
{"QuestionId":34110502,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-05T20:30:28.997","AcceptedAnswerId":34113467.0,"OwnerUserId":2584721.0,"Title":"Bilinear Tensor Product in TensorFlow","Body":"\n\n<p>I'm working on re-implementing <a href=\"http:\/\/papers.nips.cc\/paper\/5028-reasoning-with-neural-tensor-networks-for-knowledge-base-completion.pdf\" rel=\"nofollow noreferrer\">this paper<\/a> and the key operation is a bilinear tensor product. I hardly know what that means, but the paper has a nice little graphic, which I understand. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BC9gZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BC9gZ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The key operation is <strong>e_1 * W * e_2<\/strong>, and I want to know how to implement it in tensorflow, because the rest <em>should<\/em> be easy.<\/p>\n\n<p>Basically, given 3D tensor W, slice it into matrices, and for the j'th slice (a matrix), multiply it on each side by <strong>e_1<\/strong> and <strong>e_2<\/strong>, resulting in a scalar, which is the jth entry in the resulting vector (the output of this operation).<\/p>\n\n<p>So I want to perform a product of <strong>e_1<\/strong>, a d-dimensional vector, <strong>W<\/strong>, the d x d x k tensor, and <strong>e_2<\/strong>, another d-dimensional vector. Could this product be expressed concisely in TensorFlow as it is now, or would I have to define my own op somehow?<\/p>\n\n<p><strong>EARLIER EDITS<\/strong><\/p>\n\n<p>Why doesn't multiplying these tensors work, and is there some way to define it more explicitly so that it works? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The error message is <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><strong>CURRENTLY<\/strong><\/p>\n\n<p>Turns out that multiplying two 3D tensors doesn't work either with , so but  does.  will also do 3D tensors and matrices. Then I tried 3D and a vector: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34113467","CreationDate":"2015-12-06T02:54:23.773","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>You can do this with a simple reshape. For the first of the two matrix multiplies, you have k*d, length d vectors to dot product with.<\/p>\n\n<p>This should be close:<\/p>\n\n<pre><code>temp = tf.matmul(E1,tf.reshape(Wddk,[d,d*k]))\nresult = tf.matmul(E2,tf.reshape(temp,[d,k]))\n<\/code><\/pre>\n"},{"AnswerId":"49095661","CreationDate":"2018-03-04T12:59:50.653","ParentId":null,"OwnerUserId":"2502692","Title":null,"Body":"<p>You can perform a rank 3 tensor and a vector multiplication between W and e2 which yields a 2D array and then multiply the result with e1. The following function makes use of tensor product and tensor contraction to define the this product (e.g. W * e3)<\/p>\n\n<pre><code>import sympy as sp\n\ndef tensor3_vector_product(T, v):\n    \"\"\"Implements a product of a rank 3 tensor (3D array) with a \n       vector using tensor product and tensor contraction.\n\n    Parameters\n    ----------\n    T: sp.Array of dimensions n x m x k\n\n    v: sp.Array of dimensions k x 1\n\n    Returns\n    -------\n    A: sp.Array of dimensions n x m\n\n    \"\"\"\n    assert(T.rank() == 3)\n    # reshape v to ensure a 1D vector so that contraction do \n    # not contain x 1 dimension\n    v.reshape(v.shape[0], )\n    p = sp.tensorproduct(T, v)\n    return sp.tensorcontraction(p, (2, 3))\n<\/code><\/pre>\n\n<p>You can verify this multiplication by using the examples provided in <a href=\"http:\/\/epubs.siam.org\/doi\/abs\/10.1137\/07070111X?journalCode=siread\" rel=\"nofollow noreferrer\">ref<\/a>. The above function contracts the second and third axis, in your case I think you should contract (1, 2) because W is defined as d x d x k and not k x d x d as in my case.<\/p>\n"}]}
{"QuestionId":34110575,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-05T20:37:23.970","AcceptedAnswerId":34111413.0,"OwnerUserId":4229825.0,"Title":"How to simulate network from tutorial","Body":"<p>I have a code of convolutional network from tensorflow <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html#deep-mnist-for-experts\" rel=\"nofollow\">tutorial<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>How can I simulate this network? That is, I want to get a network's answer on the particular image from . Something like that:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34111413","CreationDate":"2015-12-05T22:01:45.507","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>In tensor flow every node in the graph can be evaluated. So, similarly to how you evaluate the <code>accuracy<\/code>, you can also evaluate any other node in the graph. The entire neural network is represented by its last layer, in your case <code>y_conv<\/code>. I don't have TensorFlow handy, but something along the lines of the following line should work:<\/p>\n\n<pre><code>y_conv.eval(feed_dict={x: mnist.test.images[0]})\n<\/code><\/pre>\n"}]}
{"QuestionId":34112202,"AnswerCount":1,"Tags":"<python><io><tensorflow>","CreationDate":"2015-12-05T23:32:10.740","AcceptedAnswerId":34126949.0,"OwnerUserId":5231637.0,"Title":"TensorFlow checkpoint save and read","Body":"<p>I have a TensorFlow based neural network and a set of variables. <\/p>\n\n<p>The training function is like this:<\/p>\n\n<pre><\/pre>\n\n<p>I was calling the training function like this:<\/p>\n\n<pre><\/pre>\n\n<p>I did this kind of training because I needed to feed different dataset to my model. However, if I call the train function in this way TensorFlow will generate error message indicating that it cannot read the saved model from file.<\/p>\n\n<p>After some experiments I found that this happened because the checkpoint saving was slow. Before the file was written to the disk the next train function would start reading, thus generate the error.<\/p>\n\n<p>I have tried to use time.sleep() function to make some delay between each call but it didn't work.<\/p>\n\n<p>Anyone knows how to work out this kind of write\/read error? Thank you very much!<\/p>\n","answers":[{"AnswerId":"34126949","CreationDate":"2015-12-07T05:19:33.630","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is a subtle issue in your code: each time you call the <code>train()<\/code> function, more nodes are added to the same TensorFlow graph, for all the model variables and the rest of the neural network. This means that each time you construct a <code>tf.train.Saver()<\/code>, it includes all of the variables for the previous calls to <code>train()<\/code>. Each time you recreate the model, the variables are created with an extra <code>_N<\/code> suffix to give them a unique name:<\/p>\n\n<ol>\n<li>Saver constructed with variables <code>var_a<\/code>, <code>var_b<\/code>.<\/li>\n<li>Saver constructed with variables <code>var_a<\/code>, <code>var_b<\/code>, <code>var_a_1<\/code>, <code>var_b_1<\/code>.<\/li>\n<li>Saver constructed with variables <code>var_a<\/code>, <code>var_b<\/code>, <code>var_a_1<\/code>, <code>var_b_1<\/code>, <code>var_a_2<\/code>, <code>var_b_2<\/code>.<\/li>\n<li>etc.<\/li>\n<\/ol>\n\n<p>The default behavior for a <code>tf.train.Saver<\/code> is to associate each variable with the name of the corresponding op. This means that <code>var_a_1<\/code> won't be initialized from <code>var_a<\/code>, because they end up with different names.<\/p>\n\n<p>The solution is to create a new graph each time you call <code>train()<\/code>. The simplest way to fix it is to change your main program to create a new graph for each call to <code>train()<\/code> as follows:<\/p>\n\n<pre><code># First train\nwith tf.Graph().as_default():\n    train(False, 1)\n\n# Following train\nfor i in xrange(10):\n    with tf.Graph().as_default():\n        train(True, 10)\n<\/code><\/pre>\n\n<p>...or, equivalently, you could move the <code>with<\/code> block inside the <code>train()<\/code> function.<\/p>\n"}]}
{"QuestionId":34114579,"AnswerCount":1,"Tags":"<python><ubuntu><terminal><tensorflow>","CreationDate":"2015-12-06T06:03:39.213","AcceptedAnswerId":null,"OwnerUserId":5021110.0,"Title":"Error in importing tensorflow","Body":"<p>I followed instructions in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv_install\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv_install<\/a> Virtualenv but when I type this in python in Ubuntu terminal  I got this message:\nWhat did I wrong?<\/p>\n","answers":[{"AnswerId":"34115211","CreationDate":"2015-12-06T07:47:49.183","ParentId":null,"OwnerUserId":"989659","Title":null,"Body":"<p>did you install the tenserflow module in your virtualenv?<\/p>\n\n<p>after activating the virtualenv \ntry to run the relevant pip command to install<\/p>\n\n<pre><code># Ubuntu\/Linux 64-bit, CPU only:\n(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, GPU enabled:\n(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\n\n# Mac OS X, CPU only:\n(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":34115810,"AnswerCount":1,"Tags":"<linux><install><theano>","CreationDate":"2015-12-06T09:15:41.923","AcceptedAnswerId":null,"OwnerUserId":5485347.0,"Title":"Installing Blocks on Linux","Body":"<p>I'm trying to <a href=\"http:\/\/blocks.readthedocs.org\/en\/latest\/setup.html\" rel=\"nofollow\">install Blocks<\/a> on my Linux machine and I have some troubles. How can I fix this? I'm doing the following steps:<\/p>\n\n<ol>\n<li>Installing <a href=\"https:\/\/www.continuum.io\/downloads#_unix\" rel=\"nofollow\">Anaconda for python2<\/a>.<\/li>\n<li>Theano depends on numpy 1.10.1 version. If at this step I will run  in python shell then it will work perfectly.<\/li>\n<li>Okay, it's time for Blocks. I'm installing stable version using <\/li>\n<li>After this step if I import theano it gives the following: .<\/li>\n<li>I looked at requirements.txt and see that Blocks depends on numpy 1.9.3.<\/li>\n<li>I uninstalled Theano, Blocks, downgraded numpy using , then just run again  and still while importing theano it gives me .<\/li>\n<\/ol>\n\n<p>How can I overcome this problem?<\/p>\n","answers":[{"AnswerId":"45756063","CreationDate":"2017-08-18T11:54:59.480","ParentId":null,"OwnerUserId":"2413201","Title":null,"Body":"<p>This could be a problem with the Anaconda distribution. You could try updating all of Anaconda via <code>conda update conda<\/code> and <code>conda update --all<\/code>. You could also try changing Blocks requirements.txt to refer to numpy 1.10.1 since it&#39;s likely that Blocks won&#39;t care about the changes from 1.9 to 1.10.\n                    &ndash;&nbsp;Daniel Renshaw<\/p>\n"}]}
{"QuestionId":34115948,"AnswerCount":1,"Tags":"<python><amazon-web-services><theano><keras>","CreationDate":"2015-12-06T09:34:48.060","AcceptedAnswerId":34115992.0,"OwnerUserId":1885116.0,"Title":"Error - Keras+Theano - loss function","Body":"<p>I am trying to run a Keras script on an AWS instance. While the script runs fine on my own computer (Python 2.7 - no CPU) it causes an error on AWS. I have installed the latest version of Theano - and other scripts (e.g. the mnist tutoral) do not give errors. The script that is causing the issue is a standard Keras tutoral script (<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py<\/a>). The error is copied in per below (apologies - there might be a better way to capture errors straight from the command line). Any help much appreciated<\/p>\n\n<p>First page of error message:\n<a href=\"https:\/\/i.stack.imgur.com\/E2lWX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/E2lWX.png\" alt=\"first page of the error message\"><\/a><\/p>\n\n<p>End of error message (i have not copied in the entire stack of keras\/layers errors)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MCpkH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MCpkH.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"34115992","CreationDate":"2015-12-06T09:41:14.663","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Somehow you're passing a symbolic value for the <code>border_mode<\/code> parameter. If this works fine on CPU but not on GPU then, for some reason, the CPU version of the code supports symbolic border modes but the GPU version does not.<\/p>\n\n<p>If you can, change the <code>border_mode<\/code> parameter value to be a Python literal instead of a Theano symbolic variable.<\/p>\n"}]}
{"QuestionId":34116402,"AnswerCount":2,"Tags":"<tensorflow><training-data><conv-neural-network>","CreationDate":"2015-12-06T10:32:30.180","AcceptedAnswerId":34390081.0,"OwnerUserId":2849678.0,"Title":"Tensorflow Convolution Neural Net - Training with a small dataset, applying random changes to Images","Body":"\n\n<p>Say I have a very small dataset, just 50 Images. I want to re-use the code from the tutorial at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/tutorials\/mnist\/pros\/index.md\" rel=\"nofollow noreferrer\">Red Pill<\/a>, but apply random transformations to the same set of Images in each Batch of training, say random changes to Brightness, Contrast etc. I added just one function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Small change to the old code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>First iteration is successful, thereafter it crashes:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This is exactly the same error that I get with my personal dataset with 50 training examples.     <\/p>\n","answers":[{"AnswerId":"34234606","CreationDate":"2015-12-12T00:09:38.993","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>One thing to start with:  Instead of computing y_conv and then the cross-entropy, use the merged <code>tf.softmax_cross_entropy_with_logits<\/code> operator.  This may not solve your problem, but it's more numerically stable than the naive version in the Red Pill example.<\/p>\n\n<p>Second, try printing out the cross_entropy at every iteration.<\/p>\n\n<pre><code>cross_entropy = .... (previous code here)\ncross_entropy = tf.Print(cross_entropy, [cross_entropy], \"Cross-entropy: \")\n<\/code><\/pre>\n\n<p>to get an idea if it's going to infinity as the model progresses, or if it just jumps to inf or NaN.  If it progressively blows up, then it's probably the learning rate.  If it jumps, it could be a numerical boundary condition that could be solved as above.  If it's there from the get-go, you may have an error in the way you're applying distortions that ends up feeding in horribly broken data in some way.<\/p>\n"},{"AnswerId":"34390081","CreationDate":"2015-12-21T06:33:18.690","ParentId":null,"OwnerUserId":"2849678","Title":null,"Body":"<p>Following the link: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/323#issuecomment-165855633\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/323#issuecomment-165855633<\/a> solved the issue<\/p>\n"}]}
{"QuestionId":34119072,"AnswerCount":1,"Tags":"<opencv><ubuntu><caffe>","CreationDate":"2015-12-06T15:25:22.300","AcceptedAnswerId":null,"OwnerUserId":4805839.0,"Title":"Opencv headers different version and libraries different version: ubuntu","Body":"<p>I am new to Ubuntu and OpenCV. Apologies for my amateurishness.<\/p>\n\n<p>I think I messed up OpenCV installation.<\/p>\n\n<p>I had installed OpenCV 2.4 months back. The installation was successful (built openCV from source using CMake)<\/p>\n\n<p>Couple of days back I had also installed OpenCV 3.0 using the same procedure.\nThe problem began when one of my code gave a segmentation fault for seemingly no reason. On suggestion by a friend on Caffe Neural Network issues forum (<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/3416\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/3416<\/a>), it seems that I am compiling with 3.0 headers but linking against the 2.4 libraries.<\/p>\n\n<p>I checked \/usr\/local\/include. It has the standard openCv header files ( I didn't understand what 'compiling with 3.0 header means' )<\/p>\n\n<p>In \/usr\/local\/lib I have OpenCV 3.0 libraries ( no 2.4.8 at all).<\/p>\n\n<p>Now compilation was successful, but when I try to run it, I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>Just to add<\/p>\n\n<p>\"pkg-config --modversion opencv\"\n3.0.0<\/p>\n\n<p>whereas dpkg -l | grep libopencv\ngave out<\/p>\n\n<p>i libopencv-calib3d-dev:amd64 2.4.8+dfsg1-2ubuntu1 amd64 development files for libopencv-calib3d\nii libopencv-calib3d2.4:amd64 2.4.8+dfsg1-2ubuntu1 amd64 computer vision Camera Calibration library\nii libopencv-contrib-dev:amd64 2.4.8+dfsg1-2ubuntu1 amd64 development files for libopencv-contrib\n.......<\/p>\n\n<p>q.1> why does my code link to libopencv_core.so.2.4 instead of 3.0. I have checked in \/usr\/local\/lib and did not see a single 2.4 library. I have tried manually defining links to header files and libraries in CMake but still same result.<\/p>\n\n<p>q.2> Do OpenCV headers have a version number? I don't understand how they are called '2.4 headers'. I have checked the source code but did not see any specific requirement for a version of library. Then what dictates to link to 2.4 version libraries?<\/p>\n\n<p>Sorry if this is theoretical but this would be highly beneficial to beginners like me. Please do comment for a suitable question title if the current one is not appropriate.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34121770","CreationDate":"2015-12-06T19:39:17.200","ParentId":null,"OwnerUserId":"3033937","Title":null,"Body":"<p>Your opencv2.4 libraries are located in <code>\/usr\/lib\/x86_64-linux-gnu\/<\/code> while the new opencv3 libraries are most likely located in \/usr\/local\/lib. The linker is first looking in the <code>\/usr\/lib\/x86_64-linux-gnu\/<\/code> directory, and thus finding the opencv2.4 libs first. To fix this you can tell the linker to look in <code>\/usr\/local\/lib<\/code> first by adding the following to your <code>~\/.bashrc<\/code> file<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\/usr\/local\/lib:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>Reload the bashrc file in the terminal as follows (or simply close and re-open the terminal):<\/p>\n\n<pre><code>source ~\/.bashrc\n<\/code><\/pre>\n\n<p>And update the library database:<\/p>\n\n<pre><code>sudo ldconfig\n<\/code><\/pre>\n\n<p>To check the order in which the linker will search for libraries, you can use the following:<\/p>\n\n<pre><code>ld -lopencv_core --verbose | grep attempt\n<\/code><\/pre>\n"}]}
{"QuestionId":34119185,"AnswerCount":1,"Tags":"<python><amazon-web-services><theano><keras>","CreationDate":"2015-12-06T15:36:12.527","AcceptedAnswerId":null,"OwnerUserId":1885116.0,"Title":"Error - Theano+Keras","Body":"<p>I am trying to run a Keras script on an AWS instance. While the script runs fine on my own computer (Python 2.7 - no CPU) it causes some error on a GPU enabled AWS instance. I have installed the latest version of Theano - and other scripts (e.g. the mnist tutoral) do not give errors. <\/p>\n\n<p>The script that is causing the issue is a standard Keras tutoral script (<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py<\/a>). I have already moved, based on Stack Overflow guidance, the border mode to \"valid\" - which seems to remove one issue. however - I immediately encounter the following problem (error stack below). i have run the following line at bash \n\"THEANO_FLAGS=optimizer=fast_compile,device=gpu,floatX=float32 cifar10.py\" but this didnt give more information. Maybe i should move to nolearn \/ lasagne package - but please let me know if there are easy ways to troubleshoot this.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36458346","CreationDate":"2016-04-06T17:36:11.193","ParentId":null,"OwnerUserId":"4042918","Title":null,"Body":"<p>When you change your border_mode from \"same\" to \"valid\", the dimension of your feature map will gradually shrink. At some point, the dimension of a layer becomes so small (looks like it became 1x1 already in your case) that you can't add extra layers on it anymore. <\/p>\n\n<p>If you are specifying certain strides of a layer, the size of the filter and the size of the input also need to be compatible with your stride length. <\/p>\n"}]}
{"QuestionId":34120911,"AnswerCount":3,"Tags":"<neural-network><deep-learning><tensorflow><conv-neural-network>","CreationDate":"2015-12-06T18:17:43.800","AcceptedAnswerId":null,"OwnerUserId":3492105.0,"Title":"How do we recognize text on real image after training MNIST with TensorFlow?","Body":"<p>I have just ran the code to train the MNIST model. How can we apply it to recognize handwriting in the real image?\nI'm a newbie and have just started learning this part. I've searched and couldn't find info about this.<\/p>\n","answers":[{"AnswerId":"34135601","CreationDate":"2015-12-07T14:11:23.127","ParentId":null,"OwnerUserId":"5546193","Title":null,"Body":"<p>It depends on your specific task. The MNIST model can classify character digits and so that is the data you need to feed it. <\/p>\n\n<p>If you insist on using the MNIST model (RNNs specifically <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">LSTMs<\/a> are a much better option which most OCRs use.) one approach would be to run a sliding window over your hand written text image and create a text file of the character predicted by your model. But that presents its own set of challenges like novelty detection and slide window size. Its an overkill<\/p>\n"},{"AnswerId":"34135616","CreationDate":"2015-12-07T14:12:11.470","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p>You must take into account that the MNIST data set is very much a \"toy data set\", it is just the \"Hello world!\" of image recognition, and thus, it is not very useful for real life problems. If you recreate the training images conditions and try to predict <em>that<\/em>, you will get good results, but again, it is not a real life problem, but maybe it's a fun project to start with.<\/p>\n\n<p>With that said, maybe you are not yet familiar with feature extraction, a technique that modern convolutional neural networks use. Using that <em>maybe<\/em> you could have something useful out of an MNIST trained model. <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow\">Tensorflow's CNN tutorial<\/a> would be very helpful in that sense. To further understand this tecniques I would highly recommend <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html#introducing_convolutional_networks\" rel=\"nofollow\">chapter 6 of Michael Nielsen's Deep Learning online book<\/a><\/p>\n\n<p>Hope it helps<\/p>\n"},{"AnswerId":"34811524","CreationDate":"2016-01-15T12:52:52.167","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>A neural network is essentially very good at recognizing \"stuff\" that you feed it. For the MNIST example, you are feeding it lots of 28x28x1 images of handwritten digits. <\/p>\n\n<p>Therefore, as an extension, if you want to use the exact same model for handwriting (I assume you mean characters from A-Z now), you would then have to feed it lots of 28x28x1 images of handwritten characters.<\/p>\n"}]}
{"QuestionId":34122445,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-06T20:39:35.097","AcceptedAnswerId":null,"OwnerUserId":1188374.0,"Title":"Theano function theano.tensor.shape_padaxis not found in module","Body":"<p>I'm trying to use theano.tensor.shape_padaxis as shown <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.shape_padaxis\" rel=\"nofollow\">here<\/a>.\nI get this error:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'module' object has no attribute 'shape_padaxis'<\/p>\n<\/blockquote>\n\n<p>I'm using python 2.7 and theano 0.7\nEverything else seems to work correctly.<\/p>\n\n<pre><\/pre>\n\n<p>Can anybody reproduce? Any idea where that comes from?\nThanks<\/p>\n","answers":[{"AnswerId":"34139176","CreationDate":"2015-12-07T17:06:44.900","ParentId":null,"OwnerUserId":"1188374","Title":null,"Body":"<p>Got the answer in the <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/Twq4T5V-qE4\" rel=\"nofollow\">google group<\/a> for theano users.<\/p>\n"}]}
{"QuestionId":34123291,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-06T21:55:42.823","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch - Apply function over dimension","Body":"<p>I would like to be able to apply a function which is designed for a 3D tensor to each 3D tensor in a 4D tensor, namely . For example, I can apply the function individually to two images of dimension (3,50,50) but it would be great if I could feed their 4D concatenation of (2,3,50,50).<\/p>\n\n<p>This could probably be done in a for loop but I was wondering if there was any built in function for this. Thanks.<\/p>\n","answers":[{"AnswerId":"34175121","CreationDate":"2015-12-09T09:29:05.763","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>I haven't managed to find such a function in <code>Torch<\/code>. You can, of course, define one yourself to make your life a little bit happier:<\/p>\n\n<pre><code>function apply_to_slices(tensor, dimension, func, ...)\n    for i, slice in ipairs(tensor:split(1, dimension)) do\n        func(slice, i, ...)\n    end\n    return tensor\nend\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre><code>function power_fill(tensor, i, power)\n    power = power or 1\n    tensor:fill(i ^ power)\nend\n\nA = torch.Tensor(5, 6)\n\napply_to_slices(A, 1, power_fill)\n\n 1  1  1  1  1  1\n 2  2  2  2  2  2\n 3  3  3  3  3  3\n 4  4  4  4  4  4\n 5  5  5  5  5  5\n[torch.DoubleTensor of size 5x6]\n\napply_to_slices(A, 2, power_fill, 3)\n\n   1    8   27   64  125  216\n   1    8   27   64  125  216\n   1    8   27   64  125  216\n   1    8   27   64  125  216\n   1    8   27   64  125  216\n[torch.DoubleTensor of size 5x6]\n<\/code><\/pre>\n"}]}
{"QuestionId":34125052,"AnswerCount":2,"Tags":"<opencv><ubuntu><opencv3.0><caffe><nvidia-digits>","CreationDate":"2015-12-07T01:23:48.007","AcceptedAnswerId":null,"OwnerUserId":568145.0,"Title":"Caffe with OpenCV 3 and NVIDIA Digits: OpenCV version conflict (2.4 vs 3.0)","Body":"<p><strong>Background:<\/strong><\/p>\n\n<p>I wish to use both <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe<\/a> and <a href=\"https:\/\/developer.nvidia.com\/digits\" rel=\"nofollow\">Digits<\/a> such that I can use Caffe within the Digits framework or external to it.<\/p>\n\n<p>However, for a particular project, I require that Caffe uses OpenCV 3 and not OpenCV 2.4, which Digits installs by default. This project uses Caffe external to Digits, and does not make any use of the Digits framework.<\/p>\n\n<p>It appears that by installing Digits, my OpenCV 3 installation was \"clobbered\" with OpenCV 2.4, which is now causing problems within my original Caffe installation.<\/p>\n\n<p>To make things clearer, below is a listing of the steps that I have taken.<\/p>\n\n<p>From a fresh Ubuntu 14.04 installation:<\/p>\n\n<ol>\n<li>Installed Caffe dependencies (except OpenCV) as per the <a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow\">Ubuntu\ninstall guide<\/a><\/li>\n<li>Installed OpenCV 3 from source to <\/li>\n<li>Tested OpenCV installation\n\n<ul>\n<li>Demos worked fine including OpenCV\n3-specific code   <\/li>\n<\/ul><\/li>\n<li>Compiled Caffe, setting the  to use OpenCV 3<\/li>\n<li>Tested Caffe installation\n\n<ul>\n<li>All tests passed, demos worked fine<\/li>\n<\/ul><\/li>\n<li>Installed Digits as per the <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-2.0\/docs\/WebInstall.md\" rel=\"nofollow\">install guide<\/a>\n\n<ul>\n<li>Caffe and OpenCV 2.4 were installed by default by the installer script<\/li>\n<li>OpenCV 3 clobbered by OpenCV 2.4 (?)<\/li>\n<\/ul><\/li>\n<li>Carried out the steps from the Digits <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-2.0\/docs\/GettingStarted.md\" rel=\"nofollow\">Getting Started guide<\/a>\n\n<ul>\n<li>All steps were successful<\/li>\n<\/ul><\/li>\n<li>Suspected OpenCV conflict, so tried to compile a Caffe demo\n\n<ul>\n<li>Error occurred relating to OpenCV 3.0 and 2.4 conflict -- details below.<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>Compilation command:<\/p>\n\n<blockquote>\n  <p>g++ classification.cpp -o classification -I\/home\/josh\/software\/caffe\/include\/ -L\/home\/josh\/software\/caffe\/build\/lib\/ -lcaffe -I\/usr\/local\/cuda\/include -L\/usr\/local\/cuda\/lib64 -lcuda -lcudart -lcublas -lcurand -I\/home\/josh\/software\/cudnn\/include\/ -L\/home\/josh\/software\/cudnn\/lib64\/ -lcudnn -L\/usr\/lib\/x86_64-linux-gnu\/ -lglog -L\/usr\/local\/lib -lboost_system -lopencv_core -lopencv_highgui -lopencv_imgproc -lopencv_imgcodecs -DUSE_OPENCV<\/p>\n<\/blockquote>\n\n<p>Error message:<\/p>\n\n<blockquote>\n  <p>\/usr\/bin\/ld: warning: libopencv_core.so.3.0, needed by \/home\/josh\/software\/caffe\/build\/lib\/\/libcaffe.so, may conflict with libopencv_core.so.2.4\n  \/usr\/bin\/ld: \/tmp\/ccHaWcOl.o: undefined reference to symbol '_ZN2cv6String10deallocateEv'\n  \/\/usr\/local\/lib\/libopencv_core.so.3.0: error adding symbols: DSO missing from command line\n  collect2: error: ld returned 1 exit status<\/p>\n<\/blockquote>\n\n<p><strong>Questions:<\/strong><\/p>\n\n<ol>\n<li>How do I best resolve the OpenCV version conflict without breaking either Caffe\/Digits installations? <\/li>\n<li>Do I need to remove OpenCV and reinstall Caffe and Digits?\n\n<ul>\n<li>If so, what do I need to do differently to prevent the OpenCV version conflicts while still allowing Caffe (with OpenCV 3) and Digits to work side-by-side?<\/li>\n<\/ul><\/li>\n<\/ol>\n","answers":[{"AnswerId":"43152651","CreationDate":"2017-04-01T03:54:32.420","ParentId":null,"OwnerUserId":"7800342","Title":null,"Body":"<p>if you install Caffe, using opencv3.0 is not a good ideal. Caffe is not very compatible with it. And if you persist in using it , please remove 2.4 version, using command <code>sudo apt-get autoremove libopencv-dev<\/code>.\nThe reason for this version conflict, is you install libopencv-dev (default version is 2.4) and install opencv with 3.0. So remove it, and then compile caffe.<\/p>\n"},{"AnswerId":"34211100","CreationDate":"2015-12-10T20:27:13.967","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>When OpenCV library is installed, the shared libraries are stored like libopencv_core.so.2.4.8, where 2.4.8 is the version number of OpenCV. libopencv_core.so file usually links to the last version of OpenCV installed in the system.<\/p>\n\n<p>Thus, though libopencv_core.so was initially linked to libopencv_core.so.3.0, when you installed the older version, the link got updated to link to the older version.<\/p>\n\n<p>The workaround will be to specifically call the latest OpenCV library in caffe. To do so, you will have to modify the Makefile present in caffe. \nReplace all occurrences of <code>opencv_*<\/code> to <code>:libopencv_*.so.3.0<\/code><\/p>\n"}]}
{"QuestionId":34126314,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><data-science>","CreationDate":"2015-12-07T04:11:42.487","AcceptedAnswerId":null,"OwnerUserId":2780313.0,"Title":"How to represent a linear data in TensorFlow","Body":"<p>I'm trying to model some oscilloscope-like data in TensorFlow - a linear stream of energy pulses with a duration, intensity, etc. - but otherwise performing very similar classification tasks, and I'm having trouble figuring out how best to represent it in TensorFlow.<\/p>\n\n<p>The tutorials are aimed at image classification, and the framework seems to be built around 4d tensors I'm having trouble figuring out how to represent my relatively simpler data.<\/p>\n\n<p>Specifically, I'm trying to figure out the following:<\/p>\n\n<p>1) If I have a string of 100 pulses with 4 tracks of information about them, is that equivalent to a 4x100 image or a 100x4 image?  Or perhaps a 100x1 image with 4 \"channels\"?  How the tensors are folded\/unfolded is unclear.<\/p>\n\n<p>2) How does max-pooling translate into this lower dimensional space?  Like, how do I ensure that I'm pooling across pulses instead of just within a channel, or some other non-sensical pattern?  How the \"strides\" are calculated and used is not well explained.<\/p>\n\n<p>Had anyone tried to model similar data in TensorFlow?<\/p>\n","answers":[{"AnswerId":"34126799","CreationDate":"2015-12-07T05:04:50.683","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Recurrent_neural_network\" rel=\"nofollow\">Recurrent neural networks<\/a> (RNNs) are a possible representation for sequential data like a stream of energy pulses. The TensorFlow website has a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">tutorial on building an RNN<\/a> for predicting the next word in a sentence of words, but this could possibly be adapted to predicting the next value in your scenario.<\/p>\n"}]}
{"QuestionId":34128104,"AnswerCount":3,"Tags":"<arrays><masking><tensorflow>","CreationDate":"2015-12-07T07:03:09.950","AcceptedAnswerId":34138336.0,"OwnerUserId":1002035.0,"Title":"tensorflow creating mask of varied lengths","Body":"<p>I have a tensor of lengths in tensorflow, let's say it looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I wish to create a mask of 1s and 0s whose number of 1s correspond to the entries to this tensor, padded by 0s to a total length of 8. I.e. I want to create this tensor:<\/p>\n\n<pre><\/pre>\n\n<p>How might I do this?<\/p>\n","answers":[{"AnswerId":"39386358","CreationDate":"2016-09-08T08:42:28.910","ParentId":null,"OwnerUserId":"830928","Title":null,"Body":"<p>I've got a bit shorter version, than previous answer. Not sure if it is more efficient or not<\/p>\n\n<pre><code> def mask(self, seq_length, max_seq_length):\n    return tf.map_fn(\n        lambda x: tf.pad(tf.ones([x], dtype=tf.int32), [[0, max_seq_length - x]]),\n        seq_length)\n<\/code><\/pre>\n"},{"AnswerId":"40386786","CreationDate":"2016-11-02T18:03:28.753","ParentId":null,"OwnerUserId":"4263827","Title":null,"Body":"<p>This can now be achieved by <code>tf.sequence_mask<\/code>. More details <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sequence_mask\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"34138336","CreationDate":"2015-12-07T16:24:10.783","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This can be achieved using a variety of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html\" rel=\"noreferrer\">TensorFlow transformations<\/a>:<\/p>\n\n<pre><code># Make a 4 x 8 matrix where each row contains the length repeated 8 times.\nlengths = [4, 3, 5, 2]\nlengths_transposed = tf.expand_dims(lengths, 1)\n\n# Make a 4 x 8 matrix where each row contains [0, 1, ..., 7]\nrange = tf.range(0, 8, 1)\nrange_row = tf.expand_dims(range, 0)\n\n# Use the logical operations to create a mask\nmask = tf.less(range_row, lengths_transposed)\n\n# Use the select operation to select between 1 or 0 for each value.\nresult = tf.select(mask, tf.ones([4, 8]), tf.zeros([4, 8]))\n<\/code><\/pre>\n"}]}
{"QuestionId":34129504,"AnswerCount":1,"Tags":"<lua><audio-recording><audio-processing><torch>","CreationDate":"2015-12-07T08:45:53.053","AcceptedAnswerId":null,"OwnerUserId":5118777.0,"Title":"Is it possible to record sound in torch 7?","Body":"<p>As title,<\/p>\n\n<p>Is it possible to record sound in torch 7 ?<\/p>\n\n<p>Is there a library to do this ?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"34130362","CreationDate":"2015-12-07T09:35:47.327","ParentId":null,"OwnerUserId":"227563","Title":null,"Body":"<p>I didn't try it, but it seems doing what you look for.\n<a href=\"https:\/\/github.com\/ansca\/Simple-Audio-Recorder\/blob\/master\/main.lua\" rel=\"nofollow\">https:\/\/github.com\/ansca\/Simple-Audio-Recorder\/blob\/master\/main.lua<\/a><\/p>\n"}]}
{"QuestionId":34130902,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-12-07T10:04:48.177","AcceptedAnswerId":34143927.0,"OwnerUserId":867505.0,"Title":"Create color histogram of an image using tensorflow","Body":"<p>Is there a neat way to compute a color histogram of an image? Maybe by abusing the internal code of ? From what I've seen, this code is not very modular and calls directly some C++ code.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"34133471","CreationDate":"2015-12-07T12:22:09.130","ParentId":null,"OwnerUserId":"867505","Title":null,"Body":"<p>This is what I'm using right now:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Assumption: img is a tensor of the size [img_width, img_height, 3], normalized to the range [-1, 1].\nwith tf.variable_scope('color_hist_producer') as scope:\n  bin_size = 0.2\n  hist_entries = []\n  # Split image into single channels\n  img_r, img_g, img_b = tf.split(2, 3, img)\n  for img_chan in [img_r, img_g, img_b]:\n    for idx, i in enumerate(np.arange(-1, 1, bin_size)):\n      gt = tf.greater(img_chan, i)\n      leq = tf.less_equal(img_chan, i + bin_size)\n      # Put together with logical_and, cast to float and sum up entries -&gt; gives count for current bin.\n      hist_entries.append(tf.reduce_sum(tf.cast(tf.logical_and(gt, leq), tf.float32)))\n\n  # Pack scalars together to a tensor, then normalize histogram.\n  hist = tf.nn.l2_normalize(tf.pack(hist_entries), 0)\n<\/code><\/pre>\n"},{"AnswerId":"34143927","CreationDate":"2015-12-07T21:51:16.917","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I would use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#unsorted_segment_sum\" rel=\"nofollow\"><code>tf.unsorted_segment_sum<\/code><\/a>, where the \"segment IDs\" are computed from the color values and the thing you sum is a <code>tf.ones<\/code> vector.  Note that <code>tf.unsorted_segment_sum<\/code> is probably better thought of as \"bucket sum\".  It implements <code>dest[segment] += thing_to_sum<\/code> -- exactly the operation you need for a histogram.<\/p>\n\n<p>In slightly pseudocode (meaning I haven't run this):<\/p>\n\n<pre><code>binned_values = tf.reshape(tf.floor(img_r * (NUM_BINS-1)), [-1])\nbinned_values = tf.cast(binned_values, tf.int32)\nones = tf.ones_like(binned_values, dtype=tf.int32)\ncounts = tf.unsorted_segment_sum(ones, binned_values, NUM_BINS)\n<\/code><\/pre>\n\n<p>You could accomplish this in one pass instead of separating out the r, g, and b values with a split if you wanted to cleverly construct your \"ones\" to look like \"100100...\" for red, \"010010\" for green, etc., but I suspect it would be slower overall, and harder to read.  I'd just do the split that you proposed above.<\/p>\n"},{"AnswerId":"44988054","CreationDate":"2017-07-08T15:57:32.553","ParentId":null,"OwnerUserId":"7414362","Title":null,"Body":"<pre><code>tf.histogram_fixed_width \n<\/code><\/pre>\n\n<p>might be what you are looking for...<\/p>\n\n<p>Full documentation on <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/histogram_fixed_width\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/histogram_fixed_width<\/a><\/p>\n"}]}
{"QuestionId":34131877,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-12-07T10:55:32.523","AcceptedAnswerId":34137685.0,"OwnerUserId":4882455.0,"Title":"Caffe transfer subset of parameters","Body":"<p>I want to take out a subset of <em><\/em> parameters in a given layer from a trained network <strong><\/strong>, and transfer this subset into a layer of another network <strong><\/strong>. \nIn the layer to which this subset is transferred, the remaining parameters of said layer <em>(total - transferred subset)<\/em> should be randomly initialized.<\/p>\n\n<p>Is this achievable? \n(I know it can be done with whole layers of parameters) <\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"34137685","CreationDate":"2015-12-07T15:50:32.843","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>This can be done in python. See <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">net surgery tutorial<\/a>.<br>\nBasically, you should instantiate two <code>caffe.Net<\/code> objects one with parameters of <code>A<\/code> and the other with random. Then you simply assign the subset of parameters from <code>A<\/code> to <code>B<\/code>, saves <code>B<\/code> and you are done. <\/p>\n"}]}
{"QuestionId":34135878,"AnswerCount":3,"Tags":"<python-2.7><machine-learning><neural-network><tensorflow>","CreationDate":"2015-12-07T14:25:01.807","AcceptedAnswerId":null,"OwnerUserId":5599589.0,"Title":"Error in TensorFlow program","Body":"<p>I am learning TensorFlow and I stumble upon <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3%20-%20Neural%20Networks\/multilayer_perceptron.ipynb\" rel=\"nofollow\">this<\/a> example code for creating simple multi-layer sigmoid network. The program in the link is for MNIST database and hand written digit classification.<\/p>\n\n<p>I want to train a network for regression task. I have 30 inputs(float) which is used to predict one output(float). So I tweaked the code to change the task from classification to regression. <\/p>\n\n<p>My problem is that I'm getting an error in tf.Session.run(). The code and the error log is given below.<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>runfile('\/mnt\/sdb6\/Projects\/StockML\/demo1.py', <\/p>\n  \n  <p>wdir='\/mnt\/sdb6\/Projects\/StockML')<\/p>\n  \n  <p>Reloaded modules: tensorflow.python.ops.nn_grad, <\/p>\n  \n  <p>tensorflow.python.training.momentum,<\/p>\n  \n  <p>. . . . <\/p>\n  \n  <p>tensorflow.python.util.protobuf,<\/p>\n  \n  <p>google.protobuf.internal.enum_type_wrapper,<\/p>\n  \n  <p>tensorflow.python.ops.nn_ops, tensorflow.python,<\/p>\n  \n  <p>tensorflow.python.platform.test,<\/p>\n  \n  <p>google.protobuf.internal.api_implementation, tensorflow,<\/p>\n  \n  <p>google.protobuf.internal.encoder<\/p>\n  \n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 1, in \n      runfile('\/mnt\/sdb6\/Projects\/StockML\/demo1.py', wdir='\/mnt\/sdb6\/Projects\/StockML')<\/p>\n  \n  <p>File\n  \"\/usr\/lib\/python2.7\/dist-packages\/spyderlib\/widgets\/externalshell\/sitecustomize.py\",\n  line 685, in runfile\n      execfile(filename, namespace)<\/p>\n  \n  <p>File\n  \"\/usr\/lib\/python2.7\/dist-packages\/spyderlib\/widgets\/externalshell\/sitecustomize.py\",\n  line 78, in execfile\n      builtins.execfile(filename, *where)<\/p>\n  \n  <p>File \"\/mnt\/sdb6\/Projects\/StockML\/demo1.py\", line 69, in \n      sess.run(optimizer, feed_dict={x: batch_xs, y: batch_ys})<\/p>\n  \n  <p>File\n  \"\/home\/rammak\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 345, in run\n      results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)<\/p>\n  \n  <p>File\n  \"\/home\/rammak\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 406, in _do_run\n      except tf_session.StatusNotOK as e:<\/p>\n  \n  <p>AttributeError: 'module' object has no attribute 'StatusNotOK'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53968750","CreationDate":"2018-12-29T10:35:56.220","ParentId":null,"OwnerUserId":"9110100","Title":null,"Body":"<p>Change your logging level from <em>WARN<\/em> to <em>INFO<\/em>, so that can get a better visualization of the error you're getting.<br>\nFor knowledge purpose, you should know there are 5 logging levels:  <\/p>\n\n<ol>\n<li><em>DEBUG<\/em>  <\/li>\n<li><em>INFO<\/em>  <\/li>\n<li><em>WARN<\/em>  <\/li>\n<li><em>ERROR<\/em>  <\/li>\n<li><em>FATAL<\/em>  <\/li>\n<\/ol>\n"},{"AnswerId":"34147295","CreationDate":"2015-12-08T03:08:02.603","ParentId":null,"OwnerUserId":"2879934","Title":null,"Body":"<p>Protobuf error is usually an installation issue , run it in a virtual env<\/p>\n\n<pre><code># On Mac:\n$ sudo easy_install pip  # If pip is not already installed\n$ sudo pip install --upgrade virtualenv\nNext, set up a new virtualenv environment. To set it up in the directory ~\/tensorflow, run:\n$ virtualenv --system-site-packages ~\/tensorflow\n$ cd ~\/tensorflow\nThen activate the virtualenv:\n$ source bin\/activate  # If using bash\n$ source bin\/activate.csh  # If using csh\n(tensorflow)$  # Your prompt should change\nInside the virtualenv, install TensorFlow:\n(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\nYou can then run your TensorFlow program like:\n(tensorflow)$ python tensorflow\/models\/image\/mnist\/convolutional.py\n\n# When you are done using TensorFlow:\n(tensorflow)$ deactivate  # Deactivate the virtualenv\n\n$  # Your prompt should change back\n<\/code><\/pre>\n"},{"AnswerId":"35447593","CreationDate":"2016-02-17T03:53:15.157","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>If you just begin to learn TensorFlow, I would suggest you trying out <a href=\"https:\/\/github.com\/tensorflow\/skflow\/tree\/master\/examples\" rel=\"nofollow\">examples<\/a> in <a href=\"https:\/\/github.com\/tensorflow\/skflow\" rel=\"nofollow\">TensorFlow\/skflow<\/a> first and then once you are more familiar with TensorFlow it would be fairly easy for you to insert TensorFlow code to build a custom model you want (there are also examples for this). <\/p>\n\n<p>Hope those examples for images and text understanding could get you started and let us know if you encounter any issues! (post issues or tag skflow in SO). <\/p>\n"}]}
{"QuestionId":34139673,"AnswerCount":1,"Tags":"<optimization><gradient><tensorflow>","CreationDate":"2015-12-07T17:33:18.720","AcceptedAnswerId":34140268.0,"OwnerUserId":3450762.0,"Title":"stopping gradient optimizer in TensorFlow","Body":"<p>I'm trying to build a simple neural network in Tensorflow, but I have a question about gradient optimization.<\/p>\n\n<p>It might be a naive question, but do I have to set conditions to <em>stop<\/em> the optimizer?  Below is a sample printout from my network and you can see that after iteration (batch gradient descent of all data) 66, the cost begins to increase again.  So is it up to <em>me<\/em> to make sure the optimizer stops at this point?  (note: I didn't put all the output here, but the cost begins to increase exponentially as the number of iterations increase).<\/p>\n\n<p>Thanks for any guidance.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34140268","CreationDate":"2015-12-07T18:07:43.457","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>That's correct - the TensorFlow <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#Optimizer\" rel=\"nofollow\"><code>tf.train.Optimizer<\/code><\/a> classes expose an <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#Optimizer.minimize\" rel=\"nofollow\">operation that you can run to take one (gradient descent-style) step<\/a>, but they do not monitor the current value of the cost or decide when to stop, so you may see increasing cost once the network begins to overfit. <\/p>\n"}]}
{"QuestionId":34141430,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2015-12-07T19:18:45.447","AcceptedAnswerId":37241750.0,"OwnerUserId":4975126.0,"Title":"Tensorflow Tensor reshape and pad with zeros","Body":"<p>Is there a way to reshape a tensor and pad any overflow with zeros? I know ndarray.reshape does this, but as I understand it, converting a Tensor to an ndarray would require flip-flopping between the GPU and CPU. <\/p>\n\n<p>Tensorflow's reshape() documentation says the TensorShapes need to have the same number of elements, so perhaps the best way would be a pad() and then reshape()?<\/p>\n\n<p>I'm trying to achieve:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34143633","CreationDate":"2015-12-07T21:32:28.913","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As far as I know, there's no built-in operator that does this (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#reshape\" rel=\"nofollow noreferrer\"><code>tf.reshape()<\/code><\/a> will give you an error if the shapes do not match). However, you can achieve the same result with a few different operators:<\/p>\n\n<pre><code>a = tf.constant([[1, 2], [3, 4]])\n\n# Reshape `a` as a vector. -1 means \"set this dimension automatically\".\na_as_vector = tf.reshape(a, [-1])\n\n# Create another vector containing zeroes to pad `a` to (2 * 3) elements.\nzero_padding = tf.zeros([2 * 3] - tf.shape(a_as_vector), dtype=a.dtype)\n\n# Concatenate `a_as_vector` with the padding.\na_padded = tf.concat([a_as_vector, zero_padding], 0)\n\n# Reshape the padded vector to the desired shape.\nresult = tf.reshape(a_padded, [2, 3])\n<\/code><\/pre>\n"},{"AnswerId":"37241750","CreationDate":"2016-05-15T17:56:25.027","ParentId":null,"OwnerUserId":"927972","Title":null,"Body":"<p>Tensorflow now offers the pad function which performs padding on a tensor in a number of ways(like opencv2's padding function for arrays):\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad<\/a><\/p>\n\n<pre><code>tf.pad(tensor, paddings, mode='CONSTANT', name=None)\n<\/code><\/pre>\n\n<p>example from the docs above:<\/p>\n\n<pre><code># 't' is [[1, 2, 3], [4, 5, 6]].\n# 'paddings' is [[1, 1,], [2, 2]].\n# rank of 't' is 2.\npad(t, paddings, \"CONSTANT\") ==&gt; [[0, 0, 0, 0, 0, 0, 0],\n                                  [0, 0, 1, 2, 3, 0, 0],\n                                  [0, 0, 4, 5, 6, 0, 0],\n                                  [0, 0, 0, 0, 0, 0, 0]]\n\npad(t, paddings, \"REFLECT\") ==&gt; [[6, 5, 4, 5, 6, 5, 4],\n                                 [3, 2, 1, 2, 3, 2, 1],\n                                 [6, 5, 4, 5, 6, 5, 4],\n                                 [3, 2, 1, 2, 3, 2, 1]]\n\npad(t, paddings, \"SYMMETRIC\") ==&gt; [[2, 1, 1, 2, 3, 3, 2],\n                                   [2, 1, 1, 2, 3, 3, 2],\n                                   [5, 4, 4, 5, 6, 6, 5],\n                                   [5, 4, 4, 5, 6, 6, 5]]\n<\/code><\/pre>\n"}]}
{"QuestionId":34146656,"AnswerCount":1,"Tags":"<function><machine-learning><gpgpu><theano>","CreationDate":"2015-12-08T01:53:34.197","AcceptedAnswerId":34152327.0,"OwnerUserId":3041068.0,"Title":"Is there a way to change a function's update list without re-compiling it in Theano?","Body":"<p>Indeed I want to change the learning rate at different periods of training. Something like:<\/p>\n\n<pre><\/pre>\n\n<p>Apparently recompiling a new function for every iteration is going to be too slow.\nSo I was wondering is there a better way to do it in Theano?\nThanks!<\/p>\n","answers":[{"AnswerId":"34152327","CreationDate":"2015-12-08T09:33:16.950","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can make the learning rate a symbolic variable and pass it into the training function like this:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile(input_size, hidden_size, output_size):\n    W_h = theano.shared(numpy.random.standard_normal(size=(input_size, hidden_size)).astype(theano.config.floatX))\n    b_h = theano.shared(numpy.zeros((hidden_size,), dtype=theano.config.floatX))\n    W_y = theano.shared(numpy.random.standard_normal(size=(hidden_size, output_size)).astype(theano.config.floatX))\n    b_y = theano.shared(numpy.zeros((output_size,), dtype=theano.config.floatX))\n\n    x = tt.matrix('x')\n    z = tt.ivector('z')\n    learning_rate = tt.scalar()\n    h = tt.tanh(theano.dot(x, W_h) + b_h)\n    y = tt.nnet.softmax(theano.dot(h, W_y) + b_y)\n    cost = tt.nnet.categorical_crossentropy(y, z).mean()\n    updates = [(p, p - learning_rate * tt.grad(cost, p)) for p in (W_h, b_h, W_y, b_y)]\n    return theano.function([x, z, learning_rate], outputs=cost, updates=updates)\n\n\ndef main():\n    input_size = 5\n    hidden_size = 4\n    output_size = 3\n    train = compile(input_size, hidden_size, output_size)\n    print train([[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]], [1, 2], 0.1)\n\n\nmain()\n<\/code><\/pre>\n\n<p>Note that the training function now has three parameters; the third is the learning rate.<\/p>\n"}]}
{"QuestionId":34146924,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-12-08T02:25:51.077","AcceptedAnswerId":null,"OwnerUserId":4159177.0,"Title":"How to implement object detection by Caffe and CNN","Body":"<p>I would like to implement object detection using Caffe framework and Convolution Neural Network, could you recommend some papers and demos about that?<br>\nI just need to know how to implement it.<br>\nIf you can provide the source code, it will be perfect.<\/p>\n","answers":[{"AnswerId":"34150828","CreationDate":"2015-12-08T08:03:50.300","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>A very good starting point to object detection using Deep Learning is the series of papers on R-CNN by Ross Girshick.<br>\nYou can find <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\" rel=\"noreferrer\">python impelementation<\/a> of faster R-CNN a good reference for implementation.<\/p>\n"}]}
{"QuestionId":34147244,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-08T03:02:28.067","AcceptedAnswerId":null,"OwnerUserId":5652641.0,"Title":"torch FloatTensor toString method?","Body":"<p>I have a torch.FloatTensor that is 2 rows and 200 columns. I want to print the lines to a text file. Is there a toString() method for the torch.FloatTensor? If not, how do you print the line to the file? Thanks.<\/p>\n\n<p>I can convert the FloatTensor into a Lua table: <\/p>\n\n<pre><\/pre>\n\n<p>Is there an easy way to convert the Lua table to a string, so I can write it to file? Thanks!<\/p>\n","answers":[{"AnswerId":"34151219","CreationDate":"2015-12-08T08:32:19.287","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>There is a built-in table conversion called <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/dddfc3c\/doc\/utility.md#table-torchtotableobject\" rel=\"nofollow\"><code>torch.totable<\/code><\/a>. If what you want to do is save and load your tensor, then it's even easier to use Torch native <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/dddfc3c\/doc\/serialization.md\" rel=\"nofollow\">serialization<\/a> like <code>torch.save('file.txt', tensor, 'ascii')<\/code>.<\/p>\n"}]}
{"QuestionId":34149313,"AnswerCount":1,"Tags":"<python><vector><concatenation><theano><chain>","CreationDate":"2015-12-08T06:19:39.070","AcceptedAnswerId":null,"OwnerUserId":5173166.0,"Title":"Any way to combine Theano tensor vectors similar to itertools.chain?","Body":"<p>Does anyone know a way to combine tensor variable (containing vectors) in theano into another variable that references the originals, similar to itertools.chain?<\/p>\n\n<p>Right now I can't get a result where  results where z.shape[0] = shape x.shape[0] + y.shape[0]. It always ends up with z having the same shape as x.<\/p>\n\n<p>I'd appreciate any help you can provide.<\/p>\n","answers":[{"AnswerId":"34152135","CreationDate":"2015-12-08T09:23:57.333","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's a demonstration of <code>theano.tensor.concatenate<\/code>. You can use the <code>axis<\/code> parameter to adjust which dimension gets concatenated.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\nx, y = tt.matrices('x', 'y')\nz1 = tt.concatenate([x, y])\nz2 = tt.concatenate([x, y], axis=0)\nz3 = tt.concatenate([x, y], axis=1)\nf = theano.function([x, y], [z1, z2, z3])\nfor output in f([[1, 2], [3, 4]], [[5, 6], [7, 8]]):\n    print output.shape\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>(4L, 2L)\n(4L, 2L)\n(2L, 4L)\n<\/code><\/pre>\n\n<p>So I am unable to reproduce the stated behaviour. Using <code>z = tt.concatenate([x, y])<\/code> yields a single tensor that has <code>z.shape[0] = x.shape[0] + y.shape[0]<\/code> as desired.<\/p>\n"}]}
{"QuestionId":34150921,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2015-12-08T08:11:18.623","AcceptedAnswerId":null,"OwnerUserId":3631966.0,"Title":"session.run error when implement a pairwise cnn with tensorflow","Body":"\n\n<p>I am very new to tensorflow. Now I am implementing a pairwise cnn model. While I got the error:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.pywrap_tensorflow.StatusNotOK: Not found:\n  FeedInputs: unable to find feed output Placeholder:0<\/p>\n<\/blockquote>\n\n<p>The error is attributed to the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Specifically in  function,  is a dictionary containing input data and corresponding labels.  and  are the output of the same neural networks and finally I would like to compare  and .<\/p>\n\n<p>The gist codes are:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here I create three placeholders, and do some operations like . Forward is the neural networks used in my implementation. Then in the training part, I load text data and labels with , finally run the optimization operation with . The codes is available below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><strong>NEW ERROR<\/strong><\/p>\n\n<p>I tried to change the maps1 as Tensors instead of numpy ndarray, the error above disappeared. Unfortunately new errors raised.<\/p>\n\n<blockquote>\n  <p>File \"Train.py\", line 55, in train      _,train_loss,train_feature1,train_feature2=sess.run([model.train_op,model.cost,model.feature1,model.feature2],feed)\n        File \"\/usr\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 334, in run\n          np_val = np.array(subfeed_val, dtype=subfeed_t.dtype.as_numpy_dtype)\n      ValueError: setting an array element with a sequence.<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":34152090,"AnswerCount":1,"Tags":"<lua><torch><lstm>","CreationDate":"2015-12-08T09:21:21.317","AcceptedAnswerId":null,"OwnerUserId":5653522.0,"Title":"Torch: RNN clones run out of GPU memory","Body":"<p>Karpathy's <a href=\"https:\/\/github.com\/karpathy\/char-rnn\/blob\/master\/util\/model_utils.lua#L113\" rel=\"nofollow\">char-rnn<\/a> (based on Wojciechz learning_to_execute) uses a common RNN hack:<\/p>\n\n<ul>\n<li>clone a prototype network as many times as there are time steps per sequence<\/li>\n<li>share the parameters between the clones<\/li>\n<\/ul>\n\n<p>I can watch my 5GB GPU memory run out when I clone 217 times (the threshold is likely lower), resulting in this:<\/p>\n\n<p><\/p>\n\n<p>The problem is the  function (linked above). The clones seem to point to the same physical parameters storage in the prototype, but for some reason it still explodes.<\/p>\n\n<p>Has anyone encountered this and\/or have an idea how to train really long sequences? <\/p>\n\n<p>(Same question asked here: <a href=\"https:\/\/github.com\/karpathy\/char-rnn\/issues\/108\" rel=\"nofollow\">https:\/\/github.com\/karpathy\/char-rnn\/issues\/108<\/a>)<\/p>\n","answers":[{"AnswerId":"44875296","CreationDate":"2017-07-02T21:18:50.040","ParentId":null,"OwnerUserId":"1949987","Title":null,"Body":"<p>To run the model, I had to increase memory capacity on the GPUs. With Sun's Grid Engine, use <code>-l h_vmem=8G<\/code> for 8 GB<\/p>\n\n<p>Otherwise, you can try <a href=\"https:\/\/github.com\/jcjohnson\/torch-rnn\" rel=\"nofollow noreferrer\">torch-rnn<\/a>. It uses Adam for optimization and hard-codes the RNN\/LSTM forward\/backward passes for space\/time efficiency. This also avoids headaches with cloning models. <\/p>\n"}]}
{"QuestionId":34153930,"AnswerCount":3,"Tags":"<python><machine-learning><nlp><tensorflow><linguistics>","CreationDate":"2015-12-08T10:50:57.270","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"Using RNN tensorflow language model to predict the probabilities of test sentences","Body":"<p>I was able to train a language model using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html#language-modeling\" rel=\"nofollow noreferrer\">tensorflow tutorials<\/a> , the models are saved as checkpoint files as per the <a href=\"https:\/\/stackoverflow.com\/questions\/33980496\/tensorflow-rnn-model-path\/33981028?noredirect=1#comment55719523_33981028\">code given here<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>Now I need to restore the checkpoint and use it in the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I cannot find any way of encoding the test sentences and getting sentence probability output from the trained checkpoint model.<\/p>\n\n<p>The tutorials mention following code:<\/p>\n\n<pre><\/pre>\n\n<p>but that it is for training and I cannot figure out how do I get the actual probabilities.\nI should Ideally get something like :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34160986","CreationDate":"2015-12-08T16:26:13.443","ParentId":null,"OwnerUserId":"3826327","Title":null,"Body":"<p>I had the same question and I think I found a way around it but I am not an expert so comments are welcomed!<\/p>\n\n<p>In the PTBModel class, you need to add this line:<\/p>\n\n<pre><code>    self._proba = tf.nn.softmax(logits)\n<\/code><\/pre>\n\n<p>before (or within) this loop:<\/p>\n\n<pre><code>    if not is_training:\n        return\n<\/code><\/pre>\n\n<p>and also add this property:<\/p>\n\n<pre><code>      @property\n      def proba(self):\n          return self._proba\n<\/code><\/pre>\n\n<p>Now in the run_epoch function you can get the probabilities using something like:<\/p>\n\n<pre><code>    cost, state, proba, _ = session.run([m.cost, m.final_state, m.proba, eval_op],...\n<\/code><\/pre>\n\n<p>From here you should have access to all the probabilities with proba. There may be a better way ...\nhope this help !<\/p>\n"},{"AnswerId":"52755942","CreationDate":"2018-10-11T08:48:38.150","ParentId":null,"OwnerUserId":"4794701","Title":null,"Body":"<p>There should be start symbol (SOS, or something else) and end symbol (EOS, or something else) symbols in the vocab and you can get the index of end symbol then get the corresponding probability value in proba.<\/p>\n"},{"AnswerId":"39932744","CreationDate":"2016-10-08T13:10:03.517","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>You should first know how to calculate the score. Thanks to the Markov Assumption we need not to calculate too much(based on the chain rule). What should be worked out is just the probabilities of the following few words(let's say one for convenience). Then the key becomes how to calculate the rate of the next word. <\/p>\n\n<pre><code>probab = session.run(myours.proba], feed_dict) # only the input is needed\n<\/code><\/pre>\n\n<p>You should create a model named myours as described in @Romain's answer(mine is just a complementation of that). And create your own ptb_iterator to yield only the x (first you should use raw_input or others to get your words for example in a loop). <\/p>\n\n<pre><code>for i in range(epoch_size):                   \n    x = data[:, i*num_steps:(i+1)*num_steps]  \n    yield x # the old one with y is better, use y to locate the probability of the coming word \n<\/code><\/pre>\n\n<p>Now that you have the probability you can do everything the language model can do. For example predict the next word.  <\/p>\n\n<pre><code>list(probab[0]).index(max(probab[0])) #  an id_to_word dict should be created\n<\/code><\/pre>\n\n<p>You will get n-1 scores(more precisely n-1 probability distribution of the vocabulary length, and you should choose one based on the index of the coming word) for an n-word sentence. <\/p>\n\n<p>I use this way to calculate the score(not certain if it's right or wrong yet, I encounter the same problem as <a href=\"https:\/\/stackoverflow.com\/q\/18928243\/3552975\">this one<\/a>):<\/p>\n\n<pre><code>pro += math.log(list(probab[0])[y[0]], 2)\n<\/code><\/pre>\n\n<p>PS:  <\/p>\n\n<ol>\n<li><p>To save your time you can save the variables in the session the first time you train the network and\nrestore it every time you want to have a test yourself.  <\/p>\n\n<p>save_path = saver.save( session, \".\/tmp\/model.epoch.%03d.ckpt\" % (i + 1))\nsaver.restore(session, \".\/tmp\/model.epoch.013.ckpt\") # only last one<\/p><\/li>\n<li><p>a sentence to ids function is also needed:<\/p>\n\n<p>return [word_to_id[word] if word in word_to_id else word_to_id[<code>\"&lt;unk&gt;\"<\/code>] for word in nltk.tokenize.word_tokenize(sentence)]<\/p><\/li>\n<\/ol>\n\n<p>Hope that this helps and is explanatory enough to answer your question. <\/p>\n"}]}
{"QuestionId":34154219,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-08T11:03:55.583","AcceptedAnswerId":34156154.0,"OwnerUserId":5435604.0,"Title":"Choose a number from a a given pmf in theano","Body":"<p>Say I have an array . Let  be a probability mass function for a Random Variable .  Now, I am writing a theano code, where I have a  generated at each iteration and  I also have  weight matrices. (Here.)<\/p>\n\n<p>Now, in each iteration I would want to select one of these weight matrices for further propagation. Could someone help out regarding how to write this piece of code. I am not sure I can write the exact code required to enable backpropagation(i.e. have the gradients corrected properly)<\/p>\n\n<p>Note that all the s as well as the enter  are model parameters. <\/p>\n\n<blockquote>\n  <p>Edit<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>and my grad vaiables are  Ignore the other variables as they are defined somewhere else. Now, I get the following error<\/p>\n\n<blockquote>\n  <p>\n  Traceback (most recent call last):\n    File \"trainmnist_mixture.py\", line 55, in \n      encoder.createGradientFunctions()<\/p>\n  \n  <p>File \"\/home\/amartya\/Variational-Autoencoder\/Theano\/VariationalAutoencoder_mixture.py\", line 118, in createGradientFunctions\n      derivatives = T.grad(logp,gradvariables)<\/p>\n  \n  <p>File \"\/usr\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/gradient.py\", line 543, in grad\n      grad_dict, wrt, cost_name)<\/p>\n  \n  <p>File \"\/usr\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/gradient.py\", line 1273, in _populate_grad_dict\n      rval = [access_grad_cache(elem) for elem in wrt]<\/p>\n  \n  <p>File \"\/usr\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/gradient.py\", line 1233, in access_grad_cache\n      term = access_term_cache(node)[idx]<\/p>\n  \n  <p>File \"\/usr\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/gradient.py\", line 944, in access_term_cache\n      output_grads = [access_grad_cache(var) for var in node.outputs]<\/p>\n  \n  <p>File \"\/usr\/lib\/python2.7\/site-packages\/Theano-0.6.0-py2.7.egg\/theano\/gradient.py\", line 1243, in access_grad_cache\n      term.type.why_null)<\/p>\n  \n  <p><strong>theano.gradient.NullTypeGradError: tensor.grad encountered a NaN. This variable is Null because the grad method for input 0 (Subtensor{int64:int64:}.0) of the Nonzero op is mathematically undefined<\/strong><\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34156154","CreationDate":"2015-12-08T12:39:37.783","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use the <code>choice<\/code> method of a <code>RandomStreams<\/code> instance. More on random numbers in Theano can be found in the documentation <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/shared_randomstreams.html\" rel=\"nofollow\">here<\/a> and <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html#id1\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>Here's an example:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\nimport theano.tensor.shared_randomstreams\n\nn = 6\nalpha = [1] * n\nseed = 1\nw = theano.shared(numpy.random.randn(n, 2, 2).astype(theano.config.floatX))\np = theano.shared(numpy.random.dirichlet(alpha).astype(theano.config.floatX))\nrng = tt.shared_randomstreams.RandomStreams(seed=seed)\ni = rng.choice(size=(1,), a=n, p=p)\nf = theano.function([], [p, i, w[i]])\nprint f()\nprint f()\nprint f()\nprint f()\n<\/code><\/pre>\n"}]}
{"QuestionId":34156639,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2015-12-08T13:03:37.363","AcceptedAnswerId":34157312.0,"OwnerUserId":1978504.0,"Title":"Tensorflow (python): \"ValueError: setting an array element with a sequence\" in train_step.run(...)","Body":"<p>I'm trying to implement a simple logistic regression model trained with my own set of images, but I am getting this error when I try to train the model:<\/p>\n\n<pre><\/pre>\n\n<p>The data I'm feeding to  is like this:<\/p>\n\n<ul>\n<li>batch_xs: list of tensor objects representing images of 100x100 (10,000 long tensors)<\/li>\n<li>batch_ys: list of labels as floats (1.0 or 0.0)<\/li>\n<\/ul>\n\n<p>What am I doing wrong?\nThx in advance!<\/p>\n\n<p><strong>EDIT 1<\/strong>: It seams the problem was that I had to evaluate the tensors in  before passing them to . I thought the run method would take care of that, but I guess I was wrong?\nAnyway, so once I did this before calling the function:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT 2<\/strong>: I had several issues even after doing what is suggested in the answers below. I finally fixed everything by ditching tensors and using numpy arrays.<\/p>\n\n<p>Hope this helps someone else<\/p>\n","answers":[{"AnswerId":"34161823","CreationDate":"2015-12-08T17:05:55.080","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>feed_dict<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Operation.run\" rel=\"noreferrer\"><code>Operation.run()<\/code><\/a> (also <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor.eval\" rel=\"noreferrer\"><code>Tensor.eval()<\/code><\/a>) accepts a dictionary mapping <code>Tensor<\/code> objects (usually <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> tensors) to a numpy array (or objects that can be trivially converted to a numpy array).<\/p>\n\n<p>In your case, you are passing <code>batch_xs<\/code>, which is a list of numpy arrays, and TensorFlow does not know how to convert this to a numpy array. Let's say that <code>batch_xs<\/code> is defined as follows:<\/p>\n\n<pre><code>batch_xs = [np.random.rand(100, 100),\n            np.random.rand(100, 100),\n            ...,                       # 29 rows omitted.\n            np.random.rand(100, 100)]  # len(batch_xs) == 32.\n<\/code><\/pre>\n\n<p>We can convert <code>batch_xs<\/code> into a <code>32 x 100 x 100<\/code> array using the following:<\/p>\n\n<pre><code># Convert each 100 x 100 element to 1 x 100 x 100, then vstack to concatenate.\nbatch_xs = np.vstack([np.expand_dims(x, 0) for x in batch_xs])\nprint batch_xs.shape\n# ==&gt; (32, 100, 100) \n<\/code><\/pre>\n\n<p>Note that, if <code>batch_ys<\/code> is a list of floats, this will be transparently converted into a 1-D numpy array by TensorFlow, so you should not need to convert this argument.<\/p>\n\n<p><strong>EDIT:<\/strong> mdaoust makes a valid point in the comments: If you pass a list of arrays into <code>np.array<\/code> (and therefore as the value in a <code>feed_dict<\/code>), it will automatically be <code>vstack<\/code>ed, so there should be no need to convert your input as I suggested. Instead, it sounds like you have a mismatch between the shapes of your list elements. Try adding the following:<\/p>\n\n<pre><code>assert all(x.shape == (100, 100) for x in batch_xs)\n<\/code><\/pre>\n\n<p>...before the call to <code>train_step.run()<\/code>, and this should reveal whether you have a mismatch.<\/p>\n"},{"AnswerId":"34157312","CreationDate":"2015-12-08T13:36:39.083","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>This particular error is coming out of <code>numpy<\/code>. Calling <code>np.array<\/code> on a sequence with a inconsistant dimensions can throw it.<\/p>\n\n<pre><code>&gt;&gt;&gt; np.array([1,2,3,[4,5,6]])\n\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>It looks like it's failing at the point where <code>tf<\/code> ensures that all the elements of the <code>feed_dict<\/code> are <code>numpy.array<\/code>s.<\/p>\n\n<p>Check your <code>feed_dict<\/code>.<\/p>\n"}]}
{"QuestionId":34160692,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-08T16:13:15.067","AcceptedAnswerId":null,"OwnerUserId":5654767.0,"Title":"Error using dropout in tensorflow","Body":"\n\n<p>I'm trying to use the dropout functionality in tensorflow:     <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This example is very similar to how it's done in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow noreferrer\">the tutorial<\/a>; however, I end up with the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have some trouble to understand the dtype , which seems to be the background to the problem. I've also tried to specify , but that doesn't fix anything. <\/p>\n\n<p>I also tried this Example, which works fine with :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, if I cast  instead of  I get the same error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>It seems like this problem only arises when using dropout directly on Variables, works for placeholders and for products of Variables and placeholders, Example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34162382","CreationDate":"2015-12-08T17:36:10.740","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is a bug in the implementation of <code>tf.nn.dropout<\/code> that was fixed in a recent commit, and will be included in the next release of TensorFlow. For now, to avoid the issue, either <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\">build TensorFlow from source<\/a>, or modify your program as follows:<\/p>\n\n<pre><code>#dx = tf.nn.dropout(x, keep_prob)\ndx = tf.nn.dropout(tf.identity(x), keep_prob)\n<\/code><\/pre>\n"}]}
{"QuestionId":34162549,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><neural-network><classification><tensorflow>","CreationDate":"2015-12-08T17:45:35.320","AcceptedAnswerId":34163681.0,"OwnerUserId":1266985.0,"Title":"Miminum requirements for Google tensorflow image classifier","Body":"<p>We are planning to build image classifiers using Google Tensorflow.<\/p>\n\n<p>I wonder what are the minimum and what are the optimum requirements to train a custom image classifier using a convolutional deep neural network?<\/p>\n\n<p>The questions are specifically:<\/p>\n\n<ul>\n<li>how many images per class should be provided at a minimum?<\/li>\n<li>do we need to appx. provide the same amount of training images per class or can the amount per class be disparate?<\/li>\n<li>what is the impact of wrong image data in the training data? E.g. 500 images of a tennis shoe and 50 of other shoes.  <\/li>\n<li>is it possible to train a classifier with much more classes than the recently published inception-v3 model? Let's say: 30.000.<\/li>\n<\/ul>\n","answers":[{"AnswerId":"34163681","CreationDate":"2015-12-08T18:49:22.673","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>\"how many images per class should be provided at a minimum?\"<\/p>\n\n<p>Depends how you train.<\/p>\n\n<p>If training a new model from scratch, purely supervised: For a rule of thumb on the number of images, you can look at the MNIST and CIFAR tasks. These seem to work OK with about 5,000 images per class. That's if you're training from scratch.<\/p>\n\n<p>You can probably bootstrap your network by beginning with a model trained on ImageNet. This model will already have good features, so it should be able to learn to classify new categories without as many labeled examples. I don't think this is well-studied enough to tell you a specific number.<\/p>\n\n<p>If training with unlabeled data, maybe only 100 labeled images per class. There is a lot of recent research work on this topic, though not scaling to as large of tasks as Imagenet.\nSimple to implement:<\/p>\n\n<pre><code>http:\/\/arxiv.org\/abs\/1507.00677\n<\/code><\/pre>\n\n<p>Complicated to implement:<\/p>\n\n<pre><code>http:\/\/arxiv.org\/abs\/1507.02672\nhttp:\/\/arxiv.org\/abs\/1511.06390\nhttp:\/\/arxiv.org\/abs\/1511.06440\n<\/code><\/pre>\n\n<p>\"do we need to appx. provide the same amount of training images per class or can the amount per class be disparate?\"<\/p>\n\n<p>It should work with different numbers of examples per class.<\/p>\n\n<p>\"what is the impact of wrong image data in the training data? E.g. 500 images of a tennis shoe and 50 of other shoes.\"<\/p>\n\n<p>You should use the label smoothing technique described in this paper:<\/p>\n\n<pre><code>http:\/\/arxiv.org\/abs\/1512.00567\n<\/code><\/pre>\n\n<p>Smooth the labels based on your estimate of the label error rate.<\/p>\n\n<p>\"is it possible to train a classifier with much more classes than the recently published inception-v3 model? Let's say: 30.000.\"<\/p>\n\n<p>Yes<\/p>\n"},{"AnswerId":"34162898","CreationDate":"2015-12-08T18:05:17.353","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<blockquote>\n  <p>How many images per class should be provided at a minimum?<\/p>\n  \n  <p>do we need to appx. provide the same amount of training images per class or can the amount per class be disparate?<\/p>\n  \n  <p>what is the impact of wrong image data in the training data? E.g. 500 images of a tennis shoe and 50 of other shoes.<\/p>\n<\/blockquote>\n\n<p>These three questions are not really TensorFlow specific. But the short answer is, it depends on the resiliency of your model in handling unbalanced data set and noisy labels.<\/p>\n\n<blockquote>\n  <p>is it possible to train a classifier with much more classes than the recently published inception-v3 model? Let's say: 30.000.<\/p>\n<\/blockquote>\n\n<p>Yes, definitely. This would mean a much larger classifier layer, so your training time might be longer. Other than that, there are no limitations in TensorFlow.<\/p>\n"}]}
{"QuestionId":34163953,"AnswerCount":1,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2015-12-08T19:05:17.130","AcceptedAnswerId":null,"OwnerUserId":5655900.0,"Title":"Tensorflow, binary classification","Body":"<p>Is it possible to make a binary classification? Especially for pedestrian detection, whether it is or not a pedestrian. I couldn't find anything in the API or any good tutorials for this. \nI tried to adapt the code from the deep MNIST tutorial, that was used for multi-class classification; I made the images with pedestrians in them labeled with 1, and the negatives with 0, and used 3 channels(for colours, shouldn't be a problem right?), but the accuracy just jumps all over the place.<\/p>\n\n<p>The code<\/p>\n\n<pre><\/pre>\n\n<p>The output<\/p>\n\n<pre><\/pre>\n\n<p>Help would be apriciated, thanks.<\/p>\n","answers":[{"AnswerId":"34455063","CreationDate":"2015-12-24T15:45:42.493","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p>You should definitely use tensorboard to visualize the cross entropy , biases and weights summaries. I think it will give you a much better  view of what is going on.<\/p>\n\n<p>Try with this code and then run tensorboard:<\/p>\n\n<pre><code>import dataset as input_data\nimport tensorflow as tf\n\n\ndef weight_variable(shape):\n    initial = tf.truncated_normal(shape, stddev=0.1)\n    return tf.Variable(initial)\n\n\ndef bias_variable(shape):\n    initial = tf.constant(0.1, shape=shape)\n    return tf.Variable(initial)\n\n\ndef conv2d(x, W):\n    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')\n\n\ndef max_pool_2x2(x):\n    return tf.nn.max_pool(x, ksize=[1, 3, 3, 1],\n                          strides=[1, 2, 2, 1], padding='SAME')\n\n\ndata = input_data.read_data_sets()\n\nsess = tf.InteractiveSession()\n\nx = tf.placeholder(\"float\", shape=[None, input_data.HEIGHT * input_data.WIDTH * 3])\ny_ = tf.placeholder(\"float\", shape=[None, 2])\n\nW_conv1 = weight_variable([5, 5, 3, 64])\nb_conv1 = bias_variable([64])\n\nx_image = tf.reshape(x, [-1, input_data.WIDTH, input_data.HEIGHT, 3])\n\nh_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)\nh_pool1 = max_pool_2x2(h_conv1)\nh_norm1 = tf.nn.lrn(h_pool1, 4, bias=1.0, alpha=0.001 \/ 9.0, beta=0.75)\n\nW_conv2 = weight_variable([5, 5, 64, 64])\nb_conv2 = bias_variable([64])\n\nh_conv2 = tf.nn.relu(conv2d(h_norm1, W_conv2) + b_conv2)\nh_norm2 = tf.nn.lrn(h_conv2, 4, bias=1.0, alpha=0.001 \/ 9.0, beta=0.75)\nh_pool2 = max_pool_2x2(h_norm2)\n\nW_fc1 = weight_variable([input_data.HEIGHT \/ 4 * input_data.WIDTH \/ 4 * 64, 1024])\nb_fc1 = bias_variable([1024])\n\nh_pool2_flat = tf.reshape(h_pool2, [-1, input_data.HEIGHT \/ 4 * input_data.WIDTH \/ 4 * 64])\nh_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)\n\nkeep_prob = tf.placeholder(\"float\")\nh_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)\n\nW_fc2 = weight_variable([1024, 2])\nb_fc2 = bias_variable([2])\n\ny_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)\n\n# Add summary ops to collect data\nw_fc2_hist = tf.histogram_summary(\"weights_fc2\", W_fc2)\nb_fc2_hist = tf.histogram_summary(\"bias_fc2\", b_fc2)\nw_in_hist = tf.histogram_summary(\"weights_in\", W)\nb_in_hist = tf.histogram_summary(\"bias_in\", b)\ny_hist = tf.histogram_summary(\"y_conv\", y_conv)\n\ncross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))\nce_summ = tf.scalar_summary(\"cross entropy\", cross_entropy)\ntrain_step = tf.train.AdamOptimizer(1e-6).minimize(cross_entropy)\ncorrect_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\naccuracy_summary = tf.scalar_summary(\"train accuracy\", accuracy)\n# Merge all the summaries and write them out to \/tmp\/tf\nmerged = tf.merge_all_summaries()\nwriter = tf.train.SummaryWriter(\"\/tmp\/tf\", sess.graph_def)\nsess.run(tf.initialize_all_variables())\nfor i in range(20000):\n    batch = data.train.next_batch(50)\n    feed={x: batch[0], y_: batch[1], keep_prob: 0.5}\n    result = sess.run([merged, accuracy, train_step], feed_dict=feed)        \n    if i % 100 == 0:  # Record summary data, and the accuracy\n            summary_str = result[0]\n            acc = result[1]\n            writer.add_summary(summary_str, i)\n            print(\"Accuracy at step {0}\/{1}: {2}%\".format(i, 20000, int(acc*100)))\n\nprint \"test accuracy %g\" % accuracy.eval(feed_dict={\n    x: data.test.images, y_: data.test.labels, keep_prob: 1.0})\n<\/code><\/pre>\n"}]}
{"QuestionId":34167537,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-12-08T22:47:33.837","AcceptedAnswerId":34168743.0,"OwnerUserId":3830011.0,"Title":"Can't optimize multivariate linear regression in Tensorflow","Body":"<p>I worked the single variable example from the tensorflow tutorials, but I am having trouble optimizing a multivariate linear regression problem in Tensorflow. <\/p>\n\n<p>I am using a dataset of Portland housing prices that is used <a href=\"http:\/\/openclassroom.stanford.edu\/MainFolder\/DocumentPage.php?course=MachineLearning&amp;doc=exercises\/ex3\/ex3.html\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>I am completely new to Tensorflow, I am sure there are some horrible looking things here.<\/p>\n\n<p>The optimization does not seem to be working at all. It quickly blows up to infinity. Any help is appreciated.<\/p>\n\n<pre><\/pre>\n\n<p>For output I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>I built from source and I am using python3, after the fixes went through to allow this. I doubt this has anything to do with it, but just wanted to make sure. I am sure it is lack of user knowledge.<\/p>\n","answers":[{"AnswerId":"34168743","CreationDate":"2015-12-09T00:31:49.387","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>Your learning rate is too high, so the solution is jumping back and forth, farther and farther.<\/p>\n\n<p>It's generally good practice for a problem like this to normalize your input ranges, for example, so they have mean(0) and var(1).<\/p>\n"}]}
{"QuestionId":34167578,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-08T22:50:27.390","AcceptedAnswerId":34190263.0,"OwnerUserId":904032.0,"Title":"Any status update, on an iOS example for Tensorflow?","Body":"<p>The following response on github mentions that an iOS sample is under active development. I'm just wondering if there's been any timeframe update. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16<\/a> ... Thank you.<\/p>\n","answers":[{"AnswerId":"34190263","CreationDate":"2015-12-09T22:31:12.360","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>No updates yet, but rest assured it's not forgotten or being ignored.<\/p>\n\n<p>(Update 2016-01-01:  This answer is still current;  it's still being worked on.)<\/p>\n"}]}
{"QuestionId":34168437,"AnswerCount":2,"Tags":"<tensorflow><skflow>","CreationDate":"2015-12-09T00:02:52.163","AcceptedAnswerId":null,"OwnerUserId":334757.0,"Title":"Error running skflow example","Body":"\n\n<p>I just installed skflow and TensorFlow and I'm having a problem with the example that comes with skflow. The example code is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I run this example, I get:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The failure occurs on:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Any help will be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"35402416","CreationDate":"2016-02-15T05:58:49.143","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>thanks for using skflow! We've added pandas support long time ago. You can find the specific implementation <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/io\/pandas_io.py\" rel=\"nofollow\">io\/pandas_io.py in skflow<\/a>: <\/p>\n\n<p>More of our examples now use pandas to load the data, for example, the text classification examples, like <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/examples\/text_classification.py\" rel=\"nofollow\">this one<\/a>: <\/p>\n\n<p>Hope this helps and happy using skflow!<\/p>\n"},{"AnswerId":"34182534","CreationDate":"2015-12-09T15:24:41.870","ParentId":null,"OwnerUserId":"5659897","Title":null,"Body":"\n\n<p>I think this is an issue with the interface between skflow and pandas. Try calling <code>.values<\/code> on the data frames before passing them to skflow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)\n\nlr = LogisticRegression()\nlr.fit(X_train.values, y_train.values)\nprint accuracy_score(lr.predict(X_test.values), y_test.values)\n\n# 3 layer neural network with rectified linear activation.\n\nrandom.seed(42)\nclassifier = skflow.TensorFlowDNNClassifier(hidden_units=[10, 20, 10],\n                                        n_classes=2, batch_size=128, steps=500,\n                                        learning_rate=0.05)\nclassifier.fit(X_train.values, y_train.values)\nprint accuracy_score(classifier.predict(X_test.values), y_test.values)\n<\/code><\/pre>\n"}]}
{"QuestionId":34170610,"AnswerCount":3,"Tags":"<python><classification><tensorflow>","CreationDate":"2015-12-09T04:16:34.460","AcceptedAnswerId":34171696.0,"OwnerUserId":4616159.0,"Title":"Tensorflow Image Recognition tutorial seemingly out of date?","Body":"<p>I am currently working through the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">Image Recognition Tutorial<\/a> from the Tensorflow website but am having a hard time getting started...<\/p>\n\n<p>I have installed Tensorflow through pip as per the instructions on the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#binary-installation\" rel=\"nofollow\">repo<\/a>.<\/p>\n\n<p>When I cd in to my tensorflow installation () my  folder does not contain . I cannot seem to find this python file anywhere on the web and would like to get started playing around with image classification.<\/p>\n","answers":[{"AnswerId":"37740634","CreationDate":"2016-06-10T05:37:46.547","ParentId":null,"OwnerUserId":"6349874","Title":null,"Body":"<p>Try to find this models folder in the tensorflow which is also in another tensorflow folder. <\/p>\n\n<pre><code>cd tensorflow\/tensorflow\/models\/image\/imagenet\n<\/code><\/pre>\n\n<p>In the imagenet folder, there is classify_image.py<\/p>\n"},{"AnswerId":"34173630","CreationDate":"2015-12-09T08:04:44.143","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As hsc points out, the <code>classify_image.py<\/code> script isn't in the binary release that you have installed (0.5.0). However, the script is compatible with the binary release, so you can download the script from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py\" rel=\"nofollow\">here<\/a>, and run <code>python classify_image.py<\/code> to try it out. Alternatively, you can install the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#pip_install\" rel=\"nofollow\">latest binary version of TensorFlow (0.6.0)<\/a>, which includes the <code>classify_image.py<\/code> script.<\/p>\n"},{"AnswerId":"34171696","CreationDate":"2015-12-09T05:55:51.550","ParentId":null,"OwnerUserId":"4761675","Title":null,"Body":"<p>The 'classify_image.py' isn't in the 0.5.0 released version which you have installed. But you can find the file you want <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py\" rel=\"nofollow\">here<\/a>. Actually I think this is because the tutorial is up-to-date and the file is just uploaded to github two days ago. Maybe install from github is a better option for you because it is always up to date.<\/p>\n"}]}
{"QuestionId":34172622,"AnswerCount":1,"Tags":"<python><tensorflow><instance-variables>","CreationDate":"2015-12-09T07:04:37.463","AcceptedAnswerId":34173565.0,"OwnerUserId":678572.0,"Title":"Get original value of Tensor in Tensorflow","Body":"<p>I'm learning tensorflow right now and I have made a habit in programming of printing my variables to see what I'm working with as I'm coding. I can't figure out how to print the values of  and  objects in <\/p>\n\n<p><strong>How can I get back the values that I assigned to the instances?<\/strong> <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34173565","CreationDate":"2015-12-09T08:00:43.613","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The simplest way to see the values of your tensors is to create a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session\" rel=\"nofollow\"><code>tf.Session<\/code><\/a> and use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow\"><code>Session.run<\/code><\/a> to evaluate the tensors. Thus, your code would look like:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\n\nC_int = tf.constant(134)\nV_mat = tf.Variable(tf.zeros([2,3]))\nC_str = tf.constant(\"ABC\")\n\nsess = tf.Session()\nsess.run(C_int)\n#134\nsess.run(V_mat)\n#array([[ 0.,  0.,  0.],\n#       [ 0.,  0.,  0.]])\nsess.run(C_str)\n#ABC\n\nsess.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":34175174,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><tensorflow>","CreationDate":"2015-12-09T09:31:30.003","AcceptedAnswerId":34175683.0,"OwnerUserId":853462.0,"Title":"Extract features using pre-trained (Tensorflow) CNN","Body":"<p>Deep Learning has been applied successfully on several large data sets for the classification of a handful of classes (cats, dogs, cars, planes, etc), with performances beating simpler descriptors like Bags of Features over SIFT, color histograms, etc.<\/p>\n\n<p>Nevertheless, training such a network requires a lot of data per class and a lot of training time. However, very often one doesn't have enough data or just wants to get an idea of how well a convolutional neural network might do, before spending time one designing and training such a device and gathering the training data.<\/p>\n\n<p>In this particular case, it might be ideal to have a network configured and trained using some benchmark data set used by the state of the art publications, and to simply apply it to some data set that you might have as a feature extractor.<\/p>\n\n<p>This results in a set of features for each image, which one could feed to a classical classification method like SVM's, logistic regression, neural networks, etc.<\/p>\n\n<p>In particular when one does not have enough data to train the CNN, I may expect this to outperform a pipeline where the CNN was trained on few samples.<\/p>\n\n<p>I was looking at the tensorflow tutorials, but they always seem to have a clear training \/ testing phase. I couldn't find a pickle file (or similar) with a pre-configured CNN feature extractor.<\/p>\n\n<p>My questions are: do such pre-trained networks exist and where can I find them. Alternatively: does this approach make sense? Where could I find a CNN+weights ?<\/p>\n\n<p><strong>EDIT<\/strong>\nW.r.t. @john's comment I tried using  and  and checked the outputs, which are different (:S)<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>So the max absolute difference is , and in general all the elements differ (especially since the average value of the  and  themselves lies between 0 - 0.5).<\/p>\n\n<p>I also would expect that  needs a jpeg-string, not a numpy array, why else does the name contain 'Jpeg'. @john: Could you state how\n sure you are about your comment?<\/p>\n\n<p>So I guess I'm not sure what is what, as I would expect a trained neural network to be deterministic (but chaotic at most).<\/p>\n","answers":[{"AnswerId":"34175683","CreationDate":"2015-12-09T09:53:35.033","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow team recently released a deep CNN trained on the ImageNet dataset. You can download the script that fetches the data (including the model graph and the trained weights) from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/2fa4bb61689fe0cf98a12c2cd2a531bd8f5428b0\/tutorials\/image\/imagenet\/classify_image.py\" rel=\"nofollow noreferrer\">here<\/a>. The associated <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow noreferrer\">Image Recognition tutorial<\/a> has more details about the model.<\/p>\n\n<p>While the current model isn't specifically packaged to be used in a subsequent training step, you could explore modifying the script to reuse parts of the model and the trained weights in your own network.<\/p>\n"}]}
{"QuestionId":34177624,"AnswerCount":1,"Tags":"<numpy><parallel-processing><gpgpu><theano>","CreationDate":"2015-12-09T11:24:33.933","AcceptedAnswerId":34178011.0,"OwnerUserId":3041068.0,"Title":"Scan over every other element of a tensor variable in Theano","Body":"<p>it should produce something like:<\/p>\n\n<pre><\/pre>\n\n<p>I have briefly read the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loop.html\" rel=\"nofollow\">theano.scan tutorial<\/a> but I didn't find what I'm looking for.\nThank you.:)<\/p>\n","answers":[{"AnswerId":"34178011","CreationDate":"2015-12-09T11:43:47.023","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You don't need to use <code>theano.scan<\/code>. Just use ordinary index\/slice notation as in numpy:<\/p>\n\n<p>In numpy, if<\/p>\n\n<pre><code>input = [1,2,3,4,5,6]\n<\/code><\/pre>\n\n<p>then<\/p>\n\n<pre><code>print input[::2]\n<\/code><\/pre>\n\n<p>will display<\/p>\n\n<pre><code>[1, 3, 5]\n<\/code><\/pre>\n\n<p>In Theano, this can be achieved by doing the same thing:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\ninput = tt.vector()\nf = theano.function([input], input[::2])\nprint f([1,2,3,4,5,6])\n<\/code><\/pre>\n"}]}
{"QuestionId":34182530,"AnswerCount":1,"Tags":"<python><docker><tensorflow>","CreationDate":"2015-12-09T15:24:33.840","AcceptedAnswerId":null,"OwnerUserId":3465763.0,"Title":"How do I run a python script with Tensorflow running in a Docker on Windows?","Body":"<p>Imagine I manage to install Tensorflow on Windows, using Docker as in these two links, for example:<\/p>\n\n<p><a href=\"https:\/\/caffinc.github.io\/2015\/11\/tensorflow-windows\/\">TensorFlow on Windows<\/a><\/p>\n\n<p><a href=\"http:\/\/www.netinstructions.com\/how-to-install-and-run-tensorflow-on-a-windows-pc\/\">How to install and run TensorFlow on a Windows PC<\/a><\/p>\n\n<p>In both links, they're able to use Tensorflow on a shell python (don't know exactly what version, I have Anaconda installed). <\/p>\n\n<p>But what if I want to run a script that I made on my local machine that has Tensorflow in it? How do I call the script from docker? I mean, how do I find the script (located on my desktop, for instance) from docker to run it?<\/p>\n","answers":[{"AnswerId":"34182903","CreationDate":"2015-12-09T15:41:25.077","ParentId":null,"OwnerUserId":"6309","Title":null,"Body":"<p>If you want your container (that has Tensorflow already preinstalled, since it is running from the Tensorflow  image) to access your script, you need to mount that script from your host onto a local path in your container.<\/p>\n\n<pre><code>docker run -v \/path\/to\/your\/script:\/path\/to\/script\n<\/code><\/pre>\n\n<p>See \"<a href=\"https:\/\/docs.docker.com\/engine\/userguide\/dockervolumes\/#mount-a-host-file-as-a-data-volume\" rel=\"nofollow noreferrer\">Mount a host file as a data volume<\/a>\".<\/p>\n\n<blockquote>\n  <p>The <code>-v<\/code> flag can also be used to mount a single file - instead of just directories - from the host machine.<\/p>\n<\/blockquote>\n\n<pre><code>$ docker run --rm -it -v ~\/.bash_history:\/.bash_history ubuntu \/bin\/bash\n<\/code><\/pre>\n\n<p>Then, from your container, you will access the same script in <code>\/path\/to\/script<\/code>.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/users\/1061542\/alex-pryiomka\">Alex Pryiomka<\/a> gives an example of running such a script in tensorflow with \"<a href=\"https:\/\/stackoverflow.com\/a\/44015407\/6309\">How to run Python Scripts on Mac Terminal using Docker with Tensorflow?<\/a>\"<\/p>\n"}]}
{"QuestionId":34182549,"AnswerCount":1,"Tags":"<python><csv><numpy><tensorflow>","CreationDate":"2015-12-09T15:25:05.637","AcceptedAnswerId":34185958.0,"OwnerUserId":3849791.0,"Title":"Why is TensorFlow predicting all 0's or all 1's after training?","Body":"<p>So my problem is that I am running through the beginner level code in the TensorFlow tutorial and have modified it for my needs but when I make it  it used to always print out a 1.0, now it's always guessing 0's and printing out an ~93% accuracy. When I use , it guesses all 1's and produces an ~7% accuracy rate. Add up the two and it equals 100%. I don't get how  guesses 1's and  guesses 0's. Obviously something is wrong with the code. Please have a look and let me know what I can do to fix this issue. I think that the code is going wrong during the training, but I could be wrong.<\/p>\n\n<pre><\/pre>\n\n<p>At this point, if I make it  and , this is what the results look like.<\/p>\n\n<pre><\/pre>\n\n<p>Ok, let's carry on then.<\/p>\n\n<pre><\/pre>\n\n<p>This percentage does not shift. It's guessing all zeros regardless of the onehot that I created for the 2 classes (1 or 0).<\/p>\n\n<p>Here's a look at the data-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34185958","CreationDate":"2015-12-09T18:09:13.937","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p><strong>tl;dr:  The way the sample code above posted computes the cross-entropy is not numerically robust.  Use <code>tf.nn.cross_entropy_with_logits<\/code> instead.<\/strong><\/p>\n\n<p>(in response to v1 of the question, which has changed):  I'm worried that your training is not actually getting run to completion or working, based upon the <code>nan<\/code>s in your x_train data that you showed.  I'd suggest first fixing that - and identifying why they showed up and fixing that bug, and seeing if you also have <code>nan<\/code>s in your test set.  Might be helpful to show x_test and y_test also.<\/p>\n\n<p>Finally, I believe there's a bug in the way <code>y_<\/code> is handled in relation to x.  The code is written as if <code>y_<\/code> is a one-hot matrix, but when you show <code>y_train[:10]<\/code>, it only has 10 elements, not <code>10*num_classes<\/code> categories.  I suspect a bug there.  When you argmax it on axis 1, you're always going to get a vector full of zeros (because there's only one element on that axis, so of course it's the maximum element).  Combine that with a bug producing always-zero output on the estimate, and you're always producing a \"correct\" answer. :)<\/p>\n\n<p><strong>Update for revised version<\/strong>\nIn the changed version, if you run it and print out W at the end of every execution by changing your code to look like this:<\/p>\n\n<pre><code> _, w_out, b_out = sess.run([train_step, W, b], feed_dict={x: [x_train[g]], y_: [y_train[g]]})\n<\/code><\/pre>\n\n<p>you'll observe that W is full of <code>nan<\/code>s.  To debug this, you can either stare a lot at your code to see if there's a mathematical problem you can spot, or you can instrument back through the pipeline to see where they show up.  Let's try that.  First, what's the <code>cross_entropy<\/code>?  (add <code>cross_entropy<\/code> to the list of things in the <code>run<\/code> statement and print it out)<\/p>\n\n<pre><code>Cross entropy:  inf\n<\/code><\/pre>\n\n<p>Great!  So.. why?  Well, one answer is that when:<\/p>\n\n<pre><code>y = [0, 1]\ntf.log(y) = [-inf, 0]\n<\/code><\/pre>\n\n<p>This is a valid possible output for y, but one that your computation of the cross-entropy is not robust to.  You could either manually add some epsilons to avoid the corner cases, or use <code>tf.nn.softmax_cross_entropy_with_logits<\/code> to do it for you.  I recommend the latter:<\/p>\n\n<pre><code>yprime = tf.matmul(x,W)+b\ny = tf.nn.softmax(yprime)\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(yprime, y_)\n<\/code><\/pre>\n\n<p>I don't guarantee that your model will work, but this should fix your current NaN problem.<\/p>\n"}]}
{"QuestionId":34183343,"AnswerCount":6,"Tags":"<numpy><scipy><tensorflow>","CreationDate":"2015-12-09T16:00:07.887","AcceptedAnswerId":34184489.0,"OwnerUserId":2584721.0,"Title":"How does tensorflow batch_matmul work?","Body":"<p>Tensorflow has a function called <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#batch_matmul\" rel=\"noreferrer\">batch_matmul<\/a> which multiplies higher dimensional tensors. But I'm having a hard time understanding how it works, perhaps partially because I'm having a hard time visualizing it. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xoiGQ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xoiGQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What I want to do is multiply a matrix by each slice of a 3D tensor, but I don't quite understand what the shape of tensor a is. Is z the innermost dimension? Which of the following is correct? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sm1bZ.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sm1bZ.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I would most prefer the first to be correct -- it's most intuitive to me and easy to see in the .eval() output. But I suspect the second is correct. <\/p>\n\n<p>Tensorflow says that batch_matmul performs:<\/p>\n\n<pre><\/pre>\n\n<p>What does that mean? What does that mean in the context of my example? What is being multiplied with with what? And why aren't I getting a 3D tensor the way I expected? <\/p>\n","answers":[{"AnswerId":"52609721","CreationDate":"2018-10-02T13:48:07.060","ParentId":null,"OwnerUserId":"1109026","Title":null,"Body":"<p>tf.tensordot should solve this problem. It supports batch operations, e.g., if you want to contract a 2D tensor with a 3D tensor, with the latter having a batch dimension. <\/p>\n\n<p>If a is shape [n,m] b is shape [?,m,l], then <\/p>\n\n<p>y = tf.tensordot(b, a, axes=[1, 1]) will produce a tensor of shape [?,n,l]<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tensordot\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tensordot<\/a><\/p>\n"},{"AnswerId":"43819275","CreationDate":"2017-05-06T10:22:52.343","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>First of all <code>tf.batch_matmul()<\/code> was <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/3a88ec0b8e5a9b8f3f3efc4d9a8f88d47b3f4042\" rel=\"noreferrer\">removed<\/a> and no longer available. Now you suppose to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"noreferrer\"><code>tf.matmul()<\/code><\/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>So let's assume you have the following code:<\/p>\n\n<pre><code>import tensorflow as tf\nbatch_size, n, m, k = 10, 3, 5, 2\nA = tf.Variable(tf.random_normal(shape=(batch_size, n, m)))\nB = tf.Variable(tf.random_normal(shape=(batch_size, m, k)))\ntf.matmul(A, B)\n<\/code><\/pre>\n\n<p>Now you will receive a tensor of the shape <code>(batch_size, n, k)<\/code>. Here is what is going on here. Assume you have <code>batch_size<\/code> of matrices <code>nxm<\/code> and <code>batch_size<\/code> of matrices <code>mxk<\/code>. Now for each pair of them you calculate <code>nxm X mxk<\/code> which gives you an <code>nxk<\/code> matrix. You will have <code>batch_size<\/code> of them.<\/p>\n\n<p>Notice that something like this is also valid:<\/p>\n\n<pre><code>A = tf.Variable(tf.random_normal(shape=(a, b, n, m)))\nB = tf.Variable(tf.random_normal(shape=(a, b, m, k)))\ntf.matmul(A, B)\n<\/code><\/pre>\n\n<p>and will give you a shape <code>(a, b, n, k)<\/code><\/p>\n"},{"AnswerId":"34184489","CreationDate":"2015-12-09T16:53:52.547","ParentId":null,"OwnerUserId":"5612876","Title":null,"Body":"<p>You can imagine it as doing a matmul over each training example in the batch.<\/p>\n\n<p>For example, if you have two tensors with the following dimensions:<\/p>\n\n<pre><code>a.shape = [100, 2, 5]\nb.shape = [100, 5, 2]\n<\/code><\/pre>\n\n<p>and you do a batch <code>tf.matmul(a, b)<\/code>, your output will have the shape <code>[100, 2, 2]<\/code>.<\/p>\n\n<p>100 is your batch size, the other two dimensions are the dimensions of  your data.<\/p>\n"},{"AnswerId":"41117425","CreationDate":"2016-12-13T09:22:03.293","ParentId":null,"OwnerUserId":"6384734","Title":null,"Body":"<p>The answer to this particular answer is using tf.scan function.<\/p>\n\n<p>If  a = [5,3,2]  #dimension of 5 batch, with 3X2 mat in each batch<br>\nand b = [2,3]    #  a constant matrix to be multiplied with each sample<\/p>\n\n<p>then let def fn(a,x):\n              return tf.matmul(x,b)<\/p>\n\n<p>initializer = tf.Variable(tf.random_number(3,3))<\/p>\n\n<p>h = tf.scan(fn,outputs,initializer)<\/p>\n\n<p>this h will store all the outputs.<\/p>\n"},{"AnswerId":"39823596","CreationDate":"2016-10-03T01:54:33.457","ParentId":null,"OwnerUserId":"4862162","Title":null,"Body":"<p>You can now do it using tf.einsum, starting from Tensorflow <strong>0.11.0rc0<\/strong>.<\/p>\n\n<p>For example, <\/p>\n\n<pre><code>M1 = tf.Variable(tf.random_normal([2,3,4]))\nM2 = tf.Variable(tf.random_normal([5,4]))  \nN = tf.einsum('ijk,lk-&gt;ijl',M1,M2)       \n<\/code><\/pre>\n\n<p>It multiplies the matrix M2 with every frame (3 frames) in every batch (2 batches) in M1.<\/p>\n\n<p>The output is:<\/p>\n\n<pre><code>[array([[[ 0.80474716, -1.38590837, -0.3379252 , -1.24965811],\n        [ 2.57852983,  0.05492432,  0.23039417, -0.74263287],\n        [-2.42627382,  1.70774114,  1.19503212,  0.43006262]],\n\n       [[-1.04652011, -0.32753903, -1.26430523,  0.8810069 ],\n        [-0.48935518,  0.12831448, -1.30816901, -0.01271309],\n        [ 2.33260512, -1.22395933, -0.92082584,  0.48991606]]], dtype=float32),\narray([[ 1.71076882, 0.79229093, -0.58058828, -0.23246667],\n       [ 0.20446332,  1.30742455, -0.07969904,  0.9247328 ],\n       [-0.32047141,  0.66072595, -1.12330854,  0.80426538],\n       [-0.02781649, -0.29672042,  2.17819595, -0.73862702],\n       [-0.99663496,  1.3840003 , -1.39621222,  0.77119476]], dtype=float32), \narray([[[ 0.76539308, 2.77609682, -1.79906654,  0.57580602, -3.21205115],\n        [ 4.49365759, -0.10607499, -1.64613271,  0.96234947, -3.38823152],\n        [-3.59156275,  2.03910899,  0.90939498,  1.84612727,  3.44476724]],\n\n       [[-1.52062428,  0.27325237,  2.24773455, -3.27834225,  3.03435063],\n        [ 0.02695178,  0.16020992,  1.70085776, -2.8645196 ,  2.48197317],\n        [ 3.44154787, -0.59687197, -0.12784094, -2.06931567, -2.35522676]]], dtype=float32)]\n<\/code><\/pre>\n\n<p>I have verified, the arithmetic is correct.<\/p>\n"},{"AnswerId":"37501489","CreationDate":"2016-05-28T16:42:54.033","ParentId":null,"OwnerUserId":"3050716","Title":null,"Body":"<p>It is simply like splitting on the first dimension respectively, multiply and concat them back. If you want to do 3D by 2D, you can reshape, multiply, and reshape it back. I.e. [100, 2, 5] -> [200, 5] -> [200, 2] -> [100, 2, 2]<\/p>\n"}]}
{"QuestionId":34183997,"AnswerCount":1,"Tags":"<gcc><tensorflow><bazel>","CreationDate":"2015-12-09T16:30:47.947","AcceptedAnswerId":null,"OwnerUserId":5659969.0,"Title":"Why am I getting Docker error \"C++ compilation of rule '\/\/tensorflow\/core:kernels' failed\"?","Body":"<p>I get build errors when I do the following. I need to build Tensorflow because I'm trying to add my own ops. I get the same error also without using \"-c opt\". What am I doing wrong?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34457997","CreationDate":"2015-12-24T21:20:33.453","ParentId":null,"OwnerUserId":"2138200","Title":null,"Body":"<p>I was having a similar issue with exit code 4 (which I believe is out of memory).<\/p>\n\n<p>The solution for me was to increase docker-machine to use 8GB. Something like:<\/p>\n\n<p><code>docker-machine create --driver virtualbox --virtualbox-memory 8192 --virtualbox-cpu-count 4 default<\/code><\/p>\n"}]}
{"QuestionId":34187681,"AnswerCount":0,"Tags":"<histogram><tensorflow>","CreationDate":"2015-12-09T19:46:33.767","AcceptedAnswerId":null,"OwnerUserId":5556479.0,"Title":"How To Understand Tensor Flow Weights and Gradient Histogram","Body":"<p>I am experimenting with conv nets for sentiment analysis.<br>\nI used a basic convent architecture.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ARzr1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ARzr1.png\" alt=\"Network\"><\/a><\/p>\n\n<p>I am trying to analyze the learning, I have a bad feeling about weights, Although I am early in the learning process, but the weights for fully connected layer  essentially suggest no updates\/ update on only very few component , because histogram essentially remains the same.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ENePa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ENePa.png\" alt=\"Weight and Gradient Histogram\"><\/a><\/p>\n\n<p>Any hints on what I could be missing?<\/p>\n\n<p>Regards,\nVaibhav Singh<\/p>\n","answers":[]}
{"QuestionId":34189397,"AnswerCount":1,"Tags":"<python><neural-network><theano><backpropagation><recurrent-neural-network>","CreationDate":"2015-12-09T21:32:43.813","AcceptedAnswerId":null,"OwnerUserId":5435604.0,"Title":"Theano , recurrent neural network, error is nan","Body":"<p>I am trying to duplicate the recent work on unitary evolution neural networks. Adapting from the code published by the author, I have written the following code<\/p>\n\n<pre><\/pre>\n\n<p>But after an initial error of 6.23 , I am getting all nans. Could someone explain if there is some error in the code. I posted the whole code so that I do not miss out on the erroneous section(and I dont know which one it is)<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"37016325","CreationDate":"2016-05-04T00:03:23.350","ParentId":null,"OwnerUserId":"1870254","Title":null,"Body":"<p>While Marcin Mo\u017cejko was probably right about the source of the NaNs beieng numerical problems with the logs, theano's doc gives good general advice on <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/nan_tutorial.html\" rel=\"nofollow\">dealing with NaNs<\/a>. <\/p>\n\n<p>In particular, how to use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/nanguardmode.html#nanguardmode\" rel=\"nofollow\">NaN guard<\/a> - which will throw an error as soon as a tensor variable contains NaNs:<\/p>\n\n<pre><code>from theano.compile.nanguardmode import NanGuardMode\n\n...\n... = theano.function(..., mode=NanGuardMode(nan_is_error=True,\n                                             inf_is_error=True,\n                                             big_is_error=True)\n                     )\n<\/code><\/pre>\n"}]}
{"QuestionId":34189579,"AnswerCount":1,"Tags":"<python><version><protocol-buffers><caffe><tensorflow>","CreationDate":"2015-12-09T21:44:35.690","AcceptedAnswerId":null,"OwnerUserId":2035694.0,"Title":"Caffe and TensorFlow Protobuf - Maintain different Version","Body":"<p>I have installed Caffe on my Ubuntu15.10x64 Machine with . Also, as of today, I installed TensorFlow with  but TensorFlow needs . From what I have read,  supports using different libraries for python projects. Can I safely  inside the (tensorflow) virtualenv without disrupting my Caffe install - thus having  as default, and  in my virtualenv used by TensorFlow?<\/p>\n\n<p>Or is the only solution to install with Docker or Create a Clone in a Virtual Machine? Better than nothing, but I would like to use my GPU and full power.<\/p>\n\n<p>Thanks for help and clarification.<\/p>\n\n<p>Edit:<\/p>\n\n<p>So, this is what happens when I try to update protobuf -><\/p>\n\n<pre><\/pre>\n\n<p>If i check the version this is what it returns -><\/p>\n\n<pre><\/pre>\n\n<p>Do I have to manually update protobuf? I presume that I would create a  folder inside  with the version I need?<\/p>\n","answers":[{"AnswerId":"34193234","CreationDate":"2015-12-10T03:34:25.027","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Note that your virtualenv doesn't see packages from outside, so you should not have protobuf visible in your virtualenv at all, even though you have it installed globally in the system (or in the Caffe's virtualenv, depending on your setup).<\/p>\n\n<p>It should be safe for you to run <code>pip install protobuf<\/code> inside the TensorFlow's virtualenv, it will not disrupt your global setup or any other virtualenv you have.<\/p>\n"}]}
{"QuestionId":34190298,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-12-09T22:33:59.450","AcceptedAnswerId":null,"OwnerUserId":1576602.0,"Title":"What's the best way to refresh TensorBoard after new events\/logs were added?","Body":"<p>What is the best way to quickly see the updated graph in the most recent event file in an open TensorBoard session? Re-running my Python app results in a new log file being created with potentially new events\/graph. However, TensorBoard does not seem to notice those differences, unless restarted.<\/p>\n","answers":[{"AnswerId":"37491695","CreationDate":"2016-05-27T20:23:40.590","ParentId":null,"OwnerUserId":"1576602","Title":null,"Body":"<p>It turns out that TensorBoard backend refreshes the logs every minute. This has been reported as a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2050\" rel=\"noreferrer\">TensorFlow issue<\/a>. <\/p>\n\n<p>The reload interval can be configured using the <code>--reload_interval<\/code> flag of the TensorBoard process, but this option is currently only available in master and as of version 0.8 has not been released.<\/p>\n"},{"AnswerId":"44359908","CreationDate":"2017-06-04T23:45:16.027","ParentId":null,"OwnerUserId":"6452438","Title":null,"Body":"<p>My issue is different. Each time I refresh <code>0.0.0.0:6006<\/code>, it seems the new graph keep appending to the old one, which is quite annoying.<\/p>\n\n<p>After trying kill process and delete old log several times, I realized the issue comes from <code>writer.add_graph(sess.graph)<\/code>,  because I didn't <strong>reset the graph<\/strong> in jupyter notebook.  After resetting, the tensorboard could show the newest gragh.<\/p>\n"}]}
{"QuestionId":34191321,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-12-09T23:56:13.080","AcceptedAnswerId":null,"OwnerUserId":1576602.0,"Title":"How to specify tags for scalar summary of unknown shape in TensorFlow?","Body":"<p>Let's imagine the following situation:<\/p>\n\n<pre><\/pre>\n\n<p>In this situation, the actual shape of the  and  placeholders will only be set when running the session and feeding concrete data to the placeholders. <\/p>\n\n<p>However, according to API the shape of tags needs to match the shape of . How can that be achieved?<\/p>\n","answers":[]}
{"QuestionId":34192229,"AnswerCount":1,"Tags":"<python><tensorflow><linear-algebra><matrix-multiplication><tensor>","CreationDate":"2015-12-10T01:32:48.170","AcceptedAnswerId":34192524.0,"OwnerUserId":1576602.0,"Title":"Efficient element-wise multiplication of a matrix and a vector in TensorFlow","Body":"<p>What would be the most efficient way to multiply (element-wise) a 2D tensor (matrix):<\/p>\n\n<pre><\/pre>\n\n<p>by a vertical vector:<\/p>\n\n<pre><\/pre>\n\n<p>to obtain a new matrix:<\/p>\n\n<pre><\/pre>\n\n<p>To give some context, we have  data samples in a batch that can be processed in parallel, and each -element sample must be multiplied by weights  stored in a variable to eventually pick the largest  for each row .<\/p>\n","answers":[{"AnswerId":"34192524","CreationDate":"2015-12-10T02:08:24.427","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The simplest code to do this relies on the broadcasting behavior of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/multiply\" rel=\"noreferrer\"><code>tf.multiply()<\/code><\/a><sup>*<\/sup>, which is based on <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"noreferrer\">numpy's broadcasting behavior<\/a>:<\/p>\n\n<pre><code>x = tf.constant(5.0, shape=[5, 6])\nw = tf.constant([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])\nxw = tf.multiply(x, w)\nmax_in_rows = tf.reduce_max(xw, 1)\n\nsess = tf.Session()\nprint sess.run(xw)\n# ==&gt; [[0.0, 5.0, 10.0, 15.0, 20.0, 25.0],\n#      [0.0, 5.0, 10.0, 15.0, 20.0, 25.0],\n#      [0.0, 5.0, 10.0, 15.0, 20.0, 25.0],\n#      [0.0, 5.0, 10.0, 15.0, 20.0, 25.0],\n#      [0.0, 5.0, 10.0, 15.0, 20.0, 25.0]]\n\nprint sess.run(max_in_rows)\n# ==&gt; [25.0, 25.0, 25.0, 25.0, 25.0]\n<\/code><\/pre>\n\n<p><sup>*<\/sup> In older versions of TensorFlow, <code>tf.multiply()<\/code> was called <code>tf.mul()<\/code>. You can also use the <code>*<\/code> operator (i.e. <code>xw = x * w<\/code>) to perform the same operation.<\/p>\n"}]}
{"QuestionId":34192568,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-10T02:12:38.863","AcceptedAnswerId":34192610.0,"OwnerUserId":1884420.0,"Title":"Torch - How to change tensor type?","Body":"<p>I created a permutation of the numbers from 1 to 3.<\/p>\n\n<pre><\/pre>\n\n<p>Now, I want to convert  to a . How can I do that?<\/p>\n","answers":[{"AnswerId":"34192610","CreationDate":"2015-12-10T02:16:05.987","ParentId":null,"OwnerUserId":"1884420","Title":null,"Body":"<p><code>y = y:long()<\/code> does the job. There are similar methods for other data types such as <code>int<\/code>, <code>char<\/code>, <code>float<\/code> and <code>byte<\/code><\/p>\n"}]}
{"QuestionId":34192780,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-10T02:37:15.340","AcceptedAnswerId":34193440.0,"OwnerUserId":5118777.0,"Title":"how to tackle HEX in torch7?","Body":"<p>I'm using <a href=\"https:\/\/github.com\/clementfarabet\/async\/blob\/master\/tests\/test-tcp-client.lua\" rel=\"nofollow\">async-tcp client<\/a> to connect to server and receive data(an array).<\/p>\n\n<pre><\/pre>\n\n<p>If data type is HEX, I can get data but it is all gibberish.<\/p>\n\n<p>It seems that there is something wrong with encoding.<\/p>\n\n<hr>\n\n<p>If data type is note HEX, I can also get data but it is string.<\/p>\n\n<p>I have no idea to convert the 'array string' to tensor.<\/p>\n\n<p>'0.001 0.002 0.003' ->   ??<\/p>\n\n<p>What should I do ?<\/p>\n\n<p>Thank you<\/p>\n\n<h1>==================================================<\/h1>\n\n<p>EDIT<\/p>\n\n<p><strong>string.byte<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34193440","CreationDate":"2015-12-10T03:58:10.297","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>If you know the format ahead of time, you can use the <code>match<\/code> function to get the list of values from a string, which you can then convert to the table and the Tensor:<\/p>\n\n<pre><code>local str = \"0.001 0.002 0.003\"\ntorch.Tensor({{str:match(\"(%d+%.%d*)%s+(%d+%.%d*)%s+(%d+%.%d*)\")}})\n<\/code><\/pre>\n\n<p>This returns:<\/p>\n\n<pre><code>0.001 *\n  1.0000  2.0000  3.0000\n[torch.DoubleTensor of size 1x3]\n<\/code><\/pre>\n\n<p>If the number is in the hex format, you can use <code>tonumber<\/code> function to convert, for example, <code>tonumber(\"0x12\") == 18<\/code>.<\/p>\n"}]}
{"QuestionId":34193649,"AnswerCount":2,"Tags":"<python><neural-network><theano><deep-learning><keras>","CreationDate":"2015-12-10T04:22:52.087","AcceptedAnswerId":42158235.0,"OwnerUserId":4068755.0,"Title":"Multiplying the output of two layers in keras","Body":"<p>I'm trying to implement in keras a neural(ish) net with this design:  <a href=\"http:\/\/nlp.cs.rpi.edu\/paper\/AAAI15.pdf\" rel=\"noreferrer\">http:\/\/nlp.cs.rpi.edu\/paper\/AAAI15.pdf<\/a><\/p>\n\n<p>The algorithm has, essentially, three inputs.  Input 2 and Input 3 get multiplied by the same weight matrix W1 to produce O2 and O3.  Input 1 gets multiplied by W2 to produce O1.  Then, we need to take the dot product of O1 * O2,  and O1 * O3.  <\/p>\n\n<p>I'm trying to implement this in keras.  <\/p>\n\n<p>My first thought was to use the keras  class and make W1 a shared node layer with two inputs and two outputs.  Fine so far. <\/p>\n\n<p>The problem then arises of how to take the dot products of those two outputs with O1.  <\/p>\n\n<p>I tried to define a custom function:<\/p>\n\n<pre><\/pre>\n\n<p>Then:<\/p>\n\n<pre><\/pre>\n\n<p>The problem that arises on compile, is that keras only wants to give the Lambda layer one input:<\/p>\n\n<pre><\/pre>\n\n<p>I thought an alternative might be to use a  class, which has  as a type of permitted merge.  But, the input layers for a  class have to be passed to the constructor. So, there doesn't seem to be a way of getting the outputs from the shared node into the  to add the  to the . <\/p>\n\n<p>If I was using  containers, I could feed those into the .  But, then there wouldn't be a way to implement that the two  layers need to share the same weight matrix. <\/p>\n\n<p>I thought about trying to concatenate O1, O2, and O3 together into a single vector as an output layer and then do the multiplication inside an objective function.  But, that would require the objective function to split its input, which doesn't seem to be possible in keras (the relevant Theano functions aren't passed-through to to keras API).  <\/p>\n\n<p>Any know a solution?<\/p>\n\n<p>EDIT: <\/p>\n\n<p>I thought I'd made some progress because I found that  is implementing  (even if its not in the documentation).  <\/p>\n\n<p>So I got to:<\/p>\n\n<pre><\/pre>\n\n<p>And now the error is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42158235","CreationDate":"2017-02-10T11:25:23.270","ParentId":null,"OwnerUserId":"2756668","Title":null,"Body":"<p>You can use this<\/p>\n\n<p>main_branch.add(Merge([branch_1, branch_2], mode = 'dot'))<\/p>\n"},{"AnswerId":"37085396","CreationDate":"2016-05-07T06:42:44.660","ParentId":null,"OwnerUserId":"6303101","Title":null,"Body":"<p>I'm facing the similar problem. I think out a solution, but did not try it yet.<\/p>\n\n<ol>\n<li><p>Using Convolution Layers for Sequential model A which takes both Input2 and Input3 as the input. In this way, same convolution kernel will be applied to Input2 an Input3, aka the same weight W1.<\/p><\/li>\n<li><p>Taking Input1 as the input of another Sequential model B.<\/p><\/li>\n<li><p>Using a Merge Layer to merge the output out of A and B.\nAnd dot can be also done via custom function of a merge layer.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34194151,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-10T05:12:12.380","AcceptedAnswerId":34199902.0,"OwnerUserId":1576602.0,"Title":"Best way to flatten a 2D tensor containing a vector in TensorFlow?","Body":"<p>What is the most efficient way to flatten a 2D tensor which is actually a horizontal or vertical vector into a 1D tensor? <\/p>\n\n<p>Is there a difference in terms of performance between:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>?<\/p>\n","answers":[{"AnswerId":"34199902","CreationDate":"2015-12-10T10:47:47.937","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Both <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"noreferrer\"><code>tf.reshape(w, [-1])<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/squeeze\" rel=\"noreferrer\"><code>tf.squeeze(w)<\/code><\/a> are \"cheap\" in that they operate only on the metadata (i.e. the shape) of the given tensor, and don't modify the data itself. Of the two <code>tf.reshape()<\/code> has slightly <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/8242b4dd1b36440e191fef8a07b6f37d8bcee60d\/tensorflow\/core\/kernels\/reshape_op.h#L34\" rel=\"noreferrer\">simpler logic internally<\/a>, but the performance of the two should be indistinguishable.<\/p>\n"}]}
{"QuestionId":34194382,"AnswerCount":3,"Tags":"<sentiment-analysis><tensorflow>","CreationDate":"2015-12-10T05:31:27.997","AcceptedAnswerId":null,"OwnerUserId":3899000.0,"Title":"Sentiment Analysis using tensorflow","Body":"<p>I am exploring tensorflow and would like to do sentiment analysis using the options available. I had a look at the following tutorial <a href=\"http:\/\/www.tensorflow.org\/tutorials\/recurrent\/index.html#language_modeling\">http:\/\/www.tensorflow.org\/tutorials\/recurrent\/index.html#language_modeling<\/a> <\/p>\n\n<p>I have worked woth Naive Bayes Classifier, Maximum Entropy Algorithm and Scikit Learn Classifier and would like to know if there are any better algorithms offered by tensorflow. Is this the right place to start or are there any other options?<\/p>\n\n<p>Any help pointing in the right direction would be greatly appreciated.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"39154333","CreationDate":"2016-08-25T20:36:13.720","ParentId":null,"OwnerUserId":"1751387","Title":null,"Body":"<p>A commonly used approach would be using a Convolutional Neural Network (CNN) to do sentiment analysis. You can find a great explanation\/tutorial in this <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"noreferrer\">WildML blogpost<\/a>. The accompanying TensorFlow code can be found <a href=\"https:\/\/github.com\/dennybritz\/cnn-text-classification-tf\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Another approach would be using an LSTM (or related network), you can find example implementations online, a good starting point is <a href=\"http:\/\/monik.in\/a-noobs-guide-to-implementing-rnn-lstm-using-tensorflow\/\" rel=\"noreferrer\">this blogpost<\/a>.<\/p>\n"},{"AnswerId":"44402375","CreationDate":"2017-06-07T02:09:24.350","ParentId":null,"OwnerUserId":"4883502","Title":null,"Body":"<p>I would suggest you try a character-level LSTM, it's been shown to be able to achieve state-of-the-art results in many text classification tasks one of them being sentiment analysis.<\/p>\n\n<p>I wrote a pretty lengthy article that you can find <a href=\"https:\/\/charlesashby.github.io\/2017\/06\/05\/sentiment-analysis-with-char-lstm\/\" rel=\"nofollow noreferrer\">here<\/a> where I go through it's implementation in TensorFlow line by line. The result is a model that is less than 100mb in size and that achieves an accuracy of over 80% on a test set of 80,000 tweets.<\/p>\n\n<p>Another approach that has proven to be very effective is to use a recursive neural network, you can read the paper from Stanford NLP Group <a href=\"https:\/\/nlp.stanford.edu\/~socherr\/EMNLP2013_RNTN.pdf\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"43477865","CreationDate":"2017-04-18T16:46:19.383","ParentId":null,"OwnerUserId":"3907250","Title":null,"Body":"<p>For me, the easiest tutorial to follow was: <a href=\"https:\/\/pythonprogramming.net\/data-size-example-tensorflow-deep-learning-tutorial\/?completed=\/train-test-tensorflow-deep-learning-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/pythonprogramming.net\/data-size-example-tensorflow-deep-learning-tutorial\/?completed=\/train-test-tensorflow-deep-learning-tutorial\/<\/a><\/p>\n\n<p>It walks you through<code>TensorFlow.train.AdamOptimizer().minimize(cost)<\/code> and uses Sentiment140 dataset (from Stanford, ~1 mil examples of positive and negative sentiment)<\/p>\n"}]}
{"QuestionId":34199233,"AnswerCount":12,"Tags":"<python><tensorflow><nvidia-titan>","CreationDate":"2015-12-10T10:19:51.040","AcceptedAnswerId":34200194.0,"OwnerUserId":1841986.0,"Title":"How to prevent tensorflow from allocating the totality of a GPU memory?","Body":"<p>I work in an environment in which computational resources are shared, i.e., we have a few server machines equipped with a few Nvidia Titan X GPUs each.<\/p>\n\n<p>For small to moderate size models, the 12GB of the Titan X are usually enough for 2-3 people to run training concurrently on the same GPU. If the models are small enough that a single model does not take full advantage of all the computational units of the Titan X, this can actually result in a speedup compared with running one training process after the other. Even in cases where the concurrent access to the GPU does slow down the individual training time, it is still nice to have the flexibility of having several users running things on the GPUs at once.<\/p>\n\n<p>The problem with TensorFlow is that, by default, it allocates the full amount of available memory on the GPU when it is launched. Even for a small 2-layer Neural Network, I see that the 12 GB of the Titan X are used up.<\/p>\n\n<p>Is there a way to make TensorFlow only allocate, say, 4GB of GPU memory, if one knows that that amount is enough for a given model?<\/p>\n","answers":[{"AnswerId":"56417247","CreationDate":"2019-06-02T17:15:29.323","ParentId":null,"OwnerUserId":"4780278","Title":null,"Body":"<p>You can use <\/p>\n\n<pre><code>TF_FORCE_GPU_ALLOW_GROWTH=true\n<\/code><\/pre>\n\n<p>in your environment variables.<\/p>\n\n<p>In <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/gpu\/gpu_bfc_allocator.cc\" rel=\"nofollow noreferrer\">tensorflow<\/a> code:<\/p>\n\n<pre><code>bool GPUBFCAllocator::GetAllowGrowthValue(const GPUOptions&amp; gpu_options) {\n  const char* force_allow_growth_string =\n      std::getenv(\"TF_FORCE_GPU_ALLOW_GROWTH\");\n  if (force_allow_growth_string == nullptr) {\n    return gpu_options.allow_growth();\n}\n<\/code><\/pre>\n"},{"AnswerId":"48677030","CreationDate":"2018-02-08T03:25:04.283","ParentId":null,"OwnerUserId":"4495833","Title":null,"Body":"<p>All the answers above assume execution with a <code>sess.run()<\/code> call, which is becoming the exception rather than the rule in recent versions of TensorFlow.<\/p>\n\n<p>When using the <code>tf.Estimator<\/code> framework (TensorFlow 1.4 and above) the way to pass the fraction along to the implicitly created <code>MonitoredTrainingSession<\/code> is,<\/p>\n\n<pre><code>opts = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)\nconf = tf.ConfigProto(gpu_options=opts)\ntrainingConfig = tf.estimator.RunConfig(session_config=conf, ...)\ntf.estimator.Estimator(model_fn=..., \n                       config=trainingConfig)\n<\/code><\/pre>\n\n<p>Similarly in Eager mode (TensorFlow 1.5 and above), <\/p>\n\n<pre><code>opts = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)\nconf = tf.ConfigProto(gpu_options=opts)\ntfe.enable_eager_execution(config=conf)\n<\/code><\/pre>\n\n<p><strong>Edit: 11-04-2018<\/strong>\nAs an example, if you are to use <code>tf.contrib.gan.train<\/code>, then you can use something similar to bellow:<\/p>\n\n<pre><code>tf.contrib.gan.gan_train(........, config=conf)\n<\/code><\/pre>\n"},{"AnswerId":"44128902","CreationDate":"2017-05-23T07:52:55.927","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>Shameless plug: If you install the GPU supported Tensorflow, the session will first allocate all GPU whether you set it to use only CPU or GPU. I may add my tip that even you set the graph to use CPU only you should set the same configuration(as answered above:) ) to prevent the unwanted GPU occupation. <\/p>\n\n<p>And in interactive interface like IPython you should also set that configure, otherwise it will allocate all memory and left almost none for others. This is sometimes hard to notice.  <\/p>\n"},{"AnswerId":"58252356","CreationDate":"2019-10-05T21:05:50.780","ParentId":null,"OwnerUserId":"2352507","Title":null,"Body":"<p>For <strong>Tensorflow 2.0<\/strong> this <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_memory_growth#for_example\" rel=\"nofollow noreferrer\">this solution<\/a> worked for me.  (TF-GPU 2.0, Windows 10, GeForce RTX 2070)<\/p>\n\n<pre><code>physical_devices = tf.config.experimental.list_physical_devices('GPU')\nassert len(physical_devices) &gt; 0, \"Not enough GPU hardware devices available\"\ntf.config.experimental.set_memory_growth(physical_devices[0], True)\n<\/code><\/pre>\n"},{"AnswerId":"37454574","CreationDate":"2016-05-26T07:43:45.363","ParentId":null,"OwnerUserId":"586275","Title":null,"Body":"<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth=True\nsess = tf.Session(config=config)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1578\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1578<\/a><\/p>\n"},{"AnswerId":"56632028","CreationDate":"2019-06-17T13:08:14.587","ParentId":null,"OwnerUserId":"8099519","Title":null,"Body":"<h1>Tensorflow 2.0 Beta and (probably) beyond<\/h1>\n\n<p>The API changed again. It can be now found in:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.config.experimental.set_memory_growth(\n    device,\n    enable\n)\n<\/code><\/pre>\n\n<p>Aliases:<\/p>\n\n<ul>\n<li>tf.compat.v1.config.experimental.set_memory_growth<\/li>\n<li>tf.compat.v2.config.experimental.set_memory_growth<\/li>\n<\/ul>\n\n<p><strong>References:<\/strong><\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_memory_growth\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_memory_growth<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/guide\/gpu#limiting_gpu_memory_growth\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/gpu#limiting_gpu_memory_growth<\/a><\/li>\n<\/ul>\n\n<p><strong>See also:<\/strong>\n<em>Tensorflow - Use a GPU<\/em>: <a href=\"https:\/\/www.tensorflow.org\/guide\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/gpu<\/a><\/p>\n\n<p><strong>for Tensorflow 2.0 Alpha see:<\/strong> <a href=\"https:\/\/stackoverflow.com\/a\/55541385\/8099519\">this answer<\/a><\/p>\n"},{"AnswerId":"48214084","CreationDate":"2018-01-11T18:57:16.453","ParentId":null,"OwnerUserId":"1767754","Title":null,"Body":"<p>Here is an excerpt from the Book <code>Deep Learning with TensorFlow<\/code> <\/p>\n\n<blockquote>\n  <p>In some cases it is desirable for the process to only allocate a subset of the available memory, or to only grow the memory usage as it is needed by the process. TensorFlow provides <strong>two configuration<\/strong> options on the session to control this. The first is the <code>allow_growth<\/code> option, which attempts to allocate only as much GPU memory based on runtime allocations, it starts out allocating very little memory, and as sessions get run and more GPU memory is needed, we extend the GPU memory region needed by the TensorFlow process.<\/p>\n<\/blockquote>\n\n<p><strong>1) Allow growth: (more flexible)<\/strong><\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nsession = tf.Session(config=config, ...)\n<\/code><\/pre>\n\n<p>The second method is <code>per_process_gpu_memory_fraction<\/code> option, which determines the fraction of the overall amount of memory that <code>each<\/code> visible GPU should be allocated. <strong>Note:<\/strong> No release of memory needed, it can even worsen memory fragmentation when done.<\/p>\n\n<p><strong>2) Allocate fixed memory<\/strong>:<\/p>\n\n<p>To only allocate <code>40%<\/code> of the total memory of each GPU by:<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.per_process_gpu_memory_fraction = 0.4\nsession = tf.Session(config=config, ...)\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> \nThat's only useful though if you truly want to bind the amount of GPU memory available on the TensorFlow process.<\/p>\n"},{"AnswerId":"55541385","CreationDate":"2019-04-05T18:26:37.670","ParentId":null,"OwnerUserId":"3744998","Title":null,"Body":"<h1>Updated for TensorFlow 2.0 Alpha and beyond<\/h1>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/using_gpu#limiting_gpu_memory_growth\" rel=\"noreferrer\">From the 2.0 Alpha docs, the answer is now just one line before you do anything with TensorFlow:<\/a><\/p>\n\n<pre><code>import tensorflow as tf\ntf.config.gpu.set_per_process_memory_growth(True)\n<\/code><\/pre>\n"},{"AnswerId":"34200194","CreationDate":"2015-12-10T11:00:19.170","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can set the fraction of GPU memory to be allocated when you construct a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session\" rel=\"noreferrer\"><code>tf.Session<\/code><\/a> by passing a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/08ed32dbb9e8f67eec9efce3807b5bdb3933eb2f\/tensorflow\/core\/protobuf\/config.proto\" rel=\"noreferrer\"><code>tf.GPUOptions<\/code><\/a> as part of the optional <code>config<\/code> argument:<\/p>\n\n<pre><code># Assume that you have 12GB of GPU memory and want to allocate ~4GB:\ngpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)\n\nsess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))\n<\/code><\/pre>\n\n<p>The <code>per_process_gpu_memory_fraction<\/code> acts as a hard upper bound on the amount of GPU memory that will be used by the process on each GPU on the same machine. Currently, this fraction is applied uniformly to all of the GPUs on the same machine; there is no way to set this on a per-GPU basis.<\/p>\n"},{"AnswerId":"54295016","CreationDate":"2019-01-21T17:26:59.050","ParentId":null,"OwnerUserId":"8308619","Title":null,"Body":"<p>Well I am new to tensorflow, I have Geforce 740m or something GPU with 2GB ram, I was running mnist handwritten kind of example for a native language with training data containing of 38700 images and 4300 testing images and was trying to get precision , recall , F1 using following code as sklearn was not giving me precise reults. once i added this to my existing code i started getting GPU errors.<\/p>\n\n<pre><code>TP = tf.count_nonzero(predicted * actual)\nTN = tf.count_nonzero((predicted - 1) * (actual - 1))\nFP = tf.count_nonzero(predicted * (actual - 1))\nFN = tf.count_nonzero((predicted - 1) * actual)\n\nprec = TP \/ (TP + FP)\nrecall = TP \/ (TP + FN)\nf1 = 2 * prec * recall \/ (prec + recall)\n<\/code><\/pre>\n\n<p>plus my model was heavy i guess, i was getting memory error after 147, 148 epochs, and then I thought why not create functions for the tasks so I dont know if it works this way in tensrorflow, but I thought if a local variable is used and when out of scope it may release memory and i defined the above elements for training and testing in modules, I was able to achieve 10000 epochs without any issues, I hope this will help..<\/p>\n"},{"AnswerId":"58126280","CreationDate":"2019-09-27T00:28:35.503","ParentId":null,"OwnerUserId":"3503565","Title":null,"Body":"<pre><code># allocate 60% of GPU memory \nfrom keras.backend.tensorflow_backend import set_session\nimport tensorflow as tf \nconfig = tf.ConfigProto()\nconfig.gpu_options.per_process_gpu_memory_fraction = 0.6\nset_session(tf.Session(config=config))\n<\/code><\/pre>\n"},{"AnswerId":"52828871","CreationDate":"2018-10-16T06:05:52.857","ParentId":null,"OwnerUserId":"9110143","Title":null,"Body":"<p>i tried to train unet on voc data set but because of huge image size, memory finishes. i tried all the above tips, even tried with batch size==1, yet to no improvement. sometimes TensorFlow version also causes the memory issues. try by using <\/p>\n\n<blockquote>\n  <p>pip install tensorflow-gpu==1.8.0<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":34199978,"AnswerCount":1,"Tags":"<installation><packages><caffe>","CreationDate":"2015-12-10T10:50:39.317","AcceptedAnswerId":34211372.0,"OwnerUserId":3902984.0,"Title":"Error in installation of caffe on fedora 21","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/erob5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/erob5.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have tried to build caffe for python on my fedora 21(without GPU) for quite some time with no success. Can someone please help me on this?<\/p>\n\n<p>Few of the commands that I have executed are:<\/p>\n\n<pre><\/pre>\n\n<p>On running the 'make all' command I get the result as shown on the terminal screen.\nBut OpenBLAS seems to be present at location \/opt\/openblas and I found cblas.h file at \/opt\/openblas\/include. Then why do I get such an error?\nTo try to fix this, I added the following lines to my ~\/.bashrc file:<\/p>\n\n<pre><\/pre>\n\n<p>I also executed the following on the terminal and got the output as below,<\/p>\n\n<pre><\/pre>\n\n<p>Note: I have also installed atlas. But nothing has come to my rescue.\nWhat am I missing here?<\/p>\n","answers":[{"AnswerId":"34211372","CreationDate":"2015-12-10T20:43:51.893","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Edit Make.config file to reflect the changes. You will find a flag to select the BLAS library, set its  value as 'open'. Then uncomment the library path and include path and reflect the correct values for those variables.<\/p>\n"}]}
{"QuestionId":34204551,"AnswerCount":2,"Tags":"<unit-testing><tensorflow><testing><bazel>","CreationDate":"2015-12-10T14:35:06.983","AcceptedAnswerId":34205394.0,"OwnerUserId":1067998.0,"Title":"Run Tensorflow unit tests","Body":"<p>Is there any way to run Tensorflow unit tests manually? I want to perform sanity checks while modifying TF source code.<\/p>\n\n<p>I see there are many _test.py files with classes that perform many test operations and I can't figure out how to run them. There should be an easy way?<\/p>\n","answers":[{"AnswerId":"34205394","CreationDate":"2015-12-10T15:11:55.747","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to run the TensorFlow unit tests is using Bazel, assuming you have <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\">downloaded the source from Git<\/a>:<\/p>\n\n<pre><code># All tests (for C++ changes).\n$ bazel test \/\/tensorflow\/...\n\n# All Python tests (for Python front-end changes).\n$ bazel test \/\/tensorflow\/python\/...\n\n# All tests (with GPU support).\n$ bazel test -c opt --config=cuda \/\/tensorflow\/...\n$ bazel test -c opt --config=cuda \/\/tensorflow\/python\/...\n<\/code><\/pre>\n"},{"AnswerId":"55016320","CreationDate":"2019-03-06T05:43:37.817","ParentId":null,"OwnerUserId":"7458927","Title":null,"Body":"<p>In addition to the above answer, you can run individual tests as shown instead of full packages, which can save a significant amount of time.<\/p>\n\n<pre><code>bazel run \/\/tensorflow\/python\/kernel_tests:string_split_op_test\n\nbazel run \/\/tensorflow\/python:special_math_ops_test\n<\/code><\/pre>\n\n<p>Or you can go to the individual directory and run all the tests there<\/p>\n\n<pre><code>cd python\/kernel_tests\nbazel run :one_hot_op_test\n<\/code><\/pre>\n"}]}
{"QuestionId":34207530,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-10T16:56:47.620","AcceptedAnswerId":34214946.0,"OwnerUserId":404737.0,"Title":"Tensorflow, equivalent of Theano's pydotprint?","Body":"<p>In Theano, I can use pydotprint to generate a nice graph of my model.  Very useful for debugging, and for presenting too.  Is there an equivalent for TensorFlow?<\/p>\n","answers":[{"AnswerId":"34209619","CreationDate":"2015-12-10T18:58:31.730","ParentId":null,"OwnerUserId":"5665427","Title":null,"Body":"<p>Look into <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">Tensorboard<\/a>, which ships with Tensorflow. I use it to track the performance of my models and make sure they are converging.<\/p>\n"},{"AnswerId":"34214946","CreationDate":"2015-12-11T01:55:01.063","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As @JHafdahl points out, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">TensorBoard<\/a> provides graph visualization for TensorFlow graphs, which includes support for summarizing complex nested subgraphs.<\/p>\n\n<p>To visualize a graph, build a TensorFlow graph as normal, then add the following statements to your Python program:<\/p>\n\n<pre><code>writer = tf.train.SummaryWriter(\"\/path\/to\/logs\", tf.get_default_graph().as_graph_def())\nwriter.flush()\n<\/code><\/pre>\n\n<p>Then, in a separate terminal, run TensorBoard to visualize your graph:<\/p>\n\n<pre><code>$ tensorboard --logdir=\/path\/to\/logs --port 6006\n<\/code><\/pre>\n\n<p>Finally, connect to TensorBoard by opening <code>http:\/\/localhost:6006<\/code> in your web browser. Clicking on the \"Graph\" tab will show the visualization of your graph; see the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/graph_viz\/index.html\" rel=\"nofollow\">graph visualization tutorial<\/a> for more details.<\/p>\n"}]}
{"QuestionId":34208336,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-10T17:38:41.070","AcceptedAnswerId":null,"OwnerUserId":5665427.0,"Title":"Tensorflow multivariate linear regression not converging","Body":"<p>I am trying to train a multivariate linear regression model with regularization using tensorflow. For some reason I am not able to get the training piece of the below code to calculate the error I want to use for the gradient descent update. Am I doing something wrong in setting up my graph?<\/p>\n\n<pre><\/pre>\n\n<p>my output looks like this<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43209291","CreationDate":"2017-04-04T13:55:03.227","ParentId":null,"OwnerUserId":"4511978","Title":null,"Body":"<p>it seems indeed a problem with the learning rate: <code>0.03<\/code> may be too high depending on how does your data look like. Also, you probably want to create your graph separated from the session in a more explicit way, or even use the <em>normal equations<\/em> to reach the optimal solution without having to iterate, if your dataset has a mid\/low dimensionality. <a href=\"https:\/\/stackoverflow.com\/a\/43170585\/4511978\">Here<\/a> I posted some examples that you may hopefully find helpful! Also, the <a href=\"https:\/\/www.tensorflow.org\/get_started\/get_started\" rel=\"nofollow noreferrer\">TF tutorials<\/a> cover it well (search for \"Complete program\" in that page).<\/p>\n\n<p>But regarding your code, here is a version that worked for me: I changed some deprecated functions, and basically set the learning rate to the much lower value <code>alpha=1e-8<\/code>, which (on the synthetic dataset also generated in the code) seems to converge:<\/p>\n\n<pre><code>test accuracy 2176.11\ntest accuracy 1898.6\ntest accuracy 1663.69\ntest accuracy 1458.53\ntest accuracy 1287.57\ntest accuracy 1116.9\ntest accuracy 969.474\ntest accuracy 841.028\ntest accuracy 738.592\ntest accuracy 649.891\ntest accuracy 565.188\ntest accuracy 495.33\ntest accuracy 438.351\ntest accuracy 381.161\ntest accuracy 333.213\ntest accuracy 289.575\ntest accuracy 254.394\ntest accuracy 222.836\ntest accuracy 197.36\ntest accuracy 172.788\ntest accuracy 152.251\ntest accuracy 132.664\ntest accuracy 115.982\ntest accuracy 101.021\nfinal test accuracy 90.2555\n<\/code><\/pre>\n\n<hr>\n\n<h2>CODE:<\/h2>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n\n# generate some dataset\nDIMENSIONS = 5\nDS_SIZE = 5000\nTRAIN_RATIO = 0.5 # 50% of the dataset isused for training\n_train_size = int(DS_SIZE*TRAIN_RATIO)\n_test_size = DS_SIZE - _train_size\nf = lambda(x): sum(x) # the \"true\" function: f = 0 + 1*x1 + 1*x2 + 1*x3 ...\nnoise = lambda: np.random.normal(0,10) # some noise\n# training globals\nLAMBDA = 1e6 # L2 regularization factor\n# generate the dataset, the labels and split into train\/test\nds = [[np.random.rand()*1000 for d in range(DIMENSIONS)] for _ in range(DS_SIZE)]\nds = [([1]+x, [f(x)+noise()]) for x in ds] # add x[0]=1 dimension and labels\nnp.random.shuffle(ds)\ntrain_data, train_labels = zip(*ds[0:_train_size])\ntest_data, test_labels = zip(*ds[_train_size:])\n\n\n\ndef normalize_data(matrix):\n    averages = np.average(matrix,0)\n    mins = np.min(matrix,0)\n    maxes = np.max(matrix,0)\n    ranges = maxes - mins\n    return ((matrix - averages)\/ranges)\n\ndef run_regression(X, Y, X_test, Y_test, lambda_value = 0.1, normalize=False, batch_size=10, alpha=1e-8):\n    x_train = normalize_data(X) if normalize else X\n    y_train = Y\n    x_test = X_test\n    y_test = Y_test\n    session = tf.Session()\n    # Calculate number of features for X and Y\n    x_features_length = len(X[0])\n    y_features_length = len(Y[0])\n    # Build Tensorflow graph parts\n    x = tf.placeholder('float', [None, x_features_length], name=\"X\")\n    y = tf.placeholder('float', [None, y_features_length], name=\"Y\")\n    theta = tf.Variable(tf.random_normal([x_features_length, y_features_length], stddev=0.01), name=\"Theta\")\n    lambda_val = tf.constant(lambda_value)\n    # Trying to implement this way http:\/\/openclassroom.stanford.edu\/MainFolder\/DocumentPage.php?course=MachineLearning&amp;doc=exercises\/ex5\/ex5.html\n    y_predicted = tf.matmul(x, theta, name=\"y_predicted\")\n    #regularization_cost_part = tf.cast(tf.multiply(lambda_val,tf.reduce_sum(tf.pow(theta,2)), name=\"regularization_param\"), 'float')\n    #polynomial_cost_part = tf.reduce_sum(tf.pow(tf.subtract(y_predicted, y), 2), name=\"polynomial_sum\")\n    # Set up some summary info to debug\n    with tf.name_scope('cost') as scope:\n        #cost_func = tf.multiply(tf.cast(1\/(2*batch_size), 'float'), tf.cast(tf.add(polynomial_cost_part, regularization_cost_part), 'float'))\n        cost_func = (tf.nn.l2_loss(y_predicted - y)+lambda_val*tf.nn.l2_loss(theta))\/float(batch_size)\n        #DEPRECATED*** cost_summary = tf.scalar_summary(\"cost\", cost_func)\n        cost_summary = tf.summary.scalar('cost', cost_func)# Add a scalar summary for the snapshot loss.\n    training_func = tf.train.GradientDescentOptimizer(alpha).minimize(cost_func)\n    with tf.name_scope(\"test\") as scope:\n        correct_prediction = tf.subtract(tf.cast(1, 'float'), tf.reduce_mean(tf.subtract(y_predicted, y)))\n        accuracy = tf.cast(correct_prediction, \"float\")\n        #DEPRECATED*** accuracy_summary = tf.scalar_summary(\"accuracy\", accuracy)\n        #accuracy_summary = tf.summary.scalar(\"accuracy\", accuracy)\n    saver = tf.train.Saver()\n    #DEPRECATED*** merged = tf.merge_all_summaries()\n    merged = tf.summary.merge_all()\n    #DEPRECATED*** writer = tf.train.SummaryWriter(\"\/tmp\/football_logs\", session.graph_def)\n    writer = tf.summary.FileWriter(\"\/tmp\/football_logs\", session.graph)\n    #DEPRECATED*** init = tf.initialize_all_variables()\n    init = tf.global_variables_initializer()\n    session.run(init)\n    for i in range(1, (len(x_train)\/batch_size)):\n        session.run(training_func, feed_dict={x: x_train[i*batch_size:i*batch_size+batch_size], y: y_train[i*batch_size:i*batch_size+batch_size]})\n        if i % batch_size == 0:\n            print \"test accuracy %g\"%session.run(accuracy, feed_dict={x: x_test, y: y_test})\n            #result = session.run([merged, accuracy], feed_dict={x: x_test, y: y_test})\n            # writer.add_summary(result[0], i)\n            # print \"step %d, training accuracy %g\"%(i, result[1])\n            #writer.flush()\n    print \"final test accuracy %g\"%session.run(accuracy, feed_dict={x: x_test, y: y_test})\n    # save_path = saver.save(session, \"\/tmp\/football.ckpt\")\n    # print \"Model saved in file: \", save_path\n    session.close()\n\nrun_regression(train_data, train_labels, test_data, test_labels, normalize=False, alpha=1e-8)\n<\/code><\/pre>\n\n<p>As I said, you will probably want to change the structure to favor readability and scalability, but hopefully this helps!<\/p>\n\n<p>Cheers,\nAndres<\/p>\n"}]}
{"QuestionId":34209874,"AnswerCount":1,"Tags":"<installation><caffe>","CreationDate":"2015-12-10T19:13:16.347","AcceptedAnswerId":34211276.0,"OwnerUserId":3902984.0,"Title":"Error in caffe installation","Body":"<p>I am facing a issue when i run the following command while installing caffe on fedora 21<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p>Some help please!<\/p>\n\n<p>I still can't get rid of the last two (-lcblas and -latlas) errors. I have installed atlas and OpenBLAS long ago. Also I tried the following commands but with no success<\/p>\n\n<pre><\/pre>\n\n<p>Need some inputs!!<\/p>\n","answers":[{"AnswerId":"34211276","CreationDate":"2015-12-10T20:38:05.367","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>The leveldb present in your system seems to be of a different architecture, most probably a 32bit compiled leveldb and 64bit compiled caffe, or vice-versa.<\/p>\n\n<p>The rest of the libraries are to be installed in Fedora before compiling Caffe.<\/p>\n"}]}
{"QuestionId":34210606,"AnswerCount":1,"Tags":"<neural-network><caffe><conv-neural-network>","CreationDate":"2015-12-10T19:56:03.330","AcceptedAnswerId":38025728.0,"OwnerUserId":1648356.0,"Title":"Caffe: How to train end-to-end (image to image)?","Body":"<p>We are quite new to <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">caffe<\/a>, but what we have seen so far, looks really promising.<\/p>\n\n<p>After reading a few papers (<a href=\"http:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow\">1<\/a>,<a href=\"https:\/\/github.com\/s9xie\/hed\" rel=\"nofollow\">2<\/a>), we wanted to reproduce the result of <a href=\"http:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow\">1<\/a>, specifically about a segmentation challenge <a href=\"http:\/\/brainiac2.mit.edu\/isbi_challenge\/home\" rel=\"nofollow\">4<\/a>.<\/p>\n\n<p>We downloaded the modified caffe from <a href=\"http:\/\/lmb.informatik.uni-freiburg.de\/people\/ronneber\/u-net\/\" rel=\"nofollow\">3<\/a> and were able to execute it, just to see, that the trained network didn't work with the dataset from <a href=\"http:\/\/brainiac2.mit.edu\/isbi_challenge\/home\" rel=\"nofollow\">4<\/a>.<\/p>\n\n<p>At first we thought that the network needs to be trained for the specific problem. \nWhich lead to the problem of how to do 'image-to-image (aka end-to-end) learning ' (<a href=\"http:\/\/brainiac2.mit.edu\/isbi_challenge\/home\" rel=\"nofollow\">4<\/a>, training data).<\/p>\n\n<p>This lead us to 'holistically nested edge detection' (hed, <a href=\"https:\/\/github.com\/s9xie\/hed\" rel=\"nofollow\">2<\/a>), where image-to-image learning, seems to be used.\nWith hed, we were able to retrain the network on our own. But it doesn't work (it leads to all 0 or 0.5 images - black images :-( ) if we try to train the network for the dataset of <a href=\"http:\/\/brainiac2.mit.edu\/isbi_challenge\/home\" rel=\"nofollow\">4<\/a>. For initialization we wrote a script to calculate the mean-map witch we use for the dataset of <a href=\"http:\/\/brainiac2.mit.edu\/isbi_challenge\/home\" rel=\"nofollow\">4<\/a>.<\/p>\n\n<p><strong>Our question(s) are:<\/strong><\/p>\n\n<ul>\n<li>How can we reproduce the result, mentioned in <a href=\"http:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow\">1<\/a> by running\nimage-to-image training?<\/li>\n<\/ul>\n\n<p><em>or<\/em><\/p>\n\n<ul>\n<li>How do you train networks, where we have image-to-image learning?<\/li>\n<li>Since we only have 30 image-to-image pairs, should we implement\ndeformation as mentioned in <a href=\"http:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow\">1<\/a>\/<a href=\"http:\/\/lmb.informatik.uni-freiburg.de\/people\/ronneber\/u-net\/\" rel=\"nofollow\">3<\/a> via matlab\/python or is there a\nfunctionality within caffe already? <\/li>\n<li>Are we missing something simple from <a href=\"http:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow\">1<\/a> or <a href=\"https:\/\/github.com\/s9xie\/hed\" rel=\"nofollow\">2<\/a>?<\/li>\n<\/ul>\n\n<p>Kind regards,\nKlaus and Bernhard<\/p>\n\n<p>Ps: We asked the same question at the caffe-user group and intend to post solutions at both <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/RrO0Y4e3AbQ\" rel=\"nofollow\">locations<\/a>.<\/p>\n","answers":[{"AnswerId":"38025728","CreationDate":"2016-06-25T06:33:25.237","ParentId":null,"OwnerUserId":"1648356","Title":null,"Body":"<p>After some time, and trying several different things out - i stumbled upon:<\/p>\n\n<p><a href=\"https:\/\/github.com\/naibaf7\" rel=\"nofollow\">https:\/\/github.com\/naibaf7<\/a><\/p>\n\n<p>Using that caffe fork, with caffe_neural_models and caffe_neural_tool training image(raw)-to-image(labels) can be done quite simple.<\/p>\n\n<p>Just check out 'caffe_neural_models\/net*' for different configurations.<\/p>\n"}]}
{"QuestionId":34213543,"AnswerCount":0,"Tags":"<tensorflow><demo>","CreationDate":"2015-12-10T23:17:48.850","AcceptedAnswerId":null,"OwnerUserId":5587428.0,"Title":"New Error Running CIFAR-10 Demo with TF 0.6.0","Body":"<p>I just updated TF with pip and attempted to run the cifar demo for a single gpu system. I am getting the errors below when running the single-gpu version of the demo code. The multi-gpu version of the demo code runs fine. Also previously the single-gpu demo ran fine under 0.5.0. <\/p>\n\n<p>A portion of the error output:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34215746,"AnswerCount":3,"Tags":"<python><scope><tensorflow>","CreationDate":"2015-12-11T03:42:46.823","AcceptedAnswerId":null,"OwnerUserId":1576602.0,"Title":"What is the difference between variable_scope and name_scope?","Body":"<p>What is the difference between  and ? The <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope#names_of_ops_in_tfvariable_scope\" rel=\"nofollow noreferrer\">variable scope tutorial<\/a> talks about  implicitly opening . I also noticed that creating a variable in a  automatically expands its name with the scope name as well. So, what is the difference?<\/p>\n","answers":[{"AnswerId":"52277944","CreationDate":"2018-09-11T14:12:40.350","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p><code>tf.variable_scope<\/code> is an evolution of <code>tf.name_scope<\/code> to handle <code>Variable<\/code> reuse. As you noticed, it does more than <code>tf.name_scope<\/code>, so there is no real reason to use <code>tf.name_scope<\/code>: not surprisingly, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6007#issuecomment-264535954\" rel=\"nofollow noreferrer\">a TF developper advises to just use <code>tf.variable_scope<\/code><\/a>.<\/p>\n\n<p>My understanding for having <code>tf.name_scope<\/code> still lying around is that there are subtle incompatibilities in the behavior of those two, which invalidates <code>tf.variable_scope<\/code> as a drop-in replacement for <code>tf.name_scope<\/code>.<\/p>\n"},{"AnswerId":"34232533","CreationDate":"2015-12-11T21:04:12.837","ParentId":null,"OwnerUserId":"604964","Title":null,"Body":"<p>When you create a variable with <code>tf.get_variable<\/code> instead of <code>tf.Variable<\/code>, Tensorflow will start checking the names of the vars created with the same method to see if they collide. If they do, an exception will be raised. If you created a var with <code>tf.get_variable<\/code> and you try to change the prefix of your variable names by using the <code>tf.name_scope<\/code> context manager, this won't prevent the Tensorflow of raising an exception. Only <code>tf.variable_scope<\/code> context manager will effectively change the name of your var in this case. Or if you want to reuse the variable you should call scope.reuse_variables() before creating the var the second time.<\/p>\n\n<p>In summary, <code>tf.name_scope<\/code> just add a prefix to all tensor created in that scope (except the vars created with <code>tf.get_variable<\/code>), and <code>tf.variable_scope<\/code> add a prefix to the variables created with <code>tf.get_variable<\/code>.<\/p>\n"},{"AnswerId":"43580096","CreationDate":"2017-04-24T05:00:06.303","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I had problems understanding the difference between <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"noreferrer\">variable_scope<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/name_scope\" rel=\"noreferrer\">name_scope<\/a> (they looked almost the same) before I tried to visualize everything by creating a simple example:<\/p>\n\n<pre><code>import tensorflow as tf\ndef scoping(fn, scope1, scope2, vals):\n    with fn(scope1):\n        a = tf.Variable(vals[0], name='a')\n        b = tf.get_variable('b', initializer=vals[1])\n        c = tf.constant(vals[2], name='c')\n        with fn(scope2):\n            d = tf.add(a * b, c, name='res')\n\n        print '\\n  '.join([scope1, a.name, b.name, c.name, d.name]), '\\n'\n    return d\n\nd1 = scoping(tf.variable_scope, 'scope_vars', 'res', [1, 2, 3])\nd2 = scoping(tf.name_scope,     'scope_name', 'res', [1, 2, 3])\n\nwith tf.Session() as sess:\n    writer = tf.summary.FileWriter('logs', sess.graph)\n    sess.run(tf.global_variables_initializer())\n    print sess.run([d1, d2])\n    writer.close()\n<\/code><\/pre>\n\n<p>Here I create a function that creates some variables and constants and groups them in scopes (depending by the type I provided). In this function I also print the names of all the variables. After that I executes the graph to get values of the resulting values and save event-files to investigate them in tensorboard. If you run this, you will get the following:<\/p>\n\n<pre><code>scope_vars\n  scope_vars\/a:0\n  scope_vars\/b:0\n  scope_vars\/c:0\n  scope_vars\/res\/res:0 \n\nscope_name\n  scope_name\/a:0\n  b:0\n  scope_name\/c:0\n  scope_name\/res\/res:0 \n<\/code><\/pre>\n\n<p>You see the similar pattern if you open TB (as you see <code>b<\/code> is outside of <code>scope_name<\/code> rectangular):\n<a href=\"https:\/\/i.stack.imgur.com\/K0VgJ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/K0VgJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<hr>\n\n<p><strong>This gives you the answer<\/strong>:<\/p>\n\n<p>Now you see that <code>tf.variable_scope()<\/code> adds a prefix to the names of all variables (no matter how you create them), ops, constants. On the other hand <code>tf.name_scope()<\/code> ignores variables created with <code>tf.get_variable()<\/code> because it assumes that you know which variable and in which scope you wanted to use.<\/p>\n\n<p>A good documentation on <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"noreferrer\">Sharing variables<\/a> tells you that <\/p>\n\n<blockquote>\n  <p><code>tf.variable_scope()<\/code>: Manages namespaces for names passed to <code>tf.get_variable()<\/code>.<\/p>\n<\/blockquote>\n\n<p>The same documentation provides a more details how does Variable Scope work and when it is useful.<\/p>\n"}]}
{"QuestionId":34217582,"AnswerCount":2,"Tags":"<torch>","CreationDate":"2015-12-11T06:42:55.023","AcceptedAnswerId":null,"OwnerUserId":5629831.0,"Title":"Decreasing one dimension of a tensor in Lua Torch","Body":"<p>I have a tensor X of the shape<\/p>\n\n<pre><\/pre>\n\n<p>where the three dots can be any vector of dimensions (i.e: 4,4)<\/p>\n\n<p>I would like to transform this tensor to a tensor whose shape is<\/p>\n\n<pre><\/pre>\n\n<p>In other words, I want to reduce the dimensionality of the first dimension. Of course, I will loose the information indexed by 2 but it does not matter in this case. The problem is not trivial as I don't know the number of dimensions of the tensors. But at least it's greater or equal to 2. A th code is written below:<\/p>\n\n<pre><\/pre>\n\n<p>Do you have any ideas how to do it?<\/p>\n\n<p>Many thanks in advance<\/p>\n","answers":[{"AnswerId":"34224759","CreationDate":"2015-12-11T13:33:16.093","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>I want to reduce the dimensionality of the first dimension<\/p>\n<\/blockquote>\n\n<p>If by reducing you mean <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/1d4047e\/doc\/tensor.md#self-narrowdim-index-size\" rel=\"nofollow\">narrow down<\/a> then you can do e.g.:<\/p>\n\n<pre><code>x = torch.Tensor(2, 4, 4)\nx[1]:fill(0)\nx[2]:fill(1)\n\n-- this creates tensors of size 1x4x4\n-- the first dimension is narrowed, one index at a time\na = x:narrow(1, 1, 1)\nb = x:narrow(1, 2, 1)\n<\/code><\/pre>\n\n<p>This gives:<\/p>\n\n<pre><code>&gt; print(a)\n(1,.,.) = \n  0  0  0  0\n  0  0  0  0\n  0  0  0  0\n  0  0  0  0\n[torch.DoubleTensor of size 1x4x4]\n\n&gt; print(b)\n(1,.,.) = \n  1  1  1  1\n  1  1  1  1\n  1  1  1  1\n  1  1  1  1\n[torch.DoubleTensor of size 1x4x4]\n<\/code><\/pre>\n"},{"AnswerId":"34228796","CreationDate":"2015-12-11T17:04:24.200","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>This code does what you need:<\/p>\n\n<pre><code>y = x[{{1, 1}, {}, {}}]\n<\/code><\/pre>\n\n<p>The unknown number of dimensions? Not a problem, you don't have to specify them all:<\/p>\n\n<pre><code>y = x[{{1, 1}}]\n\n(1,.,.) = \n  1  1  1  1\n  1  1  1  1\n  1  1  1  1\n  1  1  1  1\n[torch.DoubleTensor of size 1x4x4]\n<\/code><\/pre>\n"}]}
{"QuestionId":34220374,"AnswerCount":1,"Tags":"<mask><tensorflow>","CreationDate":"2015-12-11T09:43:03.380","AcceptedAnswerId":34220963.0,"OwnerUserId":867505.0,"Title":"Apply 1 channel mask to 3 channel Tensor in tensorflow","Body":"<p>I'm trying to apply a mask (binary, only one channel) to an RGB image (3 channels, normalized to [0, 1]). My current solution is, that I split the RGB image into it's channels, multiply it with the mask and concatenate these channels again:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, this seems pretty inefficient, especially since, to my understanding, none of these computations are done in-place. Is there a more efficient way for doing this?<\/p>\n","answers":[{"AnswerId":"34220963","CreationDate":"2015-12-11T10:11:45.673","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#mul\"><code>tf.mul()<\/code><\/a> operator supports <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\">numpy-style broadcasting<\/a>, which would allow you to simplify and optimize the code slightly.<\/p>\n\n<p>Let's say that <code>zero_one_mask<\/code> is an <code>m x n<\/code> tensor, and <code>output_img<\/code> is a <code>b x m x n x 3<\/code> (where <code>b<\/code> is the batch size - I'm inferring this from the fact that you split <code>output_img<\/code> on dimension 3)*. You can use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#expand_dims\"><code>tf.expand_dims()<\/code><\/a> to make <code>zero_one_mask<\/code> broadcastable to <code>channels<\/code>, by reshaping it to be an <code>m x n x 1<\/code> tensor:<\/p>\n\n<pre><code>with tf.variable_scope('apply_mask') as scope:\n  # Output mask is in range [-1, 1], bring to range [0, 1] first\n  # NOTE: Assumes `output_mask` is a 2-D `m x n` tensor.\n  zero_one_mask = tf.expand_dims((output_mask + 1) \/ 2, 2)\n  # Apply mask to all channels.\n  # NOTE: Assumes `output_img` is a 4-D `b x m x n x c` tensor.\n  output_img = tf.mul(output_img, zero_one_mask)\n<\/code><\/pre>\n\n<p>(* This would work equally if <code>output_img<\/code> were a 4-D <code>b x m x n x c<\/code> (for any number of channels <code>c<\/code>) or 3-D <code>m x n x c<\/code> tensor, due to the way broadcasting works.)<\/p>\n"}]}
{"QuestionId":34220532,"AnswerCount":9,"Tags":"<python><tensorflow><neural-network><deep-learning><variable-assignment>","CreationDate":"2015-12-11T09:51:46.307","AcceptedAnswerId":34220750.0,"OwnerUserId":3537687.0,"Title":"How to assign a value to a TensorFlow variable?","Body":"<p>I am trying to assign a new value to a tensorflow variable in python.<\/p>\n\n<pre><\/pre>\n\n<p>But the output I get is<\/p>\n\n<pre><\/pre>\n\n<p>So the value has not changed. What am I missing?<\/p>\n","answers":[{"AnswerId":"47272180","CreationDate":"2017-11-13T19:37:03.687","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Also, it has to be noted that if you're using <code>your_tensor.assign()<\/code>, then the <code>tf.global_variables_initializer<\/code> need not be called explicitly since the assign operation does it for you in the background.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>In [212]: w = tf.Variable(12)\nIn [213]: w_new = w.assign(34)\n\nIn [214]: with tf.Session() as sess:\n     ...:     sess.run(w_new)\n     ...:     print(w_new.eval())\n\n# output\n34 \n<\/code><\/pre>\n\n<p>However, this will not initialize all variables, but it will only initialize the variable on which <code>assign<\/code> was executed on.<\/p>\n"},{"AnswerId":"43580452","CreationDate":"2017-04-24T05:33:53.437","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>First of all you can assign values to variables\/constants just by feeding values into them the same way you do it with placeholders. So this is perfectly legal to do:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.Variable(0)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print sess.run(x, feed_dict={x: 3})\n<\/code><\/pre>\n\n<p>Regarding your confusion with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"noreferrer\">tf.assign()<\/a> operator. In TF nothing is executed before you run it inside of the session. So you always have to do something like this: <code>op_name = tf.some_function_that_create_op(params)<\/code> and then inside of the session you run <code>sess.run(op_name)<\/code>. Using assign as an example you will do something like this:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.Variable(0)\ny = tf.assign(x, 1)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print sess.run(x)\n    print sess.run(y)\n    print sess.run(x)\n<\/code><\/pre>\n"},{"AnswerId":"55492890","CreationDate":"2019-04-03T10:26:05.233","ParentId":null,"OwnerUserId":"7580851","Title":null,"Body":"<p>So i had a adifferent case where i needed to assign values before running a session, So this was the easiest way to do that:<\/p>\n\n<pre><code>other_variable = tf.get_variable(\"other_variable\", dtype=tf.int32,\n  initializer=tf.constant([23, 42]))\n<\/code><\/pre>\n\n<p>here i'm creating a variable as well as assigning it values at the same time<\/p>\n"},{"AnswerId":"54467346","CreationDate":"2019-01-31T18:49:27.570","ParentId":null,"OwnerUserId":"10997543","Title":null,"Body":"<p>Use Tensorflow eager execution mode which is latest.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\nmy_int_variable = tf.get_variable(\"my_int_variable\", [1, 2, 3])\nprint(my_int_variable)\n<\/code><\/pre>\n"},{"AnswerId":"34220750","CreationDate":"2015-12-11T10:01:57.190","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The statement <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.assign\" rel=\"noreferrer\"><code>x.assign(1)<\/code><\/a> does not actually assign the value <code>1<\/code> to <code>x<\/code>, but rather creates a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Operation\" rel=\"noreferrer\"><code>tf.Operation<\/code><\/a> that you have to explicitly <em>run<\/em> to update the variable.* A call to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Operation.run\" rel=\"noreferrer\"><code>Operation.run()<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a> can be used to run the operation:<\/p>\n\n<pre><code>assign_op = x.assign(1)\nsess.run(assign_op)  # or `assign_op.op.run()`\nprint(x.eval())\n# ==&gt; 1\n<\/code><\/pre>\n\n<p>(* In fact, it returns a <code>tf.Tensor<\/code>, corresponding to the updated value of the variable, to make it easier to chain assignments.)<\/p>\n"},{"AnswerId":"48489637","CreationDate":"2018-01-28T18:02:58.317","ParentId":null,"OwnerUserId":"5096199","Title":null,"Body":"<p>You can also assign a new value to a <code>tf.Variable<\/code> without adding an operation to the graph: <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.5\/api_docs\/python\/tf\/Variable#load\" rel=\"nofollow noreferrer\"><code>tf.Variable.load(value, session)<\/code><\/a>. This function can also save you adding placeholders when assigning a value from outside the graph and it is useful in case the graph is finalized.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nx = tf.Variable(0)\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nprint(sess.run(x))  # Prints 0.\nx.load(1, sess)\nprint(sess.run(x))  # Prints 1.\n<\/code><\/pre>\n\n<p>Update: This is depricated in TF2 as eager execution is default and graphs are <a href=\"https:\/\/medium.com\/tensorflow\/effective-tensorflow-2-0-best-practices-and-whats-changed-a0ca48767aff\" rel=\"nofollow noreferrer\">no longer exposed in the user-facing API<\/a>.<\/p>\n"},{"AnswerId":"54226708","CreationDate":"2019-01-16T23:04:27.820","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Here is the complete working example:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nw= tf.Variable(0, dtype=tf.float32) #good practice to set the type of the variable\ncost = 10 + 5*w + w*w\ntrain = tf.train.GradientDescentOptimizer(0.01).minimize(cost)\n\ninit = tf.global_variables_initializer()\nsession = tf.Session()\nsession.run(init)\n\nprint(session.run(w))\n\nsession.run(train)\nprint(session.run(w)) # runs one step of gradient descent\n\nfor i in range(10000):\n  session.run(train)\n\nprint(session.run(w))\n<\/code><\/pre>\n\n<p>Note the output will be:<\/p>\n\n<pre><code>0.0\n-0.049999997\n-2.499994\n<\/code><\/pre>\n\n<p>This means at the very start the Variable was 0, as defined, then after just one step of gradient decent the variable was -0.049999997, and after 10.000 more steps we are reaching -2.499994 (based on our cost function).<\/p>\n\n<p>Note: You originally used the Interactive session. Interactive session is useful when multiple different sessions needed to be run in the same script. However, I used the non interactive session for simplicity.<\/p>\n"},{"AnswerId":"42785945","CreationDate":"2017-03-14T12:24:30.953","ParentId":null,"OwnerUserId":"4219997","Title":null,"Body":"<p>There is an easier approach: <\/p>\n\n<pre><code>x = tf.Variable(0)\nx = x + 1\nprint x.eval()\n<\/code><\/pre>\n"},{"AnswerId":"56544146","CreationDate":"2019-06-11T12:48:48.933","ParentId":null,"OwnerUserId":"6721863","Title":null,"Body":"<p>I answered a similar question <a href=\"https:\/\/stackoverflow.com\/a\/56544044\/6721863\">here<\/a>.  I looked in a lot of places that always created the same problem.  Basically, I did not want to assign a value to the weights, but simply change the weights.  The short version of the above answer is:<\/p>\n\n<p><code>tf.keras.backend.set_value(tf_var, numpy_weights)<\/code><\/p>\n"}]}
{"QuestionId":34221453,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-11T10:35:56.043","AcceptedAnswerId":null,"OwnerUserId":5668239.0,"Title":"Tensorflow: how to add user custom op accepting two 1D vec tensor and output a scalar?","Body":"<p>I'm trying below but not work.<\/p>\n\n<pre><\/pre>\n\n<p>.\/test.py \nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 8\nI tensorflow\/core\/common_runtime\/direct_session.cc:60] Direct session inter op parallelism threads: 8\nF .\/tensorflow\/core\/public\/tensor.h:453] Check failed: dtype() ==  DataTypeToEnum::v() (1 vs. 2)\nAborted<\/p>\n","answers":[{"AnswerId":"34222395","CreationDate":"2015-12-11T11:22:27.043","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue is that the <code>predicts<\/code> tensor in your Python program has type <code>float<\/code>, and your op registration accepts this as a valid type for the <code>predicts<\/code> input (since <code>T1<\/code> can be <code>float<\/code> or <code>double<\/code>), <strong>but<\/strong> <code>AucOp::Compute()<\/code> assumes that the <code>predicts<\/code> input always has type <code>double<\/code> (in the call to <code>predicts_tensor.flat&lt;double&gt;()<\/code>). The <code>tensorflow::Tensor<\/code> class does not convert the type of elements in the tensor when you ask for values of a different type, and instead it raises a fatal error.<\/p>\n\n<p>There are several possible solutions:<\/p>\n\n<ol>\n<li><p>To get things working quickly, you could change the type of <code>predicts<\/code> in your Python program to <code>tf.float64<\/code> (which is a synonym for <code>double<\/code> in the Python front-end):<\/p>\n\n<pre><code>predicts = tf.constant([0.8, 0.5, 0.12], dtype=tf.float64)\n<\/code><\/pre><\/li>\n<li><p>You could start by defining a simpler op that accepts inputs of a single type only:<\/p>\n\n<pre><code>REGISTER_OP(\"Auc\")\n.Input(\"predicts: double\")\n.Input(\"labels: int32\")\n...;\n<\/code><\/pre><\/li>\n<li><p>You could add code in the <code>AucOp::Compute()<\/code> method to test the input type and access the input values as appropriate. (Use <code>this-&gt;input_type(i)<\/code> to find the type of the ith input.<\/p><\/li>\n<li><p>You could define a templated class <code>AucOp&lt;TPredict, TLabel&gt;<\/code>, then use <code>TypeConstraint&lt;&gt;<\/code> in the <code>REGISTER_KERNEL_BUILDER<\/code> call to define specializations for each of the four valid combinations of prediction and label types. This would look something like:<\/p>\n\n<pre><code>REGISTER_KERNEL_BUILDER(Name(\"Auc\")\n                            .Device(DEVICE_CPU)\n                            .TypeConstraint&lt;float&gt;(\"T1\")\n                            .TypeConstraint&lt;int32&gt;(\"T2\"),\n                       AucOp&lt;float, int32&gt;);\n\/\/ etc. for AucOp&lt;double, int32&gt;, AucOp&lt;float, int64&gt;, and AucOp&lt;double, int64&gt;.\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":34221660,"AnswerCount":2,"Tags":"<installation><linker><libraries><caffe><binaries>","CreationDate":"2015-12-11T10:46:25.840","AcceptedAnswerId":null,"OwnerUserId":3902984.0,"Title":"Make error in caffe","Body":"<p><a href=\"http:\/\/i.stack.imgur.com\/24EBq.png\" rel=\"nofollow\">Output for 'make all'<\/a><\/p>\n\n<pre><\/pre>\n\n<p>I have been trying to install caffe on fedora 21 with little success. On running the following make command I get the errors as given in the screenshot.<\/p>\n\n<p>I have tried the following to fix this but have had no success.<\/p>\n\n<pre><\/pre>\n\n<p>The contents of \/lib64\/atlas\/ are as shown below I am not sure though if these are the required libraries.<\/p>\n\n<pre><\/pre>\n\n<p>Google say's to make use of -L to link command. But I don't know which command should be used, how that command is to be used and which libraries should be linked.\nI need help please. Thanks!<\/p>\n","answers":[{"AnswerId":"43896325","CreationDate":"2017-05-10T15:06:19.560","ParentId":null,"OwnerUserId":"7992199","Title":null,"Body":"<p>I think the reason is you use make to compile, which makes caffe's python port only find libraries in this catalog. Maybe you use cmaketo compile and it could work.<\/p>\n\n<pre><code>$make clean\n$cd caffe-master\n$mkdir build\n$cd build\n$cmake ..\n$make all -j8\n<\/code><\/pre>\n\n<p>I hope I could help you!<\/p>\n"},{"AnswerId":"34227030","CreationDate":"2015-12-11T15:30:58.187","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<pre><code>cd \/lib64\/atlas\nsudo ln -s libtatlas.so \/lib\/libatlas.so\nsudo ln -s libtatlas.so \/lib\/libcblas.so\n<\/code><\/pre>\n\n<p>The new Atlas ver3 is bundled with both atlas and cblas. Thus there is no need to include both of them separately.\nEither you can follow the above method or remove calling the cblas library in your makefile.<\/p>\n"}]}
{"QuestionId":34223315,"AnswerCount":1,"Tags":"<neural-network><tensorflow><conv-neural-network>","CreationDate":"2015-12-11T12:14:16.487","AcceptedAnswerId":34269446.0,"OwnerUserId":4057461.0,"Title":"Incompatible shapes on tensorflow.equal() op for correct predictions evaluation","Body":"<p>Using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow noreferrer\">MNIST tutorial of Tensorflow<\/a>, I try to make a convolutional network for face recognition with the <a href=\"http:\/\/www.cl.cam.ac.uk\/research\/dtg\/attarchive\/facedatabase.html\" rel=\"nofollow noreferrer\">\"Database of Faces\"<\/a>.<\/p>\n\n<p>The images size are 112x92, I use 3 more convolutional layer to reduce it to 6 x 5 as adviced <a href=\"https:\/\/stackoverflow.com\/a\/33901141\/4057461\">here<\/a><\/p>\n\n<p>I'm very new at convolutional network and most of my layer declaration is made by analogy to the Tensorflow MNIST tutorial, it may be a bit clumsy, so feel free to advice me on this.<\/p>\n\n<pre><\/pre>\n\n<p>My problem appear when the session run the \"correct_prediction\" op which is<\/p>\n\n<pre><\/pre>\n\n<p>At least I think given the error message:<\/p>\n\n<pre><\/pre>\n\n<p>It looks like the y_conv output a matrix of shape <em>8 x batch_size<\/em> instead of <em>number_of_class x batch_size<\/em><\/p>\n\n<p>If I change the batch size from 20 to 10, the error message stay the same but instead <em>[8] vs. [20]<\/em> I get <em>[4] vs. [10]<\/em>. So from that I conclude that the problem may come from the <em>y_conv<\/em> declaration (last line of the code above).<\/p>\n\n<p>The loss function, optimizer, training, etc declarations is the same as in the MNIST tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for reading, have a good day<\/p>\n","answers":[{"AnswerId":"34269446","CreationDate":"2015-12-14T14:29:34.900","ParentId":null,"OwnerUserId":"4057461","Title":null,"Body":"<p>Well, after a lot debugging, I found that my issue was due to a bad instantiation of the labels. Instead of creating arrays full of zeros and replace one value by one, I created them with random value! Stupid mistake. In case someone wondering what I did wrong there and how I fix it <a href=\"https:\/\/github.com\/TonyRoussel\/orlfaces_tensorflow\/commit\/f8028429a2b90d8634e7b558c667a476c94c1eb4\" rel=\"nofollow noreferrer\">here<\/a> is the change I made.<\/p>\n\n<p>Anyway during all the debugging I made, to find this mistake, I found some useful information to debug this kind of problem:<\/p>\n\n<ol>\n<li>For the cross entropy declaration, the tensorflow's MNIST tutorial use a formula that can lead to NaN value<\/li>\n<\/ol>\n\n<p>This formula is<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))\n<\/code><\/pre>\n\n<p>Instead of this, I found two ways to declare it in a safer fashion:<\/p>\n\n<pre><code>cross_entropy = -tf.reduce_sum(y_ * tf.log(tf.clip_by_value(y_conv, 1e-10, 1.0)))\n<\/code><\/pre>\n\n<p>or also:<\/p>\n\n<pre><code>cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logit, y_))\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>As mrry says. printing the shape of the tensors can help to detect shape anomaly.<\/li>\n<\/ol>\n\n<p>To get the shape of a tensor just call his get_shape() method like this:<\/p>\n\n<pre><code>print \"W shape:\", W.get_shape()\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>user1111929 in <a href=\"https:\/\/stackoverflow.com\/questions\/33712178\/tensorflow-nan-bug\">this<\/a> question use a debug print that help me assert where the problem come from.<\/li>\n<\/ol>\n"}]}
{"QuestionId":34224826,"AnswerCount":1,"Tags":"<python><neural-network><forecasting><tensorflow><skflow>","CreationDate":"2015-12-11T13:37:01.520","AcceptedAnswerId":34341900.0,"OwnerUserId":4842101.0,"Title":"skflow regression predict multiple values","Body":"<p>I'm trying to forecast a time series: given 50 previous values, I want to predict the 5 next values.<\/p>\n\n<p>To do so, I'm using the  package (based on TensorFlow), and this problem is relatively close to the <a href=\"https:\/\/github.com\/google\/skflow\/blob\/master\/examples\/boston.py\" rel=\"noreferrer\">Boston example provided in the Github repo<\/a>.<\/p>\n\n<p>My code is as follows :<\/p>\n\n<pre><\/pre>\n\n<p>This leads to :<\/p>\n\n<blockquote>\n  <p>ValueError: y_true and y_pred have different number of output (1!=5)<\/p>\n<\/blockquote>\n\n<p>at the end of the training. <\/p>\n\n<p>And when I try to predict, I have the same kind of problem<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>(200, 1) (200, 5)<\/p>\n<\/blockquote>\n\n<p>We can therefore see that the model is somehow predicting only 1 value instead of the 5 wanted\/hoped.<\/p>\n\n<p>How could I use the same model to predict values for several values ? <\/p>\n","answers":[{"AnswerId":"34341900","CreationDate":"2015-12-17T18:38:15.243","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>I've just added support for multi-output regression into <code>skflow<\/code> since this <a href=\"https:\/\/github.com\/google\/skflow\/commit\/e443c7343e7a625f3c35874fc87a5f6eb9934250\" rel=\"nofollow\">#e443c734<\/a>, so please reinstall the package are try again. If it doesn't work, please follow up on Github.<\/p>\n\n<p>I also added an example of multioutput regression to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/skflow\/multioutput_regression.py\" rel=\"nofollow\">examples folder<\/a>:<\/p>\n\n<pre><code># Create random dataset.\nrng = np.random.RandomState(1)\nX = np.sort(200 * rng.rand(100, 1) - 100, axis=0)\ny = np.array([np.pi * np.sin(X).ravel(), np.pi * np.cos(X).ravel()]).T\n\n# Fit regression DNN model.\nregressor = skflow.TensorFlowDNNRegressor(hidden_units=[5, 5])\nregressor.fit(X, y)\nscore = mean_squared_error(regressor.predict(X), y)\nprint(\"Mean Squared Error: {0:f}\".format(score))\n<\/code><\/pre>\n"}]}
{"QuestionId":34224954,"AnswerCount":2,"Tags":"<neural-network><tensorflow>","CreationDate":"2015-12-11T13:43:51.300","AcceptedAnswerId":null,"OwnerUserId":691867.0,"Title":"Why my VAE for toy dateset doesn't learn?","Body":"<p>I played with VAE (Variational Auto-encoder) for a couple of days. It works pretty well on the demo dataset like mnist or frayface. However, when I apply it to my own dataset, it doesn't work at all.<\/p>\n\n<p>I cannot find any reference regarding the potential difficulties in training VAE, so I made a tiny test case.<\/p>\n\n<p>My toy dataset is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>And my VAE structure is like:<\/p>\n\n<pre><\/pre>\n\n<p>The input is a probability, I hope after learning, VAE can figure out that the latent variable is actually X itself.<\/p>\n\n<p>However, I tried Bernoulli as well as Gaussian loss for the output layer, neither works.<\/p>\n\n<p>The code layer is stuck with the prior N(0,1), hence the prediction is always around 0.5.<\/p>\n\n<p>The toy script is here, <a href=\"https:\/\/gist.github.com\/colinfang\/20abd925ad3140236e63\" rel=\"nofollow\">https:\/\/gist.github.com\/colinfang\/20abd925ad3140236e63<\/a>. The majority of the code comes from <a href=\"https:\/\/jmetzen.github.io\/2015-11-27\/vae.html\" rel=\"nofollow\">https:\/\/jmetzen.github.io\/2015-11-27\/vae.html<\/a><\/p>\n","answers":[{"AnswerId":"58475919","CreationDate":"2019-10-20T18:17:33.733","ParentId":null,"OwnerUserId":"931831","Title":null,"Body":"<p>I'd suggest taking a VAE that works on moderately sized datasets (no mode collapse etc) and updating\/adapting for the kind of dataset you're passionate about. I've applied a <a href=\"https:\/\/becominghuman.ai\/variational-autoencoders-for-new-fruits-with-keras-and-pytorch-6d0cfc4eeabd\" rel=\"nofollow noreferrer\">VAE for new fruits<\/a> to a 1000 element dataset of furniture (recliners) I collected, with a modicum of success (shapes and some texture emulated, but not yet a truly notable imitation). <\/p>\n"},{"AnswerId":"56199781","CreationDate":"2019-05-18T14:15:09.807","ParentId":null,"OwnerUserId":"931115","Title":null,"Body":"<p>One very common problem that occur in VAEs is 'Posterior Collapse' (see here: <a href=\"https:\/\/datascience.stackexchange.com\/questions\/48962\/what-is-posterior-collapse-phenomenon\">https:\/\/datascience.stackexchange.com\/questions\/48962\/what-is-posterior-collapse-phenomenon<\/a>).<\/p>\n\n<p>If this is indeed the case, I suggest looking at the solution here:\n<a href=\"https:\/\/www.quora.com\/How-do-you-fix-a-Variational-Autoencoder-VAE-that-suffers-from-mode-collapse\" rel=\"nofollow noreferrer\">https:\/\/www.quora.com\/How-do-you-fix-a-Variational-Autoencoder-VAE-that-suffers-from-mode-collapse<\/a><\/p>\n\n<p>It worked out well for me.<\/p>\n"}]}
{"QuestionId":34228138,"AnswerCount":1,"Tags":"<python><pandas><theano><deep-learning><keras>","CreationDate":"2015-12-11T16:26:20.610","AcceptedAnswerId":34230175.0,"OwnerUserId":4437954.0,"Title":"Doesn't work example with Keras framework","Body":"<p>I am trying to study  library and created follow script as example:<\/p>\n\n<pre><\/pre>\n\n<p>In my opinion, this script should work :), however I got next error:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n<pre><\/pre>\n  \n  <p>by setting the Theano flag 'optimizer=fast_compile'. If that does not\n  work, Theano optimizations can be disabled with 'optimizer=None'.\n      HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.<\/p>\n<\/blockquote>\n\n<p>PS. Shapes data:<\/p>\n\n<ul>\n<li>Size of X_train (9L, 784L)<\/li>\n<li>Size of X_test (9L, 784L)<\/li>\n<\/ul>\n","answers":[{"AnswerId":"34230175","CreationDate":"2015-12-11T18:26:59.650","ParentId":null,"OwnerUserId":"1391220","Title":null,"Body":"<p>Check this line in your code<\/p>\n\n<pre><code>model.add(Dense(64, input_dim=20, init='uniform'))\n<\/code><\/pre>\n\n<p>Why 20 input dimensions? MNIST has 28X28 images, i.e. an input dimension of <code>784<\/code>. The error message confirms that as well:<\/p>\n\n<pre><code>ValueError: ('shapes (9,784) and (20,64) not aligned: 784 (dim 1) != 20 (dim 0)', (9L, 784L), (20L, 64L))\n<\/code><\/pre>\n\n<p>You can further verify the size of your input <\/p>\n\n<pre><code>print \"Size of X_train\", x_train.shape\nprint \"Size of X_test\", x_test.shape \n<\/code><\/pre>\n\n<p>And accordingly change the line above to:<\/p>\n\n<pre><code>model.add(Dense(64, input_dim=784, init='uniform'))\n<\/code><\/pre>\n"}]}
{"QuestionId":34229140,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><svm><tensorflow>","CreationDate":"2015-12-11T17:26:10.957","AcceptedAnswerId":34248291.0,"OwnerUserId":4341948.0,"Title":"Choosing from different cost function and activation function of a neural network","Body":"<p>Recently I started toying with neural networks. I was trying to implement an  gate with Tensorflow. I am having trouble understanding when to use different cost and activation functions. This is a basic neural network with only input and output layers, no hidden layers.<\/p>\n\n<p>First I tried to implement it in this way. As you can see this is a poor implementation, but I think it gets the job done, at least in some way. So, I tried only the real outputs, no one hot true outputs. For activation functions, I used a sigmoid function and for cost function I used squared error cost function (I think its called that, correct me if I'm wrong). <\/p>\n\n<p>I've tried using ReLU and Softmax as activation functions (with the same cost function) and it doesn't work. I figured out why they don't work. I also tried the sigmoid function with Cross Entropy cost function, it also doesn't work.<\/p>\n\n<pre><\/pre>\n\n<p>after 5000 iterations:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Question 1<\/strong> - Is there any other activation function and cost function, that can work(learn) for the above network, without changing the parameters(meaning without changing W, x, b). <\/p>\n\n<p><strong>Question 2<\/strong> - I read from a StackOverflow post <a href=\"https:\/\/stackoverflow.com\/questions\/20850895\/neural-networks-activation-function\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>[Activation Function] selection depends on the problem.<\/p>\n<\/blockquote>\n\n<p>So there are no cost functions that can be used anywhere? I mean there is no <em>standard<\/em> cost function that can be used on any neural network. Right? Please correct me on this.<\/p>\n\n<p><br\/><\/p>\n\n<p>I also implemented the  gate with a different approach, with the output as one-hot true. As you can see the   means that the 0th index is 1, so the answer is 0. I hope you get it. <\/p>\n\n<p>Here I have used a softmax activation function, with cross entropy as cost function. Sigmoid function as activation function fails miserably. <\/p>\n\n<pre><\/pre>\n\n<p>after 5000 iteration<\/p>\n\n<pre><\/pre>\n\n<p><strong>Question 3<\/strong> So in this case what cost function and activation function can I use? How do I understand what type of cost and activation functions I should use? Is there a standard way or rule, or just experience only? Should I have to try every cost and activation function in a brute force manner? I found an answer <a href=\"https:\/\/stackoverflow.com\/questions\/20368015\/activation-function-choice-for-neural-network\">here<\/a>. But I am hoping for a more elaborate explanation. <\/p>\n\n<p><strong>Question 4<\/strong> I have noticed that it takes many iterations to converge to a near accurate prediction. I think the convergance rate depends on the learning rate (using too large of will miss the solution) and the cost function (correct me if I'm wrong). So, is there any optimal way (meaning the fastest) or cost function for converging to a correct solution?<\/p>\n","answers":[{"AnswerId":"34248291","CreationDate":"2015-12-13T05:28:47.830","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>I will answer your questions a little bit out of order, starting with more general answers, and finishing with those specific to your particular experiment.<\/p>\n\n<p><strong>Activation functions<\/strong> Different activation functions, in fact, do have different properties. Let's first consider an activation function between two layers of a neural network. The only purpose of an activation function there is to serve as an nonlinearity. If you do not put an activation function between two layers, then two layers together will serve no better than one, because their effect will still be just a linear transformation. For a long while people were using sigmoid function and tanh, choosing pretty much arbitrarily, with sigmoid being more popular, until recently, when ReLU became the dominant nonleniarity. The reason why people use ReLU between layers is because it is non-saturating (and is also faster to compute). Think about the graph of a sigmoid function. If the absolute value of <code>x<\/code> is large, then the derivative of the sigmoid function is small, which means that as we propagate the error backwards, the gradient of the error will vanish very quickly as we go back through the layers. With ReLU the derivative is <code>1<\/code> for all positive inputs, so the gradient for those neurons that fired will not be changed by the activation unit at all and will not slow down the gradient descent.<\/p>\n\n<p>For the last layer of the network the activation unit also depends on the task. For regression you will want to use the sigmoid or tanh activation, because you want the result to be between 0 and 1. For classification you will want only one of your outputs to be one and all others zeros, but there's no differentiable way to achieve precisely that, so you will want to use a softmax to approximate it.<\/p>\n\n<p><strong>Your example<\/strong>. Now let's look at your example. Your first example tries to compute the output of <code>AND<\/code> in a following form:<\/p>\n\n<pre><code>sigmoid(W1 * x1 + W2 * x2 + B)\n<\/code><\/pre>\n\n<p>Note that <code>W1<\/code> and <code>W2<\/code> will always converge to the same value, because the output for (<code>x1<\/code>, <code>x2<\/code>) should be equal to the output of (<code>x2<\/code>, <code>x1<\/code>). Therefore, the model that you are fitting is:<\/p>\n\n<pre><code>sigmoid(W * (x1 + x2) + B)\n<\/code><\/pre>\n\n<p><code>x1 + x2<\/code> can only take one of three values (0, 1 or 2) and you want to return <code>0<\/code> for the case when <code>x1 + x2 &lt; 2<\/code> and 1 for the case when <code>x1 + x2 = 2<\/code>. Since the sigmoid function is rather smooth, it will take very large values of <code>W<\/code> and <code>B<\/code> to make the output close to the desired, but because of a small learning rate they can't get to those large values fast. Increasing the learning rate in your first example will increase the speed of convergence.<\/p>\n\n<p>Your second example converges better because the <code>softmax<\/code> function is good at making precisely one output be equal to <code>1<\/code> and all others to <code>0<\/code>. Since this is precisely your case, it does converge quickly. Note that <code>sigmoid<\/code> would also eventually converge to good values, but it will take significantly more iterations (or higher learning rate).<\/p>\n\n<p><strong>What to use<\/strong>. Now to the last question, how does one choose which activation and cost functions to use. These advices will work for majority of cases:<\/p>\n\n<ol>\n<li><p>If you do classification, use <code>softmax<\/code> for the last layer's nonlinearity and <code>cross entropy<\/code> as a cost function.<\/p><\/li>\n<li><p>If you do regression, use <code>sigmoid<\/code> or <code>tanh<\/code> for the last layer's nonlinearity and <code>squared error<\/code> as a cost function.<\/p><\/li>\n<li><p>Use ReLU as a nonlienearity between layers.<\/p><\/li>\n<li><p>Use better optimizers (<code>AdamOptimizer<\/code>, <code>AdagradOptimizer<\/code>) instead of <code>GradientDescentOptimizer<\/code>, or use momentum for faster convergence,<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34230613,"AnswerCount":2,"Tags":"<python><image><graph><structure><tensorflow>","CreationDate":"2015-12-11T18:55:05.927","AcceptedAnswerId":34231094.0,"OwnerUserId":678572.0,"Title":"Display image of graph in TensorFlow?","Body":"<p>I wrote a simple script to calculate the golden ratio from 1,2,5.  Is there a way to actually produce a visual through tensorflow (possibly with the aid of  or ) of the actual graph structure? The doc of tensorflow is pretty similar to a factor graph so I was wondering:<\/p>\n\n<p><strong>How can an image of the graph structure be generated through tensorflow?<\/strong><\/p>\n\n<p>In this example below, it would be  as individual nodes, and then  would have the  operation followed by the operation that brings them together. Maybe the graph structure (nodes,edges) can be imported into ? I see that the  objects have a  attribute but I haven't found out how to actually use this for imaging purposes. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34231094","CreationDate":"2015-12-11T19:26:49.477","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can get an image of the graph using <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"noreferrer\">Tensorboard<\/a>.  You need to edit your code to output the graph, and then you can launch tensorboard and see it.  See, in particular, <a href=\"https:\/\/www.tensorflow.org\/get_started\/graph_viz\" rel=\"noreferrer\">TensorBoard: Graph Visualization<\/a>.  You create a <code>SummaryWriter<\/code> and include the <code>sess.graph_def<\/code> in it.  The graph def will be output to the log directory.<\/p>\n"},{"AnswerId":"43569991","CreationDate":"2017-04-23T10:32:38.043","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This is exactly what tensorboard was created for. You need to slightly modify your code to store the information about your graph.<\/p>\n\n<pre><code>import tensorflow as tf\nC_1 = tf.constant(5.0)\nC_2 = tf.constant(1.0)\nC_3 = tf.constant(2.0)\n\ngolden_ratio = (tf.sqrt(C_1) + C_2)\/C_3\n\nwith tf.Session() as sess:\n    writer = tf.summary.FileWriter('logs', sess.graph)\n    print sess.run(golden_ratio)\n    writer.close()\n<\/code><\/pre>\n\n<p>This will create a <code>logs<\/code> folder with event files in your working directory. After this you should run tensorboard from your command line <code>tensorboard --logdir=\"logs\"<\/code> and navigate to the url it gives you (<a href=\"http:\/\/127.0.0.1:6006\" rel=\"noreferrer\">http:\/\/127.0.0.1:6006<\/a>). In your browser go to GRAPHS tab and enjoy your graph.<\/p>\n\n<p>You will use TB a lot if you are going to do anything with TF. So it makes sense to learn about it more from <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"noreferrer\">official<\/a> <a href=\"https:\/\/www.tensorflow.org\/get_started\/graph_viz\" rel=\"noreferrer\">tutorials<\/a> and from this <a href=\"https:\/\/www.youtube.com\/watch?v=eBbEDRsCmv4\" rel=\"noreferrer\">video<\/a>.<\/p>\n"}]}
{"QuestionId":34233767,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-12-11T22:44:37.703","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Event files in Google Tensorflow","Body":"<p>I am using Tensorflow to build up the Neural Network, and I would like to show training results on the Tensorboard. So far everything works fine. But I have a question on \"event file\" for the Tensorboard. I notice that every time when I run my python script, it generates different event files. And when I run my local server using \n, it shows up error if there are more than 1 event files. It seems to be annoying since whenever I run my local server, I have to delete all previous event files to make it work. So I'm wondering if there is any solution to prevent this issue. I would really appreciate it.<\/p>\n","answers":[{"AnswerId":"35853184","CreationDate":"2016-03-07T20:17:46.617","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>The best solution from a TensorBoard perspective is to have a root directory for your experiment, e.g. ~\/tensorflow\/mnist_experiment, and then to create a new subdirectory for each run, e.g. ~\/tensorflow\/mnist_experiment\/run1\/...<\/p>\n\n<p>Then run TensorBoard against the root directory, and every time you invoke your code, setup the SummaryWriter pointing to a new subdirectory. TensorBoard will then interpret all of the event files correctly, and it will also make it easy to compare between your different runs.<\/p>\n"}]}
{"QuestionId":34234350,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-12-11T23:41:28.937","AcceptedAnswerId":34307054.0,"OwnerUserId":5228890.0,"Title":"Accuracy issue in caffe","Body":"<p>I have a network which has 4 Boolean outputs. It is not a classification problem and each of them are meaningful. I expect to get a zero or one for each of them. Right now I have used the Euclidean loss function.<\/p>\n\n<p>There are 1000000 samples. In the input file, each of them have 144 features, so there the size of the input is 1000000*144.\nI have used batch size of 50, because otherwise the processing time is too much.\nThe output file is of the size 1000000*4, i.e. there are four output per each input. <\/p>\n\n<p>When I am using the accuracy layer, it complains about the dimension of output. It needs just one Boolean output, not four. I think it is because it considers the problem as a classification problem. \nI have two questions.\nFirst, considering the error of the accuracy layer, is the Euclidean loss function suitable for this task? And How I can get the accuracy for my network?\nSecond,I gonna get the exact value of the predicted output for each of the four variable. I mean I need the exact predicted values for each test record. Now, I just have the loss value for each batch. \nPlease guide me to solve those issues.<\/p>\n\n<p>Thanks,\nAfshin<\/p>\n\n<p>The train network is:<\/p>\n\n<pre><\/pre>\n\n<p>The test network is also:<\/p>\n\n<pre><\/pre>\n\n<p>And I get this error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Without using the accuracy layer caffe gives me the loss value.<\/p>\n","answers":[{"AnswerId":"34307054","CreationDate":"2015-12-16T08:24:25.477","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h3>Should <code>\"EuclideanLoss\"<\/code> be used for predicting binary outputs?<\/h3>\n\n<p>If you are trying to predict <em>discrete<\/em> binary labels then <code>\"EuclideanLoss\"<\/code> is not a very good choice. This loss is better  suited for regression tasks where you wish to predict continuous values (e.g., estimating coordinated of bounding boxes etc.).<br>\nFor predicting discrete labels, <code>\"SoftmaxWithLoss\"<\/code> or <a href=\"https:\/\/stackoverflow.com\/a\/30497907\/1714410\"><code>\"InfogainLoss\"<\/code><\/a> are better suited. Usually, <code>\"SoftmaxWithLoss\"<\/code> is used.<br>\nFor predicting binary outputs you may also consider <a href=\"https:\/\/stackoverflow.com\/a\/33142623\/1714410\"><code>\"SigmoidCrossEntropyLoss\"<\/code><\/a>.<\/p>\n\n<h3>Why is there an error in the <code>\"Accuracy\"<\/code> layer?<\/h3>\n\n<p>In <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a>, \"<code>Accuracy\"<\/code> layers expects two inputs (\"bottom\"s): one is a prediction <em>vector<\/em> and the other is the ground truth expected <em>discrete<\/em> label.\nIn your case, you need to provide, <em>for each binary output<\/em> a vector of length 2 with the predicted probabilities of 0 and 1, and a single binary label:<\/p>\n\n<pre><code>layer {\n  name: \"acc01\"\n  type: \"Accuracy\"\n  bottom: \"predict01\"\n  bottom: \"label01\"\n  top: \"acc01\"\n}\n<\/code><\/pre>\n\n<p>In this example you measure the accuracy for a <em>single<\/em> binary output. The input <code>\"predict01\"<\/code> is a two-vector for each example in the batch (for <code>batch_size: 50<\/code> the shape of this blob should be 50-by-2). <\/p>\n\n<h3>What can you do?<\/h3>\n\n<p>You are trying to predict 4 <em>different<\/em> outputs in a single net, therefore, you need <strong>4<\/strong> <em>different<\/em> loss and accuracy layers.<br>\nFirst, you need to split (<code>\"Slice\"<\/code>) the ground truth labels into 4 scalars (instead of a single binary 4-vector):<\/p>\n\n<pre><code>layer {\n  name: \"label_split\"\n  bottom: \"label\" # name of input 4-vector\n  top: \"label01\"\n  top: \"label02\"\n  top: \"label03\"\n  top: \"label04\"\n  type: \"Slice\"\n  slice_param {\n    axis: 1\n    slice_point: 1\n    slice_point: 2\n    slice_point: 3\n  }\n}\n<\/code><\/pre>\n\n<p>Now you have to have a prediction, loss and accuracy layer for each of the binary labels<\/p>\n\n<pre><code>layer {\n  name: \"predict01\"\n  type: \"InnerProduct\"\n  bottom: \"sig2\"\n  top: \"predict01\"\n  inner_product_param {\n    num_outout: 2 # because you need to predict 2 probabilities one for False, one for True\n    ...\n}\nlayer {\n  name: \"loss01\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"predict01\"\n  bottom: \"label01\"\n  top: \"loss01\"\n}\nlayer {\n  name: \"acc01\"\n  type: \"Accuracy\"\n  bottom: \"predict01\"\n  bottom: \"label01\"\n  top: \"acc01\"\n}\n<\/code><\/pre>\n\n<p>Now you need to replicate these three layer for each of the four binary labels you wish to predict.<\/p>\n"}]}
{"QuestionId":34235225,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-12-12T01:38:15.057","AcceptedAnswerId":34235413.0,"OwnerUserId":1576602.0,"Title":"Is it possible to modify an existing TensorFlow computation graph?","Body":"<p>TensorFlow graph is usually built gradually from inputs to outputs, and then executed. Looking at the Python code, the inputs lists of operations are immutable which suggests that the inputs should not be modified. Does that mean that there is no way to update\/modify an existing graph?<\/p>\n","answers":[{"AnswerId":"44353399","CreationDate":"2017-06-04T11:07:48.740","ParentId":null,"OwnerUserId":"2436776","Title":null,"Body":"<p>Yes, <code>tf.Graph<\/code> are build in an append-only fashion as @mrry puts it.<\/p>\n\n<p><strong>But there's workaround<\/strong>:<\/p>\n\n<p>Conceptually you can modify an existing graph by cloning it and perform the modifications needed along the way. As of r1.1, Tensorflow provides a module named <code>tf.contrib.graph_editor<\/code> which implements the above idea as a set of convinient functions.<\/p>\n"},{"AnswerId":"34235413","CreationDate":"2015-12-12T02:06:24.250","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph\"><code>tf.Graph<\/code><\/a> class is an <em>append-only<\/em> data structure, which means that you can add nodes to the graph after executing part of the graph, but you cannot remove or modify existing nodes. Since TensorFlow executes only the necessary subgraph when you call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\"><code>Session.run()<\/code><\/a>, there is no execution-time cost to having redundant nodes in the graph (although they will continue to consume memory).<\/p>\n\n<p>To remove <em>all<\/em> nodes in the graph, you can create a session with a new graph:<\/p>\n\n<pre><code>with tf.Graph().as_default():  # Create a new graph, and make it the default.\n  with tf.Session() as sess:  # `sess` will use the new, currently empty, graph.\n    # Build graph and execute nodes in here.\n<\/code><\/pre>\n"},{"AnswerId":"57528005","CreationDate":"2019-08-16T16:19:05.360","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>In addition to what @zaxily and @mrry says, I want to provide an example of how to actually do a modification to the graph. In short:<\/p>\n\n<ol>\n<li>one can not modify existing operations, all ops are final and non-mutable <\/li>\n<li>one may copy an op, modify it's inputs or attributes and add new op back to the graph<\/li>\n<li>all downstream ops that depend on the new\/copied op have to be recreated. Yes, a signifficant portion of the graph would be copied copied, which is not a problem<\/li>\n<\/ol>\n\n<p>The code:<\/p>\n\n<pre><code>import tensorflow\nimport copy\nimport tensorflow.contrib.graph_editor as ge\nfrom copy import deepcopy\n\na = tf.constant(1)\nb = tf.constant(2)\nc = a+b\n\ndef modify(t): \n    # illustrate operation copy&amp;modification\n    new_t = deepcopy(t.op.node_def)\n    new_t.name = new_t.name+\"_but_awesome\"\n    new_t = tf.Operation(new_t, tf.get_default_graph())\n    # we got a tensor, let's return a tensor\n    return new_t.outputs[0]\n\ndef update_existing(target, updated):\n    # illustrate how to use new op\n    related_ops = ge.get_backward_walk_ops(target, stop_at_ts=updated.keys(), inclusive=True)\n    new_ops, mapping = ge.copy_with_input_replacements(related_ops, updated)\n    new_op = mapping._transformed_ops[target.op]\n    return new_op.outputs[0]\n\nnew_a = modify(a)\nnew_b = modify(b)\ninjection = new_a+39 # illustrate how to add another op to the graph\nnew_c = update_existing(c, {a:injection, b:new_b})\n\nwith tf.Session():\n    print(c.eval()) # -&gt; 3\n    print(new_c.eval()) # -&gt; 42\n<\/code><\/pre>\n"}]}
{"QuestionId":34235557,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-12-12T02:26:40.097","AcceptedAnswerId":34243895.0,"OwnerUserId":1576602.0,"Title":"What is the purpose of graph collections in TensorFlow?","Body":"<p>The API discusses <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#graph-collections\">Graph Collections<\/a> which judging from the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L2071\">code<\/a> are a general purpose key\/data storage. What is the purpose of those collections?<\/p>\n","answers":[{"AnswerId":"47587634","CreationDate":"2017-12-01T06:14:46.167","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>I think there are at least two benefits for me so far: <\/p>\n\n<ol>\n<li>when you distribute your program on multiple GPUs or machines it is convenient to collect losses from different devices which are in the same collection. Use the tf.add_n to add them to accumulate the loss.   <\/li>\n<li>To update a particular set of variables like weights and biases in my own way. <\/li>\n<\/ol>\n\n<p>For instance:        <\/p>\n\n<pre><code>import tensorflow as tf    \nw = tf.Variable([1,2,3], collections=[tf.GraphKeys.WEIGHTS], dtype=tf.float32)    \nw2 = tf.Variable([11,22,32], collections=[tf.GraphKeys.WEIGHTS], dtype=tf.float32)\nweight_init_op = tf.variables_initializer(tf.get_collection_ref(tf.GraphKeys.WEIGHTS))\nsess = tf.InteractiveSession()\nweight_init_op.run()\nfor vari in tf.get_collection_ref(tf.GraphKeys.WEIGHTS): \n    tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, vari.assign(0.2 * vari))\nweight_update_ops = tf.get_collection_ref(tf.GraphKeys.UPDATE_OPS)\nfor op in weight_update_ops:\n    print(op.eval())\n<\/code><\/pre>\n\n<p>The output:  <\/p>\n\n<pre><code>[0.2 0.4 0.6]\n[2.2 4.4 6.4]\n<\/code><\/pre>\n"},{"AnswerId":"34243895","CreationDate":"2015-12-12T19:17:47.517","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Remember that under the hood, Tensorflow is a system for specifying and then executing computational data flow graphs.  The graph collections are used as part of keeping track of the constructed graphs and how they must be executed.  For example, when you create certain kinds of ops, such as <code>tf.train.batch_join<\/code>, the code that adds the op will also add some queue runners to the <code>QUEUE_RUNNERS<\/code> graph collection.  Later, when you call <code>start_queue_runners()<\/code>, by default, it will look at the <code>QUEUE_RUNNERS<\/code> collection to know which runners to start.<\/p>\n"}]}
{"QuestionId":34236252,"AnswerCount":4,"Tags":"<python><numpy><machine-learning><mean><tensorflow>","CreationDate":"2015-12-12T00:21:08.420","AcceptedAnswerId":34237210.0,"OwnerUserId":678572.0,"Title":"What is the difference between np.mean and tf.reduce_mean?","Body":"<p>In the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow noreferrer\">MNIST beginner tutorial<\/a>, there is the statement <\/p>\n\n<pre><\/pre>\n\n<p> basically changes the type of tensor the object is, but what is the difference between <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"nofollow noreferrer\"><\/a> and <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.13.0\/reference\/generated\/numpy.mean.html\" rel=\"nofollow noreferrer\"><\/a>? <\/p>\n\n<p>Here is the doc on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"nofollow noreferrer\"><\/a>:<\/p>\n\n<blockquote>\n  <p><\/p>\n  \n  <p>: The tensor to reduce. Should have numeric type.<\/p>\n  \n  <p>: The dimensions to reduce. If  (the defaut), reduces all dimensions.<\/p>\n\n<pre><\/pre>\n<\/blockquote>\n\n<p>For a 1D vector, it looks like , but I don't understand what's happening in .  kind of makes sense, since mean of  and  is , but what's going on with ?<\/p>\n","answers":[{"AnswerId":"43713062","CreationDate":"2017-05-01T01:19:23.120","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The new documentation states that <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"nofollow noreferrer\"><code>tf.reduce_mean()<\/code><\/a> produces the same results as np.mean:<\/p>\n\n<blockquote>\n  <p>Equivalent to np.mean<\/p>\n<\/blockquote>\n\n<p>It also has absolutely the same parameters as <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.mean.html\" rel=\"nofollow noreferrer\">np.mean<\/a>. But here is an important difference: they produce the same results <strong>only on float values<\/strong>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom random import randint\n\nnum_dims = 10\nrand_dim = randint(0, num_dims - 1)\nc = np.random.randint(50, size=tuple([5] * num_dims)).astype(float)\n\nwith tf.Session() as sess:\n    r1 = sess.run(tf.reduce_mean(c, rand_dim))\n    r2 = np.mean(c, rand_dim)\n    is_equal = np.array_equal(r1, r2)\n    print is_equal\n    if not is_equal:\n        print r1\n        print r2\n<\/code><\/pre>\n\n<p>If you will remove type conversion, you will see different results<\/p>\n\n<hr>\n\n<p>In additional to this, many other <code>tf.reduce_<\/code> functions such as <code>reduce_all<\/code>, <code>reduce_any<\/code>, <code>reduce_min<\/code>, <code>reduce_max<\/code>, <code>reduce_prod<\/code> produce the same values as there numpy analogs. Clearly because they are operations, they can be executed only from inside of the session.<\/p>\n"},{"AnswerId":"34236253","CreationDate":"2015-12-12T00:49:39.853","ParentId":null,"OwnerUserId":"2954547","Title":null,"Body":"<p><code>1<\/code> usually refers to rows, and <code>2<\/code> usually refers to columns. Reducing \"over\" index <code>1<\/code> means to reduce rowwise.<\/p>\n\n<p><code>[1., 2.]<\/code> is just <code>[ &lt;row 1 mean&gt; , &lt;row 2 mean&gt; ]<\/code>.<\/p>\n\n<p>This index numbering convention is typical in stats software, especially R.<\/p>\n"},{"AnswerId":"34237210","CreationDate":"2015-12-12T06:55:01.373","ParentId":null,"OwnerUserId":"4341948","Title":null,"Body":"<p>The functionality of <code>numpy.mean<\/code> and <code>tensorflow.reduce_mean<\/code> are the same. They do the same thing. From the documentation, for <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.mean.html\" rel=\"noreferrer\">numpy<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"noreferrer\">tensorflow<\/a>, you can see that. Lets look at an example,<\/p>\n\n<pre><code>c = np.array([[3.,4], [5.,6], [6.,7]])\nprint(np.mean(c,1))\n\nMean = tf.reduce_mean(c,1)\nwith tf.Session() as sess:\n    result = sess.run(Mean)\n    print(result)\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<pre><code>[ 3.5  5.5  6.5]\n[ 3.5  5.5  6.5]\n<\/code><\/pre>\n\n<p>Here you can see that when <code>axis<\/code>(numpy) or <code>reduction_indices<\/code>(tensorflow) is 1, it computes mean across (3,4) and (5,6) and (6,7), so <code>1<\/code> defines across which axis the mean is computed. When it is 0, the mean is computed across(3,5,6) and (4,6,7), and so on. I hope you get the idea.<\/p>\n\n<p>Now what are the differences between them? <\/p>\n\n<p>You can compute the numpy operation anywhere on python. But in order to do a tensorflow operation, it must be done inside a tensorflow <code>Session<\/code>. You can read more about it <a href=\"https:\/\/www.tensorflow.org\/get_started\/get_started\" rel=\"noreferrer\">here<\/a>. So when you need to perform any computation for your tensorflow graph(or structure if you will), it must be done inside a tensorflow <code>Session<\/code>.<\/p>\n\n<p>Lets look at another example.<\/p>\n\n<pre><code>npMean = np.mean(c)\nprint(npMean+1)\n\ntfMean = tf.reduce_mean(c)\nAdd = tfMean + 1\nwith tf.Session() as sess:\n    result = sess.run(Add)\n    print(result)\n<\/code><\/pre>\n\n<p>We could increase mean by <code>1<\/code> in <code>numpy<\/code> as you would naturally, but in order to do it in tensorflow, you need to perform that in <code>Session<\/code>, without using <code>Session<\/code> you can't do that. In other words, when you are computing <code>tfMean = tf.reduce_mean(c)<\/code>, tensorflow doesn't compute it then. It only computes that in a <code>Session<\/code>. But numpy computes that instantly, when you write <code>np.mean()<\/code>.<\/p>\n\n<p>I hope it makes sense.<\/p>\n"},{"AnswerId":"47242212","CreationDate":"2017-11-11T19:58:05.373","ParentId":null,"OwnerUserId":"2577292","Title":null,"Body":"<p>The key here is the word reduce, a concept from functional programming, which makes it possible for reduce_mean in TensorFlow to keep a running average of the results of computations from a batch of inputs. <\/p>\n\n<p>If you are not familiar with functional programming, this can seem mysterious. So first let us see what reduce does. If you were given a list like [1,2,5,4] and were told to compute the mean, that is easy - just pass the whole array to np.mean and you get the mean. However what if you had to compute the mean of a stream of numbers? In that case, you would have to first assemble the array by reading from the stream and then call np.mean on the resulting array - you would have to write some more code. <\/p>\n\n<p>An alternative is to use the reduce paradigm. As an example, look at how we can use reduce in python to calculate the sum of numbers: \n<code>reduce(lambda x,y: x+y, [1,2,5,4])<\/code>. <\/p>\n\n<p>It works like this:<\/p>\n\n<ol>\n<li>Step 1: Read 2 digits from the list - 1,2. Evaluate lambda 1,2. reduce stores the result 3. Note - this is the only step where 2 digits are read off the list<\/li>\n<li>Step 2: Read the next digit from the list - 5. Evaluate lambda 5, 3 (3 being the result from step 1, that reduce stored). reduce stores the result 8.<\/li>\n<li>Step 3: Read the next digit from the list - 4. Evaluate lambda 8,4 (8 being the result of step 2, that reduce stored). reduce stores the result 12<\/li>\n<li>Step 4: Read the next digit from the list - there are none, so return the stored result of 12.<\/li>\n<\/ol>\n\n<p>Read more here <a href=\"https:\/\/www.python-course.eu\/lambda.php\" rel=\"noreferrer\">Functional Programming in Python<\/a><\/p>\n\n<p>To see how this applies to TensorFlow, look at the following block of code, which defines a simple graph, that takes in a float and computes the mean. The input to the graph however is not a single float but an array of floats. The reduce_mean computes the mean value over all those floats. <\/p>\n\n<pre><code>import tensorflow as tf\n\n\ninp = tf.placeholder(tf.float32)\nmean = tf.reduce_mean(inp)\n\nx = [1,2,3,4,5]\n\nwith tf.Session() as sess:\n    print(mean.eval(feed_dict={inp : x}))\n<\/code><\/pre>\n\n<p>This pattern comes in handy when computing values over batches of images. Look at <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"noreferrer\">The Deep MNIST Example<\/a> where you see code like:<\/p>\n\n<pre><code>correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n<\/code><\/pre>\n"}]}
{"QuestionId":34237349,"AnswerCount":1,"Tags":"<python><computer-vision><neural-network><theano><conv-neural-network>","CreationDate":"2015-12-12T07:17:49.023","AcceptedAnswerId":null,"OwnerUserId":5431231.0,"Title":"Theano: Train theano neural net using data in CSV File","Body":"<p>I am doing image classification where i have to detect whether images contains airplane or not.<br>\nI am done with following steps:<br>\n1. Took features as descriptors from image dataset<br>\n2. Done with  K - means clustering and generated corpus of descriptors<br>\n3. Normalized data of corpus within the range of 0 - 1 and saved it in CSV format file<br>\nThe csv contains the data like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is there anyone who knows how to Train theano neural net using data in CSV File<\/p>\n","answers":[{"AnswerId":"34476729","CreationDate":"2015-12-27T01:02:39.913","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Well I think you first need to learn a little bit more about Theano. In a nutshell, in Theano you first define a mathematical model (i.e. a computational graph) which includes the relationships between variables and operations. Then you input your data into the model and get back the answer. <\/p>\n\n<p>Just a quick note about your data: remove <strong>imgX,<\/strong> from the beginning and <strong>,label<\/strong> at the end. Then create a list <strong>y<\/strong> with the length equal to the number of data points you have and put a <strong>1<\/strong> where there is an airplane and <strong>0<\/strong> where there is other things.You may take a look at <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg\" rel=\"nofollow\">this tutorial<\/a> to get a better understanding on how to define that computational graph I just talked about.<\/p>\n"}]}
{"QuestionId":34237632,"AnswerCount":1,"Tags":"<numpy><parallel-processing><theano>","CreationDate":"2015-12-12T07:58:06.057","AcceptedAnswerId":34237998.0,"OwnerUserId":3041068.0,"Title":"Create a tensor variable of given length in Theano functions","Body":"<p>For example:<\/p>\n\n<pre><\/pre>\n\n<p>What is the best way to do this <em>give_me_zeros_of_length<\/em> in Theano?<br>\nThank you in advance.<\/p>\n\n<p>Edit:<br>\nSo I've answered my own question, the answer is written clearly in the Theano tutorial document. Lazy me.<\/p>\n","answers":[{"AnswerId":"34237998","CreationDate":"2015-12-12T08:48:25.920","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>I've found the tensor.alloc() function from <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#creating-tensor\" rel=\"nofollow\">here<\/a> <\/p>\n\n<pre><code>x = tensor.scalar(dtype=\"int32\")\nret = tensor.alloc(0, x)\nmy_zeros = theano.function(inputs=[x], outputs=ret)\nmy_zeros(5)  # gives [0 0 0 0 0]\n<\/code><\/pre>\n\n<p>Note that by default the <code>alloc<\/code> function will take 0 as of type int8, can use <code>numpy.int32(0)<\/code> to convert it into a 32-bit 0.<\/p>\n"}]}
{"QuestionId":34239134,"AnswerCount":1,"Tags":"<numpy><parallel-processing><vectorization><theano>","CreationDate":"2015-12-12T11:08:13.840","AcceptedAnswerId":34249948.0,"OwnerUserId":3041068.0,"Title":"I want to map a function to each element of a vector in Theano, can I do it without using scan?","Body":"<p>Say a function that counts the appearances of ones at each index of an array:<\/p>\n\n<pre><\/pre>\n\n<p>As said <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#minimizing-scan-usage\" rel=\"nofollow\">here<\/a>, using scan may not be efficient. Plus, theano.scan always produces \"<em>RuntimeWarning: numpy.ndarray size changed, may indicate binary incompatibility from scan_perform.scan_perform import<\/em> *\" on my machine.  <\/p>\n\n<p>So I was wondering is there a better way of mapping functions in Theano?<br>\nThanks in advance.<\/p>\n\n<p>Edit:<br>\nI just realized it is a terrible example, apparently there's a more efficient way of just looping over the vector once like:<\/p>\n\n<pre><\/pre>\n\n<p>However according to @Daniel Renshaw's answer, since<\/p>\n\n<blockquote>\n  <p>the computation in one step is dependent on the same computation at\n  some earlier step<\/p>\n<\/blockquote>\n\n<p>so actually I can't avoid using scan in this regard, right?<\/p>\n\n<p>Edit:<br>\nI thought of a way of vercotrizing it as:<\/p>\n\n<pre><\/pre>\n\n<p>But in this case I have to know the size of the input in advance (unless I can craft numpy.tri like matrices on the fly?).<br>\nAny suggestions would be welcome. :)<\/p>\n\n<p>Edit:<br>\nI benchmarked the three methods using a 512D input array and 10000 iterations, and got the following results:<\/p>\n\n<ol>\n<li>map a sum function to each element: CPU 16s GPU 140s<\/li>\n<li>loop over the array using scan: CPU 13s GPU 32s<\/li>\n<li>vectorization: CPU 0.8s GPU 0.8s (actually I don't think theano has engaged GPU to do this<\/li>\n<\/ol>\n","answers":[{"AnswerId":"34249948","CreationDate":"2015-12-13T10:07:22.557","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>In the most general case, if no assumptions are made about the function, then scan would have to be used. However, many (maybe most?) useful functions can be vectorized such that scan is not needed. As is pointed out in the question edit, the example function can certainly be applied to the input without using scan.<\/p>\n\n<p>To decided whether scan is needed will depend on the function that needs to be applied. Case that certainly <em>will<\/em> require scan are those when the computation in one step is dependent on the same computation at some earlier step.<\/p>\n\n<p>P.S. The warning about binary incompatibility can be safely ignored.<\/p>\n"}]}
{"QuestionId":34239537,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2015-12-12T11:53:51.483","AcceptedAnswerId":null,"OwnerUserId":2890858.0,"Title":"How to Update Tensorflow from source","Body":"<p>I installed the latest  from source via git clone.\nand want to update to <\/p>\n\n<pre><\/pre>\n\n<p>but the Tensorflow library in the directory  still has the version 0.5.0<\/p>\n\n<p>the version in the result of  also is <\/p>\n","answers":[{"AnswerId":"43332670","CreationDate":"2017-04-10T20:50:00.753","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\n<\/code><\/pre>\n\n<p>first 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 $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>For more detail, follow this link in <a href=\"http:\/\/e-lin.github.io\/wiki\/jekyll\/update\/2017\/02\/18\/Upgrade-the-Installed-TensorFlow.html\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>If you get the error saying  <strong>not a supported wheel on this platform<\/strong>. You might be updating tensorflow for python3. For that you will need pip3\nTry installing pip3 <\/p>\n\n<pre><code>sudo apt-get -y install python3-pip\n<\/code><\/pre>\n\n<p>Then, set the path for download if you haven't already set the path<\/p>\n\n<pre><code>$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.1.0rc0-cp35-cp35m-linux_x86_64.whl\n\n$ pip3 install --ignore-installed --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n"},{"AnswerId":"38206191","CreationDate":"2016-07-05T14:35:30.443","ParentId":null,"OwnerUserId":"315046","Title":null,"Body":"<p>To show the version:<\/p>\n\n<pre><code>python -c \"import tensorflow; print(tensorflow.__version__);\"\n<\/code><\/pre>\n\n<p>And if it is not the latest, you have uninstall it via pip uninstall:<\/p>\n\n<pre><code>sudo pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>and subsequently install it:<\/p>\n\n<pre><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.9.0-cp27-none-linux_x86_64.whl\nsudo pip install $TF_BINARY_URL\n<\/code><\/pre>\n"},{"AnswerId":"37712500","CreationDate":"2016-06-08T20:40:26.907","ParentId":null,"OwnerUserId":"3824287","Title":null,"Body":"<p><code>git pull<\/code> doesn't work for me since some local files are modified by the last build so with a slight modification I update like this:<\/p>\n\n<pre><code>git fetch --all\ngit reset --hard origin\/master\n.\/configure\nbazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\nbazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\nsudo pip install \/tmp\/tensorflow_pkg\/tensorflow-0.8.0-py2-none-any.whl\n<\/code><\/pre>\n\n<p>Tested to work as of today. The Installation from Source instructions in tensorflow docs are misleading in the sense they only include the real pip wheel installation commands for Mac and the example-trainer build command exists instead in Linux instructions.<\/p>\n"},{"AnswerId":"34361587","CreationDate":"2015-12-18T18:11:00.003","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To install the TensorFlow library from source, you need to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#create-pip\" rel=\"noreferrer\">build a PIP package and install it<\/a>. The steps are as follows:<\/p>\n\n<pre><code>$ git pull\n$ .\/configure\n\n$ bazel build -c opt \/\/tensorflow\/tools\/pip_package:build_pip_package\n# ...or, with GPU support\n$ bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n\n$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n\n# The name of the .whl file will depend on your platform.\n$ pip install \/tmp\/tensorflow_pkg\/tensorflow-0.6.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"35716690","CreationDate":"2016-03-01T07:07:16.113","ParentId":null,"OwnerUserId":"6001116","Title":null,"Body":"<p>updating tensorflow install with <code>sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl<\/code>.\nI find it in the below issue,mohamed-ali's comment. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1105\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1105<\/a><\/p>\n"}]}
{"QuestionId":34239743,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-12T12:16:45.393","AcceptedAnswerId":null,"OwnerUserId":5671966.0,"Title":"Should I consider x0, threshold when using a Tensorflow placeholder function?","Body":"<p>I am currently doing a mood classification project by supervised learning, using tensor flow.<\/p>\n\n<p>And in machine learning theory, as you know, there is always an x0 which is +1. When making a placeholder for input dataset, is the function automatically produce a x0 part? or should I designate it manually?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34246821","CreationDate":"2015-12-13T00:52:21.813","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>There are two ways of thinking about <code>x0<\/code>. Either your input has an extra dimension, which always has <code>1<\/code> in it, and then a linear regression or a fully connected layer in a neural network will be represented as:<\/p>\n\n<pre><code>out = W * in\n<\/code><\/pre>\n\n<p>where <code>*<\/code> is matrix-vector multiplication, or, which is more common, to not add that extra dimension, and instead model it as<\/p>\n\n<pre><code>out = W * in + b\n<\/code><\/pre>\n\n<p>This is, in part, to highlight the difference between <code>W<\/code>, which is how we \"weight\" the input, and <code>b<\/code>, which is how much we \"shift\" it (<code>b<\/code> is called a \"bias\" term). One other reason why this representation is more desirable is because it is common to regularize <code>W<\/code>, but not <code>b<\/code>.<\/p>\n\n<p>Now, back to your question, TensorFlow neural network library models fully connected layer in terms of a weight matrix and a bias vector, therefore you do not need to add an extra one to your input vector.<\/p>\n\n<p>If you use low-level Tensor operations instead of the high-level predefined layers, then TensorFlow makes no assumptions about your input, and if you want to model your model in terms of operations on a vector with an extra <code>1<\/code> in it, it is your responsibility to add that <code>1<\/code> to that vector, TensorFlow will not do that for you.<\/p>\n"}]}
{"QuestionId":34240703,"AnswerCount":5,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-12-12T14:03:27.000","AcceptedAnswerId":34243720.0,"OwnerUserId":4341948.0,"Title":"What is logits, softmax and softmax_cross_entropy_with_logits?","Body":"<p>I was going through the tensorflow API docs <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#softmax\" rel=\"noreferrer\">here<\/a>. In the tensorflow documentation, they used a keyword called . What is it? In a lot of methods in the API docs it is written like<\/p>\n\n<pre><\/pre>\n\n<p>If what is written is those  are only , why keeping a different name like ? <\/p>\n\n<p>Another thing is that there are two methods I could not differentiate. They were<\/p>\n\n<pre><\/pre>\n\n<p>What are the differences between them? The docs are not clear to me. I know what  does. But not the other. An example will be really helpful.<\/p>\n","answers":[{"AnswerId":"34243720","CreationDate":"2015-12-12T19:01:22.017","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Logits simply means that the function operates on the unscaled output of earlier layers and that the relative scale to understand the units is linear.  It means, in particular, the sum of the inputs may not equal 1, that the values are <em>not<\/em> probabilities (you might have an input of 5).<\/p>\n\n<p><code>tf.nn.softmax<\/code> produces just the result of applying the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Softmax_function\" rel=\"noreferrer\">softmax function<\/a> to an input tensor.  The softmax \"squishes\" the inputs so that <code>sum(input) = 1<\/code>:  it's a way of normalizing.  The shape of output of a softmax is the same as the input: it just normalizes the values.  The outputs of softmax <em>can<\/em> be interpreted as probabilities.<\/p>\n\n<pre><code>a = tf.constant(np.array([[.1, .3, .5, .9]]))\nprint s.run(tf.nn.softmax(a))\n[[ 0.16838508  0.205666    0.25120102  0.37474789]]\n<\/code><\/pre>\n\n<p>In contrast, <code>tf.nn.softmax_cross_entropy_with_logits<\/code> computes the cross entropy of the result after applying the softmax function (but it does it all together in a more mathematically careful way).  It's similar to the result of:<\/p>\n\n<pre><code>sm = tf.nn.softmax(x)\nce = cross_entropy(sm)\n<\/code><\/pre>\n\n<p>The cross entropy is a summary metric: it sums across the elements.  The output of <code>tf.nn.softmax_cross_entropy_with_logits<\/code> on a shape <code>[2,5]<\/code> tensor is of shape <code>[2,1]<\/code> (the first dimension is treated as the batch).<\/p>\n\n<p>If you want to do optimization to minimize the cross entropy <strong>AND<\/strong> you're softmaxing after your last layer, you should use <code>tf.nn.softmax_cross_entropy_with_logits<\/code> instead of doing it yourself, because it covers numerically unstable corner cases in the mathematically right way.  Otherwise, you'll end up hacking it by adding little epsilons here and there.<\/p>\n\n<p><strong>Edited 2016-02-07:<\/strong> \nIf you have single-class labels, where an object can only belong to one class, you might now  consider using <code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code> so that you don't have to convert your labels to a dense one-hot array.  This function was added after release 0.6.0.<\/p>\n"},{"AnswerId":"34272341","CreationDate":"2015-12-14T16:47:11.653","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p><code>tf.nn.softmax<\/code> computes the forward propagation through a softmax layer. You use it during <strong>evaluation<\/strong> of the model when you compute the probabilities that the model outputs.<\/p>\n\n<p><code>tf.nn.softmax_cross_entropy_with_logits<\/code> computes the cost for a softmax layer. It is only used during <strong>training<\/strong>. <\/p>\n\n<p>The logits are the <em>unnormalized log probabilities<\/em> output the model (the values output before the softmax normalization is applied to them).<\/p>\n"},{"AnswerId":"56943911","CreationDate":"2019-07-09T00:14:17.063","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>What ever goes to <code>softmax<\/code> is logit, this is what J. Hinton repeats in coursera videos all the time.<\/p>\n"},{"AnswerId":"39499486","CreationDate":"2016-09-14T20:54:01.270","ParentId":null,"OwnerUserId":"4561314","Title":null,"Body":"<p><strong>Short version:<\/strong><\/p>\n\n<p>Suppose you have two tensors, where <code>y_hat<\/code> contains computed scores for each class (for example, from y = W*x +b) and <code>y_true<\/code> contains one-hot encoded true labels. <\/p>\n\n<pre><code>y_hat  = ... # Predicted label, e.g. y = tf.matmul(X, W) + b\ny_true = ... # True label, one-hot encoded\n<\/code><\/pre>\n\n<p>If you interpret the scores in <code>y_hat<\/code> as unnormalized log probabilities, then they are <strong>logits<\/strong>.<\/p>\n\n<p>Additionally, the total cross-entropy loss computed in this manner:<\/p>\n\n<pre><code>y_hat_softmax = tf.nn.softmax(y_hat)\ntotal_loss = tf.reduce_mean(-tf.reduce_sum(y_true * tf.log(y_hat_softmax), [1]))\n<\/code><\/pre>\n\n<p>is essentially equivalent to the total cross-entropy loss computed with the function <code>softmax_cross_entropy_with_logits()<\/code>:<\/p>\n\n<pre><code>total_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_hat, y_true))\n<\/code><\/pre>\n\n<p><strong>Long version:<\/strong><\/p>\n\n<p>In the output layer of your neural network, you will probably compute an array that contains the class scores for each of your training instances, such as from a computation <code>y_hat = W*x + b<\/code>. To serve as an example, below I've created a <code>y_hat<\/code> as a 2 x 3 array, where the rows correspond to the training instances and the columns correspond to classes. So here there are 2 training instances and 3 classes.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.Session()\n\n# Create example y_hat.\ny_hat = tf.convert_to_tensor(np.array([[0.5, 1.5, 0.1],[2.2, 1.3, 1.7]]))\nsess.run(y_hat)\n# array([[ 0.5,  1.5,  0.1],\n#        [ 2.2,  1.3,  1.7]])\n<\/code><\/pre>\n\n<p>Note that the values are not normalized (i.e. the rows don't add up to 1). In order to normalize them, we can apply the softmax function, which interprets the input as unnormalized log probabilities (aka <strong>logits<\/strong>) and outputs normalized linear probabilities. <\/p>\n\n<pre><code>y_hat_softmax = tf.nn.softmax(y_hat)\nsess.run(y_hat_softmax)\n# array([[ 0.227863  ,  0.61939586,  0.15274114],\n#        [ 0.49674623,  0.20196195,  0.30129182]])\n<\/code><\/pre>\n\n<p>It's important to fully understand what the softmax output is saying. Below I've shown a table that more clearly represents the output above. It can be seen that, for example, the probability of training instance 1 being \"Class 2\" is 0.619. The class probabilities for each training instance are normalized, so the sum of each row is 1.0.<\/p>\n\n<pre><code>                      Pr(Class 1)  Pr(Class 2)  Pr(Class 3)\n                    ,--------------------------------------\nTraining instance 1 | 0.227863   | 0.61939586 | 0.15274114\nTraining instance 2 | 0.49674623 | 0.20196195 | 0.30129182\n<\/code><\/pre>\n\n<p>So now we have class probabilities for each training instance, where we can take the argmax() of each row to generate a final classification. From above, we may generate that training instance 1 belongs to \"Class 2\" and training instance 2 belongs to \"Class 1\". <\/p>\n\n<p>Are these classifications correct? We need to measure against the true labels from the training set. You will need a one-hot encoded <code>y_true<\/code> array, where again the rows are training instances and columns are classes. Below I've created an example <code>y_true<\/code> one-hot array where the true label for training instance 1 is \"Class 2\" and the true label for training instance 2 is \"Class 3\".<\/p>\n\n<pre><code>y_true = tf.convert_to_tensor(np.array([[0.0, 1.0, 0.0],[0.0, 0.0, 1.0]]))\nsess.run(y_true)\n# array([[ 0.,  1.,  0.],\n#        [ 0.,  0.,  1.]])\n<\/code><\/pre>\n\n<p>Is the probability distribution in <code>y_hat_softmax<\/code> close to the probability distribution in <code>y_true<\/code>? We can use <a href=\"https:\/\/en.wikipedia.org\/wiki\/Cross_entropy\">cross-entropy loss<\/a> to measure the error.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rODko.png\"><img src=\"https:\/\/i.stack.imgur.com\/rODko.png\" alt=\"Formula for cross-entropy loss\"><\/a><\/p>\n\n<p>We can compute the cross-entropy loss on a row-wise basis and see the results. Below we can see that training instance 1 has a loss of 0.479, while training instance 2 has a higher loss of 1.200. This result makes sense because in our example above, <code>y_hat_softmax<\/code> showed that training instance 1's highest probability was for \"Class 2\", which matches training instance 1 in <code>y_true<\/code>; however, the prediction for training instance 2 showed a highest probability for \"Class 1\", which does not match the true class \"Class 3\".<\/p>\n\n<pre><code>loss_per_instance_1 = -tf.reduce_sum(y_true * tf.log(y_hat_softmax), reduction_indices=[1])\nsess.run(loss_per_instance_1)\n# array([ 0.4790107 ,  1.19967598])\n<\/code><\/pre>\n\n<p>What we really want is the total loss over all the training instances. So we can compute:<\/p>\n\n<pre><code>total_loss_1 = tf.reduce_mean(-tf.reduce_sum(y_true * tf.log(y_hat_softmax), reduction_indices=[1]))\nsess.run(total_loss_1)\n# 0.83934333897877944\n<\/code><\/pre>\n\n<p><strong>Using softmax_cross_entropy_with_logits()<\/strong><\/p>\n\n<p>We can instead compute the total cross entropy loss using the <code>tf.nn.softmax_cross_entropy_with_logits()<\/code> function, as shown below. <\/p>\n\n<pre><code>loss_per_instance_2 = tf.nn.softmax_cross_entropy_with_logits(y_hat, y_true)\nsess.run(loss_per_instance_2)\n# array([ 0.4790107 ,  1.19967598])\n\ntotal_loss_2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_hat, y_true))\nsess.run(total_loss_2)\n# 0.83934333897877922\n<\/code><\/pre>\n\n<p>Note that <code>total_loss_1<\/code> and <code>total_loss_2<\/code> produce essentially equivalent results with some small differences in the very final digits. However, you might as well use the second approach: it takes one less line of code and accumulates less numerical error because the softmax is done for you inside of <code>softmax_cross_entropy_with_logits()<\/code>.<\/p>\n"},{"AnswerId":"45183635","CreationDate":"2017-07-19T07:25:41.583","ParentId":null,"OwnerUserId":"5287756","Title":null,"Body":"<p>Above answers have enough description for the asked question.<\/p>\n\n<p>Adding to that, Tensorflow has optimised the operation of applying the activation function then calculating cost using its own activation followed by cost functions. Hence it is a good practice to use: <code>tf.nn.softmax_cross_entropy()<\/code> over <code>tf.nn.softmax(); tf.nn.cross_entropy()<\/code><\/p>\n\n<p>You can find prominent difference between them in a resource intensive model.<\/p>\n"}]}
{"QuestionId":34243843,"AnswerCount":3,"Tags":"<c++><tensorflow>","CreationDate":"2015-12-12T19:12:27.743","AcceptedAnswerId":null,"OwnerUserId":2663583.0,"Title":"C++ - Deleting an object before returning it","Body":"<p>I am going through the TensorFlow tutorials and while digging into the code I came across this:<\/p>\n\n<pre><\/pre>\n\n<p>The Register function declaration is the following:<\/p>\n\n<pre><\/pre>\n\n<p>The above snippet seems to be creating an  object with <\/p>\n\n<p>Then it destroys that object with  within the lambda function passed as argument to . And then it returns that same object(!). Assuming that the lambda function gets called within  (and it does get called according to my understanding) this does not make sense to me.<\/p>\n\n<p>I am no C++ beginner but I have never seen this kind of practice before. What am I missing here?<\/p>\n\n<p>The full .cpp file (including the  definition) is <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/core\/framework\/op.cc\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>Your help will be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"34244263","CreationDate":"2015-12-12T19:54:33.733","ParentId":null,"OwnerUserId":"1120273","Title":null,"Body":"<p>The constructor of <code>OpDef<\/code> sets the <code>initialized<\/code> member to <code>false<\/code>. However, the function object passed to <code>Register<\/code> is only called when this member is <code>true<\/code>. That is, the registration doesn't call the function <code>delete<\/code>ing the object.<\/p>\n\n<p>Supposedly, the function gets executed later at which point the object gets cleaned up. This partice looks somewhat questionable to me but without digging a lot deeper there doesn't seem to be an obvious error.<\/p>\n"},{"AnswerId":"34244206","CreationDate":"2015-12-12T19:48:27.653","ParentId":null,"OwnerUserId":"3410396","Title":null,"Body":"<p>It is hard to say without code for <code>Register<\/code>, but most likely it does not execute lambda passed to it right away. It probably stores it to execute later, as some kind of clean-up. <\/p>\n"},{"AnswerId":"34248020","CreationDate":"2015-12-13T04:34:31.293","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I'm about 90% convinced this is an actual bug.  I'm digging into it with the help of some of the folks who're more knowledgable about the op registry.  Will update.<\/p>\n\n<p>(I believe that what's happening is that in every use of the registry, <code>initialized_<\/code> is false -- i.e., the ops are all registered before any lookups or other operations are done on the registry.  This path is obviously correct because the op is thrown into the deferred queue, so the delete does not take effect until later.  However, it smells like a bug if an op is registered later.  This is what I'm checking on and trying to create a test case for.)<\/p>\n"}]}
{"QuestionId":34246955,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-13T01:16:13.657","AcceptedAnswerId":null,"OwnerUserId":2857219.0,"Title":"Running two tensorflow graphs at a time","Body":"<p>I have trained 2 models for slightly different OCR tasks and I would like to run them both at the same time in a production system. <\/p>\n\n<p>I have a class that starts a session and loads the saved model. The problem I am having is that the 2 models have variables saved with the same name so they are conflicting when I try to load them in.<\/p>\n\n<p>Is there a way to run them each in their own session or on their own graph?<\/p>\n","answers":[{"AnswerId":"34249940","CreationDate":"2015-12-13T10:06:32.753","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The simplest option would be to build the two models in different graphs and start separate sessions for each graph. However, each session owns its devices (CPU and possibly GPU), so you would have separate thread pools for each model, and this might cause suboptimal scheduling behavior. Careful use of the <code>tf.Session<\/code> configuration options would be necessary to get good performance.<\/p>\n\n<p>Alternatively, you could combine the two models in the same graph and use a single session. As you point out, the variables for the two models would necessarily have different names. Therefore, to make this work, you would need to provide an explicit name-to-<code>Variable<\/code> mapping when you <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver.__init__\" rel=\"nofollow\">construct the <code>tf.Saver<\/code><\/a> for loading in the model.<\/p>\n"}]}
{"QuestionId":34247374,"AnswerCount":4,"Tags":"<python><string><concatenation><tensorflow>","CreationDate":"2015-12-13T02:26:17.223","AcceptedAnswerId":null,"OwnerUserId":678572.0,"Title":"Merge string tensors in TensorFlow","Body":"<p>I work with a lot of  data.  I've been trying to build a simple graph as in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#SummaryWriter\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#SummaryWriter<\/a>.  <\/p>\n\n<p>For a simple operation, I wanted to concatenate strings together using a  as in (<a href=\"https:\/\/stackoverflow.com\/questions\/33810990\/how-to-feed-a-placeholder\">How to feed a placeholder?<\/a>)<\/p>\n\n<p><strong>Does anyone know how to merge string tensors together?<\/strong>  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34696282","CreationDate":"2016-01-09T16:58:16.210","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Thanks to your question, we prioritized adding support for string concatenation in TensorFlow, and added it in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/7116cc8b39c432f64e6e82b5708d2773f3fec159\" rel=\"noreferrer\">this commit<\/a>. String concatenation is implemented using the existing <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#add\" rel=\"noreferrer\"><code>tf.add()<\/code><\/a> operator, to match the behavior of <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/reference\/generated\/numpy.add.html\" rel=\"noreferrer\">NumPy's <code>add<\/code> operator<\/a> (including broadcasting).<\/p>\n\n<p>To implement your example, you can write:<\/p>\n\n<pre><code>complete = left + middle + right\n<\/code><\/pre>\n\n<p>\u2026or, equivalently, but if you want to name the resulting tensor:<\/p>\n\n<pre><code>complete = tf.add(tf.add(left, middle), right, name=\"COMPLETE\")\n<\/code><\/pre>\n\n<p>We have not yet added support for strings in <code>tf.add_n()<\/code> (or related ops like <code>tf.reduce_sum()<\/code>) but will consider this if there are use cases for it.<\/p>\n\n<p><strong>NOTE:<\/strong> To use this functionality immediately, you will need to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\" rel=\"noreferrer\">build TensorFlow from source<\/a>. The new op will be available in the next release of TensorFlow (0.7.0).<\/p>\n"},{"AnswerId":"34272445","CreationDate":"2015-12-14T16:52:18.560","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>I believe the sparse_concat op is what you are looking for:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#sparse_concat\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#sparse_concat<\/a><\/p>\n\n<p>add_n will add numeric values together.<\/p>\n"},{"AnswerId":"34656321","CreationDate":"2016-01-07T13:26:54.057","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>I know this is not an immediate answer and don't want this to remain hidden in the comments. <\/p>\n\n<p>If you'd like to incorporate an operation that isn't covered by the existing library, you can create a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/adding_an_op\/index.html\" rel=\"nofollow\">custom Op<\/a>. To incorporate your custom Op, you'll need to:  <\/p>\n\n<ul>\n<li>Register the new Op in a C++ file. The Op registration is independent of the implementation, and describes the semantics of how the Op is invoked. For example, it defines the Op name, and specifies its inputs and outputs. <\/li>\n<li>Implement the Op in C++. This implementation is called a \"kernel\", and there can be multiple kernels for different architectures (e.g. CPUs, GPUs) or input \/ output types.<\/li>\n<li>Create a Python wrapper. This wrapper is the public API to create the Op. A default wrapper is generated from the Op registration, which can be used directly or added to.<\/li>\n<li>Optionally, write a function to compute gradients for the Op.  <\/li>\n<li>Optionally, write a function that describes the input and output shapes for the Op. This allows shape inference to work with your Op.  <\/li>\n<li>Test the Op, typically in Python. If you define gradients, you can verify them with the Python GradientChecker.  <\/li>\n<\/ul>\n\n<p>What you asked if very relevant and will probably become one of the higher Google search results in the future for using string type with TensorFlow; as such this avenue to a solution needs to made available so that others are aware it exist. <\/p>\n"},{"AnswerId":"34621768","CreationDate":"2016-01-05T21:43:58.117","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>There is not a way to do this at the present time, AFAIK.<\/p>\n\n<p>Recognizing that this is a disappointing answer, I've filed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/701\" rel=\"nofollow\">issue #701 on GitHub<\/a> to track this request.<\/p>\n"}]}
{"QuestionId":34247661,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-13T03:25:51.880","AcceptedAnswerId":null,"OwnerUserId":2525299.0,"Title":"How to evaluate lists with predicted lengths? (`tensorflow.nn.top_k` with array of `k`s from another model)","Body":"<p>I am trying to predict medications given to patients. For each medication I have a column in the predictions (through softmax) indicating the probability that the patient will get this medication.<\/p>\n\n<p>But obviously people can get several meds at once, so I have another model that tries to predict the number of different medications given.<\/p>\n\n<p>I would like to evaluate them in a single TensorFlow call (I currently have a bunch of slow NumPy hacks), but I can't pass  an array of s (one for each patient, i.e. row), only a fixed integer - which doesn't work because different patients will get different numbers of meds.<br>\nUltimately I'm trying to  between the actually prescribed medication indices and the predicted ones. And then maybe the  of it.<\/p>\n\n<pre><\/pre>\n\n<p><em>Bonus question: Would it be possible to train a model directly on this instead of two seperate cross entropies? It doesn't really look differentiable to me - or do only the underlying softmaxes need to be differentiable?<\/em><\/p>\n","answers":[{"AnswerId":"34272580","CreationDate":"2015-12-14T16:59:29.977","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>The length of the predicted list is indeed not differentiable. You need to add an extra softmax output to the model predicting the length of the list, or add many sigmoid outputs predicting which entries should be included.<\/p>\n\n<p>I wrote a paper about transcribing variable-length text sequences from images, and the appendix goes into a lot of detail with a worked example for how the math works:\n<a href=\"http:\/\/arxiv.org\/abs\/1312.6082\" rel=\"nofollow\">http:\/\/arxiv.org\/abs\/1312.6082<\/a><\/p>\n"}]}
{"QuestionId":34249025,"AnswerCount":1,"Tags":"<matlab><dataset><neural-network><deep-learning><caffe>","CreationDate":"2015-12-13T07:47:12.150","AcceptedAnswerId":34261942.0,"OwnerUserId":2693257.0,"Title":"How to training\/testing my own dataset in caffe?","Body":"<p>I started with Caffe and the mnist example ran well.<br>\nI have the train and label data as . (I have 300 training data with 30 features and labels are  that have saved in ).<\/p>\n\n<p>However, I don't quite understand how I can use caffe to implement my own dataset?<\/p>\n\n<p>Is there a step by step tutorial can teach me?<\/p>\n\n<p>Many thanks!!!! Any advice would be appreciated!<\/p>\n","answers":[{"AnswerId":"34261942","CreationDate":"2015-12-14T07:44:20.740","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I think the most straight forward way to transfer data from Matlab to caffe is via HDF5 file.<\/p>\n\n<p>First, save your data in Matlab in an HDF5 file using <a href=\"http:\/\/www.mathworks.com\/help\/matlab\/ref\/hdf5write.html\" rel=\"nofollow noreferrer\"><code>hdf5write<\/code><\/a>. I assume your training data is stored in a variable name <code>X<\/code> of size 300-by-30 and the labels are stored in <code>y<\/code> a 300-by-1 vector:<\/p>\n\n<pre><code>hdf5write('my_data.h5', '\/X', \n  single( permute(reshape(X,[300, 30, 1, 1]),[4:-1:1]) ) );\nhdf5write('my_data.h5', '\/label', \n  single( permute(reshape(y,[300, 1, 1, 1]),[4:-1:1]) ), \n  'WriteMode', 'append' );\n<\/code><\/pre>\n\n<p>Note that the data is saved as a 4D array: the first dimension is the number of features, second one is the feature's dimension and the last two are 1 (representing no spatial dimensions). Also note that the names given to the data in the HDF5 are <code>\"X\"<\/code> and <code>\"label\"<\/code> - these names should be used as the <code>\"top\"<\/code> blobs of the input data layer.<\/p>\n\n<p>Why <code>permute<\/code>? please see <a href=\"https:\/\/stackoverflow.com\/a\/34431532\/1714410\">this answer<\/a> for an explanation.<\/p>\n\n<p>You also need to prepare a text file listing the names of all hdf5 files you are using (in your case, only <code>my_data.h5<\/code>). File <code>\/path\/to\/list\/file.txt<\/code> should have a single line <\/p>\n\n<blockquote>\n  <p>\/path\/to\/my_data.h5<\/p>\n<\/blockquote>\n\n<p>Now you can add an input data layer to your train_val.prototxt<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  name: \"data\"\n  top: \"X\"     # note: same name as in HDF5\n  top: \"label\" # \n  hdf5_data_param {\n    source: \"\/path\/to\/list\/file.txt\"\n    batch_size: 20\n  }\n  include { phase: TRAIN }\n}\n<\/code><\/pre>\n\n<p>For more information regarding hdf5 input layer, you can see in <a href=\"https:\/\/stackoverflow.com\/a\/31634295\/1714410\">this answer<\/a>.<\/p>\n"}]}
{"QuestionId":34249516,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-13T09:04:30.417","AcceptedAnswerId":34274877.0,"OwnerUserId":5674125.0,"Title":"does word2vec tutorial example imply potential sub-optimal implementation?","Body":"<p><strong>Machine Specs:<\/strong><\/p>\n\n<ul>\n<li>Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz<\/li>\n<li>16GB RAM<\/li>\n<li>nvidia TitanX<\/li>\n<\/ul>\n\n<p>(btw\nThere is a bug in word2vec example(both word2vec.py and word2vec_optimized.py) that it would run into error if no device directive is given. I add a with tf.device('\/cpu:0') to work around)  <\/p>\n\n<p><strong>Hyper-Parameter Setup:<\/strong><\/p>\n\n<ul>\n<li>the default model hyper-parameter + 8 concurrent threads<\/li>\n<li>use text8 as training data and questions-words.txt as test.<\/li>\n<\/ul>\n\n<p><strong>Profiling Result:<\/strong><\/p>\n\n<ul>\n<li><p>word2vec_optimized.py: 171891 words\/sec ( 8 threads)<\/p><\/li>\n<li><p>word2vec_optimized.py: 9258 words\/sec (8 threads)<\/p><\/li>\n<\/ul>\n\n<p>(the optimized version is comparable with Mikolov's original implementation, and the standard one is significantly slower)<\/p>\n\n<p>CPU utilization during training is roughly the same (the optimized version is a little bit higher)<\/p>\n\n<p>I reviewed the code of two the implementations, and I couldn't figure out why the standard implementation (word2vec.py) is 20x slower than the C++ hand-coded kernel(word2vec_kernels.cc)<\/p>\n\n<p>The word2vec_kernel.cc does not seem to do anything special to speed up (it first do a batch random sampling and then run a loop over samples in a single batch)<\/p>\n\n<p>The word2vec.py version doesn't seem to be sub-optimal to me either, it uses the standard tensorflow operations and the batching seems to be suitable.<\/p>\n\n<p>Have you guys ever do any internal profiling to figure out what is the performance bottleneck of the standard version? Would the performance bug be a threat to the training performance of other neural networks in general?<\/p>\n","answers":[{"AnswerId":"34274877","CreationDate":"2015-12-14T19:16:47.510","ParentId":null,"OwnerUserId":"5545462","Title":null,"Body":"<p>Have you guys ever do any internal profiling to figure out what is the performance bottleneck of the standard version?\nA: Yes. We did. The profiling led our decision to write the optimized version. <\/p>\n\n<p>Would the performance bug be a threat to the training performance of other neural networks in general?\nA: It's a complex question and the answer depends on scenario. I'd not make such a generalization. In many other models (at least for some that I played with), the computation is often dominated by \"heavy operations\" like matmul, convolution, etc. In those models, the loss and its gradient's computation is relatively cheap. On the other hand, word2vec is very special in that the training step is basically embedding lookup + loss's gradient + apply gradient, because these operations are not \"fused\", executing these operations incurs much higher memory bandwidth. TF plans to develop compiler optimization techniques to fuse such operations automatically and that will avoid , to some degree, the need to manually fuse operations to optimize performance.<\/p>\n"}]}
{"QuestionId":34250574,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2015-12-13T11:26:28.643","AcceptedAnswerId":null,"OwnerUserId":3537687.0,"Title":"Cannot import tensorflow in python after source build","Body":"<p>I am trying to install tensorflow with cuda and cudnn on a linux machine. I do not have sudo access, so I am building from source. I followed instructions here: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source<\/a><\/p>\n\n<p>I got till the part where we get lots of output:<\/p>\n\n<pre><\/pre>\n\n<p>But after that, when I open python and try to import tensorflow, it says there is no such module.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"34278292","CreationDate":"2015-12-14T23:09:32.243","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You need to take a few more steps before you can import TensorFlow in a Python shell: just building <code>\/\/tensorflow\/cc:tutorials_example_trainer<\/code> does not build any of the Python front-end.<\/p>\n\n<p>The easiest way to do this from a source installation is to follow the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/get_started\/os_setup.html#create-the-pip-package-and-install\" rel=\"nofollow\">instructions for building a PIP package from source<\/a>, and then installing it (either globally on your machine, or in a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/get_started\/os_setup.html#virtualenv_installation\" rel=\"nofollow\">virtualenv<\/a>).<\/p>\n"}]}
{"QuestionId":34251933,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><nolearn>","CreationDate":"2015-12-13T14:02:29.943","AcceptedAnswerId":null,"OwnerUserId":1097347.0,"Title":"nolearn error when trying to run an AutoEncoder","Body":"<p>I am trying to run a simple auto encoder using :<\/p>\n\n<pre><\/pre>\n\n<p>For some reason I encounter this error:<\/p>\n\n<blockquote>\n  <p>ValueError: bad input shape (10000, 500)<\/p>\n<\/blockquote>\n\n<p>Can anyone explain why this error occur, and how to solve it ?<\/p>\n","answers":[{"AnswerId":"34253326","CreationDate":"2015-12-13T16:22:54.613","ParentId":null,"OwnerUserId":"991339","Title":null,"Body":"<p>Usually the second dimension corresponds to features, while the first to instances. Try:<\/p>\n\n<pre><code>[x_train.shape[1], hidden_layer, x_train.shape[1]]\n<\/code><\/pre>\n"}]}
{"QuestionId":34254838,"AnswerCount":2,"Tags":"<python><concatenation><theano><hessian-matrix>","CreationDate":"2015-12-13T18:41:07.393","AcceptedAnswerId":null,"OwnerUserId":4868180.0,"Title":"Combining scalars and vectors in Theano for computing Hessian","Body":"<p>I'm trying to use Theano to compute the hessian of a function with respect to a vector as well as a couple scalars (edit: that is, I essentially want the scalars appended to the vector that I am computing the hessian with respect to). Here's a minimal example:<\/p>\n\n<pre><\/pre>\n\n<p>My first try was:<\/p>\n\n<pre><\/pre>\n\n<p>Which fails with <\/p>\n\n<p>My second try was to combine A, b, and c with:<\/p>\n\n<pre><\/pre>\n\n<p>Which fails with <\/p>\n\n<p>What is the correct way to compute the hessian in this case?<\/p>\n\n<p>Update: To clarify on what I am looking for, suppose A is a 2 element vector. Then the Hessian would be:<\/p>\n\n<pre><\/pre>\n\n<p>which for the example function  should be:<\/p>\n\n<pre><\/pre>\n\n<p>So if we were to define a function:<\/p>\n\n<pre><\/pre>\n\n<p>then, assuming we could compute  successfully, we would expect the output:<\/p>\n\n<pre><\/pre>\n\n<p>In my actual application, A has 25 elements and  is more complicated, but the idea is the same.<\/p>\n","answers":[{"AnswerId":"34270652","CreationDate":"2015-12-14T15:24:31.690","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>If you pass <code>b,c<\/code> as vectors, it should work. The hessian operator expects 1D arrays. Even though scalars should work, too, it is probably easiest to just provide the type of input it likes.<\/p>\n\n<p>The reason why your stacking fails is that the <code>stack<\/code> operation yields a new, non-endnode variable on a different branch of the graph with respect to which you can't generally take derivatives explicitly. So theano simply doesn't permit this.<\/p>\n\n<p>This works for me:<\/p>\n\n<pre><code>import theano.tensor as T\nA = T.vector('A')\nb,c = T.vectors('b','c')\ny = T.sum(A)*b[0]*c[0]\n\nhy = T.hessian(y,[A,b,c])\n<\/code><\/pre>\n"},{"AnswerId":"34275385","CreationDate":"2015-12-14T19:49:12.780","ParentId":null,"OwnerUserId":"4868180","Title":null,"Body":"<p>Based on a suggestion from @eickenberg to combine the inputs at the numpy level, I used the following workaround:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nA,temp = T.vectors('A','T')\nb,c = T.scalars('b','c')\n\ny = T.sum(A)*b*c\ny2 = theano.clone(y,{A:temp[:-2],b:temp[-2],c:temp[-1]})\n\nhy = T.hessian(y2,[temp])\nf = theano.function([temp], hy)\n\nf([1,1,4,5])\n<\/code><\/pre>\n\n<p>gives the expected output:<\/p>\n\n<pre><code>&gt; [array([[ 0.,  0.,  5.,  4.],\n&gt;         [ 0.,  0.,  5.,  4.],\n&gt;         [ 5.,  5.,  0.,  2.],\n&gt;         [ 4.,  4.,  2.,  0.]])]\n<\/code><\/pre>\n\n<p>This works but feels rather awkward, if anyone knows of a better (more general) solution please let me know.<\/p>\n"}]}
{"QuestionId":34258043,"AnswerCount":2,"Tags":"<python><neural-network><binaryfiles><tensorflow>","CreationDate":"2015-12-14T00:20:51.010","AcceptedAnswerId":null,"OwnerUserId":3780020.0,"Title":"Getting good mixing with many input datafiles in tensorflow","Body":"<p>I'm working with tensorflow hoping to train a deep CNN to do move prediction for the game Go. The dataset I created consists of 100,000 binary data files, where each datafile corresponds to a recorded game and contains roughly 200 training samples (one for each move in the game). I believe it will be very important to get good mixing when using SGD. I'd like my batches to contain samples from different games AND samples from different stages of the games. So for example simply reading one sample from the start of 100 files and shuffling isn't good b\/c those 100 samples will all be the first move of each game. <\/p>\n\n<p>I have read the tutorial on feeding data from files but I'm not sure if their provided libraries do what I need. If I were to hard code it myself I would basically initialize a bunch of file pointers to random locations within each file and then pull samples from random files, incrementing the file pointers accordingly.<\/p>\n\n<p>So, my question is does tensorflow provide this sort of functionality or would it be easier to write my own code for creating batches?<\/p>\n","answers":[{"AnswerId":"34258214","CreationDate":"2015-12-14T00:43:47.913","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Yes - what you want is to use a combination of two things.<\/p>\n\n<p>First, randomly shuffle the order in which you input your datafiles, by reading from them using a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"noreferrer\"><code>tf.train.string_input_producer<\/code><\/a> with <code>shuffle=True<\/code> that feeds into whatever input method you use (if you can put your examples into tf.Example proto format, that's easy to use with <code>parse_example<\/code>).  To be very clear, you put the list of filenames in the <code>string_input_producer<\/code> and then read them with another method such as <code>read_file<\/code>, etc.<\/p>\n\n<p>Second, you need to mix at a finer granularity.  You can accomplish this by feeding the input examples into a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#shuffle_batch\" rel=\"noreferrer\"><code>tf.train.shuffle_batch<\/code><\/a> node with a large capacity and large value of <code>min_after_dequeue<\/code>.  One particularly nice way is to use a <code>shuffle_batch_join<\/code> that receives input from multiple files, so that you get a lot of mixing.  Set the capacity of the batch big enough to mix well without exhausting your RAM.  Tens of thousands of examples usually works pretty well.<\/p>\n\n<p>Keep in mind that the batch functions add a <code>QueueRunner<\/code> to the <code>QUEUE_RUNNERS<\/code> collection, so you need to run <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#start_queue_runners\" rel=\"noreferrer\"><code>tf.train.start_queue_runners()<\/code><\/a><\/p>\n"},{"AnswerId":"44893357","CreationDate":"2017-07-03T20:19:28.407","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>In your case it is not a problem to do some preprocessing and create one file out of all the files you have. For this type of games, where the history is not important and the position determines everything your dataset can consist just from <code>position -&gt; next_move<\/code>.<\/p>\n\n<hr>\n\n<p>For a more broad case TF provides everything to allow the shuffling you want. There are two types shuffling which serve different purposes and shuffle different things:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/string_input_producer\" rel=\"nofollow noreferrer\"><code>tf.train.string_input_producer<\/code><\/a> <em>shuffle: Boolean. If true, the strings are randomly shuffled within each epoch.<\/em>. So if you have a few files <code>['file1', 'file2', ..., 'filen']<\/code> this randomly selects a file from this list. If case of false, the files follow one after each other.<\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/shuffle_batch\" rel=\"nofollow noreferrer\"><code>tf.train.shuffle_batch<\/code><\/a> <em>Creates batches by randomly shuffling tensors.<\/em> So it takes <code>batch_size<\/code> tensors from your queue (you will need to create a queue with <code>tf.train.start_queue_runners<\/code> ) and shuffles them.<\/li>\n<\/ul>\n"}]}
{"QuestionId":34258669,"AnswerCount":1,"Tags":"<eval><tensorflow><one-hot-encoding>","CreationDate":"2015-12-14T01:59:30.407","AcceptedAnswerId":34260164.0,"OwnerUserId":5676182.0,"Title":"Tensorflow : one hot encoding","Body":"<p>The following code works fine, but uses eval() which I think would be inefficient.  Is there a better way to achieve the same ?<\/p>\n\n<pre><\/pre>\n\n<p>output<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34260164","CreationDate":"2015-12-14T05:10:50.780","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>One way to achieve it is to compute max on each row and then compare each element to that value. I don't have tensor flow installed on this machine, so can't provide you with the exact code, but it will be along the lines of this:<\/p>\n\n<pre><code>z1 = tf.equal(t, tf.reduce_max(t, reduction_indices=[1], keep_dims=True))\n<\/code><\/pre>\n"}]}
{"QuestionId":34259877,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-14T04:39:47.110","AcceptedAnswerId":null,"OwnerUserId":5676466.0,"Title":"Error in running Torch\/Lua, maybe an installation error","Body":"<p>I installed Torch following <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">http:\/\/torch.ch\/docs\/getting-started.html<\/a>.<\/p>\n\n<p>However, when I run things like \"th\" or any simple code, I got the following error message: <\/p>\n\n<p>\/Users\/JianxuChen\/torch\/install\/bin\/luajit: ...rs\/JianxuChen\/torch\/install\/share\/lua\/5.1\/trepl\/init.lua:692: attempt to call field 'setheaptracking' (a nil value)\nstack traceback:\n    ...rs\/JianxuChen\/torch\/install\/share\/lua\/5.1\/trepl\/init.lua:692: in main chunk\n    [C]: in function 'require'\n    ...Chen\/torch\/install\/lib\/luarocks\/rocks\/trepl\/scm-1\/bin\/th:104: in main chunk\n    [C]: at 0x01089f8bc0<\/p>\n\n<p>Does anyone have ever encountered similar issue, or have any idea about how to solve it?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"34907067","CreationDate":"2016-01-20T17:49:21.030","ParentId":null,"OwnerUserId":"5801374","Title":null,"Body":"<p>I met the same problem and solved it in this way. Your torch is not the latest version. What I do is NOT follow this <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">http:\/\/torch.ch\/docs\/getting-started.html<\/a>. Instead download <a href=\"https:\/\/github.com\/torch\/distro\" rel=\"nofollow\">https:\/\/github.com\/torch\/distro<\/a> and run .\/install. Then it works!<\/p>\n"}]}
{"QuestionId":34264512,"AnswerCount":4,"Tags":"<macos><pip><tensorflow>","CreationDate":"2015-12-14T10:19:01.983","AcceptedAnswerId":null,"OwnerUserId":5492392.0,"Title":"Error setuptools when installing tensorflow","Body":"<p>I followed installation instruction at: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#pip_install\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#pip_install<\/a><\/p>\n\n<p>I am using python 2.7.11 (Anaconda) on Mac OS X 10.11. I checked but setuptools 18.5 is already installed on my computer.<\/p>\n\n<p>and met the error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34294493","CreationDate":"2015-12-15T16:31:24.407","ParentId":null,"OwnerUserId":"209852","Title":null,"Body":"<p>I had similar errors on Ubuntu<\/p>\n\n<blockquote>\n<pre><code>conda install -c https:\/\/conda.anaconda.org\/anaconda setuptools\n<\/code><\/pre>\n<\/blockquote>\n\n<p>solved the problem for me<\/p>\n"},{"AnswerId":"44614176","CreationDate":"2017-06-18T10:45:22.810","ParentId":null,"OwnerUserId":"5356660","Title":null,"Body":"<p>I ran into a similar problem today and solved it by running the following:<\/p>\n\n<pre><code>sudo -H pip install --upgrade tensorflow --ignore-installed\n<\/code><\/pre>\n"},{"AnswerId":"34277044","CreationDate":"2015-12-14T21:35:19.237","ParentId":null,"OwnerUserId":"1499416","Title":null,"Body":"<p>I had an identical problem and solved it by installing the python 3 version of tensorflow. So maybe check if there's a mismatch between your version of python and your version of tensorflow?<\/p>\n"},{"AnswerId":"35217266","CreationDate":"2016-02-05T05:59:29.973","ParentId":null,"OwnerUserId":"3297280","Title":null,"Body":"<p>I had the exact same issue (using Python 2.7.11 (Anaconda) on Mac OS X 10.11). <\/p>\n\n<p>Adding <code>--ignore-installed<\/code> flag worked for me (as discussed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/622\" rel=\"noreferrer\">here<\/a>).<\/p>\n\n<pre><code>$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl --ignore-installed\n\n$ python\n&gt;&gt;&gt; import tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":34265768,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2015-12-14T11:21:11.013","AcceptedAnswerId":34269302.0,"OwnerUserId":2505295.0,"Title":"What is a tensorflow float ref?","Body":"<p>Trying to run the following basic example to run a conditional calculation I got the following error message:<\/p>\n\n<blockquote>\n  <p>'x' was passed float incompatible with expected float_ref<\/p>\n<\/blockquote>\n\n<p>what is a tensorflow float_ref and how does the code have to be modified?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53867201","CreationDate":"2018-12-20T10:50:57.267","ParentId":null,"OwnerUserId":"10814608","Title":null,"Body":"<p>change<\/p>\n\n<pre><code>a = tf.Variable(tf.constant(0.),name=\"a\")\nb = tf.Variable(tf.constant(0.),name=\"b\")\nx = tf.Variable(tf.constant(0.),name=\"x\")\n\ndef add():\n    x.assign( a + b)\n    return x\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>    a = tf.placeholder(dtype=tf.float32,name=\"a\")\n    b = tf.placeholder(dtype=tf.float32,name=\"b\")\n    x = tf.placeholder(dtype=tf.float32,name=\"x\")\n    def add():\n        x= a + b\n        return x\n<\/code><\/pre>\n"},{"AnswerId":"34269302","CreationDate":"2015-12-14T14:22:21.453","ParentId":null,"OwnerUserId":"2505295","Title":null,"Body":"<p>this doesn't explain what a float_ref is, but it fixes the issues:<\/p>\n\n<p>1) variables need to be created in the session\n2) assignment op was not what we expected<\/p>\n\n<p>this fixed code works:<\/p>\n\n<pre><code>def add():\n    print(\"add\")\n    x = a + b\n    return x\n\ndef last():\n    print(\"last\")\n    return x\n\nwith tf.Session() as s:\n    a = tf.Variable(tf.constant(0.),name=\"a\")\n    b = tf.Variable(tf.constant(0.),name=\"b\")\n    x = tf.constant(-1.)\n    calculate= cond(x.eval()==-1.,add,last)\n    val = s.run([calculate], {a: 1., b: 2.})\n    print(val) # 3\n    print(s.run([calculate],{a:3.,b:4.})) # 7\n    print(val) # 3\n<\/code><\/pre>\n"},{"AnswerId":"49479286","CreationDate":"2018-03-25T18:20:33.640","ParentId":null,"OwnerUserId":"2110869","Title":null,"Body":"<p><code>float_ref<\/code> here refers to a reference to a float, i.e. your Tensorflow float variable <code>x<\/code>.<\/p>\n\n<p>As explained <a href=\"https:\/\/stackoverflow.com\/a\/49479197\/2110869\">here<\/a> you are facing this error because <strong>you can't simultaneously assign and pass a variable as a feed_dict<\/strong> in the same session run like you are doing in this statement:<\/p>\n\n<pre><code>val = s.run([calculate], {a: 1., b: 2., x: 0.})\n<\/code><\/pre>\n\n<p>It becomes more obvious when you resolve that statement to end up with:<\/p>\n\n<pre><code>val = s.run([x.assign( a + b)], {a: 1., b: 2., x: 0.})\n<\/code><\/pre>\n"},{"AnswerId":"47573648","CreationDate":"2017-11-30T12:34:56.383","ParentId":null,"OwnerUserId":"4827363","Title":null,"Body":"<p>FYI.\nI got a similar error and mine was:<\/p>\n\n<blockquote>\n  <p>node GradientDescent\/update_input\/ApplyGradientDescent was passed float from _arg_input_0_1:0 incompatible with expected float_ref.<\/p>\n<\/blockquote>\n\n<p>This happened because somewhere in my node-tree I had a <code>tf.Variable<\/code> instead of a <code>t.fplaceholder<\/code>. After replacing the variable with the placeholder, it worked.<\/p>\n"}]}
{"QuestionId":34268779,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-14T13:55:17.303","AcceptedAnswerId":null,"OwnerUserId":2505295.0,"Title":"How to keep calculated values in a Tensorflow graph (on the GPU)?","Body":"<p>How can we make sure that a calculated value will not be copied back to CPU\/python memory, but is still available for calculations in the next step?<\/p>\n\n<p>The following code obviously doesn't do it:<\/p>\n\n<pre><\/pre>\n\n<p>Error :  Attempting to use uninitialized value _recv_b_0<\/p>\n","answers":[{"AnswerId":"34270183","CreationDate":"2015-12-14T15:02:26.180","ParentId":null,"OwnerUserId":"2505295","Title":null,"Body":"<p>The answer is to store the value in a <code>tf.Variable<\/code> by storing to it using <a href=\"https:\/\/stackoverflow.com\/questions\/34220532\/how-to-assign-value-to-a-tensorflow-variable\">the assign operation<\/a>:<\/p>\n\n<p>working code:<\/p>\n\n<pre><code>import tensorflow as tf\nwith tf.Session() as s:\n    a = tf.Variable(tf.constant(1.),name=\"a\")\n    b = tf.Variable(tf.constant(2.),name=\"b\")\n    result = a + b\n    stored  = tf.Variable(tf.constant(0.),name=\"stored_sum\")\n    assign_op=stored.assign(result)\n    val,_ = s.run([result,assign_op],{a:1.,b:2.})\n    print(val) # 3\n    val=s.run(result,{a:4.,b:5.})\n    print(val[0]) # 9\n    print(stored.eval()) # ok, still 3 \n<\/code><\/pre>\n"}]}
{"QuestionId":34269672,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-12-14T14:38:38.023","AcceptedAnswerId":34270456.0,"OwnerUserId":5435604.0,"Title":"Theano batch learning in MLP, shape of bias","Body":"<p>I am confused about this. Say I have a weight matrix W(10,20). This expects a 20 dimensional input vector(say X). So, dot(W,X) will be (10,1) and to this I can add a 10 dimensional bias vector.<\/p>\n\n<p>Now I want to do batch learning. Is it possible to multiplt a batch of input vectors(30 vectors), so that X is now(20,30) and then dot(W,X) will be (10,30). But what about bias. how should I adjust bias to this ?<\/p>\n","answers":[{"AnswerId":"34270456","CreationDate":"2015-12-14T15:15:18.627","ParentId":null,"OwnerUserId":"1375015","Title":null,"Body":"<p>I assume the bias is the same for all 30 input vectors. Let me know if this is not the case.<\/p>\n\n<p>Think of your <code>(10,30)<\/code> matrix as a collection of 30 10-dimensional result vectors. They are independent, and they are just arranged into a matrix for easier handling.<\/p>\n\n<p>If you want to add the 10-dimensional bias vector to your 30 result vectors, all you have to do is to copy your 10-dimensional bias vector 30 times into a (10,30) bias matrix. \nThen, adding the matrices will be the same as adding the bias vector to each of your 30 10-dimensional result vectos<\/p>\n\n<p>You can obtain this by multiplying <\/p>\n\n<pre><code>&lt;vertical bias vector&gt;  &lt;outer\/tensor product&gt;  &lt;horizontal vector, filled with 30 ones&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":34271493,"AnswerCount":1,"Tags":"<python><csv><deep-learning><tensorflow>","CreationDate":"2015-12-14T16:04:06.503","AcceptedAnswerId":34400568.0,"OwnerUserId":3849791.0,"Title":"What is going wrong with the training and predictions using TensorFlow?","Body":"<p>Please see the code written below.<\/p>\n\n<pre><\/pre>\n\n<p>So here we see that there are 80 features in the data with only 2 possible outputs. I set the  and the  like so.<\/p>\n\n<pre><\/pre>\n\n<p>Initialize all variables.<\/p>\n\n<pre><\/pre>\n\n<p>Then I use this code to \"train\" my Neural Network.<\/p>\n\n<pre><\/pre>\n\n<p>After training the network, it always produces the same accuracy rate regardless of how many times I train it. That accuracy rate is  and I get to that accuracy with this code-<\/p>\n\n<pre><\/pre>\n\n<p>So this is where the question comes in. Is it because I have too small of dimensions? Maybe I should break the features into a 10x8 matrix? Maybe a 4x20 matrix? etc.<\/p>\n\n<p>Then I try to get the probabilities of the actual test data producing a 0 or a 1 like so-<\/p>\n\n<pre><\/pre>\n\n<p>And print out the probabilities:<\/p>\n\n<pre><\/pre>\n\n<p>(Note: it does produce  sometimes.)<\/p>\n\n<p>I then tried to see if applying the expert methodology would produce better results. Please see the following code.<\/p>\n\n<pre><\/pre>\n\n<p>(Please note how I changed the  in order to avoid errors).<\/p>\n\n<pre><\/pre>\n\n<p>Here is where the question comes in again. I define the Tensor (vector\/matrix if you will) as 80x1 (so 1 row with 80 features in it); I continue to do that throughout the rest of the code (please see below).<\/p>\n\n<pre><\/pre>\n\n<p><strong>Second Convolutional Layer<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Densely Connected Layer<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Dropout<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Readout<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>In the above you'll see that I defined the output as 2 possible answers (also to avoid errors).<\/p>\n\n<p>Then  and the .<\/p>\n\n<pre><\/pre>\n\n<p>Start the session.<\/p>\n\n<pre><\/pre>\n\n<p>\"Train\" the neural network.<\/p>\n\n<pre><\/pre>\n\n<p>And, once again, it always produces  as the output.<\/p>\n\n<p>The probabilities on the actual data producing a 0 or a 1 are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>As you see, there is <strong>some<\/strong> variance in these probabilities, but typically just the same result.<\/p>\n\n<p>I know that this isn't a Deep Learning problem. This is obviously a training problem. I know that there should always be some variance in the training accuracy every time you reinitialize the variables and retrain the network, but I just don't know why or where it's going wrong.<\/p>\n","answers":[{"AnswerId":"34400568","CreationDate":"2015-12-21T16:47:41.353","ParentId":null,"OwnerUserId":"3849791","Title":null,"Body":"<p>The answer is 2 fold.<\/p>\n\n<p>One problem is with the dimensions\/parameters. The other problem is that the features are being placed in the wrong spot.<\/p>\n\n<pre><code>W_conv1 = weight_variable([1, 2, 1, 80])\nb_conv1 = bias_variable([80])\n<\/code><\/pre>\n\n<p>Notice the first two numbers in the <code>weight_variable<\/code> correspond to the dimensions of the input. The second two numbers correspond to the dimensions of the feature tensor. The <code>bias_variable<\/code> always takes the final number in the <code>weight_variable<\/code>.<\/p>\n\n<p><strong>Second Convolutional Layer<\/strong><\/p>\n\n<pre><code>W_conv2 = weight_variable([1, 2, 80, 160])\nb_conv2 = bias_variable([160])\n<\/code><\/pre>\n\n<p>Here the first two numbers still correspond to the dimensions of the input. The second two numbers correspond to the amount of features and the weighted network that results from the 80 previous features. In this case, we double the weighted network. 80x2=160. The <code>bias_variable<\/code> then takes the final number in the <code>weight_variable<\/code>. If you were to finish the code at this point, the last number in the <code>weight_variable<\/code> would be a 1 in order to prevent dimensional errors due to the shape of the input tensor and the output tensor. But, instead, for better predictions, let's add a third convolutional layer.<\/p>\n\n<p><strong>Third Convolutional Layer<\/strong><\/p>\n\n<pre><code>W_conv3 = weight_variable([1, 2, 160, 1])\nb_conv3 = bias_variable([1])\n<\/code><\/pre>\n\n<p>Once again, the first two numbers in the <code>weight_variable<\/code> take the shape of the input. The third number corresponds to the amount of the weighted variables we established in the Second Convolutional Layer. The last number in the <code>weight_variable<\/code> now becomes 1 so we don't run into any dimension errors on the output that we are predicting. In this case, the output has the dimensions of <code>1, 2<\/code>.<\/p>\n\n<pre><code>W_fc2 = weight_variable([80, 1024])\nb_fc2 = bias_variable([1024])\n<\/code><\/pre>\n\n<p>Here, the number of neurons is <code>1024<\/code> which is completely arbitrary, but the first number in the <code>weight_variable<\/code> needs to be something that the dimensions of our feature matrix needs to be divisible by. In this case it can be any number (such as <code>2, 4, 10, 20, 40, 80<\/code>). Once again, the <code>bias_variable<\/code> takes the last number in the <code>weight_variable<\/code>.<\/p>\n\n<p>At this point, make sure that the last number in <code>h_pool3_flat = tf.reshape(h_pool3, [-1, 80])<\/code> corresponds to the first number in the <code>W_fc2<\/code> <code>weight_variable<\/code>.<\/p>\n\n<p>Now when you run your training program you will notice that the outcome varies and won't always guess all 1's or all 0's.<\/p>\n\n<p>When you want to predict the probabilities, you have to feed <code>x<\/code> to the <code>softmax<\/code> variable-> <code>y_conv=tf.nn.softmax(tf.matmul(h_fc2_drop, W_fc3) + b_fc3)<\/code> like so-<\/p>\n\n<pre><code>ans = sess.run(y_conv, feed_dict={x: x_test_actual, keep_prob: 1.0})\n<\/code><\/pre>\n\n<p>You can alter the <code>keep_prob<\/code> variable, but keeping it at a <code>1.0<\/code> always produces the best results. Now, if you print out <code>ans<\/code> you'll have something that looks like this-<\/p>\n\n<pre><code>[[ 0.90855026  0.09144982]\n [ 0.93020624  0.06979381]\n [ 0.98385173  0.0161483 ]\n [ 0.93948185  0.06051811]\n [ 0.90705943  0.09294061]\n [ 0.95702559  0.04297439]\n [ 0.95543593  0.04456403]\n [ 0.95944828  0.0405517 ]\n [ 0.99154049  0.00845954]\n [ 0.84375167  0.1562483 ]\n [ 0.98449463  0.01550537]\n [ 0.97772813  0.02227189]\n [ 0.98341942  0.01658053]\n [ 0.93026513  0.06973486]\n [ 0.93376994  0.06623009]\n [ 0.98026556  0.01973441]\n [ 0.93210858  0.06789146]\n<\/code><\/pre>\n\n<p>Notice how the probabilities vary. Your training is now working properly.<\/p>\n"}]}
{"QuestionId":34273951,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-14T18:20:20.140","AcceptedAnswerId":34278070.0,"OwnerUserId":3559888.0,"Title":"Tensorflow: Multi-GPU single input queue","Body":"<p>In <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/models\/image\/cifar10\" rel=\"noreferrer\">tensorflow's cifar10 multi-GPU example<\/a>, it seems (correct me if I am wrong) that one queue of training images is created per GPU. Wouldn't the \"right\" way of doing things be to have a single queue feeding all of the towers? If so, is there an example available of a shared queue?<\/p>\n","answers":[{"AnswerId":"34278070","CreationDate":"2015-12-14T22:50:14.210","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You're correct that the code for the CIFAR-10 model uses multiple input queues (through multiple calls to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv_install\" rel=\"noreferrer\"><code>cifar10.distorted_inputs()<\/code><\/a> via <code>cifar10.tower_loss()<\/code>).<\/p>\n\n<p>The easiest way to use a shared queue between the GPUs would be to do the following:<\/p>\n\n<ol>\n<li><p>Increase the batch size by a factor of N, where N is the number of GPUs.<\/p><\/li>\n<li><p>Move the call to <code>cifar10.distorted_inputs()<\/code> out of <code>cifar10.tower_loss()<\/code> and outside the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/6936918ad8abd7ade445674d7c912e69157a1017\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py#L178\" rel=\"noreferrer\">loop over GPUs<\/a>.<\/p><\/li>\n<li><p>Split the <code>images<\/code> and <code>labels<\/code> tensors that are returned from <code>cifar10.distorted_inputs()<\/code> along the 0th (batch) dimension:<\/p>\n\n<pre><code>images, labels = cifar10.distorted_inputs()\nsplit_images = tf.split(0, FLAGS.num_gpus, images)\nsplit_labels = tf.split(0, FLAGS.num_gpus, labels)\n<\/code><\/pre><\/li>\n<li><p>Modify <code>cifar10.tower_loss()<\/code> to take <code>images<\/code> and <code>labels<\/code> arguments, and invoke it as follows:<\/p>\n\n<pre><code>for i in xrange(FLAGS.num_gpus):\n  with tf.device('\/gpu:%d' % i):\n    with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME, i)) as scope:\n\n      loss = tower_loss(scope, split_images[i], split_labels[i])\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":34274262,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-12-14T18:40:36.960","AcceptedAnswerId":34380667.0,"OwnerUserId":1502718.0,"Title":"In theano, why am I getting this theano.gradient.DisconnectedInputError when it's clearly connected?","Body":"<pre><\/pre>\n\n<p>What does this mean? Below is my code an an explanation for why I think this error is vacuous.<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the  node is defined and used by the cost function. I don't see how the function  is discontinuous at all. Why is this breaking?<\/p>\n","answers":[{"AnswerId":"34380667","CreationDate":"2015-12-20T11:47:18.143","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that you need to assign the result of the <code>inc_subtensor<\/code> calls back to <code>R<\/code>.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>T.inc_subtensor(R[i:i+1][j:j+1], r(i, j))\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>T.inc_subtensor(R[i+1:i+2][i:i+1], sines[i])\n<\/code><\/pre>\n\n<p>use<\/p>\n\n<pre><code>R = T.inc_subtensor(R[i:i+1][j:j+1], r(i, j))\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>R = T.inc_subtensor(R[i+1:i+2][i:i+1], sines[i])\n<\/code><\/pre>\n\n<p><code>inc_subtensor<\/code> is a symbolic operation that returns an object representing the symbolic result of incrementing the provided subtensor by the provided value.<\/p>\n"}]}
{"QuestionId":34275060,"AnswerCount":2,"Tags":"<segmentation-fault><celery><theano>","CreationDate":"2015-12-14T19:27:45.050","AcceptedAnswerId":null,"OwnerUserId":3706442.0,"Title":"Theano and Celery : Worker exited prematurely: signal 11 (SIGSEGV)","Body":"<p>I'm building a web app where I start the training of a neural network implemented with theano on the server side by sending an ajax request from the client. Obviously I don't want to wait for the server to train the network entirely to send an answer back to my client because it would be way to long. <\/p>\n\n<p>So I came up with celery which enable me to do asynchronous code on the server side. I run the celery workers with the command . Unfortunately I get the following message each time a worker is running my task (which is training my network using theano) :<\/p>\n\n<pre><\/pre>\n\n<p>I have been searching why this error might occur and from what I understand the code I'm running is suffering memory leak. What I don't understand is why my code has no problem running when not using celery but this error come up when using Celery ?<\/p>\n\n<p>And most importantly I have no clue how to resolve this. I used lldb to look at the dump file generated and here is the backtrace I have :<\/p>\n\n<pre><\/pre>\n\n<p>I don't really know how to interpret this backtrace. Thanks in advance for any help !<\/p>\n","answers":[{"AnswerId":"34360736","CreationDate":"2015-12-18T17:17:58.557","ParentId":null,"OwnerUserId":"3706442","Title":null,"Body":"<p>If anyone run into the same problem turns out a workaround is to import theano library inline in the task and not at the module level.<\/p>\n\n<p>This way :<\/p>\n\n<pre><code>import baz\nimport bar\n\n@app.task\ndef foo():\n    import theano\n\n    # do something with theano\n<\/code><\/pre>\n\n<p>Check <a href=\"https:\/\/github.com\/celery\/celery\/issues\/2964#issuecomment-165669659\" rel=\"nofollow\">here<\/a> for more explanation<\/p>\n"},{"AnswerId":"41753954","CreationDate":"2017-01-20T00:08:21.213","ParentId":null,"OwnerUserId":"6867082","Title":null,"Body":"<p>FWIW, this also happens with <code>sklearn.cluster.KMeans<\/code>.  Works fine if I create threads on my own using <code>threading.Thread<\/code>.  Gets a sig11 if I attempt to call <code>fit<\/code> under Celery workers.<\/p>\n\n<p>I've not had the same issues with <code>sklearn.linear_model.LogisticRegression<\/code>, <code>Ridge<\/code> or <code>LinearRegression<\/code>.<\/p>\n"}]}
{"QuestionId":34276280,"AnswerCount":1,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2015-12-14T20:45:58.970","AcceptedAnswerId":null,"OwnerUserId":3850875.0,"Title":"AttributeError: 'module' object has no attribute 'MutableMapping'","Body":"<p>I followed the instructions for installing Google Tensorflow and its dependencies on an Ubuntu 14.04 g2.8xlarge aws instance. While trying to run the example problems, I'm running into the error posted below. Any help would be greatly appreciated. Thanks.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34277768","CreationDate":"2015-12-14T22:25:52.763","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This sounds like an incompatibility between TensorFlow and the version of Protocol Buffers that's installed on your machine. The two best options are:<\/p>\n\n<ol>\n<li><p>Try to upgrade the Protobuf library in <code>\/usr\/local\/lib\/python2.7\/dist-packages\/google\/protobuf\/<\/code> to version 3.0.0a3 or higher.<\/p><\/li>\n<li><p>Install TensorFlow in a virtualenv, by following the instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv_install\" rel=\"noreferrer\">here<\/a>. This should install the appropriate version of protobuf alongside TensorFlow.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34277938,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-14T22:39:29.440","AcceptedAnswerId":null,"OwnerUserId":5641162.0,"Title":"Stable results with TensorFlow","Body":"<p>What would be a good way to ensure TensorFlow results persist across\ndifferent runs? I can't find much information on e.g. seeding random weight initialization, so my results tend to differ with every run. <\/p>\n","answers":[{"AnswerId":"34278169","CreationDate":"2015-12-14T22:59:17.020","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The API method <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/constant_op.html#set_random_seed\" rel=\"noreferrer\"><code>tf.set_random_seed()<\/code><\/a> can be used to set a random seed that will be used in all TensorFlow random operations (including the usual random weight initializers and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#RandomShuffleQueue\" rel=\"noreferrer\">tf.RandomShuffleQueue<\/a>).<\/p>\n"}]}
{"QuestionId":34283090,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-15T07:13:55.840","AcceptedAnswerId":34298158.0,"OwnerUserId":1841986.0,"Title":"When using tensorboard, how to summarize a loss that is computed over several minibatches?","Body":"<p>I would like to use Tensorboard to visualize the evolution of the loss over a validation sample. But the validation set is too large to compute in one minibatch. Therefore, to compute my validation loss, I have to call session.run several times over several minibatches covering the validation set. Then I sum the loss (in python) of each minibatches to obtain the full validation loss.<\/p>\n\n<p>My problem is that tf.scalar_summary seems to have to be attached to a tensorflow node. But I would need to somehow \"attach\" it to the sum of the values of a node over several run of session.run.<\/p>\n\n<p>Is there a way to do that? Maybe by directly summarizing the python float that contains the sum of the minibatch losses? But I have not seen in the docs a way to \"summarize\" for tensorboard a python value that is outside of a computation. The example in the \"How-To\" section of the doc is only concerned with losses that can be computed in a single call to session.run.<\/p>\n","answers":[{"AnswerId":"34298158","CreationDate":"2015-12-15T19:51:10.163","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>You could add a <code>Variable<\/code> that is updated on each <code>sess.Run<\/code> call and have the summary track the value of the <code>Variable<\/code>.<\/p>\n"}]}
{"QuestionId":34283995,"AnswerCount":1,"Tags":"<anaconda><theano>","CreationDate":"2015-12-15T08:15:23.620","AcceptedAnswerId":null,"OwnerUserId":3392343.0,"Title":"Theano is missing modules","Body":"<p>I am on windows 64bit, I have installed anaconda, and managed to create an environment with python 2.7<\/p>\n\n<p>I have numpy, pylearn2, theano, and every package is built properly<\/p>\n\n<p>I have been able to import all these modules, however I get some very esoteric messages when I try to complete the model, like <\/p>\n\n<pre><\/pre>\n\n<p>Then I tried to actually find the package in the installation, but inside the cuda folder, there is no module named dnn. Looking at github, I see that it should be there.<\/p>\n\n<p>Why is theano missing modules? I installed using conda install theano, and it gave some suggestions, I have managed to pick the correct one. <\/p>\n\n<p>I have uninstalled and installed theano many time, I can import it but I can never get the proper modules.<\/p>\n\n<p>What is going wrong?<\/p>\n","answers":[{"AnswerId":"34300318","CreationDate":"2015-12-15T22:06:32.213","ParentId":null,"OwnerUserId":"3392343","Title":null,"Body":"<p>Ok, after a few days of search, it seems like Theano installed from anaconda is missing a lot of modules. However, installing theano by cloning the repository with <\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>seems to resolve the issue. Since windows normally does not have git, it can be easily installed (seems to take care of the environment's path variable) from here<\/p>\n\n<pre><code>https:\/\/git-scm.com\/download\/win\n<\/code><\/pre>\n"}]}
{"QuestionId":34286689,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2015-12-15T10:27:18.777","AcceptedAnswerId":null,"OwnerUserId":5681323.0,"Title":"How can I determine several labels in parallel (in a neural network) by using a softmax-output-layer in tensorflow?","Body":"<p>Due to the project work of my master study I am implementing a neural network using the tensorflow library form Google. At that I would like to determine (at the output layer of my feed forward neural network) several labels in parallel. And as activation function of the output layer I want to use the softmax function. \nSo what I want to have specifically is a output is a Vector that looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here the first three numbers are the probabilities of the three classes of the first classification and the other three numbers are the probabilities of the three classes of the second classification. So in this case I would say that the labels are: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In a first attempt I tried to implement this by first reshapeing the (1x6) vector to a (2x3) Matrix with tf.reshape(), then apply the softmax-function on the matrix tf.nn.softmax() and finally reshape the matrix back to a vector. Unfortunately, due to the reshaping, the Gradient-Descent-Optimizer gets problems with calculating the gradient, so I tried something different.<\/p>\n\n<p>What I do now is, I take the (1x6) vector and multiply it my a matrix that has a (3x3) identity-matrix in the upper part and a (3x3) zero-matrix in the lower part. Whit this I extract the first three entries of the vector. Then I can apply the softmax function and bring it back into the old form of (1x6) by another matrix multiplication. This has to be repeated for the other three vector entries as well.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>It works so far, but I don't like this solution.\nBecause in my real problem, I not only have to determine two labels at a time but 91, I would have to repeat the procedure form above 91-times.<\/p>\n\n<p>Does anyone have an solution, how I can obtain the desired vector, where the softmax function is applied on only three entries at a time, without writing the \"same\" code 91-times?  <\/p>\n","answers":[{"AnswerId":"34292496","CreationDate":"2015-12-15T15:02:14.837","ParentId":null,"OwnerUserId":"867505","Title":null,"Body":"<p>You could apply the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#split\" rel=\"nofollow\"><code>tf.split<\/code><\/a> function to obtain 91 tensors (one for each class), then apply softmax to each of them. <\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>classes_split = tf.split(0, 91, all_in_one)\n\nfor c in classes_split:\n    softmax_class = tf.nn.softmax(c)\n    # use softmax_class to compute some loss, add it to overall loss\n<\/code><\/pre>\n\n<p>or instead of computing the loss directly, you could also concatenate them together again:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>classes_split = tf.split(0, 91, all_in_one)\n\n# softmax each split individually\nclasses_split_softmaxed = [tf.nn.softmax(c) for c in classes_split]\n# Concatenate again\nall_in_one_softmaxed = tf.concat(0, classes_split_softmaxed)\n<\/code><\/pre>\n"}]}
{"QuestionId":34289131,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-15T12:19:02.347","AcceptedAnswerId":34292393.0,"OwnerUserId":3260688.0,"Title":"Basic neural network in TensorFlow","Body":"<p>I am trying to implement a very basic neural network in TensorFlow but I am having some problems. It is a very basic network that takes as input to values (hours or sleep and hours of study) and predicts the score on a test (I found this example on you-tube). So basically I have only one hidden layer with three units, each one computes an activation function (sigmoid) and the cost function is sum of square errors and I am using Gradient descent to minimize it. So the problem is, when I train the net with the training data and try to make some predictions using the same training data, the results do not quite match and they also appear strange because the look equal each other. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>It yields:<\/p>\n\n<blockquote>\n  <p>[[ 0.51873487]\n   [ 0.51874501]\n   [ 0.51873082]]<\/p>\n<\/blockquote>\n\n<p>and I believe it should be similar to the training data results. <\/p>\n\n<p>I am quite new to neural nets and machine learning so pardon me for any mistakes, thanks in advance. <\/p>\n","answers":[{"AnswerId":"34292393","CreationDate":"2015-12-15T14:57:16.483","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The main reason that your network isn't training is that the statement:<\/p>\n\n<pre><code>sess.run(train_op, feed_dict={X: trX, Y: trY})\n<\/code><\/pre>\n\n<p>\u2026only executes once. In TensorFlow, running <code>train_op<\/code> (or whatever operation is returned from <code>Optimizer.minimize()<\/code> will only cause the network to take a single gradient descent step. You should execute it in a loop to perform iterative training, and the weights will eventually converge.<\/p>\n\n<p>Two other tips: (i) you might achieve faster convergence if you feed a <em>subset<\/em> of your training data in each step, rather than the entire dataset; and (ii) the learning rate of 0.5 is probably too high (although this depends on the data).<\/p>\n"}]}
{"QuestionId":34290669,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-15T13:35:48.037","AcceptedAnswerId":34292411.0,"OwnerUserId":802158.0,"Title":"Traversing theano vector slice by slice","Body":"<p>Suppose I have a vector  with length 30. This vector is wrapped by a theano Tensor.<\/p>\n\n<p>I want to create a new vector, with same size as , but setting each 3 size slices elements to the sum of them and then do the same for the next 3 elements and so forth. <\/p>\n\n<p>In the example, the first 3-length  slice sum to 0.4, and the second 0.7 so .<\/p>\n\n<p>How can I do that? is using <a href=\"https:\/\/theano.readthedocs.org\/en\/rel-0.6rc3\/library\/scan.html\" rel=\"nofollow\"><\/a> the only way?<\/p>\n","answers":[{"AnswerId":"34292411","CreationDate":"2015-12-15T14:58:17.760","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can do the following in numpy:<\/p>\n\n<pre><code>import numpy as np\nx = np.array([0.1, 0.2, 0.1, 0.1, 0.2, 0.4])\ny = (x.reshape(-1, 3).mean(1)[:, np.newaxis] * np.ones(3)).ravel()\n\nprint(y)\n<\/code><\/pre>\n\n<p>In theano, you can proceed in a very similar way:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nxx = T.fvector()\nyy = (xx.reshape((-1, 3)).mean(axis=1).reshape((-1, 1)) * T.ones(3)).ravel()\n\nprint(yy.eval({xx: x.astype('float32')}))\n<\/code><\/pre>\n"}]}
{"QuestionId":34290933,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><mnist>","CreationDate":"2015-12-15T13:49:08.663","AcceptedAnswerId":null,"OwnerUserId":4882455.0,"Title":"Loss CaffeNet vs LeNet on MNIST dataset","Body":"<p>I want to train <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/train_val.prototxt\" rel=\"nofollow\">CaffeNet<\/a> on the MNIST dataset in Caffe. However, I noticed that after  iterations the loss just slightly dropped (from  to ). <\/p>\n\n<p>However, when I use <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">LeNet<\/a> on MNIST, the loss goes from  to , after  iterations.<\/p>\n\n<p>Does this happen because CaffeNet has more layers, and therefore needs more training time to converge? Or is it due to something else? I made sure the solver.prototxt of the nets were the same.<\/p>\n\n<p>While I know 100 iterations is extremely short (as CaffeNet usually trains for ~300-400k iterations), I find it odd that LeNet is able to get a loss so small, so soon.<\/p>\n","answers":[{"AnswerId":"34293928","CreationDate":"2015-12-15T16:07:09.777","ParentId":null,"OwnerUserId":"3633250","Title":null,"Body":"<p>I am not familiar with architecture of these nets, but in general there are several possible reasons:<\/p>\n\n<p>1) One of the nets is really much more complicated<\/p>\n\n<p>2) One of the nets was trained with a bigger learning rate<\/p>\n\n<p>3) Or maybe it used a training with momentum while other net didn't use it?<\/p>\n\n<p>4) Also possible that they both use momentum during training but one of them had the bigger momentum coefficient specified <\/p>\n\n<p>Really, there are tons of possible explanations for that.<\/p>\n"}]}
{"QuestionId":34292636,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-15T15:08:41.070","AcceptedAnswerId":34302284.0,"OwnerUserId":691867.0,"Title":"How to directly write to summary which mimics scalar_summary?","Body":"\n\n<p>I would like to achieve something like , but not used as a  that takes a . E.g. If I call <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In tensorboard events I expect it generates a plot named tag, with a line in it.<\/p>\n\n<p>How can I do that?<\/p>\n","answers":[{"AnswerId":"34298054","CreationDate":"2015-12-15T19:46:33.160","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>There is no requirement that the argument to <code>scalar_summary<\/code> is a <code>Variable<\/code>. It just needs to be a <code>Tensor<\/code>. You could make <code>value<\/code> a placeholder and feed many different values of it in inside your for loop.<\/p>\n"},{"AnswerId":"34302284","CreationDate":"2015-12-16T01:01:16.773","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>Alternatively, if you want to generate a TensorBoard log in pure Python code, you can do the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>summary_writer = tf.train.SummaryWriter(log_dir)\nfor i in 10000:\n    value = 0.2 * i\n    summary = tf.Summary(value=[tf.Summary.Value(tag=tag, simple_value=value)])\n    summary_writer.add_summary(summary, global_step=i)\nsummary_writer.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":34293714,"AnswerCount":8,"Tags":"<tensorflow>","CreationDate":"2015-12-15T15:57:34.047","AcceptedAnswerId":34297960.0,"OwnerUserId":3559888.0,"Title":"Can I measure the execution time of individual operations with TensorFlow?","Body":"<p>I know I can measure the execution time of a call to , but is it possible to get a finer granularity and measure the execution time of individual operations?<\/p>\n","answers":[{"AnswerId":"37669051","CreationDate":"2016-06-07T01:13:40.087","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>To update this answer, we do have some functionality for CPU profiling, focused on inference. If you look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/benchmark\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/benchmark<\/a> you'll see a program you can run on a model to get per-op timings.<\/p>\n"},{"AnswerId":"48416732","CreationDate":"2018-01-24T07:11:13.033","ParentId":null,"OwnerUserId":"231742","Title":null,"Body":"<p>Recently released by Uber SBNet custom op library (<a href=\"http:\/\/www.github.com\/uber\/sbnet\" rel=\"nofollow noreferrer\">http:\/\/www.github.com\/uber\/sbnet<\/a>) has an implementation of cuda event based timers, which can be used in the following manner:<\/p>\n\n<pre><code>with tf.control_dependencies([input1, input2]):\n    dt0 = sbnet_module.cuda_timer_start()\nwith tf.control_dependencies([dt0]):\n    input1 = tf.identity(input1)\n    input2 = tf.identity(input2)\n\n### portion of subgraph to time goes in here\n\nwith tf.control_dependencies([result1, result2, dt0]):\n    cuda_time = sbnet_module.cuda_timer_end(dt0)\nwith tf.control_dependencies([cuda_time]):\n    result1 = tf.identity(result1)\n    result2 = tf.identity(result2)\n\npy_result1, py_result2, dt = session.run([result1, result2, cuda_time])\nprint \"Milliseconds elapsed=\", dt\n<\/code><\/pre>\n\n<p>Note that any portion of subgraph can be asynchronous you should be very careful with specifying all the input and output dependencies for the timer ops. Otherwise the timer might get inserted into the graph out of order and you can get erroneous time. I found both the timeline and time.time() timing of very limited utility for profiling Tensorflow graphs. Also note that cuda_timer APIs will synchronize on default stream, which is currently by design because TF uses multiple streams.<\/p>\n\n<p>Having said this I personally recommend switching to PyTorch :) Development iteration is faster, code runs faster and everything is a lot less painful.<\/p>\n\n<p>Another somewhat hacky and arcane approach to subtracting the overhead from tf.Session (which can be enormous) is to replicate the graph N times and run it for a variable N, solving for an equation of unknown fixed overhead. I.e. you'd measure around session.run() with N1=10 and N2=20 and you know that your time is t and overhead is x. So something like<\/p>\n\n<pre><code>N1*x+t = t1\nN2*x+t = t2\n<\/code><\/pre>\n\n<p>Solve for x and t. Downside is this might require a lot of memory and is not necessarily accurate :) Also make sure that your inputs are completely different\/random\/independent otherwise TF will fold the entire subgraph and not run it N times... Have fun with TensorFlow :)<\/p>\n"},{"AnswerId":"37774470","CreationDate":"2016-06-12T12:52:50.410","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>I have used the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/client\/timeline.py\"><strong><code>Timeline<\/code><\/strong> object<\/a> to get the time of execution for each node in the graph:<\/p>\n\n<ul>\n<li>you use a classic <code>sess.run()<\/code> but also specify the optional arguments <code>options<\/code> and <code>run_metadata<\/code><\/li>\n<li>you then create a <code>Timeline<\/code> object with the <code>run_metadata.step_stats<\/code> data<\/li>\n<\/ul>\n\n<hr>\n\n<p>Here is an example program that measures the performance of a matrix multiplication:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow.python.client import timeline\n\nx = tf.random_normal([1000, 1000])\ny = tf.random_normal([1000, 1000])\nres = tf.matmul(x, y)\n\n# Run the graph with full trace option\nwith tf.Session() as sess:\n    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\n    run_metadata = tf.RunMetadata()\n    sess.run(res, options=run_options, run_metadata=run_metadata)\n\n    # Create the Timeline object, and write it to a json\n    tl = timeline.Timeline(run_metadata.step_stats)\n    ctf = tl.generate_chrome_trace_format()\n    with open('timeline.json', 'w') as f:\n        f.write(ctf)\n<\/code><\/pre>\n\n<hr>\n\n<p>You can then open Google Chrome, go to the page <code>chrome:\/\/tracing<\/code> and load the <code>timeline.json<\/code> file.\nYou should see something like:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qrPyb.png\"><img src=\"https:\/\/i.stack.imgur.com\/qrPyb.png\" alt=\"timeline\"><\/a><\/p>\n"},{"AnswerId":"50134629","CreationDate":"2018-05-02T12:26:13.313","ParentId":null,"OwnerUserId":"2598661","Title":null,"Body":"<p>As of Tensorflow 1.8, there's a really good example for using the <code>tf.profile.Profiler<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/profiler\/Profiler\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"46719332","CreationDate":"2017-10-12T21:42:53.593","ParentId":null,"OwnerUserId":"4063209","Title":null,"Body":"<p>For the comments of <em>fat-lobyte<\/em> under <em>Olivier Moindrot<\/em>'s answer, if you want to gather the timeline over all sessions, you can change \"<code>open('timeline.json', 'w')<\/code>\" to \"<code>open('timeline.json', 'a')<\/code>\". <\/p>\n"},{"AnswerId":"34297960","CreationDate":"2015-12-15T19:41:17.453","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>There is not yet a way to do this in the public release. We are aware that it's an important feature and we are working on it.<\/p>\n"},{"AnswerId":"43692312","CreationDate":"2017-04-29T05:57:48.743","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You can extract this information using <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/graph_viz#runtime_statistics\" rel=\"noreferrer\">runtime statistics<\/a>. You will need to do something like this (check the full example in the above-mentioned link):<\/p>\n\n<pre><code>run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\nrun_metadata = tf.RunMetadata()\nsess.run(&lt;values_you_want_to_execute&gt;, options=run_options, run_metadata=run_metadata)\nyour_writer.add_run_metadata(run_metadata, 'step%d' % i)\n<\/code><\/pre>\n\n<hr>\n\n<p>Better than just printing it you can see it in tensorboard:<\/p>\n\n<blockquote>\n  <p>Additionally, clicking on a node will display the exact total memory,\n  compute time, and tensor output sizes.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JdoXP.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JdoXP.png\" alt=\"[Example from link\"><\/a><\/p>\n"},{"AnswerId":"47841406","CreationDate":"2017-12-16T00:18:41.547","ParentId":null,"OwnerUserId":"4495833","Title":null,"Body":"<p>Since this is high up when googling for \"Tensorflow Profiling\", note that the current (late 2017, TensorFlow 1.4) way of getting the Timeline is using a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/ProfilerHook\" rel=\"noreferrer\">ProfilerHook<\/a>. This works with the MonitoredSessions in tf.Estimator where tf.RunOptions are not available.<\/p>\n\n<pre><code>estimator = tf.estimator.Estimator(model_fn=...)\nhook = tf.train.ProfilerHook(save_steps=10, output_dir='.')\nestimator.train(input_fn=..., steps=..., hooks=[hook])\n<\/code><\/pre>\n"}]}
{"QuestionId":34295136,"AnswerCount":3,"Tags":"<python><neural-network><deep-learning><caffe>","CreationDate":"2015-12-15T17:01:28.013","AcceptedAnswerId":34308240.0,"OwnerUserId":4613543.0,"Title":"ImportError: dynamic module does not define module export function (PyInit__caffe)","Body":"<p>I install caffe with python3,but when I import caffe, I get some errors\nTraceback (most recent call last):<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>But it work well in python2.7.<\/p>\n\n<p>I had add \/path\/to\/caffe\/distrubute\/python to the PATH, but when I make pycaffe, it shows that<\/p>\n\n<pre><\/pre>\n\n<p>How can I solve this problem?\nThank you very much.<\/p>\n","answers":[{"AnswerId":"45716502","CreationDate":"2017-08-16T14:36:49.180","ParentId":null,"OwnerUserId":"1781448","Title":null,"Body":"<p>It is now possible to build Caffe for Python3, and I'm almost sure it was possible in December 16 when the question was asked.<\/p>\n\n<p>To do this, you need to remove the comments in the Makefile.config with Python3:<\/p>\n\n<pre><code># Uncomment to use Python 3 (default is Python 2)\n# Check that boost library name is correct here!!!\nPYTHON_LIBRARIES := boost_python3 python3.5m\nPYTHON_INCLUDE := \/usr\/include\/python3.5m \\\n                 \/usr\/lib\/python3.5\/dist-packages\/numpy\/core\/include\n<\/code><\/pre>\n\n<p>But therefore you will have caffe only in python3 OR python2, because of the way how caffe installs (with PYTHON_PATH, not a good way indeed).<\/p>\n\n<p>To workaround this, you can do such trick in your ~\/.bashrc:<\/p>\n\n<pre><code>alias python2=\"export PYTHONPATH=$PYTHONPATH:\/home\/undead\/reps\/caffe_py2\/python &amp;&amp; python2\"\nalias python3=\"export PYTHONPATH=$PYTHONPATH:\/home\/undead\/reps\/caffe_py3\/python &amp;&amp; python3\"\nalias python=\"export PYTHONPATH=$PYTHONPATH:\/home\/undead\/reps\/caffe_py2\/python &amp;&amp; python2\"\n<\/code><\/pre>\n\n<p>Therefore both will works.<\/p>\n"},{"AnswerId":"54838762","CreationDate":"2019-02-23T06:07:17.507","ParentId":null,"OwnerUserId":"7062433","Title":null,"Body":"<p>make sure you are using the RIGHT python version.<\/p>\n\n<pre><code>import platform\nprint(platform.python_version())\n<\/code><\/pre>\n\n<p>I met the problem in Python3, and it worked Okay with Python2.<\/p>\n"},{"AnswerId":"34308240","CreationDate":"2015-12-16T09:27:35.640","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><strong>Update<\/strong><br>\nCaffe supports python 3.3+. \nPlease checkout installation guide and <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#prerequisites\" rel=\"nofollow noreferrer\">prerequisites<\/a>.<\/p>\n\n<p><strong>Original (outdated) answer<\/strong><br>\nUsing caffe with python 3 is not currently supported:<\/p>\n\n<blockquote>\n  <p>Caffe\u2019s Python interface works with Python 2.7. Python 3 or earlier Pythons are your own adventure.<\/p>\n<\/blockquote>\n\n<p>See <a href=\"http:\/\/tutorial.caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow noreferrer\">caffe's installation tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":34297488,"AnswerCount":0,"Tags":"<python><caffe><lstm><recurrent-neural-network>","CreationDate":"2015-12-15T19:12:35.640","AcceptedAnswerId":null,"OwnerUserId":2877875.0,"Title":"Request for example: Caffe RNN\/LSTM regression for Python","Body":"<p>I have been looking almost in every resource I can find on the web to see if someone posted an example of Regression using RNN\/LSTM in Caffe (<a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/caffe-users\/regression\/caffe-users\/_NdX1tpHt5Y\/gYPCLb_m45oJ\" rel=\"nofollow noreferrer\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/28057957\/multi-label-regression-in-caffe\">here<\/a>, and <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2033\" rel=\"nofollow noreferrer\">here<\/a>). Unfortunately, it looks like such resources don't exist till now. I'm working with <strong>Jeff Donahue<\/strong> version on python. <\/p>\n\n<p>What I'm looking for is something very simple. For example, if you have a <strong>100<\/strong> data points of <strong>(x,y)<\/strong> pairs. How would you go about:<\/p>\n\n<ol>\n<li>Creating the input matrix.<\/li>\n<li>Creating the contineuation matrix (do we need it or not?)<\/li>\n<li>Creating the target matrix.<\/li>\n<li>How the prototxt file will look like?<\/li>\n<li>Can we extrapolate (do prediction) from using this model?<\/li>\n<\/ol>\n\n<p>Also, how the first three items will be if the data is multi dimensional. For example <strong>X<\/strong> is a <strong>d<\/strong>-dimensional vector and <strong>Y<\/strong> is a <strong>k<\/strong>-dimensional vector.<\/p>\n\n<p>Feel free to use your own examples as long as they cover the steps of formatting the data under Python. <\/p>\n\n<p>I just wanted to note that I also opened a Caffe User's question for this <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/n1HFQBz3uQQ\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[]}
{"QuestionId":34300596,"AnswerCount":1,"Tags":"<image><image-processing><lua><transform><torch>","CreationDate":"2015-12-15T22:24:20.803","AcceptedAnswerId":34308024.0,"OwnerUserId":4013781.0,"Title":"Pad a Tensor in Torch","Body":"<p>I have a 100x100 pixel image in a Torch Tensor and I want to implement a \"zoom out\" transformation. How can I achieve this using the Torch <a href=\"https:\/\/github.com\/torch\/image\/blob\/master\/doc\/simpletransform.md\/\/\" rel=\"nofollow\">Image<\/a> toolbox (or other)? <\/p>\n\n<p>I have already implemented \"zoom in\" by simply using image.crop followed by image.resize. <\/p>\n\n<p>In Matlab, I would calculate the mean grayscale of the image, pad the array n pixels with that colour (keeping the original image centred), and then resize to 100x100 pixels. In there a \"pad Tensor\" function for Torch?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"34308024","CreationDate":"2015-12-16T09:17:50.177","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>Is there a \"pad Tensor\" function for Torch?<\/p>\n<\/blockquote>\n\n<p>One possibility is to use the <a href=\"https:\/\/github.com\/torch\/nn\/blob\/f58632d24de1b27a113f738c1b5a9dfc00a82921\/doc\/simple.md#nn.Padding\" rel=\"nofollow\">nn.Padding<\/a> module from <a href=\"https:\/\/github.com\/torch\/nn#neural-network-package\" rel=\"nofollow\">torch\/nn<\/a>, e.g.:<\/p>\n\n<pre><code>require 'image'\nrequire 'nn'\n\nlocal x = image.lena()\n\nlocal pad  = 64\nlocal pix  = 0\nlocal ndim = x:dim()\n\nlocal s = nn.Sequential()\n  :add(nn.Padding(ndim-1,  pad, ndim, pix))\n  :add(nn.Padding(ndim-1, -pad, ndim, pix))\n  :add(nn.Padding(ndim,    pad, ndim, pix))\n  :add(nn.Padding(ndim,   -pad, ndim, pix))\n\nlocal y = s:forward(x)\n\nimage.display(y) -- this requires qlua\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>As can be seen in the <a href=\"https:\/\/github.com\/torch\/nn\/blob\/efb781d\/Padding.lua#L20-L29\" rel=\"nofollow\">implementation<\/a> padding is obtained by:<\/p>\n\n<ol>\n<li>allocating an output tensor of the expected size post-padding filled with the padding color,<\/li>\n<li>filling the area corresponding of the input tensor with the original values thanks to <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/ad73aec\/doc\/tensor.md#self-narrowdim-index-size\" rel=\"nofollow\"><code>narrow<\/code><\/a>.<\/li>\n<\/ol>\n\n<p>Toy example:<\/p>\n\n<pre><code>require 'torch'\n\nlocal input  = torch.zeros(2, 5)\nlocal dim    = 2 -- target dimension for padding\nlocal pad    = 3 -- amount of padding\nlocal pix    = 1 -- pixel value (color)\n\n-- (1) compute the expected size post-padding, allocate a large enough tensor\n--     and fill with expected color\nlocal size   = input:size()\nsize[dim]    = size[dim] + pad\nlocal output = input.new():resize(size):fill(pix)\n\n-- (2) fill the original area with original values\nlocal area   = output:narrow(dim, 1, input:size(dim)):copy(input)\n<\/code><\/pre>\n\n<p>This gives as output:<\/p>\n\n<pre><code>0  0  0  0  0  1  1  1\n0  0  0  0  0  1  1  1\n[torch.DoubleTensor of size 2x8]\n<\/code><\/pre>\n\n<p>For specific zero-padding there are other convenient possibilities like: <\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/torch\/nn\/blob\/efb781d\/doc\/convolution.md#nn.SpatialZeroPadding\" rel=\"nofollow\"><code>nn.SpatialZeroPadding<\/code><\/a>,<\/li>\n<li><code>padzero<\/code> and <code>padmirror<\/code> from <a href=\"https:\/\/github.com\/koraykv\/fex\/blob\/9ba3300\/utils.lua#L96-L132\" rel=\"nofollow\">koraykv\/fex<\/a>.<\/li>\n<\/ul>\n"}]}
{"QuestionId":34300700,"AnswerCount":2,"Tags":"<deep-learning><tensorflow>","CreationDate":"2015-12-15T22:31:23.030","AcceptedAnswerId":34305091.0,"OwnerUserId":4544800.0,"Title":"Use of matrix multiplication function in TensorFlow","Body":"<p>I am a bit confused about the use of the function  in TensorFlow. My question might be more about the theory of deep learning, though. Say you have an input X and weight matrix W (assuming zero bias), I want to compute WX as an output which could be done by . However, in the tutorial <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"noreferrer\">MNIST for beginners<\/a> it is reversed and  is used instead. On the other hand, in the next tutorial <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/tf\/index.html#tensorflow-mechanics-101\" rel=\"noreferrer\">TensorFlow Mechanics 101<\/a>,  is used. Since the matrix sizes are important for multiplication I wonder if someone can clarify this issue.<\/p>\n","answers":[{"AnswerId":"34303002","CreationDate":"2015-12-16T02:31:38.263","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I think you must be misreading the mechanics 101 tutorial - or could you point to the specific line?<\/p>\n\n<p>In general, for a network layer, I think of the inputs \"flowing through\" the weights.  To represent that, I write <code>tf.matmul(Inputs, Weights)<\/code> to produce the output of that layer.  That output may then have a bias <code>b<\/code> added to it, and the result of that fed into a nonlinear function such as a relu, and then into another <code>tf.matmul<\/code> as the input for the next layer.<\/p>\n\n<p>Second, remember that the Weights matrix may be sized to produce multiple outputs.  That's why it's a matrix, not just a vector.  For example, if you wanted two hidden units and you had five input features, you would use a shape <code>[5, 2]<\/code> weight matrix, like this (shown in numpy for ease of exposition - you can do the same thing in tensorflow):<\/p>\n\n<pre><code>import numpy as np\na = np.array([1, 2, 3, 4, 5])\nW = np.array([[.5, .6], [.7, .8], [.9, .1], [.2, .3], [.4, .5]])\n\n&gt;&gt;&gt; np.dot(a, W)\narray([ 7.4,  6.2])\n<\/code><\/pre>\n\n<p>This has the nice behavior that if you then add a batch dimension to <code>a<\/code>, it still works:\n    a = np.array[[1, 2, 3, 4, 5],\n                 [6, 7, 8, 9, 0]]<\/p>\n\n<pre><code>&gt;&gt;&gt; np.dot(a, W)\narray([[  7.4,   6.2],\n       [ 20.9,  17.7]])\n<\/code><\/pre>\n\n<p>This is exactly what you're doing when you use tf.matmul to go from input features to hidden units, or from one layer of hidden units to another.<\/p>\n"},{"AnswerId":"34305091","CreationDate":"2015-12-16T06:14:14.600","ParentId":null,"OwnerUserId":"2282381","Title":null,"Body":"<p>I don't know much about TensorFlow, but intuitively I feel that the confusion is regarding the data representation of input. When you say you want to multiply an input <code>X<\/code> with a weight <code>W<\/code> I think what you mean is that you want to multiply each dimension (feature) with its corresponding weight and take the sum. So if you have an input <code>x<\/code> with say <code>m<\/code> dimensions, you should have a weight vector <code>w<\/code> with <code>m<\/code> values (<code>m+1<\/code> if you consider the bias).<\/p>\n\n<p>Now if you choose to represent the different training instances as rows of a matrix <code>X<\/code>, you would have to perform <code>X * w<\/code>, instead if you choose to represent them as columns, you would do <code>w^T * X<\/code><\/p>\n"}]}
{"QuestionId":34302042,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2015-12-16T00:34:12.690","AcceptedAnswerId":34302199.0,"OwnerUserId":894313.0,"Title":"TensorFlow LSTM Generative Model","Body":"\n\n<p>I'm working off the LSTM language model tutorial <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"nofollow noreferrer\">discussed here<\/a>.<\/p>\n\n<p>With language models, it's common to use the model to generate a new sentence from scratch after training (i.e. sample from the model).<\/p>\n\n<p>I'm new to TensorFlow but I'm trying to use my trained model to generate new words until the end-of-sentence marker.<\/p>\n\n<p>My initial attempt:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>It fails with error:<\/p>\n\n<blockquote>\n  <p>ValueError: setting an array element with a sequence.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34302199","CreationDate":"2015-12-16T00:52:07.517","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue here seems to be the <code>m.input_data: x<\/code> mapping in the <code>feed_dict<\/code> passed <code>session.run()<\/code>. In this case, TensorFlow expects that <code>x<\/code> is a numpy array (or some object that can be implicitly converted to a numpy array), but the value is a TensorFlow <code>Tensor<\/code> (the result of <code>tf.zeros_like()<\/code>).<\/p>\n\n<p>Fortunately, the solution is simple. Replace <code>x = tf.zeros_like(m.input_data)<\/code> with the following:<\/p>\n\n<pre><code>x = tf.zeros_like(m.input_data).eval()\n<\/code><\/pre>\n\n<p>...which ensures that <code>x<\/code> is converted to a numpy array.<\/p>\n\n<p>(Note that a more direct way to achieve this would be to construct the initial <code>x<\/code> as a numpy array of the appropriate size.)<\/p>\n"}]}
{"QuestionId":34311586,"AnswerCount":3,"Tags":"<python><neural-network><keras>","CreationDate":"2015-12-16T12:05:02.787","AcceptedAnswerId":36330550.0,"OwnerUserId":4393339.0,"Title":"Cannot train a neural network solving XOR mapping","Body":"<p>I am trying to implement a simple classifier for the XOR problem in Keras. Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>I tried changing the number of epochs, learning rate, and other parameters. But the error remains constant from the first to the last epoch.<\/p>\n\n<pre><\/pre>\n\n<p>How do you train this network in Keras?<\/p>\n\n<p>Also, is there a better library for implementing neural networks? I tried PyBrain, but it has been abandoned, tried scikit-neuralnetwork but the documentation is really cryptic so couldn't figure out how to train it. And I seriously doubt if Keras even works.<\/p>\n","answers":[{"AnswerId":"34719010","CreationDate":"2016-01-11T10:19:06.653","ParentId":null,"OwnerUserId":"3759736","Title":null,"Body":"<p>Try the last perceptron in the network without an activation function. I had the same problem and it starts learning when you remove the activation function.<\/p>\n\n<p>Also, you could try to split the output layer into 2 neurons. And have the output be [0,1] for 0 and [1,0] for one. <\/p>\n\n<p>However, removing the activation function should do the trick.<\/p>\n"},{"AnswerId":"36330550","CreationDate":"2016-03-31T10:17:23.560","ParentId":null,"OwnerUserId":"5661707","Title":null,"Body":"<p>In your example, you have a Dense layer with 1 unit with a softmax activation. The value of such a unit will always be 1.0, so no information can flow from your inputs to your outputs, and the network won't do anything. Softmax is only really useful when you need to generate a prediction of a probability among n classes, where n is greater than 2.<\/p>\n\n<p>The other answers suggest changes to the code to make it work. Just removing <code>activation='softmax'<\/code> may be enough.<\/p>\n\n<p>Keras does generally work.<\/p>\n"},{"AnswerId":"35615678","CreationDate":"2016-02-24T23:58:27.220","ParentId":null,"OwnerUserId":"973469","Title":null,"Body":"<p>This code works for me:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers.core import Activation, Dense\nfrom keras.optimizers import SGD\n\nX = np.array([[1, 1], [0, 0], [1, 0], [0, 1], [1, 1], [0, 0]], dtype='uint8')\ny = np.array([[0], [0], [1], [1], [0], [0]], dtype='uint8')\n\n\nmodel = Sequential()\nmodel.add(Dense(2, input_dim=2))\nmodel.add(Activation('sigmoid'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='mean_squared_error', optimizer=sgd, class_mode=\"binary\")\n\nhistory = model.fit(X, y, nb_epoch=10000, batch_size=4, show_accuracy=True)\n\nprint\nscore = model.evaluate(X,y)\nprint\nprint score\nprint model.predict(np.array([[1, 0]]))\nprint model.predict(np.array([[0, 0]]))\n\n# X vs y comparison\nprint\npredictions = model.predict(X)\npredictions = predictions.T\npredictions = [1 if prediction &gt;= 0.5 else 0 for prediction in predictions[0]]\nprint predictions\nprint [int(n) for n in y]\n<\/code><\/pre>\n\n<p>Unfortunatelly, I'm beginner in machine learning and I don't know why my code works and your doesn't.<\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/112#issuecomment-156311068\" rel=\"nofollow\">I used this code.<\/a><\/p>\n"}]}
{"QuestionId":34311893,"AnswerCount":2,"Tags":"<python><regression><tensorflow>","CreationDate":"2015-12-16T12:20:20.457","AcceptedAnswerId":null,"OwnerUserId":562769.0,"Title":"How can I train a simple, non-linear regression model with tensor flow?","Body":"<p>I've seen <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2%20-%20Basic%20Classifiers\/linear_regression.ipynb\" rel=\"noreferrer\">this example for linear regression<\/a> and I would like to train a model<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SsHgX.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SsHgX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where <a href=\"https:\/\/i.stack.imgur.com\/7T0wM.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7T0wM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h2>What I've tried<\/h2>\n\n<pre><\/pre>\n\n<p>When I execute it, I get:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried a couple of slight variations in the input data, but I can't get it to work.<\/p>\n\n<p><strong>How can I train such a simple, non-linear regression model with Google TensorFlow?<\/strong><\/p>\n","answers":[{"AnswerId":"34316040","CreationDate":"2015-12-16T15:35:30.640","ParentId":null,"OwnerUserId":"4511070","Title":null,"Body":"<p>The shape of your data (40 dimensions) is not compatible with the shape of the matrix your multiplying it with (10 dimensions). Try changing either size.<\/p>\n"},{"AnswerId":"34316689","CreationDate":"2015-12-16T16:04:16.773","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>InvalidArgumentError<\/code> is due to the values that you are feeding (<code>train_X<\/code> and <code>train_Y<\/code>) not having the necessary shape to be multiplied by <code>W1<\/code>.<\/p>\n\n<p>There are a few issues here:<\/p>\n\n<ol>\n<li><p>The statement <code>mul = X * W1<\/code> should be <code>mul = tf.matmul(X, W1)<\/code>, since <code>*<\/code> computes an elementwise multiplication, which is not what your equation specifies.<\/p><\/li>\n<li><p>The input data <code>X<\/code> should be a one-column matrix. To handle scalar and vector data - as you have in your feed calls, you could reshape it as follows:<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32)\nreshaped_X = tf.reshape(X, [-1, 1])\n# ...\nmul = reshaped_X * W1\n<\/code><\/pre><\/li>\n<li><p>When you fetch the final cost, the first argument to <code>sess.run<\/code> should be <code>l2_loss<\/code> (and not <code>cost<\/code>):<\/p>\n\n<pre><code>print(\"cost=\", sess.run(l2_loss, feed_dict={X: train_X, Y: train_Y}),\n      \"W1=\", sess.run(W1))\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":34312889,"AnswerCount":1,"Tags":"<python><python-2.7><deep-learning><tensorflow><tensorboard>","CreationDate":"2015-12-16T13:06:00.860","AcceptedAnswerId":null,"OwnerUserId":5686697.0,"Title":"Installing tensorflow causes an error on OS X?","Body":"<p>I install the tensorflow on Python 3.5 successfully. <\/p>\n\n<p>But I can't use the tensorboard due to the  which just support Python 2.x, \nthen I install tensorflow on Python 2.7 it causes these problems:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34316716","CreationDate":"2015-12-16T16:05:49.533","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you are having permissions issues or conflicts with other installed libraries, the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv_install\" rel=\"nofollow\">Virtualenv-based installation<\/a> is the easiest way to get TensorFlow installed.<\/p>\n"}]}
{"QuestionId":34313460,"AnswerCount":2,"Tags":"<numpy><theano><deep-learning>","CreationDate":"2015-12-16T13:34:42.860","AcceptedAnswerId":34359936.0,"OwnerUserId":4280916.0,"Title":"theano - use tensordot compute dot product of two tensor","Body":"<p>I want to use tensordot to compute the dot product of a specific dim of two tensors. Like:<\/p>\n\n<p>A is a tensor, whose shape is (3, 4, 5)\nB is a tensor, whose shape is (3, 5)<\/p>\n\n<p>I want to do a dot use A's third dim and B's second dim, and get a output whose dims is (3, 4)<\/p>\n\n<p>Like below:<\/p>\n\n<pre><\/pre>\n\n<p>How to do it by tensordot?<\/p>\n","answers":[{"AnswerId":"34359936","CreationDate":"2015-12-18T16:27:59.030","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Well, do you want this in numpy or in Theano? \nIn the case, where, as you state, you would like to contract axis 3 of A against axis 2 of B, both are straightforward:<\/p>\n\n<pre><code>import numpy as np\n\na = np.arange(3 * 4 * 5).reshape(3, 4, 5).astype('float32')\nb = np.arange(3 * 5).reshape(3, 5).astype('float32')\n\nresult = a.dot(b.T)\n<\/code><\/pre>\n\n<p>in Theano this writes as<\/p>\n\n<pre><code>import theano.tensor as T\n\nA = T.ftensor3()\nB = T.fmatrix()\n\nout = A.dot(B.T)\n\nout.eval({A: a, B: b})\n<\/code><\/pre>\n\n<p>however, the output then is of shape (3, 4, 3). Since you seem to want an output of shape (3, 4), the numpy alternative uses einsum, like so<\/p>\n\n<pre><code>einsum_out = np.einsum('ijk, ik -&gt; ij', a, b)\n<\/code><\/pre>\n\n<p>However, einsum does not exist in Theano. So the specific case here can be emulated as follows<\/p>\n\n<pre><code>out = (a * b[:, np.newaxis]).sum(2)\n<\/code><\/pre>\n\n<p>which can also be written in Theano<\/p>\n\n<pre><code>out = (A * B.dimshuffle(0, 'x', 1)).sum(2)\nout.eval({A: a, B: b})\n<\/code><\/pre>\n"},{"AnswerId":"34314504","CreationDate":"2015-12-16T14:24:09.873","ParentId":null,"OwnerUserId":"325565","Title":null,"Body":"<p>In this specific case, <code>einsum<\/code> is probably easier to understand than <code>tensordot<\/code>.  For example:<\/p>\n\n<pre><code>c = np.einsum('ijk,ik-&gt;ij', a, b)\n<\/code><\/pre>\n\n<p>I'm going to over-simplify the explanation a bit to make things more immediately understandable. We have two input arrays (separated by the comma) and this yields our output array (to the right of the <code>-&gt;<\/code>).<\/p>\n\n<ul>\n<li><code>a<\/code> has shape <code>3, 4, 5<\/code> and we'll refer to it as <code>ijk<\/code><\/li>\n<li><code>b<\/code> has shape <code>3, 5<\/code> (<code>ik<\/code>)<\/li>\n<li>We want the output <code>c<\/code> to have shape <code>3, 4<\/code> (<code>ij<\/code>)<\/li>\n<\/ul>\n\n<p>Seems a bit magical, right?  Let's break that down a bit.<\/p>\n\n<ul>\n<li>The letters we \"lose\" as we cross the <code>-&gt;<\/code> are axes that will be summed over.  That's what <code>dot<\/code> is doing, as well.<\/li>\n<li>We want output with shape <code>3, 4<\/code>, so we're eliminating <code>k<\/code><\/li>\n<li>Therefore, the output <code>c<\/code> should be <code>ij<\/code><\/li>\n<li>This means we'll refer to <code>b<\/code> as <code>ik<\/code>.  <\/li>\n<\/ul>\n\n<p>As a full example:<\/p>\n\n<pre><code>import numpy as np\n\na = np.random.random((3, 4, 5))\nb = np.random.random((3, 5))\n\n# Looping through things\nc1 = []\nfor i in range(3):\n    c1.append(a[i].dot(b[i]))\nc1 = np.array(c1)\n\n# Using einsum instead\nc2 = np.einsum('ijk,ik-&gt;ij', a, b)\n\nassert np.allclose(c1, c2)\n<\/code><\/pre>\n\n<p>You can do this with <code>tensordot<\/code> as well.  I'll add an example of that as soon as I have a bit more time.  (Of course, if anyone else would like to add a <code>tensordot<\/code> example as another answer in the meantime, feel free!)<\/p>\n"}]}
{"QuestionId":34313640,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2015-12-16T13:43:16.337","AcceptedAnswerId":null,"OwnerUserId":1661672.0,"Title":"luarocks install fails?","Body":"<p>I install torch following\n<a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow noreferrer\">http:\/\/torch.ch\/docs\/getting-started.html<\/a><\/p>\n\n<p>But when I use luarocks to install libs, like itorch, I command  <\/p>\n\n<pre><\/pre>\n\n<p>gives<\/p>\n\n<blockquote>\n  <p>Error: Your user does not have write permissions in\n  \/home\/myname\/code\/torch\/install\/lib\/luarocks\/rocks\n  -- you may want to run as a privileged user or use your local tree with --local.<\/p>\n<\/blockquote>\n\n<p>Then I cmd  <\/p>\n\n<pre><\/pre>\n\n<p>gives  <\/p>\n\n<blockquote>\n  <p>sudo: luarocks: command not found<\/p>\n<\/blockquote>\n\n<p>Then I cmd  <\/p>\n\n<pre><\/pre>\n\n<p>gives  <\/p>\n\n<blockquote>\n  <p>Configuration files:<br>\n  System: \/home\/myname\/code\/torch\/install\/etc\/luarocks\/config.lua (ok)<br>\n  User  : \/home\/myname\/.luarocks\/config.lua <strong>(not found)<\/strong><br>\n  Rocks trees in use:<br>\n  \/home\/myname\/.luarocks<br>\n  \/home\/myname\/code\/torch\/install <\/p>\n<\/blockquote>\n\n<p>So is it installed successfully ?<\/p>\n","answers":[{"AnswerId":"44482260","CreationDate":"2017-06-11T09:23:39.150","ParentId":null,"OwnerUserId":"204011","Title":null,"Body":"<p>The first error message explains the problem: your user doesn't have write permission to the directory where you installed Torch. This is probably because you ran something as root that you shouldn't have.<\/p>\n\n<p>You can probably fix this by running something like:<\/p>\n\n<pre><code>sudo chown -R myname: \/home\/myname\/code\/torch\/install\/lib\/luarocks\nsudo chmod -R u+w \/home\/myname\/code\/torch\/install\/lib\/luarocks\n<\/code><\/pre>\n"}]}
{"QuestionId":34314455,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-12-16T14:22:13.113","AcceptedAnswerId":34316117.0,"OwnerUserId":3155053.0,"Title":"How to print flag descriptions in Tensorflow?","Body":"<p>Google has many examples that use flags.  They all have descriptions in the definition.  Is there a way I can print these descriptions out to the terminal?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34316117","CreationDate":"2015-12-16T15:38:44.750","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>flags<\/code> module used in TensorFlow is a wrapper around the <a href=\"https:\/\/github.com\/google\/python-gflags\" rel=\"noreferrer\"><code>python-gflags<\/code> module<\/a>. To see a list of all flags used in a Python application using <code>python-gflags<\/code>, you can run it with the <code>-h<\/code> or <code>--help<\/code> flag. For example:<\/p>\n\n<pre><code>$ tensorboard -h\nusage: tensorboard [-h] [--logdir LOGDIR] [--debug DEBUG] [--nodebug]\n                   [--host HOST] [--port PORT]\n\noptional arguments:\n  -h, --help       show this help message and exit\n  --logdir LOGDIR  logdir specifies where TensorBoard will look to find\n                   TensorFlow event files that it can display. In the simplest\n                   case, logdir is a directory containing tfevents files.\n                   TensorBoard also supports comparing multiple TensorFlow\n                   executions: to do this, you can use directory whose\n                   subdirectories contain tfevents files, as in the following\n                   example: foo\/bar\/logdir\/\n                   foo\/bar\/logdir\/mnist_1\/events.out.tfevents.1444088766\n                   foo\/bar\/logdir\/mnist_2\/events.out.tfevents.1444090064 You\n                   may also pass a comma seperated list of log directories,\n                   and you can assign names to individual log directories by\n                   putting a colon between the name and the path, as in\n                   tensorboard\n                   --logdir=name1:\/path\/to\/logs\/1,name2:\/path\/to\/logs\/2\n  --debug DEBUG    Whether to run the app in debug mode. This increases log\n                   verbosity to DEBUG.\n  --nodebug\n  --host HOST      What host to listen to. Defaults to allowing remote access,\n                   set to 127.0.0.1 to serve only on localhost.\n  --port PORT      What port to serve TensorBoard on.\n<\/code><\/pre>\n"},{"AnswerId":"45973086","CreationDate":"2017-08-31T04:07:46.660","ParentId":null,"OwnerUserId":"7923397","Title":null,"Body":"<p>A tricky way:<\/p>\n\n<pre><code>print(FLAGS.__dict__['__flags'])\n<\/code><\/pre>\n\n<p>You can find reason in file <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.3\/tensorflow\/python\/platform\/flags.py#L35\" rel=\"nofollow noreferrer\">tensorflow\/python\/platform\/flags.py<\/a><\/p>\n"}]}
{"QuestionId":34315015,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-16T14:47:01.763","AcceptedAnswerId":null,"OwnerUserId":4511070.0,"Title":"Training from CSV file - use every train example per epoch","Body":"<p>I have a CSV file with 200.000 training samples that I would like to train my network with.<\/p>\n\n<p>I'm using an  and  to get the data. I then run all the data through , where I set ,  and . <\/p>\n\n<p>I then run a loop and call  repeatedly.<\/p>\n\n<p>The question I have is that I now want to run this for several epochs. In each epoch I would like to exaust the entire training set. I don't think the current setup does this. How would I go about doing that? \nI'm not even sure, if I understood the inner workings of  and its parameters fully as of yet. <\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"34324012","CreationDate":"2015-12-16T23:20:31.713","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>The queue should block at the end of the epoch. When that happens, you will know that you have exhausted the training set. More information in this related question: <a href=\"https:\/\/stackoverflow.com\/questions\/34079781\/tensor-flow-shuffle-batch-blocks-at-end-of-epoch\">Tensor Flow shuffle_batch() blocks at end of epoch<\/a><\/p>\n"}]}
{"QuestionId":34316687,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-12-16T16:04:14.260","AcceptedAnswerId":34359679.0,"OwnerUserId":3681744.0,"Title":"Stride in Conv2D","Body":"<p>Does  or any other way by which we can form CNNs support stride, i.e., the amount of moving the filter over the input? the default value of stride for normal 2D convolution is 1 but I would like to change it.<\/p>\n","answers":[{"AnswerId":"34359679","CreationDate":"2015-12-18T16:12:26.433","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The function <code>T.nnet.conv2d<\/code> has a parameter <code>subsample<\/code> which permits you to do this.<\/p>\n"}]}
{"QuestionId":34317398,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-16T16:39:47.337","AcceptedAnswerId":null,"OwnerUserId":1973820.0,"Title":"Error on restoring Tensorflow checkpoint file","Body":"<p>I get following error when using  method in tensorflow. Any idea why this happens?<\/p>\n\n<p>I saved the model like this:\n<\/p>\n\n<p>And the error is:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Full trace:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>My code for restoring checkpoint file<\/p>\n\n<pre><\/pre>\n\n<hr>\n","answers":[{"AnswerId":"52538094","CreationDate":"2018-09-27T13:24:38.153","ParentId":null,"OwnerUserId":"5106245","Title":null,"Body":"<p>You are trying to load variable that doesn't exist in the original network, I believe that omitting <\/p>\n\n<pre><code>    v1 = tf.Variable(0)\n<\/code><\/pre>\n\n<p>will solve the problem.<\/p>\n\n<p>In case that you want add a new variable, you need to load it differently, the loading method should be like that:<\/p>\n\n<pre><code>reader = tf.train.NewCheckpointReader(os.path.join(checkpoint_dir, ckpt_name))\nrestore_dict = dict()\nfor v in tf.trainable_variables():\n    tensor_name = v.name.split(':')[0]\n    if reader.has_tensor(tensor_name):\n        print('has tensor ', tensor_name)\n        restore_dict[tensor_name] = v\n    # put the logic of the new\/modified variable here and assign to the restore_dict, i.e. \n    # restore_dict['my_var_scope\/my_var'] = get_my_variable()\n<\/code><\/pre>\n"}]}
{"QuestionId":34318918,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><softmax>","CreationDate":"2015-12-16T17:56:30.687","AcceptedAnswerId":34324056.0,"OwnerUserId":3723253.0,"Title":"Tensor Flow softmax Regression Always Predicts 1","Body":"<p>I have the following code based on the MNIST example. It is modified in two ways:<\/p>\n\n<p>1) I'm not using a one-hot-vector, so I simply use <\/p>\n\n<p>2) My results are binary: either 0 or 1<\/p>\n\n<pre><\/pre>\n\n<p>The accuracy is always 0.5 (because my test data has about as many 1s as 0s). The values of  and  always seem to increase, probably because the values of  are always a vector of all zeros.<\/p>\n\n<p>When I try and use this model for prediction, the predictions are always 1:<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong here?<\/p>\n","answers":[{"AnswerId":"34324056","CreationDate":"2015-12-16T23:24:37.253","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>You seem to be predicting a single scalar, rather than a vector. The softmax op produces a vector-valued prediction for each example. This vector must always sum to 1. When the vector only contains one element, that element must always be 1. If you want to use a softmax for this problem, you could use [1, 0] as the output target where you are currently using [0] and use [0, 1] where you are currently using [1]. Another option is you could keep using just one number, but change the output layer to sigmoid instead of softmax, and change the cost function to be the sigmoid-based cost function as well.<\/p>\n"}]}
{"QuestionId":34321467,"AnswerCount":1,"Tags":"<python><install><redhat><tensorflow>","CreationDate":"2015-12-16T20:27:09.010","AcceptedAnswerId":null,"OwnerUserId":1054435.0,"Title":"TensorFlow failed to pip-install on RedHat 6","Body":"<p>I have a RedHat 6 machine with python installed using , so<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>When I ran<\/p>\n\n<pre><\/pre>\n\n<p>I got lots of log messages including compile\/build errors. I am not dumping the whole log here, just the pieces that I think are crucial, but can add more if needed:<\/p>\n\n<pre><\/pre>\n\n<p>How do I fix this? Thx!<\/p>\n","answers":[{"AnswerId":"34321839","CreationDate":"2015-12-16T20:48:23.740","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Seems like You are missing the cblas libraries:<\/p>\n\n<pre><code>\/usr\/bin\/ld: cannot find -lcblas\n...\n distutils.errors.LinkError: Command \"cc \/tmp\/tmp72IZmg\/tmp\/tmp72IZmg\/source.o -L\/usr\/lib64 -lblas -o \/tmp\/tmp72IZmg\/a.out\" failed with exit status 1\n<\/code><\/pre>\n\n<p>try running <code>yum install blas blas-devel<\/code><\/p>\n"}]}
{"QuestionId":34322799,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2015-12-16T21:53:47.933","AcceptedAnswerId":null,"OwnerUserId":3799389.0,"Title":"Deep Learning Caffe Save Test output labels","Body":"<p>I'm wondering how I can save in Caffe the output of testing results (predicted labels) in a file?<\/p>\n\n<p>In addition to averaged accuracy reported by Caffe on the screen, I'd like to take a look at the labels in detail.<\/p>\n\n<p>Thank you,\nSaman<\/p>\n","answers":[{"AnswerId":"34327671","CreationDate":"2015-12-17T06:04:35.717","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe offers a <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers.html\" rel=\"nofollow\"><code>\"HDF5Output\"<\/code><\/a> layer that can save to hdf5 binary file the outputs of certain layers.<br>\nYou can see the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/layers\/hdf5_output_layer.hpp\" rel=\"nofollow\">header file<\/a> for more information.<\/p>\n"},{"AnswerId":"34960838","CreationDate":"2016-01-23T07:28:54.810","ParentId":null,"OwnerUserId":"5140144","Title":null,"Body":"<p>@Saman - I would suggest you to use the original version of Caffe which is designed for Linux . As they have mentioned in their website, the Windows version is unofficial. Hope it helps. <\/p>\n"}]}
{"QuestionId":34324449,"AnswerCount":1,"Tags":"<machine-learning><neural-network><tensorflow><lstm><tensorboard>","CreationDate":"2015-12-17T00:04:36.950","AcceptedAnswerId":34340682.0,"OwnerUserId":5187368.0,"Title":"Translating a TensorFlow LSTM into synapticjs","Body":"<p>I'm working on implementing an interface between a TensorFlow basic LSTM that's already been trained and a javascript version that can be run in the browser.  The problem is that in all of the literature that I've read LSTMs are modeled as mini-networks (using only connections, nodes and gates) and TensorFlow seems to have a lot more going on.<\/p>\n\n<p>The two questions that I have are: <\/p>\n\n<ol>\n<li><p>Can the TensorFlow model be easily translated into a more conventional neural network structure?<\/p><\/li>\n<li><p>Is there a practical way to map the trainable variables that TensorFlow gives you to this structure?<\/p><\/li>\n<\/ol>\n\n<p>I can get the 'trainable variables' out of TensorFlow, the issue is that they appear to only have one value for bias per LSTM node, where most of the models I've seen would include several biases for the memory cell, the inputs and the output.<\/p>\n","answers":[{"AnswerId":"34340682","CreationDate":"2015-12-17T17:26:12.430","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"\n\n<p>Internally, the <code>LSTMCell<\/code> class stores the LSTM weights as a one big matrix instead of 8 smaller ones for efficiency purposes. It is quite easy to divide it horizontally and vertically to get to the more conventional representation. However, it might be easier and more efficient if your library does the similar optimization.<\/p>\n\n<p>Here is the relevant piece of code of the <code>BasicLSTMCell<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>concat = linear([inputs, h], 4 * self._num_units, True)\n\n# i = input_gate, j = new_input, f = forget_gate, o = output_gate\ni, j, f, o = array_ops.split(1, 4, concat)\n<\/code><\/pre>\n\n<p>The <code>linear<\/code> function does the matrix multiplication to transform the concatenated input and the previous <code>h<\/code> state into 4 matrices of <code>[batch_size, self._num_units]<\/code> shape. The linear transformation uses a single matrix and bias variables that you're referring to in the question. The result is then split into different gates used by the LSTM transformation.<\/p>\n\n<p>If you'd like to explicitly get the transformations for each gate, you can split that matrix and bias into 4 blocks. It is also quite easy to implement it from scratch using 4 or 8 linear transformations.<\/p>\n"}]}
{"QuestionId":34328233,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-17T06:44:35.813","AcceptedAnswerId":34328697.0,"OwnerUserId":1017986.0,"Title":"Feeding parameters into placeholders in tensorflow","Body":"<p>I'm trying to get into tensorflow, setting up a network and then feeding data to it. For some reason I end up with the error message . I made a minimal example of what I'm trying to do:<\/p>\n\n<pre><\/pre>\n\n<p>Basically, I'm setting up a network with place holders and a sequence of input embeddings that I want to learn, and then I try to run the network, feeding the input embeddings into it. From what I can tell by searching for the error message, there might be something wrong with my , but I can't see any obvious mismatches in eg. dimensionality.<\/p>\n\n<p>So, what did I miss, or how did I get this completely backwards?<\/p>\n\n<p>EDIT: I've edited the above to clarify that the input represents embeddings that need to be learned. I guess the question can be asked more sharply as: <em>Is it possible to use placeholders for parameters?<\/em><\/p>\n","answers":[{"AnswerId":"34328697","CreationDate":"2015-12-17T07:13:35.603","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The inputs should be numpy arrays.<\/p>\n\n<p>So, instead of <code>tf.Variable(tf.random_normal([K]))<\/code>, simply write <code>np.random.randn(K)<\/code> and everything should work as expected.<\/p>\n\n<p>EDIT (The question was clarified after my answer):<\/p>\n\n<p>It is possible to use placeholders as parameters but in a slightly different way. For example:<\/p>\n\n<pre><code>lchild = tf.placeholder(tf.float32, shape=(K))\nrchild = tf.placeholder(tf.float32, shape=(K))\nparent = tf.nn.tanh(tf.add(lchild, rchild))\nloss = &lt;some loss that depends on the parent tensor or lchild\/rchild&gt;\n# Compute gradients with respect to the input variables\ngrads = tf.gradients(loss, [lchild, rchild])\n\ninputs = [np.random.randn(K), np.random.randn(K)]\nfor i in range(&lt;number of iterations&gt;):\n    np_grads = sess.run(grads, feed_dict={lchild:inputs[0], rchild:inputs[1])\n    inputs[0] -= 0.1 * np_grads[0]\n    inputs[1] -= 0.1 * np_grads[1]\n<\/code><\/pre>\n\n<p>It is not however the best or easiest way to do this. The main problem with it is that at every iteration you need to copy numpy arrays in and out of the session (which is running potentially on a different device like GPU).<\/p>\n\n<p>Placeholders generally are used to feed the data external to the model (like texts or images). The way to solve it using tensorflow utilities would be something like:<\/p>\n\n<pre><code>lchild = tf.Variable(tf.random_normal([K])\nrchild = tf.Variable(tf.random_normal([K])\nparent = tf.nn.tanh(tf.add(lchild, rchild))\nloss = &lt;some loss that depends on the parent tensor or lchild\/rchild&gt;\ntrain_op = tf.train.GradientDescentOptimizer(loss).minimize(0.1)\n\nfor i in range(&lt;number of iterations&gt;):\n    sess.run(train_op)\n\n# Retrieve the weights back to numpy:\nnp_lchild = sess.run(lchild)\n<\/code><\/pre>\n"}]}
{"QuestionId":34328467,"AnswerCount":1,"Tags":"<ipython><theano><nvcc><keras>","CreationDate":"2015-12-17T06:59:02.813","AcceptedAnswerId":null,"OwnerUserId":3704652.0,"Title":"How Can I use my GPU on Ipython Notebook?","Body":"<p>OS : Ubuntu 14.04LTS <br>\nLanguage : Python Anaconda 2.7 (keras, theano) <br>\nGPU : GTX980Ti\nCUDA : CUDA 7.5<\/p>\n\n<p>I wanna run keras python code on IPython Notebook by using my GPU(GTX980Ti)\n<Br> But I can't find it.<\/p>\n\n<p>I want to test below code. When I run it on to Ubuntu terminal,\nI command as below (It uses GPU well. It doesn't have any problem)<\/p>\n\n<p><strong>First I set the path like below<\/strong> <br><\/p>\n\n<pre><\/pre>\n\n<p><strong>Second I run the code as below<\/strong> <br><\/p>\n\n<pre><\/pre>\n\n<p><br>And it runs well. <\/p>\n\n<p><strong>But when i run the code on pycharm(python IDE) or\nWhen I run it on Ipython Notebook, It doesn't use gpu.\nIt only uses CPU<\/strong><\/p>\n\n<p>myscript.py code is as below.<\/p>\n\n<pre><\/pre>\n\n<p><strong>To solve it, I force the code use gpu as below<\/strong>\n(Insert two lines more on myscript.py)<\/p>\n\n<pre><\/pre>\n\n<p><strong>Then It generate the error like below<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>how to do it??? I spent two days.. <br>\nAnd I surely did the way of using '.theanorc' file at home directory.<\/p>\n","answers":[{"AnswerId":"35315126","CreationDate":"2016-02-10T12:19:33.220","ParentId":null,"OwnerUserId":"3212998","Title":null,"Body":"<p>I'm using theano on an ipython notebook making use of my system's GPU. This configuration seems to work fine on my system.(Macbook Pro with GTX 750M)<\/p>\n\n<p>My ~\/.theanorc file :<\/p>\n\n<pre><code>[global]\ncnmem = True\nfloatX = float32\ndevice = gpu0\n<\/code><\/pre>\n\n<p>Various environment variables (I use a virtual environment(macvnev):<\/p>\n\n<pre><code>echo $LD_LIBRARY_PATH\n\/opt\/local\/lib:\n\necho $PATH\n\/Developer\/NVIDIA\/CUDA-7.5\/bin:\/opt\/local\/bin:\/opt\/local\/sbin:\/Developer\/NVIDIA\/CUDA-7.0\/bin:\/Users\/Ramana\/projects\/macvnev\/bin:\/usr\/local\/bin:\/usr\/bin:\/bin:\/usr\/sbin:\/sbin\n\necho $DYLD_LIBRARY_PATH\n\/Developer\/NVIDIA\/CUDA-7.5\/lib:\/Developer\/NVIDIA\/CUDA-7.0\/lib:\n<\/code><\/pre>\n\n<p>How I run ipython notebook (For me, the device is gpu0)  :<\/p>\n\n<pre><code>$THEANO_FLAGS=mode=FAST_RUN,device=gpu0,floatX=float32 ipython notebook\n<\/code><\/pre>\n\n<p>Output of <code>$nvcc -V<\/code> :<\/p>\n\n<pre><code>nvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2015 NVIDIA Corporation\nBuilt on Thu_Sep_24_00:26:39_CDT_2015\nCuda compilation tools, release 7.5, V7.5.19\n<\/code><\/pre>\n\n<p>From your post, probably you've set the $PATH variable wrong. <\/p>\n"}]}
{"QuestionId":34329750,"AnswerCount":2,"Tags":"<deep-learning><caffe>","CreationDate":"2015-12-17T08:21:04.443","AcceptedAnswerId":null,"OwnerUserId":2488716.0,"Title":"Does Caffe Scale up to multiple CPU Cores?","Body":"<p>I wish to run Caffe on a 32 core machine.<\/p>\n\n<ol>\n<li>Does caffe scale up to available number of cores to utilize them the best? <\/li>\n<li>Although there are 32 cores, can I make caffe use only a selected number of cores? <\/li>\n<\/ol>\n","answers":[{"AnswerId":"34337453","CreationDate":"2015-12-17T14:50:36.183","ParentId":null,"OwnerUserId":"2915032","Title":null,"Body":"<p>Generally <strong>caffe<\/strong> doesn't support multiple CPUs\/cores in its source code, but it uses <strong>BLAS<\/strong> routines.<\/p>\n\n<p>Thus answers to your questions are the following:<\/p>\n\n<ol>\n<li>Yes, but only through <strong>BLAS<\/strong> configuration, i. e. your <strong>BLAS<\/strong> version should be compiled with multithreading support (see related discussions: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1492\">here<\/a> or <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/439\">here<\/a> - at the second link you can also find some modifications for <strong>caffe<\/strong> itself).<\/li>\n<li>Also through <strong>BLAS<\/strong> (if it was compiled with <strong>openmp<\/strong> support, you can define <code>OMP_NUM_THREADS<\/code> to desired value).<\/li>\n<\/ol>\n"},{"AnswerId":"40154934","CreationDate":"2016-10-20T12:47:30.657","ParentId":null,"OwnerUserId":"4568702","Title":null,"Body":"<p>caffe does not, but you can you Intel caffe which is optimized for CPU and supports multi node\n<a href=\"https:\/\/github.com\/intel\/caffe\/wiki\/Multinode-guide\" rel=\"nofollow\">https:\/\/github.com\/intel\/caffe\/wiki\/Multinode-guide<\/a><\/p>\n"}]}
{"QuestionId":34330318,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><caffe><pycaffe>","CreationDate":"2015-12-17T08:55:33.627","AcceptedAnswerId":null,"OwnerUserId":5690129.0,"Title":"BVLC\/caffe giving same prediction everytime","Body":"<p>I am trying to run <strong>BVLC\/caffe<\/strong> model (CPU only). \nI have completed all installation. \nWhen i run below command to run:<\/p>\n\n<p><\/p>\n\n<p>Then its giving below output:<\/p>\n\n<pre><\/pre>\n\n<p>Above output is same for any image.<\/p>\n\n<p>classify.py FILE:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34407818","CreationDate":"2015-12-22T03:01:22.457","ParentId":null,"OwnerUserId":"4977638","Title":null,"Body":"<p>I have tried the same code using a somewhat similar classify.py. Why don't you give a try?<\/p>\n\n<pre><code>#Import all the nessesary libraries\nimport numpy as np\nimport matplotlib\nmatplotlib.use('Agg')\nimport matplotlib.pyplot as plt\nimport caffe\nimport os\nfrom os import listdir\nimport time\n\n#Set caffe to cpu or gpu mode, either is caffe.set_mode_cpu() or caffe.set_mode_gpu()\ncaffe.set_mode_cpu()\n\n#Define variable for location of required files\nMODEL_FILE = '..\/models\/bvlc_reference_caffenet\/deploy.prototxt'\nPRETRAINED = '..\/models\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel'\nMEAN_FILE = '.\/caffe\/imagenet\/ilsvrc_2012_mean.npy'\nLABEL_FILE = '..\/data\/ilsvrc12\/synset_words.txt'\n\n#Load the BVLC Reference Caffenet models\nnet = caffe.Classifier(MODEL_FILE, PRETRAINED,\n                       mean=np.load(MEAN_FILE).mean(1).mean(1),\n                       channel_swap=(2,1,0),\n                       raw_scale=255,\n                       image_dims=(256, 256))\n\n'''net = caffe.Classifier(MODEL_FILE, PRETRAINED)\nnet.set_channel_swap('data',(2,1,0))\nnet.set_raw_scale('data',255)\nnet.set_mean('data',np.load(MEAN_FILE))'''\n\npred_features = []\n\ninput_image = caffe.io.load_image('..\/examples\/images\/cat.jpg')\n\n'''\n#Let plot out the image\nplt.imshow(input_image)\nplt.savefig('..\/examples\/images\/cat.jpg')\nplt.close()\n'''\n\n#Predict class\nstart = time.time()\nprediction = net.predict([input_image])\nprint(\"Done in %.2f s.\" % (time.time() - start))\nprint 'prediction shape:', prediction[0].shape[0]\nprint 'predicted class:', prediction[0].argmax()\n\n#Predict label\nfi = open(LABEL_FILE)\nlabels = fi.readlines()\nprint 'predicted name:', labels[prediction[0].argmax()],\n\n'''\n\n#Plot the polygon frequency\nplt.plot(prediction[0])\nplt.savefig('..\/examples\/images\/cat.jpg')\nplt.close()\n'''\n<\/code><\/pre>\n\n<p>You just need to modify this as this code doesn't take input image as a command line argument.<\/p>\n"}]}
{"QuestionId":34331448,"AnswerCount":1,"Tags":"<python><virtualenv><osx-elcapitan><tensorflow>","CreationDate":"2015-12-17T09:50:39.513","AcceptedAnswerId":34335789.0,"OwnerUserId":5690264.0,"Title":"Virtualenv can not inherit GetSitePackages() Attribute","Body":"<p>I just installed TensorFlow under VirtualEnv on Mac OSX El Capitan. Now I am trying to understand the structure by following examples given in tensorflow.org website. <\/p>\n\n<p>I am new to python and its syntax. But as far as I can figure out the attribute called getsitepackages() is kind of important in order to list modules' attributes easily. But with its default python and virtualenv version on el capitan, it seems that virtualenv can not inherit  attribute of the module called \"site\".<\/p>\n\n<p>Hence I couldn't run the simple example command () to locate tensorflow libraries.<\/p>\n\n<p>I guess this is a known bug but I couldn't find a way to solve this issue. I just wonder if anyone has already came up with and solved this problem? <\/p>\n\n<p>P.S. Outside the virtualenv  just works fine. But in the virtualenv I get the following error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34335789","CreationDate":"2015-12-17T13:28:29.547","ParentId":null,"OwnerUserId":"3119756","Title":null,"Body":"<p>This seems to be a problem with <code>sites.py<\/code> which dates all the way back to 2012. As mentioned <a href=\"https:\/\/github.com\/pypa\/virtualenv\/issues\/355\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>Have a stab at trying to create the virtualenv using a different python version. For example:<\/p>\n\n<pre><code>virtuelnv -p python3 virtualenvname\n<\/code><\/pre>\n\n<p>It's worth checking what python version you're running (<code>python --version<\/code>). This seems to only be a problem with <code>python2.7<\/code> - earlier versions like <code>python2.6<\/code> do not experience this problem however they lack a lot of useful packages that were added in <code>python2.7<\/code>.<\/p>\n\n<p>My recommendation would be to run it under <code>python3<\/code> or <code>python3.4<\/code>.\nTensorFlow seems to support <code>python3<\/code> with the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/0.6.0\" rel=\"nofollow\">0.6.0 release<\/a>.<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":34337788,"AnswerCount":1,"Tags":"<python><windows><docker><tensorflow>","CreationDate":"2015-12-17T15:06:05.347","AcceptedAnswerId":null,"OwnerUserId":4684826.0,"Title":"Location of tensorflow\/models.. in Windows","Body":"<p>I have installed tensorflow on Windows using Docker, I want to go to the folder \"tensorflow\/models\/image\/imagenet\" that contains \"classify_image.py\" python file..\nCan someone please how to reach this mentioned path?  <\/p>\n","answers":[{"AnswerId":"34340617","CreationDate":"2015-12-17T17:22:23.767","ParentId":null,"OwnerUserId":"364150","Title":null,"Body":"<p>If you're using one of the devel tags (<code>:latest-devel<\/code> or <code>:latest-devel-gpu<\/code>), the file should be in <code>\/tensorflow\/tensorflow\/models\/image\/imagenet\/classify_image.py<\/code>.<\/p>\n\n<p>If you're using the base container (<code>b.gcr.io\/tensorflow\/tensorflow:latest<\/code>), it's not included -- that image just has the binary installed, not a full source distribution, and <code>classify_image.py<\/code> isn't included in the binary distribution.<\/p>\n"}]}
{"QuestionId":34338365,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-17T15:34:01.137","AcceptedAnswerId":34340346.0,"OwnerUserId":1441121.0,"Title":"Caching Computations in TensorFlow","Body":"<p>Is there a canonical way to reuse computations from a previously-supplied placeholder in TensorFlow? My specific use case:<\/p>\n\n<ul>\n<li>supply many inputs (using one placeholder) simultaneously, all of which are fed through a network to obtain smaller representations<\/li>\n<li>define a loss based on various combinations of these smaller representations<\/li>\n<li>train on one batch at a time, where each batch uses some subset of the inputs, <em>without recomputing the smaller representations<\/em><\/li>\n<\/ul>\n\n<p>Here is the goal in code, but which is defective because the same computations are carried out again and again:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"34340346","CreationDate":"2015-12-17T17:07:34.483","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>The canonical way to share computed values across sess.Run() calls is to use a <code>Variable<\/code>. In this case, you could set up your graph so that when the Placeholders are fed, they compute a new value of the representation that is saved into a Variable. A separate portion of the graph reads those Variables to compute the loss. This will not work if you need to compute gradients through the part of the graph that computes the representation. Computing those gradients will require recomputing every Op in the encoder.<\/p>\n"},{"AnswerId":"34375887","CreationDate":"2015-12-19T22:07:39.587","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This is the kind of thing that should be solved automatically with CSE (common subexpression elimination). Not sure what the support in TensorFlow right now, might be kind of spotty, but there's <code>optimizer_do_cse<\/code> flag for Graph options which is defaulting to false, and you can set it to true using GraphConstructorOptions. Here's a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/direct_session.cc#L500\" rel=\"nofollow\">C++ example<\/a> of using GraphConstructorOptions (sorry, couldn't find a Python one)<\/p>\n\n<p>If that doesn't work, you could do \"manual CSE\", ie, figure out which part is being needlessly recomputed, factor it out into separate Tensor, and reference that tensor in all the calculations.<\/p>\n"}]}
{"QuestionId":34338710,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2015-12-17T15:49:49.940","AcceptedAnswerId":34422721.0,"OwnerUserId":4618605.0,"Title":"TensorFlow no attribute 'make_template'","Body":"<p>I am trying to get familiar with tensorflow and NNs. I recently crashed into this problem when I tried to use . I first noticed that there were no auto-complete option in the IDE I use, and then I just tried to call the function from the python cmd:<\/p>\n\n<pre><\/pre>\n\n<p>I installed tensorflow from sources, and the protobuf version installed is:<\/p>\n\n<pre><\/pre>\n\n<p>I haven't faced any similar behaviour with other tensorflow functions so far.\nAny ideas about what's the issue causing this one?<\/p>\n","answers":[{"AnswerId":"34422721","CreationDate":"2015-12-22T19:02:50.753","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As noted in the comments, this issue was due to a mismatch between the installed version of TensorFlow (0.5.0) and the downloaded source (0.6.0).<\/p>\n\n<p>To upgrade to the latest development version of TensorFlow, follow the instructions to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\" rel=\"nofollow\">install from source<\/a>, then <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#create-pip\" rel=\"nofollow\">build and install the PIP package<\/a> based on that source.<\/p>\n"}]}
{"QuestionId":34339489,"AnswerCount":2,"Tags":"<lua><neural-network><torch>","CreationDate":"2015-12-17T16:23:00.073","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"In Torch, how to add momentum to my stochastic descent?","Body":"<p>I'm implementing a multi-layer perceptron neural network in Torch, by following <a href=\"http:\/\/torch5.sourceforge.net\/manual\/newbieTutorial.html\" rel=\"nofollow\">this example<\/a> (from the \"Example: training a neural network\" to the \"Torch basics: testing your neural network\" section).<\/p>\n\n<p>Now I have to add some <a href=\"https:\/\/en.wikipedia.org\/wiki\/Gradient_descent#The_momentum_method\" rel=\"nofollow\">momentum<\/a> to this neural network model, but I cannot understand how to do it.\nThe gradient descent part is:<\/p>\n\n<pre><\/pre>\n\n<p>I was wishing there was some command like: trainer.momentum = 1, but there's no such thing.<\/p>\n\n<p>Does anyone know how to add momentum to my Torch neural network model?<\/p>\n","answers":[{"AnswerId":"34821828","CreationDate":"2016-01-15T23:51:29.347","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>You can do one of the two things:<\/p>\n\n<ul>\n<li>implement SGD by your own - it is very basic algorithm, really. It will take just few lines of torch code and then add momentum<\/li>\n<li>use other library for training, like <code>optim<\/code> package from <code>torch<\/code> <a href=\"https:\/\/github.com\/torch\/optim\/blob\/master\/doc\/index.md\" rel=\"nofollow\">https:\/\/github.com\/torch\/optim\/blob\/master\/doc\/index.md<\/a> which has multiple training methods, including SGD with momentum<\/li>\n<\/ul>\n"},{"AnswerId":"34965355","CreationDate":"2016-01-23T15:37:34.593","ParentId":null,"OwnerUserId":"5407700","Title":null,"Body":"<p>Using the package <code>optim<\/code> (see <a href=\"https:\/\/github.com\/torch\/optim\" rel=\"nofollow\">https:\/\/github.com\/torch\/optim<\/a>) for optimisation you can make a Lua table for the parameters. Here is an example;<\/p>\n\n<pre><code>optimState = { \nlearningRate = 0.001,\nweightDecay = 0,\nmomentum = 0.9,\nlearningRateDecay = 0 \n}\n<\/code><\/pre>\n\n<p>Then you can pass this table of parameters into the <code>optim.method<\/code> function (which optimises your loss function), such as;<\/p>\n\n<pre><code>optim.method(func, x, optimState)\n<\/code><\/pre>\n"}]}
{"QuestionId":34340064,"AnswerCount":1,"Tags":"<image><lua><torch>","CreationDate":"2015-12-17T16:51:37.973","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"image.display window in lua\/torch","Body":"<p>I am displaying images in torch using the  function. However each time I call the function it creates a new window but I would like it to replace the existing one if that is possible. Possibly its something to do with the  function but I can't figure it out.\nThanks.<\/p>\n","answers":[{"AnswerId":"34352402","CreationDate":"2015-12-18T09:33:52.903","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<pre><code>$ qlua\n&gt; require 'image'\n&gt; w = image.display(image.lena()) -- with positional arguments mode\n&gt; image.display{image=image.fabio(), win=w} -- with named arguments mode\n<\/code><\/pre>\n"}]}
{"QuestionId":34340489,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-12-17T17:14:59.480","AcceptedAnswerId":36947632.0,"OwnerUserId":2594660.0,"Title":"Tensorflow read images with labels","Body":"<p>I am building a standard image classification model with Tensorflow. For this I have input images, each assigned with a label (number in {0,1}). The Data can hence be stored in a list using the following format:<\/p>\n\n<pre><\/pre>\n\n<p>I want to use TensorFlow's queuing system to read my data and feed it to my model. Ignoring the labels, one can easily achieve this by using  and . Here the code:<\/p>\n\n<pre><\/pre>\n\n<p>However, the labels are lost in that process as the image data is purposely shuffled as part of the input pipeline. What is the easiest way of pushing the labels together with the image data through the input queues?<\/p>\n","answers":[{"AnswerId":"44894131","CreationDate":"2017-07-03T21:33:11.483","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>In addition to the answers provided there are few other things you can do:<\/p>\n\n<p><strong>Encode your label into the filename.<\/strong> If you have N different categories you can rename your files to something like: <code>0_file001, 5_file002, N_file003<\/code>. Afterwards when you read the data from a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/WholeFileReader\" rel=\"nofollow noreferrer\">reader<\/a> <code>key, value = reader.read(filename_queue)<\/code> your key\/value are: <\/p>\n\n<blockquote>\n  <p>The output of Read will be a filename (key) and the contents of that file (value)<\/p>\n<\/blockquote>\n\n<p>Then parse your filename, extract the label and convert it to int. This will require a little bit of preprocessing of the data.<\/p>\n\n<p><strong>Use <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">TFRecords<\/a><\/strong> which will allow you to store the data and labels at the same file.<\/p>\n"},{"AnswerId":"36947632","CreationDate":"2016-04-29T21:20:28.780","ParentId":null,"OwnerUserId":"2594660","Title":null,"Body":"<p>Using <code>slice_input_producer<\/code> provides a solution which is much cleaner. Slice Input Producer allows us to create an Input Queue containing arbitrarily many separable values. This snippet of the question would look like this:<\/p>\n\n<pre><code>def read_labeled_image_list(image_list_file):\n    \"\"\"Reads a .txt file containing pathes and labeles\n    Args:\n       image_list_file: a .txt file with one \/path\/to\/image per line\n       label: optionally, if set label will be pasted after each line\n    Returns:\n       List with all filenames in file image_list_file\n    \"\"\"\n    f = open(image_list_file, 'r')\n    filenames = []\n    labels = []\n    for line in f:\n        filename, label = line[:-1].split(' ')\n        filenames.append(filename)\n        labels.append(int(label))\n    return filenames, labels\n\ndef read_images_from_disk(input_queue):\n    \"\"\"Consumes a single filename and label as a ' '-delimited string.\n    Args:\n      filename_and_label_tensor: A scalar string tensor.\n    Returns:\n      Two tensors: the decoded image, and the string label.\n    \"\"\"\n    label = input_queue[1]\n    file_contents = tf.read_file(input_queue[0])\n    example = tf.image.decode_png(file_contents, channels=3)\n    return example, label\n\n# Reads pfathes of images together with their labels\nimage_list, label_list = read_labeled_image_list(filename)\n\nimages = ops.convert_to_tensor(image_list, dtype=dtypes.string)\nlabels = ops.convert_to_tensor(label_list, dtype=dtypes.int32)\n\n# Makes an input queue\ninput_queue = tf.train.slice_input_producer([images, labels],\n                                            num_epochs=num_epochs,\n                                            shuffle=True)\n\nimage, label = read_images_from_disk(input_queue)\n\n# Optional Preprocessing or Data Augmentation\n# tf.image implements most of the standard image augmentation\nimage = preprocess_image(image)\nlabel = preprocess_label(label)\n\n# Optional Image and Label Batching\nimage_batch, label_batch = tf.train.batch([image, label],\n                                          batch_size=batch_size)\n<\/code><\/pre>\n\n<p>See also the <a href=\"https:\/\/github.com\/TensorVision\/TensorVision\/blob\/master\/examples\/inputs\/generic_input.py\">generic_input_producer<\/a> from the <a href=\"https:\/\/github.com\/TensorVision\/TensorVision\">TensorVision<\/a> examples for full input-pipeline.<\/p>\n"},{"AnswerId":"34345827","CreationDate":"2015-12-17T22:54:53.867","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are three main steps to solving this problem:<\/p>\n\n<ol>\n<li><p>Populate the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"noreferrer\"><code>tf.train.string_input_producer()<\/code><\/a> with a list of strings containing the original, space-delimited string containing the filename and the label.<\/p><\/li>\n<li><p>Use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#read_file\" rel=\"noreferrer\"><code>tf.read_file(filename)<\/code><\/a> rather than <code>tf.WholeFileReader()<\/code> to read your image files. <code>tf.read_file()<\/code> is a stateless op that consumes a single filename and produces a single string containing the contents of the file. It has the advantage that it's a pure function, so it's easy to associate data with the input and the output. For example, your <code>read_my_file_format<\/code> function would become:<\/p>\n\n<pre><code>def read_my_file_format(filename_and_label_tensor):\n  \"\"\"Consumes a single filename and label as a ' '-delimited string.\n\n  Args:\n    filename_and_label_tensor: A scalar string tensor.\n\n  Returns:\n    Two tensors: the decoded image, and the string label.\n  \"\"\"\n  filename, label = tf.decode_csv(filename_and_label_tensor, [[\"\"], [\"\"]], \" \")\n  file_contents = tf.read_file(filename)\n  example = tf.image.decode_png(file_contents)\n  return example, label\n<\/code><\/pre><\/li>\n<li><p>Invoke the new version of <code>read_my_file_format<\/code> by passing a single dequeued element from the <code>input_queue<\/code>:<\/p>\n\n<pre><code>image, label = read_my_file_format(input_queue.dequeue())         \n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>You can then use the <code>image<\/code> and <code>label<\/code> tensors in the remainder of your model.<\/p>\n"}]}
{"QuestionId":34343259,"AnswerCount":6,"Tags":"<protocol-buffers><tensorflow>","CreationDate":"2015-12-17T20:01:06.000","AcceptedAnswerId":34343517.0,"OwnerUserId":2116766.0,"Title":"Is there an example on how to generate protobuf files holding trained TensorFlow graphs","Body":"<p>I am looking at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">Google's example<\/a> on how to deploy and use a pre-trained Tensorflow graph (model) on Android. This example uses a  file at: <\/p>\n\n<blockquote>\n  <p><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<\/blockquote>\n\n<p><em>which is a link to a file that downloads automatically<\/em>.<\/p>\n\n<p>The example shows how to load the  file to a Tensorflow session and use it to perform classification, but it doesn't seem to mention how to generate such a  file, after a graph is trained (e.g., in Python).<\/p>\n\n<p>Are there any examples on how to do that?<\/p>\n","answers":[{"AnswerId":"36512748","CreationDate":"2016-04-09T04:24:54.450","ParentId":null,"OwnerUserId":"359431","Title":null,"Body":"<p>Here's another take on @Mostafa's answer. A somewhat cleaner way to run the <code>tf.assign<\/code> ops is to store them in a <code>tf.group<\/code>.  Here's my Python code:<\/p>\n\n<pre><code>  ops = []\n  for v in tf.trainable_variables():\n    vc = tf.constant(v.eval())\n    ops.append(tf.assign(v, vc));\n  tf.group(*ops, name=\"assign_trained_variables\")\n<\/code><\/pre>\n\n<p>And in C++:<\/p>\n\n<pre><code>  std::vector&lt;tensorflow::Tensor&gt; tmp;\n  status = session.Run({}, {}, { \"assign_trained_variables\" }, &amp;tmp);\n  if (!status.ok()) {\n    \/\/ Handle error\n  }\n<\/code><\/pre>\n\n<p>This way you have only one named op to run on the C++ side, so you don't have to mess around with iterating over nodes.<\/p>\n"},{"AnswerId":"34343517","CreationDate":"2015-12-17T20:18:18.450","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>EDIT:<\/strong> The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"noreferrer\"><code>freeze_graph.py<\/code><\/a> script, which is part of the TensorFlow repository, now serves as a tool that generates a protocol buffer representing a \"frozen\" trained model, from an existing TensorFlow <code>GraphDef<\/code> and a saved checkpoint. It uses the same steps as described below, but it much easier to use.<\/p>\n\n<hr>\n\n<p>Currently the process isn't very well documented (and subject to refinement), but the approximate steps are as follows:<\/p>\n\n<ol>\n<li>Build and train your model as a <code>tf.Graph<\/code> called <code>g_1<\/code>.<\/li>\n<li>Fetch the final values of each of the variables and store them as numpy arrays (using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a>).<\/li>\n<li><strong>In a new <code>tf.Graph<\/code><\/strong> called <code>g_2<\/code>, create <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/constant_op.html#constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> tensors for each of the variables, using the value of the corresponding numpy array fetched in step 2.<\/li>\n<li><p>Use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#import_graph_def\" rel=\"noreferrer\"><code>tf.import_graph_def()<\/code><\/a> to copy nodes from <code>g_1<\/code> into <code>g_2<\/code>, and use the <code>input_map<\/code> argument to replace each variable in <code>g_1<\/code> with the corresponding <code>tf.constant()<\/code> tensors created in step 3. You may also want to use <code>input_map<\/code> to specify a new input tensor (e.g. replacing an <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html\" rel=\"noreferrer\">input pipeline<\/a> with a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a>). Use the <code>return_elements<\/code> argument to specify the name of the predicted output tensor.<\/p><\/li>\n<li><p>Call <code>g_2.as_graph_def()<\/code> to get a protocol buffer representation of the graph.<\/p><\/li>\n<\/ol>\n\n<p>(<strong>NOTE:<\/strong> The generated graph will have extra nodes in the graph for training. Although it is not part of the public API, you may wish to use the internal <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html\" rel=\"noreferrer\"><code>graph_util.extract_sub_graph()<\/code><\/a> function to strip these nodes from the graph.)<\/p>\n"},{"AnswerId":"35027060","CreationDate":"2016-01-27T01:07:57.787","ParentId":null,"OwnerUserId":"214488","Title":null,"Body":"<p>Just found this post and it was very useful thanks! I'm also going with @Mostafa's method, though my C++ code is a bit different:<\/p>\n\n<pre><code>    std::vector&lt;string&gt; names;\n    int node_count = graph.node_size();\n    cout &lt;&lt; node_count &lt;&lt; \" nodes in graph\" &lt;&lt; endl;\n\n    \/\/ iterate all nodes\n    for(int i=0; i&lt;node_count; i++) {\n        auto n = graph.node(i);\n        cout &lt;&lt; i &lt;&lt; \":\" &lt;&lt; n.name() &lt;&lt; endl;\n\n        \/\/ if name contains \"var_hack\", add to vector\n        if(n.name().find(\"var_hack\") != std::string::npos) {\n            names.push_back(n.name());\n            cout &lt;&lt; \"......bang\" &lt;&lt; endl;\n        }\n    }\n    session.Run({}, names, {}, &amp;outputs);\n<\/code><\/pre>\n\n<p>NB I use \"var_hack\" as my variable name in python<\/p>\n"},{"AnswerId":"38853802","CreationDate":"2016-08-09T14:38:36.507","ParentId":null,"OwnerUserId":"169275","Title":null,"Body":"<p>Alternatively to my previous answer using <code>freeze_graph()<\/code>, which is only good if you call it as a script, there is a very nice function that will do all the heavy lifting for you and is suitable to be called from your normal model training code.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v0.10.0rc0\/tensorflow\/python\/framework\/graph_util.py#L193\"><code>convert_variables_to_constants()<\/code><\/a> does two things:<\/p>\n\n<ul>\n<li>It freezes the weights by replacing variables with constants<\/li>\n<li>It removes nodes which are not related to feedforward prediction<\/li>\n<\/ul>\n\n<p>Assuming <code>sess<\/code> is your <code>tf.Session()<\/code> and <code>\"output\"<\/code> is the name of your prediction node, the following code will serialize your minimal graph both into textual and binary protobuf.<\/p>\n\n<hr>\n\n<pre><code>from tensorflow.python.framework.graph_util import convert_variables_to_constants\n\nminimal_graph = convert_variables_to_constants(sess, sess.graph_def, [\"output\"])\n\ntf.train.write_graph(minimal_graph, '.', 'minimal_graph.proto', as_text=False)\ntf.train.write_graph(minimal_graph, '.', 'minimal_graph.txt', as_text=True)\n<\/code><\/pre>\n"},{"AnswerId":"34820022","CreationDate":"2016-01-15T21:20:11.120","ParentId":null,"OwnerUserId":"5790995","Title":null,"Body":"<p>I could not figure out how to implement the method described by mrry. But here how I solved it. I'm not sure if that is the best way of solving the problem but at least it solves it.<\/p>\n\n<p>As write_graph can also store the values of the constants, I added the following code to the python just before writing the graph with write_graph function:<\/p>\n\n<pre><code>for v in tf.trainable_variables():\n    vc = tf.constant(v.eval())\n    tf.assign(v, vc, name=\"assign_variables\")\n<\/code><\/pre>\n\n<p>This creates constants that store variables' values after being trained and then create tensors \"<strong><em>assign_variables<\/em><\/strong>\" to assign them to the variables. Now, when you call write_graph, it will store the variables' values in the file in form of constants.<\/p>\n\n<p>The only remaining part is to call these tensors \"<strong><em>assign_variables<\/em><\/strong>\" in the c code to make sure that your variables are assigned with the constants values that are stored in the file. Here is a one way to do it:<\/p>\n\n<pre><code>      Status status = NewSession(SessionOptions(), &amp;session);\n      std::vector&lt;tensorflow::Tensor&gt; outputs;\n      char name[100];\n      for(int i = 0;status.ok(); i++) {\n        if (i==0)\n            sprintf(name, \"assign_variables\");\n        else\n            sprintf(name, \"assign_variables_%d\", i);\n\n        status = session-&gt;Run({}, {name}, {}, &amp;outputs);\n      }\n<\/code><\/pre>\n"},{"AnswerId":"38852398","CreationDate":"2016-08-09T13:39:04.353","ParentId":null,"OwnerUserId":"169275","Title":null,"Body":"<p>I found a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow\"><code>freeze_graph()<\/code><\/a> function in the Tensorflow codebase that might be helpful when doing this. From what I understand it swaps variables with constants before serializing the GraphDef and so when you then load this graph from C++ it has no variables that need to be set anymore, and you can directly use it for predictions.<\/p>\n\n<p>There is also a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph_test.py\" rel=\"nofollow\">test<\/a> for it and some description in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/how_tos\/tool_developers\/index.html#freezing\" rel=\"nofollow\">Guide<\/a>.<\/p>\n\n<p>This seems like the cleanest option here.<\/p>\n"}]}
{"QuestionId":34346839,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-18T00:40:34.260","AcceptedAnswerId":34380488.0,"OwnerUserId":1248073.0,"Title":"Theano: change `base_compiledir` to save compiled files in another directory","Body":"<p> refers to the directory where the compiled files are stored. <\/p>\n\n<p>Is there a way where I could permanently set  to a different location, perhaps by modifying the content of some internal theano files ? <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a> does explain ways for configuring  in some aspect but I still couldn't address my question.<\/p>\n\n<p>I am using Ubuntu.<\/p>\n\n<p>Thanks &amp; Cheers!<\/p>\n","answers":[{"AnswerId":"34380488","CreationDate":"2015-12-20T11:25:48.917","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>As the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"noreferrer\">documentation<\/a> explains, you can set this, or any other Theano config flag, permanently by altering either the <code>THEANO_FLAGS<\/code> environment variable (e.g. in your <code>~\/.bashrc<\/code> file) or using a <code>~\/.theanorc<\/code> file.<\/p>\n\n<p>For the former, add a line like this to your <code>~\/.bashrc<\/code> file:<\/p>\n\n<pre><code>export THEANO_FLAGS=\"base_compiledir=\/some\/path\"\n<\/code><\/pre>\n\n<p>For the latter, create a <code>~\/.theanorc<\/code> file with contents that look like this:<\/p>\n\n<pre><code>[global]\nbase_compiledir=\/some\/path\n<\/code><\/pre>\n\n<p>In either case, you'll probably want to add a bunch of other flags such as <code>device=<\/code>, <code>floatX=<\/code>, etc.<\/p>\n\n<p>If you alter your <code>~\/.bashrc<\/code> file the changes won't take effect in any active terminals unless you run <code>source ~\/.bashrc<\/code> in each one or simply close the terminals and start new ones.<\/p>\n"}]}
{"QuestionId":34347568,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2015-12-18T02:17:56.363","AcceptedAnswerId":34359648.0,"OwnerUserId":4280916.0,"Title":"theano - Operate Two Matrix By Rule like T.dot","Body":"<p>I want to operate two matrix by rule as T.dot, like:<\/p>\n\n<pre><\/pre>\n\n<p>and return the matrix:<\/p>\n\n<pre><\/pre>\n\n<p>But, instead dot between two vectors, I want the op is the function to compute cosine similarity.<\/p>\n\n<p>Is  there any function can do this in theano?<\/p>\n\n<p>=======<\/p>\n\n<p>The cosine similarity function is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34359648","CreationDate":"2015-12-18T16:10:50.610","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You should probably do this with the matrices directly:<\/p>\n\n<p>Define<\/p>\n\n<pre><code>V = (v1, v2, v3)\nU = (u1, u2, u3)\n<\/code><\/pre>\n\n<p>Then<\/p>\n\n<pre><code>import theano.tensor as T\nimport numpy as np\n\nU = T.fmatrix()\nV = T.fmatrix()\n\ncos_sim = T.dot(U, V.T) \/ (T.sqrt((U ** 2).sum(0)) * T.sqrt((V ** 2).sum(0).reshape((-1, 1))))\n\nu = np.arange(9.).reshape(3, 3)\n\ncos_sim.eval({U: u.astype('float32'), V: u.astype('float32')})\n<\/code><\/pre>\n"}]}
{"QuestionId":34349316,"AnswerCount":1,"Tags":"<optimization><machine-learning><tensorflow>","CreationDate":"2015-12-18T05:56:45.777","AcceptedAnswerId":34361377.0,"OwnerUserId":2272798.0,"Title":"Synchronous vs asynchronous computation in Tensorflow","Body":"<p>In the Tensorflow CIFAR tutorial it talks about using multiple GPUs and gives this warning:<\/p>\n\n<p>\"Naively employing asynchronous updates of model parameters leads to sub-optimal training performance because an individual model replica might be trained on a stale copy of the model parameters. Conversely, employing fully synchronous updates will be as slow as the slowest model replica.\"<\/p>\n\n<p>What does this mean? Could someone provide a very simple example that illustrates this warning?<\/p>\n","answers":[{"AnswerId":"34361377","CreationDate":"2015-12-18T17:57:58.963","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>Suppose you have n workers.<\/p>\n\n<p>Asynchronous means that each worker just reads parameters, computes updates, and writes updated parameters, without any locking mechanism at all. The workers can overwrite each other's work freely.\nSuppose worker 1 is slow for some reason. Worker 1 reads parameters at time t, and then tries to write updated parameters at time t+100. In the meantime, workers 2-n have all done a lot of updates at time step t+1, t+2, etc. When the slow worker 1 finally does its write, it overwrites all of the progress the other workers have made.<\/p>\n\n<p>Fully synchronous means that all the workers are coordinated. Every worker reads the parameters, computes a gradient, and then waits for the other workers to finish. Then the learning algorithm computes the average of all of the gradients they computed, and does an update based on that one average. If worker 1 is very slow and takes 100 time steps to finish, but workers 2-n all finish on time step 2, then most of the workers will spend most of the time sitting doing nothing waiting for worker 1.<\/p>\n"}]}
{"QuestionId":34349967,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-18T06:51:38.233","AcceptedAnswerId":34361526.0,"OwnerUserId":5693945.0,"Title":"Tensorflow - The source code of the tensorflow_inception_graph.pb","Body":"<p>Does anyone know the original source code of the tensorflow_inception_graph.pb.<\/p>\n\n<p>I really wan't to know the operation in the example project<\/p>\n\n<p>The tensorflow\/example\/android  - Read me.\nTensorflow Android Camera Demo<\/p>\n\n<p><a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/examples\/android\/\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/examples\/android\/<\/a>\n<a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip<\/a><\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34361526","CreationDate":"2015-12-18T18:07:24.927","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This code has not yet been released. As the <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msg\/discuss\/kl3d9DspOMw\/YIpdtXerCQAJ\" rel=\"nofollow\">release announcement for the latest ImageNet model<\/a> mentions, the team is working on releasing the complete Python training framework for this type of model, including the code for building the graph.<\/p>\n"},{"AnswerId":"42724133","CreationDate":"2017-03-10T17:17:41.317","ParentId":null,"OwnerUserId":"758836","Title":null,"Body":"<p>I think you the rb and inception labels assets directly from <a href=\"https:\/\/github.com\/miyosuda\/TensorFlowAndroidDemo\/tree\/master\/app\/src\/main\/assets\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":34351935,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2015-12-18T09:04:42.570","AcceptedAnswerId":null,"OwnerUserId":5640428.0,"Title":"Caffe large files of type lmdb","Body":"<p>I'm wondering if there's a reason why lmdb files using in <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> are so much larger than the file containing the original images. Could you give me an explanation please?<\/p>\n","answers":[{"AnswerId":"34352325","CreationDate":"2015-12-18T09:29:22.130","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It is hard to give a concrete answer to such an abstract question, but I'll give it a try:<br>\nImage files are usually compressed: a <code>.png<\/code> or <code>.jpg<\/code> of size <code>h<\/code> by <code>w<\/code> by <code>3<\/code> takes far less disk space than <code>h*w*3<\/code> bytes due to compression. On the other hand, for processing the image in a neural network (or any other ML software for that matter) you need to work with the un-compressed representation of the image. Therefore, <code>lmdb<\/code>, <code>leveldb<\/code>, <code>hdf5<\/code> datasets used by caffe stores the input images in an <strong>uncompressed<\/strong> manner using 32bit float numbers for each pixels (instead of <code>uint8<\/code>) thus the drastic increase in file size.<\/p>\n"}]}
{"QuestionId":34352593,"AnswerCount":1,"Tags":"<deep-learning><keras>","CreationDate":"2015-12-18T09:44:18.960","AcceptedAnswerId":null,"OwnerUserId":1474615.0,"Title":"keras dense input layer","Body":"<p>I am mainly a caffe user, I wonder how can I input my color image in a fully connected layer and then flatten it ( I know this is not the best solution but I need it).\nThis is a sample that is not working:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestion? I think I have to set the input layer and then flatten it but I don't know how.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"34352662","CreationDate":"2015-12-18T09:48:02.213","ParentId":null,"OwnerUserId":"1474615","Title":null,"Body":"<p>Just found the solution...<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Flatten(input_shape=(img_channels, img_rows, img_cols)))\nmodel.add(Dense(1000))\nmodel.add(Activation('relu'))\n<\/code><\/pre>\n\n<p>It was obvious though. :(<\/p>\n"}]}
{"QuestionId":34353016,"AnswerCount":1,"Tags":"<torch><luarocks>","CreationDate":"2015-12-18T10:07:45.073","AcceptedAnswerId":null,"OwnerUserId":3902984.0,"Title":"Luarocks error on ubuntu","Body":"<p>I am trying to run Neuraltalk2 on Ubuntu. But I am getting an error as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I followed the following steps uptill now:<\/p>\n\n<pre><\/pre>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"34400265","CreationDate":"2015-12-21T16:30:09.873","ParentId":null,"OwnerUserId":"1793220","Title":null,"Body":"<p>Try running this all without <code>sudo<\/code>. The last line, especially, <code>sudo source ~\/.bashrc<\/code> does not work because <code>source<\/code> is meant to operate on the shell you are currently running. If you run it with <code>sudo<\/code>, it will load <code>.bashrc<\/code> into the temporary subshell created by <code>sudo<\/code> (in practice having no effect).<\/p>\n\n<p>Your error message indicates that <code>luarocks<\/code> was installed correctly, but it failed to find the rock. Make sure the name of the rock is correct, try searching it with the <code>luarocks search<\/code> command, and check your configuration running <code>luarocks<\/code> with no arguments (it will display the name of your config files in use, helping you to troubleshoot the issue).<\/p>\n"}]}
{"QuestionId":34353160,"AnswerCount":1,"Tags":"<protocol-buffers><tensorflow>","CreationDate":"2015-12-18T10:14:23.973","AcceptedAnswerId":34402643.0,"OwnerUserId":5693561.0,"Title":"How can I execute a TensorFlow graph from a protobuf in C++?","Body":"<p>I got a simple code form tutorial and output it to .pb file as below:<\/p>\n\n<p>mnist_softmax_train.py<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>In C++, I load the same graph, and feed in fake data for testing:<\/p>\n\n<pre><\/pre>\n\n<p>However, this fails with the error:<\/p>\n\n<pre><\/pre>\n\n<p>The graph runs correctly in Python. How can I run it correctly in C++?<\/p>\n","answers":[{"AnswerId":"34402643","CreationDate":"2015-12-21T19:07:09.103","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue here is that you are running the <code>\"train_step\"<\/code> target, which performs much more work than just inference. In particular, it attempts to update the variables <code>W<\/code> and <code>b<\/code> with the result of the gradient descent step. The error message<\/p>\n\n<pre><code>Invalid argument: Input 0 of node train_step\/update_W\/ApplyGradientDescent was passed float from _recv_W_0:0 incompatible with expected float_ref.\n<\/code><\/pre>\n\n<p>...means that one of the nodes you attempted to run (<code>\"train_step\/update_W\/ApplyGradientDescent\"<\/code>) expected a mutable input (with type <code>float_ref<\/code>) but it got an immutable input (with type <code>float<\/code>) because the value was fed in.<\/p>\n\n<p>There are (at least) two possible solutions:<\/p>\n\n<ol>\n<li><p>If you only want to see predictions for a given input and given weights, fetch <code>\"softmax:0\"<\/code> instead of <code>\"train_step\"<\/code> in the call to <code>Session::Run()<\/code>.<\/p><\/li>\n<li><p>If you want to perform training in C++, do not feed <code>W<\/code> and <code>b<\/code>, but instead assign values to those variables, then continue to execute <code>\"train_step\"<\/code>. You may find it easier to create a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver\"><code>tf.train.Saver<\/code><\/a> when you build the graph in Python, and then invoke the operations that it produces to save and restore values from a checkpoint.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34354126,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-18T11:02:00.700","AcceptedAnswerId":34360715.0,"OwnerUserId":2576346.0,"Title":"Tensorflow MaxPool doesn't accept float64","Body":"<p>I am using Tensorflow to train a CNN. I am currently basing my calculations on Float32 which is kind of default at the time of initialising variables.<\/p>\n\n<p>I guessed that by using as my dtype I can get more accurate results so I changed the initilization of my varibales as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But I get following error in the operation:\nI chacked the documentation and it accept value type as follow:\n<\/p>\n\n<p>But I get the following error. Is this a bug or I'm doing something wrong?<\/p>\n\n<p><\/p>\n\n<hr>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34360715","CreationDate":"2015-12-18T17:16:14.867","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Support for <code>tf.nn.max_pool()<\/code> for types other than single-precision floating point values is currently not implemented, and the documentation is incorrect. (I have updated it upstream, and it should appear on GitHub and the website soon.)<\/p>\n\n<p>The reason for the incompatibility is that TensorFlow has a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/maxpooling_op_gpu.cu.cc\" rel=\"nofollow\">specialized implementation of max-pooling<\/a> on GPUs for performance reasons, and we almost always use <code>tf.float32<\/code> when training deep networks, so there isn't equivalent support for the other types. It would be possible to add, so contributions are welcome: see the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/547\" rel=\"nofollow\">GitHub issue<\/a> for more discussion.<\/p>\n"}]}
{"QuestionId":34356828,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-18T13:27:57.563","AcceptedAnswerId":34360634.0,"OwnerUserId":3399066.0,"Title":"Tensorflow image reading empty","Body":"<p>This question is based on: <a href=\"https:\/\/stackoverflow.com\/questions\/33648322\/tensorflow-image-reading-display\">Tensorflow image reading &amp; display<\/a><\/p>\n\n<p>Following their code we have the following:<\/p>\n\n<pre><\/pre>\n\n<p>Why does  have no dimensions? (Dimension(3) is only because of the argument )<\/p>\n\n<p>Does this mean that the image is not properly loaded? ( does load that image).<\/p>\n","answers":[{"AnswerId":"34360634","CreationDate":"2015-12-18T17:11:18.193","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The image will be properly loaded, but TensorFlow doesn't have enough information to infer the image's shape until the op is run. This arises because <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#decode_jpeg\" rel=\"noreferrer\"><code>tf.image.decode_jpeg()<\/code><\/a> can produce tensors of different shapes (heights and widths), depending on the contents of the string tensor <code>value<\/code>. This enables you to build <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html\" rel=\"noreferrer\">input pipelines<\/a> using a collection of images with different sizes.<\/p>\n\n<p>The <code>Dimension(None)<\/code> in the shape means <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Dimension.value\" rel=\"noreferrer\">\"unknown\" rather than \"empty\"<\/a>.\nIf you happen to know that all images read by this operation will have the same size, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor.set_shape\" rel=\"noreferrer\"><code>Tensor.set_shape()<\/code><\/a> to provide this information, and doing so will help to validate the shapes of later parts of the graph:<\/p>\n\n<pre><code>my_img = tf.image.decode_jpeg(value, channels=3)    \nKNOWN_HEIGHT = 28\nKNOWN_WIDTH = 28\nmy_img.set_shape([KNOWN_HEIGHT, KNOWN_WIDTH, 3])\n\nprint(my_img)\n# Output: Tensor(\"DecodeJpeg:0\", shape=TensorShape([Dimension(28), Dimension(28), Dimension(3)]), dtype=uint8)\n<\/code><\/pre>\n"}]}
{"QuestionId":34357430,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-18T14:03:54.763","AcceptedAnswerId":34358522.0,"OwnerUserId":1908423.0,"Title":"Tensorflow: No shape function registered for standard op: ExtractGlimpse. Where do I add my code for the shape function?","Body":"<p>I am trying to build a tensorflow graph using . <\/p>\n\n<p>Unfortunately I think there is a bug in the API itself. I am receiving the error <\/p>\n\n<p>There is actually the following code in  :<\/p>\n\n<pre><\/pre>\n\n<p>For some reason this function isn't being used properly, however it's not entirely clear from the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/adding_an_op\/index.html#implement-a-shape-function-in-python\" rel=\"nofollow\">documentation<\/a> where exactly this function is called. <\/p>\n\n<p>In which python file is this function supposed to be called, and how does the call need to be used?<\/p>\n\n<p>Thanks in advance <\/p>\n","answers":[{"AnswerId":"34358522","CreationDate":"2015-12-18T15:05:36.830","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This looks like a bug in TensorFlow: the shape function is defined in the correct place, but the code in <code>attention_ops.py<\/code> is never executed, so the shape function is never registered.<\/p>\n\n<p>I will fix it upstream, but in the meantime you can fix it by adding the following line to your program:<\/p>\n\n<pre><code>from tensorflow.python.ops import attention_ops\n<\/code><\/pre>\n"}]}
{"QuestionId":34361030,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><mnist>","CreationDate":"2015-12-18T17:35:07.777","AcceptedAnswerId":34361223.0,"OwnerUserId":2411173.0,"Title":"TensorFlow create dataset from numpy array","Body":"<p>TensorFlow as build it a nice way to store data. This is for example used to store the MNIST data in the example:<\/p>\n\n<pre><\/pre>\n\n<p>Suppose to have a input and output numpy arrays. <\/p>\n\n<pre><\/pre>\n\n<p>How can I transform them in a  dataset? <\/p>\n\n<p>I want to use functions like <\/p>\n","answers":[{"AnswerId":"34361223","CreationDate":"2015-12-18T17:47:13.247","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>The Dataset object is only part of the MNIST tutorial, not the main TensorFlow library.<\/p>\n\n<p>You can see where it is defined here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a5d8217c4ed90041bea2616c14a8ddcf11ec8c03\/tensorflow\/examples\/tutorials\/mnist\/input_data.py\" rel=\"noreferrer\">GitHub Link<\/a><\/p>\n\n<p>The constructor accepts an images and labels argument so presumably you can pass your own values there.<\/p>\n"},{"AnswerId":"47111900","CreationDate":"2017-11-04T14:42:17.817","ParentId":null,"OwnerUserId":"8885688","Title":null,"Body":"<p>As a alternative, you may use the function <code>tf.train.batch()<\/code> to create a batch of your data and at the same time eliminate the use of <code>tf.placeholder<\/code>. Refer to the documentation for more details.<\/p>\n\n<pre><code>&gt;&gt;&gt; images = tf.constant(X, dtype=tf.float32) # X is a np.array\n&gt;&gt;&gt; labels = tf.constant(y, dtype=tf.int32)   # y is a np.array\n&gt;&gt;&gt; batch_images, batch_labels = tf.train.batch([images, labels], batch_size=32, capacity=300, enqueue_many=True)\n<\/code><\/pre>\n"},{"AnswerId":"49806414","CreationDate":"2018-04-12T21:52:04.590","ParentId":null,"OwnerUserId":"2543480","Title":null,"Body":"<p>Recently, Tensorflow added a feature to its dataset api to consume numpy array. See <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/datasets#consuming_numpy_arrays\" rel=\"nofollow noreferrer\">here<\/a> for details.<\/p>\n\n<p>Here is the snippet that I copied from there:<\/p>\n\n<pre><code># Load the training data into two NumPy arrays, for example using `np.load()`.\nwith np.load(\"\/var\/data\/training_data.npy\") as data:\n  features = data[\"features\"]\n  labels = data[\"labels\"]\n\n# Assume that each row of `features` corresponds to the same row as `labels`.\nassert features.shape[0] == labels.shape[0]\n\nfeatures_placeholder = tf.placeholder(features.dtype, features.shape)\nlabels_placeholder = tf.placeholder(labels.dtype, labels.shape)\n\ndataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder))\n# [Other transformations on `dataset`...]\ndataset = ...\niterator = dataset.make_initializable_iterator()\n\nsess.run(iterator.initializer, feed_dict={features_placeholder: features,\n                                          labels_placeholder: labels})\n<\/code><\/pre>\n"}]}
{"QuestionId":34361769,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-12-18T18:23:35.133","AcceptedAnswerId":34380411.0,"OwnerUserId":5485347.0,"Title":"Why do we need Theano reshape?","Body":"<p>I don't understand why do we need  function in <strong>Theano<\/strong>. It is said in the documentation:<\/p>\n\n<blockquote>\n  <p>Returns a view of this tensor that has been reshaped as in\n  numpy.reshape.<\/p>\n<\/blockquote>\n\n<p>As far as I understood,  is some entity that is used for computation graphs creation. And it is absolutely independent of shapes. For instance when you create your function you can pass there matrix 2x2 or matrix 100x200. As I thought reshape somehow restricts this variety. But it is not. Suppose the following example:<\/p>\n\n<pre><\/pre>\n\n<p>As I understood, it should give an error since I passed matrix 2x2 not 3x3, but it computes element-wise squares perfectly.<\/p>\n\n<p><em>So what is the <strong>shape<\/strong> of the theano tensor variable and where should we use it?<\/em><\/p>\n","answers":[{"AnswerId":"34380411","CreationDate":"2015-12-20T11:17:02.730","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>There is an error in the provided code though Theano fails to point this out.<\/p>\n\n<p>Instead of<\/p>\n\n<pre><code>f = theano.function([X_resh], Y)\n<\/code><\/pre>\n\n<p>you should really use<\/p>\n\n<pre><code>f = theano.function([X], Y)\n<\/code><\/pre>\n\n<p>Using the original code you are actually providing the tensor <em>after the reshape<\/em> so the reshape command never gets executed. This can be seen by adding<\/p>\n\n<pre><code>theano.printing.debugprint(f)\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>Elemwise{sqr,no_inplace} [id A] ''   0\n |&lt;TensorType(float64, matrix)&gt; [id B]\n<\/code><\/pre>\n\n<p>Note that there is no reshape operation in this compiled execution graph.<\/p>\n\n<p>If one changes the code so that <code>X<\/code> is used as the input instead of <code>X_resh<\/code> then Theano throws an error including the message<\/p>\n\n<blockquote>\n  <p>ValueError: total size of new array must be unchanged Apply node that\n  caused the error: Reshape{2}(X, TensorConstant{(2L,) of 3})<\/p>\n<\/blockquote>\n\n<p>This is expected because one cannot reshape a tensor with shape <code>(2, 2)<\/code> (i.e. 4 elements) into a tensor with shape <code>(3, 3)<\/code> (i.e. 9 elements).<\/p>\n\n<p>To address the broader question, we can use symbolic expressions in the target shape and those expressions can be functions of the input tensor's symbolic shape. Here's some examples:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor\n\nX = theano.tensor.matrix('X')\nX_vector = X.reshape((X.shape[0] * X.shape[1],))\nX_row = X.reshape((1, X.shape[0] * X.shape[1]))\nX_column = X.reshape((X.shape[0] * X.shape[1], 1))\nX_3d = X.reshape((-1, X.shape[0], X.shape[1]))\nf = theano.function([X], [X_vector, X_row, X_column, X_3d])\n\nfor output in f(numpy.array([[1, 2], [3, 4]])):\n    print output.shape, output\n<\/code><\/pre>\n"}]}
{"QuestionId":34362193,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-12-18T18:53:35.527","AcceptedAnswerId":34362271.0,"OwnerUserId":12981.0,"Title":"How to explicitly broadcast a tensor to match another's shape in tensorflow?","Body":"<p>I have three tensors,  in tensorflow,  and  are both of shape ,  is a binary tensor of shape .<\/p>\n\n<p>I want to select elements from either A or B based on the value of . The obvious tool is , however that does not have broadcasting semantics, so I need to first explicitly broadcast  to the same shape as A and B.<\/p>\n\n<p>This would be my first attempt at how to do this, but it doesn't like me mixing a tensor () into the shape list.  <\/p>\n\n<pre><\/pre>\n\n<p>What's the correct approach here?<\/p>\n","answers":[{"AnswerId":"44679609","CreationDate":"2017-06-21T14:56:25.120","ParentId":null,"OwnerUserId":"6639245","Title":null,"Body":"<p>Here's a dirty hack:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef broadcast(tensor, shape):\n    return tensor + tf.zeros(shape, dtype=tensor.dtype)\n\nA = tf.random_normal([20, 100, 10])\nB = tf.random_normal([20, 100, 10])\nC = tf.random_normal([20, 100, 1])\n\nC = broadcast(C, A.shape)\nD = tf.select(C, A, B)\n<\/code><\/pre>\n"},{"AnswerId":"46300920","CreationDate":"2017-09-19T12:49:35.943","ParentId":null,"OwnerUserId":"6639245","Title":null,"Body":"<pre><code>import tensorflow as tf\n\ndef broadcast(tensor, shape):\n     \"\"\"Broadcasts ``x`` to have shape ``shape``.\n                                                                   |\n     Uses ``tf.Assert`` statements to ensure that the broadcast is\n     valid.\n\n     First calculates the number of missing dimensions in \n     ``tf.shape(x)`` and left-pads the shape of ``x`` with that many \n     ones. Then identifies the dimensions of ``x`` that require\n     tiling and tiles those dimensions appropriately.\n\n     Args:\n         x (tf.Tensor): The tensor to broadcast.\n         shape (Union[tf.TensorShape, tf.Tensor, Sequence[int]]): \n             The shape to broadcast to.\n\n     Returns:\n         tf.Tensor: ``x``, reshaped and tiled to have shape ``shape``.\n\n     \"\"\"\n     with tf.name_scope('broadcast') as scope:\n         shape_x = tf.shape(x)\n         rank_x = tf.shape(shape0)[0]\n         shape_t = tf.convert_to_tensor(shape, preferred_dtype=tf.int32)\n         rank_t = tf.shape(shape1)[0]\n\n         with tf.control_dependencies([tf.Assert(\n             rank_t &gt;= rank_x,\n             ['len(shape) must be &gt;= tf.rank(x)', shape_x, shape_t],\n             summarize=255\n         )]):\n             missing_dims = tf.ones(tf.stack([rank_t - rank_x], 0), tf.int32)\n\n         shape_x_ = tf.concat([missing_dims, shape_x], 0)\n         should_tile = tf.equal(shape_x_, 1)\n\n         with tf.control_dependencies([tf.Assert(\n             tf.reduce_all(tf.logical_or(tf.equal(shape_x_, shape_t), should_tile),\n             ['cannot broadcast shapes', shape_x, shape_t],\n             summarize=255\n         )]):\n             multiples = tf.where(should_tile, shape_t, tf.ones_like(shape_t))\n             out = tf.tile(tf.reshape(x, shape_x_), multiples, name=scope)\n\n         try:\n             out.set_shape(shape)\n         except:\n             pass\n\n         return out\n\nA = tf.random_normal([20, 100, 10])\nB = tf.random_normal([20, 100, 10])\nC = tf.random_normal([20, 100, 1])\n\nC = broadcast(C, A.shape)\nD = tf.select(C, A, B)\n<\/code><\/pre>\n"},{"AnswerId":"34362271","CreationDate":"2015-12-18T19:00:15.877","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>EDIT:<\/strong> In all versions of TensorFlow since 0.12rc0, the code in the question works directly. TensorFlow will automatically stack tensors and Python numbers into a tensor argument. The solution below using <code>tf.pack()<\/code> is only needed in versions prior to 0.12rc0. Note that <code>tf.pack()<\/code> was renamed to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"noreferrer\"><code>tf.stack()<\/code><\/a> in TensorFlow 1.0.<\/p>\n\n<hr>\n\n<p>Your solution is very close to working. You should replace the line:<\/p>\n\n<pre><code>C = tf.tile(C, [1,1,tf.shape(C)[2]])\n<\/code><\/pre>\n\n<p>...with the following:<\/p>\n\n<pre><code>C = tf.tile(C, tf.pack([1, 1, tf.shape(A)[2]]))\n<\/code><\/pre>\n\n<p>(The reason for the issue is that TensorFlow won't implicitly convert a list of tensors and Python literals into a tensor. <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#pack\" rel=\"noreferrer\"><code>tf.pack()<\/code><\/a> takes a list of tensors, so it will convert each of the elements in its input (<code>1<\/code>, <code>1<\/code>, and <code>tf.shape(C)[2]<\/code>) to a tensor. Since each element is a scalar, the result will be a vector.)<\/p>\n"}]}
{"QuestionId":34363549,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2015-12-18T20:35:59.237","AcceptedAnswerId":34363605.0,"OwnerUserId":2251083.0,"Title":"Tensorflow backprop through rnn ValueError","Body":"<p>I am using a recurrent neural network in tensorflow with BasicLSTMCells. Basically, I have an input sequence of word ids, I convert each id to word embeddings, pass the word embeddings one at a time through the rnn, and then make a prediction for a single word after reading the whole sequence. My embedding matrix is of dimension V x H where V is the size of my vocabulary and H is the number of hidden units in my rnn. In order to make a prediction for the next word, I multiply my hidden vector by a weight matrix of size H x V and then compute a softmax. With the setup I described, everything seems to work as expected. I'm able to train on some examples and make reasonable predictions.\nHowever, I've noticed that if I try to use the transpose of the embedding matrix, which will be a matrix of size H x V, instead of a separate matrix for the softmax layer, tensorflow raises a value error claiming that the dimensions of something it's not specifying don't have the same rank. I've verified that the dimensions of my embedding matrix (well, its transpose) are the same as those of the separate softmax matrix I'm creating. Changing just the one line of code from using my embedding matrix vs a separate softmax weight matrix causes the error.\nI created a relatively small program to demonstrate what I'm trying to do and to show what causes the error. I was not able to make the error occur on a smaller network when I tried with just a single hidden layer network. <\/p>\n\n<pre><\/pre>\n\n<p>The error I see is ValueError: Shapes TensorShape([Dimension(100)]) and TensorShape([Dimension(17), Dimension(100)]) must have the same rank\nwhen uncommenting the line I mentioned above. What is the cause of the error I'm seeing? Why is the embedding matrix not treated the same way as the matrix self.soft_w?<\/p>\n","answers":[{"AnswerId":"34363605","CreationDate":"2015-12-18T20:39:28.287","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The 0.6.0 (and earlier) release of TensorFlow had a bug in the implementation of gradients for <code>tf.nn.embedding_lookup()<\/code> and <code>tf.gather()<\/code> when the indices argument (<code>self._input_data<\/code> in your code) had more than one dimension.<\/p>\n\n<p>Upgrading to the latest source release should fix this error. Otherwise, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/f2eae4b3d27a4dc6d1f591f55a50fb3e1984a287\" rel=\"nofollow\">this commit<\/a> has the relevant change (to <code>array_grad.py<\/code>) that will enable your program to work.<\/p>\n"}]}
{"QuestionId":34364673,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-18T22:08:00.813","AcceptedAnswerId":34366060.0,"OwnerUserId":5696754.0,"Title":"Not seeing option of uploading graph file in Tensorboard using FireFox","Body":"<p>I am trying to get the graph using TensorBoard for MNIST example.\nI checked other tutorial, questions and it seems uploading file option is not coming under graph menu for me but I haven't found the reason for this.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/DGDI3.png\" alt=\"Screenshot of TensorBoard on my laptopre\"><\/p>\n\n<p>P.S. I am able to generate graph file using write_graph which I want to display on TensorBoard. <\/p>\n","answers":[{"AnswerId":"34366060","CreationDate":"2015-12-19T00:44:59.797","ParentId":null,"OwnerUserId":"604964","Title":null,"Body":"<p>Try using it on chrome instead of Firefox.  <\/p>\n\n<p>It isn't working on my Firefox too.  <\/p>\n\n<p>See this discussion for more info:<br>\n<a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/Lc7Y_6UEFOM\" rel=\"nofollow\">Is anyone else having trouble viewing the tensorboard\/graph tab in firefox? <\/a><\/p>\n"}]}
{"QuestionId":34366387,"AnswerCount":2,"Tags":"<loops><for-loop><lua><iterator><torch>","CreationDate":"2015-12-19T01:42:19.363","AcceptedAnswerId":34366613.0,"OwnerUserId":4013781.0,"Title":"Iterating over directories in Lua Torch","Body":"<p>In Torch, I am iterating over a folder full of subfolders as follows:<\/p>\n\n<pre><\/pre>\n\n<p>When I print subfolders, the subfolders appear to have been accessed in a random order. Instead, want to iterate over them in name order. How do I do this? Thanks!<\/p>\n","answers":[{"AnswerId":"34366613","CreationDate":"2015-12-19T02:22:42.267","ParentId":null,"OwnerUserId":"4013781","Title":null,"Body":"<p>Solved using the following: <\/p>\n\n<pre><code>subfolders = {}\ncounter = 0\n\nlocal dirs = paths.dir(root-directory)\ntable.sort(dirs)\n\nfor i = 1, 447 do\n    counter = counter + 1\n    subfolders[counter] = dirs[i]\nend\n<\/code><\/pre>\n"},{"AnswerId":"45620673","CreationDate":"2017-08-10T18:05:32.197","ParentId":null,"OwnerUserId":"1830334","Title":null,"Body":"<p>I needed a more robust solution than Chris's answer, which doesn't exclude plain files, parent directories (<code>..<\/code>) or the current directory (<code>.<\/code>). I'm also not sure what the magic number <code>447<\/code> in his code is. <a href=\"https:\/\/stackoverflow.com\/questions\/2833675\/using-lua-check-if-file-is-a-directory\">Standard Lua doesn't have a way to check if a file is a directory<\/a>, so this only works on Linux\/OSX.<\/p>\n\n<pre><code>function isSubdir(path)\n    noError, result = pcall(isDir, path)\n    if noError then return result else return false end\nend\n\n-- Credit: https:\/\/stackoverflow.com\/a\/3254007\/1830334\nfunction isDir(path)\n    local f = io.open(path, 'r')\n    local ok, err, code = f:read(1)\n    f:close()\n    return code == 21\nend\n\nfunction getSubdirs(rootDir)\n    subdirs = {}\n    counter = 0\n    local dirs = paths.dir(rootDir)\n    table.sort(dirs)\n    for i = 1, #dirs do\n        local dir = dirs[i]\n        if dir ~= nil and dir ~= '.' and dir ~= '..' then\n            local path = rootDir .. '\/' .. dir\n            if isSubdir(path) then\n                counter = counter + 1\n                subdirs[counter] = path\n            end\n        end\n    end\n    return subdirs\nend\n\nlocal subdirs = getSubdirs('\/your\/root\/path\/here')\nprint(subdirs)\n<\/code><\/pre>\n"}]}
{"QuestionId":34366798,"AnswerCount":1,"Tags":"<caffe><tensorflow>","CreationDate":"2015-12-19T03:00:55.193","AcceptedAnswerId":34401816.0,"OwnerUserId":5692514.0,"Title":"convert Caffe train.txt to Tensorflow","Body":"<p>I am trying to adapting my Caffe code to tensorflow. I wonder what is the best way to convert my train.txt and test.txt in order to work for tensorflow.<\/p>\n\n<p>In my train.txt, the file looks like:<\/p>\n\n<pre><\/pre>\n\n<p>The first column is the image name and the second column is the class label<\/p>\n\n<p>Thanks!!<\/p>\n","answers":[{"AnswerId":"34401816","CreationDate":"2015-12-21T18:07:52.640","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I'm assuming that you want to obtain a batch of identically-sized images with numeric labels. We'll use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#decode_csv\" rel=\"nofollow\"><code>tf.decode_csv()<\/code><\/a> to parse the text, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#read_file\" rel=\"nofollow\"><code>tf.read_file()<\/code><\/a> to load the JPEG data as a string, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#decode_jpeg\" rel=\"nofollow\"><code>tf.image.decode_jpeg()<\/code><\/a> to parse it into a dense tensor, and finally <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#batch\" rel=\"nofollow\"><code>tf.train.batch()<\/code><\/a> to build the parsed data into a batch of images. Many of these functions have a lot of options to configure, so see the documentation for further customization details.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Set options here for whether to repeat, etc.\nfilename_producer = tf.string_input_producer([\"train.txt\"], ...)\n\n# Read lines from the file, one at a time.\nline_reader = tf.TextLineReader()\nnext_line = line_reader.read(filename_producer)\n\n# Parse line into a filename and an integer label.\nimage_filename, label = tf.decode_csv(\n    next_line, [tf.constant([], dtype=tf.string), tf.constant([], dtype=tf.int32)],\n    field_delim=\" \")\n\n# Read the image as a string.\nimage_bytes = tf.read_file(image_filename)\n\n# Convert the image into a 3-D tensor (height x width x channels).\nimage_tensor = tf.image.decode_jpeg(image_bytes, ...)\n\n# OPTIONAL: Resize your image to a standard size if they are not already.\nHEIGHT = ...\nWIDTH = ...\nimage_tensor = tf.image.resize_image_with_crop_or_pad(image_tensor, HEIGHT, WIDTH)\n\n# Create a batch of images.\nBATCH_SIZE = 32\nimages, labels = tf.train.batch([image_tensor, label], BATCH_SIZE, ...)\n\n# [...build the rest of your model...]\n<\/code><\/pre>\n\n<p>This example makes extensive use of TensorFlow prefetching to load the examples. The TensorFlow documentation has a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html\" rel=\"nofollow\">how-to that explains how to use the prefetching feature<\/a>, but the most important thing to note is that you must call <code>tf.train.start_queue_runners()<\/code> at the start of your session to begin prefetching:<\/p>\n\n<pre><code>sess = tf.Session()\n\n# You must execute this statement to begin prefetching data.\ntf.train.start_queue_runners(sess)\n<\/code><\/pre>\n"}]}
{"QuestionId":34369522,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2015-12-19T10:08:27.063","AcceptedAnswerId":34380243.0,"OwnerUserId":4393339.0,"Title":"Dynamically creating symbolic expressions in theano","Body":"<p>I am trying to implement a cost function in theano for a feed forward neural network with multiple hidden layers. The cost function is<\/p>\n\n<pre><\/pre>\n\n<p>However I decide the number of hidden at runtime through a constructor to the network class. So the number of Ws and bs is decided at runtime and hence the expression for cost has to be created at runtime. I can compute the sums of Ws and bs outside the theano function and simply pass the scalar values. But I need the symbolic expression for computing gradients later. How do I make symbolic expression at runtime? <\/p>\n","answers":[{"AnswerId":"34380243","CreationDate":"2015-12-20T10:57:13.920","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use regular Python loops to construct the cost for a dynamic number of layers. Note that Theano 'run time' and Python 'run time' are two different things. Theano's 'compile time' happens during Python's 'run time' so you can use Python code to construct dynamic Theano expressions that depend on parameters known only when the Python code is running.<\/p>\n\n<p>The cost you give is only L2 regularization of the network parameters. You presumably have additional components for the full cost. Here's a full example.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile(input_size, hidden_sizes, output_size, reg_lambda, learning_rate):\n    ws, bs = [], []\n    x = tt.matrix('x')\n    x.tag.test_value = numpy.random.standard_normal(size=(2, input_size))\\\n        .astype(theano.config.floatX)\n    previous_size = input_size\n    h = x\n    for hidden_size in hidden_sizes:\n        w = theano.shared(\n                numpy.random.standard_normal(size=(previous_size, hidden_size))\n                          .astype(theano.config.floatX))\n        b = theano.shared(numpy.zeros((hidden_size,), dtype=theano.config.floatX))\n        h = tt.tanh(tt.dot(h, w) + b)\n        ws.append(w)\n        bs.append(b)\n        previous_size = hidden_size\n    w = theano.shared(numpy.random.standard_normal(size=(previous_size, output_size))\n                      .astype(theano.config.floatX))\n    b = theano.shared(numpy.zeros((output_size,), dtype=theano.config.floatX))\n    y = tt.nnet.softmax(tt.dot(h, w) + b)\n    ws.append(w)\n    bs.append(b)\n    z = tt.ivector('z')\n    z.tag.test_value = numpy.random.randint(output_size, size=(2,))\n    cost = tt.nnet.categorical_crossentropy(y, z).mean()\n    for w, b in zip(ws, bs):\n        cost += tt.sum(w ** 2) * reg_lambda\n        cost += tt.sum(b ** 2) * reg_lambda\n    updates = [(p, p - learning_rate * tt.grad(cost, p)) for p in ws + bs]\n    return theano.function([x, z], outputs=[cost], updates=updates)\n\n\ntheano.config.compute_test_value = 'raise'\ncompile(10, [8, 6, 4, 8, 16], 32, 0.1, 0.01)\n<\/code><\/pre>\n\n<p>Note the second <code>for<\/code> loop which adds L2 regularization components to the <code>cost<\/code> for each of the layers. The number of layers is passed as a parameter to the function.<\/p>\n"}]}
{"QuestionId":34370040,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-19T11:14:31.313","AcceptedAnswerId":34377778.0,"OwnerUserId":1259390.0,"Title":"How to find duplicated elements in a 1D Tensor","Body":"<p>I want to get elements that appear more than one time in a 1D tensor. Precisely, I want to create a function that does the opposite of . For example if  I need the output to be  and at the same time also retrieve the index of those elements in the source tensor.\nMy final goal is to get examples in a batch that labels appear more than one times. <\/p>\n","answers":[{"AnswerId":"34377778","CreationDate":"2015-12-20T03:47:33.370","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can do it using existing Tensorflow operations in a slightly round-about way, by counting the unique items to create a dense set of indexes of the unique items, and then counting them using <code>tf.unsorted_segment_sum<\/code>.  Once you have the count, select the items with <code>&gt; N<\/code> using <code>tf.greater<\/code>, and gather them back into a dense list:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant([8, 7, 8, 1, 3, 4, 5, 9, 5, 0, 5])\ninit = tf.initialize_all_variables()\n\nunique_a_vals, unique_idx = tf.unique(a)\ncount_a_unique = tf.unsorted_segment_sum(tf.ones_like(a),                   \n                                         unique_idx,                        \n                                         tf.shape(a)[0])                    \n\nmore_than_one = tf.greater(count_a_unique, 1)                               \nmore_than_one_idx = tf.squeeze(tf.where(more_than_one))                     \nmore_than_one_vals = tf.squeeze(tf.gather(unique_a_vals, more_than_one_idx))\n\n# If you want the original indexes:                                         \nnot_duplicated, _ = tf.listdiff(a, more_than_one_vals)                      \ndups_in_a, indexes_in_a = tf.listdiff(a, not_duplicated)                    \n\nwith tf.Session() as s:                                                     \n    s.run(init)                                                             \n    a, dupvals, dupidxes, dia = s.run([a, more_than_one_vals,                    \n                                  indexes_in_a, dups_in_a])                            \n    print \"Input: \", a                                                      \n    print \"Duplicate values: \", dupvals                                     \n    print \"Indexes of duplicates in a: \", dupidxes\n    print \"Dup vals with dups: \", dia\n<\/code><\/pre>\n\n<blockquote>\n  <p>Input:  [8 7 8 1 3 4 5 9 5 0 5]<\/p>\n  \n  <p>Duplicate values:  [8 5]<\/p>\n  \n  <p>Indexes of duplicates in a:  [ 0  2  6  8 10]<\/p>\n  \n  <p>Dup vals with dups:  [8 8 5 5 5]<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":34371896,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2015-12-19T14:36:25.027","AcceptedAnswerId":34372717.0,"OwnerUserId":2594660.0,"Title":"Tensorflow Tutorial: Duplicated Shuffling in the Input Pipeline","Body":"<p>In the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#batching\" rel=\"noreferrer\">Tensorflow reading data tutorial<\/a> an example input pipeline is given. In that pipeline the data is shuffled twice, inside the  as well as in the . Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>Does the second shuffle serve any useful purpose? The shuffle batch generator has the disadvantage that  examples are always stored pre-loaded in the memory to allow a useful shuffle. I do have image data which is quite heavy in memory consumption. That is why I am considering to use a  instead. Is there any advantage in shuffling the data twice?<\/p>\n\n<p>Edit: Additional Question, why is the  initialized only with default capacity of 32? Wouldn't it be advantageous to have a multiple of batch_size as capacity?<\/p>\n","answers":[{"AnswerId":"34372717","CreationDate":"2015-12-19T16:12:58.170","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Yes - this is a common pattern, and it's shown in the most general way.  The <code>string_input_producer<\/code> shuffles the order in which the data files are read.  Each data file typically contains many examples, for efficiency.  (Reading a million small files is very slow;  it's better to read 1000 large files with 1000 examples each.)<\/p>\n\n<p>Therefore, the examples from the files are read into a shuffling queue, where they are shuffled at a much finer granularity, so that examples from the same file aren't always trained in the same order, and to get mixing across the input files.<\/p>\n\n<p>For more details, see <a href=\"https:\/\/stackoverflow.com\/questions\/34258043\/getting-good-mixing-with-many-input-datafiles-in-tensorflow\/34258214#34258214\">Getting good mixing with many input datafiles in tensorflow<\/a><\/p>\n\n<p>If your files each contain only one input example, you don't need to shuffle multiple times and could get away with only a <code>string_input_producer<\/code>, but note that you still will likely benefit from having a queue that holds a few images after reading, so that you can overlap the inputting and training of your network.  The <code>queue_runner<\/code> for a <code>batch<\/code> or <code>shuffle_batch<\/code> will run in a separate thread, ensuring that the I\/O is happening in the background and that images are always available for training.  And, of course, it's typically nice for speed to create minibatches to train on.<\/p>\n"},{"AnswerId":"44868089","CreationDate":"2017-07-02T06:36:03.600","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Both of the shuffles serves different purposes and shuffle different things:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/string_input_producer\" rel=\"nofollow noreferrer\"><code>tf.train.string_input_producer<\/code><\/a> <em>shuffle: Boolean. If true, the strings are randomly shuffled within each epoch.<\/em>. So if you have a few files <code>['file1', 'file2', ..., 'filen']<\/code> this randomly selects a file from this list. If case of false, the files follow one after each other.<\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/shuffle_batch\" rel=\"nofollow noreferrer\"><code>tf.train.shuffle_batch<\/code><\/a> <em>Creates batches by randomly shuffling tensors.<\/em> So it takes <code>batch_size<\/code> tensors from your queue <code>read_my_file_format<\/code> and shuffles them.<\/li>\n<\/ul>\n\n<p>Because both shuffles do different things, there is an advantage to shuffle the data two times. Even if you use a batch of 256 images, and each image is 256x256 pixels, you will consume less than 100 Mb of memory. If at some point you will see memory problems, you can try decreasing your batch size.<\/p>\n\n<p>Regarding the default capacity - it is <a href=\"https:\/\/stackoverflow.com\/a\/44857104\/1090562\">model specific<\/a>. It makes sense to have it bigger than batch_size and to make sure it is never empty during the training.<\/p>\n"},{"AnswerId":"48325842","CreationDate":"2018-01-18T16:21:22.660","ParentId":null,"OwnerUserId":"7704013","Title":null,"Body":"<p>To answer the additional question, the <code>string_input_producer<\/code> returns a queue containing the <strong>name of the files<\/strong> containing the samples, not the samples themselves. This filenames are then used by <code>shuffle_batch<\/code> to load the data. The number of samples that are loaded is thus related to the <code>capacity<\/code> parameter of the <strong><code>shuffle_batch<\/code><\/strong> function, not <code>string_input_producer<\/code>. <\/p>\n"}]}
{"QuestionId":34374959,"AnswerCount":2,"Tags":"<text><tensorflow><text-classification>","CreationDate":"2015-12-19T20:15:33.863","AcceptedAnswerId":null,"OwnerUserId":4975751.0,"Title":"TensorFlow example for text classification - how to evaluate your own text?","Body":"<p>Does any one have full steps  and example for TensorFlow example for passing in your own text files and getting them evaluated against the existing model that comes with examples - using  as documented?<\/p>\n\n<p>Also, if I wanted to train on different input set of say 1000 text files of my own samples, and then use that model for new text files? I know there is documentation but is terse for someone who is not familiar with text classification process.<\/p>\n\n<p>I was able to run image example against my own images as that was only requiring to swap out one image  file name for myh new image file, but for text it seems to be more involved.<\/p>\n\n<p>Thanks <\/p>\n","answers":[{"AnswerId":"47028155","CreationDate":"2017-10-31T04:57:40.603","ParentId":null,"OwnerUserId":"8854376","Title":null,"Body":"<p>You need to load the vocabulary file saved during training and process your new text with that. See the eval.py file <a href=\"https:\/\/github.com\/dennybritz\/cnn-text-classification-tf\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Change the data parameters with your input text and proceed.<\/p>\n"},{"AnswerId":"37933552","CreationDate":"2016-06-20T23:51:03.087","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Here is an example: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/skflow\/text_classification.py\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/skflow\/text_classification.py<\/a><\/p>\n\n<p>You can set the flag test_with_fake_data to use the fake data in text_train.csv (training samples) and text_test.csv (testing samples) <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/data\/\" rel=\"nofollow\">here<\/a>. Next, you can modify these two files to include whatever data you'd like to have. You will need to do some preprocessing if your existing text files are in a different format.<\/p>\n"}]}
{"QuestionId":34375871,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-19T22:05:44.513","AcceptedAnswerId":34376231.0,"OwnerUserId":2272798.0,"Title":"Overriding device scope in Tensorflow","Body":"<p>How is device scope handled in the following situation, where an outer device scope is overridden by an inner device scope:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>My intuition is as follows:<\/p>\n\n<p>1) \"a\" is computed first on \"cpu:0\"<\/p>\n\n<p>2) \"b\" and \"c\" and computed in parallel on \"gpu:0\" and \"gpu:1\", respectively.<\/p>\n\n<p>3) \"d\" waits for \"b\" and \"c\" because it depends on them, and when their values become available, \"d\" gets computed on \"cpu:0\"<\/p>\n\n<p>Is my intuition correct?<\/p>\n","answers":[{"AnswerId":"34376240","CreationDate":"2015-12-19T22:53:39.177","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Yes.<\/p>\n\n<p>PS, to check your intution, you could do something like this<\/p>\n\n<pre><code>with tf.device(\"\/cpu:0\"):\n  a = tf.placeholder(dtype=tf.int32, name=\"a\")\n  with tf.device(\"\/gpu:0\"):\n    b = tf.placeholder(dtype=tf.int32, name=\"b\")\n    with tf.device(\"\/gpu:1\"):\n      c = tf.placeholder(dtype=tf.int32, name=\"c\")\n      d = a+b+c\n\nprint d.graph.as_graph_def()\n<\/code><\/pre>\n\n<p>This gives the underlying graph definition that TensorFlow system will run<\/p>\n"},{"AnswerId":"34376231","CreationDate":"2015-12-19T22:52:41.903","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Mostly, with a few subtle points:<\/p>\n\n<p>(a)  \"b\" and \"c\" <em>could<\/em> be computed in parallel, provided there are no control flow dependencies or data dependencies in what they're doing.  But whether they actually <em>are<\/em> executed truly at the same time is unpredictable from this example.  (I assume that was already obvious, but I wanted to be sure it was to others who might read this later.)<\/p>\n\n<p>Note also that as specified, <code>b<\/code> and <code>c<\/code> don't explicitly depend on <code>a<\/code>, so it's possible that all three of them could be executed concurrently.  It is not the case that <code>a<\/code> <em>must<\/em> be executed first.<\/p>\n\n<p>(b)  By default, if you don't supply any configuration options, device placement is \"soft\" -- the runtime can override things if the op can't be executed on the specific device.  For example, a CPU-only op could be moved from a GPU back to <code>\/cpu:0<\/code>;  or an op pinned to <code>\/gpu:1<\/code> could be moved to <code>\/gpu:0<\/code> if the graph was run on a machine that had only a single GPU.<\/p>\n\n<p>You can control the hard-vs-soft placement by supplying a configuration to the <code>tf.Session<\/code>:<\/p>\n\n<pre><code>with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)):\n<\/code><\/pre>\n"}]}
{"QuestionId":34376859,"AnswerCount":1,"Tags":"<c++><database><caffe><lmdb>","CreationDate":"2015-12-20T00:32:57.477","AcceptedAnswerId":null,"OwnerUserId":4358739.0,"Title":"writing to LMDB failed","Body":"<p>I was trying write data to lmdb database (took reference from caffe\/tools\/convert_imageset.cpp). There were no errors in console but the file size of .mdb file remains zero i.e no data is written to the file but data and lock file are created.\nThe part of the code that I tried is  : <\/p>\n\n<pre><\/pre>\n\n<p>I was not able to figure out if I am missing something. Any references to how LMBD works would be helpful. <\/p>\n","answers":[{"AnswerId":"37646072","CreationDate":"2016-06-05T19:46:03.333","ParentId":null,"OwnerUserId":"5129319","Title":null,"Body":"<p>Make sure the key is not empty before calling <code>Put<\/code>.<br>\nYou could use <code>mdb_dump<\/code> and print the contents of the database.<\/p>\n"}]}
{"QuestionId":34376944,"AnswerCount":1,"Tags":"<tensorflow><one-hot-encoding>","CreationDate":"2015-12-20T00:48:13.807","AcceptedAnswerId":null,"OwnerUserId":5676182.0,"Title":"adding one hot encoding throws error in previously working code in Tensorflow","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>If I introduce one hot encoding on the logits here, the program gives error later although both returns the same dimensions. <strong>If I comment out this line, the program does not give any error.<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Rest of code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>---> 23     grads,norm = tf.clip_by_global_norm(tf.gradients(cost,tvars),5)<\/p>\n\n<blockquote>\n  <p>ValueError: List argument 'values' to 'Pack' Op with length 0 shorter\n  than minimum length 1.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34378049","CreationDate":"2015-12-20T04:44:39.650","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Neural networks can only be trained if all the operations they perform are differentiable. The \"one-hot\" step you apply is not differentiable, and hence such a neural network cannot be trained using any gradient descent-based optimizer (= any optimizer that tensor flow implements).<\/p>\n\n<p>The general approach is to use <code>softmax<\/code> (which is differentiable) during training to approximate one-hot encoding (and your model already has <code>softmax<\/code> following computing logits, so commenting out the \"one-hot\" is actually all you need to do).<\/p>\n"}]}
{"QuestionId":34377797,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2015-12-20T03:52:46.597","AcceptedAnswerId":34389676.0,"OwnerUserId":2116766.0,"Title":"In Tensorflow's C++ TensorShape API, what is the equivalent of Python's None?","Body":"<p>Suppose I have a tensor created with the Python API of Tensorflow, as follows,<\/p>\n\n<pre><\/pre>\n\n<p>I would like to create a tensorflow::Tensor equivalent of that in C++, so that I can run a trained graph that takes x as the input. What should I do with the first dimension of the input shape, which in C++ is of type tensorflow::TensorShape? <\/p>\n\n<p>If I do: <\/p>\n\n<pre><\/pre>\n\n<p>it doesn't seem to work, because num_elements becomes 0, which is not the value of inputLen that I expect. <\/p>\n","answers":[{"AnswerId":"34389676","CreationDate":"2015-12-21T05:50:02.507","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>Update<\/strong>: There is now a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a5074383617a9947f248a0ddd56b94f9fb0f970b\/tensorflow\/core\/framework\/partial_tensor_shape.h\" rel=\"nofollow noreferrer\"><code>tensorflow::PartialTensorShape<\/code><\/a> class that can represent shapes with unknown dimensions or an unknown rank. The value <code>-1<\/code> is used to represent an unknown value (i.e. what <code>None<\/code> represents in Python). It is used in the C++ shape inference code, and can be specified in a shape-typed attr or a <code>tf.TensorShape<\/code> proto.<\/p>\n\n<hr>\n\n<p><strong>TL;DR:<\/strong> There is no equivalent in C++ because the C++ part of TensorFlow only checks shapes at runtime, when they are fully defined; whereas the Python part checks shapes at graph-construction time, when they might not be fully defined.<\/p>\n\n<p>There is no equivalent of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Dimension\" rel=\"nofollow noreferrer\"><code>tf.Dimension(None)<\/code><\/a> (i.e., an unknown dimension) in the C++ <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassTensorShape.html#class-tensorflow-tensorshape\" rel=\"nofollow noreferrer\"><code>tensorflow::TensorShape<\/code><\/a> class. This is because the (C++) <code>tensorflow::TensorShape<\/code> class describes the shape of a (C++) <code>tensorflow::Tensor<\/code>, which represents a <strong>concrete<\/strong> value for a tensor, and therefore must have a fully defined shape. The Python <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor\" rel=\"nofollow noreferrer\"><code>tf.Tensor<\/code><\/a> class represents a <strong>symbolic<\/strong> tensor&mdash;representing the output of an operation that has yet to be run&mdash;and so it can have a shape that is unknown in one or more dimensions.<\/p>\n\n<p>If you are using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassSession.html#class-tensorflow-session\" rel=\"nofollow noreferrer\">C++ API<\/a> to feed a placeholder, you should simply create a new <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassTensor.html#class-tensorflow-tensor\" rel=\"nofollow noreferrer\"><code>tensorflow::Tensor<\/code><\/a> with a fully defined shape for each different value that you feed to the placeholder (in a <code>Session::Run()<\/code> call). Note however that the C++ API does not check shapes of placeholders, so you should manually ensure that the shape matches the expected shape for the placeholder.<\/p>\n\n<p>If you are building a graph using the C++ API, and you want to <em>define<\/em> a placeholder with an unknown size in one or more dimensions, you should define a placeholder node with its <code>shape<\/code> attr set to <code>tensorflow::TensorShape({})<\/code>. Although this is equivalent to a scalar, for historical reasons this is treated as the shape being completely unconstrained.<\/p>\n"}]}
{"QuestionId":34378258,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-12-20T05:27:22.567","AcceptedAnswerId":null,"OwnerUserId":5697948.0,"Title":"Can not import theano after installation","Body":"<p>The code is like this<\/p>\n\n<pre><\/pre>\n\n<p>I am sorry that I am still a freshman both in this site and linux.<\/p>\n","answers":[{"AnswerId":"34389279","CreationDate":"2015-12-21T05:08:01.500","ParentId":null,"OwnerUserId":"5697948","Title":null,"Body":"<p>That may due to the the two versions of python, so firstly I removed one and add the path for python to get the packages installed by pip, but it still does not help. Finally I came to to github to download it and unpackaged it, aftering test for around 30 minutes, it came out no errors, which means I setup it successfully.<\/p>\n"},{"AnswerId":"39095089","CreationDate":"2016-08-23T07:36:47.663","ParentId":null,"OwnerUserId":"6742292","Title":null,"Body":"<p>After removing the previous version, I installed it as follows:\n(<a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html<\/a>)<\/p>\n\n<pre><code>git clone git:\/\/github.com\/Theano\/Theano.git\ncd Theano\npython setup.py develop --user\ncd ..\n<\/code><\/pre>\n"}]}
{"QuestionId":34381208,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-20T12:56:30.987","AcceptedAnswerId":34383291.0,"OwnerUserId":4761675.0,"Title":"Session.run() \/Tensor.eval() of Tensorflow run for a crazy long time","Body":"<p>I am trying to learn tenforflow by following the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html\" rel=\"noreferrer\">Convolutional Neural Networks<\/a> tutorial, but when I was trying to figure out how  loads data from , I encountered a problem that  executes for a very long time or runs forever without a result.\nThe code is like this:<\/p>\n\n<pre><\/pre>\n\n<p>The code is basically from  and the file  is\nextracted from \u3002<\/p>\n\n<p>Normally, I can observe a tensor using its  method. But in this case it runs continually for a longer time than ever\uff08I waited for nearly an hour and it was still running). Is there something wrong in my code? <\/p>\n","answers":[{"AnswerId":"34383291","CreationDate":"2015-12-20T16:43:34.293","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>1)  As a basic sanity check:  <code>ls -al \/Users\/me\/Downloads\/cifar-10-batches-bin\/data_batch_1.bin<\/code><\/p>\n\n<p>2) Don't forget to:<\/p>\n\n<pre><code>init = tf.initialize_all_variables()\nsess.run(init)\n<\/code><\/pre>\n\n<p>3) <code>tf.train.start_queue_runners()<\/code>  (after creating your session)<\/p>\n\n<p>It's probably #3.  The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"noreferrer\"><code>string_input_producer<\/code><\/a> adds a queue runner to the <code>QUEUE_RUNNERS<\/code> collection, which needs to be started.<\/p>\n"}]}
{"QuestionId":34388428,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-12-21T03:11:12.223","AcceptedAnswerId":null,"OwnerUserId":5697948.0,"Title":"Error when running deepmind","Body":"<p>It took me two days to install the requirements of deepQ(python version),then I tried to run it today but I faced this problem, and the code are as followed.<\/p>\n\n<pre><\/pre>\n\n<p>I have already tested theano ,numpy, scipy and there was no errors coming out. But when I ran it, it said dnn not available. So I came to find dnn, and the code is like this <\/p>\n\n<pre><\/pre>\n\n<p>Just hope someone can help me.<\/p>\n","answers":[{"AnswerId":"34392653","CreationDate":"2015-12-21T09:40:06.237","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>Did you install CUDA and cuDNN? \nLasagne is build on top of Theano and is, in some cases, relying on cuda code (e.g. here), rather than abstracting it away.<\/p>\n\n<p>This can be seen from the import:<\/p>\n\n<pre><code>from theano.sandbox.cuda import dnn\n<\/code><\/pre>\n\n<p>Also see: <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/issues\/242\" rel=\"nofollow\">https:\/\/github.com\/Lasagne\/Lasagne\/issues\/242<\/a><\/p>\n\n<p>To get cuDNN you need to register at NVidia as a developer, see:\n<a href=\"https:\/\/developer.nvidia.com\/accelerated-computing\" rel=\"nofollow\">https:\/\/developer.nvidia.com\/accelerated-computing<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n\n<p>Cheers,\nMichael<\/p>\n"}]}
{"QuestionId":34389945,"AnswerCount":1,"Tags":"<multithreading><tensorflow>","CreationDate":"2015-12-21T06:19:49.903","AcceptedAnswerId":34390517.0,"OwnerUserId":829624.0,"Title":"Changing the number of threads in TensorFlow on Cifar10","Body":"<p>Whenever I run the cifar10_eval.py, in creates 32 threads as following:<\/p>\n\n<p><strong>I tensorflow\/core\/common_runtime\/local_device.cc:25] Local device intra op parallelism threads: 32<\/strong><\/p>\n\n<p>I think this number of threads is the number of threads running on CPUs, but when I check the usage, only 400-500% of CPUs are used. Is there anyway to change this number of threads?<\/p>\n","answers":[{"AnswerId":"34390517","CreationDate":"2015-12-21T07:09:31.683","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To configure this value, you can pass a <code>tf.ConfigProto<\/code> argument when constructing the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/Session\" rel=\"noreferrer\"><code>tf.Session<\/code><\/a>:<\/p>\n\n<pre><code>NUM_THREADS = \u2026\nsess = tf.Session(config=tf.ConfigProto(\n    intra_op_parallelism_threads=NUM_THREADS))\n<\/code><\/pre>\n"}]}
{"QuestionId":34391581,"AnswerCount":0,"Tags":"<python><numpy><theano><conv-neural-network>","CreationDate":"2015-12-21T08:31:21.690","AcceptedAnswerId":null,"OwnerUserId":2845263.0,"Title":"Indexing error in Theano for 3D matrices","Body":"<p>I am trying to build a 3D CNN in Theano and I am getting a index of bound error when I am running this nested loops :<\/p>\n\n<pre><\/pre>\n\n<p>Where  = 61 and  = 20 and  is the following definition:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this code I am getting the following errors : <\/p>\n\n<pre><\/pre>\n\n<p>If anyone is interested in the entire code it can be found <a href=\"https:\/\/github.com\/abhijith2693\/iitmFinalYearProject\/blob\/master\/CNNGlioma\/runCNNGlioma.py\" rel=\"nofollow\">here<\/a><\/p>\n","answers":[]}
{"QuestionId":34391704,"AnswerCount":0,"Tags":"<model><neural-network><caffe>","CreationDate":"2015-12-21T08:40:49.203","AcceptedAnswerId":null,"OwnerUserId":2693257.0,"Title":"How to training\/testing my custom model (net) in caffe?","Body":"<p>I just started using caffe and was amazed by its ease for training and testing models. I just wanted to train my own model( net) and keep using it.<\/p>\n\n<p>For more explanations, I have a Specific multi layer model like as bellow picture. (My input data is breast-cancer data set.):<br>\n <img src=\"https:\/\/i.stack.imgur.com\/fF3E8.png\" alt=\"My specific multi layer model, Click it please\"><\/p>\n\n<p>(This model is not fully connected neural network, and some weights are omitted as you can see in picture.)<\/p>\n\n<p><strong>My certain question is, how to implement this model in caffe?!<\/strong><\/p>\n\n<p>Many thanks!!!! Any advice would be appreciated!<\/p>\n","answers":[]}
{"QuestionId":34393876,"AnswerCount":2,"Tags":"<python><convolution><deep-learning><tensorflow><conv-neural-network>","CreationDate":"2015-12-21T10:47:24.820","AcceptedAnswerId":34652349.0,"OwnerUserId":2576346.0,"Title":"Difference between local and dense layers in CNNs","Body":"<p>What is the difference between a \"Local\" layer and a \"Dense\" layer in a convolutional neural network?  I am trying to understand the CIFAR-10 code in TensorFlow, and I see it uses \"Local\" layers instead of regular dense layers. Is there any class in TF that supports implementing \"Local\" layers?<\/p>\n","answers":[{"AnswerId":"34568528","CreationDate":"2016-01-02T17:18:52.997","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>I am quoting user2576346's comments under the question: <\/p>\n\n<blockquote>\n  <p>As I understand, either it should be densely connected or be a convolutional layer ...<\/p>\n<\/blockquote>\n\n<p>No this is not true. A more accurate way to phrase that statement would be that layers are either fully connected (dense) or locally connected.<\/p>\n\n<p>A convolutional layer is an example of a locally connected layer. In general a locally connected layer is a layer in which each of its units is only connected to a local portion of the input. A convolutional layer is a special type of local layer which exhibits a spatial translation invariance as each convolutional feature detector is strided across the entire image in local receptive windows, e.g. of size 3x3 or 5x5 for example.<\/p>\n"},{"AnswerId":"34652349","CreationDate":"2016-01-07T10:07:58.697","ParentId":null,"OwnerUserId":"1716303","Title":null,"Body":"<p>Quoting from <a href=\"https:\/\/code.google.com\/p\/cuda-convnet\/wiki\/LayerParams#Locally-connected_layer_with_unshared_weights\" rel=\"noreferrer\">cuda-convnet<\/a>:<\/p>\n\n<blockquote>\n  <p>Locally-connected layer with unshared-weight: This kind of layer is just like a convolutional layer, but without any weight-sharing. That is to say, a different set of filters is applied at every (x, y) location in the input image. Aside from that, it behaves exactly as a convolutional layer.<\/p>\n<\/blockquote>\n\n<p>In the TensorFlow CIFAR-10 example, although the two layers are named <code>local3<\/code> and <code>local4<\/code>, they are actually fully-connected layer, not locally-connected layer as specified in <a href=\"https:\/\/code.google.com\/p\/cuda-convnet\/wiki\/LayerParams#Locally-connected_layer_with_unshared_weights\" rel=\"noreferrer\">cuda-convnet<\/a> (you can see that the output from <code>pool2<\/code> is flattened into the input of <code>local3<\/code> layer). <\/p>\n"}]}
{"QuestionId":34395942,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2015-12-21T12:40:43.083","AcceptedAnswerId":34414172.0,"OwnerUserId":1619749.0,"Title":"Elementwise updates of shared matrix in Theano","Body":"<p>I am trying to calculate the gradient of the energy function, E( phi, theta, psi), where phi, theta, psi are Euler angles around Z, Y and X axis respectively. \nR is a rotation matrix converted from Euler angles. \nWhere the columns are rotation vectors.\nFor a point cloud of a indoor room, I have normals N for each point in the point cloud. \nN has shape of (numPoints x 3)<\/p>\n\n<p>Now E = sum(1-max(N.dot(R))<\/p>\n\n<p>I would like to find the gradient of E with respect to Euler angles.<\/p>\n\n<p>To begin with the problem I am trying to calculate rotation matrix R from Euler angles using Theano.\nPlease find the code below.<\/p>\n\n<pre><\/pre>\n\n<p>But this gives me error at two different stage, <\/p>\n\n<pre><\/pre>\n\n<p>If I chage the Eurler angles to scalar values I get error as below<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone please help me with this problem?<\/p>\n","answers":[{"AnswerId":"34414172","CreationDate":"2015-12-22T10:53:18.757","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The first error is exactly as the message indicates: you can't supply shared variables as inputs to Theano functions. Either use shared variables with no input parameters or regular tensors as input parameters.<\/p>\n\n<p>The second error is due to incorrect <code>updates<\/code>. Updates should be a list of pairs. Each pair should consist of the shared variable to be updated and a symbolic expression describing how the update should be computed.<\/p>\n\n<p>In this case the update might be computed like this (untested):<\/p>\n\n<pre><code>Rnew = T.set_subtensor(R[0, 0],  T.cos(theta)*T.cos(phi))\nRnew = T.set_subtensor(Rnew[1, 0],  T.cos(theta)*T.sin(phi))\nRnew = T.set_subtensor(Rnew[2, 0],  -T.sin(theta))\n\nRnew = T.set_subtensor(Rnew[0, 1],  T.sin(psi)*T.sin(theta)*T.cos(phi) - T.cos(psi)*T.sin(phi))\nRnew = T.set_subtensor(Rnew[1, 1],  T.sin(psi)*T.sin(theta)*T.sin(phi) + T.cos(psi)*T.cos(phi))\nRnew = T.set_subtensor(Rnew[2, 1],  T.sin(psi)*T.cos(theta))\n\nRnew = T.set_subtensor(Rnew[0, 2],  T.cos(psi)*T.sin(theta)*T.cos(phi) + T.sin(psi)*T.sin(phi))\nRnew = T.set_subtensor(Rnew[1, 2],  T.cos(psi)*T.sin(theta)*T.sin(phi) - T.sin(psi)*T.cos(phi))\nRnew = T.set_subtensor(Rnew[2, 2],  T.cos(psi)*T.cos(theta))\n\nf = theano.function([phi, theta, psi], updates=[(R, Rnew)])\n<\/code><\/pre>\n"}]}
{"QuestionId":34397120,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><theano>","CreationDate":"2015-12-21T13:42:27.660","AcceptedAnswerId":34453371.0,"OwnerUserId":4393339.0,"Title":"MLP classifier in theano settles at local minima","Body":"<p>I wrote an MLP classifier using theano. The training function using back propagation algorithm is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to train the classifier for the XOR problem. The implementation is <\/p>\n\n<pre><\/pre>\n\n<p>The initialize() method just compiles all the functions at the backend i.e. the back propagation function,a forward pass function for calculating the predictions and a few other theano functions. Now, when i run this code, the training settles at a local minima. <\/p>\n\n<pre><\/pre>\n\n<p>At the start of the training, the loss was about 0.92. It steadily decreased upto the above value and stopped there. I tried changing the values of alpha and the momentum. What am i doing wrong?<\/p>\n\n<p>P.S.\nThe whole code is here:\nnetworks.py<\/p>\n\n<pre><\/pre>\n\n<p>layers.py<\/p>\n\n<pre><\/pre>\n\n<p>backend.py<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34453371","CreationDate":"2015-12-24T13:14:39.407","ParentId":null,"OwnerUserId":"4393339","Title":null,"Body":"<p>Finally figured it out! In the NetworkBackend, while calculating the cost, I am calculating the cross entropy between the expected outputs and the predictions passed as arguments to the theano function instead of using the predictions computed by the activate function. Thus, the theano graph does not contain the forward pass. Consequently, theano.tensor.grad only finds the gradient for the regularization function and not the actual cost function! so the proper implementation should be:<\/p>\n\n<pre><code>inputs=T.dmatrix()\ntemp=self.layers[0].forwardPass(inputs)\nfor i in range(1,len(self.layers[:-1])):\n    temp=self.layers[i].forwardPass(T.dot\n    (temp,self.weights[i-1].transpose())+self.bias[i-1])\n    output=self.layers[-1].forwardPass(T.dot(temp,self.weights[-1].\n    transpose())+self.bias[-1])\nself.activate=theano.function([inputs],output)\n\nlabel=T.dmatrix()\nreg_lambda=T.dscalar()\nalpha=T.dscalar()\nmomentum=T.dscalar()\ncost=T.nnet.categorical_crossentropy(output,label).mean()\nfor i,j in zip(self.weights,self.bias):\n    cost+=T.sum(i**2)*reg_lambda\n    cost+=T.sum(j**2)*reg_lambda\nparameters=self.weights+self.bias\nrates=[(alpha*T.grad(cost,parameter)+momentum*prev_rate) \nfor parameter,prev_rate in zip(parameters,self.prev_rate)]\nupdates=[(weight,weight-rate) for weight,rate\n in zip(parameters,rates)]+[(prev_rate,rate) \nfor prev_rate,rate in zip(self.prev_rate,rates)]\nself.backprop=theano.function([inputs,label,reg_lambda,alpha,momentum],\ncost,updates=updates)\n<\/code><\/pre>\n\n<p>so instead of declaring a new matrix for predictions, i am taking the inputs and computing the predictions in the train function using the same equation used in the activation function. This completes the theano graph and the theano.tensor.grad() now calculates the gradient for the cost function as well as the reqularization.<\/p>\n"},{"AnswerId":"34415240","CreationDate":"2015-12-22T11:46:57.637","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This is probably caused by parameter initialization.  The following code example implements a basic XOR learner using a neural network with a single hidden layer.<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile(input_size, hidden_size):\n    w_h = theano.shared(numpy.random.standard_normal(size=(input_size, hidden_size)).astype(theano.config.floatX))\n    b_h = theano.shared(numpy.zeros((hidden_size,), dtype=theano.config.floatX))\n    w_y = theano.shared(numpy.zeros((hidden_size,), dtype=theano.config.floatX))\n    b_y = theano.shared(numpy.zeros(1, dtype=theano.config.floatX), broadcastable=(True,))\n    x = tt.matrix()\n    z = tt.ivector()\n    learning_rate = tt.scalar()\n    h = tt.tanh(tt.dot(x, w_h) + b_h)\n    y = tt.nnet.sigmoid(tt.dot(h, w_y) + b_y)\n    cost = tt.nnet.binary_crossentropy(y, z).mean()\n    updates = [(p, p - learning_rate * tt.grad(cost, p)) for p in [w_h, b_h, w_y, b_y]]\n    return theano.function([x, z, learning_rate], outputs=cost, updates=updates), theano.function([x], outputs=y)\n\n\ndef main():\n    numpy.random.seed(5)\n    train, test = compile(2, 2)\n    for _ in xrange(100000):\n        print train([[1, 1], [1, 0], [0, 1], [0, 0]], [0, 1, 1, 0], 0.1)\n    print test([[1, 1], [1, 0], [0, 1], [0, 0]])\n\n\nmain()\n<\/code><\/pre>\n\n<p>Note the random number generator seed value. With a seed of <code>5<\/code> the learner converges on a good solution and looks like it's tending towards a perfect solution, given enough time. However, if the seed is changed to <code>1<\/code> the network gets stuck in a local optimum; it is able to discriminate the second dimension but not the first.<\/p>\n\n<p>Different random initialization methods may yield better results, i.e. be less sensitive to the RNG seed.<\/p>\n"}]}
{"QuestionId":34398328,"AnswerCount":1,"Tags":"<sequence><tensorflow>","CreationDate":"2015-12-21T14:45:34.857","AcceptedAnswerId":null,"OwnerUserId":5676182.0,"Title":"Tensorflow seq2seq weight sharing","Body":"<pre><\/pre>\n\n<p>I want to create two RNN models, one for training and another for testing. For that, I can call the function twice passing the feed_previous to True or False. <\/p>\n\n<pre><\/pre>\n\n<p>But if I call the above function twice, wouldn't it create two different RNNs ? I am wondering if they would be able to share the weights.<\/p>\n","answers":[{"AnswerId":"34399476","CreationDate":"2015-12-21T15:48:28.730","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Both functions operate on the same default graph and so can reuse the variables, check out <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/variable_scope\/index.html\" rel=\"nofollow\">variable scopes tutorial<\/a> and see if your variables are created with <code>reuse=True<\/code> parameter<\/p>\n\n<p>As a sanity check, try following snippet to list all variables in the default graph:<\/p>\n\n<pre><code>[v.name for v in tf.get_default_graph().as_graph_def().node if v.op=='Variable']\n<\/code><\/pre>\n"}]}
{"QuestionId":34399588,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-21T15:54:37.697","AcceptedAnswerId":34399966.0,"OwnerUserId":691867.0,"Title":"How to rename a variable which respects the name scope?","Body":"<p>Given  are tensors, I know I can do <\/p>\n\n<pre><\/pre>\n\n<p>So that  is named .<\/p>\n\n<p>I am wondering if there exists a function  which assign the name directly in the following case:<\/p>\n\n<pre><\/pre>\n\n<p>The stupid  I am using now is <\/p>\n","answers":[{"AnswerId":"34399966","CreationDate":"2015-12-21T16:14:48.407","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you want to \"rename\" an op, there is no way to do that directly, because a <code>tf.Operation<\/code> (or <code>tf.Tensor<\/code>) is immutable once it has been created. The typical way to rename an op is therefore to use <code>tf.identity()<\/code>, which has almost no runtime cost:<\/p>\n\n<pre><code>with tf.name_scope(\"abc\"):\n    z = x + y\n    z = tf.identity(z, name=\"z\")\n<\/code><\/pre>\n\n<p>Note however that the recommended way to structure your name scope is to assign the name of the scope itself to the \"output\" from the scope (if there is a single output op):<\/p>\n\n<pre><code>with tf.name_scope(\"abc\") as scope:\n    # z will get the name \"abc\". x and y will have names in \"abc\/...\" if they\n    # are converted to tensors.\n    z = tf.add(x, y, name=scope)\n<\/code><\/pre>\n\n<p>This is how the TensorFlow libraries are structured, and it tends to give the best visualization in TensorBoard.<\/p>\n"},{"AnswerId":"46844995","CreationDate":"2017-10-20T08:17:56.830","ParentId":null,"OwnerUserId":"8069423","Title":null,"Body":"<p>It seems it works also without <code>tf.name_scope<\/code> only with <code>z = tf.identity(z, name=\"z_name\")<\/code>. If you run additionally <code>z = tf.identity(z, name=\"z_name_new\")<\/code> then you can access the same tensor using both names: <code>tf.get_default_graph().get_tensor_by_name(\"z_name:0\")<\/code> or <code>tf.get_default_graph().get_tensor_by_name(\"z_name_new:0\")<\/code><\/p>\n"}]}
{"QuestionId":34401714,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-21T18:00:11.657","AcceptedAnswerId":34402043.0,"OwnerUserId":620625.0,"Title":"Is tensorflow lazy?","Body":"<p>Let's say you have some piece of code like this<\/p>\n\n<pre><\/pre>\n\n<p>b is a boolean tensor of the same shape as e and f. \nIf all the elements of b evaluate to true, we don't need\nf and the result will just be (or be equal to) e. <\/p>\n\n<p>The question: when the session is run with <em>result<\/em>, and the elements of e are all true, is f evaluated?<\/p>\n","answers":[{"AnswerId":"34402043","CreationDate":"2015-12-21T18:24:01.440","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> TensorFlow is strict, so both <code>e<\/code> and <code>f<\/code> will be evaluated before the <code>tf.select()<\/code> node executes.<\/p>\n\n<p>This has caused some confusion. TensorFlow first prunes the dataflow graph based on which operations are <em>statically<\/em> required to produce the values that are fetched (i.e. the arguments to <code>sess.run()<\/code>). Once the graph has been pruned, however, the runtime uses strict execution, whereby all of the inputs to an operation (such as <code>tf.select()<\/code>) must have been computed before that operation can execute.<\/p>\n\n<p>There is experimental support for conditional execution in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py\"><code>tf.control_flow_ops<\/code><\/a> module, using the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0f7dbaa217b9a990c43f6a0c702844c1738d4361\/tensorflow\/python\/ops\/control_flow_ops.py#L595\"><code>tf.control_flow_ops.cond()<\/code><\/a> function, but this is sparsely documented at present.<\/p>\n"}]}
{"QuestionId":34403447,"AnswerCount":2,"Tags":"<gcc><cuda><g++><caffe><nvcc>","CreationDate":"2015-12-21T20:00:36.217","AcceptedAnswerId":34407088.0,"OwnerUserId":1321680.0,"Title":"Caffe compilation fails due to unsupported gcc compiler version","Body":"<p>I struggle with <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow noreferrer\">Caffe<\/a> compilation. Unfortunately I failed to compile it.<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#cmake-compilation\" rel=\"nofollow noreferrer\">Steps<\/a> I followed:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>Running  fails with the following error message:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>Software version:<\/p>\n\n<ul>\n<li>OS: .<\/li>\n<li> version: .<\/li>\n<li> version: .<\/li>\n<li> result:<\/li>\n<\/ul>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<hr>\n\n<h2>Attempts to solve the problem<\/h2>\n\n<h3><em>1st try<\/em><\/h3>\n\n<p>Simple solutions are often best ones, so (as suggested <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/516939\/cuda4-1-gcc-4-6\/\" rel=\"nofollow noreferrer\">here<\/a>) I tried to comment out macro checking  version from  (line 82). Unfortunately it doesn't work and compilation fails badly:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<h3><em>2nd try<\/em><\/h3>\n\n<p>I tried to run:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>but it fails with <em>exactly<\/em> the same error message (even though  should be accepted).<\/p>\n\n<h3><em>3rd try<\/em><\/h3>\n\n<p>I've found <a href=\"https:\/\/stackoverflow.com\/questions\/28009785\/cuda-6-5-complains-about-not-supporting-gcc-4-9-what-to-do\">similar problem<\/a> (though not related to <em>Caffe<\/em>) and I tried to solve it as suggested in the accepted answer.<\/p>\n\n<p>What I did:<\/p>\n\n<ol>\n<li>I've ran  command and found <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/cmake\/Cuda.cmake\" rel=\"nofollow noreferrer\"><\/a> file which has  in line <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/cmake\/Cuda.cmake#L225\" rel=\"nofollow noreferrer\">225<\/a>.<\/li>\n<li>I added  to <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/cmake\/Cuda.cmake\" rel=\"nofollow noreferrer\"><\/a>, line before line: .<\/li>\n<li>I removed everything from  directory and ran  and  again - with and without .<\/li>\n<\/ol>\n\n<p>Unfortunately result is exactly the same. Caffe probably overwrites it somehow - I didn't figure it out <em>how<\/em>.<\/p>\n\n<p> returns nothing.<\/p>\n\n<p>What's interesting,  prints command that fails, which is:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>when I add  flag manually, it prints error:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>which may be related to <a href=\"https:\/\/cmake.org\/Bug\/view.php?id=13674\" rel=\"nofollow noreferrer\">this<\/a> bug report.<\/p>\n\n<p><strong>Edit:<\/strong> I didn't notice that  and  are the same options, and the latter is already set in above command that failed. When I changed  to  in above command that failed, it completes successfully. Now I need to find option in Caffe's CMake file that overwrite  in all subsequent Caffe's CMakes. Looking at  seems to be good way to go.<\/p>\n\n<hr>\n\n<p>How can I successfully complete my compilation? Any help is appreciated.<\/p>\n\n<p>Related SO questions:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/33886654\/host-config-hunsupported-gnu-version-gcc-versions-later-than-4-9-are-not-suppo\"><em>host_config.h:unsupported GNU version! gcc versions later than 4.9 are not supported<\/em>.<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/28009785\/cuda-6-5-complains-about-not-supporting-gcc-4-9-what-to-do\"><em>CUDA 6.5 complains about not supporting gcc 4.9 - what to do?<\/em>.<\/a><\/li>\n<\/ul>\n","answers":[{"AnswerId":"35161168","CreationDate":"2016-02-02T18:28:35.903","ParentId":null,"OwnerUserId":"5615276","Title":null,"Body":"<p>My machine runs Ubuntu 15.10, and my default compiler version is gcc 5.2.1 .<\/p>\n\n<p>Commenting out the #error directive in line 115 of file<\/p>\n\n<pre><code>\/usr\/local\/cuda-7.5\/include\/host_config.h \n<\/code><\/pre>\n\n<p>(or whatever the path on your system is) did the trick for me. Caffe compiled fine, all tests ran smoothly.<\/p>\n\n<p>On the other hand, if one chooses to ignore this and proceed to compile part of the project with one compiler version, part of the project with another (for me it was gcc-4.8 and gcc-5.2.1), linking problems will arise. The linking problems of protobuf and libgflags another answer mentions are <em>not<\/em> unrelated to this.<\/p>\n"},{"AnswerId":"34407088","CreationDate":"2015-12-22T01:22:24.207","ParentId":null,"OwnerUserId":"1321680","Title":null,"Body":"<p><code>cmake -D CUDA_NVCC_FLAGS=\"-ccbin gcc-4.8\" .. &amp;&amp; make<\/code> causes <strong>successful compilation<\/strong>.<\/p>\n\n<hr>\n\n<p>Now another problem showed up: linking Google's <code>libgflags<\/code> or <code>libprotobuf<\/code> fails probably due to fact that it was compiled with newer <code>gcc<\/code> version but it's not related to asked question.<\/p>\n"}]}
{"QuestionId":34403873,"AnswerCount":1,"Tags":"<benchmarking><tensorflow>","CreationDate":"2015-12-21T20:32:10.007","AcceptedAnswerId":34404789.0,"OwnerUserId":867505.0,"Title":"How can I create\/run benchmarks for custom kernels in tensorflow?","Body":"<p>There is already some functionality in tensorflow to create benchmarks which can be seen in action for example <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/adjust_contrast_op_benchmark_test.cc\" rel=\"nofollow\">in the adjust contrast op benchmark<\/a>. If I run this on my machine, however, I just get an empty output:<\/p>\n\n<pre><\/pre>\n\n<p>Is my invocation of wrong? <\/p>\n","answers":[{"AnswerId":"34404789","CreationDate":"2015-12-21T21:40:50.633","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To invoke the benchmarks, run the following command (passing <code>--benchmarks=all<\/code> as the final argument):<\/p>\n\n<pre><code>$ bazel run -c opt \/\/tensorflow\/core:kernels_adjust_contrast_op_benchmark_test \\\n      --test_output=all --cache_test_results=no -- --benchmarks=all\n<\/code><\/pre>\n\n<p>To run GPU benchmarks, you must pass <code>--config=cuda<\/code> to <code>bazel<\/code> and append <code>_gpu<\/code> to the name of the test target. For example:<\/p>\n\n<pre><code>$ bazel run -c opt --config=cuda \\ \n     \/\/tensorflow\/core:kernels_adjust_contrast_op_benchmark_test_gpu \\\n     --test_output=all --cache_test_results=no -- --benchmarks=all\n<\/code><\/pre>\n"}]}
{"QuestionId":34404253,"AnswerCount":2,"Tags":"<lua><row><apply><torch><elementwise-operations>","CreationDate":"2015-12-21T21:01:27.590","AcceptedAnswerId":null,"OwnerUserId":1001930.0,"Title":"Apply function to each row in Torch","Body":"<p>I know that tensors have an  method, but this only applies a function to each element. Is there an elegant way to do row-wise operations? For example, can I multiply each row by a different value?<\/p>\n\n<p>Say<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>and I want to multiply each element in the th row of A by the th element of B to get<\/p>\n\n<pre><\/pre>\n\n<p>how would I do that?<\/p>\n","answers":[{"AnswerId":"34411404","CreationDate":"2015-12-22T08:26:03.380","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>One possibility is to <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#result-expandresult-sizes\" rel=\"nofollow\">expand<\/a> <code>B<\/code> as follow:<\/p>\n\n<pre><code> 1  1  1\n 2  2  2\n 3  3  3\n[torch.DoubleTensor of size 3x3]\n<\/code><\/pre>\n\n<p>Then you can use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#res-torchcmulres-tensor1-tensor2\" rel=\"nofollow\">element-wise multiplication<\/a> directly:<\/p>\n\n<pre><code>local A = torch.Tensor{{1,2,3},{4,5,6},{7,8,9}}\nlocal B = torch.Tensor{1,2,3}\nlocal C = A:cmul(B:view(3,1):expand(3,3))\n<\/code><\/pre>\n"},{"AnswerId":"34437715","CreationDate":"2015-12-23T14:23:12.233","ParentId":null,"OwnerUserId":"1001930","Title":null,"Body":"<p>See this link: <a href=\"https:\/\/stackoverflow.com\/questions\/34123291\/torch-apply-function-over-dimension\">Torch - Apply function over dimension<\/a><\/p>\n\n<p>(Thanks to Alexander Lutsenko for providing it. I just moved it to the answer.)<\/p>\n"}]}
{"QuestionId":34405665,"AnswerCount":0,"Tags":"<python><theano><deep-learning>","CreationDate":"2015-12-21T22:48:26.313","AcceptedAnswerId":null,"OwnerUserId":1626564.0,"Title":"How to write a custom pooling layer in Theano?","Body":"<p>I am new to Theano, and I try to learn Theano's source code. However, the Theano implementation on <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/signal\/downsample.py\" rel=\"nofollow\">MaxPooling<\/a> completely confused me. Especially, in its pooling's grad definition, it goes for another MaxPoolGrad, and this MaxPoolGrad's grad method further goes for DownsampleFactorMaxGradGrad. <\/p>\n\n<p>Can anyone explain me why Theano has to go that deep to implement this gradient function? Please correct me if I am wrong: assuming (2,2) max-pooling, I thought all what we need to do is to compute gx from gz (see the grad def below), i.e. getting the gradient of x by upscaling gz to the shape of x and only populates gz elements to those gx elements whose corresponding x elements are maximum in a (2,2) neighborhood. <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34406792,"AnswerCount":1,"Tags":"<tensorflow><roi>","CreationDate":"2015-12-22T00:44:47.790","AcceptedAnswerId":34421867.0,"OwnerUserId":5705364.0,"Title":"How to to find location (ROI) of a recognized object in tensorflow","Body":"<p>I was using the tensorflow android example which has a placeholder for the location of a ROI (see tensorflow\/examples\/android\/src\/...\/Classifier.java), but it is not used.  I would like to be able to locate the ROI of the classified object.  I looked through the code, but it wasn't obvious how the location could be determined.  Does tensorflow support this?<\/p>\n","answers":[{"AnswerId":"34421867","CreationDate":"2015-12-22T18:02:02.043","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>Unfortunately that placeholder is not populated, and actually determining the region of interest of an object is still a hard research problem. We don't currently have a released example of a model that can do that in TensorFlow, but if you want to look through the research, the term to look for is 'localization'.<\/p>\n\n<p>There's a related question you might find helpful here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/28178054\/does-convolutional-neural-network-possess-localization-abilities-on-images\">Does Convolutional Neural Network possess localization abilities on images?<\/a><\/p>\n"}]}
{"QuestionId":34408330,"AnswerCount":1,"Tags":"<python><python-2.7><numpy><tensorflow><conda>","CreationDate":"2015-12-22T04:14:45.480","AcceptedAnswerId":null,"OwnerUserId":5705743.0,"Title":"Cannot import name random\/multiarray in conda environment","Body":"<p>I'm trying to run tensorflow in a conda environment. I started off by creating a python 2.7 environment with  and then activated it. Within the environment, I ran , which has tensorflow and numpy in the package, so hypothetically there shouldn't be any issues running numpy. <\/p>\n\n<p>When I open up the python console within the environment, however, I'm continually getting  and  (I can import random with no issues, but then I get the multiarray issue) no matter how many times I uninstall\/reinstall numpy\/matplotlib (at one point I even uninstalled\/reinstalled python) and no matter what versions of these I try to use, I keep on getting the same issue. What should I do? <\/p>\n","answers":[{"AnswerId":"35170707","CreationDate":"2016-02-03T07:15:20.003","ParentId":null,"OwnerUserId":"1887427","Title":null,"Body":"<p>There is an answer <a href=\"https:\/\/stackoverflow.com\/questions\/33959028\/trouble-creating-a-conda-env-with-working-numpy-importerror-cannot-import-nam\">here<\/a>.<\/p>\n\n<p>Shortly: that issue has something with the version of numpy which is upgraded by another package by whatever reason. Try to specify version: <code>conda create -n NAME numpy=1.9.3 other_package<\/code>.<\/p>\n\n<p>If that doesn't work, check if you have files in your working directory which names matches the names of some packages. For example, I had a similar problem after renaming <code>numpy.py.txt<\/code> (which is a sort of handmade cheatsheet) into just <code>numpy.py<\/code> and trying to <code>import numpy<\/code> within Python shell when I was in that directory.<\/p>\n"}]}
{"QuestionId":34410654,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-22T07:35:10.483","AcceptedAnswerId":34420827.0,"OwnerUserId":1716733.0,"Title":"tensorflow: ValueError: setting an array element with a sequence","Body":"<p>I am playing with the fixed code from <a href=\"https:\/\/stackoverflow.com\/questions\/34311893\/how-can-i-train-a-simple-non-linear-regression-model-with-tensor-flow\">this question<\/a>. I am getting the above error. Googling suggests it might be some kind of dimension mismatch, though my diagnostics does not show any:<\/p>\n\n<pre><\/pre>\n\n<p>Looking at the dimensions, everything is alright: <\/p>\n\n<pre><\/pre>\n\n<p>Error:<\/p>\n\n<pre><\/pre>\n\n<p>Any debugging tips?<\/p>\n","answers":[{"AnswerId":"34420827","CreationDate":"2015-12-22T16:56:11.603","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This&mdash;not very helpful&mdash;error is raised when one of the values in the <code>feed_dict<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>tf.Session.run()<\/code><\/a> is a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a> object (in this case, the result of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#reshape\" rel=\"noreferrer\"><code>tf.reshape()<\/code><\/a>).<\/p>\n\n<p>The values in <code>feed_dict<\/code> must be numpy arrays, or some value <code>x<\/code> that can be implicitly converted to a numpy array using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/reference\/generated\/numpy.array.html\" rel=\"noreferrer\"><code>numpy.array(x)<\/code><\/a>. <code>tf.Tensor<\/code> objects cannot be <em>implicitly<\/em> converted, because doing so might require a lot of work: instead you have to call <code>sess.run(t)<\/code> to convert a tensor <code>t<\/code> to a numpy array.<\/p>\n\n<p>As you noticed in your answer, using <code>np.reshape(_y_, [-1, 1])<\/code> works, because it produces a numpy array (and because <code>_y_<\/code> is a numpy array to begin with). In general, you should always prepare data to be fed using numpy and other pure-Python operations.<\/p>\n"},{"AnswerId":"34410725","CreationDate":"2015-12-22T07:39:51.060","ParentId":null,"OwnerUserId":"1716733","Title":null,"Body":"<p>replacing <code>tf<\/code> reshape with plain <code>numpy<\/code> one helped:<\/p>\n\n<pre><code>        _y_ = np.reshape(_y_, [-1, 1])\n<\/code><\/pre>\n\n<p>the actual reason why is still unclear, but it works.<\/p>\n"}]}
{"QuestionId":34411448,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-22T08:29:30.387","AcceptedAnswerId":null,"OwnerUserId":5706386.0,"Title":"TensorFlow - optimization with normalization constraints","Body":"<p>Is there any implementation using TensorFlow to optimize a vector under a constraint such as |w|^2==1 ?<\/p>\n","answers":[{"AnswerId":"34421417","CreationDate":"2015-12-22T17:32:08.117","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>You can add a soft constraint to your loss: <code>some_constant * (norm(w)- 1)^2<\/code> but, as far as I know, there are no functionalities specifically for constrained optimization.<\/p>\n"}]}
{"QuestionId":34413575,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2015-12-22T10:25:23.253","AcceptedAnswerId":34414455.0,"OwnerUserId":3487835.0,"Title":"How to handle empty sequence with theano.scan()?","Body":"<p>The code snippet is like:<\/p>\n\n<pre><\/pre>\n\n<p>the scan process will iterate along t. However, when K&lt;=1, t will be an empty range, then theano.scan() will crash. Is there any way to fix this?<\/p>\n","answers":[{"AnswerId":"34414455","CreationDate":"2015-12-22T11:07:14.153","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can use <code>theano.ifelse.ifelse<\/code> to compute the scan only when the sequence has some elements in it. For example:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\nimport theano.ifelse\n\n\ndef step(x_t, s_tm1):\n    return s_tm1 + x_t\n\n\ndef compile():\n    K = tt.lscalar()\n    t = tt.arange(1, K)\n    zero = tt.constant(0, dtype='int64')\n    outputs, _ = theano.scan(step, sequences=[t], outputs_info=[zero])\n    output = theano.ifelse.ifelse(tt.gt(K, 1), outputs[-1], zero)\n    return theano.function([K], outputs=[output])\n\n\ndef main():\n    f = compile()\n    print f(3)\n    print f(2)\n    print f(1)\n    print f(0)\n    print f(-1)\n\n\nmain()\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre><code>[array(3L, dtype=int64)]\n[array(1L, dtype=int64)]\n[array(0L, dtype=int64)]\n[array(0L, dtype=int64)]\n[array(0L, dtype=int64)]\n<\/code><\/pre>\n"}]}
{"QuestionId":34416496,"AnswerCount":1,"Tags":"<memory-management><neural-network><gpu><theano><large-data>","CreationDate":"2015-12-22T12:59:26.223","AcceptedAnswerId":null,"OwnerUserId":1914080.0,"Title":"How to fit a very large convolutional neural network into the GPU with Theano","Body":"<p>I'm trying to program a large deep neural network in Theano (several big convolutional layers for image recognition). The whole net does not fit into the GPU, therefore I need to divide the net some way. My question is: <\/p>\n\n<p>Does anybody know where to find tutorials or sample code to address this problem? There are a lot of good tutorials and code for \"small networks\" but I don't find anything for big networks. <\/p>\n\n<p>This is not a trivial problem, reducing the batch of images during training or reducing the number or size of the layers is not a solution.<\/p>\n\n<p>I would appreciate any idea about how to design the network or managing the GPU.<\/p>\n\n<p>Thank you very much.<\/p>\n","answers":[{"AnswerId":"34467045","CreationDate":"2015-12-25T22:42:03.957","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Although it degrades the performance, but you can use <code>theano.tensor._shared<\/code> instead of <code>theano.shared<\/code>. <code>theano.tensor._shared<\/code> forces the data to be stored on the CPU RAM. However, you should first compute the amount of memory required by the network using the following method and then separate the network parameters into GPU and CPU RAM accordingly.<\/p>\n\n<pre><code>import numpy as np\nimport theano.tensor as T\nT.config.floatX = 'float32'\ndataPoints = np.random.random((5000, 65536)).astype(T.config.floatX)\n#float32 data type requires 4 bytes\nsizeinGBs = 5000 * 65546 * 4 \/ 1024 \/ 1024 \/ 1024 + (some small over-head constant which we can ignore safely)\n<\/code><\/pre>\n"}]}
{"QuestionId":34416702,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorboard>","CreationDate":"2015-12-22T13:11:21.197","AcceptedAnswerId":null,"OwnerUserId":5707306.0,"Title":"Tensorboard Graph Visualiyation Error using Python 3","Body":"<p>I cleared out the differences of python 2to3 in the files tensorboard.py and tensorboard_handler.py but Tensorboard still has Visualization error.<\/p>\n\n<p>Error on Browser side:<\/p>\n\n<blockquote>\n  <p>Graph Visualization failed:Error:Failure parsing graph definition<\/p>\n<\/blockquote>\n\n<p>The Server side error log is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34459035","CreationDate":"2015-12-25T00:20:26.023","ParentId":null,"OwnerUserId":"1661471","Title":null,"Body":"<p><code>tensorboard_handler.py<\/code> uses <code>StringIO<\/code>, but then passes bytes to it. When working with Python 3, it should probably use <code>BytesIO<\/code> instead (which is new in Python 3). This question is related: <a href=\"https:\/\/stackoverflow.com\/questions\/11914472\/stringio-in-python3\">StringIO in python3<\/a>. <\/p>\n\n<p>Unfortunately, Tensorboard does not yet support python 3, but if you happen to fix this issue, send a pull request.<\/p>\n"}]}
{"QuestionId":34417569,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2015-12-22T13:59:30.310","AcceptedAnswerId":null,"OwnerUserId":2505295.0,"Title":"Minimal RNN example in tensorflow","Body":"<p>Trying to implement a minimal toy RNN example in tensorflow.\nThe goal is to learn a mapping from the input data to the target data, similar to this wonderful concise <a href=\"https:\/\/github.com\/lmjohns3\/theanets\/blob\/master\/examples\/recurrent-sinusoid.py\" rel=\"noreferrer\">example in theanets<\/a>.<\/p>\n\n<p><em>Update<\/em>: We're getting there. The only part remaining is to make it converge (and less convoluted). Could someone help to turn the following into running code or provide a simple example?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34428867","CreationDate":"2015-12-23T04:52:50.383","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>I think there are a few problems with your code, but the idea is right.<\/p>\n\n<p>The main issue is that you're using a single tensor for inputs and outputs, as in:<br>\n <code>inputs = tf.placeholder(tf.int32, [batch_size, num_steps])<\/code>.<\/p>\n\n<p>In TensorFlow the RNN functions take a list of tensors (because num_steps can vary in some models). So you should construct inputs like this:<br>\n<code>inputs = [tf.placeholder(tf.int32, [batch_size, 1]) for _ in xrange(num_steps)]<\/code><\/p>\n\n<p>Then you need to take care of the fact that your inputs are int32s, but a RNN cell works on float vectors - that's what embedding_lookup is for.<\/p>\n\n<p>And finally you'll need to adapt your feed to put in the input list.<\/p>\n\n<p>I think the ptb tutorial is a reasonable place to look, but if you want an even more minimal example of an out-of-the-box RNN you can take a look at some of the rnn unit tests, e.g., here.\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/kernel_tests\/rnn_test.py#L164\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/kernel_tests\/rnn_test.py#L164<\/a><\/p>\n"}]}
{"QuestionId":34418027,"AnswerCount":1,"Tags":"<matlab><neural-network><hdf5><deep-learning><caffe>","CreationDate":"2015-12-22T14:24:19.147","AcceptedAnswerId":34431532.0,"OwnerUserId":2693257.0,"Title":"[caffe]: check fails: Check failed: hdf_blobs_[i]->shape(0) == num (200 vs. 6000)","Body":"<p>I have the train and label data as data.mat. (I have 200 training data with 6000 features and labels are (-1, +1) that have saved in data.mat).<\/p>\n\n<p>I am trying to convert my data in hdf5 and run Caffe using:<\/p>\n\n<pre><\/pre>\n\n<p>And my layer.prototxt (just data layer) is:<\/p>\n\n<pre><\/pre>\n\n<p>but, i have an error:\n( Check failed: hdf_blobs_[i]->shape(0) == num (200 vs. 6000))<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Many thanks!!!! Any advice would be appreciated!<\/p>\n","answers":[{"AnswerId":"34431532","CreationDate":"2015-12-23T08:32:32.943","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h3>The problem<\/h3>\n\n<p>It seems like there is indeed a conflict with the order of elements in arrays: matlab arranges the elements from the first dimension to the last (like fortran), while caffe and hdf5 stores the arrays from last dimension to first:<br>\nSuppose we have <code>X<\/code> of shape <code>n<\/code>x<code>c<\/code>x<code>h<\/code>x<code>w<\/code> then the \"second element of <code>X<\/code>\" is <code>X[2,1,1,1]<\/code> in  matlab but <code>X[0,0,0,1]<\/code> in C (1-based vs 0-based indexing doesn't make life easier at all).<br>\nTherefore, when you save an array of <code>size=[200, 6000, 1, 1]<\/code> in Matlab, what hdf5 and caffe are actually seeing is as array of <code>shape=[6000,200]<\/code>.<\/p>\n\n<p>Using the <a href=\"https:\/\/www.hdfgroup.org\/HDF5\/doc\/RM\/Tools.html#Tools-Ls\" rel=\"nofollow\"><code>h5ls<\/code><\/a> command line tool can help you spot the problem.<br>\nIn matlab you saved<\/p>\n\n<pre><code>&gt;&gt; hdf5write('my_data.h5', '\/new_train_x', \n  single( reshape(new_train_x,[200, 6000, 1, 1]) );\n&gt;&gt; hdf5write('my_data.h5', '\/label_train', \n  single( reshape(label_train,[200, 1, 1, 1]) ),\n  'WriteMode', 'append' );\n<\/code><\/pre>\n\n<p>Now you can inspect the resulting <code>my_data.h5<\/code> using <code>h5ls<\/code> (in Linux terminal):<\/p>\n\n<pre><code>user@host:~\/$ h5ls .\/my_data.h5\n  label_train              Dataset {200}\n  new_train_x              Dataset {6000, 200}\n<\/code><\/pre>\n\n<p>As you can see, the arrays are written \"backwards\".    <\/p>\n\n<h3>Solution<\/h3>\n\n<p>Taking this conflict into account when exporting data from Matlab, you should <a href=\"http:\/\/www.mathworks.com\/help\/matlab\/ref\/permute.html\" rel=\"nofollow\"><code>permute<\/code><\/a>:<\/p>\n\n<pre><code>load data.mat\nhdf5write('my_data.h5', '\/new_train_x', \n  single( permute(reshape(new_train_x,[200, 6000, 1, 1]),[4:-1:1] ) );\nhdf5write('my_data.h5', '\/label_train', \n  single( permute(reshape(label_train,[200, 1, 1, 1]), [4:-1:1] ) ),\n  'WriteMode', 'append' );\n<\/code><\/pre>\n\n<p>Inspect the resulting <code>my_data.h5<\/code> using <code>h5ls<\/code> now results with:<\/p>\n\n<pre><code>user@host:~\/$ h5ls .\/my_data.h5\n  label_train              Dataset {200, 1, 1, 1}\n  new_train_x              Dataset {200, 6000, 1, 1}\n<\/code><\/pre>\n\n<p>Which is what you expected in the first place.<\/p>\n"}]}
{"QuestionId":34418556,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><skflow>","CreationDate":"2015-12-22T14:54:02.720","AcceptedAnswerId":null,"OwnerUserId":1795377.0,"Title":"Cannot import name NotFittedError on Ubuntu","Body":"<p>When i try to run titanic.py script of TensorFlow and Scikit Flow example on TensorFlow i got this error:<\/p>\n\n<pre><\/pre>\n\n<p>I am using pip, python 2.7.9 and Ubuntu 15 <\/p>\n\n<p>Can anyone please help me to solve this error?<\/p>\n","answers":[{"AnswerId":"34418917","CreationDate":"2015-12-22T15:15:41.227","ParentId":null,"OwnerUserId":"1060350","Title":null,"Body":"<p>You need to <strong>learn to debug such problems yourself<\/strong>.<\/p>\n\n<p>See: every installation is a bit different. There are hundreds of packages installed, and apparently you have a <strong>version conflict<\/strong>. But <em>we<\/em> do not have access to your computer, and we cannot tell which versions you have installed.<\/p>\n\n<p>You are missing the import <code>NotFittedError<\/code><\/p>\n\n<p>Now you need to:<\/p>\n\n<ol>\n<li>find out <strong>where<\/strong> it is imported (the error message is very helpful there)<\/li>\n<li>where it is imported <em>from<\/em> (you need to look at the file on <em>your<\/em> computer...)<\/li>\n<li>figure out why this is not found<\/li>\n<li>find out when <code>NotFittedError<\/code> was added or removed<\/li>\n<\/ol>\n\n<p>As far as I can tell, your tensorflow version is incompatible with your sklearn version.<\/p>\n"},{"AnswerId":"34419983","CreationDate":"2015-12-22T16:10:57.317","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>EDIT:<\/strong> This has now been <a href=\"https:\/\/github.com\/google\/skflow\/commit\/35862b95ffeab2d0d741f0489ae5dcf51436c713\" rel=\"nofollow\">fixed<\/a> in skflow. Upgrading to the latest version of skflow will fix the issue.<\/p>\n\n<p>The offending import is in <a href=\"https:\/\/github.com\/google\/skflow\/blob\/5dbcc5afdb94675ab019ea08658c9a16b8db1162\/skflow\/estimators\/base.py#L25\" rel=\"nofollow\"><code>skflow\/estimators\/base.py<\/code><\/a>:<\/p>\n\n<pre><code>from sklearn.utils.validation import NotFittedError\n<\/code><\/pre>\n\n<p>It looks like this class was moved in a (relatively) <a href=\"https:\/\/github.com\/scikit-learn\/scikit-learn\/commit\/e3afc0e8c9de522324a2f0cf0f7a17fee5147195\" rel=\"nofollow\">recent commit to scikit-learn<\/a>. It would probably be easiest to downgrade to a previous version of scikit-learn (e.g. the 0.17 release seems to be compatible). If you're feeling adventurous, you could try editing line 25 of <code>\"build\/bdist.linux-x86_64\/egg\/skflow\/estimators\/base.py\"<\/code> to read:<\/p>\n\n<pre><code>from sklean.exceptions import NotFittedError\n<\/code><\/pre>\n"},{"AnswerId":"35485231","CreationDate":"2016-02-18T15:12:24.510","ParentId":null,"OwnerUserId":"4102822","Title":null,"Body":"<pre><code>pip install --upgrade sklearn\n<\/code><\/pre>\n\n<p>fixed the problem<\/p>\n"}]}
{"QuestionId":34418989,"AnswerCount":1,"Tags":"<python><firefox><tensorflow><tensorboard>","CreationDate":"2015-12-22T15:19:00.287","AcceptedAnswerId":null,"OwnerUserId":286579.0,"Title":"Not found : \/lib\/python2.7\/site-packages\/tensorflow\/tensorboard\/lib\/svg\/summary-icon.sv","Body":"<p>I am using tensorboard to visualize the graphs and minimization of objective function. I can visualize the the loss function but can not visualize the graph. I get a not found error <\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GnRVy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GnRVy.png\" alt=\"When clicked on GRAPH\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/ad7uc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ad7uc.png\" alt=\"loss minimization curve\"><\/a><\/p>\n\n<p>[EDIT]<\/p>\n\n<pre><\/pre>\n\n<p>[EDIT2]\nWhen I switched to another browser, (chromium), it works. But does not work with Firefox.\n<a href=\"https:\/\/i.stack.imgur.com\/5D8VC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5D8VC.png\" alt=\"Graph with Chromium\"><\/a><\/p>\n","answers":[{"AnswerId":"34420290","CreationDate":"2015-12-22T16:28:13.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> This issue was fixed in version 0.6.0 of TensorFlow, so upgrading to that version should fix it.<\/p>\n\n<p>The file <code>summary-icon.svg<\/code> was <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/f7918e1dcd5b0c1f8114f488fc35a63a81e94535\" rel=\"nofollow\">removed from TensorFlow<\/a> in version 0.6.0, and replaced with an embedded image. Is it possible that you have an installation that mixes versions 0.5.0 and 0.6.0 of TensorFlow (or some intermediate version that is built from source)?<\/p>\n"}]}
{"QuestionId":34419645,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-22T15:52:11.117","AcceptedAnswerId":34420169.0,"OwnerUserId":5094001.0,"Title":"Asynchronous computation in TensorFlow","Body":"<p>Recently I've been toying with <strong>TensorFlow<\/strong> and I mentioned that the framework is not able to use all my available computational resources. In <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html\" rel=\"noreferrer\">Convolutional Neural Networks<\/a> tutorial they mention that<\/p>\n\n<blockquote>\n  <p>Naively employing asynchronous updates of model parameters leads to sub-optimal training performance because an individual model replica might be trained on a stale copy of the model parameters. Conversely, employing fully synchronous updates will be as slow as the slowest model replica.<\/p>\n<\/blockquote>\n\n<p>Although they mention it in both in the tutorial and in a <a href=\"http:\/\/download.tensorflow.org\/paper\/whitepaper2015.pdf\" rel=\"noreferrer\">whitepaper<\/a> I did not really find a way to do the asynchronous parallel computation on a local machine. Is it even possible? Or is it part of the distributed to-be-released version of TensorFlow. If it is, then how?<\/p>\n","answers":[{"AnswerId":"34420169","CreationDate":"2015-12-22T16:21:02.003","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Asynchronous gradient descent is supported in the open-source release of TensorFlow, without even modifying your graph. The easiest way to do it is to execute multiple <strong>concurrent steps<\/strong> in parallel:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loss = ...\n\n# Any of the optimizer classes can be used here.\ntrain_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\n\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\n\ndef train_function():\n  # TODO: Better termination condition, e.g. using a `max_steps` counter.\n  while True:\n    sess.run(train_op)\n\n# Create multiple threads to run `train_function()` in parallel\ntrain_threads = []\nfor _ in range(NUM_CONCURRENT_STEPS):\n  train_threads.append(threading.Thread(target=train_function))\n\n# Start the threads, and block on their completion.\nfor t in train_threads:\n  t.start()\nfor t in train_threads:\n  t.join()\n<\/code><\/pre>\n\n<p>This example sets up <code>NUM_CONCURRENT_STEPS<\/code> calls to <code>sess.run(train_op)<\/code>. Since there is no coordination between these threads, they proceed asynchronously.<\/p>\n\n<p>It's actually more challenging to achieve <em>synchronous<\/em> parallel training (at present), because this requires additional coordination to ensure that all replicas read the same version of the parameters, and that all of their updates become visible at the same time. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"noreferrer\">multi-GPU example for CIFAR-10 training<\/a> performs synchronous updates by making multiple copies of the \"tower\" in the training graph with shared parameters, and explicitly averaging the gradients across the towers before applying the update. <\/p>\n\n<hr>\n\n<p><strong>N.B.<\/strong> The code in this answer places all computation on the same device, which will not be optimal if you have multiple GPUs in your machine. If you want to use all of your GPUs, follow the example of the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"noreferrer\">multi-GPU CIFAR-10 model<\/a>, and create multiple \"towers\" with their operations pinned to each GPU. The code would look roughly as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_ops = []\n\nfor i in range(NUM_GPUS):\n  with tf.device(\"\/gpu:%d\" % i):\n    # Define a tower on GPU `i`.\n    loss = ...\n\n    train_ops.append(tf.train.GradientDescentOptimizer(0.01).minimize(loss))\n\ndef train_function(train_op):\n  # TODO: Better termination condition, e.g. using a `max_steps` counter.\n  while True:\n    sess.run(train_op)\n\n\n# Create multiple threads to run `train_function()` in parallel\ntrain_threads = []\nfor train_op in train_ops:\n  train_threads.append(threading.Thread(target=train_function, args=(train_op,))\n\n\n# Start the threads, and block on their completion.\nfor t in train_threads:\n  t.start()\nfor t in train_threads:\n  t.join()\n<\/code><\/pre>\n\n<p>Note that you might find it convenient to use a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/how_tos\/variable_scope\/index.html\" rel=\"noreferrer\">\"variable scope\"<\/a> to facilitate variable sharing between the towers.<\/p>\n"}]}
{"QuestionId":34421017,"AnswerCount":1,"Tags":"<python><memory-management><gpu><gpgpu><theano>","CreationDate":"2015-12-22T17:07:45.227","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"How to quantify how much data is copied into the GPU in a Theano script?","Body":"<p>As a general rule, it's often preferable to avoid to repeatedly copy data from the RAM to the GPU memory, if the data can instead stay in the GPU memory. As a result, some codes store data sets in the GPU memory as a shared variable (<a href=\"https:\/\/github.com\/yoonkim\/CNN_sentence\/blob\/master\/conv_net_sentence.py#L186\" rel=\"nofollow\">example<\/a>).<\/p>\n\n<p>In Theano, is there any way to quantify how much data is copied into the GPU when a script is executed? My goal is to estimate to what extent not declaring the data set as a shared variable is slowing the program due to copying data from the RAM to the GPU memory. (declaring the data set as a shared variable in the script I am currently profiling would require quite some code changes, so I would prefer to get the estimate before diving into the code).<\/p>\n","answers":[{"AnswerId":"34467009","CreationDate":"2015-12-25T22:35:35.063","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>As far as I know you cannot compute the amount of memory required by a function or script. Assuming you know the number of elements to be stored on a GPU you can easily compute the amount of memory required to store those elements.<\/p>\n\n<p>A simple example:<\/p>\n\n<pre><code>import numpy as np\nimport theano.tensor as T\nT.config.floatX = 'float32'\ndataPoints = np.random.random((5000, 256 * 256)).astype(T.config.floatX)\n#float32 data type requires 4 bytes\nsizeinGBs = 5000 * 256 * 256 * 4 \/ 1024. \/ 1024 \/ 1024 + (some small over-head constant which we can ignore safely)\n<\/code><\/pre>\n\n<p>Which is roughly 1.22 GBs<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I found the answer to your question. You can easily get the amount of free memory on the GPU using the <code>CUDA<\/code> libray. Below is the code that can help you compute how much GPU memory has been dedicated to a task before and after a function is executed.<\/p>\n\n<pre><code>import theano.sandbox.cuda.basic_ops as sbcuda\nimport numpy as np\nimport theano.tensor as T\nT.config.floatX = 'float32'\nfreeGPUMemInGBs = sbcuda.cuda_ndarray.cuda_ndarray.mem_info()[0]\/1024.\/1024\/1024\nprint \"Your GPU has %s GBs of free memory\" % str(freeGPUMemInGBs)\n#An operation is to be executed below\ntestData = shared(np.random.random((5000, 256 * 256)).astype(T.config.floatX), borrow = True)\nprint \"The tasks above used %s GBs of your GPU memory. The available memory is %s GBs\" % (str(freeGPUMemInGBs - sbcuda.cuda_ndarray.cuda_ndarray.mem_info()[0]\/1024.\/1024\/1024), str(sbcuda.cuda_ndarray.cuda_ndarray.mem_info()[0]\/1024.\/1024\/1024))\n<\/code><\/pre>\n\n<p>Then the output (for my machine) is:<\/p>\n\n<pre><code>&gt;&gt;&gt; Your GPU has 11.2557678223 GBs of free memory\n&gt;&gt;&gt; The tasks above used 1.22077941895 GBs of your GPU memory. The available memory is 10.0349884033 GBs\n<\/code><\/pre>\n"}]}
{"QuestionId":34425161,"AnswerCount":1,"Tags":"<python-2.7><neural-network><gradient><theano><automatic-differentiation>","CreationDate":"2015-12-22T22:02:10.043","AcceptedAnswerId":34485198.0,"OwnerUserId":2838606.0,"Title":"How efficient\/intelligent is Theano in computing gradients?","Body":"<p>Suppose I have an artificial neural networks with 5 hidden layers. For the moment, forget about the details of the neural network model such as biases, the activation functions used, type of data and so on ... . Of course, the activation functions are differentiable.<\/p>\n\n<p>With symbolic differentiation, the following computes the gradients of the objective function with respect to the layers' weights:<\/p>\n\n<pre><\/pre>\n\n<p>This way, to compute the gradients w.r.t <em>w1<\/em> the gradients w.r.t w2, w3, w4 and w5 must first be computed. Similarly to compute the gradients w.r.t <em>w2<\/em> the gradients w.r.t w3, w4 and w5 must be computed first.<\/p>\n\n<p>However, I could the following code also computes the gradients w.r.t to each weight matrix:<\/p>\n\n<pre><\/pre>\n\n<p>I was wondering, is there any difference between these two methods in terms of performance? Is Theano intelligent enough to avoid re-computing the gradients using the second method? By intelligent I mean to compute w3_grad, Theano should [preferably] use the pre-computed gradients of w_output_grad, w5_grad and w4_grad instead of computing them again.<\/p>\n","answers":[{"AnswerId":"34485198","CreationDate":"2015-12-27T21:51:47.537","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Well it turns out Theano does not take the previously-computed gradients to compute the gradients in lower layers of a computational graph. Here's a dummy example of a neural network with 3 hidden layers and an output layer. However, it's <strong>not<\/strong> going to be a big deal at all since computing the gradients is a once-in-a-life-time operation unless you have to compute the gradient on each iteration. Theano returns a symbolic expression for the derivatives as a <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/symbolic_graphs.html#theano-graphs\" rel=\"nofollow\">computational graph<\/a> and you can simply use it as a function from that point on. From that point on we simply use the function derived by Theano to compute <strong>numerical<\/strong> values and update the weights using those.<\/p>\n\n<pre><code>import theano.tensor as T\nimport time\nimport numpy as np\nclass neuralNet(object):\n    def __init__(self, examples, num_features, num_classes):\n        self.w = shared(np.random.random((16384, 5000)).astype(T.config.floatX), borrow = True, name = 'w')\n        self.w2 = shared(np.random.random((5000, 3000)).astype(T.config.floatX), borrow = True, name = 'w2')\n        self.w3 = shared(np.random.random((3000, 512)).astype(T.config.floatX), borrow = True, name = 'w3')\n        self.w4 = shared(np.random.random((512, 40)).astype(T.config.floatX), borrow = True, name = 'w4')\n        self.b = shared(np.ones(5000, dtype=T.config.floatX), borrow = True, name = 'b')\n        self.b2 = shared(np.ones(3000, dtype=T.config.floatX), borrow = True, name = 'b2')\n        self.b3 = shared(np.ones(512, dtype=T.config.floatX), borrow = True, name = 'b3')\n        self.b4 = shared(np.ones(40, dtype=T.config.floatX), borrow = True, name = 'b4')\n        self.x = examples\n\n        L1 = T.nnet.sigmoid(T.dot(self.x, self.w) + self.b)\n        L2 = T.nnet.sigmoid(T.dot(L1, self.w2) + self.b2)\n        L3 = T.nnet.sigmoid(T.dot(L2, self.w3) + self.b3)\n        L4 = T.dot(L3, self.w4) + self.b4\n        self.forwardProp = T.nnet.softmax(L4)\n        self.predict = T.argmax(self.forwardProp, axis = 1)\n\n    def loss(self, y):\n        return -T.mean(T.log(self.forwardProp)[T.arange(y.shape[0]), y])\n\nx = T.matrix('x')\ny = T.ivector('y')\n\nnnet = neuralNet(x)\nloss = nnet.loss(y)\n\ndiffrentiationTime = []\nfor i in range(100):\n    t1 = time.time()\n    gw, gw2, gw3, gw4, gb, gb2, gb3, gb4 = T.grad(loss, [nnet.w, nnet.w2, logReg.w3, nnet.w4, nnet.b, nnet.b2, nnet.b3, nnet.b4])\n    diffrentiationTime.append(time.time() - t1)\nprint 'Efficient Method: Took %f seconds with std %f' % (np.mean(diffrentiationTime), np.std(diffrentiationTime))\n\ndiffrentiationTime = []\nfor i in range(100):\n    t1 = time.time()\n    gw = T.grad(loss, [nnet.w])\n    gw2 = T.grad(loss, [nnet.w2])\n    gw3 = T.grad(loss, [nnet.w3])\n    gw4 = T.grad(loss, [nnet.w4])\n    gb = T.grad(loss, [nnet.b])\n    gb2 = T.grad(loss, [nnet.b2])\n    gb3 = T.grad(loss, [nnet.b3])\n    gb4 = T.grad(loss, [nnet.b4])\n    diffrentiationTime.append(time.time() - t1)\nprint 'Inefficient Method: Took %f seconds with std %f' % (np.mean(diffrentiationTime), np.std(diffrentiationTime))\n<\/code><\/pre>\n\n<p>This will print out the followings:<\/p>\n\n<pre><code>Efficient Method: Took 0.061056 seconds with std 0.013217\nInefficient Method: Took 0.305081 seconds with std 0.026024\n<\/code><\/pre>\n\n<p>This shows that Theano uses a dynamic-programming approach to compute gradients for the efficient method.<\/p>\n"}]}
{"QuestionId":34426268,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2015-12-22T23:36:30.847","AcceptedAnswerId":null,"OwnerUserId":2615676.0,"Title":"Restricting number of cores used","Body":"<p>I'm trying to restrict the number of cores that a tf session uses but it's not working. This is how I'm initializing the session:<\/p>\n\n<pre><\/pre>\n\n<p>The system has 12 cores \/ 24 threads, and I can see that 40-60% of them are being used at any given point in time. The system also has 8 GPUs, but I construct the whole graph with .<\/p>\n\n<p>UPDATE: To clarify, the graph itself is a simple LSTM-RNN, that hews very closely to the examples in the tf source code. For completeness here's the full graph:<\/p>\n\n<pre><\/pre>\n\n<p>One important thing to note is that if the first time I call , I pass in the appropriate parameters, then the session <em>does<\/em> only run on a single core. The problem is that in subsequent runs, I am unable to change the behavior, even though I use  which is supposed to specifically allow for session-specific settings. I.e. even after I close the session using  and start a new one with new options, the original behavior remains unchanged unless I restart the python kernel (which is very costly because it takes it nearly an hour to load my data).<\/p>\n","answers":[{"AnswerId":"42768473","CreationDate":"2017-03-13T16:16:51.027","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>TensorFlow does an optimization where the first time a <code>DirectSession<\/code> is created it will create static thread pools which then will be reused. If you want to change this, specify multiple different thread pools in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto#L209\" rel=\"nofollow noreferrer\">session_inter_op_thread_pool<\/a> flag and specify which one you want to use.<\/p>\n"},{"AnswerId":"44945433","CreationDate":"2017-07-06T09:47:31.180","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p><code>use_per_session_threads<\/code> will only affect the <code>inter_op_parallelism_threads<\/code> but not the <code>intra_op_parallelism_threads<\/code>. The <code>intra_op_parallelism_threads<\/code> will be used for the Eigen thread pool (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/local_device.cc\" rel=\"nofollow noreferrer\">here<\/a>) which is always global, thus subsequent sessions will not influence this anymore.<\/p>\n\n<p>Note that there are other TF functions which can also trigger the initialization of the Eigen thread pool, so it can happen that it's already initialized before you create the first <code>tf.Session<\/code>. One example is <code>tensorflow.python.client.device_lib.list_local_devices()<\/code>.<\/p>\n\n<p>I solve this in a way that very early in my Python script, I create a dummy session with the appropriate values.<\/p>\n"}]}
{"QuestionId":34427209,"AnswerCount":5,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-23T01:30:04.380","AcceptedAnswerId":34427582.0,"OwnerUserId":2272798.0,"Title":"Save Tensorflow graph for viewing in Tensorboard without summary operations","Body":"<p>I have a rather complicated Tensorflow graph that I'd like to visualize for optimization purposes. Is there a function that I can call that will simply save the graph for viewing in Tensorboard without needing to annotate variables?<\/p>\n\n<p>I Tried this:<\/p>\n\n<pre><\/pre>\n\n<p>But no output was produced. This is using the 0.6 wheel.<\/p>\n\n<p>This appears to be related:\n<a href=\"https:\/\/stackoverflow.com\/questions\/33773271\/graph-visualisaton-is-not-showing-in-tensorboard-for-seq2seq-model\">Graph visualisaton is not showing in tensorboard for seq2seq model<\/a><\/p>\n","answers":[{"AnswerId":"42093630","CreationDate":"2017-02-07T15:27:38.303","ParentId":null,"OwnerUserId":"5023129","Title":null,"Body":"<p>This worked for me:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    ... build graph (without annotations) ...\nwriter = tf.summary.FileWriter(logdir='logdir', graph=graph)\nwriter.flush()\n<\/code><\/pre>\n\n<p>The graph is loaded automatically when launching tensorboard with \"--logdir=logdir\/\". No \"upload\" button needed.<\/p>\n"},{"AnswerId":"36178694","CreationDate":"2016-03-23T12:49:38.927","ParentId":null,"OwnerUserId":"6104317","Title":null,"Body":"<p>For all clarity, this is how I used the <code>.flush()<\/code> method and resolved the issue:<\/p>\n\n<p>Initialize the writer with:<\/p>\n\n<pre><code>writer = tf.train.SummaryWriter(\"\/home\/rob\/Dropbox\/ConvNets\/tf\/log_tb\", sess.graph_def)\n<\/code><\/pre>\n\n<p>and use the writer with:<\/p>\n\n<pre><code>writer.add_summary(summary_str, i)\n    writer.flush()\n<\/code><\/pre>\n"},{"AnswerId":"34440319","CreationDate":"2015-12-23T16:52:13.693","ParentId":null,"OwnerUserId":"5711971","Title":null,"Body":"<p>You can also dump the graph as a GraphDef protobuf and load that directly in TensorBoard.  You can do this without starting a session or running the model.<\/p>\n\n<pre><code>## ... create graph ...\n&gt;&gt;&gt; graph_def = tf.get_default_graph().as_graph_def()\n&gt;&gt;&gt; graphpb_txt = str(graph_def)\n&gt;&gt;&gt; with open('graphpb.txt', 'w') as f: f.write(graphpb_txt)\n<\/code><\/pre>\n\n<p>This will output a file that looks something like this, depending on the specifics of your model.<\/p>\n\n<pre><code>node {\n  name: \"W\"\n  op: \"Const\"\n  attr {\n    key: \"dtype\"\n    value {\n      type: DT_FLOAT\n    }\n  }\n...\nversion 1\n<\/code><\/pre>\n\n<p>In TensorBoard you can then use the \"Upload\" button to load it from disk.<\/p>\n"},{"AnswerId":"34427582","CreationDate":"2015-12-23T02:15:52.573","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>For efficiency, the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#SummaryWriter\" rel=\"noreferrer\"><code>tf.train.SummaryWriter<\/code><\/a> logs asynchronously to disk. To ensure that the graph appears in the log, you must call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#SummaryWriter.close\" rel=\"noreferrer\"><code>close()<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/adding_summaries_to_event_files#SummaryWriter.flush\" rel=\"noreferrer\"><code>flush()<\/code><\/a> on the writer before the program exits.<\/p>\n"},{"AnswerId":"55676941","CreationDate":"2019-04-14T15:26:57.827","ParentId":null,"OwnerUserId":"429476","Title":null,"Body":"<p>Nothing worked for me except this<\/p>\n\n<pre><code># Helper for Converting Frozen graph from Disk to TF serving compatible Model\ndef get_graph_def_from_file(graph_filepath):\n  tf.reset_default_graph()\n  with ops.Graph().as_default():\n    with tf.gfile.GFile(graph_filepath, 'rb') as f:\n      graph_def = tf.GraphDef()\n      graph_def.ParseFromString(f.read())\n      return graph_def\n\n#let us get the output nodes from the graph\ngraph_def =get_graph_def_from_file('\/coding\/ssd_inception_v2_coco_2018_01_28\/frozen_inference_graph.pb')\n\nwith tf.Session(graph=tf.Graph()) as session:\n    tf.import_graph_def(graph_def, name='')\n    writer = tf.summary.FileWriter(logdir='\/coding\/log_tb\/1', graph=session.graph)\n    writer.flush()\n\n<\/code><\/pre>\n\n<p>Then using TB worked<\/p>\n\n<pre><code>#ssh -L 6006:127.0.0.1:6006 root@&lt;remoteip&gt; # for tensor board - in your local machine type 127.0.0.1\n!tensorboard --logdir '\/coding\/log_tb\/1'\n<\/code><\/pre>\n"}]}
{"QuestionId":34428850,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-23T04:51:24.067","AcceptedAnswerId":34471586.0,"OwnerUserId":859840.0,"Title":"Variables on CPU, training\/gradients on GPU","Body":"<p>On the CIFAR-10 tutorial, I noticed that the variables are placed in CPU memory, but it is stated in <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10_train.py\" rel=\"noreferrer\"><\/a> that it is trained with a single GPU.<\/p>\n\n<p>I'm quite confused.. are the layer\/activations stored in GPU? Or alternatively, are the gradients stored in the GPU? Otherwise, it would seem storing variables on CPU would not make use of the GPU at all - everything is stored in CPU memory, so only the CPU is used for forward\/backward propagation.<\/p>\n\n<p>If the GPU was used for f\/b propagation, wouldn't that be a waste due to latency shuffling data CPU &lt;-> GPU?<\/p>\n","answers":[{"AnswerId":"34471586","CreationDate":"2015-12-26T13:04:58.370","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>Indeed, in cifar10-train the activations and gradients are on GPU, only the parameters are on CPU. You are right that this is not optimal for single-GPU training due to the cost of copying parameters between CPU and GPU. I suspect the reason it is done this way is to have a single library for single-GPU and multi-GPU models, as in the multi-GPU case, it is probably faster to have parameters on CPU. You can test easily what speedup you can get by moving all variables to GPU, just remove the \"with tf.device('\/cpu:0')\" in \"_variable_on_cpu\" in cifar10.py.<\/p>\n"}]}
{"QuestionId":34430007,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2015-12-23T06:39:48.230","AcceptedAnswerId":null,"OwnerUserId":4977638.0,"Title":"How to train a caffemodel on our own dataset?","Body":"<p>I tried using pre-trained <em>bvlc_reference_caffenet.caffemodel<\/em> for object recognition from images. I got good results for images containing only a single object. For images with multiple objects, I removed the  term from prediction which gives the class label with the maximum probability.<\/p>\n\n<p>Still, the accuracy is very less for the labels which I am getting. So, I am thinking of training the same <em>caffemodel<\/em> on my own dataset (containing images with multiple objects). How should I proceed? Is there any way to retrain a pre-trained <em>caffemodel<\/em> with the different dataset?<\/p>\n","answers":[{"AnswerId":"34431339","CreationDate":"2015-12-23T08:16:29.920","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>What you are after is called \"finetuning\": taking a deep net trained for task A, reusing its weights and re-train it to accomplish task B.<br>\nYou can start with <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow noreferrer\">this tutorial<\/a>, but you will find much more information simply by googling \"finetune caffe model\".<\/p>\n\n<p>You may also be interested in <a href=\"https:\/\/stackoverflow.com\/a\/33142623\/1714410\">this post<\/a> regarding training caffe with mutiple categories per input image.<\/p>\n"}]}
{"QuestionId":34434719,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2015-12-23T11:30:26.117","AcceptedAnswerId":34474693.0,"OwnerUserId":4117331.0,"Title":"TypeError from theano While using 3D numpy array","Body":"<p>I am trying something similar to code below<\/p>\n\n<pre><\/pre>\n\n<p>Theano is giving me a TypeError with msg:<\/p>\n\n<blockquote>\n  <p>Cannot convert Type TensorType(float64, 3D) (of Variable Subtensor{int64:int64:}.0) into Type TensorType(float64, matrix). You can try to manually convert Subtensor{int64:int64:}.0 into a TensorType(float64, matrix)<\/p>\n<\/blockquote>\n\n<p>What I am doing wrong here?<\/p>\n\n<p>Edit<\/p>\n\n<p>As mentioned by daniel changing x to dtensor3 will result in another error.<\/p>\n\n<blockquote>\n  <p>ValueError: Input dimension mis-match. (input[0].shape[1] = 5, input[1].shape[1] = 90)<\/p>\n  \n  <p>Apply node that caused the error: Elemwise{add,no_inplace}(Sum{axis=[1], acc_dtype=float64}.0, DimShuffle{x,0}.0)       <\/p>\n<\/blockquote>\n\n<p>Another way is to modify my train function but then I won't be able to do batch learning.<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to implement  RBM with integer values for visible units.<\/p>\n","answers":[{"AnswerId":"34439159","CreationDate":"2015-12-23T15:44:38.210","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is that <code>datax<\/code> is a 3D tensor and <code>datax[index*5:(index+1)*5]<\/code> is also a 3D tensor but you're trying to assign that to <code>x<\/code> which is a 2D tensor (i.e. a matrix).<\/p>\n\n<p>Changing<\/p>\n\n<pre><code>x = T.dmatrix('x')\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>x = T.dtensor3('x')\n<\/code><\/pre>\n\n<p>solves this problem but creates a new one because the dimensions of W and x don't match up to perform the dot product. It's unclear what the desired outcome is.<\/p>\n"},{"AnswerId":"34474693","CreationDate":"2015-12-26T19:52:29.417","ParentId":null,"OwnerUserId":"4117331","Title":null,"Body":"<p>Solved it after few hit and trials.<\/p>\n\n<p>What I needed was to change <\/p>\n\n<pre><code>x=T.dmatrix('x')\nw_v_bias=T.dot(W,x).sum(axis=2).sum(axis=1)+Hb \n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>x=T.dtensor3('x')\nw_v_bias=T.dot(x,W).sum(axis=3).sum(axis=1)+Hb\n<\/code><\/pre>\n\n<p>Now it produces (5,90) array after adding Hb elementwise to each of the five vectors of dot product.<\/p>\n"}]}
{"QuestionId":34439045,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2015-12-23T15:38:24.737","AcceptedAnswerId":34440598.0,"OwnerUserId":1971872.0,"Title":"Tensorflow Setup for Distributed Computing","Body":"<p>Can anyone provide guidance on how to setup tensorflow to work on many CPUs across a network? All of the examples I have found thus far use only one local box and multi-gpus at best. I have found that I can pass in a list of targets in the session_opts, but I'm not sure how to setup tensorflow on each box to listen for networked nodes\/tasks. Any example would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"34440598","CreationDate":"2015-12-23T17:10:19.147","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The open-source version (currently 0.6.0) of TensorFlow supports single-process execution only: in particular, the only valid target in the <code>tensorflow::SessionOptions<\/code> is the empty string, which means \"current process.\"<\/p>\n\n<p>The <a href=\"http:\/\/download.tensorflow.org\/paper\/whitepaper2015.pdf\" rel=\"nofollow\">TensorFlow whitepaper<\/a> describes the structure of the distributed implementation (see Figure 3) that we use inside Google. The basic idea is that the Session interface can be implemented using RPC to a <em>master<\/em>; and the master can partition the computation across a set of devices in multiple <em>worker<\/em> processes, which also communicate using RPC. Alas, the current version depends heavily on Google-internal technologies (like <a href=\"http:\/\/static.googleusercontent.com\/media\/research.google.com\/en\/\/pubs\/archive\/43438.pdf\" rel=\"nofollow\">Borg<\/a>), so a lot of work remains to make it ready for external consumption. We are currently working on this, and you can follow the progress on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23\" rel=\"nofollow\">this GitHub issue<\/a>.<\/p>\n\n<p><strong>EDIT on 2\/26\/2016:<\/strong> Today we released an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/distributed_runtime\" rel=\"nofollow\">initial version of the distributed runtime<\/a> to GitHub. It supports multiple machines and multiple GPUs.<\/p>\n"}]}
{"QuestionId":34442420,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-12-23T19:26:09.097","AcceptedAnswerId":34442629.0,"OwnerUserId":4359986.0,"Title":"Changing Tensorflow MNIST code with interactive session into session","Body":"<p>So, I have been learning tensorflow, and I have tried to change the code on the documentation from being run on an interactive session to being run in a regular session, so that I can run the python file containing the code from command line. The relevant tensorflow code is here: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html<\/a><\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting error message when I try to run the code:<\/p>\n\n<pre><\/pre>\n\n<p>I have no idea what is causing this problem; most likely it is an incorrect usage of the session. Could someone give me some help? Thanks.<\/p>\n","answers":[{"AnswerId":"34442780","CreationDate":"2015-12-23T19:53:52.427","ParentId":null,"OwnerUserId":"4359986","Title":null,"Body":"<p>The following code works:<\/p>\n\n<pre><code>import input_data\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n\nimport tensorflow as tf\n\nwith tf.Session() as sess:\n\n    x = tf.placeholder(\"float\", shape=[None, 784])\n    y_ = tf.placeholder(\"float\", shape=[None, 10])\n    W = tf.Variable(tf.zeros([784,10]))\n    b = tf.Variable(tf.zeros([10]))\n    y = tf.nn.softmax(tf.matmul(x,W) + b)\n    init_op = tf.initialize_all_variables()\n    init_op.run()\n    cross_entropy = -tf.reduce_sum(y_*tf.log(y))\n    train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)\n    for i in range(1000):\n        batch = mnist.train.next_batch(50)\n        train_step.run(feed_dict={x: batch[0], y_: batch[1]})\n\n    correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n    print(accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels}))\n<\/code><\/pre>\n"},{"AnswerId":"34442629","CreationDate":"2015-12-23T19:41:58.300","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are just two things you need to change to make this work:<\/p>\n\n<ol>\n<li><p>Initialise the variables before running the first training steps.<\/p>\n\n<pre><code>init_op = tf.initialize_all_variables()\ninit_op.run()\nfor i in range(1000):\n    # \u2026\n<\/code><\/pre>\n\n<p>This will fix the first error you are seeing, and is an important first step in any TensorFlow program that uses variables.<\/p><\/li>\n<li><p>Inline the bodies of <code>train()<\/code> and <code>eval()<\/code> in the <code>with tf.Session() as sess:<\/code> block. Your <code>eval()<\/code> function uses local variables from <code>train()<\/code> so the code is not valid Python as written. (Note that the <code>sess.run()<\/code> around <code>train()<\/code> and <code>eval()<\/code> is incorrect too\u2014those functions don't have a return value, so this is equivalent to calling <code>sess.run(None)<\/code>, which will raise an error.)<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34445890,"AnswerCount":1,"Tags":"<python><scipy><theano><deep-learning><keras>","CreationDate":"2015-12-24T00:53:59.410","AcceptedAnswerId":36970643.0,"OwnerUserId":1452257.0,"Title":"Gradient from Theano expression for filter visualization in Keras","Body":"<p>For a ConvNet it can be interesting to find the norm-bounded input that maximizes the activity of a single conv. filter as a way to <a href=\"http:\/\/igva2012.wikispaces.asu.edu\/file\/view\/Erhan+2009+Visualizing+higher+layer+features+of+a+deep+network.pdf\" rel=\"nofollow\">visualize the filters<\/a>. I'd like to do this in the deep learning package Keras. This could be done using a black box optimization algorithm with the code from the <a href=\"http:\/\/keras.io\/faq\/#how-can-i-visualize-the-output-of-an-intermediate-layer\" rel=\"nofollow\">FAQ<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>However, it would be a substantially easier optimization task if I had the gradient. How can I extract the gradient from the Theano expression and input it into a Python optimization library such as Scipy?<\/p>\n","answers":[{"AnswerId":"36970643","CreationDate":"2016-05-01T18:55:15.937","ParentId":null,"OwnerUserId":"1397061","Title":null,"Body":"<p>You can print out the gradient as described <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html\" rel=\"nofollow noreferrer\">here<\/a> and hand-code it into Scipy.  You can also do the optimization in Theano - see <a href=\"https:\/\/stackoverflow.com\/a\/29643793\/1397061\">this question<\/a>.  <\/p>\n\n<p>However, probably the most straight-forward approach is to create a function <code>get_gradients()<\/code> that uses <code>theano.grad()<\/code> to return the gradients of the filters with respect to an input, then call <code>scipy.optimize.minimize<\/code> with <code>jac=get_gradients<\/code>.  According to the <a href=\"http:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.optimize.minimize.html\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>jac : bool or callable, optional Jacobian (gradient) of objective\n  function. [...] jac can also be a callable returning the gradient of\n  the objective. In this case, it must accept the same arguments as fun.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":34448582,"AnswerCount":1,"Tags":"<neural-network><keras><data-science>","CreationDate":"2015-12-24T06:53:55.807","AcceptedAnswerId":34570714.0,"OwnerUserId":4984897.0,"Title":"How to input data into Keras? Specifically what is the x_train and y_train if I have more than 2 columns?","Body":"<p>How can I input data into keras? What is the structure? Specifically what is the x_train and y_train if I have more than 2 columns?<\/p>\n\n<p>This is the data I want to input:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2XvBc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2XvBc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am trying to define Xtrain in this example Multi Layer Perceptron Neural Network code Keras has in its documentation. (<a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow noreferrer\">http:\/\/keras.io\/examples\/<\/a>) Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT (additional information):<\/p>\n\n<p>Looking here: <a href=\"https:\/\/stackoverflow.com\/questions\/30383404\/what-is-data-type-for-python-keras-deep-learning-package?rq=1\">What is data type for Python Keras deep learning package?<\/a><\/p>\n\n<blockquote>\n  <p>Keras uses numpy arrays containing the theano.config.floatX floating point type. This can be configured in your .theanorc file. Typically, it will be float64 for CPU computations and float32 for GPU computations, although you can also set it to float32 when working on the CPU if you prefer. You can create a zero-filled array of the proper type by the command<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>Question: Step 1 looks like create a floating point numpy array using my above data from the excel file. What do I do with the winner column?<\/p>\n","answers":[{"AnswerId":"34570714","CreationDate":"2016-01-02T21:04:31.730","ParentId":null,"OwnerUserId":"5540600","Title":null,"Body":"<p>It all depends on your need.<\/p>\n\n<p>It looks like that you want to predict the winner based on the parameters shown in column A - N. Then you should define <code>input_dim<\/code> to be 14, and  <code>X_train<\/code> should be an (N,14) numpy array like this:<\/p>\n\n<pre><code>[\n   [9278,  37.9, ...],\n   [18594, 36.3, ...],\n   ...\n]\n<\/code><\/pre>\n\n<p>It seems that your prediction set only contains 2 items ( 2 president candidates LOL), so you should encode the answer <code>Y_train<\/code> in an (N,2) numpy array like this:<\/p>\n\n<pre><code>[\n   [1, 0],\n   [1, 0],\n   ...\n   [0, 1],\n   [0, 1],\n   ...\n]\n<\/code><\/pre>\n\n<p>where <code>[1,0]<\/code> indicates that Barack Obama is the winner and vice versa.<\/p>\n"}]}
{"QuestionId":34453929,"AnswerCount":1,"Tags":"<neural-network><caffe><pycaffe><weighted-average><matcaffe>","CreationDate":"2015-12-24T13:59:24.857","AcceptedAnswerId":34454068.0,"OwnerUserId":3106889.0,"Title":"How to set some weights in a layer to a particulr number?","Body":"<p>I would like to ask if  it is possible to set some weights in a layer to particular number or zero of pretrained model.<\/p>\n\n<p>For example, I want to download Lecun model and set some weights in the last layer to number e.x  and calculate the accuracy.<\/p>\n\n<p>How can I do that?<\/p>\n","answers":[{"AnswerId":"34454068","CreationDate":"2015-12-24T14:11:48.800","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are after <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">\"net surgery\"<\/a>. In python you can load the net and get direct access to the stored weights. Then you can tweak them as you pleased and save the modified net.<\/p>\n"}]}
{"QuestionId":34454278,"AnswerCount":1,"Tags":"<neural-network><deep-learning><hdf5><caffe>","CreationDate":"2015-12-24T14:30:26.980","AcceptedAnswerId":34454893.0,"OwnerUserId":2693257.0,"Title":"Check fail: how to use hdf5 data layer in deep layer?","Body":"<p>I have the train and label data as . (I have 200 training data with 6000 features and labels are (-1, +1) that have saved in data.mat).<\/p>\n\n<p>I am trying to convert my data (train and test) in  and run Caffe using:<\/p>\n\n<pre><\/pre>\n\n<p>(See <a href=\"https:\/\/stackoverflow.com\/q\/34249025\/1714410\">this thread<\/a> regarding converting mat-files to hdf5 in Matlab).<\/p>\n\n<p>My  is:<\/p>\n\n<pre><\/pre>\n\n<p>But I have a problem. It seems, it cannot read my input data.<\/p>\n\n<pre><\/pre>\n\n<p>Now, if i change loss layer like this:<\/p>\n\n<pre><\/pre>\n\n<p>I have this problem:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Many thanks!!!! Any advice would be appreciated!<\/strong><\/p>\n","answers":[{"AnswerId":"34454893","CreationDate":"2015-12-24T15:27:34.787","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your data layer is only defined for <code>phase: TRAIN<\/code> I believe the error occurs when caffe attempts to construct the test-time net (i.e., the <code>phase: TEST<\/code> net).<br>\nYou should have an additional layer with test data:<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  name: \"data\"\n  top: \"new_train_x\"     # note: same name as in HDF5\n  top: \"label_train\"     # \n  hdf5_data_param {\n    source: \"test_file.txt\"\n    batch_size: 20\n  }\n  include { phase: TEST } # do not forget TEST phase\n}\n<\/code><\/pre>\n\n<p>BTW, if you do not want to test your net during training, you can switch this option off. See <a href=\"https:\/\/stackoverflow.com\/a\/34387104\/1714410\">this thread<\/a> for more information.<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong><br>\nForgive me for being blunt, but you are making quite a mess.<\/p>\n\n<ol>\n<li><code>\"TanH\"<\/code> is not a loss layer - it's a neuron\/activation layer. It serves as a non-linarity applied to a linear layer (conv\/inner-product). As such, it accepts a single input (bottom blob) and outputs a single blob (top).<br>\nA loss layer computes a <strong>scalar<\/strong> loss value and usually requires two inputs: prediction and ground truth to compare to.  <\/li>\n<li>You did change your net and added a <code>\"HDF5Data\"<\/code> layer for the <code>TEST<\/code> phase as well, but this layer outputs a <code>top: \"test_x\"<\/code>, no layer in your net expects a <code>bottom: \"test_x\"<\/code> you only have layers expecting <code>\"new_train_x\"<\/code>... same goes for <code>\"label_text\"<\/code>.<\/li>\n<\/ol>\n\n<p>I suggest you re-write your hdf5 files with more generic names (e.g., <code>x<\/code> and <code>label<\/code>) for <em>both<\/em> train and test. Just use different file names to distinguish between them. This way your net works with <code>\"x\"<\/code> and <code>\"label\"<\/code> in both phases and only loads the appropriate dataset according to phase.<\/p>\n"}]}
{"QuestionId":34454721,"AnswerCount":8,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-24T15:10:54.183","AcceptedAnswerId":37411464.0,"OwnerUserId":1978504.0,"Title":"How to \"reset\" tensorboard data after killing tensorflow instance","Body":"<p>I'm testing different hyperparameters for a cnn model I built, but I'm having a small annoyance when viewing the summaries in Tensorboard. The problem seems to be that the data is just \"added\" in consecutive runs, so the functions result in a weird superposition unless I see the information as \"relative\" instead of \"by step\". See here:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hZDiS.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hZDiS.png\" alt=\"X Type: Step\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kef2k.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kef2k.png\" alt=\"X Type: Relative\"><\/a><\/p>\n\n<p>I've tried killing tensorboard's process and erasing the log files, but it seems it is not enough.<\/p>\n\n<p><strong>So the question is, how do I reset this information?<\/strong><\/p>\n\n<p>Thanks!!<\/p>\n","answers":[{"AnswerId":"57521302","CreationDate":"2019-08-16T08:30:17.897","ParentId":null,"OwnerUserId":"3252625","Title":null,"Body":"<p>Had a similar issue, which threw an error: '<em>You must feed a value for placeholder tensor 'dense_input' with dtype float and shape<\/em>'. This specific issue was discussed <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/10074\" rel=\"nofollow noreferrer\">here<\/a>.\nThe suggestion that worked for me used:<\/p>\n\n<pre><code>from keras.backend import clear_session\n# Before instantiating a tf.data.Dataset obj &amp; before model creation, call:\nclear_session()\n<\/code><\/pre>\n\n<p>Also make sure you're using unique names for each model's TensorBoard log_dir. Then to view all models, run TensorBoard from terminal at the parent dir of the logs files, I.e.:<\/p>\n\n<pre><code>tensorboard --logdir &lt;projDir&gt;\/logs\/tf_log\n<\/code><\/pre>\n"},{"AnswerId":"49207968","CreationDate":"2018-03-10T10:26:07.710","ParentId":null,"OwnerUserId":"7473896","Title":null,"Body":"<p>I had a similar problem, but with a duplication of computational graphs: they've just added in tensorboard when I called <\/p>\n\n<pre><code>writer.add_graph(graph=sess.graph)\n<\/code><\/pre>\n\n<p>In my case, it wasn't about log files but about Jupyter Notebook context.<\/p>\n\n<p>I figured out that after multiple runs of a Jupyter cell with a Graph definition the graph hasn't been reset but appeared in the context as a duplicate, so I added <\/p>\n\n<pre><code>tf.reset_default_graph()\n<\/code><\/pre>\n\n<p>before the start of building a computational graph.<\/p>\n\n<p>Hope it will help.<\/p>\n"},{"AnswerId":"34559901","CreationDate":"2016-01-01T20:46:20.907","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p>Ok, for some reason it didn't work before but now it did:<\/p>\n\n<p><strong>You must erase Tensorboard's log files AND kill its process<\/strong><\/p>\n\n<p>After killing the process run <code>fuser 6006\/tcp -k<\/code> to free port 6006 (if you're in linux) and fire tensorboard again.<\/p>\n"},{"AnswerId":"40286712","CreationDate":"2016-10-27T14:03:59.437","ParentId":null,"OwnerUserId":"4865809","Title":null,"Body":"<p>I just figured out the solution to this problem. Just put each Events.out file in a separate folder inside your log directory. And you will get a nice visualization in tensorboard with each run in a different color.<\/p>\n"},{"AnswerId":"45091720","CreationDate":"2017-07-13T22:10:18.840","ParentId":null,"OwnerUserId":"2386301","Title":null,"Body":"<p>This automatically deletes the log directory.<\/p>\n\n<pre><code>import shutil\n\nshutil.rmtree('log_dir', ignore_errors=True)\n<\/code><\/pre>\n"},{"AnswerId":"42922149","CreationDate":"2017-03-21T08:47:10.793","ParentId":null,"OwnerUserId":"5861256","Title":null,"Body":"<p>Add the following snippet to your code and it should automatically reset your tensorboard.<\/p>\n\n<pre><code>if tf.gfile.Exists(dirpath):\n   tf.gfile.DeleteRecursively(dirpath) \n<\/code><\/pre>\n\n<p>This will delete the previous logs.<\/p>\n"},{"AnswerId":"37411464","CreationDate":"2016-05-24T10:52:40.577","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>Note: The solution you've posted (erase TensorBoard's log files and kill the process) will work, but it isn't preferred, because it destroys historical information about your training.<\/p>\n\n<p>Instead, you can have each new training job write to a new subdirectory (of your top-level log directory). Then, TensorBoard will consider each job a new \"run\" and will create a nice comparison view so you can see how the training differed between iterations of your model. <\/p>\n\n<p>For an example, see: <a href=\"https:\/\/www.tensorflow.org\/tensorboard\/\">https:\/\/www.tensorflow.org\/tensorboard\/<\/a><\/p>\n"},{"AnswerId":"40164869","CreationDate":"2016-10-20T21:28:00.240","ParentId":null,"OwnerUserId":"3020535","Title":null,"Body":"<p>Yes, I believe ultimately this aspect is positive.<br>\nAs an example, in my script I automate new run logs via <code>datetime<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from datetime import datetime\nnow = datetime.now()\nlogdir = \"tf_logs\/...\/\" + now.strftime(\"%Y%m%d-%H%M%S\") + \"\/\"\n<\/code><\/pre>\n\n<p>Then when running TensorBoard you can click the different runs on and off provided you ran TensorBoard in the parent directory.<\/p>\n\n<p>If you <strong>know<\/strong> you don't care about a previous run and want it out of your life, then yes, you need to remove the event files and endure the unusually long process of killing and restarting TensorBoard.<\/p>\n"}]}
{"QuestionId":34454901,"AnswerCount":1,"Tags":"<variables><tensorflow><inference>","CreationDate":"2015-12-24T15:28:22.223","AcceptedAnswerId":34458812.0,"OwnerUserId":1067998.0,"Title":"Transfer parameters from training to inference graph","Body":"\n\n<p>To not to carry optimizer and gradient nodes into inference environment, I'm trying to create two versions of graph - one with training nodes and the other one without.<\/p>\n\n<p>And idea was to use  to pass variables from train graph version into inference graph version.<\/p>\n\n<p>So I've tried the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In this case  calls  and adds optimizer and loss on top of it.<\/p>\n\n<p>For some reason,  constructor doesn't add  node into the inference graph (or I just don't understand what  option does). I've tried empty constructor and<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>failed with error<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>What is the proper way to achieve this?<\/p>\n","answers":[{"AnswerId":"34458812","CreationDate":"2015-12-24T23:36:50.737","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>When you construct your inference graph, you should be able to construct a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver.__init__\" rel=\"noreferrer\"><code>tf.train.Saver()<\/code><\/a> with no arguments, and it will construct the appropriate save and restore ops for you. You should then be able to call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver.restore\" rel=\"noreferrer\"><code>saver.restore(sess, filename)<\/code><\/a> to restore the variables from a file.<\/p>\n\n<p><strong>N.B.<\/strong> For the constructor to work with no arguments, (i) the variables in the inference graph (i.e. the result of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#all_variables\" rel=\"noreferrer\"><code>tf.all_variables()<\/code><\/a>) must be a subset of the variables in the training graph, and (ii) the corresponding variables must have exactly the same names. If either of these conditions doesn't hold, you will need to specify a variable name map to the saver constructor. (However, if <code>self.CreateTrainingGraph()<\/code> calls <code>self.CreateInferenceGraph()<\/code> before creating any other variables, and doesn't do anything different with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#name_scope\" rel=\"noreferrer\"><code>tf.name_scope()<\/code><\/a>, then this should be fine.)<\/p>\n\n<p>(The <code>saver_def<\/code> argument is infrequently  used when you load in a graph&mdash;for example using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#import_graph_def\" rel=\"noreferrer\"><code>tf.import_graph_def()<\/code><\/a>&mdash;that already contains the save and restore ops from a previously created Saver. It will then create a Saver in your Python program that reuses those ops, and you will get a mysterious error if the graph does not contain those ops.)<\/p>\n"}]}
{"QuestionId":34458189,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-24T21:49:33.363","AcceptedAnswerId":null,"OwnerUserId":1092746.0,"Title":"Interpreting Tensorflow\/Tensorboard \"subtraction\" operation","Body":"<p>The following is code adapted from a simple learning example, that I have bent out of shape to understand the Tensorboard graph visualizations:<\/p>\n\n<pre><\/pre>\n\n<p>A sample output of the print statements is:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly, the subtractions are working properly-- the inputs to the subtraction are in different order, and yield different outputs.  However, the graph visualization is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/d6DqH.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d6DqH.jpg\" alt=\"Tensorboard Screenshot\"><\/a><\/p>\n\n<p>Notice the \"Sub\" operators, which appear not to reverse the order of the operands as the code does.  (Highlighting either operator yields no additional insight.)  Am I missing something obvious, or do the node visualizations completely obscure order of operands? <\/p>\n","answers":[{"AnswerId":"34598814","CreationDate":"2016-01-04T19:56:19.860","ParentId":null,"OwnerUserId":"1092746","Title":null,"Body":"<p>After futzing around with this, my considered answer to my own question is, \"Yes, this is working as intended.\"  The inputs to the nodes show only what the inputs are, not any particular relationships to the operation or the node or themselves; indeed, if one added a variable to itself in an operation node, the input variable would show up only once. <\/p>\n\n<p>This is not a design choice I would have made, but that does seem to be the intent. <\/p>\n\n<p>I still encourage others who may have more insight to comment or fully answer. <\/p>\n"}]}
{"QuestionId":34458306,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><theano>","CreationDate":"2015-12-24T22:10:24.810","AcceptedAnswerId":null,"OwnerUserId":797880.0,"Title":"Python one hot vector encoding","Body":"<p>I'm trying to convert my categorical data to one hot vector with using python, preferably with numpy or theano. I don't want to use sklearn(I can not install it).<\/p>\n\n<p>My data is like that:<\/p>\n\n<pre><\/pre>\n\n<p>My data have 7 columns and each column can have values 0,1, or 2. At end of operation I want to have one hot vector which will have 3^7-1 zeros and one 1 values.<\/p>\n\n<p>I try to use <\/p>\n\n<pre><\/pre>\n\n<p>but did not work and I'm not sure if it is suitable for this task.<\/p>\n\n<p>Most of the examples I see at internet suitable only for single column. <\/p>\n","answers":[{"AnswerId":"34460517","CreationDate":"2015-12-25T06:19:09.227","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>I don't think there's a built-in function for this, because you have to tell numpy\/theano that your values can only be from {0,1,2}. \nHere is a simple numpy implementation, where<br>\n0000000 is mapped to 10...0,<br>\n1000000 is mapped to 01...0,<br>\n2222222 is mapped to 00...1.<\/p>\n\n<pre><code>data=[\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n[ 0., 2., 2., 0., 2., 0., 2.],\n]\ndata_num = 7\n\none_hot_dat = np.zeros((data_num, 3**7))\n# vector for making indices\nvec = np.asarray([3**i for i in range(7)])\n# compute the corresponding index for each data point\nhot_idx = np.sum(np.asarray(data)*vec, axis=1).astype(int)\none_hot_dat[range(data_num), hot_idx] = 1\n# one_hot_dat[0] should give the first data point in one-hot\n<\/code><\/pre>\n"}]}
{"QuestionId":34458603,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-24T22:57:23.043","AcceptedAnswerId":34458742.0,"OwnerUserId":2615676.0,"Title":"Selectively registering the backward pass of a set of ops on the GPU","Body":"<p>I have a set of ops that are faster on CPUs than GPUs, both in terms of the forward and backward (gradient) computations. However they're only a small fraction of the whole model, most of which is better run on the GPU. Currently, if I just use  when specifying the forward model, and I let TF decide where to place the optimizer (e.g.  op), then it puts all the backward pass computations on the GPU, which is suboptimal. Is there some way of specifying that an op and its gradients should be registered on the GPU?<\/p>\n","answers":[{"AnswerId":"34458742","CreationDate":"2015-12-24T23:24:50.123","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>Currently there's no good way to customize the device assignment for ops in the (automatically generated) gradient computation. However, one thing you can do is to register a \"device function\" using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#device\" rel=\"nofollow noreferrer\"><code>with tf.device():<\/code><\/a>,  (though the documentation for <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph.device\" rel=\"nofollow noreferrer\">this function<\/a> applies and is more comprehensive). A \"device function\" is a function that takes a newly-constructed <code>tf.Operation<\/code> and returns a device name, and TensorFlow assigns the operation to that device. This enables you to do the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># These are almost certainly faster on GPU, but are just shown as an example.\nOPS_ON_CPU = set([\"AvgPool\", \"AvgPoolGrad\"])\n\ndef _device_function(op):\n  if op.type in OPS_ON_CPU:\n    return \"\/cpu:0\"\n  else:\n    # Other ops will be placed on GPU if available, otherwise CPU.\n    return \"\"\n\nwith tf.device(_device_function):\n  # Build model in here.\n  # ...\n  loss = ...\n  train_op = tf.train.AdamOptimizer(0.01).minimize(loss) \n<\/code><\/pre>\n\n<p>...which will place all ops with type <code>\"AvgPool\"<\/code> or <code>\"AvgPoolGrad\"<\/code> on the CPU.<\/p>\n"}]}
{"QuestionId":34459087,"AnswerCount":0,"Tags":"<macos><boost><homebrew><boost-python><caffe>","CreationDate":"2015-12-25T00:34:26.910","AcceptedAnswerId":null,"OwnerUserId":5715640.0,"Title":"Mac OS X 10.11 brew install boost-python end up in dead loop","Body":"<p>I am trying to get caffe, which needs  installed on my Mac.<\/p>\n\n<p>When I installed  with , it ended in dead loops and never finished. I could only interrupt with Ctrl+C.<\/p>\n\n<p>Here are the error messages. The following message just loop endlessly...<\/p>\n\n<pre><\/pre>\n\n<p>I have tried using other versions of boost-python but there's no luck...<\/p>\n\n<p>I'm new to OS X and don't know what to do. Googling yielded some similar cases while people are actually using boost, but it I couldn't find any fixes for installing boost... Any help is greatly appreciated. <\/p>\n\n<p>OS condition:<\/p>\n\n<ul>\n<li>Just formatted hard drive and reinstalled Mac OS X<\/li>\n<li>installed Xcode, homebrew<\/li>\n<li>installed pyenv with homebrew<\/li>\n<li>installed anaconda with pyenv<\/li>\n<li>installed hdf5, opencv with homebrew<\/li>\n<li>edited formulas for other caffe dependencies with:<\/li>\n<\/ul>\n\n<p>.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34461249,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><neural-network><theano>","CreationDate":"2015-12-25T08:25:36.570","AcceptedAnswerId":null,"OwnerUserId":4393339.0,"Title":"How to use a Neural Network for face detection?","Body":"<p>I'm trying to build a face detection system using a neural network written in theano. I am a bit confused as to what should be the expected output against which i would have to calculate the crossentropy. I don't want to know whether the face is present or not, i need to highlight the face in an image (find the location of the face). The size of the images is constant. But the size of the faces in the image is not. How do i go about that? Also, my webcam currently captures 480x640 images. Creating that number of neurons in the input layer would be very heavy on the system, how do i compress the images without losing any features? <\/p>\n","answers":[{"AnswerId":"34463798","CreationDate":"2015-12-25T14:48:15.300","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>There are many possible solutions, one of the easiest ones is to perform a sliding window search and ask a network \"is there a face in this part of an image?\" - and this is quite \"standard\" approach. In particular, you do it hierarchicaly - split image into 9 overlapping squares (I assume the image is square) and ask in each of them \"is there a face in it?\" by rescaling it to your network input. Next you again split the one answering \"yes\" into 9 squares and repeat. This way you can find face kind fast. Another would be to perform supervised segmentation where you try to predict which part of image (pixels\/superpixels) belong to face and which do not. This is not exhaustive list, but should give you general idea how to proceed.<\/p>\n\n<blockquote>\n  <p>how do i compress the images without losing any features?<\/p>\n<\/blockquote>\n\n<p>You do not. It is not possible. You will always lose some data when downscaling (lossless compression exists but it destroys structure, thus making classification extremely hard).<\/p>\n"},{"AnswerId":"34467099","CreationDate":"2015-12-25T22:51:14.413","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>You should first create a training set from the images received through the web-cam. The training set must contain face and non-face images (such as apple, car and ...). For better generalization you may use some off-the-shelve data sets. After you trained the network on the images you can use the network to classify unseen images.<\/p>\n\n<p>This approach is suitable if your goal is to only detect whether an image contains a face. However, if you want to identify faces (e.g. this face belongs to John and not other people) you need to train the network with the images of the people you want to do identification for. The number of classes in such network is equivalent to the number of distinct people. <\/p>\n"}]}
{"QuestionId":34465340,"AnswerCount":1,"Tags":"<python><performance><hdf5><caffe>","CreationDate":"2015-12-25T18:30:59.587","AcceptedAnswerId":34465820.0,"OwnerUserId":2527680.0,"Title":"Improve speed of I\/O operations ? For HDF5 data creation for caffe","Body":"<p>The basic objective of my program is to read images and make a hd5 format file.\nI'm splitting the hd5 data files into parts of 1000 for manageability.<\/p>\n\n<p>The program reads and resizes the images and then writes to file.<\/p>\n\n<p>I dont think that using a multi-threading would be improving the speed of this, but I might be wrong.<\/p>\n\n<p>My dataset is around  15 million images.<\/p>\n\n<p>I use a powerful pc with a 4GB gpu and 32 GB ram and a Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz<\/p>\n\n<p>P.S I could try using some other image transformation package like opencv, But has no basis for comparison.<\/p>\n\n<p>As of now the program has been running for 3 days non stop and almost 80% done.\nI would like to avoid this problem in the future when I do something similar.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34465820","CreationDate":"2015-12-25T19:39:58.273","ParentId":null,"OwnerUserId":"5658989","Title":null,"Body":"<p>I am quite sure you can improve your performance with a Multi-threading approach, you haven't been spending 3 days on loading data from disk (you would need an unrealistic amount of disk space to read from), so you seem to be waiting for the resize process on CPU.<\/p>\n\n<p>You could fx do: 1 Reader that read data in large chucks and puts single images in a queue. Some Workers that take image from queue, resize it and put it another queue. 1 Writer that takes the resized images out of second queues, and write them to disk when it gathered many (Reader and Writer can probably be the same process without efficiency loss, assuming you read \/ write to the same disk anyway).<\/p>\n\n<p>My guess is 1 worker per HW thread (16 in your case), minus 2 for the core you put the reader and writer on (so 14), should be a good starting point.<\/p>\n\n<p>This way you will isolate waits for IO access from the CPU work, and minimize the IO access overhead by doing large chunks of work per time you initialize a read \/ write.<\/p>\n"}]}
{"QuestionId":34465679,"AnswerCount":2,"Tags":"<sequence><tensorflow><lstm>","CreationDate":"2015-12-25T19:21:18.633","AcceptedAnswerId":45965988.0,"OwnerUserId":3623648.0,"Title":"Sequence Labeling in TensorFlow","Body":"<p>I have managed to train a word2vec with tensorflow and I want to feed those results into an rnn with lstm cells for sequence labeling.<\/p>\n\n<p>1) It's not really clear on how to use your trained word2vec model for a rnn. (How to feed the result?)<\/p>\n\n<p>2) I don't find much documentation on how to implement a sequence labeling lstm. (How do I bring in my labels?)<\/p>\n\n<p>Could someone point me in the right direction on how to start with this task?<\/p>\n","answers":[{"AnswerId":"45965988","CreationDate":"2017-08-30T16:59:09.227","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>I realize this was posted a while ago, but I found <a href=\"https:\/\/gist.github.com\/danijar\/61f9226f7ea498abce36187ddaf51ed5\" rel=\"nofollow noreferrer\">this Gist about sequence labeling<\/a> and <a href=\"https:\/\/gist.github.com\/danijar\/d11c77c5565482e965d1919291044470\" rel=\"nofollow noreferrer\">this Gist about variable sequence labeling<\/a> really helpful for figuring out sequence labeling. The basic outline (the gist of the Gist):<\/p>\n\n<ol>\n<li>Use <code>dynamic_rnn<\/code> to handle unrolling your network for training and prediction. This method has moved around some in the API, so you may have to find it for your version, but just Google it.<\/li>\n<li>Arrange your data into batches of size <code>[batch_size, sequence_length, num_features]<\/code>, and your labels into batches of size <code>[batch_size, sequence_length, num_classes]<\/code>. Note that you want a label for every time step in your sequence.<\/li>\n<li>For variable-length sequences, pass a value to the <code>sequence_length<\/code> argument of the <code>dynamic_rnn<\/code> wrapper for each sequence in your batch.<\/li>\n<li>Training the RNN is very similar to training any other neural network once you have the network structure defined: feed it training data and target labels and watch it learn!<\/li>\n<\/ol>\n\n<p>And some caveats:<\/p>\n\n<ol>\n<li>With variable-length sequences, you will need to build masks for calculating your error metrics and stuff. It's all in the second link above, but don't forget when you make your own error metrics! I ran in to this a couple of times and it made my networks look like they were doing much worse on variable-length sequences.<\/li>\n<li>You might want to add a regularization term to your loss function. I had some convergence issues without this.<\/li>\n<li>I recommend using <code>tf.train.AdamOptimizer<\/code> with the default settings at first. Depending on your data, this may not converge and you will need to adjust the settings. <a href=\"http:\/\/ruder.io\/optimizing-gradient-descent\/\" rel=\"nofollow noreferrer\">This article<\/a> does a good job of explaining what the different knobs do. Start reading from the beginning, some of the knobs are explained before the Adam section.<\/li>\n<\/ol>\n\n<p>Hopefully these links are helpful to others in the future!<\/p>\n"},{"AnswerId":"34471743","CreationDate":"2015-12-26T13:23:51.250","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>I suggest you start by reading the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"nofollow noreferrer\">RNN tutorial<\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html\" rel=\"nofollow noreferrer\">sequence-to-sequence tutorial<\/a>. They explain how to build LSTMs in TensorFlow. Once you're comfortable with that, you'll have to find the right embedding Variable and assign it using your pre-trained word2vec model.<\/p>\n"}]}
{"QuestionId":34466356,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-25T20:58:23.323","AcceptedAnswerId":34585563.0,"OwnerUserId":2138200.0,"Title":"Error running label_image with output_layer=pool_3","Body":"\n\n<p>The tensorflow exercise in the image recognition tutorial suggests running the c++ example with \n. I have tried running this and am getting an error:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>What am I missing?<\/p>\n","answers":[{"AnswerId":"34585563","CreationDate":"2016-01-04T06:08:57.200","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue here is that the TensorFlow code in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">image recognition tutorial<\/a> requires additional modification before the <code>--output_layer=pool_3<\/code> option will work:<\/p>\n\n<blockquote>\n  <p>One can specify this by setting <code>--output_layer=pool_3<\/code> in the C++ API example and then <strong>changing the output tensor handling<\/strong>.<\/p>\n<\/blockquote>\n\n<p>To change the output tensor handling, you will need to modify the code <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2bde1e1308767ca240422596143832cbb095196b\/tensorflow\/examples\/label_image\/main.cc#L308\" rel=\"nofollow\">below this line in <code>label_image\/main.cc<\/code><\/a>. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2bde1e1308767ca240422596143832cbb095196b\/tensorflow\/examples\/label_image\/main.cc#L208\" rel=\"nofollow\"><code>PrintTopLabels()<\/code><\/a> function calls <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2bde1e1308767ca240422596143832cbb095196b\/tensorflow\/examples\/label_image\/main.cc#L183\" rel=\"nofollow\"><code>GetTopLabels()<\/code><\/a>, which takes a single 2-D (batch x classes) tensor&mdash;assumed to be the output of a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#softmax\" rel=\"nofollow\"><code>tf.nn.softmax()<\/code><\/a> containing a probability distribution for the labels in a batch of images&mdash;and builds a small TensorFlow graph using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#top_k\" rel=\"nofollow\"><code>tf.nn.top_k()<\/code><\/a> op . The <code>pool_3<\/code> layer outputs a four-dimensional (batch x height x width x depth) tensor, which will require additional processing.<\/p>\n\n<p>The additional processing has been left as an exercise for the reader. However, there are some things you could try:<\/p>\n\n<ul>\n<li><p>Reshape the output into a two-dimensional (batch x features) matrix, and train a fully connected layer (or more) to recognize features in your own training data.<\/p><\/li>\n<li><p>Visualize the output of the pooling layer by slicing it along the depth dimension, and encoding the slices into images using the  <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#encode_png\" rel=\"nofollow\"><code>tf.image.encode_png()<\/code><\/a> op.<\/p><\/li>\n<\/ul>\n\n<p><strong>N.B.<\/strong> Since the documentation is better, I've provided links to the Python documentation, rather than the corresponding C++ API. You might find it (much!) easier to modify the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2bde1e1308767ca240422596143832cbb095196b\/tensorflow\/models\/image\/imagenet\/classify_image.py\" rel=\"nofollow\">Python code for Inception inference<\/a> instead.<\/p>\n"}]}
{"QuestionId":34466375,"AnswerCount":0,"Tags":"<python><memory-management><machine-learning><gpgpu><theano>","CreationDate":"2015-12-25T21:01:15.353","AcceptedAnswerId":null,"OwnerUserId":2838606.0,"Title":"Dealing with memory fragmentation of GPUs in Theano","Body":"<p>To allocate space to a variable on GPU memory there must be enough space on continuous memory region. In other words you cannot have fragmented memory regions allocated to a variable on GPUS, unlike RAM. Having different shared variables stored on the GPU memory and continuously updating them could cause memory fragmentation. Therefore, even if there is enough free memory (in terms of bytes) on the GPU, you may not be able to use those memory regions as they are not in a continuous block.<\/p>\n\n<p>My question is how does Theano deal with such problem?<\/p>\n\n<p>Does  release all the memory assigned to that shared variable so that the next update () will only allocate the amount of memory to the shared variable and therefore avoid memory fragmentation?<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/faq.html#out-of-memory-but-not-really\" rel=\"nofollow\">Here<\/a> is is explained that updating a shared variable may still cause memory fragmentation. So I wonder whether the problem persists if the parameters  or  (in theanorc) is set to ?<\/p>\n\n<p>How one can keep track of the amount of free memory in a block (continuous) on a GPU?<\/p>\n","answers":[]}
{"QuestionId":34466901,"AnswerCount":2,"Tags":"<deep-learning><caffe><conv-neural-network><pycaffe><nvidia-digits>","CreationDate":"2015-12-25T22:18:21.627","AcceptedAnswerId":34935528.0,"OwnerUserId":2515628.0,"Title":"Multiple pretrained networks in Caffe","Body":"<p>Is there a simple way (e.g. without modifying <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> code) to load wights from multiple pretrained networks into one network? The network contains some layers with same dimensions and names as both pretrained networks.<\/p>\n\n<p>I am trying to achieve this using NVidia DIGITS and Caffe.<\/p>\n\n<p><strong>EDIT<\/strong>: I thought it wouldn't be possible to do it directly from DIGITS, as confirmed by answers. Can anyone suggest a simple way to modify the DIGITS code to be able to select multiple pretrained networks? I checked the code a bit, and thought the <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/f9ad5ced48fa49c677066f1d6b315601ccfae163\/digits\/model\/tasks\/train.py#L49\" rel=\"noreferrer\">training script<\/a> would be a good place to start, but I don't have in-depth knowledge of Caffe, so I'm not sure what the best\/quickest way to achieve this would be.<\/p>\n","answers":[{"AnswerId":"34478057","CreationDate":"2015-12-27T06:10:43.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>AFAIK there is no straight forward way of doing so.<br>\nHowever, you can use <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">net surgery<\/a> to load the pretrained models and manually assign their weights to the target net. Once you have a single net with all the weights initialized according to the various pretrained models, you can save it and use it as a single pretrained model for the rest of your work. <\/p>\n"},{"AnswerId":"34935528","CreationDate":"2016-01-21T22:25:58.103","ParentId":null,"OwnerUserId":"2515628","Title":null,"Body":"<p>As Shai suggested, there was no way of doing this, so I decided to clone the official repository and make the appropriate changes. I changed the code so that multiple pretrained networks can be loaded by using a colon as separator.<\/p>\n\n<p>I created a pull request on the official repository and my changes were then merged with the main branch of DIGITS, meaning it is now possible to use this functionality in DIGITS.<\/p>\n"}]}
{"QuestionId":34468584,"AnswerCount":1,"Tags":"<config><theano>","CreationDate":"2015-12-26T04:49:48.700","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"theano.config.floatX doesnt give any output","Body":"<p>print (theano.config.floatX) doesnt work.<\/p>\n\n<p>my .theanorc is as below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34468631","CreationDate":"2015-12-26T05:00:57.620","ParentId":null,"OwnerUserId":"5583416","Title":null,"Body":"<p>Found the solution at :<a href=\"https:\/\/stackoverflow.com\/questions\/22675274\/python-wont-find-variable-in-module\">Python won&#39;t find variable in module<\/a><\/p>\n\n<p>You have to go to the Python interpreter settings in Eclipse preferences and add theano.configdefaults to the list of built-ins <\/p>\n"}]}
{"QuestionId":34470479,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-26T10:30:07.663","AcceptedAnswerId":null,"OwnerUserId":2411173.0,"Title":"most efficient way to evaluate tensorflow Tensor on multiple input","Body":"<p>Suppose the following example where the tensor  is evaluated for several values of <\/p>\n\n<pre><\/pre>\n\n<p>I am wondering if the several calls to  can slow down the process and if there is a more efficient way to do this. <\/p>\n","answers":[{"AnswerId":"34471695","CreationDate":"2015-12-26T13:17:19.100","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>I think the best way to do evaluation for multiple values is to put them in one batch and evaluate them in a single call to eval(). In your code you create 100 minibatches of size 200 -- if you want a fast evaluation, why not use a single batch of 100*200? In some cases that might cause memory problems, but I think that would be the first thing to try for speed.<\/p>\n"}]}
{"QuestionId":34471563,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2015-12-26T13:02:27.213","AcceptedAnswerId":null,"OwnerUserId":3468216.0,"Title":"Logging training and validation loss in tensorboard","Body":"<p>I'm trying to learn how to use tensorflow and tensorboard. I have a test project based on the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"noreferrer\">MNIST neural net tutorial<\/a>.<\/p>\n\n<p>In my code, I construct a node that calculates the fraction of digits in a data set that are correctly classified, like this:<\/p>\n\n<pre><\/pre>\n\n<p>Here, is the inference part of the graph, and  is a placeholder that contains the correct labels.<\/p>\n\n<p>Now, what I would like to do is evaluate the accuracy for both the training set and the validation set as training proceeds. I can do this by running the accuracy node twice, with different feed_dicts:<\/p>\n\n<pre><\/pre>\n\n<p>This works as intended. I can print the values, and I can see that initially, the two accuracies will both increase, and eventually the validation accuracy will flatten out while the training accuracy keeps increasing.<\/p>\n\n<p>However, I would also like to get graphs of these values in tensorboard, and I can not figure out how to do this. If I simply add a  to , the logged values will not distinguish between training set and validation set. <\/p>\n\n<p>I also tried creating two identical  nodes with different names and running one on the training set and one on the validation set. I then add a  to each of these nodes. This does give me two graphs in tensorboard, but instead of one graph showing the training set accuracy and one showing the validation set accuracy, they are both showing identical values that do not match either of the ones printed to the terminal.<\/p>\n\n<p>I am probably misunderstanding how to solve this problem. What is the recommended way of separately logging the output from a single node for different inputs?<\/p>\n","answers":[{"AnswerId":"34474533","CreationDate":"2015-12-26T19:31:54.327","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are several different ways  you could achieve this, but you're on the right track with creating different <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/summary\/scalar\" rel=\"noreferrer\"><code>tf.summary.scalar()<\/code><\/a> nodes. Since you must explicitly call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/summary\/FileWriter#add_summary\" rel=\"noreferrer\"><code>SummaryWriter.add_summary()<\/code><\/a> each time you want to log a quantity to the event file, the simplest approach is probably to fetch the appropriate summary node each time you want to get the training or validation accuracy:<\/p>\n\n<pre><code>accuracy = tf.reduce_mean(correct)\n\ntraining_summary = tf.summary.scalar(\"training_accuracy\", accuracy)\nvalidation_summary = tf.summary.scalar(\"validation_accuracy\", accuracy)\n\n\nsummary_writer = tf.summary.FileWriter(...)\n\nfor step in xrange(NUM_STEPS):\n\n  # Perform a training step....\n\n  if step % LOG_PERIOD == 0:\n\n    # To log training accuracy.\n    train_acc, train_summ = sess.run(\n        [accuracy, training_summary], \n        feed_dict={images : training_set.images, labels : training_set.labels})\n    writer.add_summary(train_summ, step) \n\n    # To log validation accuracy.\n    valid_acc, valid_summ = sess.run(\n        [accuracy, validation_summary],\n        feed_dict={images : validation_set.images, labels : validation_set.labels})\n    writer.add_summary(valid_summ, step)\n<\/code><\/pre>\n\n<p>Alternatively, you could create a single summary op whose tag is a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"noreferrer\"><code>tf.placeholder(tf.string, [])<\/code><\/a> and feed the string <code>\"training_accuracy\"<\/code> or <code>\"validation_accuracy\"<\/code> as appropriate.<\/p>\n"},{"AnswerId":"49108094","CreationDate":"2018-03-05T10:18:40.530","ParentId":null,"OwnerUserId":"6344985","Title":null,"Body":"<p>Another way to do it, is to use a second file writer. So you are able to use the merge_summaries command.<\/p>\n\n<pre><code>train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '\/train',\n                                      sess.graph)\ntest_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '\/test')\ntf.global_variables_initializer().run()\n<\/code><\/pre>\n\n<p>Here is the complete documentation. This works for me fine : <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">TensorBoard: Visualizing Learning<\/a> <\/p>\n"}]}
{"QuestionId":34476239,"AnswerCount":1,"Tags":"<python><theano><recurrent-neural-network>","CreationDate":"2015-12-26T23:29:03.510","AcceptedAnswerId":34550501.0,"OwnerUserId":557478.0,"Title":"Issue with computing gradient for Rnn in Theano","Body":"<p>I am playing with vanilla Rnn's, training with gradient descent (non-batch version), and I am having an issue with the gradient computation for the (scalar) cost; here's the relevant portion of my code:<\/p>\n\n<pre><\/pre>\n\n<p>For some reasons I don't want to pass complete (training) data to the train_model(..), instead I want to pass individual examples at a time. Now the problem is that each call to train_model(..) returns me the cost (negative log-likelihood) of that particular example and then I have to aggregate all the cost (of the complete (training) data-set) and then take derivative and perform the relevant update to the weight parameters in the sgd_step(..), and for obvious reasons with my current implementation I am getting this error: <strong>theano.gradient.DisconnectedInputError: grad method was asked to compute the gradient with respect to a variable that is not part of the computational graph of the cost, or is used only by a non-differentiable operator: W_xh<\/strong>. Now I don't understand how to make '<strong>cost<\/strong>' a part of computational graph (as in my case when I have to wait for it to be aggregated) or is there any better\/elegant way to achieve the same thing ?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"34550501","CreationDate":"2015-12-31T18:47:21.860","ParentId":null,"OwnerUserId":"557478","Title":null,"Body":"<p>It turns out one cannot bring the symbolic variable into Theano graph if they are not part of computational graph. Therefore, I have to change the way to pass data to the train_model(..); passing the complete training data instead of individual example fix the issue.<\/p>\n"}]}
{"QuestionId":34477889,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-27T05:35:28.893","AcceptedAnswerId":34478044.0,"OwnerUserId":488517.0,"Title":"Holding variables constant during optimizer","Body":"<p>I have a TensorFlow computational graph for a loss tensor L that depends on 2 tf.Variables, A and B. <\/p>\n\n<p>I'd like to run gradient ascent on variable A (A+=gradient of L wrt A) while holding B fixed, and vice versa - running gradient ascent on B (B+=gradient of L wrt B) while holding A fixed. How do I do this?<\/p>\n","answers":[{"AnswerId":"34478044","CreationDate":"2015-12-27T06:07:13.293","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"\n\n<p><code>tf.stop_gradient(tensor)<\/code> might be what you are looking for. The tensor will be treated as constant for gradient computation purposes. You can create two losses with different parts treated as constants.<\/p>\n\n<p>The other option (and often better) would be to create 2 optimizers but explicitly optimize only subsets of variables, e.g.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_a = tf.train.GradientDescentOptimizer(0.1).minimize(loss_a, var_list=[A])\ntrain_b = tf.train.GradientDescentOptimizer(0.1).minimize(loss_b, var_list=[B])\n<\/code><\/pre>\n\n<p>and you can iterate between them on the updates.<\/p>\n"}]}
{"QuestionId":34478687,"AnswerCount":2,"Tags":"<deep-learning><tensorflow>","CreationDate":"2015-12-27T08:07:20.140","AcceptedAnswerId":34516450.0,"OwnerUserId":4563793.0,"Title":"Port TensorFlow code to Android","Body":"<p>I have written a script for sequence classification using TensorFlow in Python. I would like to port this code to Android. I have seen the example on the TensorFlow github page regarding Android but that is for images.<\/p>\n\n<p>Is there any way to directly port my TensorFlow Python code on Android?<\/p>\n","answers":[{"AnswerId":"34495029","CreationDate":"2015-12-28T14:05:43.877","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p>There is <a href=\"https:\/\/search.yahoo.com\/yhs\/search?p=python+android&amp;ei=UTF-8&amp;hspart=mozilla&amp;hsimp=yhs-001\" rel=\"nofollow\">QPython or Kivy<\/a>.<\/p>\n\n<p>QPython - Android Apps on <a href=\"https:\/\/play.google.com\/store\/apps\/details?id=org.qpython.qpy\" rel=\"nofollow\">GooglePlay<\/a>. It's a script engine that runs Python on android devices. It lets your android device run Python scripts and projects. It contains the Python interpreter and some other stuff like <code>pip<\/code>, but there's no compiler available, so only pure-python packages will work.<\/p>\n\n<p><a href=\"https:\/\/github.com\/kivy\/python-for-android\" rel=\"nofollow\">Python for Android<\/a> - lets you compile a Python application into an Android APK together with additional packages both pure-python and those that need compiling.<\/p>\n"},{"AnswerId":"34516450","CreationDate":"2015-12-29T18:02:17.927","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The typical way to do this is to build (and train) your model using Python, save the <code>GraphDef<\/code> proto to a file using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#write_graph\" rel=\"noreferrer\"><code>tf.train.write_graph()<\/code><\/a>, and then write an app using the JNI to call the C++ TensorFlow API (see a complete example <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"noreferrer\">here<\/a>).<\/p>\n\n<p>When you build your graph in Python, you should take note of the names of the tensors that will represent (i) the input data to be classified, and (ii) the predicted output values. Then you will be able to run a step by feeding a value for (i), and fetching the value for (ii).<\/p>\n\n<p>One final concern is how to represent the model parameters in your exported graph. There are several ways to do this, including shipping a TensorFlow checkpoint (written by a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver\" rel=\"noreferrer\"><code>tf.train.Saver<\/code><\/a>) as part of your app, and running the restore ops to reload it. One method, which has been used in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"noreferrer\">released InceptionV3 model<\/a> is to rewrite the graph so that the model parameters are replaced with <code>\"Const\"<\/code> nodes, and the model graph becomes self contained.<\/p>\n"}]}
{"QuestionId":34478775,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2015-12-27T08:24:23.763","AcceptedAnswerId":34572161.0,"OwnerUserId":5720143.0,"Title":"Compilation error in the process of installing fast-rcnn","Body":"<pre><\/pre>\n","answers":[{"AnswerId":"34572161","CreationDate":"2016-01-03T00:24:59.200","ParentId":null,"OwnerUserId":"4378596","Title":null,"Body":"<p>Try following <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\" rel=\"nofollow\">these<\/a> instructions on Rob's site for installing fast rcnn.<\/p>\n\n<p>Also you could check to see if the issue is with your cudnn config by following <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Install-Caffe-on-EC2-from-scratch-(Ubuntu,-CUDA-7,-cuDNN)\" rel=\"nofollow\">these<\/a> instructions (assuming you're using EC2 or Ubuntu 14.04).<\/p>\n\n<p>If its still an issue, would be helpful if you could share the steps you took which give you this error and what machine OS you're building to.<\/p>\n"},{"AnswerId":"36258895","CreationDate":"2016-03-28T08:45:58.817","ParentId":null,"OwnerUserId":"5107003","Title":null,"Body":"<p>I have met this trouble recently. It occurs when I compile the project of <a href=\"https:\/\/github.com\/rbgirshick\/fast-rcnn\" rel=\"nofollow\">fast-rcnn<\/a>, but doesn't occur for <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\" rel=\"nofollow\">py-faster-rcnn<\/a> and <a href=\"https:\/\/github.com\/ShaoqingRen\/faster_rcnn\" rel=\"nofollow\">enter link description here<\/a>. After I read the caffe installation guide which is presented on web: <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">Caffe Installation<\/a>, I think the version of cuDNN(the former is v3, the latter is v4) may result in the problem. My installation version of cuDNN is v4.0.7, and it runs well for the latter two projects.<\/p>\n"}]}
{"QuestionId":34479872,"AnswerCount":4,"Tags":"<javascript><performance><optimization><neural-network><tensorflow>","CreationDate":"2015-12-27T11:15:54.003","AcceptedAnswerId":null,"OwnerUserId":3830367.0,"Title":"Why is Tensorflow 100x slower than convnetjs in this simple NN example?","Body":"<p>I've been working with convnetjs for 1 year and now I want to move on to more powerful and fast libraries. I thought Tensorflow would be orders of magnitude faster than a JS library, so I wrote a simple neural network for both libraries and did some tests. It is a 3-5-5-1 neural network, trained on one single example for a certain number of epochs with SGD and RELU layers.<\/p>\n\n<p>Tensorflow code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>JS code:<\/p>\n\n<pre class=\"lang-js prettyprint-override\"><\/pre>\n\n<p>The results are that convnetjs trains for 100'000 epochs in 3 seconds, while Tensorflow trains for 1000 epochs in 2.5 seconds. Is this expected?<\/p>\n","answers":[{"AnswerId":"34484579","CreationDate":"2015-12-27T20:36:39.927","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>There could be many reasons why:<\/p>\n\n<ul>\n<li><p>The data input is so small that most of the time is spent in just conversion between python and the C++ core, while the JS is just one language. <\/p><\/li>\n<li><p>You are using only one core in Tensorflow while the JS could potentially leverage more than one<\/p><\/li>\n<li><p>the JS library is able to create a highly optimized JIT version of the program.<\/p><\/li>\n<\/ul>\n\n<p>The real benefits of Tensorflow will come when the distributed version will be public. Then the ability to run big networks on many nodes will be more important than the speed of a single node.<\/p>\n"},{"AnswerId":"34491273","CreationDate":"2015-12-28T09:57:16.700","ParentId":null,"OwnerUserId":"3301357","Title":null,"Body":"<p>As for now(version 0.6), it doesn't matter if you use CPU or GPU for tensorflow, tensorflow will be slow on GPU also.<\/p>\n\n<p><a href=\"https:\/\/github.com\/soumith\/convnet-benchmarks\/issues\/66\">Here are corresponding benchmarks<\/a><\/p>\n\n<p><strong>Tensorflow may be slower than torch, convnetjs, etc on CPU due to:<\/strong><\/p>\n\n<ol>\n<li>You may use non-optimized calculation graph.<\/li>\n<li>TF is not so mature as torch, convnetjs, etc. It's simply not so optimized. I hope <strong>yet<\/strong>.<\/li>\n<li><p>According to <strong>rumors<\/strong> google doesn't care about optimization for a single machine. Bare in mind, that<\/p>\n\n<p>3a) we live in the cluster age<\/p>\n\n<p>3b) you <a href=\"http:\/\/www.colfaxdirect.com\/store\/pc\/viewPrd.asp?idproduct=2642&amp;gclid=CPrwm9yk7ckCFUHicgodm0YCFQ#details\">can buy 57-core processor for 195$<\/a> (however I haven't tested if TF works with this hardware<\/p>\n\n<p>3c) Here's <a href=\"http:\/\/www.extremetech.com\/extreme\/219160-googles-quantum-computer-is-100-million-times-faster-than-a-conventional-system\">what google says about their quantum computer<\/a>. 100 million times faster than a conventional system.<\/p><\/li>\n<\/ol>\n\n<p><strong>TensorFlow is slower than caffe, torch, etc on GPU because of:<\/strong><\/p>\n\n<ol>\n<li>TF (as for 0.6) doesn't <strong>fully<\/strong> support cuda 7.5. <\/li>\n<li><p>TF (as for 0.6) doesn't support cudnn v3 and cudnn v4.<\/p>\n\n<p>and this makes TF 0.6 several orders slower than its competitors on 'machine learning desktops\/amateurs'.<\/p>\n\n<p>However, there is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/54\">an issue<\/a> to address cuda 7.5 and cudnn v3. Yet, it's closed as duplicated with  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20\">another issue, which is much less concrete<\/a> (IMHO). The latter issue, which is still opened,  doesn't obligate to support cuda 7.5 and cudnn v3\/v4(yep, I'm a pessimist).<\/p><\/li>\n<\/ol>\n\n<p>So, we can only either<\/p>\n\n<ol>\n<li>Hope and wait for google to solve these issues(add cuda 7.5 and cudnn v3\/v4 support and keep TF up to date all the way)<\/li>\n<li>Contribute. As TF is opensource. Or wait somebody to contribute :)<\/li>\n<\/ol>\n\n<p>I had the same confusion as an author of this question has. I hope my answer helped. <\/p>\n"},{"AnswerId":"36713107","CreationDate":"2016-04-19T08:42:37.330","ParentId":null,"OwnerUserId":"3481173","Title":null,"Body":"<p>The problem might be with your loss function. Why not try this instead?<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))\n<\/code><\/pre>\n"},{"AnswerId":"34501472","CreationDate":"2015-12-28T22:05:36.453","ParentId":null,"OwnerUserId":"5724808","Title":null,"Body":"<p>Yes, for tiny models this is expected.<\/p>\n\n<p>Tensorflow is not optimized for tiny neural nets with single item batches because making that regime faster is a waste of time. Those models are not expensive so there is no point. If you made the minibatch size larger (64 cases maybe) and the model somewhat larger (hundreds of hidden units) I would expect tensorflow to be much faster when compared to the other library.<\/p>\n\n<p>Imagine implementing the neural net naively in python using numpy. A naive numpy implementation would also be slow for this model.<\/p>\n"}]}
{"QuestionId":34480630,"AnswerCount":2,"Tags":"<python><numpy><lua><torch>","CreationDate":"2015-12-27T13:00:49.397","AcceptedAnswerId":34480871.0,"OwnerUserId":5720571.0,"Title":"Simple Torch7 equivalent to numpy.roll","Body":"<p>Is there any simple way of rolling a tensor in torch7 like numpy.roll and numpy.rollaxis in python?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"34480871","CreationDate":"2015-12-27T13:28:28.040","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<p>You can achieve the effects of numpy's <code>rollaxis<\/code> with torch's <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-permutedim1-dim2--dimn\" rel=\"nofollow\">permute<\/a>. While <code>rollaxis<\/code> requires the start and end position of the one axis to move, <code>permute<\/code> requires the new positions of all axis. E.g. for a 3-dimensional tensor <code>np.rollaxis(x, 0, 3)<\/code> (move the 1st axis to the end) would be equivalent to <code>x:permute(2, 3, 1)<\/code>.<\/p>\n\n<p>I'm not aware of an easy replacement for numpy's <code>roll<\/code>, but <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-scatterdim-index-srcval\" rel=\"nofollow\">scatter<\/a> seems like a decent candidate. Call it with the required dimension and the new order of elements after the shift. (Requires one new order of elements for each single row.)\nThe following example shifts each row of <code>x<\/code> (containing 2 rows and 4 columns with random values) 2 to the right along the last axis:<\/p>\n\n<pre><code>th&gt; x = torch.zeros(2, 4):uniform(0, 10)\nth&gt; y = torch.zeros(2, 4):scatter(2, torch.LongTensor{{3, 4, 1, 2}, {3, 4, 1, 2}}, x)\nth&gt; x\n 0.7295  3.2218  7.3979  5.5500\n 8.4354  3.6722  5.5463  3.4323\n[torch.DoubleTensor of size 2x4]\nth&gt; y\n 7.3979  5.5500  0.7295  3.2218\n 5.5463  3.4323  8.4354  3.6722\n[torch.DoubleTensor of size 2x4]\n<\/code><\/pre>\n"},{"AnswerId":"36897097","CreationDate":"2016-04-27T17:40:44.320","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>Maybe you could try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a> so you can use numpy array in python and then pass it to torch module and convert it back.<\/p>\n"}]}
{"QuestionId":34483277,"AnswerCount":2,"Tags":"<python><theano><symbolic-computation>","CreationDate":"2015-12-27T18:04:47.093","AcceptedAnswerId":34484383.0,"OwnerUserId":5106953.0,"Title":"5D tensor in Theano","Body":"<p>I was wondering how to make a 5D tensor in Theano.<\/p>\n\n<p>Specifically, I tried . However, the only issue is that returns: <\/p>\n\n<p>Whereas if I used a standard tensor type like, I don't get this issue when I call \nIs there a way to have this not be an issue with a 5D tensor in Theano?<\/p>\n","answers":[{"AnswerId":"34484383","CreationDate":"2015-12-27T20:12:38.780","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Theano variables do not have explicit shape information since they are <strong>symbolic<\/strong> variables, not numerical. Even <code>dtensor3 = T.tensor3(T.config.floatX)<\/code> does not have an explicit shape. When you type <code>dtensor3.shape<\/code> you'll get an object <code>Shape.0<\/code> but when you do <code>dtensor3.shape.eval()<\/code> to get its value you'll get an error.<\/p>\n\n<p>For both cases however, <code>dtensor.ndim<\/code> works and prints out <code>5<\/code> and <code>3<\/code> respectively.<\/p>\n"},{"AnswerId":"37423699","CreationDate":"2016-05-24T20:55:03.543","ParentId":null,"OwnerUserId":"6378189","Title":null,"Body":"<pre><code>dtensor = T.TensorType('float32',(False,)*5) \n<\/code><\/pre>\n\n<p>only calls the function <code>TensorType<\/code>. In order to use the attribute <code>dtensor.shape<\/code> you need to make it an object. You can do it by:<\/p>\n\n<pre><code>dtensor = T.TensorType('float32',(False,)*5) ()\n<\/code><\/pre>\n\n<p>You can specify the name inside the parenthesis at the end if you wish. <\/p>\n"}]}
{"QuestionId":34484148,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2015-12-27T19:43:35.610","AcceptedAnswerId":34497833.0,"OwnerUserId":5692514.0,"Title":"Feeding image data in tensorflow for transfer learning","Body":"<p>I am trying to use tensorflow for transfer learning. I downloaded the pre-trained model inception3 from the tutorial. In the code, for prediction: <\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to feed the png image. I tried changing  to  but it did not work. Beside, what should I change if I want to feed decoded image file like a numpy array or a batch of arrays?<\/p>\n\n<p>Thanks!!<\/p>\n","answers":[{"AnswerId":"34497833","CreationDate":"2015-12-28T17:19:51.860","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>The shipped InceptionV3 graph used in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py\" rel=\"nofollow noreferrer\"><code>classify_image.py<\/code><\/a> only supports JPEG images out-of-the-box. There are two ways you could use this graph with PNG images:<\/p>\n\n<ol>\n<li><p>Convert the PNG image to a <code>height<\/code> x <code>width<\/code> x 3 (channels) Numpy array, for example using <a href=\"http:\/\/www.pythonware.com\/products\/pil\/\" rel=\"nofollow noreferrer\">PIL<\/a>, then feed the <code>'DecodeJpeg:0'<\/code> tensor:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom PIL import Image\n# ...\n\nimage = Image.open(\"example.png\")\nimage_array = np.array(image)[:, :, 0:3]  # Select RGB channels only.\n\nprediction = sess.run(softmax_tensor, {'DecodeJpeg:0': image_array})\n<\/code><\/pre>\n\n<p>Perhaps confusingly, <code>'DecodeJpeg:0'<\/code> is the <em>output<\/em> of the <code>DecodeJpeg<\/code> op, so by feeding this tensor, you are able to feed raw image data.<\/p><\/li>\n<li><p>Add a <code>tf.image.decode_png()<\/code> op to the imported graph. Simply switching the name of the fed tensor from <code>'DecodeJpeg\/contents:0'<\/code> to <code>'DecodePng\/contents:0'<\/code> does not work because there is no <code>'DecodePng'<\/code> op in the shipped graph. You can add such a node to the graph by using the <code>input_map<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#import_graph_def\" rel=\"nofollow noreferrer\"><code>tf.import_graph_def()<\/code><\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>png_data = tf.placeholder(tf.string, shape=[])\ndecoded_png = tf.image.decode_png(png_data, channels=3)\n# ...\n\ngraph_def = ...\nsoftmax_tensor = tf.import_graph_def(\n    graph_def,\n    input_map={'DecodeJpeg:0': decoded_png},\n    return_elements=['softmax:0'])\n\nsess.run(softmax_tensor, {png_data: ...})\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"43964988","CreationDate":"2017-05-14T14:22:28.360","ParentId":null,"OwnerUserId":"662770","Title":null,"Body":"\n\n<p>The following code should handle of both cases.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom PIL import Image\n\nimage_file = 'test.jpeg'\nwith tf.Session() as sess:\n\n    #     softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')\n    if image_file.lower().endswith('.jpeg'):\n        image_data = tf.gfile.FastGFile(image_file, 'rb').read()\n        prediction = sess.run('final_result:0', {'DecodeJpeg\/contents:0': image_data})\n    elif image_file.lower().endswith('.png'):\n        image = Image.open(image_file)\n        image_array = np.array(image)[:, :, 0:3]\n        prediction = sess.run('final_result:0', {'DecodeJpeg:0': image_array})\n\n    prediction = prediction[0]    \n    print(prediction)\n<\/code><\/pre>\n\n<p>or shorter version with direct strings:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>image_file = 'test.png' # or 'test.jpeg'\nimage_data = tf.gfile.FastGFile(image_file, 'rb').read()\nph = tf.placeholder(tf.string, shape=[])\n\nwith tf.Session() as sess:        \n    predictions = sess.run(output_layer_name, {ph: image_data} )\n<\/code><\/pre>\n"}]}
{"QuestionId":34494430,"AnswerCount":4,"Tags":"<ubuntu><linker><caffe>","CreationDate":"2015-12-28T13:27:02.410","AcceptedAnswerId":null,"OwnerUserId":1972060.0,"Title":"How to solve link error in Ubuntu?","Body":"<p>I want to build a Caffe project using make, and during build receive following error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm on Ubuntu 14.04 and trying to build Caffe with Anaconda.  I have installed everything according to installation tutorial.  In fact I have used it before, and I can't understand what problem is now.  I'm not a very experienced Linux-user so I don't know how to trace it.<\/p>\n\n<p>Can you give any hints how to solve it? <\/p>\n","answers":[{"AnswerId":"36719241","CreationDate":"2016-04-19T13:01:27.743","ParentId":null,"OwnerUserId":"6225037","Title":null,"Body":"<p>just let people know the solution of \n\"Just export LD_LIBRARY_PATH=\"\/lib\/x86_64-linux-gnu:$LD_LIBRARY_PATH\"\" works for me. it is a pity to figure this out after 30 mins searching with ignoring this one at first sight..<\/p>\n"},{"AnswerId":"38897225","CreationDate":"2016-08-11T13:03:51.707","ParentId":null,"OwnerUserId":"6705165","Title":null,"Body":"<p>Suggestions:\n2 things worth to try:\nIn <code>~\/.bashrc<\/code> file :<\/p>\n\n<p><strong>(1)<\/strong><\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\"\/lib\/x86_64-linux-gnu:$LD_LIBRARY_PATH\"\n<\/code><\/pre>\n\n<p><strong>(2)<\/strong>  comment out<\/p>\n\n<pre><code>export PATH=\"\/home\/jackie\/anaconda2\/bin:$PATH\".  \n<\/code><\/pre>\n\n<p>Explain: <\/p>\n\n<blockquote>\n  <p>Sometime cmake will try to find library dependency base on python\n  execute (\/home\/jackie\/anaconda2\/bin\/). This will cause some\n  ananconda's 3-rd library sneak in while linking.<\/p>\n<\/blockquote>\n\n<p>P.S. It is also good to use <code>make VERBOSE=1<\/code> to make the cmake project, which can show the gcc compiling command line.<\/p>\n"},{"AnswerId":"43547612","CreationDate":"2017-04-21T16:21:44.583","ParentId":null,"OwnerUserId":"2449404","Title":null,"Body":"<p>I had a similar problem on Ubuntu 16.04. This solved it:<\/p>\n\n<pre><code>apt install lzma-dev\n<\/code><\/pre>\n\n<p>adding <strong>-llzma<\/strong> to link libraries.<\/p>\n"},{"AnswerId":"34540248","CreationDate":"2015-12-31T02:35:28.517","ParentId":null,"OwnerUserId":"5730706","Title":null,"Body":"<p>I get a solution:<\/p>\n\n<p>Just export <code>LD_LIBRARY_PATH=\"\/lib\/x86_64-linux-gnu:$LD_LIBRARY_PATH\"<\/code><\/p>\n"}]}
{"QuestionId":34494796,"AnswerCount":1,"Tags":"<deep-learning><caffe><word2vec><lstm><recurrent-neural-network>","CreationDate":"2015-12-28T13:51:50.743","AcceptedAnswerId":null,"OwnerUserId":4977638.0,"Title":"How to generate a sentence from feature vector or words?","Body":"<p>I used VGG 16-Layer Caffe model for image captions and I have several captions per image. Now, I want to generate a sentence from those captions (words).<br>\nI read in a paper on LSTM that I should remove the SoftMax layer from the training network and provide the 4096 feature vector from  layer directly to LSTM.<br>\nI am new to LSTM and RNN stuff.<br>\nWhere should I begin? Is there any tutorial showing how to generate sentence by sequence labeling?<\/p>\n","answers":[{"AnswerId":"34494925","CreationDate":"2015-12-28T13:59:55.933","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>AFAIK the master branch of BVLC\/caffe does not yet support a recurrent layer architecture.<\/p>\n\n<p>You should pull branch <code>recurrent<\/code> from <a href=\"https:\/\/github.com\/jeffdonahue\/caffe\/tree\/recurrent\" rel=\"nofollow\">jeffdonahue\/caffe<\/a>. This branch supports RNN and LSTM.<br>\nIt also contains a detailed <a href=\"https:\/\/github.com\/jeffdonahue\/caffe\/tree\/recurrent\/examples\/coco_caption\" rel=\"nofollow\">example<\/a> on how to generate image captions trained using <a href=\"http:\/\/mscoco.org\/\" rel=\"nofollow\">MS COCO<\/a> data.<\/p>\n"}]}
{"QuestionId":34498267,"AnswerCount":1,"Tags":"<python><docker><pygame><tensorflow>","CreationDate":"2015-12-28T17:50:14.890","AcceptedAnswerId":null,"OwnerUserId":2245776.0,"Title":"Getting pygame running in docker","Body":"<p>I'm trying to get PyGame running in a docker (so I can eventually work with PyGame and TensorFlow in windows). But when I try run:<\/p>\n\n<pre><\/pre>\n\n<p>I get back<\/p>\n\n<pre><\/pre>\n\n<p>Pip install does work with other packages, e.g. pybrain. Also is successfully installed on my windows machine. But I want to install it within docker so I can use it with TensorFlow.<\/p>\n\n<p>Does anyone have any experience of getting pygame running in docker? Is it even possible?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"34498924","CreationDate":"2015-12-28T18:42:18.023","ParentId":null,"OwnerUserId":"1703772","Title":null,"Body":"<p>You can try installing the security extras:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/29202163\/1703772\">https:\/\/stackoverflow.com\/a\/29202163\/1703772<\/a><\/p>\n\n<p>Some questions:<\/p>\n\n<ul>\n<li><strike>Can you pip install other packages?<\/strike><\/li>\n<li><strike>Is it just pygame that fails?<\/strike><\/li>\n<li>Is installing a pip package locally a viable solution for you? <a href=\"https:\/\/scicomp.stackexchange.com\/questions\/2987\/what-is-the-simplest-way-to-do-a-user-local-install-of-a-python-package\">https:\/\/scicomp.stackexchange.com\/questions\/2987\/what-is-the-simplest-way-to-do-a-user-local-install-of-a-python-package<\/a><\/li>\n<\/ul>\n\n<p>The fact that it reports no versions of pygame accessible is a red flag.\nYou could try installing using a windows installer: <a href=\"https:\/\/bitbucket.org\/pygame\/pygame\/downloads\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/pygame\/pygame\/downloads<\/a> (from <a href=\"https:\/\/scicomp.stackexchange.com\/questions\/2987\/what-is-the-simplest-way-to-do-a-user-local-install-of-a-python-package\">https:\/\/scicomp.stackexchange.com\/questions\/2987\/what-is-the-simplest-way-to-do-a-user-local-install-of-a-python-package<\/a>)<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>From my non-exhaustive searching, <code>Windows<\/code> + <code>Docker<\/code> + <code>pip<\/code> to install <code>pygame<\/code> <strong>will not work at this time.<\/strong> Installing using a windows installer seems to be the accepted solution by many. <\/p>\n\n<p>Here are some solutions you can try:<\/p>\n\n<ul>\n<li>Installing via <code>pip wheel<\/code> locally (64-bit and 32-bit wheels)\n\n<ul>\n<li><strong>Instructions<\/strong>: <a href=\"https:\/\/www.webucator.com\/blog\/2015\/03\/installing-the-windows-64-bit-version-of-pygame\/\" rel=\"nofollow noreferrer\">https:\/\/www.webucator.com\/blog\/2015\/03\/installing-the-windows-64-bit-version-of-pygame\/<\/a><\/li>\n<li><strong>Downloads<\/strong>: <a href=\"http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#pygame\" rel=\"nofollow noreferrer\">http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/#pygame<\/a><\/li>\n<\/ul><\/li>\n<li>Installing Pygame from a windows binary (32-bit only, <code>.msi<\/code>): <a href=\"http:\/\/www.pygame.org\/download.shtml\" rel=\"nofollow noreferrer\">http:\/\/www.pygame.org\/download.shtml<\/a><\/li>\n<li>If Docker is anything like Vagrant (uses vms as applications), you can bring in local wheel files as part of bringing up the application.<\/li>\n<li>Upload the wheels to PyPi and use pip to install that package, but then there are licensing issues you'd have to consider and abide by.<\/li>\n<\/ul>\n\n<p>The current pygame on pypi is 1.7.1 and supports up to Windows 2000\/NT: <a href=\"https:\/\/pypi.python.org\/pypi\/Pygame\/1.7.1\" rel=\"nofollow noreferrer\">https:\/\/pypi.python.org\/pypi\/Pygame\/1.7.1<\/a>? - perhaps your windows version is too new. <\/p>\n\n<p>I hope this has answered your question.<\/p>\n"}]}
{"QuestionId":34498527,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2015-12-28T18:10:19.077","AcceptedAnswerId":34499207.0,"OwnerUserId":5724271.0,"Title":"Pycaffe: How to create custom weights in a python layer?","Body":"<p>I've been searching the net and the caffe source code for a while without any solutions to speak of, but in a custom application neural net, I am building a few <a href=\"https:\/\/stackoverflow.com\/a\/41481539\/1714410\">custom layers in python<\/a>. Forward passes and backward passes are functionally working well, and I can create custom weight parameters in my setup routine, but try as I might I cannot get caffe to set up \"official\" weights for my layer. This would of course allow better snapshotting, easier solver implementation, etc.<\/p>\n\n<p>Any idea what I am missing here?<\/p>\n\n<p>[EDIT: Code from layer shown below. Removed some things for brevity. The purpose of this layer is to add color to the flattened, activated filters from a convolutional layer]<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34499207","CreationDate":"2015-12-28T19:05:10.877","ParentId":null,"OwnerUserId":"5724271","Title":null,"Body":"<p>It's me from the future! Here's how to solve your question:<\/p>\n\n<p>Recently blob adding was implemented Python in Caffe. Here's an example layer that does that:<\/p>\n\n<pre><code>class Param(caffe.Layer):\n    def setup(self, bottom, top):\n        self.blobs.add_blob(1,2,3)\n        self.blobs[0].data[...] = 0\n\n    def reshape(self, bottom, top):\n        top[0].reshape(10)\n\n    def forward(self, bottom, top):\n        print(self.blobs[0].data)\n        self.blobs[0].data[...] += 1\n\n    def backward(self, top, propagate_down, bottom):\n        pass\n<\/code><\/pre>\n\n<p>To access the diffs, just use self.blobs[0].diff[...] and you'll be all set. The solver will take care of the rest. For more info, see <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2944\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/2944<\/a><\/p>\n"}]}
{"QuestionId":34500052,"AnswerCount":1,"Tags":"<python><scikit-learn><tensorflow>","CreationDate":"2015-12-28T20:11:24.167","AcceptedAnswerId":34500690.0,"OwnerUserId":1716733.0,"Title":"tensorflow: saving and restoring session","Body":"<p>I am trying to implement a suggestion from answers:\n<a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-restore-a-previously-saved-model-python\">Tensorflow: how to save\/restore a model?<\/a><\/p>\n\n<p>I have an object which wraps a  model in a  style. <\/p>\n\n<pre><\/pre>\n\n<p>When I run:<\/p>\n\n<pre><\/pre>\n\n<p>I get output:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I try to recover the parameters (even without killing the object):\n<\/p>\n\n<p>I get strange results. First of all, the loaded value does not correspond the saved one.<\/p>\n\n<pre><\/pre>\n\n<p>What is the right way to load, and probably first inspect the saved variables?<\/p>\n","answers":[{"AnswerId":"34500690","CreationDate":"2015-12-28T20:59:34.070","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> You should try to rework this class so that <code>self.create_network()<\/code> is called (i) only once, and (ii) before the <code>tf.train.Saver()<\/code> is constructed.<\/p>\n\n<p>There are two subtle issues here, which are due to the code structure, and the default behavior of the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver.__init__\" rel=\"noreferrer\"><code>tf.train.Saver<\/code> constructor<\/a>. When you construct a saver with no arguments (as in your code), it collects the current set of variables in your program, and adds ops to the graph for saving and restoring them. In your code, when you call <code>tflasso()<\/code>, it will construct a saver, and there will be no variables (because <code>create_network()<\/code> has not yet been called). As a result, the checkpoint should be empty.<\/p>\n\n<p>The second issue is that&mdash;by default&mdash;the format of a saved checkpoint is a map from the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.name\" rel=\"noreferrer\"><code>name<\/code> property of a variable<\/a> to its current value. If you create two variables with the same name, they will be automatically \"uniquified\" by TensorFlow:<\/p>\n\n<pre><code>v = tf.Variable(..., name=\"weights\")\nassert v.name == \"weights\"\nw = tf.Variable(..., name=\"weights\")\nassert v.name == \"weights_1\"  # The \"_1\" is added by TensorFlow.\n<\/code><\/pre>\n\n<p>The consequence of this is that, when you call <code>self.create_network()<\/code> in the second call to <code>tfl.fit()<\/code>, the variables will all have different names from the names that are stored in the checkpoint&mdash;or would have been if the saver had been constructed after the network. (You can avoid this behavior by passing a name-<code>Variable<\/code> dictionary to the saver constructor, but this is usually quite awkward.)<\/p>\n\n<p>There are two main workarounds:<\/p>\n\n<ol>\n<li><p>In each call to <code>tflasso.fit()<\/code>, create the whole model afresh, by defining a new <code>tf.Graph<\/code>, then in that graph building the network and creating a <code>tf.train.Saver<\/code>.<\/p><\/li>\n<li><p><strong>RECOMMENDED<\/strong> Create the network, then the <code>tf.train.Saver<\/code> in the <code>tflasso<\/code> constructor, and reuse this graph on each call to <code>tflasso.fit()<\/code>. Note that you might need to do some more work to reorganize things (in particular, I'm not sure what you do with <code>self.X<\/code> and <code>self.xlen<\/code>) but it should be possible to achieve this with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\">placeholders<\/a> and feeding.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34500905,"AnswerCount":1,"Tags":"<lua><arguments><torch>","CreationDate":"2015-12-28T21:17:49.980","AcceptedAnswerId":34501589.0,"OwnerUserId":4380945.0,"Title":"lua - invalid argument type","Body":"<p>I am a newbie to Lua. Currently getting the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I am sure it is something pretty easy to solve, but cannot manage to find the solution. <\/p>\n\n<p>The 'model' is a file lm_checkpoint_epoch50.00_2.7196.t7, which is in the directory <\/p>\n\n<pre><\/pre>\n\n<p>I am running the program from the parent directory (\/home\/ubuntu\/xxx\/nn)<\/p>\n\n<p>I have tried out the following options to run the program (from one directory above the one the model is saved):<\/p>\n\n<pre><\/pre>\n\n<p>Also, the program has a torch.CmdLine() object where :argument equals '\/cv\/lm_checkpoint_epoch50.00_2.7196.t7'. The program prints the parameters, so that you see the following output on the screen:<\/p>\n\n<pre><\/pre>\n\n<p>so it finds a value for argument 'model', which is picked up from the .lua file, not the parameter in the command line. This file is a valid mode.<\/p>\n\n<p>Pretty lost, hope someone relates to this issue. Thanks.<\/p>\n","answers":[{"AnswerId":"34501589","CreationDate":"2015-12-28T22:17:12.953","ParentId":null,"OwnerUserId":"4380945","Title":null,"Body":"<p>found the issue - it was a bug as smhx suggested. I inadvertently changed the source code from:<\/p>\n\n<pre><code>require 'torch'\ncmd = torch.CmdLine()\ncmd:argument('-model','model checkpoint to use for sampling')\n<\/code><\/pre>\n\n<p>Note that there is no argument in the source code. To:<\/p>\n\n<pre><code>cmd:argument('-model','\/cv\/model lm_chelm_checkpoint_epoch50.00_2.5344.t7'\n'model checkpoint to use for sampling')\n<\/code><\/pre>\n\n<p>So the argument must be passed through the command line, not the source code. With parameters, it is different - you can include them in the source code.<\/p>\n\n<p>So if I change back the source code and run the following from the command line:<\/p>\n\n<pre><code>th sample.lua cv\/lm_chelm_checkpoint_epoch50.00_2.5344.t7\n<\/code><\/pre>\n\n<p>it works. <\/p>\n"}]}
{"QuestionId":34500992,"AnswerCount":4,"Tags":"<machine-learning><classification><tensorflow>","CreationDate":"2015-12-28T21:25:29.243","AcceptedAnswerId":null,"OwnerUserId":5492392.0,"Title":"Tensorflow classification with extremely unbalanced dataset","Body":"<p>I am using TensorFlow LinearClassifier and also DNN to classify two - classes dataset.<\/p>\n\n<p>However, the problem is the dataset contains 96% of Positive output, and 4% of negative output, and my program always return the prediction as Positive. Of course, in this case I will achieved the accuracy of 96%, but it does not make sense at all.<\/p>\n\n<p>What is the good way to deal with this kind of situation?<\/p>\n","answers":[{"AnswerId":"34501073","CreationDate":"2015-12-28T21:32:46.387","ParentId":null,"OwnerUserId":"4231985","Title":null,"Body":"<p>You could try changing the cost function so that a false positive output would be penalized more heavily than a false negative. <\/p>\n"},{"AnswerId":"34501571","CreationDate":"2015-12-28T22:14:38.073","ParentId":null,"OwnerUserId":"5724808","Title":null,"Body":"<p>Here is what you can do, with the simplest solutions first:<\/p>\n\n<ol>\n<li>You can construct minibatches that sample the classes equally and then re-calibrate the model at test time.<\/li>\n<li>You can reweight the examples to favor the negatives<\/li>\n<li>You can use a hinge loss instead of a log loss that might be more robust to the unbalanced data since it would get no gradient when an example is correct beyond the margin<\/li>\n<li>You can explore some other loss function that treats the different types of errors asymetrically<\/li>\n<\/ol>\n"},{"AnswerId":"34501801","CreationDate":"2015-12-28T22:37:07.190","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>You may learn an auto-encoder on the negative examples you have (if their number is kind of large) and then generate examples using an inference technique such as <a href=\"http:\/\/arxiv.org\/abs\/1312.6114\" rel=\"nofollow\">variational Bayes<\/a> or <a href=\"http:\/\/jmlr.csail.mit.edu\/papers\/volume15\/alain14a\/alain14a.pdf\" rel=\"nofollow\">Markov Chain Monte Carlo<\/a>. This way you can increase the number of samples for the negative examples and kind of move towards a more balanced data set.<\/p>\n"},{"AnswerId":"34512234","CreationDate":"2015-12-29T13:36:38.193","ParentId":null,"OwnerUserId":"1286398","Title":null,"Body":"<p>You can check this paper for different sampling techniques for mitigating class imbalance problem <a href=\"http:\/\/www.machinelearning.org\/proceedings\/icml2007\/papers\/62.pdf\" rel=\"nofollow\">http:\/\/www.machinelearning.org\/proceedings\/icml2007\/papers\/62.pdf<\/a>. Simple random minority oversampling usually works better.<\/p>\n"}]}
{"QuestionId":34501887,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2015-12-28T22:47:48.883","AcceptedAnswerId":34502146.0,"OwnerUserId":4459702.0,"Title":"Building a graph with custom ops in tensorflow","Body":"<p>I'm trying to explore use of tensorflow with custom ops. I build a simple switch op and verified it as suggested in tensorflow document. Now I'm trying to build the graph and then call  method in a tensorflow \nSession. Below is my code. I get the following error. Can someone help what should I do to fix it. Do I need to re-install tensorflow everytime I add a new custom op to ?<\/p>\n\n<pre><\/pre>\n\n<p>When executed in python interpreter I get the following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"tensorflow-switch.py\", line 14, in \n      out = tf.simple_switch(input1, input2, input3)\n  AttributeError: 'module' object has no attribute 'simple_switch'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34502146","CreationDate":"2015-12-28T23:12:38.830","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>After <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/adding_an_op\/index.html\" rel=\"nofollow\">adding a user-defined op<\/a> (in TensorFlow 0.6.0 or earlier), to use it in the Python interpreter you must reinstall from the source repository. The easiest way to do this is to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#create-pip\" rel=\"nofollow\">build and install a PIP package<\/a> using Bazel. (The unit test would pass because running <code>bazel test<\/code> would cause TensorFlow to be rebuilt, and the rebuilt version to be used when running the tests.)<\/p>\n\n<p><strong>NOTE:<\/strong> This feature is experimental, and an improved workflow for adding user-defined ops is in development.<\/p>\n"}]}
{"QuestionId":34503471,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><precision><theano>","CreationDate":"2015-12-29T02:09:12.373","AcceptedAnswerId":null,"OwnerUserId":2751512.0,"Title":"NumPy precision when doing dot product","Body":"<p>I am using Theano\/NumPy doing some deep learning stuff. I found a very annoying problem. I got a weights matrix A(suppose to be 50*2048), and a feature vector b(2048 dim).<\/p>\n\n<p>A is initialized using<\/p>\n\n<pre><\/pre>\n\n<p>b is a 2048 dim numpy.ndarrary from theano.<\/p>\n\n<p>The problem is<\/p>\n\n<pre><\/pre>\n\n<p>Some rows of X and Y are not strictly equal. I compared them and found that the difference is in 1e-6 to 1e-7.<\/p>\n\n<p>Currently I prefer to use the second to computed the dot product since it seems that it can learn better weights. But the first is much faster. So I'm wondering why there is such a big difference. Is it caused by different implementations of dot(matrix, vector) and dot(vector, vector)? Thanks a lot!<\/p>\n\n<p>--edit\nAs uhoh mentioned, this is the code that you can reproduce it.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34517319","CreationDate":"2015-12-29T19:00:55.080","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Well there is usually a trade-off between performance and precision. You may have to compensate one in favor or the other. Although I personally do not believe a difference of 0.0000001 is a big deal in most applications.  If you seek higher precision you'd better go with <code>float64<\/code>, but note that <code>float64<\/code> operations are extremely slow on the GPUs, especially NVIDIA 9xx series GPUs.<\/p>\n\n<p>I may note that the mentioned issue seems to depend on your hardware settings too cause I do not encounter such problem on my machine.<\/p>\n\n<p>You may also use <code>np.allclose(x, y)<\/code> to see if the difference is tangible.<\/p>\n"}]}
{"QuestionId":34508431,"AnswerCount":5,"Tags":"<python><dll><installation><theano><pdb>","CreationDate":"2015-12-29T09:45:57.313","AcceptedAnswerId":null,"OwnerUserId":1256369.0,"Title":"Installing Theano on Windows - DLL load failed","Body":"<p>I am trying to install Theano on Windwos 8<\/p>\n\n<p>Have followed <a href=\"https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\/98833\" rel=\"noreferrer\">these<\/a> steps.<\/p>\n\n<p>I try to test using:<\/p>\n\n<pre><\/pre>\n\n<p>I am aware that the script is comparison between numpy and theano computation time.<\/p>\n\n<p>But somehow, some dll is not found. Pl find the following log:<\/p>\n\n<pre><\/pre>\n\n<p>being new to python world, I am not able to find out which Dll in not found. I would install it if missing and would add the path the system path variable. But how do I get to know which dll it is. I tried using  and set a breakpoint in  at various locations, but none gets hit. <\/p>\n\n<ol>\n<li>Are you familiar to this kind of error and its resolution?<\/li>\n<li>Otherwise, help me find that missing dll name, and I'll do the rest.<\/li>\n<li>I do not mind a walking a completely different path in installing . I came across few posts which suggest microsoft compiler works just fine. But I am already sitting on a problem since past 14 hours continuously and would want a method which works. Ultimately I want  on my system.<\/li>\n<\/ol>\n\n<p>BTW, do not have a . I am trying on CPU only.<\/p>\n","answers":[{"AnswerId":"35204973","CreationDate":"2016-02-04T15:27:55.243","ParentId":null,"OwnerUserId":"868972","Title":null,"Body":"<p>OK, I finally found the reason at least for my installation using WinPython, python3.5.1.1, Win 7 64bit and mingw64:<\/p>\n\n<p>In the .theanorc[.txt] file, I ha, after installing blas, a section [blas] with a line including the openblas include path. I removed that line and left the area below [blas] blank. Restarted WinPython\/Spider, opened a new console and it worked.<\/p>\n\n<p>Testet afterwards with running mnist.py from Lasagne using Theano.<\/p>\n"},{"AnswerId":"40630956","CreationDate":"2016-11-16T11:26:23.800","ParentId":null,"OwnerUserId":"6135446","Title":null,"Body":"<p>I was struggling with getting theano to work on Windows 10 x64 as well, and after reading several help guides, what worked for me in the end was the following:<\/p>\n\n<ol>\n<li>Install Intel Distribution for Python (make sure c:\\intelpython27\\ and c:\\intelpython27\\scripts are in the PATH)<\/li>\n<li>Open a command line and do: <code>conda install mingw libpython<\/code><\/li>\n<li>Then do: <code>pip install theano<\/code><\/li>\n<li>Download the precompiled libopenblas.dll from  <a href=\"http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/\" rel=\"nofollow noreferrer\">http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/<\/a> (I got OpenBLAS-v0.2.14-Win64-int32.zip)<\/li>\n<li>Download mingw64_dll.zip as well from <a href=\"http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/mingw64_dll.zip\/download\" rel=\"nofollow noreferrer\">http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/mingw64_dll.zip\/download<\/a><\/li>\n<li>extract OpenBLAS-v0.2.14-Win64-int32.zip to c:\\openblas, for example<\/li>\n<li>extract mingw64_dll.zip to c:\\openblas\\bin<\/li>\n<li>Add c:\\openblas\\lib and c:\\openblas\\bin to PATH<\/li>\n<li>Create .theanorc.txt in c:\\Users\\{username}<\/li>\n<li><p>Put the following in .theanorc.txt:<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = cpu\n\n[blas]\nldflags = -LC:\\\\openblas\\\\bin -LC:\\\\openblas\\\\lib -lopenblas\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Hope this helps someone.<\/p>\n"},{"AnswerId":"40666101","CreationDate":"2016-11-17T22:21:02.253","ParentId":null,"OwnerUserId":"5710883","Title":null,"Body":"<p>Just to add my experience with Win10 64bit and theano (and keras). I was getting the  DLL load fail error as well. Tried all the stuff I have found on theano win install  webpage and in the different forum and blog posts (mentioned here in Uradium's post).<\/p>\n\n<p>It didn't work for me, but here is what worked:<\/p>\n\n<ol>\n<li><p>I installed TDM-GCC-64 and openblas, using regedit I have added their \/bin directories to PATH (and \/lib directory of openblas as well)<\/p><\/li>\n<li><p>made clean venv under miniconda3 - installed prerequisites for theano and keras (scipy, numpy, pyyaml), python 3.5.2<\/p><\/li>\n<li><p><code>conda install mingw libpython<\/code><\/p><\/li>\n<li><p><code>conda install theano<\/code> (installs all the mingw libraries as well)<\/p><\/li>\n<li><p>at this point - <strong>it worked<\/strong> - so I didnt use any <code>.theanorc.txt<\/code> file configuration, no fiddling with Windows environment variables. It even worked for keras based python scripts, even-though I had not installed keras into my conda venv it was working in...(I have keras installed in my \"normal\" non-conda python3, and later I pip installed keras under my conda venv as well).<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"37191753","CreationDate":"2016-05-12T15:46:49.483","ParentId":null,"OwnerUserId":"4824639","Title":null,"Body":"<p>This may be a little bit late. I followed the exact same guide as you, and ran into the same error.<\/p>\n\n<p>Turns out that it can be fixed by adding the openblas directory to system path. In your example, add \"E:\\Things_Installed_Here\\Theano\\openblas\" to system path variable.<\/p>\n\n<p>Note that depending on where you put the DLL files, you either need to add \"E:\\Things_Installed_Here\\Theano\\openblas\" or \"E:\\Things_Installed_Here\\Theano\\openblas\\bin\".<\/p>\n\n<p>Windows 10 64 bit, Python 3.5.1 (Anaconda 2.5), Theano 0.8.2.<\/p>\n\n<p>For those who need it, I am able to install Theano in Windows 10 thanks to these three guides <a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/3hkv2b\/most_recent_way_to_install_theano_for_windows_10\/\" rel=\"noreferrer\">guide1<\/a>, <a href=\"http:\/\/rosinality.ncity.net\/doku.php?id=python:installing_theano\" rel=\"noreferrer\">guide2<\/a> and the guide attached by OP.<\/p>\n"},{"AnswerId":"43072566","CreationDate":"2017-03-28T14:40:08.837","ParentId":null,"OwnerUserId":"5714313","Title":null,"Body":"<p>Quite late, but for future reference. I stumbled upon this thread trying to run an OpenBLAS test script on the CPU. It worked fine on the GPU but the CPU gives me the same errors as the OP. I tried the suggestions here, but that didn't work. Turns out that the line in .theanorc:<\/p>\n\n<pre><code>ldflags=-LE:\\OpenBLAS-v0.2.14-Win64-int32\/bin -lopenblas\n<\/code><\/pre>\n\n<p>shouldn't have the space after bin and before -l. So it should be:<\/p>\n\n<pre><code>ldflags=-LE:\\OpenBLAS-v0.2.14-Win64-int32\/bin-lopenblas\n<\/code><\/pre>\n\n<p>Now everything works. Tried it on:<\/p>\n\n<pre><code>Windows 10\nPython 3.5\nTheano 0.9\nCuda 8 (with CudaMEM and CudaDNN)\nVisual Studio 2015 Community\n<\/code><\/pre>\n"}]}
{"QuestionId":34509830,"AnswerCount":0,"Tags":"<python><c++><gcc><g++><theano>","CreationDate":"2015-12-29T11:14:45.600","AcceptedAnswerId":null,"OwnerUserId":272963.0,"Title":"Get rid of c++ compiler in Theano","Body":"<p>I need to distribute my application, which is using Theano, across multiple machines.\nSo I want to compile all Theano functions once and cache them and make my app independent of g++ for distributions convenience.<\/p>\n\n<p>My first attempt was to save Theano functions in app's folder:<\/p>\n\n<pre><\/pre>\n\n<p>Then change  option, so it will looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>And I get this exception:<\/p>\n\n<pre><\/pre>\n\n<p>I spent a lot of time trying to solve this problem. So, is it possible at all?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":34511680,"AnswerCount":3,"Tags":"<algorithm><optimization><matrix><vectorization><theano>","CreationDate":"2015-12-29T13:03:31.810","AcceptedAnswerId":34518383.0,"OwnerUserId":3041068.0,"Title":"Loop over (or vectorize) variable length matrices in Theano","Body":"<p>I have a list of matrices , where each item  is a  matrix ( is a variable,  is a constant).  <\/p>\n\n<p>I want to compute the sum of  for all items in  ( is the transpose of ) as the following Python code does:<\/p>\n\n<pre><\/pre>\n\n<p>Actually I want to implement this in Theano (which doesn't support variable length multidimensional arrays), and I don't want to pad all matrices to the same size because that will waste too much space (some of the matrices can be very large).<\/p>\n\n<p>Is there a better way to do this?<\/p>\n\n<p><strong>Edit<\/strong>:<\/p>\n\n<p> is known before the Theano compilation.<\/p>\n\n<p><strong>Edit<\/strong>:<\/p>\n\n<p>received two excellent answers from @DanielRenshaw and @Divakar, emotionally difficult to choose one to accept.<\/p>\n","answers":[{"AnswerId":"34518383","CreationDate":"2015-12-29T20:16:55.670","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>You can just pad the input arrays along the first axis that is all the <code>x<\/code>'s add up. Thus, we would end up with a tall <code>(X,n)<\/code> array, where <code>X =x1+x2+x3+....<\/code>. This can be transposed and its dot product with its self would be the desired output of shape <code>(n,n)<\/code>. All of this is achieved with a pure vectorized solution leveraging the powerful dot product. Thus, the implementation would be -<\/p>\n\n<pre><code># Concatenate along axis=0\nLcat = np.concatenate(L,axis=0)\n\n# Perform dot product of the transposed version with self\nout = Lcat.T.dot(Lcat)\n<\/code><\/pre>\n\n<p>Runtime tests and verify output -<\/p>\n\n<pre><code>In [116]: def vectoized_approach(L):\n     ...:   Lcat = np.concatenate(L,axis=0)\n     ...:   return Lcat.T.dot(Lcat)\n     ...: \n     ...: def original_app(L):\n     ...:   n = L[0].shape[1]\n     ...:   res = np.zeros((n,n))\n     ...:   for M in L:\n     ...:     res += np.dot(M.T, M)\n     ...:   return res\n     ...: \n\nIn [117]: # Input\n     ...: L = [np.random.rand(np.random.randint(1,9),5)for iter in range(1000)]\n\nIn [118]: np.allclose(vectoized_approach(L),original_app(L))\nOut[118]: True\n\nIn [119]: %timeit original_app(L)\n100 loops, best of 3: 3.84 ms per loop\n\nIn [120]: %timeit vectoized_approach(L)\n1000 loops, best of 3: 632 \u00b5s per loop\n<\/code><\/pre>\n"},{"AnswerId":"34542846","CreationDate":"2015-12-31T07:31:23.570","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>In addition to @DanielRenshaw's answer, if we increase the number of matrices to 1000, the <code>compile_theano_version1<\/code> function will yield <code>RuntimeError: maximum recursion depth exceeded<\/code>, and <code>compile_theano_version2<\/code> seems to take forever to compile.<\/p>\n\n<p>There's a fix for this by using <code>typed_list<\/code>:<\/p>\n\n<pre><code>def compile_theano_version4(number_of_matrices, n):\n    import theano.typed_list\n    L = theano.typed_list.TypedListType(tt.TensorType(theano.config.floatX, broadcastable=(None, None)))()\n    res, _ = theano.scan(fn=lambda i: tt.dot(L[i].T, L[i]),\n                         sequences=[theano.tensor.arange(number_of_matrices, dtype='int64')])\n    return theano.function([L], res.sum(axis=0))\n<\/code><\/pre>\n\n<p>Moreover, I set the data type of all relevant variables to <code>float32<\/code> and ran @DanielRenshaw's script on GPU, it turned out that @Divakar's suggestion (<code>theano_version3<\/code>) is the most efficient in this case. Though as @DanielRenshaw said, using a huge matrix may not always be a good practice. <\/p>\n\n<p>The followings are settings and outputs on my machine.<\/p>\n\n<pre><code>iteration_count = 100\nnumber_of_matrices = 200\nn = 300\nmin_x = 20\ndtype = 'float32'\ntheano.config.floatX = dtype\n\n\nnumpy_version1 5.30542397499\nnumpy_version2 3.96656394005\ntheano_version1 5.26742005348\ntheano_version2 1.76983904839\ntheano_version3 1.03577589989\ntheano_version4 5.58366179466\n<\/code><\/pre>\n"},{"AnswerId":"34512763","CreationDate":"2015-12-29T14:08:32.130","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Given that the number of matrices is known before the Theano compilation needs to happen, one can simply use regular Python lists of Theano matrices.<\/p>\n\n<p>Here's a complete example showing the difference between numpy and Theano versions.<\/p>\n\n<p>This code has been updated to include a comparison with @Divakar's vectorized approach which performs better. Two vectorized approaches are possible for Theano, one where Theano performs the concatenation, and one where numpy does the concatenation the result of which is then passed to Theano.<\/p>\n\n<pre><code>import timeit\nimport numpy as np\nimport theano\nimport theano.tensor as tt\n\n\ndef compile_theano_version1(number_of_matrices, n, dtype):\n    assert number_of_matrices &gt; 0\n    assert n &gt; 0\n    L = [tt.matrix() for _ in xrange(number_of_matrices)]\n    res = tt.zeros(n, dtype=dtype)\n    for M in L:\n        res += tt.dot(M.T, M)\n    return theano.function(L, res)\n\n\ndef compile_theano_version2(number_of_matrices):\n    assert number_of_matrices &gt; 0\n    L = [tt.matrix() for _ in xrange(number_of_matrices)]\n    concatenated_L = tt.concatenate(L, axis=0)\n    res = tt.dot(concatenated_L.T, concatenated_L)\n    return theano.function(L, res)\n\n\ndef compile_theano_version3():\n    concatenated_L = tt.matrix()\n    res = tt.dot(concatenated_L.T, concatenated_L)\n    return theano.function([concatenated_L], res)\n\n\ndef numpy_version1(*L):\n    assert len(L) &gt; 0\n    n = L[0].shape[1]\n    res = np.zeros((n, n), dtype=L[0].dtype)\n    for M in L:\n        res += np.dot(M.T, M)\n    return res\n\n\ndef numpy_version2(*L):\n    concatenated_L = np.concatenate(L, axis=0)\n    return np.dot(concatenated_L.T, concatenated_L)\n\n\ndef main():\n    iteration_count = 100\n    number_of_matrices = 20\n    n = 300\n    min_x = 400\n    dtype = 'float64'\n    theano_version1 = compile_theano_version1(number_of_matrices, n, dtype)\n    theano_version2 = compile_theano_version2(number_of_matrices)\n    theano_version3 = compile_theano_version3()\n    L = [np.random.standard_normal(size=(x, n)).astype(dtype)\n         for x in range(min_x, number_of_matrices + min_x)]\n\n    start = timeit.default_timer()\n    numpy_res1 = np.sum(numpy_version1(*L)\n                        for _ in xrange(iteration_count))\n    print 'numpy_version1', timeit.default_timer() - start\n\n    start = timeit.default_timer()\n    numpy_res2 = np.sum(numpy_version2(*L)\n                        for _ in xrange(iteration_count))\n    print 'numpy_version2', timeit.default_timer() - start\n\n    start = timeit.default_timer()\n    theano_res1 = np.sum(theano_version1(*L)\n                         for _ in xrange(iteration_count))\n    print 'theano_version1', timeit.default_timer() - start\n\n    start = timeit.default_timer()\n    theano_res2 = np.sum(theano_version2(*L)\n                         for _ in xrange(iteration_count))\n    print 'theano_version2', timeit.default_timer() - start\n\n    start = timeit.default_timer()\n    theano_res3 = np.sum(theano_version3(np.concatenate(L, axis=0))\n                         for _ in xrange(iteration_count))\n    print 'theano_version3', timeit.default_timer() - start\n\n    assert np.allclose(numpy_res1, numpy_res2)\n    assert np.allclose(numpy_res2, theano_res1)\n    assert np.allclose(theano_res1, theano_res2)\n    assert np.allclose(theano_res2, theano_res3)\n\n\nmain()\n<\/code><\/pre>\n\n<p>When run this prints (something like)<\/p>\n\n<pre><code>numpy_version1 1.47830819649\nnumpy_version2 1.77405482179\ntheano_version1 1.3603150303\ntheano_version2 1.81665318145\ntheano_version3 1.86912039489\n<\/code><\/pre>\n\n<p>The asserts pass, showing that the Theano and numpy versions both compute the same result to high degree of accuracy. Clearly this accuracy will reduce if using <code>float32<\/code> instead of <code>float64<\/code>.<\/p>\n\n<p>The timing results show that the vectorized approach may not be preferable, it depends on the matrix sizes. In the example above the matrices are large and the non-concatenation approach is faster but if the <code>n<\/code> and <code>min_x<\/code> parameters are changed in the <code>main<\/code> function to be much smaller then the concatenation approach is quicker. Other results may hold when running on a GPU (Theano versions only).<\/p>\n"}]}
{"QuestionId":34514324,"AnswerCount":1,"Tags":"<gpgpu><tensorflow>","CreationDate":"2015-12-29T15:45:53.633","AcceptedAnswerId":null,"OwnerUserId":5654767.0,"Title":"Error using Tensorflow with GPU","Body":"<p>I've tried a bunch of different Tensorflow examples, which works fine on the CPU but generates the same error when I'm trying to run them on the GPU. One little example is this:<\/p>\n\n<pre><\/pre>\n\n<p>The error is always the same, CUDA_ERROR_OUT_OF_MEMORY:<\/p>\n\n<pre><\/pre>\n\n<p>I guess that the problem has to do with my configuration rather than the memory usage of this tiny example. Does anyone have any idea? <\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>I've found out that the problem may be as simple as someone else running a job on the same GPU, which would explain the little amount of free memory. In that case: sorry for taking up your time...<\/p>\n","answers":[{"AnswerId":"34514932","CreationDate":"2015-12-29T16:20:06.413","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There appear to be two issues here:<\/p>\n\n<ol>\n<li><p>By default, TensorFlow allocates a large fraction (95%) of the available GPU memory (on each GPU device) when you create a <code>tf.Session<\/code>. It uses a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/266083148c9a1544cd7e3d1881111133045b030b\/tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc#L507\" rel=\"noreferrer\">heuristic<\/a> that reserves 200MB of GPU memory for \"system\" uses, but doesn't set this aside if the amount of free memory is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/266083148c9a1544cd7e3d1881111133045b030b\/tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc#L550\" rel=\"noreferrer\">smaller than that<\/a>.<\/p><\/li>\n<li><p>It looks like you have very little free GPU memory on either of your GPU devices (105.73MiB and 133.48MiB). This means that TensorFlow will attempt to allocate memory that should probably be reserved for the system, and hence the allocation fails. <\/p><\/li>\n<\/ol>\n\n<p>Is it possible that you have another TensorFlow process (or some other GPU-hungry code) running while you attempt to run this program? For example, a Python interpreter with an open session&mdash;even if it is not using the GPU&mdash;will attempt to allocate almost the entire GPU memory.<\/p>\n\n<p>Currently, the only way to restrict the amount of GPU memory that TensorFlow uses is the following configuration option (from <a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory\">this question<\/a>):<\/p>\n\n<pre><code># Assume that you have 12GB of GPU memory and want to allocate ~4GB:\ngpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)\n\nsess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))\n<\/code><\/pre>\n"}]}
{"QuestionId":34519423,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2015-12-29T21:33:13.250","AcceptedAnswerId":34523795.0,"OwnerUserId":1878136.0,"Title":"TensorFlow error: logits and labels must be same size","Body":"<p>I've been trying to learn TensorFlow by implementing ApproximatelyAlexNet based on various examples on the internet. Basically extending the AlexNet example <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/alexnet.py\" rel=\"noreferrer\">here<\/a> to take 224x224 RGB images (rather than 28x28 grayscale images), and adding a couple more layers, changing kernel sizes, strides, etc, per other AlexNet implementations I've found online.<\/p>\n\n<p>Have worked through a number of mismatched shape type errors, but this one has me stumped:<\/p>\n\n<pre><\/pre>\n\n<p>The 49 dimension is particularly puzzling. For debugging, my batch size is currently 1, if I increase it to 2 then the 49 becomes 98.<\/p>\n\n<p>If I log the shape of the x and y that I pass to <\/p>\n\n<pre><\/pre>\n\n<p>I get<\/p>\n\n<pre><\/pre>\n\n<p>Which is as expected: 150528 = 224 * 224 RGB pixels, and a one-hot vector representing my labels.<\/p>\n\n<p>Would appreciate any help in figuring this out!<\/p>\n\n<p><strong>Update:<\/strong> code exhibiting the fault here:<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/j4m3z0r\/e70096d0f7bd4bd24c42\" rel=\"noreferrer\">https:\/\/gist.github.com\/j4m3z0r\/e70096d0f7bd4bd24c42<\/a><\/p>\n","answers":[{"AnswerId":"34523795","CreationDate":"2015-12-30T06:00:59.997","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Thanks for sharing your code as a Gist. There are two changes that are necessary to make the shapes agree:<\/p>\n\n<ol>\n<li><p>The line:<\/p>\n\n<pre><code>fc1 = tf.reshape(pool5, [-1, wd1Shape[0]])\n<\/code><\/pre>\n\n<p>...is responsible for the erroneous <code>49<\/code> in the batch dimension. The input is 1 x 7 x 7 x 256, and it is reshaped to be 49 x 256, because <code>wd1Shape[0]<\/code> is 256. One possible replacement is the following:<\/p>\n\n<pre><code>pool5Shape = pool5.get_shape().as_list()\nfc1 = tf.reshape(pool5, [-1, pool5Shape[1] * pool5Shape[2] * pool5Shape[3]])\n<\/code><\/pre>\n\n<p>...which will give  <code>fc1<\/code> the shape 1 x 12544.<\/p><\/li>\n<li><p>After making this change, the size of the <code>'wd1'<\/code> weight matrix (256 x 4096) doesn't match the number of nodes in <code>fc1<\/code>. You could change the definition of this matrix as follows:<\/p>\n\n<pre><code>    'wd1': tf.Variable(tf.random_normal([12544, 4096])),\n<\/code><\/pre>\n\n<p>...although you may want to modify the other weights, or perform additional pooling to reduce the size of this matrix.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"48667431","CreationDate":"2018-02-07T15:26:03.887","ParentId":null,"OwnerUserId":"8520208","Title":null,"Body":"<p>This issue is because your class variable and label is not matching.<\/p>\n\n<p>ex:- \nIn your code, You declared class variable as 10\nBut label might not be 10.<\/p>\n\n<p>Once you made your class variable and label same dimension. This issue will be resolved.<\/p>\n"},{"AnswerId":"34521012","CreationDate":"2015-12-29T23:57:18.180","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>Given that you didn't provide an actual code you are using it's hard to exactly say what's wrong.<\/p>\n\n<p>Here are some general tips for debugging such problems:<\/p>\n\n<ul>\n<li><p>Add <code>print(tensor.get_shape())<\/code> to the places related to the problem (in your case dense2, out, _weights['out'], _biases['out'] are suspects).<\/p><\/li>\n<li><p>Make sure your matrix multiplications are in right order (e.g. dense2 by _weights['out'], should result in batch_size x 10 matrix).<\/p><\/li>\n<\/ul>\n\n<p>If you modified the code in AlexNet you linked, you probably have changed next lines:<\/p>\n\n<pre><code> dense1 = tf.reshape(norm3, [-1, _weights['wd1'].get_shape().as_list()[0]]) # Reshape conv3 output to fit dense layer input\n dense1 = tf.nn.relu(tf.matmul(dense1, _weights['wd1']) + _biases['bd1'], name='fc1') # Relu activation\n dense2 = tf.nn.relu(tf.matmul(dense1, _weights['wd2']) + _biases['bd2'], name='fc2') # Relu activation\n out = tf.matmul(dense2, _weights['out']) + _biases['out']\n<\/code><\/pre>\n\n<p>Probably shape of <code>dense2<\/code> is [49, 1024] in your case. You can check by adding a print <code>dense2.get_shape()<\/code>. You should prints for shapes for all the tensors until you find one that gets 49. I can only guess what you changed, but it can be one of reshapes.<\/p>\n"}]}
{"QuestionId":34519627,"AnswerCount":2,"Tags":"<python><tensorflow><cost-based-optimizer>","CreationDate":"2015-12-29T21:50:33.783","AcceptedAnswerId":34520066.0,"OwnerUserId":2411173.0,"Title":"sum over a list of tensors in tensorflow","Body":"<p>I have a deep neural network where the weights between layers are stored in a list. <\/p>\n\n<p> I want to incluse the ridge penalty in my cost function. I need then to use something like \n i.e. the squared sum of all the weights. <\/p>\n\n<p>In particular the weights are defined as:<\/p>\n\n<pre><\/pre>\n\n<p>How can I do that in tensorflow ?<\/p>\n","answers":[{"AnswerId":"34520066","CreationDate":"2015-12-29T22:25:05.580","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The standard way to sum a list of tensors is to use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/add_n\" rel=\"nofollow noreferrer\"><code>tf.add_n()<\/code><\/a> operation, which takes a list of tensors (each having the same size and shape) and produces a single tensor containing the sum.<\/p>\n\n<p>For the particular problem that you have, I am assuming that each <code>layers[j].weights<\/code> could have a different size. Therefore you will need reduce each element down to a scalar before summing, e.g. using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/l2_loss\" rel=\"nofollow noreferrer\"><code>tf.nn.l2_loss()<\/code><\/a> function itself:<\/p>\n\n<pre><code>weights = [layers[j].weights for j in range(self.n_layers)]\nlosses = [tf.nn.l2_loss(w) for w in weights]\ntotal_loss = tf.add_n(losses)\n<\/code><\/pre>\n\n<p>(Note however that when the values to be added are large, you may find it more efficient to calculate a sequence of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/add\" rel=\"nofollow noreferrer\"><code>tf.add()<\/code><\/a> operations, since TensorFlow keeps the values of each of the <code>add_n<\/code> arguments in memory until <em>all<\/em> of them have been computed. A chain of <code>add<\/code> ops allows some of the computation to happen earlier.)<\/p>\n"},{"AnswerId":"35615274","CreationDate":"2016-02-24T23:23:43.600","ParentId":null,"OwnerUserId":"1816995","Title":null,"Body":"<p>The <code>tf.nn.l2_loss()<\/code> function returns a tensor with 0 dimensions. <\/p>\n\n<p>But it's nice to not need to manually apply that to each weight tensor, so storing the weight tensors in a list is one way to solve the problem (as @mrry noted).<\/p>\n\n<p>But rather than needing to write that out every time, what you could do is use the following function<\/p>\n\n<pre><code>def l2_loss_sum(list_o_tensors):\n    return tf.add_n([tf.nn.l2_loss(t) for t in list_o_tensors])\n<\/code><\/pre>\n\n<p>In your case this would look like: <\/p>\n\n<pre><code>total_loss = l2_loss_sum([layers[j].weights for j in range(self.n_layers)])\n<\/code><\/pre>\n\n<p>Also, <code>tf.nn.l2_loss()<\/code> implicitly applies the squaring operation to the values as well as multiplying all the squared values by 1\/2 , so were you use something like <code>tf.nn.l2_loss(layers[j].weights**2 for j in range(self.n_layers))<\/code> you would actually be raising the weights to the 4th power. As a result your derivative of this loss term would be strange: it wouldn't cancel the 1\/2 to 1 (but would implicitly be doubling your \u03b2) and the weights would be cubed.<\/p>\n"}]}
{"QuestionId":34520471,"AnswerCount":1,"Tags":"<python><performance><optimization><gpu><theano>","CreationDate":"2015-12-29T23:05:02.167","AcceptedAnswerId":null,"OwnerUserId":498892.0,"Title":"How to force Theano to parallelize an operation on GPU (test case: numpy.bincount)","Body":"<p>I am looking for possibility to speed up computation of bincount using GPU.<\/p>\n\n<p>Reference code in numpy:<\/p>\n\n<pre><\/pre>\n\n<p>I want to measure only speed of operation, not the time spent on passing array, so I create a shared variable:<\/p>\n\n<pre><\/pre>\n\n<p>This operation is of course highly parallelizable, but in practice on GPU this code is times slower than CPU version:<\/p>\n\n<pre><\/pre>\n\n<p>So my questions are:<\/p>\n\n<ol>\n<li>What could be the reason for such low performance?<\/li>\n<li>Can I write parallel version of bincount using theano?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"34520841","CreationDate":"2015-12-29T23:39:24.693","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>I think you cannot increase this operation on the GPU further unless you can somehow manually tell Theano to do in in a parallelized manner, which seems not to be possible. On the GPU, the computations that are not to be done in parallel will be done at the same speed or slower compared to CPU.<\/p>\n\n<p>Quote from <a href=\"https:\/\/stackoverflow.com\/users\/127480\/daniel-renshaw\">Daniel Renshaw<\/a>:<\/p>\n\n<blockquote>\n  <p>To an extent, Theano expects you to focus more on what you want\n  computed rather than on how you want it computed. The idea is that the\n  Theano optimizing compiler will automatically parallelize as much as\n  possible (either on GPU or on CPU using OpenMP).<\/p>\n<\/blockquote>\n\n<p>And another quote:<\/p>\n\n<blockquote>\n  <p>You need to be able to specify your computation in terms of Theano operations. If those operations can be parallelized on the GPU, they should be parallelized automatically.<\/p>\n<\/blockquote>\n\n<p>Quote from Theano's webpage:<\/p>\n\n<blockquote>\n  <ul>\n  <li>Indexing, dimension-shuffling and constant-time reshaping will be\n  equally fast on GPU as on CPU. <\/li>\n  <li>Summation over rows\/columns of tensors\n  can be a little slower on the GPU than on the CPU.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>I think the only thing you can do is to set the <code>openmp<\/code> flag to <code>True<\/code> in your <code>.theanorc<\/code> file.<\/p>\n\n<p>Anyway I tried an idea. It does not work for now, but hopefully someone can help us make it work. If worked, you might be able to parallelize the operation on the GPU. The code below tries to do EVERYTHING in the GPU with CUDA API. However, there are two bottle-necks not allowing the operation take place: 1) <strong>Currently (as of Jan. 4th, 2016)<\/strong> Theano and CUDA do not support any operations on any data type rather than <strong>float32<\/strong> and 2) <code>T.extra_ops.bincount()<\/code> only works with <code>int<\/code> data types. So it might be the bottleneck for Theano not being able to fully parallelize the operation.<\/p>\n\n<pre><code>import theano.tensor as T\nfrom theano import shared, Out, function\nimport numpy as np\nimport theano.sandbox.cuda.basic_ops as sbasic\n\nshared_var = shared(np.random.randint(0, 1000, 1000000).astype(T.config.floatX), borrow = True)\nx = T.vector('x');\ncomputeFunc = T.extra_ops.bincount(sbasic.as_cuda_ndarray_variable(T.cast(x, 'int16')))\nfunc = function([], Out(sbasic.gpu_from_host(computeFunc), borrow = True), givens = {x: shared_var})\n<\/code><\/pre>\n\n<p><strong>Sources<\/strong><\/p>\n\n<p>1- <a href=\"https:\/\/stackoverflow.com\/questions\/31777135\/how-do-i-set-many-elements-in-parallel-in-theano\">How do I set many elements in parallel in theano<\/a><\/p>\n\n<p>2- <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#what-can-be-accelerated-on-the-gpu\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#what-can-be-accelerated-on-the-gpu<\/a><\/p>\n\n<p>3- <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/multi_cores.html<\/a><\/p>\n"}]}
{"QuestionId":34522571,"AnswerCount":1,"Tags":"<python><neural-network><classification><hdf5><caffe>","CreationDate":"2015-12-30T03:37:56.737","AcceptedAnswerId":34624313.0,"OwnerUserId":2476920.0,"Title":"Caffe: predict multiple labels at once (in Python)","Body":"<p>In caffe, I would like to be able to predict multiple labels at once, such as keyboard arrow keys: two keys could be pressed at the same time. I am trying to drive a virtual F1 car in the  game with a convolutional neural network, I plan to gather and shape the training data very soon and I would like to know if I do things right. <\/p>\n\n<p>I think this post would give a good example on how to do such classification in Python, I did not found any satisfying examples on how to do that yet. <\/p>\n\n<p>Could someone please validate that this way to gather and represent the data in the neural network will work as I expect to: <\/p>\n\n<h2>HDF5 python code<\/h2>\n\n<pre><\/pre>\n\n<h2>Info about HDF5 data shapes<\/h2>\n\n<p>Input: <\/p>\n\n<pre><\/pre>\n\n<p>Output: <\/p>\n\n<p><em>Note: I use numpy's \"int_\" to get a label class to be classified.<\/em><\/p>\n\n<pre><\/pre>\n\n<h2>Convolutional neural network architecture<\/h2>\n\n<p>I have put some semi-related comments here, I would appreciate any opinion on the architecture of the network too if it can make it more performant :)<\/p>\n\n<pre><\/pre>\n\n<p>Moreover, I would like to have only one of the left arrow key or right arrow key pressed at once. Instead of doing it programactically afterwards, would it be a good idea to fusion label_right and label_left like this considering I will use some : <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34624313","CreationDate":"2016-01-06T01:40:37.317","ParentId":null,"OwnerUserId":"2476920","Title":null,"Body":"<p>Finally, what I did was the right thing to do for the task, except that the concat layer might not work because of the difference in the joined layers' shape.\nI tested this with the cifar-100 dataset where there are both coarse and fine labels, and it worked well. <\/p>\n"}]}
{"QuestionId":34522668,"AnswerCount":2,"Tags":"<python><neural-network><concat><convolution><caffe>","CreationDate":"2015-12-30T03:54:19.563","AcceptedAnswerId":35422360.0,"OwnerUserId":2476920.0,"Title":"Caffe: concatenation layer in python (L.Concat)","Body":"<p>I wonder how to perform a concatenation of two layers into one in python. More specifically, I want to join the output of a pooling (subsampling) layer with not-visual data to then put a fully connected layer on top of that. <\/p>\n\n<p>Let's say the subsampling layer will output neurons with shape 64*2*2 (if we ignore the caffe batch_size) and that the data layer I want to join on contains only 1 feature (a speed float that ranges from 0 to 1). <\/p>\n\n<p>Here is some dummy code to put you in context: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35422360","CreationDate":"2016-02-16T01:59:48.197","ParentId":null,"OwnerUserId":"2888303","Title":null,"Body":"<p>I'm not sure if you have figured out the answer to your question, but if you haven't then you may want to try the following:<\/p>\n\n<pre><code>bottom_layers = [n.relu4, n.data_speed]\nn.join_speed = L.Concat(*bottom_layers)\n<\/code><\/pre>\n\n<p>The above should allow you to call Concat layer via pycaffe\/python layer.<\/p>\n"},{"AnswerId":"38973341","CreationDate":"2016-08-16T11:08:57.477","ParentId":null,"OwnerUserId":"3817456","Title":null,"Body":"<p>I am similarly trying to do a python-generated deconv layer<\/p>\n\n<pre><code>n.deconv8 = L.Deconvolution(n.conv7)\n<\/code><\/pre>\n\n<p>works ok but <\/p>\n\n<pre><code>n.deconv8 = L.Deconvolution(n.conv7,kernel_size=2)\n<\/code><\/pre>\n\n<p>hits<\/p>\n\n<pre><code>AttributeError: 'LayerParameter' object has no attribute 'kernel_size'\n<\/code><\/pre>\n\n<p>and similarly for <\/p>\n\n<pre><code>n.deconv8 = L.Deconvolution(n.conv7,num_output=1024)\n<\/code><\/pre>\n\n<p>so is there some new syntax for indicating these parameters (also weight_filler, bias_filler,stride)<\/p>\n"}]}
{"QuestionId":34529927,"AnswerCount":2,"Tags":"<linux><machine-learning><tensorflow><system-requirements><tensorboard>","CreationDate":"2015-12-30T12:52:42.733","AcceptedAnswerId":null,"OwnerUserId":3252956.0,"Title":"Tensorflow recommended system specifications?","Body":"<p>I am getting started with installation of Tensorflow on my RHEL 6.5 box. But it turns out that Tensorflow needs glibc >= 2.17 and the default glibc on rhel 6.5 is 2.12. <\/p>\n\n<p>I was wondering if anybody could help me with minimum\/recommended system specifications for tensorflow?<\/p>\n","answers":[{"AnswerId":"34534045","CreationDate":"2015-12-30T17:17:00.553","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow requirements are listed <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#requirements\" rel=\"nofollow\">here<\/a>, but these do not recommend a particular operating system or glibc version.<\/p>\n\n<p>The best-supported operating systems are Ubuntu 14.04 64-bit, and Mac OS X 10.10 (Yosemite) and later. The current limiting factor is the set of <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"nofollow\">supported operating systems for Bazel<\/a>, which we use to make the binary packages. You may be able to <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"nofollow\">install Bazel from source<\/a>, and then <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\" rel=\"nofollow\">install TensorFlow from source<\/a>, to get around these issues. Many users find it easier to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#docker_install\" rel=\"nofollow\">install TensorFlow in a Docker container<\/a> to avoid this problem.<\/p>\n"},{"AnswerId":"37077405","CreationDate":"2016-05-06T16:37:27.203","ParentId":null,"OwnerUserId":"1303432","Title":null,"Body":"<p>Although your system is old and you can not install tensorflow from pip directly, you can install it from source. \nI installed tensorflow from source successfully. My system is Redhat enterprises 6.0, gcc version is 6.1.0, bazel version is 0.1.4.\nFirst you need to install a new gcc version, then you need to compile bazel, which is a tool to compile tensorflow. DO NOT use the newest version of bazel !!!! The 0.1.4 is fine. You can follow this page to find some information.<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110<\/a><\/p>\n"}]}
{"QuestionId":34533424,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2015-12-30T16:38:15.020","AcceptedAnswerId":34533701.0,"OwnerUserId":1716733.0,"Title":"tensorflow concatenation error: a cryptic string formatting issue","Body":"<p>I am trying to implement L1 norm on bias + weights. For this I try to concatenate them together and take a mean. <\/p>\n\n<p>Namely, I have a bias  (shape: ) and weights  (shape: ). So I tried na\u00efvely to concatenate along 1st dimension:<\/p>\n\n<pre><\/pre>\n\n<p>However I am getting:<\/p>\n\n<pre><\/pre>\n\n<p>Adding a name argument does not help.<\/p>\n\n<p>As an alternative,  works with me, but what is the problem with ?<\/p>\n","answers":[{"AnswerId":"34533701","CreationDate":"2015-12-30T16:55:24.950","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue seems to be caused by a bug on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/266083148c9a1544cd7e3d1881111133045b030b\/tensorflow\/python\/ops\/array_ops.py#L302\">this line of the <code>tf.concat()<\/code> implementation<\/a>. At present&mdash;and unlike most other variadic TensorFlow operators&mdash;<code>tf.concat()<\/code> only accepts <em>lists<\/em> of tensors, and does not handle <em>tuples<\/em> correctly. If you change your code to the following, it should work:<\/p>\n\n<pre><code> l1_penalty = tf.reduce_mean(tf.abs(tf.concat(1, [self.W1, self.b1]))) \n<\/code><\/pre>\n\n<p>In the meantime, I'll fix this bug and get it submitted upstream, so that it's handled in the next version.<\/p>\n"}]}
{"QuestionId":34536340,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><reinforcement-learning><q-learning>","CreationDate":"2015-12-30T19:56:12.037","AcceptedAnswerId":34544424.0,"OwnerUserId":5547331.0,"Title":"How to use Tensorflow Optimizer without recomputing activations in reinforcement learning program that returns control after each iteration?","Body":"<p>EDIT(1\/3\/16): <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/672\" rel=\"nofollow noreferrer\">corresponding github issue<\/a><\/p>\n\n<p>I'm using Tensorflow (Python interface) to implement a q-learning agent with function approximation trained using stochastic gradient-descent.  At each iteration of the experiment a step function in the agent is called that updates the parameters of the approximator based on the new reward and activation, and then chooses a new action to perform.<\/p>\n\n<p>Here is the problem(with reinforcement learning jargon):<\/p>\n\n<ul>\n<li>The agent computes its state-action value predictions to choose an action.<\/li>\n<li>Then gives control back another program which simulates a step in the environment.<\/li>\n<li>Now the agent's step function is called for the next iteration.  I want to use Tensorflow's Optimizer class to compute the gradients for me.  However, this requires both the state-action value predictions that I computed last step, AND their graph.  So:\n\n<ul>\n<li>If I run the optimizer on the whole graph, then it has to recompute the state-action value predictions.<\/li>\n<li>But, if I store the prediction (for the chosen action) as a variable, then feed it to the optimizer as a placeholder, it no longer has the graph necessary to compute the gradients.<\/li>\n<li>I can't just run it all in the same sess.run() statement, because I have to give up control and return the chosen action in order to get the next observation and reward (to use in the target for the loss function).<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>So, is there a way that I can (without reinforcement learning jargon):<\/p>\n\n<ol>\n<li>Compute part of my graph, returning value1.<\/li>\n<li>Return value1 to the calling program to compute value2<\/li>\n<li>In the next iteration, use value2 to as part of my loss function for gradient descent WITHOUT recomputing the the part of the graph that computes value1.<\/li>\n<\/ol>\n\n<p>Of course, I've considered the obvious solutions:<\/p>\n\n<ol>\n<li><p>Just hardcode the gradients:  This would be easy for the really simple approximators I'm using now, but would be really inconvenient if I were experimenting with different filters and activation functions in a big convolutional network.  I'd really like to use the Optimizer class if possible.<\/p><\/li>\n<li><p>Call the environment simulation from within the agent:  <a href=\"https:\/\/github.com\/asrivat1\/DeepLearningVideoGames\/blob\/master\/deep_q_network.py#L135\" rel=\"nofollow noreferrer\">This system<\/a> does this, but it would make mine more complicated, and remove a lot of the modularity and structure.  So, I don't want to do this.<\/p><\/li>\n<\/ol>\n\n<p>I've read through the API and whitepaper several times, but can't seem to come up with a solution.  I was trying to come up with some way to feed the target into a graph to calculate the gradients, but couldn't come up with a way to build that graph automatically.<\/p>\n\n<p>If it turns out this isn't possible in TensorFlow yet, do you think it would be very complicated to implement this as a new operator? (I haven't used C++ in a couple of years, so the TensorFlow source looks a little intimidating.)  Or would I be better off switching to something like Torch, which has the imperative differentiation Autograd, instead of symbolic differentiation?<\/p>\n\n<p>Thanks for taking the time to help me out on this.  I was trying to make this as concise as I could.<\/p>\n\n<p>EDIT:  After doing some further searching I came across <a href=\"https:\/\/stackoverflow.com\/questions\/32082506\/neural-network-reinforcement-learning-requiring-next-state-propagation-for-backp\">this previously asked question<\/a>.  It's a little different than mine (they are trying to avoid updating an LSTM network twice every iteration in Torch), and doesn't have any answers yet.<\/p>\n\n<p>Here is some code if that helps:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34544424","CreationDate":"2015-12-31T09:47:07.260","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>Right now what you want to do is very difficult in Tensorflow (0.6). Your best bet is to bite the bullet and call run multiple times at the cost of recomputing the activations. However, we are very aware of this issue internally. A prototype \"partial run\" solution is in the works, but there is no timeline for its completion right now. Since a truly satisfactory answer might require modifying tensorflow itself, you could also make a github issue for this and see if anyone else has anything to say on this there.<\/p>\n\n<p>Edit: Experimental support for partial_run is now in. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/client\/session.py#L317\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/client\/session.py#L317<\/a><\/p>\n"}]}
{"QuestionId":34542274,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><tensorflow>","CreationDate":"2015-12-31T06:40:20.017","AcceptedAnswerId":34552299.0,"OwnerUserId":829624.0,"Title":"Reading multiple bytes into a single value in TensorFlow","Body":"<p>\nI'm trying to read the label in a similar way as described in cifar10 example in TensorFlow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The problem is, if  is bigger than 1 (e.g.,2),  seems to become a tensor of two elements (each of which is 1-byte). I just want to represent the consecutive  of bytes into a single value. How do I do that?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34552299","CreationDate":"2015-12-31T23:06:52.050","ParentId":null,"OwnerUserId":"4332314","Title":null,"Body":"<p>Create a second decoder, decode int16 with it and take the first element as your label<\/p>\n\n<pre><code>shorts = tf.decode_raw(value, tf.int16)\nresult.label = tf.cast(shorts[0], tf.int32)\n<\/code><\/pre>\n\n<p>There's probably a better solution but it works.<\/p>\n"}]}
{"QuestionId":34546792,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><conv-neural-network>","CreationDate":"2015-12-31T12:52:36.710","AcceptedAnswerId":null,"OwnerUserId":4393339.0,"Title":"Convolution Layer output changes to 1. after passing it through a fully connected layer for every image","Body":"<p>I built a CNN using theano. The code for the Convolution and Hidden layers is:<\/p>\n\n<pre><\/pre>\n\n<p>My neural network has 5 layers. But for checking what was going wrong, I stripped off all the other hidden layers and just kept the one which is connected to the convolution layer. I observed that the output of the convolution layer was fine, but after passing it through the fully connected layer, it became:<\/p>\n\n<pre><\/pre>\n\n<p>here 2,5000 is the shape of the output matrix of the hidden layer. The network implementation is:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what might be causing the output of convolution layers to change to all 1.'s after passing through the hidden layer?<\/p>\n","answers":[{"AnswerId":"34547825","CreationDate":"2015-12-31T14:32:35.007","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>Your nonlinearity is tanh which has saturating values of -1 (corresponding to a very negative input) and +1 (corresponding to a very large positive input). Perhaps the activations of your layers are giving you very large values in magnitude? You can try global contrast normalization to your input (zero center and divide by the standard deviation) to better condition your input values. Also try switching to a ReLU activation or even linear activation to check what sort of activations values you are getting from the particular layer you are referencing.<\/p>\n"}]}
{"QuestionId":34547402,"AnswerCount":0,"Tags":"<matlab><caffe><matcaffe>","CreationDate":"2015-12-31T13:52:15.620","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Write caffe architecture in matlab","Body":"<p>Has anyone an idea how one could write caffe architectures using matlab?\nIn python there are caffe mods (eg <a href=\"http:\/\/apollocaffe.com\/\" rel=\"nofollow\">ApolloCaffe<\/a>) where you can get the  corresponding prototxt string by typing for example:<\/p>\n\n<pre><\/pre>\n\n<p>Like this one can easily build various caffe architectures without hardcoding them which has the danger of making copy-paste errors.<\/p>\n\n<p>Is there something similar for Matlab?  <\/p>\n","answers":[]}
{"QuestionId":34547795,"AnswerCount":2,"Tags":"<python><anaconda><macports><caffe>","CreationDate":"2015-12-31T14:30:13.710","AcceptedAnswerId":34557806.0,"OwnerUserId":1430898.0,"Title":"Macports caffe wrong python","Body":"<p>I installed caffe using Macports . However, Macports didn't use my anaconda python, which I would like to use for development. I can only  in Macports' .<\/p>\n\n<p>Is there a way to show anaconda python where to look or do I have to reinstall for anaconda python? Either way, I would be grateful for a hint how to do it.<\/p>\n","answers":[{"AnswerId":"34549989","CreationDate":"2015-12-31T17:57:44.183","ParentId":null,"OwnerUserId":"3257992","Title":null,"Body":"<p>Install pyvenv ... it is easy to do on a Mac - then you can use whatever you want to.<\/p>\n"},{"AnswerId":"34557806","CreationDate":"2016-01-01T16:22:52.330","ParentId":null,"OwnerUserId":"2127896","Title":null,"Body":"<p>MacPorts never installs modules for use with other Python versions than MacPorts' own version. As a consequence, there is no switch to select the Python version to build against.<\/p>\n\n<p>You'll have to install <code>caffe<\/code> for your Anaconda Python yourself, e.g. in a virtualenv, or use MacPorts Python.<\/p>\n"}]}
{"QuestionId":34549743,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2015-12-31T17:35:59.080","AcceptedAnswerId":37138706.0,"OwnerUserId":1714410.0,"Title":"Caffe: how to get the phase of a Python layer?","Body":"<p>I created a  layer  in caffe, and use it in the net  I insert the layer like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now, my layer only participates in the ing phase of the net,<br>\nhow can I know that in my layer's  function??<\/p>\n\n<pre><\/pre>\n\n<p>PS,<br>\nI posted this question on <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/uiZ2jy0QOdI\" rel=\"noreferrer\">\"Caffe Users\" google group<\/a> as well. I'll udpdate if anything pops there.<\/p>\n","answers":[{"AnswerId":"34588801","CreationDate":"2016-01-04T10:13:44.107","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>As pointed out by <a href=\"https:\/\/stackoverflow.com\/a\/37138706\/1714410\">galloguille<\/a>, caffe is now exposing <code>phase<\/code> to the python layer class. This new feature makes this answer a bit redundant. Still it is useful to know about the <code>param_str<\/code> in caffe python layer for passing other parameters to the layer.<\/p>\n\n<h3>Original answer:<\/h3>\n\n<p>AFAIK there is no trivial way of getting the phase. However, one can pass arbitrary parameters from the net prototxt to python. This can be done using the <code>param_str<\/code> parameters of the <code>python_param<\/code>.<br>\nHere's how it's done:<\/p>\n\n<pre><code>layer {\n  type: \"Python\"\n  ...\n  python_param {\n    ...\n    param_str: '{\"phase\":\"TRAIN\",\"numeric_arg\":5}' # passing params as a STRING\n<\/code><\/pre>\n\n<p>In python, you get <code>param_str<\/code> in the layer's <code>setup<\/code> function:<\/p>\n\n<pre><code>import caffe, json\nclass myLayer(caffe.Layer):\n  def setup(self, bottom, top):\n    param = json.loads( self.param_str ) # use JSON to convert string to dict\n    self.phase = param['phase']\n    self.other_param = int( param['numeric_arg'] ) # I might want to use this as well...\n<\/code><\/pre>\n"},{"AnswerId":"37138706","CreationDate":"2016-05-10T12:33:16.120","ParentId":null,"OwnerUserId":"1738214","Title":null,"Body":"<p>This is a very good workaround, but if you are only interested in passing the <code>phase<\/code> as a parameter, now you can access the phase as an attribute of the layer. This feature was merged just 6 days ago <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3995\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/pull\/3995<\/a>. <\/p>\n\n<p>Specific commit: <a href=\"https:\/\/github.com\/BVLC\/caffe\/commit\/de8ac32a02f3e324b0495f1729bff2446d402c2c\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/commit\/de8ac32a02f3e324b0495f1729bff2446d402c2c<\/a><\/p>\n\n<p>With this new feature you just need to use the attribute <code>self.phase<\/code>. For example you can do the following:<\/p>\n\n<pre><code>class PhaseLayer(caffe.Layer):\n\"\"\"A layer for checking attribute `phase`\"\"\"\n\ndef setup(self, bottom, top):\n    pass\n\ndef reshape(self, bootom, top):\n    top[0].reshape()\n\ndef forward(self, bottom, top):\n    top[0].data[()] = self.phase\n<\/code><\/pre>\n"}]}
{"QuestionId":34551942,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2015-12-31T22:01:13.183","AcceptedAnswerId":null,"OwnerUserId":290918.0,"Title":"How to write summaries for multiple runs in Tensorflow","Body":"<p>If you look at the Tensorboard <a href=\"https:\/\/www.tensorflow.org\/tensorboard\/cifar.html\">dashboard<\/a> for the cifar10 demo, it shows data for multiple runs. I am having trouble finding a good example showing how to set the graph up to output data in this fashion. I am currently doing something similar to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\">this<\/a>, but it seems to be combining data from runs and whenever a new run starts I see the warning on the console:<\/p>\n\n<blockquote>\n  <p>WARNING:root:Found more than one graph event per run.Overwritting the graph with the newest event<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34551943","CreationDate":"2015-12-31T22:01:13.183","ParentId":null,"OwnerUserId":"290918","Title":null,"Body":"\n\n<p>The solution turned out to be simple (and probably a bit obvious), but I'll answer anyway. The writer is instantiated like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>writer = tf.train.SummaryWriter(FLAGS.log_dir, sess.graph_def)\n<\/code><\/pre>\n\n<p>The events for the current run are written to the specified directory. Instead of having a fixed value for the <code>logdir<\/code> parameter, just set a variable that gets updated for each run and use that as the name of a sub-directory inside the log directory:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>writer = tf.train.SummaryWriter('%s\/%s' % (FLAGS.log_dir, run_var), sess.graph_def)\n<\/code><\/pre>\n\n<p>Then just specify the root <code>log_dir<\/code> location when starting tensorboard via the <code>--logdir<\/code> parameter.<\/p>\n"},{"AnswerId":"40690175","CreationDate":"2016-11-19T07:07:54.183","ParentId":null,"OwnerUserId":"808257","Title":null,"Body":"<p>As mentioned in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.11\/tensorflow\/tensorboard\/README.md#runs-comparing-different-executions-of-your-model\" rel=\"nofollow noreferrer\">documentation<\/a>, you can specify multiple log directories when running tensorboard. Alternatively, you can create multiple <code>run<\/code> subfolder in the log directory to visualize different plots in the same graph.<\/p>\n"}]}
{"QuestionId":34552584,"AnswerCount":0,"Tags":"<python><memory-management><memory-leaks><theano>","CreationDate":"2016-01-01T00:07:19.460","AcceptedAnswerId":null,"OwnerUserId":5635159.0,"Title":"Is there any chance I got memory leak with python or theano specifically?","Body":"<p>Happy new year to all of you!\nI trained a deep learning model with theano(newest develop version). After one or two epochs it consumes the memory of 24% and I ran four of them at the same time, so they consume the whole memory. And when I run three of them, they will consume 33% memory. In the trainning it is just a loop statement and I do not accumulate any thing in a list or dict or something. I paste the main loop as follow for it may helps. <\/p>\n\n<pre><\/pre>\n\n<p>I tried to use gc.collect it still didn't help. And I ran it on our old server which setuped a old version of theano (0.6.0), there is no that kind of problem.<\/p>\n\n<p>Is anyone got a clue? I would be quite appreciate it.<\/p>\n\n<p>I got a alias question here.(<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/vX6sPnGhRCg\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/vX6sPnGhRCg<\/a>)<\/p>\n","answers":[]}
{"QuestionId":34556645,"AnswerCount":2,"Tags":"<python><class><lua><torch>","CreationDate":"2016-01-01T13:46:57.347","AcceptedAnswerId":34558734.0,"OwnerUserId":4380945.0,"Title":"how to get the class type in lua \/ translation from python","Body":"<p>I am trying to find something analogous to the class in lua. In python, I would do:<\/p>\n\n<pre><\/pre>\n\n<p>So I have the object vocab in lua. When I print the object, I get:<\/p>\n\n<pre><\/pre>\n\n<p>how can I get lua to spit the object, something analogous to the type() in python? - which will give you the class of the object<\/p>\n","answers":[{"AnswerId":"34557440","CreationDate":"2016-01-01T15:30:58.110","ParentId":null,"OwnerUserId":"734069","Title":null,"Body":"<p>It is said that Lua defines \"mechanisms, not policies\". Classes would be a \"policy\". You can implement the effect of classes however you like; Lua provides many mechanisms that could be used to do so. But Lua has no single way of declaring such a construct.<\/p>\n\n<p>Lua's <code>type<\/code> standard method will only return the broad category of a Lua value: number, string, table, etc. And since Lua has only one kind of data structure (table), every kind of class (not generated from C) would be of the \"table\" type.<\/p>\n\n<p>This has the side effect of making it difficult in Lua to use someone else's policies. For example, if Torch has a way to distinguish its own \"classes\" from regular tables, that would work. But it wouldn't be able to distinguish some <em>other<\/em> class implementation from a regular table.<\/p>\n\n<p>So there's no general way to distinguish a table from some form of \"class\".<\/p>\n"},{"AnswerId":"34558734","CreationDate":"2016-01-01T18:19:12.597","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p><a href=\"http:\/\/www.lua.org\/pil\/2.html\" rel=\"noreferrer\">There are 8 types in <code>Lua<\/code><\/a>: nil, boolean, number, string, function, thread, <strong>table<\/strong> and <strong>userdata<\/strong>. You can find out which of these basic types your object belongs to using built-in <code>type()<\/code> function:<\/p>\n\n<pre><code>type('Hello world')                    == 'string'\ntype(3.14)                             == 'number'\ntype(true)                             == 'boolean'\ntype(nil)                              == 'nil'\ntype(print)                            == 'function'\ntype(coroutine.create(function() end)) == 'thread'\ntype({})                               == 'table'\ntype(torch.Tensor())                   == 'userdata'\n<\/code><\/pre>\n\n<p>Note that the type of <code>torch.Tensor<\/code> is userdata. That makes sense since <code>torch<\/code> library is written in C.<\/p>\n\n<blockquote>\n  <p>The type userdata is provided to allow arbitrary C data to be stored\n  in Lua variables. A userdata value is a pointer to a block of raw\n  memory. Userdata has no predefined operations in Lua, except\n  assignment and identity test.<\/p>\n  \n  <p>The metatable for the userdata is put in the registry, and the __index\n  field points to the table of methods so that the object:method()\n  syntax will work.<\/p>\n<\/blockquote>\n\n<p>So, dealing with a userdata object, we do not know what it is but have a list of its methods and can invoke them. <\/p>\n\n<p>It would be great if custom objects had some mechanism (a method or something) to see their custom types. And guess what? Torch objects have one:<\/p>\n\n<pre><code>t = torch.Tensor()\ntype(t)       == 'userdata' # Because the class was written in C\ntorch.type(t) == 'torch.DoubleTensor'\n# or\nt:type()      == 'torch.DoubleTensor'\n<\/code><\/pre>\n\n<p>Speaking of <code>Torch<\/code>. It has its own object system emulator, and you are free to create some torch classes yourself and check their types the same way. For <code>Lua<\/code>, however, such classes\/objects are nothing more than ordinary tables.<\/p>\n\n<pre><code>local A = torch.class('ClassA')\nfunction A:__init(val)\n    self.val = val\nend\n\nlocal B, parent = torch.class('ClassB', 'ClassA')\nfunction B:__init(val)\n    parent.__init(self, val)\nend\n\nb = ClassB(5)\ntype(b)       == 'table' # Because the class was written in Lua\ntorch.type(b) == 'ClassB'\nb:type() # exception; Custom Torch classes have no :type() method by defauld\n<\/code><\/pre>\n"}]}
{"QuestionId":34556932,"AnswerCount":1,"Tags":"<machine-learning><neural-network><tensorflow>","CreationDate":"2016-01-01T14:23:48.983","AcceptedAnswerId":34557466.0,"OwnerUserId":4341948.0,"Title":"XOR gate with a neural network","Body":"\n\n<p>I was trying to implement an XOR gate with tensorflow. I succeeded in implementing that, but i don't fully understand why it works. I got help from stackoverflow posts <a href=\"https:\/\/stackoverflow.com\/questions\/33747596\/problems-implementing-an-xor-gate-with-neural-nets-in-tensorflow\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/33997823\/tensorflow-mlp-not-training-xor\/34022791#34022791\">here<\/a>. So both with  and  outputs. Here is the network as i understood, in order to set things clear.<a href=\"https:\/\/i.stack.imgur.com\/YmD1V.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YmD1V.png\" alt=\"neural network visualization\"><\/a><\/p>\n\n<p><strong>My Question #1:<\/strong>\nNotice the  function and  function. Why we need that(specifically the  function)? You may say that in order to achieve non linearity.  I understand how  achieves non-linearity. I got the answer from <a href=\"https:\/\/stats.stackexchange.com\/questions\/141960\/deep-neural-nets-relus-removing-non-linearity\">here<\/a>. Now from what I understand the difference between using  and without using  is this(see the picture).[I tested the  function. The output is like this]<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/i1OhT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/i1OhT.png\" alt=\"RELU\"><\/a><\/p>\n\n<p>Now, if the first function works, why not the second function? From my perspective  achieves non-linearity by combining multiple linear functions. So both is linear function(upper two). If first one achieves non linearity, 2nd one should too, shouldn't it? The question is that, without using the  why the network gets stuck?<\/p>\n\n<p><strong>XOR gate with one hot true outputs<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The code you see above implements the XOR gate with one hot true outputs. If i take out , the network gets stuck. Why?<\/p>\n\n<p><strong>My Question #2:<\/strong>\nHow can I understand if a network is going to get stuck on some local minima[or some value]? Is it from the plot of cost function (or loss function)? Say, for the network designed above, I used cross entropy as the loss function. I could not find the plotting of cross entropy function. (If you can provide this, this would be very helpful.)<\/p>\n\n<p><strong>My Question #3:<\/strong>\nNotice on the code there is a line . It means that i have set the number of neurons in the hidden layer . Reducing the number of neurons to  makes the network to get stuck. So what should be the number of neurons on hidden layer? <\/p>\n\n<p>The output when the network works the way it is supposed to : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The output when the network gets stuck:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34557466","CreationDate":"2016-01-01T15:33:37.537","ParentId":null,"OwnerUserId":"1149632","Title":null,"Body":"<h2>Question 1<\/h2>\n\n<p>Both the ReLU and Sigmoid function is non-linear. On the contrary, the function drawn to the right of the ReLU function is linear. Applying multiple linear activation functions will still make the network linear. <\/p>\n\n<p>Therefore, the network gets stuck when trying to perform linear regression on a non-linear problem.<\/p>\n\n<h2>Question 2<\/h2>\n\n<p>Yes, you will have to pay attention to the progression of the error rate. In larger problem instances, you would typically pay attention to the development of the error function on your test set. This is done by measuring the accuracy of the network after a period of training.<\/p>\n\n<h2>Question 3<\/h2>\n\n<p>The XOR problem requires at least 2 input, 2 hidden, and 1 output node, that is: five nodes are required to correctly model the XOR problem with a simple neural network.<\/p>\n"}]}
{"QuestionId":34559990,"AnswerCount":1,"Tags":"<machine-learning><theano><rbm>","CreationDate":"2016-01-01T20:58:35.167","AcceptedAnswerId":null,"OwnerUserId":4659277.0,"Title":"Restricted Boltzman Machines for continious inputs","Body":"<p>There is an implementation for RBMs. Thr original RBM implementation is for the discrete data such as images, my data is a binary data, does the code work for real data too? I read somewhere that there is a gaussin version of the typical RBM that works for that, is it also implemented in that module?<\/p>\n","answers":[{"AnswerId":"41244269","CreationDate":"2016-12-20T14:08:54.097","ParentId":null,"OwnerUserId":"4500078","Title":null,"Body":"<p>In short, an RBM is simply a Markov random field on a bipartite graph. So therefore you can have any probability distribution to describe the relationship between nodes. <\/p>\n\n<p>In terms of code, you don't really need to copy things explicitly. The role that the probability function selected will play is in the contrastive divergence algorithm. You should only have to change the way that the samples are selected. The parts of the code that need to be changed are copied below.<\/p>\n\n<pre><code>def sample_h_given_v(self, v0_sample):\n    ''' This function infers state of hidden units given visible units '''\n    # compute the activation of the hidden units given a sample of\n    # the visibles\n    pre_sigmoid_h1, h1_mean = self.propup(v0_sample)\n    # get a sample of the hiddens given their activation\n    # Note that theano_rng.binomial returns a symbolic sample of dtype\n    # int64 by default. If we want to keep our computations in floatX\n    # for the GPU we need to specify to return the dtype floatX\n    h1_sample = self.theano_rng.binomial(size=h1_mean.shape,\n                                         n=1, p=h1_mean,\n                                         dtype=theano.config.floatX)\n    return [pre_sigmoid_h1, h1_mean, h1_sample]\n\ndef propdown(self, hid):\n    '''This function propagates the hidden units activation downwards to\n    the visible units\n\n    Note that we return also the pre_sigmoid_activation of the\n    layer. As it will turn out later, due to how Theano deals with\n    optimizations, this symbolic variable will be needed to write\n    down a more stable computational graph (see details in the\n    reconstruction cost function)\n\n    '''\n    pre_sigmoid_activation = T.dot(hid, self.W.T) + self.vbias\n    return [pre_sigmoid_activation, T.nnet.sigmoid(pre_sigmoid_activation)]\n\ndef sample_v_given_h(self, h0_sample):\n    ''' This function infers state of visible units given hidden units '''\n    # compute the activation of the visible given the hidden sample\n    pre_sigmoid_v1, v1_mean = self.propdown(h0_sample)\n    # get a sample of the visible given their activation\n    # Note that theano_rng.binomial returns a symbolic sample of dtype\n    # int64 by default. If we want to keep our computations in floatX\n    # for the GPU we need to specify to return the dtype floatX\n    v1_sample = self.theano_rng.binomial(size=v1_mean.shape,\n                                         n=1, p=v1_mean,\n                                         dtype=theano.config.floatX)\n    return [pre_sigmoid_v1, v1_mean, v1_sample]\n<\/code><\/pre>\n"}]}
{"QuestionId":34562115,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-02T02:48:25.030","AcceptedAnswerId":34564457.0,"OwnerUserId":5737155.0,"Title":"How to combine two tensor in theano","Body":"<p>I am wondering how is it possible to combine two variables, similar to append in python? For example, we have two variables (after feeding with data):<\/p>\n\n<p>x: with size 1*3\ny: with size 1*3<\/p>\n\n<p>now I want to have a variable, which combine x and y to a size of 1*3*2<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"34564457","CreationDate":"2016-01-02T09:35:24.473","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>One can use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.stack\" rel=\"noreferrer\">theano.tensor.stack<\/a> to achieve this. Here's a working example:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\nx = tt.matrix()\ny = tt.matrix()\nz = tt.stack([x, y], axis=2)\nf = theano.function([x, y], z)\nprint f([[1, 2, 3]], [[4, 5, 6]])\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>[[[ 1.  4.]\n  [ 2.  5.]\n  [ 3.  6.]]]\n<\/code><\/pre>\n"}]}
{"QuestionId":34562557,"AnswerCount":1,"Tags":"<google-app-engine><google-cloud-platform><tensorflow><gae-module>","CreationDate":"2016-01-02T04:18:55.193","AcceptedAnswerId":34563293.0,"OwnerUserId":1464495.0,"Title":"Can I use TensorFlow in a Google App Engine module?","Body":"<p>I have a data processing module as part of my GAE application. TensorFlow has separate installers for Linux and OSX, which indicates to me that it might not play nicely with the GAE sandbox. Is it possible to use TensorFlow in a GAE module?<\/p>\n","answers":[{"AnswerId":"34563293","CreationDate":"2016-01-02T06:32:10.393","ParentId":null,"OwnerUserId":"196964","Title":null,"Body":"<p>No, you won't be able to install TensorFlow on standard instances.<\/p>\n\n<p>You can, however, install whatever you like on <a href=\"https:\/\/cloud.google.com\/appengine\/docs\/managed-vms\/\" rel=\"noreferrer\">managed VMs<\/a> instances. Though it may be easier to use compute engine instances instead.<\/p>\n"}]}
{"QuestionId":34563976,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-02T08:19:34.247","AcceptedAnswerId":34568455.0,"OwnerUserId":4242336.0,"Title":"How to use the trained algorithm for tensorflow","Body":"<p>I have installed the tensorflow and follow the tutorial here<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/tf\/index.html#tensorflow-mechanics-101\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/tf\/index.html#tensorflow-mechanics-101<\/a><\/p>\n\n<p>and build it successfully, I can get the evaluation result for the same size dataset, like 1000X784 for training set, and 1000X784 for testing set.<\/p>\n\n<p>but what if i want to test one data, 1X784, and find out what's the output, using the algorithm trained above.<\/p>\n\n<p>I am now to tensorflow, and new to Machine Learning, I hope that I have described my self.<\/p>\n","answers":[{"AnswerId":"34568455","CreationDate":"2016-01-02T17:11:32.730","ParentId":null,"OwnerUserId":"2251083","Title":null,"Body":"<p>It's not clear to me which part you're having trouble with, but I think what you're asking is how to use batch size 1000 for training, but only predict on a single input. I assume you already know how to predict on batches of size 1000.<\/p>\n\n<p>If the first dimension of your model's input placeholder, which is usually the batch size, is set to be None, the size is inferred when you provide an input. So, if you change the 1000 to be None, you should then be able to pass an input of size 1 by 784 to make predictions. <\/p>\n"},{"AnswerId":"37110131","CreationDate":"2016-05-09T07:38:04.740","ParentId":null,"OwnerUserId":"6171593","Title":null,"Body":"<p>The solution that you found to feed a 1*784 is a great solution to just get a quick feed back , however in bigger networks which they need a lot of time (around hours) for training your solution is not feasible.<\/p>\n\n<p>Tensorflow they have a new feature it's name is Tensorflow serving which you give it a train model then you  interact with your model as a client.<\/p>\n\n<p>Here is their website for more information : <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/serving<\/a><\/p>\n"}]}
{"QuestionId":34564430,"AnswerCount":1,"Tags":"<caffe><mnist>","CreationDate":"2016-01-02T09:31:43.120","AcceptedAnswerId":null,"OwnerUserId":4968861.0,"Title":"Why Caffe's Accuracy layer's bottoms consist of InnerProduct and Label?","Body":"<p>I'm new to <strong>caffe<\/strong>, in the <em>MNIST<\/em> example, I was thought that label should compared with softmax layers, but it was not the situation in .\nI wonder why use  result and label to get the accuracy, it seems unreasonable. Was it because I missed something in the layers?<\/p>\n","answers":[{"AnswerId":"34565976","CreationDate":"2016-01-02T12:40:35.610","ParentId":null,"OwnerUserId":"2693257","Title":null,"Body":"<p>The output dimension of the last inner product layer is 10, which corresponds to the number of classes (digits 0~9) of your problem.\nThe loss layer, takes two blobs, the first one being the prediction(ip2) and the second one being the label provided by the data layer.<\/p>\n\n<pre><code>loss_layer = SoftmaxLossLayer(name=\"loss\", bottoms=[:ip2,:label])\n<\/code><\/pre>\n\n<p>It does not produce any outputs - all it does is to compute the loss function value, report it when backpropagation starts, and initiates the gradient with respect to ip2. This is where all magic starts.\nAfter training ( in TEST phase), In the last layer desire results come from multiply weights and ip1( that are computed in last layer ); And each of class( one of 10 neurons) has max value is choosen.<\/p>\n"}]}
{"QuestionId":34565479,"AnswerCount":2,"Tags":"<python-2.7><machine-learning><classification><theano><supervised-learning>","CreationDate":"2016-01-02T11:42:09.773","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"Can i give a tuple or list to a givens in the theano","Body":"<p>In theano, my funciton is showed below:<\/p>\n\n<pre><\/pre>\n\n<p>but it doesn't work. (the x in the givens is a tuple,is it right?)\nThanks<\/p>\n","answers":[{"AnswerId":"34565809","CreationDate":"2016-01-02T12:20:49.907","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The <code>x<\/code> value should be a symbolic expression that is of the same dimensionality as <code>x<\/code>'s dimensionality. For example, if <code>x<\/code> is a matrix then the symbolic expression here should also describe a matrix.<\/p>\n"},{"AnswerId":"34573331","CreationDate":"2016-01-03T04:29:31.100","ParentId":null,"OwnerUserId":"5737155","Title":null,"Body":"<p>I solved this problem by another way, instead of assigning a tuple, i add the parameters,i.e. there are x1 and x2 in the function.<\/p>\n"}]}
{"QuestionId":34567260,"AnswerCount":0,"Tags":"<python><machine-learning><theano>","CreationDate":"2016-01-02T14:58:24.377","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"cannot convert Type TensorType(float32, matrix) (of Variable Subtensor{::,int64:int64:}.0) into Type TensorType(float32, 3D)","Body":"<p>I have looked up the problem on the internet and haven't been able to find the solution for. It is said that the y and dataset_y don't match, but when I print the y.type and train_y.type, both of them are: TensorType(int32,vector). The code I've written is as follow:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34567454,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><artificial-intelligence><theano>","CreationDate":"2016-01-02T15:19:59.623","AcceptedAnswerId":34568142.0,"OwnerUserId":497600.0,"Title":"Theano: How to give training data to a neural network","Body":"<p>I'm trying to create a simple multilayer Perceptron (MLP) for \"logical and\" in Theano.\nThere is one layer between input and output. The structure is this:<\/p>\n\n<p>2 value input -> multiply with weights, add bias -> softmax -> 1 value output<\/p>\n\n<p>The change in dimension is caused by the weights matrix.<\/p>\n\n<p>The implementation is based on this tutorial: <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a><\/p>\n\n<p>This is my class for the Layer:<\/p>\n\n<pre><\/pre>\n\n<p>The class is meant to be modular. Instead of just one layer, I want to be able to add multiple layers.\nTherefore the functions for prediction, cost and errors are outside of the class (as opposed to the tutorial):<\/p>\n\n<pre><\/pre>\n\n<p>The logical and has 4 trainingcases:<\/p>\n\n<ul>\n<li>[0,0] -> 0<\/li>\n<li>[1,0] -> 0<\/li>\n<li>[0,1] -> 0<\/li>\n<li>[1,1] -> 1<\/li>\n<\/ul>\n\n<p>Here are the setup of the classifier and the functions for training and evaluating:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to follow the conventions. Each row in the data matrix is a training sample. Each training sample is matched to the correct output.\nUnfortunately the code breaks. I can't interpret the error message. What did I do wrong ?\nError:<\/p>\n\n<blockquote>\n  <p>TypeError: Cannot convert Type TensorType(int32, scalar) (of Variable Subtensor{int64}.0) into Type TensorType(int32, vector). You can try to manually convert Subtensor{int64}.0 into a TensorType(int32, vector).<\/p>\n<\/blockquote>\n\n<p>This error occurs deep in the Theano code. The conflicting line in my program is:<\/p>\n\n<pre><\/pre>\n\n<p>Apparently there is a mismatch between the dimensions of y and the training data.\nMy complete code on pastebin: <a href=\"http:\/\/pastebin.com\/U5jYitk2\" rel=\"nofollow\">http:\/\/pastebin.com\/U5jYitk2<\/a>\nThe complete error message on pastebin: <a href=\"http:\/\/pastebin.com\/hUQJhfNM\" rel=\"nofollow\">http:\/\/pastebin.com\/hUQJhfNM<\/a><\/p>\n\n<p><strong>Concise question:<\/strong>\nWhat is the correct way to give training data to an mlp in theano ?\nWhere is my mistake ?<\/p>\n\n<p>I copied most of the code of the tutorial. Notable changes (probable causes of the error) are:<\/p>\n\n<ul>\n<li>training data for y is not a matrix. I think this is right, because the output of my network is just a scalar value<\/li>\n<li>input of the first layer is a vector. This variable is named x.<\/li>\n<li>Access of the training data does not use slicing. In the tutorial the training data is very complex and I find it hard to read the data access code. I believe that x should be a row of the data-matrix. This is how I implemented it.<\/li>\n<\/ul>\n\n<p><strong>UPDATE:<\/strong>\nI used the code of Amir. Looks very good, thank you.<\/p>\n\n<p>But it creates an error, too. The last loop is out-of-bounds:<\/p>\n\n<blockquote>\n  <p>\/usr\/bin\/python3.4 \/home\/lhk\/programming\/sk\/mlp\/mlp\/Layer.py Traceback\n  (most recent call last):   File\n  \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 595, in <strong>call<\/strong>\n      outputs = self.fn() ValueError: y_i value out of bounds<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>Traceback (most recent call last):   File\n  \"\/home\/lhk\/programming\/sk\/mlp\/mlp\/Layer.py\", line 113, in \n      train_model(i)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 606, in <strong>call<\/strong>\n      storage_map=self.fn.storage_map)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/gof\/link.py\", line 206,\n  in raise_with_op\n      raise exc_type(exc_value).with_traceback(exc_trace)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 595, in <strong>call<\/strong>\n      outputs = self.fn() ValueError: y_i value out of bounds Apply node that caused the error: CrossentropySoftmaxArgmax1HotWithBias(Dot22.0,\n  b, Elemwise{Cast{int32}}.0) Inputs types: [TensorType(float64,\n  matrix), TensorType(float64, vector), TensorType(int32, vector)]\n  Inputs shapes: [(1, 1), (1,), (1,)] Inputs strides: [(8, 8), (8,),\n  (4,)] Inputs values: [array([[ 0.]]), array([ 0.]), array([1],\n  dtype=int32)]<\/p>\n  \n  <p>HINT: Re-running with most Theano optimization disabled could give you\n  a back-trace of when this node was created. This can be done with by\n  setting the Theano flag 'optimizer=fast_compile'. If that does not\n  work, Theano optimizations can be disabled with 'optimizer=None'.\n  HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint\n  and storage map footprint of this apply node.<\/p>\n<\/blockquote>\n\n<p>Line 113 is this one:<\/p>\n\n<pre><\/pre>\n\n<p>I believe this is because the indexing of the training data uses  and . Why is that necessary ? One row should be one training sample. And one row is <\/p>\n\n<p>Edit: I debugged the code. Without slicing it returns a 1d-array, with slicing it's 2d. 1d should be incompatible to the matrix x.<\/p>\n\n<p>But while I did this, I found another strange problem:\nI added this code to look at the effect of the training:<\/p>\n\n<pre><\/pre>\n\n<p>This makes sense, since the first three samples have 0 as correct output.\nIf I change the order and move the training sample (1,1),1 to the front, the program crashes.<\/p>\n\n<blockquote>\n  <p>before [[ 0.]  [ 0.]] [ 0.] Traceback (most recent call last):   File\n  \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 595, in <strong>call<\/strong>\n      outputs = self.fn() ValueError: y_i value out of bounds<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>Traceback (most recent call last):   File\n  \"\/home\/lhk\/programming\/sk\/mlp\/mlp\/Layer.py\", line 121, in \n      train_model(i)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 606, in <strong>call<\/strong>\n      storage_map=self.fn.storage_map)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/gof\/link.py\", line 206,\n  in raise_with_op\n      raise exc_type(exc_value).with_traceback(exc_trace)   File \"\/usr\/local\/lib\/python3.4\/dist-packages\/theano\/compile\/function_module.py\",\n  line 595, in <strong>call<\/strong>\n      outputs = self.fn() ValueError: y_i value out of bounds Apply node that caused the error: CrossentropySoftmaxArgmax1HotWithBias(Dot22.0,\n  b, Elemwise{Cast{int32}}.0) Inputs types: [TensorType(float64,\n  matrix), TensorType(float64, vector), TensorType(int32, vector)]\n  Inputs shapes: [(1, 1), (1,), (1,)] Inputs strides: [(8, 8), (8,),\n  (4,)] Inputs values: [array([[ 0.]]), array([ 0.]), array([1],\n  dtype=int32)]<\/p>\n  \n  <p>HINT: Re-running with most Theano optimization disabled could give you\n  a back-trace of when this node was created. This can be done with by\n  setting the Theano flag 'optimizer=fast_compile'. If that does not\n  work, Theano optimizations can be disabled with 'optimizer=None'.\n  HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint\n  and storage map footprint of this apply node.<\/p>\n<\/blockquote>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I installed Python2.7 with Theano and tried to run the code again. The same error occurs. And I added verbose exception handling. Here is the output:<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>I looked at the training data again. Any sample with 1 as label will produce the above error.<\/p>\n\n<pre><\/pre>\n\n<p>The above sample labels will crash for every train_model(i) for i in (0,1,2,3).\nApparently there is an interference between indexing the samples and the sample content.<\/p>\n\n<p><strong>UPDATE:<\/strong>\nThe problem is indeed, like Amir's contact indicated, the dimension of the output layer. I had the misconception that I could train the network to encode the output of the function \"logical and\" directly in the output neuron. While this is certainly possible, this training approach uses the y value indexing to choose the output node which should have the highest value. After changing the output size to two, the code works. And with enough training, the errors for all cases indeed become zero.<\/p>\n","answers":[{"AnswerId":"40622249","CreationDate":"2016-11-16T00:45:26.440","ParentId":null,"OwnerUserId":"956730","Title":null,"Body":"<p>You can try my MLP class:<\/p>\n\n<p>Lasagne \/ Theano based MultiLayer Perceptron MLP which accepts both sparse and dense input matrices and is very easy to use with scikit-learn api similarity.<\/p>\n\n<p>It has drop-out configurable \/ sparse input \/ can be changed to logistic regression, easy to change the cost function and l1\/l2\/elasticnet regularisations.<\/p>\n\n<p>The code is <a href=\"https:\/\/github.com\/afshinrahimi\/sparsemultilayerperceptron\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"34568142","CreationDate":"2016-01-02T16:37:13.653","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Here is the working code for your problem. There were quite a few little bugs in your code. The one causing the error you were getting was due to defining <code>b<\/code> as a <code>n_in<\/code> by <code>n_out<\/code> matrix instead of simply defining it as an 'n_out' vector. The update part was defined in brackets <code>[]<\/code> as opposed to parenthesis <code>()<\/code>. <\/p>\n\n<p>Also, the index was defined as an <code>int32<\/code> symbolic scalar (this is not very important). The other import change was to define the functions given the correct indexing. The way you had used the <code>index<\/code> to compile your functions would not let the function to compile for some reason. You had also declared your input as a vector. This way, you will not be able to train the model using mini-batches or full batch. So it's safe to declare it as a symbolic matrix. And to use a vector, you need to store your inputs as vectors instead of a matrix on the shared variable to make the program run. So, there will be such a headache declaring it as a vector. In the end, you had use <code>classifier.errors(y)<\/code> to compile your validation function although you had removed the function <code>errors<\/code> from the <code>Layer<\/code> class.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy\n\n\nclass Layer(object):\n    \"\"\"\n    this is a layer in the mlp\n    it's not meant to predict the outcome hence it does not compute a loss.\n    apply the functions for negative log likelihood = cost on the output of the last layer\n    \"\"\"\n\n    def __init__(self, input, n_in, n_out):\n        self.x = input\n        self.W = theano.shared(\n                value=numpy.zeros(\n                        (n_in, n_out),\n                        dtype=theano.config.floatX\n                ),\n                name=\"W\",\n                borrow=True\n        )\n        self.b = theano.shared(\n                value=numpy.zeros(n_out,\n                                  dtype=theano.config.floatX),\n                name=\"b\",\n                borrow=True\n        )\n\n        self.output = T.nnet.softmax(T.dot(self.x, self.W) + self.b)\n        self.params = [self.W, self.b]\n        self.input = input\n\n\ndef y_pred(output):\n    return T.argmax(output, axis=1)\n\n\ndef negative_log_likelihood(output, y):\n    return -T.mean(T.log(output)[T.arange(y.shape[0]), y])\n\n\ndef errors(output, y):\n    # check if y has same dimension of y_pred\n    if y.ndim != y_pred(output).ndim:\n        raise TypeError(\n                'y should have the same shape as self.y_pred',\n                ('y', y.type, 'y_pred', y_pred(output).type)\n        )\n    # check if y is of the correct datatype\n    if y.dtype.startswith('int'):\n        # the T.neq operator returns a vector of 0s and 1s, where 1\n        # represents a mistake in prediction\n        return T.mean(T.neq(y_pred(output), y))\n    else:\n        raise NotImplementedError()\n\ndata_x = numpy.matrix([[0, 0],\n                       [1, 0],\n                       [0, 1],\n                       [1, 1]])\n\ndata_y = numpy.array([0,\n                      0,\n                      0,\n                      1])\n\ntrain_set_x = theano.shared(numpy.asarray(data_x,\n                         dtype=theano.config.floatX),\n                         borrow=True)\n\ntrain_set_y = T.cast(theano.shared(numpy.asarray(data_y,\n                         dtype=theano.config.floatX),\n                         borrow=True),\"int32\")\n\nx = T.matrix(\"x\")  # data\ny = T.ivector(\"y\")  # labels\n\nclassifier = Layer(input=x, n_in=2, n_out=1)\n\ncost = negative_log_likelihood(classifier.output, y)\n\ng_W = T.grad(cost=cost, wrt=classifier.W)\ng_b = T.grad(cost=cost, wrt=classifier.b)\nindex = T.iscalar()\n\nlearning_rate = 0.15\n\nupdates = (\n    (classifier.W, classifier.W - learning_rate * g_W),\n    (classifier.b, classifier.b - learning_rate * g_b)\n)\n\ntrain_model = theano.function(\n        inputs=[index],\n        outputs=cost,\n        updates=updates,\n        givens={\n            x: train_set_x[index:index + 1],\n            y: train_set_y[index:index + 1]\n        }\n)\nvalidate_model = theano.function(\n        inputs=[index],\n        outputs=errors(classifier.output, y),\n        givens={\n            x: train_set_x[index:index + 1],\n            y: train_set_y[index:index + 1]\n        }\n)\n\n#train the model\nfor i in range(train_set_x.shape[0].eval()):\n    train_model(i)\n<\/code><\/pre>\n\n<p><strong>Here's the updated code<\/strong>. Please note that the main difference between the code above and the code below is that the later is suitable for a binary problem while the other only works if you have a multi-class problem, which is not the case here. The reason I putting both code snippets here is for educational purposes. Please read the comments to get a sense of the problem with the code above and how I went about resolving it.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy\n\n\nclass Layer(object):\n    \"\"\"\n    this is a layer in the mlp\n    it's not meant to predict the outcome hence it does not compute a loss.\n    apply the functions for negative log likelihood = cost on the output of the last layer\n    \"\"\"\n\n    def __init__(self, input, n_in, n_out):\n        self.x = input\n        self.W = theano.shared(\n                value=numpy.zeros(\n                        (n_in, n_out),\n                        dtype=theano.config.floatX\n                ),\n                name=\"W\",\n                borrow=True\n        )\n        self.b = theano.shared(\n                value=numpy.zeros(n_out,\n                                  dtype=theano.config.floatX),\n                name=\"b\",\n                borrow=True\n        )\n\n        self.output = T.reshape(T.nnet.sigmoid(T.dot(self.x, self.W) + self.b), (input.shape[0],))\n        self.params = [self.W, self.b]\n        self.input = input\n\n\ndef y_pred(output):\n    return output\n\n\ndef negative_log_likelihood(output, y):\n    return T.mean(T.nnet.binary_crossentropy(output,y))\n\n\ndef errors(output, y):\n    # check if y has same dimension of y_pred\n    if y.ndim != y_pred(output).ndim:\n        raise TypeError(\n                'y should have the same shape as self.y_pred',\n                ('y', y.type, 'y_pred', y_pred(output).type)\n        )\n    # check if y is of the correct datatype\n    if y.dtype.startswith('int'):\n        # the T.neq operator returns a vector of 0s and 1s, where 1\n        # represents a mistake in prediction\n        return T.mean(T.neq(y_pred(output), y))\n    else:\n        raise NotImplementedError()\n\ndata_x = numpy.matrix([[0, 0],\n                       [1, 0],\n                       [0, 1],\n                       [1, 1]])\n\ndata_y = numpy.array([0,\n                      0,\n                      0,\n                      1])\n\ntrain_set_x = theano.shared(numpy.asarray(data_x,\n                         dtype=theano.config.floatX),\n                         borrow=True)\n\ntrain_set_y = T.cast(theano.shared(numpy.asarray(data_y,\n                         dtype=theano.config.floatX),\n                         borrow=True),\"int32\")\n\nx = T.matrix(\"x\")  # data\ny = T.ivector(\"y\")  # labels\n\nclassifier = Layer(input=x, n_in=2, n_out=1)\n\ncost = negative_log_likelihood(classifier.output, y)\n\ng_W = T.grad(cost=cost, wrt=classifier.W)\ng_b = T.grad(cost=cost, wrt=classifier.b)\nindex = T.iscalar()\n\nlearning_rate = 0.15\n\nupdates = (\n    (classifier.W, classifier.W - learning_rate * g_W),\n    (classifier.b, classifier.b - learning_rate * g_b)\n)\n\ntrain_model = theano.function(\n        inputs=[index],\n        outputs=cost,\n        updates=updates,\n        givens={\n            x: train_set_x[index:index+1],\n            y: train_set_y[index:index+1]\n        }\n)\nvalidate_model = theano.function(\n        inputs=[index],\n        outputs=errors(classifier.output, y),\n        givens={\n            x: train_set_x[index:index + 1],\n            y: train_set_y[index:index + 1]\n        }\n)\n\n#train the model\nfor i in range(train_set_x.shape[0].eval()):\n    train_model(i)\n<\/code><\/pre>\n"}]}
{"QuestionId":34568675,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-02T17:31:37.407","AcceptedAnswerId":34569074.0,"OwnerUserId":1716733.0,"Title":"Parsing `summary_str` byte string evaluated on tensorflow summary object","Body":"<p>Currently tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/470\" rel=\"noreferrer\">tensorboard is not compatible with python3<\/a>. Therefore and generally, I am looking for a way to print out the summary readouts once in 100 epochs. <\/p>\n\n<p>Is there a function to parse the   byte string produced in the following lines into a dictionary of floats?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34569074","CreationDate":"2016-01-02T18:11:35.487","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>You can get a textual representation of <code>summary_str<\/code> by parsing it into a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/summary.proto\" rel=\"nofollow noreferrer\"><code>tf.Summary<\/code><\/a> protocol buffer as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>summary_proto = tf.Summary()\nsummary_proto.ParseFromString(summary_str)\nprint(summary_proto)\n<\/code><\/pre>\n\n<p>You can then convert it into a dictionary mapping string tags to floats:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>summaries = {}\nfor val in summary_proto.value:\n    # Assuming all summaries are scalars.\n    summaries[val.tag] = val.simple_value\n<\/code><\/pre>\n"}]}
{"QuestionId":34569804,"AnswerCount":1,"Tags":"<matlab><neural-network><regression><caffe><conv-neural-network>","CreationDate":"2016-01-02T19:27:04.883","AcceptedAnswerId":34580263.0,"OwnerUserId":3698878.0,"Title":"Single Label Regression(Finetuning) With Input Artificial Features In Caffe","Body":"<p>I have,say,  images and for each of them I have additional 2 artificial(made-up) features, and image-labels are single dimensional integers. <\/p>\n\n<p>I want to fine-tune Image-net on my dataset, but I do not know how to handle these 2 additional features as input, how should I feed the data to caffe? Please help!<\/p>\n\n<p>EDIT: The 2 features can be any 2 numbers (1 dimensional) say two numbers representing what class an image falls into, and how many images fall into that class.<\/p>\n\n<p>Say, I have 'cat.jpg', then the features are say, 5 and 2000, where 5 is the feature 1 representing the class and 2000 is the total images in that class. \nIn short, the 2 features can be any two integers.<\/p>\n","answers":[{"AnswerId":"34580263","CreationDate":"2016-01-03T18:51:02.050","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I think the most straight forward way for you is to use <code>\"HDF5Data\"<\/code> input layer, where you can store both the input images, the additional two \"features\" and the expected output value (for regression).<\/p>\n\n<p>You can see an example <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">here<\/a> for creating HDF5 data in python. A Matlab example can be found <a href=\"https:\/\/stackoverflow.com\/a\/34261942\/1714410\">here<\/a>. <\/p>\n\n<p>Your HDF5 should have 4 \"datasets\": one is the input images (or the image descriptors of dim 4096). <code>n<\/code> dimensional array of images\/descriptors.<br>\nAnother dataset is <code>\"feat_1\"<\/code> an <code>n<\/code> by 1 array, and <code>\"feat_2\"<\/code> and <code>n<\/code> by 1 array.<br>\nFinally you should have another input <code>\"target\"<\/code> an <code>n<\/code> by 1 array of the expected output you wish to learn.<\/p>\n\n<p>Once you have an HDF5 file ready with these datasets in it, you should have<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  top: \"data\" # name of dataset with images\/imagenet features\n  top: \"feat_1\"\n  top: \"feat_2\"\n  top: \"target\"\n  hdf5_data_param {\n    source: \"\/path\/to\/list\/file.txt\"\n  }\n}\n<\/code><\/pre>\n\n<p>As you can see a <em>single<\/em> <code>\"HDF5Data\"<\/code> layer can produce several \"top\"s.<\/p>\n"}]}
{"QuestionId":34570177,"AnswerCount":2,"Tags":"<python><list><machine-learning><tensorflow>","CreationDate":"2016-01-02T20:06:00.607","AcceptedAnswerId":34570503.0,"OwnerUserId":5738942.0,"Title":"Tensorflow: List of Tensors for Cost","Body":"<p>I am trying to work with LSTMs in Tensor Flow. I found a tutorial online where a set of sequences is taken in and the objective function is composed of the last output of the LSTM and the known values. However, I would like to have my objective function use information from each output. Specifically, I am trying to have the LSTM learn the set of sequences (i.e. learn all the letters in words in a sentence).:<\/p>\n\n<pre><\/pre>\n\n<p>Doing this results in the error:<\/p>\n\n<pre><\/pre>\n\n<p>How should I fix this issue? Does the  take in a list of tensor values, or is there some special tensor object that takes them?<\/p>\n","answers":[{"AnswerId":"34570503","CreationDate":"2016-01-02T20:38:51.767","ParentId":null,"OwnerUserId":"2251083","Title":null,"Body":"<p>In your code, <code>losses<\/code> is a Python list. TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#reduce_mean\" rel=\"nofollow\"><code>reduce_mean()<\/code><\/a> expects a single tensor, not a Python list.<\/p>\n\n<pre><code>losses = tf.reshape(tf.concat(1, losses), [-1, size])\n<\/code><\/pre>\n\n<p>where size is the number of values you're taking a softmax over should do what you want. See <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#concat\" rel=\"nofollow\">concat()<\/a><\/p>\n\n<p>But, one thing I notice in your code that seems a bit odd, is that you have a list of placeholders for your inputs, whereas the code in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">the TensorFlow Tutorial<\/a> uses an order 3 tensor for inputs. Your input is a list of order 2 tensors. I recommend looking over the code in the tutorial, because it does almost exactly what you're asking about. <\/p>\n\n<p>One of the main files in that tutorial is <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow\" title=\"here\">here<\/a>. In particular, line 139 is where they create their cost. \nRegarding your input, lines 90 and 91 are where the input and target placeholders are setup. The main takeaway in those 2 lines is that an entire sequence is passed in in a single placeholder rather than with a list of placeholders.<\/p>\n\n<p>See line 120 in the ptb_word_lm.py file to see where they do their concatenation.<\/p>\n"},{"AnswerId":"37303795","CreationDate":"2016-05-18T15:17:38.993","ParentId":null,"OwnerUserId":"4305018","Title":null,"Body":"<p>Working example, check <a href=\"https:\/\/github.com\/ragulpr\/jupyter-notebooks\/blob\/master\/lstm_as_list_of_steps.ipynb\" rel=\"nofollow noreferrer\">notebook<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.models.rnn import rnn, rnn_cell\nprint(tf.__version__) \n#&gt; 0.8.0\n\nbatch_size  = 2\noutput_size = input_size  = 2\nseq_len     = 10\nnum_units   = 2\n\ncell = rnn_cell.BasicLSTMCell(num_units)\ninputs = [tf.placeholder(tf.float32, shape=[batch_size,input_size ]) for _ in xrange(seq_len)]\nresult = [tf.placeholder(tf.float32, shape=[batch_size,output_size]) for _ in xrange(seq_len)]\n\nW_o = tf.Variable(tf.random_normal([num_units,input_size], stddev=0.01))     \nb_o = tf.Variable(tf.random_normal([input_size],           stddev=0.01))\n\noutputs, states = rnn.rnn(cell, inputs, dtype=tf.float32)   \n\nlosses = []\n\nfor i in xrange(seq_len):\n    final_transformed_val = tf.matmul(outputs[i],W_o) + b_o\n    losses.append(tf.squared_difference(result[i],final_transformed_val)) \n\nlosses = tf.reshape(tf.concat(1, losses), [-1, seq_len])\ncost = tf.reduce_mean(losses) \n<\/code><\/pre>\n\n<p>To see this in action, you can feed the graph in a hacky way:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\n\nstep = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost)\nsess = tf.InteractiveSession()\n\nsess.run(tf.initialize_all_variables())\n\ncosts = []\n\n# EXAMPLE\n#  Learn cumsum over each sequence in x\n# | t        | 0 | 1 | 2 | 3 | 4  | ...|\n# |----------|---|---|---|---|----|----|\n# | x[:,0,0] | 1 | 1 | 1 | 1 | 1  | ...|\n# | x[:,0,1] | 1 | 1 | 1 | 1 | 1  | ...|\n# |          |   |   |   |   |    |    |\n# | y[:,0,0] | 1 | 2 | 3 | 4 | 5  | ...|\n# | y[:,0,1] | 1 | 2 | 3 | 4 | 5  | ...|\n\nn_iterations = 300\nfor _ in xrange(n_iterations):\n    x  = np.random.uniform(0,1,[seq_len,batch_size,input_size])\n    y  = np.cumsum(x,axis=0)\n\n    x_list = {key: value for (key, value) in zip(inputs, x)}\n    y_list = {key: value for (key, value) in zip(result, y)}\n\n    err,_ = sess.run([cost, step], feed_dict=dict(x_list.items()+y_list.items()))\n    costs.append(err)\n\nplt.plot(costs)\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pFbw7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pFbw7.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As tensorflow-beginner I have yet to find a unified way\/best practices way of handling RNN's but as mentioned above I'm sure this isn't recommended. Liked your script as a very nice intro, thanks for the snippets. Also, there's things going on w.r.g to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn.py\" rel=\"nofollow noreferrer\">implementation of scan and RNN-tuple-friendliness<\/a> right now so be careful<\/p>\n"}]}
{"QuestionId":34570614,"AnswerCount":2,"Tags":"<python><machine-learning><computer-vision><tensorflow><conv-neural-network>","CreationDate":"2016-01-02T20:52:59.603","AcceptedAnswerId":null,"OwnerUserId":4165403.0,"Title":"Face Recognition system in Python Tensorflow","Body":"<p>So I decided to go further on the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"nofollow\">MNIST tutorial<\/a> in Google's Tensorflow and try to create a rudimentary face recognition system. <\/p>\n\n<p>The Directories: <\/p>\n\n<p>amar -> contains all the target images<\/p>\n\n<p>test -> contains all the testing images with negatives<\/p>\n\n<p>train -> contains all the training images<\/p>\n\n<p>There are 60 image files in each directory. I am using the directory names as the image labels. <\/p>\n\n<p>At this point I am able to extract the images intensities and everything is done but I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The following is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34570754","CreationDate":"2016-01-02T21:09:16.457","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>batch_xs and batch_ys (and in general inputs to the feed_dict) are expected to be numpy arrays and not lists.<\/p>\n"},{"AnswerId":"45648063","CreationDate":"2017-08-12T08:14:03.347","ParentId":null,"OwnerUserId":"7831265","Title":null,"Body":"<p>In your Batch function you are returning image,label and both are list types.\nMy Suggestion is that you convert this images and labels into np.array right before returning it.\nsomething like this <\/p>\n\n<pre><code>def batch(no):\n     do the batch calculation and prepare the img,label set\n     # now before returning convert then to np array\n     img = np.array(img)\n     return img, label\n<\/code><\/pre>\n\n<p>I tried your batch function in my code and this is how I was able to solve the error.<\/p>\n\n<p>Hope this will solve your error.<\/p>\n"}]}
{"QuestionId":34571812,"AnswerCount":1,"Tags":"<python-3.x><pycharm><anaconda><theano><conda>","CreationDate":"2016-01-02T23:30:19.787","AcceptedAnswerId":34572343.0,"OwnerUserId":3834415.0,"Title":"Integrating Theano Installation with Anaconda install","Body":"<p>I installed Theano according to the instructions on the website:<\/p>\n\n<ol>\n<li>clone the git <\/li>\n<li>run the command <\/li>\n<li>...<\/li>\n<\/ol>\n\n<p>I am using Pycharm with an Anaconda installation as my interpreter.  How do I integrate my new Theano installation with my Anaconda interpreter and make it part of the \"Anaconda environment\"?  (Theano is in , Anaconda is in )<\/p>\n\n<p><strong>My Theano installation actually shows up in my Pycharm -> Settings -> Interpreter window (location: ) But when I reference it in code, it is not recognized.<\/strong><\/p>\n","answers":[{"AnswerId":"34572343","CreationDate":"2016-01-03T00:56:29.457","ParentId":null,"OwnerUserId":"837534","Title":null,"Body":"<p>You can to install Theano with <code>conda<\/code>:<\/p>\n\n<pre><code>conda install -c https:\/\/conda.anaconda.org\/ideas theano\n<\/code><\/pre>\n\n<p>This gives you 0.7 for Windows 64. It should work with Anaconda.<\/p>\n"}]}
{"QuestionId":34572414,"AnswerCount":1,"Tags":"<python><neural-network><theano><conv-neural-network><lasagne>","CreationDate":"2016-01-03T01:12:29.197","AcceptedAnswerId":34644185.0,"OwnerUserId":3510129.0,"Title":"Error while using Conv2DLayer with lasagne NeuralNet","Body":"<p>I have windows 8.1 64bit and use recommended here <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#installing-theano\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#installing-theano<\/a> python win-python distribution (python 3.4). I've went through every step of tutorial (excluding CUDA stuff and GPU config), uninstalled everything and did it again but my problem persists. I am trying to build convolutional neural network using Lasagne. Every layer I've tested so far is working - only Conv2DLayer throws errors. Code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Below I have pasted the error that I get. I suppose that something is wrong with gcc (I got exact version from tutorial) but can't really figure out what is the problem. I have bleeding edge versions of lasagne\/theano. I've tried both Anaconda and WinPython distributions. I've tried both with python 2.7 and 3.4 and can not really find any solution to this problem. Nets with only dense\/maxPooling layers are running just fine. I would be grateful for any suggestions that would help me to resolve this issue.<\/p>\n\n<p>Error Message:<\/p>\n\n<pre>\nProblem occurred during compilation with the command line below:\ng++.exe -shared -g -O3 -fno-math-errno -Wno-unused-label -Wno-unused-variable -Wno-write-strings -march=haswell -mmmx -mno-3dnow -msse -msse2 -msse3 -mssse3 -mno-sse4a -mcx16 -msahf -mmovbe -maes -mno-sha -mpclmul -mpopcnt -mabm -mno-lwp -mfma -mno-fma4 -mno-xop -mbmi -mbmi2 -mno-tbm -mavx -mavx2 -msse4.2 -msse4.1 -mlzcnt -mrtm -mhle -mrdrnd -mf16c -mfsgsbase -mno-rdseed -mno-prfchw -mno-adx -mfxsr -mxsave -mxsaveopt -mno-avx512f -mno-avx512er -mno-avx512cd -mno-avx512pf -mno-prefetchwt1 --param l1-cache-size=32 --param l1-cache-line-size=64 --param l2-cache-size=6144 -mtune=haswell -D NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION -m64 -DMS_WIN64 -IC:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\numpy\\core\\include -IC:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\include -IC:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof -o C:\\Users\\Michal\\AppData\\Local\\Theano\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\\tmpgdh7ov2i\\mf217e5b3a6b61b4ef70844368439f6cb.pyd C:\\Users\\Michal\\AppData\\Local\\Theano\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\\tmpgdh7ov2i\\mod.cpp -LC:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\libs -LC:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64 -lpython34\nTraceback (most recent call last):\n\n  File \"\", line 1, in \n    net2.fit(X, y)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\nolearn\\lasagne\\base.py\", line 457, in fit\n    self.initialize()\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\nolearn\\lasagne\\base.py\", line 303, in initialize\n    self.y_tensor_type,\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\nolearn\\lasagne\\base.py\", line 435, in _create_iter_funcs\n    allow_input_downcast=True,\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\compile\\function.py\", line 317, in function\n    output_keys=output_keys)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\compile\\pfunc.py\", line 526, in pfunc\n    output_keys=output_keys)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\compile\\function_module.py\", line 1778, in orig_function\n    defaults)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\compile\\function_module.py\", line 1642, in create\n    input_storage=input_storage_lists, storage_map=storage_map)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\link.py\", line 690, in make_thunk\n    storage_map=storage_map)[:3]\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\vm.py\", line 1037, in make_all\n    no_recycling))\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\op.py\", line 932, in make_thunk\n    no_recycling)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\op.py\", line 850, in make_c_thunk\n    output_storage=node_output_storage)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cc.py\", line 1207, in make_thunk\n    keep_lock=keep_lock)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cc.py\", line 1152, in __compile__\n    keep_lock=keep_lock)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cc.py\", line 1602, in cthunk_factory\n    key=key, lnk=self, keep_lock=keep_lock)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 1174, in module_from_key\n    module = lnk.compile_cmodule(location)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cc.py\", line 1513, in compile_cmodule\n    preargs=preargs)\n\n  File \"C:\\scisoft\\WinPython-64bit-3.4.3.7\\python-3.4.3.amd64\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 2187, in compile_str\n    (status, compile_stderr.replace('\\n', '. ')))\n\nException: ('The following error happened while compiling the node', CorrMM{valid, (1, 1)}(input.input, Subtensor{::, ::, ::int64, ::int64}.0), '\\n', \"Compilation failed (return status=1): C:\\\\Users\\\\Michal\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\\\\tmpgdh7ov2i\\\\mod.cpp: In member function 'int {anonymous}::__struct_compiled_op_mf217e5b3a6b61b4ef70844368439f6cb::run()':\\r. C:\\\\Users\\\\Michal\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\\\\tmpgdh7ov2i\\\\mod.cpp:947:16: warning: converting to non-pointer type 'int' from NULL [-Wconversion-null]\\r.              kH = NULL;\\r.                 ^\\r. C:\\\\Users\\\\Michal\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\\\\tmpgdh7ov2i\\\\mod.cpp:958:16: warning: converting to non-pointer type 'int' from NULL [-Wconversion-null]\\r.              kW = NULL;\\r.                 ^\\r. C:\\\\Users\\\\Michal\\\\AppData\\\\Local\\\\Temp\\\\cc67su6o.o: In function `corrMM(tagPyArrayObject*, tagPyArrayObject*, tagPyArrayObject*, int, int, int, int, int)':\\r. C:\/Users\/Michal\/AppData\/Local\/Theano\/compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\/tmpgdh7ov2i\/mod.cpp:431: undefined reference to `dgemm_'\\r. C:\/Users\/Michal\/AppData\/Local\/Theano\/compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\/tmpgdh7ov2i\/mod.cpp:528: undefined reference to `dgemm_'\\r. C:\/Users\/Michal\/AppData\/Local\/Theano\/compiledir_Windows-8-6.2.9200-Intel64_Family_6_Model_60_Stepping_3_GenuineIntel-3.4.3-64\/tmpgdh7ov2i\/mod.cpp:483: undefined reference to `dgemm_'\\r. collect2.exe: error: ld returned 1 exit status\\r. \", '[CorrMM{valid, (1, 1)}(input.input, )]')\n<\/pre>\n","answers":[{"AnswerId":"34644185","CreationDate":"2016-01-06T22:37:08.890","ParentId":null,"OwnerUserId":"3510129","Title":null,"Body":"<p>I managed to solve my issue by following steps given by IshanAntony on this website <a href=\"https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\/98833\" rel=\"nofollow\">https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\/98833<\/a> . Apparently Theano had problem with linking BLAS correctly so I downloaded and used open-BLAS and error vanished (links and steps needed to download and configure OPEN-BLAS can be found in IshanAntony's answer).<\/p>\n"}]}
{"QuestionId":34574577,"AnswerCount":3,"Tags":"<lua><torch>","CreationDate":"2016-01-03T08:13:09.937","AcceptedAnswerId":34577511.0,"OwnerUserId":3106889.0,"Title":"Find a smallest positive real number in an Tensor7?","Body":"<p>If I have  4D Tensor and the Tensor contains real numbers <\/p>\n\n<p>Could you please tell how to  find the smallest positive number greater than zero in the tensor <\/p>\n\n<p>for example I have:<\/p>\n\n<pre><\/pre>\n\n<p>Here the smallest positive number greater than zero  is 0.3.<\/p>\n","answers":[{"AnswerId":"34574964","CreationDate":"2016-01-03T09:14:32.550","ParentId":null,"OwnerUserId":"3735873","Title":null,"Body":"<p>One possibility:<\/p>\n\n<pre><code>t = { 0,  -1,  -3,\n      6,   5,   0,\n      0.3, 0.6, 0.9 }\n\ntemp = {}\nfor _,n in ipairs(t) do\n  if n &gt; 0 then               --keep only positives\n    temp[#temp+1] = n\n  end\nend\ntable.sort(temp)              --sorting them will bring the smallest first\n\nprint('Answer',temp[1])\n<\/code><\/pre>\n\n<p>UPDATE: To also find the position at which the lowest value occurs, modify the above to this:<\/p>\n\n<pre><code>t = { 0,  -1,  -3,\n      6,   5,   0,\n      0.3, 0.6, 0.9 }\n\ntemp = {}\nfor i,n in ipairs(t) do\n  if n &gt; 0 then               --keep only positives\n    temp[#temp+1] = { n = n, p = i}\n  end\nend\ntable.sort(temp,function(a,b) return a.n &lt; b.n end)      --sorting them will bring the smallest first\n\nprint('Answer '.. temp[1].n ..' at position '.. temp[1].p)\n<\/code><\/pre>\n"},{"AnswerId":"34574995","CreationDate":"2016-01-03T09:19:43.667","ParentId":null,"OwnerUserId":"3735873","Title":null,"Body":"<p>Another possibility:<\/p>\n\n<pre><code>t = { 0,  -1,  -3,\n      6,   5,   0,\n      0.3, 0.6, 0.9 }\n\nfor _,n in ipairs(t) do\n  if n &gt; 0 then               --only for positives\n    if ans == nil then\n      ans = n                 --first positive assumed lowest\n    else\n      if n &lt; ans then ans = n end  --if a lower value is found, replaces previous one\n    end\n  end\nend\n\nprint('Answer',ans)\n<\/code><\/pre>\n\n<p>UPDATE: To also find the position at which the lowest value occurs, modify the above to this:<\/p>\n\n<pre><code>t = { 0,  -1,  -3,\n      6,   5,   0,\n      0.3, 0.6, 0.9 }\n\nfor i,n in ipairs(t) do\n  if n &gt; 0 then               --only for positives\n    if ans == nil then\n      ans = n                 --first positive assumed lowest\n      pos = i\n    else\n      if n &lt; ans then         --if a lower value is found, replaces previous one\n        ans = n\n        pos = i               --keep position\n      end\n    end\n  end\nend\n\nprint('Answer '.. ans ..' at position '.. pos)\n<\/code><\/pre>\n"},{"AnswerId":"34577511","CreationDate":"2016-01-03T14:20:18.277","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<pre><code>t = torch.Tensor({{0, -1, -3}, {6, 5, 0}, {0.3, 0.6, 0.9}})\nminpos = torch.min(t[t:gt(0)])\n\n0.3\n<\/code><\/pre>\n\n<p><strong>How to get index(es) of desired element(s):<\/strong><\/p>\n\n<p>1) Create the mask<\/p>\n\n<pre><code>mask = t:eq(minpos)\n\n 0  0  0\n 0  0  0\n 1  0  0\n[torch.ByteTensor of size 3x3]\n<\/code><\/pre>\n\n<p>2) Somehow get the indexes of non-zero elements of the mask. For example, using this function:<\/p>\n\n<pre><code>function indexesOf(mask)\n    local lin_indices = torch.linspace(1, mask:nElement(), mask:nElement())[mask]\n    if lin_indices:nElement() == 0 then return nil end\n\n    local sp_indices = torch.LongTensor(mask:nDimension(), lin_indices:nElement())\n    sp_indices[1] = lin_indices - 1\n    local divisor = mask:nElement()\n    for d = 1, mask:nDimension() - 1 do\n        divisor = divisor \/ mask:size(d)\n        local fdiv = torch.div(sp_indices[d], divisor)\n        sp_indices[d + 1] = sp_indices[d] - fdiv * divisor\n        sp_indices[d] = fdiv\n    end\n    return sp_indices:t() + 1\nend\n\nindexes = indexesOf(mask)\n\n 3  1\n[torch.LongTensor of size 1x2]\n<\/code><\/pre>\n"}]}
{"QuestionId":34575046,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-03T09:27:13.623","AcceptedAnswerId":34576796.0,"OwnerUserId":5737155.0,"Title":"Get the value of tensor in theano","Body":"<p>In my program, I want to use the shape value of a tensor. For example, the tensor x has shape of (3,4,5). I want to get the value of 4, i tried the way below:<\/p>\n\n<pre><\/pre>\n\n<p>how can i get a value of 4, what i want is a type of int, not scalar or other types.<\/p>\n","answers":[{"AnswerId":"34576796","CreationDate":"2016-01-03T12:59:37.083","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Try this instead:<\/p>\n\n<pre><code>import numpy as np\nint(np.asscalar(x.shape[1].eval()))\n<\/code><\/pre>\n"}]}
{"QuestionId":34577898,"AnswerCount":1,"Tags":"<macos><tensorflow><bazel>","CreationDate":"2016-01-03T15:00:19.330","AcceptedAnswerId":null,"OwnerUserId":5413507.0,"Title":"Installing Bazel Release 0.1.2 on a Mac","Body":"<p>I'm installing tensorflow on a Mac OS X 10.11.12 with Xcode 7.2 and JDK 8 already installed.<\/p>\n\n<p>I'm following the instructions here - <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"nofollow\">http:\/\/bazel.io\/docs\/install.html<\/a><\/p>\n\n<pre><\/pre>\n\n<p>And have downloaded the file <\/p>\n\n<p>how do I type the aforementioned instructions, given the filename for the downloaded file?<\/p>\n\n<p>Thanks.<\/p>\n\n<p>George<\/p>\n","answers":[{"AnswerId":"34577926","CreationDate":"2016-01-03T15:03:56.337","ParentId":null,"OwnerUserId":"1747493","Title":null,"Body":"<p>Simply open a terminal, go to the directory where you downloaded the file using the cd command and execute the instructions.<\/p>\n\n<pre><code>chmod +x bazel-0.1.2-installer-darwin-x86_64.sh\n.\/bazel-0.1.2-installer-darwin-x86_64.sh --user\n<\/code><\/pre>\n"}]}
{"QuestionId":34579979,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-03T18:23:30.470","AcceptedAnswerId":null,"OwnerUserId":3834415.0,"Title":"Error: NVCC compiler not in $path","Body":"<p>I've got Theano working, and I've installed Cuda. <\/p>\n\n<p>The nvcc compiler is in the windows path variable. <\/p>\n\n<p>Yet, when I run theano, I get the following error: <\/p>\n\n<blockquote>\n  <p>ERROR (theano.sandbox.cuda): nvcc compiler not found on $PATH. Check your nvcc installation and try again.<\/p>\n<\/blockquote>\n\n<p>I'm not really sure what could be going on.  Could it be that theano is looking at a different \"$PATH\"?<\/p>\n\n<p>Theanorc file:<\/p>\n\n<pre><\/pre>\n\n<p>(I get the same error regardless of whether I include )<\/p>\n","answers":[{"AnswerId":"39794333","CreationDate":"2016-09-30T14:51:19.083","ParentId":null,"OwnerUserId":"6905301","Title":null,"Body":"<p>Try add the following section in Thranorc file:\n[cuda]\nroot = \/usr\/local\/cuda-7.5<\/p>\n\n<p>I encountered same issue and solved by add these 2 lines.<\/p>\n"}]}
{"QuestionId":34581230,"AnswerCount":2,"Tags":"<python><neural-network><theano>","CreationDate":"2016-01-03T20:25:52.267","AcceptedAnswerId":44004794.0,"OwnerUserId":3834415.0,"Title":"Error in Theano loop: Outputs_info dim and fn output dim mismatch in scan","Body":"<p>I am a beginner with Theano, and I am working from an example of another's code that, presumably, worked at some point (however, I have modified it...but I'm pretty sure my modifications have nothing to do with what is going wrong at the moment).   <\/p>\n\n<p>Anyhow, I am trying to debug a Theano Scan...and I think what I am observing is a fundamental error in the scan function. <\/p>\n\n<pre><\/pre>\n\n<p> is an  matrix,  is an  matrix, and  is an  matrix...and  is a scalar (4).   but I don't think the internals of my function are what are failing at the moment.<\/p>\n\n<p>The error I get is:<\/p>\n\n<blockquote>\n  <p>ValueError: When compiling the inner function of scan the following error has been encountered: The initial state ( in scan nomenclature) of variable IncSubtensor{Set;:int64:}.0 (argument number 1) has 2 dimension(s), while the result of the inner function () has 2 dimension(s) (should be one less than the initial state).<\/p>\n<\/blockquote>\n\n<p>I have tried altering the dimensions of outputs_info, and the return dimensions of , but nothing seems to work so far.  <\/p>\n\n<p>I am currently looking into the documentation...but, from the documentation, it seems like what I am doing is correct (below is the example from the documentation):<\/p>\n\n<pre><\/pre>\n\n<p>And here is the documentation scan:<\/p>\n\n<pre><\/pre>\n\n<p>The return of the function is: '[x_t,y_t]' and the  is ...<\/p>\n\n<p>While in my implementation, the return of the function is:  and the  is ...which makes sense, since I have no reason to pass my output into the function...<\/p>\n","answers":[{"AnswerId":"43507353","CreationDate":"2017-04-19T22:48:28.753","ParentId":null,"OwnerUserId":"7670710","Title":null,"Body":"<p>Try the following? Note the difference with <code>(self.hidden_dim, )<\/code> and <code>(self.hidden_dim)<\/code><\/p>\n\n<pre><code>outputs_info=[None, dict(initial=T.zeros((self.hidden_dim, )))],\n<\/code><\/pre>\n"},{"AnswerId":"44004794","CreationDate":"2017-05-16T14:49:56.303","ParentId":null,"OwnerUserId":"8020011","Title":null,"Body":"<p>I had exactly the same problem when applying RNN for NLP task. This error occurs because of the type of <code>x_t<\/code> argument of <code>forward_prop_step<\/code> function, which is <strong>scalar<\/strong> because of iterating through the ivector <code>x<\/code>. <\/p>\n\n<p>Solution here is to use a <strong>vector<\/strong>. Here, for example, <code>x_tv<\/code> is a vector which have all zeros and 1 at <code>x_t<\/code> index.<\/p>\n\n<pre><code>def forward_prop_step(x_t, s_t_prev, U, V, W):\n    x_tv = T.eye(1, m=input_size, k=x_t)[0]\n    s_t = T.tanh(U.dot(x_tv) + V.dot(s_t_prev))\n    o_t = T.tanh(W.dot(s_t))\n    return [o_t, s_t]\n<\/code><\/pre>\n"}]}
{"QuestionId":34581971,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-01-03T21:46:15.337","AcceptedAnswerId":39615834.0,"OwnerUserId":2515628.0,"Title":"Rename Caffe layers through pycaffe","Body":"<p>Is there a simple way of renaming layers in a <a href=\"\/questions\/tagged\/caffe\" class=\"post-tag\" title=\"show questions tagged &#39;caffe&#39;\" rel=\"tag\">caffe<\/a> network by using the <em>pycaffe<\/em> interface?<\/p>\n\n<p>I have looked through the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"noreferrer\">net surgery<\/a> example, but I cannot find an example of what I need.<\/p>\n\n<p>For example, I would like to load a trained Caffe model and change the name of  layer and its corresponding blob to .<\/p>\n","answers":[{"AnswerId":"39615834","CreationDate":"2016-09-21T11:53:14.210","ParentId":null,"OwnerUserId":"2835427","Title":null,"Body":"<p>I don't know a direct way to do it, but here is a workaround:  <\/p>\n\n<p>Given a pretrained Caffe model <code>my_model.caffemodel<\/code> and its net architecture <code>net.prototxt<\/code>. Make a copy of <code>net.prototxt<\/code> (say <code>net_new.prototxt<\/code>), and change the name of <code>conv1<\/code> layer to <code>new-conv1<\/code> (you can change the names of <code>bottom<\/code> and <code>top<\/code> if you want).<\/p>\n\n<pre><code>import caffe\nnet_old = caffe.Net('net.prototxt','my_model.caffemodel',caffe.TEST)\nnet_new = caffe.Net('net_new.prototxt','my_model.caffemodel',caffe.TEST)\nnet_new.params['new-conv1'][0].data[...] = net_old.params['conv1'][0].data[...]  #copy filter across 2 nets\nnet_new.params['new-conv1'][1].data[...] = net_old.params['conv1'][1].data[...]  #copy bias\nnet_new.save('my_model_new.caffemodel')\n<\/code><\/pre>\n"}]}
{"QuestionId":34582669,"AnswerCount":0,"Tags":"<machine-learning><neural-network><caffe><pycaffe><matcaffe>","CreationDate":"2016-01-03T23:06:02.580","AcceptedAnswerId":null,"OwnerUserId":3879031.0,"Title":"How to generate the predicted label in caffe from the output of the last layer?","Body":"<p>I have trained my own dataset of images (traffic light images 11x27) with LeNet, using caffe and DIGITS interface. I get 99% accuracy and when I give new images via DIGITS, it predicts the good label, so the network seems to work very well.<\/p>\n\n<p>However, I struggle to predict the labels through Python\/Matlab API for caffe. The last layer output (ip2) is a vector with 2 elements (I have 2 classes), which looks like [4.8060, -5.2608] for example (the first component is always positive, the second always negative and the absolute values range from 4 to 20). I know it from many tests in Python, Matlab and DIGITS.<\/p>\n\n<p>My problem is :<\/p>\n\n<ul>\n<li>Argmax can't work directly on this layer (it always gives 0)<\/li>\n<li>If I use a softmax function, it will always give me [1, 0] (and that's actually the value of  or  in the python interface, no matter the class of my image)<\/li>\n<\/ul>\n\n<p>So, how can I get the good label predicted ?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":34585232,"AnswerCount":1,"Tags":"<python><machine-learning><gpgpu><theano><theano-cuda>","CreationDate":"2016-01-04T05:34:03.277","AcceptedAnswerId":34605596.0,"OwnerUserId":4935271.0,"Title":"Understanding Theano Example in terms of GPU cores\/threads","Body":"<p>I am just getting started with Theano and Deep Learning.  I was experimenting with an example from the Theano tutorial (<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#returning-a-handle-to-device-allocated-data\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#returning-a-handle-to-device-allocated-data<\/a>).  The example code is shown here:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to understand the expression defining 'vlen',<\/p>\n\n<pre><\/pre>\n\n<p>I can't find anywhere in the text that refers to the number of GPU cores specified in this example and why 30 was selected. Nor can I find why the value of 768 threads was used.  My GPU (GeForce 840M) has 384 cores.  Can I assume that if I substitute 384 in the place of the value of 30, that I will be using all 384 cores ?  Also should the value of 768 threads remain fixed ?<\/p>\n","answers":[{"AnswerId":"34605596","CreationDate":"2016-01-05T06:37:29.703","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>I believe the logic is as follows.  Looking at <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#returning-a-handle-to-device-allocated-data\" rel=\"nofollow\">the referenced page<\/a>, we see that there is mention of a GTX 275 GPU.  So the GPU being used for that tutorial may have been a very old CUDA GPU from the cc1.x generation (no longer supported by CUDA 7.0 and 7.5).  In the comment, the developer seems to be using the word \"core\" to refer to a GPU SM (multiprocessor).<\/p>\n\n<p>There were a number of GPUs in that family that had 30 SMs (a cc1.x SM was a very different animal than a cc 2+ SM), including <a href=\"http:\/\/www.geforce.com\/hardware\/desktop-gpus\/geforce-gtx-275\/specifications\" rel=\"nofollow\">GTX 275<\/a> (240 CUDA cores = 30SMs * 8cores\/SM in the cc1.x generation).  So the 30 number is derived from the number of SMs in the GPU being used at the time.<\/p>\n\n<p>Furthermore, if you review <a href=\"https:\/\/en.wikipedia.org\/wiki\/CUDA#Version_features_and_specifications\" rel=\"nofollow\">old documentation<\/a> for CUDA versions that supported such GPUs, you will find that cc1.0 and cc1.1 GPUs supported a max of 768 threads per multiprocessor (SM).  So I believe this is where the 768 number comes from.<\/p>\n\n<p>Finally, a good CUDA code will oversubscribe the GPU (total number of threads is more than what the GPU can instantaneously handle).  So I believe the factor of 10 is just to ensure \"oversubscription\".<\/p>\n\n<p>There is no magic to a particular number -- it is just the length of an array (<code>vlen<\/code>).  The length of this array, after it flows through the theano framework, will ultimately determine the number of threads in CUDA kernel launch.  This code isn't really a benchmark or other performance indicator.  It's stated purpose is just to demonstrate that the GPU is being used.<\/p>\n\n<p>So I wouldn't read too much into that number.  It was a casual choice by the developer that followed a certain amount of logic pertaining to the GPU at hand.<\/p>\n"}]}
{"QuestionId":34586330,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-01-04T07:16:44.223","AcceptedAnswerId":34626725.0,"OwnerUserId":5193828.0,"Title":"How to set learning rate in a big net for fine-tuning in Caffe?","Body":"<p>I have a big net with many layers. I add a new full-connected layer in the net and want to do a fine-tuning. However, it's so difficult to set  in every layer except the new one, since there are many layers in the net.<br>\nIf there is a good way to solve these problem?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"34626725","CreationDate":"2016-01-06T06:10:39.863","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>How about, instead of setting <code>lr_mult: 0<\/code> for all parameters to all layers prior to the new fully connected layer, just stop the back propagation after the new layer?<\/p>\n\n<p>You can do that by setting <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L333\" rel=\"nofollow\"><code>propagate_down: false<\/code><\/a>.<br>\nFor example:<\/p>\n\n<pre><code>layer {\n  name: \"new_layer\"\n  type: \"InnerProduct\"\n  ...\n  inner_product_param {\n    ...\n  }\n  propagate_down: false # do not continue backprop after this layer\n}\n<\/code><\/pre>\n\n<hr>\n\n<p>Alternatively, you can use <a href=\"http:\/\/linux.die.net\/man\/1\/sed\" rel=\"nofollow\"><code>sed<\/code><\/a>, a command line utility, to directly change all entries in your prototxt file:<\/p>\n\n<pre><code>~$ sed -i -E 's\/lr_mult *: *[0-9]+\/lr_mult: 0\/g' train_val.prototxt\n<\/code><\/pre>\n\n<p>This one line will change all <code>lr_mult<\/code> in your <code>train_val.prototxt<\/code> to zero. You'll only need to <em>manually<\/em> set the <code>lr_mult<\/code> for the new layer.<\/p>\n"}]}
{"QuestionId":34587199,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><caffe><lstm><pycaffe>","CreationDate":"2016-01-04T08:30:35.057","AcceptedAnswerId":null,"OwnerUserId":3902984.0,"Title":"How do I use 4096D vector as input to LSTM?","Body":"<p>I am trying to build a Deep Learning system that generates a sentence description for a image. For this I have used a trained CNN caffe model to get a 4096 dimensional feature vector. From what I understand, I need to give this 4096 dimensional vector as an input to LSTM. How can this be done in caffe or some other framework?<\/p>\n","answers":[]}
{"QuestionId":34587251,"AnswerCount":1,"Tags":"<python><numpy><parallel-processing><gpgpu><theano>","CreationDate":"2016-01-04T08:34:04.667","AcceptedAnswerId":34589235.0,"OwnerUserId":3041068.0,"Title":"Speedup for the reduce operation in Theano","Body":"<p><strong>Edit<\/strong>:<br>\nSo sorry, it turns out that I had other processes running on my GPU while I did the test, I've updated the timing results on an idle GPU, and the speedup becomes noticeable for larger matrices.<\/p>\n\n<p>Original Post:<\/p>\n\n<p>As posted in <a href=\"https:\/\/stackoverflow.com\/questions\/34511680\/loop-over-or-vectorize-variable-length-matrices-using-theano\">this question<\/a>,  is a list of matrices, where each item  is a  matrix ( is a variable,  is fixed).  <\/p>\n\n<p>I want to compute the sum of  for all items in  ( is the transpose of ) as the following Python code does.<\/p>\n\n<pre><\/pre>\n\n<p>Followings are some examples of Numpy and Theano implementations (for executable script please refer to @DanielRenshaw's answer to the <a href=\"https:\/\/stackoverflow.com\/a\/34512763\/3041068\">previous question<\/a>).<\/p>\n\n<pre><\/pre>\n\n<p>I ran the Numpy version on CPU, and Theano versions on GPU with different  settings, it seems that the Theano versions are always proportionally slower than the Numpy version (regardless of the number and size of matices).   <\/p>\n\n<p>But I was expecting there could be some optimization w.r.t GPU as it is a simple reduce operation.  <\/p>\n\n<p>Could someone help me understand what's going on under the hood?  <\/p>\n\n<p>Edit:<br>\nFollowings are the script (from @DanielRenshaw) for generating data, settings I've tired and results.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34589235","CreationDate":"2016-01-04T10:37:03.933","ParentId":null,"OwnerUserId":"764322","Title":null,"Body":"<p>The problem you are having, is not the <em>number<\/em> of matrices, is the <em>size<\/em> of them. <\/p>\n\n<p>Your test example, creates matrices of size dependent on the number of matrices you have, thus, the more matrices you have the larger the matrices are, but also the larger the python loop overhead is (in reduce operations), and thus, it makes harder to detect speed improvements.<\/p>\n\n<p>I've sightly modify your matrix generation in order to make some new tests:<\/p>\n\n<pre><code>S = 1000 # Size of the matrices\nN = 10 # Number of matrices\n\nL = [np.random.standard_normal(size=(np.random.randint(S\/\/2, S*2), S)).astype(np.float32) for _ in range(N)]\n<\/code><\/pre>\n\n<p>This generates only 10 matrices of size <code>(x, 1000)<\/code> where <code>x<\/code> is some value in the range of <code>[S\/\/2, S*2] == [500, 2000]<\/code>.<\/p>\n\n<pre><code>f1 = compile_theano_version1(N, S, np.float32)\nf2 = compile_theano_version2(N, S)\n<\/code><\/pre>\n\n<hr>\n\n<ul>\n<li>Now some tests with <code>N = 10<\/code> big matrices:<\/li>\n<\/ul>\n\n<p>For <code>S = 1000<\/code>, <code>N = 10<\/code>:<\/p>\n\n<pre><code> %timeit numpy_version1(*L)   # 10 loops, best of 3: 131 ms per loop\n %timeit f1(*L)               # 10 loops, best of 3: 37.3 ms per loop\n %timeit f2(L)                # 10 loops, best of 3: 68.7 ms per loop\n<\/code><\/pre>\n\n<p>where theano functions have a <code>x4<\/code> and <code>x2<\/code> speedup in a laptop with a pretty nice <code>i7<\/code> and a decent NVIDIA 860M (which means that you should get some nicer speedups here).<\/p>\n\n<p>For <code>S = 5000<\/code>, <code>N = 10<\/code>:<\/p>\n\n<pre><code> %timeit numpy_version1(*L)   # 1 loops, best of 3: 4 s per loop\n %timeit f1(*L)               # 1 loops, best of 3: 907 ms per loop\n %timeit f2(L)                # 1 loops, best of 3: 1.77 s per loop\n<\/code><\/pre>\n\n<p>So, overall, with this setup the larger the <code>S<\/code> the larger the speedup theano gets over the CPU. <\/p>\n\n<hr>\n\n<ul>\n<li>Some tests with <code>N = 100<\/code> big matrices: theano seems faster<\/li>\n<\/ul>\n\n<p>For <code>S = 1000<\/code>, <code>N = 100<\/code>:<\/p>\n\n<pre><code>%timeit numpy_version1(*L)   # 1 loops, best of 3: 1.46 s per loop\n%timeit f1(*L)               # 1 loops, best of 3: 408 ms per loop\n%timeit f2(L)                # 1 loops, best of 3: 724 s per loop\n<\/code><\/pre>\n\n<p>For <code>S = 2000<\/code>, <code>N = 100<\/code>:<\/p>\n\n<pre><code>%timeit numpy_version1(*L)   # 1 loops, best of 3: 11.3 s per loop\n%timeit f1(*L)               # 1 loops, best of 3: 2.72 s per loop\n%timeit f2(L)                # 1 loops, best of 3: 4.01 s per loop\n<\/code><\/pre>\n\n<hr>\n\n<ul>\n<li>Tests with <code>N = 100<\/code> small matrices: numpy seems faster<\/li>\n<\/ul>\n\n<p>For <code>S = 50<\/code>, <code>N = 100<\/code>:<\/p>\n\n<pre><code>%timeit numpy_version1(*L)   # 100 loops, best of 3: 1.17 ms per loop\n%timeit f1(*L)               # 100 loops, best of 3: 4.21 ms per loop\n%timeit f2(L)                # 100 loops, best of 3: 7.42 ms per loop\n<\/code><\/pre>\n\n<hr>\n\n<p>Specifications for the tests:<\/p>\n\n<ul>\n<li>Processor: i7 4710HQ<\/li>\n<li>GPU: NVIDIA GeForce GTX 860M<\/li>\n<li>Numpy: version 1.10.2 built with intel MKT<\/li>\n<li>Theano: version 0.70; floatX = float32; using GPU<\/li>\n<\/ul>\n"}]}
{"QuestionId":34589275,"AnswerCount":1,"Tags":"<python><parameters><neural-network><theano><deep-learning>","CreationDate":"2016-01-04T10:39:58.880","AcceptedAnswerId":34593906.0,"OwnerUserId":5583416.0,"Title":"why are layer parameters added for multilayer perceptron example at deeplearning.net","Body":"<p>For the multilayer perceptron example at<\/p>\n\n<p>[<a href=\"http:\/\/www.deeplearning.net\/tutorial\/mlp.html#mlp][1]\" rel=\"nofollow\">http:\/\/www.deeplearning.net\/tutorial\/mlp.html#mlp][1]<\/a><\/p>\n\n<p>why are the params (weights &amp; biases) added.  See the fifth code section in the above page.<\/p>\n","answers":[{"AnswerId":"34593906","CreationDate":"2016-01-04T14:58:44.413","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>They are not really \"<em>added<\/em>\". The parameters of the layer are in a list and sum of lists is still a list. So the operation is not a mathematical summation.<\/p>\n\n<p>Here's the code to describe the idea:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>param1 = range(2, 6)\nparam2 = range(12, 15)\nprint param1\nprint param2\nprint \"param1 + param2: %s\" % ''.join(str(param1 + param2))\n<\/code><\/pre>\n\n<p>Will print<\/p>\n\n<pre><code>[2, 3, 4, 5]\n[12, 13, 14]\nparam1 + param2: [2, 3, 4, 5, 12, 13, 14]\n<\/code><\/pre>\n"}]}
{"QuestionId":34589335,"AnswerCount":1,"Tags":"<machine-learning><neural-network><tensorflow>","CreationDate":"2016-01-04T10:43:25.127","AcceptedAnswerId":34594851.0,"OwnerUserId":628096.0,"Title":"How does the distorted_inputs() function in the TensorFlow CIFAR-10 example tutorial get 128 images per batch?","Body":"\n\n<p>I was going through the CIFAR-10 example at <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html\" rel=\"nofollow noreferrer\">TensorFlow getting started guide for CNN<\/a><\/p>\n\n<p>Now in the train function in <strong>cifar10_train.py<\/strong> we get images as <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In the  function we generate the filenames in a queue and then read a single record as <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I add debugging code, the  variable contains only 1 record with an image and its height, width, and label name.<\/p>\n\n<p>The example then applies some distortion to the read image\/record and then passes it to the  function.<\/p>\n\n<p>This function then returns a 4D Tensor of shape   where . <\/p>\n\n<p>The above function utilizes the  function when returns the batch. <\/p>\n\n<p>My question is where do the extra records come from in the  function? We are not passing it any filename or  reader object. <\/p>\n\n<p>Can someone shed some light on how we go from 1 record to 128 records? I looked into the documentation but didn't understand.<\/p>\n","answers":[{"AnswerId":"34594851","CreationDate":"2016-01-04T15:50:15.013","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#shuffle_batch\"><code>tf.train.shuffle_batch()<\/code><\/a> function can be used to produce (one or more) tensors containing a batch of inputs. Internally, <code>tf.train.shuffle_batch()<\/code> creates a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#RandomShuffleQueue\"><code>tf.RandomShuffleQueue<\/code><\/a>, on which it calls <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#QueueBase.enqueue\"><code>q.enqueue()<\/code><\/a> with the image and label tensors to enqueue a single element (image-label pair). It then returns the result of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\"><code>q.dequeue_many(batch_size)<\/code><\/a>, which concatenates <code>batch_size<\/code> randomly selected elements (image-label pairs) into a batch of images and a batch of labels.<\/p>\n\n<p>Note that, although it looks from the code like <code>read_input<\/code> and <code>filename_queue<\/code> have a functional relationship, there is an additional wrinkle. Simply evaluating the result of <code>tf.train.shuffle_batch()<\/code> will block forever, because no elements have been added to the internal queue. To simplify this, when you call <code>tf.train.shuffle_batch()<\/code>, TensorFlow will add a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#QueueRunner\"><code>QueueRunner<\/code><\/a> to an internal collection in the graph. A later call to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#start_queue_runners\"><code>tf.train.start_queue_runners()<\/code><\/a> (e.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2d5f0ac61fe4e4160fbb68d8031418528111dae9\/tensorflow\/models\/image\/cifar10\/cifar10_train.py#L97\">here in <code>cifar10_train.py<\/code><\/a>) will start a thread that adds elements to the queue, and enables training to proceed. The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/threading_and_queues\/index.html\">Threading and Queues HOWTO<\/a> has more information on how this works.<\/p>\n"}]}
{"QuestionId":34589492,"AnswerCount":1,"Tags":"<python><artificial-intelligence><classification><theano>","CreationDate":"2016-01-04T10:53:05.043","AcceptedAnswerId":34790870.0,"OwnerUserId":497600.0,"Title":"Theano: How to implement the distance between desired output (1d) and label as cost function","Body":"<p>I would like to train a neural network to represent a function from R^n to R.\nThe neuronal network has only one layer, the input neurons are the function parameters, the output is the function value.\nFor example, the function could be \"logical and\": Two input values, one output.<\/p>\n\n<p>In order to train such a network, I need to define a cost function which can then be derived with Theano's support for gradients.\nThe problem is:\nUsually you would use a neuronal network for classification. A training sample is (input, y) where y is the desired output = index of the output neuron which should have maximum likelihood.<\/p>\n\n<p>In this case that is impossible, I only have one output neuron and need to compare it with the label => The label is not used for indexing.\nIn pseudocode:<\/p>\n\n<pre><\/pre>\n\n<p>With this approach, the cost would have to be recomputed with every sample. Since the formula for computing the cost depends on the value of y.<\/p>\n\n<p>I believe it is necessary to implement the choice inside of a theano formula.\nSomething like in this pseudocode:<\/p>\n\n<pre><\/pre>\n\n<p>In the theano tutorials indexing is used in combination with differentiation, so this should work.<\/p>\n\n<p><strong>Actual question:<\/strong>\nHow does blocking work in theano ?<\/p>\n\n<p>The types of the symbols in my code: output is a matrix, y is a vector. The samples are created like this, a row is a sample.<\/p>\n\n<pre><\/pre>\n\n<p>The complete code on pastebin: <a href=\"http:\/\/pastebin.com\/PByUyvMQ\" rel=\"nofollow\">http:\/\/pastebin.com\/PByUyvMQ<\/a>\nIt is mostly exactly like this tutorial: <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a><\/p>\n","answers":[{"AnswerId":"34790870","CreationDate":"2016-01-14T13:28:30.723","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>I'm not quite sure what blocking exactly means, but for simply concatenating two tensor variables, there're <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\"><code>theano.tensor.concatenate()<\/code><\/a> and <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.stack\" rel=\"nofollow\"><code>theano.tensor.stack()<\/code><\/a>.  <\/p>\n\n<p>Moreover, there's another formulation of your problem that may save from conditioning and blocking:<\/p>\n\n<p><code>cost = -(1-y)*output -y*(1-output)<\/code><\/p>\n"}]}
{"QuestionId":34589573,"AnswerCount":2,"Tags":"<macos><unix><boost><caffe>","CreationDate":"2016-01-04T10:57:54.647","AcceptedAnswerId":null,"OwnerUserId":4127806.0,"Title":"error installing caffe","Body":"<p>I am trying to install caffe on my mac. I have installed boost but I get this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47663383","CreationDate":"2017-12-05T22:01:57.003","ParentId":null,"OwnerUserId":"781938","Title":null,"Body":"<p>If you're using anaconda, you can also provide the boost dependency using the conda-forge anaconda channel:<\/p>\n\n<p><code>$ conda install -c conda-forge boost<\/code><\/p>\n\n<p>And create the symlinks, as @Marqin said, if they're not available. <\/p>\n\n<p>The conda-forge boost might soon include the <code>-mt<\/code> libraries. I opened an issue for it on <a href=\"https:\/\/github.com\/conda-forge\/boost-cpp-feedstock\/issues\/18\" rel=\"nofollow noreferrer\">their github repo<\/a>. <\/p>\n"},{"AnswerId":"34817424","CreationDate":"2016-01-15T18:25:26.690","ParentId":null,"OwnerUserId":"898435","Title":null,"Body":"<p>You have probably installed boost without boost-thread, or have old installer that installed boost_thread-mt as boost_thread. If it's the second then you can just make symbolic link from <code>libboost_thread.a<\/code> to <code>libboost_thread-mt.a<\/code>.<\/p>\n\n<p>If it's the first cause, then you have to reinstall brew with boost-thread, and the easiest way will be to <a href=\"http:\/\/brew.sh\/\" rel=\"noreferrer\">get Homebrew<\/a> and install boost with it:<\/p>\n\n<pre class=\"lang-bash prettyprint-override\"><code>brew install boost\n<\/code><\/pre>\n\n<p>Which installs proper libs:<\/p>\n\n<pre><code>-&gt; % ls  \/usr\/local\/lib\/libboost_thread-mt.* \n\/usr\/local\/lib\/libboost_thread-mt.a     \/usr\/local\/lib\/libboost_thread-mt.dylib\n<\/code><\/pre>\n"}]}
{"QuestionId":34589693,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-01-04T11:04:56.093","AcceptedAnswerId":34600950.0,"OwnerUserId":5583416.0,"Title":"How to add two shared variables in theano?","Body":"<pre><\/pre>\n\n<p> gives <\/p>\n\n<p>I want the value of C. How do I get it?<\/p>\n","answers":[{"AnswerId":"34600950","CreationDate":"2016-01-04T22:22:24.607","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Your code as it is will not work since the shapes of your shared variables do not align.<\/p>\n\n<p>Correcting your example, you can do<\/p>\n\n<pre><code>import theano\nimport numpy as np\n\nA = theano.shared(np.random.randn(3, 4))\nB = theano.shared(np.random.randn(3, 4))\nC = A + B\n<\/code><\/pre>\n\n<p>Then this is correctly evaluable. If you are working in command line, then <code>C.eval()<\/code> will do the trick. However, the more general and comprehensive approach to this is to create a theano function.<\/p>\n\n<pre><code>f = theano.function([], C)\n<\/code><\/pre>\n\n<p>Then you can call <code>f()<\/code> and get the value of <code>C<\/code>. If your computation depended on other (non-shared) symbolic variables, you would provide the necessary values as arguments to the function (this also works with <code>eval<\/code> by specifying a dictionary with the relevant entries).<\/p>\n"}]}
{"QuestionId":34591195,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-01-04T12:28:48.520","AcceptedAnswerId":null,"OwnerUserId":286579.0,"Title":"Multitask deep learning with Tensorflow","Body":"<p>Have someone tried doing multitask deep learning with TensorFlow? That is, sharing the bottom layers while not sharing the top layers.  An example with simple illustration would help a lot. <\/p>\n","answers":[{"AnswerId":"40656464","CreationDate":"2016-11-17T13:40:56.603","ParentId":null,"OwnerUserId":"2991813","Title":null,"Body":"<p>There is an similar question <a href=\"https:\/\/stackoverflow.com\/questions\/37861112\/how-to-do-multitask-deep-learning-with-tensorflow\">here<\/a>, the answer used keras. <\/p>\n\n<p>It's similar when just using tensorflow. The idea is this: we can define multiple outputs of a network, and thus multiple loss functions (objectives). We then tell optimizer to minimize a combined loss function, usually using a linear combination.<\/p>\n\n<p>A concept diagram <a href=\"https:\/\/i.stack.imgur.com\/Jcumv.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Jcumv.png\" alt=\"Multi Task Deep Learning Diagram\"><\/a><\/p>\n\n<p>This diagram is drawn according to this <a href=\"https:\/\/www.google.com.hk\/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=1&amp;ved=0ahUKEwj54p_W86_QAhUDHJQKHeKoAXEQFggeMAA&amp;url=http%3A%2F%2Fresearch.google.com%2Fpubs%2Farchive%2F42241.pdf&amp;usg=AFQjCNER9_KAZYR4LCjFbKP9FL_cmXeHgg\" rel=\"noreferrer\">paper<\/a>.<\/p>\n\n<p>Let's say we are training a classifier that predict the digit in the image, with maximum 5 digits per image. Here we defined 6 output layer: <code>digit1<\/code>, <code>digit2<\/code>, <code>digit3<\/code>, <code>digit4<\/code>, <code>digit5<\/code>, <code>length<\/code>. The <code>digit<\/code> layer should output 0~9 if there is such a digit, or <code>X<\/code>(<em>substitute it with an real number in practice<\/em>) if there isn't any digit <strong>in its position<\/strong>. Same thing for <code>length<\/code>, it should output 0~5 if the image contains 0~5 digit, or <code>X<\/code> if it contains more than 5 digits.<\/p>\n\n<p>Now to train it, we just add up all the cross entropy loss of each softmax function:<\/p>\n\n<pre><code># Define loss and optimizer\nlossLength = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(length_logits, true_length)), 1e-37, 1e+37))\nlossDigit1 = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(digit1_logits, true_digit1)), 1e-37, 1e+37))\nlossDigit2 = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(digit2_logits, true_digit2)), 1e-37, 1e+37))\nlossDigit3 = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(digit3_logits, true_digit3)), 1e-37, 1e+37))\nlossDigit4 = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(digit4_logits, true_digit4)), 1e-37, 1e+37))\nlossDigit5 = tf.log(tf.clip_by_value(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(digit5_logits, true_digit5)), 1e-37, 1e+37))\n\ncost = tf.add(\n        tf.add(\n        tf.add(\n        tf.add(\n        tf.add(cL,lossDigit1),\n        lossDigit2),\n        lossDigit3),\n        lossDigit4),\n        lossDigit5)\n\noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)\n<\/code><\/pre>\n"}]}
{"QuestionId":34592001,"AnswerCount":2,"Tags":"<c++><machine-learning><deep-learning><caffe><leveldb>","CreationDate":"2016-01-04T13:13:04.213","AcceptedAnswerId":34595915.0,"OwnerUserId":5572849.0,"Title":"How to write or convert float-type data to leveldb in caffe","Body":"<p>Now I am making the leveldb to train caffe framework.So I use \"<a href=\"https:\/\/github.com\/Wangyida\/caffe\/blob\/f23813f3c779e4ec35c0b0b75a4ed574d0c086cd\/tools\/convert_imageset.cpp\" rel=\"nofollow\">convert_imageset.cpp<\/a>\". This cpp file writes the char-type data only to leveldb.\nBut I have the float data to write it to leveldb. This data is pre-proceed image  data so it is float type data.\nhow can I write or convert this float data to leveldb.\nThis float data is a set of vector with 4096 dimensions.\nPlease help me.\nOr not how to convert it to HDF5Data?<\/p>\n","answers":[{"AnswerId":"34607714","CreationDate":"2016-01-05T09:01:41.223","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<pre><code>def del_and_create(dname):\n    if os.path.exists(dname):\n        shutil.rmtree(dname)\n    os.makedirs(dname)\n\ndef get_img_datum(image_fn):\n    img = cv.imread(image_fn, cv.IMREAD_COLOR)\n    img = img.swapaxes(0, 2).swapaxes(1, 2)\n    datum = caffe.io.array_to_datum(img, 0)\n    return datum\n\ndef get_jnt_datum(joint_fn):\n    joint = np.load(joint_fn)\n    datum = caffe.io.caffe_pb2.Datum()\n    datum.channels = len(joint)\n    datum.height = 1\n    datum.width = 1\n    datum.float_data.extend(joint.tolist())\n\n    return datum\n\ndef create_dataset():\n    img_db_fn = 'img.lmdb'\n    del_and_create(img_db_fn)\n    img_env = lmdb.Environment(img_db_fn, map_size=1099511627776)\n    img_txn = img_env.begin(write=True, buffers=True)\n\n    jnt_db_fn = 'joint.lmdb'\n    del_and_create(jnt_db_fn)\n    jnt_env = lmdb.Environment(jnt_db_fn, map_size=1099511627776)\n    jnt_txn = jnt_env.begin(write=True, buffers=True)\n\n    img_fns = glob.glob('imageData\/*.jpg')\n    fileCount = len(img_fns)\n    print 'A total of ', fileCount, ' images.'\n    jnt_fns = glob.glob('jointData\/*.npy')\n    jointCount = len(jnt_fns)\n    if(fileCount != jointCount):\n        print 'The file counts doesnot match'\n        exit()\n\n    keys = np.arange(fileCount)\n    np.random.shuffle(keys)\n\n    for i, (img_fn, jnt_fn) in enumerate( zip(sorted(img_fns), sorted(jnt_fns)) ):\n        img_datum = get_img_datum(img_fn)\n        jnt_datum = get_jnt_datum(jnt_fn)\n        key = '%010d' % keys[i]\n\n        img_txn.put(key, img_datum.SerializeToString())\n        jnt_txn.put(key, jnt_datum.SerializeToString())\n\n        if i % 10000 == 0:\n            img_txn.commit()\n            jnt_txn.commit()\n            jnt_txn = jnt_env.begin(write=True, buffers=True)\n            img_txn = img_env.begin(write=True, buffers=True)\n\n        print '%d'%(i), os.path.basename(img_fn), os.path.basename(jnt_fn)\n\n    img_txn.commit()\n    jnt_txn.commit()\n    img_env.close()\n    jnt_env.close()\n<\/code><\/pre>\n\n<p>The above code expects images from a given path, and the labels of each image as .npy file.<\/p>\n\n<p>Credits: <a href=\"https:\/\/github.com\/mitmul\/deeppose\/blob\/caffe\/scripts\/dataset.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mitmul\/deeppose\/blob\/caffe\/scripts\/dataset.py<\/a><\/p>\n\n<p>Note: I had seen <a href=\"https:\/\/stackoverflow.com\/users\/1714410\/shai\">Shai<\/a>'s <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/773226\">answer<\/a> to a question, which claims that lmdb doesnot support float-type data. But, it does work for me with the latest version of Caffe and LMDB and using this code snippet. As his answer is way too old, its highly likely that older versions may not have supported float-type data.<\/p>\n"},{"AnswerId":"34595915","CreationDate":"2016-01-04T16:51:30.403","ParentId":null,"OwnerUserId":"1537990","Title":null,"Body":"<p>HDF5 stands for hierarchical data format. You can manipulate such data format for example with <code>R<\/code> (<a href=\"https:\/\/www.bioconductor.org\/packages\/3.3\/bioc\/manuals\/rhdf5\/man\/rhdf5.pdf\" rel=\"nofollow\">RHDF5 documentation<\/a>)<\/p>\n\n<p>Other software that can process HDF5 are <code>Matlab<\/code> and <code>Mathematica<\/code>.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>A new set of tools called <code>HDFql<\/code> has been recently released to simplify \"managing HDF files through a high-level language like C\/C++\". You can check it out <a href=\"http:\/\/www.hdfql.com\/\" rel=\"nofollow\">here<\/a><\/p>\n"}]}
{"QuestionId":34592172,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-04T13:22:38.020","AcceptedAnswerId":34606240.0,"OwnerUserId":5743655.0,"Title":"About names of variable scope in tensorflow","Body":"\n\n<p>Recently I have been trying to learn to use TensorFlow, and I do not understand how variable scopes work exactly. In particular, I have the following problem:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Where does the   in  come from? How does it work when later using  to get the same variable again?<\/p>\n\n<p>edit: I think I find a related problem:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The output is<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'm confused...<\/p>\n","answers":[{"AnswerId":"41368131","CreationDate":"2016-12-28T18:58:32.343","ParentId":null,"OwnerUserId":"2990262","Title":null,"Body":"<p>The answer above is somehow misguiding.<\/p>\n\n<p>Let me answer why you got two different scope names, even though it looks like that you defined two identical functions: <code>creating<\/code> and <code>creating_mod<\/code>.<\/p>\n\n<p>This is simply because you used <code>tf.Variable(0.0, name=s)<\/code> to create the variable in the function <code>creating_mod<\/code>.<\/p>\n\n<p><strong>ALWAYS use <code>tf.get_variable<\/code><\/strong>, if you want your variable to be recognized by scope!<\/p>\n\n<p>Check out this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1325\" rel=\"noreferrer\">issue<\/a> for more details.<\/p>\n\n<p>Thanks!<\/p>\n"},{"AnswerId":"34606240","CreationDate":"2016-01-05T07:24:08.647","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"\n\n<p>The <code>\"_2\"<\/code> in <code>\"BasicLSTMCell_2\"<\/code> relates to the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph.name_scope\" rel=\"nofollow noreferrer\"><em>name scope<\/em><\/a> in which the op <code>outpts[2]<\/code> was created. Every time you create a new name scope (with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#name_scope\" rel=\"nofollow noreferrer\"><code>tf.name_scope()<\/code><\/a>) or variable scope (with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#variable_scope\" rel=\"nofollow noreferrer\"><code>tf.variable_scope()<\/code><\/a>) a unique suffix is added to the current name scope, based on the given string, possibly with an additional suffix to make it unique.  The call to <code>rnn.rnn(...)<\/code> has the following pseudocode (simplified and using public API methods for clarity):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs = []\nwith tf.variable_scope(\"RNN\"):\n  for timestep, input_t in enumerate(inputs):\n    if timestep &gt; 0:\n      tf.get_variable_scope().reuse_variables()\n    with tf.variable_scope(\"BasicLSTMCell\"):\n      outputs.append(...)\nreturn outputs\n<\/code><\/pre>\n\n<p>If you look at the names of the tensors in <code>outpts<\/code>, you'll see that they are the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; print [o.name for o in outpts]\n[u'RNN\/BasicLSTMCell\/mul_2:0',\n u'RNN\/BasicLSTMCell_1\/mul_2:0',\n u'RNN\/BasicLSTMCell_2\/mul_2:0',\n u'RNN\/BasicLSTMCell_3\/mul_2:0',\n u'RNN\/BasicLSTMCell_4\/mul_2:0']\n<\/code><\/pre>\n\n<p>When you enter a new name scope (by entering a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#name_scope\" rel=\"nofollow noreferrer\"><code>with tf.name_scope(\"...\"):<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#variable_scope\" rel=\"nofollow noreferrer\"><code>with tf.variable_scope(\"...\"):<\/code><\/a> block), TensorFlow creates a new, <em>unique<\/em> name for the scope. The first time the <code>\"BasicLSTMCell\"<\/code> scope is entered, TensorFlow uses that name verbatim, because it hasn't been used before (in the <code>\"RNN\/\"<\/code> scope). The next time, TensorFlow appends <code>\"_1\"<\/code> to the scope to make it unique, and so on up to <code>\"RNN\/BasicLSTMCell_4\"<\/code>.<\/p>\n\n<p>The main difference between variable scopes and name scopes is that a variable scope also has a set of name-to-<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow noreferrer\"><code>tf.Variable<\/code><\/a> bindings. By calling <code>tf.get_variable_scope().reuse_variables()<\/code>, we instruct TensorFlow to <em>reuse<\/em> rather than create variables for the <code>\"RNN\/\"<\/code> scope (and its children), after timestep 0. This ensures that the weights are correctly shared between the multiple RNN cells.<\/p>\n"}]}
{"QuestionId":34594198,"AnswerCount":2,"Tags":"<python><multithreading><latency><tensorflow><prefetch>","CreationDate":"2016-01-04T15:14:04.837","AcceptedAnswerId":34596212.0,"OwnerUserId":1104696.0,"Title":"How to prefetch data using a custom python function in tensorflow","Body":"<p>I am trying to prefetch training data to hide I\/O latency. I would like to write custom Python code that loads data from disk and preprocesses the data (e.g. by adding a context window). In other words, one thread does data preprocessing and the other does training. Is this possible in TensorFlow?<\/p>\n\n<p>Update: I have a working example based on @mrry's example.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44208648","CreationDate":"2017-05-26T19:10:10.097","ParentId":null,"OwnerUserId":"796120","Title":null,"Body":"<blockquote>\n  <p>In other words, one thread does data preprocessing and the other does training. Is this possible in TensorFlow?<\/p>\n<\/blockquote>\n\n<p>Yes, it is. mrry's solution works, but simpler exists.<\/p>\n\n<h1>Fetching data<\/h1>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\"><code>tf.py_func<\/code><\/a> wraps a python function and uses it as a TensorFlow operator. So we can load the data at <code>sess.run()<\/code> each time. The problem with this approach is that data is loaded during <code>sess.run()<\/code> via the main thread.<\/p>\n\n<p>A minimal example:<\/p>\n\n<pre><code>def get_numpy_tensor():\n  return np.array([[1,2],[3,4]], dtype=np.float32)\ntensorflow_tensor = tf.py_func(get_numpy_tensor, [], tf.float32)\n<\/code><\/pre>\n\n<p>A more complex example:<\/p>\n\n<pre><code>def get_numpy_tensors():\n  # Load data from the disk into numpy arrays.\n  input = np.array([[1,2],[3,4]], dtype=np.float32)\n  target = np.int32(1)\n  return input, target\ntensorflow_input, tensorflow_target = tf.py_func(get_numpy_tensors, [], [tf.float32, tf.int32])\n\ntensorflow_input, tensorflow_target = 2*tensorflow_input, 2*tensorflow_target\n\nsess = tf.InteractiveSession()\nnumpy_input, numpy_target = sess.run([tensorflow_input, tensorflow_target])\nassert np.all(numpy_input==np.array([[2,4],[6,8]])) and numpy_target==2\n<\/code><\/pre>\n\n<h1>Prefetching data in another thread<\/h1>\n\n<p>To queue our data in another thread (so that <code>sess.run()<\/code> won't have to wait for the data), we can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/batch\" rel=\"nofollow noreferrer\"><code>tf.train.batch()<\/code><\/a> on our operators from <code>tf.py_func()<\/code>.<\/p>\n\n<p>A minimal example:<\/p>\n\n<pre><code>tensor_shape = get_numpy_tensor().shape\ntensorflow_tensors = tf.train.batch([tensorflow_tensor], batch_size=32, shapes=[tensor_shape])\n# Run `tf.train.start_queue_runners()` once session is created.\n<\/code><\/pre>\n\n<p>We can omit the argument <code>shapes<\/code> if <code>tensorflow_tensor<\/code> has its shape specified:<\/p>\n\n<pre><code>tensor_shape = get_numpy_tensor().shape\ntensorflow_tensor.set_shape(tensor_shape)\ntensorflow_tensors = tf.train.batch([tensorflow_tensor], batch_size=32)\n# Run `tf.train.start_queue_runners()` once session is created.\n<\/code><\/pre>\n\n<p>A more complex example:<\/p>\n\n<pre><code>input_shape, target_shape = (2, 2), ()\ndef get_numpy_tensors():\n  input = np.random.rand(*input_shape).astype(np.float32)\n  target = np.random.randint(10, dtype=np.int32)\n  print('f', end='')\n  return input, target\ntensorflow_input, tensorflow_target = tf.py_func(get_numpy_tensors, [], [tf.float32, tf.int32])\nbatch_size = 2\ntensorflow_inputs, tensorflow_targets = tf.train.batch([tensorflow_input, tensorflow_target], batch_size, shapes=[input_shape, target_shape], capacity=2)\n# Internal queue will contain at most `capasity=2` times `batch_size=2` elements `[tensorflow_input, tensorflow_target]`.\n\ntensorflow_inputs, tensorflow_targets = 2*tensorflow_inputs, 2*tensorflow_targets\n\nsess = tf.InteractiveSession()\ntf.train.start_queue_runners() # Internally, `tf.train.batch` uses a QueueRunner, so we need to ask tf to start it.\nfor _ in range(10):\n  numpy_inputs, numpy_targets = sess.run([tensorflow_inputs, tensorflow_targets])\n  assert numpy_inputs.shape==(batch_size, *input_shape) and numpy_targets.shape==(batch_size, *target_shape)\n  print('r', end='')\n\n# Prints `fffffrrffrfrffrffrffrffrffrffrf`.\n<\/code><\/pre>\n\n<p>In case <code>get_numpy_tensor()<\/code> returns a batch of tensors, then <code>tf.train.batch(..., enqueue_many=True)<\/code> will help.<\/p>\n"},{"AnswerId":"34596212","CreationDate":"2016-01-04T17:08:34.510","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is a common use case, and most implementations use TensorFlow's <em>queues<\/em> to decouple the preprocessing code from the training code. There is <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/threading_and_queues\/index.html\" rel=\"noreferrer\">a tutorial on how to use queues<\/a>, but the main steps are as follows:<\/p>\n\n<ol>\n<li><p>Define a queue, <code>q<\/code>, that will buffer the preprocessed data. TensorFlow supports the simple <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#FIFOQueue\" rel=\"noreferrer\"><code>tf.FIFOQueue<\/code><\/a> that produces elements in the order they were enqueued, and the more advanced <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#RandomShuffleQueue\" rel=\"noreferrer\"><code>tf.RandomShuffleQueue<\/code><\/a> that produces elements in a random order. A queue element is a tuple of one or more tensors (which can have different types and shapes). All queues support single-element (<code>enqueue<\/code>, <code>dequeue<\/code>) and batch (<code>enqueue_many<\/code>, <code>dequeue_many<\/code>) operations, but to use the batch operations you must specify the shapes of each tensor in a queue element when constructing the queue.<\/p><\/li>\n<li><p>Build a subgraph that enqueues preprocessed elements into the queue. One way to do this would be to define some <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> ops for tensors corresponding to a single input example, then pass them to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#QueueBase.enqueue\" rel=\"noreferrer\"><code>q.enqueue()<\/code><\/a>. (If your preprocessing produces a batch at once, you should use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#QueueBase.enqueue_many\" rel=\"noreferrer\"><code>q.enqueue_many()<\/code><\/a> instead.) You might also include TensorFlow ops in this subgraph.<\/p><\/li>\n<li><p>Build a subgraph that performs training. This will look like a regular TensorFlow graph, but will get its input by calling <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\" rel=\"noreferrer\"><code>q.dequeue_many(BATCH_SIZE)<\/code><\/a>.<\/p><\/li>\n<li><p>Start your session.<\/p><\/li>\n<li><p>Create one or more threads that execute your preprocessing logic, then execute the enqueue op, feeding in the preprocessed data. You may find the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#Coordinator\" rel=\"noreferrer\"><code>tf.train.Coordinator<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#QueueRunner\" rel=\"noreferrer\"><code>tf.train.QueueRunner<\/code><\/a> utility classes useful for this.<\/p><\/li>\n<li><p>Run your training graph (optimizer, etc.) as normal.<\/p><\/li>\n<\/ol>\n\n<p><strong>EDIT:<\/strong> Here's a simple <code>load_and_enqueue()<\/code> function and code fragment to get you started:<\/p>\n\n<pre><code># Features are length-100 vectors of floats\nfeature_input = tf.placeholder(tf.float32, shape=[100])\n# Labels are scalar integers.\nlabel_input = tf.placeholder(tf.int32, shape=[])\n\n# Alternatively, could do:\n# feature_batch_input = tf.placeholder(tf.float32, shape=[None, 100])\n# label_batch_input = tf.placeholder(tf.int32, shape=[None])\n\nq = tf.FIFOQueue(100, [tf.float32, tf.int32], shapes=[[100], []])\nenqueue_op = q.enqueue([feature_input, label_input])\n\n# For batch input, do:\n# enqueue_op = q.enqueue_many([feature_batch_input, label_batch_input])\n\nfeature_batch, label_batch = q.dequeue_many(BATCH_SIZE)\n# Build rest of model taking label_batch, feature_batch as input.\n# [...]\ntrain_op = ...\n\nsess = tf.Session()\n\ndef load_and_enqueue():\n  with open(...) as feature_file, open(...) as label_file:\n    while True:\n      feature_array = numpy.fromfile(feature_file, numpy.float32, 100)\n      if not feature_array:\n        return\n      label_value = numpy.fromfile(feature_file, numpy.int32, 1)[0]\n\n      sess.run(enqueue_op, feed_dict={feature_input: feature_array,\n                                      label_input: label_value})\n\n# Start a thread to enqueue data asynchronously, and hide I\/O latency.\nt = threading.Thread(target=load_and_enqueue)\nt.start()\n\nfor _ in range(TRAINING_EPOCHS):\n  sess.run(train_op)\n<\/code><\/pre>\n"}]}
{"QuestionId":34597316,"AnswerCount":4,"Tags":"<machine-learning><neural-network><deep-learning><tensorflow>","CreationDate":"2016-01-04T18:17:25.413","AcceptedAnswerId":34597667.0,"OwnerUserId":4341948.0,"Title":"Why input is scaled in tf.nn.dropout in tensorflow?","Body":"<p>I can't understand why dropout works like this in tensorflow. The blog of <a href=\"http:\/\/cs231n.github.io\/neural-networks-2\/\">CS231n<\/a> says that,  Also you can see this from picture(Taken from the same site)\n<a href=\"https:\/\/i.stack.imgur.com\/SbXq1.jpg\"><img src=\"https:\/\/i.stack.imgur.com\/SbXq1.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From tensorflow site, <\/p>\n\n<p>Now, why the input element is scaled up by ? Why not keep the input element as it is with probability and not scale it with ? <\/p>\n","answers":[{"AnswerId":"58374585","CreationDate":"2019-10-14T10:12:59.797","ParentId":null,"OwnerUserId":"8468326","Title":null,"Body":"<p>Here is a quick experiment to disperse any remaining confusion. <\/p>\n\n<p>Statistically the weights of a NN-layer follow a distribution that is <em>usually<\/em> close to normal (but not necessarily), but even in the case when trying to sample a perfect normal distribution in practice, there are always computational errors.<\/p>\n\n<p>Then consider the following experiment:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>DIM = 1_000_000                      # set our dims for weights and input\nx = np.ones((DIM,1))                 # our input vector\n#x = np.random.rand(DIM,1)*2-1.0     # or could also be a more realistic normalized input\n\nprobs = [1.0, 0.7, 0.5, 0.3]         # define dropout probs\n\nW = np.random.normal(size=(DIM,1))   # sample normally distributed weights\nprint(\"W-mean = \", W.mean())         # note the mean is not perfect --&gt; sampling error!\n\n# DO THE DRILL\nh = defaultdict(list)\nfor i in range(1000):\n  for p in probs:\n    M = np.random.rand(DIM,1)\n    M = (M &lt; p).astype(int)\n    Wp = W * M\n    a = np.dot(Wp.T, x)\n    h[str(p)].append(a)\n\nfor k,v in h.items():\n  print(\"For drop-out prob %r the average linear activation is %r (unscaled) and %r (scaled)\" % (k, np.mean(v), np.mean(v)\/float(k)))\n<\/code><\/pre>\n\n<p>Sample output:<\/p>\n\n<pre><code>x-mean =  1.0\nW-mean =  -0.001003985674840264\nFor drop-out prob '1.0' the average linear activation is -1003.985674840258 (unscaled) and -1003.985674840258 (scaled)\nFor drop-out prob '0.7' the average linear activation is -700.6128015029908 (unscaled) and -1000.8754307185584 (scaled)\nFor drop-out prob '0.5' the average linear activation is -512.1602655283492 (unscaled) and -1024.3205310566984 (scaled)\nFor drop-out prob '0.3' the average linear activation is -303.21194422742315 (unscaled) and -1010.7064807580772 (scaled)\n<\/code><\/pre>\n\n<p>Notice that the unscaled activations diminish due to the <em>statistically imperfect normal distribution<\/em>.<\/p>\n\n<p>Can you spot an obvious correlation between the <code>W-mean<\/code> and the average linear activation means?<\/p>\n"},{"AnswerId":"56576486","CreationDate":"2019-06-13T08:32:16.170","ParentId":null,"OwnerUserId":"9492673","Title":null,"Body":"<p>If you keep reading in <a href=\"http:\/\/cs231n.github.io\/neural-networks-2\/\" rel=\"nofollow noreferrer\">cs231n<\/a>, the difference between <strong>dropout<\/strong> and <strong>inverted dropout<\/strong> is explained. <\/p>\n\n<p>Since we want to leave the forward pass at test time untouched (and tweak our network just during training), <code>tf.nn.dropout<\/code> directly implements <strong>inverted dropout<\/strong>, scaling the values.<\/p>\n"},{"AnswerId":"47665238","CreationDate":"2017-12-06T01:07:59.107","ParentId":null,"OwnerUserId":"1499598","Title":null,"Body":"<p>Let's say the network had <code>n<\/code> neurons and we applied dropout rate <code>1\/2<\/code><\/p>\n\n<p><strong>Training phase<\/strong>, we would be left with <code>n\/2<\/code> neurons. So if you were expecting output <code>x<\/code> with all the neurons, now you will get on <code>x\/2<\/code>. So for every batch, the network weights are trained according to this x\/2 <\/p>\n\n<p><strong>Testing\/Inference\/Validation phase<\/strong>, we dont apply any dropout so the output is x. So, in this case, the output would be with x and not x\/2, which would give you the incorrect result. So what you can do is scale it to x\/2 during testing. <\/p>\n\n<p>Rather than the above scaling specific to Testing phase. What Tensorflow's dropout layer does is that whether it is with dropout or without (Training or testing), it scales the output so that the sum is constant.<\/p>\n"},{"AnswerId":"34597667","CreationDate":"2016-01-04T18:41:34.537","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This scaling enables the same network to be used for training (with <code>keep_prob &lt; 1.0<\/code>) and evaluation (with <code>keep_prob == 1.0<\/code>). From the <a href=\"http:\/\/www.cs.toronto.edu\/~rsalakhu\/papers\/srivastava14a.pdf\" rel=\"noreferrer\">Dropout paper<\/a>:<\/p>\n\n<blockquote>\n  <p>The idea is to use a single neural net at test time without dropout. The weights of this network are scaled-down versions of the trained weights. If a unit is retained with probability <em>p<\/em> during training, the outgoing weights of that unit are multiplied by <em>p<\/em> at test time as shown in Figure 2.<\/p>\n<\/blockquote>\n\n<p>Rather than adding ops to scale down the weights by <code>keep_prob<\/code> at test time, the TensorFlow implementation adds an op to scale up the weights by <code>1. \/ keep_prob<\/code> at training time. The effect on performance is negligible, and the code is simpler (because we use the same graph and treat <code>keep_prob<\/code> as a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> that is fed a different value depending on whether we are training or evaluating the network).<\/p>\n"}]}
{"QuestionId":34598010,"AnswerCount":1,"Tags":"<regression><caffe><conv-neural-network><loss><matcaffe>","CreationDate":"2016-01-04T19:01:55.723","AcceptedAnswerId":34665002.0,"OwnerUserId":3698878.0,"Title":"Custom Loss Function Caffe (Spearman Coefficient) while FInetuning (Regression)","Body":"<p>I am finetuning imagenet for a regression problem in caffe. At present I am using Eucliden Loss, but I don't think it is any good in my case.<br>\nI want the loss values to be the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Spearman%27s_rank_correlation_coefficient\" rel=\"nofollow\">spearman coefficient<\/a> between predicted labels and actual labels. How can I do so?  <\/p>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"34665002","CreationDate":"2016-01-07T20:56:51.183","ParentId":null,"OwnerUserId":"3698878","Title":null,"Body":"<p>As cleared in the comment section, since the loss function needs to be differentiable, and spearman coefficient isn't, we can't use it as a loss function.<\/p>\n"}]}
{"QuestionId":34598204,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-04T19:15:15.963","AcceptedAnswerId":34600671.0,"OwnerUserId":4975126.0,"Title":"Check if values are contained within a Tensor","Body":"<p>Unfortunately I was unable to find a function that achieved the following:<\/p>\n\n<p>Inputs:<\/p>\n\n<ul>\n<li>test: Tensor of values that may exist within target<\/li>\n<li>target: Tensor of values<\/li>\n<\/ul>\n\n<p>Outputs:<\/p>\n\n<ul>\n<li>output: Tensor of boolean, same shape as test. <\/li>\n<\/ul>\n\n<hr>\n\n<p>Put plainly: I need to test whether the elements of a Tensor are contained within another Tensor.<\/p>\n","answers":[{"AnswerId":"34600671","CreationDate":"2016-01-04T22:01:36.130","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You want to use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#listdiff\" rel=\"noreferrer\"><code>tf.listdiff<\/code><\/a>.<\/p>\n\n<p>Given two tensors <code>test<\/code> and <code>target<\/code>,<\/p>\n\n<pre><code>c, _ = tf.listdiff(test, target)\n<\/code><\/pre>\n\n<p><code>c<\/code> will contain the items from test that <em>are not<\/em> in the target.  You can then listdiff it again to find the ones that are in there:<\/p>\n\n<pre><code>output, _ = tf.listdiff(test, c)\n<\/code><\/pre>\n\n<p><code>output<\/code> contains items from <code>test<\/code> that were successfully found in the <code>target<\/code> tensor.<\/p>\n"}]}
{"QuestionId":34598371,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-01-04T19:26:24.400","AcceptedAnswerId":34599220.0,"OwnerUserId":4975126.0,"Title":"Count number of \"True\" values in boolean Tensor","Body":"<p>I understand that  will return the locations of  values, so that I could use the result's  to get the number of s. <\/p>\n\n<p>However, when I try and use this, the dimension is unknown (which makes sense as it needs to be computed at runtime). So my question is, how can I access a dimension and use it in an operation like a sum?<\/p>\n\n<p>For example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34599220","CreationDate":"2016-01-04T20:23:49.417","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>You can cast the values to floats and compute the sum on them:\n<code>tf.reduce_sum(tf.cast(myOtherTensor, tf.float32))<\/code><\/p>\n\n<p>Depending on your actual use case you can also compute sums per row\/column if you specify the reduce dimensions of the call.<\/p>\n"},{"AnswerId":"34604457","CreationDate":"2016-01-05T05:00:11.740","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Rafal's answer is almost certainly the simplest way to count the number of <code>true<\/code> elements in your tensor, but the other part of your question asked:<\/p>\n\n<blockquote>\n  <p>[H]ow can I access a dimension and use it in an operation like a sum?<\/p>\n<\/blockquote>\n\n<p>To do this, you can use TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#shapes_and_shaping\" rel=\"noreferrer\">shape-related operations<\/a>, which act on the runtime value of the tensor. For example, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#size\" rel=\"noreferrer\"><code>tf.size(t)<\/code><\/a> produces a scalar <code>Tensor<\/code> containing the number of elements in <code>t<\/code>, and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#shape\" rel=\"noreferrer\"><code>tf.shape(t)<\/code><\/a> produces a 1D <code>Tensor<\/code> containing the size of <code>t<\/code> in each dimension.<\/p>\n\n<p>Using these operators, your program could also be written as:<\/p>\n\n<pre><code>myOtherTensor = tf.constant([[True, True], [False, True]])\nmyTensor = tf.where(myOtherTensor)\ncountTrue = tf.shape(myTensor)[0]  # Size of `myTensor` in the 0th dimension.\n\nsess = tf.Session()\nsum = sess.run(countTrue)\n<\/code><\/pre>\n"},{"AnswerId":"46845155","CreationDate":"2017-10-20T08:27:44.913","ParentId":null,"OwnerUserId":"6538933","Title":null,"Body":"<p>There is a tensorflow function to count non-zero values <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/count_nonzero\" rel=\"nofollow noreferrer\"><code>tf.count_nonzero<\/code><\/a>. The function also accepts an <code>axis<\/code> and <code>keep_dims<\/code> arguments.<\/p>\n\n<p>Here is a simple example:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\na = tf.constant(np.random.random(100))\nwith tf.Session() as sess:\n    print(sess.run(tf.count_nonzero(tf.greater(a, 0.5))))\n<\/code><\/pre>\n"},{"AnswerId":"47324693","CreationDate":"2017-11-16T08:18:35.047","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>I think this is the easiest way to do it: <\/p>\n\n<pre><code>In [38]: myOtherTensor = tf.constant([[True, True], [False, True]])\n\nIn [39]: if_true = tf.count_nonzero(myOtherTensor)\n\nIn [40]: sess.run(if_true)\nOut[40]: 3\n<\/code><\/pre>\n"}]}
{"QuestionId":34600941,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2016-01-04T22:21:47.233","AcceptedAnswerId":null,"OwnerUserId":4112409.0,"Title":"Tutorial: theano.grad() function not calculating gradient","Body":"<p>I have problem running a simple linear regression using Theano grad() function. It is a very simple example based on the following tutorial:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/1_linear_regression.py\" rel=\"nofollow\">https:\/\/github.com\/Newmu\/Theano-Tutorials\/blob\/master\/1_linear_regression.py<\/a><\/p>\n\n<p>which I found a little too easy and wanted to add additional input variable (and therefore made weights an array of 2 variables). However, I cannot make grad() work as it does not converge and both W1 shared variable (weights matrix) and cost function stay the same.<\/p>\n\n<p>Basically the example should take input matrix 5,2 (5 samples, each having 2 features) and create weight matrix (1,2) given the output vector (5 samples).<\/p>\n\n<p><strong>EDIT: (I forgot to add updates=update in grad())<\/strong><\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34600996","CreationDate":"2016-01-04T22:26:18.257","ParentId":null,"OwnerUserId":"4112409","Title":null,"Body":"<p>Ok, stupid mistake: I forgot to set updates=update in grad() invocation.<\/p>\n"}]}
{"QuestionId":34611584,"AnswerCount":1,"Tags":"<python><image-processing><neural-network><deep-learning><caffe>","CreationDate":"2016-01-05T12:20:25.357","AcceptedAnswerId":34611744.0,"OwnerUserId":2527680.0,"Title":"caffe resize image make all values 0","Body":"<p>I encountered this problem when working with caffe to generate HDF5 data files from images.<\/p>\n\n<p>the  loads images into a varible in the normalized range 0-1.<\/p>\n\n<p><strong>resizes but all values in  converted to zero<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>the output I get is <\/p>\n\n<pre><\/pre>\n\n<p>the values of the  is  changed for some values to all 0's\ncould any one help me fix this.<\/p>\n\n<p>I would like the same float32 changing the order of the  command gives the correct output\n<strong>Gives real non zero values to var <\/strong>\n<strong>but order and type is not one I need<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>output <\/p>\n\n<pre><\/pre>\n\n<p><strong>I need the shape to be 3,58,58 without converting all values to zeros ie real data<\/strong> <\/p>\n","answers":[{"AnswerId":"34611744","CreationDate":"2016-01-05T12:28:38.500","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h3>What went wrong?<\/h3>\n\n<p>your <code>img<\/code> is a 3-D array, but there is a different meaning to the different dimensions. The first two <code>(240,320)<\/code> are <em>spatial<\/em> dimensions while the third <code>(3)<\/code> is the \"channel\" (or \"feature\") dimension. When <code>resize<\/code>-ing an image you want to change the <em>spatial<\/em> dimensions and even mix (interpolate) neighboring values, but you always want to keep the channels intact: you do not want to mix \"red\" values in order to decide on \"green\" values for neighboring pixels.<\/p>\n\n<p>In caffe, in order to take an input image of shape <code>(240,320,3)<\/code> and transform it into shape <code>(3,58,58)<\/code> you need to do <strong>two<\/strong> things:<br>\n 1. Change the <em>spatial<\/em> dimensions from <code>(240,320)<\/code> to <code>(58,58)<\/code> using <code>caffe.io.resize_image<\/code>. This stage will result with <code>img<\/code> of shape <code>(58,58,3)<\/code>.<br>\n 2. The second stage is to shift\/<code>transpose<\/code> the dimensions of the <code>img<\/code> so that the channels dimension will be the first: <code>img.transpose((2,0,1))<\/code>.<\/p>\n\n<h3>Solution:<\/h3>\n\n<pre><code># first, resize the spatial dimensions, do not touch the channels\nimg = caffe.io.resize_image( img, (SIZE,SIZE), interp_order=3 )\n# transpose the dimensions from H-W-C to C-H-W\nimg = img.transpose( (2,0,1) )\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L305\" rel=\"nofollow\"><code>resize_image<\/code><\/a> for more details.<\/p>\n\n<p>PS,<br>\nIt is also very common to change RGB image to BGR. You skipped this stage. Is that intentional?<\/p>\n"}]}
{"QuestionId":34612786,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-01-05T13:23:15.690","AcceptedAnswerId":34623864.0,"OwnerUserId":2835670.0,"Title":"How to clamp small values to zero in theano","Body":"<p>I know the theano clip function, but what I want to do is different: I have a vector and if an entry is small, but not zero, I want to make it zero. Entries greater than the given threshold will remain unchanged.\nIs there a way to do that in theano?<\/p>\n","answers":[{"AnswerId":"34623864","CreationDate":"2016-01-06T00:46:35.650","ParentId":null,"OwnerUserId":"557478","Title":null,"Body":"<p>You can clip the values of a vector below a specific threshold to zero with the following code snippet:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\nx = T.ivector('x')\nthreshold = 5  # change accordingly\nx_clipped = x * (x &gt; threshold)\n\nf = theano.function(inputs=[x], outputs=x_clipped)\n\nprint(f([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]))\n<\/code><\/pre>\n"}]}
{"QuestionId":34613261,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><tensorflow><tensorboard>","CreationDate":"2016-01-05T13:48:00.723","AcceptedAnswerId":null,"OwnerUserId":5743410.0,"Title":"Cost and activation functions for multiple independent labels","Body":"<p>After completing the mnist\/cifar tutorials I thought I'd experiment with tensorflow by making my own 'large' set of data and for simplicity's sake, I settled on a black-on-white oval shape that changes its height and width independently on a 0.0-1.0 scale as a 28x28 pixel image (of which I have 5000 training images, 1000 test images).  <\/p>\n\n<p>My code uses the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow noreferrer\">'MNIST expert'<\/a> tutorial as a basis (scaled back for speed) but I switched in a squared-error based cost function and, based on advice from <a href=\"https:\/\/stackoverflow.com\/questions\/34229140\/choosing-from-different-cost-function-and-activation-function-of-a-neural-networ\">here<\/a>, swapped in a sigmoid function for the final activation layer, given that this isn't a classification but rather a 'best-fit' between two tensors, y_ and y_conv.<\/p>\n\n<p>However over the course of >100k iterations, the loss output quickly settles into an oscilation between 400 and 900 (or, consequently, 0.2-0.3 of any given label averaged over 2 labels in a batch of 50) so I imagine I'm just getting noise.  Perhaps I'm mistaken but I was hoping to use Tensorflow to convolve images so as to deduce maybe 10 or more independent labelled variables.  Am I missing something fundamental here? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'm most concerned over how tensor board seems to show that the weights are barely changing whatsoever, even after hours and hours of training  and a decaying learning rate (although not shown in the code).  My understanding of machine learning is that, when convolving images, the layers effectively amounts to layers of edge detection.....so I'm confused as to why they should barely change.<\/p>\n\n<p>My theories currently are:<br>\n1. I've overlooked\/misunderstood something regarding the loss function.<br>\n2. I've misunderstood how weights are initialized\/updated<br>\n3. I've grossly underestimated how long the process should take...although,   again, the loss seems to simply be oscillating.<\/p>\n\n<p>Any help would be greatly appreciated, thanks!<\/p>\n","answers":[{"AnswerId":"35929955","CreationDate":"2016-03-11T00:18:22.290","ParentId":null,"OwnerUserId":"2642845","Title":null,"Body":"<p>From what I can see, your cost function is not the usual mean square error.<br>\nYou are optimizing <code>tf.reduce_sum(tf.abs(tf.sub(y_,y_conv)))<\/code> squared. This function is not differentiable in 0 (it is the square of the l1 norm). This can cause some stability issues (especially in the back propagation steps, I don't know what kind of sub gradient they use in this case).  <\/p>\n\n<p>The usual mean square error can be written as <\/p>\n\n<pre><code>residual = tf.sub(y_, y_conv)\nerror = tf.reduce_mean(tf.reduce_sum(residual*residual, reduction_indices=[1]))\n<\/code><\/pre>\n\n<p>(the use of mean and sum is to avoid having value dependent on the batch size). This is differentiable and should give you better behavior.<\/p>\n"}]}
{"QuestionId":34618207,"AnswerCount":3,"Tags":"<python><set><tensorflow>","CreationDate":"2016-01-05T17:54:49.400","AcceptedAnswerId":34618648.0,"OwnerUserId":5451067.0,"Title":"Determining if A Value is in a Set in TensorFlow","Body":"<p>The , , and  functions are very useful.<\/p>\n\n<p>However, suppose you have value , and you wanted to see if it was in a . In python you would simply write:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I can tell, the only way to do this in TensorFlow, is to have nested 'tf.logical_or' along with 'tf.equal'. I provided just one iteration of this concept below:<\/p>\n\n<pre><\/pre>\n\n<p>I feel that there must be an easier way to do this in TensorFlow. Is there?<\/p>\n","answers":[{"AnswerId":"34618648","CreationDate":"2016-01-05T18:22:02.907","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>Take a look at this related question: <a href=\"https:\/\/stackoverflow.com\/questions\/34598371\/count-number-of-true-values-in-boolean-tensor\/\">Count number of &quot;True&quot; values in boolean Tensor<\/a><\/p>\n\n<p>You should be able to build a tensor consisting of [a, b, c, d, e] and then check if any of the rows is equal to x using <code>tf.equal(.)<\/code><\/p>\n"},{"AnswerId":"48117911","CreationDate":"2018-01-05T16:50:49.407","ParentId":null,"OwnerUserId":"4121413","Title":null,"Body":"<p>To provide a more concrete answer, say you want to check whether the last dimension of the tensor <code>x<\/code> contains any value from a 1D tensor <code>s<\/code>, you could do the following:<\/p>\n\n<pre><code>tile_multiples = tf.concat([tf.ones(tf.shape(tf.shape(x)), dtype=tf.int32), tf.shape(s)], axis=0)\nx_tile = tf.tile(tf.expand_dims(x, -1), tile_multiples)\nx_in_s = tf.reduce_any(tf.equal(x_tile, s), -1))\n<\/code><\/pre>\n\n<p>For example, for <code>s<\/code> and <code>x<\/code>: <\/p>\n\n<pre><code>s = tf.constant([3, 4])\nx = tf.constant([[[1, 2, 3, 0, 0], \n                  [4, 4, 4, 0, 0]], \n                 [[3, 5, 5, 6, 4], \n                  [4, 7, 3, 8, 9]]])\n<\/code><\/pre>\n\n<p><code>x<\/code> has shape <code>[2, 2, 5]<\/code> and <code>s<\/code> has shape <code>[2]<\/code> so <code>tile_multiples = [1, 1, 1, 2]<\/code>, meaning we will tile the last dimension of <code>x<\/code> 2 times (once for each element in <code>s<\/code>) along a new dimension. So, <code>x_tile<\/code> will look like:<\/p>\n\n<pre><code>[[[[1 1]\n   [2 2]\n   [3 3]\n   [0 0]\n   [0 0]]\n\n  [[4 4]\n   [4 4]\n   [4 4]\n   [0 0]\n   [0 0]]]\n\n [[[3 3]\n   [5 5]\n   [5 5]\n   [6 6]\n   [4 4]]\n\n  [[4 4]\n   [7 7]\n   [3 3]\n   [8 8]\n   [9 9]]]]\n<\/code><\/pre>\n\n<p>and <code>x_in_s<\/code> will compare each of the tiled values to one of the values in <code>s<\/code>. <code>tf.reduce_any<\/code> along the last dim will return true if any of the tiled values was in <code>s<\/code>, giving the final result:<\/p>\n\n<pre><code>[[[False False  True False False]\n  [ True  True  True False False]]\n\n [[ True False False False  True]\n  [ True False  True False False]]]\n<\/code><\/pre>\n"},{"AnswerId":"58469142","CreationDate":"2019-10-20T00:36:27.250","ParentId":null,"OwnerUserId":"304926","Title":null,"Body":"<p>Here's two solutions, we want to check if <code>query<\/code> is in <code>whitelist<\/code><\/p>\n\n<pre><code>whitelist = tf.constant([\"CUISINE\", \"DISH\", \"RESTAURANT\", \"ADDRESS\"])\nquery = \"RESTAURANT\"\n\n#use broadcasting for element-wise tensor operation\nbroadcast_equal = tf.equal(whitelist, query)\n\n#method 1: using tensor ops\nbroadcast_equal_int = tf.cast(broadcast_equal, tf.int8)\nbroadcast_sum = tf.reduce_sum(broadcast_equal_int)\n\n#method 2: using some tf.core API\nnz_cnt = tf.count_nonzero(broadcast_equal)\n\nsess.run([broadcast_equal, broadcast_sum, nz_cnt])\n#=&gt; [array([False, False,  True, False]), 1, 1]\n<\/code><\/pre>\n\n<p>So if the output is <code>&gt; 0<\/code> then the item is in the set.<\/p>\n"}]}
{"QuestionId":34619177,"AnswerCount":5,"Tags":"<neural-network><tensorflow>","CreationDate":"2016-01-05T18:51:56.717","AcceptedAnswerId":44103248.0,"OwnerUserId":4341948.0,"Title":"What does tf.nn.conv2d do in tensorflow?","Body":"<p>I was looking at the docs of tensorflow about  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"noreferrer\">here<\/a>. But I can't understand what it does or what it is trying to achieve. It says on the docs, <\/p>\n\n<blockquote>\n  <p>#1 : Flattens the filter to a 2-D matrix with shape <\/p>\n  \n  <p>.<\/p>\n<\/blockquote>\n\n<p>Now what does that do? Is that element-wise multiplication or just plain matrix multiplication? I also could not understand the other two points mentioned in the docs. I have written them below :<\/p>\n\n<blockquote>\n  <p># 2: Extracts image patches from the the input tensor to form a virtual tensor of shape <\/p>\n  \n  <p>.<\/p>\n  \n  <p># 3: For each patch, right-multiplies the filter matrix and the image patch vector.<\/p>\n<\/blockquote>\n\n<p>It would be really helpful if anyone could give an example, a piece of code (extremely helpful) maybe and explain what is going on there and why the operation is like this.<\/p>\n\n<p>I've tried coding a small portion and printing out the shape of the operation. Still, I can't understand. <\/p>\n\n<p>I tried something like this:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I understand bits and pieces of convolutional neural networks. I studied them <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"noreferrer\">here<\/a>. But the implementation on tensorflow is not what I expected. So it raised the question.<\/p>\n\n<p><strong>EDIT<\/strong>:\nSo, I implemented a much simpler code. But I can't figure out what's going on. I mean how the results are like this. It would be extremely helpful if anyone could tell me what process yields this output.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>output<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"50709573","CreationDate":"2018-06-05T22:02:59.797","ParentId":null,"OwnerUserId":"8190089","Title":null,"Body":"<p>In addition to other answers, conv2d operation is operating in c++ (cpu) or cuda for gpu machines that requires to flatten and reshape data in certain way and use gemmBLAS or cuBLAS(cuda) matrix multiplication.<\/p>\n"},{"AnswerId":"46454888","CreationDate":"2017-09-27T18:32:08.180","ParentId":null,"OwnerUserId":"8685078","Title":null,"Body":"<p>I tried to implement conv2d (for my studying). Well, I wrote that:<\/p>\n\n<pre><code>def conv(ix, w):\n   # filter shape: [filter_height, filter_width, in_channels, out_channels]\n   # flatten filters\n   filter_height = int(w.shape[0])\n   filter_width = int(w.shape[1])\n   in_channels = int(w.shape[2])\n   out_channels = int(w.shape[3])\n   ix_height = int(ix.shape[1])\n   ix_width = int(ix.shape[2])\n   ix_channels = int(ix.shape[3])\n   filter_shape = [filter_height, filter_width, in_channels, out_channels]\n   flat_w = tf.reshape(w, [filter_height * filter_width * in_channels, out_channels])\n   patches = tf.extract_image_patches(\n       ix,\n       ksizes=[1, filter_height, filter_width, 1],\n       strides=[1, 1, 1, 1],\n       rates=[1, 1, 1, 1],\n       padding='SAME'\n   )\n   patches_reshaped = tf.reshape(patches, [-1, ix_height, ix_width, filter_height * filter_width * ix_channels])\n   feature_maps = []\n   for i in range(out_channels):\n       feature_map = tf.reduce_sum(tf.multiply(flat_w[:, i], patches_reshaped), axis=3, keep_dims=True)\n       feature_maps.append(feature_map)\n   features = tf.concat(feature_maps, axis=3)\n   return features\n<\/code><\/pre>\n\n<p>Hope I did it properly. Checked on MNIST, had very close results (but this implementation is slower). I hope this helps you.<\/p>\n"},{"AnswerId":"45079451","CreationDate":"2017-07-13T11:35:31.523","ParentId":null,"OwnerUserId":"7378751","Title":null,"Body":"<p>Just to add to the other answers, you should think of the parameters in <\/p>\n\n<pre><code>filter = tf.Variable(tf.random_normal([3,3,5,7]))\n<\/code><\/pre>\n\n<p>as '5' corresponding to the number of channels in each filter. Each filter is a 3d cube, with a depth of 5. Your filter depth must correspond to your input image's depth. The last parameter, 7, should be thought of as the number of filters in the batch. Just forget about this being 4D, and instead imagine that you have a set or a batch of 7 filters. What you do is create 7 filter cubes with dimensions (3,3,5).<\/p>\n\n<p>It is a lot easier to visualize in the Fourier domain since convolution becomes point-wise multiplication. For an input image of dimensions (100,100,3) you can rewrite the filter dimensions as<\/p>\n\n<pre><code>filter = tf.Variable(tf.random_normal([100,100,3,7]))\n<\/code><\/pre>\n\n<p>In order to obtain one of the 7 output feature maps, we simply perform the point-wise multiplication of the filter cube with the image cube, then we sum the results across the channels\/depth dimension (here it's 3), collapsing to a 2d (100,100) feature map. Do this with each filter cube, and you get 7 2D feature maps.<\/p>\n"},{"AnswerId":"44103248","CreationDate":"2017-05-22T00:59:24.497","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>2D convolution is computed in a similar way one would calculate <a href=\"http:\/\/www.riptutorial.com\/tensorflow\/example\/30750\/math-behind-1d-convolution-with-advanced-examples-in-tf\" rel=\"noreferrer\">1D convolution<\/a>: you slide your kernel over the input, calculate the element-wise multiplications and sum them up. But instead of your kernel\/input being an array, here they are matrices.<\/p>\n\n<hr>\n\n<p>In the most basic example there is no padding and stride=1. Let's assume your <code>input<\/code> and <code>kernel<\/code> are:\n<a href=\"https:\/\/i.stack.imgur.com\/yTCl8.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yTCl8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When you use your kernel you will receive the following output: <a href=\"https:\/\/i.stack.imgur.com\/TPhBi.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TPhBi.png\" alt=\"enter image description here\"><\/a>, which is calculated in the following way:<\/p>\n\n<ul>\n<li>14 = 4 * 1 + 3 * 0 + 1 * 1 + 2 * 2 + 1 * 1 + 0 * 0 + 1 * 0 + 2 * 0 + 4 * 1<\/li>\n<li>6  = 3 * 1 + 1 * 0 + 0 * 1 + 1 * 2 + 0 * 1 + 1 * 0 + 2 * 0 + 4 * 0 + 1 * 1<\/li>\n<li>6  = 2 * 1 + 1 * 0 + 0 * 1 + 1 * 2 + 2 * 1 + 4 * 0 + 3 * 0 + 1 * 0 + 0 * 1<\/li>\n<li>12 = 1 * 1 + 0 * 0 + 1 * 1 + 2 * 2 + 4 * 1 + 1 * 0 + 1 * 0 + 0 * 0 + 2 * 1<\/li>\n<\/ul>\n\n<p>TF's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"noreferrer\">conv2d<\/a> function calculates convolutions in batches and uses a slightly different format. For an input it is <code>[batch, in_height, in_width, in_channels]<\/code> for the kernel it is <code>[filter_height, filter_width, in_channels, out_channels]<\/code>. So we need to provide the data in the correct format:<\/p>\n\n<pre><code>import tensorflow as tf\nk = tf.constant([\n    [1, 0, 1],\n    [2, 1, 0],\n    [0, 0, 1]\n], dtype=tf.float32, name='k')\ni = tf.constant([\n    [4, 3, 1, 0],\n    [2, 1, 0, 1],\n    [1, 2, 4, 1],\n    [3, 1, 0, 2]\n], dtype=tf.float32, name='i')\nkernel = tf.reshape(k, [3, 3, 1, 1], name='kernel')\nimage  = tf.reshape(i, [1, 4, 4, 1], name='image')\n<\/code><\/pre>\n\n<p>Afterwards the convolution is computed with:<\/p>\n\n<pre><code>res = tf.squeeze(tf.nn.conv2d(image, kernel, [1, 1, 1, 1], \"VALID\"))\n# VALID means no padding\nwith tf.Session() as sess:\n   print sess.run(res)\n<\/code><\/pre>\n\n<p>And will be equivalent to the one we calculated by hand. <\/p>\n\n<hr>\n\n<p>For <a href=\"http:\/\/www.riptutorial.com\/tensorflow\/example\/30755\/some-padding--strides-1\" rel=\"noreferrer\">examples with padding\/strides, take a look here<\/a>.<\/p>\n"},{"AnswerId":"34698115","CreationDate":"2016-01-09T19:45:15.640","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>Ok I think this is about the simplest way to explain it all.<\/p>\n\n<hr>\n\n<p>Your example is 1 image, size 2x2, with 1 channel. You have 1 filter, with size 1x1, and 1 channel (size is height x width x channels x number of filters). <\/p>\n\n<p>For this simple case the resulting 2x2, 1 channel image (size 1x2x2x1, number of images x height x width x x channels) is the result of multiplying the filter value by each pixel of the image.<\/p>\n\n<hr>\n\n<p>Now let's try more channels:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,3,3,5]))\nfilter = tf.Variable(tf.random_normal([1,1,5,1]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='VALID')\n<\/code><\/pre>\n\n<p>Here the 3x3 image and the 1x1 filter each have 5 channels. The resulting image will be 3x3 with 1 channel (size 1x3x3x1), where the value of each pixel is the dot product across channels of the filter with the corresponding pixel in the input image.<\/p>\n\n<hr>\n\n<p>Now with a 3x3 filter<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,3,3,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,1]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='VALID')\n<\/code><\/pre>\n\n<p>Here we get a 1x1 image, with 1 channel (size 1x1x1x1). The value is the sum of the 9, 5-element dot products. But you could just call this a 45-element dot product.<\/p>\n\n<hr>\n\n<p>Now with a bigger image<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,5,5,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,1]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='VALID')\n<\/code><\/pre>\n\n<p>The output is a 3x3 1-channel image (size 1x3x3x1). \nEach of these values is a sum of 9, 5-element dot products. <\/p>\n\n<p>Each output is made by centering the filter on one of the 9 center pixels of the input image, so that none of the filter sticks out. The <code>x<\/code>s below represent the filter centers for each output pixel.<\/p>\n\n<pre><code>.....\n.xxx.\n.xxx.\n.xxx.\n.....\n<\/code><\/pre>\n\n<hr>\n\n<p>Now with \"SAME\" padding:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,5,5,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,1]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>This gives a 5x5 output image (size 1x5x5x1). This is done by centering the filter at each position on the image. <\/p>\n\n<p>Any of the 5-element dot products where the filter sticks out past the edge of the image get a value of zero. <\/p>\n\n<p>So the corners are only sums of 4, 5-element dot products.<\/p>\n\n<hr>\n\n<p>Now with multiple filters.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,5,5,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,7]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>This still gives a 5x5 output image, but with 7 channels (size 1x5x5x7). Where each channel is produced by one of the filters in the set.<\/p>\n\n<hr>\n\n<p>Now with strides 2,2:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([1,5,5,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,7]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 2, 2, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>Now the result still has 7 channels, but is only 3x3 (size 1x3x3x7).<\/p>\n\n<p>This is because instead of centering the filters at every point on the image, the filters are centered at every other point on the image, taking steps (strides) of width 2. The <code>x<\/code>'s below represent the filter center for each output pixel, on the input image.<\/p>\n\n<pre><code>x.x.x\n.....\nx.x.x\n.....\nx.x.x\n<\/code><\/pre>\n\n<hr>\n\n<p>And of course the first dimension of the input is the number of images so you can apply it over a batch of 10 images, for example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.Variable(tf.random_normal([10,5,5,5]))\nfilter = tf.Variable(tf.random_normal([3,3,5,7]))\n\nop = tf.nn.conv2d(input, filter, strides=[1, 2, 2, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>This performs the same operation, for each image independently, giving a stack of 10 images as the result (size 10x3x3x7)  <\/p>\n"}]}
{"QuestionId":34622460,"AnswerCount":1,"Tags":"<python><machine-learning><regression><theano><theano-cuda>","CreationDate":"2016-01-05T22:32:54.737","AcceptedAnswerId":34623724.0,"OwnerUserId":4112409.0,"Title":"Theano simple linear regression runs on CPU instead of GPU","Body":"<p>I created a simple python script (using Theano) performing linear regression which should be run on GPU. When code starts it says \"using gpu device\", but (according to the profiler) all operations are CPU-specific (ElemWise, instead of GpuElemWise, no GpuFromHost etc.).<\/p>\n\n<p>I checked the variables, THEANO_FLAGS, everything seems right and I cannot see the catch (especially when Theano tutorials with the same settings are correctly run on GPU :)).<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <ul>\n  <li>THEANO_FLAGS=cuda.root=\/usr\/local\/cuda<\/li>\n  <li>device=gpu<\/li>\n  <li>floatX=float32<\/li>\n  <li>lib.cnmem=.5<\/li>\n  <li>profile=True<\/li>\n  <li>CUDA_LAUNCH_BLOCKING=1<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34623724","CreationDate":"2016-01-06T00:29:42.393","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>As mentioned in the comments although you have set the <code>allow_input_downcast<\/code> parameter to <code>True<\/code>, but you need to make sure all the data to be assigned to shared variables are in <code>float32<\/code>. As of <strong>Jan. 06, 2016<\/strong> Theano still <strong>cannot<\/strong> work with any other data type rather than <code>float32<\/code> to do the computations on GPU, as mentioned <a href=\"https:\/\/stackoverflow.com\/a\/34520841\/2838606\">here<\/a> in a more details. So you have to have to cast your data into 'float32' format.<\/p>\n\n<p>Therefore, here should be the code you need to use:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as T\n\n\ninput_data = numpy.matrix([[28, 1], [35, 2], [18, 1], [56, 2], [80, 3]])\noutput_data = numpy.matrix([1600, 2100, 1400, 2500, 3200])\n\nTS = theano.shared(input_data.astype('float32'), \"training-set\")\nE = theano.shared(output_data.astype('float32'), \"expected\")\nW1 = theano.shared(numpy.zeros((1, 2), dtype = 'float32'))\n\nO = T.dot(TS, W1.T)\ncost = T.mean(T.sqr(E - O.T))\ngradient = T.grad(cost=cost, wrt=W1)\nupdate = [[W1, W1 - gradient * 0.0001]]\ntrain = theano.function([], cost, updates=update, allow_input_downcast=True, profile = True)\n\nfor i in range(1000):\n    train()\n\ntrain.profile.print_summary()\n<\/code><\/pre>\n\n<p>And here will be the profiling result:<\/p>\n\n<pre><code>Message: LearnTheano.py:18\n  Time in 1000 calls to Function.__call__: 2.642968e-01s\n  Time in Function.fn.__call__: 2.460811e-01s (93.108%)\n  Time in thunks: 1.877530e-01s (71.039%)\n  Total compile time: 2.483290e+01s\n    Number of Apply nodes: 17\n    Theano Optimizer time: 2.818849e-01s\n       Theano validate time: 3.435850e-03s\n    Theano Linker time (includes C, CUDA code generation\/compiling): 2.453926e+01s\n       Import time 1.241469e-02s\n\nTime in all call to theano.grad() 1.206994e-02s\nClass\n---\n&lt;% time&gt; &lt;sum %&gt; &lt;apply time&gt; &lt;time per call&gt; &lt;type&gt; &lt;#call&gt; &lt;#apply&gt; &lt;Class name&gt;\n  34.8%    34.8%       0.065s       3.27e-05s     C     2000       2   theano.sandbox.cuda.blas.GpuGemm\n  28.8%    63.5%       0.054s       1.80e-05s     C     3000       3   theano.sandbox.cuda.basic_ops.GpuElemwise\n  12.9%    76.4%       0.024s       2.42e-05s     C     1000       1   theano.sandbox.cuda.basic_ops.GpuCAReduce\n  10.3%    86.7%       0.019s       1.93e-05s     C     1000       1   theano.sandbox.cuda.basic_ops.GpuFromHost\n   7.2%    93.9%       0.014s       1.36e-05s     C     1000       1   theano.sandbox.cuda.basic_ops.HostFromGpu\n   1.8%    95.7%       0.003s       1.13e-06s     C     3000       3   theano.sandbox.cuda.basic_ops.GpuDimShuffle\n   1.5%    97.2%       0.003s       2.81e-06s     C     1000       1   theano.tensor.elemwise.Elemwise\n   1.1%    98.4%       0.002s       1.08e-06s     C     2000       2   theano.compile.ops.Shape_i\n   1.1%    99.5%       0.002s       1.02e-06s     C     2000       2   theano.sandbox.cuda.basic_ops.GpuSubtensor\n   0.5%   100.0%       0.001s       9.96e-07s     C     1000       1   theano.tensor.opt.MakeVector\n   ... (remaining 0 Classes account for   0.00%(0.00s) of the runtime)\n\nOps\n---\n&lt;% time&gt; &lt;sum %&gt; &lt;apply time&gt; &lt;time per call&gt; &lt;type&gt; &lt;#call&gt; &lt;#apply&gt; &lt;Op name&gt;\n  25.3%    25.3%       0.047s       4.74e-05s     C     1000        1   GpuGemm{no_inplace}\n  12.9%    38.1%       0.024s       2.42e-05s     C     1000        1   GpuCAReduce{pre=sqr,red=add}{1,1}\n  12.8%    51.0%       0.024s       2.41e-05s     C     1000        1   GpuElemwise{mul,no_inplace}\n  10.3%    61.3%       0.019s       1.93e-05s     C     1000        1   GpuFromHost\n   9.5%    70.8%       0.018s       1.79e-05s     C     1000        1   GpuGemm{inplace}\n   8.2%    79.0%       0.015s       1.55e-05s     C     1000        1   GpuElemwise{Composite{((i0 \/ i1) \/ i2)}}[(0, 0)]\n   7.7%    86.7%       0.014s       1.44e-05s     C     1000        1   GpuElemwise{Composite{((i0 * i1) \/ i2)}}[(0, 1)]\n   7.2%    93.9%       0.014s       1.36e-05s     C     1000        1   HostFromGpu\n   1.5%    95.4%       0.003s       2.81e-06s     C     1000        1   Elemwise{Cast{float32}}\n   1.1%    96.5%       0.002s       1.02e-06s     C     2000        2   GpuSubtensor{int64}\n   1.0%    97.5%       0.002s       9.00e-07s     C     2000        2   GpuDimShuffle{x,x}\n   0.8%    98.3%       0.002s       1.59e-06s     C     1000        1   GpuDimShuffle{1,0}\n   0.7%    99.1%       0.001s       1.38e-06s     C     1000        1   Shape_i{0}\n   0.5%    99.6%       0.001s       9.96e-07s     C     1000        1   MakeVector\n   0.4%   100.0%       0.001s       7.76e-07s     C     1000        1   Shape_i{1}\n   ... (remaining 0 Ops account for   0.00%(0.00s) of the runtime)\n\nApply\n------\n&lt;% time&gt; &lt;sum %&gt; &lt;apply time&gt; &lt;time per call&gt; &lt;#call&gt; &lt;id&gt; &lt;Apply name&gt;\n  25.3%    25.3%       0.047s       4.74e-05s   1000     3   GpuGemm{no_inplace}(expected, TensorConstant{-1.0}, &lt;CudaNdarrayType(float32, matrix)&gt;, GpuDimShuffle{1,0}.0, TensorConstant{1.0})\n  12.9%    38.1%       0.024s       2.42e-05s   1000     5   GpuCAReduce{pre=sqr,red=add}{1,1}(GpuGemm{no_inplace}.0)\n  12.8%    51.0%       0.024s       2.41e-05s   1000    13   GpuElemwise{mul,no_inplace}(GpuDimShuffle{x,x}.0, GpuDimShuffle{x,x}.0)\n  10.3%    61.3%       0.019s       1.93e-05s   1000     7   GpuFromHost(Elemwise{Cast{float32}}.0)\n   9.5%    70.8%       0.018s       1.79e-05s   1000    16   GpuGemm{inplace}(&lt;CudaNdarrayType(float32, matrix)&gt;, TensorConstant{-9.99999974738e-05}, GpuElemwise{Composite{((i0 * i1) \/ i2)}}[(0, 1)].0, training-set, TensorConstant{1.0})\n   8.2%    79.0%       0.015s       1.55e-05s   1000    12   GpuElemwise{Composite{((i0 \/ i1) \/ i2)}}[(0, 0)](GpuCAReduce{pre=sqr,red=add}{1,1}.0, GpuSubtensor{int64}.0, GpuSubtensor{int64}.0)\n   7.7%    86.7%       0.014s       1.44e-05s   1000    15   GpuElemwise{Composite{((i0 * i1) \/ i2)}}[(0, 1)](CudaNdarrayConstant{[[-2.]]}, GpuGemm{no_inplace}.0, GpuElemwise{mul,no_inplace}.0)\n   7.2%    93.9%       0.014s       1.36e-05s   1000    14   HostFromGpu(GpuElemwise{Composite{((i0 \/ i1) \/ i2)}}[(0, 0)].0)\n   1.5%    95.4%       0.003s       2.81e-06s   1000     6   Elemwise{Cast{float32}}(MakeVector.0)\n   0.8%    96.3%       0.002s       1.59e-06s   1000     0   GpuDimShuffle{1,0}(training-set)\n   0.7%    97.0%       0.001s       1.38e-06s   1000     2   Shape_i{0}(expected)\n   0.7%    97.7%       0.001s       1.30e-06s   1000     8   GpuSubtensor{int64}(GpuFromHost.0, Constant{0})\n   0.6%    98.3%       0.001s       1.08e-06s   1000    11   GpuDimShuffle{x,x}(GpuSubtensor{int64}.0)\n   0.5%    98.8%       0.001s       9.96e-07s   1000     4   MakeVector(Shape_i{0}.0, Shape_i{1}.0)\n   0.4%    99.2%       0.001s       7.76e-07s   1000     1   Shape_i{1}(expected)\n   0.4%    99.6%       0.001s       7.40e-07s   1000     9   GpuSubtensor{int64}(GpuFromHost.0, Constant{1})\n   0.4%   100.0%       0.001s       7.25e-07s   1000    10   GpuDimShuffle{x,x}(GpuSubtensor{int64}.0)\n   ... (remaining 0 Apply instances account for 0.00%(0.00s) of the runtime)\n<\/code><\/pre>\n"}]}
{"QuestionId":34626535,"AnswerCount":2,"Tags":"<machine-learning><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2016-01-06T05:55:29.447","AcceptedAnswerId":null,"OwnerUserId":4977638.0,"Title":"How do I reduce 4096-dimensional feature vector to 1024-dimensional vector in CNN Caffemodel?","Body":"<p>I used 16-layers VGGnet to extract features from an image. It outputs a 4096-dimensional feature vector. However, I need a 1024-dimensional vector. How do I further reduce this 4096-vector into 1024-vector? Do I need to add a new layer on top of ?<\/p>\n","answers":[{"AnswerId":"34626953","CreationDate":"2016-01-06T06:28:19.327","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Yes.\u200f\u200f\u200f \u200f\u200f \u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f\u200f<\/p>\n"},{"AnswerId":"35745663","CreationDate":"2016-03-02T11:16:33.600","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Yes, you need to add another layer on top of <code>fc7<\/code>. This is how your last few layers should be like<\/p>\n\n<pre><code>layers {\n  bottom: \"fc7\"\n  top: \"fc7\"\n  name: \"relu7\"\n  type: RELU\n}\nlayers {\n  bottom: \"fc7\"\n  top: \"fc7\"\n  name: \"drop7\"\n  type: DROPOUT\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayers {\n  name: \"fc8\"\n  bottom: \"fc7\"\n  top: \"fc8\"\n  type: INNER_PRODUCT\n  inner_product_param {\n    num_output: 1024\n  }\n  blobs_lr: 0\n  blobs_lr: 0\n}\nlayers {\n  name: \"loss\"\n  type: SOFTMAX_LOSS\n  bottom: \"fc8\"\n  bottom: \"label\"\n  top: \"loss\/loss\"\n}\nlayers {\n  name: \"accuracy\/top1\"\n  type: ACCURACY\n  bottom: \"fc8\"\n  bottom: \"label\"\n  top: \"accuracy@1\"\n  include: { phase: TEST }\n  accuracy_param {\n    top_k: 1\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":34629613,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-01-06T09:29:36.837","AcceptedAnswerId":null,"OwnerUserId":3632556.0,"Title":"How to profile TensorFlow networks?","Body":"<p>Is there a way to calculate the time cost for each node in a TensorFlow network?<br>\nI find it hard to locate the performance bottlenecks.<\/p>\n\n<p>EDIT: The  profiler is really awesome (<a href=\"https:\/\/stackoverflow.com\/a\/37774470\/3632556\">https:\/\/stackoverflow.com\/a\/37774470\/3632556<\/a>).<\/p>\n","answers":[{"AnswerId":"35902269","CreationDate":"2016-03-09T20:48:35.703","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>Right now the best external way is to compile with a CPU\/GPU profiler, but this has to be done manually by changing options in the <code>BUILD<\/code> and <code>tensorflow.bzl<\/code> files (not sure where).  <\/p>\n\n<p>You can then get profile information about which functions are taking the most time, etc.  It's up to you, you can run by using <a href=\"https:\/\/github.com\/gperftools\/gperftools\" rel=\"nofollow noreferrer\">gperftools<\/a> and visualizing by <a href=\"https:\/\/github.com\/google\/pprof\" rel=\"nofollow noreferrer\">pprof<\/a><\/p>\n\n<p>Since TensorFlow's code is linked via <code>swig<\/code>, i'm not 100% sure how easy it will be to get the symbols when reading the profile via <code>pprof<\/code>.  You may be able to point it to the swig <code>.so<\/code> file that was built.<\/p>\n\n<p>Give it a try and let us know if it works for you!<\/p>\n"},{"AnswerId":"44379742","CreationDate":"2017-06-06T00:42:35.937","ParentId":null,"OwnerUserId":"7583767","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/profiler\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/profiler<\/a><\/p>\n\n<p>Major Features:<\/p>\n\n<p>Measure model parameters, float operations, tensor shapes.<\/p>\n\n<p>Profile op execution times, requested memory size and device placement.<\/p>\n\n<p>Inspect checkpoint tensors' shapes and their values.<\/p>\n\n<p>Selectively group, filter, account and order ops.<\/p>\n"},{"AnswerId":"43733392","CreationDate":"2017-05-02T08:29:45.623","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>If you want to find how much time was spent on each operation at TF, you can do this in tensorboard using <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/graph_viz#runtime_statistics\" rel=\"noreferrer\">runtime statistics<\/a>. You will need to do something like this (check the full example in the above-mentioned link):<\/p>\n\n<pre><code>run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\nrun_metadata = tf.RunMetadata()\nsess.run(&lt;values_you_want_to_execute&gt;, options=run_options, run_metadata=run_metadata)\nyour_writer.add_run_metadata(run_metadata, 'step%d' % i)\n<\/code><\/pre>\n\n<hr>\n\n<p>Better than just printing it you can see it in tensorboard:<\/p>\n\n<blockquote>\n  <p>Additionally, clicking on a node will display the exact total memory,\n  compute time, and tensor output sizes.<\/p>\n<\/blockquote>\n\n<p>Also now tensorflow has a <a href=\"https:\/\/i.stack.imgur.com\/JdoXP.png\" rel=\"noreferrer\">debugger<\/a>. Here is a tutorial of how to use it.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JdoXP.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JdoXP.png\" alt=\"[Example from link\"><\/a><\/p>\n"}]}
{"QuestionId":34631852,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-06T11:18:15.207","AcceptedAnswerId":34637731.0,"OwnerUserId":628096.0,"Title":"Attach a queue to a numpy array in tensorflow for data fetch instead of files?","Body":"<p>I have read the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html\" rel=\"noreferrer\">CNN Tutorial on the TensorFlow<\/a> and I am trying to use the same model for my project. \nThe problem is now in data reading. I have around 25000 images for training and around 5000 for testing and validation each. The files are in png format and I can read them and convert them into the numpy.ndarray. <\/p>\n\n<p>The CNN example in the tutorials use a queue to fetch the records from the file list provided. I tried to create my own such binary file by reshaping my images into 1-D array and attaching a label value in the front of it. So my data looks like this <\/p>\n\n<pre><\/pre>\n\n<p>The single row of the above array is of length <strong>22501<\/strong> size where the first element is the label.<\/p>\n\n<p>I dumped the file to using pickle and the tried to read from the file using the \ntf.FixedLengthRecordReader to read from the file as <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10_input.py#150\" rel=\"noreferrer\">demonstrated in example<\/a><\/p>\n\n<p>I am doing the same things as given in the <strong>cifar10_input.py<\/strong> to read the binary file and putting them into the record object.<\/p>\n\n<p>Now when I read from the files the labels and the image values are different. I can understand the reason for this to be that pickle dumps the extra information of braces and brackets also in the binary file and they change the fixed length record size. <\/p>\n\n<p>The above example uses the filenames and pass it to a queue to fetch the files and then the queue to read a single record from the file. <\/p>\n\n<p>I want to know if I can pass the numpy array as defined above instead of the filenames to some reader and it can fetch records one by one from that array instead of the files.<\/p>\n","answers":[{"AnswerId":"44209178","CreationDate":"2017-05-26T19:49:21.917","ParentId":null,"OwnerUserId":"796120","Title":null,"Body":"<blockquote>\n  <p>I want to know if I can pass the numpy array as defined above instead\n  of the filenames to some reader and it can fetch records one by one\n  from that array instead of the files.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\"><code>tf.py_func<\/code><\/a>, that wraps a python function and uses it as a TensorFlow operator, might help. Here's an <a href=\"https:\/\/stackoverflow.com\/a\/44208648\/796120\">example<\/a>. <\/p>\n\n<p>However, since you've mentioned that your images are stored in png files, I think the simplest solution would be to replace <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_input.py#L79\" rel=\"nofollow noreferrer\">this<\/a>:<\/p>\n\n<pre><code>reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)\nresult.key, value = reader.read(filename_queue)\n<\/code><\/pre>\n\n<p>with this:<\/p>\n\n<pre><code>result.key, value = tf.WholeFileReader().read(filename_queue))\nvalue = tf.image.decode_jpeg(value)\n<\/code><\/pre>\n"},{"AnswerId":"34637731","CreationDate":"2016-01-06T16:17:26.967","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Probably the easiest way to make your data work with the CNN example code is to make a modified version of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/411f57e291839094108afdaa9c43094f44979eaa\/tensorflow\/models\/image\/cifar10\/cifar10_input.py#L41\"><code>read_cifar10()<\/code><\/a> and use it instead:<\/p>\n\n<ol>\n<li><p>Write out a binary file containing the contents of your numpy array.<\/p>\n\n<pre><code>import numpy as np\nimages_and_labels_array = np.array([[...], ...],  # [[1,12,34,24,53,...,102],\n                                                  #  [12,112,43,24,52,...,98],\n                                                  #  ...]\n                                   dtype=np.uint8)\n\nimages_and_labels_array.tofile(\"\/tmp\/images.bin\")\n<\/code><\/pre>\n\n<p>This file is similar to the format used in CIFAR10 datafiles. You might want to generate multiple files in order to get read parallelism. Note that <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/reference\/generated\/numpy.ndarray.tofile.html\"><code>ndarray.tofile()<\/code><\/a> writes binary data in row-major order with no other metadata; pickling the array will add Python-specific metadata that TensorFlow's parsing routines do not understand.<\/p><\/li>\n<li><p>Write a modified version of <code>read_cifar10()<\/code> that handles your record format.<\/p>\n\n<pre><code>def read_my_data(filename_queue):\n\n  class ImageRecord(object):\n    pass\n  result = ImageRecord()\n\n  # Dimensions of the images in the dataset.\n  label_bytes = 1\n  # Set the following constants as appropriate.\n  result.height = IMAGE_HEIGHT\n  result.width = IMAGE_WIDTH\n  result.depth = IMAGE_DEPTH\n  image_bytes = result.height * result.width * result.depth\n  # Every record consists of a label followed by the image, with a\n  # fixed number of bytes for each.\n  record_bytes = label_bytes + image_bytes\n\n  assert record_bytes == 22501  # Based on your question.\n\n  # Read a record, getting filenames from the filename_queue.  No\n  # header or footer in the binary, so we leave header_bytes\n  # and footer_bytes at their default of 0.\n  reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)\n  result.key, value = reader.read(filename_queue)\n\n  # Convert from a string to a vector of uint8 that is record_bytes long.\n  record_bytes = tf.decode_raw(value, tf.uint8)\n\n  # The first bytes represent the label, which we convert from uint8-&gt;int32.\n  result.label = tf.cast(\n      tf.slice(record_bytes, [0], [label_bytes]), tf.int32)\n\n  # The remaining bytes after the label represent the image, which we reshape\n  # from [depth * height * width] to [depth, height, width].\n  depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),\n                           [result.depth, result.height, result.width])\n  # Convert from [depth, height, width] to [height, width, depth].\n  result.uint8image = tf.transpose(depth_major, [1, 2, 0])\n\n  return result\n<\/code><\/pre><\/li>\n<li><p>Modify <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/411f57e291839094108afdaa9c43094f44979eaa\/tensorflow\/models\/image\/cifar10\/cifar10_input.py#L133\"><code>distorted_inputs()<\/code><\/a> to use your new dataset:<\/p>\n\n<pre><code>def distorted_inputs(data_dir, batch_size):\n  \"\"\"[...]\"\"\"\n  filenames = [\"\/tmp\/images.bin\"]  # Or a list of filenames if you\n                                   # generated multiple files in step 1.\n  for f in filenames:\n    if not gfile.Exists(f):\n      raise ValueError('Failed to find file: ' + f)\n\n  # Create a queue that produces the filenames to read.\n  filename_queue = tf.train.string_input_producer(filenames)\n\n  # Read examples from files in the filename queue.\n  read_input = read_my_data(filename_queue)\n  reshaped_image = tf.cast(read_input.uint8image, tf.float32)\n\n  # [...] (Maybe modify other parameters in here depending on your problem.)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>This is intended to be a minimal set of steps, given your starting point. It may be more efficient to do the PNG decoding using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#decode_png\">TensorFlow ops<\/a>, but that would be a larger change.<\/p>\n"},{"AnswerId":"34640057","CreationDate":"2016-01-06T18:22:35.733","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In your question, you specifically asked:<\/p>\n\n<blockquote>\n  <p>I want to know if I can pass the numpy array as defined above instead of the filenames to some reader and it can fetch records one by one from that array instead of the files.<\/p>\n<\/blockquote>\n\n<p>You <em>can<\/em> feed the numpy array to a queue directly, but it will be a more invasive change to the <code>cifar10_input.py<\/code> code than my <a href=\"https:\/\/stackoverflow.com\/a\/34637731\/3574081\">other answer<\/a> suggests.<\/p>\n\n<p>As before, let's assume you have the following array from your question:<\/p>\n\n<pre><code>import numpy as np\nimages_and_labels_array = np.array([[...], ...],  # [[1,12,34,24,53,...,102],\n                                                  #  [12,112,43,24,52,...,98],\n                                                  #  ...]\n                                   dtype=np.uint8)\n<\/code><\/pre>\n\n<p>You can then define a queue that contains the entire data as follows:<\/p>\n\n<pre><code>q = tf.FIFOQueue([tf.uint8, tf.uint8], shapes=[[], [22500]])\nenqueue_op = q.enqueue_many([image_and_labels_array[:, 0], image_and_labels_array[:, 1:]])\n<\/code><\/pre>\n\n<p>...then call <code>sess.run(enqueue_op)<\/code> to populate the queue.<\/p>\n\n<hr>\n\n<p>Another&mdash;more efficient&mdash;approach would be to <em>feed<\/em> records to the queue, which you could do from a parallel thread (see <a href=\"https:\/\/stackoverflow.com\/a\/34596212\/3574081\">this answer<\/a> for more details on how this would work):<\/p>\n\n<pre><code># [With q as defined above.]\nlabel_input = tf.placeholder(tf.uint8, shape=[])\nimage_input = tf.placeholder(tf.uint8, shape=[22500])\n\nenqueue_single_from_feed_op = q.enqueue([label_input, image_input])\n\n# Then, to enqueue a single example `i` from the array.\nsess.run(enqueue_single_from_feed_op,\n         feed_dict={label_input: image_and_labels_array[i, 0],\n                    image_input: image_and_labels_array[i, 1:]})\n<\/code><\/pre>\n\n<p>Alternatively, to enqueue a batch at a time, which will be more efficient:<\/p>\n\n<pre><code>label_batch_input = tf.placeholder(tf.uint8, shape=[None])\nimage_batch_input = tf.placeholder(tf.uint8, shape=[None, 22500])\n\nenqueue_batch_from_feed_op = q.enqueue([label_batch_input, image_batch_input])\n\n# Then, to enqueue a batch examples `i` through `j-1` from the array.\nsess.run(enqueue_single_from_feed_op,\n         feed_dict={label_input: image_and_labels_array[i:j, 0],\n                    image_input: image_and_labels_array[i:j, 1:]})\n<\/code><\/pre>\n"}]}
{"QuestionId":34637573,"AnswerCount":1,"Tags":"<android><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-01-06T16:08:55.173","AcceptedAnswerId":34637842.0,"OwnerUserId":5617366.0,"Title":"Caffe: Check failed: Network should have exactly one output","Body":"<p>I'm using the <a href=\"https:\/\/github.com\/sh1r0\/caffe-android-lib\" rel=\"nofollow\">Caffe android lib<\/a> (based on the C++ api). I'm trying it with a GoogleNet model.<br>\nUnfortunatly it's giving me this error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I'm having the feeling I'm missing something stupid. Any thoughts?<\/p>\n","answers":[{"AnswerId":"34637842","CreationDate":"2016-01-06T16:24:14.583","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>What implementation of GoogLeNet are you using?<br>\nDepending on the implementation, you might indeed get more than a single output.\nThe original GoogLeNet (published in CVPR 2015) had three \"output\" losses for training, so you might be using a similat prototxt resulting with several output layers.<\/p>\n\n<p>Considering <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt\" rel=\"nofollow\">this <code>deploy.prototxt<\/code><\/a> - it has only one output <code>\"prob\"<\/code>. However, the corresponding <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/train_val.prototxt\" rel=\"nofollow\"><code>train_val.prototxt<\/code><\/a> has three outputs: <code>\"loss1\/loss1\"<\/code>, <code>\"loss2\/loss1\"<\/code> and <code>\"loss3\/loss3\"<\/code>.<\/p>\n"}]}
{"QuestionId":34642595,"AnswerCount":4,"Tags":"<python><neural-network><convolution><tensorflow><conv-neural-network>","CreationDate":"2016-01-06T20:56:26.040","AcceptedAnswerId":34643081.0,"OwnerUserId":3908247.0,"Title":"Tensorflow Strides Argument","Body":"<p>I am trying to understand the <strong>strides<\/strong> argument in tf.nn.avg_pool, tf.nn.max_pool, tf.nn.conv2d. <\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#max_pool\" rel=\"noreferrer\">documentation<\/a> repeatedly says <\/p>\n\n<blockquote>\n  <p>strides: A list of ints that has length >= 4. The stride of the sliding window for each dimension of the input tensor.<\/p>\n<\/blockquote>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>What do each of the 4+ integers represent?<\/li>\n<li>Why must they have strides[0] = strides[3] = 1 for convnets?<\/li>\n<li>In <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3%20-%20Neural%20Networks\/convolutional_network.ipynb\" rel=\"noreferrer\">this example<\/a> we see . Why -1?<\/li>\n<\/ol>\n\n<p>Sadly the examples in the docs for reshape using -1 don't translate too well to this scenario.<\/p>\n","answers":[{"AnswerId":"34642991","CreationDate":"2016-01-06T21:20:22.440","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The inputs are 4 dimensional and are of form: <code>[batch_size, image_rows, image_cols, number_of_colors]<\/code><\/p>\n\n<p>Strides, in general, define an overlap between applying operations. In the case of conv2d, it specifies what is the distance between consecutive applications of convolutional filters. The value of 1 in a specific dimension means that we apply the operator at every row\/col, the value of 2 means every second, and so on.<\/p>\n\n<p><strong>Re 1)<\/strong> The values that matter for convolutions are 2nd and 3rd and they represent the overlap in the application of the convolutional filters along rows and columns. The value of [1, 2, 2, 1] says that we want to apply the filters on every second row and column.<\/p>\n\n<p><strong>Re 2)<\/strong> I don't know the technical limitations (might be CuDNN requirement) but typically people use strides along the rows or columns dimensions. It doesn't necessarily make sense to do it over batch size. Not sure of the \nlast dimension.<\/p>\n\n<p><strong>Re 3)<\/strong> Setting -1 for one of the dimension means, \"set the value for the first dimension so that the total number of elements in the tensor is unchanged\". In our case, the -1 will be equal to the batch_size.<\/p>\n"},{"AnswerId":"51624458","CreationDate":"2018-08-01T02:09:35.890","ParentId":null,"OwnerUserId":"3901871","Title":null,"Body":"<p>@dga has done a wonderful job explaining and I can't be thankful enough how helpful it has been. In the like manner, I will like to share my findings on how <code>stride<\/code> works in 3D convolution.<\/p>\n\n<p>According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv3d\" rel=\"nofollow noreferrer\">TensorFlow documentation<\/a> on conv3d, the shape of the input must be in this order: <\/p>\n\n<p><code>[batch, in_depth, in_height, in_width, in_channels]<\/code><\/p>\n\n<p>Let's explain the variables from the extreme right to the left using an example. Assuming the input shape is \n<code>input_shape = [1000,16,112,112,3]<\/code><\/p>\n\n<pre><code>input_shape[4] is the number of colour channels (RGB or whichever format it is extracted in)\ninput_shape[3] is the width of the image\ninput_shape[2] is the height of the image\ninput_shape[1] is the number of frames that have been lumped into 1 complete data\ninput_shape[0] is the number of lumped frames of images we have.\n<\/code><\/pre>\n\n<p>Below is a summary documentation for how stride is used.<\/p>\n\n<blockquote>\n  <p>strides: A list of ints that has length >= 5. 1-D tensor of length 5.\n  The stride of the sliding window for each dimension of input. Must\n  have <code>strides[0] = strides[4] = 1<\/code><\/p>\n<\/blockquote>\n\n<p>As indicated in many works, strides simply mean how many steps away a window or kernel jumps away from the closest element, be it a data frame or pixel (this is paraphrased by the way).<\/p>\n\n<p>From the above documentation, a stride in 3D will look like this strides = (1,<strong>X<\/strong>,<strong>Y<\/strong>,<strong>Z<\/strong>,1). <\/p>\n\n<p>The documentation emphasizes that <code>strides[0] = strides[4] = 1<\/code>. <\/p>\n\n<pre><code>strides[0]=1 means that we do not want to skip any data in the batch \nstrides[4]=1 means that we do not want to skip in the channel \n<\/code><\/pre>\n\n<p>strides[X] means how many skips we should make in the lumped frames. So for example, if we have 16 frames, X=1 means use every frame. X=2 means use every second frame and it goes and on<\/p>\n\n<p>strides[y] and strides[z] follow the explanation by <a href=\"https:\/\/stackoverflow.com\/a\/34643081\/3901871\">@dga<\/a> so I will not redo that part.<\/p>\n\n<p>In keras however, you only need to specify a tuple\/list of 3 integers, specifying the strides of the convolution along each spatial dimension, where spatial dimension is stride[x], strides[y] and strides[z]. strides[0] and strides[4] is already defaulted to 1.<\/p>\n\n<p>I hope someone finds this helpful!<\/p>\n"},{"AnswerId":"44103110","CreationDate":"2017-05-22T00:32:01.347","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<h3>Let's start with what stride does in 1-dim case.<\/h3>\n\n<p>Let's assume your <code>input = [1, 0, 2, 3, 0, 1, 1]<\/code> and <code>kernel = [2, 1, 3]<\/code> the result of the convolution is <code>[8, 11, 7, 9, 4]<\/code>, which is calculated by sliding your kernel over the input, performing element-wise multiplication and summing everything. <a href=\"http:\/\/www.riptutorial.com\/tensorflow\/example\/30750\/math-behind-1d-convolution-with-advanced-examples-in-tf\" rel=\"noreferrer\">Like this<\/a>:<\/p>\n\n<ul>\n<li>8  = 1 * 2 + 0 * 1 + 2 * 3<\/li>\n<li>11 = 0 * 2 + 2 * 1 + 3 * 3<\/li>\n<li>7  = 2 * 2 + 3 * 1 + 0 * 3<\/li>\n<li>9  = 3 * 2 + 0 * 1 + 1 * 3<\/li>\n<li>4  = 0 * 2 + 1 * 1 + 1 * 3<\/li>\n<\/ul>\n\n<p>Here we slide by one element, but nothing stops you by using any other number. This number is your stride. You can think about it as downsampling the result of the 1-strided convolution by just taking every s-th result.<\/p>\n\n<p>Knowing the input size <strong>i<\/strong>, kernel size <strong>k<\/strong>, stride <strong>s<\/strong> and padding <strong>p<\/strong> you can easily calculate the output size of the convolution as:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OC5Tk.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OC5Tk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here || operator means ceiling operation. For a pooling layer s = 1.<\/p>\n\n<hr>\n\n<h2>N-dim case.<\/h2>\n\n<p>Knowing the math for a 1-dim case, n-dim case is easy once you see that each dim is independent. So you just slide each dimension separately. Here is an <a href=\"http:\/\/www.riptutorial.com\/tensorflow\/example\/30756\/padding-and-strides--the-most-general-case-\" rel=\"noreferrer\">example for 2-d<\/a>. Notice that you do not need to have the same stride at all the dimensions. So for an N-dim input\/kernel you should provide N strides.<\/p>\n\n<hr>\n\n<h2>So now it is easy to answer all your questions:<\/h2>\n\n<ol>\n<li><strong>What do each of the 4+ integers represent?<\/strong>. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"noreferrer\">conv2d<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/max_pool\" rel=\"noreferrer\">pool<\/a> tells you that this list represents the strides among each dimension. Notice that the length of strides list is the same as the rank of kernel tensor.<\/li>\n<li><strong>Why must they have strides[0] = strides<a href=\"http:\/\/www.riptutorial.com\/tensorflow\/example\/30756\/padding-and-strides--the-most-general-case-\" rel=\"noreferrer\">3<\/a> = 1 for convnets?<\/strong>. The first dimension is batch size, the last is channels. There is no point of skipping neither batch nor channel. So you make them 1. For width\/height you can skip something and that's why they might be not 1.<\/li>\n<li><strong>tf.reshape(_X,shape=[-1, 28, 28, 1]). Why -1?<\/strong> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"noreferrer\">tf.reshape<\/a> has it covered for you: \n\n<blockquote>\n  <p>If one component of shape is the special value -1, the size of that dimension is computed so that the total size remains constant. In particular, a shape of [-1] flattens into 1-D. At most one component of shape can be -1.<\/p>\n<\/blockquote><\/li>\n<\/ol>\n"},{"AnswerId":"34643081","CreationDate":"2016-01-06T21:25:57.910","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The pooling and convolutional ops slide a \"window\" across the input tensor.  Using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#conv2d\"><code>tf.nn.conv2d<\/code><\/a> as an example: If the input tensor has 4 dimensions:  <code>[batch, height, width, channels]<\/code>, then the convolution operates on a 2D window on the <code>height, width<\/code> dimensions.<\/p>\n\n<p><code>strides<\/code> determines how much the window shifts by in each of the dimensions.  The typical use sets the first (the batch) and last (the depth) stride to 1.<\/p>\n\n<p>Let's use a very concrete example:  Running a 2-d convolution over a 32x32 greyscale input image.  I say greyscale because then the input image has depth=1, which helps keep it simple.  Let that image look like this:<\/p>\n\n<pre><code>00 01 02 03 04 ...\n10 11 12 13 14 ...\n20 21 22 23 24 ...\n30 31 32 33 34 ...\n...\n<\/code><\/pre>\n\n<p>Let's run a 2x2 convolution window over a single example (batch size = 1).  We'll give the convolution an output channel depth of 8.<\/p>\n\n<p>The input to the convolution has <code>shape=[1, 32, 32, 1]<\/code>.<\/p>\n\n<p>If you specify <code>strides=[1,1,1,1]<\/code> with <code>padding=SAME<\/code>, then the output of the filter will be [1, 32, 32, 8].<\/p>\n\n<p>The filter will first create an output for:<\/p>\n\n<pre><code>F(00 01\n  10 11)\n<\/code><\/pre>\n\n<p>And then for:<\/p>\n\n<pre><code>F(01 02\n  11 12)\n<\/code><\/pre>\n\n<p>and so on.  Then it will move to the second row, calculating:<\/p>\n\n<pre><code>F(10, 11\n  20, 21)\n<\/code><\/pre>\n\n<p>then<\/p>\n\n<pre><code>F(11, 12\n  21, 22)\n<\/code><\/pre>\n\n<p>If you specify a stride of [1, 2, 2, 1] it won't do overlapping windows.  It will compute:<\/p>\n\n<pre><code>F(00, 01\n  10, 11)\n<\/code><\/pre>\n\n<p>and then<\/p>\n\n<pre><code>F(02, 03\n  12, 13)\n<\/code><\/pre>\n\n<p>The stride operates similarly for the pooling operators.<\/p>\n\n<p><strong>Question 2:  Why strides [1, x, y, 1] for convnets<\/strong><\/p>\n\n<p>The first 1 is the batch:  You don't usually want to skip over examples in your batch, or you shouldn't have included them in the first place. :)<\/p>\n\n<p>The last 1 is the depth of the convolution:  You don't usually want to skip inputs, for the same reason.<\/p>\n\n<p>The conv2d operator is more general, so you <em>could<\/em> create convolutions that slide the window along other dimensions, but that's not a typical use in convnets.  The typical use is to use them spatially.<\/p>\n\n<p><strong>Why reshape to -1<\/strong>  -1 is a placeholder that says \"adjust as necessary to match the size needed for the full tensor.\"  It's a way of making the code be independent of the input batch size, so that you can change your pipeline and not have to adjust the batch size everywhere in the code.<\/p>\n"}]}
{"QuestionId":34646471,"AnswerCount":1,"Tags":"<machine-learning><classification><tensorflow><mnist>","CreationDate":"2016-01-07T02:34:24.367","AcceptedAnswerId":34658802.0,"OwnerUserId":1176802.0,"Title":"Tensorflow MNIST beginners need some understanding evaluation step","Body":"<p>I went thru the elementary example in Tensorflow of evaluating the trained model. Here is what it says:<\/p>\n\n<pre><\/pre>\n\n<p>I did not follow this code, where is the trained 'model'? or is it <strong>tf.reduce_mean(....)<\/strong> ? checking the trained model.<\/p>\n","answers":[{"AnswerId":"34658802","CreationDate":"2016-01-07T15:20:46.547","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p>As \"Guy Coder\" says, maybe you should check other online resources or MOOCs before starting with tensorflow.<\/p>\n\n<p>But anyway, maybe you will get a clearer picture with this...<\/p>\n\n<p>There are two parts into training a model in tensorflow. <\/p>\n\n<ol>\n<li>First you declare the structure of your model, with the different layers and variables. Tensorflow will make a graph out of that, but there is no computation happened yet. <\/li>\n<li>Then you ask tensorflow to \"run\" and optimize the model. What you do here is tell tensorflow that you want to reduce the cross entropy, or whatever loss function you define, so you provide the input data and labels the graph needs to compute that.<\/li>\n<\/ol>\n\n<p>After this you come up with a trained model. Maybe you will want to save the model and reuse it later, but that's another story.<\/p>\n\n<p>So, <strong>during training, or after it is finished<\/strong> you may call <br><code>print(sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels}))<\/code>. <\/p>\n\n<p>This is telling tensorflow to calculate the accuracy using the graph with the current value of the variables (maybe you are in the middle of training). And you are feeding this accuracy function the images and the labels. Tensorflow will take the x values and try to predict y_, and the accuracy will be the result of how well he did.<\/p>\n\n<p>The connection to your trained model comes from the <code>correct_prediction<\/code> function which should compare the correct output with your model's prediction, i.e. <strong>y_ vs y<\/strong> <\/p>\n\n<p>Hope this helps<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I will answer based on your comments, but be aware that your question is very poorly explained... as pointed out by S_kar<\/p>\n\n<p>To save a model you do it like this:<\/p>\n\n<pre><code># model declared before this line\nwith tf.Session() as sess:\n    # Merge all the summaries and write them out to \/tmp\/tf\n    merged = tf.merge_all_summaries()\n    writer = tf.train.SummaryWriter(\"\/tmp\/tf\", sess.graph_def)\n    tf.initialize_all_variables().run()\n\n    saver = tf.train.Saver()\n\n    \"\"\"\n    train the model...\n    \"\"\"\n\n    print \"Model succesfuly trained\"\n\n    # now save the model in a subdirectory called \"model\"\n    checkpoint_path = os.getcwd() + \"\/model\/model.ckpt\"\n    saver.save(sess, checkpoint_path)\n    print \"Model saved\"\n<\/code><\/pre>\n\n<p>To restore the model look into <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-restore-a-previously-saved-model-python\">this question<\/a><\/p>\n"}]}
{"QuestionId":34648760,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2016-01-07T06:37:13.417","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"how to switch a tensor in theano to numpy.array","Body":"<p>I have a tensor T (shape:300) and a array A(shape:300), what i want to do is combine them into a new array [T,A] with the shape (600). I tried the solutiona below:<\/p>\n\n<p>1 combine directly,use function: np.concatenate((T,A)), the result is:zero-dimensional arrays cannot be concatenated<\/p>\n\n<p>2 switch one type to another, try to switch the T to the type of numpy.array: i use: a=np.array(T), but when print a.shape, it is (), nothing in the bracket.<\/p>\n\n<p>Besides, when i print T.shape and A.shape, T.shape is ([300]) and A.shape is (300,)what is the difference?<\/p>\n","answers":[{"AnswerId":"34650739","CreationDate":"2016-01-07T08:48:31.023","ParentId":null,"OwnerUserId":"5737155","Title":null,"Body":"<p>when we want to get a numpy.array from a tensor T, it can be done by T.eval(), i tried a lot and found this way. But i haven't found the way switched from numpy.array to tensor T yet.  Anyone can help?<\/p>\n"}]}
{"QuestionId":34650307,"AnswerCount":2,"Tags":"<python><machine-learning><conv-neural-network><keras>","CreationDate":"2016-01-07T08:21:26.943","AcceptedAnswerId":34662626.0,"OwnerUserId":3200890.0,"Title":"Can deeper convolutional neural nets result in lesser parameters?","Body":"<p>I trained two convolutional neural nets in Keras. The first is net is as below<\/p>\n\n<pre><\/pre>\n\n<p>The second net<\/p>\n\n<pre><\/pre>\n\n<p>When I call the  method, the first net results in 604035 parameter and the second net results in 336387.<\/p>\n\n<p>How is this possible? The second net is deeper and should contain more parameters. Is there any mistake?<\/p>\n","answers":[{"AnswerId":"34654016","CreationDate":"2016-01-07T11:29:23.583","ParentId":null,"OwnerUserId":"4253485","Title":null,"Body":"<p>The depth of the network is not the only thing that affects its number of parameters. The number of parameters per layer has a huge effect. This means that for each convolutional layer, the size of the filter and the number of filters (features learned) would make a huge difference. Have a look at the paper of the VGG group at this link <a href=\"http:\/\/arxiv.org\/pdf\/1409.1556.pdf\" rel=\"nofollow\">http:\/\/arxiv.org\/pdf\/1409.1556.pdf<\/a><\/p>\n\n<p>Also please look at the paper \"<a href=\"http:\/\/arxiv.org\/pdf\/1512.03385v1.pdf\" rel=\"nofollow\">Deep Residual Learning for image recognition<\/a>\" which represents networks that have up to 152 layers i.e. 8x deeper than VGG nets while still having lower complexity.<\/p>\n"},{"AnswerId":"34662626","CreationDate":"2016-01-07T18:33:55.120","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p><strong>Yes, deeper neural network can have less parameters.<\/strong> It doesn't matter if they are CNNs. You might be confused, because in the graphical representation one tends to focus on the neurons. However, what gets learned are the weights, which are on the edges between neurons.<\/p>\n\n<p>Besides the link to \"Deep Residual Learning for image recognition\" (please upvote Midos answer for that), I would like to give a toy example of a multilayer perceptron (MLP).<\/p>\n\n<h2>Toy example: Bottleneck Features<\/h2>\n\n<p>The first MLP has an input layer of 784 neurons, two hidden layers of 2000 neurons each and an output layer of 10 neurons (short: 784:2000:2000:10). This results in a network with <a href=\"https:\/\/i.stack.imgur.com\/coA8f.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/coA8f.png\" alt=\"enter image description here\"><\/a> \nneurons. Now consider a network with the architecture 784:2000:50:2000:10. This has <a href=\"https:\/\/i.stack.imgur.com\/APiMH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/APiMH.png\" alt=\"enter image description here\"><\/a> neurons.<\/p>\n\n<p>This means adding another layer, even without reducing any of the layer, reduced the size of the network to 32% of the previous size!<\/p>\n"}]}
{"QuestionId":34651159,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2016-01-07T09:11:17.220","AcceptedAnswerId":34717765.0,"OwnerUserId":5545794.0,"Title":"error in caffe.io.load_image(), it cannot load image","Body":"<p>when I run the command in python:<\/p>\n\n<pre><\/pre>\n\n<p>it raised an error:<\/p>\n\n<blockquote>\n  <p>ValueError: Could not load \"bird.jpg\"\n  Please see documentation at: <a href=\"http:\/\/pillow.readthedocs.org\/en\/latest\/installation.html#external-libraries\" rel=\"nofollow\">http:\/\/pillow.readthedocs.org\/en\/latest\/installation.html#external-libraries<\/a><\/p>\n<\/blockquote>\n\n<p>I find some solutions at <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/ltdeqU4hiaw\" rel=\"nofollow\">caffe users<\/a> and <a href=\"http:\/\/delbert.me\/archives\/753.html\" rel=\"nofollow\">Delbert's Blog<\/a>, but they don't work, How can I fix it?<\/p>\n","answers":[{"AnswerId":"34717765","CreationDate":"2016-01-11T09:16:11.170","ParentId":null,"OwnerUserId":"5545794","Title":null,"Body":"<p>I solved my problem, here is the details below:<\/p>\n\n<p>First, I recheck this error imformation:<\/p>\n\n<blockquote>\n  <p>File \"\/usr\/local\/lib\/python2.7\/dist-packages\/scikit_image-0.11.3-py2.7-linux-x86_64.egg\/skimage\/io\/_plugins\/pil_plugin.py\", line 52, in imread\n  raise ValueError('Could not load \"%s\"\\nPlease see documentation at: %s' % (fname, site))\n  ValueError: Could not load \"bird.jpg\"\n  Please see documentation at: <a href=\"http:\/\/pillow.readthedocs.org\/en\/latest\/installation.html#external-libraries\" rel=\"nofollow noreferrer\">http:\/\/pillow.readthedocs.org\/en\/latest\/installation.html#external-libraries<\/a><\/p>\n<\/blockquote>\n\n<p>it is a error with <strong>imread<\/strong>, I checked the file <strong>pil_plugin.py<\/strong>, I found it got thiss error when it came to<\/p>\n\n<pre><code>im = Image.open(fname)\ntry:\n    # this will raise an IOError if the file is not readable\n    im.getdata()[0]\nexcept IOError:\n    site = \"http:\/\/pillow.readthedocs.org\/en\/latest\/installation.html#external-libraries\"\n    raise ValueError('Could not load \"%s\"\\nPlease see documentation at: %s' % (fname, site))\nelse:\n    return pil_to_ndarray(im, dtype=dtype, img_num=img_num)\n<\/code><\/pre>\n\n<p>so it is most probably wrong with <strong>im.getdata()<\/strong>, then I run it and it raise an error:<\/p>\n\n<blockquote>\n  <p>im.getdata() IOError: broken data stream when reading image file.<\/p>\n<\/blockquote>\n\n<p>then I followed <a href=\"https:\/\/stackoverflow.com\/questions\/6245619\/error-broken-data-stream-when-reading-image-file-truetype-fonts-libjpeg-and\">Matt W-D's solution<\/a> to reinstall jpeg6,  delete the <strong>PIL<\/strong> package installed by Pillow, and reinstall PIL by downloading file from <a href=\"http:\/\/effbot.org\/downloads#pil\" rel=\"nofollow noreferrer\">http:\/\/effbot.org\/downloads#pil<\/a>. <\/p>\n\n<p>At last, I linked the libjpeg<\/p>\n\n<pre><code>ln -s \/usr\/lib\/x86_64-linux-gnu\/libjpeg.so \/usr\/lib\nln -s \/usr\/lib\/x86_64-linux-gnu\/libfreetype.so \/usr\/lib\nln -s \/usr\/lib\/x86_64-linux-gnu\/libz.so \/usr\/lib\nsudo ldconfigthe \n<\/code><\/pre>\n\n<p>The error disappeared<\/p>\n"}]}
{"QuestionId":34651620,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-01-07T09:32:38.087","AcceptedAnswerId":null,"OwnerUserId":5756552.0,"Title":"Pycaffe: Index out of range when using original train.prototxt as the pretrained_file in Caffe.classifier","Body":"<p>When using ipython:<\/p>\n\n<pre><\/pre>\n\n<p>I got an exception when I use the original train.prototxt as the pretrained_file(PRETRAINED)<\/p>\n\n<pre><\/pre>\n\n<p>I have no idea why  the index  is out of range, Here I know a specified deploy.prototxt may work, but I don't know how.<\/p>\n\n<p>really hope some can answer my question!<\/p>\n","answers":[{"AnswerId":"34657177","CreationDate":"2016-01-07T14:07:09.250","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You get this error because your <code>train_val.prototxt<\/code> has no \"external\" inputs: There are no \"bottom\"s that are expected by the net and not provided by the net.<br>\nThis is because the <code>train_val.prototxt<\/code> has a specialized data layers that take care of the input (usually using pre computed datasets like lmdb\/leveldb\/hdf5 etc.).<br>\nIf you look closely at your <code>deploy.prototxt<\/code> you'll notice it has no input layer, instead it has a declaration:<\/p>\n\n<pre><code>net: \"some name\"\ninput: \"data\"\ninput_shape {\n  dim: 10\n  dim: 3\n  dim: 224\n  dim: 224\n}\n<\/code><\/pre>\n\n<p>This statement declares an \"external\" input, that is the resulting net will expect an input of name \"data\" with shape <code>(10,3,224,224)<\/code>.<br>\nYour <code>train_val.prototxt<\/code> has no such declaration.<\/p>\n"}]}
{"QuestionId":34652624,"AnswerCount":0,"Tags":"<python><windows><boost-python><caffe>","CreationDate":"2016-01-07T10:20:51.123","AcceptedAnswerId":null,"OwnerUserId":213615.0,"Title":"Using faster-rcnn pycaffe in Windows fails with boost python error","Body":"<p>I'm trying to run <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\" rel=\"nofollow\">py-faster-rcnn<\/a> in Windows, over static library build of Caffe with CPU only configuration.<br>\nIt fails in GetPythonLayer in the following line:  <\/p>\n\n<pre><\/pre>\n\n<p>with the following error:  <\/p>\n\n<blockquote>\n  <p>TypeError: No to_python (by-value) converter found for C++ type: class\n  caffe::LayerParameter<\/p>\n<\/blockquote>\n\n<p>Boost version: 1.58.\nVC-14, x64, python 2.7 from Anaconda, Win8.1.<\/p>\n\n<p>Any idea what could be the cause?<\/p>\n\n<p>Much thanks! <\/p>\n","answers":[]}
{"QuestionId":34652915,"AnswerCount":1,"Tags":"<python><class><undefined><theano>","CreationDate":"2016-01-07T10:33:41.453","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"theano : undefined variable in class","Body":"<p>minimalist code:<\/p>\n\n<pre><\/pre>\n\n<p>In the line: <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>I get an error in eclipse:<\/p>\n\n<blockquote>\n  <p>\"undefined variable y\"<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"34655154","CreationDate":"2016-01-07T12:28:09.813","ParentId":null,"OwnerUserId":"774513","Title":null,"Body":"<p>Well, you are passing a variable <code>y<\/code>, which isn't defined anywhere.<br>\nEclipse is looking for a definition of a variable y, for example, <code>y=10<\/code>, in the scope of the constructor and then in the global scope.<br>\nSince it can not find it, it warns you.  <\/p>\n\n<p>Did you possibly meant to pass <code>p_y_given_x<\/code> variable?<\/p>\n"}]}
{"QuestionId":34653643,"AnswerCount":1,"Tags":"<machine-learning><classification><deep-learning><tensorflow>","CreationDate":"2016-01-07T11:10:33.283","AcceptedAnswerId":null,"OwnerUserId":4164870.0,"Title":"tensorflow cifar10 code modification for reading images","Body":"<p>I'm trying to modify the code of cifar10.py in order to be able to feed images to the network.<\/p>\n\n<p>I'm actually able to run the code and to start the training process but after some time, if I run tensorboard, under the section \"images\" I always have the same image. \nMoreover the crossentropy goes to zero.\nI think that I'm loading the images wrong.<\/p>\n\n<p>Here's the code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The image reading part comes from <a href=\"https:\/\/github.com\/HamedMP\/ImageFlow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/HamedMP\/ImageFlow<\/a>\nThe custom reader comes from <a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\/34345827#34345827\">Tensorflow read images with labels<\/a> and the relative function is implemented as the following<\/p>\n\n <pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35893658","CreationDate":"2016-03-09T13:58:26.513","ParentId":null,"OwnerUserId":"5577321","Title":null,"Body":"<p>You can use this piece of code created by me, for my classification problem:<\/p>\n\n<pre><code>        resized_image = cv2.resize(image, (WIDTH, HEIGHT))\n        label = np.uint8(nclass)\n\n        arr = np.uint8([0 for x in range(image_bytes)])\n        #  fill the label:\n        arr[0] = label\n        arr_cnt = 1\n\n        #  fill the image (row-major order). first R values, then G values then B values\n        for y in range(0, HEIGHT):\n            for x in range(0, WIDTH):\n                arr[arr_cnt] = np.uint8(resized_image[x, y, 2])  # R\n                arr[arr_cnt + 1024] = np.uint8(resized_image[x, y, 1])  # G\n                arr[arr_cnt + 2048] = np.uint8(resized_image[x, y, 0])  # B\n\n                arr_cnt += 1\n\n        print \"train arr:\", arr[0], arr[3072]\n        train_arr = np.append(train_arr, arr)\n        #print train_arr[file_in_dir*3073]\n    else:\n        invalids_cnt += 1\n        #print \"image\", files_in_dir[file_in_dir], \"is invalid\"\n\n    #  Write array to train.bin file:\nwith open('data_batch_%d.bin' % nclass, 'wb') as f:\n        f.write(train_arr)\n        f.close()\n<\/code><\/pre>\n\n<p>Here, the resized image is the resized version of one input image \"image\". Next, I create an array with 3073 bytes: 1st byte = label, next 1024 bytes = red values of the image, next 1024 bytes = green values of the image, next 1024 bytes = blue values of the image.<\/p>\n\n<p>I do this for every single input image and then concatenate it into a big binary array, which is written in a binary file \"data_batch_%d\"<\/p>\n\n<p>I have posted my complete script (maybe harder to understand for general purposes) in this gist: <a href=\"https:\/\/gist.github.com\/twimnox\/27abbd245d403636efe5\" rel=\"nofollow\">gist<\/a><\/p>\n"}]}
{"QuestionId":34654220,"AnswerCount":1,"Tags":"<python><image><tensorflow>","CreationDate":"2016-01-07T11:39:12.243","AcceptedAnswerId":null,"OwnerUserId":3399066.0,"Title":"tensorflow check image is in reader","Body":"<p>I'm not sure whether TensorFlow actually has actually decoded an image when the output is only: <\/p>\n\n<p><strong>How can I show the image and label from the Tensor object?<\/strong><\/p>\n\n<p>(code partly from: <a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\">Tensorflow read images with labels<\/a>)<\/p>\n\n<pre><\/pre>\n\n<p>What can I do to show the actual image in  and also show the ? Because whether the image is indeed in the same folder or not doesn't change the output from  and .<\/p>\n\n<h2>Edit<\/h2>\n\n<p>The following code (partly from: <a href=\"https:\/\/stackoverflow.com\/questions\/33648322\/tensorflow-image-reading-display\">Tensorflow image reading &amp; display<\/a>) shows an image on my friends Mac, but not on my Ubuntu 14.04:<\/p>\n\n<pre><\/pre>\n\n<p>and this results in the following <strong>error<\/strong>:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34657937","CreationDate":"2016-01-07T14:43:10.693","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p>It seems there is code omitted in your second code section, no?<\/p>\n\n<p>Try the solution suggested <a href=\"http:\/\/andyljones.tumblr.com\/post\/133267887103\/tensorflow-placeholder-queue-errors\" rel=\"nofollow noreferrer\">here<\/a>. You have to initialize the placeholders before you start the queue runner. That fixed a similar problem for me.<\/p>\n\n<p>And I guess you already figured out that you must call <code>.eval()<\/code> over your tensor in order to get its actual value. Look in this <a href=\"https:\/\/stackoverflow.com\/questions\/34097281\/tensorflow-convert-a-tensor-into-a-numpy-array-in-python\">question I made<\/a><\/p>\n"}]}
{"QuestionId":34654263,"AnswerCount":1,"Tags":"<windows><batch-file><theano>","CreationDate":"2016-01-07T11:41:45.227","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"How do add a line to msys.bat:","Body":"<p>I am given the following advice on how to resolve an issue and i have no idea how to do that in windows<\/p>\n\n<p>Add the line to the beginning of msys.bat:<\/p>\n\n<p>call \u201cc:\\Program Files (x86)\\Microsoft Visual Studio 11.0\\VC\\bin\\x86_amd64\\vcvarsx86_amd64.bat\u201d<\/p>\n\n<p>it comes from a guide stating :<\/p>\n\n<pre><\/pre>\n\n<p>I adjusted it since i used a different version of visual studio, still i get the same mentioned error message.<\/p>\n","answers":[{"AnswerId":"34684426","CreationDate":"2016-01-08T19:15:54.450","ParentId":null,"OwnerUserId":"5463995","Title":null,"Body":"<p>Perhaps this?\nObviously in whatever path <code>mySys.bat<\/code> is located. Feel free to backup <code>mySys.bat<\/code> just to be careful.<\/p>\n\n<pre><code>&gt;&gt; mySysTemp.bat echo call \"c:\\Program Files (x86)\\Microsoft Visual Studio 11.0\\VC\\bin\\x86_amd64\\vcvarsx86_amd64.bat\"\nfor \/f \"tokens=* delims=\" %%G in (mySys.bat) do (\n    echo %%G\n) &gt;&gt; mySysTemp.bat\ndel mySys.bat &amp;&amp; ren mySysTemp.bat mySys.bat\n<\/code><\/pre>\n\n<p>This just echo's your line into a dummy file, copy line by line <code>mySys.bat<\/code> into the dummy file, then delete <code>mySys.bat<\/code> and rename the dummy file to replace it.<\/p>\n"}]}
{"QuestionId":34660337,"AnswerCount":1,"Tags":"<python><keras><autoencoder>","CreationDate":"2016-01-07T16:31:12.013","AcceptedAnswerId":35109634.0,"OwnerUserId":5561337.0,"Title":"Keras autoencoder accuracy\/loss doesn't change","Body":"<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>X has a shape (537621, 256). I'm trying to find a way to compress the vectors of size 256 to 100, then to 70, then to 50. I have done this is Lasagne but in Keras it seems to be easier to work w\/ Autoencoders.<\/p>\n\n<p>Here is the output:<\/p>\n\n<p>Epoch 1\/500\n537621\/537621 [==============================] - 27s - loss: 0.1339 - acc: 0.0036<br>\nEpoch 2\/500\n537621\/537621 [==============================] - 32s - loss: 0.1339 - acc: 0.0036<br>\nEpoch 3\/500\n252336\/537621 [=============>................] - ETA: 14s - loss: 0.1339 - acc: 0.0035<\/p>\n\n<p>And it continues like this on and on..<\/p>\n","answers":[{"AnswerId":"35109634","CreationDate":"2016-01-31T02:23:42.020","ParentId":null,"OwnerUserId":"1815451","Title":null,"Body":"<p>It's now fixed on master:) openning issues is sometimes best choice \n<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1604\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/issues\/1604<\/a><\/p>\n"}]}
{"QuestionId":34661818,"AnswerCount":1,"Tags":"<python-2.7><speech-recognition><theano><mfcc><lstm>","CreationDate":"2016-01-07T17:47:27.627","AcceptedAnswerId":34667051.0,"OwnerUserId":2220487.0,"Title":"Building Speech Dataset for LSTM binary classification","Body":"<p>I'm trying to do binary LSTM classification using theano. \nI have gone through the example code however I want to build my own.<\/p>\n\n<p>I have a small set of \"Hello\" &amp; \"Goodbye\" recordings that I am using. I preprocess these by extracting the MFCC features for them and saving these features in a text file. I have 20 speech files(10 each) and I am generating a text file for each word, so 20 text files that contains the MFCC features. Each file is a 13x56 matrix.<\/p>\n\n<p>My problem now is: How do I use this text file to train the LSTM?<\/p>\n\n<p>I am relatively new to this. I have gone through some literature on it as well but not found really good understanding of the concept.<\/p>\n\n<p>Any simpler way using LSTM's would also be welcome.<\/p>\n","answers":[{"AnswerId":"34667051","CreationDate":"2016-01-07T23:28:39.840","ParentId":null,"OwnerUserId":"432021","Title":null,"Body":"<p>There are many existing implementation for example <a href=\"https:\/\/github.com\/mozilla\/DeepSpeech\" rel=\"noreferrer\">Tensorflow Implementation<\/a>, <a href=\"https:\/\/github.com\/srvk\/eesen\" rel=\"noreferrer\">Kaldi-focused implementation with all the scripts<\/a>, it is better to check them first.<\/p>\n\n<p>Theano is too low-level, you might try with <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"noreferrer\">keras<\/a> instead, as described in <a href=\"http:\/\/danielhnyk.cz\/predicting-sequences-vectors-keras-using-rnn-lstm\/\" rel=\"noreferrer\">tutorial<\/a>. You can run tutorial \"as is\" to understand how things goes.<\/p>\n\n<p>Then, you need to prepare a dataset. You need to turn your data into sequences of data frames and for every data frame in sequence you need to assign an output label. <\/p>\n\n<p>Keras supports two types of RNNs - layers returning sequences and layers returning simple values. You can experiment with both, in code you just use <code>return_sequences=True<\/code> or <code>return_sequences=False<\/code><\/p>\n\n<p>To train with sequences you can assign dummy label for all frames except the last one where you can assign the label of the word you want to recognize. You need to place input and output labels to arrays. So it will be:<\/p>\n\n<pre><code>X = [[word1frame1, word1frame2, ..., word1framen],[word2frame1, word2frame2,...word2framen]]\n\nY = [[0,0,...,1], [0,0,....,2]]\n<\/code><\/pre>\n\n<p>In X every element is a vector of 13 floats. In Y every element is just a number - 0 for intermediate frames and word ID for final frame.<\/p>\n\n<p>To train with just labels you need to place input and output labels to arrays and output array is simpler. So the data will be:<\/p>\n\n<pre><code>X = [[word1frame1, word1frame2, ..., word1framen],[word2frame1, word2frame2,...word2framen]]\n\nY = [[0,0,1], [0,1,0]]\n<\/code><\/pre>\n\n<p>Note that output is vectorized (np_utils.to_categorical) to turn it to vectors instead of just numbers.<\/p>\n\n<p>Then you create network architecture. You can have 13 floats for input, a vector for output. In the middle you might have one fully connected layer followed by one lstm layer. Do not use too big layers, start with small ones.<\/p>\n\n<p>Then you feed this dataset into <code>model.fit<\/code> and it trains you the model. You can estimate model quality on heldout set after training.<\/p>\n\n<p>You will have a problem with convergence since you have just 20 examples. You need way more examples, preferably thousands to train LSTM, you will only be able to use very small models.<\/p>\n"}]}
{"QuestionId":34662443,"AnswerCount":2,"Tags":"<convolution><tensorflow><conv-neural-network>","CreationDate":"2016-01-07T18:24:23.923","AcceptedAnswerId":34664566.0,"OwnerUserId":3908247.0,"Title":"Fully-connected layer weight dimensions in TensorFlow ConvNet","Body":"<p>I've been coding along <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3%20-%20Neural%20Networks\/convolutional_network.ipynb\" rel=\"noreferrer\">this example<\/a> of a convolution net in TensorFlow and I'm mystified by this allocation of weights:<\/p>\n\n<pre><\/pre>\n\n<p><strong>How do we know the 'wd1' weight matrix should have 7 x 7 x 64 rows?<\/strong><\/p>\n\n<p>It's later used to reshape the output of the second convolution layer:<\/p>\n\n<pre><\/pre>\n\n<p>By my math, pooling layer 2 (conv2 output) has 4 x 4 x 64 neurons.<\/p>\n\n<p><strong>Why are we reshaping to [-1, 7*7*64]?<\/strong><\/p>\n","answers":[{"AnswerId":"39841888","CreationDate":"2016-10-03T23:19:41.637","ParentId":null,"OwnerUserId":"2373206","Title":null,"Body":"<p>This question requires you have a good understanding of deep learning convolutions.<\/p>\n\n<p>Basically, each convolution layer your model has will reduce the convolutional pyramid transversal area. This reduction is made by the <strong>convolution stride<\/strong> and <strong>max_pooling stride<\/strong>. And to complicate things we have two options based on the PADDING.  <\/p>\n\n<p>Option 1 - <code>PADDING='SAME'<\/code><\/p>\n\n<pre><code>out_height = ceil(float(in_height) \/ float(strides[1]))\nout_width  = ceil(float(in_width) \/ float(strides[2]))\n<\/code><\/pre>\n\n<p>Option 2 - <code>PADDING='VALID'<\/code><\/p>\n\n<pre><code>out_height = ceil(float(in_height - filter_height + 1) \/ float(strides[1]))\nout_width  = ceil(float(in_width - filter_width + 1) \/ float(strides[2]))\n<\/code><\/pre>\n\n<p>For <strong>EACH<\/strong> convolution and max pooling call you will have to calculate a new <code>out_height<\/code> and <code>out_width<\/code>. Then, in the end of the convolutions you multiply <code>out_height<\/code>, <code>out_width<\/code> and the <strong>depth<\/strong> of your last convolution layer. The result of this multiplication is the output feature map size which is the input of your first fully connected layer.<\/p>\n\n<p>So, in your example you probably had only <code>PADDING='SAME'<\/code>, a convolution stride of 1 and a max pooling stride of 2, <strong>twice<\/strong>. In the end you just had to divide everything by 4 (1,2,1,2).  <\/p>\n\n<p>More info at <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/nn.html#convolution\" rel=\"nofollow noreferrer\">tensorflow API<\/a><\/p>\n"},{"AnswerId":"34664566","CreationDate":"2016-01-07T20:30:03.487","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Working from the start:<\/p>\n\n<p>The input, <code>_X<\/code> is of size <code>[28x28x1]<\/code> (ignoring the batch dimension).  A 28x28 greyscale image.<\/p>\n\n<p>The first convolutional layer uses <code>PADDING=same<\/code>, so it outputs a 28x28 layer, which is then passed to a <code>max_pool<\/code> with <code>k=2<\/code>, which reduces each dimension by a factor of two, resulting in a 14x14 spatial layout.  conv1 has 32 outputs -- so the full per-example tensor is now <code>[14x14x32]<\/code>.<\/p>\n\n<p>This is repeated in <code>conv2<\/code>, which has 64 outputs, resulting in a <code>[7x7x64]<\/code>.<\/p>\n\n<p>tl;dr:  The image starts as 28x28, and each maxpool reduces it by a factor of two in each dimension.  28\/2\/2 = 7.<\/p>\n"}]}
{"QuestionId":34662855,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-07T18:47:40.343","AcceptedAnswerId":34663362.0,"OwnerUserId":1828879.0,"Title":"Is there an identity function?","Body":"<p>Suppose I have a Theano symbol  and I run the following code.<\/p>\n\n<pre><\/pre>\n\n<p>Of course  is then . <strong>Is there an identity function that lets me do something like the following?<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The expected behavior is that  is now seen to be a function of , and they are both properly named. Of course the Theano compiler would simply merge the symbols because  is just .<\/p>\n\n<p>The reason I am asking this is because I have a situation like the following, where  and  are Theano symbols and  is either  or .<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that in the case of  being , my  symbol gets the name .<\/p>\n\n<p>I can fix this by working around the problem:<\/p>\n\n<pre><\/pre>\n\n<p>But an identity function would be much more elegant:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34663362","CreationDate":"2016-01-07T19:17:49.067","ParentId":null,"OwnerUserId":"1828879","Title":null,"Body":"<p>The <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.copy\" rel=\"nofollow\"><code>copy(name=None)<\/code> function<\/a> on tensors is what you want.<\/p>\n\n<p>The first example becomes this:<\/p>\n\n<pre><code>x.name = \"x\"\ny = x.copy(\"y\")\n<\/code><\/pre>\n\n<p>The second example becomes this:<\/p>\n\n<pre><code>activation = T.dot(feature, filter)\nactivation.name = \"activation\"\nresponse = T.nnet.sigmoid(activation) if nonlinear else activation.copy()\nresponse.name = \"response\"\n<\/code><\/pre>\n"}]}
{"QuestionId":34663737,"AnswerCount":1,"Tags":"<python><neural-network><theano><convolution><conv-neural-network>","CreationDate":"2016-01-07T19:39:31.273","AcceptedAnswerId":null,"OwnerUserId":5548166.0,"Title":"CNN does not work well on my dataset in Theano","Body":"<p>I want to apply the convolutional_mlp in theano to my own dataset. But the CNN cannot work well to my dataset, and I cannot find its reason. In the following, I will describe my datset in details, and can someone help me? Thanks so much!<\/p>\n\n<p>The dataset:\nMy goal is to use the CNN to classify my images (whose size is 28*28 = 784) into two types, 0 and 1. \nMy original data are the csv files, the datax.csv and datay.csv. In the datax.csv, each row stores the gray values of each pixel of each image, and in the datay.csv, each row is the label of each image. \ndatax.csv (1000 rows, and 784 columns)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VZirs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VZirs.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>datay.csv (1000 rows)<\/p>\n\n<pre><\/pre>\n\n<p>Based on the above csv data, I construct my dataset ('myDataset.pkl.gz') by the Python code as follows:<\/p>\n\n<pre><\/pre>\n\n<p>After constructing my dataset, I use it in the convolutional_mlp.py, by replacing the mnist.pkl.gz. \nIn my dataset, there are 1000 examples, 500 examples, and 500 examples in the training, validating, and testing datasets respectively. So in the convolutional_mlp.py, I set the batch_size as 500. Then there are 2 mini-batches, 1 mini-batch, and 1 mini-batch in the training, validating, and testing datasets respectively. Also I let the n_out be 2. Then I run the code, and print the training error of each epoch. The results are as follows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sGf0F.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sGf0F.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>No matter how many epochs it runs, the validation error never changes. In fact, I have changed the learning rate from 0.1, 0.01, 0.001,\u2026 But the results never changes. \nI will want to know why this will happen, can someone help me? Thank you so much!<\/p>\n","answers":[{"AnswerId":"40271516","CreationDate":"2016-10-26T20:24:00.737","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>It looks like your training, validation and test datasets are entirely different from each other. I mean, it may be that your dataset is sorted on  some feature values or output, such that the training set and validation set do not have features common with each other. One example of such scenario is when you only have values from 0-3 in training set, 4-6 in validation set and 7-9 in test set for MNIST data partition.<\/p>\n\n<p>It will work better if you read all the three datasets, shuffle them randomly and then split them into three sets. Generally, people use either 7:2:1 or 8:1:1 ratio of trianing size, validation size and test size. The model will learn better after such splitting and it will improve in accuracy.<\/p>\n\n<p>Also, I would recommend you to use a smaller minibatch size, may be 32 or 64. That will lead to higher improvement in accuracy with each epoch. I hope it helps.<\/p>\n"}]}
{"QuestionId":34664602,"AnswerCount":1,"Tags":"<tensorflow><standards><bazel>","CreationDate":"2016-01-07T20:32:25.667","AcceptedAnswerId":34665355.0,"OwnerUserId":1243762.0,"Title":"Building new TensorFlow Op, is there a build system standard?","Body":"<p>After watching this <a href=\"https:\/\/stackoverflow.com\/q\/34247374\/1243762\">question<\/a> I decided to give writing a new op for TensorFlow a try. <\/p>\n\n<p>Since the requirements of C++, Python and likely a *nix system are not my primary tools, I would like to avoid being at a point where I have to back out and make a system\/tool changes just because I did not ask.<\/p>\n\n<p>Is there a standard or preferred system and or tools used by those working or TensorFlow? <\/p>\n\n<p>I know that recommendation questions are not allowed here; I am not asking for a personal recommendation, I am asking for the standard used by or what the TensorFlow group finds that works.<\/p>\n","answers":[{"AnswerId":"34665355","CreationDate":"2016-01-07T21:18:47.040","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Really, anything where you can get Bazel and the required libraries up and running.  But since you're starting from scratch:  Ubuntu's a very safe bet and (I haven't measured this, but this is a solid estimate) probably gets the most testing and development by the tf team.  But there are many options that all work -- you can develop inside a virtualenv on many environments.  Things like GPU support get a little more platform-specific, and that's where Ubuntu starts to become the easiest choice if you don't have any other constraints.<\/p>\n\n<p>The key requirements are outlined in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">installing Tensorflow from sources<\/a>.<\/p>\n"}]}
{"QuestionId":34668144,"AnswerCount":2,"Tags":"<compilation><redhat><glibc><tensorflow><bazel>","CreationDate":"2016-01-08T01:35:20.697","AcceptedAnswerId":null,"OwnerUserId":1104696.0,"Title":"How to install bazel and tensorflow on Red Hat 6.7","Body":"<p>I would like to install bazel from source, and use bazel to compile tensorflow on a cluster running redhat 6.7. When I try to install bazel, the glibc version (2.12) is too old. I do not have root access to the cluster. Is it possible to install tensorflow in this case?<\/p>\n\n<p>My system information:<\/p>\n\n<pre><\/pre>\n\n<p>The system has newer gcc installed as well. I tried using it, bazel still won't compile.<\/p>\n\n<pre><\/pre>\n\n<p>When I was compiling bazel, I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Some people also reported this issue:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/110<\/a>\nand <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/527\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/527<\/a><\/p>\n\n<p>How can I install the missing dependency locally, and have bazel pick up the right library?<\/p>\n","answers":[{"AnswerId":"43991470","CreationDate":"2017-05-16T02:12:50.487","ParentId":null,"OwnerUserId":"2528241","Title":null,"Body":"<p>In case anyone need to do this manually:<\/p>\n\n<ol>\n<li><p>Compile up-to-date glibc, gcc and all their dependencies from source with the option <code>--disable-rpath<\/code> to avoid glibc path being hard coded to system default. Do NOT add glibc to <code>LD_LIBRARY_PATH<\/code> directly, or all executables including <code>rm<\/code> will stop working.<\/p><\/li>\n<li><p>Compile python with your gcc, install pip and the official wheel<\/p><\/li>\n<\/ol>\n\n<p><code>.\/configure --prefix=$PWD\/build --enable-unicode=ucs4 --with-cxx-main=g++ &amp;&amp; make &amp;&amp; make install<\/code><\/p>\n\n<ol start=\"3\">\n<li>Start python with the following command to load the correct glibc, where ${GLIBC_PATH} is your install location to glibc, make an alias<\/li>\n<\/ol>\n\n<p><code>alias tensorflow='${GLIBC_PATH}\/lib\/ld-2.23.so --library-path ${GLIBC_PATH}\/lib:${LD_LIBRARY_PATH}<\/code>which python<code>'<\/code><\/p>\n\n<ol start=\"4\">\n<li>Import tensorflow to check no error occur<\/li>\n<\/ol>\n\n<p>After able to import tensorflow module without any error, you can use any other computer (possibly ubuntu VM on your PC) to compile a custom wheel with machine specific option, and copy that to your cluster following the guide<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --copt=-msse4.1 \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n"},{"AnswerId":"34678259","CreationDate":"2016-01-08T13:29:00.080","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p>You should be able to compile from source with a newer version of Bazel exporting LD_FLAGS, CXX and CC, and tweaking the tools\/cpp\/CROSSTOOL file of Bazel. Please open a github issue on Bazel (<a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\" rel=\"nofollow\">https:\/\/github.com\/bazelbuild\/bazel\/issues<\/a>) if you have further questions.<\/p>\n\n<p>I am currently working on making all that easier. Sorry for the mess.<\/p>\n"}]}
{"QuestionId":34670112,"AnswerCount":5,"Tags":"<python><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-01-08T05:29:21.463","AcceptedAnswerId":34675264.0,"OwnerUserId":4763311.0,"Title":"How to deal with batches with variable-length sequences in TensorFlow?","Body":"<p>I was trying to use an RNN (specifically, LSTM) for sequence prediction. However, I ran into an issue with variable sequence lengths. For example,<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to predicting the next word after the current one with a simple RNN based on this <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow noreferrer\">Benchmark for building a PTB LSTM model<\/a>.<\/p>\n\n<p>However, the  parameter (used for unrolling to the previous hidden states), should remain the same in each Tensorflow's epoch. Basically, batching sentences is not possible as the sentences vary in length. <\/p>\n\n<pre><\/pre>\n\n<p>Here,  need to be changed in my case for every sentence. I have tried several hacks, but nothing seems working.<\/p>\n","answers":[{"AnswerId":"52395594","CreationDate":"2018-09-18T22:48:04.587","ParentId":null,"OwnerUserId":"6666226","Title":null,"Body":"<p>Sorry to post on a dead issue but I just submitted a PR for a better solution. <code>dynamic_rnn<\/code> is extremely flexible but abysmally slow. It works if it is your only option but CuDNN is much faster. This PR adds support for variable lengths to <code>CuDNNLSTM<\/code>, so you will hopefully be able to use that soon.<\/p>\n\n<p>You need to sort sequences by descending length. Then you can <code>pack_sequence<\/code>, run your RNNs, then <code>unpack_sequence<\/code>.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/22308\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/22308<\/a><\/p>\n"},{"AnswerId":"34675264","CreationDate":"2016-01-08T10:52:13.373","ParentId":null,"OwnerUserId":"2378490","Title":null,"Body":"<p>You can use the ideas of bucketing and padding which are described in:  <\/p>\n\n<p>&nbsp;&nbsp;&nbsp;&nbsp;<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html\" rel=\"noreferrer\">Sequence-to-Sequence Models<\/a><\/p>\n\n<p>Also, the rnn function which creates RNN network accepts parameter <strong>sequence_length.<\/strong><\/p>\n\n<p>As an example, you can create buckets of sentences of the same size, pad them with the necessary amount of zeros, or placeholders which stand for zero word and afterwards feed them along with seq_length = len(zero_words).<\/p>\n\n<pre><code>seq_length = tf.placeholder(tf.int32)\noutputs, states = rnn.rnn(cell, inputs, initial_state=initial_state, sequence_length=seq_length)\n\nsess = tf.Session()\nfeed = {\n    seq_length: 20,\n    #other feeds\n}\nsess.run(outputs, feed_dict=feed)\n<\/code><\/pre>\n\n<p>Take a look at this reddit thread as well:  <\/p>\n\n<p>&nbsp;&nbsp;&nbsp;<a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/3sok8k\/tensorflow_basic_rnn_example_with_variable_length\/\" rel=\"noreferrer\">Tensorflow basic RNN example with 'variable length' sequences<\/a><\/p>\n"},{"AnswerId":"34742097","CreationDate":"2016-01-12T11:11:00.187","ParentId":null,"OwnerUserId":"4763311","Title":null,"Body":"<p>You can use ideas of bucketing and padding which are described in  <\/p>\n\n<p>&nbsp;&nbsp;&nbsp;<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html\" rel=\"nofollow noreferrer\">Sequence-to-Sequence Models<\/a><\/p>\n\n<p>Also rnn function which creates RNN network accepts parameter sequence_length.<\/p>\n\n<p>As example you can create buckets of sentances of the same size, padd them with necessary amount of zeros, or placeholdres which stands for zero word and afterwards feed them along with seq_length = len(zero_words).<\/p>\n\n<pre><code>seq_length = tf.placeholder(tf.int32)\noutputs, states = rnn.rnn(cell, inputs,initial_state=initial_state,sequence_length=seq_length)\n\nsess = tf.Session()\nfeed = {\nseq_lenght: 20,\n#other feeds\n       }\nsess.run(outputs, feed_dict=feed)\n<\/code><\/pre>\n\n<p>Here , the most important thing is , if you want to make use of the states obtained by one sentence as , the state for the next sentence , when you are providing sequence_length , ( lets say 20 and sentence after padding is 50 ) . You want the state obtained at the 20th time step . For that , do<\/p>\n\n<pre><code>tf.pack(states)\n<\/code><\/pre>\n\n<p>After that call <\/p>\n\n<pre><code>for i in range(len(sentences)):\nstate_mat   = session.run([states],{\n            m.input_data: x,m.targets: y,m.initial_state: state,     m.early_stop:early_stop })\nstate = state_mat[early_stop-1,:,:]\n<\/code><\/pre>\n"},{"AnswerId":"41692307","CreationDate":"2017-01-17T08:32:23.327","ParentId":null,"OwnerUserId":"4999362","Title":null,"Body":"<p>You can use <code>dynamic_rnn<\/code> instead and specify length of every sequence even within one batch via passing array to <code>sequence_length<\/code> parameter.\nExample is below:<\/p>\n\n<pre><code>def length(sequence):\n    used = tf.sign(tf.reduce_max(tf.abs(sequence), reduction_indices=2))\n    length = tf.reduce_sum(used, reduction_indices=1)\n    length = tf.cast(length, tf.int32)\n    return length\n\nfrom tensorflow.nn.rnn_cell import GRUCell\n\nmax_length = 100\nframe_size = 64\nnum_hidden = 200\n\nsequence = tf.placeholder(tf.float32, [None, max_length, frame_size])\noutput, state = tf.nn.dynamic_rnn(\n    GRUCell(num_hidden),\n    sequence,\n    dtype=tf.float32,\n    sequence_length=length(sequence),\n)\n<\/code><\/pre>\n\n<p>Code is taken from a <a href=\"https:\/\/danijar.com\/variable-sequence-lengths-in-tensorflow\/\" rel=\"noreferrer\">perfect article<\/a> on the topic, please also check it. <\/p>\n\n<p>Update: Another <a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"noreferrer\">great post<\/a> on <code>dynamic_rnn<\/code> vs <code>rnn<\/code> you can find <\/p>\n"},{"AnswerId":"39901739","CreationDate":"2016-10-06T16:58:43.527","ParentId":null,"OwnerUserId":"2005810","Title":null,"Body":"<p>You can limit the maximum length of your input sequences, pad the shorter ones to that length, record the length of each sequence and use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/nn.html#dynamic_rnn\" rel=\"nofollow\">tf.nn.dynamic_rnn<\/a> . It processes input sequences as usual, but after the last element of a sequence, indicated by <code>seq_length<\/code>, it just copies the cell state through, and for output it outputs zeros-tensor.<\/p>\n"}]}
{"QuestionId":34670271,"AnswerCount":2,"Tags":"<python><machine-learning><conv-neural-network><keras>","CreationDate":"2016-01-08T05:44:00.287","AcceptedAnswerId":34757886.0,"OwnerUserId":3200890.0,"Title":"Deep Convolutional Neural Network in Keras","Body":"<p>Hi im trying to increase the depth of an existing convolutional nets in keras. Below is the existing network:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to add the depth of the network by adding in few convolution layers as below:<\/p>\n\n<pre><\/pre>\n\n<p>Keras is giving me an error. Im not sure what is wrong but seems like the tensor shape is wrong. Below is the error.<\/p>\n\n<pre><\/pre>\n\n<p>My input is 28 by 28 pixel image.\nCan anyone point me to what is wrong with my model?<\/p>\n","answers":[{"AnswerId":"38929008","CreationDate":"2016-08-13T03:52:16.747","ParentId":null,"OwnerUserId":"1556826","Title":null,"Body":"<p>I am answering this because I don't have enough rep to comment. The formula for calculating the output volume is having a slight issue. K*((W\u2212F+2P)\/S+1) should be (K*((W\u2212F+2P)\/S))+1. <\/p>\n"},{"AnswerId":"34757886","CreationDate":"2016-01-13T03:26:10.730","ParentId":null,"OwnerUserId":"3200890","Title":null,"Body":"<p>The answer most probably has to do with the image size. MY image size is a 28x28 image. When we perform convolution and pooling (without zero padding), the size of the feature maps will reduce. Therefore the number of convolution and pooling layer will be limited depending on the dimension of the input image.<\/p>\n\n<p>Following formula from <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow\">http:\/\/cs231n.github.io\/convolutional-networks\/<\/a><\/p>\n\n<p>feature map dimension = K*((W\u2212F+2P)\/S+1), \nwhere W - input volume size, \nF the receptive field size of the Conv Layer neurons, \nS - the stride with which they are applied, \nP - the amount of zero padding used on the border, \nK - the depth of conv layer<\/p>\n\n<p>Lets no worry about the depth of the conv layer for now. We just want to calculate the height and width of the resulting feature maps after each <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation.<\/p>\n\n<p>For the first network on the top I applied <code>[CONV -&gt; CONV -&gt; POOL]<\/code> twice. Lets calculate the resulting feature map. <\/p>\n\n<p>Given F=3, P=0, S=1, W=28 the output of the first <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation is:<\/p>\n\n<pre><code>[CONV]\n<\/code><\/pre>\n\n<p>feature map dimension = W\u2212F+2P)\/S+1 = (28 - 3 + 0)\/1 = 26<\/p>\n\n<pre><code>[CONV]\n<\/code><\/pre>\n\n<p>feature map dimension = W\u2212F+2P)\/S+1 = (26-3 + 0)\/1 = 23<\/p>\n\n<pre><code>[POOL]\n<\/code><\/pre>\n\n<p>applying pooling operation results in 23\/2 = 11<\/p>\n\n<p>This implies that after the first <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation, the feature map now has 11x11 pixels<\/p>\n\n<p>Lets apply the second <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation to the 11x11 feature map. We find that we will end up with a feature map of 2x2 pixels.<\/p>\n\n<p>Now if we try to apply the third <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation as what I wanted to do in the second network. We find that the dimension of the 2x2 feature map is too small for another <code>[CONV -&gt; CONV -&gt; POOL]<\/code> operation.<\/p>\n\n<p>I guess this is the reason for the error. <\/p>\n\n<p>Applying my speculation above i tried to train the second network with bigger image and the error does not appear.<\/p>\n"}]}
{"QuestionId":34672480,"AnswerCount":1,"Tags":"<python-2.7><deep-learning><caffe><pycaffe>","CreationDate":"2016-01-08T08:26:05.870","AcceptedAnswerId":34702804.0,"OwnerUserId":3106889.0,"Title":"Calling particular function at each input on testing or on prediction?","Body":"<p>I want to call a function for every input in test time for MNIST dataset. For example, if the test data in MNIST is x1, ... , xn for every input xi at test time I want to call the function. \nI modified the code in \"<a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/tutorial\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/tutorial\/examples\/01-learning-lenet.ipynb<\/a>\" such that I comment the line \" #solver.step(1)  # SGD by Caffe\" becuase I do not want any training of the network. But I think this is not right way as the accuracy decrease comparing to the orginal code. Any help please for how to implement it.<\/p>\n\n<p>The code which I think once can change.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34702804","CreationDate":"2016-01-10T06:36:14.277","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can add a python Layer for <code>include { phase: TEST }<\/code> only that will transform your test inputs.<\/p>\n"}]}
{"QuestionId":34673164,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras>","CreationDate":"2016-01-08T09:08:51.620","AcceptedAnswerId":34677655.0,"OwnerUserId":4984897.0,"Title":"How to train and tune an artificial multilayer perceptron neural network using Keras?","Body":"<p>I am building my first artificial multilayer perceptron neural network using Keras.<\/p>\n\n<p>This is my input data:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jCpMX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jCpMX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is my code which I used to build my initial model which basically follows the Keras example code:<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p>How do I train and tune this model and get my code to output my best predictive model? I am new to neural networks and am just wholly confused as to what is the next step after building the model. I know I want to optimize it, but I'm not sure which features to tweak or if I am supposed to do it manually or how to write code to do so.<\/p>\n","answers":[{"AnswerId":"34677655","CreationDate":"2016-01-08T12:55:32.460","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<p>Some things that you could do are:<\/p>\n\n<ul>\n<li>Change your loss function from <code>mean_squared_error<\/code> to <code>binary_crossentropy<\/code>. <code>mean_squared_error<\/code> is intended for regression, but you want to classify your data.<\/li>\n<li>Add <code>show_accuracy=True<\/code> to your <code>fit()<\/code> function, which outputs the accuracy of your model at every epoch. That information is probably more useful to you than just the loss value.<\/li>\n<li>Add <code>validation_split=0.2<\/code> to your <code>fit()<\/code> function. Currently you are only training on a training set and validating on nothing. That's a no-go in machine learning as you can't be sure that your model hasn't simply memorized the correct answers for your dataset (without really understanding why these answers are correct).<\/li>\n<li>Change from Obama\/Romney to Democrat\/Republican and add data from previous elections. ~1200 examples is a pretty small dataset for neural networks. Also add columns with valuable information, like unemployment rate or population <em>density<\/em>. Note that quite some of the values (like population number) are probably similar to providing the name of the state, so e.g. your net will likely learn that Texas means <code>Republican<\/code>.<\/li>\n<li>If you haven't done that already, normalize all your values to the range of 0 to 1 (by subtracting from each value the minimum of the column and then dividing by the (max - min) of the column). Neural networks can handle normalized data better than unnormalized data.<\/li>\n<li>Try <code>Adam<\/code> and <code>Adagrad<\/code> instead of <code>SGD<\/code>. Sometimes they perform better. (See <a href=\"http:\/\/keras.io\/optimizers\/\">documentation about optimizers<\/a>.)<\/li>\n<li>Try <code>Activation('relu')<\/code>, <code>LeakyReLU<\/code>, <code>PReLU<\/code> and <code>ELU<\/code> instead of <code>Activation('tanh')<\/code>. Tanh is rarely the best choice. (See <a href=\"http:\/\/keras.io\/layers\/advanced_activations\/\">advanced activation functions<\/a>.)<\/li>\n<li>Try increasing\/decreasing your dense layers sizes (e.g. from <code>64<\/code> to <code>128<\/code>). Also try adding\/removing layers.<\/li>\n<li>Try adding <code>BatchNormalization<\/code> layers (before the <code>Activation<\/code> layers). (See <a href=\"http:\/\/keras.io\/layers\/normalization\/\">documentation<\/a>.)<\/li>\n<li>Try changing the dropout rates (e.g. from <code>0.5<\/code> to <code>0.25<\/code>).<\/li>\n<\/ul>\n"}]}
{"QuestionId":34673396,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><data-science>","CreationDate":"2016-01-08T09:22:31.707","AcceptedAnswerId":37979686.0,"OwnerUserId":4984897.0,"Title":"What does the standard Keras model output mean? What is epoch and loss in Keras?","Body":"<p>I have just built my first model using Keras and this is the output. It looks like the standard output you get after building any Keras artificial neural network. Even after looking in the documentation, I do not fully understand what the epoch is and what the loss is which is printed in the output.<\/p>\n\n<p><strong>What is epoch and loss in Keras?<\/strong> <\/p>\n\n<p>(I know it's probably an extremely basic question, but I couldn't seem to locate the answer online, and if the answer is really that hard to glean from the documentation I thought others would have the same question and thus decided to post it here.)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37979686","CreationDate":"2016-06-22T23:10:45.150","ParentId":null,"OwnerUserId":"4727781","Title":null,"Body":"<p>Just to answer the questions more specifically, here's a definition of epoch and loss:<\/p>\n\n<p><strong>Epoch<\/strong>: A full pass over all of your <em>training<\/em> data. <\/p>\n\n<p>For example, in your view above, you have 1213 observations. So an epoch concludes when it has finished a training pass over all 1213 of your observations. <\/p>\n\n<p><strong>Loss<\/strong>: A scalar value that we attempt to minimize during our training of the model. The lower the loss, the closer our predictions are to the true labels. <\/p>\n\n<p>This is usually Mean Squared Error (MSE) as David Maust said above, or often in Keras, <a href=\"http:\/\/keras.io\/backend\/#categorical_crossentropy\" rel=\"noreferrer\" title=\"Categorical Cross-Entropy\">Categorical Cross Entropy<\/a><\/p>\n\n<hr>\n\n<p>What you'd expect to see from running fit on your Keras model, is a decrease in loss over n number of epochs. Your training run is rather abnormal, as your loss is actually increasing. This <em>could<\/em> be due to a learning rate that is too large, which is causing you to overshoot optima. <\/p>\n\n<p>As jaycode mentioned, you will want to look at your model's performance on unseen data, as this is the general use case of Machine Learning. <\/p>\n\n<p>As such, you should include a list of metrics in your compile method, which could look like:<\/p>\n\n<pre><code>model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>As well as run your model on validation during the fit method, such as: <\/p>\n\n<pre><code>model.fit(data, labels, validation_split=0.2)\n<\/code><\/pre>\n\n<hr>\n\n<p>There's a lot more to explain, but hopefully this gets you started.<\/p>\n"},{"AnswerId":"37865292","CreationDate":"2016-06-16T17:01:28.207","ParentId":null,"OwnerUserId":"278191","Title":null,"Body":"<p>One epoch ends when your model had run the data through all nodes in your network and ready to update the weights to reach optimal loss value. That is, smaller is better. In your case, as there are higher loss scores on higher epoch, it \"seems\" the model is better on first epoch.<\/p>\n\n<p>I said \"seems\" since we can't actually tell for sure yet as the model has not been tested using proper cross validation method i.e. it is evaluated only against its training data.<\/p>\n\n<p>Ways to improve your model:<\/p>\n\n<ul>\n<li>Use cross validation in your Keras model in order to find out how the model actually perform, does it generalize well when predicting new data it has never seen before?<\/li>\n<li>Adjust your learning rate, structure of neural network model, number of hidden units \/ layers, init, optimizer, and activator parameters used in your model among myriad other things.<\/li>\n<\/ul>\n\n<p>Combining sklearn's GridSearchCV with Keras can automate this process.<\/p>\n"}]}
{"QuestionId":34681276,"AnswerCount":1,"Tags":"<python><anaconda><torch>","CreationDate":"2016-01-08T16:06:25.493","AcceptedAnswerId":34709776.0,"OwnerUserId":2876684.0,"Title":"Anaconda and Torch7 Install Issue Mac 10.10","Body":"<p>I had Anaconda installed prior to install Torch7.\nI experienced no errors upon install Torch7 <a href=\"http:\/\/torch.ch\/docs\/getting-started.html#_\" rel=\"nofollow\">using these instructions<\/a>.<\/p>\n\n<p>I'm running OS 10.10.5<\/p>\n\n<p>My .bash_profile is the following:<\/p>\n\n<pre><\/pre>\n\n<p>After I updating the .bash_profile I ran<\/p>\n\n<pre><\/pre>\n\n<p>And I get...<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas? Should I remove Anaconda and try the installs again?<\/p>\n","answers":[{"AnswerId":"34709776","CreationDate":"2016-01-10T19:21:14.103","ParentId":null,"OwnerUserId":"2673526","Title":null,"Body":"<p><code>torch_activate<\/code> is not a directory. It should be sourced into your shell instead.<\/p>\n\n<p>Try updating your <code>.bash_profile<\/code> to include<\/p>\n\n<pre><code>source \/Users\/myusername\/torch\/install\/bin\/torch-activate\n<\/code><\/pre>\n"}]}
{"QuestionId":34682174,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-01-08T16:55:08.910","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Iterate queue element until outofrange","Body":"<p>I want to iterate every element of a data queue once () but I don't know in advance how many items I have. Queues have  option and can throw an  error. How can I catch this in a while loop ?<\/p>\n","answers":[]}
{"QuestionId":34685947,"AnswerCount":4,"Tags":"<indexing><addition><tensorflow>","CreationDate":"2016-01-08T20:58:51.647","AcceptedAnswerId":34686952.0,"OwnerUserId":5451067.0,"Title":"Adjust Single Value within Tensor -- TensorFlow","Body":"<p>I feel embarrassed asking this, but how do you adjust a single value within a tensor? Suppose you want to add '1' to only one value within your tensor?<\/p>\n\n<p>Doing it by indexing doesn't work:<\/p>\n\n<pre><\/pre>\n\n<p>One approach would be to build an identically shaped tensor of 0's. And then adjusting a 1 at the position you want. Then you would add the two tensors together. Again this runs into the same problem as before.<\/p>\n\n<p>I've read through the API docs several times and can't seem to figure out how to do this. Thanks in advance! <\/p>\n","answers":[{"AnswerId":"41418052","CreationDate":"2017-01-01T18:47:36.950","ParentId":null,"OwnerUserId":"6845877","Title":null,"Body":"<p>How about <code>tf.scatter_update(ref, indices, updates)<\/code> or <code>tf.scatter_add(ref, indices, updates)<\/code>?<\/p>\n\n<pre><code>ref[indices[...], :] = updates\nref[indices[...], :] += updates\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sparse_variable_updates#scatter_update\" rel=\"noreferrer\">this<\/a>.<\/p>\n"},{"AnswerId":"55954490","CreationDate":"2019-05-02T14:20:51.617","ParentId":null,"OwnerUserId":"11442270","Title":null,"Body":"<p><code>tf.scatter_update<\/code> has no gradient descent operator assigned and will generate an error while learning with at least <code>tf.train.GradientDescentOptimizer<\/code>. You have to implement bit manipulation with low level functions.<\/p>\n"},{"AnswerId":"34686952","CreationDate":"2016-01-08T22:09:26.140","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>UPDATE:<\/strong> TensorFlow 1.0 includes a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_nd\" rel=\"noreferrer\"><code>tf.scatter_nd()<\/code><\/a> operator, which can be used to create <code>delta<\/code> below without creating a <code>tf.SparseTensor<\/code>.<\/p>\n\n<hr>\n\n<p>This is actually surprisingly tricky with the existing ops! Perhaps somebody can suggest a nicer way to wrap up the following, but here's one way to do it.<\/p>\n\n<p>Let's say you have a <code>tf.constant()<\/code> tensor:<\/p>\n\n<pre><code>c = tf.constant([[0.0, 0.0, 0.0],\n                 [0.0, 0.0, 0.0],\n                 [0.0, 0.0, 0.0]])\n<\/code><\/pre>\n\n<p>...and you want to add <code>1.0<\/code> at location [1, 1]. One way you could do this is to define a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#SparseTensor\" rel=\"noreferrer\"><code>tf.SparseTensor<\/code><\/a>, <code>delta<\/code>, representing the change:<\/p>\n\n<pre><code>indices = [[1, 1]]  # A list of coordinates to update.\n\nvalues = [1.0]  # A list of values corresponding to the respective\n                # coordinate in indices.\n\nshape = [3, 3]  # The shape of the corresponding dense tensor, same as `c`.\n\ndelta = tf.SparseTensor(indices, values, shape)\n<\/code><\/pre>\n\n<p>Then you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#sparse_tensor_to_dense\" rel=\"noreferrer\"><code>tf.sparse_tensor_to_dense()<\/code><\/a> op to make a dense tensor from <code>delta<\/code> and add it to <code>c<\/code>:<\/p>\n\n<pre><code>result = c + tf.sparse_tensor_to_dense(delta)\n\nsess = tf.Session()\nsess.run(result)\n# ==&gt; array([[ 0.,  0.,  0.],\n#            [ 0.,  1.,  0.],\n#            [ 0.,  0.,  0.]], dtype=float32)\n<\/code><\/pre>\n"},{"AnswerId":"58295960","CreationDate":"2019-10-09T01:42:12.557","ParentId":null,"OwnerUserId":"1699075","Title":null,"Body":"<p>I feel like the fact that <code>tf.assign<\/code>, <code>tf.scatter_nd<\/code>, <code>tf.scatter_update<\/code> functions only work on <code>tf.Variables<\/code> is not stressed enough. So there it is. <\/p>\n\n<p>And in later versions of TensorFlow (tested with 1.14), you can use indexing on a <code>tf.Variable<\/code> to assign values to specific indices (again this only works on <code>tf.Variable<\/code> objects).<\/p>\n\n<pre><code>v = tf.Variable(tf.constant([[1,1],[2,3]]))\nchange_v = v[0,0].assign(4)\nwith tf.Session() as sess:\n  tf.global_variables_initializer().run()\n  print(sess.run(change_v))\n<\/code><\/pre>\n"}]}
{"QuestionId":34687761,"AnswerCount":2,"Tags":"<python><tensorflow><gradient>","CreationDate":"2016-01-08T23:24:38.503","AcceptedAnswerId":36613269.0,"OwnerUserId":1576818.0,"Title":"Efficiently grab gradients from TensorFlow?","Body":"<p>I'm trying to implement an asynchronous parameter server, DistBelief style using TensorFlow. I found that minimize() is split into two functions, compute_gradients and apply_gradients, so my plan is to insert a network boundary between them. I have a question about how to evaluate all the gradients simultaneously and pull them out all at once. I understand that eval only evaluates the subgraph necessary, but it also only returns one tensor, not the chain of tensors required to compute that tensor.<\/p>\n\n<p>How can I do this more efficiently? I took the Deep MNIST example as a starting point:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I think this last for loop is actually recalculating the last gradient several times, whereas the first gradient is computed only once? How can I grab all the gradients without recomputing them?<\/p>\n","answers":[{"AnswerId":"38131508","CreationDate":"2016-06-30T19:14:13.273","ParentId":null,"OwnerUserId":"3167448","Title":null,"Body":"<p>I coded up a very simple example with comments (inspired from the above answer) that is runnable to see gradient descent in action:<\/p>\n\n<pre><code>import tensorflow as tf\n\n#funciton to transform gradients\ndef T(g, decay=1.0):\n    #return decayed gradient\n    return decay*g\n\n# x variable\nx = tf.Variable(10.0,name='x')\n# b placeholder (simualtes the \"data\" part of the training)\nb = tf.placeholder(tf.float32)\n# make model (1\/2)(x-b)^2\nxx_b = 0.5*tf.pow(x-b,2)\ny=xx_b\n\nlearning_rate = 1.0\nopt = tf.train.GradientDescentOptimizer(learning_rate)\n# gradient variable list = [ (gradient,variable) ]\ngv = opt.compute_gradients(y,[x])\n# transformed gradient variable list = [ (T(gradient),variable) ]\ndecay = 0.1 # decay the gradient for the sake of the example\ntgv = [(T(g,decay=decay),v) for (g,v) in gv] #list [(grad,var)]\n# apply transformed gradients (this case no transform)\napply_transform_op = opt.apply_gradients(tgv)\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    epochs = 10\n    for i in range(epochs):\n        b_val = 1.0 #fake data (in SGD it would be different on every epoch)\n        print '----'\n        x_before_update = x.eval()\n        print 'before update',x_before_update\n\n        # compute gradients\n        grad_vals = sess.run([g for (g,v) in gv], feed_dict={b: b_val})\n        print 'grad_vals: ',grad_vals\n        # applies the gradients\n        result = sess.run(apply_transform_op, feed_dict={b: b_val})\n\n        print 'value of x should be: ', x_before_update - T(grad_vals[0], decay=decay)\n        x_after_update = x.eval()\n        print 'after update', x_after_update\n<\/code><\/pre>\n\n<p>you can observe the change in the variable as its trained and also the value of the gradient. Note that the only reason T decays the gradient because otherwise it reaches the global minimum in 1 step.<\/p>\n\n<hr>\n\n<p>As an extra bonus, if you want to see it work with tensorboard, here you go! :)<\/p>\n\n<pre><code>## run cmd to collect model: python quadratic_minimizer.py --logdir=\/tmp\/quaratic_temp\n## show board on browser run cmd: tensorboard --logdir=\/tmp\/quaratic_temp\n## browser: http:\/\/localhost:6006\/\n\nimport tensorflow as tf\n\n#funciton to transform gradients\ndef T(g, decay=1.0):\n    #return decayed gradient\n    return decay*g\n\n# x variable\nx = tf.Variable(10.0,name='x')\n# b placeholder (simualtes the \"data\" part of the training)\nb = tf.placeholder(tf.float32)\n# make model (1\/2)(x-b)^2\nxx_b = 0.5*tf.pow(x-b,2)\ny=xx_b\n\nlearning_rate = 1.0\nopt = tf.train.GradientDescentOptimizer(learning_rate)\n# gradient variable list = [ (gradient,variable) ]\ngv = opt.compute_gradients(y,[x])\n# transformed gradient variable list = [ (T(gradient),variable) ]\ndecay = 0.9 # decay the gradient for the sake of the example\ntgv = [ (T(g,decay=decay), v) for (g,v) in gv] #list [(grad,var)]\n# apply transformed gradients (this case no transform)\napply_transform_op = opt.apply_gradients(tgv)\n\n(dydx,_) = tgv[0]\nx_scalar_summary = tf.scalar_summary(\"x\", x)\ngrad_scalar_summary = tf.scalar_summary(\"dydx\", dydx)\n\nwith tf.Session() as sess:\n    merged = tf.merge_all_summaries()\n    tensorboard_data_dump = '\/tmp\/quaratic_temp'\n    writer = tf.train.SummaryWriter(tensorboard_data_dump, sess.graph)\n\n    sess.run(tf.initialize_all_variables())\n    epochs = 14\n    for i in range(epochs):\n        b_val = 1.0 #fake data (in SGD it would be different on every epoch)\n        print '----'\n        x_before_update = x.eval()\n        print 'before update',x_before_update\n\n        # get gradients\n        #grad_list = [g for (g,v) in gv]\n        (summary_str_grad,grad_val) = sess.run([merged] + [dydx], feed_dict={b: b_val})\n        grad_vals = sess.run([g for (g,v) in gv], feed_dict={b: b_val})\n        print 'grad_vals: ',grad_vals\n        writer.add_summary(summary_str_grad, i)\n\n        # applies the gradients\n        [summary_str_apply_transform,_] = sess.run([merged,apply_transform_op], feed_dict={b: b_val})\n        writer.add_summary(summary_str_apply_transform, i)\n\n        print 'value of x after update should be: ', x_before_update - T(grad_vals[0], decay=decay)\n        x_after_update = x.eval()\n        print 'after update', x_after_update\n<\/code><\/pre>\n"},{"AnswerId":"36613269","CreationDate":"2016-04-14T03:18:38.837","ParentId":null,"OwnerUserId":"1749667","Title":null,"Body":"<p>Just give you a simple example. Understand it and try your specific task out.<\/p>\n\n<p>Initialize required symbols.<\/p>\n\n<pre><code>x = tf.Variable(0.5)\ny = x*x\nopt = tf.train.AdagradOptimizer(0.1)\ngrads = opt.compute_gradients(y)\ngrad_placeholder = [(tf.placeholder(\"float\", shape=grad[1].get_shape()), grad[1] for grad in grads]\napply_placeholder_op = opt.apply_gradients(grad_placeholder)\ntransform_grads = [(function1(grad[0]), grad[1]) for grad in grads]\napply_transform_op = opt.apply_gradients(transform_grads)\n<\/code><\/pre>\n\n<p>Initialize<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.initialize_all_variables())\n<\/code><\/pre>\n\n<p>Get all gradients<\/p>\n\n<pre><code>grad_vals = sess.run([grad[0] for grad in grads])\n<\/code><\/pre>\n\n<p>Apply gradients<\/p>\n\n<pre><code>feed_dict = {}\nfor i in xrange(len(grad_placeholder)):\n    feed_dict[grad_placeholder[i][0]] = function2(grad_vals[i])\nsess.run(apply_placeholder_op, feed_dict=feed_dict)\nsess.run(apply_transform_op)\n<\/code><\/pre>\n\n<p>Note: the code hasn't been tested by myself, but I confirm the code is legal except minor code errors.\nNote: function1 and function2 is kind of computation, such as 2*x, x^e or e^x and so on.<\/p>\n\n<p>Refer: <a href=\"https:\/\/stackoverflow.com\/questions\/34910452\/tensorflow-apply-gradients-remotely\">TensorFlow apply_gradients remotely<\/a><\/p>\n"}]}
{"QuestionId":34689114,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-09T02:34:17.463","AcceptedAnswerId":null,"OwnerUserId":5765533.0,"Title":"Where is the external code (unsupported\/Eigen\/CXX11\/Tensor) for TensorFlow github:","Body":"<p>I would like to get the TensorFlow GitHub code and compile it on my local Windows machine.<\/p>\n\n<p>I currently have problems with the following include:  <\/p>\n\n<pre><\/pre>\n\n<p>Where can this file be found?<\/p>\n","answers":[{"AnswerId":"34753456","CreationDate":"2016-01-12T20:46:33.300","ParentId":null,"OwnerUserId":"3309044","Title":null,"Body":"<p>Eigen is downloaded from bitbucket by bazel. Version ce5a455 corresponds to the <a href=\"http:\/\/bitbucket.org\/eigen\/eigen\/get\/3.3-beta1.tar.bz2\" rel=\"nofollow\">3.3-beta1 release<\/a> of the Eigen codebase, which had some issues. As of today, TensorFlow is pulling <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/get\/5651786.tar.gz\" rel=\"nofollow\">this newer version<\/a>. Let me know if you still experience problems and I'll take a look.<\/p>\n"},{"AnswerId":"42405691","CreationDate":"2017-02-23T01:58:54.460","ParentId":null,"OwnerUserId":"7346290","Title":null,"Body":"<p>please run the shell below first<\/p>\n\n<p>tensorflow\/contrib\/makefile\/build_all_%os_name%.sh<\/p>\n"}]}
{"QuestionId":34689668,"AnswerCount":0,"Tags":"<c++><multithreading><qt><blocking><caffe>","CreationDate":"2016-01-09T04:27:44.553","AcceptedAnswerId":null,"OwnerUserId":568145.0,"Title":"Caffe feature extraction blocks other executing threads (Qt \/ C++)","Body":"<p><strong>Background<\/strong><\/p>\n\n<p>I am developing a Qt application with three threads: , , and .<\/p>\n\n<ul>\n<li><p> creates, starts, and displays the results of  and\n, while also iteratively feeding them input.<\/p><\/li>\n<li><p> performs an intensive computation (~1s) once every <em>n<\/em>\ninputs, ignoring all other inputs. The majority of this time is spent on feature extraction\nusing the <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">Caffe framework<\/a>.<\/p><\/li>\n<li><p> performs fast computations (~20ms) on every input. But\nevery <em>n+1<\/em> inputs depend on the output of feeding input <em>n<\/em> to .<\/p><\/li>\n<\/ul>\n\n<p>During thread execution,  appears to block  when extracting features using the Caffe network. However,  does <strong>not<\/strong> block  during other processing steps (e.g. network input preprocessing).<\/p>\n\n<p>At first, I thought this was caused by an unmet dependency: i.e.  \"blocks\"  because input (for example) <em>2n + 1<\/em> is ready to be processed by  but input <em>2n<\/em> has not been fully processed by .<\/p>\n\n<p>However, from analysing the execution flow, I noticed this \"blocking\" behaviour was occurring whilst dependencies were met: i.e. let <em>n<\/em> = 10,  would pause execution at input 15 while  was extracting Caffe features from input 20.<\/p>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>How do I prevent  from blocking  during Caffe feature extraction?<\/p>\n\n<p><strong>Code<\/strong><\/p>\n\n<p>Below is a stripped-down version of my code, which shows the key components and logic of my program.<\/p>\n\n<p>I have highlighted the problem in comments , which can be found in  and .<\/p>\n\n<p>main.cpp:<\/p>\n\n<pre><\/pre>\n\n<p>thread1worker.cpp:<\/p>\n\n<pre><\/pre>\n\n<p>thread2worker.cpp:<\/p>\n\n<pre><\/pre>\n\n<p>featureengine.cpp:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34690906,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras>","CreationDate":"2016-01-09T07:30:22.537","AcceptedAnswerId":null,"OwnerUserId":4815753.0,"Title":"cant load mnist dataset in keras","Body":"<p>im trying to load the mnist dataset by :<\/p>\n\n<pre><\/pre>\n\n<p>but i get this error :<\/p>\n\n<pre><\/pre>\n\n<p>what should i do?<\/p>\n\n<p>thanks<\/p>\n","answers":[{"AnswerId":"54419806","CreationDate":"2019-01-29T11:14:34.510","ParentId":null,"OwnerUserId":"10159144","Title":null,"Body":"<p><strong>MNIST database of handwritten digits<\/strong><\/p>\n\n<p>Dataset of 60,000 28x28 grayscale images of the 10 digits, along with a test set of 10,000 images.<\/p>\n\n<p>Usage:<\/p>\n\n<pre><code>  from keras.datasets import mnist\n\n  (x_train, y_train), (x_test, y_test) = mnist.load_data()\n<\/code><\/pre>\n\n<p>Returns:<\/p>\n\n<p>2 tuples:\nx_train, x_test: uint8 array of grayscale image data with shape (num_samples, 28, 28).<\/p>\n\n<p>y_train, y_test: uint8 array of digit labels (integers in range 0-9) with shape (num_samples,).\nArguments:<\/p>\n\n<p>Note(path: if you do not have the index file locally (at '~\/.keras\/datasets\/' + path), it will be downloaded to this location.)<\/p>\n"}]}
{"QuestionId":34693708,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-09T12:55:58.440","AcceptedAnswerId":null,"OwnerUserId":1716733.0,"Title":"Tensorflow summary: how to get items before evaluation","Body":"<p>I would like to get a list of summaries created with a model in . I know that <a href=\"https:\/\/stackoverflow.com\/questions\/34568675\/parsing-summary-str-byte-string-evaluated-on-tensorflow-summary-object\/34569074#34569074\">one can get a dictionary of key:value of evaluated summaries<\/a>. <\/p>\n\n<p>Can one get a list of summary keys before evaluating them from  object? I need it to initialize a dictionary of lists where I will store summaries from each epoch, instead of storing list of dictionaries.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34714698","CreationDate":"2016-01-11T05:25:35.157","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It would probably be easiest to initialize your dictionary of lists on demand, using the following code:<\/p>\n\n<pre><code>train_op = ...\nsummary_op = tf.merge_all_summaries()\n\nsummaries = {}\n\nsess = tf.Session()\n\nfor _ in range(NUM_EPOCHS):\n  _, summary_str = sess.run([train_op, summary_op], feed_dict=feed_dict)\n  summary_proto = tf.Summary()\n  summary_proto.ParseFromString(summary_str)\n\n  for val in summary_proto.value:\n    try:\n      list_for_tag = summaries[val.tag]\n    except KeyError:\n      list_for_tag = []\n      summaries[val.tag] = list_for_tag\n\n    # Assuming all summaries are scalars.\n    list_for_tag.append(val.simple_value)\n<\/code><\/pre>\n\n<hr>\n\n<p>However, to answer your original question, it is possible to get the individual tags by evaluating the <code>tag<\/code> <strong>inputs<\/strong> to the individual summary ops (which most likely do not depend on the result of training):<\/p>\n\n<pre><code>summaries = {}\n\nsess = tf.Session()\n\nall_summary_tensors = tf.get_collection(tf.GraphKeys.SUMMARIES)\n\nfor summary_t in all_summary_tensors:\n  tag_input = summary_t.op.inputs[0]  # The tag input is the 0th input.\n  tags = sess.run(tag_input)\n\n  if isinstance(tags, str):\n    summaries[tags] = []\n  else:\n    for tag in tags.flatten():\n      summaries[tag] = []\n<\/code><\/pre>\n"}]}
{"QuestionId":34694701,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-09T14:32:29.313","AcceptedAnswerId":34695116.0,"OwnerUserId":5766785.0,"Title":"TensorFlow - which Docker image to use?","Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html\" rel=\"nofollow\">TensorFlow Download and Setup<\/a> under  <\/p>\n\n<p>Docker installation I see:<\/p>\n\n<ol>\n<li>b.gcr.io\/tensorflow\/tensorflow        latest              4ac133eed955        653.1 MB<\/li>\n<li>b.gcr.io\/tensorflow\/<strong>tensorflow        latest-devel<\/strong>        6a90f0a0e005        2.111 GB<\/li>\n<li>b.gcr.io\/tensorflow\/<strong>tensorflow-full   latest<\/strong>              edc3d721078b        2.284 GB<\/li>\n<\/ol>\n\n<p>I know 2. &amp; 3. are with source code and I am using 2. for now. <\/p>\n\n<p>What is the difference between 2. &amp; 3. ? \nWhich one is recommended for \"normal\" use? <\/p>\n\n<p>TLDR:<\/p>\n\n<p>First of all - thanks for Docker images! They are the easiest and cleanest way to start with TF. <\/p>\n\n<p>Few aside things about images<\/p>\n\n<ul>\n<li>there is no PIL <\/li>\n<li>there is no nano (but there is vi) and apt-get cannot find it. yes i probable can configure repos for it, but why not out of the box<\/li>\n<\/ul>\n","answers":[{"AnswerId":"34695116","CreationDate":"2016-01-09T15:09:18.573","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>There are four images:<\/p>\n\n<ol>\n<li>b.gcr.io\/tensorflow\/tensorflow: TensorFlow CPU binary image.  <\/li>\n<li>b.gcr.io\/tensorflow\/tensorflow:latest-devel: CPU Binary image plus source code.  <\/li>\n<li>b.gcr.io\/tensorflow\/tensorflow:latest-gpu: TensorFlow GPU binary image.<\/li>\n<li>gcr.io\/tensorflow\/tensorflow:latest-devel-gpu: GPU Binary image plus source code.<\/li>\n<\/ol>\n\n<p>And the two properties of concern are:<br>\n 1. <code>CPU<\/code> or <code>GPU<\/code><br>\n 2. <code>no source<\/code> or <code>plus source<\/code>  <\/p>\n\n<p><code>CPU<\/code> or <code>GPU<\/code>: <code>CPU<\/code> <\/p>\n\n<p>For a first time user it is highly recommended to avoid the GPU version as they can be any where from difficult to impossible to use. The reason is that not all machines have an NVidia graphic chip that meet the requirements. You should first get TensorFlow working to understand it then move onto using the GPU version if you want\/need.<\/p>\n\n<p>From TensorFlow Build Instructions  <\/p>\n\n<blockquote>\n  <p><strong>Optional: Install CUDA (GPUs on Linux)<\/strong><\/p>\n  \n  <p>In order to build or run TensorFlow with GPU support, both Cuda\n  Toolkit 7.0 and CUDNN 6.5 V2 from NVIDIA need to be installed.<\/p>\n  \n  <p>TensorFlow GPU support requires having a GPU card with \n  <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">NVidia Compute Capability >= 3.5<\/a>. Supported cards include but are not limited to:  <\/p>\n  \n  <p>NVidia Titan<br>\n  NVidia Titan X<br>\n  NVidia K20<br>\n  NVidia K40 <\/p>\n<\/blockquote>\n\n<p><code>no source<\/code> or <code>plus source<\/code>: <code>no source<\/code><\/p>\n\n<p>The docker images will work without needing the source. You should only want or need the source if you need to rebuild TensorFlow for some reason such as adding a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/how_tos\/adding_an_op\/index.html\" rel=\"nofollow noreferrer\">new OP<\/a>. <\/p>\n\n<p>The standard recommendation for someone new to using TensorFlow is to start with the CPU version without the source.<\/p>\n"}]}
{"QuestionId":34694920,"AnswerCount":1,"Tags":"<java><lru><caffeine>","CreationDate":"2016-01-09T14:51:15.053","AcceptedAnswerId":34696626.0,"OwnerUserId":2051659.0,"Title":"LRU with Caffeine","Body":"<p>I'm trying to use Caffeine as LRU cache, so entries that were added first will be evicted first.\nRan this code:<\/p>\n\n<pre><\/pre>\n\n<p>Which prints :<\/p>\n\n<pre><\/pre>\n\n<p>But I expected <\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"34696626","CreationDate":"2016-01-09T17:26:56.323","ParentId":null,"OwnerUserId":"786223","Title":null,"Body":"<p>Caffeine does not implement LRU as its cache eviction policy.  Instead, Caffeine uses a policy called <a href=\"http:\/\/arxiv.org\/pdf\/1512.00727.pdf\" rel=\"noreferrer\">TinyLFU<\/a>.  The Caffeine documentation includes a page on <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/wiki\/Efficiency\" rel=\"noreferrer\">Efficiency<\/a>, which discusses the rationale for this design choice.  Quoting that page:<\/p>\n\n<blockquote>\n  <p><code>TinyLfu<\/code> relies on a frequency sketch to probabilistically estimate the historic usage of an entry.<\/p>\n<\/blockquote>\n\n<p>Since Caffeine does not in fact implement LRU, I don't think you can reliably expect it to exhibit strict LRU behavior when you inspect the entries in the cache.<\/p>\n\n<p>If you absolutely must have LRU behavior, then the JDK standard <a href=\"http:\/\/docs.oracle.com\/javase\/7\/docs\/api\/java\/util\/LinkedHashMap.html\" rel=\"noreferrer\"><code>LinkedHashMap<\/code><\/a> is a good, straightforward choice.  You would need to subclass it and override <a href=\"http:\/\/docs.oracle.com\/javase\/7\/docs\/api\/java\/util\/LinkedHashMap.html#removeEldestEntry(java.util.Map.Entry)\" rel=\"noreferrer\"><code>removeEldestEntry<\/code><\/a> with logic to signal when the cache has grown larger than you want it.  If multi-threaded usage is necessary, then you'll need to wrap operations with appropriate synchronization.<\/p>\n\n<p>Caffeine was heavily inspired by the <a href=\"https:\/\/code.google.com\/p\/guava-libraries\/wiki\/CachesExplained\" rel=\"noreferrer\">Guava Cache<\/a>, which similarly provides concurrent access and has approximate LRU behavior.  A quick test of your code against the Guava cache shows similar results.  I am not aware of any standard library that would provide predictable, externally observable LRU results and true concurrent access without a coarse-grained lock.<\/p>\n\n<p>You might reconsider whether it's really a requirement to have strict, externally observable LRU results.  By its nature, a cache is fast temporary storage to provide optimized lookups.  I would not expect program behavior to change drastically dependent on whether the cache implements strict LRU, an approximation of LRU, LFU, or some other eviction policy.<\/p>\n\n<p>This prior question also has a great discussion of LRU cache options in Java.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/221525\/how-would-you-implement-an-lru-cache-in-java-6\">How would you implement an LRU cache in Java?<\/a><\/p>\n"}]}
{"QuestionId":34694965,"AnswerCount":2,"Tags":"<python><image-recognition><tensorflow>","CreationDate":"2016-01-09T14:55:29.783","AcceptedAnswerId":38526641.0,"OwnerUserId":1055357.0,"Title":"Image recognition using TensorFlow","Body":"<p>I'm new to TensorFlow and I am looking for help on image recognition. Is there an example that showcases how to use TensorFlow to train your own digital images for image recognition like the image-net model used in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html#image-recognition\">TensorFlow image recognition tutorial<\/a><\/p>\n\n<p>I looked at the CIFAR-10 model training but it doesn't seem to provide examples for training your own images.<\/p>\n","answers":[{"AnswerId":"44535991","CreationDate":"2017-06-14T05:17:57.110","ParentId":null,"OwnerUserId":"5248003","Title":null,"Body":"<p>This repo gives final layer image retraining for your own images.<\/p>\n\n<p>Add your images to the training_\"dataset folder\" <\/p>\n\n<p>Add more and more images(atleast 30 images for each class required.)<\/p>\n\n<p>First execute the bash script to download the Imagenet files.<\/p>\n\n<p>Next, run the python script to retrain your images which outputs validation accuracy etc.,<\/p>\n\n<p><a href=\"https:\/\/github.com\/shivakrishna9\/tensorflow-retrain\" rel=\"nofollow noreferrer\">https:\/\/github.com\/shivakrishna9\/tensorflow-retrain<\/a><\/p>\n"},{"AnswerId":"38526641","CreationDate":"2016-07-22T12:37:43.710","ParentId":null,"OwnerUserId":"5463996","Title":null,"Body":"<p>I would recommend using Google's trained Inception model to do image recognition.\nPlease refer to the example \"How to Retrain Inception's Final Layer for New Categories\" on tensorflow website. It is at <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/how_tos\/image_retraining\/index.html\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.9\/how_tos\/image_retraining\/index.html<\/a>.<\/p>\n\n<p>Using trained model is easy and can achieve reasonable accuracy. You just simply feed the model with your own data set. The last classication layer of Google's inception will be modified and we only train the last layer. For several thousand images with in several categories, it only take several hours to finish training.\nPlease note: in order to use the example, you have to build tensorflow from source.<\/p>\n\n<p>I am using the transfer learning feature and achieving very good results. To illustrate the benefit of transfer learning, I am comparing \"Transfer Learning on Trained GoogleNet\" with \"Build and train a 5-layer-Convnet from scratch\". The classification task is done on 5000 images with 5 categories.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7XqIq.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7XqIq.png\" alt=\"Build a 5-layer-convet and train it from scratch\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HXEp8.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HXEp8.png\" alt=\"Use the transfer learning on the trained GoogleNet\"><\/a><\/p>\n\n<p>See this simple example: <a href=\"https:\/\/www.youtube.com\/watch?v=QfNvhPx5Px8\" rel=\"noreferrer\">https:\/\/www.youtube.com\/watch?v=QfNvhPx5Px8<\/a> (Build a TensorFlow Image Classifier in 5 Min)<\/p>\n"}]}
{"QuestionId":34696845,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-09T17:47:56.877","AcceptedAnswerId":34846882.0,"OwnerUserId":5766785.0,"Title":"How to see multiple images through tf.image_summary","Body":"<p>Problem - only one image is shown at TensorBoard<\/p>\n\n<p>Inspired by this \n<a href=\"https:\/\/stackoverflow.com\/questions\/33783672\/how-can-i-visualize-the-weightsvariables-in-cnn-in-tensorflow\/33794463#33794463\">How can I visualize the weights(variables) in cnn in Tensorflow?<\/a><\/p>\n\n<p>Here is code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here is output: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Looks ok <\/p>\n\n<ul>\n<li>5 [image label] pairs were delivered<\/li>\n<li>in case I uncomment <strong>print \"imss\",imss<\/strong> I can see 5 different buffers each with own png image <\/li>\n<li>op graph looks ok in TB<\/li>\n<\/ul>\n\n<p>However only one image in TB. I suspect I have missed something important about how TF is working -.i.e. what caused what at graph execution time. <\/p>\n\n<p>Second question: what I need to do to see result i.e. <strong>img2 = img+img<\/strong> in TB?<\/p>\n","answers":[{"AnswerId":"41228796","CreationDate":"2016-12-19T18:30:03.527","ParentId":null,"OwnerUserId":"5307226","Title":null,"Body":"<p>As of r0.12, <code>tf.image_summary<\/code> has been replaced with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/image\" rel=\"nofollow noreferrer\"><code>tf.summary.image<\/code><\/a><\/p>\n\n<pre><code>tf.summary.image(name, tensor, max_outputs=3, collections=None)\n<\/code><\/pre>\n"},{"AnswerId":"34846882","CreationDate":"2016-01-18T03:32:31.817","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>You are right that you will only see one image. You are calling the image summary op once in each for loop, and each time you call it, you are passing it a single image.<\/p>\n\n<p>What you could do to see all images that you want to see, would be to compile these images into a single tensor. If we refer to TensorFlow API (link always changes so find the latest one)<\/p>\n\n<blockquote>\n  <p><s>tf.image_summary(tag, tensor, max_images=3, collections=None,\n  name=None) <\/s><\/p>\n<\/blockquote>\n\n<p>As of TF 1.0.0, it's this:<\/p>\n\n<blockquote>\n  <p>tf.summary.image(name, tensor, max_outputs=3, collections=None)<\/p>\n<\/blockquote>\n\n<p>Put your \"multiple image tensor\" in, set max_images to the number of images you have, and you should be able to see all the images in TensorBoard.<\/p>\n\n<p>Let me know if there are still problems.<\/p>\n"}]}
{"QuestionId":34697519,"AnswerCount":1,"Tags":"<image-processing><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2016-01-09T18:48:35.997","AcceptedAnswerId":34703380.0,"OwnerUserId":978912.0,"Title":"Caffe: variable input-image size","Body":"<p>I am trying out Google's <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"nofollow\">deepdream code<\/a> which makes use of Caffe.\nThey use the GoogLeNet model pre-trained on ImageNet, as provided by the ModelZoo. That means the network was trained on images cropped to the size 224x224 pixel. From the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/train_val.prototxt\" rel=\"nofollow\">train_val.prototext<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>The <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt\" rel=\"nofollow\">deploy.prototext<\/a> used for processing also defines an input layer with the size of 224x224x3x10 (RGB images of size 224x224, batchsize 10).<\/p>\n\n<pre><\/pre>\n\n<p>However I can use this net to process images of any size (the example above used one of 1024x574 pixel).<\/p>\n\n<ol>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_googlenet\/deploy.prototxt\" rel=\"nofollow\">deploy.prototext<\/a> does not configure caffe to use cropping.<\/li>\n<li>The preprocessing in the <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"nofollow\">deepdream code<\/a> only does demeaning, also no cropping here<\/li>\n<\/ol>\n\n<p>How is it possible that I can run on images which are too big for the input layer?<\/p>\n\n<hr>\n\n<p>complete code can be found <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"nofollow\">here<\/a><\/p>\n","answers":[{"AnswerId":"34703380","CreationDate":"2016-01-10T08:07:21.463","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>DeepDream does not crop the input image at all.<br>\nIf you pay close attention you'll notice that it operates on a mid-level layer: it's <code>end=<\/code> argument is set to <code>'inception_4c\/output'<\/code> or <code>end='inception_3b\/5x5_reduce'<\/code>, but NEVER <code>end='loss3\/classifier'<\/code>.  The reason for this is that the GoogLeNet up to these layers is a <em>fully-convolutional<\/em> net, that is, it can take any sized input image and produce outputs of sizes proportional to the input size (output size is usually affected by conv padding and pooling).<\/p>\n\n<p>To adjust the net to the different sizes of inputs the function <code>deepdream<\/code> has the line <\/p>\n\n<pre><code>src.reshape(1,3,h,w) # resize the network's input image size\n<\/code><\/pre>\n\n<p>This line adjusts the net's layers to accommodate input of shape <code>(1,3,h,w)<\/code>. <\/p>\n"}]}
{"QuestionId":34699069,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2016-01-09T21:19:18.473","AcceptedAnswerId":34770965.0,"OwnerUserId":1243762.0,"Title":"Building TensorFlow from source on Ubuntu 14.04 LTS: gcc: internal compiler error: Killed (program cc1plus)","Body":"<p>I have successfully built TensorFlow from source under Debian but at present cannot get it to build starting with a new virtual machine using <a href=\"http:\/\/www.ubuntu.com\/download\/desktop\" rel=\"nofollow\">Ubuntu 14.04 LTS<\/a>. IIRC  for Debian I tried g++\/gcc 5.2 but had to downgrade to g++\/gcc 4.9 and it worked. Following the instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">Installing from sources<\/a> if I install g++ the version is 4.8 and it failed .<\/p>\n\n<blockquote>\n  <p>gcc: internal compiler error: Killed (program cc1plus)<\/p>\n<\/blockquote>\n\n<p>I have not tired 4.9 yet.<\/p>\n\n<p>I checked the info on the last <a href=\"http:\/\/ci.tensorflow.org\/job\/tensorflow-master\/lastBuild\/\" rel=\"nofollow\">Jenkins build<\/a> but could not find anything listed for the tools and their versions. Even opened issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/736\" rel=\"nofollow\">Build tools and versions listed in Jenkins build log<\/a><\/p>\n\n<p>What version(s) of g++\/gcc is know to work?<br>\nWhat version of g++\/gcc do the build machines use?  <\/p>\n\n<p>EDIT <\/p>\n\n<p>Found this: <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/tools\/ci_build\/README.md\" rel=\"nofollow\">TensorFlow.org Continuous Integration<\/a><\/p>\n","answers":[{"AnswerId":"34770965","CreationDate":"2016-01-13T15:45:29.057","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>The problem is not with the g++\/gcc version but the number of CPU cores Bazel uses to build TensorFlow.<\/p>\n\n<p>In running multiple builds on a VMware Workstation 7.1 with a fresh install of Ubuntu 14.04 LTS with one CPU core, 2G ram, 2G swap partition and 2G swap file the builds run the fastest. This may not be the best setup, but is the best one I have found so far that consistently works. If I allow 4 cores via VMware and build with Bazel it fails. If I limit the resources with the Bazel option <a href=\"http:\/\/bazel.io\/docs\/bazel-user-manual.html#flag--local_resources\" rel=\"nofollow\">--local_resources<\/a> using <\/p>\n\n<pre><code>--local_resources 2048,2.0,1.0\n<\/code><\/pre>\n\n<p>builds successfully<\/p>\n\n<pre><code>INFO: Elapsed time: 11683.908s, Critical Path: 11459.26s\n<\/code><\/pre>\n\n<p>using<\/p>\n\n<pre><code>--local_resources 4096,2.0,1.0\n<\/code><\/pre>\n\n<p>builds successfully<\/p>\n\n<pre><code>INFO: Elapsed time: 39765.257s, Critical Path: 39578.52s\n<\/code><\/pre>\n\n<p>using<\/p>\n\n<pre><code>--local_resources 4096,1.0,1.0\n<\/code><\/pre>\n\n<p>builds successfully<\/p>\n\n<pre><code>INFO: Elapsed time: 6562.744s, Critical Path: 6443.80s\n<\/code><\/pre>\n\n<p>using<\/p>\n\n<pre><code>--local_resources 6144,1.0,1.0\n<\/code><\/pre>\n\n<p>builds successfully<\/p>\n\n<pre><code>INFO: Elapsed time: 2810.509s, Critical Path: 2654.90s\n<\/code><\/pre>\n\n<p>In summary more memory and less CPU cores works best for my environment.<\/p>\n\n<p>TLDR;<\/p>\n\n<p>While keeping an eye during the build process I noticed that certain source files would take a long time to compile and appeared to tie down the flow rate while building. It is as if they are in competition for a resource with other source files and that Bazel does not know about this critical resource so it allows the competing files to compile at the same time. Thus the more files competing with the unknown resource the slower the build.<\/p>\n"}]}
{"QuestionId":34699260,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-09T21:40:57.497","AcceptedAnswerId":34699391.0,"OwnerUserId":62695.0,"Title":"How to serialize both the graph and values in a protobuf file?","Body":"<p>The Android example that comes with Tensorflow downloads a protobuf file for InceptionV3 which contains both the graph and the values from the model. In the docs, I could only find how to serialize the graph () or save the variable values with a . How can you save everything to a single file, as done for that example?<\/p>\n","answers":[{"AnswerId":"34699391","CreationDate":"2016-01-09T21:57:37.530","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I answered a similar question on this topic: <a href=\"https:\/\/stackoverflow.com\/a\/34343517\/3574081\">Is there an example on how to generate protobuf files holding trained Tensorflow graphs?<\/a><\/p>\n\n<p>The basic idea is to use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#import_graph_def\" rel=\"nofollow noreferrer\"><code>tf.import_graph_def()<\/code><\/a> to replace the variables in the original (training) graph with constants, and then write out the resulting <code>GraphDef<\/code> using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph.as_graph_def\" rel=\"nofollow noreferrer\"><code>tf.Graph.as_graph_def()<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":34702041,"AnswerCount":2,"Tags":"<python><machine-learning><keras><data-science>","CreationDate":"2016-01-10T04:23:16.660","AcceptedAnswerId":34705152.0,"OwnerUserId":4984897.0,"Title":"How to tell which Keras model is better?","Body":"<p>I don't understand which accuracy in the output to use to compare my 2 Keras models to see which one is better. <\/p>\n\n<p>Do I use the \"acc\" (from the training data?) one or the \"val acc\" (from the validation data?) one?<\/p>\n\n<p>There are different accs and val accs for each epoch. How do I know the acc or val acc for my model as a whole? Do I average all of the epochs accs or val accs to find the acc or val acc of the model as a whole?<\/p>\n\n<p><strong>Model 1 Output<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Model 2 Output<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34705152","CreationDate":"2016-01-10T12:05:43.957","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<blockquote>\n  <p>Do I use the \"acc\" (from the training data?) one or the \"val acc\" (from the validation data?) one?<\/p>\n<\/blockquote>\n\n<p>If you want to estimate the ability of your model to generalize to new data (which is probably what you want to do), then you look at the validation accuracy, because the validation split contains only data that the model never sees during the training and therefor cannot just memorize.<\/p>\n\n<p>If your training data accuracy (\"acc\") keeps improving while your validation data accuracy (\"val_acc\") gets worse, you are likely in an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Overfitting\">overfitting<\/a> situation, i.e. your model starts to basically just memorize the data.<\/p>\n\n<blockquote>\n  <p>There are different accs and val accs for each epoch. How do I know the acc or val acc for my model as a whole? Do I average all of the epochs accs or val accs to find the acc or val acc of the model as a whole?<\/p>\n<\/blockquote>\n\n<p>Each epoch is a training run over all of your data. During that run the parameters of your model are adjusted according to your loss function. The result is a set of parameters which have a certain ability to generalize to new data. That ability is reflected by the validation accuracy. So think of every epoch as its own model, which can get better or worse if it is trained for another epoch. Whether it got better or worse is judged by the change in validation accuracy (better = validation accuracy increased). Therefore pick the model of the epoch with the highest validation accuracy. Don't average the accuracies over different epochs, that wouldn't make much sense. You can use the Keras callback <code>ModelCheckpoint<\/code> to automatically save the model with the highest validation accuracy (see <a href=\"http:\/\/keras.io\/callbacks\/\">callbacks documentation<\/a>).<\/p>\n\n<p>The highest accuracy in model 1 is <code>0.7737<\/code> and the highest one in model 2 is <code>0.7572<\/code>. Therefore you should view model 1 (at epoch 3) as better. Though it is possible that the <code>0.7737<\/code> was just a random outlier.<\/p>\n"},{"AnswerId":"36359633","CreationDate":"2016-04-01T15:00:49.597","ParentId":null,"OwnerUserId":"627042","Title":null,"Body":"<p>You need to key on decreasing val_loss or increasing val_acc, ultimately it doesn't matter much.   The differences are well within random\/rounding errors.<\/p>\n\n<p>In practice, the training loss can drop significantly due to over-fitting, which is why you want to look at validation loss.<\/p>\n\n<p>In your case, you can see that your training loss is not dropping - which means you are learning nothing after each epoch.   It look like there's nothing to learn in this model, aside from some trivial linear-like fit or cutoff value.<\/p>\n\n<p>Also, when learning nothing, or a trivial linear thing, you should a similar performance on training and validation (trivial learning is always generalizable).  You should probably shuffle your data before using the validation_split feature.<\/p>\n"}]}
{"QuestionId":34703953,"AnswerCount":0,"Tags":"<matlab><hdf5><deep-learning><caffe><conv-neural-network>","CreationDate":"2016-01-10T09:34:48.750","AcceptedAnswerId":null,"OwnerUserId":3698878.0,"Title":"Creating HDF5 from rgb images matlab","Body":"<p>I intend to feed my test and train images to caffe through hdf5. But before that some pre-processing is required, below is my hdf5 conversion code. My images are already 256*256.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>And then, \n<\/p>\n\n<p>I am not sure about the correctness of this method. Please help! (Also I am not doing oversampling here, as I don't know how to handle that in hdf5.)<\/p>\n","answers":[]}
{"QuestionId":34706950,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-10T15:07:18.077","AcceptedAnswerId":34708998.0,"OwnerUserId":1886138.0,"Title":"How many processes does TensorFlow open?","Body":"<p>I am using torque to run some CNN-based learning using  library. (1 CPU per task)<\/p>\n\n<p>When I run  on my server, I noticed that: <\/p>\n\n<p>I create a session like this: <\/p>\n\n<p>So my question is there some place in documentation where I can read when and how many processes TensorFlow uses.<\/p>\n","answers":[{"AnswerId":"34708998","CreationDate":"2016-01-10T18:10:09.807","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The current version of TensorFlow (0.6.0) runs in a <strong>single process<\/strong>: the same process in which you create the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session\" rel=\"nofollow\"><code>tf.Session<\/code><\/a> (or call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassSession.html#class-tensorflow-session\" rel=\"nofollow\"><code>tensorflow::NewSession()<\/code><\/a>, if you're using the C++ interface).<\/p>\n\n<p>However, TensorFlow by default uses <strong>multiple threads<\/strong>, which could account for the high load averages you are seeing. By default TensorFlow will choose the number of threads for two threadpools based on an appropriate number for your system. Currently the default behavior is to allocate the same number of threads as cores, for both the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f9514a917265f0e98c8fb44abc51158685f72ca6\/tensorflow\/core\/common_runtime\/direct_session.cc#L63\" rel=\"nofollow\">\"inter-op\" threadpool<\/a> (which determines the number of ops that can execute in parallel), and the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f9514a917265f0e98c8fb44abc51158685f72ca6\/tensorflow\/core\/common_runtime\/local_device.cc#L39\" rel=\"nofollow\">\"intra-op\" threadpool<\/a> (which determines the number of threads available for parallelizing within an op). Another source of threads is in the Python-based <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#QueueRunner\" rel=\"nofollow\"><code>QueueRunner<\/code><\/a>, which starts at least one thread per queue (typically in an input pipeline).<\/p>\n\n<p>Therefore, if you have a large number of cores, and your TensorFlow program has a lot of available parallelism (or a complex input pipeline), you can end up seeing such high load averages.<\/p>\n"}]}
{"QuestionId":34708722,"AnswerCount":1,"Tags":"<theano><conv-neural-network>","CreationDate":"2016-01-10T17:44:55.290","AcceptedAnswerId":34709371.0,"OwnerUserId":5583416.0,"Title":"Confused about the size of the output of convolution layer in theano deep learning tutorial","Body":"<p>[<a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet]\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet]<\/a><\/p>\n\n<p>in the above link it says \n Construct the first convolutional pooling layer:\n filtering reduces the image size to (28-5+1 , 28-5+1) = (24, 24)<\/p>\n\n<p>convolution of data of size a with filter of size b gives output of size a+b-1. so here the data size 28*28, filter size is 5*5. so the output size should be (28+5-1,28+5-1). it is given as (28-5+1,28-5+1)<\/p>\n","answers":[{"AnswerId":"34709371","CreationDate":"2016-01-10T18:43:54.390","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>It depends on the <code>border_mode<\/code><\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/conv.html\" rel=\"nofollow\"><code>conv2d<\/code><\/a> uses <code>border_mode='valid'<\/code> by default which means (from the <a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.16.0\/reference\/generated\/scipy.signal.convolve2d.html\" rel=\"nofollow\">scipy documentation<\/a>)<\/p>\n\n<blockquote>\n  <p>The output consists only of those elements that do not rely on the\n  zero-padding.<\/p>\n<\/blockquote>\n\n<p>So with <code>border_mode='valid'<\/code> and a <code>(5,5)<\/code> filter the output is going to be the same size as the input minus a two pixel border, i.e. <code>image_shape - filter_shape + 1<\/code>, hence with input size <code>(28,28)<\/code> the output is going to be <code>(24,24)<\/code>.<\/p>\n\n<p>The alternative, <code>border_mode='full'<\/code> will zero pad the input such that the output is of shape <code>image_shape + filter_shape - 1<\/code>.<\/p>\n"}]}
{"QuestionId":34708950,"AnswerCount":1,"Tags":"<module><lua><neural-network><torch>","CreationDate":"2016-01-10T18:06:35.487","AcceptedAnswerId":34733005.0,"OwnerUserId":4380945.0,"Title":"Torch, neural nets - forward function on gmodule object - nngraph class","Body":"<p>I am a newbie to torch and lua (as anyone who has been following my latest posts could attest :) and have the following question on the forward function for the gmodule object (class nngraph).<\/p>\n\n<p>as per the source code (<a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/Module.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/Module.lua<\/a> - as class gmodule inherits from nn.module) the syntax is:<\/p>\n\n<pre><\/pre>\n\n<p>However, I have found cases where a table is passed as input, as in:<\/p>\n\n<pre><\/pre>\n\n<p>where:<\/p>\n\n<pre><\/pre>\n\n<p>is itself a gmodule object. In turn, rnn_state[t-1] is a table with 4 tensors. So in the end, we have something akin to <\/p>\n\n<pre><\/pre>\n\n<p>The question is, depending on the network architecture, can you pass input - formatted as table - not only to the input layer but also to the hidden layers? <\/p>\n\n<p>In that case, you have to check that you pass exactly one input per layer? (with the exception of the output layer)<\/p>\n\n<p>Thanks so much<\/p>\n","answers":[{"AnswerId":"34733005","CreationDate":"2016-01-11T23:25:16.800","ParentId":null,"OwnerUserId":"4380945","Title":null,"Body":"<p>I finally found the answer. The module class (as well as the inherited class gmodule) has an input and an output.<\/p>\n\n<p>However, the input (as well as the output) needs not be a vector, but it could be a collection of vectors - that depends on the neural net configuration, in this particular case it is a pretty complex recursive neural net.<\/p>\n\n<p>So if the net has more than one input vector, you can do:<\/p>\n\n<pre><code>result_var = gmodule:forward{[1]=tensor_1,[2]=tensor_2,[3]=tensor_3,...,[5]=tensor_5}\n<\/code><\/pre>\n\n<p>where each tensor\/vector is one of the input vectors. Only one of those vectors is the X vector, or the feature vector. The others could serve as input to other intermediate nodes.<\/p>\n\n<p>In turn, result_var (which is the output) can have one output as tensor (the prediction) or a collection of tensors as output (a collection of tensors), depending on the network configuration.<\/p>\n\n<p>If the latter is the case, one of those output tensors is the prediction, and the reminder are usually used as input to the intermediate nodes in the next time step - but that again depends on the net configuration.<\/p>\n"}]}
{"QuestionId":34709181,"AnswerCount":1,"Tags":"<python><gpgpu><theano>","CreationDate":"2016-01-10T18:26:19.497","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Theano takes a few minutes to start a script running on GPU","Body":"<p>I start a script running on GPU as follows:<\/p>\n\n<pre><\/pre>\n\n<p>It sometimes takes a few minutes before it even starts: the first line of the script  is a print statement (viz. ), and it will not get displayed before a few minutes.<\/p>\n\n<p>If using the CPU, the print statement immediately appears:<\/p>\n\n<pre><\/pre>\n\n<p>I killed all Python processes beforehand using . Running  shows that before running my script, no process uses any of the GPUs. (I use nVidia GPUs, specifically nVidia GeForce GTX Titan X)<\/p>\n\n<p>I have noticed the issue arises in a screen session (using <a href=\"https:\/\/en.wikipedia.org\/wiki\/GNU_Screen\" rel=\"nofollow\">GNU Screen<\/a>), when it has been created a while ago (> half a day it seems). Killing the screen session and creating a new one solves the issue. However, I don't understand why it fixes the issue.<\/p>\n\n<p>What could explain the issue?<\/p>\n\n<hr>\n\n<p>Some timing:<\/p>\n\n<p>Using the Python script:<\/p>\n\n<pre><\/pre>\n\n<p>run with the command:<\/p>\n\n<pre><\/pre>\n\n<p>I get:<\/p>\n\n<pre><\/pre>\n\n<p>221 seconds is quite a long time to import a library. If I kill the screen and create a new one, the import time is around 1.1 seconds. <\/p>\n\n<p>The slowness does seem reproducible: even if I run the import theano more than once, it still takes ~220 seconds (unless I kill and create a new screen). Also, I don't empty .<\/p>\n","answers":[{"AnswerId":"34855275","CreationDate":"2016-01-18T12:52:43.037","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you empty the directory pointed to by the <code>base_compiledir<\/code> Theano flag and start your script you should expect <code>import theano<\/code> to take an unusually long time because the first import of theano will cause a bunch of compilation to occur behind the scenes. Subsequent runs should <code>import theano<\/code> much quicker because the setup compilation has already happened, as long as the <code>base_compiledir<\/code> has not been cleared in the interim.<\/p>\n\n<p>Since Theano has to check that all the setup compilation is present in the cache, <code>import theano<\/code> can take a while to run even after the first time. This is especially true if the <code>base_compiledir<\/code> is on a networked, or otherwise slow, filesystem.<\/p>\n"}]}
{"QuestionId":34709435,"AnswerCount":1,"Tags":"<matlab><neural-network><deep-learning><caffe><hdf5>","CreationDate":"2016-01-10T18:50:02.360","AcceptedAnswerId":34710019.0,"OwnerUserId":3698878.0,"Title":"Caffe not reading all h5 files","Body":"<p>I have the following h5 files listed in  which I am giving to the hdf5 data layer.<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I have 3 datasets - ,  and  in these files. Initially, I kept all these in 1 h5 file, but since caffe can't handle h5 files bigger than 2 GB, I had to divide X (say X consists of 5000 samples), in 5 parts. In the first h5, I have Meta and Labels stored along with the first part, i.e; 1000 samples of X, and in the remaining 4 h5 files, I have 1000 samples each. \n   When I start finetuning, caffe crashes with the following error message<\/p>\n\n<pre><\/pre>\n\n<p>The main thing, according to me is 'Check failed: hdf_blobs_[i]->num() == num (5000 vs. 1000)' <br>\nFrom which I assume that caffe is reading only the first h5 file, how can I make it read all 5 h5 files? Please help!<\/p>\n","answers":[{"AnswerId":"34710019","CreationDate":"2016-01-10T19:46:13.063","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>How do you expect <code>caffe<\/code> synchronize all your input data across all the files? Do you expect it to read <code>X<\/code> from the second file and <code>Meta<\/code> from the first?<br>\nIf you were to implement <code>\"HDF5Data\"<\/code> layer, how would <em>you<\/em> expect the data to be laid out for you?<\/p>\n\n<p>The way things are implemented in caffe at the moment, ALL variables should be divided between the HDF5 files in the same manner. That is if you decided that  <code>X<\/code> will be divided into 5 files, with e.g., 1000 samples in the first file, 1234 samples in the second etc. Then you <strong>must<\/strong> divide <code>Meta<\/code> and <code>Labels<\/code> in the same manner: <code>train_data1.h5<\/code> will have 1000 samples of <code>X<\/code>, <code>Meta<\/code> and <code>Label<\/code>; <code>train_data2.h5<\/code> will have 1234 samples of <code>X<\/code>, <code>Meta<\/code> and <code>Label<\/code> and so forth.<\/p>\n\n<p>Caffe does not load all the data into memory, it only fetches the batch it needs for the SGD iteration. Therefore, it makes no sense to split the variables across different files. Moreover, It might help if you make the number of samples stored at each file an integer multiplicity of your <code>batch_size<\/code>.<\/p>\n"}]}
{"QuestionId":34709950,"AnswerCount":1,"Tags":"<numpy><theano><flatten><conv-neural-network>","CreationDate":"2016-01-10T19:39:33.047","AcceptedAnswerId":34710378.0,"OwnerUserId":5583416.0,"Title":"Clarification about flatten function in Theano","Body":"<p>in [<a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet]\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html#lenet]<\/a> it says:<\/p>\n\n<pre><\/pre>\n\n<p>when I use flatten function on a numpy 3d array I get a 1D array. but here it says I get a matrix. How does flatten(2) work in theano?<\/p>\n\n<p>A similar example on numpy produces 1D array:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34710378","CreationDate":"2016-01-10T20:23:35.713","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>numpy doesn't support flattening only some dimensions but Theano does.<\/p>\n\n<p>So if <code>a<\/code> is a numpy array, <code>a.flatten(2)<\/code> doesn't make any sense. It runs without error but only because the <code>2<\/code> is passed as the <code>order<\/code> parameter which seems to cause numpy to stick with the default order of <code>C<\/code>.<\/p>\n\n<p>Theano's <code>flatten<\/code> <em>does<\/em> support axis specification. <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.flatten\" rel=\"nofollow\">The documentation<\/a> explains how it works.<\/p>\n\n<pre><code>Parameters:\n    x (any TensorVariable (or compatible)) \u2013 variable to be flattened\n    outdim (int) \u2013 the number of dimensions in the returned variable\n\nReturn type:\n    variable with same dtype as x and outdim dimensions\n\nReturns:\n    variable with the same shape as x in the leading outdim-1 dimensions,\n    but with all remaining dimensions of x collapsed into the last dimension.\n<\/code><\/pre>\n\n<blockquote>\n  <p>For example, if we flatten a tensor of shape (2, 3, 4, 5) with\n  flatten(x, outdim=2), then we\u2019ll have the same (2-1=1) leading\n  dimensions (2,), and the remaining dimensions are collapsed. So the\n  output in this example would have shape (2, 60).<\/p>\n<\/blockquote>\n\n<p>A simple Theano demonstration:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as tt\n\n\ndef compile():\n    x = tt.tensor3()\n    return theano.function([x], x.flatten(2))\n\n\ndef main():\n    a = numpy.arange(2 * 3 * 4).reshape((2, 3, 4))\n    f = compile()\n    print a.shape, f(a).shape\n\n\nmain()\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre><code>(2L, 3L, 4L) (2L, 12L)\n<\/code><\/pre>\n"}]}
{"QuestionId":34710744,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-10T21:03:21.153","AcceptedAnswerId":34711711.0,"OwnerUserId":5738942.0,"Title":"Tensorflow: Dynamically Make Letter Prediction","Body":"<p>I am trying to use the LSTM module in Tensorflow to use a one-hot letter prediction at time (t-1) as input into the next state at time (t). I am doing something along the lines of:<\/p>\n\n<pre><\/pre>\n\n<p>However, I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any way these one-hot letters can be created dynamically with my dictionary from the argmax values to the one-hot letter encoding?<\/p>\n","answers":[{"AnswerId":"34711711","CreationDate":"2016-01-10T22:51:09.297","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are several ways you could achieve this.<\/p>\n\n<p>The most direct adaptation of your code would be to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#gather\" rel=\"nofollow\"><code>tf.gather()<\/code><\/a> operation to select rows from an identity matrix, as follows:<\/p>\n\n<pre><code># Build the matrix [[1., 0., 0.], [0., 1., 0.], [0., 0., 1.]].\nidentity_matrix = tf.diag(tf.ones([3]))\n\nfor ...:\n\n  # Compute a vector of predicted letter indices.\n  argmax_val = ...\n\n  # For each element of `argmax_val`, select the corresponding row\n  # from `identity_matrix` and concatenate them into matrix.\n  input = tf.gather(identity_matrix, argmax_val)\n<\/code><\/pre>\n\n<p>For the case you showed where there are only 3 different letters, the performance probably isn't critical. However, if the number of letters (and hence the size of <code>identity_matrix<\/code>) is much larger&mdash;compared to the batch size&mdash;you could achieve better memory efficiency by building a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#SparseTensor\" rel=\"nofollow\"><code>tf.SparseTensor<\/code><\/a> and using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#sparse_tensor_to_dense\" rel=\"nofollow\"><code>tf.sparse_tensor_to_dense()<\/code><\/a> op to build <code>input<\/code>.<\/p>\n"}]}
{"QuestionId":34711264,"AnswerCount":2,"Tags":"<java><scala><apache-spark><guava><tensorflow>","CreationDate":"2016-01-10T21:58:55.403","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Pure Java\/Scala code for writing Tensorflow TFRecords data file","Body":"<p>I'm trying to write a pure Java\/Scala implementation of the Tensorflow RecordWriter class in order to convert Spark DataFrame into <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/python_io.html#tfrecords-format-details\" rel=\"nofollow noreferrer\">TFRecords<\/a> file. According to the documentation, in TFRecords, each record is formated as follow:<\/p>\n\n<pre><\/pre>\n\n<p>And the CRC mask<\/p>\n\n<pre><\/pre>\n\n<p>Currently, I compute the CRC with guava implementation with the following code: <\/p>\n\n<pre class=\"lang-scala prettyprint-override\"><\/pre>\n\n<p>The rest of my code is: <\/p>\n\n<p>The data encoding part is done with the following piece of code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The record are generated with protocol buffer:<\/p>\n\n<pre><\/pre>\n\n<p>And here is an example of how I use things together in order to generate my TFRecords file:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to read the file I got inside Tensorflow with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#TFRecordReader\" rel=\"nofollow noreferrer\">TFRecordReader<\/a>, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I suspect that the CRC mask computation is not correct or the endianness \nbetween java and c++ generated file are not the same.<\/p>\n","answers":[{"AnswerId":"43992484","CreationDate":"2017-05-16T04:25:44.277","ParentId":null,"OwnerUserId":"845843","Title":null,"Body":"<p>FWIW, the Tensorflow team has provided utility code for reading\/writing TFRecords, which can be <a href=\"https:\/\/github.com\/tensorflow\/ecosystem\/tree\/master\/hadoop\/src\/main\/java\/org\/tensorflow\/hadoop\/util\" rel=\"nofollow noreferrer\">found in the ecosystem repo<\/a><\/p>\n"},{"AnswerId":"34730418","CreationDate":"2016-01-11T20:24:48.570","ParentId":null,"OwnerUserId":"1259390","Title":null,"Body":"<p>The problem with the my implementation is the computation of the CRC mask. Here is fix the I found:<\/p>\n\n<pre><code>def mask(crc: Int): Int ={\n    ((crc &gt;&gt;&gt; 15) | (crc &lt;&lt; 17)) + kMaskDelta\n}\n<\/code><\/pre>\n\n<p>The key is the use of the unsigned shift bitwise operator <code>&gt;&gt;&gt;<\/code> instead of <code>&gt;&gt;<\/code><\/p>\n"}]}
{"QuestionId":34711857,"AnswerCount":1,"Tags":"<ubuntu><compiler-errors><installation><makefile><caffe>","CreationDate":"2016-01-10T23:09:02.090","AcceptedAnswerId":34767422.0,"OwnerUserId":4386025.0,"Title":"Caffe Installation - Can't Start Compilation","Body":"<p>I am installing it without cuda support... I don't have nVidia graphics card...\n<\/p>\n\n<p>I have gcc-4.7 and gcc-4.8. I tried with both using alternative options.... Then i tried cmake<\/p>\n\n<p>cmake :<\/p>\n\n<p>```<\/p>\n\n<pre><\/pre>\n\n<p>make all```<\/p>\n\n<p>```<\/p>\n\n<pre><\/pre>\n\n<p>```<\/p>\n\n<p>Please anyone give the solution... Where is mistake ! \nI am attaching ny Makefile.config\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/files\/84085\/make_config.txt\" rel=\"nofollow\">make_config.txt<\/a><\/p>\n","answers":[{"AnswerId":"34767422","CreationDate":"2016-01-13T12:59:36.070","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>It looks like the compiler is missing a linking instructions.<\/p>\n\n<p><a href=\"http:\/\/codeyarns.com\/2014\/05\/02\/undefined-reference-error-with-__cxa_free_exceptioncxxabi_1-3\/\" rel=\"nofollow\">This<\/a> SO question deals with a similar linking error.<\/p>\n\n<p>And <a href=\"http:\/\/codeyarns.com\/2014\/05\/02\/undefined-reference-error-with-__cxa_free_exceptioncxxabi_1-3\/\" rel=\"nofollow\">this<\/a> blog post also.<\/p>\n\n<p>The quick solutions is passing <code>-lstdc++<\/code> to the linker.<\/p>\n"}]}
{"QuestionId":34712933,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras>","CreationDate":"2016-01-11T01:37:24.447","AcceptedAnswerId":null,"OwnerUserId":5771418.0,"Title":"Neural network generating incorrect results that are around the average of outputs","Body":"<p>I'm working on a neural network system to perform <a href=\"http:\/\/www.sedfitting.org\/SED08\/Welcome.html\" rel=\"noreferrer\">SED fitting<\/a> as part of a studentship project at the University of Western Australia.<\/p>\n\n<p>I have created a set of around 20,000 runs through an SED fitting program known as <a href=\"http:\/\/www.iap.fr\/magphys\/magphys\/MAGPHYS_files\/readme.pdf\" rel=\"noreferrer\">MAGPHYS<\/a>. Each run has the 42 input values and 32 output values that we're interested in (there are more outputs from the system, but we don't need them)<\/p>\n\n<p>I've been experimenting around with the Keras neural network package in order to create a network to learn this function. <\/p>\n\n<p>My current network design uses 4 hidden layers, fully interconnected, with 30 connections between each layer. Each layer is using TanH activation functions. I also have a 42 dimension input layer and 32 dimension output layer, both also using TanH activation, for a total of 6 layers.<\/p>\n\n<pre><\/pre>\n\n<p>I have been using min\/max normalisation of my input and output data to squash all of the values between 0 and 1. I'm using a stochastic gradient descent optimiser and I've experimented with various loss functions such as mean squared error, mean absolute error, mean absolute percentage error etc.<\/p>\n\n<p><strong>The main issue is that regardless of how I structure my network, it simply generates output values that are around the average of all of the training output values.<\/strong> It does not appear as through the network has actually learned the function correctly, it just generates values around the average. Worse still, some network designs I've experimented with, particularly those that use linear activation functions, will generate ONLY the average of the output values and will not vary at all.<\/p>\n\n<p>Example (for one of the 32 outputs):<\/p>\n\n<pre><\/pre>\n\n<p>Notice how all of the outputs are just around the 9.2 - 9.4 value, even though these values are quite incorrect.<\/p>\n\n<p><strong>With all of this in mind, what causes a network such as mine to generate these sorts of outputs that are all around the average?<\/strong> <\/p>\n\n<p><strong>What sort of things can I try to remedy this problem and create a network, of some sort, to actually generate correct outputs?<\/strong><\/p>\n","answers":[{"AnswerId":"38405900","CreationDate":"2016-07-15T22:35:18.950","ParentId":null,"OwnerUserId":"4727781","Title":null,"Body":"<p>I just want to throw out some thoughts on this specific problem, in addition to CAFEBABE's comment: <\/p>\n\n<ol>\n<li><p>42 input features is not a ton of features to work with. Not something you can necessarily fix, but it means that you'll want to have wider hidden layers (ie more nodes), to help with separability of the classes\/labels. Furthermore, 20K observations isn't exactly a large dataset. If you can get more data, you should. This is pretty much always the case. <\/p><\/li>\n<li><p>If you have a concrete reason for min\/max normalization, then disregard this point, but you could consider BatchNormalizing your input, which tends to help the network's ability to predict accurately. This essentially allows the activation to have inputs occurring closer to the middle of the function, rather than the ends.<\/p><\/li>\n<li><p>You should experiment more with your optimization. For example: <\/p>\n\n<ul>\n<li>How many epochs are you running the training for? One pass over the data isn't enough. <\/li>\n<li>Try different optimizers, like <code>rmsprop<\/code> or <code>adam<\/code>, or learning rates. <\/li>\n<li>You could also use a learning rate scheduler, like <a href=\"http:\/\/keras.io\/callbacks\/#learningratescheduler\" rel=\"noreferrer\">here<\/a>, where you decrease your learning rate over the epochs, so that you don't shoot past minima of your loss function. <\/li>\n<\/ul><\/li>\n<li><p>Try some different activation functions. Recent research includes: ReLU, ELU, PReLU, SReLU. All available in keras. <\/p><\/li>\n<li><p>Also try including some regularization, to avoid overfitting. Look into Dropout, or L2\/L1 <\/p><\/li>\n<li><p>While having a deeper model (ie more layers) does often help, reducing the data dimensions from 42 features, down to 30, is likely hurting your ability to separate the data. Try something bigger, like 100, or 500, or 1000.  <\/p><\/li>\n<\/ol>\n\n<p>An example model you could try would be something like: <\/p>\n\n<pre><code># imports \nfrom sklearn.cross_validation import train_test_split\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Activation, Dropout\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.layers.advanced_activations import ELU     \n\n# data shapes\nn_obs, n_feat = 20000, 42\nn_hidden = 500 # play with this, bigger tends to give better separability\nn_class = 32\n\n# instantiate model\nmodel = Sequential()\n\n# first layer --- input\nmodel.add(Dense(input_dim = n_feat, output_dim = n_hidden))\nmodel.add(BatchNormalization())\nmodel.add(ELU())\nmodel.add(Dropout(p=0.2)) # means that 20% of the nodes are turned off, randomly\n\n# second layer --- hidden\nmodel.add(Dense(input_dim = n_hidden, output_dim = n_hidden))\nmodel.add(BatchNormalization())\nmodel.add(ELU())\nmodel.add(Dropout(p=0.2))\n\n# third layer --- output\nmodel.add(Dense(input_dim = n_hidden, output_dim = n_class))\nmodel.add(BatchNormalization())\nmodel.add(Activation('softmax'))\n\n# configure optimization\nmodel.compile(optimizer = 'rmsprop', loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\n# split your data, so you test it on validation data\nX_train, X_test, Y_train, Y_test = train_test_split(data, targets)\n\n# train your model\nmodel.fit(X_train, Y_train, validation_data = (X_test, Y_test))\n<\/code><\/pre>\n\n<p>Best of luck! <\/p>\n"}]}
{"QuestionId":34713299,"AnswerCount":2,"Tags":"<python><memory-management><machine-learning><gpgpu><theano>","CreationDate":"2016-01-11T02:22:55.963","AcceptedAnswerId":34713981.0,"OwnerUserId":2093984.0,"Title":"How to calculate GPU memory usage in Theano?","Body":"<p>I am experimenting with different Theano models and use a curriculum of ever increasing sequence length.  How can I predict ahead of time how big to make the batch size for any given sequence length and model in order to fill the GPU's memory?<\/p>\n\n<p>To make matters worse, if I ever accidentally use too much memory, I get a MemoryError and the memory on the GPU is not freed, requiring me to restart the process to free the memory, and lose my network, before trying a new batch size.  Because this error is unrecoverable, it is difficult to just increase batch size until exception and then back down.<\/p>\n","answers":[{"AnswerId":"34713876","CreationDate":"2016-01-11T03:41:04.753","ParentId":null,"OwnerUserId":"145279","Title":null,"Body":"<p>It seems that Theano doesn't have any built-in way to estimate the memory size of a model. Your best bet is to create a small subset of your model with a known size, and use the memory estimation techniques described <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/python-memory-management.html\" rel=\"nofollow\">here<\/a> in the Theano manual.<\/p>\n\n<p>We'll also need to take into account how our objects are represented inside the GPU (are we using <code>float32<\/code> or <code>float64<\/code>, for example, and how many bytes does each take up inside the GPU).<\/p>\n\n<p>Once you can estimate the size of a small model, you can then project these estimates to the size of a much larger model with reasonable accuracy. You should be able to write your own memory estimation function that can take a number of features and observations, or tensors, or graph nodes as parameters and returns a memory usage amount.<\/p>\n"},{"AnswerId":"34713981","CreationDate":"2016-01-11T03:55:47.313","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Assuming you know the number of elements to be stored on a GPU you can easily compute the amount of memory required to store those elements.<\/p>\n\n<p>A simple example:<\/p>\n\n<pre><code>import numpy as np\nimport theano.tensor as T\nT.config.floatX = 'float32'\ndataPoints = np.random.random((5000, 256 * 256)).astype(T.config.floatX)\n#float32 data type requires 4 bytes\nsizeinGBs = 5000 * 256 * 256 * 4 \/ 1024. \/ 1024 \/ 1024 + (some small over-head constant)\nprint \"Data will need %2f GBs of free memory\" % sizeInGB\n<\/code><\/pre>\n\n<p>Assuming the over-head constant is 0 will print:<\/p>\n\n<pre><code>&gt;&gt;&gt; Data will need 1.22 GBs of free memory\n<\/code><\/pre>\n\n<p>If you are using an NVIDIA graphics card and have installed CUDA on your machine then you can easily get the total amount of free memory on your GPU  using the following line of code:<\/p>\n\n<pre><code>import theano.sandbox.cuda.basic_ops as sbcuda\nimport numpy as np\nimport theano.tensor as T\nT.config.floatX = 'float32'\nGPUFreeMemoryInBytes = sbcuda.cuda_ndarray.cuda_ndarray.mem_info()[0]\nfreeGPUMemInGBs = GPUFreeMemoryInBytes\/1024.\/1024\/1024\nprint \"Your GPU has %s GBs of free memory\" % str(freeGPUMemInGBs)\n#An operation is to be executed below\ntestData = shared(np.random.random((5000, 256 * 256)).astype(T.config.floatX), borrow = True)\nprint \"The tasks above used %s GBs of your GPU memory. The available memory is %s GBs\" % (str(freeGPUMemInGBs - GPUFreeMemoryInBytes\/1024.\/1024\/1024), str(GPUFreeMemoryInBytes\/1024.\/1024\/1024))\n<\/code><\/pre>\n\n<p>Then the output is in the following format (for my machine here):<\/p>\n\n<pre><code>&gt;&gt;&gt; Your GPU has 11.2557678223 GBs of free memory\n&gt;&gt;&gt; The tasks above used 1.22077941895 GBs of your GPU memory. The available memory is 10.0349884033 GBs\n<\/code><\/pre>\n\n<p>By monitoring the amount of free memory and calculating the size of your model\/data you can better use the GPU memory. However, be aware of the <a href=\"https:\/\/stackoverflow.com\/questions\/34466375\/dealing-with-memory-fragmentation-of-gpus-in-theano\">memory fragmentation<\/a> issue as it may cause <code>MemoryError<\/code> unexpectedly.<\/p>\n"}]}
{"QuestionId":34714092,"AnswerCount":1,"Tags":"<python><computer-vision><tensorflow>","CreationDate":"2016-01-11T04:12:45.120","AcceptedAnswerId":34714543.0,"OwnerUserId":5771792.0,"Title":"How to compute gradients to fool an image classifier?","Body":"<p>I am trying to get the standard trick-an-image-classifier example working in .  <\/p>\n\n<p>(That is, adjust the input image by following the gradient so it is misclassified, e.g., <a href=\"https:\/\/codewords.recurse.com\/issues\/five\/why-do-neural-networks-think-a-panda-is-a-vulture\" rel=\"nofollow\">https:\/\/codewords.recurse.com\/issues\/five\/why-do-neural-networks-think-a-panda-is-a-vulture<\/a> .)  <\/p>\n\n<p>I have downloaded the inception-v3 model from <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html<\/a> , and have it working to classify images.  <\/p>\n\n<p>But I am having a hard time computing the gradients to adjust the input image.  I was hoping I could get some help as to how it works in .  <\/p>\n\n<p>This is the basic idea for what I have been trying: <\/p>\n\n<pre><\/pre>\n\n<p>Even though  works as it should,  is just returning [None].  This is obviously a very beginner question, but can anyone point me toward where I am going wrong here?  Why isn't gradients doing anything?  <\/p>\n","answers":[{"AnswerId":"34714543","CreationDate":"2016-01-11T05:10:38.847","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The reason that <code>tf.gradients()<\/code> returns <code>[None]<\/code> is that <code>input_tensor<\/code> is subjected to a non-differentiable transformation (i.e., JPEG decoding and a cast) before it is fed into the Inception network.  Instead, you should operate on the <strong>result<\/strong> of the JPEG decoding (<strong>EDIT:<\/strong> and cast), as follows:<\/p>\n\n<pre><code># This tensor is the result of the DecodeJpeg op.\ndecoded_input_tensor = sess.graph.get_tensor_by_name('Cast:0')\n\ngrad = tf.gradients(softmax_tensor, decoded_input_tensor)[0]\n\nreal_grad = grad.eval(feed_dict)\n<\/code><\/pre>\n\n<p>After you have generated your image that tricks the classifier, you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#encode_jpeg\" rel=\"nofollow\"><code>tf.image.encode_jpeg()<\/code><\/a> op to transform it back into a JPEG image.<\/p>\n"}]}
{"QuestionId":34715055,"AnswerCount":1,"Tags":"<machine-learning><gpu><gpgpu><tensorflow>","CreationDate":"2016-01-11T05:57:40.967","AcceptedAnswerId":34715397.0,"OwnerUserId":2771184.0,"Title":"Choosing between GeForce or Quadro GPUs to do machine learning via TensorFlow","Body":"<p>Is there any noticeable difference in TensorFlow performance if using Quadro GPUs vs GeForce GPUs? <\/p>\n\n<p><strong>e.g. does it use double precision operations or something else that would cause a drop in GeForce cards?<\/strong><\/p>\n\n<p>I am about to buy a GPU for TensorFlow, and wanted to know if a GeForce would be ok. Thanks and appreciate your help<\/p>\n","answers":[{"AnswerId":"34715397","CreationDate":"2016-01-11T06:28:49.467","ParentId":null,"OwnerUserId":"1957102","Title":null,"Body":"<p>I think GeForce TITAN is great and is widely used in Machine Learning (ML). In ML, single precision is enough in most of cases. <\/p>\n\n<p>More detail on the performance of the GTX line (currently GeForce 10) can be found in Wikipedia, <a href=\"https:\/\/en.wikipedia.org\/wiki\/GeForce_10_series\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Other sources around the web support this claim. Here is a quote <a href=\"http:\/\/doc-ok.org\/?p=304\" rel=\"noreferrer\">from doc-ok in 2013<\/a> (<a href=\"https:\/\/perma.cc\/GH7A-8UUX\" rel=\"noreferrer\">permalink<\/a>).<\/p>\n\n<blockquote>\n  <p>For comparison, an \u201centry-level\u201d $700 Quadro 4000 is significantly slower than a $530 high-end GeForce GTX 680, at least according to my measurements using several Vrui applications, and the closest performance-equivalent to a GeForce GTX 680 I could find was a Quadro 6000 for a whopping $3660. <\/p>\n<\/blockquote>\n\n<p>Specific to ML, including deep learning, there is a <a href=\"https:\/\/www.kaggle.com\/general\/11332\" rel=\"noreferrer\">Kaggle forum discussion dedicated to this subject<\/a> (Dec 2014, <a href=\"https:\/\/perma.cc\/RBZ6-YX9F\" rel=\"noreferrer\">permalink<\/a>), which goes over comparisons between the Quadro, GeForce, and Tesla series:<\/p>\n\n<blockquote>\n  <p>Quadro GPUs aren't for scientific computation, Tesla GPUs are.  Quadro\n  cards are designed for accelerating CAD, so they won't help you to\n  train neural nets. They can probably be used for that purpose just\n  fine, but it's a waste of money.<\/p>\n  \n  <p>Tesla cards are for scientific computation, but they tend to be pretty\n  expensive. The good news is that many of the features offered by Tesla\n  cards over GeForce cards are not necessary to train neural networks.<\/p>\n  \n  <p>For example, Tesla cards usually have ECC memory, which is nice to\n  have but not a requirement. They also have much better support for\n  double precision computations, but single precision is plenty for\n  neural network training, and they perform about the same as GeForce\n  cards for that.<\/p>\n  \n  <p>One useful feature of Tesla cards is that they tend to have is a lot\n  more RAM than comparable GeForce cards. More RAM is always welcome if\n  you're planning to train bigger models (or use RAM-intensive\n  computations like FFT-based convolutions).<\/p>\n  \n  <p>If you're choosing between Quadro and GeForce, definitely pick\n  GeForce. If you're choosing between Tesla and GeForce, pick GeForce,\n  unless you have a lot of money and could really use the extra RAM.<\/p>\n<\/blockquote>\n\n<p><strong>NOTE:<\/strong> Be careful what platform you are working on and what the default precision is in it. For example, <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/959375\/cuda-programming-and-performance\/performance-issues-with-cuda-and-python-r\/1\" rel=\"noreferrer\">here in the CUDA forums<\/a> (August 2016), one developer owns two Titan X's (GeForce series) and doesn't see a performance gain in any of their R or Python scripts. This is diagnosed as a result of R being defaulted to double precision, and has a worse performance on new GPU than their CPU (a Xeon processor). Tesla GPUs are cited as the best performance for double precision. In this case, converting all numbers to float32 increases performance from 12.437s with nvBLAS 0.324s with gmatrix+float32s on one TITAN X (see first benchmark). Quoting from this forum discussion:<\/p>\n\n<blockquote>\n  <p>Double precision performance of Titan X is pretty low.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":34715091,"AnswerCount":3,"Tags":"<python><pycharm><caffe><pycaffe>","CreationDate":"2016-01-11T06:00:55.327","AcceptedAnswerId":null,"OwnerUserId":5299360.0,"Title":"Pycharm Couldn't connect to console process","Body":"<p>I recently installed Caffe along with OpenCV and other dependencies. Pycharm worked well for 2 days and today I can't start Python Console in Pycharm anymore. I can start python in terminal with no problem... Everytime I try to start a Python Console it will take a while to \"connecting to console\" and then this is what I get after:<\/p>\n\n<blockquote>\n  <p>\/Users\/myname\/anaconda\/bin\/python \"\/Applications\/PyCharm CE.app\/Contents\/helpers\/pydev\/pydevconsole.py\" 56229 56230<\/p>\n  \n  <p>Python 2.7.10 |Anaconda 2.2.0 (x86_64)| (default, May 28 2015, 17:04:42) \n  Type \"copyright\", \"credits\" or \"license\" for more information.<\/p>\n  \n  <p>IPython 3.0.0 -- An enhanced Interactive Python.\n  Anaconda is brought to you by Continuum Analytics.\n  Please check out: <a href=\"http:\/\/continuum.io\/thanks\" rel=\"nofollow\">http:\/\/continuum.io\/thanks<\/a> and <a href=\"https:\/\/binstar.org\" rel=\"nofollow\">https:\/\/binstar.org<\/a><\/p>\n  \n  <p>Couldn't connect to console process.<\/p>\n  \n  <p>Process finished with exit code 137<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42366289","CreationDate":"2017-02-21T11:39:22.357","ParentId":null,"OwnerUserId":"3074185","Title":null,"Body":"<p>I had this problem running interpreter in Docker. Checking \"Import credentials from Docker Machine\" in Build->Execution->Deployment->Docker solved that. <\/p>\n"},{"AnswerId":"56665162","CreationDate":"2019-06-19T10:07:37.747","ParentId":null,"OwnerUserId":"8031358","Title":null,"Body":"<p>I solved going on Settings\/Project: [YourProjectName]\/Project Interpreter and changing it to python.exe from the dropdown menu. I don't know why but there was a different one by default.<\/p>\n"},{"AnswerId":"51240448","CreationDate":"2018-07-09T07:48:18.593","ParentId":null,"OwnerUserId":"7861213","Title":null,"Body":"<p>this might be your firewall prevent the pycharm, please check the firewall.<\/p>\n"}]}
{"QuestionId":34716454,"AnswerCount":7,"Tags":"<python><keras><neural-network><data-science><batch-normalization>","CreationDate":"2016-01-11T07:47:53.143","AcceptedAnswerId":37979391.0,"OwnerUserId":4984897.0,"Title":"Where do I call the BatchNormalization function in Keras?","Body":"<p>If I want to use the BatchNormalization function in Keras, then do I need to call it once only at the beginning?<\/p>\n\n<p>I read this documentation for it: <a href=\"http:\/\/keras.io\/layers\/normalization\/\">http:\/\/keras.io\/layers\/normalization\/<\/a><\/p>\n\n<p>I don't see where I'm supposed to call it. Below is my code attempting to use it:<\/p>\n\n<pre><\/pre>\n\n<p>I ask because if I run the code with the second line including the batch normalization and if I run the code without the second line I get similar outputs. So either I'm not calling the function in the right place, or I guess it doesn't make that much of a difference.<\/p>\n","answers":[{"AnswerId":"37979391","CreationDate":"2016-06-22T22:40:51.113","ParentId":null,"OwnerUserId":"4727781","Title":null,"Body":"<p>Just to answer this question in a little more detail, and as Pavel said, Batch Normalization is just another layer, so you can use it as such to create your desired network architecture. <\/p>\n\n<p>The general use case is to use BN between the linear and non-linear layers in your network, because it normalizes the input to your activation function, so that you're centered in the linear section of the activation function (such as Sigmoid). There's a small discussion of it <a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/2x0bq8\/some_questions_regarding_batch_normalization\/?su=ynbwk&amp;st=iprg6e3w&amp;sh=88bcbe40\">here<\/a><\/p>\n\n<p>In your case above, this might look like:<\/p>\n\n<hr>\n\n<pre><code># import BatchNormalization\nfrom keras.layers.normalization import BatchNormalization\n\n# instantiate model\nmodel = Sequential()\n\n# we can think of this chunk as the input layer\nmodel.add(Dense(64, input_dim=14, init='uniform'))\nmodel.add(BatchNormalization())\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\n\n# we can think of this chunk as the hidden layer    \nmodel.add(Dense(64, init='uniform'))\nmodel.add(BatchNormalization())\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\n\n# we can think of this chunk as the output layer\nmodel.add(Dense(2, init='uniform'))\nmodel.add(BatchNormalization())\nmodel.add(Activation('softmax'))\n\n# setting up the optimization of our weights \nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='binary_crossentropy', optimizer=sgd)\n\n# running the fitting\nmodel.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose = 2)\n<\/code><\/pre>\n\n<hr>\n\n<p>Hope this clarifies things a bit more.<\/p>\n"},{"AnswerId":"45782379","CreationDate":"2017-08-20T12:54:02.197","ParentId":null,"OwnerUserId":"4653650","Title":null,"Body":"<p>This thread has some considerable debate about whether BN should be applied before non-linearity of current layer or to the activations of the previous layer. <\/p>\n\n<p>Although there is no correct answer, the authors of Batch Normalization say that \n<strong>It should be applied immediately before the non-linearity of the current layer.<\/strong> The reason ( quoted from original paper) - <\/p>\n\n<blockquote>\n  <p>\"We add the BN transform immediately before the\n  nonlinearity, by normalizing x = Wu+b. We could have\n  also normalized the layer inputs u, but since u is likely\n  the output of another nonlinearity, the shape of its distribution\n  is likely to change during training, and constraining\n  its first and second moments would not eliminate the covariate\n  shift. In contrast, Wu + b is more likely to have\n  a symmetric, non-sparse distribution, that is \u201cmore Gaussian\u201d\n  (Hyv\u00a8arinen &amp; Oja, 2000); normalizing it is likely to\n  produce activations with a stable distribution.\"<\/p>\n<\/blockquote>\n"},{"AnswerId":"41147012","CreationDate":"2016-12-14T16:02:37.207","ParentId":null,"OwnerUserId":"2845263","Title":null,"Body":"<p>It's almost become a trend now to have a <code>Conv2D<\/code> followed by a <code>ReLu<\/code> followed by a <code>BatchNormalization<\/code> layer. So I made up a small function to call all of them at once. Makes the model definition look a whole lot cleaner and easier to read. <\/p>\n\n<pre><code>def Conv2DReluBatchNorm(n_filter, w_filter, h_filter, inputs):\n    return BatchNormalization()(Activation(activation='relu')(Convolution2D(n_filter, w_filter, h_filter, border_mode='same')(inputs)))\n<\/code><\/pre>\n"},{"AnswerId":"34751249","CreationDate":"2016-01-12T18:31:08.973","ParentId":null,"OwnerUserId":"579980","Title":null,"Body":"<p>It is another type of layer, so you should add it as a layer in an appropriate place of your model<\/p>\n\n<pre><code>model.add(keras.layers.normalization.BatchNormalization())\n<\/code><\/pre>\n\n<p>See an example here: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/kaggle_otto_nn.py\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/kaggle_otto_nn.py<\/a><\/p>\n"},{"AnswerId":"55587940","CreationDate":"2019-04-09T08:08:34.427","ParentId":null,"OwnerUserId":"7218537","Title":null,"Body":"<p>Batch Normalization is used to normalize the input layer as well as hidden layers by adjusting mean and scaling of the activations. Because of this normalizing effect with additional layer in deep neural networks, the network can use higher learning rate without vanishing or exploding gradients. Furthermore, batch normalization regularizes the network such that it is easier to generalize, and it is thus unnecessary to use dropout to mitigate overfitting. <\/p>\n\n<p>Right after calculating the linear function using say, the Dense() or Conv2D() in Keras, we use BatchNormalization() which calculates the linear function in a layer and then we add the non-linearity to the layer using Activation().<\/p>\n\n<pre><code>from keras.layers.normalization import BatchNormalization\nmodel = Sequential()\nmodel.add(Dense(64, input_dim=14, init='uniform'))\nmodel.add(BatchNormalization(epsilon=1e-06, mode=0, momentum=0.9, weights=None))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, init='uniform'))\nmodel.add(BatchNormalization(epsilon=1e-06, mode=0, momentum=0.9, weights=None))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(2, init='uniform'))\nmodel.add(BatchNormalization(epsilon=1e-06, mode=0, momentum=0.9, weights=None))\nmodel.add(Activation('softmax'))\n\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='binary_crossentropy', optimizer=sgd)\nmodel.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, \nvalidation_split=0.2, verbose = 2)\n<\/code><\/pre>\n\n<p><strong>How is Batch Normalization applied?<\/strong><\/p>\n\n<p>Suppose we have input a[l-1] to a layer l. Also we have weights W[l] and bias unit b[l] for the layer l. Let a[l] be the activation vector calculated(i.e. after adding the non-linearity) for the layer l and z[l] be the vector before adding non-linearity<\/p>\n\n<ol>\n<li>Using a[l-1] and W[l] we can calculate z[l] for the layer l<\/li>\n<li>Usually in feed-forward propagation we will add bias unit to the z[l] at this stage like this z[l]+b[l], but in Batch Normalization this step of addition of b[l] is not required and no b[l] parameter is used.<\/li>\n<li>Calculate z[l] means and subtract it from each element<\/li>\n<li>Divide  (z[l] - mean) using standard deviation. Call it Z_temp[l]<\/li>\n<li><p>Now define new parameters \u03b3 and \u03b2 that will change the scale of the hidden layer as follows:<\/p>\n\n<p>z_norm[l] = \u03b3.Z_temp[l] + \u03b2<\/p><\/li>\n<\/ol>\n\n<p>In this code excerpt, the Dense() takes the a[l-1], uses W[l] and calculates z[l]. Then the immediate BatchNormalization() will perform the above steps to give z_norm[l]. And then the immediate Activation() will calculate tanh(z_norm[l]) to give a[l] i.e.<\/p>\n\n<pre><code>a[l] = tanh(z_norm[l])\n<\/code><\/pre>\n"},{"AnswerId":"41375256","CreationDate":"2016-12-29T07:42:36.713","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Keras now supports the <code>use_bias=False<\/code> option, so we can save some computation by writing like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add(Dense(64, use_bias=False))\nmodel.add(BatchNormalization(axis=bn_axis))\nmodel.add(Activation('tanh'))\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add(Convolution2D(64, 3, 3, use_bias=False))\nmodel.add(BatchNormalization(axis=bn_axis))\nmodel.add(Activation('relu'))\n<\/code><\/pre>\n"},{"AnswerId":"45624249","CreationDate":"2017-08-10T22:17:24.913","ParentId":null,"OwnerUserId":"7039439","Title":null,"Body":"<p>This thread is misleading.  Tried commenting on Lucas Ramadan's answer, but I don't have the right privileges yet, so I'll just put this here.<\/p>\n\n<p>Batch normalization works best after the activation function, and <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"noreferrer\">here<\/a> or <a href=\"https:\/\/standardfrancis.wordpress.com\/2015\/04\/16\/batch-normalization\/\" rel=\"noreferrer\">here<\/a> is why: it was developed to prevent internal covariate shift. Internal covariate shift occurs when the distribution of the <em>activations<\/em> of a layer shifts significantly throughout training.  Batch normalization is used so that the distribution of the inputs (and these inputs are literally the result of an activation function) to a specific layer doesn't change over time due to parameter updates from each batch (or at least, allows it to change in an advantageous way).  It uses batch statistics to do the normalizing, and then uses the batch normalization parameters (gamma and beta in the original paper) \"to make sure that the transformation inserted in the network can represent the identity transform\" (quote from original paper).  But the point is that we're trying to normalize the inputs to a layer, so it should always go immediately before the next layer in the network.  Whether or not that's after an activation function is dependent on the architecture in question.<\/p>\n"}]}
{"QuestionId":34716573,"AnswerCount":2,"Tags":"<python><neural-network><keras><data-science>","CreationDate":"2016-01-11T07:58:49.493","AcceptedAnswerId":34878787.0,"OwnerUserId":4984897.0,"Title":"How do you change the number of units in a layer when using Keras?","Body":"<p>The below code works perfectly okay. If I try to change all the 64s to 128s then I get an error about shape. Do I need to change the input data shape if I change the number of layers in an artificial neural network when using Keras? I didn't think so because it asks for input_dim which is correct. <\/p>\n\n<p>Works: <\/p>\n\n<pre><\/pre>\n\n<p>Doesn't Work:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34852290","CreationDate":"2016-01-18T10:20:37.137","ParentId":null,"OwnerUserId":"2046964","Title":null,"Body":"<p>You do not change the number of layers at all in between your code-snippets.\nWhat you change is the number of units.<\/p>\n\n<p>Dense(64) is a fully-connected layer with 64 hidden units. Since it is fully connected, the number of your input units also changes to that number, and hence also the requirements for the inputs.<\/p>\n"},{"AnswerId":"34878787","CreationDate":"2016-01-19T14:01:00.340","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>You can use a different number of hidden layers and a different number of units\/neurons for each of them on the same input.<\/p>\n\n<p>Each <code>Dense<\/code> except the last one can be seen as a hidden layer. The last <code>Dense<\/code> should have a number of outputs equal to your desired output dimension (in your case the dimension of <code>y<\/code> seems to be 64).<\/p>\n\n<p>Try this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(128, input_dim=14, init='uniform'))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(128, init='uniform'))\nmodel.add(Activation('tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, init='uniform'))\nmodel.add(Activation('softmax'))\n\nsgd3 = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='binary_crossentropy', optimizer=sgd3)\nmodel.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose = 2)\n<\/code><\/pre>\n"}]}
{"QuestionId":34717241,"AnswerCount":3,"Tags":"<python><machine-learning><neural-network><keras><data-science>","CreationDate":"2016-01-11T08:44:13.823","AcceptedAnswerId":34721948.0,"OwnerUserId":4984897.0,"Title":"How to use advanced activation layers in Keras?","Body":"<p>This is my code that works if I use other activation layers like tanh:<\/p>\n\n<pre><\/pre>\n\n<p>In this case, it doesn't work and says \"TypeError: 'PReLU' object is not callable\" and the error is called at the model.compile line. Why is this the case? All the non-advanced activation functions works. However, neither of the advanced activation functions, including this one, works.<\/p>\n","answers":[{"AnswerId":"43779606","CreationDate":"2017-05-04T09:55:29.650","ParentId":null,"OwnerUserId":"4146521","Title":null,"Body":"<p>If using the <code>Model<\/code> API in Keras you can call directly the function inside the Keras <code>Layer<\/code>. Here's an example:<\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Dense, Input\n# using prelu?\nfrom keras.layers.advanced_activations import PReLU\n\n# Model definition\n# encoder\ninp = Input(shape=(16,))\nlay = Dense(64, kernel_initializer='uniform',activation=PReLU(),\n            name='encoder')(inp)\n#decoder\nout = Dense(2,kernel_initializer='uniform',activation=PReLU(), \n            name='decoder')(lay)\n\n# build the model\nmodel = Model(inputs=inp,outputs=out,name='cae')\n<\/code><\/pre>\n"},{"AnswerId":"34721948","CreationDate":"2016-01-11T12:47:09.327","ParentId":null,"OwnerUserId":"179372","Title":null,"Body":"<p>The correct way to use the advanced activations like PReLU is to use it with <code>add()<\/code> method and not wrapping it using <code>Activation<\/code> class. Example:<\/p>\n\n<pre><code>model = Sequential()\nact = keras.layers.advanced_activations.PReLU(init='zero', weights=None)\nmodel.add(Dense(64, input_dim=14, init='uniform'))\nmodel.add(act)\n<\/code><\/pre>\n"},{"AnswerId":"48568208","CreationDate":"2018-02-01T17:09:33.120","ParentId":null,"OwnerUserId":"3597858","Title":null,"Body":"<p>For Keras functional API I think the correct way to combine Dense and PRelu (or any other advanced activation) is to use it like this:<\/p>\n\n<pre><code>focus_tns =focus_lr(enc_bidi_tns)\n\nenc_dense_lr = k.layers.Dense(units=int(hidden_size))\nenc_dense_tns = k.layers.PReLU()(enc_dense_lr(focus_tns))\n\ndropout_lr = k.layers.Dropout(0.2)\ndropout_tns = dropout_lr(enc_dense_tns)\n\nenc_dense_lr2 = k.layers.Dense(units=int(hidden_size\/4))\nenc_dense_tns2 = k.layers.PReLU()(enc_dense_lr2(dropout_tns)) \n<\/code><\/pre>\n\n<p>of course one should parametrize layers according to the problem<\/p>\n"}]}
{"QuestionId":34718477,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision><softmax><kaggle>","CreationDate":"2016-01-11T09:54:47.380","AcceptedAnswerId":null,"OwnerUserId":3419997.0,"Title":"Facial Keypoints Detection using Softmax Regression","Body":"<p>I am trying to build a basic Softmax Regression model using Tensorflow for <a href=\"https:\/\/www.kaggle.com\/c\/facial-keypoints-detection\" rel=\"nofollow noreferrer\">Kaggle Facial Keypoints Detection competition<\/a>.<\/p>\n\n<p>I followed Tensorflow beginner MNIST example for Softmax Regression model and Daniel Nouri's <a href=\"http:\/\/danielnouri.org\/notes\/2014\/12\/17\/using-convolutional-neural-nets-to-detect-facial-keypoints-tutorial\/\" rel=\"nofollow noreferrer\">blog<\/a> for data structuring.<\/p>\n\n<p>The issue which I am facing is that the script always predicts <strong>accuracy as 0.0<\/strong>.<\/p>\n\n<p>I also followed similar questions but with no luck :-<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/33883382\/tensorflow-why-doesnt-this-sofmax-regression-learn-anything?rq=1\">TensorFlow - why doesn&#39;t this sofmax regression learn anything?<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/33641799\/why-does-tensorflow-example-fail-when-increasing-batch-size?lq=1\">Why does TensorFlow example fail when increasing batch size?<\/a><\/li>\n<\/ul>\n\n<p><em>Note :- I am dropping all rows from dataset which have missing values in them.<\/em><\/p>\n\n<pre><\/pre>\n\n<p><strong>Terminal logs :<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>Seems like the issue is with the cost\/cross_entropy, changing it to following fixed the issue :<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34718736,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-11T10:06:11.477","AcceptedAnswerId":34725458.0,"OwnerUserId":2213720.0,"Title":"Dynamic size for tf.zeros() (for use with placeholders with None dimensions)","Body":"<p>Consider the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This will produce the following output, as expected:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, what if the placeholder has a dynamic dimension that is determined at run-time by the value passed to , as placeholders often do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This will produce an error for . Apparently  is not allowed for :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have figured out that it does not produce an error if I set the first dimension of my zeros tensor to non-None, such as 1:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>but then the resulting  tensor is truncated to this size:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>How can I pad the placeholder tensor with zeros so I get a tensor of shape  in this example?<\/p>\n\n<p>The only \"solutions\" I found so far is either something like the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Or simply declaring  as a placeholder and always passing a zero array of the right size.<\/p>\n\n<p>But neither looks like a clean solution to the problem and hacks like that get out of hand quickly in an application more complex than this simple example..<\/p>\n\n<p>I'm using .<\/p>\n","answers":[{"AnswerId":"34725458","CreationDate":"2016-01-11T15:40:35.883","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The recommended way to make a zero tensor with the <strong>same<\/strong> shape as another tensor is to use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/zeros_like\" rel=\"noreferrer\"><code>tf.zeros_like()<\/code><\/a> op:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.placeholder(tf.float32, shape=[None, 4])\ny = tf.zeros_like(x)\n<\/code><\/pre>\n\n<p>The resulting tensor <code>y<\/code> appears to have the shape <code>[None, None]<\/code> according to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#get_shape\" rel=\"noreferrer\"><code>Tensor.get_shape()<\/code><\/a>, but at runtime it will expand to the  same shape as <code>x<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print y.get_shape()\n# ==&gt; TensorShape([Dimension(None), Dimension(None)])\n\nsess = tf.Session()\ny_result = sess.run(y, feed_dict={x: np.random.rand(4, 4)})\n\nprint y_result.shape\n# ==&gt; (4, 4)\n<\/code><\/pre>\n\n<p>The <code>[None, None]<\/code> static shape is returned because shape inference hasn't been specialized for <code>tf.zeros_like()<\/code>. I've filed a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/744\" rel=\"noreferrer\">GitHub issue for that<\/a> and it should be fixed soon.<\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong> In your comment, you asked how to deal with the case where the zero tensor had a shape based on, but <strong>different from<\/strong> the original tensor. This is also possible, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/shape\" rel=\"noreferrer\"><code>tf.shape()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"noreferrer\"><code>tf.stack()<\/code><\/a> to build the dimensions, and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/fill\" rel=\"noreferrer\"><code>tf.fill()<\/code><\/a> to produce the zero tensor:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=[None, 4])\n\n# Use tf.shape() to get the runtime size of `x` in the 0th dimension.\nzeros_dims = tf.stack([tf.shape(x)[0], 7])\n\ny = tf.fill(zeros_dims, 0.0)\n\nsess = tf.Session()\ny_result = sess.run(y, feed_dict={x: np.random.rand(4, 4)})\nprint y_result.shape\n# ==&gt; (4, 7)\n<\/code><\/pre>\n"}]}
{"QuestionId":34725157,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-01-11T15:26:46.237","AcceptedAnswerId":34725213.0,"OwnerUserId":2687955.0,"Title":"How to reshape a vector to TensorFlow's filters?","Body":"<p>I want to transfer some weights trained by another network to TensorFlow, the weights are stored in a single vector like this:<\/p>\n\n<p><\/p>\n\n<p>By using numpy, I can reshape it to two 3 by 3 filters like this:<\/p>\n\n<pre><\/pre>\n\n<p>Thus, the shape of my filters are . However, in TensorFlow, the shape of filters are :<\/p>\n\n<pre><\/pre>\n\n<p>After reshaping the tf_weights to the expected shape, the weight becomes a mess and I can't get the expected convolution result.<\/p>\n\n<p>To be specific, when the shape of an image or filter is [number,channel,size,size], I wrote a convolution function and it gives the correct answer,but it's too slow:<\/p>\n\n<pre><\/pre>\n\n<p>Instead, by using tensorflow's conv2d like this:<\/p>\n\n<pre><\/pre>\n\n<p>The feature map I got is wrong.<\/p>\n","answers":[{"AnswerId":"43021828","CreationDate":"2017-03-25T20:52:28.137","ParentId":null,"OwnerUserId":"7601910","Title":null,"Body":"<h2>Sample Tensor Manipulations<\/h2>\n\n<p>I dont know if this might be of help. Consider the Reshape ,Gather, Dynamic_partition and Split operations and adapt this to your needs.\nIn what comes below is the illustration of these operations that can be adapted to use in your situation. I copied this from my git repo. I will believe if you run this examples in ipython you can figure out what you really want and get even better insight.<\/p>\n\n<h3>Reshape ,Gather, Dynamic_partition and Split<\/h3>\n\n<h3>Gather Operation ( tf.gather( ) )<\/h3>\n\n<p>Generate an array and test the gather operation. Note this approach for fast prototyping:<\/p>\n\n<ul>\n<li>We generate an array in Numpy and test the operations of tensor flow on it.<\/li>\n<\/ul>\n\n<p>Use: Gather slices from params according to indices.<\/p>\n\n<p>indices must be an integer tensor of any dimension (usually 0-D or 1-D). This is best illustrated by an example:<\/p>\n\n<pre><code>array = np.array([[1,2,3],[4,9,6],[2,3,4],[7,8,0]])\n\narray.shape\n\n\n(4, 3)\n\nIn [27]:\n\ngather_output0  = tf.gather(array,1)\ngather_output01  = tf.gather(array,2)\ngather_output02  = tf.gather(array,3)\n\ngather_output11  = tf.gather(array,[1,2])\ngather_output12  = tf.gather(array,[1,3])\ngather_output13  = tf.gather(array,[3,2])\n\n\n\n\ngather_output  = tf.gather(array,[1,0,2])\ngather_output1  = tf.gather(array,[1,1,2])\ngather_output2  = tf.gather(array,[1,2,1])\n\nIn [28]:\n\nwith tf.Session() as sess:\n    print (gather_output0.eval());print(\"\\n\")\n    print (gather_output01.eval());print(\"\\n\")\n    print (gather_output02.eval());print(\"\\n\")  \n    print (gather_output11.eval());print(\"\\n\")\n    print (gather_output12.eval());print(\"\\n\")\n    print (gather_output13.eval());print(\"\\n\")\n\n    print (gather_output.eval());print(\"\\n\")\n    print (gather_output1.eval());print(\"\\n\")\n    print (gather_output2.eval());print(\"\\n\")\n    #print (gather_output2.eval());print(\"\\n\")\n\n[4 9 6]\n\n\n[2 3 4]\n\n\n[7 8 0]\n\n\n[[4 9 6]\n [2 3 4]]\n\n\n[[4 9 6]\n [7 8 0]]\n\n\n[[7 8 0]\n [2 3 4]]\n\n\n[[4 9 6]\n [1 2 3]\n [2 3 4]]\n\n\n[[4 9 6]\n [4 9 6]\n [2 3 4]]\n\n\n[[4 9 6]\n [2 3 4]\n [4 9 6]]\n<\/code><\/pre>\n\n<p>And looking at this simple example:<\/p>\n\n<ul>\n<li>Initialise  simple array<\/li>\n<li><p>test gather operation<\/p>\n\n<p>In [11]:<\/p>\n\n<pre><code>array_simple = np.array([1,2,3])\n\nIn [15]:\n\nprint \"shape of simple array is: \", array_simple.shape\n\nshape of simple array is:  (3,)\n\nIn [57]:\n\ngather1  = tf.gather(array1,[0])\ngather01 = tf.gather(array1,[1])\ngather02 = tf.gather(array1,[2])\n\ngather2 = tf.gather(array1,[1,2])\ngather3 = tf.gather(array1,[0,1])\n\nwith tf.Session() as sess:\n    print (gather1.eval());print(\"\\n\")\n    print (gather01.eval());print(\"\\n\")\n    print (gather02.eval());print(\"\\n\")\n    print (gather2.eval());print(\"\\n\")\n    print (gather3.eval());print(\"\\n\")\n\n[1]\n\n\n[2]\n\n\n[3]\n\n\n[2 3]\n\n\n[1 2]\n\n\ntf.reshape( )\n\nNote:\n\n*  Use the same array that was initiated\n*  Do reshape using tf.reshape( )\n\nIn [64]:\n\narray.shape # Confirm array shape\n\nOut[64]:\n\n(4, 3)\n\nIn [74]:\n\nprint (\"This is the array\\n\" ,array) # see the output and compare with the initial array,\n\nThis is the array\n [[1 2 3]\n [4 9 6]\n [2 3 4]\n [7 8 0]]\n\nIn [84]:\n\nreshape_ops= tf.reshape(array,[-1,4]) # Note the parameters in reshpe\nreshape_ops1= tf.reshape(array,[-1,3]) # Note the parameters in reshpe\nreshape_ops2= tf.reshape(array,[-1,6]) # Note the parameters in reshpe\n\nreshape_ops_back1= tf.reshape(array,[6,-1]) # Note the parameters in reshpe\nreshape_ops_back2= tf.reshape(array,[3,-1]) # Note the parameters in reshpe\nreshape_ops_back3= tf.reshape(array,[4,-1]) # Note the parameters in reshpe\n\nIn [86]:\n\nwith tf.Session() as sess:\n    print(reshape_ops.eval());print(\"\\n\")\n    print(reshape_ops1.eval());print(\"\\n\")\n    print(reshape_ops2.eval());print(\"\\n\")\n    print (\"Output when we reverse the parameters:\");print(\"\\n\")\n    print(reshape_ops_back1.eval());print(\"\\n\")\n    print(reshape_ops_back2.eval());print(\"\\n\")\n    print(reshape_ops_back3.eval());print(\"\\n\")\n\n[[1 2 3 4]\n [9 6 2 3]\n [4 7 8 0]]\n\n\n[[1 2 3]\n [4 9 6]\n [2 3 4]\n [7 8 0]]\n\n\n[[1 2 3 4 9 6]\n [2 3 4 7 8 0]]\n\n\nOutput when we reverse the parameters:\n\n\n[[1 2]\n [3 4]\n [9 6]\n [2 3]\n [4 7]\n [8 0]]\n\n\n[[1 2 3 4]\n [9 6 2 3]\n [4 7 8 0]]\n\n\n[[1 2 3]\n [4 9 6]\n [2 3 4]\n [7 8 0]]\n<\/code><\/pre>\n\n<p>Note: The input size and output size must be the same. ---otherwise it gives error. Simple way to check this out is to make sure the input can be paritioned into the the reshape parameters by doing simple multiplications.<\/p><\/li>\n<\/ul>\n\n<h3>Dynamic_cell_partitions<\/h3>\n\n<pre><code>This is declared as :\n\ntf.dynamic_partition (array, partitions, num_partitions, name=None)\n\nNote:\n\n* we decalare number_partitions --- number of partitions\n* Use our array initialised earlier\n* We declare the partition as [0 1 0 1] . This signifies the partitions we want 0's fall to one partition and 1 the other partitions given that we have two num_partitions=2.\n\n* The output is a list\n\nIn [96]:\n\n    print (\"This is the array\\n\" ,array) # This is output array\n\n    This is the array\n     [[1 2 3]\n     [4 9 6]\n     [2 3 4]\n     [7 8 0]]\n\n    We show how to make two and three partitions below\n    In [123]:\n\n    num_partitions = 2\n    num_partitions1 = 3\n\n    partitions = [0, 0, 1, 1]\n    partitions1 = [0 ,1 ,1, 2 ]\n\n    In [119]:\n\n    dynamic_ops =tf.dynamic_partition(array, partitions, num_partitions, name=None) # 2 partitions\n    dynamic_ops1 =tf.dynamic_partition(array, partitions1, num_partitions1, name=None) # 3 partitions\n\n    In [125]:\n\n    with tf.Session() as sess:\n        run = sess.run(dynamic_ops)\n        run1 = sess.run(dynamic_ops1)\n        print(\"Output for 2 partitions: \")\n        print (run[0]);print(\"\\n\")\n        print(run[1]) ;print(\"\\n\")# Compare result with initial array. Out is list\n        print(\"Output for three partitions: \")\n\n        print (run1[0]);print(\"\\n\")\n        print (run1[1]);print(\"\\n\")\n        print (run1[2]);print(\"\\n\")\n\n    Output for 2 partitions: \n    [[1 2 3]\n     [4 9 6]]\n\n\n    [[2 3 4]\n     [7 8 0]]\n\n\n    Output for three partitions: \n    [[1 2 3]]\n\n\n    [[4 9 6]\n     [2 3 4]]\n\n\n    [[7 8 0]]\n<\/code><\/pre>\n\n<h3>tf.split( )<\/h3>\n\n<p>Make sure you use an up to date tensorflow version. Otherwise in older versions, this implemetation will give error<\/p>\n\n<p>This is specified in the documentation as below:<\/p>\n\n<p>tf.split(value, num_or_size_splits, axis=0, num=None, name='split').<\/p>\n\n<p>It splits a tensor into subtensors. This is best illustrated by an example:<\/p>\n\n<pre><code>* we define (5,30) aray in numpy\n* we split the array along axis 1\n* We  specify the number of splits as 1-Dimen Tensor along axis 1. So we have 3 splits.\n\nSpecify an array\n\n    Create a (5 by 30) numpy array. The syntax using numpy is shown below\n    In [2]:\n\n    ArrayBeforeSplitting = np.arange(150).reshape(5,30) \n    print (\"Array shape without split operation is : \" ,ArrayBeforeSplitting.shape)\n\n    ('Array shape without split operation is : ', (5, 30))\n\n    specify number of splits\n    In [3]:\n\n    split_1D = tf.Variable([8,13,9])\n    print(\"specify number of partions using 1-Dimen Variable:\" , tf.shape(split_1D))\n\n    ('specify number of partions using 1-Dimen Variable:', &lt;tf.Tensor 'Shape:0' shape=(1,) dtype=int32&gt;)\n\n    Use tf.split\n\n    Make 3 splits aong y axis so that we have (5,8) ,(5,13),(5,9) splits. The axis 1 add up to give 30-- we can see axis 1 has 30 elements so the partition along that axis should add up to 30 otherwise it gives error.\n    In [6]:\n\n    split1,split2,split3 = tf.split(ArrayBeforeSplitting,split_1D,1)\n    # we have 3 splits along axis 1 specified spcifically\n    # by the split_1D . That is split axis 1D (with 30 elements) into partions with 8 ,13, and 9 elements while the x axis\n    #remains constant\n\n    In [7]:\n\n    #INitialise global variables. because split_ID is a variable and needs to be initialised before being\n    #used in a computational graph\n    init_op = tf.global_variables_initializer()\n\n    In [16]:\n\n    with tf.Session() as sess:\n        sess.run(init_op) # run variable initialisation.\n        result=split1.eval();print(\"\\n\")\n        print(result)\n        print(\"the shape of the first split operation is : \",result.shape)\n        result2=split2.eval();print(\"\\n\")\n        print(result2)\n        print(\"the shape of the second split operation is : \",result2.shape)\n\n        result3=split3.eval();print(\"\\n\")\n        print(result3)\n        print(\"the shape of the third split operation is : \",result3.shape)\n\n\n    [[  0   1   2   3   4   5   6   7]\n     [ 30  31  32  33  34  35  36  37]\n     [ 60  61  62  63  64  65  66  67]\n     [ 90  91  92  93  94  95  96  97]\n     [120 121 122 123 124 125 126 127]]\n    ('the shape of the first split operation is : ', (5, 8))\n\n\n    [[  8   9  10  11  12  13  14  15  16  17  18  19  20]\n     [ 38  39  40  41  42  43  44  45  46  47  48  49  50]\n     [ 68  69  70  71  72  73  74  75  76  77  78  79  80]\n     [ 98  99 100 101 102 103 104 105 106 107 108 109 110]\n     [128 129 130 131 132 133 134 135 136 137 138 139 140]]\n    ('the shape of the second split operation is : ', (5, 13))\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"},{"AnswerId":"34725213","CreationDate":"2016-01-11T15:29:35.907","ParentId":null,"OwnerUserId":"525169","Title":null,"Body":"<p>Don't use <code>np.reshape<\/code>. It might <a href=\"https:\/\/stackoverflow.com\/a\/27286794\/525169\">mess up the order of your values<\/a>.<\/p>\n\n<p>Use <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/reference\/generated\/numpy.rollaxis.html\" rel=\"nofollow noreferrer\"><code>np.rollaxis<\/code><\/a> instead:<\/p>\n\n<pre><code>&gt;&gt;&gt; a = np.array([1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18])\n&gt;&gt;&gt; a = a.reshape((1,2,3,3))\n&gt;&gt;&gt; a\narray([[[[ 1,  2,  3],\n         [ 4,  5,  6],\n         [ 7,  8,  9]],\n\n        [[10, 11, 12],\n         [13, 14, 15],\n         [16, 17, 18]]]])\n&gt;&gt;&gt; b = np.rollaxis(a, 1, 4)\n&gt;&gt;&gt; b.shape\n(1, 3, 3, 2)\n&gt;&gt;&gt; b = np.rollaxis(b, 0, 4)\n&gt;&gt;&gt; b.shape\n(3, 3, 2, 1)\n<\/code><\/pre>\n\n<p>Note that the order of the two axes with size 3 haven't changed. If I were to label them, the two <code>rollaxis<\/code> operations caused the shapes to change as (1, 2, 3<sup>1<\/sup>, 3<sup>2<\/sup>) -> (1, 3<sup>1<\/sup>, 3<sup>2<\/sup>, 2) -> (3<sup>1<\/sup>, 3<sup>2<\/sup>, 2, 1). Your final array looks like:<\/p>\n\n<pre><code>&gt;&gt;&gt; b\narray([[[[ 1],\n         [10]],\n\n        [[ 2],\n         [11]],\n\n        [[ 3],\n         [12]]],\n\n\n       [[[ 4],\n         [13]],\n\n        [[ 5],\n         [14]],\n\n        [[ 6],\n         [15]]],\n\n\n       [[[ 7],\n         [16]],\n\n        [[ 8],\n         [17]],\n\n        [[ 9],\n         [18]]]])\n<\/code><\/pre>\n"}]}
{"QuestionId":34727431,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter>","CreationDate":"2016-01-11T17:20:25.997","AcceptedAnswerId":34771937.0,"OwnerUserId":3225904.0,"Title":"TensorFlow on Jupyter: Can't restore variables","Body":"<p>I can't seem to be able to restore saved variables when using TensorFlow in a Jupyter notebook. I train an ANN, then I run  then I train it again, save the new result  but when I run  my model doesn't load the values saved on  and keeps those in .<\/p>\n\n<p>If I run the model, save it on , then close and halt, then try to load it again, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>My code for training is:<\/p>\n\n<pre><\/pre>\n\n<p>Am I doing anything wrong? Why can't I ever restore my variables?<\/p>\n","answers":[{"AnswerId":"34771937","CreationDate":"2016-01-13T16:28:59.737","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like you are using Jupyter to build your model. One possible issue, when constructing a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow noreferrer\"><code>tf.Saver<\/code><\/a> with the default arguments is that it will use the (auto-generated) names for the variables as the keys in your checkpoint. Since in Jupyter its easy to re-execute code cells multiple times, you might be ending up with multiple copies of the variable nodes in the session that you save. See <a href=\"https:\/\/stackoverflow.com\/questions\/34500052\/tensorflow-saving-and-restoring-session\/34500690#34500690\">my answer to this question<\/a> for an explanation of what can go wrong.<\/p>\n\n<p>There are a few possible solutions. Here are the easiest:<\/p>\n\n<ul>\n<li><p>Call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#reset_default_graph\" rel=\"nofollow noreferrer\"><code>tf.reset_default_graph()<\/code><\/a> before you build your model (and the <code>Saver<\/code>). This will ensure that the variables get the names you intended, but it will invalidate previously-created graphs.<\/p><\/li>\n<li><p>Use explicit arguments to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow noreferrer\"><code>tf.train.Saver()<\/code><\/a> to specify the persistent names for the variables. For your example this shouldn't be too hard (though it becomes unwieldy for larger models):<\/p>\n\n<pre><code>saver = tf.train.Saver(var_list={\"b1\": b1, \"W1\": W1, \"b2\": b2, \"W2\": W2})\n<\/code><\/pre><\/li>\n<li><p>Create a new <code>tf.Graph()<\/code> and make it the default each time you create the model. This can be tricky in Jupyter, since it forces you to put all of the model building code in one cell, but it works well for scripts:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n  # Model building and training\/evaluation code goes here.\n<\/code><\/pre><\/li>\n<\/ul>\n"}]}
{"QuestionId":34727792,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-11T17:41:55.730","AcceptedAnswerId":34729874.0,"OwnerUserId":691867.0,"Title":"What's difference between `variable.name` and `variable.op.name`?","Body":"<p>Suppose  is a , I saw use of  like  in the tutorial. I am wondering if I can replace  with  in general.<\/p>\n\n<p>What's the difference between the two? Are they interchangeable?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34729874","CreationDate":"2016-01-11T19:51:30.713","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Right now the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.name\" rel=\"noreferrer\"><code>Variable.name<\/code><\/a> property maps to the name of the mutable <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor\" rel=\"noreferrer\"><code>Tensor<\/code><\/a> in which that variable is stored (principally because a <code>Variable<\/code> can be used wherever a <code>Tensor<\/code> is expected). Tensor names are generated from the name of the operation that produces them (a <code>Variable<\/code> op in this case) and the index of the output to which that tensor corresponds.<\/p>\n\n<p>You can feel free to use <code>tf.scalar_summary(x.name, x)<\/code> in place of <code>tf.scalar_summary(x.op.name, x)<\/code>, but the resulting visualizations will contain a redundant <code>\":&lt;N&gt;\"<\/code> in the tags.<\/p>\n"}]}
{"QuestionId":34734054,"AnswerCount":1,"Tags":"<python><argparse><tensorflow>","CreationDate":"2016-01-12T01:19:20.757","AcceptedAnswerId":34747727.0,"OwnerUserId":2594660.0,"Title":"Tensorflow ArgumentError Running CIFAR-10 example","Body":"<p>I am trying to run the CIFAR-10 example of Tensorflow. However when executing  I am getting the error attached below. <\/p>\n\n<p>I have installed Version 0.6.0 of the Tensorflow package using pip. The framework is working fine on other models including the MNIST tutorial and some self developed networks. Does anybody has an idea about the origin of the problem? Do you think I should open an issue on github?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34747727","CreationDate":"2016-01-12T15:35:29.720","ParentId":null,"OwnerUserId":"2594660","Title":null,"Body":"<p>Solved the problem. One is not supposed to execute <code>python cifar10.py<\/code> directly, but one should run <code>python cifar10_train.py<\/code> or <code>python cifar10_eval.py<\/code> instead.<\/p>\n"}]}
{"QuestionId":34739151,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><conv-neural-network><keras>","CreationDate":"2016-01-12T08:54:02.407","AcceptedAnswerId":34739417.0,"OwnerUserId":3200890.0,"Title":"Calculate dimension of feature maps in convolutional neural network","Body":"<p>I have convolutional neural network in Keras.\nI need to know the dimensions of the feature maps in each layer.\nMy input is 28 by 28 pixel image. I know theres a way to calculate this I not sure how. Below is my code snippet using Keras.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FypHL.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FypHL.jpg\" alt=\"What i want to draw\"><\/a><\/p>\n\n<p>At the end of the day, this is what i want to draw. Thank you.<\/p>\n","answers":[{"AnswerId":"34739417","CreationDate":"2016-01-12T09:08:30.350","ParentId":null,"OwnerUserId":"4229825","Title":null,"Body":"<p>Check <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/#conv\" rel=\"nofollow\">this<\/a> article. <\/p>\n\n<p>Formula for spatial size of the output volume: <strong>K*((W\u2212F+2P)\/S+1)<\/strong>, where <strong>W<\/strong> - input volume size, <strong>F<\/strong> the receptive field size of the Conv Layer neurons, <strong>S<\/strong> - the stride with which they are applied, <strong>P<\/strong> - the amount of zero padding used on the border, <strong>K<\/strong> - the depth of conv layer.<\/p>\n"}]}
{"QuestionId":34742330,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-01-12T11:21:31.490","AcceptedAnswerId":null,"OwnerUserId":492372.0,"Title":"Why this difference between the local response norm paper equation and tensorflow implementation?","Body":"<p>I came across this equation in Alex's <a href=\"http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow noreferrer\">paper<\/a> on local response normalization:\n  <a href=\"https:\/\/i.stack.imgur.com\/QYvwf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QYvwf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As we see above, the power is raised after we calculate the sum, multiply it with alpha and then add once k is added.<\/p>\n\n<p>However, I see in TensorFlow documentation it is showed as <\/p>\n\n<pre><\/pre>\n\n<p>where the beta is raised only for the sum.<\/p>\n\n<p>Why is there a discrepancy here? <\/p>\n\n<p>Also, when I looked into the TensorFlow code itself, I saw this:<\/p>\n\n<pre><\/pre>\n\n<p>which looks correct? <\/p>\n\n<p>I am kind of confused here. Can someone correct me please?<\/p>\n","answers":[{"AnswerId":"38061809","CreationDate":"2016-06-27T19:30:40.937","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>This was a bug in the Tensorflow documentation.<\/p>\n\n<p>It was fixed by this commit on Feb 26:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/ab48dbd4ac2095548a5bc8505e08e751d409727f#diff-632987400e5affbcdba4533444460b0e\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/commit\/ab48dbd4ac2095548a5bc8505e08e751d409727f#diff-632987400e5affbcdba4533444460b0e<\/a><\/p>\n"}]}
{"QuestionId":34742424,"AnswerCount":2,"Tags":"<java><caching><caffeine>","CreationDate":"2016-01-12T11:26:13.760","AcceptedAnswerId":null,"OwnerUserId":5778519.0,"Title":"Max eden space size in caffeine cache","Body":"<p>Is there a way to configure the maximum eden space size in caffeine cache? If not, what is the default size? <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/blob\/master\/simulator\/src\/main\/java\/com\/github\/benmanes\/caffeine\/cache\/simulator\/policy\/sketch\/WindowTinyLfuPolicy.java\" rel=\"nofollow\">https:\/\/github.com\/ben-manes\/caffeine\/blob\/master\/simulator\/src\/main\/java\/com\/github\/benmanes\/caffeine\/cache\/simulator\/policy\/sketch\/WindowTinyLfuPolicy.java<\/a><\/p>\n","answers":[{"AnswerId":"34752136","CreationDate":"2016-01-12T19:24:47.643","ParentId":null,"OwnerUserId":"19450","Title":null,"Body":"<p>The term <code>eden<\/code> was borrowed from JVM generational garbage collectors, where most objects die young and are not promoted. We later decided on <code>window<\/code> when coming up with a name for the eviction policy (<code>Window TinyLFU<\/code>).<\/p>\n\n<p>In the <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/wiki\/Simulator\" rel=\"nofollow\">simulator<\/a> you can adjust the window size by modifying the <code>percent-main<\/code> in the <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/blob\/master\/simulator\/src\/main\/resources\/reference.conf#L171\" rel=\"nofollow\">configuration file<\/a>. That makes it easy to experiment with different trace files when evaluating policies and the impact of tuning options.<\/p>\n\n<p>For the production facing cache, the window size is not a configuration parameter. This value is hard coded to 1%, which was near optimal across a variety of workloads and avoids TinyLFU's degredation scenarios. In some workloads (OLTP disk traces) a larger value has a small gain, but not enough to expose a configuration option. Doing so wouldn't carry its conceptual weight by confusing most users and leak implementation details.<\/p>\n"},{"AnswerId":"34742909","CreationDate":"2016-01-12T11:49:42.017","ParentId":null,"OwnerUserId":"575556","Title":null,"Body":"<p>you can add eden space declaration at the start of JVM using the parameters:<\/p>\n\n<p>-XX:NewSize=size<\/p>\n\n<p>-XX:MaxNewSize=size<\/p>\n\n<p>Ideally Eden space = 3\/8 total heap size. <\/p>\n"}]}
{"QuestionId":34743847,"AnswerCount":0,"Tags":"<python><tensorflow><loss>","CreationDate":"2016-01-12T12:34:34.577","AcceptedAnswerId":null,"OwnerUserId":3312142.0,"Title":"TensorFlow convolution huge loss function values","Body":"<p>I've been playing around with the TensorFlow Cifar10 demo (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks<\/a>), and I ran into the following funny-looking output when implementing AlexNet (227x227 images) with batch size 1 inside Google's demo code:<\/p>\n\n<p><\/p>\n\n<p>Why is the loss value so huge after a few batches? The network also does not seem to converge to anything above random (I'm training it to determine how good the color balance is in a photo). I will only post the code that I changed - with the exception of small directory name changes and image size parameters, as the rest is available on the TensorFlow website. Let me know if you need to see more of the code though.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34745196,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-12T13:37:38.760","AcceptedAnswerId":34745690.0,"OwnerUserId":3419997.0,"Title":"Weights in Convolutional network?","Body":"<p>I am currently following the TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html#build-a-multilayer-convolutional-network\" rel=\"noreferrer\">Multilayer Convolutional Network<\/a> tutorial.<\/p>\n\n<p>In various layers weight is initialised as followed :<\/p>\n\n<ul>\n<li><p>First Convolutional Layer:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Second Convolutional Layer:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Densely Connected Layer:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Readout Layer:<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n\n<p>So I am having doubts in <strong>how is the shape of the above weight variables known to us ?<\/strong><\/p>\n\n<p><strong>Is their any math used to find the shape for them ?<\/strong> <\/p>\n","answers":[{"AnswerId":"34745690","CreationDate":"2016-01-12T13:59:32.577","ParentId":null,"OwnerUserId":"1391220","Title":null,"Body":"<p>The answer is explained on the same page:<\/p>\n\n<blockquote>\n  <p>The convolutional will compute 32 features for each 5x5 patch. Its\n  weight tensor will have a shape of [5, 5, 1, 32]<\/p>\n<\/blockquote>\n\n<p>There is no involved math par say, but these terms need explanation<\/p>\n\n<ol>\n<li>The size of convolution kernel is <code>5X5<\/code>. That means there is a <code>5X5<\/code> matrix that is <em>convolved<\/em> with an input image by moving it around the image. Check <a href=\"http:\/\/setosa.io\/ev\/image-kernels\/\">this link<\/a> for an explanation of how a small <code>5X5<\/code> matrix moves over a <code>28X28<\/code> image and multiplies different cells of the image matrix with itself. This gives us first two dimentsions of <code>[5, 5, 1, 32]<\/code><\/li>\n<li>The size of input channels is <code>1<\/code>. These are BW images, hence one input channel. Most colored images have 3 channels, so expect a <code>3<\/code> in some other convolution networks working on images. Indeed, for the second layer, <code>W_conv2<\/code>, the number of input channels is <code>32<\/code>, same as number of output channels of layer 1.<\/li>\n<li>The last dimension of the weight matrix is perhaps hardest to visualize. Imagine your <code>5X5<\/code> matrix, and replicate it 32 times!. Each of these 32 things are called <code>channels<\/code>. To complete the discussion, each of these 32 <code>5X5<\/code> matrices are initialized with random weights and trained independently during forward\/back propagation of the network. More channels learn different aspects of the image and hence give extra power to your network.<\/li>\n<\/ol>\n\n<p>If you summarize these 3 points, you get desired dimensions of layer 1. Subsequent layers are an extension - First two dimensions are kernel sizes (5X5) in this case. Third dimension is equal to size of input channel, which is equal to size of output channel of previous layer. (32, since we declared 32 output channels  of layer 1). Final dimension is the size of output channel of current layer (64, even lager for second layer!. Again, keeping a large number of independent 5X5 kernels helps!). <\/p>\n\n<p>Finally, last two layers: Final dense layer is the only thing that involves some calculation: <\/p>\n\n<ol>\n<li>For each convolution layer, final size = initial size<\/li>\n<li>For pooling layer of size kXk, final size = initial size \/ k<\/li>\n<\/ol>\n\n<p>So,<\/p>\n\n<ol>\n<li>For conv1, size remains <code>28 X 28<\/code><\/li>\n<li>pool1 reduces size to <code>14 X 14<\/code><\/li>\n<li>For conv2, size remains <code>14 X 14<\/code><\/li>\n<li>pool2 reduces size to <code>7 X 7<\/code><\/li>\n<\/ol>\n\n<p>And ofcourse, we have <code>64<\/code> channels due to conv2 - pooling doesn't affect them. Hence, we get a final dense input of <code>7X7X64<\/code>. We then create fully connected <code>1024<\/code> hidden layers and add <code>10<\/code> output classes for <code>10<\/code> digits.<\/p>\n"}]}
{"QuestionId":34746777,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-12T14:51:13.487","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Why do I get ValueError('\\'image\\' must be fully defined.') when transforming image in Tensorflow?","Body":"<p>I want to do real time data augmentation by chaining different image transformation operators in tensorflow. My code begins with image decoding and then runs different transformations but it throw a . Here is an example to reproduce this error :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34747345","CreationDate":"2016-01-12T15:18:58.853","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This error arises because the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#random_flip_left_right\" rel=\"noreferrer\"><code>tf.image.random_flip_left_right()<\/code><\/a> op checks the static shape of its input when you build the graph, and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/image.html#decode_jpeg\" rel=\"noreferrer\"><code>tf.image.decode_jpeg()<\/code><\/a> produces tensors that have a data dependency on the contents of <code>image_raw<\/code> so it the shape isn't statically known. Currently the only way to work around this is to set the static shape of the <code>decoded<\/code> tensor using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor.set_shape\" rel=\"noreferrer\"><code>Tensor.set_shape()<\/code><\/a>, as follows:<\/p>\n\n<pre><code>decoded = tf.image.decode_jpeg(image_raw)\ndecoded.set_shape([IMAGE_HEIGHT, IMAGE_WIDTH, NUM_CHANNELS])\nflipped = tf.image.random_flip_left_right(decoded)\n<\/code><\/pre>\n\n<p>The downside of this is that all images must now have the same size (and number of channels).<\/p>\n\n<p>Many of the image ops don't follow the same gradual and dynamic shape inference as the rest of TensorFlow (which allows you to have unknown shapes or dimensions, assumes that the program is correct as you build the  graph, and checks the real shapes at runtime). This is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/521\" rel=\"noreferrer\">considered a bug<\/a> at the present time, and we'll figure out a way to fix it.<\/p>\n"}]}
{"QuestionId":34747982,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-12T15:47:29.733","AcceptedAnswerId":34750336.0,"OwnerUserId":1940138.0,"Title":"Tensorflow: share value for two different variables within same operation","Body":"<p>\nI have been experimenting with TensorFlow (TF) lately and I came across this problem: say I want to compute the value and the gradient of the function <\/p>\n\n<p><img src=\"https:\/\/latex.codecogs.com\/gif.latex?f%28%5Cmathbf%7Bx%7D%29%20%3D%20%5Csum_%7Bijk%7D%20J_%7Bijk%7D%20x_i%20x_j%20x_k\" alt=\"f(x) = \\sum_{ijk} x_i x_j x_k\"><\/p>\n\n<p>where the x's are indexed differently but all refer to the same vector <img src=\"https:\/\/latex.codecogs.com\/gif.latex?%5Cmathbf%7Bx%7D\" alt=\"bold x\"> and the J's are random constants (in physics this is a spin glass model). The gradient wrt <img src=\"https:\/\/latex.codecogs.com\/gif.latex?x_k\" alt=\"x_k\"> is then simply<\/p>\n\n<p><img src=\"https:\/\/latex.codecogs.com\/gif.latex?%5Cfrac%7B%5Cpartial%20f%28%5Cmathbf%7Bx%7D%29%7D%7B%5Cpartial%20x_k%7D%20%3D%20%5Csum_%7Bij%7D%20J_%7Bijk%7D%20x_i%20x_j\" alt=\"grad_k(x) = sum_ij x_i*x_j\"><\/p>\n\n<p>hence  sums over N^3 terms and  sums N times over N^2 terms. I have implemented  by generating all the terms of the sum as a rank 3 tensor and sum-reducing over all the entries. Then to differentiate I apply<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>where f is the loss function and xk a variable. Here's a MWE where assume all J's to be 1<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I need to call  many times independently and I want to reduce the number of variable assignments to just one since x1, x2 refer to the same vector. <\/p>\n\n<p>Some profiling on a related example for  (on a GeForce GTX 650) showed that<\/p>\n\n<ul>\n<li>cuMemcpyDtoHAsync takes 63% of the time<\/li>\n<li>cuMemcpyHtoDAsync 18% and <\/li>\n<li>cuEventRecord 18%. <\/li>\n<\/ul>\n\n<p>(results are similar for this mwe)<\/p>\n\n<p>Hence assignment is the most expensive operation when performing the computation on GPUs. Obviously the overhead gets worse for increasing , hence partially neutralising the benefit of using GPUs.<\/p>\n\n<p>Any suggestion on how I could be able to do reduce overhead by transferring x only once? <\/p>\n\n<p>Also any other suggestion on how to reduce any other overhead would be immensely appreciated.<\/p>\n\n<h2>EDIT<\/h2>\n\n<p>To show the problem in action I'll follow the suggestion by <em>mrry<\/em>.\nIf I were to replace all instances of x2 with x1 then the MWE would look like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>and the second assertion would fail because each entry for the gradient would be 300 instead of 100, as it should be. The reason is that while xi, xj, xk all refer to the same vector, they are symbolically distinct: replacing all x with the same variable would result in the derivative of x^3, which is 3*x^2, hence the result of the second MWE.<\/p>\n\n<p>P.S. I have also explicitly assigned x1 for clarity<\/p>\n","answers":[{"AnswerId":"34750336","CreationDate":"2016-01-12T17:39:44.817","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>One way to achieve your desired outcome is to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#stop_gradient\" rel=\"nofollow\"><code>tf.stop_gradient()<\/code><\/a> op to make an efficient copy of the variable <code>x1<\/code> without it contributing to the gradient:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\n# First define the variable.\nn = 10 # size of x                                                                                                                                                                               \nx1 = tf.Variable(tf.zeros([n], dtype=tf.float64))\nx2 = tf.stop_gradient(x1)\n\n# Now define the cost function                                                                                                                                                              \nf_tensor = tf.mul(tf.mul(tf.reshape(x1, [n]),\n                         tf.reshape(x2, [n,1])),\n                  tf.reshape(x2, [n,1,1]))\nf = tf.reduce_sum(f_tensor)\n\nsession = tf.Session()\ninit = tf.initialize_all_variables()\nsession.run(init)\n\n# Run on test array                                                                                                                                                                            \nxtest = np.ones(n)\nres = session.run([f, tf.gradients(f, x1)[0]],\n                  feed_dict={x1 : xtest})\n\nassert res[0] == 1000\nfor g in res[1]:\n    assert g == 100\n<\/code><\/pre>\n"},{"AnswerId":"34936123","CreationDate":"2016-01-21T23:08:48.167","ParentId":null,"OwnerUserId":"938911","Title":null,"Body":"<p>I couldn't comment above (not enough reputation), but note that the analytical gradient should be<\/p>\n\n<p>$$\n\\frac{\\partial f}{\\partial x_k} = \\sum_{ij} J_{ijk} x_i x_j + \\sum_{ij} J_{ikj} x_i x_j + \\sum_{ij} J_{kij} x_i x_j.\n$$<\/p>\n"}]}
{"QuestionId":34748234,"AnswerCount":1,"Tags":"<syntax><lua><torch>","CreationDate":"2016-01-12T15:57:58.353","AcceptedAnswerId":34770817.0,"OwnerUserId":350664.0,"Title":"Double curly-brace syntax in Torch","Body":"<p>I am trying to better understand this syntax in Lua when using Torch:<\/p>\n\n<pre><\/pre>\n\n<p>Specifically the line<\/p>\n\n<pre><\/pre>\n\n<p>When I run the code, it prints<\/p>\n\n<pre><\/pre>\n\n<p>I am familiar with table literals and their basic syntax, but what is the purpose of wrapping the '2' in double curly-braces, and how does it work in the engine?<\/p>\n","answers":[{"AnswerId":"34770817","CreationDate":"2016-01-13T15:39:35.280","ParentId":null,"OwnerUserId":"350664","Title":null,"Body":"<p>The following answer was posted on GitHub when I asked the same question:\n<a href=\"https:\/\/github.com\/torch\/torch7\/issues\/501#issuecomment-171290546\" rel=\"nofollow\">https:\/\/github.com\/torch\/torch7\/issues\/501#issuecomment-171290546<\/a><\/p>\n\n<hr>\n\n<p>Have a look at <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor--dim1dim2--or--dim1sdim1e-dim2sdim2e-\" rel=\"nofollow\">this part of the documentation<\/a>.<\/p>\n\n<ul>\n<li>When you have single curly-braces, you are creating a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/03c04c6ec77275c16f30ea86d3d7b11bc3233e21\/generic\/Tensor.c#L935-L947\" rel=\"nofollow\">selection of the tensor<\/a>. So ten[{1}] is <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/03c04c6ec77275c16f30ea86d3d7b11bc3233e21\/generic\/Tensor.c#L882-L902\" rel=\"nofollow\">equivalent<\/a> to ten[1], which is in turn equivalent to ten:select(1,1). If you have several indices like ten[{1,2}], this is also equivalent to the slower ten[1][2] (because the latter returns 2 times a tensor, whereas the former only returns it once). If you select on a 1D tensor, your output is a number.<\/li>\n<li>When you have double curly-braces, it returns a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/03c04c6ec77275c16f30ea86d3d7b11bc3233e21\/generic\/Tensor.c#L948-L972\" rel=\"nofollow\">narrow of the tensor<\/a>, and a narrowed tensor is always a tensor (even if it only has one element). With double curly-braces, you can specify a range in which the tensor will be narrowed, which is not possible with single curly-braces. For example, you can do ten[{{1,2},1}], which will be a 1D tensor of dimension 2, and if your do ten[{{1,2},{2}}] it will return a 2D tensor of size 2x1.<\/li>\n<\/ul>\n\n<p>For more information, have a look at <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-selectdim-index\" rel=\"nofollow\">select<\/a> and <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#self-narrowdim-index-size\" rel=\"nofollow\">narrow<\/a>.\nOne last note, the doc for select is not precisely correct, as it's possible to do a select on a 1D tensor, and the output is a number.<\/p>\n"}]}
{"QuestionId":34748720,"AnswerCount":2,"Tags":"<numpy><theano><lasagne><nolearn>","CreationDate":"2016-01-12T16:22:03.177","AcceptedAnswerId":34749845.0,"OwnerUserId":4127806.0,"Title":"Theano TypeError","Body":"<p>I am reading jpg images and then reshaping them into a tensor. I am casting the images as float32: <\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting this error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34749845","CreationDate":"2016-01-12T17:13:59.007","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This is a <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/ZEafOobLFes\" rel=\"nofollow\">cross-post to the theano-users mailing list<\/a>.<\/p>\n\n<p>Doug provided an answer there:<\/p>\n\n<blockquote>\n  <p>The theano variable you are using is defined as integer, but you\n  passed in a float, hence the error 'TensorType(int32, vector) cannot\n  store a value of dtype float32...'.  You can either modify your data\n  loading code to cast it as int32, or change the symbolic variable to\n  something that supports float32.<\/p>\n<\/blockquote>\n\n<p>So somewhere you have a line that looks something like:<\/p>\n\n<pre><code>x = T.ivector()\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>x = T.vector(dtype='int32')\n<\/code><\/pre>\n\n<p>It looks like you need to change this to something like<\/p>\n\n<pre><code>x = T.tensor4()\n<\/code><\/pre>\n\n<p>where the <code>dtype<\/code> has been changed to equal <code>theano.config.floatX<\/code> and the dimensionality as been changed to 4 to match the 4-dimensional nature of <code>X<\/code>.<\/p>\n"},{"AnswerId":"34980493","CreationDate":"2016-01-24T19:36:11.363","ParentId":null,"OwnerUserId":"5817373","Title":null,"Body":"<p>If you didn't figure it out, I had a similar error and here's how I fixed it:\nCast your y as int32. The x values can be floatx, but the y MUST be int32 in nolearn for classification.<\/p>\n"}]}
{"QuestionId":34750268,"AnswerCount":3,"Tags":"<lua><torch>","CreationDate":"2016-01-12T17:36:25.473","AcceptedAnswerId":34762233.0,"OwnerUserId":2452617.0,"Title":"Extracting the top-k value-indices from a 1-D Tensor","Body":"<p>Given a 1-D tensor in Torch (torch.Tensor), containing values which can be compared (say floating point), how can we extract the indices of the top-<em>k<\/em> values in that tensor?<\/p>\n\n<p>Apart from the brute-force method, I am looking for some API call, that Torch\/lua provides, which can perform this task efficiently.<\/p>\n","answers":[{"AnswerId":"34750580","CreationDate":"2016-01-12T17:53:35.340","ParentId":null,"OwnerUserId":"350664","Title":null,"Body":"<p>Just loop through the tensor and run your compare:<\/p>\n\n<pre><code>require 'torch'\n\ndata = torch.Tensor({1,2,3,4,505,6,7,8,9,10,11,12})\nidx  = 1\nmax  = data[1]\n\nfor i=1,data:size()[1] do\n   if data[i]&gt;max then\n      max=data[i]\n      idx=i\n   end\nend\n\nprint(idx,max)\n<\/code><\/pre>\n\n<p>--EDIT--\nResponding to your edit: Use the torch.max operation documented here: <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torchmaxresval-resind-x-dim\" rel=\"nofollow\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torchmaxresval-resind-x-dim<\/a> ...<\/p>\n\n<pre><code>y, i = torch.max(x, 1) returns the largest element in each column (across rows) of x, and a Tensor i of their corresponding indices in x\n<\/code><\/pre>\n"},{"AnswerId":"34762233","CreationDate":"2016-01-13T08:57:29.490","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>As of pull request <a href=\"https:\/\/github.com\/torch\/torch7\/pull\/496\" rel=\"noreferrer\">#496<\/a> Torch now includes a built-in API named <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/03c04c6\/doc\/maths.md#torchtopkresval-resind-x-k-dim-dir-sort\" rel=\"noreferrer\"><code>torch.topk<\/code><\/a>. Example:<\/p>\n\n<pre><code>&gt; t = torch.Tensor{9, 1, 8, 2, 7, 3, 6, 4, 5}\n\n-- obtain the 3 smallest elements\n&gt; res = t:topk(3)\n&gt; print(res)\n 1\n 2\n 3\n[torch.DoubleTensor of size 3]\n\n-- you can also get the indices in addition\n&gt; res, ind = t:topk(3)\n&gt; print(ind)\n 2\n 4\n 6\n[torch.LongTensor of size 3]\n\n-- alternatively you can obtain the k largest elements as follow\n-- (see the API documentation for more details)\n&gt; res = t:topk(3, true)\n&gt; print(res)\n 9\n 8\n 7\n[torch.DoubleTensor of size 3]\n<\/code><\/pre>\n\n<p>At the time of writing the CPU implementation follows a <a href=\"https:\/\/github.com\/wickedfoo\/torch7\/blob\/ef019670474b69629a8b3d50eb426d5858bd5c45\/lib\/TH\/generic\/THTensorMath.c#L1757-L1769\" rel=\"noreferrer\">sort and narrow approach<\/a> (there are plans to improve it in the future). That being said an optimized GPU implementation for cutorch is currently being <a href=\"https:\/\/github.com\/torch\/cutorch\/pull\/296\" rel=\"noreferrer\">reviewed<\/a>.<\/p>\n"},{"AnswerId":"55802422","CreationDate":"2019-04-22T23:38:32.837","ParentId":null,"OwnerUserId":"3650983","Title":null,"Body":"<p>You can use <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.topk\" rel=\"nofollow noreferrer\">topk<\/a> function.<\/p>\n\n<p>for example:<\/p>\n\n<pre><code>import torch\n\nt = torch.tensor([5.7, 1.4, 9.5, 1.6, 6.1, 4.3])\n\nvalues,indices = t.topk(2)\n\nprint(values)\nprint(indices)\n<\/code><\/pre>\n\n<p>the result:<\/p>\n\n<pre><code>tensor([9.5000, 6.1000])\ntensor([2, 4])\n<\/code><\/pre>\n"}]}
{"QuestionId":34750327,"AnswerCount":0,"Tags":"<3d><theano><convolution>","CreationDate":"2016-01-12T17:39:25.353","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"How to use conv3D in theano?","Body":"<p>I need an example each for theano.tensor.nnet.Conv3D &amp; theano.tensor.nnet.conv3d2d.conv3d<\/p>\n\n<p>my input data is single channel, 3D.<\/p>\n","answers":[]}
{"QuestionId":34750543,"AnswerCount":4,"Tags":"<python><python-2.7><numpy><tensorflow>","CreationDate":"2016-01-12T17:51:34.373","AcceptedAnswerId":null,"OwnerUserId":4862365.0,"Title":"Tensorflow is not installed properly","Body":"<p>I am a beginner and I tried installing tensor flow on Macbook according to the instructions given on the site.<\/p>\n\n<pre><\/pre>\n\n<p>After installation , when I am trying to import tensor flow , I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am new to python and I have tried a few solutions on internet but none of them have worked so far. Perhaps the error is due to the pip installation. I am stuck on this since 2 days. Thanks for help!<\/p>\n","answers":[{"AnswerId":"38225761","CreationDate":"2016-07-06T13:51:48.953","ParentId":null,"OwnerUserId":"2340065","Title":null,"Body":"<p>I think I came across the same problem. This was caused by an earlier version of factory-included numpy that on the factory-installed Python on Max OSX. <\/p>\n\n<p>I tried <code>sudo pip uninstall numpy<\/code> and the factor-included numpy still exists. To solve this problem, I upgraded the factory-included numpy using <code>easy_install<\/code>, then used <code>pip<\/code> to install <code>tensorflow<\/code> and it worked:<\/p>\n\n<pre><code>sudo easy_install -U numpy\nsudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py2-none-any.whl\n<\/code><\/pre>\n"},{"AnswerId":"35091443","CreationDate":"2016-01-29T18:23:41.370","ParentId":null,"OwnerUserId":"4862365","Title":null,"Body":"<p>So Finally what I did was removed tensorflow ,pip and everything. The problem was that I was using Mac OSX python due to which I was having the problem. I used homebrew to install python, then created virtualenv and used pip to install tensorflow again. After that it was working.<\/p>\n"},{"AnswerId":"39784278","CreationDate":"2016-09-30T05:28:23.237","ParentId":null,"OwnerUserId":"6902958","Title":null,"Body":"<p>To add onto the previous answer - on my <code>OSX<\/code> machine (10.11), I found two factory installations of <code>numpy<\/code> - one in <code>\/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python\/numpy<\/code> and the other in <code>\/Library\/Python\/2.7\/site-packages\/numpy\/<\/code>.  As far as I can tell, pip will upgrade the one in \/Library, but python kept trying to use the one in \/System\/Library\/ first.  <\/p>\n\n<p>As a temporary solution, I just renamed the <code>numpy<\/code> folder in <code>\/System\/Library<\/code> to <code>numpy.old<\/code>.  <code>Tensorflow<\/code> imported correctly after that.<\/p>\n"},{"AnswerId":"44190987","CreationDate":"2017-05-25T22:52:27.403","ParentId":null,"OwnerUserId":"5009287","Title":null,"Body":"<h1>Summary<\/h1>\n\n<p>This resolved my issue:<\/p>\n\n<pre><code>$ sudo rm -rf \/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python\/numpy\n<\/code><\/pre>\n\n<h1>Details<\/h1>\n\n<p><strong>Environment and symptoms<\/strong><\/p>\n\n<p>I have a 3 pythons installed on my Mac OS X Yosemite:<\/p>\n\n<ul>\n<li>2.6 which came with OS<\/li>\n<li>2.7 which I have installed <\/li>\n<li>3.4 which I have installed <\/li>\n<\/ul>\n\n<p>Of course I did a poor job at setting things properly and the python binaries and libs are all over the place on the mac.<\/p>\n\n<p>I got this error when attempting to work with <a href=\"https:\/\/github.com\/Ning-Ding\/Implementation-CVPR2015-CNN-for-ReID\/tree\/master\/CUHK03\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Ning-Ding\/Implementation-CVPR2015-CNN-for-ReID\/tree\/master\/CUHK03<\/a> <\/p>\n\n<pre><code>$ python .\/main.py --dataset cuhk-03.h5\n\nUsing TensorFlow backend.\nRuntimeError: module compiled against API version 0xa but this version of numpy is 0x9\nTraceback (most recent call last):\n  File \".\/main.py\", line 14, in &lt;module&gt;\n    from model import generate_model,compile_model\n  File \"\/Users\/rv186016\/repositories\/Implementation-CVPR2015-CNN-for-ReID\/CUHK03\/model.py\", line 12, in &lt;module&gt;\n    from keras.layers import Input\n  File \"\/Library\/Python\/2.7\/site-packages\/keras\/__init__.py\", line 3, in &lt;module&gt;\n    from . import activations\n  File \"\/Library\/Python\/2.7\/site-packages\/keras\/activations.py\", line 4, in &lt;module&gt;\n    from . import backend as K\n  File \"\/Library\/Python\/2.7\/site-packages\/keras\/backend\/__init__.py\", line 73, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"\/Library\/Python\/2.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n  File \"\/Library\/Python\/2.7\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/__init__.py\", line 51, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 52, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 41, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/Library\/Python\/2.7\/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: numpy.core.multiarray failed to import\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>Solution<\/strong><\/p>\n\n<p>After looking at <strong><em>se7en007<\/em><\/strong>'s comment on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/559\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/559<\/a>, I ran these commands which loved my issue. See the diff in the <code>numpy<\/code> version and path before and after the <code>rm<\/code> command.<\/p>\n\n<pre><code>$ python\nPython 2.7.10 (default, Jul 14 2015, 19:46:27)\n[GCC 4.2.1 Compatible Apple LLVM 6.0 (clang-600.0.39)] on darwin\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import numpy\n&gt;&gt;&gt; numpy.__version__\n'1.8.0rc1'\n&gt;&gt;&gt; numpy.__path__\n['\/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python\/numpy']\n&gt;&gt;&gt; exit()\n\n$ sudo rm -rf \/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python\/numpy\nMUSRV186016-382:CUHK03 rv186016$ python\nPython 2.7.10 (default, Jul 14 2015, 19:46:27)\n[GCC 4.2.1 Compatible Apple LLVM 6.0 (clang-600.0.39)] on darwin\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import numpy\n&gt;&gt;&gt; numpy.__path__\n['\/Library\/Python\/2.7\/site-packages\/numpy']\n&gt;&gt;&gt; numpy.__version__\n'1.12.0'\n&gt;&gt;&gt; ^D\n\n$ python .\/main.py --dataset cuhk-03.h5\nUsing TensorFlow backend.\n____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to\n====================================================================================================\ninput_1 (InputLayer)             (None, 160, 60, 3)    0\n____________________________________________________________________________________________________\ninput_2 (InputLayer)             (None, 160, 60, 3)    0\n____________________________________________________________________________________________________\nconv2d_1 (Conv2D)                (None, 156, 56, 20)   1520        input_1[0][0]\n                                                                   input_2[0][0]\n...\n<\/code><\/pre>\n"}]}
{"QuestionId":34750832,"AnswerCount":1,"Tags":"<matlab><input><hdf5><caffe><conv-neural-network>","CreationDate":"2016-01-12T18:06:07.547","AcceptedAnswerId":40595239.0,"OwnerUserId":3698878.0,"Title":"Caffe - Create deploy.prototxt from train_val.prototxt","Body":"<p>I have finetuned the imagenet pretrained model on my dataset, following are the relevant changes made in train_val.prototxt (Also I have not done oversampling, just the centre crop while preparing hdf5)<\/p>\n\n<pre><\/pre>\n\n<p>The rest is same till <\/p>\n\n<pre><\/pre>\n\n<p>The question is I don't get how to modify my deploy.prototxt and my feature extraction code (which is given below) to test this model, now with the extra meta info as input features<\/p>\n\n<pre><\/pre>\n\n<p>Here prepare_image does rgb to bgr, permute and centre crop stuff, basically preprocessing.<\/p>\n\n<p>So, it boils down to how should I modify 'caffe('forward', input_data)'; and make a deploy.prototxt so that while testing caffe is also fed the Meta (n*2) features. Thanks for your patience and help!<\/p>\n\n<p>I think deploy.prototxt beginning should be something like<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40595239","CreationDate":"2016-11-14T18:17:11.543","ParentId":null,"OwnerUserId":"3698878","Title":null,"Body":"<p>The variable input_data can be a vector of 4-D blobs. The function prepare_data gives a 4D blob of image data similarly one can reshape adequately meta or any other kind of input to a 4D blob and then pass it to the forward function.<\/p>\n\n<p><code>input_data = {prepare_data(im),prepare_meta()};<\/code><\/p>\n\n<p><code>scores = caffe('forward', input_data')<\/code>;<\/p>\n"}]}
{"QuestionId":34752142,"AnswerCount":1,"Tags":"<typeerror><theano>","CreationDate":"2016-01-12T19:25:02.810","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"Type-error in shared variable in theano","Body":"<p>I get TypeError: an integer is required on line 5:<\/p>\n\n<pre><\/pre>\n\n<p>However if i change b_shape to b_shp=(2,0), it gives no error.<\/p>\n","answers":[{"AnswerId":"34754055","CreationDate":"2016-01-12T21:25:03.397","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>My guess is that <code>W<\/code> is a symbolic tensor (e.g. a shared variable). For example,<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\nrng1 = np.random.RandomState(1234)\ndtensor = T.TensorType(dtype=theano.config.floatX, broadcastable=(False,) * 5)\nV = dtensor('V')\nW = theano.shared(np.asarray(rng1.uniform(low=-5, high=5, size=(2, 3)), dtype=V.dtype), name='W', borrow=True)\nb_shp = (W.shape[0],)\nb = theano.shared(np.asarray(rng1.uniform(low=-5, high=5, size=b_shp), dtype=V.dtype), name='b', borrow=True)\n<\/code><\/pre>\n\n<p>If this assumption is correct, the error <code>TypeError: an integer is required<\/code> occurs because <code>W.shape[0]<\/code> is a symbolic expression but numpy needs a concrete integer. You could probably change it to <code>W.get_value().shape[0]<\/code> to solve the problem though this is messy and a cleaner solution would depend on what you're trying to achieve.<\/p>\n"}]}
{"QuestionId":34752969,"AnswerCount":0,"Tags":"<deep-learning><tensorflow>","CreationDate":"2016-01-12T20:15:42.107","AcceptedAnswerId":null,"OwnerUserId":669363.0,"Title":"Simple Multilayer Perceptron model does not converge in TensorFlow","Body":"<p>I am new to TensorFlow. Today I tried to implement my first model in TF but it returned strange results. I know that I am missing something here but I was  not able to figure it out. Here is the story.<\/p>\n\n<h2>Model<\/h2>\n\n<p>I have a simple Multilayer Perceptron model with only a single hidden layer applied on MNIST databse. Layers are defined like [input(784) , hidden_layer(470) , output_layer(10)] with  as non-linearity for hidden layer and  as the loss for output layer. The optimizer I am using is <strong>Gradient Descent<\/strong> Algorithm with learning rate of . My mini batch size is 1 (I am training model with samples one by one). <\/p>\n\n<p>My implementations :<\/p>\n\n<ol>\n<li>First I implemented my model in C++ and got around 96% accuracy.Here is the repository : <a href=\"https:\/\/github.com\/amin2ros\/Artificog\" rel=\"nofollow\">https:\/\/github.com\/amin2ros\/Artificog<\/a><\/li>\n<li>I implemented the exact model in TensorFlow but surprisingly the model didn't converge at all. Here is the code.<\/li>\n<\/ol>\n\n<h2>Code:<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I am very curious why this happens. Any help is appreciated. <\/p>\n\n<h2>Edit:<\/h2>\n\n<p>As commented below definition of cost function was incorrect so it should be like <\/p>\n\n<pre><\/pre>\n\n<p>Now model converges :)<\/p>\n","answers":[]}
{"QuestionId":34754598,"AnswerCount":1,"Tags":"<python><machine-learning><caffe><pycaffe>","CreationDate":"2016-01-12T22:02:07.340","AcceptedAnswerId":34754730.0,"OwnerUserId":2527680.0,"Title":"Decrease the level of detail in caffe training output?","Body":"<p>I have compiled caffe with Debug flag on.\nNow when I run <\/p>\n\n<pre><\/pre>\n\n<p>I get output <\/p>\n\n<pre><\/pre>\n\n<p><a href=\"http:\/\/pastebin.com\/cbXTH5HH\" rel=\"nofollow\">http:\/\/pastebin.com\/cbXTH5HH<\/a><\/p>\n\n<p>I Would like output without the read time and prefetch time. Without recompiling.<\/p>\n\n<pre><\/pre>\n\n<p>For eg:- <a href=\"http:\/\/pastebin.com\/F5c3Yutu\" rel=\"nofollow\">http:\/\/pastebin.com\/F5c3Yutu<\/a><\/p>\n","answers":[{"AnswerId":"34754730","CreationDate":"2016-01-12T22:12:55.343","ParentId":null,"OwnerUserId":"33258","Title":null,"Body":"<p>What about<\/p>\n\n<pre><code>.\/examples\/mnist\/train_lenet.sh | grep -v \"Read time:\" | grep -v \"Prefetch batch:\"\n<\/code><\/pre>\n\n<p>which gives me<\/p>\n\n<pre><code>I0112 22:50:49.671303 114020 data_layer.cpp:104] Transform time: 0.791 ms.\nI0112 22:50:49.672757 114020 data_layer.cpp:104] Transform time: 0.767 ms.\nI0112 22:50:49.674334 114020 data_layer.cpp:104] Transform time: 0.836 ms.\nI0112 22:50:49.675853 114020 data_layer.cpp:104] Transform time: 0.806 ms.\nI0112 22:50:49.677273 114020 data_layer.cpp:104] Transform time: 0.762 ms.\nI0112 22:50:49.678861 114020 data_layer.cpp:104] Transform time: 0.861 ms.\nI0112 22:50:49.680376 114020 data_layer.cpp:104] Transform time: 0.821 ms.\nI0112 22:50:49.681077 113921 solver.cpp:409]     Test net output #0: accuracy = 0.9902\nI0112 22:50:49.681115 113921 solver.cpp:409]     Test net output #1: loss = 0.0292544 (* 1 = 0.0292544 loss)\nI0112 22:50:49.681125 113921 solver.cpp:326] Optimization Done.\nI0112 22:50:49.681133 113921 caffe.cpp:215] Optimization Done.\nI0112 22:50:49.681948 114020 data_layer.cpp:104] Transform time: 0.829 ms.\n<\/code><\/pre>\n\n<p>from your sample input.<\/p>\n"}]}
{"QuestionId":34756277,"AnswerCount":0,"Tags":"<machine-learning><computer-vision><neural-network><converter><caffe>","CreationDate":"2016-01-13T00:18:27.487","AcceptedAnswerId":null,"OwnerUserId":825785.0,"Title":"Converting a Cuda-Convnet checkpoint to a Caffe model binary protobuf","Body":"<p>I have an old checkpoint (model specification) from a Cuda Convnet model trained by someone else a couple years ago where the training data is no longer available.  I would like to find a way to convert this exact model to a Caffe model file.  Is there a tool (which is currently available and supported) which does this?  I would still interested even if the conversion is only to another ML framework which can export Caffe models (i.e. using Theano, Torch7, etc.) as a bridge.<\/p>\n","answers":[]}
{"QuestionId":34756903,"AnswerCount":1,"Tags":"<tensorflow><gradient>","CreationDate":"2016-01-13T01:27:40.860","AcceptedAnswerId":34758575.0,"OwnerUserId":5587428.0,"Title":"tf.Print() causing an error with gradients","Body":"<p>I was trying to use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#Print\" rel=\"nofollow\">tf.Print<\/a> debug statements to better understand the format of the reported gradients and variables from compute_gradients() but ran into an unexpected problem. The training routine and the debug routine (gvdebug) are as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This code works fine (but doesn't print), but if I uncomment the two tf.Print lines in gvdebug() I get an error message from apply_gradients: 'TypeError: Variable must be a tf.Variable'. I thought tf.Print just passed through tensors -- what am I doing wrong?<\/p>\n","answers":[{"AnswerId":"34758575","CreationDate":"2016-01-13T04:45:53.937","ParentId":null,"OwnerUserId":"1589147","Title":null,"Body":"<p><strong>TL;DR<\/strong><\/p>\n\n<p>Don't try to <code>tf.Print<\/code> <code>gv[1]<\/code> because it is a <code>tf.Variable<\/code>. It is like a pointer back to the variable which created the <code>gradient<\/code> in <code>gv[0]<\/code>.<\/p>\n\n<p><strong>Further Information<\/strong><\/p>\n\n<p>When you run <code>compute_gradients<\/code> it returns a list of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/411f57e291839094108afdaa9c43094f44979eaa\/tensorflow\/python\/ops\/gradients.py#L322\" rel=\"nofollow\"><code>gradients<\/code><\/a> and their corresponding <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/411f57e291839094108afdaa9c43094f44979eaa\/tensorflow\/python\/training\/optimizer.py#L211\" rel=\"nofollow\"><code>tf.Variable<\/code>s<\/a>.<\/p>\n\n<p>Each element of <code>grads_and_vars<\/code> is a <code>Tensor<\/code> and a <code>tf.Variable<\/code>. It is important to note that it is <strong>not<\/strong> the value of the variable.<\/p>\n\n<p>Your code works for me after removing <code>v = tf.Print(v,[v],'V: ')<\/code><\/p>\n"}]}
{"QuestionId":34758978,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-01-13T05:24:30.540","AcceptedAnswerId":34761247.0,"OwnerUserId":3487835.0,"Title":"Is there any way to cache the theano compiling result?","Body":"<p>I ask this question because the compiling process of theano is quite time consuming, and what making it worse is that each time I run the code, the function will be recompiled. So, is there any way to make theano cache the previous compiling result if there is no code change?<\/p>\n","answers":[{"AnswerId":"34761247","CreationDate":"2016-01-13T08:01:43.170","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Theano already caches compilation results. The cache is stored in <code>~\/.theano<\/code> by default but the location and format of the contents can be adjusted via the config flags <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.base_compiledir\" rel=\"nofollow\"><code>config.base_compiledir<\/code><\/a>, <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.compiledir_format\" rel=\"nofollow\"><code>config.compiledir_format<\/code><\/a>, and <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.compiledir\" rel=\"nofollow\"><code>config.compiledir<\/code><\/a>.<\/p>\n\n<p>If you create a very large computation graph then the first run should be substantially slower than subsequent runs.<\/p>\n"}]}
{"QuestionId":34759227,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-13T05:46:34.933","AcceptedAnswerId":34763642.0,"OwnerUserId":5703903.0,"Title":"TensorFlow CIFAR10 Example","Body":"<p>I am trying to run the entire CIFAR10 as is, with data from SVHN.<\/p>\n\n<p><a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"nofollow noreferrer\">http:\/\/ufldl.stanford.edu\/housenumbers\/<\/a><\/p>\n\n<p>I formatted the data in the exact format as the bin file from Alex Krizhevsky's website.<\/p>\n\n<p><a href=\"http:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">http:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a><\/p>\n\n<p>I did not edit the code, other than changing a few variable names to make it work in another directory. It gives me an error now.<\/p>\n\n<pre><\/pre>\n\n<p>Specifically, the line that fails in cifar.py is:<\/p>\n\n<pre><\/pre>\n\n<p>I have checked this solution too, it does not work.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/33933846\/tensorflow-indices-are-not-valid-out-of-bounds\">TensorFlow Indices are not valid (out of bounds)<\/a><\/p>\n\n<p>Anyone has any idea on how to make it work?<\/p>\n","answers":[{"AnswerId":"34763642","CreationDate":"2016-01-13T10:03:34.687","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>I realized the mistake. The SVHN dataset gave the number 0 a value of 10, instead of 0. I made this fatal assumption from the start and it wasted a lot of my time.<\/p>\n\n<p>Given 10 classes, the labels should range from 0-9, inclusive. The error happened because the labels ranged from 1-10.<\/p>\n\n<p><a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"noreferrer\">http:\/\/ufldl.stanford.edu\/housenumbers\/<\/a><\/p>\n\n<p>Do remember to read overviews in future!<\/p>\n"}]}
{"QuestionId":34759914,"AnswerCount":1,"Tags":"<string><split><lua><torch>","CreationDate":"2016-01-13T06:35:54.657","AcceptedAnswerId":34764488.0,"OwnerUserId":4013781.0,"Title":"Splitting String Characters into a Torch Tensor","Body":"<p>I have 2 Strings in Lua: <\/p>\n\n<pre><\/pre>\n\n<p>I want to split them into characters and load them into a Torch Tensor, like this: <\/p>\n\n<pre><\/pre>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"34764488","CreationDate":"2016-01-13T10:40:55.377","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<pre><code>function split(str)\n    tbl = {}\n    for c in str:gmatch('.') do\n         table.insert(tbl, c)\n    end\n    return tbl\nend\n\na = '01234'\nb = '12345'\n\ntens = torch.Tensor{split(a), split(b)} \n<\/code><\/pre>\n"}]}
{"QuestionId":34760042,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-01-13T06:45:03.443","AcceptedAnswerId":34760276.0,"OwnerUserId":5583416.0,"Title":"when do use borrow=True for theano shared variables?","Body":"<p>In theano when do we use borrow =true for shared variables?<\/p>\n\n<p>In theano deep learning tutorial, I found that in some examples borrow=True is used while in others it isn't. In the logistic regression and the MLP tutorial, borrow=True is used, while in the Convolutional neural network tutorial borrow=True is not used.<\/p>\n","answers":[{"AnswerId":"34760276","CreationDate":"2016-01-13T07:01:12.007","ParentId":null,"OwnerUserId":"5583416","Title":null,"Body":"<p>Found answer at theano google group <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/XWDOoKdpn5I\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/XWDOoKdpn5I<\/a><\/p>\n"}]}
{"QuestionId":34760677,"AnswerCount":1,"Tags":"<theano><keras>","CreationDate":"2016-01-13T07:25:05.057","AcceptedAnswerId":34762483.0,"OwnerUserId":2687955.0,"Title":"Does convolution in Theano rotate the filters?","Body":"<p>I have an 3-channel 5-by-5 image like this:<\/p>\n\n<pre><\/pre>\n\n<p>And a 3-channel 3-by-3 filter like this:<\/p>\n\n<pre><\/pre>\n\n<p>When convolve the image with the filter, I am expecting this output:<\/p>\n\n<pre><\/pre>\n\n<p>However, Theano (using keras) gives me this output:<\/p>\n\n<pre><\/pre>\n\n<p>It seems the output is rotated 180 degrees, I wonder why this happens and how can I get the correct answer. Here is my test code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34762483","CreationDate":"2016-01-13T09:10:30.353","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>This is the definition of convolution. It has the advantage that if you convolve an image that consists of only zeros except for one single 1 somewhere, the convolution will place a copy of the filter at that position.<\/p>\n\n<p>Theano does exactly these convolutions, as defined mathematically. This implies flipping the filters (the operation is <code>filter[:, :, ::-1, ::-1]<\/code>) before taking dot products with the image patches. Note that these are not rotations by 180 degrees, at least not in general.<\/p>\n\n<p>It appears that what you are looking for is cross-correlation, which is taking dot products with the non-flipped versions of the filters at each point of the image.<\/p>\n\n<p>See also <a href=\"https:\/\/stackoverflow.com\/questions\/25619458\/perform-the-exact-same-convolution-as-in-theanos-conv2d\/26234696#26234696\">this answer<\/a> in which <code>theano.tensor.nnet.conv2d<\/code> is shown to do exactly the same thing as the <code>scipy<\/code> counterpart.<\/p>\n"}]}
{"QuestionId":34760765,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-01-13T07:29:55.663","AcceptedAnswerId":null,"OwnerUserId":5118777.0,"Title":"Torch7 - not enough memory: you tried to allocate 0GB. Buy new RAM","Body":"<p>I tried to load a large audio dataset and implement audio.spectrogram.<\/p>\n\n<p>I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>Does torch7 have the memory limit ?<\/p>\n","answers":[{"AnswerId":"50598753","CreationDate":"2018-05-30T07:26:54.240","ParentId":null,"OwnerUserId":"1944004","Title":null,"Body":"<p><strong>No<\/strong>, Torch does not have a memory limit but it requires that certain conditions are met when allocating memory.<\/p>\n\n<p>If you take a look at <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/3864f3782a74715e35d98176f9014a74524a868a\/lib\/TH\/THGeneral.c#L226-L273\" rel=\"nofollow noreferrer\"><code>THGeneral.c<\/code><\/a> (where the error comes from), you'll see that this error is raised when the allocation using <code>THAllocInternal<\/code> has failed.  From your output I guess you are on a Unix system and I also guess that you are trying to allocate a lot of memory (but less than 1GB).  In the case where you want to align more than 5120 bytes, <code>THAllocInternal<\/code> will call <a href=\"http:\/\/man7.org\/linux\/man-pages\/man3\/posix_memalign.3.html\" rel=\"nofollow noreferrer\"><code>posix_memalign<\/code><\/a> for 64 byte aligned memory otherwise it will call standard <a href=\"http:\/\/en.cppreference.com\/w\/c\/memory\/malloc\" rel=\"nofollow noreferrer\"><code>malloc<\/code><\/a>.  That is to say that the error you see comes from either of those functions which are provided by your operating system.  You will have to check there.  You could also try to recompile Torch with the flag <code>DISABLE_POSIX_MEMALIGN<\/code> to rule that out.<\/p>\n"}]}
{"QuestionId":34760981,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-13T07:42:53.253","AcceptedAnswerId":null,"OwnerUserId":5782844.0,"Title":"RBM implementation with tensorflow","Body":"<p>I'm trying to implement RBM with tensorflow, here is the code:<\/p>\n\n<p><strong>rbm.py<\/strong><\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>rbm_MNIST_test.py<\/strong><\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>but here comes the error:<\/p>\n\n<blockquote>\n  <p>File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 1626, in as_graph_def\n      raise ValueError(\"GraphDef cannot be larger than 2GB.\") ValueError: GraphDef cannot be larger than 2GB.<\/p>\n<\/blockquote>\n\n<p>Can someone help me to solve this?<\/p>\n","answers":[{"AnswerId":"34772776","CreationDate":"2016-01-13T17:06:57.460","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow does have a limit of 2GB on the <code>GraphDef<\/code> protos, which stems from a limitation of the protocol buffers implementation. You can quickly reach that limit if you have large constant tensors in your graph. In particular, if you use the <strong>same<\/strong> numpy array multiple times, TensorFlow will add multiple constant tensors to your graph.<\/p>\n\n<p>In your case, <code>mnist.train.images<\/code> returned by <code>input_data.read_data_sets<\/code> is a numpy floating point array with shape <code>(55000, 784)<\/code>, so it is about <code>164 MB<\/code>. You pass that numpy array to <code>rbm_layer.cd1<\/code>, and inside that function, every time you use <code>visibles<\/code>, a TensorFlow <code>Const<\/code> node gets created from the numpy array. You use <code>visibiles<\/code> in 3 locations, so every call to <code>cd1<\/code> is increasing the graph size by approximately <code>492 MB<\/code>, so you easily exceed the limit. The solution is to create a TensorFlow constant once and pass that constant to the <code>cd1<\/code> function like so :<\/p>\n\n<pre><code>trX_constant = tf.constant(trX)\nfor i in range(10):\n    print \"RBM CD: \", i\n    rbm_layer.cd1(trX_constant)\n<\/code><\/pre>\n\n<p>BTW, I am not sure what your intention is in the above loop. Note that the <code>cd1<\/code> function simply adds the <code>assign_add<\/code> nodes to the graph, and does NOT actually perform the assigns. If you really want those assigns to happen while you train, you should consider chaining those assigns via control dependencies to your final <code>train_op<\/code> node.<\/p>\n"},{"AnswerId":"35446666","CreationDate":"2016-02-17T02:10:59.963","ParentId":null,"OwnerUserId":"1749667","Title":null,"Body":"<p>To fulfill @keveman's question, I think you are trying to implement the <code>CD-k<\/code> (Contrastive Divergence) step by using that loop.<\/p>\n\n<p>But I'm afraid the code is so far from appropriate since the <code>CD-k<\/code> is the function which should take the position of <em>Automatic Differentiation<\/em> in <code>RBM<\/code>. That means the <code>cost<\/code> and <code>train_op<\/code> is not the correct way to use with <em>Gradient Descent<\/em> in <code>RBM<\/code> (which is because of the special role of CD-k). By the way, the <em>RBM<\/em> layer should be trained one by one without <em>fully-connected layer<\/em> which is not in your code.<\/p>\n\n<p>I'm new to tensorflow, and I also want to get the implementation. I think I'd rather not use <em>Gradient Descent<\/em> provided by tensorflow since I need <code>CD-k<\/code> for special differentiation. Wish I could find the solution soon.<\/p>\n\n<p><strong>Update:<\/strong>\nI've worked on this implementation for a whole work day. So, here's the current status. I've implement a simple straightforward version but It just get wrong result.\nPlease refer to <a href=\"https:\/\/gist.github.com\/myme5261314\/005ceac0483fc5a581cc\" rel=\"nofollow\">code and result<\/a><\/p>\n\n<p>I just refer the specific approach from <a href=\"https:\/\/github.com\/rasmusbergpalm\/DeepLearnToolbox\" rel=\"nofollow\">DeepLearnToolbox<\/a>. I think the procedure that I'm trying to implement through <code>tensorflow<\/code> is ok, but don't know what's going wrong about the practical code.<\/p>\n\n<p><strong>Update 2:<\/strong> I've revised the code and now I've implemented the simplest rbm through tensorflow. see <code>code and result<\/code> link above.<\/p>\n"}]}
{"QuestionId":34762463,"AnswerCount":1,"Tags":"<tensorflow><tensorboard><bazel>","CreationDate":"2016-01-13T09:09:35.470","AcceptedAnswerId":34772167.0,"OwnerUserId":4844184.0,"Title":"Install Tensorflow pip wheel without internet","Body":"<p>I do not have internet access on my linux computer therefore I installed TF from source by following <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">TensorFlow Get Started<\/a>.<br>\nI ran into a few trouble to build trainer_example due to the lack of internet connection hopefully someone from tensorflow helped me through it by creating local repositories for re2, gemmlowp, jpegsrc v9a, libpng and six and modifying WORKSPACE accordingly.<br>\nWhen I try to bazel build pip_package to create the wheel then I think I run into the same problem but :  <\/p>\n\n<p>-the list of repositories is insanely long (to manually install each of them) even if they seem to be mostly part of PolymerElements<\/p>\n\n<p>Is there an easy workaround ?<\/p>\n","answers":[{"AnswerId":"34772167","CreationDate":"2016-01-13T16:39:05.320","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you are happy to create a PIP package <strong>without TensorBoard<\/strong>, you should be able to avoid rewriting the Polymer dependencies by removing <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/03b5bef1ddbb5841aaf059a8b77267becbe8ea21\/tensorflow\/tools\/pip_package\/BUILD#L29\" rel=\"nofollow\">this line<\/a> (<code>\"\/\/tensorflow\/tensorbaord\"<\/code> in the <code>build_pip_package<\/code> dependencies) from <code>tensorflow\/tools\/pip_package\/BUILD<\/code>.<\/p>\n"}]}
{"QuestionId":34762505,"AnswerCount":1,"Tags":"<deep-learning><tensorflow><conv-neural-network>","CreationDate":"2016-01-13T09:11:42.663","AcceptedAnswerId":null,"OwnerUserId":5758500.0,"Title":"Inception style convolution","Body":"<p>I have the necessity to keep the model as small as possible to deploy an image classifier that can run efficiently on an app (the accuracy is not really relevant for me)<\/p>\n\n<p>I recently approached deep learning and I haven't great experience, hence I'm currently playing with the cifar-10 example.\nI tried to replace the first two 5x5 convolutional layers with two 3x3 convolution each, as described in the <a href=\"http:\/\/arxiv.org\/abs\/1512.00567\" rel=\"nofollow\">inception paper<\/a>.<\/p>\n\n<p>Unluckily, when I'm going to classify the test set, I got around 0.1 correct classification (random choice)<\/p>\n\n<p>This is the modified code of the first layer (the second is similar):<\/p>\n\n<pre><\/pre>\n\n<p>Is it correct?<\/p>\n","answers":[{"AnswerId":"34772346","CreationDate":"2016-01-13T16:47:15.367","ParentId":null,"OwnerUserId":"3167182","Title":null,"Body":"<p>It seems you're attempting to compute 64 features (for each 3x3 patch) in the first convolutional layer and feed this directly into the second convolutional layer, with no intermediate pooling layer. Convolutional neural networks typically have a structure of stacked convolutional layers, followed by contrast normalization and max pooling.<\/p>\n\n<p>To reduce processing overheads researchers have experimented in moving from fully connected to <a href=\"http:\/\/arxiv.org\/pdf\/1409.4842v1.pdf\" rel=\"nofollow\">sparsely connected architectures<\/a>, and hence the creation of inception architecture. However, whilst these yield good results for high dimensional inputs, you may be expecting too much from the 32x32 pixels of Cifar10 in TensorFlow.<\/p>\n\n<p>Therefore, I think the issue is less around patch size of and more to do with overall architecture. This <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow\">code<\/a> is a known good starting point. Get this working and start reducing parameters until it breaks.<\/p>\n"}]}
{"QuestionId":34763915,"AnswerCount":0,"Tags":"<python-2.7><machine-learning><caffe>","CreationDate":"2016-01-13T10:14:36.153","AcceptedAnswerId":null,"OwnerUserId":4614885.0,"Title":"Pycaffe Installation in windows","Body":"<p>I had to work around with caffe in python so i tried to Build caffe for windows using opencv and vs2013\nand for cpu only mode the process was sucessfull and build completed with few warnings and no Errors\nafter that i copied the build file into lib\/site packages in my anaconda package so that i will be able to use it \nbut after that i tried to <strong>import caffe<\/strong>  it shows error<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for Help in Advance \nCan anyone get me through this error and import caffe successfully<\/p>\n","answers":[]}
{"QuestionId":34769494,"AnswerCount":1,"Tags":"<machine-learning><ubuntu-14.04><caffe><multi-gpu><pycaffe>","CreationDate":"2016-01-13T14:42:05.190","AcceptedAnswerId":34770034.0,"OwnerUserId":2527680.0,"Title":"Caffe crashes after first iteration?","Body":"<p>Caffe crashes when I train a net.  <\/p>\n\n<p>The <a href=\"http:\/\/pastebin.com\/ZbstXD8J\" rel=\"nofollow\">solver<\/a> and <a href=\"http:\/\/pastebin.com\/Ny2bUYHz\" rel=\"nofollow\">model<\/a>.<\/p>\n\n<p>In this case I have only used GPU 0. Here is the error trace:<\/p>\n\n<pre><\/pre>\n\n<p>The full output of the train is <a href=\"http:\/\/pastebin.com\/rkYQA7Df\" rel=\"nofollow\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"34770034","CreationDate":"2016-01-13T15:04:41.593","ParentId":null,"OwnerUserId":"1391220","Title":null,"Body":"<p>Your solver file has this line<\/p>\n\n<pre><code>lr_policy: \"fixed\"\n<\/code><\/pre>\n\n<p>But the Caffe output has this line<\/p>\n\n<pre><code>lr_policy: \"step\"\n<\/code><\/pre>\n\n<p>If you use stepwise learning, you must define <code>stepsize<\/code>. Again, your solver file indicates that you have defined stepsize, but Caffe output shows no stepsize. Please check your solver file again, and add this line back<\/p>\n\n<pre><code>stepsize: 10000\n<\/code><\/pre>\n"}]}
{"QuestionId":34770468,"AnswerCount":2,"Tags":"<python><install><tensorflow>","CreationDate":"2016-01-13T15:24:08.520","AcceptedAnswerId":34773809.0,"OwnerUserId":2490541.0,"Title":"Running tensorflow on GPU cluster in virtualenv","Body":"<p>I installed the GPU version of tensorflow in a virtualenv following these <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv-installation\" rel=\"nofollow\">instructions<\/a>. The problem is, I am getting a segmentation fault upon starting a session. That is, this code:<\/p>\n\n<pre><\/pre>\n\n<p>exits with the following error: <\/p>\n\n<pre><\/pre>\n\n<p>I tried to dig deeper using gdb but only got the following additional outputs:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas what is happening here and how to fix it?<\/p>\n\n<p>Here is the output of nvidia-smi:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for any help on this issue! <\/p>\n","answers":[{"AnswerId":"34773809","CreationDate":"2016-01-13T18:00:00.293","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>It's not finding CuDNN -<\/p>\n\n<blockquote>\n  <p>I tensorflow\/stream_executor\/dso_loader.cc:93] Couldn't open CUDA library > libcudnn.so.6.5. LD_LIBRARY_PATH: :\/vol\/cuda\/7.0.28\/lib64\n  I tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:1382] Unable to load cuDNN DSO<\/p>\n<\/blockquote>\n\n<p>You need to have it installed.  Please see <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#optional-install-cuda-gpus-on-linux\" rel=\"nofollow\">the TensorFlow CUDA installation instructions<\/a><\/p>\n"},{"AnswerId":"40958155","CreationDate":"2016-12-04T11:34:02.720","ParentId":null,"OwnerUserId":"5046896","Title":null,"Body":"<p>After untar the cudnn<\/p>\n\n<pre><code>[root@localhost cudnn]# cd include\/\n[root@localhost include]# mv cudnn.h \/usr\/local\/cuda\/include\/\n[root@localhost include]# cd ..\/lib64\/\n[root@localhost lib64]# mv * \/usr\/local\/cuda\/lib\n<\/code><\/pre>\n\n<p>And it is ok<\/p>\n\n<pre><code>[root@localhost ~]# python\nPython 2.7.5 (default, Sep 15 2016, 22:37:39) \n[GCC 4.8.5 20150623 (Red Hat 4.8.5-4)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow as f\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcublas.so.8.0 locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcudnn.so.5 locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcufft.so.8.0 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.8.0 locally\n&gt;&gt;&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":34775522,"AnswerCount":1,"Tags":"<gpu><tensorflow>","CreationDate":"2016-01-13T19:36:19.350","AcceptedAnswerId":34776814.0,"OwnerUserId":5765409.0,"Title":"Tensorflow multiple sessions with multiple GPUs","Body":"<p>I have a workstation with 2 GPUs and I am trying to run multiple tensorflow jobs at the same time, so I can train more than one model at once, etc.<\/p>\n\n<p>For example, I've tried to separate the sessions into different resources via the python API using in script1.py:<\/p>\n\n<pre><\/pre>\n\n<p>in script2.py:<\/p>\n\n<pre><\/pre>\n\n<p>in script3.py<\/p>\n\n<pre><\/pre>\n\n<p>If I run each script by itself I can see that it is using the specified device. (Also the models fit very well into a single GPU and doesn't use another one even if both are available.)<\/p>\n\n<p>However, if one script is running and I try to run another, I always get this error:<\/p>\n\n<pre><\/pre>\n\n<p>It seems each tensorflow process is trying to grab all of the GPUs on the machine when it loads even if not all devices are going to be used to run the model.<\/p>\n\n<p>I see there is an option to limit the amount of GPU each process uses<\/p>\n\n<pre><\/pre>\n\n<p>...I haven't tried it, but this seems like it would make two processes try to share 50% of each GPU instead of running each process on a separate GPU...<\/p>\n\n<p>Does anyone know how to configure tensorflow to use only one GPU and leave the other available for another tensorflow process?<\/p>\n","answers":[{"AnswerId":"34776814","CreationDate":"2016-01-13T20:53:25.210","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow will attempt to use (an equal fraction of the memory of) all GPU devices that are visible to it. If you want to run different sessions on different GPUs, you should do the following.<\/p>\n\n<ol>\n<li>Run each session in a different Python process.<\/li>\n<li><p>Start each process with a different value for the <a href=\"http:\/\/www.acceleware.com\/blog\/cudavisibledevices-masking-gpus\"><code>CUDA_VISIBLE_DEVICES<\/code> environment variable<\/a>. For example, if your script is called <code>my_script.py<\/code> and you have 4 GPUs, you could run the following:<\/p>\n\n<pre><code>$ CUDA_VISIBLE_DEVICES=0 python my_script.py  # Uses GPU 0.\n$ CUDA_VISIBLE_DEVICES=1 python my_script.py  # Uses GPU 1.\n$ CUDA_VISIBLE_DEVICES=2,3 python my_script.py  # Uses GPUs 2 and 3.\n<\/code><\/pre>\n\n<p>Note the GPU devices in TensorFlow will still be numbered from zero (i.e. <code>\"\/gpu:0\"<\/code> etc.), but they will correspond to the devices that you have made visible with <code>CUDA_VISIBLE_DEVICES<\/code>.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34778963,"AnswerCount":2,"Tags":"<image-processing><classification><tensorflow><conv-neural-network>","CreationDate":"2016-01-13T23:21:05.197","AcceptedAnswerId":null,"OwnerUserId":1119500.0,"Title":"Image classification with TensorFlow CNN","Body":"<p>I'm very new to Image Classification with CNNs and I followed the Tutorial for googles CNN: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/images\/cnn<\/a><\/p>\n\n<p>And this works very well but suddenly stops after training and evaluation of the model. I don't know why they stop :(<\/p>\n\n<p>Does anybody know if there is a python function like: <\/p>\n\n<pre><\/pre>\n\n<p>that return me the class and if possible all pixels of the object that was found. Or whats the way to do this?<\/p>\n","answers":[{"AnswerId":"43898280","CreationDate":"2017-05-10T16:38:46.103","ParentId":null,"OwnerUserId":"7984359","Title":null,"Body":"<pre><code>#to save a model\nsess = tf.Session()\nsaver.save(sess, '\/path\/to\/my\/saved\/model\/my_test_model)\n<\/code><\/pre>\n\n<h1>To load saved model in tensor flow<\/h1>\n\n<pre><code>with tf.Session() as sess:    \n    saver = \n    tf.train.import_meta_graph('\/path\/to\/my\/saved\/model\/my_test_model-\n    1000.meta',clear_devices=True)\n    saver.restore(sess,tf.train.latest_checkpoint('.\/'))\n<\/code><\/pre>\n\n<h1>predict on loaded model<\/h1>\n\n<pre><code>sess.run(tf.argmax(y, 1), feed_dict={x: [img]})\n<\/code><\/pre>\n"},{"AnswerId":"34822097","CreationDate":"2016-01-16T00:25:46.557","ParentId":null,"OwnerUserId":"2619740","Title":null,"Body":"<p>First you need to be more specific about \"stop to work\". What is the problem? Do you get the error or it freezes somewhere?\nSecondly, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">this<\/a> maybe useful for you<\/p>\n"}]}
{"QuestionId":34779015,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-13T23:27:03.613","AcceptedAnswerId":34780927.0,"OwnerUserId":557478.0,"Title":"column_stack equivalent in Theano","Body":"<p>I am trying to stack a tensor as columns to an existing tensor (basically equivalent of <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/reference\/generated\/numpy.column_stack.html\" rel=\"nofollow\">column_stack<\/a> in numpy)<\/p>\n\n<pre><\/pre>\n\n<p>I need the same for tensor:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34780927","CreationDate":"2016-01-14T02:56:47.623","ParentId":null,"OwnerUserId":"1606479","Title":null,"Body":"<p>Use the <strong>concatenate<\/strong> function<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\n\nx = T.fmatrix()\ny = T.fmatrix()\na = np.asarray([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=theano.config.floatX)\nb = np.asarray([[11, 12, 13], [14, 15, 16], [17, 18, 19]], dtype=theano.config.floatX)\nz = T.concatenate([x, y], axis=1)\nt = theano.function([x, y], z)\nprint t(a, b)\n<\/code><\/pre>\n"}]}
{"QuestionId":34779638,"AnswerCount":1,"Tags":"<ipython><tensorflow><gnu-screen>","CreationDate":"2016-01-14T00:29:37.193","AcceptedAnswerId":34779794.0,"OwnerUserId":2561747.0,"Title":"TensorFlow: GNU Screen for IPython causes import error","Body":"<p>I've installed TensorFlow with GPU from source on an AWS EC2 Ubuntu 14.04 system.\nUsing  results in successful tensorflow imports and training\/testing with multiple GPUs.\nHowever when I use , I get the error:<\/p>\n\n<p>.<\/p>\n\n<p>I suspected this was a problem with environment variables, and found that indeed <\/p>\n\n<pre><\/pre>\n\n<p>returns  without screen and gets a  with screen.<\/p>\n\n<p>So I created  with contents<\/p>\n\n<pre><\/pre>\n\n<p>and now  gives nearly the same environment variables (except for some that relate to terminal display that appear unrelated) but I still get the same import error as before.<\/p>\n","answers":[{"AnswerId":"34779794","CreationDate":"2016-01-14T00:46:21.590","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Screen unsets various environment variables (including <code>LD_LIBRARY_PATH<\/code>, see <a href=\"https:\/\/unix.stackexchange.com\/questions\/43729\/how-does-screen-impact-environment-variables\">this Unix StackExchange answer for details<\/a>). As you've discovered, setting <code>os.environ['LD_LIBRARY_PATH']<\/code> doesn't fix the <code>ImportError<\/code> because this environment variable must be set when the process launches. The easiest way to fix this is to modify your <code>screen<\/code> invocation as follows:<\/p>\n\n<pre><code>$ screen env LD_LIBRARY_PATH=$LD_LIBRARY_PATH ipython\n<\/code><\/pre>\n\n<p>This will set <code>LD_LIBRARY_PATH<\/code> within the screen session and it will take effect before <code>ipython<\/code> launches, as is needed.<\/p>\n"}]}
{"QuestionId":34781192,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-14T03:26:25.417","AcceptedAnswerId":34782186.0,"OwnerUserId":5703903.0,"Title":"TensorFlow. Changes in code in CIFAR10 example are not reflected when I run cifar10_train.py","Body":"<p>For example, I wanted to change the distortions in distorted_inputs. I commented out the random crops and the random flips. After doing so, I saved the file, and ran cifar10_train.py. I then ran TensorBoard, and viewed the images in the image visualizer. I realized that they are still flipped and cropped. This is something strange, as I have already edited the code, so it should not be happening. Any fixes?<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34782186","CreationDate":"2016-01-14T05:16:24.373","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>I found the problem. Changing the lines at the top of each .py file, such as:<\/p>\n\n<pre><code>from tensorflow.models.image.cifar10 import cifar10\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>import cifar10\n<\/code><\/pre>\n\n<p>did the trick. I'm not too sure myself on why it works, but it's probably because doing such an import goes into python dist-packages. Looking forward to an answer from someone who knows exactly how this works!<\/p>\n"}]}
{"QuestionId":34782201,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-14T05:17:55.853","AcceptedAnswerId":34785645.0,"OwnerUserId":5703903.0,"Title":"TensorFlow CIFAR10 cifar10_eval.py throws error: Compute status: Invalid argument: Assign requires shapes of both tensors to match","Body":"<p>I am running the <a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"nofollow\">SVHN<\/a> data set on the CIFAR10 example provided in the TensorFlow packages. All I did was just to change the source directories for the data, and modify a few lines of code here and there. I can successfully train the network.<\/p>\n\n<p>However, when I run svhn_eval.py (the equivalent of cifar10_eval.py, names changed so I know how to organize my files), I get this error of assign requires shape of both tensors to match. I guess that the problem could be due to <\/p>\n\n<pre><\/pre>\n\n<p>as the trace ends there and goes deep into the other files of TensorFlow. Does anyone know how to solve this?<\/p>\n\n<pre><\/pre>\n\n<p>EDIT 1: The lines of code that I changed are in distorted_inputs. In the the original CIFAR10, there was random crop from a 32x32 to a 24x24 picture. However, in the SVHN implementation, I input 32x32 images. Based on the output error, we can sort of figure out what's wrong.<\/p>\n\n<pre><\/pre>\n\n<p>The question we have to ask ourselves now is, why multiply by 4?<\/p>\n","answers":[{"AnswerId":"34785645","CreationDate":"2016-01-14T09:18:22.713","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>The solution is that cifar10.py has a variable called IMAGE_SIZE. I left it as 24, because I thought it would not affect anything. However, what happens is that when you try and run the test set, the inputs are cropped to a size of IMAGE_SIZE x IMAGE_SIZE.<\/p>\n\n<p>Therefore, when that wasn't changed, the tensor dimensions do not match. Changing that variable to 32 will do the trick.<\/p>\n"}]}
{"QuestionId":34783030,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-14T06:30:17.627","AcceptedAnswerId":null,"OwnerUserId":2959589.0,"Title":"Saving image files in Tensorflow","Body":"<p>I'm just starting with Tensorflow and I have a newbie question.<\/p>\n\n<p>I know that Tensorflow is all about neural nets but I'm starting with just the mechanics of it. I'm trying to get it to load, resize, flip, and save two images. Should be a simple operation, right, and it gets me started with the basics.<\/p>\n\n<p>Here's my code so far:<\/p>\n\n<pre><\/pre>\n\n<p>It works fine, resizing the two images and saving them. But it always ends with an error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm obviously doing something wrong. If I take off the num_epochs=1, then it ends with no error.<\/p>\n\n<p>I have a few questions:<\/p>\n\n<p>How do I do this correctly?<\/p>\n\n<p>Also, if I want to preserve the original file names all the way from the filename_queue through to the end so I can save them with the original names, how do I do that? And how do I know how many files I need to save? Let's say I'm making the list of file names by reading a directory. I tried many different things but I could never find out how I know when I reach the end.<\/p>\n\n<p>It seems strange to me that I'm calling resized_encoded.eval() twice.<\/p>\n\n<p>Thank you and I'm sure this is a very basic question but I'm not understanding how this works.<\/p>\n\n<p>Edit: I created an even simpler demonstration of the behavior:<\/p>\n\n<pre><\/pre>\n\n<p>This gives the same warning. I can't understand why.<\/p>\n","answers":[{"AnswerId":"34805482","CreationDate":"2016-01-15T06:31:17.933","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>This warning is perfectly normal. As stated in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"noreferrer\">TensorFlow API<\/a><\/p>\n\n<blockquote>\n  <p>num_epochs: An integer (optional). If specified, string_input_producer\n  produces each string from string_tensor num_epochs times before\n  generating an OutOfRange error. If not specified,\n  string_input_producer can cycle through the strings in string_tensor\n  an unlimited number of times.<\/p>\n<\/blockquote>\n\n<p>Why is this important, you may ask. I have refactored your code into something perhaps more understandable, in my opinion. Let me explain.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nimport os\nfrom PIL import Image\n\ncur_dir = os.getcwd()\nprint(\"resizing images\")\nprint(\"current directory:\",cur_dir)\n\ndef modify_image(image):\n    resized = tf.image.resize_images(image, 180, 180, 1)\n    resized.set_shape([180,180,3])\n    flipped_images = tf.image.flip_up_down(resized)\n    return flipped_images\n\ndef read_image(filename_queue):\n    reader = tf.WholeFileReader()\n    key,value = reader.read(filename_queue)\n    image = tf.image.decode_jpeg(value)\n    return image\n\ndef inputs():\n    filenames = ['img1.jpg', 'img2.jpg' ]\n    filename_queue = tf.train.string_input_producer(filenames,num_epochs=2)\n    read_input = read_image(filename_queue)\n    reshaped_image = modify_image(read_input)\n    return reshaped_image\n\nwith tf.Graph().as_default():\n    image = inputs()\n    init = tf.initialize_all_variables()\n    sess = tf.Session()\n    sess.run(init)\n    tf.train.start_queue_runners(sess=sess)\n    for i in xrange(2):\n        img = sess.run(image)\n        img = Image.fromarray(img, \"RGB\")\n        img.save(os.path.join(cur_dir,\"foo\"+str(i)+\".jpeg\"))\n<\/code><\/pre>\n\n<p>In the code above, if you explicitly put num_epochs=2, then as the API suggests, string_input_producer cycles through the strings in string_tensor 2 times. Since the string_tensor has only 2 filenames, the queue is filled with 4 filenames. If I change the for loop to be:<\/p>\n\n<pre><code>for i in xrange(5)\n<\/code><\/pre>\n\n<p>then this will bug out. However, if I leave it at 4, then it will be fine. Take yet another example. If I do not put num_epochs, then as suggested, it can cycle through unlimited number of times. Putting:<\/p>\n\n<pre><code>for i in xrange(100)\n<\/code><\/pre>\n\n<p>thus does not bug out. I hope this answers your question.<\/p>\n\n<p>EDIT: I realized you have more questions.<\/p>\n\n<blockquote>\n  <p>Also, if I want to preserve the original file names all the way from\n  the filename_queue through to the end so I can save them with the\n  original names, how do I do that? And how do I know how many files I\n  need to save? Let's say I'm making the list of file names by reading a\n  directory. I tried many different things but I could never find out\n  how I know when I reach the end.<\/p>\n<\/blockquote>\n\n<p>If you want to preserve the original file names, then your method needs to return the file name. Here's the code below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nimport os\nfrom PIL import Image\n\ncur_dir = os.getcwd()\nprint(\"resizing images\")\nprint(\"current directory:\",cur_dir)\n\ndef modify_image(image):\n    resized = tf.image.resize_images(image, 180, 180, 1)\n    resized.set_shape([180,180,3])\n    flipped_images = tf.image.flip_up_down(resized)\n    return flipped_images\n\ndef read_image(filename_queue):\n    reader = tf.WholeFileReader()\n    key,value = reader.read(filename_queue)\n    image = tf.image.decode_jpeg(value)\n    return key,image\n\ndef inputs():\n    filenames = ['img1.jpg', 'img2.jpg' ]\n    filename_queue = tf.train.string_input_producer(filenames)\n    filename,read_input = read_image(filename_queue)\n    reshaped_image = modify_image(read_input)\n    return filename,reshaped_image\n\nwith tf.Graph().as_default():\n    image = inputs()\n    init = tf.initialize_all_variables()\n    sess = tf.Session()\n    sess.run(init)\n    tf.train.start_queue_runners(sess=sess)\n    for i in xrange(10):\n        filename,img = sess.run(image)\n        print (filename)\n        img = Image.fromarray(img, \"RGB\")\n        img.save(os.path.join(cur_dir,\"foo\"+str(i)+\".jpeg\"))\n<\/code><\/pre>\n\n<p>To know how many files you need to save, you could just call something along the lines of:<\/p>\n\n<pre><code>os.listdir(os.getcwd())\n<\/code><\/pre>\n\n<p>This lists all files in the directory. Check the API of os.listdir to filter specifically JPG, PNG file types. Once you get this, you can call a simple length operation and do:<\/p>\n\n<pre><code>for i in xrange(len(number_of_elements))\n<\/code><\/pre>\n"}]}
{"QuestionId":34785414,"AnswerCount":11,"Tags":"<c++><windows><visual-studio><machine-learning><tensorflow>","CreationDate":"2016-01-14T09:06:14.083","AcceptedAnswerId":34788819.0,"OwnerUserId":4564882.0,"Title":"How to install TensorFlow on Windows?","Body":"<p>I am starting to work with TensorFlow library for deep learning, <a href=\"https:\/\/www.tensorflow.org\/\">https:\/\/www.tensorflow.org\/<\/a>. <\/p>\n\n<p>I found a explicit guide to work on it on linux and Mac but I did not find how to work with it under Windows. I try over the net, but the information are lacking. <\/p>\n\n<p>I use Visual Studio 2015 for my projects, and I am trying to compile the library with Visual studio Compiler VC14.<\/p>\n\n<p>How to install it and to use it under Windows?<\/p>\n\n<p>Can I use <a href=\"https:\/\/github.com\/dslomov\/bazel-windows\">Bazel for Windows <\/a> for production use?<\/p>\n","answers":[{"AnswerId":"41336556","CreationDate":"2016-12-26T21:43:33.553","ParentId":null,"OwnerUserId":"7343594","Title":null,"Body":"<p>As of writing this answer, I wasn't able to get tensorflow to install properly with python version 3.5.2.  Reverting to <a href=\"https:\/\/www.python.org\/downloads\/release\/python-350\/\" rel=\"nofollow noreferrer\">python 3.5.0<\/a> did the trick.<\/p>\n\n<p>Then I was able to install with <\/p>\n\n<blockquote>\n  <p>C:> pip install tensorflow<\/p>\n<\/blockquote>\n"},{"AnswerId":"34787383","CreationDate":"2016-01-14T10:37:44.823","ParentId":null,"OwnerUserId":"1540501","Title":null,"Body":"<p>Following may work for you: install Virtual Box, create Linux VM and install Linux into it. I'd recommend Ubuntu, because Google often uses it internally. Then, install TensorFlow in Linux VM.<\/p>\n"},{"AnswerId":"40882362","CreationDate":"2016-11-30T07:13:21.337","ParentId":null,"OwnerUserId":"3054427","Title":null,"Body":"<p><em>Sorry for the excavation, but this question is quite popular, and now it has a different answer.<\/em><\/p>\n\n<p>Google officially announced the addition of Windows (7, 10, and Server 2016) support for TensorFlow:\n<a href=\"https:\/\/developers.googleblog.com\/2016\/11\/tensorflow-0-12-adds-support-for-windows.html\" rel=\"nofollow noreferrer\">developers.googleblog.com<\/a><\/p>\n\n<p>The Python module can be installed using pip with a single command:<\/p>\n\n<pre><code>C:\\&gt; pip install tensorflow\n<\/code><\/pre>\n\n<p>And if you need GPU support:<\/p>\n\n<pre><code> C:\\&gt; pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup.html#pip-installation-on-windows\" rel=\"nofollow noreferrer\">TensorFlow manual - How to install pip on windows<\/a><\/p>\n\n<p>Another useful information are included in release notes:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/releases<\/a><\/p>\n\n<p><strong>UPD:<\/strong> As @m02ph3u5 right mentioned in the comments TF for windows supports only Python 3.5.x <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows#installing_with_native_pip\" rel=\"nofollow noreferrer\">Installing TensorFlow on Windows with native pip<\/a><\/p>\n"},{"AnswerId":"41734178","CreationDate":"2017-01-19T05:08:28.830","ParentId":null,"OwnerUserId":"3253000","Title":null,"Body":"<p>If you have already installed anaconda on your windows, there is an easier way as I found out:<\/p>\n\n<pre><code>conda create --name snakes python=3\n<\/code><\/pre>\n\n<p>Then<\/p>\n\n<pre><code>activate snakes\n<\/code><\/pre>\n\n<p>Then<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>This is similar to virtualenv and I found this helpful.<\/p>\n"},{"AnswerId":"39902815","CreationDate":"2016-10-06T18:04:06.047","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<h1>Installing TensorFlow<\/h1>\n\n<p>TensorFlow currently supports only Python 3.5 64-bit. Both CPU and GPU are supported. Here are some installation instructions assuming you do not have Python 3.5 64-bit:<\/p>\n\n<ol>\n<li>Download and install Microsoft Visual C++ 2015 Redistributable Update 3:\n<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> (required by Python 3.5 and TensorFlow)<\/li>\n<li>Download and install Python 3.5 64-bit: <a href=\"https:\/\/www.python.org\/ftp\/python\/3.5.2\/python-3.5.2-amd64.exe\" rel=\"nofollow noreferrer\">https:\/\/www.python.org\/ftp\/python\/3.5.2\/python-3.5.2-amd64.exe<\/a><\/li>\n<li>Install pip as follows: download <a href=\"https:\/\/bootstrap.pypa.io\/get-pip.py\" rel=\"nofollow noreferrer\">https:\/\/bootstrap.pypa.io\/get-pip.py<\/a>, then run <code>python get-pip.py<\/code><\/li>\n<li>Install TensorFlow with either <code>pip install tensorflow<\/code> (CPU version) or <code>pip install tensorflow-gpu<\/code> (GPU version --> requires CUDA to be installed).<\/li>\n<\/ol>\n\n<hr>\n\n<h1>Testing TensorFlow<\/h1>\n\n<p>You can now run something like following to test whether TensorFlow is working fine:<\/p>\n\n<pre><code>import tensorflow as tf\nhello = tf.constant('Hello, TensorFlow!')\nsess = tf.Session()\nprint(sess.run(hello))\na = tf.constant(10)\nb = tf.constant(32)\nprint(sess.run(a + b))\n<\/code><\/pre>\n\n<p>TensorFlow comes with a few models, which are located in <code>C:\\Python35\\Lib\\site-packages\\tensorflow\\models\\<\/code> (assuming you installed python in <code>C:\\Python35<\/code>). For example, you can run in the console: <\/p>\n\n<pre><code>python -m tensorflow.models.image.mnist.convolutional\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>python C:\\Python35\\Lib\\site-packages\\tensorflow\\models\\image\\mnist\\convolutional.py\n<\/code><\/pre>\n\n<hr>\n\n<h1>Limitations of TensorFlow on Windows<\/h1>\n\n<p>Initial support for building TensorFlow on Microsoft Windows was added on  2016-10-05 in commit <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/2098b9abcf20d2c9694055bbfd6997bc00b73578\" rel=\"nofollow noreferrer\">2098b9abcf20d2c9694055bbfd6997bc00b73578<\/a>:<\/p>\n\n<blockquote>\n  <p>This PR contains an initial version of support for building TensorFlow\n  (CPU only) on Windows using CMake. It includes documentation for\n  building with CMake on Windows, platform-specific code for\n  implementing core functions on Windows, and CMake rules for building\n  the C++ example trainer program and a PIP package (Python 3.5 only).\n  The CMake rules support building TensorFlow with Visual Studio 2015.<\/p>\n  \n  <p>Windows support is a work in progress, and we welcome your feedback\n  and contributions.<\/p>\n  \n  <p>For full details of the features currently supported and instructions\n  for how to build TensorFlow on Windows, please see the file\n  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/cmake\/README.md\" rel=\"nofollow noreferrer\"><code>tensorflow\/contrib\/cmake\/README.md<\/code><\/a>.<\/p>\n<\/blockquote>\n\n<p>The Microsoft Windows support was introduced in TensorFlow in version 0.12 RC0 (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/tag\/0.12.0-rc0\" rel=\"nofollow noreferrer\">release notes<\/a>):<\/p>\n\n<blockquote>\n  <p>TensorFlow now builds and runs on Microsoft Windows (tested on Windows 10, Windows 7, and Windows Server 2016). Supported languages include Python (via a pip package) and C++. CUDA 8.0 and cuDNN 5.1 are supported for GPU acceleration. Known limitations include: It is not currently possible to load a custom op library. The GCS and HDFS file systems are not currently supported. The following ops are not currently implemented: DepthwiseConv2dNative, DepthwiseConv2dNativeBackpropFilter, DepthwiseConv2dNativeBackpropInput, Dequantize, Digamma, Erf, Erfc, Igamma, Igammac, Lgamma, Polygamma, QuantizeAndDequantize, QuantizedAvgPool, QuantizedBatchNomWithGlobalNormalization, QuantizedBiasAdd, QuantizedConcat, QuantizedConv2D, QuantizedMatmul, QuantizedMaxPool, QuantizeDownAndShrinkRange, QuantizedRelu, QuantizedRelu6, QuantizedReshape, QuantizeV2, RequantizationRange, and Requantize.<\/p>\n<\/blockquote>\n"},{"AnswerId":"40890603","CreationDate":"2016-11-30T14:14:46.680","ParentId":null,"OwnerUserId":"1389756","Title":null,"Body":"<p>Now Tensorflow is officially supported in Windows, you can install it using <code>pip<\/code> command of Python 3.5 without compile it yourself<\/p>\n\n<p>CPU Version<\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-0.12.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p><em>cp35<\/em> indicates <em>python 3.5<\/em> wheel, <em>0.12.0<\/em> the version, you can edit these according your preference, or to install latest CPU version available you can use<\/p>\n\n<pre><code>pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>GPU Version<\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-0.12.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p><em>cp35<\/em> indicates <em>python 3.5<\/em> wheel, <em>0.12.0<\/em> the version, you can edit these according your preference, or to install latest GPU version available you can use<\/p>\n\n<pre><code>pip install --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup.html#pip-installation-on-windows\" rel=\"nofollow noreferrer\">More Info<\/a> <\/p>\n"},{"AnswerId":"37172799","CreationDate":"2016-05-11T20:36:35.540","ParentId":null,"OwnerUserId":"4473230","Title":null,"Body":"<p>I can confirm that it works in the Windows Subsystem for Linux!\nAnd it is also very straightforward.<\/p>\n\n<p>In the Ubuntu Bash on Windows 10, first update the package index:<\/p>\n\n<pre><code>apt-get update\n<\/code><\/pre>\n\n<p>Then install pip for Python 2:<\/p>\n\n<pre><code>sudo apt-get install python-pip python-dev\n<\/code><\/pre>\n\n<p>Install tensorflow:<\/p>\n\n<pre><code>sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.8.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>The package is now installed an you can run the CNN sample on the MNIST set:<\/p>\n\n<pre><code>cd \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/models\/image\/mnist\n\npython convolutional.py\n<\/code><\/pre>\n\n<p>I just tested the CPU package for now.<\/p>\n\n<p>I blogged about it: <a href=\"http:\/\/blog.mosthege.net\/2016\/05\/11\/running-tensorflow-with-native-linux-binaries-in-the-windows-subsystem-for-linux\/\" rel=\"noreferrer\">http:\/\/blog.mosthege.net\/2016\/05\/11\/running-tensorflow-with-native-linux-binaries-in-the-windows-subsystem-for-linux\/<\/a><\/p>\n\n<p>cheers<\/p>\n\n<p>~michael<\/p>\n"},{"AnswerId":"34788819","CreationDate":"2016-01-14T11:44:45.850","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<blockquote>\n  <p>How to install TensorFlow and to use it under Windows?<\/p>\n<\/blockquote>\n\n<p><strong>Updated on 8\/4\/16<\/strong> <\/p>\n\n<p>Windows 10 now has a <a href=\"http:\/\/www.ubuntu.com\/\" rel=\"noreferrer\">Ubuntu<\/a> Bash environment, AKA <a href=\"https:\/\/msdn.microsoft.com\/en-us\/commandline\/wsl\/about\" rel=\"noreferrer\">Bash on Ubuntu on Windows<\/a>, available as a standard option (as opposed to <a href=\"https:\/\/developer.microsoft.com\/en-us\/windows\/downloads\/windows-10-developer-preview\" rel=\"noreferrer\">Insider Preview updates for developers<\/a>). (StackOverflow tag <a href=\"https:\/\/stackoverflow.com\/questions\/tagged\/wsl\">wsl<\/a>) This option came with the <a href=\"https:\/\/support.microsoft.com\/en-us\/help\/12387\/windows-10-update-history\" rel=\"noreferrer\">Windows 10 anniversary update<\/a> (Version 1607) released on 8\/2\/2016. This allows the use of <a href=\"https:\/\/help.ubuntu.com\/community\/AptGet\/Howto\" rel=\"noreferrer\">apt-get<\/a> to install software packages such as <a href=\"https:\/\/www.python.org\/\" rel=\"noreferrer\">Python<\/a> and <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"noreferrer\">TensorFlow<\/a>.<\/p>\n\n<p>Note: Bash on Ubuntu on Windows does not have access to the GPU, so all of the GPU options for installing TensorFlow will not work.<\/p>\n\n<p>The dated <a href=\"https:\/\/msdn.microsoft.com\/en-us\/commandline\/wsl\/install_guide\" rel=\"noreferrer\">installation instructions<\/a> for Bash on Ubuntu on Windows are basically correct, but only these steps are necessary:<br>\nPrerequisites<br>\nEnable the Windows Subsystem for Linux feature (GUI)<br>\nReboot when prompted<br>\nRun Bash on Windows  <\/p>\n\n<p>Steps no longer needed:<br>\nTurn on Developer Mode<br>\nEnable the Windows Subsystem for Linux feature (command-line)  <\/p>\n\n<p>Then install TensorFlow using apt-get<\/p>\n\n<pre><code>sudo apt-get install python3-pip python3-dev\nsudo pip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.8.0-cp34-cp34m-linux_x86_64.whl \n<\/code><\/pre>\n\n<p>and now test TensorFlow<\/p>\n\n<pre><code>$ python3\n...\n&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; hello = tf.constant('Hello, TensorFlow!')\n&gt;&gt;&gt; sess = tf.Session()\n&gt;&gt;&gt; print(sess.run(hello))\nHello, TensorFlow!\n&gt;&gt;&gt; a = tf.constant(10)\n&gt;&gt;&gt; b = tf.constant(32)\n&gt;&gt;&gt; print(sess.run(a + b))\n42\n&gt;&gt;&gt; exit()\n<\/code><\/pre>\n\n<p>and run an actual neural network<\/p>\n\n<pre><code>python3 -m tensorflow.models.image.mnist.convolutional\n<\/code><\/pre>\n\n<h2>Earlier Answer<\/h2>\n\n<p>After learning about the developer preview of Bash on Windows.<\/p>\n\n<p>See <a href=\"http:\/\/www.hanselman.com\/blog\/PlayingWithTensorFlowOnWindows.aspx\" rel=\"noreferrer\">Playing with TensorFlow on Windows<\/a> by Scott Hanselman which uses <a href=\"http:\/\/www.hanselman.com\/blog\/DevelopersCanRunBashShellAndUsermodeUbuntuLinuxBinariesOnWindows10.aspx\" rel=\"noreferrer\">Bash on Windows 10<\/a> <\/p>\n\n<h2>Original Answer<\/h2>\n\n<p><strong>Bazel is the problem<\/strong><\/p>\n\n<p>TensorFlow is not made with <a href=\"https:\/\/en.wikipedia.org\/wiki\/List_of_build_automation_software\" rel=\"noreferrer\">build automation tools<\/a> such as <a href=\"https:\/\/en.wikipedia.org\/wiki\/Make_(software)\" rel=\"noreferrer\">make<\/a>, but with Google's in-house build tool <a href=\"http:\/\/bazel.io\/faq.html\" rel=\"noreferrer\">Bazel<\/a>. Bazel only works on systems based on <a href=\"http:\/\/www.unix.org\/\" rel=\"noreferrer\">Unix<\/a> such as <a href=\"https:\/\/en.wikipedia.org\/wiki\/Linux\" rel=\"noreferrer\">Linux<\/a> and <a href=\"http:\/\/www.apple.com\/osx\/what-is\/\" rel=\"noreferrer\">OS X<\/a>.  <\/p>\n\n<p>Since the current published\/known means to build TensorFlow uses <a href=\"http:\/\/bazel.io\/\" rel=\"noreferrer\">Bazel<\/a> and Bazel does not work on Windows, one can not install or run TensorFlow natively on Windows.  <\/p>\n\n<p>From <a href=\"http:\/\/bazel.io\/faq.html\" rel=\"noreferrer\">Bazel FAQ<\/a><\/p>\n\n<blockquote>\n  <p>What about Windows?<\/p>\n  \n  <p>Due to its UNIX heritage, porting Bazel to Windows is significant\n  work. For example, Bazel uses symlinks extensively, which has varying\n  levels of support across Windows versions.<\/p>\n  \n  <p>We are currently actively working on improving Windows support, but\n  it's still ways from being usable.<\/p>\n<\/blockquote>\n\n<p><em>Status<\/em><\/p>\n\n<p>See: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"noreferrer\">TensorFlow issue #17<\/a><br>\nSee: <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/276\" rel=\"noreferrer\">Bazel issue #276<\/a><\/p>\n\n<h2>Solutions<\/h2>\n\n<p>The solutions are listed in the order of complexity and work needed; from about an hour to may not even work.<\/p>\n\n<ol>\n<li><em><a href=\"https:\/\/www.docker.com\/what-docker\" rel=\"noreferrer\">Docker<\/a><\/em><br>\n~ 1 hour<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#docker-installation\" rel=\"noreferrer\">Docker installation<\/a><\/p>\n\n<p>Docker is a system to build self contained versions of a Linux operating system running on your machine. When you install and run TensorFlow via Docker it completely isolates the installation from pre-existing packages on your machine.<\/p>\n\n<p>Also look at <a href=\"https:\/\/stackoverflow.com\/q\/34694701\/1243762\">TensorFlow - which Docker image to use?<\/a><\/p>\n\n<ol start=\"2\">\n<li><em>OS X<\/em><br>\n~ 1 hour<\/li>\n<\/ol>\n\n<p>If you have a current Mac running OS X then see: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installation-for-mac-os-x\" rel=\"noreferrer\">Installation for Mac OS X<\/a><\/p>\n\n<ol start=\"3\">\n<li><em>Linux<\/em><\/li>\n<\/ol>\n\n<p>The <a href=\"https:\/\/stackoverflow.com\/questions\/34664602\/building-new-tensorflow-op-is-there-a-build-system-standard\">recommend Linux system<\/a> tends to be <a href=\"http:\/\/www.ubuntu.com\/\" rel=\"noreferrer\">Ubuntu 14.04 LTS<\/a> (<a href=\"http:\/\/www.ubuntu.com\/download\" rel=\"noreferrer\">Download page<\/a>).<\/p>\n\n<p>a. Virtual Machine - <a href=\"https:\/\/en.wikipedia.org\/wiki\/Hardware_virtualization#Full_virtualization\" rel=\"noreferrer\">Hardware Virtualization - Full Virtualization<\/a><br>\n     ~ 3 hours<\/p>\n\n<p>Download and install a virtual machine such as the commercial <a href=\"http:\/\/www.vmware.com\/\" rel=\"noreferrer\">VMware<\/a> or the free <a href=\"https:\/\/www.virtualbox.org\/wiki\/Downloads\" rel=\"noreferrer\">Virtual Box<\/a>, after which you can install Linux and then install TensorFlow.  <\/p>\n\n<p>When you go to install TensorFlow you will be using <a href=\"https:\/\/en.wikipedia.org\/wiki\/Pip_(package_manager)\" rel=\"noreferrer\">Pip<\/a> - Python's package management system. Visual Studio users should think NuGet. The packages are known as <a href=\"http:\/\/pythonwheels.com\/\" rel=\"noreferrer\">wheels<\/a>.<\/p>\n\n<p>See: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#overview\" rel=\"noreferrer\">Pip Installation<\/a><\/p>\n\n<p>If you need to build from the source then see: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\" rel=\"noreferrer\">Installing From Sources<\/a><br>\n~ 4 hours<\/p>\n\n<p>Note: If you plan on using a Virtual Machine and have never done so before, consider using the Docker option instead, since Docker is the Virtual Machine, OS and TensorFlow all packaged together.<\/p>\n\n<p>b. <a href=\"https:\/\/en.wikipedia.org\/wiki\/Multi-booting\" rel=\"noreferrer\">Dual boot<\/a><br>\n     ~ 3 hours<\/p>\n\n<p>If you want to run TensorFlow on the same machine that you have Windows and make use of the GPU version then you will most likely have to use this option as running on a hosted Virtual Machine, <a href=\"https:\/\/en.wikipedia.org\/wiki\/Hypervisor\" rel=\"noreferrer\">Type 2 hypervisor<\/a>, will not allow you access to the GPU. <\/p>\n\n<ol start=\"4\">\n<li><em>Remote machine<\/em><br>\n~ 4 hours<\/li>\n<\/ol>\n\n<p>If you have <a href=\"https:\/\/en.wikipedia.org\/wiki\/Remote_desktop_software\" rel=\"noreferrer\">remote access<\/a> to another machine that you can install the Linux OS and TensorFlow software on and allow remote connections to, then you can use your Windows machine to present the remote machine as an application running on Windows.<\/p>\n\n<ol start=\"5\">\n<li><em>Cloud Service<\/em><br>\nI have no experience with this. Please edit answer if you know.<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Cloud_computing\" rel=\"noreferrer\">Cloud<\/a> services such as <a href=\"https:\/\/aws.amazon.com\/\" rel=\"noreferrer\">AWS<\/a> are being used.  <\/p>\n\n<p>From <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"noreferrer\">TensorFlow Features<\/a><\/p>\n\n<blockquote>\n  <p>Want to run the model as a service in the cloud?\n  Containerize with Docker and TensorFlow just works.<\/p>\n<\/blockquote>\n\n<p>From <a href=\"https:\/\/www.docker.com\/\" rel=\"noreferrer\">Docker<\/a><\/p>\n\n<blockquote>\n  <p>Running Docker on AWS provides a highly reliable, low-cost way to\n  quickly build, ship, and run distributed applications at scale. Deploy\n  Docker using AMIs from the AWS Marketplace.<\/p>\n<\/blockquote>\n\n<ol start=\"6\">\n<li>Wait for Bazel to work on Windows.<\/li>\n<\/ol>\n\n<p>Currently it appears the only hold up is Bazel, however <a href=\"http:\/\/bazel.io\/roadmap.html\" rel=\"noreferrer\">Bazel's roadmap<\/a> list working on Windows should be available this year.<\/p>\n\n<p>There are two features listed for Windows:<\/p>\n\n<pre><code>2016\u201102  Bazel can bootstrap itself on Windows without requiring admin privileges.  \n\n2016\u201112  Full Windows support for Android: Android feature set is identical for Windows and Linux\/OS X.\n<\/code><\/pre>\n\n<ol start=\"7\">\n<li>Build TensorFlow by hand.<br>\nA few days or more depending on you skill level. I gave up on this one; too many subprojects to build and files to locate.<\/li>\n<\/ol>\n\n<p>Remember that Bazel is only used to build TensorFlow. If you get the commands Bazel runs and the correct source code and libraries you should be able to build TensorFlow on Windows. See: <a href=\"https:\/\/stackoverflow.com\/q\/33983711\/1243762\">How do I get the commands executed by Bazel<\/a>.  <\/p>\n\n<p>While I have not researched this more, you can look at the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Continuous_integration\" rel=\"noreferrer\">continuous integration<\/a> info for needed files and info on how to they build it for testing. (<a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/tools\/ci_build\/README.md\" rel=\"noreferrer\">Readme<\/a>) (<a href=\"http:\/\/ci.tensorflow.org\/\" rel=\"noreferrer\">site<\/a>)<\/p>\n\n<ol start=\"8\">\n<li>Build Bazel on Windows<br>\nA few days or more depending on you skill level. I gave up on this one also; could not find the necessary source files needed for Windows.<\/li>\n<\/ol>\n\n<p>There is a public experimental source code version of <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/blob\/master\/site\/docs\/windows.md\" rel=\"noreferrer\">Bazel that boots on Windows<\/a>. You may be able to leverage this into getting Bazel to work on Windows, etc.  <\/p>\n\n<p>Also these solutions require the use of <a href=\"https:\/\/cygwin.com\/\" rel=\"noreferrer\">Cygwin<\/a> or <a href=\"http:\/\/www.mingw.org\/\" rel=\"noreferrer\">MinGW<\/a> which adds another layer of complexity.<\/p>\n\n<ol start=\"9\">\n<li>Use alternative build system such as Make<br>\nIf you get this one to work I would like to see in on GitHub.<\/li>\n<\/ol>\n\n<p>This currently does not exist for TensorFlow. It is a feature request.<\/p>\n\n<p>See: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/380\" rel=\"noreferrer\">TensorFlow issue 380<\/a><\/p>\n\n<ol start=\"10\">\n<li>Cross Build<br>\nIf you get this one to work I would like to see in on GitHub.<\/li>\n<\/ol>\n\n<p>You build TensorFlow on Linux using Bazel but change the build process to output a wheel that can be installed on Windows. This will require detailed knowledge of Bazel to change the configuration, and locating the source code and libraries that work with Windows. An option I would only suggest as a last resort. It may not even be possible.<\/p>\n\n<ol start=\"11\">\n<li>Run on the new Windows Subsystem for Linux.<\/li>\n<\/ol>\n\n<p>See: <a href=\"https:\/\/blogs.msdn.microsoft.com\/wsl\/2016\/04\/22\/windows-subsystem-for-linux-overview\/\" rel=\"noreferrer\">Windows Subsystem for Linux Overview<\/a><\/p>\n\n<p>You will know as much as I do by reading the referenced article. <\/p>\n\n<blockquote>\n  <p>Can I use Bazel for Windows for production use?<\/p>\n<\/blockquote>\n\n<p>Since it is experimental software I would not use on a production machine.<\/p>\n\n<p>Remember that you only need Bazel to build TensorFlow. So use the experimental code on a non production machine to build the <a href=\"http:\/\/pythonwheels.com\/\" rel=\"noreferrer\">wheel<\/a>, then install the wheel on a production machine. See: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#pip-installation\" rel=\"noreferrer\">Pip Installation<\/a><\/p>\n\n<p>TLDR;<\/p>\n\n<p>Currently I have several versions for learning. Most use a VMWare 7.1 Workstation to host Ubuntu 14.04 LTS or Ubuntu 15 or Debian. I also have one dual boot of Ubuntu 14.04 LTS on my Windows machine to access the GPU as the machine with VMware does not have the proper GPU. I would recommend that you give these machines at least 8G of memory either as RAM or RAM and swap space as I have run out of memory a few times.<\/p>\n"},{"AnswerId":"34786601","CreationDate":"2016-01-14T10:02:34.123","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You can't at the moment. The problem is that tensorflow uses the bazel build <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"nofollow\">another Google internal tool that has been exposed as an open source project<\/a> and it has only support for mac and unix.\nUntil bazel is ported to windows or another build system is added to tensorflow there is a little chance to run tensorflow natively on windows.<\/p>\n\n<p>That said you can install virtualbox and then install docker-machine and run a linux container with tensorflow inside it.<\/p>\n"},{"AnswerId":"40917999","CreationDate":"2016-12-01T18:35:13.807","ParentId":null,"OwnerUserId":"4913418","Title":null,"Body":"<p>Follow this <a href=\"http:\/\/bottomstack.com\/installing-tensorflow-windows-visual-studio\/\" rel=\"nofollow noreferrer\">link<\/a> to install Tensorflow on Windows and you can also use it in Visual Studio<\/p>\n"},{"AnswerId":"35255133","CreationDate":"2016-02-07T15:08:58.097","ParentId":null,"OwnerUserId":"4832010","Title":null,"Body":"<p>I managed to install TensorFlow on Win8.1 without Docker using advice from\n<a href=\"https:\/\/discussions.udacity.com\/t\/windows-tensorflow-and-visual-studio-2015\/45636\" rel=\"nofollow noreferrer\">https:\/\/discussions.udacity.com\/t\/windows-tensorflow-and-visual-studio-2015\/45636<\/a><\/p>\n\n<p>I tried a lot of stuff before that, and i won't try to install it twice but here is what i did:\n- install VS2015 (make sure Visual C++ installed as well)\n- install Python Tools for VS2015\n- install Python2.7 with Anaconda2\n- install pip and conda for Python\n- install numpy with pip inside VS2015\n- install tensorflow with pip inside VS2015<\/p>\n\n<p>i didn't manage to do it with Python3.5<\/p>\n\n<p>I managed also to install on Win8.1 via Cloud9\nThere is a video tutorial on Youtube.<\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=kMtrOIPLpR0\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=kMtrOIPLpR0<\/a><\/p>\n\n<p>EDIT: actually for the above, (not Cloud9 which is fine) i have problems:\nTensorFlow LOOKS LIKE it's installed (i can see it in the list of modules installed in VS2015 when clicking in Solution Explorer on Python 64-bit 2.7)\n<em>but<\/em> if i type in a script or in Python Interactive <code>import tensorflow as TF<\/code> then i get an error message <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\__init__.py\", line 23, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 50, in &lt;module&gt;\n    from tensorflow.python.framework.framework_lib import *\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\framework\\framework_lib.py\", line 62, in &lt;module&gt;\n    from tensorflow.python.framework.ops import Graph\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 40, in &lt;module&gt;\n    from tensorflow.python.framework import versions\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\framework\\versions.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow = swig_import_helper()\n  File \"C:\\Users\\Fagui\\Anaconda2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 20, in swig_import_helper\n    import _pywrap_tensorflow\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wBE21.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wBE21.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":34785629,"AnswerCount":0,"Tags":"<python><matrix-multiplication><theano><deep-learning><autoencoder>","CreationDate":"2016-01-14T09:17:34.937","AcceptedAnswerId":null,"OwnerUserId":5173166.0,"Title":"Merging and training Theano autoencoders","Body":"<p>I'm trying to use Theano's autoencoders to discover context-specific features from two different types of data.<\/p>\n\n<p>The first type has 13 features and the second one has 60.<\/p>\n\n<pre><\/pre>\n\n<p>Both have their own independent stack of autoencoders.<\/p>\n\n<p>I merge the outputs of the topmost layers and feed these into a regression layer for supervised training.<\/p>\n\n<pre><\/pre>\n\n<p>Pre-training for each context seems to work correctly, however I hit a snag during finetuning using the standard training function in the tutorials.<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I believe this has something to do with how the Theano nodes are processed during training. Seems like the training batch (5, 73), is being applied straight into the output nodes separately starting with the 1st context (13, 20).<\/p>\n","answers":[]}
{"QuestionId":34790159,"AnswerCount":1,"Tags":"<machine-learning><neural-network><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-01-14T12:53:39.947","AcceptedAnswerId":null,"OwnerUserId":1472128.0,"Title":"Stacked RNN model setup in TensorFlow","Body":"<p>I'm kind of lost in building up a stacked LSTM model for text classification in TensorFlow.<\/p>\n\n<p>My input data was something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The skeleton of my code looks like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The problem is that no matter how large the dataset is, the loss and accuracy has no sign of improvement (looks completely stochastic). Am I doing anything wrong?<\/p>\n\n<p>Update:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34805936","CreationDate":"2016-01-15T07:10:54.607","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>Here are few reasons it may not be training and suggestions to try:<\/p>\n\n<ul>\n<li><p>You are not allowing to update word vectors, space of pre-learned vectors may be not working properly.<\/p><\/li>\n<li><p>RNNs really need gradient clipping when trained. You can try adding something like <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/trainer.py#L80\" rel=\"nofollow\">this<\/a>.<\/p><\/li>\n<li><p>Unit scale initialization seems to work better, as it accounts for the size of the layer and allows gradient to be scaled properly as it goes deeper.<\/p><\/li>\n<li><p>You should try removing dropout and second layer - just to check if your data passing is correct and your loss is going down at all.<\/p><\/li>\n<\/ul>\n\n<p>I also can recommend trying this example with your data: <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/examples\/text_classification.py\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/examples\/text_classification.py<\/a><\/p>\n\n<p>It trains word vectors from scratch, already has gradient clipping and uses GRUCells which usually are easier to train. You can also see nice visualizations for loss and other things by running <code>tensorboard logdir=\/tmp\/tf_examples\/word_rnn<\/code>.<\/p>\n"}]}
{"QuestionId":34793978,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-14T15:56:25.197","AcceptedAnswerId":null,"OwnerUserId":5641162.0,"Title":"Tensorflow complaining about placeholder after model restore","Body":"\n\n<p>I am having a problem with Tensorflow restoring models. I have a script that generates several models, based on a set of training files. These models are created with their own variable scopings, using <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>After training, I am able to restore the models using<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>with model_vars computed as<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>While the models do seem to load, running them produces a Placeholder-error (see below,  is my prefix).<\/p>\n\n<p>I am pretty sure I do not skip any placeholders. The model just has two ( and ) and these are used by the  call I make:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here is the error trace:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Any help very much appreciated!<\/p>\n","answers":[]}
{"QuestionId":34795032,"AnswerCount":1,"Tags":"<tensorflow><mnist><tensorboard>","CreationDate":"2016-01-14T16:44:05.103","AcceptedAnswerId":34812828.0,"OwnerUserId":4844184.0,"Title":"Wrong dtype for a feed to the placeholder x-input TensorFlow","Body":"<p>I want to implement a simple logistic regression on MNIST with TF that I just installed and want to monitor the progress of the minibatch-SGD with TensorBoard.  <\/p>\n\n<p>I first did without tensorboard it compiled and got 0.9166 accuracy on testset.<\/p>\n\n<p>However when I added tensorboard to see what was going on I couldn't even compile it anymore I got:  <\/p>\n\n<p> but all my arrays are np arrays with dtype float !  <\/p>\n\n<p>If you can point the problem(s) in my code that would be amazing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The line where it happens is in the feed to merged however as I feed exactly the same way as I fed train_step I am at a loss...<\/p>\n","answers":[{"AnswerId":"34812828","CreationDate":"2016-01-15T14:06:32.033","ParentId":null,"OwnerUserId":"4844184","Title":null,"Body":"<p>It turns out you cannot run the same script over and over when I reopened a new spyder and launched the program it worked !!!\nMind=blown<\/p>\n"}]}
{"QuestionId":34796404,"AnswerCount":2,"Tags":"<c++><model><tensorflow>","CreationDate":"2016-01-14T17:51:25.253","AcceptedAnswerId":34817402.0,"OwnerUserId":5790995.0,"Title":"Not found: FeedInputs: unable to find feed output TensorFlow","Body":"<p>I was trying this example of using Tensorflow saved model in c++ in this website: \n<a href=\"https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.ji310n4zo\" rel=\"noreferrer\">https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.ji310n4zo<\/a><\/p>\n\n<p>It works well. But it does not save the values of the variables <strong><em>a<\/em><\/strong> and <strong><em>b<\/em><\/strong> as it only saves the graph not the variables. I tried to replace the following line:<\/p>\n\n<pre><\/pre>\n\n<p>with <\/p>\n\n<pre><\/pre>\n\n<p>of course after creating the saver object. It does not work and it outputs:<\/p>\n\n<p><strong><em>Not found: FeedInputs: unable to find feed output a<\/em><\/strong><\/p>\n\n<p>I checked the nodes the Nodes that are loaded and they are only:<\/p>\n\n<blockquote>\n  <p>_SOURCE<\/p>\n  \n  <p>_SINK<\/p>\n<\/blockquote>\n\n<p>while in the write_graph function and then load the model in C++, I got the following nodes loaded:<\/p>\n\n<blockquote>\n  <p>_SOURCE<\/p>\n  \n  <p>_SINK<\/p>\n  \n  <p>save\/restore_slice_1\/shape_and_slice<\/p>\n  \n  <p>save\/restore_slice_1\/tensor_name<\/p>\n  \n  <p>save\/restore_slice\/shape_and_slice<\/p>\n  \n  <p>save\/restore_slice\/tensor_name<\/p>\n  \n  <p>save\/save\/shapes_and_slices<\/p>\n  \n  <p>save\/save\/tensor_names<\/p>\n  \n  <p>save\/Const<\/p>\n  \n  <p>save\/restore_slice_1<\/p>\n  \n  <p>save\/restore_slice<\/p>\n  \n  <p>b<\/p>\n  \n  <p>save\/Assign_1<\/p>\n  \n  <p>b\/read<\/p>\n  \n  <p>b\/initial_value<\/p>\n  \n  <p>b\/Assign<\/p>\n  \n  <p>a<\/p>\n  \n  <p>save\/Assign<\/p>\n  \n  <p>save\/restore_all<\/p>\n  \n  <p>save\/save<\/p>\n  \n  <p>save\/control_dependency<\/p>\n  \n  <p>a\/read<\/p>\n  \n  <p>c<\/p>\n  \n  <p>a\/initial_value<\/p>\n  \n  <p>a\/Assign<\/p>\n  \n  <p>init<\/p>\n  \n  <p>Tensor<\/p>\n<\/blockquote>\n\n<p>and even the graph file that is created by saver.save() is much smaller, 165B, compared to the one created by write_graph, 1.9KB.<\/p>\n","answers":[{"AnswerId":"34817402","CreationDate":"2016-01-15T18:24:09.857","ParentId":null,"OwnerUserId":"5790995","Title":null,"Body":"<p>I'm not sure if that is the best way of solving the problem but at least it solves it.<\/p>\n\n<p>As write_graph can also store the values of the constants, I added the following code to the python just before writing the graph with write_graph function:<\/p>\n\n<pre><code>for v in tf.trainable_variables():\n    vc = tf.constant(v.eval())\n    tf.assign(v, vc, name=\"assign_variables\")\n<\/code><\/pre>\n\n<p>This creates constants that store variables' values after being trained and then create tensors \"<strong><em>assign_variables<\/em><\/strong>\" to assign them to the variables. Now, when you call write_graph, it will store the variables' values in the file.<\/p>\n\n<p>The only remaining part is to call these tensors \"<strong><em>assign_variables<\/em><\/strong>\" in the c code to make sure that your variables are assigned with the constants values that are stored in the file. Here is a one way to do it:<\/p>\n\n<pre><code>      Status status = NewSession(SessionOptions(), &amp;session);\n      std::vector&lt;tensorflow::Tensor&gt; outputs;\n      for(int i = 0;status.ok(); i++) {\n        char name[100];\n        if (i==0)\n            sprintf(name, \"assign_variables\");\n        else\n            sprintf(name, \"assign_variables_%d\", i);\n\n        status = session-&gt;Run({}, {name}, {}, &amp;outputs);\n      }\n<\/code><\/pre>\n"},{"AnswerId":"36162107","CreationDate":"2016-03-22T18:12:36.677","ParentId":null,"OwnerUserId":"640898","Title":null,"Body":"<p>There is another way of restoring the variables, by calling the <code>save\/restore_all<\/code> operation, that should be present in the graph:<\/p>\n\n<pre><code>std::vector&lt;tensorflow::Tensor&gt; outputs;\nTensor checkpoint_filepath(DT_STRING, TensorShape());\ncheckpoint_filepath.scalar&lt;std::string&gt;()() = \"path to the checkpoint file\";\nstatus = session-&gt;Run( {{ \"save\/Const\", checkpoint_filepath },}, \n                       {}, {\"save\/restore_all\"}, &amp;outputs);\n<\/code><\/pre>\n"}]}
{"QuestionId":34796427,"AnswerCount":0,"Tags":"<theano><pymc><pymc3>","CreationDate":"2016-01-14T17:53:01.527","AcceptedAnswerId":null,"OwnerUserId":5790244.0,"Title":"TypeError: MaxAndArgmax needs a constant axis for arbitrary \/ custom distribution function","Body":"<p>I am trying to define a multivariate, continuous, custom\/arbitrary distribution function that is quite complex and sample from it using the NUTS. I am receiving the following error when I do:    <\/p>\n\n<pre><\/pre>\n\n<p>I have made sure that the output of the custom logp function is a theano variable scalar and, outside of my custom distribution declaration, I am able to successfully compile a theano function that returns the logp and then also the gradient with respect to the multivariate input. My custom distribution function does not use a scan op which seems to play a part in generating the above error. I also receive an infer shape warning as well with regards to a scan op (I can follow up here with that error message).<\/p>\n\n<p>From the research I've done, it seems to me that the problem is possibly coming from the calculation of the hessian by the NUTS, but I'm not sure. I'm inexperienced with pymc3 and theano so any insight here would be greatly appreciated.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Alex<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>I'm now fairly confident that the hessian is causing this error and so this becomes more of a theano issue than a pymc3 one. I was able to test this by evaluating the  function outside of pymc3 entirely, and I received exactly the same error above. Also, akin to this <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/theano-users\/Maxandargmax$20hessian\/theano-users\/Qp9dF0ZFo9k\/uRXYJiX7vsgJ\" rel=\"nofollow\">issue<\/a>, I was using a  in evaluating my likelihood function which caused the above error. However, I replaced min() with sum() to double check and theano failed to compile my function still but didn't throw an error (just blocked)... but only on the gpu. Still keeping the \"fake\" sum() operation in there, the hessian successfully evaluated on my cpu. Anyway, I think there are issues with my likelihood function as well with regards to continuity and convexity so won't be travelling down this path.<\/p>\n","answers":[]}
{"QuestionId":34796596,"AnswerCount":2,"Tags":"<machine-learning><deep-learning><tensorflow>","CreationDate":"2016-01-14T18:02:18.210","AcceptedAnswerId":34799795.0,"OwnerUserId":1978504.0,"Title":"Machine learning: how to identify if there is no object of trained classes in image","Body":"<p>if I train a deep neural network with (say) 10 classes and I feed the network a completely different image, is it reasonable to expect that the output layer's cells will not activate much, so I will know there is no object of the trained classes in the image?<\/p>\n\n<p>My intuition says \"Yes\", but is it so? And what would be the best approach to this?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"34799795","CreationDate":"2016-01-14T21:09:53.540","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>During the supervised training you usually assume that during training you get complete representation of the future types of objects. Typically - these are all labeled instances. In your case - there are also \"noise\" instances, thus there are basically two main approaches:<\/p>\n\n<ul>\n<li>As multi-class neural network has K output neurons, each representing the probability of being a member of a particular class, you can simply condition on these distribution to say that the new object does not belong to any of them. One particular approach is to check if <code>min(p(y|x))&lt;T<\/code> (where p(y|x) is the activation of output neuron) for some threshold <code>T<\/code>. You can either set this value by hand, or through analysis of you \"noise\" instances (with you do have some for training). Simply pass them through your network and compare what value of T gives you best recognition rate<\/li>\n<li>Add another one-class classifier (anomaly detector) before your network - so you will end up with the sequence of two classifiers, first is able to recognize if it is a noise or element of any of your classes (notice, that this can be trained without access to noise samples, see <em>one-class classification<\/em> or <em>anomaly detection<\/em> techniques.<\/li>\n<\/ul>\n\n<p>You could also add another output to the network to represent noise, but this probably will not work well, as you will force your network to generate consistent internal representation for both noise vs data and inter-class decisions.<\/p>\n"},{"AnswerId":"34799900","CreationDate":"2016-01-14T21:17:12.147","ParentId":null,"OwnerUserId":"3167182","Title":null,"Body":"<p>The answer to your question very much depends on your network architecture and the parameters used to train it. If you are trying to protect against false positives, we can typically set an arbitrary threshold value on the relevant output nodes.<\/p>\n\n<p>More generally, learning algorithms mostly take the form of \u201cclosed set\u201d recognition, where all testing classes are known at training time. However, a more realistic scenario for vision applications is \u201copen set\u201d recognition, where an incomplete knowledge of the world is present at training time and unknown classes can be submitted during testing.<\/p>\n\n<p>This is an on-going area of research - please see this <a href=\"http:\/\/www.wjscheirer.com\/projects\/openset-recognition\/\" rel=\"noreferrer\">Open Set Recognition<\/a> web page for plenty of resources on the subject.<\/p>\n"}]}
{"QuestionId":34798385,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-14T19:47:15.240","AcceptedAnswerId":null,"OwnerUserId":1287458.0,"Title":"TensorFlow saver error: resources exhausted","Body":"<p>Does anybody know what might cause an InternalError when I use saver.save(sess, save_path, global_step=step)? I get an error generating my (tmp) checkpoint file and the error says I have \"resources exhausted: <em>temp filename<\/em>\".<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
{"QuestionId":34799850,"AnswerCount":2,"Tags":"<neural-network><tensorflow><tensorboard>","CreationDate":"2016-01-14T21:13:35.610","AcceptedAnswerId":null,"OwnerUserId":3988198.0,"Title":"Tensorboard- superimpose 2 plots","Body":"<p>In tensorboard, I would like to superimpose 2 plots on the same graph (training and validation losses of a neural network).<\/p>\n\n<p>I can see 2 separate plots, but not one plot with 2 superimposed curves. Otherwise, I get one plot in zigzag.<\/p>\n\n<p>How can I do?<\/p>\n","answers":[{"AnswerId":"38718948","CreationDate":"2016-08-02T11:16:51.577","ParentId":null,"OwnerUserId":"4807429","Title":null,"Body":"<p>It is possible to superimpose two plots in Tensorboard. You'll have to satisfy both of the following:<\/p>\n\n<ol>\n<li><p>Create two separate <code>tf.train.SummaryWriter<\/code> objects such that it outputs in two folders.<\/p><\/li>\n<li><p>Create two summaries (e.g. <code>tf.scalar_summary<\/code>) with the <strong>same name<\/strong>.<\/p><\/li>\n<\/ol>\n\n<p>For example to plot training and validation loss:<\/p>\n\n<pre><code># Before training\ntrain_summary = tf.scalar_summary('Loss', train_loss)\nvali_summary = tf.scalar_summary('Loss', vali_loss)\ntrain_writer = tf.train.SummaryWriter('\/tmp\/train'), sess.graph)\nvali_writer = tf.train.SummaryWriter('\/tmp\/vali'), sess.graph)\n\n# And then later\ntrain_writer.add_summary(...)\nvali_writer.add_summary(...)\n<\/code><\/pre>\n"},{"AnswerId":"34800283","CreationDate":"2016-01-14T21:41:20.110","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>If you point tensorboard at the directory containing the tf events for both runs, you should be able to see them.  (if you have them in subdirectories, point it to the parent of both).<\/p>\n"}]}
{"QuestionId":34800715,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-14T22:07:50.587","AcceptedAnswerId":null,"OwnerUserId":5766785.0,"Title":"Tensorflow: custom data load + asynchronous computation","Body":"<p>This is how-to which I believe is missed from TF examples. <\/p>\n\n<p>Task: <\/p>\n\n<ol>\n<li>samples for each class are given in separate dir and thus labels are indirect (i.e. by dir) <\/li>\n<li>decoupled load and computations in TF<\/li>\n<\/ol>\n\n<p>Each separate bit could be found, however I think have them all together in one place will help to save a lot of time for TF beginners (like myself).<\/p>\n\n<p>Lets tackle <strong>1.<\/strong> in my case it is two sets of images:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Now we have all file-names and one-hot labels preloaded. <\/p>\n\n<p>Lets move to the <strong>2.<\/strong> <\/p>\n\n<p>It is based on <a href=\"https:\/\/stackoverflow.com\/questions\/34594198\/how-to-prefetch-data-using-a-custom-python-function-in-tensorflow\/34596212#34596212\" title=\"How to prefetch data using a custom python function in tensorflow\">How to prefetch data using a custom python function in tensorflow<\/a>. In short it has: <\/p>\n\n<ul>\n<li>custom image-reader (replace with yours)<\/li>\n<li>queue <strong>fnl_q<\/strong> with [filename label] which is used by reader to feed<\/li>\n<li>queue <strong>proc_q<\/strong> with [sample label] which is used to feed processing <strong>some_op<\/strong><\/li>\n<li>thread which perform read_op to get [sample label] and enqueue_op to put pair into <strong>proc_q<\/strong>. Thread is controlled by tf.Coordinator<\/li>\n<li><strong>some_op<\/strong> which first get data from <strong>proc_q<\/strong> by dequeue_many() and rest of computation (also could be put in separate thread).<\/li>\n<\/ul>\n\n<p>Notes: <\/p>\n\n<ul>\n<li>feature_read_op and label_read_op are two separate ops. <\/li>\n<li>I use sleep() to slow down and control op - only for test purposes<\/li>\n<li>i have separated \"feeding\" and \"calculation\" parts - in real case just run them in parallel <\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Typical output:\n<\/p>\n\n<pre><\/pre>\n\n<p>All as expected <\/p>\n\n<ul>\n<li>batch of pairs [sample label] are created  <\/li>\n<li>pairs are shuffled <\/li>\n<\/ul>\n\n<p>Only thing left is to apply TF as it is intended to be used by replacing <strong>some_op<\/strong> :)<\/p>\n\n<p><strong>And a question:<\/strong> \none observed problem problem - in case I use  for file-names and  for samples - shuffling doesn't happen. However other way around (as it code above) it does shuffle perfectly.   <\/p>\n\n<p>Any problem with shuffling for<br>\n ?<\/p>\n\n<hr>\n\n<p><strong>ADD: \nThe version with two threads:<\/strong><\/p>\n\n<ul>\n<li>one for re-fill\/update\/change file name queue <\/li>\n<li>second for fill samples to processing queue.<\/li>\n<\/ul>\n\n<p>Also added correct way to stop threads.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[]}
{"QuestionId":34801342,"AnswerCount":7,"Tags":"<python><tensorflow>","CreationDate":"2016-01-14T22:56:24.597","AcceptedAnswerId":34813809.0,"OwnerUserId":3988198.0,"Title":"tensorflow: how to rotate an image for data augmentation?","Body":"<p>In tensorflow, I would like to rotate an image from a random angle, for data augmentation. But I don't find this transformation in the tf.image module. <\/p>\n","answers":[{"AnswerId":"34813809","CreationDate":"2016-01-15T14:55:05.353","ParentId":null,"OwnerUserId":"1978504","Title":null,"Body":"<p><strong>Update<\/strong>: see @astromme's answer below. Tensorflow now supports rotating images natively.<\/p>\n\n<p>What you can do while there is no native method in tensorflow is something like this:<\/p>\n\n<pre><code>from PIL import Image\nsess = tf.InteractiveSession()\n\n# Pass image tensor object to a PIL image\nimage = Image.fromarray(image.eval())\n\n# Use PIL or other library of the sort to rotate\nrotated = Image.Image.rotate(image, degrees)\n\n# Convert rotated image back to tensor\nrotated_tensor = tf.convert_to_tensor(np.array(rotated))\n<\/code><\/pre>\n"},{"AnswerId":"53392066","CreationDate":"2018-11-20T11:30:17.527","ParentId":null,"OwnerUserId":"3766568","Title":null,"Body":"<p>For rotating an image or a batch of images counter-clockwise by multiples of 90 degrees, you can use <code>tf.image.rot90(image,k=1,name=None)<\/code>.<\/p>\n\n<p><code>k<\/code> denotes the number of 90 degrees rotations you want to make.<\/p>\n\n<p>In case of a single image, <code>image<\/code> is a <code>3-D Tensor of shape [height, width, channels]<\/code> and in case of a batch of images, <code>image<\/code> is a <code>4-D Tensor of shape [batch, height, width, channels]<\/code><\/p>\n"},{"AnswerId":"57476195","CreationDate":"2019-08-13T10:50:13.440","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>for tensorflow 2.0:<\/p>\n\n<pre><code>import tensorflow_addons as tfa\ntfa.image.transform_ops.rotate(image, radian)\n<\/code><\/pre>\n"},{"AnswerId":"43793868","CreationDate":"2017-05-04T22:42:53.113","ParentId":null,"OwnerUserId":"407650","Title":null,"Body":"<p>Here's the <a href=\"https:\/\/stackoverflow.com\/a\/40483687\/407650\">@zimmermc<\/a> answer updated to Tensorflow v0.12<\/p>\n\n<p>Changes:<\/p>\n\n<ul>\n<li><code>pack()<\/code> is now <code>stack()<\/code><\/li>\n<li><p>order of <code>split<\/code> parameters reversed<\/p>\n\n<pre><code>def rotate_image_tensor(image, angle, mode='white'):\n    \"\"\"\n    Rotates a 3D tensor (HWD), which represents an image by given radian angle.\n\n    New image has the same size as the input image.\n\n    mode controls what happens to border pixels.\n    mode = 'black' results in black bars (value 0 in unknown areas)\n    mode = 'white' results in value 255 in unknown areas\n    mode = 'ones' results in value 1 in unknown areas\n    mode = 'repeat' keeps repeating the closest pixel known\n    \"\"\"\n    s = image.get_shape().as_list()\n    assert len(s) == 3, \"Input needs to be 3D.\"\n    assert (mode == 'repeat') or (mode == 'black') or (mode == 'white') or (mode == 'ones'), \"Unknown boundary mode.\"\n    image_center = [np.floor(x\/2) for x in s]\n\n    # Coordinates of new image\n    coord1 = tf.range(s[0])\n    coord2 = tf.range(s[1])\n\n    # Create vectors of those coordinates in order to vectorize the image\n    coord1_vec = tf.tile(coord1, [s[1]])\n\n    coord2_vec_unordered = tf.tile(coord2, [s[0]])\n    coord2_vec_unordered = tf.reshape(coord2_vec_unordered, [s[0], s[1]])\n    coord2_vec = tf.reshape(tf.transpose(coord2_vec_unordered, [1, 0]), [-1])\n\n    # center coordinates since rotation center is supposed to be in the image center\n    coord1_vec_centered = coord1_vec - image_center[0]\n    coord2_vec_centered = coord2_vec - image_center[1]\n\n    coord_new_centered = tf.cast(tf.stack([coord1_vec_centered, coord2_vec_centered]), tf.float32)\n\n    # Perform backward transformation of the image coordinates\n    rot_mat_inv = tf.dynamic_stitch([[0], [1], [2], [3]], [tf.cos(angle), tf.sin(angle), -tf.sin(angle), tf.cos(angle)])\n    rot_mat_inv = tf.reshape(rot_mat_inv, shape=[2, 2])\n    coord_old_centered = tf.matmul(rot_mat_inv, coord_new_centered)\n\n    # Find nearest neighbor in old image\n    coord1_old_nn = tf.cast(tf.round(coord_old_centered[0, :] + image_center[0]), tf.int32)\n    coord2_old_nn = tf.cast(tf.round(coord_old_centered[1, :] + image_center[1]), tf.int32)\n\n    # Clip values to stay inside image coordinates\n    if mode == 'repeat':\n        coord_old1_clipped = tf.minimum(tf.maximum(coord1_old_nn, 0), s[0]-1)\n        coord_old2_clipped = tf.minimum(tf.maximum(coord2_old_nn, 0), s[1]-1)\n    else:\n        outside_ind1 = tf.logical_or(tf.greater(coord1_old_nn, s[0]-1), tf.less(coord1_old_nn, 0))\n        outside_ind2 = tf.logical_or(tf.greater(coord2_old_nn, s[1]-1), tf.less(coord2_old_nn, 0))\n        outside_ind = tf.logical_or(outside_ind1, outside_ind2)\n\n        coord_old1_clipped = tf.boolean_mask(coord1_old_nn, tf.logical_not(outside_ind))\n        coord_old2_clipped = tf.boolean_mask(coord2_old_nn, tf.logical_not(outside_ind))\n\n        coord1_vec = tf.boolean_mask(coord1_vec, tf.logical_not(outside_ind))\n        coord2_vec = tf.boolean_mask(coord2_vec, tf.logical_not(outside_ind))\n\n    coord_old_clipped = tf.cast(tf.transpose(tf.stack([coord_old1_clipped, coord_old2_clipped]), [1, 0]), tf.int32)\n\n    # Coordinates of the new image\n    coord_new = tf.transpose(tf.cast(tf.stack([coord1_vec, coord2_vec]), tf.int32), [1, 0])\n\n    image_channel_list = tf.split(image, s[2], 2)\n\n    image_rotated_channel_list = list()\n    for image_channel in image_channel_list:\n        image_chan_new_values = tf.gather_nd(tf.squeeze(image_channel), coord_old_clipped)\n\n        if (mode == 'black') or (mode == 'repeat'):\n            background_color = 0\n        elif mode == 'ones':\n            background_color = 1\n        elif mode == 'white':\n            background_color = 255\n\n        image_rotated_channel_list.append(tf.sparse_to_dense(coord_new, [s[0], s[1]], image_chan_new_values,\n                                                             background_color, validate_indices=False))\n\n    image_rotated = tf.transpose(tf.stack(image_rotated_channel_list), [1, 2, 0])\n\n    return image_rotated\n<\/code><\/pre><\/li>\n<\/ul>\n"},{"AnswerId":"45663250","CreationDate":"2017-08-13T17:52:41.597","ParentId":null,"OwnerUserId":"1004331","Title":null,"Body":"<p>This can be done in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/rotate\" rel=\"noreferrer\">tensorflow now<\/a>:<\/p>\n\n<pre><code>tf.contrib.image.rotate(images, degrees * math.pi \/ 180, interpolation='BILINEAR')\n<\/code><\/pre>\n"},{"AnswerId":"50906665","CreationDate":"2018-06-18T09:44:27.437","ParentId":null,"OwnerUserId":"6772741","Title":null,"Body":"<h2>Rotation and cropping in TensorFlow<\/h2>\n\n<p>I personally needed image rotation and cropping out black borders functions in TensorFlow as below.\n<a href=\"https:\/\/i.stack.imgur.com\/Guj8U.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Guj8U.jpg\" alt=\"Example\"><\/a>\nAnd I could implement this function as below.<\/p>\n\n<pre><code>def _rotate_and_crop(image, output_height, output_width, rotation_degree, do_crop):\n    \"\"\"Rotate the given image with the given rotation degree and crop for the black edges if necessary\n    Args:\n        image: A `Tensor` representing an image of arbitrary size.\n        output_height: The height of the image after preprocessing.\n        output_width: The width of the image after preprocessing.\n        rotation_degree: The degree of rotation on the image.\n        do_crop: Do cropping if it is True.\n    Returns:\n        A rotated image.\n    \"\"\"\n\n    # Rotate the given image with the given rotation degree\n    if rotation_degree != 0:\n        image = tf.contrib.image.rotate(image, math.radians(rotation_degree), interpolation='BILINEAR')\n\n        # Center crop to ommit black noise on the edges\n        if do_crop == True:\n            lrr_width, lrr_height = _largest_rotated_rect(output_height, output_width, math.radians(rotation_degree))\n            resized_image = tf.image.central_crop(image, float(lrr_height)\/output_height)    \n            image = tf.image.resize_images(resized_image, [output_height, output_width], method=tf.image.ResizeMethod.BILINEAR, align_corners=False)\n\n    return image\n\ndef _largest_rotated_rect(w, h, angle):\n    \"\"\"\n    Given a rectangle of size wxh that has been rotated by 'angle' (in\n    radians), computes the width and height of the largest possible\n    axis-aligned rectangle within the rotated rectangle.\n    Original JS code by 'Andri' and Magnus Hoff from Stack Overflow\n    Converted to Python by Aaron Snoswell\n    Source: http:\/\/stackoverflow.com\/questions\/16702966\/rotate-image-and-crop-out-black-borders\n    \"\"\"\n\n    quadrant = int(math.floor(angle \/ (math.pi \/ 2))) &amp; 3\n    sign_alpha = angle if ((quadrant &amp; 1) == 0) else math.pi - angle\n    alpha = (sign_alpha % math.pi + math.pi) % math.pi\n\n    bb_w = w * math.cos(alpha) + h * math.sin(alpha)\n    bb_h = w * math.sin(alpha) + h * math.cos(alpha)\n\n    gamma = math.atan2(bb_w, bb_w) if (w &lt; h) else math.atan2(bb_w, bb_w)\n\n    delta = math.pi - alpha - gamma\n\n    length = h if (w &lt; h) else w\n\n    d = length * math.cos(alpha)\n    a = d * math.sin(alpha) \/ math.sin(delta)\n\n    y = a * math.cos(gamma)\n    x = y * math.tan(gamma)\n\n    return (\n        bb_w - 2 * x,\n        bb_h - 2 * y\n    )\n<\/code><\/pre>\n\n<p>If you need further implementation of example and visualization in TensorFlow, you can use <a href=\"https:\/\/github.com\/kobiso\/Image-Rotation-and-Cropping-tensorflow\" rel=\"nofollow noreferrer\">this repository<\/a>.\nI hope this could be helpful to other people.<\/p>\n"},{"AnswerId":"40483687","CreationDate":"2016-11-08T09:48:58.677","ParentId":null,"OwnerUserId":"6375313","Title":null,"Body":"<p>Because I wanted to be able to rotate tensors I came up with the following piece of code, which rotates a [height, width, depth] tensor by a given angle:<\/p>\n\n<pre><code>def rotate_image_tensor(image, angle, mode='black'):\n    \"\"\"\n    Rotates a 3D tensor (HWD), which represents an image by given radian angle.\n\n    New image has the same size as the input image.\n\n    mode controls what happens to border pixels.\n    mode = 'black' results in black bars (value 0 in unknown areas)\n    mode = 'white' results in value 255 in unknown areas\n    mode = 'ones' results in value 1 in unknown areas\n    mode = 'repeat' keeps repeating the closest pixel known\n    \"\"\"\n    s = image.get_shape().as_list()\n    assert len(s) == 3, \"Input needs to be 3D.\"\n    assert (mode == 'repeat') or (mode == 'black') or (mode == 'white') or (mode == 'ones'), \"Unknown boundary mode.\"\n    image_center = [np.floor(x\/2) for x in s]\n\n    # Coordinates of new image\n    coord1 = tf.range(s[0])\n    coord2 = tf.range(s[1])\n\n    # Create vectors of those coordinates in order to vectorize the image\n    coord1_vec = tf.tile(coord1, [s[1]])\n\n    coord2_vec_unordered = tf.tile(coord2, [s[0]])\n    coord2_vec_unordered = tf.reshape(coord2_vec_unordered, [s[0], s[1]])\n    coord2_vec = tf.reshape(tf.transpose(coord2_vec_unordered, [1, 0]), [-1])\n\n    # center coordinates since rotation center is supposed to be in the image center\n    coord1_vec_centered = coord1_vec - image_center[0]\n    coord2_vec_centered = coord2_vec - image_center[1]\n\n    coord_new_centered = tf.cast(tf.pack([coord1_vec_centered, coord2_vec_centered]), tf.float32)\n\n    # Perform backward transformation of the image coordinates\n    rot_mat_inv = tf.dynamic_stitch([[0], [1], [2], [3]], [tf.cos(angle), tf.sin(angle), -tf.sin(angle), tf.cos(angle)])\n    rot_mat_inv = tf.reshape(rot_mat_inv, shape=[2, 2])\n    coord_old_centered = tf.matmul(rot_mat_inv, coord_new_centered)\n\n    # Find nearest neighbor in old image\n    coord1_old_nn = tf.cast(tf.round(coord_old_centered[0, :] + image_center[0]), tf.int32)\n    coord2_old_nn = tf.cast(tf.round(coord_old_centered[1, :] + image_center[1]), tf.int32)\n\n    # Clip values to stay inside image coordinates\n    if mode == 'repeat':\n        coord_old1_clipped = tf.minimum(tf.maximum(coord1_old_nn, 0), s[0]-1)\n        coord_old2_clipped = tf.minimum(tf.maximum(coord2_old_nn, 0), s[1]-1)\n    else:\n        outside_ind1 = tf.logical_or(tf.greater(coord1_old_nn, s[0]-1), tf.less(coord1_old_nn, 0))\n        outside_ind2 = tf.logical_or(tf.greater(coord2_old_nn, s[1]-1), tf.less(coord2_old_nn, 0))\n        outside_ind = tf.logical_or(outside_ind1, outside_ind2)\n\n        coord_old1_clipped = tf.boolean_mask(coord1_old_nn, tf.logical_not(outside_ind))\n        coord_old2_clipped = tf.boolean_mask(coord2_old_nn, tf.logical_not(outside_ind))\n\n        coord1_vec = tf.boolean_mask(coord1_vec, tf.logical_not(outside_ind))\n        coord2_vec = tf.boolean_mask(coord2_vec, tf.logical_not(outside_ind))\n\n    coord_old_clipped = tf.cast(tf.transpose(tf.pack([coord_old1_clipped, coord_old2_clipped]), [1, 0]), tf.int32)\n\n    # Coordinates of the new image\n    coord_new = tf.transpose(tf.cast(tf.pack([coord1_vec, coord2_vec]), tf.int32), [1, 0])\n\n    image_channel_list = tf.split(2, s[2], image)\n\n    image_rotated_channel_list = list()\n    for image_channel in image_channel_list:\n        image_chan_new_values = tf.gather_nd(tf.squeeze(image_channel), coord_old_clipped)\n\n        if (mode == 'black') or (mode == 'repeat'):\n            background_color = 0\n        elif mode == 'ones':\n            background_color = 1\n        elif mode == 'white':\n            background_color = 255\n\n        image_rotated_channel_list.append(tf.sparse_to_dense(coord_new, [s[0], s[1]], image_chan_new_values,\n                                                             background_color, validate_indices=False))\n\n    image_rotated = tf.transpose(tf.pack(image_rotated_channel_list), [1, 2, 0])\n\n    return image_rotated\n<\/code><\/pre>\n"}]}
{"QuestionId":34804274,"AnswerCount":0,"Tags":"<python><theano><openblas>","CreationDate":"2016-01-15T04:11:53.333","AcceptedAnswerId":null,"OwnerUserId":4086036.0,"Title":"CPU (with direct Theano binding to blas) is slower?","Body":"<p>I have followed the steps <a href=\"https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits\" rel=\"nofollow\">https:\/\/www.kaggle.com\/c\/otto-group-product-classification-challenge\/forums\/t\/13973\/a-few-tips-to-install-theano-on-windows-64-bits<\/a>  to install theano on my windows machine, and run theano.misc.check_blas.test() to test blas speed, it runs around 10s. <\/p>\n\n<pre><\/pre>\n\n<p>However, if I take out these lines from my .theanorc.txt file, <\/p>\n\n<pre><\/pre>\n\n<p>the result will be (shown only the last output line): <\/p>\n\n<pre><\/pre>\n\n<p>Why is direct Theano binding to blas much slower than without direct binding?? Is the blas I was using wrong? <\/p>\n\n<p>(I followed my above link's steps to download and use the openblas - OpenBLAS-v0.2.14-Win64-int32.zip (which can be downloaded from <a href=\"http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/OpenBLAS-v0.2.14-Win64-int32.zip\/download\" rel=\"nofollow\">http:\/\/sourceforge.net\/projects\/openblas\/files\/v0.2.14\/OpenBLAS-v0.2.14-Win64-int32.zip\/download<\/a>), and saved to my local C:\\\\openblas)<\/p>\n\n<p>I have also done a test using below scripts: <\/p>\n\n<pre><\/pre>\n\n<p>And the result (binding to openblas) is slower than NP as well:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34805329,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-15T06:16:06.877","AcceptedAnswerId":34820266.0,"OwnerUserId":5793199.0,"Title":"How to get the gradients of loss with respect to activations in Tensorflow","Body":"<p>In the  example, the gradients of loss with respect to parameters can be computed as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any way to get the gradients of loss with respect to activations (not the parameters)\uff0c and watch them in Tensorboard?<\/p>\n","answers":[{"AnswerId":"34820266","CreationDate":"2016-01-15T21:36:54.077","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#gradients\" rel=\"noreferrer\"><code>tf.gradients()<\/code><\/a> function to compute the gradient of any scalar tensor with respect to any other tensor (assuming the gradients are defined for all of the ops between those two tensors):<\/p>\n\n<pre><code>activations = ...\nloss = f(..., activations)  # `loss` is some function of `activations`.\n\ngrad_wrt_activations, = tf.gradients(loss, [activation])\n<\/code><\/pre>\n\n<p>Visualizing this in TensorBoard is tricky in general, since <code>grad_wrt_activation<\/code> is (typically) a tensor with the same shape as <code>activation<\/code>. Adding a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#histogram_summary\" rel=\"noreferrer\"><code>tf.histogram_summary()<\/code><\/a> op is probably the easiest way to visualize<\/p>\n\n<pre><code># Adds a histogram of `grad_wrt_activations` to the graph, which will be logged\n# with the other summaries, and shown in TensorBoard.\ntf.histogram_summary(\"Activation gradient\", grad_wrt_activations)\n<\/code><\/pre>\n"}]}
{"QuestionId":34807319,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><caffe>","CreationDate":"2016-01-15T08:51:18.377","AcceptedAnswerId":35030637.0,"OwnerUserId":654005.0,"Title":"Input data amount for Caffe","Body":"<p>I want to use Caffe and the googlenet structure coming with Caffe to train a model based on my own image data. \nI have 14 categories for classification. But I do have only around 250 images for training and 80 for testing. Is this enough? Are there means to find out how many images I need per class? <\/p>\n","answers":[{"AnswerId":"35030637","CreationDate":"2016-01-27T06:56:57.303","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>Solution 1:\nJust finetune the top layer since you only have such few data. By this way, you can think the network as a feature extractor and you just train a classifier on top this features.<\/p>\n\n<p>Solution 2:\nTry aggressive data augmentation. For example you can try random translation, scaling, rotation of your data. In this way, you can get a lot of images from one training image.<\/p>\n\n<p>Solution 3:\nThe most effective way is to try to get more real data. Data is very important for deep learning. As a rule of thumb, at least 1000 images for one class.<\/p>\n"}]}
{"QuestionId":34809795,"AnswerCount":3,"Tags":"<python-2.7><neural-network><deep-learning><tensorflow><google-image-search>","CreationDate":"2016-01-15T11:13:15.483","AcceptedAnswerId":34852037.0,"OwnerUserId":4758347.0,"Title":"Tensorflow return similar images","Body":"<p>I want to use Google's Tensorflow to return similar images to an input image.<\/p>\n\n<p>I have installed Tensorflow from <a href=\"http:\/\/www.tensorflow.org\" rel=\"noreferrer\">http:\/\/www.tensorflow.org<\/a> (using PIP installation - pip and python 2.7) on Ubuntu14.04 on a virtual machine CPU.<\/p>\n\n<p>I have downloaded the trained model Inception-V3 (inception-2015-12-05.tgz) from <a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz\" rel=\"noreferrer\">http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz<\/a> that is trained on ImageNet Large Visual Recognition Challenge using the data from 2012, but I think it has both the Neural network and the classifier inside it (as the task there was to predict the category). I have also downloaded the file classify_image.py that classifies an image in 1 of the 1000 classes in the model.<\/p>\n\n<p>So I have a random image image.jpg that I an running to test the model. when I run the command:<\/p>\n\n<pre><\/pre>\n\n<p>I get the below output: (Classification is done using softmax)<\/p>\n\n<pre><\/pre>\n\n<p>Now, the task at hand is to find images that are similar to the input image (image.jpg) out of a database of 60,000 images (jpg format, and kept in a folder at \/home\/amit\/images). I believe this can be done by removing the final classification layer from the inception-v3 model, and using the feature set of the input image to find cosine distance from the feature set all the 60,000 images, and we can return the images having less distance (cos 0 = 1)<\/p>\n\n<p>Please suggest me the way forward for this problem and how do I do this using Python API.<\/p>\n","answers":[{"AnswerId":"41384475","CreationDate":"2016-12-29T17:27:53.233","ParentId":null,"OwnerUserId":"2425370","Title":null,"Body":"<p>Tensorflow now has a nice tutorial on how to get the activations before the final layer and retrain a new classification layer with different categories:\n<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><\/p>\n\n<p>The example code:\n<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>In your case, yes, you can get the activations from pool_3 the layer below the softmax layer (or the so-called bottlenecks) and send them to other operations as input:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/s76Fr.png\" alt=\"retrain imagenet\"><\/p>\n\n<p>Finally, about finding similar images, I don't think imagenet's bottleneck activations are very pertinent representation for image search. You could consider to use an autoencoder network with direct image inputs. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mi9z6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mi9z6.png\" alt=\"autoencoder\"><\/a><br>\n<sub>(source: <a href=\"https:\/\/deeplearning4j.org\/img\/deep_autoencoder.png\" rel=\"nofollow noreferrer\">deeplearning4j.org<\/a>)<\/sub>  <\/p>\n"},{"AnswerId":"35261747","CreationDate":"2016-02-08T02:54:04.510","ParentId":null,"OwnerUserId":"2138200","Title":null,"Body":"<p>Your problem sounds similar to this <a href=\"https:\/\/github.com\/AKSHAYUBHAT\/VisualSearchServer\" rel=\"nofollow\">visual search project<\/a><\/p>\n"},{"AnswerId":"34852037","CreationDate":"2016-01-18T10:08:27.260","ParentId":null,"OwnerUserId":"4758347","Title":null,"Body":"<p>I think I found an answer to my question:<\/p>\n\n<p>In the file classify_image.py that classifies the image using the pre trained model (NN + classifier), I made the below mentioned changes (statements with #ADDED written next to them): <\/p>\n\n<pre><code>def run_inference_on_image(image):\n  \"\"\"Runs inference on an image.\n  Args:\n    image: Image file name.\n  Returns:\n    Nothing\n  \"\"\"\n  if not gfile.Exists(image):\n    tf.logging.fatal('File does not exist %s', image)\n  image_data = gfile.FastGFile(image, 'rb').read()\n\n  # Creates graph from saved GraphDef.\n  create_graph()\n\nwith tf.Session() as sess:\n # Some useful tensors:\n # 'softmax:0': A tensor containing the normalized prediction across\n #   1000 labels.\n # 'pool_3:0': A tensor containing the next-to-last layer containing 2048\n #   float description of the image.\n # 'DecodeJpeg\/contents:0': A tensor containing a string providing JPEG\n #   encoding of the image.\n # Runs the softmax tensor by feeding the image_data as input to the graph.\n softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')\n feature_tensor = sess.graph.get_tensor_by_name('pool_3:0') #ADDED\n predictions = sess.run(softmax_tensor,\n                        {'DecodeJpeg\/contents:0': image_data})\n predictions = np.squeeze(predictions)\n feature_set = sess.run(feature_tensor,\n                        {'DecodeJpeg\/contents:0': image_data}) #ADDED\n feature_set = np.squeeze(feature_set) #ADDED\n print(feature_set) #ADDED\n # Creates node ID --&gt; English string lookup.\n node_lookup = NodeLookup()\n\n 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>I ran the pool_3:0 tensor by feeding in the image_data to it. Please let me know if I am doing a mistake. If this is correct, I believe we can use this tensor for further calculations.<\/p>\n"}]}
{"QuestionId":34809817,"AnswerCount":1,"Tags":"<java><spring><spring-cache><jcache><caffeine>","CreationDate":"2016-01-15T11:14:17.187","AcceptedAnswerId":34810859.0,"OwnerUserId":3424394.0,"Title":"How to create a Jcache in Spring Java config?","Body":"<p>I have a problem setting up a jcache with <a href=\"http:\/\/docs.spring.io\/spring\/docs\/current\/spring-framework-reference\/html\/cache.html\" rel=\"nofollow noreferrer\">spring cache abstraction<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>I'm using <a href=\"https:\/\/github.com\/ben-manes\/caffeine\" rel=\"nofollow noreferrer\">Caffein<\/a> as Jcache. I just cannot understand what i'm doing wrong. Would you explain how to do it properly?<\/p>\n\n<p>What i'm getting is  at line <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34810859","CreationDate":"2016-01-15T12:13:27.860","ParentId":null,"OwnerUserId":"2797275","Title":null,"Body":"<p>Read more about <code>Caffeine<\/code> configuration. Its <code>JCache<\/code> adapter uses default specification (<code>JCache<\/code>) settings which states that entries never expires and are stored by value (copied when put\/retrieved from cache). (<a href=\"https:\/\/github.com\/ben-manes\/caffeine\/wiki\/JCache\" rel=\"nofollow\">reference here<\/a>).<\/p>\n\n<p>When cache is set to copying instances, you should choose adequate <code>Copier<\/code> which will handle that. So you can just write:<\/p>\n\n<pre><code>caffeineConfiguration.setCopierFactory(JavaSerializationCopier::new);\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>caffeineConfiguration.setCopierFactory(Copier::identity);\n<\/code><\/pre>\n\n<p>It depends on\u00a0if you want to be safe in case of cache key\/value mutations.<\/p>\n\n<p>Nevertheless, <strong>the recommended option is<\/strong> to use <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/blob\/master\/jcache\/src\/main\/resources\/reference.conf\" rel=\"nofollow\">default settings<\/a> where <code>store-by-value<\/code> option is disabled and copier is necessary:<\/p>\n\n<pre><code>Config config = ConfigFactory.load();\nCaffeineConfiguration&lt;String, List&lt;Product&gt;&gt; caffeineConfiguration = TypesafeConfigurator.defaults(config);\ncaffeineConfiguration.setExpiryPolicyFactory(factoryOf(new AccessedExpiryPolicy(new Duration(\n            TimeUnit.MINUTES, 60))));\n<\/code><\/pre>\n\n<p>Alternatively, you can just disable <code>store-by-value<\/code> setting by:<\/p>\n\n<pre><code>caffeineConfiguration.setStoreByValue(false);\n<\/code><\/pre>\n"}]}
{"QuestionId":34811799,"AnswerCount":1,"Tags":"<python><hdf5><caffe><h5py><pycaffe>","CreationDate":"2016-01-15T13:08:07.577","AcceptedAnswerId":null,"OwnerUserId":5790263.0,"Title":"Python layer can't read hdf5 file in caffe framework","Body":"<p>I write python layer for caffe that can read hdf5 with some manipulation needs for me. But I have some issue when open and read hdf5 file in python setup method. Here it is.\nWhen I used tables with the\ncode in setup:<\/p>\n\n<pre><\/pre>\n\n<p>I have this error when run net:<\/p>\n\n<pre><\/pre>\n\n<p>When I used h5py<\/p>\n\n<pre><\/pre>\n\n<p>I have the same error:<\/p>\n\n<pre><\/pre>\n\n<p>When I used deepdish:<\/p>\n\n<pre><\/pre>\n\n<p>I have error too:<\/p>\n\n<pre><\/pre>\n\n<p>But when I read simply *.txt file, all is ok. Also I can read this file from console and use it in hdf5 layer in caffe. Please help me, how I can read hdf5 file from python layer?<\/p>\n","answers":[{"AnswerId":"40409201","CreationDate":"2016-11-03T18:53:36.977","ParentId":null,"OwnerUserId":"3333554","Title":null,"Body":"<p>Try to install another version of h5py. I just solved it with:<\/p>\n\n<p><code>pip install h5py==prev_version<\/code><\/p>\n\n<p>I guess it has something to do with the linking but it would be interesting to know which is the exact cause for this problem.<\/p>\n"}]}
{"QuestionId":34812542,"AnswerCount":1,"Tags":"<python><neural-network><mathematical-optimization><theano><gradient-descent>","CreationDate":"2016-01-15T13:49:36.277","AcceptedAnswerId":null,"OwnerUserId":5635159.0,"Title":"Why does this really weird bug with theano.function updates happen in my program?","Body":"<p>I found my model went wrong when I use adadelta which is implemented like this:<\/p>\n\n<pre><\/pre>\n\n<p>after several epochs of updates, the params becomes nan. But if I use only simple gradient descent which is like this:<\/p>\n\n<pre><\/pre>\n\n<p>it didn't go wrong.<\/p>\n\n<p>The most weird thing is if I add grads to the output of the theano.function which is like this:<\/p>\n\n<pre><\/pre>\n\n<p>adadelta doesn't go wrong either.<\/p>\n\n<p>And when I set the optimizer is None, the problem also does not exist<\/p>\n\n<p>What is the problem? Is the grads in adadelta collected by garbage collector?<\/p>\n","answers":[{"AnswerId":"41224436","CreationDate":"2016-12-19T14:10:22.090","ParentId":null,"OwnerUserId":"5890647","Title":null,"Body":"<p>I think there might be a bug with the 'updates' input of a theano.function. <\/p>\n\n<p>The problem seems to appear whenever the function to compute the updated value of a theano variable x is defined independently of x. On the second iteration of a theano function, even if the same values are given to the shared variable, it can screw up. Sigh :(<\/p>\n"}]}
{"QuestionId":34814622,"AnswerCount":1,"Tags":"<java><spring><spring-cache><jcache><caffeine>","CreationDate":"2016-01-15T15:36:43.113","AcceptedAnswerId":34847729.0,"OwnerUserId":3424394.0,"Title":"Why I have cache misses in Service using Spring Cache","Body":"<p>I have configured my cache as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Also I have the  that uses it in following way:<\/p>\n\n<pre><\/pre>\n\n<p>So I have the next behavior.<\/p>\n\n<ol>\n<li>When I'm calling the service method first time it takes 5 seconds\nthen doing some work as expected.<\/li>\n<li>Calling method the second time with same parameters - > caching works -> returns result immediately<\/li>\n<li>Calling the third time with same parameters again results in <\/li>\n<\/ol>\n\n<p>And all over again.<\/p>\n\n<p>How to solve this ? Is that the issue about proxying ? What did I miss ?<\/p>\n","answers":[{"AnswerId":"34847729","CreationDate":"2016-01-18T05:18:31.940","ParentId":null,"OwnerUserId":"19450","Title":null,"Body":"<p>As discussed in the comments, this was a bug in the JCache adapter. Thank you for letting me know about this problem. I released version <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/releases\/tag\/v2.1.0\" rel=\"nofollow noreferrer\">2.1.0<\/a> which includes this fix. That release also includes friendlier initial settings for <code>CaffeineConfiguration<\/code> which you identified in another <a href=\"https:\/\/stackoverflow.com\/q\/34809817\/19450\">post<\/a>.<\/p>\n\n<p>While the core library is heavily tested, the JCache adapters relied too heavily on the JSR's TCK (test compatibility kit). Unfortunately that test suite isn't very effective, so I added tests to help avoid these types of mistakes in the future.<\/p>\n\n<p>This issue only occurred in JCache because its version of expiration is not supported by Caffeine's core library. Caffeine prefers to use an O(1) design that eagerly cleans up expired entries by using fixed durations. JCache uses per-entry lazy expiration and the specification authors assume a capacity constraint is used to eventually discard expired entries. I added a <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/wiki\/JCache\" rel=\"nofollow noreferrer\">warning<\/a> to the documentation about this feature, as it could be error prone. While none of the other JCache implementations go beyond this, a pending task is to decide on a mechanism to help mitigate this JCache design flaw.<\/p>\n\n<p>Thanks again for reporting this issue. As always, feel free to reach out if you have any other issues or feedback to share.<\/p>\n"}]}
{"QuestionId":34817456,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-15T18:27:18.887","AcceptedAnswerId":34817932.0,"OwnerUserId":4975126.0,"Title":"Using Tensors in control flow and conditional statements","Body":"<p>I'm interested how I can use a Tensor (without a session) in control flow statements.<\/p>\n\n<hr>\n\n<p>One example of what I mean by this is:<\/p>\n\n<pre><\/pre>\n\n<p>Can I pass numLoops (a tensor) to xrange()? If not, is there another way to do this?<\/p>\n\n<hr>\n\n<p>Another example would be:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>My Question:<\/strong> Can Tensors (without a specific session) be used in the ways outlined above?<\/p>\n\n<p>If I had a session I could simply evaluate those single-element tensors and then use them. In the event that I don't know the specific value until runtime, there must be a way to use the values.<\/p>\n\n<p><strong>Problems I could foresee:<\/strong> maybe the loop would make generating the graph impossible as you don't know how many times the contained operations will be executed? But it would seem trivial to store the loop's operations and simply apply them the correct number of times, at runtime.<\/p>\n\n<p>If anything lacks clarity, let me know and I'll provide more detail.<\/p>\n","answers":[{"AnswerId":"34817932","CreationDate":"2016-01-15T19:00:44.903","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow contains experimental support for in-graph control flow in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py\" rel=\"nofollow\">the <code>tensorflow.python.ops.control_flow_ops<\/code> module<\/a>. (<strong>N.B.<\/strong> The interfaces to these ops are subject to change! Use at your own risk!)<\/p>\n\n<hr>\n\n<p>Conditional support (your second case) will appear in the next release, and was recently added to the public API (if you build from source). It's currently being used in some of the libraries that ship with TensorFlow, such as the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7ddf1ebfdd120409c19bd6727c18a3f283ada42c\/tensorflow\/python\/ops\/rnn.py#L109\" rel=\"nofollow\">RNN cell<\/a> (which use it for early stopping when the sequence length is known).<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7ddf1ebfdd120409c19bd6727c18a3f283ada42c\/tensorflow\/python\/ops\/control_flow_ops.py#L599\" rel=\"nofollow\"><code>tf.cond()<\/code><\/a> operator takes a boolean <code>Tensor<\/code>, which acts as the predicate, and two lambda expressions; and returns one or more values. <\/p>\n\n<p>Your program would look something like:<\/p>\n\n<pre><code>if_true = lambda: ...  # Value to return if `myCondition` is true\nif_false = lambda: ...  # Value to return if `myCondition` is false\n\nresult = tf.cond(myCondition, if_true, if_false)\n<\/code><\/pre>\n\n<hr>\n\n<p>Iteration (your first case) can be handled using the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7ddf1ebfdd120409c19bd6727c18a3f283ada42c\/tensorflow\/python\/ops\/control_flow_ops.py#L1071\" rel=\"nofollow\"><code>While()<\/code><\/a> higher-order operator. Essentially, you would write your program as (in pseudocode):<\/p>\n\n<pre><code>i = 0\nwhile i &lt; tf.rank(myTensor):\n  # ...\n<\/code><\/pre>\n\n<p>...which would approximately be represented as:<\/p>\n\n<pre><code>from tensorflow.python.ops import control_flow_ops as cfo\n\ni = tf.constant(0)\ncondition = lambda i: tf.less(i, tf.rank(myTensor))\nbody = lambda x: ...\ninital_accumulator = ...  # will be bound to `x` in first iteration of `body`.\n\n# `result` will contain the value returned from the final iteration of `body`.\nresult = cfo.While(condition, body, [initial_accumulator])\n<\/code><\/pre>\n\n<p>Note that the <code>While()<\/code> operator requires you to specify initial values for all of the loop variables (just <code>initial_accumulator<\/code> in this case). Loop constants will be captured as in a normal Python lambda.<\/p>\n"}]}
{"QuestionId":34817572,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-01-15T18:35:33.437","AcceptedAnswerId":34821026.0,"OwnerUserId":2807033.0,"Title":"How to train end-to-end CNN?","Body":"<p>While using caffe as <\/p>\n\n<blockquote>\n  <p>.\/build\/tools\/caffe train --solver=models\/Handmade\/solver.prototxt<\/p>\n<\/blockquote>\n\n<p>caffe also gets into \"phase: TEST\" but I have no test data. I only want to train the parameters on my training data, so I haven't used \"phase: Test\" in \"train.prototxt\", which causes error. What should I do?<\/p>\n","answers":[{"AnswerId":"34821026","CreationDate":"2016-01-15T22:36:01.283","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>I don't know if you can completely omit the test phase but it's possible to train your model without needing a separate test set. It's also possible to prevent the solver from ever switching to the test phase.<\/p>\n\n<p>Reuse your training data for the test phase. You can do so by duplicating your data layer and specifying it for the test phase.<\/p>\n\n<p>To limit computations to the training phase only increase the value of <code>test_interval<\/code> in your solver definition to a number larger than your training set or, better, larger than <code>max_iter<\/code>. This prevents the solver from ever switching to the test phase.<\/p>\n\n<p>I find it a bit odd to train a model without wanting to know how to does on a separate set of data points.<\/p>\n"}]}
{"QuestionId":34819588,"AnswerCount":2,"Tags":"<matlab><input><hdf5><caffe><conv-neural-network>","CreationDate":"2016-01-15T20:46:52.813","AcceptedAnswerId":34819919.0,"OwnerUserId":3698878.0,"Title":"Caffe using hdf5 layer and imagedata input layer together in train_val.proto","Body":"<p>I am currently using hdf5 data layer for reading as input the images(around 30000) that I have and some metadata about the images. <\/p>\n\n<p>I am unable to do the crop-flip data augmentation, as since when I just use the centre crops, storing data of about 1500 images only leads to around 1.5 GB h5 file, the total the size of the hdf5 dataset(30 h5 files) becomes ~40 GB, and so I can not use augmentation as then the hdf5 dataset would be too large. <\/p>\n\n<p>So, I was thinking that if I could use the Imagedata layer for reading images and the hdf5 data layer for metadata, my problem can be solved. But I didn't find any material on this. Is it possible to do so?<\/p>\n","answers":[{"AnswerId":"34831493","CreationDate":"2016-01-16T20:00:50.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I'm glad you <a href=\"https:\/\/stackoverflow.com\/a\/34819919\/1714410\">figured it out by yourself<\/a>. However, the file size limitation on HDF5 files is not an issue at all. In caffe, <code>\"HDF5Data\"<\/code> layer gets as an input a <em>text<\/em> file as <code>source<\/code> listing <em>as many<\/em> hdf5 files as you wish! Thus, if you encounter a large hdf5 file, you can split it into several smaller ones and list them all in the text file.<\/p>\n\n<p>For example, you may have <code>\/path\/to\/hdf5_list.txt<\/code> file listing the following files<\/p>\n\n<blockquote>\n<pre><code>\/path\/to\/not_too_large_file.h5\n\/path\/to\/split_large_file_001.h5\n\/path\/to\/split_large_file_002.h5\n\/path\/to\/split_large_file_003.h5\n\/path\/to\/split_large_file_004.h5\n<\/code><\/pre>\n<\/blockquote>\n\n<p>And so on...<br>\nThen your input layer would look like<\/p>\n\n<pre><code>layer {\n  name: \"input\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label\"\n  hdf5_data_param {\n    source: \"\/path\/to\/hdf5_list.txt\"\n  }\n  include { phase: TRAIN }\n}\n<\/code><\/pre>\n"},{"AnswerId":"34819919","CreationDate":"2016-01-15T21:12:31.237","ParentId":null,"OwnerUserId":"3698878","Title":null,"Body":"<p>Yes it is, I thought I was getting an error but it was due to something else in my proto file, below is my input layers for proto<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  name: \"data\"\n  top: \"Meta\" \n  hdf5_data_param {\n    source: \"\/path\/to\/train.txt\"\n    batch_size: 50\n  }\n  include { phase: TRAIN }\n}\nlayer {\n  name: \"data\"\n  type: \"ImageData\"\n  top: \"X\"\n  top: \"Labels\"\n  include {\n    phase: TRAIN\n  }\n  transform_param {\n    mirror: true\n    crop_size: 227\n    mean_file: \"data\/ilsvrc12\/imagenet_mean.binaryproto\"\n  }\n  image_data_param {\n    source: \"\/path\/to\/train.txt\"\n    batch_size: 50\n    new_height: 256\n    new_width: 256\n  }\n}\nlayer {\n  type: \"HDF5Data\"\n  name: \"data\"\n  top: \"Meta\" \n  hdf5_data_param {\n    source: \"\/path\/to\/val.txt\"\n    batch_size: 50\n  }\n  include { phase: TEST }\n}\nlayer {\n  name: \"data\"\n  type: \"ImageData\"\n  top: \"X\"\n  top: \"Labels\"\n  include {\n    phase: TEST\n  }\n  transform_param {\n    mirror: false\n    crop_size: 227\n    mean_file: \"data\/ilsvrc12\/imagenet_mean.binaryproto\"\n  }\n  image_data_param {\n    source: \"\/path\/to\/val.txt\"\n    batch_size: 50\n    new_height: 256\n    new_width: 256\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":34821002,"AnswerCount":1,"Tags":"<machine-learning><neural-network><torch><conv-neural-network>","CreationDate":"2016-01-15T22:33:22.760","AcceptedAnswerId":34821792.0,"OwnerUserId":4013781.0,"Title":"Torch7 using weights with unbalanced training sets","Body":"<p>I am using a CrossEntropyCriterion with my convnet. I have 150 classes and the number of training files per class is <strong><em>very unbalanced<\/em><\/strong> (5 to 2000 files). According to the documentation, I can compensate for this using weights:<\/p>\n\n<pre><\/pre>\n\n<p>\"If provided, the optional argument weights should be a 1D Tensor assigning weight to each of the classes. This is particularly useful when you have an unbalanced training set.\"<\/p>\n\n<p>What format should the weights be in? Eg: number training files in class n \/ total number of training files.<\/p>\n","answers":[{"AnswerId":"34821792","CreationDate":"2016-01-15T23:47:46.600","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>I assume you want to balance your training in this meaning, that small class becomes more important. In general <strong>there are infinitely many possible weightings<\/strong> leading to various results. One of the simpliest ones, which simply assumes that each class should be equally important (thus efficiently you drop the empirical prior) is to put weight proportional to<\/p>\n\n<pre><code>1 \/ # samples_in_class\n<\/code><\/pre>\n\n<p>for example<\/p>\n\n<pre><code>weight_of_class_y = # all_samples \/ # samples_in_y\n<\/code><\/pre>\n\n<p>This way if you have 5:2000 dissproportion, the smaller class becomes 400 times more important for the model.<\/p>\n"}]}
{"QuestionId":34821102,"AnswerCount":1,"Tags":"<machine-learning><neural-network><torch><conv-neural-network>","CreationDate":"2016-01-15T22:41:35.513","AcceptedAnswerId":36380562.0,"OwnerUserId":4013781.0,"Title":"Torch7 alternative to MultiLabelMarginCriterion","Body":"<p>I have a multi class label problem to solve. Ie: each test image can be allocated 1-10 non-exclusive labels.<\/p>\n\n<p>However, I am having problems with MultiLabelMarginCriterion because it's not supported by cunn. So, I am looking for alternative approaches. Would either of these be effective?<\/p>\n\n<ol>\n<li><p>Calculate each permutation of labels in the training set (about 150) and train a classifier to identify those 150 classes. However, I do not think that new permutations of labels in the test set (those not found in the training set) will not be recognised.<\/p><\/li>\n<li><p>Train 10 separate binary classifiers using BCECriterion. Ie: one classifier for each label. Run each test image through each classifier and combine the results. However, training lots of CNNs is time-consuming.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"36380562","CreationDate":"2016-04-03T01:10:34.560","ParentId":null,"OwnerUserId":"4013781","Title":null,"Body":"<p>The best solution I found was to use MSECriterion where the targets and predictions are arrays of 1s and -1s, indicating the presence or absence of the label. <\/p>\n"}]}
{"QuestionId":34822152,"AnswerCount":1,"Tags":"<python><machine-learning><caffe><pycaffe><lmdb>","CreationDate":"2016-01-16T00:32:13.503","AcceptedAnswerId":34874508.0,"OwnerUserId":2527680.0,"Title":"how to append data to existing LMDB?","Body":"<p>I have around 1 million images to put in this dataset 10000 at a time appended to the set.<\/p>\n\n<p>I\"m sure the map_size is wrong with ref from this <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\">article<\/a><\/p>\n\n<p>used  this line to create the set <\/p>\n\n<pre><\/pre>\n\n<p>use this line every 10000 sample to write data to file where X and Y are placeholders for the data to be put in the LMDB.<\/p>\n\n<pre><\/pre>\n\n<p>How can I edit this code such that new data is added to this LMDB and not replaced as this present method replaces it in the same position.\nI have check the length after generation with the env.stat().<\/p>\n","answers":[{"AnswerId":"34874508","CreationDate":"2016-01-19T10:34:57.797","ParentId":null,"OwnerUserId":"1391220","Title":null,"Body":"<p>Le me expand on my comment above.<\/p>\n\n<p>All entries in LMDB are stored according to unique keys and your database already contains keys for <code>i = 0, 1, 2, ...<\/code>. You need a way to find unique keys for each <code>i<\/code>. The simplest way to do that is to find the largest key in existing DB and keep adding to it.<\/p>\n\n<p>Assuming that existing keys are consecutive, <\/p>\n\n<pre><code>max_key = env.stat()[\"entries\"]\n<\/code><\/pre>\n\n<p>Otherwise, a more thorough approach is to iterate over all keys. (<a href=\"https:\/\/stackoverflow.com\/questions\/34420375\/count-number-of-records-in-lmdb-databse-with-python\">Check this.<\/a>)<\/p>\n\n<pre><code>max_key = 0\nfor key, value in env.cursor():\n    max_key = max(max_key, key)\n<\/code><\/pre>\n\n<p>Finally, simply replace line 7 of your <code>for<\/code> loop,<\/p>\n\n<pre><code>str_id = '{:08}'.format(i)\n<\/code><\/pre>\n\n<p>by<\/p>\n\n<pre><code>str_id = '{:08}'.format(max_key + 1 + i)\n<\/code><\/pre>\n\n<p>to append to the existing database.<\/p>\n"}]}
{"QuestionId":34822219,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-16T00:41:01.147","AcceptedAnswerId":34831956.0,"OwnerUserId":2084314.0,"Title":"Tensorflow Incompatible Shapes Error in Tutorial","Body":"<p>I've been trying to create the convolutional network from the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow noreferrer\">Tensorflow tutorial<\/a>, but I've been having trouble.  For some reason, I'm getting errors where the size of y_conv is 4x larger than the size of y_, and I have no idea why.  I found <a href=\"https:\/\/stackoverflow.com\/questions\/34223315\/incompatible-shapes-on-tensorflow-equal-op-for-correct-predictions-evaluation\">this question<\/a>, but it appears to be a different problem than mine, though it looks similar.<\/p>\n\n<p>To be clear, the batch size in the below code is 50, but the error it's coming up with is<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors.InvalidArgumentError: Incompatible shapes: [200] vs. [50]<\/p>\n<\/blockquote>\n\n<p>and when I change the batch size to 10, I get<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors.InvalidArgumentError: Incompatible shapes: [40] vs. [10]<\/p>\n<\/blockquote>\n\n<p>so it's related to the batch size somehow, but I can't figure that out.  Can anybody tell me what's wrong with this code?  It's pretty much straight from the tutorial linked above.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"34831956","CreationDate":"2016-01-16T20:46:26.837","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>The reshapes with the -1's are clues. It's not the batch size that's wrong it's the image size. You're flattening it out into the batch dimension.<\/p>\n\n<p>Why is the image the wrong size? <\/p>\n\n<p>On the second conv you're passing <code>conv1<\/code> instead of <code>pool1<\/code>\n<code>conv2d(h_conv1, w_conv2)<\/code>.<\/p>\n\n<p>Personally for pipelines like this I like to use 1 name for the data as it flows through.<\/p>\n\n<p>Start using a debugger, it's worth it!<\/p>\n"}]}
{"QuestionId":34825167,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras>","CreationDate":"2016-01-16T08:49:17.487","AcceptedAnswerId":null,"OwnerUserId":5797920.0,"Title":"Keras AttributeError: 'module' object has no attribute 'relu'","Body":"<p>When I try to use  in Python, PyCharm tells me:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know why? <\/p>\n","answers":[{"AnswerId":"34825280","CreationDate":"2016-01-16T09:05:37.647","ParentId":null,"OwnerUserId":"4952130","Title":null,"Body":"<p>This has been marked as an <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/825\" rel=\"nofollow\"><em>issue<\/em><\/a> on their issue tracker and is caused by the fact that the version on <code>pip<\/code> is not the most up to date.\nAs most users on there suggested, re-installing theano by cloning the repository and using <code>setup.py<\/code> solves the issue.<\/p>\n\n<p>P.s: It is also <em>explicitly<\/em> mentioned in <a href=\"https:\/\/github.com\/fchollet\/keras#installation\" rel=\"nofollow\">their installation guide<\/a>:<\/p>\n\n<blockquote>\n  <p>Note: <em>You should use the latest version of Theano, not the PyPI<\/em> version. Install it with:<\/p>\n\n<pre><code>sudo pip install git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Along with all other dependencies, of course.<\/p>\n"}]}
{"QuestionId":34826736,"AnswerCount":2,"Tags":"<python><python-2.7><cluster-computing><tensorflow><slurm>","CreationDate":"2016-01-16T12:03:59.827","AcceptedAnswerId":36803148.0,"OwnerUserId":3314143.0,"Title":"Running TensorFlow on a Slurm Cluster?","Body":"<p>I could get access to a computing cluster, specifically one node with two 12-Core CPUs, which is running with <a href=\"https:\/\/en.wikipedia.org\/wiki\/Slurm_Workload_Manager\" rel=\"noreferrer\">Slurm Workload Manager<\/a>.<\/p>\n\n<p>I would like to run <a href=\"https:\/\/en.wikipedia.org\/wiki\/TensorFlow\" rel=\"noreferrer\">TensorFlow<\/a> on that system but unfortunately I were not able to find any information about how to do this or if this is even possible. I am new to this but as far as I understand it, I would have to run TensorFlow by creating a Slurm job and can not directly execute python\/tensorflow via ssh. <\/p>\n\n<p>Has anyone an idea, tutorial or any kind of source on this topic?<\/p>\n","answers":[{"AnswerId":"47431952","CreationDate":"2017-11-22T10:13:05.737","ParentId":null,"OwnerUserId":"8201139","Title":null,"Body":"<p>You can simply pass a batch script to slurm with the <code>sbatch<\/code> command like such<\/p>\n\n<pre><code>sbatch --partition=part start.sh\n<\/code><\/pre>\n\n<p>listing available partitions can be done with <code>sinfo<\/code>.<\/p>\n\n<p><strong>start.sh<\/strong> (possible configuration)<strong>:<\/strong> <\/p>\n\n<pre><code>#!\/bin\/sh\n#SBATCH -N 1      # nodes requested\n#SBATCH -n 1      # tasks requested\n#SBATCH -c 10      # cores requested\n#SBATCH --mem=32000  # memory in Mb\n#SBATCH -o outfile  # send stdout to outfile\n#SBATCH -e errfile  # send stderr to errfile\npython run.py\n<\/code><\/pre>\n\n<p>whereas run.py contains the script you want to be executed with slurm i.e. your tensorflow code. <\/p>\n\n<p>You can look up the details here: <a href=\"https:\/\/slurm.schedmd.com\/sbatch.html\" rel=\"nofollow noreferrer\">https:\/\/slurm.schedmd.com\/sbatch.html<\/a> <\/p>\n"},{"AnswerId":"36803148","CreationDate":"2016-04-22T20:53:30.070","ParentId":null,"OwnerUserId":"3398929","Title":null,"Body":"<p>It's relatively simple.<\/p>\n\n<p>Under the simplifying assumptions that you request one process per host, slurm will provide you with all the information you need in environment variables, specifically SLURM_PROCID, SLURM_NPROCS and SLURM_NODELIST.<\/p>\n\n<p>For example, you can initialize your task index, the number of tasks and the nodelist as follows:<\/p>\n\n<pre><code>from hostlist import expand_hostlist\ntask_index  = int( os.environ['SLURM_PROCID'] )\nn_tasks     = int( os.environ['SLURM_NPROCS'] )\ntf_hostlist = [ (\"%s:22222\" % host) for host in\n                expand_hostlist( os.environ['SLURM_NODELIST']) ]  \n<\/code><\/pre>\n\n<p>Note that slurm gives you a host list in its compressed format (e.g., \"myhost[11-99]\"), that you need to expand. I do that with module hostlist by \n Kent Engstr\u00f6m, available here <a href=\"https:\/\/pypi.python.org\/pypi\/python-hostlist\" rel=\"noreferrer\">https:\/\/pypi.python.org\/pypi\/python-hostlist<\/a><\/p>\n\n<p>At that point, you can go right ahead and create your TensorFlow cluster specification and server with the information you have available, e.g.:<\/p>\n\n<pre><code>cluster = tf.train.ClusterSpec( {\"your_taskname\" : tf_hostlist } )\nserver  = tf.train.Server( cluster.as_cluster_def(),\n                           job_name   = \"your_taskname\",\n                           task_index = task_index )\n<\/code><\/pre>\n\n<p>And you're set! You can now perform TensorFlow node placement on a specific host of your allocation with the usual syntax:<\/p>\n\n<pre><code>for idx in range(n_tasks):\n   with tf.device(\"\/job:your_taskname\/task:%d\" % idx ):\n       ...\n<\/code><\/pre>\n\n<p>A flaw with the code reported above is that all your jobs will instruct Tensorflow to install servers listening at fixed port 22222. If multiple such jobs happen to be scheduled to the same node, the second one will fail to listen to 22222. <\/p>\n\n<p>A better solution is to let slurm reserve ports for each job. You need to bring your slurm administrator on board and ask him to configure slurm so it allows you to ask for ports with the --resv-ports option. In practice, this requires asking them to add a line like the following in their slurm.conf:<\/p>\n\n<pre><code>MpiParams=ports=15000-19999\n<\/code><\/pre>\n\n<p>Before you bug your slurm admin, check what options are already configured, e.g., with:<\/p>\n\n<pre><code>scontrol show config | grep MpiParams\n<\/code><\/pre>\n\n<p>If your site already uses an old version of OpenMPI, there's a chance an option like this is already in place.<\/p>\n\n<p>Then, amend my first snippet of code as follows:<\/p>\n\n<pre><code>from hostlist import expand_hostlist\ntask_index  = int( os.environ['SLURM_PROCID'] )\nn_tasks     = int( os.environ['SLURM_NPROCS'] )\nport        = int( os.environ['SLURM_STEP_RESV_PORTS'].split('-')[0] )\ntf_hostlist = [ (\"%s:%s\" % (host,port)) for host in\n                expand_hostlist( os.environ['SLURM_NODELIST']) ]  \n<\/code><\/pre>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":34830050,"AnswerCount":0,"Tags":"<python><numpy><theano><lapack><blas>","CreationDate":"2016-01-16T17:42:10.087","AcceptedAnswerId":null,"OwnerUserId":1699711.0,"Title":"Linking NumPy with BLAS on windows","Body":"<p>I'm trying to install Theano on a Windows system and need to install BLAS and LAPACK. I've got the .dll files for these in my System32 folder. When I run numpy.<strong>config<\/strong>.show() from Anaconda, the paths to the libraries are displayed properly. However, this does not happen on IDLE and the same command tells prints NOT AVAILABLE for the paths to libraries for LAPACK and BLAS.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/25148639\/how-do-i-use-numpy-in-python-idle\">Another thread<\/a> suggests that Anaconda and my original Python 2.7 distributions are treated as separate environments.\nMy question is, if Anaconda says that BLAS is installed, what am I missing in the other environment? I re-compiled the libraries and force-reinstalled NumPy, but this doesn't seem to help.<\/p>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[]}
{"QuestionId":34830393,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-16T18:12:45.947","AcceptedAnswerId":null,"OwnerUserId":4020719.0,"Title":"Suggestions for building tensorflow seq2seq 'lite' model\/dataset","Body":"<p>I want to run the tensorflow seq2seq tutorial but don't want to use large amounts of memory. I'm not concerned with performance right now - I'm more interested in debugging. Suggestions?<\/p>\n\n<p>For background, I'm working with the seq2seq tutorial <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html\" rel=\"nofollow\">here<\/a> (and hopefully making modifications):<\/p>\n","answers":[{"AnswerId":"35901025","CreationDate":"2016-03-09T19:37:41.123","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>Run it with much smaller input files, vocab sizes, embedding size and\/or num of layers. <\/p>\n\n<p>Try using smaller training files in [your_data_directory]...<\/p>\n\n<pre><code>&gt; head -10000 giga-fren.release2.fr &gt; giga-fren.release2.fr\n&gt; head -10000 giga-fren.release2.en &gt; giga-fren.release2.en\n<\/code><\/pre>\n\n<p>Try a smaller vocab size like 1000...<\/p>\n\n<pre><code>python translate.py\n--data_dir [your_data_directory] --train_dir [checkpoints_directory]\n--en_vocab_size=1000 --fr_vocab_size=1000\n<\/code><\/pre>\n\n<p>Try a smaller embedding size and\/or fewer layers<\/p>\n\n<pre><code>python translate.py\n--data_dir [your_data_directory] --train_dir [checkpoints_directory]\n--size=128 --num_layers=1 --steps_per_checkpoint=50\n<\/code><\/pre>\n\n<p>The model will perform poorly, but run much faster with a smaller footprint.<\/p>\n"}]}
{"QuestionId":34830499,"AnswerCount":1,"Tags":"<python><machine-learning><regression><theano><lasagne>","CreationDate":"2016-01-16T18:22:44.103","AcceptedAnswerId":null,"OwnerUserId":5575869.0,"Title":"Linear Regression Lasagne \/ Theano","Body":"<p>I'm trying to make a simple multivariate linear Regression with Lasagne.\nThis is my Input:<\/p>\n\n<pre><\/pre>\n\n<p>For this two data points the network should be able to learn  perfectly.<\/p>\n\n<p>Here is the model:<\/p>\n\n<pre><\/pre>\n\n<p>After executing multiple times<\/p>\n\n<pre><\/pre>\n\n<p>the cost explodes to infinity. Any idea what is going wrong here?<\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"34839499","CreationDate":"2016-01-17T14:33:37.223","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The network has too much capacity for a single training instance. You would need to apply some strong regularization to prevent the training diverging. Alternatively, and hopefully more realistically, give it more complex training data (many instances).<\/p>\n\n<p>With a single instance the task can be solved using just one input, instead of 20, and with the <code>DenseLayer<\/code>'s bias disabled:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport lasagne\nimport theano.tensor as T\n\n\ndef compile():\n    x, z = T.matrices('x', 'z')\n    lh = lasagne.layers.InputLayer(shape=(None, 1), input_var=x)\n    ly = lasagne.layers.DenseLayer(lh, 6, nonlinearity=lasagne.nonlinearities.linear,\n                                   b=None)\n    y = lasagne.layers.get_output(ly)\n    params = lasagne.layers.get_all_params(ly, trainable=True)\n    cost = T.sum(lasagne.objectives.squared_error(y, z))\n    updates = lasagne.updates.sgd(cost, params, learning_rate=0.0001)\n    return theano.function([x, z], [y, cost], updates=updates)\n\n\ndef main():\n    f_train = compile()\n\n    x_train = np.array([[37.93]]).astype(theano.config.floatX)\n    y_train = np.array([[37.92, 138.25, 324.66, 16.28, 16.27, 16.28]])\\\n        .astype(theano.config.floatX)\n\n    for _ in xrange(100):\n        print f_train(x_train, y_train)\n\n\nmain()\n<\/code><\/pre>\n\n<p>Note that the learning rate also needs to be reduced a lot to prevent divergence.<\/p>\n"}]}
{"QuestionId":34831038,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><neural-network><torch><conv-neural-network>","CreationDate":"2016-01-16T19:15:41.767","AcceptedAnswerId":34831435.0,"OwnerUserId":4341948.0,"Title":"Trouble understanding Convolutional Neural Network","Body":"<p>I read about convolutional neural networks from <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\">here<\/a>. Then I started playing with torch7. I am having confusion with the convolutional layer of a CNN. <\/p>\n\n<p>From the tutorial,<\/p>\n\n<h1>1<\/h1>\n\n<p><\/p>\n\n<h1>2<\/h1>\n\n<p><\/p>\n\n<h1>3<\/h1>\n\n<p>if the input layer is [32x32x3], <\/p>\n\n<p>I started playing with what a CONV layer might do to an image. I did that in torch7. Here is my implementation,<\/p>\n\n<pre><\/pre>\n\n<p><strong>output<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/7KJB4.png\"><img src=\"https:\/\/i.stack.imgur.com\/7KJB4.png\" alt=\"cnn transformation\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>Now lets see the structure of a CNN <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/I3I9b.png\"><img src=\"https:\/\/i.stack.imgur.com\/I3I9b.png\" alt=\"CNN\"><\/a><\/p>\n\n<p>So, my questions are,<\/p>\n\n<h1>Question 1<\/h1>\n\n<p>Is the convolution done like this - lets say we take an image 32x32x3. And there is 5x5 filter. Then the 5x5 filter will pass through the whole 32x32 image and produce the convoluted images? Okay, so sliding 5x5 filter across the whole image, we get one image, if there are 10 output layers, we get 10 images(as you see from the output). How do we get these? (see the image for clarification if required)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JiNIa.png\"><img src=\"https:\/\/i.stack.imgur.com\/JiNIa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h1>Question 2<\/h1>\n\n<p>What is the number of neurons in the conv layer? Is it the number of output layers? In the code I've written above, . Is it 10? (no. of output layers?)<\/p>\n\n<p>If so the point number 2 does not make any sense. According to that  So what will be the weight here? I am greatly confused in this. In the model defined in torch, there is no weight. So how the weight is playing a role here?<\/p>\n\n<p>Can someone explain what is going on?<\/p>\n","answers":[{"AnswerId":"34886078","CreationDate":"2016-01-19T20:12:41.260","ParentId":null,"OwnerUserId":"5812469","Title":null,"Body":"<p>Disclaimer:\nThe information I have provided below is mostly extracted from the following papers:\nInformation Processing in Cats Visual Cortex\nGradient Based Learning Applied to Document Recognition\nNeocortigan\nReceptive Fields in the Cat's Visual Cortex<\/p>\n\n<blockquote>\n  <p>Is the convolution done like this - lets say we take an image 32x32x3. And there is 5x5 filter. Then the 5x5 filter will pass through the whole 32x32 image and produce the convoluted images?<\/p>\n<\/blockquote>\n\n<p>Yes, a 5x5 filter will pass through the whole image creating a 28x28 RGB image. Each unit in the so called \"feature map\" receives 5x5x3 inputs connected to a 5x5 area in the input image ( this 5x5 neighborhood is called the unit's \"local receptive field\"). Receptive fields of neighboring(adjacent) units in a feature map are centered on neighboring(adjacent) units in the previous layer. <\/p>\n\n<blockquote>\n  <p>Okay, so sliding 5x5 filter across the whole image, we get one image, if there are 10 output layers, we get 10 images(as you see from the output). How do we get these? (see the image for clarification if required)<\/p>\n<\/blockquote>\n\n<p>Note that units on the feature map layer share the same set of wights and perform the same operation on different parts of image.( That is if you shift the original image, the output on feature map will also shift by the same amount). That is, for every feature map you constraint the set of weights to be the same for every unit; you only have 5x5x3 unknown weights.<\/p>\n\n<p>Due to this constraint, and since we want to extract as much information as possible from the image, we add more layers, features maps: having multiple feature maps helps us extract multiple features at every pixel. <\/p>\n\n<p>Unfortunately I am not familiar with Torch7.<\/p>\n"},{"AnswerId":"34831435","CreationDate":"2016-01-16T19:55:28.467","ParentId":null,"OwnerUserId":"3760780","Title":null,"Body":"<blockquote>\n  <p>Is the convolution done like this - lets say we take an image 32x32x3. And there is 5x5 filter. Then the 5x5 filter will pass through the whole 32x32 image and produce the convoluted images?<\/p>\n<\/blockquote>\n\n<p>For a 32x32x3 input image a 5x5 filter would iterate over every single pixel and for each pixel look at the 5x5 neighborhood. That neighborhood contains 5*5*3=75 values. Below is an example image for a 3x3 filter over a single input channel, i.e. one with a neighborhood of 3*3*1 values (<a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html\">source<\/a>).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9FHml.png\"><img src=\"https:\/\/i.stack.imgur.com\/9FHml.png\" alt=\"convolution\"><\/a><\/p>\n\n<p>For each individual neighbor the filter will have one parameter (aka weight), so 75 parameters. Then to calculate one single output value (value at pixel x, y) it reads those neighbor values, multiplies each one with the respective parameter\/weight and adds those up at the end (see <a href=\"https:\/\/en.wikipedia.org\/wiki\/Convolution#Discrete_convolution\">discrete convolution<\/a>). The optimal weights have to be learned during the training.<\/p>\n\n<p>So one filter will iterate over the image and generate a new output, pixel by pixel. If you have multiple filters (i.e. the second parameter in <code>SpatialConvolutionMM<\/code> is >1) you get multiple outputs (\"planes\" in torch).<\/p>\n\n<blockquote>\n  <p>Okay, so sliding 5x5 filter across the whole image, we get one image, if there are 10 output layers, we get 10 images(as you see from the output). How do we get these? (see the image for clarification if required)<\/p>\n<\/blockquote>\n\n<p>Each output plane gets generated by its own filter. Each filter has its own parameters (5*5*3 parameters in your example). The process for multiple filters is exactly the same as it is for one.<\/p>\n\n<blockquote>\n  <p>What is the number of neurons in the conv layer? Is it the number of output layers? In the code I've written above, model:add(nn.SpatialConvolutionMM(3, 10, 5, 5)). Is it 10? (no. of output layers?)<\/p>\n<\/blockquote>\n\n<p>You should call them weights or parameters, \"neurons\" doesn't really fit for convolutional layers. The number of parameters is, as described, 5*5*3=75 per filter in your example. As you have 10 filters (\"output planes\") you have 750 parameters total. If you add a second layer to your network with <code>model:add(nn.SpatialConvolutionMM(10, 10, 5, 5))<\/code> you would have an additional 5*5*10=250 parameters per filter and 250*10=2500 total. Notice how that number can quickly grow (512 filters\/output planes in one layer operating on 256 input planes is nothing uncommon).<\/p>\n\n<p>For further reading you should look at <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html\">http:\/\/neuralnetworksanddeeplearning.com\/chap6.html<\/a> . Scroll down to the chapter \"Introducing convolutional networks\". Under \"Local receptive fields\" there are visualizations which will probably help you understand what a filter does (one was shown above).<\/p>\n"}]}
{"QuestionId":34832161,"AnswerCount":1,"Tags":"<input><hdf5><caffe><conv-neural-network><matcaffe>","CreationDate":"2016-01-16T21:09:03.913","AcceptedAnswerId":null,"OwnerUserId":3698878.0,"Title":"Caffe - Doing forward pass with multiple input blobs","Body":"<p>Following are the input layers of my fine-tuned model:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see it has one imagedata input layer and 1 hdf5 input layer, if there was only 1 type of layer say imagedata, I could have done:<\/p>\n\n<pre><\/pre>\n\n<p>and then \n     \nBut here I have to give two types of input data, how can I do this? Please help!<\/p>\n","answers":[{"AnswerId":"35722283","CreationDate":"2016-03-01T11:57:08.593","ParentId":null,"OwnerUserId":"3698878","Title":null,"Body":"<p>I had to check matcaffe.cpp (and recompile with make matcaffe) and print the check variables for the 'Invalid Input Size' condition I was failing, to get the idea of transposing input_data which works.<\/p>\n\n<pre><code>input_data = {prepare_images(im),prepare_other_data()};\nscores = caffe('forward', input_data');\n<\/code><\/pre>\n\n<p>Thus taking the transpose worked for me.<\/p>\n"}]}
{"QuestionId":34833796,"AnswerCount":0,"Tags":"<python><image><deep-learning><caffe>","CreationDate":"2016-01-17T00:34:57.800","AcceptedAnswerId":null,"OwnerUserId":5760459.0,"Title":"how to rename image names in finder with a custom list of names from an .txt\/.mat\/excel, etc file (OS X)?","Body":"<p>I download this training data from \n<a href=\"http:\/\/ai.stanford.edu\/~jkrause\/cars\/car_dataset.html\" rel=\"nofollow\">http:\/\/ai.stanford.edu\/~jkrause\/cars\/car_dataset.html<\/a>. However, the training data is not named per category.  <\/p>\n\n<p>The data is named ,,, , etc. I do have a  data file that has (1) class and (2) fname. How can I map these class names + respective file name over the actual image names in mac? <\/p>\n\n<p>I tried manually renaming the files one at a time, but there are ~2500 images. <\/p>\n","answers":[]}
{"QuestionId":34834714,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><gpu><tensorflow>","CreationDate":"2016-01-17T03:13:09.417","AcceptedAnswerId":34838710.0,"OwnerUserId":829624.0,"Title":"Does TensorFlow by default use all available GPUs in the machine?","Body":"<p>I have 3 GTX Titan GPUs in my machine. I run the example provided in Cifar10 with cifar10_train.py and got the following output:<\/p>\n\n<pre><\/pre>\n\n<p>It looks to me that TensorFlow is trying to initialize itself on two devices (gpu0 and gpu1). <\/p>\n\n<p>My question is why it only does that on two devices and is there any way to prevent that? (I only want it to run as if there is a single GPU)<\/p>\n","answers":[{"AnswerId":"34838710","CreationDate":"2016-01-17T13:13:28.613","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>See: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/using_gpu\" rel=\"nofollow noreferrer\">Using GPUs<\/a><\/p>\n\n<p><strong>Manual device placement<\/strong><\/p>\n\n<p>If you would like a particular operation to run on a device of your choice instead of what's automatically selected for you, you can use with <code>tf.device<\/code> to create a device context such that all the operations within that context will have the same device assignment. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Creates a graph.\nwith tf.device('\/cpu:0'):\n\u00a0 a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n\u00a0 b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\nc = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>You will see that now a and b are assigned to <code>cpu:0<\/code>. Since a device was not explicitly specified for the <code>MatMul<\/code> operation, the TensorFlow runtime will choose one based on the operation and available devices (gpu:0 in this example) and automatically copy tensors between devices if required.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><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\/cpu:0\na: \/job:localhost\/replica:0\/task:0\/cpu:0\nMatMul: \/job:localhost\/replica:0\/task:0\/gpu:0\n[[ 22. \u00a028.]\n\u00a0[ 49. \u00a064.]]\n<\/code><\/pre>\n\n<hr>\n\n<p>Earlier Answer 2.<\/p>\n\n<p>See: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/using_gpu\/index.html\" rel=\"nofollow noreferrer\">Using GPUs<\/a><\/p>\n\n<p><strong>Using a single GPU on a multi-GPU system<\/strong><\/p>\n\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\n<pre class=\"lang-py prettyprint-override\"><code># Creates a graph.\nwith tf.device('\/gpu:2'):\n  a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n  b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n  c = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint sess.run(c)\n<\/code><\/pre>\n\n<hr>\n\n<p>Earlier Answer 1.<\/p>\n\n<p>From <a href=\"http:\/\/acceleware.com\/blog\/cudavisibledevices-masking-gpus\" rel=\"nofollow noreferrer\">CUDA_VISIBLE_DEVICES \u2013 Masking GPUs<\/a><\/p>\n\n<blockquote>\n  <p>Does your CUDA application need to target a specific GPU? If you are\n  writing GPU enabled code, you would typically use a device query to\n  select the desired GPUs. However, a quick and easy solution for\n  testing is to use the environment variable CUDA_VISIBLE_DEVICES to\n  restrict the devices that your CUDA application sees. This can be\n  useful if you are attempting to share resources on a node or you want\n  your GPU enabled executable to target a specific GPU.<\/p>\n  \n  <p>Environment Variable Syntax<\/p>\n  \n  <p>Results<\/p>\n  \n  <p>CUDA_VISIBLE_DEVICES=1 Only device 1 will be seen \n  CUDA_VISIBLE_DEVICES=0,1 Devices 0 and 1 will be visible \n  CUDA_VISIBLE_DEVICES=\u201d0,1\u201d Same as above, quotation marks are optional\n  CUDA_VISIBLE_DEVICES=0,2,3 Devices 0, 2, 3 will be visible; device 1\n  is masked <\/p>\n  \n  <p>CUDA will enumerate the visible devices starting at zero. In the last\n  case, devices 0, 2, 3 will appear as devices 0, 1, 2. If you change\n  the order of the string to \u201c2,3,0\u201d, devices 2,3,0 will be enumerated\n  as 0,1,2 respectively. If CUDA_VISIBLE_DEVICES is set to a device that\n  does not exist, all devices will be masked. You can specify a mix of\n  valid and invalid device numbers. All devices before the invalid value\n  will be enumerated, while all devices after the invalid value will be\n  masked.<\/p>\n  \n  <p>To determine the device ID for the available hardware in your system,\n  you can run NVIDIA\u2019s deviceQuery executable included in the CUDA SDK.\n  Happy programming!<\/p>\n  \n  <p>Chris Mason<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":34835503,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-17T05:42:09.220","AcceptedAnswerId":34843397.0,"OwnerUserId":5738942.0,"Title":"Tensorflow: Where is tf.nn.conv2d Actually Executed?","Body":"<p>I am curious about the Tensorflow implementation of . To call it, one simply runs . However, I'm going down the rabbit hole trying to see where it is executed. The code is as follows (where the arrow indicates the function it ultimately calls):<\/p>\n\n<pre><\/pre>\n\n<p>I am familiar with Tensorflow's implementation of LSTMs and the ability to easily manipulate them as one deems fit. Is the function that performs the  calculation written in Python, and if so, where is it? Can I see where and how the strides are executed?<\/p>\n","answers":[{"AnswerId":"34843397","CreationDate":"2016-01-17T20:40:12.447","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> The implementation of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#conv2d\"><code>tf.nn.conv2d()<\/code><\/a> is written in C++, which invokes optimized code using either Eigen (on CPU) or the cuDNN library (on GPU). You can find the implementation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/8eaf671025e8cd5358278f91f7e89e2fbbe6a26b\/tensorflow\/core\/kernels\/conv_ops.cc#L115\">here<\/a>.<\/p>\n\n<p>The chain of functions that you mentioned in the question (from <code>tf.nn.conv2d()<\/code> down) are Python functions for <strong>building<\/strong> a TensorFlow graph, but these do not invoke the implementation. Recall that, in TensorFlow, you first <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/resources\/faq.html#why_does_c_tfmatmula_b_not_execute_the_matrix_multiplication_immediately\">build a symbolic graph, then execute it<\/a>.<\/p>\n\n<p>The implementation of <code>tf.nn.conv2d()<\/code> is only executed happens when you call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\"><code>Session.run()<\/code><\/a> passing a <code>Tensor<\/code> whose value depends on the result of some convolution. For example:<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32)\nfilter = tf.Variable(tf.truncated_normal([5, 5, 3, 32], stddev=0.1)\nconv = tf.nn.conv2d(input, filter, strides=[1, 1, 1, 1], padding='SAME')\n\nresult = sess.run(conv, feed_dict={input: ...})  # &lt;== Execution happens here.\n<\/code><\/pre>\n\n<p>Invoking <code>sess.run(...)<\/code> tells TensorFlow to run all the ops that are neeeded to compute the value of <code>conv<\/code>, including the convolution itself. The path from here to the implementation is somewhat complicated, but goes through the following steps:<\/p>\n\n<ol>\n<li><code>sess.run()<\/code> calls the TensorFlow backend to fetch the value of <code>conv<\/code>.<\/li>\n<li>The backend prunes the computation graph to work out what nodes must be executed, and places the nodes on the appropriate devices (CPU or GPU).<\/li>\n<li>Each device is instructed to execute its subgraph, using an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/executor.h\">executor<\/a>.<\/li>\n<li>The executor eventually invokes the <code>tensorflow::OpKernel<\/code> that corresponds to the convolution operator, by calling its <code>Compute()<\/code> method.<\/li>\n<\/ol>\n\n<p>The <code>\"Conv2D\"<\/code> OpKernel is implemented <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/8eaf671025e8cd5358278f91f7e89e2fbbe6a26b\/tensorflow\/core\/kernels\/conv_ops.cc#L94\">here<\/a>, and its <code>Compute()<\/code> method is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/8eaf671025e8cd5358278f91f7e89e2fbbe6a26b\/tensorflow\/core\/kernels\/conv_ops.cc#L115\">here<\/a>. Because this op is performance critical for many workloads, the implementation is quite complicated, but the basic idea is that the computation is offloaded to either the Eigen Tensor library (if running on CPU), or cuDNN's optimized GPU implementation.<\/p>\n"},{"AnswerId":"44961869","CreationDate":"2017-07-07T02:59:41.300","ParentId":null,"OwnerUserId":"8268280","Title":null,"Body":"<p>TensorFlow programs as consisting of two discrete sections:<\/p>\n\n<ul>\n<li>Building the computational graph.<\/li>\n<\/ul>\n\n<p>tf.nn.conv2d(...) -> tf.nn_ops.conv2d(...) -> tf.gen_nn_ops.conv2d(...) -> _op_def_lib.apply_op(\"Conv2D\", ...) -> graph.create_op -> register op into graph<\/p>\n\n<ul>\n<li>Running the computational graph.<\/li>\n<\/ul>\n\n<p>sess = tf.Session(target) -> sess.run(conv2d) -> master prune full graph to client graph -> master split client graph by task to graph partition -> register graph partition to worker -> worker split subgraph by device to graph partition -> then master notify all workers to run graph partitions -> worker notify all devices to run graph partitions -> executor will run ops by topological sort on device.<\/p>\n\n<p>For one of op, the executor will invoke kernel implement to compute for the op.<\/p>\n\n<p>The kernel implement of tf.nn.conv2d() is written in C++, which invokes optimized code using either Eigen (on CPU) or the cuDNN library (on GPU). <\/p>\n"}]}
{"QuestionId":34844795,"AnswerCount":0,"Tags":"<machine-learning><neural-network><torch><conv-neural-network>","CreationDate":"2016-01-17T22:58:03.193","AcceptedAnswerId":null,"OwnerUserId":4013781.0,"Title":"Installing version of cunn from GitHub and using multilabelmargin criterion","Body":"<p>I need to use multilabelmargin criterion in torch7 to train a CNN. It is not supported by the current cunn. I have tried installing a different version of cunn, as here: <a href=\"https:\/\/github.com\/torch\/cunn\/issues\/152\" rel=\"nofollow\">source<\/a><\/p>\n\n<pre><\/pre>\n\n<p>However, when I do, torch stops working and I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>I subsequently needed to reinstall torch7 using ezinstall. How can I correctly install the version of cunn required, without affecting my torch installation, and therefore use multilabelmargin?<\/p>\n","answers":[]}
{"QuestionId":34844843,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-17T23:03:17.723","AcceptedAnswerId":null,"OwnerUserId":4164870.0,"Title":"Running a Multi Scale network with Tensor Flow","Body":"<p>I want to try to build a multi-scale CNN using tensorflow from the cifar10 code.<\/p>\n\n<p>For what I understood I should take the output of the first conv layer and merge it with the output of the second conv layer to feed the first fully connected layer. Is that right? If yes, how to actually do this?<\/p>\n\n<p>I have almost the same first layers as for the cifar10 except for the norm1 and the pool1 layers that are switched<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Then I try to merge the norm1 layer with the pool2 layer using concat.\nHere's how I do this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'm not even sure that this is the right procedure because the loss is now 17 when in the single scale case I had the initial loss set around 3.\nIs this common? <\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":34847524,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-18T04:52:57.113","AcceptedAnswerId":34852892.0,"OwnerUserId":5803551.0,"Title":"TypeError: Trying to increment a 0-dimensional subtensor with a 1-dimensional value","Body":"<p>I'm a new guy in python, I'm coding theano these days and suffered a problem that bothered me many days, my code is like follows:<\/p>\n\n<pre><\/pre>\n\n<p>what I am trying to do is assigning value based on string1 and string2 to some cell in dis_matrix, I use a condition T.switch(T.eq(string1[a_location[0]-1], string2[a_location[1]-1]) to decide value of temp, so that I can put temp into dis_matrix, but when I ran this code, I got error: <\/p>\n\n<blockquote>\n  <p>TypeError: Trying to increment a 0-dimensional subtensor with a 1-dimensional value.<\/p>\n<\/blockquote>\n\n<p>It seems that  returns not a scalar but a 1-dim vector to me, if I change  to , this code works well, so I think it must be somewhere T.switch gets wrong.<\/p>\n\n<p>Could any help me? Thanks a lot!<\/p>\n","answers":[{"AnswerId":"34852892","CreationDate":"2016-01-18T10:51:35.760","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>I have not verified this proposition by writing and running some code but the problem is probably that <code>string1[a_location[0]-1]<\/code> and\/or <code>string2[a_location[1]-1]<\/code> are not scalars and are actually vectors (maybe vectors with only one entry). Consequently the results of <code>T.eq<\/code> and <code>T.switch<\/code> are also vectors (with the true\/false scalar values in <code>T.switch<\/code> being broadcasted). If <code>temp<\/code> is a vector and <code>dis_matrix[a_location[0], a_location[1]]<\/code> is a scalar then <code>inc_subtensor<\/code> will probably yield the error you see.<\/p>\n"}]}
{"QuestionId":34853070,"AnswerCount":1,"Tags":"<python-2.7><anaconda><caffe><jupyter-notebook><pycaffe>","CreationDate":"2016-01-18T11:00:24.973","AcceptedAnswerId":34865340.0,"OwnerUserId":2736559.0,"Title":"Caffe install getting ImportError: DLL load failed: The specified module could not be found","Body":"<p>I am trying to compile and run the snippets posted <a href=\"http:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">here<\/a>, which basically is going to let me visualize the network internals(feature maps).<br>\nI have successfully compiled  and  using the  branch, And I have copied the caffe folder, into  folder.\nYet still, when I try to run this snippet of code in jupyter notebook : <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error : <\/p>\n\n<pre><\/pre>\n\n<p>Whats wrong here? <\/p>\n\n<p>Notes:<br>\nI'm using   <\/p>\n","answers":[{"AnswerId":"34865340","CreationDate":"2016-01-18T22:35:27.243","ParentId":null,"OwnerUserId":"134077","Title":null,"Body":"<p>There's most likely a more specific dependency issue you are not seeing (Protobuf \/ OpenCV). First try using the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/cpp_classification.html\" rel=\"nofollow\">C++ API<\/a> to load an example and make sure all the DLL's load. Then you can more confidently narrow things down to the Python side. I recommend the more recent windows caffe instructions based off the branch you're using:<\/p>\n\n<p><a href=\"https:\/\/initialneil.wordpress.com\/2015\/01\/11\/build-caffe-in-windows-with-visual-studio-2013-cuda-6-5-opencv-2-4-9\/\" rel=\"nofollow\">https:\/\/initialneil.wordpress.com\/2015\/01\/11\/build-caffe-in-windows-with-visual-studio-2013-cuda-6-5-opencv-2-4-9\/<\/a><\/p>\n\n<p>I had to do a complete rebuild as detailed above (note that some dependencies are easier to find with NuGet). Also be on the lookout for the right protobuf binaries in various 3rdParty.zip files throughout the above blog.<\/p>\n\n<p>If you are okay with a snapshot version of Caffe and you don't need to modify the project itself, the following binaries are <em>much<\/em> easier to install and get working:<\/p>\n\n<p><a href=\"https:\/\/initialneil.wordpress.com\/2015\/07\/15\/caffe-vs2013-opencv-in-windows-tutorial-i\/\" rel=\"nofollow\">https:\/\/initialneil.wordpress.com\/2015\/07\/15\/caffe-vs2013-opencv-in-windows-tutorial-i\/<\/a><\/p>\n"}]}
{"QuestionId":34853756,"AnswerCount":0,"Tags":"<python><numpy><gradient><theano><gradient-descent>","CreationDate":"2016-01-18T11:34:38.817","AcceptedAnswerId":null,"OwnerUserId":3271237.0,"Title":"Gradient descent fails with NaN values in all parameters","Body":"<p>I'm using to minimise some function with stochastic gradient descent<\/p>\n\n<pre><\/pre>\n\n<p> and  - some numpy matrices.<\/p>\n\n<p>So, after several thousands iterations, I've got that cost_ij becomes NaN. Then I checked , and found that after last gradient descent step every value in this matrices becomes NaN. And on previous step values in this matrices wasn't so small or large. I belive that troubles in this  line. So, How I can check  or  numeric values in this step (or previous) of gradient descent?<\/p>\n","answers":[]}
{"QuestionId":34855529,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2016-01-18T13:06:15.687","AcceptedAnswerId":null,"OwnerUserId":5801374.0,"Title":"init.lua:692: attempt to call field 'setheaptracking' (a nil value)","Body":"<p>I am installing torch7. After installation, when I run .\/th, there is error shown below. Do anyone know how to solve this?<\/p>\n\n<p>\/Lua\/torch\/install\/share\/lua\/5.1\/trepl\/init.lua:692: attempt to call field 'setheaptracking' (a nil value)\nstack traceback:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34857988,"AnswerCount":2,"Tags":"<caffe><pycaffe>","CreationDate":"2016-01-18T15:10:45.487","AcceptedAnswerId":null,"OwnerUserId":5805925.0,"Title":"How to calculate time spent on every layer in CNN with caffe interface?","Body":"<p>I was trying to calculate the time spent on every layer in fast-rcnn network. I found out that caffe cmd interface  can do it. But the fast-rcnn is based on the python code, is there any function in the python interface can help me calculate the time ?<\/p>\n","answers":[{"AnswerId":"34858251","CreationDate":"2016-01-18T15:24:35.897","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>python has profiling tools. You can look at <a href=\"https:\/\/docs.python.org\/2\/library\/timeit.html\" rel=\"nofollow\"><code>timeit<\/code><\/a> and the <a href=\"https:\/\/docs.python.org\/2\/library\/profile.html\" rel=\"nofollow\"><code>cProfile<\/code><\/a> tool.<br>\nIf you are looking for something more basic, you can use <code>time<\/code> module to measure time:<\/p>\n\n<pre><code>import time\nt = time.time()\nrun_rcnn_script()\ndt = time.time()-t\nprint \"time spent in run_rcnn_script: \", dt, \" sec.\"\n<\/code><\/pre>\n"},{"AnswerId":"42689337","CreationDate":"2017-03-09T07:36:29.153","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>To calculate time spent at each layer<\/p>\n\n<pre><code>import timeit\n\nt1=timeit.default_timer()\nnet.forward(start='start_layer_name',end='end_layer_name')\nt2=timeit.default_timer()\n\nprint 'time is {}'.format(t2-t1)\n<\/code><\/pre>\n"}]}
{"QuestionId":34858459,"AnswerCount":1,"Tags":"<deep-learning><tensorflow>","CreationDate":"2016-01-18T15:34:25.583","AcceptedAnswerId":34869168.0,"OwnerUserId":3399066.0,"Title":"Tensorflow not sharing variables","Body":"<p>We have modified the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/index.html\">CIFAR-10 tutorial<\/a> (Convolution Neural Network) to run on the Adience database for Gender classification on faces. We read <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\">here<\/a> that \"<strong>Parameter Sharing<\/strong>\" is useful as assumption that one patch feature is useful regardless of location in the image. Except:<\/p>\n\n<blockquote>\n  <p>Note that sometimes the parameter sharing assumption may not make sense. This is especially the case when the input images to a ConvNet have some specific centered structure, where we should expect, for example, that completely different features should be learned on one side of the image than another. One practical example is when the input are <strong>faces<\/strong> that have been centered in the image.<\/p>\n<\/blockquote>\n\n<p><strong>Objective:<\/strong> Therefore we would like to turn off parameter sharing for our CNN.<\/p>\n\n<h2>Code<\/h2>\n\n<p>I think the CIFAR-10 tutorial uses <em>parameter sharing<\/em>? and this part of the code in the  function seems to have to do something with it:<\/p>\n\n<pre><\/pre>\n\n<p>Which calls:<\/p>\n\n<pre><\/pre>\n\n<h2>Question<\/h2>\n\n<ul>\n<li>Is <em>parameter sharing<\/em> indeed happening in the CIFAR-10 tutorial?<\/li>\n<li>Could you tell us whether we're looking at the right piece of code for turning off <em>parameter sharing<\/em> or where else to look?<\/li>\n<li>Any other help \/ suggestions are welcome, because we don't know where to start. <\/li>\n<\/ul>\n","answers":[{"AnswerId":"34869168","CreationDate":"2016-01-19T05:35:01.627","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"noreferrer\">CIFAR-10 model<\/a> from the tutorial uses \"parameter sharing\" in the first two layers (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b63c57620d08385975a22f2784e1006af9e17d78\/tensorflow\/models\/image\/cifar10\/cifar10.py#L192\" rel=\"noreferrer\"><code>'conv1'<\/code><\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b63c57620d08385975a22f2784e1006af9e17d78\/tensorflow\/models\/image\/cifar10\/cifar10.py#L209\" rel=\"noreferrer\"><code>'conv2'<\/code><\/a>). The sharing is implied by the use of the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#conv2d\" rel=\"noreferrer\"><code>tf.nn.conv2d()<\/code><\/a> operator, which effectively extracts patches from the input image and applies the same filter (i.e. a shared parameter) to each patch.<\/p>\n\n<p>It's not trivial to \"turn off\" parameter sharing when you have a set of convolutional layers: instead you have to replace them with a different type of layer. The simplest change might be to replace the convolutional layers with a fully connected layer, e.g. by using <code>tf.nn.relu_layer()<\/code> (as in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b63c57620d08385975a22f2784e1006af9e17d78\/tensorflow\/models\/image\/cifar10\/cifar10.py#L233\" rel=\"noreferrer\"><code>'local3'<\/code><\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b63c57620d08385975a22f2784e1006af9e17d78\/tensorflow\/models\/image\/cifar10\/cifar10.py#L241\" rel=\"noreferrer\"><code>'local4'<\/code><\/a> layers), which internally performs a matrix multiplication and maintains separate parameters for each input neuron.<\/p>\n\n<p><strong>N.B.<\/strong> Fully connected layers are often over-parameterized for vision tasks, and a more appropriate middle ground would be to use a <a href=\"http:\/\/robotics.stanford.edu\/~ang\/papers\/nips11-SelectingReceptiveFields.pdf\" rel=\"noreferrer\">\"local receptive field\"<\/a>, which (informally) maintains separate parameters for each input (as in a fully connected layer), but only combines values from \"nearby\" inputs to produce an output (as in a convolution). Unfortunately, TensorFlow doesn't yet contain an implementation of local receptive fields, but <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/adding_an_op\/index.html#adding-a-new-op\" rel=\"noreferrer\">adding support for them<\/a> would be a useful project.<\/p>\n"}]}
{"QuestionId":34860281,"AnswerCount":1,"Tags":"<python><gpu><tensorflow>","CreationDate":"2016-01-18T17:07:34.060","AcceptedAnswerId":34868531.0,"OwnerUserId":1104696.0,"Title":"unable to run tensorflow on multiple GPUs","Body":"<p>I am running the cifar10 multi-GPU example from the tensorflow repository. I am able to utilize more than one GPUs. My ubuntu PC has two Titan X's, I see memory are fully occupied by the process on both GPUs. However, only one GPU is actually computing. I obtain no speedup. I have tried tensorflow 0.5.0 and 0.6.0 pip binaries. I have also tried compiled from source.<\/p>\n\n<p>EDIT:\nThe problem disappeared after I installed an older version of nvidia driver. <\/p>\n","answers":[{"AnswerId":"34868531","CreationDate":"2016-01-19T04:28:37.173","ParentId":null,"OwnerUserId":"1104696","Title":null,"Body":"<p>The problem disappeared after I installed an older version (352.55) of nvidia driver.<\/p>\n"}]}
{"QuestionId":34866455,"AnswerCount":5,"Tags":"<cuda><nvidia><tensorflow>","CreationDate":"2016-01-19T00:15:18.197","AcceptedAnswerId":null,"OwnerUserId":418595.0,"Title":"Unable to get cuda to work in tensorflow","Body":"<p>I'm trying to use cuda to accelerate tensorflow. I'm running tensorflow using the docker image.<\/p>\n\n<p>Firstly, when I launch the gpu image, it has a mismatch in the LT_LIBRARY_PATH environment variable:<\/p>\n\n<pre><\/pre>\n\n<p>There's no nvidia directory there. When I try to run the convolutional.py demo, it can't initialise the cuda support:<\/p>\n\n<pre><\/pre>\n\n<p>It then goes on to train using cpu only.<\/p>\n\n<pre><\/pre>\n\n<p>So in the docker image, there's only the gnu cpu cuda implementation. No NVIDIA stuff. In the host ubuntu 15.10 session, I have libcuda.so installed:<\/p>\n\n<pre><\/pre>\n\n<p>So these seem to be stubs ... not sure why.<\/p>\n\n<p>Is there some trick to getting this to work?<\/p>\n","answers":[{"AnswerId":"38055383","CreationDate":"2016-06-27T13:34:12.557","ParentId":null,"OwnerUserId":"2799796","Title":null,"Body":"<p>I had the same problem with running tensorflow on a Ubuntu machine after I upgraded my driver to 352.63 and 352.93. (I remember it works with 346.* but when I try to install 346., it installs 352. automatically for some reason).<\/p>\n\n<p>I finally figured out that it's caused by permission issue. (I can run it with root) So, I changed the permission of the libcuda.so.352-63 file to executable by anyone and it works well now.<\/p>\n\n<p>Hope this will be helpful to those still struggling with this issue.<\/p>\n\n<p>I didn't try the docker one, but I guess it's also caused by permission setting.<\/p>\n"},{"AnswerId":"39776678","CreationDate":"2016-09-29T17:36:30.813","ParentId":null,"OwnerUserId":"2693559","Title":null,"Body":"<p>Try this command<\/p>\n\n<pre><code>sudo apt-get install nvidia-modprobe\n<\/code><\/pre>\n\n<p>As mentioned here: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/394\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/394<\/a>\nand \n<a href=\"http:\/\/kkjkok.blogspot.in\/2016_08_01_archive.html\" rel=\"nofollow\">http:\/\/kkjkok.blogspot.in\/2016_08_01_archive.html<\/a><\/p>\n"},{"AnswerId":"42418815","CreationDate":"2017-02-23T14:40:20.537","ParentId":null,"OwnerUserId":"1351629","Title":null,"Body":"<p>After I updated NVIDIA driver to  378.09 on Ubuntu 14.10 I had the same error, \nalthough all the right for lib files were set correctly.<br>\nThanks to @PhoenixQ, I tried to run with <code>sudo<\/code> and it worked. <\/p>\n\n<p>After that I tried to run without <code>sudo<\/code> one more time and error disappeared. I'm not sure what ecxactly happened, but maybe something was configured during call with <code>sudo<\/code>, which was not possible withous <code>sudo<\/code>. <\/p>\n\n<p>So the <strong>solution:<\/strong>  <\/p>\n\n<ol>\n<li>Try to run the same thing with <code>sudo<\/code>.  <\/li>\n<li>After this. Tryu running without <code>sudo<\/code>. Worked for me.<\/li>\n<\/ol>\n"},{"AnswerId":"34890208","CreationDate":"2016-01-20T01:53:28.397","ParentId":null,"OwnerUserId":"2404064","Title":null,"Body":"<p>Try rebuilding the Docker image directly from the Tensorflow repository (i.e. don't rely on the image on the container registry) and use <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\" rel=\"nofollow\">https:\/\/github.com\/NVIDIA\/nvidia-docker<\/a> to run the container (the Docker command described in the Tensorflow documentation is not portable).<\/p>\n"},{"AnswerId":"35501644","CreationDate":"2016-02-19T09:23:23.297","ParentId":null,"OwnerUserId":"5950288","Title":null,"Body":"<p>I had a similar problem, though not in docker. The libcuda.so in \/usr\/local\/cuda\/lib64\/stubs was a broken sym link. When I searched for libcuda.so it only turned up a file in a lib32 folder. <\/p>\n\n<p>It seems that the problem was how I originally installed the NVIDIA device driver. At some point in the driver install process you're given the option to install the lib32 drivers. I had thought this meant in addition to lib64 drivers so I selected it. Turns out it only installs lib32 and not lib64 drivers. <\/p>\n\n<p>I reinstalled the NIVDIA device driver, this time not selecting the lib32 'option'. Now tensorflow finds libcuda.so.<\/p>\n"}]}
{"QuestionId":34870614,"AnswerCount":8,"Tags":"<python><tensorflow><deep-learning><word-embedding><natural-language-processing>","CreationDate":"2016-01-19T07:14:40.633","AcceptedAnswerId":34877590.0,"OwnerUserId":5808490.0,"Title":"What does tf.nn.embedding_lookup function do?","Body":"<pre><\/pre>\n\n<p>I cannot understand the duty of this function. Is it like a lookup table? Which means to return the parameters corresponding to each id (in ids)?<\/p>\n\n<p>For instance, in the  model if we use , then for each  it finds the correspond embedding?<\/p>\n","answers":[{"AnswerId":"48438325","CreationDate":"2018-01-25T08:14:07.070","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Yes, the purpose of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/embedding_lookup\" rel=\"noreferrer\"><strong><code>tf.nn.embedding_lookup()<\/code><\/strong><\/a> function is to perform a <em>lookup<\/em> in the <em>embedding matrix<\/em> and return the embeddings (or in simple terms the vector representation) of words.<\/p>\n\n<p>A simple embedding matrix (of shape: <strong><code>vocabulary_size x embedding_dimension<\/code><\/strong>) would look like below. (i.e. each <em>word<\/em> will be represented by a <em>vector<\/em> of numbers; hence the name <em>word2vec<\/em>)<\/p>\n\n<hr>\n\n<p><strong>Embedding Matrix<\/strong><\/p>\n\n<pre><code>the 0.418 0.24968 -0.41242 0.1217 0.34527 -0.044457 -0.49688 -0.17862\nlike 0.36808 0.20834 -0.22319 0.046283 0.20098 0.27515 -0.77127 -0.76804\nbetween 0.7503 0.71623 -0.27033 0.20059 -0.17008 0.68568 -0.061672 -0.054638\ndid 0.042523 -0.21172 0.044739 -0.19248 0.26224 0.0043991 -0.88195 0.55184\njust 0.17698 0.065221 0.28548 -0.4243 0.7499 -0.14892 -0.66786 0.11788\nnational -1.1105 0.94945 -0.17078 0.93037 -0.2477 -0.70633 -0.8649 -0.56118\nday 0.11626 0.53897 -0.39514 -0.26027 0.57706 -0.79198 -0.88374 0.30119\ncountry -0.13531 0.15485 -0.07309 0.034013 -0.054457 -0.20541 -0.60086 -0.22407\nunder 0.13721 -0.295 -0.05916 -0.59235 0.02301 0.21884 -0.34254 -0.70213\nsuch 0.61012 0.33512 -0.53499 0.36139 -0.39866 0.70627 -0.18699 -0.77246\nsecond -0.29809 0.28069 0.087102 0.54455 0.70003 0.44778 -0.72565 0.62309 \n<\/code><\/pre>\n\n<hr>\n\n<p>I split the above embedding matrix and loaded only the <em>words<\/em> in <code>vocab<\/code> which will be our vocabulary and the corresponding vectors in <code>emb<\/code> array.<\/p>\n\n<pre><code>vocab = ['the','like','between','did','just','national','day','country','under','such','second']\n\nemb = np.array([[0.418, 0.24968, -0.41242, 0.1217, 0.34527, -0.044457, -0.49688, -0.17862],\n   [0.36808, 0.20834, -0.22319, 0.046283, 0.20098, 0.27515, -0.77127, -0.76804],\n   [0.7503, 0.71623, -0.27033, 0.20059, -0.17008, 0.68568, -0.061672, -0.054638],\n   [0.042523, -0.21172, 0.044739, -0.19248, 0.26224, 0.0043991, -0.88195, 0.55184],\n   [0.17698, 0.065221, 0.28548, -0.4243, 0.7499, -0.14892, -0.66786, 0.11788],\n   [-1.1105, 0.94945, -0.17078, 0.93037, -0.2477, -0.70633, -0.8649, -0.56118],\n   [0.11626, 0.53897, -0.39514, -0.26027, 0.57706, -0.79198, -0.88374, 0.30119],\n   [-0.13531, 0.15485, -0.07309, 0.034013, -0.054457, -0.20541, -0.60086, -0.22407],\n   [ 0.13721, -0.295, -0.05916, -0.59235, 0.02301, 0.21884, -0.34254, -0.70213],\n   [ 0.61012, 0.33512, -0.53499, 0.36139, -0.39866, 0.70627, -0.18699, -0.77246 ],\n   [ -0.29809, 0.28069, 0.087102, 0.54455, 0.70003, 0.44778, -0.72565, 0.62309 ]])\n\n\nemb.shape\n# (11, 8)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Embedding Lookup in TensorFlow<\/strong><\/p>\n\n<p>Now we will see how can we perform <em>embedding lookup<\/em> for some arbitrary input sentence.<\/p>\n\n<pre><code>In [54]: from collections import OrderedDict\n\n# embedding as TF tensor (for now constant; could be tf.Variable() during training)\nIn [55]: tf_embedding = tf.constant(emb, dtype=tf.float32)\n\n# input for which we need the embedding\nIn [56]: input_str = \"like the country\"\n\n# build index based on our `vocabulary`\nIn [57]: word_to_idx = OrderedDict({w:vocab.index(w) for w in input_str.split() if w in vocab})\n\n# lookup in embedding matrix &amp; return the vectors for the input words\nIn [58]: tf.nn.embedding_lookup(tf_embedding, list(word_to_idx.values())).eval()\nOut[58]: \narray([[ 0.36807999,  0.20834   , -0.22318999,  0.046283  ,  0.20097999,\n         0.27515   , -0.77126998, -0.76804   ],\n       [ 0.41800001,  0.24968   , -0.41242   ,  0.1217    ,  0.34527001,\n        -0.044457  , -0.49687999, -0.17862   ],\n       [-0.13530999,  0.15485001, -0.07309   ,  0.034013  , -0.054457  ,\n        -0.20541   , -0.60086   , -0.22407   ]], dtype=float32)\n<\/code><\/pre>\n\n<p>Observe how we got the <em>embeddings<\/em> from our original embedding matrix (with words) using the <em>indices of words<\/em> in our vocabulary.<\/p>\n\n<p>Usually, such an embedding lookup is performed by the first layer (called <em>Embedding layer<\/em>) which then passes these embeddings to RNN\/LSTM\/GRU layers for further processing.<\/p>\n\n<hr>\n\n<p><em>Side Note<\/em>: Usually the vocabulary will also have a special <code>unk<\/code> token. So, if a token from our input sentence is not present in our vocabulary, then the index corresponding to <strong><code>unk<\/code><\/strong> will be looked up in the embedding matrix.<\/p>\n\n<hr>\n\n<p><strong>P.S.<\/strong> Note that <code>embedding_dimension<\/code> is a hyperparameter that one has to tune for their application but popular models like <strong><a href=\"https:\/\/en.wikipedia.org\/wiki\/Word2vec\" rel=\"noreferrer\">Word2Vec<\/a><\/strong> and <strong><a href=\"https:\/\/nlp.stanford.edu\/projects\/glove\/\" rel=\"noreferrer\">GloVe<\/a><\/strong> uses <code>300<\/code> dimension vector for representing each word.<\/p>\n\n<p><strong>Bonus Reading<\/strong> <a href=\"http:\/\/mccormickml.com\/2016\/04\/19\/word2vec-tutorial-the-skip-gram-model\/\" rel=\"noreferrer\">word2vec skip-gram model<\/a><\/p>\n"},{"AnswerId":"48572802","CreationDate":"2018-02-01T22:28:32.923","ParentId":null,"OwnerUserId":"4585592","Title":null,"Body":"<p>Since I was also intrigued by this function, I'll give my two cents.<\/p>\n\n<p>The way I see it in the 2D case is just as a matrix multiplication (it's easy to generalize to other dimensions).<\/p>\n\n<p>Consider a vocabulary with N symbols.\nThen, you can represent a symbol <strong><em>x<\/em><\/strong> as a vector of dimensions Nx1, one-hot-encoded.<\/p>\n\n<p>But you want a representation of this symbol not as a vector of Nx1, but as one with dimensions Mx1, called <strong><em>y<\/em><\/strong>.<\/p>\n\n<p>So, to transform <strong><em>x<\/em><\/strong> into <strong><em>y<\/em><\/strong>, you can use and embedding matrix <strong>E<\/strong>, with dimensions MxN: <\/p>\n\n<p><strong><em>y<\/em><\/strong> = <strong>E<\/strong> <strong><em>x<\/em><\/strong>.<\/p>\n\n<p>This is essentially what tf.nn.embedding_lookup(params, ids, ...) is doing, with the nuance that <em>ids<\/em> are just one number that represents the position of the 1 in the one-hot-encoded vector <strong><em>x<\/em><\/strong>.<\/p>\n"},{"AnswerId":"41922877","CreationDate":"2017-01-29T16:03:01.607","ParentId":null,"OwnerUserId":"4949974","Title":null,"Body":"<p>Yes, this function is hard to understand, until you get the point.<\/p>\n\n<p>In its simplest form, it is similar to <code>tf.gather<\/code>. It returns the elements of <code>params<\/code> according to the indexes specified by <code>ids<\/code>.<\/p>\n\n<p>For example (assuming you are inside <code>tf.InteractiveSession()<\/code>)<\/p>\n\n<pre><code>params = tf.constant([10,20,30,40])\nids = tf.constant([0,1,2,3])\nprint tf.nn.embedding_lookup(params,ids).eval()\n<\/code><\/pre>\n\n<p>would return <code>[10 20 30 40]<\/code>, because the first element (index 0) of params is <code>10<\/code>, the second element of params (index 1) is <code>20<\/code>, etc.<\/p>\n\n<p>Similarly, <\/p>\n\n<pre><code>params = tf.constant([10,20,30,40])\nids = tf.constant([1,1,3])\nprint tf.nn.embedding_lookup(params,ids).eval()\n<\/code><\/pre>\n\n<p>would return <code>[20 20 40]<\/code>.<\/p>\n\n<p>But <code>embedding_lookup<\/code> is more than that. The <code>params<\/code> argument can be a <strong>list<\/strong> of tensors, rather than a single tensor.<\/p>\n\n<pre><code>params1 = tf.constant([1,2])\nparams2 = tf.constant([10,20])\nids = tf.constant([2,0,2,1,2,3])\nresult = tf.nn.embedding_lookup([params1, params2], ids)\n<\/code><\/pre>\n\n<p>In such a case, the indexes, specified in <code>ids<\/code>, correspond to elements of tensors according to a <strong>partition strategy<\/strong>, where the default partition strategy is 'mod'.<\/p>\n\n<p>In the 'mod' strategy, index 0 corresponds to the first element of the first tensor in the list. Index 1 corresponds to the <strong>first<\/strong> element of the <strong>second<\/strong> tensor. Index 2 corresponds to the <strong>first<\/strong> element of the <strong>third<\/strong> tensor, and so on. Simply index <code>i<\/code> corresponds to the first element of the (i+1)th tensor , for all the indexes <code>0..(n-1)<\/code>, assuming params is a list of <code>n<\/code> tensors.<\/p>\n\n<p>Now, index <code>n<\/code> cannot correspond to tensor n+1, because the list <code>params<\/code> contains only <code>n<\/code> tensors. So index <code>n<\/code> corresponds to the <strong>second<\/strong> element of the first tensor. Similarly, index <code>n+1<\/code> corresponds to the second element of the second tensor, etc.<\/p>\n\n<p>So, in the code<\/p>\n\n<pre><code>params1 = tf.constant([1,2])\nparams2 = tf.constant([10,20])\nids = tf.constant([2,0,2,1,2,3])\nresult = tf.nn.embedding_lookup([params1, params2], ids)\n<\/code><\/pre>\n\n<p>index 0 corresponds to the first element of the first tensor: 1<\/p>\n\n<p>index 1 corresponds to the first element of the second tensor: 10<\/p>\n\n<p>index 2 corresponds to the second element of the first tensor: 2<\/p>\n\n<p>index 3 corresponds to the second element of the second tensor: 20<\/p>\n\n<p>Thus, the result would be:<\/p>\n\n<pre><code>[ 2  1  2 10  2 20]\n<\/code><\/pre>\n"},{"AnswerId":"45724451","CreationDate":"2017-08-16T23:31:13.600","ParentId":null,"OwnerUserId":"3907250","Title":null,"Body":"<p>Adding to Asher Stern's answer,\n <code>params<\/code> is\n  interpreted as a <strong>partitioning<\/strong> of a large embedding tensor. It can be a  single tensor representing the complete embedding tensor,\n      or a list of X tensors all of same shape except for the first dimension,\n      representing sharded embedding tensors. <\/p>\n\n<p>The function <code>tf.nn.embedding_lookup<\/code> is written considering the fact that embedding (params) will be large. Therefore we need <code>partition_strategy<\/code>.<\/p>\n"},{"AnswerId":"46965588","CreationDate":"2017-10-27T00:05:57.550","ParentId":null,"OwnerUserId":"6126756","Title":null,"Body":"<p>Another way to look at it is , assume that you flatten out the tensors to one dimensional array, and then you are performing a lookup<\/p>\n\n<p>(eg) Tensor0=[1,2,3], Tensor1=[4,5,6], Tensor2=[7,8,9]<\/p>\n\n<p>The flattened out tensor will be as follows\n[1,4,7,2,5,8,3,6,9]<\/p>\n\n<p>Now when you do a lookup of [0,3,4,1,7] it will yeild [1,2,5,4,6]<\/p>\n\n<p>(i,e) if lookup value is 7 for example , and we have 3 tensors (or a tensor with 3 rows) then, <\/p>\n\n<p>7 \/ 3 : (Reminder is 1, Quotient is 2) So 2nd element of Tensor1 will be shown, which is 6 <\/p>\n"},{"AnswerId":"47482287","CreationDate":"2017-11-25T03:15:07.273","ParentId":null,"OwnerUserId":"8185944","Title":null,"Body":"<p>When the params tensor is in high dimensions, the ids only refers to top dimension. Maybe it's obvious to most of people but I have to run the following code to understand that:<\/p>\n\n<pre><code>embeddings = tf.constant([[[1,1],[2,2],[3,3],[4,4]],[[11,11],[12,12],[13,13],[14,14]],\n                          [[21,21],[22,22],[23,23],[24,24]]])\nids=tf.constant([0,2,1])\nembed = tf.nn.embedding_lookup(embeddings, ids, partition_strategy='div')\n\nwith tf.Session() as session:\n    result = session.run(embed)\n    print (result)\n<\/code><\/pre>\n\n<p>Just trying the 'div' strategy and for one tensor, it makes no difference.<\/p>\n\n<p>Here is the output:<\/p>\n\n<pre><code>[[[ 1  1]\n  [ 2  2]\n  [ 3  3]\n  [ 4  4]]\n\n [[21 21]\n  [22 22]\n  [23 23]\n  [24 24]]\n\n [[11 11]\n  [12 12]\n  [13 13]\n  [14 14]]]\n<\/code><\/pre>\n"},{"AnswerId":"34877590","CreationDate":"2016-01-19T13:05:06.647","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p><code>embedding_lookup<\/code> function retrieves rows of the <code>params<\/code> tensor. The behavior is similar to using indexing with arrays in numpy. E.g.<\/p>\n\n<pre><code>matrix = np.random.random([1024, 64])  # 64-dimensional embeddings\nids = np.array([0, 5, 17, 33])\nprint matrix[ids]  # prints a matrix of shape [4, 64] \n<\/code><\/pre>\n\n<p><code>params<\/code> argument can be also a list of tensors in which case the <code>ids<\/code> will be distributed among the tensors. For example, given a list of 3 tensors <code>[2, 64]<\/code>, the default behavior is that they will represent <code>ids<\/code>: <code>[0, 3]<\/code>, <code>[1, 4]<\/code>, <code>[2, 5]<\/code>. <\/p>\n\n<p><code>partition_strategy<\/code> controls the way how the <code>ids<\/code> are distributed among the list. The partitioning is useful for larger scale problems when the matrix might be too large to keep in one piece.<\/p>\n"},{"AnswerId":"51694027","CreationDate":"2018-08-05T11:46:33.763","ParentId":null,"OwnerUserId":"1699075","Title":null,"Body":"<p>Here's an image depicting the process of embedding lookup. <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/5wFii.jpg\" alt=\"Image: Embedding lookup process\">\n<br\/>\n<br\/>\nConcisely, it gets the corresponding rows of a embedding layer, specified by a list of IDs and provide that as a tensor. It is achieved through the following process.<\/p>\n\n<ol>\n<li>Define a placeholder <code>lookup_ids = tf.placeholder([10])<\/code><\/li>\n<li>Define a embedding layer <code>embeddings = tf.Variable([100,10],...)<\/code><\/li>\n<li>Define the tensorflow operation <code>embed_lookup = tf.embedding_lookup(embeddings, lookup_ids)<\/code><\/li>\n<li>Get the results by running <code>lookup = session.run(embed_lookup, feed_dict={lookup_ids:[95,4,14]})<\/code><\/li>\n<\/ol>\n"}]}
{"QuestionId":34871429,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-01-19T08:06:53.143","AcceptedAnswerId":null,"OwnerUserId":3355076.0,"Title":"How do you actually apply a trained model?","Body":"<p>I've been slowly going through the tensorflow tutorials, and I assume I will have to again. I don't have a background in ML but am slowly pushing my way up. <\/p>\n\n<p>Anyway, after reading through the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/index.html\" rel=\"noreferrer\">RNN tutorial<\/a> and running the training code, I am confused. <\/p>\n\n<p>How does one actually apply the trained model so that it can be used to make language predictions?<\/p>\n\n<p>I know this is a terrible noobish and simple question, but I believe it will be of use to others, as I have seen it asked and not answered in a satisfactory way.<\/p>\n","answers":[{"AnswerId":"34874233","CreationDate":"2016-01-19T10:23:02.190","ParentId":null,"OwnerUserId":"1479778","Title":null,"Body":"<p>In general, when you train a model, you first do a forward pass, and then a backward pass. The forward pass makes a prediction based on your input data, and the backward pass adjust your model based on how correct your prediction was.<\/p>\n\n<p>So when you want to apply your model, you just do a forward pass with your new data as input.<\/p>\n\n<p>In your particular example, using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"noreferrer\">this code<\/a>, you can see how it's done by looking at how they run the test set, starting line 286.<\/p>\n\n<pre><code># They instantiate the model with is_training=False\nmtest = PTBModel(is_training=False, config=eval_config)\n\n# Then they can do a forward pass\ntest_perplexity = run_epoch(session, mtest, test_data, tf.no_op())\nprint(\"Test Perplexity: %.3f\" % test_perplexity)\n<\/code><\/pre>\n\n<p>And if you want the actual prediction and not the perplexity, it is the state in the run_epoch function :<\/p>\n\n<pre><code>cost, state, _ = session.run([m.cost, m.final_state, eval_op],\n                             {m.input_data: x,\n                              m.targets: y,\n                              m.initial_state: state})\n<\/code><\/pre>\n"}]}
{"QuestionId":34872174,"AnswerCount":0,"Tags":"<python-3.x><ssl><openssl><anaconda><keras>","CreationDate":"2016-01-19T08:46:53.020","AcceptedAnswerId":null,"OwnerUserId":1256369.0,"Title":"Python 3.5 - ssl.SSLEOFError: EOF occurred in violation of protocol (_ssl.c:645)","Body":"<p>I am using  and trying to load  dataset.<\/p>\n\n<p>Link to the website is: <a href=\"https:\/\/s3.amazonaws.com\/img-datasets\/mnist.pkl.gz\" rel=\"nofollow\">https:\/\/s3.amazonaws.com\/img-datasets\/mnist.pkl.gz<\/a><\/p>\n\n<p>It works in a browser. But when I call 'mnist.load_data()', I get the exception: <\/p>\n\n<pre><\/pre>\n\n<p>There are many solutions already provided, I tried almost all of them. <\/p>\n\n<p>Below is the stacktrace: <\/p>\n\n<pre><\/pre>\n\n<p>My 'openssl version' on console yeilds:<\/p>\n\n<pre><\/pre>\n\n<p>In python console:<\/p>\n\n<pre><\/pre>\n\n<p>I am using Python 3.5 that came with anaconda<\/p>\n\n<p>I also upgraded OpenSSL using \nbut the error persists.<\/p>\n\n<p>I also tried a patch:<\/p>\n\n<pre><\/pre>\n\n<p>But still no resolution. How can I get rid of this error?<\/p>\n","answers":[]}
{"QuestionId":34872524,"AnswerCount":1,"Tags":"<python><matlab><theano><conv-neural-network>","CreationDate":"2016-01-19T09:04:04.197","AcceptedAnswerId":35216269.0,"OwnerUserId":null,"Title":"How to solve: \" ValueError: Not enough dimensions on Elemwise{neg,no_inplace}.0 to reduce on axis 5000 \"?","Body":"<p>I am trying to classify images of size 166x166 (Matlab cropped) using the CNN code  (Convolutional_mlp.py )  which is taught in the DeepLearning tutorial. The input is a .mat file with X_2k.mat for the input image pixel values and y_2k.mat with the output labels for those images. For the time being there are 2115 images and their labels which we are giving as training data, test data and validation data. I have converted the numpy nd array obtained from the .mat files, into the corresponding Tensor shared variables of the appropriate data type i.e. converted from uint8 matrix to float64 matrix for X and into an int32 vector for Y. (Also commented this part in the code), in order to solve the error \"Cannot convert tensor type (uint8,matrix) into tensor type(float 64,matrix) in case of X \", which occured when giving .mat input.  There are 3 classes for the image, 0, 1 and 2. But need help in resolving the error: \"ValueError: Not enough dimensions on Elemwise{neg,no_inplace}.0 to reduce on axis 5000\"<\/p>\n\n<pre><\/pre>\n\n<p>OUTPUT: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35216269","CreationDate":"2016-02-05T04:24:28.537","ParentId":null,"OwnerUserId":"2446478","Title":null,"Body":"<p>There is a builtin function called ,'min' which is getting overridden because you have imported everything from tensor. On line 339, try to use __ builtin__ .min as a function call.<\/p>\n"}]}
{"QuestionId":34873645,"AnswerCount":1,"Tags":"<theano><convolution>","CreationDate":"2016-01-19T09:57:39.950","AcceptedAnswerId":34907245.0,"OwnerUserId":5583416.0,"Title":"theano: conv3d2d error while doing 3d convolution","Body":"<p>i tried to use conv3d2d for making 3d CNN. I get error below(my code is at <a href=\"http:\/\/pastebin.com\/YaJ9yD4T\" rel=\"nofollow\">enter link description here<\/a>):<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\/home\/shome\/workspace_temp\/conv3d_test\/conv3d_test.py\", line 1, in  from convnet3d import ConvLayer<\/p>\n  \n  <p>File \"\/home\/shome\/softwares\/theano\/Theano-3D-ConvNet-master\/convnet3d\/convnet3d.py\", line 12, in  from conv3d2d import conv3d<\/p>\n  \n  <p>File \"\/home\/shome\/softwares\/theano\/Theano-3D-ConvNet-master\/convnet3d\/conv3d2d.py\", line 298, in  make_gpu_optimizer(DiagonalSubtensor, [0])<\/p>\n  \n  <p>File \"\/home\/shome\/softwares\/theano\/Theano-3D-ConvNet-master\/convnet3d\/conv3d2d.py\", line 266, in make_gpu_optimizer @theano.gof.local_optimizer([])<\/p>\n  \n  <p>File \"\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/gof\/opt.py\", line 948, in decorator\n      raise ValueError, <\/p>\n  \n  <p>(\"Use None instead of an empty list to apply to all nodes.\", <\/p>\n  \n  <p>ValueError: (<strong>'Use None instead of an empty list to apply to all nodes.', 'conv3d2d', 'local_to_gpu'<\/strong>).<\/p>\n<\/blockquote>\n\n<p>My CNN construction is as below:<\/p>\n\n<pre><\/pre>\n\n<p>I think the error is in instantiating the Convlayer. Can anyone help?<\/p>\n","answers":[{"AnswerId":"34907245","CreationDate":"2016-01-20T18:00:04.770","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>Your <code>conv3d2d.py<\/code> files seems to have been copied from an earlier version of Theano, and the syntax for <code>@theano.gof.local_optimizer<\/code> has changed since.<\/p>\n\n<p>If you look at the <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nnet\/conv3d2d.py#L314\" rel=\"nofollow\">updated version<\/a> in the latest master, the call to the decorator has been changed from <code>@theano.gof.local_optimizer([])<\/code> to <code>@theano.gof.local_optimizer([op, cuda.gpu_from_host])<\/code>.<\/p>\n\n<p>Applying that change alone may not be enough, though, so you may be better off importing <code>conv3d2d<\/code> from Theano rather than your repository, or updating the whole file.<\/p>\n"}]}
{"QuestionId":34875881,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-19T11:40:38.257","AcceptedAnswerId":34877160.0,"OwnerUserId":1495939.0,"Title":"Initializing new variables in tensorflow","Body":"<p>I have built and trained a model. On second phase I want to replace the last two layers and retrain them using different data.\nI constantly get the errors for not initializing variables even though I did run initialization on the new vars:<\/p>\n\n<p><\/p>\n\n<p>I understand I have to initialize the new optimizer (ADAMSolever) as well, but \nnot sure how to do that.<\/p>\n\n<p>Assuming I want to replace the optimizer (and other variables) in the middle how do I initialize it without trashing already trained variables?<\/p>\n","answers":[{"AnswerId":"34877160","CreationDate":"2016-01-19T12:44:37.890","ParentId":null,"OwnerUserId":"3568055","Title":null,"Body":"<p>You can get all the trainable variables using tf.trainable_variables(), and exclude the variables which should be restored from the pretrained model. Then you can initialize the other variables.<\/p>\n"}]}
{"QuestionId":34875944,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-01-19T11:44:35.850","AcceptedAnswerId":37573411.0,"OwnerUserId":5576325.0,"Title":"How to write a custom loss function in Tensorflow?","Body":"<p>I am new to . I want to write my own custom loss function. Is there any tutorial about this? For example, the hinge loss or a sum_of_square_loss(though this is already in tf)?\nCan I do it directly in python or I have to write the cpp code?<\/p>\n","answers":[{"AnswerId":"43960630","CreationDate":"2017-05-14T05:16:22.243","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Almost in all tensorflow tutorials they use custom functions. For example in the very <a href=\"https:\/\/www.tensorflow.org\/get_started\/get_started\" rel=\"noreferrer\">beginning tutorial<\/a> they write a custom function: <\/p>\n\n<blockquote>\n  <p>sums the squares of the deltas between the current model and the provided data<\/p>\n<\/blockquote>\n\n<pre><code>squared_deltas = tf.square(linear_model - y)\nloss = tf.reduce_sum(squared_deltas)\n<\/code><\/pre>\n\n<p>In the next <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"noreferrer\">MNIST for beginners they use a cross-entropy<\/a>:<\/p>\n\n<pre><code>cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))\n<\/code><\/pre>\n\n<hr>\n\n<p>As you see it is not that hard at all: you just need to encode your function in a tensor-format and use their basic functions.<\/p>\n\n<p>For example here is how <strong>you can implement F-beta score<\/strong> (a general approach to <a href=\"https:\/\/en.wikipedia.org\/wiki\/F1_score\" rel=\"noreferrer\">F1 score<\/a>). Its formula is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PUuJP.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PUuJP.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The only thing we will need to do is to find how to calculate true_positive, false_positive, false_negative for boolean or 0\/1 values. If you have vectors of 0\/1 values, you can calculate each of the values as:<\/p>\n\n<pre><code>TP = tf.count_nonzero(actual, predicted)\nFP = tf.count_nonzero((actual - 1) * predicted)\nFN = tf.count_nonzero((predicted - 1) * actual)\n<\/code><\/pre>\n\n<p>Now once you know these values you can easily get your <\/p>\n\n<pre><code>denom = (1 + b**2) * TP + b**2 TN + FP\nFb = (1 + b**2) * TP \/ denom\n<\/code><\/pre>\n"},{"AnswerId":"34877704","CreationDate":"2016-01-19T13:11:06.247","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>In addition to the other answer, you can write a loss function in Python if it can be represented as a composition of existing functions.<\/p>\n\n<p>Take a look, for example, at the implementation of <code>sigmoid_cross_entropy_with_logits<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/nn.py#L238\">link<\/a>, which is implemented using basic transformations.<\/p>\n"},{"AnswerId":"37573411","CreationDate":"2016-06-01T15:50:09.563","ParentId":null,"OwnerUserId":"6371086","Title":null,"Body":"<p>We need to write down the loss function. For example, we can use basic mean square error as our loss function for predicted y and target y_:<\/p>\n\n<pre><code> loss_mse = 1\/n(Sum((y-y_)^2))\n<\/code><\/pre>\n\n<p>There are basic functions for tensors like <code>tf.add(x,y)<\/code>, <code>tf.sub(x,y)<\/code>, <code>tf.square(x)<\/code>, <code>tf.reduce_sum(x)<\/code>,<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/index.html\" rel=\"noreferrer\"> etc<\/a>.<\/p>\n\n<p>Then we can define our loss function in Tensorflow like:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.square(tf.sub(y,y_)))\n<\/code><\/pre>\n\n<p>Note: y and y_ are tensors.<\/p>\n\n<p>Moreover, we can define any other loss functions if we can write down the equations. For some training operators (minimizers), the loss function should satisfy some conditions (smooth, differentiable ...). <\/p>\n\n<p>In one word, Tensorflow define arrays, constants, variables into tensors, define calculations using tf functions, and use session to run though graph. We can define whatever we like and run it in the end.<\/p>\n"}]}
{"QuestionId":34876494,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-19T12:11:07.500","AcceptedAnswerId":34887370.0,"OwnerUserId":1783952.0,"Title":"Tensorflow: var.eval() vs. var.initialized_value().eval()","Body":"<p>When I run the following Tensorflow commands I get different results. Anybody knows why?<\/p>\n\n<pre><\/pre>\n\n<p>Produces:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34887370","CreationDate":"2016-01-19T21:35:29.923","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As you correctly <a href=\"https:\/\/stackoverflow.com\/a\/34876842\/3574081\">worked out<\/a>, the difference between evaluating a variable directly and evaluating <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.initialized_value\" rel=\"nofollow noreferrer\"><code>var.initialized_value()<\/code><\/a> is that evaluating <code>var.initialized_value()<\/code> will re-execute <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.initializer\" rel=\"nofollow noreferrer\"><code>var.initializer<\/code><\/a>, and&mdash;as a side effect&mdash;modify the value stored in <code>var<\/code>.<\/p>\n\n<p>This leaves the question: why do we have <code>initialized_value<\/code> at all? The reason is that it helps when defining one variable in terms of another. For example, let's say we want to initialize two variables to the same random value. By default tf.global_variables_initializer does not specify the order in which variables are initialized. Therefore, if the initial value of a variable depends on another variable's value, it's likely that you'll get an error. The following code will not work reliably:<\/p>\n\n<pre><code>v1 = tf.Variable(tf.truncated_normal([20, 20]))\nv2 = tf.Variable(v1)\n\ninit_op = tf.global_variables_initializer()\nsess = tf.Session()\n# The following line will non-deterministically crash with an error about\n# using an uninitialized value.\nsess.run(init_op)\n<\/code><\/pre>\n\n<p>Instead, you should define <code>v2<\/code> in terms of <code>v1.initialized_value()<\/code>. This imposes an order on how the initializers may be executed, and ensures that <code>v1<\/code> is initialized first:<\/p>\n\n<pre><code>v1 = tf.Variable(tf.truncated_normal([20, 20]))\nv2 = tf.Variable(v1.initialized_value())\n\ninit_op = tf.global_variables_initializer()\nsess = tf.Session()\n# The following line will succeed.\nsess.run(init_op)\n<\/code><\/pre>\n"},{"AnswerId":"34876842","CreationDate":"2016-01-19T12:27:45.437","ParentId":null,"OwnerUserId":"1783952","Title":null,"Body":"<p>Ok - easy. The reason for the different outputs is that the initialized_value() re-runs the initialization operation. Thus different results.<\/p>\n\n<p>The initialized_value() method returns the value used to initialize the variable (in this case a random number).<\/p>\n\n<p>I was looking to copy the values of a variable after some operations and ran into the initialized_value() method as a way to copy variables. However, this seems to only be true if you want to copy the initial values. It sounds kind of obvious from the name though...<\/p>\n\n<pre><code>import tensorflow as tf\n\nsess = tf.Session()\nvar = tf.Variable([[1.234]])\nsess.run(tf.initialize_all_variables())\n\nprint var.eval(session=sess)\nprint var.initialized_value().eval(session=sess)\n\nvar.assign_add([[2]]).eval(session=sess)\n\nprint var.eval(session=sess)\nprint var.initialized_value().eval(session=sess)\n<\/code><\/pre>\n\n<p>produces:<\/p>\n\n<pre><code>[[ 1.23399997]]\n[[ 1.23399997]]\n\n[[ 3.23399997]]\n[[ 1.23399997]]\n<\/code><\/pre>\n"}]}
{"QuestionId":34877523,"AnswerCount":8,"Tags":"<python><tensorflow>","CreationDate":"2016-01-19T13:01:41.973","AcceptedAnswerId":34881060.0,"OwnerUserId":1441121.0,"Title":"In TensorFlow, what is tf.identity used for?","Body":"<p>I've seen  used in a few places, such as the official CIFAR-10 tutorial and the batch-normalization implementation on stackoverflow, but I don't see why it's necessary.<\/p>\n\n<p>What's it used for? Can anyone give a use case or two?<\/p>\n\n<p>One proposed answer is that it can be used for transfer between the CPU and GPU. This is not clear to me. Extension to the question, based on <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"noreferrer\">this<\/a>:  is under the GPU block, which suggests to me that all operators defined in  are mapped to the GPU. Then, at the end of , we see  before it's returned. Why? What would be the flaw with not using  here?<\/p>\n","answers":[{"AnswerId":"55795328","CreationDate":"2019-04-22T13:41:05.410","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>I see this kind of hack to check assert:<\/p>\n\n<pre><code>assertion = tf.assert_equal(tf.shape(image)[-1], 3, message=\"image must have 3 color channels\")\nwith tf.control_dependencies([assertion]):\n    image = tf.identity(image)\n<\/code><\/pre>\n\n<p>Also it's used just to give a name:<\/p>\n\n<pre><code>image = tf.identity(image, name='my_image')\n<\/code><\/pre>\n"},{"AnswerId":"39141615","CreationDate":"2016-08-25T09:33:22.640","ParentId":null,"OwnerUserId":"3312142","Title":null,"Body":"<p>I came across another use case that is not completely covered by the other answers.<\/p>\n\n<pre><code>def conv_layer(input_tensor, kernel_shape, output_dim, layer_name, decay=None, act=tf.nn.relu):\n    \"\"\"Reusable code for making a simple convolutional layer.\n    \"\"\"\n    # Adding a name scope ensures logical grouping of the layers in the graph.\n    with tf.name_scope(layer_name):\n        # This Variable will hold the state of the weights for the layer\n        with tf.name_scope('weights'):\n            weights = weight_variable(kernel_shape, decay)\n            variable_summaries(weights, layer_name + '\/weights')\n        with tf.name_scope('biases'):\n            biases = bias_variable([output_dim])\n            variable_summaries(biases, layer_name + '\/biases')\n        with tf.name_scope('convolution'):\n            preactivate = tf.nn.conv2d(input_tensor, weights, strides=[1, 1, 1, 1], padding='SAME')\n            biased = tf.nn.bias_add(preactivate, biases)\n            tf.histogram_summary(layer_name + '\/pre_activations', biased)\n        activations = act(biased, 'activation')\n        tf.histogram_summary(layer_name + '\/activations', activations)\n        return activations\n<\/code><\/pre>\n\n<p>Most of the time when constructing a convolutional layer, you just want the activations returned so you can feed those into the next layer. Sometimes, however - for example when building an auto-encoder - you want the pre-activation values.<\/p>\n\n<p>In this situation an elegant solution is to pass <code>tf.identity<\/code> as the activation function, effectively not activating the layer.<\/p>\n"},{"AnswerId":"34881060","CreationDate":"2016-01-19T15:47:12.883","ParentId":null,"OwnerUserId":"1441121","Title":null,"Body":"<p>After some stumbling I think I've noticed a single use case that fits all the examples I've seen. If there are other use cases, please elaborate with an example.<\/p>\n\n<p>Use case:<\/p>\n\n<p>Suppose you'd like to run an operator every time a particular Variable is evaluated. For example, say you'd like to add one to <code>x<\/code> every time the variable <code>y<\/code> is evaluated. It might seem like this will work:<\/p>\n\n<pre><code>x = tf.Variable(0.0)\nx_plus_1 = tf.assign_add(x, 1)\n\nwith tf.control_dependencies([x_plus_1]):\n    y = x\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as session:\n    init.run()\n    for i in xrange(5):\n        print(y.eval())\n<\/code><\/pre>\n\n<p>It doesn't: it'll print 0, 0, 0, 0, 0. Instead, it seems that we need to add a new node to the graph within the <code>control_dependencies<\/code> block. So we use this trick:<\/p>\n\n<pre><code>x = tf.Variable(0.0)\nx_plus_1 = tf.assign_add(x, 1)\n\nwith tf.control_dependencies([x_plus_1]):\n    y = tf.identity(x)\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as session:\n    init.run()\n    for i in xrange(5):\n        print(y.eval())\n<\/code><\/pre>\n\n<p>This works: it prints 1, 2, 3, 4, 5.<\/p>\n\n<p>If in the CIFAR-10 tutorial we dropped <code>tf.identity<\/code>, then <code>loss_averages_op<\/code> would never run.<\/p>\n"},{"AnswerId":"48695926","CreationDate":"2018-02-08T22:37:25.217","ParentId":null,"OwnerUserId":"8755319","Title":null,"Body":"<p>When our input data is serialized in bytes, and we want to extract features from this dataset. We can do so in key-value format and then get a placeholder for it. Its benefits are more realised when there are multiple features and each feature has to be read in different format.<\/p>\n\n<pre><code>  #read the entire file in this placeholder      \n  serialized_tf_example = tf.placeholder(tf.string, name='tf_example')\n\n  #Create a pattern in which data is to be extracted from input files\n  feature_configs = {'image': tf.FixedLenFeature(shape=[256], dtype=tf.float32),\/\n                     'text': tf.FixedLenFeature(shape=[128], dtype=tf.string),\/\n                     'label': tf.FixedLenFeature(shape=[128], dtype=tf.string),}\n\n  #parse the example in key: tensor dictionary\n  tf_example = tf.parse_example(serialized_tf_example, feature_configs)\n\n  #Create seperate placeholders operation and tensor for each feature\n  image = tf.identity(tf_example['image'], name='image')\n  text  = tf.identity(tf_example['text'], name='text')\n  label = tf.identity(tf_example['text'], name='label')\n<\/code><\/pre>\n"},{"AnswerId":"47958446","CreationDate":"2017-12-24T06:07:44.440","ParentId":null,"OwnerUserId":"5307226","Title":null,"Body":"<p>In addition to the above, I simply use it when I need to assign a name to ops that do not have a name argument, just like when initializing a state in RNN's:<\/p>\n\n<pre><code>rnn_cell = tf.contrib.rnn.MultiRNNCell([cells])\n# no name arg\ninitial_state = rnn_cell.zero_state(batch_size,tf.float32)\n# give it a name with tf.identity()\ninitial_state = tf.identity(input=initial_state,name=\"initial_state\")\n<\/code><\/pre>\n"},{"AnswerId":"34877802","CreationDate":"2016-01-19T13:15:11.870","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p><code>tf.identity<\/code> is useful when you want to explicitly transport tensor between devices (like, from GPU to a CPU).\nThe op adds send\/recv nodes to the graph, which make a copy when the devices of the input and the output are different.<\/p>\n\n<p>A default behavior is that the send\/recv nodes are added implicitly when the operation happens on a different device but you can imagine some situations (especially in a multi-threaded\/distributed settings) when it might be useful to fetch the value of the variable multiple times within a single execution of the <code>session.run<\/code>. <code>tf.identity<\/code> allows for more control with regard to when the value should be read from the source device. Possibly a more appropriate name for this op would be <code>read<\/code>.<\/p>\n\n<p>Also, please note that in the implementation of <code>tf.Variable<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/variables.py#L205\" rel=\"noreferrer\">link<\/a>, the identity op is added in the constructor, which makes sure that all the accesses to the variable copy the data from the source only once. Multiple copies can be expensive in cases when the variable lives on a GPU but it is read by multiple CPU ops (or the other way around). Users can change the behavior with multiple calls to <code>tf.identity<\/code> when desired.<\/p>\n\n<p>EDIT: Updated answer after the question was edited.<\/p>\n\n<p>In addition, <code>tf.identity<\/code> can be used used as a dummy node to update a reference to the tensor. This is useful with various control flow ops. In the CIFAR case we want to enforce that the ExponentialMovingAverageOp will update relevant variables before retrieving the value of the loss. This can be implemented as:<\/p>\n\n<pre><code>with tf.control_dependencies([loss_averages_op]):\n  total_loss = tf.identity(total_loss)\n<\/code><\/pre>\n\n<p>Here, the <code>tf.identity<\/code> doesn't do anything useful aside of marking the <code>total_loss<\/code> tensor to be ran after evaluating <code>loss_averages_op<\/code>.<\/p>\n"},{"AnswerId":"44719124","CreationDate":"2017-06-23T10:25:16.777","ParentId":null,"OwnerUserId":"3333803","Title":null,"Body":"<p>I found another application of tf.identity in Tensorboard.\nIf you use tf.shuffle_batch, it returns multiple tensors at once, so you see messy picture when visualizing the graph, you can't split tensor creation pipeline from actiual input tensors: <a href=\"https:\/\/i.stack.imgur.com\/UwAaH.png\" rel=\"nofollow noreferrer\">messy<\/a><\/p>\n\n<p>But with tf.identity you can create duplicate nodes, which don't affect computation flow: <a href=\"https:\/\/i.stack.imgur.com\/LJs84.png\" rel=\"nofollow noreferrer\">nice<\/a><\/p>\n"},{"AnswerId":"47969514","CreationDate":"2017-12-25T14:29:40.623","ParentId":null,"OwnerUserId":"2017102","Title":null,"Body":"<p>In distribution training, we should use tf.identity or the workers will hang at waiting for initialization of the chief worker:<\/p>\n\n<pre><code>vec = tf.identity(tf.nn.embedding_lookup(embedding_tbl, id)) * mask\nwith tf.variable_scope(\"BiRNN\", reuse=None):\n    out, _ = tf.nn.bidirectional_dynamic_rnn(fw, bw, vec, sequence_length=id_sz, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>For details, without identity, the chief worker would treat some variables as local variables inappropriately and the other workers wait for an initialization operation that can not end<\/p>\n"}]}
{"QuestionId":34877998,"AnswerCount":1,"Tags":"<python><pip><theano>","CreationDate":"2016-01-19T13:24:51.837","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"ImportError: No module named theano","Body":"<p>I use the introduction in the deeplearning tutorials. <\/p>\n\n<pre><\/pre>\n\n<p>There is no problem with numpy and scipy.test:<\/p>\n\n<pre><\/pre>\n\n<p>but when i use the the theano test:<\/p>\n\n<pre><\/pre>\n\n<p>It told me \"no module named theano\" which confused me a lot.<\/p>\n","answers":[{"AnswerId":"43921607","CreationDate":"2017-05-11T17:00:46.723","ParentId":null,"OwnerUserId":"5810495","Title":null,"Body":"<p>In order to be certain of installing the version that will match your interpreter's and not pip's version of Python, you could try the following:<\/p>\n\n<pre><code>sudo -H python -m pip install Theano\n<\/code><\/pre>\n\n<p>Which would load the pip module corresponding to your interpreter.<\/p>\n\n<p>If you're not a system administrator of your computer, or you just prefer to avoid installing python packages system-wide, you can do the following<\/p>\n\n<pre><code>pip install --user Theano\n<\/code><\/pre>\n\n<p>Third possibility would be to use a <a href=\"https:\/\/virtualenv.pypa.io\/en\/stable\/\" rel=\"nofollow noreferrer\">virtualenv<\/a><\/p>\n\n<pre><code>virtualenv env\nsource env\/bin\/activate\npip install Theano\n<\/code><\/pre>\n"}]}
{"QuestionId":34880558,"AnswerCount":4,"Tags":"<python><python-2.7><python-3.x><theano>","CreationDate":"2016-01-19T15:22:51.683","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"Installing Python 2.x and python 3.x on the same computer","Body":"<p>There are python 2.7 and python 3.2 on my computer. The default version is 2.7 because using python -V gives 2.7 as the version.  <\/p>\n\n<p>But when I use apt-get install numpy, scipy,pip why it install them into the python3.2 folder. After that I used pip to install the module into the 3.2 folder.   <\/p>\n\n<p>I also installed Theano this way but in the end it showed a message saying that there is no module named Theano installed although it is in the python 3.2 folder.<\/p>\n","answers":[{"AnswerId":"49637196","CreationDate":"2018-04-03T18:50:31.630","ParentId":null,"OwnerUserId":"6470761","Title":null,"Body":"<p>You can also use Anaconda for maintaining two versions of Python. Anaconda  consists of various libraries so you don't have to install them and after switching it to the different version of Python you can easily install them :<\/p>\n\n<ol>\n<li><p>Download Anaconda for both Python versions<\/p><\/li>\n<li><p>Open <code>.bashrc<\/code><\/p><\/li>\n<li><p>Add the path to new Anaconda you have installed for, e.g.:<\/p>\n\n<pre><code>export PATH=\"\/home\/paras\/anaconda3\/bin:$PATH\"\n<\/code><\/pre><\/li>\n<li><p>Now there will be 2 export paths: one for Python 2 and one for Python 3. Comment the one which you don't want.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"34886270","CreationDate":"2016-01-19T20:23:11.490","ParentId":null,"OwnerUserId":"3957677","Title":null,"Body":"<p>Depending on what you're doing with python, it's often a good idea to run in a virtual environment, this lets you have several different versions of python with several different sets of installed packages on the same system. . . <\/p>\n\n<p>See <a href=\"http:\/\/docs.python-guide.org\/en\/latest\/dev\/virtualenvs\/\" rel=\"nofollow\">http:\/\/docs.python-guide.org\/en\/latest\/dev\/virtualenvs\/<\/a> for the details.<\/p>\n"},{"AnswerId":"52030730","CreationDate":"2018-08-26T22:34:03.263","ParentId":null,"OwnerUserId":"1619003","Title":null,"Body":"<p>First install different versions of python or whichever python version you would like to use<\/p>\n\n<pre><code>sudo update-alternatives --install \/usr\/bin\/python python \/usr\/bin\/python2.6 10\nsudo update-alternatives --install \/usr\/bin\/python python \/usr\/bin\/python2.7 20\nsudo update-alternatives --install \/usr\/bin\/python python \/usr\/bin\/python3.3 30\nsudo update-alternatives --install \/usr\/bin\/python python \/usr\/bin\/python3.4 40\n<\/code><\/pre>\n\n<p>Now create virtualenv like this and give the path of python version you want to use inside the virtualenv.<\/p>\n\n<pre><code>virtualenv -p \/usr\/bin\/python3.6 &lt;foldername&gt;\nvirtualenv -p \/usr\/bin\/python3.4 &lt;foldername&gt;\n<\/code><\/pre>\n"},{"AnswerId":"34882530","CreationDate":"2016-01-19T16:54:10.170","ParentId":null,"OwnerUserId":"2867317","Title":null,"Body":"<p>You haven't told on which OS you're running this, but it look likes a debian base linux, maybe ubuntu?<\/p>\n\n<p>If so, I'd try with:\n<code>sudo apt-get install python3-numpy<\/code> or \n<code>sudo apt-get install python2-numpy<\/code>.<br>\nThis would also work with <code>python-pip2<\/code> and <code>python-pip3<\/code>.  <\/p>\n\n<p>After this, you could effectively use \"pip2\" or \"pip3\" to install your packages without having to go through the OS \"prebuild\" modules (but the os version of the packages are usually my prefered way to install them, if the exists in the repo)<\/p>\n"}]}
{"QuestionId":34880748,"AnswerCount":1,"Tags":"<python><pickle><theano>","CreationDate":"2016-01-19T15:33:02.963","AcceptedAnswerId":37006422.0,"OwnerUserId":5631237.0,"Title":"Theano's pkl_utils Dump Function not available in Theano 0.7?","Body":"<p>I want to save a model that I have trained.\nSince it is using shared variables (like Weights, bias and so on) and since it should be readable on machines without Theano installed, I wanted to use the theano.misc.pkl_utils.dump() function.\nHowever, it seems as if that is only installed in bleeding edge installations (the current github file looks different than my local one).<\/p>\n\n<p>Is that really the case? And why is the description in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/misc\/pkl_utils.html#theano.misc.pkl_utils.dump\" rel=\"nofollow\">docs<\/a> then?<\/p>\n\n<p>I am using theano 0.7.0 and I'm seriously confused about this.\nIf that feature is not yet available (I can't install bleeding edge right now), what are other ways? I'm sure that I am not the only one trying to save a trained model the easiest way possible ;-)<\/p>\n\n<p>Thank you a lot,<\/p>\n\n<p>Roman<\/p>\n","answers":[{"AnswerId":"37006422","CreationDate":"2016-05-03T14:02:49.527","ParentId":null,"OwnerUserId":"6194082","Title":null,"Body":"<p>If you train your model using Theano, the parameters of the model will be eventually <code>shared<\/code> variables (probably a list of <code>shared<\/code> variables if the network consists of several layers). It is possible to <em>pickle<\/em> the list of <code>shared<\/code> variables and <em>unpickle<\/em> it later. However you might have problems to <em>unpickle<\/em> such variables in another machine, e.g. with no Theano installation or if you train in a GPU-capable machine that generates <em>CudaNdarrays<\/em> and then you want to load back the model in a non-GPU-capable machine. What I recommend you is the following: convert every shared variable of the list of parameters into a numpy <code>ndarray<\/code>:<\/p>\n\n<p><code>params_numpy = [numpy.asarray(p.get_value()) for p in params]<\/code><\/p>\n\n<p>where <code>params<\/code> is a list of <code>shared<\/code> variables. Then you can safely <em>pickle<\/em>\/<em>unpickle<\/em> <code>params_numpy<\/code>.<\/p>\n"}]}
{"QuestionId":34881572,"AnswerCount":1,"Tags":"<optimization><machine-learning><tensorflow><word2vec>","CreationDate":"2016-01-19T16:10:11.450","AcceptedAnswerId":null,"OwnerUserId":5188903.0,"Title":"Bad performance for hogwild-style SGD in tensorflow's word2vec","Body":"<p>Tensorflow's gradient decent optimizer GradientDescentOptimizer does not use locking by default. If the optimizer is invoked in multiple threads in parallel - as for example <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/embedding\/word2vec.py\" rel=\"nofollow\">google's word2vec example<\/a> - we get a hogwild-style lock-free asynchronous optimization schema. <\/p>\n\n<p>In theory the number of updates per time (e.g. words per second in word2vec) should scale pretty much linearly. However, in the word2vec example, there is no performance gain for more than, say, three or four CPUs. I got almost no improvement by going to 16 or 32 cores. Also with large batch_sizes such as 512. CPU usage goes to up but is in parts spent on kernel operations according to htop.<\/p>\n\n<p>Now I am wondering where the overhead comes from? If things do not lock, why don't we get at least a small performance gain?<\/p>\n\n<p>PS: Let's keep performance in terms of accuracy aside. Large batch sizes for word2vec.py are a bad idea, I just take it as an example...<\/p>\n","answers":[{"AnswerId":"34885294","CreationDate":"2016-01-19T19:23:58.123","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>I'm assuming that you are increasing the number of 'concurrent steps' when you talk about 'increasing the number of cores'.  Even running one concurrent step in TensorFlow will try to use all the cores on your system.  Let me know if that's not what you meant.<\/p>\n\n<p>It's not guaranteed that you will get better performance by increasing the number of concurrent steps.   Concurrent steps only really helps when there are gaps in the execution of a single graph that can be filled by work from other graphs.  If all cores are always busy, increasing concurrent steps won't improve the number of samples per second processed.<\/p>\n\n<p>It may be the case that the kernels used in word2vec could be either better parallelized (more effectively using the cores available on the system), or simply more optimized in general.<\/p>\n"}]}
{"QuestionId":34883029,"AnswerCount":0,"Tags":"<ubuntu><caffe>","CreationDate":"2016-01-19T17:17:31.383","AcceptedAnswerId":null,"OwnerUserId":1556826.0,"Title":"I am facing some errors while installing Caffe on Ubuntu 14.04","Body":"<p>I am trying to install Caffe on a system running Ubuntu 14.04. When I run <em>make all<\/em>, it goes seemingly well till I face the error shown in the screenshot. Any help on resolving it would be much appreciated.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/KoPiT.png\" alt=\"Screenshot containing the error\"><\/p>\n","answers":[]}
{"QuestionId":34883771,"AnswerCount":1,"Tags":"<cmake><caffe><atlas><linuxbrew>","CreationDate":"2016-01-19T17:55:54.320","AcceptedAnswerId":34885978.0,"OwnerUserId":1859485.0,"Title":"linuxbrew installed cmake on a cluster cannot find lib in \/usr\/lib64","Body":"<p>I'm trying to build caffe on a GGPU cluster. I've installed a lot of dependencies in a subfolder of my home using linuxbrew. One of those dependencies is an updated version of cmake(needed).<\/p>\n\n<p>When i launch the command <\/p>\n\n<pre><\/pre>\n\n<p>i get this output<\/p>\n\n<pre><\/pre>\n\n<p>and it seems there is nothing i can do to link the ATLAS libs correctly.\nThe missing libs are under <\/p>\n\n<pre><\/pre>\n\n<p>Is it possible that there exists a limitation such that my linuxbrewed cmake cannot link to upper folders?<\/p>\n\n<p>In the error message it says that:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to modify that file adding the correct paths:<\/p>\n\n<pre><\/pre>\n\n<p>and nothing happened.   <\/p>\n\n<p>i tried to change the PATH and LD_LIBRARY_PATH:<\/p>\n\n<pre><\/pre>\n\n<p>and nothing happened.<\/p>\n\n<p>I'm not a linux veteran so i apologize if i missed some important information. Any suggestion is appreciated.<\/p>\n\n<p>EDIT:\nthis is the content of \/usr\/lib64\/atlas3-sse3:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34885978","CreationDate":"2016-01-19T20:05:58.853","ParentId":null,"OwnerUserId":"1859485","Title":null,"Body":"<p>It seems that lapack is bundled in atlas, for this reason FindAtlas.cmake is not able to find  alapack_r, alapck or lapack_atlas.<\/p>\n\n<p>It is enough to edit caffe_folder\/cmake\/Module\/FindAtlas.cmake, look for Atlas_LAPACK_LIBRARY NAMES and add clapack to the list of libs.<\/p>\n\n<p>It completely solve the problem.<\/p>\n"}]}
{"QuestionId":34883885,"AnswerCount":3,"Tags":"<python><eclipse><macos><pydev><tensorflow>","CreationDate":"2016-01-19T18:02:46.900","AcceptedAnswerId":34887922.0,"OwnerUserId":373386.0,"Title":"TensorFlow with Eclipse","Body":"<p>I'm trying to run the TensorFlow Python examples for with Eclipse Mars for Mac OS X, but I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The error is raised on the 3rd line: <\/p>\n\n<pre><\/pre>\n\n<p>I'm new to Python and TensorFlow. I'm using PyDev for Eclipse. I read somewhere that I can add code to the \"External Libraries\" folder but I'm not sure where the TensorFlow library is.<\/p>\n\n<p>I installed TensorFlow following the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html\" rel=\"nofollow\">installation directions on the TensorFlow website<\/a>.<\/p>\n\n<p>What am I missing? <\/p>\n","answers":[{"AnswerId":"36362964","CreationDate":"2016-04-01T18:04:15.190","ParentId":null,"OwnerUserId":"6146290","Title":null,"Body":"<ol>\n<li>install tensorflow.<\/li>\n<li>in eclipse , assume you have pydev installed, go to preferences>pydev>python interpreter.<\/li>\n<li>in the libraries tab add the path to the lib folder of your tensorflow lib folder.<\/li>\n<\/ol>\n\n<p>here is screenshot:\n<a href=\"https:\/\/i.stack.imgur.com\/BCZQH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BCZQH.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"42428258","CreationDate":"2017-02-23T23:42:13.297","ParentId":null,"OwnerUserId":"5484902","Title":null,"Body":"<p>I also had import issues. I fixed it by creating a new project and specifying interpreter as ~\/tensorflow\/lib\/python2.7<\/p>\n\n<p>Eclipse seems to then point to the right interpreter and offers to import other relevant files - result is similar to answer above.<\/p>\n\n<p>The answer from @mrry above was helpful - only the link to the 5 step guide seems to be unavailable. \nHope this helps! <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FsEKe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FsEKe.png\" alt=\"Screenshot\"><\/a><\/p>\n"},{"AnswerId":"34887922","CreationDate":"2016-01-19T22:11:02.970","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It sounds like you are running into a problem with your Python environment. The easiest way to deal with this is to follow these steps:<\/p>\n\n<ol>\n<li><p>Install TensorFlow into a virtualenv, by following the instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#virtualenv-installation\" rel=\"noreferrer\">here<\/a>. This avoids the need to acquire root privileges, and makes it easier to experiment with different versions of TensorFlow. (It also avoids conflicts with other system-installed versions of different libraries.)<\/p><\/li>\n<li><p>Set up PyDev to use your new virtualenv, by following step 5 onwards in <a href=\"http:\/\/ntraft.com\/eclipse-with-pydev-and-virtualenv\/\" rel=\"noreferrer\">this guide<\/a>. This involves setting up a new \"interpreter\" whose library will be the set of libraries in the virtualenv. You may need to recreate your Eclipse project to use this interpreter.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":34886605,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><tensorflow><conv-neural-network>","CreationDate":"2016-01-19T20:45:15.323","AcceptedAnswerId":34886873.0,"OwnerUserId":3314143.0,"Title":"TensorFlow network not training?","Body":"<p>I had a really weird time with TensorFlow the last days and can not think of what's going wrong atm.<\/p>\n\n<p>I have built this network: <a href=\"http:\/\/pastebin.com\/GZaX4aYN\" rel=\"noreferrer\">link<\/a>. It is a copy of TensorFlow's MNIST example.<\/p>\n\n<p>Basically, what I did, is altering the network from taking 28x28x1 images (MNIST greyscale) to 600x800x1 images (some images I took on my own, webcam with a relatively clean background and one object at different locations).<\/p>\n\n<p>What I wanted to do is playing around with such a CNN and have it output the x-location of the object in the image. So one single output neuron.<\/p>\n\n<p>However, no matter what I tried, the network is always outputting 1.0. Even (when you look at my testing section at the end of the code) when I feed all ones, all zeros or all random numbers into the network.<\/p>\n\n<p>Of course, since I have only 21 labeled training and 7 labeled test pictures I expected the performance to be really bad (since 800x600 pixel images are huge for neural networks and locating an object isn't easy).<\/p>\n\n<p>...but I have no idea at all why the network always outputs 1.0 even if it is fed with nonsense. Any ideas?<\/p>\n","answers":[{"AnswerId":"34886873","CreationDate":"2016-01-19T21:02:40.617","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Looking at your source code, it appears that your final fully connected layer before the softmax (L.104 in <a href=\"http:\/\/pastebin.com\/GZaX4aYN\" rel=\"noreferrer\">the code you shared<\/a>) reduces each example down to a <strong>single output class<\/strong> before computing the softmax. Since there is only one class for each example, the result of the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#softmax\" rel=\"noreferrer\"><code>tf.nn.softmax()<\/code><\/a> op (<code>y_conv<\/code>) will be a batch_size x 1 matrix containing 1.0 in every element, and the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#argmax\" rel=\"noreferrer\"><code>tf.argmax()<\/code><\/a> of that will  containing 0 for every element, since there is only one value. Similarly, applying <code>tf.argmax()<\/code> to y_train (which is a batch_size x 1 matrix) will yield 0 for every element, so the \"accuracy\" will be 100%.<\/p>\n\n<p>To fix this, you should probably (i) increase the number of output units from the final fully connected layer to be the number of classes, and (ii) encode each row <code>y_train<\/code> as a one-hot vector representing the true class of each example.<\/p>\n"}]}
{"QuestionId":34888235,"AnswerCount":2,"Tags":"<python><random><tensorflow>","CreationDate":"2016-01-19T22:33:06.813","AcceptedAnswerId":34888586.0,"OwnerUserId":5812408.0,"Title":"How do I generate a random vector in TensorFlow and maintain it for further use?","Body":"<p>I am trying to generate a random variable and use it twice. However, when I use it the second time, the generator creates a second random variable that is not identical to the first. Here is code to demonstrate:<\/p>\n\n<pre><\/pre>\n\n<p>I want  and  to be equal. I think this is because the  op gets called twice. Is there a way to use TensorFlow (without using NumPy) to achieve this? <\/p>\n\n<p>(My use case is more complicated, but this is the distilled question.)<\/p>\n","answers":[{"AnswerId":"34888295","CreationDate":"2016-01-19T22:37:56.127","ParentId":null,"OwnerUserId":"442945","Title":null,"Body":"<p>Your question has the same issue as <a href=\"https:\/\/stackoverflow.com\/questions\/13538266\/python-copy-of-a-variable\">this question<\/a>, in that if you call <code>random_uniform<\/code> twice you will get two results, and as such you need to set your second variable to the value of the first. That means that, assuming you are not changing <code>rand_var_1<\/code> later, you can do this:<\/p>\n\n<pre><code>rand_var_1 = tf.random_uniform([5],0,10, dtype = tf.int32, seed = 0)\nrand_var_2 = rand_var_1\n<\/code><\/pre>\n\n<p>But, that said, if you want <code>z1<\/code> and <code>z2<\/code> to be equal, why have separate variables at all? Why not do:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\n# A random variable\nrand_var = tf.random_uniform([5],0,10, dtype = tf.int32, seed = 0)\nop = tf.add(rand_var,rand_var)\n\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    z1_op = sess.run(op)\n    z2_op = sess.run(op)\n    print(z1_op,z2_op)\n<\/code><\/pre>\n"},{"AnswerId":"34888586","CreationDate":"2016-01-19T23:01:31.243","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The current version of your code will randomly generate a new value for <code>rand_var_1<\/code> and <code>rand_var_2<\/code> on each call to <code>sess.run()<\/code> (although since you set the seed to 0, they will have the same value within a single call to <code>sess.run()<\/code>).<\/p>\n\n<p>If you want to retain the value of a randomly-generated tensor for later use, you should assign it to a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a>:<\/p>\n\n<pre><code>rand_var_1 = tf.Variable(tf.random_uniform([5], 0, 10, dtype=tf.int32, seed=0))\nrand_var_2 = tf.Variable(tf.random_uniform([5], 0, 10, dtype=tf.int32, seed=0))\n\n# Or, alternatively:\nrand_var_1 = tf.Variable(tf.random_uniform([5], 0, 10, dtype=tf.int32, seed=0))\nrand_var_2 = tf.Variable(rand_var_1.initialized_value())\n\n# Or, alternatively:\nrand_t = tf.random_uniform([5], 0, 10, dtype=tf.int32, seed=0)\nrand_var_1 = tf.Variable(rand_t)\nrand_var_2 = tf.Variable(rand_t)\n<\/code><\/pre>\n\n<p>...then <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#initialize_all_variables\" rel=\"noreferrer\"><code>tf.initialize_all_variables()<\/code><\/a> will have the desired effect:<\/p>\n\n<pre><code># Op 1\nz1 = tf.add(rand_var_1, rand_var_2)\n\n# Op 2\nz2 = tf.add(rand_var_1, rand_var_2)\n\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(init)        # Random numbers generated here and cached.\n    z1_op = sess.run(z1)  # Reuses cached values for rand_var_1, rand_var_2.\n    z2_op = sess.run(z2)  # Reuses cached values for rand_var_1, rand_var_2.\n    print(z1_op, z2_op)   # Will print two identical vectors.\n<\/code><\/pre>\n"}]}
{"QuestionId":34888735,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-19T23:14:47.780","AcceptedAnswerId":34888832.0,"OwnerUserId":2615676.0,"Title":"Does Saver not support string variables?","Body":"<p>It appears that the builtin checkpointing tools do not support saving  variables? For example this fails:<\/p>\n\n<pre><\/pre>\n\n<p>If so is there any reason for this?<\/p>\n","answers":[{"AnswerId":"34888832","CreationDate":"2016-01-19T23:23:08.637","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>String support was added to (the ops used in) <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow\"><code>tf.train.Saver<\/code><\/a> earlier this month, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/107853df74593ee0d88e9634b4fc21431d9404a7\" rel=\"nofollow\">in this commit<\/a>. This support is not available in the current (0.6.0) release, but it will appear in the next release. If you want to use it in the meantime, you should <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">install TensorFlow from source<\/a>.<\/p>\n"}]}
{"QuestionId":34889260,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-01-20T00:07:22.580","AcceptedAnswerId":null,"OwnerUserId":1525479.0,"Title":"Caffe framework: How can I solve this error?","Body":"<p>I started to use Caffe framework on Linux as non-admin, Caffe was installed on my account. I went to usr\/local\/caffe path and started to run a cifar10 example by .But I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>I have checked the build directory and there is file named there. How can I solve it?<\/p>\n","answers":[{"AnswerId":"35617169","CreationDate":"2016-02-25T02:33:47.897","ParentId":null,"OwnerUserId":"5977856","Title":null,"Body":"<p>You should execute the script 'create_cifar10.sh' from the $CAFFE_ROOT directory.<\/p>\n\n<p>cd $CAFFE_ROOT<\/p>\n\n<p>.\/examples\/cifar10\/create_cifar10.sh<\/p>\n"}]}
{"QuestionId":34889605,"AnswerCount":2,"Tags":"<android><tensorflow><bazel>","CreationDate":"2016-01-20T00:46:34.037","AcceptedAnswerId":34902781.0,"OwnerUserId":3280780.0,"Title":"Running a Tensorflow model on Android","Body":"<p>I'm trying to figure out the workflow for training and deploying a Tensorflow model on Android.  I'm aware of the other questions similar to this one on StackOverflow, but none of them seem to address the problems I've run into.  <\/p>\n\n<p>After studying the Android example from the Tensorflow repository, this is what I think the workflow should be:<\/p>\n\n<ol>\n<li>Build and train Tensorflow model in Python.<\/li>\n<li>Create a new graph, and transfer all relevant nodes (i.e. not the nodes responsible for training) to this new graph.  Trained weight variables are imported as constants so that the C++ API can read them.<\/li>\n<li>Develop Android GUI in Java, using the native keyword to stub out a call to the Tensorflow model.<\/li>\n<li>Run javah to generate the C\/C++ stub code for the Tensorflow native call.<\/li>\n<li>Fill in the stub by using the Tensorflow C++ API to read in and access the trained\/serialized model.<\/li>\n<li>Use Bazel to build BOTH the Java app, the native Tensorflow interface (as a .so file), and generate the APK.<\/li>\n<li><p>Use adb to deploy the APK.<\/p>\n\n<p>Step 6 is the problem.  Bazel will happily compile a native (to OSX) .dylib that I can call from Java via the JNI.  Android Studio, likewise, will generate a whole bunch of XML code that makes the GUI I want.  However, Bazel wants all of the java app code to be inside the 'WORKSPACE' top-level directory (in the Tensorflow repo), and Android Studio immediately links in all sorts of external libraries from the SDK to make GUIs (I know because my Bazel compile run fails when it can't find these resources).  The only way I can find to force Bazel to cross-compile a .so file is by making it a dependent rule of an Android rule.  Directly cross-compiling a native lib is what I'd prefer to porting my A.S. code to a Bazel project.<\/p>\n\n<p>How do I square this?  Bazel will supposedly compile Android code, but Android Studio generates code that Bazel can't compile.  All the examples from Google simply give you code from a repo without any clue as to how it was generated.  As far as I know, the XML that's part of an Android Studio app is supposed to be generated, not made by hand.  If it can be made by hand, how do I avoid the need for all those external libraries? <\/p>\n\n<p>Maybe I'm getting the workflow wrong, or there's some aspect of Bazel\/Android Studio that I'm not understanding.  Any help appreciated.  <\/p><\/li>\n<\/ol>\n\n<p>Thanks!<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>There were several things that I ended up doing that might have contributed to the library building successfully:<\/p>\n\n<ol>\n<li>I upgraded to the latest Bazel.<\/li>\n<li>I rebuilt TensorFlow from source.<\/li>\n<li><p>I implemented the recommended Bazel BUILD file below, with a few additions (taken from the Android example):<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>I haven't verified that this library can be loaded and used in Android yet; Android Studio 1.5 seems to be very finicky about acknowledging the presence of native libs.<\/p>\n","answers":[{"AnswerId":"34902781","CreationDate":"2016-01-20T14:30:31.047","ParentId":null,"OwnerUserId":"4410907","Title":null,"Body":"<p>After setting up an Android NDK in your WORKSPACE file, Bazel can cross-compile a .so for Android, like this:<\/p>\n\n<pre><code>cc_binary(\n    name = \"libfoo.so\",\n    srcs = [\"foo.cc\"],\n    deps = [\":bar\"],\n    linkstatic = 1,\n    linkshared = 1,\n)\n\n$ bazel build foo:libfoo.so \\\n    --crosstool_top=\/\/external:android\/crosstool --cpu=armeabi-v7a \\\n    --host_crosstool_top=@bazel_tools\/\/tools\/cpp:toolchain\n$ file bazel-bin\/foo\/libfoo.so\nbazel-bin\/foo\/libfoo.so: ELF 32-bit LSB  shared object, ARM, EABI5 version 1 (SYSV), dynamically linked (uses shared libs), not stripped\n<\/code><\/pre>\n\n<blockquote>\n  <p>Bazel wants all of the java app code to be inside the 'WORKSPACE'\n  top-level directory (in the Tensorflow repo)<\/p>\n<\/blockquote>\n\n<p>When 0.1.4 is released (pushing it right now) and we have pushed some fixes to TensorFlow and Protobuf, you can start using the TensorFlow repo as a remote repository. After setting it up in your WORKSPACE file, you can then refer to TensorFlow rules using <code>@tensorflow\/\/foo\/bar<\/code> labels.<\/p>\n"},{"AnswerId":"42735706","CreationDate":"2017-03-11T13:21:33.510","ParentId":null,"OwnerUserId":"6474756","Title":null,"Body":"<pre><code>git clone --recurse-submodules https:\/\/github.com\/tensorflow\/tensorflow.git\n<\/code><\/pre>\n\n<p>Note: --recurse-submodules is important to pull submodules.<\/p>\n\n<p>Install Bazel from here. Bazel is the primary build system for TensorFlow.\nNow, edit the WORKSPACE, we can find the WORKSPACE file in the root directory of the TensorFlow that we have cloned earlier.<\/p>\n\n<pre><code># Uncomment and update the paths in these entries to build the Android demo.\n#android_sdk_repository(\n#    name = \"androidsdk\",\n#    api_level = 23,\n#    build_tools_version = \"25.0.1\",\n#    # Replace with path to Android SDK on your system\n#    path = \"&lt;PATH_TO_SDK&gt;\",\n#)\n#\n#android_ndk_repository(\n#    name=\"androidndk\",\n#    path=\"&lt;PATH_TO_NDK&gt;\",\n#    api_level=14)\n<\/code><\/pre>\n\n<p>Like below with our sdk and ndk path:<\/p>\n\n<pre><code>android_sdk_repository(\n    name = \"androidsdk\",\n    api_level = 23,\n    build_tools_version = \"25.0.1\",\n    # Replace with path to Android SDK on your system\n    path = \"\/Users\/amitshekhar\/Library\/Android\/sdk\/\",\n)\nandroid_ndk_repository(\n    name=\"androidndk\",\n    path=\"\/Users\/amitshekhar\/Downloads\/android-ndk-r13\/\",\n    api_level=14)\n<\/code><\/pre>\n\n<p>Then build the .so file.<\/p>\n\n<pre><code>bazel build -c opt \/\/tensorflow\/contrib\/android:libtensorflow_inference.so \\\n   --crosstool_top=\/\/external:android\/crosstool \\\n   --host_crosstool_top=@bazel_tools\/\/tools\/cpp:toolchain \\\n   --cpu=armeabi-v7a\n<\/code><\/pre>\n\n<p>Replacing armeabi-v7a with our desired target architecture.\nThe library will be located at:<\/p>\n\n<pre><code>bazel-bin\/tensorflow\/contrib\/android\/libtensorflow_inference.so\n<\/code><\/pre>\n\n<p>To build the Java counterpart:<\/p>\n\n<pre><code>bazel build \/\/tensorflow\/contrib\/android:android_tensorflow_inference_java\n<\/code><\/pre>\n\n<p>We can find the JAR file at:<\/p>\n\n<pre><code>bazel-bin\/tensorflow\/contrib\/android\/libandroid_tensorflow_inference_java.jar\n<\/code><\/pre>\n\n<p>Now we have both jar and .so file. I have already built both .so file and jar, you can directly use from the <a href=\"https:\/\/github.com\/MindorksOpenSource\/AndroidTensorFlowMachineLearningExample\" rel=\"nofollow noreferrer\">project<\/a>.<\/p>\n\n<p>Put libandroid_tensorflow_inference_java.jar in libs folder and right click and add as library.<\/p>\n\n<pre><code>compile files('libs\/libandroid_tensorflow_inference_java.jar')\n<\/code><\/pre>\n\n<p>Create jniLibs folder in main directory and put libtensorflow_inference.so in jniLibs\/armeabi-v7a\/ folder.<\/p>\n\n<p>Now, we will be able to call TensorFlow Java API.<\/p>\n\n<p>The TensorFlow Java API has exposed all the required methods through a class TensorFlowInferenceInterface.<\/p>\n\n<p>Now, we have to call the TensorFlow Java API with the model path and load it.<\/p>\n\n<p>I have written a complete blog <a href=\"https:\/\/blog.mindorks.com\/android-tensorflow-machine-learning-example-ff0e9b2654cc\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":34890899,"AnswerCount":1,"Tags":"<python><numpy><neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-01-20T03:11:38.950","AcceptedAnswerId":34892643.0,"OwnerUserId":5813490.0,"Title":"How to add dynamically created input images to a RandomShuffleQueue using a QueueRunner in tensor flow","Body":"<p>I am trying to train a CNN with images created during program execution. I have a game environment (not created by me) that generates screen images that depend on actions taken in the game. The actions are controlled by the learnt CNN. <\/p>\n\n<p>These images are then pushed into a RandomShuffleQueue, from which mini batches are dequeued and used to train the CNN on the correct action.\nI would like to do this (game play and training) asynchronously, where the game is being played and screens of it are added to the RandomShuffleQueue in a separate thread that the one used to train the model. Here is a very simplified version of what I am trying.<\/p>\n\n<pre><\/pre>\n\n<p>Meanwhile, the game_env looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the game environment is really simple as of now: every time a screen grab is performed, a counter is incremented and an image filled with the counter (of the correct size) is returned. <\/p>\n\n<p>It should be noted that I wrote the above class just for testing and in general grab_screen can return any numpy nd-array. Moreover it is not written by me so I can just call grab_screen and not make any changes inside the real thing.<\/p>\n\n<p>Now, the problem is experience queue seems to be holding only tensors of ones (i.e. the counter only gets updated once!!)<\/p>\n\n<p>Sample output:<\/p>\n\n<pre><\/pre>\n\n<p>and so on.\nMy question is: how do I dynamically create the input image to be enqueued to the RandomShuffleQueue like this?\nThanks!<\/p>\n","answers":[{"AnswerId":"34892643","CreationDate":"2016-01-20T06:04:55.350","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem can be traced to this line, which defines the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#QueueRunner\" rel=\"noreferrer\"><code>tf.train.QueueRunner<\/code><\/a>:<\/p>\n\n<pre><code>experience_runner = tf.train.QueueRunner(\n    experience, [perceive(game()) for num in range(available_threads)])\n<\/code><\/pre>\n\n<p>This creates four (<code>available_threads<\/code>) ops that, each time any of them runs, will enqueue a tensor filled with 1.0 to the <code>experience<\/code> queue.\nStepping through what happens in the list comprehension should make this clearer. The following happens four times:<\/p>\n\n<ol>\n<li>A <code>game<\/code> object is constructed.<\/li>\n<li>It is passed to <code>perceive()<\/code>.<\/li>\n<li><code>perceive()<\/code> calls <code>game.grab_screen()<\/code> once, which increments the counter, and returns a tensor <code>1 * tf.ones(self.screen_size)<\/code><\/li>\n<li><code>percieve()<\/code> passes this tensor to <code>experience.enqueue()<\/code> and returns the resulting op.<\/li>\n<\/ol>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#QueueRunner.create_threads\" rel=\"noreferrer\"><code>QueueRunner.create_threads()<\/code><\/a> call creates one thread per enqueue op, and these run in an infinite loop (blocking when the queue reaches capacity).<\/p>\n\n<p>To have the desired effect, you should use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#feeding\" rel=\"noreferrer\">feed mechanism and a placeholder<\/a> to pass a different value for the grabbed screen each time you enqueue an experience. It depends on how your <code>game<\/code> class is implemented, but you probably also want to initialize a single instance of that class as well. Finally, it's not clear whether you want multiple enqueuing threads, but let's assume that <code>game.grab_screen()<\/code> is thread-safe and permits some concurrency. Given all this, a plausible version looks like the following (note that you'll need to create a custom thread rather than a <code>QueueRunner<\/code> to use feeding):<\/p>\n\n<pre><code>import tensorflow as tf\nfrom game_env import game\n\nexperience = tf.RandomShuffleQueue(10000,\n                                   1000, tf.float32,\n                                   shapes=[32,32],  \n                                   name='experience_replay')\n\nscreen_placeholder = tf.placeholder(tf.float32, [32, 32])\n# You can create a single enqueue op and dequeued tensor, and reuse these from\n# multiple threads.\nenqueue_op = experience.enqueue(screen_placeholder)\ndequeued_t = experience.dequeue()\n# ...\n\ninit_op = tf.initialize_all_variables()\n\ngame_obj = game()\n\nsess = tf.Session()\ncoord = tf.train.Coordinator()\n\n# Define a custom thread for running the enqueue op that grabs a new\n# screen in a loop and feeds it to the placeholder.\ndef enqueue_thread():\n    with coord.stop_on_exception():\n        while not coord.should_stop():\n            screen_val = game_obj.grab_screen()\n            # Run the same op, but feed a different value for the screen.\n            sess.run(enqueue_op, feed_dict={screen_placeholder: screen_val}) \n\navailable_threads = 4\nfor _ in range(available_threads):\n    threading.Thread(target=enqueue_thread).start()\n\n\nwhile True:\n    # N.B. It's more efficient to reuse the same dequeue op in a loop.\n    print sess.run(dequeued_t)\n    time.sleep(0.5)\n<\/code><\/pre>\n"}]}
{"QuestionId":34890925,"AnswerCount":1,"Tags":"<python><iteration><theano>","CreationDate":"2016-01-20T03:14:46.750","AcceptedAnswerId":34960747.0,"OwnerUserId":5128473.0,"Title":"How to map one matrix value to another in theano function","Body":"<p>I want to implement the following function in theano function,<\/p>\n\n<pre><\/pre>\n\n<p>I got the error \nDo I have to use scan? or is there any simpler way?\nThanks!<\/p>\n","answers":[{"AnswerId":"34960747","CreationDate":"2016-01-23T07:17:18.883","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Since b is of type <code>ndarray<\/code>, I'll assume every <code>b_row<\/code> has the same length.  <\/p>\n\n<p>If I understood correctly the code swaps the order of columns in <code>b<\/code> according to <code>dictx<\/code>, and pads the non-specified columns with zeros.<\/p>\n\n<p>The main problem is Theano doesn't have a dictionary-like data structure (please let me know if there's one).  <\/p>\n\n<p>Because in your example the dictionary keys and values are integers within <code>range(len(b_row))<\/code>, one way to work around this is to construct a vector that uses indices as keys (if some index should not be contained in the dictionary, make its value -1).  <\/p>\n\n<p>The same idea should apply for mapping elements of a matrix in general, and there're certainly other (better) ways of doing this.<\/p>\n\n<p>Here is the code.<br>\nNumpy:<\/p>\n\n<pre><code>dictx = {0:1,1:2}\nb = numpy.asarray([[1,2,3],\n                [4,5,6],\n                [7,8,9]])\na = numpy.array([[b_row[dictx[idx]] if idx in dictx else 0 for idx in range(len(b_row))] for b_row in b])\nprint a\n<\/code><\/pre>\n\n<p>Theano:<\/p>\n\n<pre><code>dictx = theano.shared(numpy.asarray([1,2,-1]))\nb = tensor.matrix()\na = tensor.switch(tensor.eq(dictx, -1), tensor.zeros_like(b), b[:,dictx])\nfn = theano.function([b],a)\nprint fn(numpy.asarray([[1,2,3],\n                        [4,5,6],\n                        [7,8,9]]))\n<\/code><\/pre>\n\n<p>They both print:  <\/p>\n\n<pre><code>[[2 3 0]  \n [5 6 0]  \n [8 9 0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":34892006,"AnswerCount":0,"Tags":"<python><machine-learning><theano><tensorflow>","CreationDate":"2016-01-20T05:09:44.210","AcceptedAnswerId":null,"OwnerUserId":5738942.0,"Title":"Tensorflow: Diagonal Subtensor for 3D Convolutional NN","Body":"<p>Theano uses a workaround for a 3D convolutional neural (CNN) network using a variation on a 2D CNN:<\/p>\n\n<p><a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nnet\/conv3d2d.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nnet\/conv3d2d.py<\/a><\/p>\n\n<p>It relies on being able to grab the diagonal of the matrix and reshape it. From the website above:<\/p>\n\n<pre><\/pre>\n\n<p>Can this be done in Tensorflow? This seems like a relatively easy task that can still use the vast majority of Tensorflow's CNN framework while handling 3D datatypes.<\/p>\n","answers":[]}
{"QuestionId":34893578,"AnswerCount":1,"Tags":"<c++><cmake><undefined-reference><caffe><leveldb>","CreationDate":"2016-01-20T07:07:23.083","AcceptedAnswerId":null,"OwnerUserId":5814306.0,"Title":"Undefined reference to leveldb when compiling Caffe","Body":"<p>My os is ubuntu15.10. When I using CMake to compile Caffe I got those error:<\/p>\n\n<pre><\/pre>\n\n<p>I had installed libleveldb-dev yet. <\/p>\n\n<p>And if I make USE_LEVELDB \"Build with levelDB\" become OFF in CMakeLists, compile would succeed, but the examples can't be run.<\/p>\n","answers":[{"AnswerId":"34898434","CreationDate":"2016-01-20T11:10:39.037","ParentId":null,"OwnerUserId":"5814306","Title":null,"Body":"<p>Ok... 2 hours ago I fixed this problem by myself. I give up using CMake and made some changes in <em>makefile.configure<\/em> : <\/p>\n\n<pre><code>INCLUDE_DIRS := $(PYTHON_INCLUDE) \/usr\/local\/include \/usr\/lib\/x86_64-linux-gnu\/hdf5\/serial\/include \/home\/vic\/leveldb\/include\/leveldb \/opt\/OpenBLAS\/include\nLIBRARY_DIRS := $(PYTHON_LIB) \/usr\/local\/lib \/usr\/lib \/home\/vic\/leveldb \/usr\/lib\/x86_64-linux-gnu \/usr\/lib\/x86_64-linux-gnu\/hdf5\/serial \/opt\/OpenBLAS\/lib\n<\/code><\/pre>\n\n<p>I just add a series of addresses of libraries which I thought may be. \n And finally I successfully compile Caffe by <em>make all<\/em> with levelDB and successfully run mnist. She may really can't find the libraries if they were not at the same place as earlier versions.<\/p>\n"}]}
{"QuestionId":34894653,"AnswerCount":0,"Tags":"<plot><caffe><training-data><pycaffe>","CreationDate":"2016-01-20T08:09:40.743","AcceptedAnswerId":null,"OwnerUserId":2693257.0,"Title":"Calculate accuracy on both train and test","Body":"<p>I have 200 training data with 6000 features and labels are (0, +1) that have been converted to  format. (Like that, i have 77 testing data).<\/p>\n\n<p>I am working on plotting statistics like the accuracy over iterations. I use parse_log and gnuplot for that. Now my problem is that I need accuracy for both testset and trainset. In other words, I want that the TEST phase will be applied on both databases and as a result give me two accuracy numbers in order to see how the two courves evolve.\nMy  in <\/p>\n\n<blockquote>\n  <p>train_val.prototxt<\/p>\n<\/blockquote>\n\n<p>is:<\/p>\n\n<pre><\/pre>\n\n<p>And My in solver.prototxt is:<\/p>\n\n<pre><\/pre>\n\n<p>And accuracy for both train and test is:<\/p>\n\n<blockquote>\n  <p>(test_accuracy)<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DhfK5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DhfK5.png\" alt=\"test_acc\"><\/a><\/p>\n\n<blockquote>\n  <p>And train Accuracy is:<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1MUSG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1MUSG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>I wonder why train accuracy is not change in iterations (in TRAIN phase)<\/strong>.<\/p>\n","answers":[]}
{"QuestionId":34897622,"AnswerCount":2,"Tags":"<machine-learning><statistics><neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-01-20T10:35:22.403","AcceptedAnswerId":34900153.0,"OwnerUserId":3314143.0,"Title":"TensorFlow: best method for non-one-hot vectors?","Body":"<p>Currently I have a CNN predicting classes fed with one-hot  vectors like this:<\/p>\n\n<pre><\/pre>\n\n<p>So  would return 5, since this is the right class.<\/p>\n\n<p>Unfortunately I have far to less training data for the network to classify my test data correctly. The training error gets better over time, but I think this is overfitting. The test error is always really bad.<\/p>\n\n<p>I thought a normal distribution would make the network's life easier, so instead of saying class 5 is the only right class and all other classes are equally wrong, I want to give the network a high penalty if it predicts class 0 instead of 5 but just a small penalty if it predicts 4 instead of 5. I thought  for vectors encoding class 5 could look like this then:<\/p>\n\n<pre><\/pre>\n\n<p>The question is, <strong>which method for calculating the cost would be appropriate?<\/strong><\/p>\n\n<p>Would:<\/p>\n\n<pre><\/pre>\n\n<p>Still produce reasonable results, even if the  vectors aren't one-hot encoded anymore?<\/p>\n\n<p>Or does this require stronger architecture changes? Currently I have two Conv\/Pool-Layers and two fully connected ones. The output is simply , since  applies the softmax.<\/p>\n\n<p>My actual architecture has an output of 800 dimensional output vectors, so having very few training data for now the network nearly never manages to hit one single class out of 800 correctly in the test data, only in the training data (with overfitting).<\/p>\n\n<p>However, I would be totally ok if the network predicts the class in a range like +\/- 20. So if class 400 is the right one, predicting a class between 380 and 420 would be enough for me.<\/p>\n","answers":[{"AnswerId":"36680139","CreationDate":"2016-04-17T17:58:10.720","ParentId":null,"OwnerUserId":"492849","Title":null,"Body":"<blockquote>\n  <p>However, I would be totally ok if the network predicts the class in a range like +\/- 20. So if class 400 is the right one, predicting a class between 380 and 420 would be enough for me.<\/p>\n<\/blockquote>\n\n<p>I think the obvious thing to do here is to make your labels more coarse, i.e. if you're predicting depth values you can have 100 different values instead of 800. <\/p>\n"},{"AnswerId":"34900153","CreationDate":"2016-01-20T12:30:59.080","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The <code>softmax_cross_entropy_with_logits<\/code> should support your use case, according to documentation.<\/p>\n\n<p>Since you are fine with predicting in a range +\/- 20 and not necessarily the exact labels, then maybe a better loss would be appropriate. For example, MSE between the target <em>value<\/em> (e.g. 400) and your prediction (e.g. 420)? This part sounds more like a research question than related to TensorFlow.<\/p>\n"}]}
{"QuestionId":34898247,"AnswerCount":1,"Tags":"<python><machine-learning><theano><deep-learning><keras>","CreationDate":"2016-01-20T11:02:30.377","AcceptedAnswerId":35017314.0,"OwnerUserId":1256369.0,"Title":"Theano - Keras - No Module named `pool`","Body":"<p>I have installed a , and the following packages in following order:<\/p>\n\n<p>gfortran:<\/p>\n\n<pre><\/pre>\n\n<p>OpenBLAS:<\/p>\n\n<pre><\/pre>\n\n<p>Anaconda, first downloaded Anaconda3-2.4.1-Linux-x86_64.sh, and then:<\/p>\n\n<pre><\/pre>\n\n<p>Then, pydot (after updating):<\/p>\n\n<pre><\/pre>\n\n<p>Them I cloned and installed Theano: <\/p>\n\n<pre><\/pre>\n\n<p>I moved from  to  and got very happy that theano is installed.<\/p>\n\n<p>I run a small script, to verify it is indeed working correctly.<\/p>\n\n<pre><\/pre>\n\n<p>So, it is working, may be a trivial example to prove a point. <\/p>\n\n<p>After that, comes <\/p>\n\n<pre><\/pre>\n\n<p>Then I change to examples directory of keras, and simply type in <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Now, what the hell....<strong>Am I missing some package?<\/strong> <\/p>\n\n<p>I think this is not issue of  but rather <strong>problem with .<\/strong> <\/p>\n\n<p>I went ahead and tried a dirty trick, , and then rerun the above example, but I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>I can provide a stack trace as well, if needed. <\/p>\n\n<p>I have struggled a lot in getting  going, eagerly want to get started... before entire energy drains in the installations, <\/p>\n","answers":[{"AnswerId":"35017314","CreationDate":"2016-01-26T15:20:14.223","ParentId":null,"OwnerUserId":"1537597","Title":null,"Body":"<p>It seems it cannot compile the file <code>theano\/tensor\/signal\/pool.py<\/code> there is a issue and fix on <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/3894\" rel=\"nofollow\">github<\/a><\/p>\n"}]}
{"QuestionId":34900246,"AnswerCount":2,"Tags":"<python><python-2.7><python-multiprocessing><tensorflow>","CreationDate":"2016-01-20T12:34:48.727","AcceptedAnswerId":34900386.0,"OwnerUserId":1978504.0,"Title":"Tensorflow: Passing a session to a python multiprocess","Body":"<p>I'm using tensorflow to preprocess some large images. I was having a problem where the memory was rapidly collapsing. I turned to use multiprocessing in python so the memory would free up entirely whenever I want.<\/p>\n\n<p>The thing is, I'm using python's multiprocess queues and for some reason unknown I can't pass my tensorflow session from my parent process to the children. Using some advanced debugging techniques (i.e. printing something every few lines) I noticed that python just goes idle inside the line where I make use of the session, it doesn't throw an error message.<\/p>\n\n<p>My code looks something like this:<\/p>\n\n<pre><\/pre>\n\n<p>What could be the problem?\nMany thanks!<\/p>\n","answers":[{"AnswerId":"42310383","CreationDate":"2017-02-18T02:23:03.530","ParentId":null,"OwnerUserId":"523503","Title":null,"Body":"<p>All you need is distributed tensorflow.<\/p>\n\n<ol>\n<li>Create the graph and session in the parent process. Place some of the operators(especially, variables) to workers while constructing graph.<\/li>\n<li>Create child processes, and run them<\/li>\n<\/ol>\n"},{"AnswerId":"34900386","CreationDate":"2016-01-20T12:41:59.880","ParentId":null,"OwnerUserId":"86","Title":null,"Body":"<p>I don't think you can share <a href=\"https:\/\/docs.python.org\/2\/library\/multiprocessing.html#sharing-state-between-processes\" rel=\"nofollow\">\"state\"<\/a> as in the <code>tf.Session()<\/code> between processes like that.\nI would think that each process needed it's own session.<\/p>\n"}]}
{"QuestionId":34900719,"AnswerCount":1,"Tags":"<c++><visual-studio><deep-learning><caffe>","CreationDate":"2016-01-20T12:58:13.043","AcceptedAnswerId":34903164.0,"OwnerUserId":4564882.0,"Title":"Unknown pooling method when testing caffe with cuda but not cudnn","Body":"<p>I built the caffe deep learning library in windows as shown in this link:<\/p>\n\n<p><a href=\"https:\/\/initialneil.wordpress.com\/2015\/07\/15\/caffe-vs2013-opencv-in-windows-tutorial-i\/\" rel=\"nofollow noreferrer\">https:\/\/initialneil.wordpress.com\/2015\/07\/15\/caffe-vs2013-opencv-in-windows-tutorial-i\/<\/a><\/p>\n\n<p>I deactivated the cuDNN because my nvidia card didnot support this and changed the targert architecture to fermi architecture.<\/p>\n\n<p>I built caffe as static library to use it in the test project shown below:<\/p>\n\n<pre><\/pre>\n\n<p>But I get an error when accessing the file:  in the function described below:<\/p>\n\n<pre><\/pre>\n\n<p>And get the message  as shown in the window below:\n<a href=\"https:\/\/i.stack.imgur.com\/l4izh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/l4izh.png\" alt=\"Execution of the testing project to classify one given image\"><\/a><\/p>\n\n<p>The normal execution of my project is described in the image below:\n<a href=\"https:\/\/i.stack.imgur.com\/Cxk2b.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Cxk2b.png\" alt=\"enter image description here\"><\/a>\nCould someone give me an idea about the possible solution?<\/p>\n","answers":[{"AnswerId":"34903164","CreationDate":"2016-01-20T14:47:56.320","ParentId":null,"OwnerUserId":"3082081","Title":null,"Body":"<p>The pooling layer which by default should be <code>max pooling<\/code> was translated into some other layers. You might add a breakpoint at <a href=\"https:\/\/github.com\/initialneil\/caffe-vs2013\/blob\/master\/caffe\/src\/caffe\/layers\/pooling_layer.cu#L163\" rel=\"nofollow\">pooling_layer.cu (line 163)<\/a> or add <code>cout &lt;&lt; this-&gt;layer_param_.pooling_param().pool() &lt;&lt; endl;<\/code> before that line to see what pooling layer it was using. I guess it doesn't equal to <code>PoolingParameter_PoolMethod_MAX<\/code> here.<\/p>\n\n<p>I'm not sure why it happened, maybe there some error in the <em>prototxt<\/em> file or the <em>protobuf<\/em>. A <em>brutal trick<\/em> would be overlapping <code>line 206<\/code> with <a href=\"https:\/\/github.com\/initialneil\/caffe-vs2013\/blob\/master\/caffe\/src\/caffe\/layers\/pooling_layer.cu#L165-L176\" rel=\"nofollow\">line 165-176<\/a> in order to force using <code>max pooling<\/code>.<\/p>\n"}]}
{"QuestionId":34902782,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-20T14:30:35.063","AcceptedAnswerId":34910583.0,"OwnerUserId":2213720.0,"Title":"Interpolated sampling of points in an image with TensorFlow","Body":"<p>Given is a grayscale image <strong>I<\/strong> as 2D Tensor (Dimension W,H) and a Tensor of coordinates <strong>C<\/strong> (Dim. None,2). I want to interpret the rows of <strong>C<\/strong> as coordinates in <strong>I<\/strong>, sample <strong>I<\/strong> at those coordinates using some kind of interpolation (bilinear would probably be fine for my use case), and store the resulting values in a new Tensor <strong>P<\/strong> (of dimension None, i.e. 1-dimensional with as many entries as <strong>C<\/strong> has rows).<\/p>\n\n<p>Is this possible (efficiently) with TensorFlow? All I can find are functions for resizing (equidistant resampling if you like) of images. But I can't find anything out-of-the-box to sample at a list of coordinates.<\/p>\n\n<p>I.e. I would have expected to find something like a tf.interpolate() function:<\/p>\n\n<pre><\/pre>\n\n<p>Ideally I would be looking for a solution that would allow me to interpolate in an N dimensional tensor <strong>I<\/strong> along M dimensions using a <strong>C<\/strong> with shape (None, M) and produce an N-M+1 dimensional output, as indicated by the \"axis\" parameter in the code above.<\/p>\n\n<p>(The \"image\" in my application isn't a picture btw., it's sampled data from a physical model (when used as placeholder) or an alternative learned model (when used as variable). Right now this physical model has 2 degrees of freedom, thus interpolating in an \"image\" is sufficient for now, but I might look into higher dimensional models in the future.)<\/p>\n\n<p>In case something like that is not possible with existing TensorFlow features: Where should I start when I'd like to implement something like this tf.interpolate() operator? (documentation and\/or simple example code)<\/p>\n","answers":[{"AnswerId":"34910583","CreationDate":"2016-01-20T21:10:03.703","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is no built-in op that performs this kind of interpolation, but you should be able to do it using a composition of existing TensorFlow ops. I'd suggest the following strategy for the bilinear case:<\/p>\n\n<ol>\n<li><p>From your tensor <code>C<\/code> of indices, compute integer tensors corresponding to the four corner points. For example (with names assuming that the origin is at the top left):<\/p>\n\n<pre><code>top_left = tf.cast(tf.floor(C), tf.int32)\n\ntop_right = tf.cast(\n    tf.concat(1, [tf.floor(C[:, 0:1]), tf.ceil(C[:, 1:2])]), tf.int32)\n\nbottom_left = tf.cast(\n    tf.concat(1, [tf.ceil(C[:, 0:1]), tf.floor(C[:, 1:2])]), tf.int32)\n\nbottom_right = tf.cast(tf.ceil(C), tf.int32)\n<\/code><\/pre><\/li>\n<li><p>From each tensor representing a particular corner point, extract a vector of values from <code>I<\/code> at those points. For example, for the following function does this for the 2-D case:<\/p>\n\n<pre><code>def get_values_at_coordinates(input, coordinates):\n  input_as_vector = tf.reshape(input, [-1])\n  coordinates_as_indices = (coordinates[:, 0] * tf.shape(input)[1]) + coordinates[:, 1]\n  return tf.gather(input_as_vector, coordinates_as_indices)\n\nvalues_at_top_left = get_values_at_coordinates(I, top_left)\nvalues_at_top_right = get_values_at_coordinates(I, top_right)\nvalues_at_bottom_left = get_values_at_coordinates(I, bottom_left)\nvalues_at_bottom_right = get_values_at_coordinates(I, bottom_right)\n<\/code><\/pre><\/li>\n<li><p>Compute the interpolation in the horizontal direction first:<\/p>\n\n<pre><code># Varies between 0.0 and 1.0.\nhorizontal_offset = C[:, 0] - tf.cast(top_left[:, 0], tf.float32)\n\nhorizontal_interpolated_top = (\n    ((1.0 - horizontal_offset) * values_at_top_left)\n    + (horizontal_offset * values_at_top_right))\n\nhorizontal_interpolated_bottom = (\n    ((1.0 - horizontal_offset) * values_at_bottom_left)\n    + (horizontal_offset * values_at_bottom_right))\n<\/code><\/pre><\/li>\n<li><p>Now compute the interpolation in the vertical direction:<\/p>\n\n<pre><code>vertical_offset = C[:, 1] - tf.cast(top_left[:, 1], tf.float32)\n\ninterpolated_result = (\n    ((1.0 - vertical_offset) * horizontal_interpolated_top)\n    + (vertical_offset * horizontal_interpolated_bottom))\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"34910627","CreationDate":"2016-01-20T21:12:39.847","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This turned out to be tricky for nearest neighbor given that TF doesn't have Numpy slicing generality yet (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow noreferrer\">github issue #206<\/a>), and the fact that <code>gather<\/code> only works on first dimension. But here's a way to work around it by using gather->transpose->gather->extract diagonal<\/p>\n\n<pre><code>def identity_matrix(n):\n  \"\"\"Returns nxn identity matrix.\"\"\"\n  # note, if n is a constant node,\u00a0this assert node won't be executed,\n  # this error will be caught during shape analysis \n  assert_op = tf.Assert(tf.greater(n, 0), [\"Matrix size must be positive\"])\n  with tf.control_dependencies([assert_op]):\n    ones = tf.fill(n, 1)\n    diag = tf.diag(ones)\n  return diag\n\ndef extract_diagonal(tensor):\n  \"\"\"Extract diagonal of a square matrix.\"\"\"\n\n  shape = tf.shape(tensor)\n  n = shape[0]\n  assert_op = tf.Assert(tf.equal(shape[0], shape[1]), [\"Can't get diagonal of \"\n                                                       \"a non-square matrix\"])\n\n  with tf.control_dependencies([assert_op]):\n    return tf.reduce_sum(tf.mul(tensor, identity_matrix(n)), [0])\n\n\n# create sample matrix\nsize=4\nI0=np.zeros((size,size), dtype=np.int32)\nfor i in range(size):\n  for j in range(size):\n    I0[i, j] = 10*i+j\n\nI = tf.placeholder(dtype=np.int32, shape=(size,size))\nC = tf.placeholder(np.int32, shape=[None, 2])\nC0 = np.array([[0, 1], [1, 2], [2, 3]])\nrow_indices = C[:, 0]\ncol_indices = C[:, 1]\n\n# since gather only supports dim0, have to transpose\nI1 = tf.gather(I, row_indices)\nI2 = tf.gather(tf.transpose(I1), col_indices)\nI3 = extract_diagonal(tf.transpose(I2))\n\nsess = create_session()\nprint sess.run([I3], feed_dict={I:I0, C:C0})\n<\/code><\/pre>\n\n<p>So starting with a matrix like this:<\/p>\n\n<pre><code>array([[ 0,  1,  2,  3],\n       [10, 11, 12, 13],\n       [20, 21, 22, 23],\n       [30, 31, 32, 33]], dtype=int32)\n<\/code><\/pre>\n\n<p>This code extracts diagonal one above the main<\/p>\n\n<pre><code>[array([ 1, 12, 23], dtype=int32)]\n<\/code><\/pre>\n\n<p>There's some magic happening with [] operators getting turned into <code>Squeeze<\/code> and <code>Slice<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9y2Z5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9y2Z5.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":34903916,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2016-01-20T15:23:05.943","AcceptedAnswerId":34906637.0,"OwnerUserId":2954187.0,"Title":"How to refer to a built-in op when writing a custom user op","Body":"<p>I'm writing a custom op in tensorflow, and I would like to refer to an op that already exists, either one of my own user ops, or one of the built-in ops. Is there any way to do this without just copy pasting the code from the other op that I'm referring to?<\/p>\n","answers":[{"AnswerId":"34906637","CreationDate":"2016-01-20T17:28:50.607","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The answer to this is \"it depends.\" Only some of the built-in ops are factored in a reusable way. For example, the component-wise ops are implemented as <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/04de87f1204103cf9b753858ecae33099aafb4c2\/tensorflow\/core\/kernels\/cwise_ops.h#L573\" rel=\"nofollow\">reusable functors<\/a>, so you may be able to <code>#include \"tensorflow\/core\/kernels\/cwise_ops.h\"<\/code> and use the <code>UnaryFunctor<\/code> or <code>BinaryFunctor<\/code> classes if your needs to borrow some of their functionality.<\/p>\n\n<p>Most other ops have their implementations inlined into the <code>Compute()<\/code> method of the corresponding <code>OpKernel<\/code>. For your own custom ops, you are free to define the implementation in free functions, or to use inheritance to define common functionality that is shared between ops.<\/p>\n"}]}
{"QuestionId":34903975,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><training-data>","CreationDate":"2016-01-20T15:25:32.900","AcceptedAnswerId":34922529.0,"OwnerUserId":4838304.0,"Title":"How to Create CaffeDB training data for siamese networks out of image directory","Body":"<p>I need some help to create a CaffeDB for siamese CNN out of a plain directory with images and label-text-file. Best would be a python-way to do it.<br>\nThe problem is not to walk through the directory and making pairs of images. My problem is more of making a CaffeDB out of those pairs.<br>\nSo far I only used <a href=\"https:\/\/stackoverflow.com\/a\/31431716\/1714410\"><\/a> to create a CaffeDB out of an image directory.<br>\nThanks for help!<\/p>\n","answers":[{"AnswerId":"34922529","CreationDate":"2016-01-21T11:25:58.777","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Why don't you simply make two datasets using good old <code>convert_imagest<\/code>? <\/p>\n\n<pre><code>layer {\n  name: \"data_a\"\n  top: \"data_a\"\n  top: \"label_a\"\n  type: \"Data\"\n  data_param { source: \"\/path\/to\/first\/data_lmdb\" }\n  ...\n}\nlayer {\n  name: \"data_b\"\n  top: \"data_b\"\n  top: \"label_b\"\n  type: \"Data\"\n  data_param { source: \"\/path\/to\/second\/data_lmdb\" }\n  ...\n}\n<\/code><\/pre>\n\n<p>As for the loss, since every example has a class label you need to convert <code>label_a<\/code> and <code>label_b<\/code> into a <code>same_not_same_label<\/code>. I suggest you do this \"on-the-fly\" using a python layer. In the <code>prototxt<\/code> add the call to python layer:<\/p>\n\n<pre><code>layer {\n  name: \"a_b_to_same_not_same_label\"\n  type: \"Python\"\n  bottom: \"label_a\"\n  bottom: \"label_b\"\n  top: \"same_not_same_label\"\n  python_param { \n    # the module name -- usually the filename -- that needs to be in $PYTHONPATH\n    module: \"siamese\"\n    # the layer name -- the class name in the module\n    layer: \"SiameseLabels\"\n  }\n  propagate_down: false\n}\n<\/code><\/pre>\n\n<p>Create <code>siamese.py<\/code> (make sure it is in your <code>$PYTHONPATH<\/code>). In <code>siamese.py<\/code> you should have the layer class:<\/p>\n\n<pre><code>import sys, os\nsys.path.insert(0,os.environ['CAFFE_ROOT'] + '\/python')\nimport caffe\nclass SiameseLabels(caffe.Layer):\n  def setup(self, bottom, top):\n    if len(bottom) != 2:\n       raise Exception('must have exactly two inputs')\n    if len(top) != 1:\n       raise Exception('must have exactly one output')\n  def reshape(self,bottom,top):\n    top[0].reshape( *bottom[0].shape )\n  def forward(self,bottom,top):\n    top[0].data[...] = (bottom[0].data == bottom[1].data).astype('f4')\n  def backward(self,top,propagate_down,bottom):\n      # no back prop\n      pass\n<\/code><\/pre>\n\n<p>Make sure you shuffle the examples in the two sets in a different manner, so you get non-trivial pairs. Moreover, if you construct the first and second data sets with <em>different<\/em> number of examples, then you will see different pairs at each epoch ;)    <\/p>\n\n<hr>\n\n<p>Make sure you construct the network to share the weights of the duplicated layers, see <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html\" rel=\"nofollow\">this tutorial<\/a> for more information.<\/p>\n"}]}
{"QuestionId":34907657,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2016-01-20T18:24:07.443","AcceptedAnswerId":null,"OwnerUserId":2797644.0,"Title":"scale the loss value according to \"badness\" in caffe","Body":"<p>I want to scale the loss value of each image based on how close\/far is the \"current prediction\" to the \"correct label\" during the training. For example if the correct label is \"cat\" and the network think it is \"dog\" the penalty (loss) should be less than the case if the network thinks it is a \"car\".<\/p>\n\n<p>The way that I am doing is as following:<\/p>\n\n<p>1- I defined a matrix of the distance between the labels,<br>\n2- pass that matrix as a bottom to the  layer,<br>\n3- multiply each log(prob) to this value to scale the loss according to badness in <\/p>\n\n<p>However I do not know what should I do in the  part. I understand the gradient (bottom_diff) has to be changed but not quite sure, how to incorporate the scale value here. According to the math I have to scale the gradient by the scale (because it is just an scale) but don't know how.<\/p>\n\n<p>Also, seems like there is loosLayer in caffe called  that does very similar job if I am not mistaken, however the backward part of this layer is a little confusing:<\/p>\n\n<pre><\/pre>\n\n<p>I am not sure why  is divide by  rather than being multiply by! If I use identity matrix for  isn't it supposed to act like softmax loss in both forward and backward?<\/p>\n\n<p>It will be highly appreciated if someone can give me some pointers.<\/p>\n","answers":[{"AnswerId":"34917052","CreationDate":"2016-01-21T06:46:20.097","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are correct in observing that the scaling you are doing for the <code>log(prob)<\/code> is exactly what <code>\"InfogainLoss\"<\/code> layer is doing (You can read more about it <a href=\"https:\/\/stackoverflow.com\/q\/27632440\/1714410\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/a\/30497907\/1714410\">here<\/a>).<\/p>\n\n<p>As for the derivative (back-prop): the loss computed by this layer is <\/p>\n\n<pre><code>L = - sum_j infogain_mat[label * dim + j] * log( prob(j) )\n<\/code><\/pre>\n\n<p>If you differentiate this expression with respect to <code>prob(j)<\/code> (which is the input variable to this layer), you'll notice that the derivative of <code>log(x)<\/code> is <code>1\/x<\/code> this is why you see that<\/p>\n\n<pre><code>dL\/dprob(j) = - infogain_mat[label * dim + j] \/ prob(j) \n<\/code><\/pre>\n\n<p>Now, <em>why don't you see similar expression in the back-prop of <code>\"SoftmaxWithLoss\"<\/code> layer?<\/em><br>\nwell, as the name of that layer suggests it is actually a  combination of two layers: softmax that computes class probabilities from classifiers outputs <em>and<\/em> a log loss layer on top of it. Combining these two layer enables a more numerically robust estimation of the gradients.<br>\nWorking a little with <code>\"InfogainLoss\"<\/code> layer I noticed that sometimes <code>prob(j)<\/code> can have a very small value leading to unstable estimation of the gradients. <\/p>\n\n<p>Here's a detailed computation of the forward and backward passes of <code>\"SoftmaxWithLoss\"<\/code> and <code>\"InfogainLoss\"<\/code> layers with respect to the <em>raw<\/em> predictions (<code>x<\/code>), rather than the \"softmax\" probabilities derived from these predictions using a softmax layer. You can use these equations to create a <code>\"SoftmaxWithInfogainLoss\"<\/code> layer that is more numerically robust than computing infogain loss on top of a softmax layer:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Epnzb.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Epnzb.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>PS,<br>\nNote that if you are going to use infogain loss for weighing, you should feed <code>H<\/code> (the <code>infogain_mat<\/code>) with label <em>similarities<\/em>, rather than distances.  <\/p>\n\n<p><strong>Update:<\/strong><br>\nI recently implemented this robust gradient computation and created <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3855\" rel=\"nofollow noreferrer\">this pull request<\/a>. This PR was merged to master branch on April, 2017. <\/p>\n"}]}
{"QuestionId":34908033,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-20T18:46:30.263","AcceptedAnswerId":34908326.0,"OwnerUserId":2225493.0,"Title":"ValueError when performing matmul with Tensorflow","Body":"<p>I'm a total beginner to TensorFlow, and I'm trying to multiply two matrices together, but I keep getting an exception that says: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here's minimal example code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Confusingly, the following very similar code works fine:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Can anyone point to what the issue is? I must be missing something obvious here..<\/p>\n","answers":[{"AnswerId":"34908326","CreationDate":"2016-01-20T19:03:25.693","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"noreferrer\"><code>tf.matmul()<\/code><\/a> op requires that both of its inputs are matrices (i.e. 2-D tensors)<sup>*<\/sup>, and doesn't perform any automatic conversion. Your <code>T1<\/code> variable is a matrix, but your <code>x<\/code> placeholder is a length-2 vector (i.e. a 1-D tensor), which is the source of the error.<\/p>\n\n<p>By contrast, the <code>*<\/code> operator (an alias for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/multiply\" rel=\"noreferrer\"><code>tf.multiply()<\/code><\/a>) is a broadcasting element-wise multiplication. It will convert the vector argument to a matrix by following <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"noreferrer\">NumPy broadcasting rules<\/a>.<\/p>\n\n<p>To make your matrix multiplication work, you can either require that <code>x<\/code> is a matrix:<\/p>\n\n<pre><code>data = np.array([[0.1], [0.2]])\nx = tf.placeholder(tf.float32, shape=[2, 1])\nT1 = tf.Variable(tf.ones([2, 2]))\nl1 = tf.matmul(T1, x)\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    sess.run(l1, feed_dict={x: data})\n<\/code><\/pre>\n\n<p>...or you could use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims\" rel=\"noreferrer\"><code>tf.expand_dims()<\/code><\/a> op to convert the vector to a matrix:<\/p>\n\n<pre><code>data = np.array([0.1, 0.2])\nx = tf.placeholder(tf.float32, shape=[2])\nT1 = tf.Variable(tf.ones([2, 2]))\nl1 = tf.matmul(T1, tf.expand_dims(x, 1))\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    # ...\n<\/code><\/pre>\n\n<p><sup>*<\/sup> This was true when I posted the answer at first, but now <code>tf.matmul()<\/code> also supports batched matrix multiplications. This requires both arguments to have <em>at least<\/em> 2 dimensions. See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"noreferrer\">the documentation<\/a> for more details.<\/p>\n"}]}
{"QuestionId":34908928,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2016-01-20T19:34:23.800","AcceptedAnswerId":34963059.0,"OwnerUserId":2835670.0,"Title":"How to write update in theano function","Body":"<p>I am new to theano so maybe this is a simple question.\nIf I have a function<\/p>\n\n<pre><\/pre>\n\n<p>and y depends on w that I want to update using<\/p>\n\n<pre><\/pre>\n\n<p>I can write<\/p>\n\n<pre><\/pre>\n\n<p>and it will update the function f using the update rule.<\/p>\n\n<p>But if y depends on another variable, say z, that I want to update using a different rule, say<\/p>\n\n<pre><\/pre>\n\n<p>How do I combine the two rules together? <\/p>\n","answers":[{"AnswerId":"34960053","CreationDate":"2016-01-23T05:38:52.400","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p><code>update += [(z, z + tr_rate*(x - vis))]<\/code><br>\ndoes this work?<\/p>\n"},{"AnswerId":"34963059","CreationDate":"2016-01-23T11:47:47.240","ParentId":null,"OwnerUserId":"2835670","Title":null,"Body":"<p>I found my own answer and I am posting it if it can help other people.\nYou can create a variable update and then use the .append function to define new rules.<\/p>\n\n<p>So, instead of <\/p>\n\n<pre><code>wparameters = [w]\nupdate = [(wparam, \n           wparam + tr_rate * (pos_associations-neg_associations)) for wparam in wparameters]\n<\/code><\/pre>\n\n<p>you can append a new rule and write:<\/p>\n\n<pre><code>wparameters  = [w]\nzparameters = [z]\n\n\nupdate = []\nfor wparam, zparam in zip(wparameters, bparameters):\n    update.append((wparam, wparam + tr_rate*(pos_associations - neg_associations)))\n    update.append((zparam, zparam + tr_rate*(x - vis)) \n<\/code><\/pre>\n"}]}
{"QuestionId":34909772,"AnswerCount":0,"Tags":"<lua><neural-network><torch>","CreationDate":"2016-01-20T20:23:42.553","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch - fractional max pooling (fmp) issue","Body":"<p>I am stuck on implementing more than one fractional max pooling layer in torch. The source code is below;<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/SpatialFractionalMaxPooling.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/SpatialFractionalMaxPooling.lua<\/a><\/p>\n\n<p>I can get the following model working (this is a torch print out of the model);<\/p>\n\n<pre><\/pre>\n\n<p>As soon as I add more conv\/fmp layers I get;<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":34910327,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-01-20T20:54:39.850","AcceptedAnswerId":null,"OwnerUserId":4332827.0,"Title":"How can theano.scan execute external code","Body":"<p>We know that theano scan compile and execute the code with it's code in the symbolic graph. How to execute external (or irrelevant) code, for example:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is, x() will not be executed. How to solve it?<\/p>\n","answers":[]}
{"QuestionId":34910452,"AnswerCount":1,"Tags":"<gradient><tensorflow>","CreationDate":"2016-01-20T21:02:52.013","AcceptedAnswerId":34911086.0,"OwnerUserId":1576818.0,"Title":"TensorFlow apply_gradients remotely","Body":"<p>I'm trying to split up the minimize function over two machines. On one machine, I'm calling \"compute_gradients\", on another I call \"apply_gradients\" with gradients that were sent over the network. The issue is that calling apply_gradients(...).run(feed_dict) doesn't seem to work no matter what I do. I've tried inserting placeholders in place of the tensor gradients for apply_gradients, <\/p>\n\n<pre><\/pre>\n\n<p>then later when I receive the gradients I call<\/p>\n\n<pre><\/pre>\n\n<p>However, when I do this, I get <\/p>\n\n<p>ValueError: setting an array element with a sequence.<\/p>\n\n<p>This is only the latest thing I've tried, I've also tried the same solution without placeholders, as well as waiting to create the apply_gradients operation until I receive the gradients, which results in non-matching graph errors. <\/p>\n\n<p>Any help on which direction I should go with this?<\/p>\n","answers":[{"AnswerId":"34911086","CreationDate":"2016-01-20T21:37:36.803","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Assuming that each <code>gradients[i]<\/code> is a NumPy array that you've fetched using some out-of-band mechanism, the fix is simply to remove the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#convert_to_tensor\" rel=\"noreferrer\"><code>tf.convert_to_tensor()<\/code><\/a> invocation when building <code>feed_dict<\/code>:<\/p>\n\n<pre><code>feed_dict = {}\nfor i, grad_var in enumerate(compute_gradients):\n    feed_dict[placeholder_gradients[i][0]] = gradients[i]\napply_gradients.run(feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>Each value in a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>feed_dict<\/code><\/a> should be a NumPy array (or some other object that is trivially convertible to a NumPy array). In particular, a <code>tf.Tensor<\/code> is not a valid value for a <code>feed_dict<\/code>.<\/p>\n"}]}
{"QuestionId":34911276,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-20T21:49:08.770","AcceptedAnswerId":34911503.0,"OwnerUserId":4562374.0,"Title":"Cannot gather gradients for GradientDescentOptimizer in TensorFlow","Body":"<p>I've been trying to gather the gradient steps for each step of the GradientDescentOptimizer within TensorFlow, however I keep running into a TypeError when I try to pass the result of  to . The code I'm trying to run is:<\/p>\n\n<pre><\/pre>\n\n<p>Note that if I replace the last line with , where , the error is not raised. My confusion arises from the fact that  calls  with exactly the same arguments as its first step. Can someone explain why this behavior occurs?<\/p>\n","answers":[{"AnswerId":"34911483","CreationDate":"2016-01-20T22:01:37.623","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>minimize calls compute_gradients followed by apply_gradients: it's possible you're missing the second step.<\/p>\n\n<p>compute_gradients just returns the grads \/ variables, but doesn't apply the update rule to them.<\/p>\n\n<p>Here is an example: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f2bd0fc399606d14b55f3f7d732d013f32b33dd5\/tensorflow\/python\/training\/optimizer.py#L69\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f2bd0fc399606d14b55f3f7d732d013f32b33dd5\/tensorflow\/python\/training\/optimizer.py#L69<\/a><\/p>\n"},{"AnswerId":"34911503","CreationDate":"2016-01-20T22:03:03.940","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#Optimizer.compute_gradients\"><code>Optimizer.compute_gradients()<\/code><\/a> method returns a list of (<code>Tensor<\/code>, <code>Variable<\/code>) pairs, where each tensor is the gradient with respect to the corresponding variable.<\/p>\n\n<p><code>Session.run()<\/code> expects a list of <code>Tensor<\/code> objects (or objects convertible to a <code>Tensor<\/code>) as its first argument. It does not understand how to handle a list of pairs, and hence you get a <code>TypeError<\/code> which you try to run <code>sess.run(gradstep, ...)<\/code><\/p>\n\n<p>The correct solution depends on what you are trying to do. If you want to fetch all of the gradient values, you can do the following:<\/p>\n\n<pre><code>grad_vals = sess.run([grad for grad, _ in gradstep], feed_dict={x: batch_x, y: batch_y})\n\n# Then, e.g., nuild a variable name-to-gradient dictionary.\nvar_to_grad = {}\nfor grad_val, (_, var) in zip(grad_vals, gradstep):\n    var_to_grad[var.name] = grad_val\n<\/code><\/pre>\n\n<p>If you also want to fetch the variables, you can execute the following statement separately:<\/p>\n\n<pre><code>sess.run([var for _, var in gradstep])\n<\/code><\/pre>\n\n<p>...though note that&mdash;without further modification to your program&mdash;this will just return the initial values for each variable.\nYou will have to run the optimizer's training step (or otherwise call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#Optimizer.apply_gradients\"><code>Optimizer.apply_gradients()<\/code><\/a>) to update the variables.<\/p>\n"}]}
{"QuestionId":34913762,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-01-21T01:22:06.570","AcceptedAnswerId":34935094.0,"OwnerUserId":2084314.0,"Title":"How to expand a Tensorflow Variable","Body":"<p>Is there any way to make a Tensorflow Variable larger?  Like, let's say I wanted to add a neuron to a layer of a neural network in the middle of training.  How would I go about doing that?  An answer in <a href=\"https:\/\/stackoverflow.com\/questions\/33654754\/how-can-i-change-the-shape-of-a-variable-in-tensorflow\">This question<\/a> told me how to change the shape of the variable, to expand it to fit another row of weights, but I don't know how to initialize those new weights.<\/p>\n\n<p>I figure another way of going about this might involve combining variables, as in initializing the weights first in a second variable and then adding that in as a new row or column of the first variable, but I can't find anything that lets me do that either.<\/p>\n","answers":[{"AnswerId":"34916437","CreationDate":"2016-01-21T06:02:46.293","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>There are various ways you could accomplish this.<\/p>\n\n<p>1) The second answer in that post (<a href=\"https:\/\/stackoverflow.com\/a\/33662680\/5548115\">https:\/\/stackoverflow.com\/a\/33662680\/5548115<\/a>) explains how you can change the shape of a variable by calling 'assign' with validate_shape=False.  For example, you could do something like<\/p>\n\n<pre><code># Assume var is [m, n] \n# Add the new 'data' of shape [1, n] with new values\nnew_neuron = tf.constant(...)  \n\n# If concatenating to add a row, concat on the first dimension.\n# If new_neuron was [m, 1], you would concat on the second dimension.\nnew_variable_data = tf.concat(0, [var, new_neuron])  # [m+1, n]\n\nresize_var = tf.assign(var, new_variable_data, validate_shape=False)\n<\/code><\/pre>\n\n<p>Then when you run resize_var, the data pointed to by 'var' will now have the updated data.<\/p>\n\n<p>2) You could also create a large initial variable, and call tf.slice on different regions of the variable as training progresses, since you can dynamically change the 'begin' and 'size' attributes of slice.<\/p>\n"},{"AnswerId":"34935094","CreationDate":"2016-01-21T21:57:10.253","ParentId":null,"OwnerUserId":"2084314","Title":null,"Body":"<p>Figured it out.  It's kind of a roundabout process, but it's the only one I can tell that actually functions.  You need to first unpack the variables, then append the new variable to the end, then pack them back together.<\/p>\n\n<p>If you're expanding along the first dimension, it's rather short: only 7 lines of actual code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#the first variable is 5x3\nv1 = tf.Variable(tf.zeros([5, 3], dtype=tf.float32), \"1\")\n\n#the second variable is 1x3\nv2 = tf.Variable(tf.zeros([1, 3], dtype=tf.float32), \"2\")\n\n#unpack the first variable into a list of size 3 tensors\n#there should be 5 tensors in the list\nchange_shape = tf.unpack(v1)\n\n#unpack the second variable into a list of size 3 tensors\n#there should be 1 tensor in this list\nchange_shape_2 = tf.unpack(v2)\n\n#for each tensor in the second list, append it to the first list\nfor i in range(len(change_shape_2)):\n  change_shape.append(change_shape_2[i])\n\n#repack the list of tensors into a single tensor\n#the shape of this resultant tensor should be [6, 3]\nfinal = tf.pack(change_shape)\n<\/code><\/pre>\n\n<p>If you want to expand along the second dimension, it gets somewhat longer.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#First variable, 5x3\nv3 = tf.Variable(tf.zeros([5, 3], dtype=tf.float32))\n\n#second variable, 5x1\nv4 = tf.Variable(tf.zeros([5, 1], dtype=tf.float32))\n\n#unpack tensors into lists of size 3 tensors and size 1 tensors, respectively\n#both lists will hold 5 tensors\nchange = tf.unpack(v3)\nchange2 = tf.unpack(v4)\n\n#for each tensor in the first list, unpack it into its own list\n#this should make a 2d array of size 1 tensors, array will be 5x3\nchangestep2 = []\nfor i in range(len(change)):\n  changestep2.append(tf.unpack(change[i]))\n\n#do the same thing for the second tensor\n#2d array of size 1 tensors, array will be 5x1\nchange2step2 = []\nfor i in range(len(change2)):\n  change2step2.append(tf.unpack(change2[i]))\n\n  #for each tensor in the array, append it onto the corresponding array in the first list\n  for j in range(len(change2step2[i])):\n    changestep2[i].append(change2step2[i][j])\n\n  #pack the lists in the array back into tensors\n  changestep2[i] = tf.pack(changestep2[i])\n\n#pack the list of tensors into a single tensor\n#the shape of this resultant tensor should be [5, 4]\nfinal2 = tf.pack(changestep2)\n<\/code><\/pre>\n\n<p>I don't know if there's a more efficient way of doing this, but this works, as far as it goes.  Changing further dimensions would require more layers of lists, as necessary.<\/p>\n"},{"AnswerId":"40077350","CreationDate":"2016-10-17T01:36:12.233","ParentId":null,"OwnerUserId":"7012810","Title":null,"Body":"<p>Simply using tf.concat for expand a Tensorflow Variable,you can see the api_docs \nfor detail.<\/p>\n\n<pre><code>    v1 = tf.Variable(tf.zeros([5,3]),dtype=tf.float32)\n    v2 = tf.Variable(tf.zeros([1,3]),dtype=tf.float32)\n    v3 = tf.concat(0,[v1, v2])\n<\/code><\/pre>\n"}]}
{"QuestionId":34915378,"AnswerCount":1,"Tags":"<python><machine-learning><theano>","CreationDate":"2016-01-21T04:24:34.833","AcceptedAnswerId":null,"OwnerUserId":3627108.0,"Title":"Theano dmatrix contains newaxis raise dimension mismatch","Body":"<p>The following is the snippet of the definition of theano variable and function for RBFNN.  <\/p>\n\n<pre><\/pre>\n\n<p>Then I feed the data<\/p>\n\n<pre><\/pre>\n\n<p>where X, y is the train data and the centers and sigmas is the kmeans centers and the std of each centers.<\/p>\n\n<p>It finally raise the error<\/p>\n\n<pre><\/pre>\n\n<p>Did self.X and self.center has the shape [50, newaxis, 4] and [newaxis, 10, 4]? Why it turns out [50, 4] and [10, 4]? Could you tell me how to solve it?<\/p>\n","answers":[{"AnswerId":"34935312","CreationDate":"2016-01-21T22:10:05.990","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>You can use <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/debug_faq.html#using-test-values\" rel=\"nofollow\">test values<\/a> to try and debug the issue while building the graph.<\/p>\n"}]}
{"QuestionId":34917283,"AnswerCount":1,"Tags":"<eclipse><pydev><theano>","CreationDate":"2016-01-21T07:00:48.843","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"Eclipse with pydev cant import theano","Body":"<p>I have installed theano. ipython detects theano, however eclipse fails to do so. i have installed pydev &amp; normal python cods using numpy work fine.<\/p>\n\n<p>i have set theanorc as follows:<\/p>\n\n<pre><\/pre>\n\n<p>what settings do i have to do to allow pydev detect theano?<\/p>\n","answers":[{"AnswerId":"39543871","CreationDate":"2016-09-17T07:04:06.960","ParentId":null,"OwnerUserId":"5808396","Title":null,"Body":"<p>After installing Theano when I restarted eclipse, it automatically displayed a popup asking my permission to add theano to PYTHONPATH. I have installed theano and eclipse in both Windows and Linux PC and both are working fine.<\/p>\n\n<p>My steps are as follows:<\/p>\n\n<ol>\n<li>Install Anaconda (Python distribution from Continuum Analytics)<\/li>\n<li>Setup Eclipse to use it (<a href=\"https:\/\/www.rose-hulman.edu\/class\/csse\/resources\/Eclipse\/eclipse-python-configuration.htm\" rel=\"nofollow\">link<\/a>)<\/li>\n<li><p>In terminal (cmd for Windows) execute following two commands:<\/p>\n\n<p>a) pip install Theano<\/p>\n\n<p>b) conda install pydot-ng<\/p><\/li>\n<li>Start\/Restart eclipse and it will automatically dispay the popup asking your permission to add Theano to PYTHONPATH.<\/li>\n<\/ol>\n"}]}
{"QuestionId":34917819,"AnswerCount":1,"Tags":"<machine-learning><pickle><theano><keras><mnist>","CreationDate":"2016-01-21T07:33:57.573","AcceptedAnswerId":null,"OwnerUserId":3921440.0,"Title":"How to use pickled file as dataset for keras","Body":"<p>I have build my own dataset for digit classification and it worked well with convolutional network model developed by lisa lab (<a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/blob\/8e9db1a5494ca22ee01f53da34df26ea766ebaee\/code\/convolutional_mlp.py\" rel=\"nofollow\">Here<\/a>). I wanted to visualize the weights and i wanted to do it through keras.<\/p>\n\n<p>Keras documentation tries to load mnist data like this:<\/p>\n\n<pre><\/pre>\n\n<p>But i want my pickled dataset to load instead of mnist default data. Where does mnist module for keras load it's dataset ? And, how can i pass my own dataset instead of that to use mnist module from keras ?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"34932200","CreationDate":"2016-01-21T19:03:14.683","ParentId":null,"OwnerUserId":"2014584","Title":null,"Body":"<p>You can get a lot of info about this method by reading the source: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/datasets\/mnist.py\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/datasets\/mnist.py<\/a><\/p>\n\n<p>In this case, the dataset is a pickle file loaded from an Amazon S3 bucket.<\/p>\n\n<p>You could write a copy of this function and use it yourself to load up a different pickled dataset.<\/p>\n"}]}
{"QuestionId":34918851,"AnswerCount":1,"Tags":"<python><macos><tensorflow>","CreationDate":"2016-01-21T08:35:48.597","AcceptedAnswerId":null,"OwnerUserId":373386.0,"Title":"TensorFlow installation results in ImportError: No module named tensorflow","Body":"<p>I'm trying to get  to work on my Mac (OSX El Capitan 10.11.2). I tried the  from the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#download-and-setup\" rel=\"nofollow noreferrer\">setup guide<\/a> and also followed the instructions in <a href=\"https:\/\/stackoverflow.com\/questions\/33730259\/tensorflow-installation-error-on-mac\">the accepted answer here<\/a>. <\/p>\n\n<p>In both cases I am able to successfully activate the  and my prompt changes to   works fine, I'm able to do simple computations on the . But when I try to import tensorflow: <\/p>\n\n<pre><\/pre>\n\n<p>I repeatedly get this error: <\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated. <\/p>\n","answers":[{"AnswerId":"37838472","CreationDate":"2016-06-15T14:35:03.927","ParentId":null,"OwnerUserId":"2370817","Title":null,"Body":"<p>I had the issue reported in the original question.  Python worked fine, I followed the installation steps on the Tensforflow website, and got \"No module named tensorflow.\"<\/p>\n\n<p>Re-installing python via brew, and re-installing pip as a result, fixed it.  I didn't need to uninstall anything, just the line below along with the normal pip install from the tensorflow installation document afterwards:<\/p>\n\n<pre><code>brew install python\n<\/code><\/pre>\n"}]}
{"QuestionId":34920998,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-21T10:20:50.107","AcceptedAnswerId":null,"OwnerUserId":867505.0,"Title":"Tensorflow network diverges if reading\/preprocessing is done on cpu","Body":"<p>I have some code that reads images from a tensor record, preprocesses them (cropping, apply random hue\/saturation etc., all through tensorflows own methods) and then uses  to generate batches.<\/p>\n\n<pre><\/pre>\n\n<p>This works well and leads to converging network when placing all operations on gpu. However, this is right now the bottleneck of my code and I'd like to speed it up by placing it on the cpu by wrapping this block in a . With this I have faster iterations (about 1\/5th), but the network diverges after about 10 iterations, leading to a loss of NaN. When visually inspecting the samples created in the tensorboard, there is no apparent difference.<\/p>\n\n<p>Why is the convergence behavior different on cpu vs gpu? How could I further investigate into this weird behavior?<\/p>\n","answers":[{"AnswerId":"43269084","CreationDate":"2017-04-07T03:23:16.410","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>I was having a very similar issue. In my case I had pinned the image transformation ops to the GPU, but the data was being fed from queues on the CPU. For me it was a mistake to pin the ops to the GPU, so I pinned them to the CPU with <code>with tf.device('\/cpu:0')<\/code> and my numerical instability issues went away. <\/p>\n\n<p>Notably, I had also run these image preprocessing steps on the GPU previously but was not using queues to load the data. I simply feed the data directly to the GPU via a placeholder and everything ran just fine.<\/p>\n\n<p>It was only when I started using queues, and loaded those queues from independent threads that I started seeing this issue (I was able to use the queues in a sequential manner when I didn't load them from separate threads). <\/p>\n\n<p>I don't have an exact answer as to what happened yet, but it does seem like making sure that the data and the image pre-processing ops are consistently placed on the same device is quite critical.<\/p>\n"}]}
{"QuestionId":34921847,"AnswerCount":0,"Tags":"<python><machine-learning><theano><logistic-regression><deep-learning>","CreationDate":"2016-01-21T10:56:47.920","AcceptedAnswerId":null,"OwnerUserId":4535284.0,"Title":"What does negative log likelihood of logistic regression in theano look like?","Body":"<p>I have been reading the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">theano's logistic regression tutorial<\/a>. I was trying to understand how the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#defining-a-loss-function\" rel=\"nofollow\">negative log likelihood<\/a> is calculated.<\/p>\n\n<pre><\/pre>\n\n<p>On pretty printing printing  it returned<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone explain what is this  doing with a minuscule example?<\/p>\n\n<p>After this they calculated the <\/p>\n\n<p>Any help is appreciated! :)<\/p>\n","answers":[]}
{"QuestionId":34929254,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow><reinforcement-learning><q-learning>","CreationDate":"2016-01-21T16:34:33.470","AcceptedAnswerId":null,"OwnerUserId":1783952.0,"Title":"Tensorflow implementation of loss of Q-network with slicing","Body":"<p>I'm implementing a Q-network as described in Human-level control through deep reinforcement learning (Mnih et al. 2015) in TensorFlow.<\/p>\n\n<p>To approximate the Q-function they use a neural network. The Q-function maps a state and an action to a scalar value, known as the Q-value. I.e. it's a function like Q(s,a) = qvalue.<\/p>\n\n<p>But instead of taking both state and action as input, they only take the state as input and output a vector with one element per legal action in a given order. Thus Q(s,a) becomes Q'(s) = , where  is Q(s,a1). <\/p>\n\n<p>This raises the question of how to modify the loss function. The loss function is a L2 loss function computed on the difference of a target (y) and Q(s,a).<\/p>\n\n<p>My idea is to create a new TF operation and use a binary mask indicating which action I want to train on and multiply it with the output of the network. Effectively producing a vector like  if the action in question is . <\/p>\n\n<p>And then feed the result of the new operation to the loss operation, which TF is then minimizing.<\/p>\n\n<p>Questions:<\/p>\n\n<ol>\n<li><p>Is this a sound idea? Or is there a better way of solving this?<\/p><\/li>\n<li><p>How can this be solved with TensorFlow?<\/p>\n\n<p>There is a SO thread on something similar (<a href=\"https:\/\/stackoverflow.com\/questions\/34685947\/adjust-single-value-within-tensor-tensorflow\/34686952#34686952\">Adjust Single Value within Tensor -- TensorFlow<\/a>), but I would like to choose the column value with the help of a  that I can feed to the network at runtime. It doesn't seem to work when just replacing the static lists in that examples with placeholders.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"34932441","CreationDate":"2016-01-21T19:16:18.417","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>There are a few implementations out there of doing DeepQ learning in TensorFlow that might be useful references to check out:<\/p>\n\n<p><a href=\"https:\/\/github.com\/asrivat1\/DeepLearningVideoGames\" rel=\"nofollow\">https:\/\/github.com\/asrivat1\/DeepLearningVideoGames<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/nivwusquorum\/tensorflow-deepq\" rel=\"nofollow\">https:\/\/github.com\/nivwusquorum\/tensorflow-deepq<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/mrkulk\/deepQN_tensorflow\" rel=\"nofollow\">https:\/\/github.com\/mrkulk\/deepQN_tensorflow<\/a><\/p>\n\n<p>I'm not sure what the best idea is without digging more deeply, but you can definitely apply a mask in a few different ways.<\/p>\n\n<p>If you have your binary mask as a boolean vector e.g., [False, False, True, False] set up already, then you can do:<\/p>\n\n<pre><code>val_array = ...\nbinary_mask = tf.constant([False, False, True, False])\nresult = tf.select(binary_mask, val_array, tf.zeros_like(val_array))\n<\/code><\/pre>\n\n<p>This selects the entry from val_array whereever binary_mask is True, and zeros otherwise.<\/p>\n\n<p>If your mask is not boolean but already a numeric type of the same type as val_array (e.g., 0.0s and 1.0s), then you can do a tf.mul(mask, val_array).<\/p>\n"}]}
{"QuestionId":34931121,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-21T18:07:23.440","AcceptedAnswerId":34934543.0,"OwnerUserId":2615676.0,"Title":"Can cond support TF ops with side effects?","Body":"<p>The (source code) documentation for  is unclear on whether the functions to be performed when the predicate is evaluated can have side effects or not. I've done some tests but I'm getting conflicting results. For example the code below does not work:<\/p>\n\n<pre><\/pre>\n\n<p>I.e. no matter what value the predicate evaluates to, both functions are getting run, and so the net result is a subtraction of 1. On the other hand, this code snippet does work, where the only difference is that I add new ops to the graph every time  is called:<\/p>\n\n<pre><\/pre>\n\n<p>Not sure why creating new ops every time works while the other case doesn't, but I'd obviously rather not be adding nodes as the graph will eventually become too big.<\/p>\n","answers":[{"AnswerId":"34934543","CreationDate":"2016-01-21T21:18:11.207","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Your second version&mdash;where the <code>assign_add()<\/code> and <code>assign_sub()<\/code> ops are creating inside the lambdas passed to <code>cond()<\/code>&mdash;is the correct way to do this. Fortunately, each of the two lambdas is only evaluated once, during the call to <code>cond()<\/code>, so your graph will not grow without bound.<\/p>\n\n<p>Essentially what <code>cond()<\/code> does is the following:<\/p>\n\n<ol>\n<li><p>Create a <code>Switch<\/code> node, which forwards its input to only one of two outputs, depending on the value of <code>pred<\/code>. Let's call the outputs <code>pred_true<\/code> and <code>pred_false<\/code>. (They have the same value as <code>pred<\/code> but that's unimportant since this is never directly evaluated.)<\/p><\/li>\n<li><p>Build the subgraph corresponding to the <code>if_true<\/code> lambda, where all of the nodes have a control dependency on <code>pred_true<\/code>.<\/p><\/li>\n<li><p>Build the subgraph corresponding to the <code>if_false<\/code> lambda, where all of the nodes have a control dependency on <code>pred_false<\/code>.<\/p><\/li>\n<li><p>Zip together the lists of return values from the two lambdas, and create a <code>Merge<\/code> node for each of these. A <code>Merge<\/code> node takes two inputs, of which only one is expected to be produced, and forwards it to its output.<\/p><\/li>\n<li><p>Return the tensors that are the outputs of the <code>Merge<\/code> nodes.<\/p><\/li>\n<\/ol>\n\n<p>This means you can run your second version, and be content that the graph remains a fixed size, regardless of how many steps you run.<\/p>\n\n<p>The reason your first version doesn't work is that, when a <code>Tensor<\/code> is captured (like <code>adder<\/code> or <code>subtractor<\/code> in your example), an additional <code>Switch<\/code> node is added to enforce the logic that the value of the tensor is only forwarded to the branch that actually executes. This is an artifact of how TensorFlow combines feed-forward dataflow and control flow  in its execution model. The result is that the captured tensors (in this case the results of the <code>assign_add<\/code> and <code>assign_sub<\/code>) will always be evaluated, even if they aren't used, and you'll see their side effects. This is something we need to document better, and <a href=\"https:\/\/stackoverflow.com\/a\/34934516\/3574081\">as Michael says<\/a>, we're going to make this more usable in future.<\/p>\n"},{"AnswerId":"34934516","CreationDate":"2016-01-21T21:16:23.240","ParentId":null,"OwnerUserId":"5823391","Title":null,"Body":"<p>The second case works because you have added the ops within the cond: this causes them to conditionally execute. <\/p>\n\n<p>The first case it is analogous to saying:<\/p>\n\n<pre><code>adder = (count += 1)\nsubtractor = (count -= 2)\nif (cond) { adder } else { subtractor }\n<\/code><\/pre>\n\n<p>Since adder and subtractor are outside the conditional, they are always executed.<\/p>\n\n<p>The second case is more like saying<\/p>\n\n<pre><code>if (cond) { adder = (count += 1) } else { subtractor = (count -= 2) }\n<\/code><\/pre>\n\n<p>which in this case does what you expected.<\/p>\n\n<p>We realize that the interaction between side effects and (somewhat) lazy evaluation is confusing, and we have a medium-term goal to make things more uniform. But the important thing to understand for now is that we do not do true lazy evaluation: the conditional acquires a dependency on every quantity defined outside the conditional that is used within either branch.<\/p>\n"}]}
{"QuestionId":34934303,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-01-21T21:05:01.407","AcceptedAnswerId":null,"OwnerUserId":388951.0,"Title":"Renormalize weight matrix using TensorFlow","Body":"<p>I'd like to add a max norm constraint to several of the weight matrices in my TensorFlow graph, ala Torch's <a href=\"http:\/\/torch7.readthedocs.org\/en\/rtd\/maths\/#torchrenormres-x-p-dim-maxnorm\" rel=\"noreferrer\"><\/a> method.<\/p>\n\n<p>If the L2 norm of any neuron's weight matrix exceeds , I'd like to scale its weights down so that their L2 norm is exactly .<\/p>\n\n<p>What's the best way to express this using TensorFlow?<\/p>\n","answers":[{"AnswerId":"34934673","CreationDate":"2016-01-21T21:26:57.157","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#clip_by_norm\" rel=\"nofollow\">clip_by_norm<\/a> function, which does exactly this. It takes a single tensor as input and returns a scaled down tensor.<\/p>\n"},{"AnswerId":"34952915","CreationDate":"2016-01-22T17:46:15.413","ParentId":null,"OwnerUserId":"388951","Title":null,"Body":"<p>Using Rafa\u0142's suggestion and TensorFlow's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a5d8217c4ed90041bea2616c14a8ddcf11ec8c03\/tensorflow\/python\/ops\/clip_ops.py#L90-L95\" rel=\"nofollow\">implementation<\/a> of <code>clip_by_norm<\/code>, here's what I came up with:<\/p>\n\n<pre><code>def renorm(x, axis, max_norm):\n    '''Renormalizes the sub-tensors along axis such that they do not exceed norm max_norm.'''\n    # This elaborate dance avoids empty slices, which TF dislikes.\n    rank = tf.rank(x)\n    bigrange = tf.range(-1, rank + 1)\n    dims = tf.slice(\n                tf.concat(0, [tf.slice(bigrange, [0], [1 + axis]),\n                              tf.slice(bigrange, [axis + 2], [-1])]),\n                [1], rank - [1])\n\n    # Determine which columns need to be renormalized.\n    l2norm_inv = tf.rsqrt(tf.reduce_sum(x * x, dims, keep_dims=True))\n    scale = max_norm * tf.minimum(l2norm_inv, tf.constant(1.0 \/ max_norm))\n\n    # Broadcast the scalings\n    return tf.mul(scale, x)\n<\/code><\/pre>\n\n<p>It seems to have the desired behavior for 2-dimensional matrices and should \ngeneralize to tensors:<\/p>\n\n<pre><code>&gt; x = tf.constant([0., 0., 3., 4., 30., 40., 300., 400.], shape=(4, 2))\n&gt; print x.eval()\n[[   0.    0.]  # rows have norms of 0, 5, 50, 500\n [   3.    4.]  # cols have norms of ~302, ~402\n [  30.   40.]\n [ 300.  400.]]\n&gt; print renorm(x, 0, 10).eval()\n[[ 0.          0.        ]  # unaffected\n [ 3.          4.        ]  # unaffected\n [ 5.99999952  7.99999952]  # rescaled\n [ 6.00000048  8.00000095]] # rescaled\n&gt; print renorm(x, 1, 350).eval()\n[[   0.            0.        ]  # col 0 is unaffected\n [   3.            3.48245788]  # col 1 is rescaled\n [  30.           34.82457733]\n [ 300.          348.24578857]]\n<\/code><\/pre>\n"},{"AnswerId":"39028553","CreationDate":"2016-08-18T22:37:21.533","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>Here is a possible implementation:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef maxnorm_regularizer(threshold, axes=1, name=\"maxnorm\", collection=\"maxnorm\"):\n    def maxnorm(weights):\n        clipped = tf.clip_by_norm(weights, clip_norm=threshold, axes=axes)\n        clip_weights = tf.assign(weights, clipped, name=name)\n        tf.add_to_collection(collection, clip_weights)\n        return None # there is no regularization loss term\n    return maxnorm\n<\/code><\/pre>\n\n<p>Here's how you would use it:<\/p>\n\n<pre><code>from tensorflow.contrib.layers import fully_connected\nfrom tensorflow.contrib.framework import arg_scope\n\nwith arg_scope(\n        [fully_connected],\n        weights_regularizer=max_norm_regularizer(1.5)):\n    hidden1 = fully_connected(X, 200, scope=\"hidden1\")\n    hidden2 = fully_connected(hidden1, 100, scope=\"hidden2\")\n    outputs = fully_connected(hidden2, 5, activation_fn=None, scope=\"outs\")\n\nmax_norm_ops = tf.get_collection(\"max_norm\")\n\n[...]\n\nwith tf.Session() as sess:\n    sess.run(init)\n    for epoch in range(n_epochs):\n        for X_batch, y_batch in load_next_batch():\n            sess.run(training_op, feed_dict={X: X_batch, y: y_batch})\n            sess.run(max_norm_ops)\n<\/code><\/pre>\n\n<p>This creates a 3 layer neural network and trains it with max norm regularization at every layer (with a threshold of 1.5). I just tried it, seems to work. Hope this helps! Suggestions for improvements are welcome. :)<\/p>\n\n<p><strong>Notes<\/strong><\/p>\n\n<p>This code is based on <code>tf.clip_by_norm()<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; x = tf.constant([0., 0., 3., 4., 30., 40., 300., 400.], shape=(4, 2))\n&gt;&gt;&gt; print(x.eval())\n[[   0.    0.]\n [   3.    4.]\n [  30.   40.]\n [ 300.  400.]]\n&gt;&gt;&gt; clip_rows = tf.clip_by_norm(x, clip_norm=10, axes=1)\n&gt;&gt;&gt; print(clip_rows.eval())\n[[ 0.          0.        ]\n [ 3.          4.        ]\n [ 6.          8.        ]  # clipped!\n [ 6.00000048  8.        ]] # clipped!\n<\/code><\/pre>\n\n<p>You can also clip columns if you need to:<\/p>\n\n<pre><code>&gt;&gt;&gt; clip_cols = tf.clip_by_norm(x, clip_norm=350, axes=0)\n&gt;&gt;&gt; print(clip_cols.eval())\n[[   0.            0.        ]\n [   3.            3.48245788]\n [  30.           34.82457733]\n [ 300.          348.24578857]]\n                # clipped!\n<\/code><\/pre>\n"}]}
{"QuestionId":34934437,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-01-21T21:12:18.470","AcceptedAnswerId":null,"OwnerUserId":2280020.0,"Title":"tensorboard not displaying anything","Body":"<p>The nodes that I have to work on don't have many outside internet connections. When on one of these nodes, after running tensorboard and visiting <a href=\"http:\/\/localhost:6006\/\" rel=\"nofollow\">http:\/\/localhost:6006\/<\/a>, I don't see anything - just a blank page - for a long time, finally I get something. <\/p>\n\n<p>On a node with internet access, things came up faster. Is tensorboard going to outside URL's for styling and fonts and things like that? On the internal node, I see all these messages that have \/external in them, like<\/p>\n\n<pre><\/pre>\n\n<p>There are also warnings like<\/p>\n\n<pre><\/pre>\n\n<p>I don't see any simple way to force tensorboard to stay local - or do you think I am barking up the wrong tree here? Maybe there is some other problem, but I didn't get those messages on the internet connected node.<\/p>\n","answers":[{"AnswerId":"37932562","CreationDate":"2016-06-20T22:04:32.963","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>TensorBoard does try to load font data externally (it makes a GET request here: <a href=\"https:\/\/fonts.googleapis.com\/css?family=Roboto:400,300,300italic,400italic,500,500italic,700,700italic\" rel=\"nofollow\">https:\/\/fonts.googleapis.com\/css?family=Roboto:400,300,300italic,400italic,500,500italic,700,700italic<\/a> ). However, if the font request fails, it shouldn't block the rest of the page from loading. The \/external\/ refers to external to the TensorFlow repository at build time, but it is all cloned locally.<\/p>\n\n<p>The summary svg issue suggests to me that you may be using an old version of TensorBoard, although that shouldn't matter for this issue.<\/p>\n\n<p>I guess that there is some other issue on your non-connected nodes that is causing the initial TensorBoard asset load (which loads a large html file, and a lot of small open-source assets) to be slow. I recommend trying using Google Chrome if you are not already, I have seen issues with Firefox being slow because it loads all the assets in sequence rather than in parallel. If you are using Chrome, try investigating the \"network\" pane in the Chrome Debugger and reply back with what you find.<\/p>\n"},{"AnswerId":"41921714","CreationDate":"2017-01-29T14:01:51.780","ParentId":null,"OwnerUserId":"7483498","Title":null,"Body":"<p>Be sure that you include the full path when starting TensorBoard<\/p>\n\n<pre><code> tensorboard --logdir='.\/somedir\/logdir'\n<\/code><\/pre>\n\n<p>I\u00b4m using safari, crome and firefox without problems to connect to localhost:6006<\/p>\n"}]}
{"QuestionId":34935464,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-21T22:20:28.883","AcceptedAnswerId":null,"OwnerUserId":4455672.0,"Title":"Update a subset of weights in TensorFlow","Body":"<p>Does anyone know how to update a subset (i.e. only some indices) of the weights that are used in the forward propagation?<\/p>\n\n<p>My guess is that I might be able to do that after applying compute_gradients as follows:<\/p>\n\n<pre><\/pre>\n\n<p>...and then do something with the list of tuples in .<\/p>\n","answers":[{"AnswerId":"36472763","CreationDate":"2016-04-07T10:00:22.377","ParentId":null,"OwnerUserId":"6091891","Title":null,"Body":"<p>Easiest way is to pull the <code>tf.Variable<\/code> into python (as a numpy array) using <code>npvar = sess.run(tfvar)<\/code>, then perform some operation on it such as <code>npvar[1, 2] = -10<\/code>.  Then you can upload the modified data back into tensorflow using <code>sess.run(tfvar.assign(npvar))<\/code>.<\/p>\n\n<p>Obviously this is very slow and not really useful for training but it does work.<\/p>\n"},{"AnswerId":"34937627","CreationDate":"2016-01-22T01:41:57.573","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You could use a combination of <code>gather<\/code> and <code>scatter_update<\/code>. Here's an example that doubles the values at position <code>0<\/code> and <code>2<\/code><\/p>\n\n<pre><code>indices = tf.constant([0,2])\ndata = tf.Variable([1,2,3])\ndata_subset = tf.gather(data, indices)\nupdated_data_subset = 2*data_subset\nsparse_update = tf.scatter_update(data, indices, updated_data_subset)\ninit_op = tf.initialize_all_variables()\n\nsess = tf.Session()\nsess.run([init_op])\nprint \"Values before:\", sess.run([data])\nsess.run([sparse_update])\nprint \"Values after:\", sess.run([data])\n<\/code><\/pre>\n\n<p>You should see<\/p>\n\n<pre><code>Values before: [array([1, 2, 3], dtype=int32)]\nValues after: [array([2, 2, 6], dtype=int32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":34936243,"AnswerCount":1,"Tags":"<python><gpu><tensorflow>","CreationDate":"2016-01-21T23:17:39.907","AcceptedAnswerId":null,"OwnerUserId":1104696.0,"Title":"How to safely terminate a tensorflow program running on multiple GPUs","Body":"<p>I have implement a network using tensorflow. The network is trained on 4 GPUs. When I hit <kbd>ctrl+c<\/kbd>, the program crashed the nvidia driver and created zombie process named \"python\". I am not able to kill the zombie process, neither can I reboot ubuntu system by . <\/p>\n\n<p>I am using a FIFO queue and a thread to read data from binary files.<\/p>\n\n<pre><\/pre>\n\n<p>After I call , the program won't stop and I see:<\/p>\n\n<pre><\/pre>\n\n<p>It seems GPU resource is not released. If I open another terminal,  command won't work. Then I have to brutally restart system by:<\/p>\n\n<pre><\/pre>\n\n<p>I know  may be too brutal. So I tried using dequeue operation to empty the FIFO queue, then:<\/p>\n\n<pre><\/pre>\n\n<p>This method does not work either. Basically, the program can not terminate after max training iteration is reached.<\/p>\n","answers":[{"AnswerId":"35952621","CreationDate":"2016-03-12T01:25:43.490","ParentId":null,"OwnerUserId":"6052502","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/658\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/658<\/a><\/p>\n\n<p>This solved the problem:<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0\n<\/code><\/pre>\n"}]}
{"QuestionId":34941620,"AnswerCount":2,"Tags":"<tensorflow><bazel>","CreationDate":"2016-01-22T07:52:51.287","AcceptedAnswerId":34951823.0,"OwnerUserId":5703903.0,"Title":"Unable to build TensorFlow from source with bazel. 22nd January 2016","Body":"<p>I wanted to install the latest TensorFlow from source. I pulled the repository, and then ran<\/p>\n\n<p><\/p>\n\n<p>I got this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34941621","CreationDate":"2016-01-22T07:52:51.287","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>Turns out that bazel was outdated. The previous version of my bazel was 0.1.2, which was last updated in December. Bazel has since then been updated to 0.1.4. For those who have this error, updating bazel would do the trick. You can then build as per normal.<\/p>\n\n<p>I guess in future, if it can't build, you might want to check if bazel is updated to the latest version.<\/p>\n"},{"AnswerId":"34951823","CreationDate":"2016-01-22T16:46:10.927","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There was a recent change to the TensorFlow <code>BUILD<\/code> files that causes the dependency on Bazel 0.1.4. This was <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/?utm_medium=email&amp;utm_source=footer#!msg\/discuss\/qUech_YUtCM\/h4IDS38yCQAJ\" rel=\"noreferrer\">announced<\/a> on the <code>discuss@tensorflow.org<\/code> mailing list, but hasn't made it into the documentation on the website yet.<\/p>\n\n<p>As you have already discovered, upgrading Bazel to <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/releases\/tag\/0.1.4\" rel=\"noreferrer\">version 0.1.4<\/a> should fix this build error.<\/p>\n"}]}
{"QuestionId":34941627,"AnswerCount":0,"Tags":"<caffe><pycaffe>","CreationDate":"2016-01-22T07:53:10.937","AcceptedAnswerId":null,"OwnerUserId":654005.0,"Title":"During classification I sometimes get \"nan\" values","Body":"<p>I've trained a caffemodel with GoogLeNet. During testing I have a very high accuracy:<\/p>\n\n<pre><\/pre>\n\n<p>No I have a python classifier which looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I ran the classifier through all my validation data. The accuarcy is very high. But I get some \"nan\" probability values for some input images. What's the reason for that? What does \"nan\" mean? Is it \"I did not recognize any class\"? <\/p>\n\n<p>Edit:\nThis question is not a <a href=\"https:\/\/stackoverflow.com\/questions\/33962226\/\">duplication<\/a> since it refers to the classification and not to the training<\/p>\n\n<p>Thank you. <\/p>\n","answers":[]}
{"QuestionId":34942745,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-01-22T09:05:07.040","AcceptedAnswerId":34951694.0,"OwnerUserId":3661568.0,"Title":"Preprocess a Tensorflow tensor in Numpy","Body":"<p>I have set up a CNN in Tensorflow where I read my data with a TFRecordReader. It works well but I would like to do some more preprocessing and data augmentation than offered by the  functions. I would specifically like to do some randomized scaling.<\/p>\n\n<p>Is it possible to process a Tensorflow tensor in Numpy? Or do I need to drop the TFRecordReader and rather do all my preprocessing in Numpy and feed data using the feed_dict? I suspect that the feed_dict method is slow when training on images, but I might be wrong?<\/p>\n","answers":[{"AnswerId":"34951694","CreationDate":"2016-01-22T16:40:05.283","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you could create a custom I\/O pipeline that fetches intermediate results back from TensorFlow using one or more threads, applies arbitrary Python logic, and then feeds them into a queue for subsequent processing. The resulting program would be somewhat more complicated, but I suggest you look at the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/threading_and_queues\/index.html\">threading and queues HOWTO<\/a> for information on how to get started.<\/p>\n\n<hr>\n\n<p>There is an <strong>experimental<\/strong> feature that might make this easier, if you <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#installing-from-sources\">install from source<\/a>.<\/p>\n\n<p>If you have already built a preprocessing pipeline using TensorFlow ops, the easiest way to add some custom Python code is to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/script_ops.html#py_func\"><code>tf.py_func()<\/code><\/a> operator, which takes a list of <code>Tensor<\/code> objects, and a Python function that maps one or more NumPy arrays to one or more NumPy arrays.<\/p>\n\n<p>For example, let's say you have a pipeline like this:<\/p>\n\n<pre><code>reader = tf.TFRecordReader(...)\nimage_t = tf.image.decode_png(tf.parse_single_example(reader.read(), ...))\n<\/code><\/pre>\n\n<p>...you could use <code>tf.py_func()<\/code> to apply some custom NumPy processing as follows:<\/p>\n\n<pre><code>from scipy import ndimage\ndef preprocess(array):\n  # `array` is a NumPy array containing.\n  return ndimage.rotate(array, 45)\n\nimage_t = tf.py_func(preprocess, [image_t], [tf.float32])\n<\/code><\/pre>\n"}]}
{"QuestionId":34945554,"AnswerCount":6,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2016-01-22T11:22:10.507","AcceptedAnswerId":34948185.0,"OwnerUserId":5825952.0,"Title":"How to set layer-wise learning rate in Tensorflow?","Body":"<p>I am wondering if there is a way that I can use different learning rate for different layers like what is in Caffe. I am trying to modify a pre-trained model and use it for other tasks. What I want is to speed up the training for new added layers and keep the trained layers at low learning rate in order to prevent them from being distorted. for example, I have a 5-conv-layer pre-trained model. Now I add a new conv layer and fine tune it. The first 5 layers would have learning rate of 0.00001 and the last one would have 0.001. Any idea how to achieve this?<\/p>\n","answers":[{"AnswerId":"34951945","CreationDate":"2016-01-22T16:52:02.620","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><strong>Update Jan 22<\/strong>: recipe below is only a good idea for <code>GradientDescentOptimizer<\/code> , other optimizers that keep a running average will apply learning rate before the parameter update, so recipe below won't affect that part of the equation<\/p>\n\n<p>In addition to Rafal's approach, you could use <code>compute_gradients<\/code>, <code>apply_gradients<\/code> interface of <code>Optimizer<\/code>. For instance, here's a toy network where I use 2x the learning rate for second parameter<\/p>\n\n<pre><code>x = tf.Variable(tf.ones([]))\ny = tf.Variable(tf.zeros([]))\nloss = tf.square(x-y)\nglobal_step = tf.Variable(0, name=\"global_step\", trainable=False)\n\nopt = tf.GradientDescentOptimizer(learning_rate=0.1)\ngrads_and_vars = opt.compute_gradients(loss, [x, y])\nygrad, _ = grads_and_vars[1]\ntrain_op = opt.apply_gradients([grads_and_vars[0], (ygrad*2, y)], global_step=global_step)\n\ninit_op = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init_op)\nfor i in range(5):\n  sess.run([train_op, loss, global_step])\n  print sess.run([x, y])\n<\/code><\/pre>\n\n<p>You should see<\/p>\n\n<pre><code>[0.80000001, 0.40000001]\n[0.72000003, 0.56]\n[0.68800002, 0.62400001]\n[0.67520005, 0.64960003]\n[0.67008007, 0.65984005]\n<\/code><\/pre>\n"},{"AnswerId":"34948185","CreationDate":"2016-01-22T13:45:45.657","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>It can be achieved quite easily with 2 optimizers:<\/p>\n\n<pre><code>var_list1 = [variables from first 5 layers]\nvar_list2 = [the rest of variables]\ntrain_op1 = GradientDescentOptimizer(0.00001).minimize(loss, var_list=var_list1)\ntrain_op2 = GradientDescentOptimizer(0.0001).minimize(loss, var_list=var_list2)\ntrain_op = tf.group(train_op1, train_op2)\n<\/code><\/pre>\n\n<p>One disadvantage of this implementation is that it computes tf.gradients(.) twice inside the optimizers and thus it might not be optimal in terms of execution speed. This can be mitigated by explicitly calling tf.gradients(.), splitting the list into 2 and passing corresponding gradients to both optimizers.<\/p>\n\n<p>Related question: <a href=\"https:\/\/stackoverflow.com\/questions\/34477889\/holding-variables-constant-during-optimizer\/34478044#34478044\">Holding variables constant during optimizer<\/a><\/p>\n\n<p>EDIT: Added more efficient but longer implementation:<\/p>\n\n<pre><code>var_list1 = [variables from first 5 layers]\nvar_list2 = [the rest of variables]\nopt1 = tf.train.GradientDescentOptimizer(0.00001)\nopt2 = tf.train.GradientDescentOptimizer(0.0001)\ngrads = tf.gradients(loss, var_list1 + var_list2)\ngrads1 = grads[:len(var_list1)]\ngrads2 = grads[len(var_list1):]\ntran_op1 = opt1.apply_gradients(zip(grads1, var_list1))\ntrain_op2 = opt2.apply_gradients(zip(grads2, var_list2))\ntrain_op = tf.group(train_op1, train_op2)\n<\/code><\/pre>\n\n<p>You can use <code>tf.trainable_variables()<\/code> to get all training variables and decide to select from them.\nThe difference is that in the first implementation <code>tf.gradients(.)<\/code> is called twice inside the optimizers. This may cause some redundant operations to be executed (e.g. gradients on the first layer can reuse some computations for the gradients of the following layers).<\/p>\n"},{"AnswerId":"53367277","CreationDate":"2018-11-19T01:45:05.580","ParentId":null,"OwnerUserId":"9319536","Title":null,"Body":"<p>A slight variation of Sergey Demyanov answer, where you only have to specify the learning rates you would like to change<\/p>\n\n<pre><code>from collections import defaultdict\n\nself.learning_rates = defaultdict(lambda: 1.0)\n...\nx = tf.layers.Dense(3)(x)\nself.learning_rates[x.op.name] = 2.0\n...\noptimizer = tf.train.MomentumOptimizer(learning_rate=1e-3, momentum=0.9)\ngrads_and_vars = optimizer.compute_gradients(loss)\ngrads_and_vars_mult = []\nfor grad, var in grads_and_vars:\n    grad *= self.learning_rates[var.op.name]\n    grads_and_vars_mult.append((grad, var))\ntrain_op = optimizer.apply_gradients(grads_and_vars_mult, tf.train.get_global_step())\n<\/code><\/pre>\n"},{"AnswerId":"50388264","CreationDate":"2018-05-17T09:42:52.803","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>Tensorflow 1.7 introduced <code>tf.custom_gradient<\/code> that greatly simplifies setting learning rate multipliers, in a way that is now compatible with any optimizer, including those accumulating gradient statistics. For example,<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef lr_mult(alpha):\n  @tf.custom_gradient\n  def _lr_mult(x):\n    def grad(dy):\n      return dy * alpha * tf.ones_like(x)\n    return x, grad\n  return _lr_mult\n\nx0 = tf.Variable(1.)\nx1 = tf.Variable(1.)\nloss = tf.square(x0) + tf.square(lr_mult(0.1)(x1))\n\nstep = tf.train.GradientDescentOptimizer(learning_rate=0.1).minimize(loss)\n\nsess = tf.InteractiveSession()\ntf.global_variables_initializer().run()\ntf.local_variables_initializer().run()\n\nfor _ in range(5):\n  sess.run([step])\n  print(sess.run([x0, x1, loss]))\n<\/code><\/pre>\n"},{"AnswerId":"37825314","CreationDate":"2016-06-15T02:51:05.270","ParentId":null,"OwnerUserId":"586275","Title":null,"Body":"<p>Collect learning rate multipliers for each variable like:<\/p>\n\n<pre><code>self.lr_multipliers[var.op.name] = lr_mult\n<\/code><\/pre>\n\n<p>and then apply them during before applying the gradients like:<\/p>\n\n<pre><code>def _train_op(self):\n  tf.scalar_summary('learning_rate', self._lr_placeholder)\n  opt = tf.train.GradientDescentOptimizer(self._lr_placeholder)\n  grads_and_vars = opt.compute_gradients(self._loss)\n  grads_and_vars_mult = []\n  for grad, var in grads_and_vars:\n    grad *= self._network.lr_multipliers[var.op.name]\n    grads_and_vars_mult.append((grad, var))\n    tf.histogram_summary('variables\/' + var.op.name, var)\n    tf.histogram_summary('gradients\/' + var.op.name, grad)\n  return opt.apply_gradients(grads_and_vars_mult)\n<\/code><\/pre>\n\n<p>You can find the whole example <a href=\"https:\/\/github.com\/sdemyanov\/tensorflow-worklab\/\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"47518070","CreationDate":"2017-11-27T19:05:01.077","ParentId":null,"OwnerUserId":"3293004","Title":null,"Body":"<blockquote>\n  <p>The first 5 layers would have learning rate of 0.00001 and the last one would have 0.001. Any idea how to achieve this?<\/p>\n<\/blockquote>\n\n<p>There is an easy way to do that using tf.stop_gradient.\nHere is an example with 3 layers:<\/p>\n\n<pre><code>x = layer1(input)\nx = layer2(x)\noutput = layer3(x)\n<\/code><\/pre>\n\n<p>You can shrink your gradient in the first two layers by a ratio of 1\/100:<\/p>\n\n<pre><code>x = layer1(input)\nx = layer2(x)\nx = 1\/100*x + (1-1\/100)*tf.stop_gradient(x)\noutput = layer3(x)\n<\/code><\/pre>\n\n<p>On the layer2, the \"flow\" is split in two branches: one which has a contribution of 1\/100 computes its gradient regularly but with a gradient magnitude shrinked by a proportion of 1\/100, the other branch provides the remaining \"flow\" without contributing to the gradient because of the tf.stop_gradient operator. As a result, if you use a learning rate of 0.001 on your model optimizer, the first two layers will virtually have a learning rate of 0.00001.<\/p>\n"}]}
{"QuestionId":34946937,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2016-01-22T12:38:04.260","AcceptedAnswerId":34968198.0,"OwnerUserId":4838304.0,"Title":"How to enforce feature vector representing label probability with Caffe siamese CNN?","Body":"<p>related to <a href=\"https:\/\/stackoverflow.com\/questions\/34903975\/how-to-create-caffedb-training-data-for-siamese-networks-out-of-image-directory\">How to Create CaffeDB training data for siamese networks out of image directory<\/a><\/p>\n\n<p>If I have  labels. How can I enforce, that the feature vector of size  right before the contrastive loss layer represents some kind of probability for each class? Or comes that automatically with the siamese net design?<\/p>\n","answers":[{"AnswerId":"34968198","CreationDate":"2016-01-23T20:01:37.103","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you only use contrastive loss in a Siamese network, there is no way of forcing the net to classify into the correct label - because the net is only trained using \"same\/not same\" information and does not know the semantics of the different classes.<\/p>\n\n<p>What you can do is train with multiple loss layers.<br>\nYou should aim at training a feature representation that is reach enough for your domain, so that looking at the trained feature vector of some input (in some high dimension) you should be able to easily classify that input to the correct class. Moreover, given that feature representation of two inputs one should be able to easily say if they are \"same\" or \"not same\".<br>\nTherefore, I recommend that you train your deep network with two loss layer with <code>\"bottom\"<\/code> as the output of one of the <code>\"InnerProduct\"<\/code> layers. One loss is the contrastive loss. The other loss should have another <code>\"InnerProduct\"<\/code> layer with <code>num_output: N<\/code> and a <code>\"SoftmaxWithLoss\"<\/code> layer.   <\/p>\n\n<p>A similar concept was used in this work:\n<a href=\"http:\/\/papers.nips.cc\/paper\/5416-deep-learning-face-representation-by-joint-identification-verification.pdf\" rel=\"nofollow\">Sun, Chen, Wang and Tang <em>Deep Learning Face Representation by Joint Identification-Verification<\/em> NIPS 2014<\/a>.<\/p>\n"}]}
{"QuestionId":34950930,"AnswerCount":0,"Tags":"<lua><neural-network><training-data><torch>","CreationDate":"2016-01-22T16:03:15.733","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, why does my neural network model loaded from file not predict any negative elements?","Body":"<p>I've been working on an artificial neural network model in Torch \/ Lua, able to learn how to recognize true couples (ones) and false couples (zeros) of elements for a particular biological project.<\/p>\n\n<p>The algorithm behaves this way: after the optimization phase in which the software select the best combination of hyper-parameters for the neural network, it runs a training phase, in which the software trains the neural network with the selected combination of best hyper-parameters, and finally saves the trained model to a file (<a href=\"https:\/\/stackoverflow.com\/questions\/33288006\/torch-lua-how-to-save-a-trained-neural-network-model-to-file\">as explained here<\/a>). During the training, the model predicts both true couples and false couples.<\/p>\n\n<p>Everything works so far...<\/p>\n\n<p>Afterwards, in the testing phase, my software loads the model previously saved to a file, and applies it to new datasets.\nHere go the problems: <strong>during testing, the neural network surprisingly predicts only ones (\"yes\" couples) and never predicts any zeros (\"false\" couples).<\/strong><\/p>\n\n<p>I've been working on this problem for weeks and I'm really not able to understand why this happens... <strong>the neural network model correctly predicts ones and zeros during training, so why does it not predict any zeros during testing?<\/strong>\nPerhaps something unexplained happens in the model saving.<\/p>\n\n<p>Any idea about what's happening?<\/p>\n\n<p>Any suggestions would be very appreciated!<\/p>\n","answers":[]}
{"QuestionId":34951592,"AnswerCount":2,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2016-01-22T16:35:15.503","AcceptedAnswerId":null,"OwnerUserId":4720623.0,"Title":"Theano: How to take a \"matrix outer product\" where the elements are matrices","Body":"<p>Basically, I have 2 tensors: A, where A.shape = , and B, where B.shape = . What I would like to do is to get a tensor, C, with shape  such that :<\/p>\n\n<pre><\/pre>\n\n<p>Can this be done efficiently in Theano?<\/p>\n\n<p>Side note: The actual end result that I would like to achieve is to have a tensor, E, of shape  such that :<\/p>\n\n<pre><\/pre>\n\n<p>So, if there is a way to get this directly, I would prefer it (saves on space hopefully). <\/p>\n","answers":[{"AnswerId":"34952148","CreationDate":"2016-01-22T17:01:43.773","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>One approach could be suggested that uses <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow\"><code>broadcasting<\/code><\/a> -<\/p>\n\n<pre><code>(A[:,None]*B).sum(2)\n<\/code><\/pre>\n\n<p>Please note that the intermediate array being created would be of shape <code>(N, K, H, D)<\/code> before sum-reduction on <code>axis=2<\/code> reduces it to <code>(N,K,D).<\/code><\/p>\n"},{"AnswerId":"39436393","CreationDate":"2016-09-11T13:03:52.870","ParentId":null,"OwnerUserId":"4798943","Title":null,"Body":"<p>You can get the final three dimensional result <code>E<\/code> without creating the large intermediate array using <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.batched_dot\" rel=\"nofollow\"><code>batched_dot<\/code><\/a>:\n<\/p>\n\n<pre><code>import theano.tensor as tt\nA = tt.tensor3('A')  # A.shape = (D, N, H)\nB = tt.tensor3('B')  # B.shape = (D, H, K)\nE = tt.batched_dot(A, B)  # E.shape = (D, N, K)\n<\/code><\/pre>\n\n<p>Unfortunately this requires you to permute the dimensions on your input and output arrays. Though this can be done with <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.dimshuffle\" rel=\"nofollow\"><code>dimshuffle<\/code><\/a> in Theano it seems <code>batched_dot<\/code> can't cope with arbitrarily strided arrays and so the following raises a <code>ValueError: Some matrix has no unit stride<\/code> when <code>E<\/code> is evaluated:<\/p>\n\n<pre><code>import theano.tensor as tt\nA = tt.tensor3('A')  # A.shape = (N, H, D)\nB = tt.tensor3('B')  # B.shape = (K, H, D)\nA_perm = A.dimshuffle((2, 0, 1))  # A_perm.shape = (D, N, H)\nB_perm = B.dimshuffle((2, 1, 0))  # B_perm.shape = (D, H, K)\nE_perm = tt.batched_dot(A_perm, B_perm)  # E_perm.shape = (D, N, K)\nE = E_perm.dimshuffle((1, 2, 0))  # E.shape = (N, K, D)\n<\/code><\/pre>\n\n<p><code>batched_dot<\/code> uses <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\"><code>scan<\/code><\/a> along the first (size <code>D<\/code>) dimension. As <code>scan<\/code> is performed sequentially this could be computationally less efficient than computing all the products in parallel if running on a GPU. <\/p>\n\n<p>You can tradeoff between the memory efficiency of the <code>batched_dot<\/code> approach and parallelism in the broadcasting approach using <code>scan<\/code> explicitly. Idea would be to calculate the full product <code>C<\/code> for batches of size <code>M<\/code> in parallel (assuming <code>M<\/code> is an exact factor of <code>D<\/code>), iterating over batches with <code>scan<\/code>:\n<\/p>\n\n<pre><code>import theano as th\nimport theano.tensor as tt\nA = tt.tensor3('A')  # A.shape = (N, H, D)\nB = tt.tensor3('B')  # B.shape = (K, H, D)\nA_batched = A.reshape((N, H, M, D \/ M))\nB_batched = B.reshape((K, H, M, D \/ M))\nE_batched, _ = th.scan(\n    lambda a, b: (a[:, :, None, :] * b[:, :, :, None]).sum(1),\n    sequences=[A_batched.T, B_batched.T]\n)\nE = E_batched.reshape((D, K, N)).T  # E.shape = (N, K, D)\n<\/code><\/pre>\n"}]}
{"QuestionId":34959853,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-23T05:08:17.600","AcceptedAnswerId":34964654.0,"OwnerUserId":2084314.0,"Title":"How to make an if statement using a boolean Tensor","Body":"<p>How do I make an if statement using a boolean tensor?  To be more precise, I'm trying to compare a tensor of size 1 to a constant, checking to see if the value in the tensor is less than the constant.  I figured out I have to make the constant its own size 1 tensor and use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#less\" rel=\"noreferrer\">this<\/a> method to check if the first tensor is less than the second tensor, but I'm not sure how to make the resulting boolean tensor fit correctly into an if statement.  Just putting it in as the query for the if statement makes if statement always return true.<\/p>\n\n<p>EDIT:  This is more or less what the code looked like.  However, I'm getting the error  regardless of whether it has parameters or not, which makes me think the problem is instead that it's not returning a TensorFlow object.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34961322","CreationDate":"2016-01-23T08:29:04.933","ParentId":null,"OwnerUserId":"15161","Title":null,"Body":"<p>use <code>reshape(t, [])<\/code> to obtain the value and use that in your if-statement<\/p>\n"},{"AnswerId":"34964654","CreationDate":"2016-01-23T14:26:36.607","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> You need to use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a> to get a Python boolean, but there are other ways to achieve the same result that might be more efficient.<\/p>\n\n<p>It looks like you've already figured out how to get a boolean tensor from your value, but for the benefit of other readers, it would look something like this:<\/p>\n\n<pre><code>computed_val = ...\nconstant_val = tf.constant(37.0)\npred = tf.less(computed_val, constant_val)  # N.B. Types of the two args must match\n<\/code><\/pre>\n\n<p>The next part is how to use it as a conditional. The simplest thing to do is to use a Python <code>if<\/code> statement, but to do that you must <strong>evaluate<\/strong> the tensor <code>pred<\/code> using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a>:<\/p>\n\n<pre><code>sess = tf.Session()\n\nif sess.run(pred):\n  # Do something.\nelse:\n  # Do something else.\n<\/code><\/pre>\n\n<p>One caveat about using a Python <code>if<\/code> statement is that you have to evaluate the whole expression up to <code>pred<\/code>, which makes it tricky to reuse intermediate values that have already been computed. I'd like to draw your attention to two other ways you can compute conditional expressions using TensorFlow, which don't require you to evaluate the predicate and get a Python value back. <\/p>\n\n<p>The first way uses the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/control_flow_ops\/comparison_operators#select\" rel=\"noreferrer\"><code>tf.select()<\/code><\/a> op to conditionally pass through values from two tensors passed as arguments:<\/p>\n\n<pre><code>pred = tf.placeholder(tf.bool)  # Can be any computed boolean expression.\nval_if_true = tf.constant(28.0)\nval_if_false = tf.constant(12.0)\nresult = tf.select(pred, val_if_true, val_if_false)\n\nsess = tf.Session()\nsess.run(result, feed_dict={pred: True})   # ==&gt; 28.0\nsess.run(result, feed_dict={pred: False})  # ==&gt; 12.0\n<\/code><\/pre>\n\n<p>The <code>tf.select()<\/code> op works element-wise on all of its arguments, which allows you to combine values from the two input tensors. See <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/control_flow_ops\/comparison_operators#select\" rel=\"noreferrer\">its documentation<\/a> for more details. The drawback of <code>tf.select()<\/code> is that it evaluates both <code>val_if_true<\/code> and <code>val_if_false<\/code> before computing the result, which might be expensive if they are complicated expressions.<\/p>\n\n<p>The second way uses the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2c1ab21a6ff9d267d6d93db149720bd1e2ad13c9\/tensorflow\/python\/ops\/control_flow_ops.py#L1103\" rel=\"noreferrer\"><code>tf.cond()<\/code><\/a> op, which conditionally evaluates one of two expressions. This is particularly useful if the expressions are expensive, and it is essential if they <a href=\"https:\/\/stackoverflow.com\/questions\/34931121\/can-cond-support-tf-ops-with-side-effects\">have side effects<\/a>. The basic pattern is to specify two Python functions (or lambda expressions) that build subgraphs that will execute on the true or false branches:<\/p>\n\n<pre><code># Define some large matrices\na = ...\nb = ...\nc = ...\n\npred = tf.placeholder(tf.bool)\n\ndef if_true():\n  return tf.matmul(a, b)\n\ndef if_false():\n  return tf.matmul(b, c)\n\n# Will be `tf.cond()` in the next release.\nfrom tensorflow.python.ops import control_flow_ops\n\nresult = tf.cond(pred, if_true, if_false)\n\nsess = tf.Session()\nsess.run(result, feed_dict={pred: True})   # ==&gt; executes only (a x b)\nsess.run(result, feed_dict={pred: False})  # ==&gt; executes only (b x c)\n<\/code><\/pre>\n"}]}
{"QuestionId":34961088,"AnswerCount":1,"Tags":"<python><ubuntu><caffe>","CreationDate":"2016-01-23T08:01:53.920","AcceptedAnswerId":null,"OwnerUserId":5829424.0,"Title":"Setting up caffe on Ubuntu 14.04 but facing errors when running classify.py","Body":"<p>I'm installing Caffe on an Ubuntu 14.04 virtual server with CUDA installed (without driver) using <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-14.04-VirtualBox-VM<\/a> as inspiration. I've installed all the necessary dependencies and have followed all the instructions step by step but get the error below when I try to test the installation. <\/p>\n\n<p>In regards to the step:\n\" Modify python\/classify.py to add the --print_results option\"<\/p>\n\n<p>I had amended the code via in classify.py to be identical to the official caffe distribution. Not sure if that is the step that is causing the problem but thought I'd add that extra piece of information just in case. <\/p>\n\n<p>The error i get is as below:<\/p>\n\n<pre><\/pre>\n\n<p>Any input would be greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"34962867","CreationDate":"2016-01-23T11:28:23.307","ParentId":null,"OwnerUserId":"3159253","Title":null,"Body":"<p>Check your current shell environment, these variables: <code>LC_ALL<\/code>, <code>LC_CTYPE<\/code>, <code>LANG<\/code>, <code>LANGUAGE<\/code>:<\/p>\n\n<pre><code>set | grep -a '^LC_ALL=.'; set | grep -a '^LC_CTYPE=.';\nset | grep -a '^LANG=.'; set | grep -a '^LANGUAGE=.'\n<\/code><\/pre>\n\n<p>Likely the first printed line will contain one of these variables equal to 'UTF-8'. This is wrong.<\/p>\n\n<p>A proper locale should have language and optional country and encoding specification, like 'en' or 'en_US' or 'ru_RU.UTF-8'. There's a special locale 'C' used as fallback.<\/p>\n\n<p>So you may redefine your locale permanently in <code>\/etc\/default\/locale<\/code> (don't forget to reload your settings, the easiest way is to logoff\/logon), or simply override it for a particular command:<\/p>\n\n<pre><code>&gt; LC_ALL=C python python\/classify.py --print_results examples\/images\/cat.jpg foo\n<\/code><\/pre>\n\n<p>Regarding libdc1394: if you don't need interaction with FireWire check <a href=\"https:\/\/stackoverflow.com\/questions\/12689304\/ctypes-error-libdc1394-error-failed-to-initialize-libdc1394\">ctypes error: libdc1394 error: Failed to initialize libdc1394<\/a> <\/p>\n"}]}
{"QuestionId":34961102,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><keras>","CreationDate":"2016-01-23T08:02:58.343","AcceptedAnswerId":null,"OwnerUserId":1472128.0,"Title":"How to predict using Keras?","Body":"<p>I'm using model.predict() to make predictions batch by batch.<\/p>\n\n<p>By doing so, I find that Keras seems to normalize the input on a batch basis which means if I combine the results, they will become meaningless because they are not normalized together and the probabilities vary a lot from batch to batch.<\/p>\n\n<p>So what should I do to tackle this? Loading all the data into one batch is not an option as the dataset is too large. Should I pass in one sample per time at the cost of performance?<\/p>\n","answers":[]}
{"QuestionId":34961725,"AnswerCount":0,"Tags":"<amazon-ec2><theano><pymc><pymc3>","CreationDate":"2016-01-23T09:20:16.533","AcceptedAnswerId":null,"OwnerUserId":1509797.0,"Title":"How to benefit from GPU with PYMC3","Body":"<p>I see zero difference in PYMC3 speed when using GPU vs. CPU.<\/p>\n\n<p>I am fitting a model that requires 500K+ samples to converge. Obviously it is very slow, so I tried to speed things up with GPU (using GPU instance on EC2).\nTheano reports to be using GPU, so I believe CUDA\/Theano are configured correctly. However, I strongly suspect that Pymc3 is not utilising GPU.<\/p>\n\n<ul>\n<li>do I need to set my variables to TensorType(float32, scalar) explicitly? Currently, they are float64.<\/li>\n<li>Are only some samplers\/likelihoods can benefit from CUDA? I am fitting Poisson-based model and so using Metropolis sampler, not NUTS<\/li>\n<li>is there a way to check that pymc3 is using GPU?<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":34964018,"AnswerCount":1,"Tags":"<optimization><lua><neural-network><deep-learning><torch>","CreationDate":"2016-01-23T13:25:27.000","AcceptedAnswerId":35116491.0,"OwnerUserId":3654968.0,"Title":"Can Torch optim package support multiple inputs","Body":"<p>I'm trying to use the torch7 <a href=\"https:\/\/github.com\/torch\/optim\/blob\/master\/adam.lua\" rel=\"nofollow\" title=\"optim package adam algorithm implementation\">optim package adam algorithm implementation<\/a>\nfor optimizing a neural network which takes two independent inputs. Can this be done? The code seems to only support a single input vector. Is there some other implementation which can take a generic table of inputs?\nThe reference usage I saw, upon which I based my code is <a href=\"https:\/\/github.com\/jcjohnson\/neural-style\/blob\/master\/neural_style.lua\" rel=\"nofollow\">here<\/a><\/p>\n","answers":[{"AnswerId":"35116491","CreationDate":"2016-01-31T16:53:52.067","ParentId":null,"OwnerUserId":"217802","Title":null,"Body":"<p>In principle, if you flatten all your parameters and gradients into two tensors then optim can handle it. See <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/Module.lua#L285\" rel=\"nofollow\">getParameters()<\/a> function.<\/p>\n"}]}
{"QuestionId":34965139,"AnswerCount":2,"Tags":"<numpy><image-processing><lua><torch>","CreationDate":"2016-01-23T15:14:49.170","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Reading .nrrd file into Lua","Body":"<p>I have a .nrrd file which I can read into python and save as a np array. I want to use the resulting array in lua\/torch, how would I go about this? Or is there a way of reading a .nrrd file straight into lua? Thanks.<\/p>\n","answers":[{"AnswerId":"37023122","CreationDate":"2016-05-04T09:04:07.473","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>Now you could also try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a>, essentially, you can use torch and any lua libraries in python. <\/p>\n\n<p>The conversion will be done with <code>torch.fromNumpyArray(arr)<\/code>, and you will get a torch tensor. there is another function <code>tensor.asNumpyArray()<\/code>, helps you convert back to numpy array. <\/p>\n\n<pre><code>import lutorpy as lua\nimport numpy as np\n\nxn = np.random.randn(100)\n## convert the numpy array into torch tensor\ntensor_xn = torch.fromNumpyArray(xn)\n\n# you can use torch tensor as well\nt = torch.DoubleTensor(10,3)\nprint(t._size()) # the corresponding lua version is t:size()\n\n## convert torch tensor to numpy array\narr = t.asNumpyArray()\nprint(arr.shape)\n<\/code><\/pre>\n\n<p>Mostly, the conversion is done instantly with shared memory between your numpy array and torch tensor, no disk saving and even <strong>no memory copy<\/strong>.<\/p>\n"},{"AnswerId":"35303139","CreationDate":"2016-02-09T22:06:08.687","ParentId":null,"OwnerUserId":"4509141","Title":null,"Body":"<p>As in comments, <a href=\"https:\/\/github.com\/htwaijry\/npy4th\" rel=\"nofollow\" title=\"npy4th\">npy4th<\/a> should work for what you want. Anyways, i also found a python library that looks easier for this.<\/p>\n\n<p>You can export the result into Lua in many ways, like outputting to a text file. I drop some URLs that can be helpful. As I've seen, it is pretty simple to use in Python, just with an <code>import nrrd<\/code> and a <code>frames, options = nrrd.read(\"test.nrrd\")<\/code>.<\/p>\n\n<p><a href=\"http:\/\/github.com\/mhe\/pynrrd#pynnrd\" rel=\"nofollow\" title=\"GitHub Documentation\">GitHub documentation for Pynrrd<\/a><\/p>\n"}]}
{"QuestionId":34966614,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2016-01-23T17:36:27.810","AcceptedAnswerId":null,"OwnerUserId":2824962.0,"Title":"Sequence labeling in Keras","Body":"<p>I'm working on sentence labeling problem. I've done embedding and padding by myself and my inputs look like:<\/p>\n\n<pre><\/pre>\n\n<p>For every word in sentence I want to predict one of four classes, so my desired output should look like:<\/p>\n\n<pre><\/pre>\n\n<p>My simple network architecture is:<\/p>\n\n<pre><\/pre>\n\n<p>It shows approximately 95% while training, but when I'm trying to predict new sentences using trained model results are really bad. It looks like model just learnt some classes for first words and shows it every time. I think the problem can is:<\/p>\n\n<ol>\n<li><p>Written by myself padding (zero vectors in the end of the sentence), can it make learning worse?<\/p><\/li>\n<li><p>I should try to learn sentences of different length, without padding (if yes, can you help me how train such kind of a model in Keras?)<\/p><\/li>\n<li><p>Wrong objective of learning, but I tried mean squared error, binary cross entropy and others, it doesn't change.<\/p><\/li>\n<li><p>Something with  and , I think, that I've got how it works, but still not 100% sure.<\/p><\/li>\n<\/ol>\n\n<p>I'll be glad to see any hint or help regarding to this problem, thank you!<\/p>\n","answers":[{"AnswerId":"41395316","CreationDate":"2016-12-30T11:12:48.143","ParentId":null,"OwnerUserId":"1360133","Title":null,"Body":"<p>I personally think that you misunderstand what \"sequence labeling\" means. <\/p>\n\n<p>Do you mean:<\/p>\n\n<ol>\n<li><code>X<\/code> is a list of sentences, each element <code>X[i]<\/code> is a word sequence of arbitrary length?<\/li>\n<li><code>Y[i]<\/code> is the category of <code>X[i]<\/code>, and the one hot form of <code>Y[i]<\/code> is a <code>[0, 1, 0, 0]<\/code> like array?<\/li>\n<\/ol>\n\n<p>If it is, then it's not a sequence labeling problem, it's a classification problem. <\/p>\n\n<p>Don't use <code>TimeDistributedDense<\/code>, and if it is a multi-class classification problem, i.e., <code>len(Y[i]) &gt; 2<\/code>, then use \"<code>categorical_crossentropy<\/code>\" instead of \"<code>binary_crossentropy<\/code>\"<\/p>\n"}]}
{"QuestionId":34967312,"AnswerCount":1,"Tags":"<python><speech-recognition><keras><lstm><recurrent-neural-network>","CreationDate":"2016-01-23T18:44:00.447","AcceptedAnswerId":34968916.0,"OwnerUserId":5444351.0,"Title":"How to stack LSTM layers to classify speech files","Body":"<p>I have been trying to implement a LSTM based classifier to classify descrete speech. I have created feature vectors with 13 mfcc. For a given file have 2D vector of [99, 13]. After following the mnist_irnn example I could set up single layer RNN to classify my speech files. But now I want to add more layers to the network. Therefore, I have been trying to implement the network with two LSTM layers and softmax layer as the output layer. After going through number of posts here I could set up the network as follows, where it doesn't throw any exceptions during model building time. <\/p>\n\n<pre><\/pre>\n\n<p>I have been trying different values at different points. (For the moment I have been trying with a small sample, therefore values are very small) But, now it is throwing exception during training. Some dimension mismatch.<\/p>\n\n<pre><\/pre>\n\n<p>I would like to know what I am doing wrong here. I have been going through the code whole day, but still I am unable to figure out the reason for dimension mismatch.<\/p>\n\n<p>Further, I would be really thankful if someone can explain what it is meant by output_dim. (Is that the shape of the vector output by a single node, when we have n number of nodes in a given layer? Should it be equal to the number of nodes in the next layer? )<\/p>\n","answers":[{"AnswerId":"34968916","CreationDate":"2016-01-23T21:11:20.663","ParentId":null,"OwnerUserId":"432021","Title":null,"Body":"<p>You have a problem with <code>Y<\/code>dimension, the output should be something like <code>(100, 99, 10)<\/code>, that is a set of sequences of outputs, same as features, just 1 in output. It seems your <code>Y<\/code> vector is different. Method <code>to_categorical<\/code> is not really applicable to a sequences, it expects a vector.<\/p>\n\n<p>Alternatively you can output a single vector and feed it into a dense layer in in the last LSTM layer with <code>return_sequences=False<\/code><\/p>\n\n<p>You do not need stateful network as well.<\/p>\n"}]}
{"QuestionId":34968899,"AnswerCount":1,"Tags":"<python><tensorflow><parameterized>","CreationDate":"2016-01-23T21:10:04.053","AcceptedAnswerId":34969301.0,"OwnerUserId":1815451.0,"Title":"Parameterized model over output functions using tensorflow","Body":"<p>Is it possible to make parameterized model over output functions using tensorflow? And how? <\/p>\n\n<p>example: \nI want model with used softmax function at the end, so<\/p>\n\n<pre><\/pre>\n\n<p>but then I would like to try sigmoidal output and mean squares, in that case:<\/p>\n\n<pre><\/pre>\n\n<p>my question is, how can one make parametrized model with specification of cost function? I don't have tensorflow placeholders before calling constructor for such model. <\/p>\n","answers":[{"AnswerId":"34969301","CreationDate":"2016-01-23T21:49:55.647","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Put both functions in the graph, and then during evaluation call either one or the other. The evaluation framework will select subset of the graph that's needed for the particular version of cost during evaluation<\/p>\n\n<pre><code>cost1 = tf.reduce_mean(\n    tf.nn.softmax_cross_entropy_with_logits(output_potentials, y))\ncost2 = tf.reduce_sum(\n    tf.pow(tf.sub(tf.nn.sigmoid(output_potentials), y), 2.0))\nsess.run([cost1])\nsess.run([cost2])\n<\/code><\/pre>\n"}]}
{"QuestionId":34970582,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-24T00:04:40.923","AcceptedAnswerId":null,"OwnerUserId":2383830.0,"Title":"Using a variable for num_splits for tf.split()","Body":"<p>Is it possible to use a placeholder input for the num_split parameter to tf.split()?<\/p>\n\n<p>I would ideally like to do something like:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>TypeError: Expected int for argument 'num_split' not\n  .<\/p>\n<\/blockquote>\n\n<p>There is possibly something wrong with my approach. I am looking to enumerate across a dimension in a variable shape tensor. Thanks!<\/p>\n","answers":[{"AnswerId":"56015552","CreationDate":"2019-05-07T04:25:43.443","ParentId":null,"OwnerUserId":"5834394","Title":null,"Body":"<p>A solution to a similar problem in an unrelated circumstance is using a placeholder for a fixed length of maximum size. Suppose I have a sequence of length 40.     <code>t = tf.range(40)<\/code>. \nNow  at run time, I get a split,  (say <code>x = [6,9,10,5,1]<\/code>). Now follow the steps <\/p>\n\n<p><strong>Step 1<\/strong>: \nDetermine the maximum number of splits there can be, say: <strong>19<\/strong> <\/p>\n\n<p><strong>Step 2<\/strong>: <\/p>\n\n<pre><code>    num_splits = tf.placeholder(tf.int32, [19]) \n    y= tf.split(t, num_or_size_splits=num_splits, axis=0)\n<\/code><\/pre>\n\n<p>This will break the sequence into run-time determined split sizes<\/p>\n\n<p><strong>Step 4<\/strong>: \nAt run time : <\/p>\n\n<pre><code>    x = [6,9,10,5,1] \n    x += [40-sum(x)] + [0 for i in range(19-1-len(x))]\n<\/code><\/pre>\n\n<p>First line means the actual split  sizes we need<br>\nSplit requires the split sizes should sum upto total split size, i.e. 40  in this case, and 0 are the split sizes for the left over splits.<\/p>\n\n<p><code>session.run(y, feed_dict={num_splits:x})<\/code> will show the results like :   <\/p>\n\n<pre><code>[0, 1, 2, 3, 4, 5]  \n[ 6,  7,  8,  9, 10, 11, 12, 13, 14]  \n[15, 16, 17, 18, 19, 20, 21, 22, 23, 24]  \n[25, 26, 27, 28, 29]  \n[30]  \n[31, 32, 33, 34, 35, 36, 37, 38, 39]  \n[]\n.\n.\n[]  \n<\/code><\/pre>\n\n<p><strong>Step 5<\/strong>: \n(Optional, preferred) pad with zeros till the max length of the sequence <\/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\n\n    m = []\n    for t1 in y :\n      t1=tf.reshape(t1,[1,-1])\n      t_padded = pad_up_to(t1, [1,15], 0)\n      session.run(t_padded  , feed_dict={num_splits:x})  \n      m+=  [t_padded]\n    m= tf.concat(m,0)\n<\/code><\/pre>\n\n<p>This will pad the chunks with zeros to create equal sized chunks. <\/p>\n\n<p>NOTE: <em>The above methodology helped me in converting sequences into sentences  (variable number of sentences) for NLP related tasks<\/em> <\/p>\n\n<p>funciton : pad_up_to() is  taken from\n<a href=\"https:\/\/stackoverflow.com\/questions\/42334646\/tensorflow-pad-unknown-size-tensor-to-a-specific-size\">Q:42334646<\/a><\/p>\n"},{"AnswerId":"34971110","CreationDate":"2016-01-24T01:15:54.910","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>There's a general philosophy of \"tensor in-tensor out\" for core graph ops, so it may simplify things if you can restructure your computation to deal with single tensor of variable size instead of variable number of tensors.<\/p>\n\n<p>Ops like <code>pack<\/code>, <code>unpack<\/code>, <code>split<\/code> deal with multiple tensors but they compile into \"tensor-in\/tensor-out\" ops during graph construction time, which is why <code>num_splits<\/code> needs to be fixed. Ops like <code>dynamic_partition<\/code>, <code>dynamic_stitch<\/code>, <code>dequeue_many<\/code> take over some of that functionality for single tensors with variable <code>0<\/code>-th dimension.<\/p>\n\n<p>If you really need to deal with variable number of tensors, typical approach is break computation into multiple <code>session.run<\/code> calls, with one input tensor per <code>run<\/code> call, and tie things together using queues. There's a <code>slice_input_producer<\/code> which splits variable sized input along 0'th dimension and produces a tensor for each row, so if you wanted to evaluate <code>myfunction<\/code> in a loop on each row of <code>inputs<\/code> you could do this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def myfunction(vector):\n  result = tf.reduce_sum(vector)\n  print_result = tf.Print(result, [result], \"myfunction called \")\n  return print_result\n\nMAX_ROWS = 10\n\n# input matrix with 2 columns and unknown number of rows (&lt;MAX_ROWS)\ninputs = tf.placeholder(tf.int32, [None, 2])\n# copy of inputs, will need to have a persistent copy of it because we will\n# be fetching rows in different session.run calls\ndata = tf.Variable(inputs, validate_shape=False)\n# input producer that iterates over the rows and pushes them onto Queue\nrow = tf.train.slice_input_producer([data], num_epochs=1, shuffle=False)[0]\nmyfunction_op = myfunction(row)\n\n# this op will save placeholder values into the variable\ninit_op = tf.initialize_all_variables()\n\n# Coordinator is not necessary in this case, but you'll need it if you have\n# more than one Queue in order to close all queues together\nsess = tf.Session()\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\nsess.run([init_op], feed_dict={inputs:[[0, 0], [1, 1], [2, 2]]})\n\ntry:\n  for i in range(MAX_ROWS):\n    sess.run([myfunction_op])\nexcept tf.errors.OutOfRangeError:\n  print('Done iterating')\nfinally:\n  # When done, ask other threads to stop.\n  coord.request_stop()\n<\/code><\/pre>\n\n<p>If you run this, you should see<\/p>\n\n<pre><code>myfunction called [0]\nmyfunction called [2]\nmyfunction called [4]\nDone iterating\n<\/code><\/pre>\n"}]}
{"QuestionId":34970681,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-01-24T00:16:23.537","AcceptedAnswerId":34973370.0,"OwnerUserId":4380945.0,"Title":"deep learning - a number of naive questions about caffe","Body":"<p>I am trying to understand the basics of caffe, in particular to use with python.<\/p>\n\n<p>My understanding is that the model definition (say a given neural net architecture) must be included in the  file.<\/p>\n\n<p>And that when you train the model on data using the , you save the weights\/model parameters to a  file<\/p>\n\n<p>Also, there is a difference between the  file used for training (which includes learning rate and regularization parameters) and the one used for testing\/deployment, which does not include them.<\/p>\n\n<p><strong>Questions:<\/strong><\/p>\n\n<ol>\n<li>is it correct that the  is the basis for training and that\nthe  is the result of training (weights), using the\n on the training data?<\/li>\n<li>is it correct that there is a  for training and one for\ntesting, and that there are only slight differences (learning rate\nand regularization factors on training), but that the nn\narchitecture (assuming you use neural nets) is the same?<\/li>\n<\/ol>\n\n<p>Apologies for such basic questions and possibly some very incorrect assumptions, I am doing some online research and the lines above summarize my understanding to date.<\/p>\n","answers":[{"AnswerId":"34973370","CreationDate":"2016-01-24T07:46:58.950","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Let's take a look at one of the examples provided with BVLC\/caffe: <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_reference_caffenet\" rel=\"nofollow noreferrer\"><code>bvlc_reference_caffenet<\/code><\/a>.<br>\nYou'll notice that in fact there are <strong>3<\/strong> <code>'.prototxt'<\/code> files:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/train_val.prototxt\" rel=\"nofollow noreferrer\"><code>train_val.prototxt<\/code><\/a>: this file describe the net architecture for the training phase.  <\/li>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt\" rel=\"nofollow noreferrer\"><code>depoly.prototxt<\/code><\/a>: this file describe the net architecture for test time (\"deployment\").  <\/li>\n<li><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/solver.prototxt\" rel=\"nofollow noreferrer\"><code>solver.prototxt<\/code><\/a>: this file is very small and contains \"meta parameters\" for training. For example, the <a href=\"https:\/\/stackoverflow.com\/questions\/30033096\/what-is-lr-policy-in-caffe\/30045244#30045244\">learning rate policy<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/32177764\/what-is-weight-decay-meta-parameter-in-caffe\">regulariztion<\/a> etc.<\/li>\n<\/ul>\n\n<p>The net architecture represented by <code>train_val.prototxt<\/code> and <code>deploy.prototxt<\/code> should be mostly similar. There are few main difference between the two:<\/p>\n\n<ul>\n<li><p>Input data: during training one usually use a predefined set of inputs for training\/validation. Therefore, <code>train_val<\/code> usually contains an explicit input layer, e.g., <code>\"HDF5Data\"<\/code> layer or a <code>\"Data\"<\/code> layer. On the other hand, <code>deploy<\/code> usually does not know in advance what inputs it will get, it only contains a statement:<\/p>\n\n<pre><code>input: \"data\"\ninput_shape {\n  dim: 10\n  dim: 3\n  dim: 227\n  dim: 227\n}\n<\/code><\/pre>\n\n<p>that declares what input the net expects and what should be its dimensions.<br>\nAlternatively, One can put an <a href=\"http:\/\/caffe.help\/manual\/layers\/input.html\" rel=\"nofollow noreferrer\"><code>\"Input\"<\/code><\/a> layer:  <\/p>\n\n<pre><code>layer {\n  name: \"input\"\n  type: \"Input\"\n  top: \"data\"\n  input_param { shape { dim: 10 dim: 3 dim: 227 dim: 227 } }\n}\n<\/code><\/pre><\/li>\n<li>Input labels: during training we supply the net with the \"ground truth\" expected outputs, this information is obviously not available during <code>deploy<\/code>.   <\/li>\n<li>Loss layers: during training one must define a loss layer. This layer tells the solver in what direction it should tune the parameters at each iteration. This loss compares the net's current prediction to the expected \"ground truth\". The gradient of the loss is back-propagated to the rest of the net and this is what drives the learning process. During <code>deploy<\/code> there is no loss and no back-propagation.<\/li>\n<\/ul>\n\n<p>In caffe, you supply a <code>train_val.prototxt<\/code> describing the net, the train\/val datasets and the loss. In addition, you supply a <code>solver.prototxt<\/code> describing the meta parameters for training. The <em>output<\/em> of the training process is a <code>.caffemodel<\/code> binary file containing the trained parameters of the net.<br>\nOnce the net was trained, you can use the <code>deploy.prototxt<\/code> with the <code>.caffemodel<\/code> parameters to predict outputs for new and unseen inputs.<\/p>\n"},{"AnswerId":"34973419","CreationDate":"2016-01-24T07:54:19.223","ParentId":null,"OwnerUserId":"222216","Title":null,"Body":"<p>Yes but, there is diffrent types of  .prototxt files\nfor example<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt<\/a><\/p>\n\n<p>this is for the training and testing network<\/p>\n\n<p>for commandline training ypu can use a solver file which is also  .prototxt file for example<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt<\/a><\/p>\n"}]}
{"QuestionId":34970818,"AnswerCount":1,"Tags":"<python><neural-network><time-series><keras><recurrent-neural-network>","CreationDate":"2016-01-24T00:34:50.717","AcceptedAnswerId":45060104.0,"OwnerUserId":589451.0,"Title":"Training, testing, and validation sets for bidirectional LSTM (BLSTM)","Body":"<p>When it comes to normal ANNs, or any of the standard machine learning techniques, I understand what the training, testing, and validation sets should be (both conceptually, and the rule-of-thumb ratios). However, for a bidirectional LSTM (BLSTM) net, how to split the data is confusing me.<\/p>\n\n<p>I am trying to improve prediction on individual subject data that consists of monitored health values. In the simplest case, for each subject, there is one long time series of values (>20k values), and contiguous parts of that time series are labeled from a set of categories, depending on the current health of the subject. For a BLSTM, the net is trained on all of the data going forwards and backwards simultaneously. The problem then is, how does one split a time series for one subject? <\/p>\n\n<ul>\n<li>I can't just take the last 2,000 values (for example), because they might all fall into a single category.<\/li>\n<li>And I can't chop the time series up randomly, because then both the learning and testing phases would be made of disjointed chunks.<\/li>\n<\/ul>\n\n<p>Finally, each of the subjects (as far as I can tell) has slightly different (but similar) characteristics. So, maybe, since I have thousands of subjects, do I train on some, test on some, and validate on others? However, since there are inter-subject differences, how would I set up the tests if I was only considering one subject to start?<\/p>\n","answers":[{"AnswerId":"45060104","CreationDate":"2017-07-12T14:11:36.147","ParentId":null,"OwnerUserId":"8296355","Title":null,"Body":"<p>I think this has more to do with your particular dataset than Bi-LSTMs in general.<\/p>\n\n<p>You're confusing splitting a dataset for training\/testing vs. splitting a sequence in a particular sample. It seems like you have many different subjects, which constitute a different sample. For a standard training\/testing split, you would split your dataset between subjects, as you suggested in the last paragraph.<\/p>\n\n<p>For any sort of RNN application, you do NOT split along your temporal sequence; you input your entire sequence as a single sample to your Bi-LSTM. So the question really becomes whether such a model is well-suited to your problem, which has multiple labels at specific points in the sequence. You can use a sequence-to-sequence variant of the LSTM model to predict which label each time point in the sequence belongs to, but again you would NOT be splitting the sequence into multiple parts.<\/p>\n"}]}
{"QuestionId":34971710,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2016-01-24T02:57:00.360","AcceptedAnswerId":34971733.0,"OwnerUserId":3765905.0,"Title":"TensorFlow with Docker","Body":"<p>I've created a Docker machine to run TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>Then I try to launch a container with the TesnorFlow dev image:<\/p>\n\n<pre><\/pre>\n\n<p>But I get the following printout and error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure what to do here... <strong>Will someone explain what this means and how I should proceed?<\/strong><\/p>\n\n<p>Extra info: I'm on Mac OSX 10.9,<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34971733","CreationDate":"2016-01-24T03:02:02.593","ParentId":null,"OwnerUserId":"610569","Title":null,"Body":"<p>Instead of the default docker create, try this:<\/p>\n\n<pre><code>docker-machine create --driver virtualbox --virtualbox-disk-size \"20000\" tensorflow\n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>eval $(docker-machine env tensorflow)\ndocker run -it b.gcr.io\/tensorflow\/tensorflow:latest-devel\n<\/code><\/pre>\n\n<p><strong>Note<\/strong>: <\/p>\n\n<ul>\n<li>See <a href=\"https:\/\/stackoverflow.com\/questions\/32485723\/docker-increase-disk-space\">Docker increase disk space<\/a><\/li>\n<li>Make sure you have 20 GB on your physical disk<\/li>\n<\/ul>\n"}]}
{"QuestionId":34971964,"AnswerCount":0,"Tags":"<python><segmentation-fault><tensorflow>","CreationDate":"2016-01-24T03:44:09.170","AcceptedAnswerId":null,"OwnerUserId":3568055.0,"Title":"tensorflow: segmentation error in multiple threads","Body":"<p>I build a RNN network using the python ops rnn provided. But it sometimes runs into the error  and sometimes performs well. The output message says , but actually it creates 25 threads. And when I run it on another computer with fewer threads (parallelism threads: 4, actual threads 9), it never occurs this error. I wonder if this error is related to multiple threads and how can I solve it?<\/p>\n\n<p>I used a python module 'faulthandler' to print more error messages:<\/p>\n\n<p>`Thread 0x00007f1047bf9700 (most recent call first):\n  File \"\/usr\/lib\/python2.7\/threading.py\", line 339 in wait\n  File \"\/usr\/lib\/python2.7\/Queue.py\", line 168 in get\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/summary_io.py\", line 197 in run\n  File \"\/usr\/lib\/python2.7\/threading.py\", line 810 in __bootstrap_inner\n  File \"\/usr\/lib\/python2.7\/threading.py\", line 783 in __bootstrap<\/p>\n\n<p>Thread 0x00007f10db79b740 (most recent call first):\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 428 in _do_run\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 368 in run`<\/p>\n","answers":[]}
{"QuestionId":34972881,"AnswerCount":1,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2016-01-24T06:24:13.200","AcceptedAnswerId":34975293.0,"OwnerUserId":5832218.0,"Title":"How to convert trained convolutional network (CNN) to use as a detector?","Body":"<p>I am new in the computer vision and ML. After several exiting weeks with Python and TensorFlow  I have some questions.<\/p>\n\n<p>How to convert trained convolutional network (CNN) to use it as a detector? (get a heatmap)<\/p>\n\n<h2>For ex.:<\/h2>\n\n<h3>We have<\/h3>\n\n<p>trained MNIST model<\/p>\n\n<p>MODEL = [None, 28**2] > [CONV3+POOL] > [CONV3+POOL] > [DENSELAYER] > [None, 10]<\/p>\n\n<h3>I want:<\/h3>\n\n<p>[None, WH] > [MODEL] > [None, w, h, 10]<\/p>\n\n<p>I read that to do so it is necessary to convert the [DENSELAYER] into convolution. Am i right?<\/p>\n\n<p>Thank you in advance for your help and sorry for my english :) the old code looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I think the first fully connected layer should be a 7x7 conv, but I'm not sure how to handle the second fc layer (the output layer):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34975293","CreationDate":"2016-01-24T11:39:52.550","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>Yes. Reshaping the first FC layer to 7x7 is what you want. \nFor the output layer what you're looking for is a <a href=\"https:\/\/www.facebook.com\/yann.lecun\/posts\/10152820758292143\" rel=\"nofollow\">1x1 convolution<\/a>, which basically runs the fully connected output layer on each pixel.<\/p>\n\n<pre><code>wc4 = tf.reshape(_weights['out'], [1, 1, d1, out])\nconv4 = conv2d_valid(conv3, wc4, _biases['out'])\n<\/code><\/pre>\n"}]}
{"QuestionId":34973435,"AnswerCount":0,"Tags":"<ubuntu><caffe>","CreationDate":"2016-01-24T07:58:03.200","AcceptedAnswerId":null,"OwnerUserId":1556826.0,"Title":"Error during make test while installing Caffe","Body":"<p>I am trying to get Caffe up and running on my Ubuntu 14.04 system. After downloading Caffe and installing all the dependencies (no GPU), I run make all, which runs smoothly. When I run make test, I get a mountain of errors.<\/p>\n\n<pre><\/pre>\n\n<p>I figured out that I had multiple versions of OpenCV on my system. I tried removing one from the bash file. I modified the Makefile as per the instructions given here (<a href=\"https:\/\/stackoverflow.com\/questions\/34125052\/caffe-with-opencv-3-and-nvidia-digits-opencv-version-conflict-2-4-vs-3-0\/34211100\">Caffe with OpenCV 3 and NVIDIA Digits: OpenCV version conflict (2.4 vs 3.0)<\/a>), but that is not working. I get another set of errors like this<\/p>\n\n<pre><\/pre>\n\n<p>How do I rectify these errors? <\/p>\n","answers":[]}
{"QuestionId":34973483,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-01-24T08:05:08.920","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"Cannot convert Type TensorType(float64, matrix) into Type TensorType(float32, matrix)","Body":"<p>I wrote a python program execute correctly in a computer. But when i put it in another computer, It has such a problem:   <\/p>\n\n<pre><\/pre>\n\n<p>There is no bugs in the codes because i can run it correctly in my computer. But why it has such a problem in another computer?<\/p>\n","answers":[{"AnswerId":"34973967","CreationDate":"2016-01-24T09:08:44.737","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>If you don't explicitly specify the datatype of a theano tensor, it will attribute <code>theano.config.floatX<\/code>, of which the default can vary according to systems.<\/p>\n\n<p>The error you have here usually comes from using a numpy array of dtype <code>float64<\/code> and trying to feed it into a theano function that was built for <code>float32<\/code> (either <code>floatX == 'float32'<\/code> or you specified it explicitly).<\/p>\n\n<p>A remedy for this is to be explicit at every step about dtype. If you want to use a GPU at some point, sticking to <code>float32<\/code> is best at the moment. Just declare every single tensor and numpy array as <code>float32<\/code>.<\/p>\n"}]}
{"QuestionId":34975884,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2016-01-24T12:38:48.717","AcceptedAnswerId":null,"OwnerUserId":5833020.0,"Title":"How to invoke tf.initialize_all_variables() in C++? tensorflow","Body":"<p>I'm newbie in tensorflow.\nBtw, how to execute ops in C++ such as tf.initialize_all_variables() or some ops returned by optimizers\nself.train_op = self.optimizer.apply_gradients(gradients)\n? I've tried <\/p>\n\n<pre><\/pre>\n\n<p>in python and then in C++:<\/p>\n\n<pre><\/pre>\n\n<p>so I got<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"34980560","CreationDate":"2016-01-24T19:42:26.827","ParentId":null,"OwnerUserId":"5833020","Title":null,"Body":"<p>I've changed <\/p>\n\n<pre><code>status = session-&gt;Run(inputs, {\"init_all_vars_op\"}, {}, &amp;outputs);\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>status = session-&gt;Run(inputs, {}, {\"init_all_vars_op\"}, &amp;outputs);\n<\/code><\/pre>\n\n<p>And now it works fine.<\/p>\n\n<p>According to\n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassSession.html#virtual_Status_tensorflow_Session_Run\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/cc\/ClassSession.html#virtual_Status_tensorflow_Session_Run<\/a>\nSecond parameter is for operations which returns output tensors and third for operations which don't return tensors.<\/p>\n"}]}
{"QuestionId":34982492,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2016-01-24T22:45:31.080","AcceptedAnswerId":35004257.0,"OwnerUserId":1981911.0,"Title":"Restoring TensorFlow model","Body":"<p>I'm trying to restore TensorFlow model. I followed this example:\n<a href=\"http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/\" rel=\"noreferrer\">http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/<\/a><\/p>\n\n<p>At the end of the code in the example I added these lines:<\/p>\n\n<pre><\/pre>\n\n<p>Two files were created: checkpoint and model.ckpt.<\/p>\n\n<p>In a new python file (tomas_bees_predict.py), I have this code:<\/p>\n\n<pre><\/pre>\n\n<p>However when I execute the code, I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>ValueError: No variables to save<\/p>\n\n<p>Is there a way to read mode.ckpt file and see what variables are saved?\nOr maybe someone can help with saving the model and restoring it based on the example described above? <\/p>\n\n<p><strong>EDIT 1:<\/strong><\/p>\n\n<p>I think I tried running the same code in order to recreate model structure and I was getting the error. I think it could be related to the fact that code described here isn't using named variables:\n<a href=\"http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/\" rel=\"noreferrer\">http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/<\/a><\/p>\n\n<pre><\/pre>\n\n<p>So I did this experiment. I wrote two versions of the code (with and without named variables) to save the model and the code to restore the model.<\/p>\n\n<p><strong>tensor_save_named_vars.py<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p><strong>tensor_save_not_named_vars.py:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>tensor_restore.py:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Here is what I get when I execute this code:<\/p>\n\n<pre><\/pre>\n\n<p>So perhaps the original code (see the external link above) could be modified to something like this:<\/p>\n\n<pre><\/pre>\n\n<p>But then the question I have: is restoring weight_var and bias_var variables sufficient to implement the prediction? I did the training on the powerful machine with GPU and I would like to copy the model to the less powerful computer without GPU to run predictions.<\/p>\n","answers":[{"AnswerId":"35004257","CreationDate":"2016-01-25T23:12:28.383","ParentId":null,"OwnerUserId":"1981911","Title":null,"Body":"<p>I think I tried running the same code in order to recreate model structure and I was getting the error. I think it could be related to the fact that code described here isn't using named variables:\n<a href=\"http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/\" rel=\"nofollow\">http:\/\/nasdag.github.io\/blog\/2016\/01\/19\/classifying-bees-with-google-tensorflow\/<\/a><\/p>\n\n<pre><code>def weight_variable(shape):\n  initial = tf.truncated_normal(shape, stddev=0.1)\n  return tf.Variable(initial)\n\ndef bias_variable(shape):\n  initial = tf.constant(0.1, shape=shape)\n  return tf.Variable(initial)\n<\/code><\/pre>\n\n<p>So I did this experiment. I wrote two versions of the code (with and without named variables) to save the model and the code to restore the model.<\/p>\n\n<p><strong>tensor_save_named_vars.py<\/strong>:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Create some variables.\nv1 = tf.Variable(1, name=\"v1\")\nv2 = tf.Variable(2, name=\"v2\")\n\n# Add an op to initialize the variables.\ninit_op = tf.initialize_all_variables()\n\n# Add ops to save and restore all the variables.\nsaver = tf.train.Saver()\n\n# Later, launch the model, initialize the variables, do some work, save the\n# variables to disk.\nwith tf.Session() as sess:\n  sess.run(init_op)\n  print \"v1 = \", v1.eval()\n  print \"v2 = \", v2.eval()\n  # Save the variables to disk.\n  save_path = saver.save(sess, \"\/tmp\/model.ckpt\")\n  print \"Model saved in file: \", save_path\n<\/code><\/pre>\n\n<p><strong>tensor_save_not_named_vars.py:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\n\n# Create some variables.\nv1 = tf.Variable(1)\nv2 = tf.Variable(2)\n\n# Add an op to initialize the variables.\ninit_op = tf.initialize_all_variables()\n\n# Add ops to save and restore all the variables.\nsaver = tf.train.Saver()\n\n# Later, launch the model, initialize the variables, do some work, save the\n# variables to disk.\nwith tf.Session() as sess:\n  sess.run(init_op)\n  print \"v1 = \", v1.eval()\n  print \"v2 = \", v2.eval()\n  # Save the variables to disk.\n  save_path = saver.save(sess, \"\/tmp\/model.ckpt\")\n  print \"Model saved in file: \", save_path\n<\/code><\/pre>\n\n<p><strong>tensor_restore.py:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\n\n# Create some variables.\nv1 = tf.Variable(0, name=\"v1\")\nv2 = tf.Variable(0, 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  print \"v1 = \", v1.eval()\n  print \"v2 = \", v2.eval()\n<\/code><\/pre>\n\n<p>Here is what I get when I execute this code:<\/p>\n\n<pre><code>$ python tensor_save_named_vars.py \n\nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 4\nI tensorflow\/core\/common_runtime\/direct_session.cc:58] Direct session inter op parallelism threads: 4\nv1 =  1\nv2 =  2\nModel saved in file:  \/tmp\/model.ckpt\n\n$ python tensor_restore.py \n\nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 4\nI tensorflow\/core\/common_runtime\/direct_session.cc:58] Direct session inter op parallelism threads: 4\nModel restored.\nv1 =  1\nv2 =  2\n\n$ python tensor_save_not_named_vars.py \n\nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 4\nI tensorflow\/core\/common_runtime\/direct_session.cc:58] Direct session inter op parallelism threads: 4\nv1 =  1\nv2 =  2\nModel saved in file:  \/tmp\/model.ckpt\n\n$ python tensor_restore.py \nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 4\nI tensorflow\/core\/common_runtime\/direct_session.cc:58] Direct session inter op parallelism threads: 4\nW tensorflow\/core\/common_runtime\/executor.cc:1076] 0x7ff953881e40 Compute status: Not found: Tensor name \"v2\" not found in checkpoint files \/tmp\/model.ckpt\n     [[Node: save\/restore_slice_1 = RestoreSlice[dt=DT_INT32, preferred_shard=-1, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](_recv_save\/Const_0, save\/restore_slice_1\/tensor_name, save\/restore_slice_1\/shape_and_slice)]]\nW tensorflow\/core\/common_runtime\/executor.cc:1076] 0x7ff953881e40 Compute status: Not found: Tensor name \"v1\" not found in checkpoint files \/tmp\/model.ckpt\n     [[Node: save\/restore_slice = RestoreSlice[dt=DT_INT32, preferred_shard=-1, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](_recv_save\/Const_0, save\/restore_slice\/tensor_name, save\/restore_slice\/shape_and_slice)]]\nTraceback (most recent call last):\n  File \"tensor_restore.py\", line 14, in &lt;module&gt;\n    saver.restore(sess, \"\/tmp\/model.ckpt\")\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 891, in restore\n    sess.run([self._restore_op_name], {self._filename_tensor_name: save_path})\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 368, in run\n    results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 444, in _do_run\n    e.code)\ntensorflow.python.framework.errors.NotFoundError: Tensor name \"v2\" not found in checkpoint files \/tmp\/model.ckpt\n     [[Node: save\/restore_slice_1 = RestoreSlice[dt=DT_INT32, preferred_shard=-1, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](_recv_save\/Const_0, save\/restore_slice_1\/tensor_name, save\/restore_slice_1\/shape_and_slice)]]\nCaused by op u'save\/restore_slice_1', defined at:\n  File \"tensor_restore.py\", line 8, in &lt;module&gt;\n    saver = tf.train.Saver()\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 713, in __init__\n    restore_sequentially=restore_sequentially)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 432, in build\n    filename_tensor, vars_to_save, restore_sequentially, reshape)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 191, in _AddRestoreOps\n    values = self.restore_op(filename_tensor, vs, preferred_shard)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 106, in restore_op\n    preferred_shard=preferred_shard)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/io_ops.py\", line 189, in _restore_slice\n    preferred_shard, name=name)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_io_ops.py\", line 271, in _restore_slice\n    preferred_shard=preferred_shard, name=name)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/op_def_library.py\", line 664, in apply_op\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1834, in create_op\n    original_op=self._default_original_op, op_def=op_def)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1043, in __init__\n    self._traceback = _extract_stack()\n<\/code><\/pre>\n\n<p>So perhaps the original code (see the external link above) could be modified to something like this:<\/p>\n\n<pre><code>def weight_variable(shape):\n  initial = tf.truncated_normal(shape, stddev=0.1)\n  weight_var = tf.Variable(initial, name=\"weight_var\")\n  return weight_var\n\ndef bias_variable(shape):\n  initial = tf.constant(0.1, shape=shape)\n  bias_var = tf.Variable(initial, name=\"bias_var\")\n  return bias_var\n<\/code><\/pre>\n\n<p>But then the question I have: is restoring weight_var and bias_var variables sufficient to implement the prediction? I did the training on the powerful machine with GPU and I would like to copy the model to the less powerful computer without GPU to run predictions.<\/p>\n"},{"AnswerId":"48089875","CreationDate":"2018-01-04T06:33:48.190","ParentId":null,"OwnerUserId":"9147492","Title":null,"Body":"<p>This issue should be caused by the name scope variants when double creating the same network.<\/p>\n\n<p>put the command: <\/p>\n\n<p>tf.reset_default_graph()<\/p>\n\n<p>before creating the network<\/p>\n"},{"AnswerId":"42972981","CreationDate":"2017-03-23T10:11:18.493","ParentId":null,"OwnerUserId":"6400639","Title":null,"Body":"<p>make sure the declaration of tf.train.Saver() is in with tf.Session() as sess<\/p>\n"},{"AnswerId":"51474426","CreationDate":"2018-07-23T08:22:38.667","ParentId":null,"OwnerUserId":"9354751","Title":null,"Body":"<p>If a problem like this occurs then try to restart your kernel as the current variable overwrites the previous causing conflict between them, thus it shows notFoundError and other issues come up.<\/p>\n\n<p>I encountered the same type of problem and restarting the kernel worked for me.\n(Caution: Try avoiding running your kernel multiple times as it can ruin your model file recreating variables that overwrite the existing one thus end up changing the original values.)<\/p>\n"},{"AnswerId":"34982715","CreationDate":"2016-01-24T23:06:37.593","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>There's a similar question here: <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-restore-a-previously-saved-model-python\/\">Tensorflow: how to save\/restore a model?<\/a>\nTLDR; you need to recreate model structure using same sequence of TensorFlow API commands before using Saver object to restore the weights<\/p>\n\n<p>This is suboptimal, follow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/696\" rel=\"nofollow noreferrer\">Github issue #696<\/a> for progress on making this easier<\/p>\n"}]}
{"QuestionId":34983207,"AnswerCount":1,"Tags":"<python><initialization><tensorflow>","CreationDate":"2016-01-25T00:08:51.840","AcceptedAnswerId":34983516.0,"OwnerUserId":1815451.0,"Title":"Initializing variable with another variable using tensorflow","Body":"<p>I am building neural network using denoising stacked autoencoders. I train autoencoder and then I would like to take the matrix of weights W and copy\/initialize\/clone it's values into new variable which is used in supervised optimization. How can I do such thing? <\/p>\n\n<p>.initialized_value() doesn't work for me :\/ <\/p>\n","answers":[{"AnswerId":"34983516","CreationDate":"2016-01-25T00:48:01.367","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Use <code>var.assign<\/code>, ie<\/p>\n\n<pre><code>vara = tf.Variable(0)\nvarb = tf.Variable(0)\ninit_op = tf.initialize_all_variables()\nsess = tf.InteractiveSession()\nsess.run([init_op])\nsess.run([vara.assign_add(1)])\nprint 'variable a', vara.eval()\nprint 'variable b', varb.eval()\nsess.run([varb.assign(vara)])\nprint 'variable b', varb.eval()\n<\/code><\/pre>\n\n<p>You should see<\/p>\n\n<pre><code>variable a 1\nvariable b 0\nvariable b 1\n<\/code><\/pre>\n"}]}
{"QuestionId":34983983,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2016-01-25T01:56:32.960","AcceptedAnswerId":null,"OwnerUserId":3487835.0,"Title":"The right way to define a function in theano?","Body":"<h1>Background:<\/h1>\n\n<p>Usually I will define a theano function with input like 'x = fmatrix()', however, during modifying keras (a deep learning library based on theano) to make it work with CTC cost, I noticed a very weird problem: if one input of the cost function is declared as<\/p>\n\n<pre><\/pre>\n\n<p>instead of <\/p>\n\n<pre><\/pre>\n\n<p>the training process will converge much faster.<\/p>\n\n<h1>A simplified problem:<\/h1>\n\n<p>The whole codes above are quite big. So I try to simplify the problem like the following: say a function for computing Levenshtein edit distance as<\/p>\n\n<pre><\/pre>\n\n<p>then I define two functions f1 and f2 like:<\/p>\n\n<pre><\/pre>\n\n<p>When computing with f1 and f2, the results are different:<\/p>\n\n<pre><\/pre>\n\n<p>f1 gives the right result, but f2 doen't.<\/p>\n\n<p>So my problem is: what is the right way to define a theano function? And, what actually went wrong about f2?<\/p>\n\n<h1>Update:<\/h1>\n\n<p>I'm using theano of version 0.8.0.dev0. I just tried theano 0.7.0, both f1 and f2 give correct result. Maybe this is a bug of theano?<\/p>\n\n<h1>Update_1st 1-27-2016:<\/h1>\n\n<p>According to the explanation of @lamblin on this issue (<a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/3925#issuecomment-175088918\" rel=\"nofollow\">https:\/\/github.com\/Theano\/Theano\/issues\/3925#issuecomment-175088918<\/a>), this was actually a bug of theano, and has been fixed in the latest (1-26-2016) version. For convenience, lamblin's explanation is quoted here:<\/p>\n\n<blockquote>\n  <p>The first way is the most natural one, but in theory both should be equivalent.\n  x3 and x4 are created as the output of an \"alloc\" operation, the input of which would be the constant 3, rather than free inputs like x1 and x2, but that should not matter since you pass [x3, x4] as inputs to theano.function, which should cut the computation graph right there.<\/p>\n  \n  <p>My guess is that scan is optimizing prematurely, believing that x3 or x4 is guaranteed to always be the constant 0, and does some simplifications that proved incorrect when values are provided for them. That would be an actual bug in scan.\"<\/p>\n<\/blockquote>\n\n<h1>Update_2nd 1-27-2016:<\/h1>\n\n<p>Unfortunately the bug is not totally fixed yet. In the background section I mentioned if one input of the cost function is declared as tensor.zeros() the convergence process will be much faster, I've found the reason: when input declared as tensor.zeros(), the cost function gave incorrect result, though mysteriously this helped the convergence. \nI managed a simplified problem reproduction demo here (<a href=\"https:\/\/github.com\/daweileng\/TheanoDebug\" rel=\"nofollow\">https:\/\/github.com\/daweileng\/TheanoDebug<\/a>), run the ctc_bench.py and you can see the results.<\/p>\n","answers":[{"AnswerId":"34988307","CreationDate":"2016-01-25T08:45:15.520","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p><code>theano.tensor.zeros(...)<\/code> can't take any other value than 0.<\/p>\n\n<p>Unless you add nodes to the graph of course and modify parts of the zeros tensor using <code>theano.tensor.set_subtensor<\/code>.<\/p>\n\n<p>The input tensor <code>theano.tensor.fmatrix<\/code> can take any value you input.<\/p>\n"}]}
{"QuestionId":34984065,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-25T02:07:41.013","AcceptedAnswerId":null,"OwnerUserId":5834943.0,"Title":"How to Fine-tuning a Pretrained Network in Tensorflow?","Body":"<p>Can anyone give an example of how to fine tune a pretrained imagenet network with new data and different classes similar to this:<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/03-fine-tuning.ipynb\" rel=\"noreferrer\">Fine-tuning a Pretrained Network for Style Recognition<\/a><\/p>\n","answers":[{"AnswerId":"50111180","CreationDate":"2018-05-01T04:00:54.017","ParentId":null,"OwnerUserId":"2130551","Title":null,"Body":"<p>This <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">TensorFlow tutorial<\/a> describes how to retrain a image classifier for new data and new classes.<\/p>\n"}]}
{"QuestionId":34985751,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-01-25T05:33:56.883","AcceptedAnswerId":null,"OwnerUserId":1176802.0,"Title":"tensorflow import images code does not work correctly","Body":"<p>Using <a href=\"https:\/\/stackoverflow.com\/questions\/33648322\/tensorflow-image-reading-display\/34985237#34985237\">TensorFlow image reading &amp; display<\/a> <\/p>\n\n<p>I ran the above code which works correctly on the first run.<br>\nI have only one jpg file &lt; 1MB.<br>\nSubsequent runs just hang; any idea or input?<\/p>\n","answers":[]}
{"QuestionId":34986911,"AnswerCount":2,"Tags":"<deep-learning><tensorflow>","CreationDate":"2016-01-25T07:09:20.767","AcceptedAnswerId":null,"OwnerUserId":5782844.0,"Title":"How to set weight cost strength in TensorFlow?","Body":"<p>I'm trying to use TensorFlow with my deep learning project.  <\/p>\n\n<p>When I use Momentum Gradient Descent, how is the weight cost strength set?<br>\n(The \u03bb in this <a href=\"http:\/\/i.stack.imgur.com\/305RE.jpg\" rel=\"noreferrer\">formula<\/a>.) <\/p>\n","answers":[{"AnswerId":"50780889","CreationDate":"2018-06-10T05:40:49.393","ParentId":null,"OwnerUserId":"2366315","Title":null,"Body":"<p>Note that the formula you show does not actually present true \"weight decay\", but instead L2-regularization. Many people mix these up, including well-known professors. Let me explain.<\/p>\n\n<p>When using pure SGD (without momentum) as an optimizer, weight decay is the same thing as adding a L2-regularization term to the loss. <strong>When using any other optimizer, including Momentum, this is not true.<\/strong><\/p>\n\n<p>Weight decay (don't know how to TeX here, so excuse my pseudo-notation):<\/p>\n\n<pre><code>w[t+1] = w[t] - learning_rate * dw - weight_decay * w\n<\/code><\/pre>\n\n<p>L2-regularization:<\/p>\n\n<pre><code>loss = actual_loss + lambda * 1\/2 sum(||w||_2 for w in network_params)\n<\/code><\/pre>\n\n<p>Computing the gradient of the extra term in L2-regularization gives <code>lambda * w<\/code> and thus inserting it into the SGD update equation<\/p>\n\n<pre><code>dloss_dw = dactual_loss_dw + lambda * w\nw[t+1] = w[t] - learning_rate * dw\n<\/code><\/pre>\n\n<p>gives the same as weight decay, but mixes <code>lambda<\/code> with the <code>learning_rate<\/code>. Any other optimizer, even SGD with momentum, gives a different update rule for weight decay as for L2-regularization! See the paper <a href=\"\/https:\/\/arxiv.org\/abs\/1711.05101\">Fixing weight decay in Adam<\/a> for more details. (Edit: AFAIK, <a href=\"http:\/\/www.cs.toronto.edu\/~hinton\/absps\/parle.pdf\" rel=\"nofollow noreferrer\">this 1987 Hinton paper<\/a> introduced \"weight decay\", literally as \"each time the weights are updated, their magnitude is also decremented by 0.4%\" at page 10)<\/p>\n\n<p>That being said, there doesn't seem to be support for \"proper\" weight decay in TensorFlow yet. There are a few issues discussing it, specifically because of above paper.<\/p>\n\n<p>One possible way to implement it is by writing an op that does the decay step manually after every optimizer step. A different way, which is what I'm currently doing, is using an additional SGD optimizer just for the weight decay, and \"attaching\" it to your <code>train_op<\/code>. Both of these are just crude work-arounds, though. My current code:<\/p>\n\n<pre><code># In the network definition:\nwith arg_scope([layers.conv2d, layers.dense],\n               weights_regularizer=layers.l2_regularizer(weight_decay)):\n    # define the network.\n\nloss = # compute the actual loss of your problem.\ntrain_op = optimizer.minimize(loss, global_step=global_step)\nif args.weight_decay not in (None, 0):\n    with tf.control_dependencies([train_op]):\n        sgd = tf.train.GradientDescentOptimizer(learning_rate=1.0)\n        train_op = sgd.minimize(tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)))\n<\/code><\/pre>\n\n<p>This somewhat makes use of TensorFlow's provided bookkeeping. Note that the <code>arg_scope<\/code> takes care of appending an L2-regularization term for every layer to the <code>REGULARIZATION_LOSSES<\/code> graph-key, which I then all sum up and optimize using SGD which, as shown above, corresponds to actual weight-decay.<\/p>\n\n<p>Hope that helps, and if anyone gets a nicer code snippet for this, or TensorFlow implements it better (i.e. in the optimizers), please share.<\/p>\n\n<p><strong>Edit:<\/strong> see also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/17438\" rel=\"nofollow noreferrer\">this PR<\/a> which just got merged into TF.<\/p>\n"},{"AnswerId":"34996719","CreationDate":"2016-01-25T15:52:20.000","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The term for the weight cost\/decay is not part of the optimizers in TensorFlow.<\/p>\n\n<p>It is easy to include, however, by adding the extra penalty to the cost function with the L2 loss on the weights:<\/p>\n\n<pre><code>C = &lt;your initial cost function&gt;\nl2_loss = tf.add_n([tf.nn.l2_loss(v) for v in tf.trainable_variables()])\nC = C + lambda * l2_loss\n<\/code><\/pre>\n\n<p><code>tf.nn.l2_loss(v)<\/code> <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/nn.html#l2_loss\">link<\/a> is simply <code>0.5 * tf.reduce_sum(v * v)<\/code> and the gradients for individual weights will be equal to <code>lambda * w<\/code>, which should be equivalent to your linked equation.<\/p>\n"}]}
{"QuestionId":34987509,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><max><tensor>","CreationDate":"2016-01-25T07:49:27.570","AcceptedAnswerId":34988069.0,"OwnerUserId":5813490.0,"Title":"TensorFlow: Max of a tensor along an axis","Body":"<p>My question is in two connected parts:<\/p>\n\n<ol>\n<li><p>How do I calculate the max along a certain axis of a tensor? For example, if I have<\/p>\n\n<pre><\/pre>\n\n<p>I want something like <\/p>\n\n<pre><\/pre>\n\n<p>I know there is a  and a , but neither give the maximum value along an axis of a single tensor. For now I have a workaround:<\/p>\n\n<pre><\/pre>\n\n<p>But it looks less than optimal. Is there a better way to do this?<\/p><\/li>\n<li><p>Given the indices of an  of a tensor, how do I index into another tensor using those indices? Using the example of  above, how do I do something like the following:<\/p>\n\n<pre><\/pre>\n\n<p>I know slicing, like the last line, does not exist in TensorFlow yet (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"noreferrer\">#206<\/a>). <\/p>\n\n<p>My question is: <em>what is the best workaround for my specific case (maybe using other methods like gather, select, etc.)?<\/em><\/p>\n\n<p>Additional information: I know  and  are going to be two dimensional tensors only!<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"52784988","CreationDate":"2018-10-12T18:10:06.427","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>As of <strong>TensorFlow 1.10.0<\/strong>-<em>dev20180626<\/em>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_max\" rel=\"nofollow noreferrer\"><strong><code>tf.reduce_max<\/code><\/strong><\/a> accepts <code>axis<\/code> and <code>keepdims<\/code> keyword arguments offering the similar functionality of <code>numpy.max<\/code>.<\/p>\n\n<pre><code>In [55]: x = tf.constant([[1,220,55],[4,3,-1]])\n\nIn [56]: tf.reduce_max(x, axis=1).eval() \nOut[56]: array([220,   4], dtype=int32)\n<\/code><\/pre>\n\n<p>To have a resultant tensor of the same dimension as the input tensor, use <code>keepdims=True<\/code><\/p>\n\n<pre><code>In [57]: tf.reduce_max(x, axis=1, keepdims=True).eval()Out[57]: \narray([[220],\n       [  4]], dtype=int32)\n<\/code><\/pre>\n\n<hr>\n\n<p>If the <code>axis<\/code> argument is not explicitly specified then the tensor level maximum element is returned (i.e. all axes are reduced).<\/p>\n\n<pre><code>In [58]: tf.reduce_max(x).eval()\nOut[58]: 220\n<\/code><\/pre>\n"},{"AnswerId":"34988069","CreationDate":"2016-01-25T08:30:27.960","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#reduce_max\"><code>tf.reduce_max()<\/code><\/a> operator provides exactly this functionality. By default it computes the global maximum of the given tensor, but you can specify a list of <code>reduction_indices<\/code>, which has the same meaning as <code>axis<\/code> in NumPy. To complete your example:<\/p>\n\n<pre><code>x = tf.constant([[1, 220, 55], [4, 3, -1]])\nx_max = tf.reduce_max(x, reduction_indices=[1])\nprint sess.run(x_max)  # ==&gt; \"array([220,   4], dtype=int32)\"\n<\/code><\/pre>\n\n<p>If you compute the argmax using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#argmax\"><code>tf.argmax()<\/code><\/a>, you could obtain the the values from a different tensor <code>y<\/code> by flattening <code>y<\/code> using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#reshape\"><code>tf.reshape()<\/code><\/a>, converting the argmax indices into vector indices as follows, and using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#gather\"><code>tf.gather()<\/code><\/a> to extract the appropriate values:<\/p>\n\n<pre><code>ind_max = tf.argmax(x, dimension=1)\ny = tf.constant([[1, 2, 3], [6, 5, 4]])\n\nflat_y = tf.reshape(y, [-1])  # Reshape to a vector.\n\n# N.B. Handles 2-D case only.\nflat_ind_max = ind_max + tf.cast(tf.range(tf.shape(y)[0]) * tf.shape(y)[1], tf.int64)\n\ny_ = tf.gather(flat_y, flat_ind_max)\n\nprint sess.run(y_) # ==&gt; \"array([2, 6], dtype=int32)\"\n<\/code><\/pre>\n"}]}
{"QuestionId":34989070,"AnswerCount":0,"Tags":"<theano><keras><cudnn>","CreationDate":"2016-01-25T09:29:06.427","AcceptedAnswerId":null,"OwnerUserId":2687955.0,"Title":"Theano running slowly","Body":"<p>I implemented a 9 layer convolutional neural network in Keras (with Theano back end) and translated pre-trained weights from another network. But while it only takes less than 1s in the original network which is implemented in C, Theano takes like 10 seconds for a single image. <\/p>\n\n<p>I am running on a 3 GB Quadro K4000 and has installed  and . Did I do something wrong?<\/p>\n","answers":[]}
{"QuestionId":34995530,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-25T14:54:47.167","AcceptedAnswerId":34996139.0,"OwnerUserId":2084314.0,"Title":"How to process gradients with a Dimension size of None","Body":"<p>Using AdamOptimizer, when I get the gradients of a 2d variable, the second dimension's size ends up being None, while the first dimension is the same size as the variable's first dimension.  This makes it difficult to process the gradients, since a size of None isn't compatible with other sizes for most functions.  When I get the gradients of a 1d variable, the gradient's dimension size is the same as the variable's.  I haven't tried variables with more than 2 dimensions.<\/p>\n\n<p>Is this a bug?  Is there a way to specify what the size of the gradient should be through the compute_gradients function?  Is there a way to process the gradient that gets around the size None issue?<\/p>\n","answers":[{"AnswerId":"34996139","CreationDate":"2016-01-25T15:24:31.073","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> It shouldn't matter, and  you can process the gradients using the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#AdamOptimizer\" rel=\"nofollow\"><code>tf.train.AdamOptimizer<\/code><\/a> as normal. If you are seeing shape-related errors, this most likely arises from one of the known dimensions not matching.<\/p>\n\n<p>The presence of <code>None<\/code> in a gradient tensor's <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#TensorShape\" rel=\"nofollow\">shape<\/a> simply means that the size in that dimension could not be statically inferred. This is not necessarily a bug: the shapes of many operators depend on their data inputs, and the TensorFlow Python front-end uses a simple heuristic (i.e., only compute a limited set of ops with constant inputs) to decide what data inputs to evaluate. Almost all of the TensorFlow ops&mdash;excluding some image processing ops&mdash;will work on inputs whose shape is unknown (or only partially known), and perform checks at runtime instead.<\/p>\n\n<p>The main way to process gradients is using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#Optimizer.compute_gradients\" rel=\"nofollow\"><code>Optimizer.apply_gradients()<\/code><\/a>, which defers shape checking to the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a27d844e05447e65aa279ae5269a2d75590f46f6\/tensorflow\/python\/training\/training_ops.py#L61\" rel=\"nofollow\">shape function for the <code>ApplyAdam<\/code> operator<\/a>. This shape function asserts that the variable and gradient have the same shape, but the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#TensorShape.merge_with\" rel=\"nofollow\"><code>TensorShape.merge_with()<\/code><\/a> method allows false positives in the presence of <code>None<\/code> in either of the shapes.<\/p>\n\n<p>Finally, if you need to process the gradients at graph construction time, and your processing somehow depends on the gradients having known shapes, you can always use the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Tensor.set_shape\" rel=\"nofollow\"><code>Tensor.set_shape()<\/code><\/a> method to copy the shape of the variable to the shape of the gradient, as these must be equivalent:<\/p>\n\n<pre><code>var = tf.Variable(...)\nloss = ...\ngrad = tf.gradients(loss, [var])[0]\n\n# `grad` and `var` must have the same shape.\ngrad.set_shape(var.get_shape())\n<\/code><\/pre>\n"}]}
{"QuestionId":34996075,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-01-25T15:20:53.380","AcceptedAnswerId":34996628.0,"OwnerUserId":3130081.0,"Title":"caffe data layer example step by step","Body":"<p>I want to find a caffe python data layer example to learn.\nI know that Fast-RCNN has a python data layer, but it's rather complicated since I\nam not familiar with object detection.<br>\nSo my question is, is there a python data layer example where I can learn how to define my own data preparation procedure?<br>\nFor example, how to do define a python data layer do much more data augmentation\n(such as translation, rotation etc.) than caffe .<\/p>\n\n<p>Thank you very much<\/p>\n","answers":[{"AnswerId":"34996628","CreationDate":"2016-01-25T15:47:22.967","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can use a <code>\"Python\"<\/code> layer: a layer implemented in python to feed data into your net. (See an example for adding a <code>type: \"Python\"<\/code> layer <a href=\"https:\/\/stackoverflow.com\/a\/34922529\/1714410\">here<\/a>).<\/p>\n\n<pre><code>import sys, os\nsys.path.insert(0, os.environ['CAFFE_ROOT']+'\/python')\nimport caffe\nclass myInputLayer(caffe.Layer):\n  def setup(self,bottom,top):\n    # read parameters from `self.param_str`\n    ...\n  def reshape(self,bottom,top):\n    # no \"bottom\"s for input layer\n    if len(bottom)&gt;0:\n      raise Exception('cannot have bottoms for input layer')\n    # make sure you have the right number of \"top\"s\n    if len(top)!= ...\n       raise ...\n    top[0].reshape( ... ) # reshape the outputs to the proper sizes\n\n  def forward(self,bottom,top): \n    # do your magic here... feed **one** batch to `top`\n    top[0].data[...] = one_batch_of_data\n\n\n  def backward(self, top, propagate_down, bottom):\n    # no back-prop for input layers\n    pass\n<\/code><\/pre>\n\n<p>For more information on <code>param_str<\/code> see <a href=\"https:\/\/stackoverflow.com\/a\/34588801\/1714410\">this thread<\/a>.<br>\nYou can find a sketch of a data loading layer with pre-fetch <a href=\"https:\/\/stackoverflow.com\/a\/48065550\/1714410\">here<\/a>.<\/p>\n"},{"AnswerId":"35017732","CreationDate":"2016-01-26T15:38:32.037","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>@Shai's answer is great. At the same time, I find another detailed example about python data layer in one PR of caffe-master. <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3471\/files\" rel=\"noreferrer\">https:\/\/github.com\/BVLC\/caffe\/pull\/3471\/files<\/a>\nI hope this detailed example be helpful for anyone else.<\/p>\n"}]}
{"QuestionId":34997810,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2016-01-25T16:46:38.587","AcceptedAnswerId":null,"OwnerUserId":3610694.0,"Title":"Caffe without GPU: make runtest error","Body":"<p>I'm trying to compile caffe without GPU support. I successfully ran \"make all\" and \"make test\", but when I try to run \"make runtest\", I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on how to fix?<\/p>\n","answers":[{"AnswerId":"35003622","CreationDate":"2016-01-25T22:27:07.323","ParentId":null,"OwnerUserId":"5839341","Title":null,"Body":"<p>This is because caffe library is not linked by default after make all, make test. I manually did the following in terminal:<\/p>\n\n<pre><code>export DYLD_LIBRARY_PATH=\/YourPathToCaffeMasterFolder\/.build_release\/lib:$DYLD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>It fixed the problem for me, for now. Caffe's installation is so much pain.<\/p>\n"},{"AnswerId":"35216464","CreationDate":"2016-02-05T04:48:26.907","ParentId":null,"OwnerUserId":"3610694","Title":null,"Body":"<p>Copying the lib files from $CAFFE_ROOT\/.build_release\/lib to \/usr\/local\/lib fixed the problem for me!<\/p>\n"}]}
{"QuestionId":35001027,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-25T19:47:40.880","AcceptedAnswerId":35008834.0,"OwnerUserId":2615676.0,"Title":"Does batching queue (tf.train.batch) not preserve order?","Body":"<p>I've set up a filename-producing queue using  with the  option set to , coupled to a batching queue using  (i.e. non-shuffling). Looking at the list of examples being read, while the ordering is <em>almost<\/em> perfectly preserved, it is not strictly so. For example the first few sample are 4, 2, 1, 3, 5, 6, 7, 8, 9, 11, 10, ..., where the number corresponds to the position of the sample within the first file read. After that the ordering is almost prefect for several hundred samples, but it occasionally switches adjacent samples. Is this expected behavior? Is there some way to enforce that the ordering is preserved, so that one does not have to keep track of what file got read when, etc?<\/p>\n\n<p>I should say that I conditionally discard some samples by enqueuing either 0 or 1 sample at a time, and setting  to  in the batching queue. None of the samples above are being skipped however and so this shouldn't in principle be an issue.<\/p>\n","answers":[{"AnswerId":"35008834","CreationDate":"2016-01-26T07:34:24.257","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>As Yaroslav has mentioned in the comments, a single thread would do the trick. In addition to a single thread, you should set num_epochs = 1. If you don't, it will keep producing batches and it may seem like order is not preserved as it loops from the start again. I hope this works.<\/p>\n\n<p>Having said that though, I hope someone can come up with a better answer to solving this!<\/p>\n"}]}
{"QuestionId":35003741,"AnswerCount":0,"Tags":"<c++><macos><gcc><tensorflow><binutils>","CreationDate":"2016-01-25T22:34:39.170","AcceptedAnswerId":null,"OwnerUserId":1525945.0,"Title":"TensorFlow compilation error: Expecting string instruction after 'rep'","Body":"<p>I'm trying to compile <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow\">TensorFlow<\/a> from source on OSX 10.10, following these <a href=\"https:\/\/gist.github.com\/Mistobaan\/dd32287eeb6859c6668d\" rel=\"nofollow\">instructions<\/a> so that I can use the GPU. I'm using GCC 4.8.5.<\/p>\n\n<p>This is the error that I get during compilation:\n<\/p>\n\n<p>From what I've read, this seems to be a bug in an older version of binutils (at least on Linux), so I installed binutils 2.25.1 with Homebrew, but it didn't help. I've also tried including the compiler flag .<\/p>\n\n<p>I can compile without errors with GCC 5.1.0, but I can't use this compiler as CUDA does not support it.<\/p>\n","answers":[]}
{"QuestionId":35004565,"AnswerCount":0,"Tags":"<python><c++><theano><pymc3>","CreationDate":"2016-01-25T23:40:50.203","AcceptedAnswerId":null,"OwnerUserId":3413239.0,"Title":"PyMC3: Theano's cxxflags not properly configured","Body":"<p>I have been running find_MAP() from pymc3. It has been giving me same error though out. <\/p>\n\n<pre><\/pre>\n\n<p>This is python's exception:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried various steps as mentioned here : <a href=\"https:\/\/stackoverflow.com\/a\/10894096\/3413239\">https:\/\/stackoverflow.com\/a\/10894096\/3413239<\/a> by trying out various cxxflags but nothing worked.<\/p>\n\n<p>Since I am not familiar with Theano and C++ I am unable to understand what exactly is the problem. Any help will be highly appreciated.<\/p>\n\n<p>what does  mean? What instructions are these? I have lot of directories in  but not files which look like : . What is the purpose of these files?<\/p>\n\n<p>I installed PyMC3 and Theano was installed along with it. Can there be a problem in installation of Theano via PyMC3?<\/p>\n\n<h2>EDIT<\/h2>\n\n<p>There was a problem with numpy and scipy configurations. What I have done so far:<\/p>\n\n<ul>\n<li>Uninstalled numpy and scipy<\/li>\n<li>Rebuilt numpy with proper configuration and links to OpenBlas. <a href=\"https:\/\/stackoverflow.com\/questions\/11443302\/compiling-numpy-with-openblas-integration\/14391693#14391693\">Compiling numpy with OpenBLAS integration<\/a><\/li>\n<li>Reinstalled scipy, theano, pymc3 via pip\nIts still giving me same problems.<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":35004619,"AnswerCount":1,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2016-01-25T23:46:13.310","AcceptedAnswerId":35004791.0,"OwnerUserId":5825952.0,"Title":"reading a large dataset in tensorflow","Body":"<p>I am not quite sure about how file-queue works. I am trying to use a large dataset like imagenet as input. So preloading data is not the case, so I am wondering how to use the file-queue. According to the tutorial, we can convert data to TFRecords file as input. Now we have a single big TFRecords file. So when we specify a FIFO queue for the reader, does it mean the program would fetch a batch of data each time and feed the graph instead of loading the whole file of data?<\/p>\n","answers":[{"AnswerId":"35004791","CreationDate":"2016-01-26T00:03:42.047","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>The amount of pre-fetching depends on your queue capacity. If you use <code>string_input_producer<\/code> for your filenames and <code>batch<\/code> for batching, you will have 2 queues - filename queue, and prefetching queue created by batch. Queue created by <code>batch<\/code> has default capacity of <code>32<\/code>, controlled by <code>batch(...,capacity=)<\/code> argument, therefore it can prefetch up to <code>32<\/code> images. If you follow outline in TensorFlow official howto's, processing examples (everything after <code>batch<\/code>) will happen in main Python thread, whereas filling up the queue will happen in threads created\/started by <code>batch\/start_queue_runners<\/code>, so prefetching new data and running prefetched data through the network will occur concurrently, blocking when the queue gets full or empty.<\/p>\n"}]}
{"QuestionId":35005691,"AnswerCount":2,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2016-01-26T01:39:45.650","AcceptedAnswerId":null,"OwnerUserId":5825952.0,"Title":"How to use image_summary to view images from different batches in Tensorflow?","Body":"<p>I am curious about how image_summary works. There is a parameter called max_images, which controls how many images would be shown. However it seems the summary only displays images from one batch. If we use bigger value of max_iamges, we will just view more images from the batch. Is there a way I can view for example one image from each batch?<\/p>\n","answers":[{"AnswerId":"35005934","CreationDate":"2016-01-26T02:06:34.980","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To view one image from each batch, you need to fetch the result of the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#image_summary\"><code>tf.image_summary()<\/code><\/a> op every time you run a step. For example, it you have the following setup:<\/p>\n\n<pre><code>images = ...\nloss = ...\noptimizer = ...\n\ntrain_op = optimizer.minimize(loss)\ninit_op = tf.initialize_all_variables()\nimage_summary_t = tf.image_summary(images.name, images, max_images=1)\n\nsess = tf.Session()\nsummary_writer = tf.train.SummaryWriter(...)\nsess.run(init_op)\n<\/code><\/pre>\n\n<p>...you could set up your training loop to capture one image per iteration as follows:<\/p>\n\n<pre><code>for _ in range(10000):\n    _, image_summary = sess.run([train_op, image_summary_t])\n    summary_writer.add_summary(image_summary)\n<\/code><\/pre>\n\n<p>Note that capturing summaries on each batch might be inefficient, and you should probably only capture the summary periodically for faster training.<\/p>\n\n<p><strong>EDIT:<\/strong> The above code writes a separate summary for each image, so your log will contain all of the images, but they will not all be visualized in TensorBoard. If you want to combine your summaries to visualize images from multiple batches, you could do the following:<\/p>\n\n<pre><code>combined_summary = tf.Summary()\nfor i in range(10000):\n    _, image_summary = sess.run([train_op, image_summary_t])\n    combined_summary.MergeFromString(image_summary)\n    if i % 10 == 0:\n        summary_writer.add_summary(combined_summary)\n        combined_summary = tf.Summary()\n<\/code><\/pre>\n"},{"AnswerId":"36635449","CreationDate":"2016-04-14T22:57:22.877","ParentId":null,"OwnerUserId":"262271","Title":null,"Body":"<p>I was able to solve this by creating a new <code>image_summary<\/code> op for each batch. i.e. I went from something that looked like:<\/p>\n\n<pre><code>train_writer = tf.train.SummaryWriter('summary_dir')\nimg = tf.image_summary(\"fooImage\", img_data)\nfor i in range(N_BATCHES):\n  summary, _ = sess.run([img, train_step])\n  train_writer.add_summary(summary, i)\n<\/code><\/pre>\n\n<p>(Which, frustratingly, was not doing what I expected.) To...<\/p>\n\n<pre><code>train_writer = tf.train.SummaryWriter('summary_dir')\nfor i in range(N_BATCHES):\n  # Images are sorted in lexicographic order, so zero-pad the name\n  img = tf.image_summary(\"fooImage{:06d}\".format(i), img_data)\n  summary, _ = sess.run([img, train_step])\n  train_writer.add_summary(summary)\n<\/code><\/pre>\n"}]}
{"QuestionId":35007207,"AnswerCount":1,"Tags":"<python><python-3.x><virtualenv><anaconda><tensorflow>","CreationDate":"2016-01-26T04:47:11.423","AcceptedAnswerId":35089013.0,"OwnerUserId":3781180.0,"Title":"Installing TensorFlow in python 3.5 Anaconda distro conda env","Body":"<p><strong><em>What I did to try and solve this issue:<\/em><\/strong><\/p>\n\n<hr>\n\n<ol>\n<li>Using Mac OS X<\/li>\n<li>Using Anaconda distro currently in my computer as the path <\/li>\n<li>Used conda instead of virtualenv because of my distribution of python. I referred to the guide here and tried to download the same with conda's package manager: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#download-and-setup\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#download-and-setup<\/a><\/li>\n<li>Tensorflow worked correctly when I created an anaconda env using the command  to state I'm using 2.7 python, then used this command  gathered from the Anaconda Cloud to attempt to download it. It worked when I said<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><strong><em>However, it doesn't really fix the issue:<\/em><\/strong><\/p>\n\n<hr>\n\n<p>However, I want to use the 3.5 distribution of python, simply to keep up with the times and not use outdated pythons. <\/p>\n\n<ol start=\"5\">\n<li>I attempted to download it using the same way in python=3.5, but it raised this error when I tried to run the command  :<\/li>\n<\/ol>\n\n<pre class=\"lang-sh prettyprint-override\"><\/pre>\n\n<p>And I ran the  to see what was going on and I got:<\/p>\n\n<pre class=\"lang-html prettyprint-override\"><\/pre>\n\n<p>This is really frustrating me and I'm not sure what to do. If there's no work around for this, I think I'll have to use Anaconda's python 2.7 distribution for TensorFlow experiments. If anyone has any idea how to solve this compatibility issue(or some other kind of issue), I'd be extremely grateful. Thanks.<\/p>\n\n<p>EDIT: I'm pretty sure TensorFlow supports 3.5 because in their documentations they say they support 2.7 python and 3.3+. If there were no clues as to whether they support 3.5, I would have already given up and used 2.7 by now.<\/p>\n","answers":[{"AnswerId":"35089013","CreationDate":"2016-01-29T16:11:55.123","ParentId":null,"OwnerUserId":"5858079","Title":null,"Body":"<p>The version of TensorFlow packaged <a href=\"https:\/\/conda.anaconda.org\/jjhelmus\" rel=\"nofollow\">here<\/a> is version 0.5.<\/p>\n\n<p>Python 3.x support was introduced in TensorFlow 0.6, so you need to figure out how to install the newest version into Anaconda.<\/p>\n\n<p>Usually you can install packages into anaconda using <code>pip<\/code>, but I haven't succeeded with TensorFlow.<\/p>\n\n<p>EDIT: I just noticed that the documentation hasn't updated the url to the pip-wheel.<\/p>\n\n<p>To install tensorflow in python 3.5 via pip, use the following command:<\/p>\n\n<pre><code>$ sudo easy_install --upgrade six\n$ sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.5.0-py2-none-any.whl\" rel=\"nofollow\">This links<\/a> to <em>...tensorflow-0.<strong>5<\/strong>.0-py<strong>2<\/strong>-none-any.whl<\/em>, which is an older version of tensorflow for python 2.x.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/get_started\/os_setup.html#pip_install\" rel=\"nofollow\">Tensorflow get started page link<\/a>.<\/p>\n"}]}
{"QuestionId":35007634,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-26T05:39:17.320","AcceptedAnswerId":null,"OwnerUserId":2084314.0,"Title":"Rebuilding the Graph while keeping the variables","Body":"<p>I have variables that change size over the course of the program, and variables that should get added into the middle of the Graph, but when I run the program, those variables remain unchanged.<\/p>\n\n<p>I found <a href=\"https:\/\/stackoverflow.com\/questions\/34235225\/is-it-possible-to-modify-an-existing-tensorflow-computation-graph\">this question<\/a> which tells me it's not possible to change variables after the Graph has been created, and suggests a method of creating a session with a new graph.  When I tried that, I got issues with having uninitialized Variables.  I tried saving the variables I needed and then restoring them in the new Session, but it still claimed they were uninitialized.<\/p>\n\n<p>How can I add Variables to a Session while keeping their previous values?  Or, better if possible, how can I change the size of Variables in a Graph, or force the Graph to restructure without reinitializing the Variables?<\/p>\n","answers":[]}
{"QuestionId":35012392,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><theano><deep-learning>","CreationDate":"2016-01-26T11:06:12.743","AcceptedAnswerId":null,"OwnerUserId":5841128.0,"Title":"What should I do for using theano scan in deep Learning","Body":"<p>I want to use then scan for a tensor3 calculation in my DNN, we can say this tensor3 is a 3D matrix, I want each layer of this matrix go through the  calculation, the , and  should be the fixed, so I could do the update for later.  The following is my coding:<\/p>\n\n<pre><\/pre>\n\n<p>but the results show this error:<\/p>\n\n<pre><\/pre>\n\n<p>I check the  type is:  array but the , and  is  type.\nSo what should I do to modify this code. <\/p>\n","answers":[{"AnswerId":"35013927","CreationDate":"2016-01-26T12:31:18.840","ParentId":null,"OwnerUserId":"5841128","Title":null,"Body":"<p>get the solution by my mentor, coding it like following:<\/p>\n\n<pre><code>XS = T.tensor3(\"XS\") \nw_o = init_weights((1089, 10))\nb_o = init_weights((10,))\n\nresults, updates = theano.scan(lambda XS: T.nnet.softmax(T.dot(XS,w_o) + b_o),  \n                               sequences=[XS],  \n                               outputs_info=None)\n\nresult = results  \nMutiply = theano.function(inputs=[XS], outputs=[result]) \n<\/code><\/pre>\n"}]}
{"QuestionId":35012529,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-01-26T11:14:14.350","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch - Randomly choose dimension of tensor to slice","Body":"<p>I have a torch tensor of size  and I would like to randomly choose one of the last 3 dimensions to randomly slice at s and then do. For example it could output the below tensors with equal probability;<\/p>\n\n<pre><\/pre>\n\n<p>I realise I could just do a few if else statements but I am curious if there is a \"neater\" option using a function like  to select the dimension.<\/p>\n","answers":[{"AnswerId":"35013345","CreationDate":"2016-01-26T11:57:20.200","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>assuming that you want to narrow a block of s elements randomly, out of n elements.<\/p>\n\n<p>Let's use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#self-narrowdim-index-size\" rel=\"nofollow\">:narrow<\/a>.<\/p>\n\n<pre><code>n = 100\ns = 20\nx = torch.randn(1, n, n, n)\ny = x:narrow(torch.random(2, 4), torch.random(1, n - s + 1), s)\n<\/code><\/pre>\n"}]}
{"QuestionId":35013080,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-01-26T11:43:25.443","AcceptedAnswerId":35018188.0,"OwnerUserId":1681590.0,"Title":"Tensorflow: How to get all variables from rnn_cell.BasicLSTM & rnn_cell.MultiRNNCell","Body":"<p>I have a setup where I need to initialize an LSTM after the main initialization which uses . I.e. I want to call  <\/p>\n\n<p>Is there way to collect all the internal trainable variables for both:<\/p>\n\n<ul>\n<li>rnn_cell.BasicLSTM<\/li>\n<li>rnn_cell.MultiRNNCell<\/li>\n<\/ul>\n\n<p>so that I can initialize <strong>JUST<\/strong> these parameters?<\/p>\n\n<p>The main reason I want this is because I do not want to re-initialize some trained values from earlier.<\/p>\n","answers":[{"AnswerId":"35018188","CreationDate":"2016-01-26T15:58:52.233","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The easiest way to solve your problem is to use variable scope. The names of the variables within a scope will be prefixed with its name. Here is a short snippet:<\/p>\n\n<pre><code>cell = rnn_cell.BasicLSTMCell(num_nodes)\n\nwith tf.variable_scope(\"LSTM\") as vs:\n  # Execute the LSTM cell here in any way, for example:\n  for i in range(num_steps):\n    output[i], state = cell(input_data[i], state)\n\n  # Retrieve just the LSTM variables.\n  lstm_variables = [v for v in tf.all_variables()\n                    if v.name.startswith(vs.name)]\n\n# [..]\n# Initialize the LSTM variables.\ntf.initialize_variables(lstm_variables)\n<\/code><\/pre>\n\n<p>It would work the same way with <code>MultiRNNCell<\/code>.<\/p>\n\n<p>EDIT: changed <code>tf.trainable_variables<\/code> to <code>tf.all_variables()<\/code><\/p>\n"},{"AnswerId":"36931096","CreationDate":"2016-04-29T06:32:08.510","ParentId":null,"OwnerUserId":"1628308","Title":null,"Body":"<p>You can also use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#get_collection\" rel=\"noreferrer\"><code>tf.get_collection()<\/code><\/a>:<\/p>\n\n<pre><code>cell = rnn_cell.BasicLSTMCell(num_nodes)\nwith tf.variable_scope(\"LSTM\") as vs:\n  # Execute the LSTM cell here in any way, for example:\n  for i in range(num_steps):\n    output[i], state = cell(input_data[i], state)\n\n  lstm_variables = tf.get_collection(tf.GraphKeys.VARIABLES, scope=vs.name)\n<\/code><\/pre>\n\n<p>(partly copied from Rafal's answer)<\/p>\n\n<p>Note that the last line is equivalent to the list comprehension in Rafal's code. <\/p>\n\n<p>Basically, tensorflow stores a global collection of variables, which can be fetched by either <code>tf.all_variables()<\/code> or <code>tf.get_collection(tf.GraphKeys.VARIABLES)<\/code>. If you specify <code>scope<\/code> (scope name) in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#get_collection\" rel=\"noreferrer\"><code>tf.get_collection()<\/code><\/a> function, then you only fetch tensors (variables in this case) in the collection whose scopes are under the specified scope.<\/p>\n\n<p>EDIT:\nYou can also use <code>tf.GraphKeys.TRAINABLE_VARIABLES<\/code> to get trainable variables only. But since vanilla BasicLSTMCell does not initialize any non-trainable variable, both will be functionally equivalent. For a complete list of default graph collections, check <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/framework.html#GraphKeys\" rel=\"noreferrer\">this<\/a> out.<\/p>\n"}]}
{"QuestionId":35014913,"AnswerCount":0,"Tags":"<android><android-ndk><convolution><caffe><conv-neural-network>","CreationDate":"2016-01-26T13:28:02.690","AcceptedAnswerId":null,"OwnerUserId":3009514.0,"Title":"How to get the confidence of my prediction on caffe-mobile?","Body":"<p>I'm trying to get the confidence of my prediction on caffe-mobile. In python its easy to get it, I just did:<\/p>\n\n<pre><\/pre>\n\n<p>And I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>I have four classes, so these are the confidence of my classes. <\/p>\n\n<p>On android if I do this:<\/p>\n\n<pre><\/pre>\n\n<p>My output is just the predicted class.<\/p>\n\n<pre><\/pre>\n\n<p>I think that I have to implement this method, because the size of  is 1. But how do I change and what files must I change to get it work on android?<\/p>\n\n<p>So, how can I do this using <a href=\"https:\/\/github.com\/sh1r0\/caffe-android-lib\" rel=\"nofollow\">caffe-mobile<\/a> for android?<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>I added the following to :<\/p>\n\n<pre><\/pre>\n\n<p>Then I added this to :<\/p>\n\n<pre><\/pre>\n\n<p>I added this to caffe android class:<\/p>\n\n<pre><\/pre>\n\n<p>And called this method<\/p>\n\n<pre><\/pre>\n\n<p>But, it only return a rounded value ( I think ).<\/p>\n\n<pre><\/pre>\n\n<p>The output if , ( I have 4 classes );<\/p>\n\n<p>So:<\/p>\n\n<pre><\/pre>\n\n<p>If try to classify an image labelled as 0, I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>or if try to classify an image labelled as 2, this is the output:<\/p>\n\n<pre><\/pre>\n\n<p>Am I doing something wrong, or is it not possible?<\/p>\n","answers":[]}
{"QuestionId":35015196,"AnswerCount":1,"Tags":"<python><numpy><matrix><linear-algebra><theano>","CreationDate":"2016-01-26T13:42:08.663","AcceptedAnswerId":35016071.0,"OwnerUserId":3041068.0,"Title":"Does Theano\/Numpy's matrix inverse function use GPU at all?","Body":"<p>I ran the following code to compute the pseudo inverse of a matrix, but it seems that it makes no difference whether I switch on the GPU or not.<\/p>\n\n<pre><\/pre>\n\n<p>Then I looked at Theano's source code of , and found it just calls Numpy's  in the following code (I'll leave out the comments).<\/p>\n\n<pre><\/pre>\n\n<p>I'm not very familar with how Numpy is implemented, can it run on GPU?<br>\nIf not, does that mean that every time I want to compute matrix inverse in Theano, I have to go back from GPU to CPU?<\/p>\n","answers":[{"AnswerId":"35016071","CreationDate":"2016-01-26T14:23:02.637","ParentId":null,"OwnerUserId":"3005167","Title":null,"Body":"<p>See the article <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"nofollow\">Using the GPU<\/a> in the theano docs.<\/p>\n\n<blockquote>\n  <p>Note that we use the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/shared.html\" rel=\"nofollow\"><code>shared<\/code><\/a> function to make sure that the input x is stored on the graphics device.<\/p>\n<\/blockquote>\n\n<p>You have to make sure that the data is stored in graphics memory. Otherwise, I imagine, theano falls back to using numpy routines.<\/p>\n\n<p>Numpy normally does not run on the GPU. I'm not sure if it is possible to link it against CudaBLAS for doing so, but I guess that is out of scope here.<\/p>\n"}]}
{"QuestionId":35016712,"AnswerCount":1,"Tags":"<visual-studio><python-2.7><registry><windows-7-x64><theano>","CreationDate":"2016-01-26T14:51:54.200","AcceptedAnswerId":null,"OwnerUserId":618045.0,"Title":"Theano compilation hangs due to reg query \"HKCU\\SOFTWARE\\Microsoft\\VisualStudio\\SxS\\VS7\" \/v \"10.0\" fail","Body":"<p>I try to run Python 2.7.11 (Anaconda 2.4.1) code which utilizes Theano 0.8.0 computations. Simple GPU test from Theano documentations passes but my code fails.\nAfter a lot of compilation it hangs on <\/p>\n\n<pre><\/pre>\n\n<p>according to Process explorer. If try to execute in command line:<\/p>\n\n<pre><\/pre>\n\n<p>How this problem can be solved?<\/p>\n","answers":[{"AnswerId":"35017215","CreationDate":"2016-01-26T15:15:54.123","ParentId":null,"OwnerUserId":"618045","Title":null,"Body":"<p>It seems manual adding registry entry <\/p>\n\n<pre><code>HKEY_CURRENT_USER\\SOFTWARE\\Microsoft\\VisualStudio\\SxS\\VS7\n    10.0    REG_SZ    C:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\\n<\/code><\/pre>\n\n<p>solved this problem.<\/p>\n\n<p>UPDATE:\nAnother time there was missing:<\/p>\n\n<pre><code>HKEY_LOCAL_MACHINE\\SOFTWARE\\Microsoft\\VisualStudio\\SxS\\VS7\n    10.0    REG_SZ    C:\\Program Files (x86)\\Microsoft Visual Studio 10.0\\VC\\\n<\/code><\/pre>\n"}]}
{"QuestionId":35017213,"AnswerCount":0,"Tags":"<python><opencv><neural-network><deep-learning><caffe>","CreationDate":"2016-01-26T15:15:50.777","AcceptedAnswerId":null,"OwnerUserId":5389418.0,"Title":"why net.blobs['fc7'].data[0] giving all zeros","Body":"<p>I want to extract face descriptors from photos of people. This is what I've done so far:<\/p>\n\n<ol>\n<li>First detected faces from photos using opencv library in python.<\/li>\n<li>Saved those faces in another image.<\/li>\n<li>Next I have to extract descriptor from face image.<\/li>\n<\/ol>\n\n<p>For this I have downloaded  CNN from here: <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/vgg_face\/\" rel=\"nofollow\">http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/vgg_face\/<\/a><\/p>\n\n<p>To extract descriptor, first I did this:<\/p>\n\n<pre><\/pre>\n\n<p>But it gives dimension mismatch error that data should be of dimension  instead of <\/p>\n\n<p>Then I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>Here when I , it displays all zeros. Why is it so?<\/p>\n","answers":[]}
{"QuestionId":35020357,"AnswerCount":2,"Tags":"<python><centos><glibc><tensorflow><binutils>","CreationDate":"2016-01-26T17:44:51.563","AcceptedAnswerId":35081986.0,"OwnerUserId":5272213.0,"Title":"Installing Tensorflow on centos 5","Body":"<p>I'm trying to install Tensorflow on Centos 5.\nUnfortunatelly, after having installed it with pip as explained in the official doc<\/p>\n\n<blockquote>\n  <p>sudo pip install --upgrade\n  <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl\" rel=\"noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.5.0-cp27-none-linux_x86_64.whl<\/a><\/p>\n<\/blockquote>\n\n<p>I have the following error when I try to import tensorflow on a python shell<\/p>\n\n<blockquote>\n  <p>ImportError:\n  \/home\/users\/caohao\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so:\n  ELF file OS ABI invalid<\/p>\n<\/blockquote>\n\n<p>The following issue on github gave me a lead:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/350\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/350<\/a><\/p>\n\n<p>Actually, tensorflow requires glibc version to be at least 2.17 <\/p>\n\n<p>When I type <\/p>\n\n<blockquote>\n  <p>ldd --version<\/p>\n<\/blockquote>\n\n<p>I have the following output: <\/p>\n\n<pre><\/pre>\n\n<p>So I need to upgrade glibc. \nI would like to perform a \"parallel\" installation, because upgrading glibc could be dangerous for my machine\nSo I downloaded the source, and I use the command<\/p>\n\n<blockquote>\n  <p>..\/glibc-2.18\/configure --prefix=\/opt\/glibc-2.18<\/p>\n<\/blockquote>\n\n<p>to install it in an other directory<\/p>\n\n<p>The problem is that I have the following error when I launch configure<\/p>\n\n<blockquote>\n  <p>configure: error: assembler too old, .cfi_personality support missing<\/p>\n<\/blockquote>\n\n<p>I think I have to update binutils as well, so I downloaded the sources and install it in \/opt\/binutils-2.26\/ to not harm the system.<\/p>\n\n<p>Yet I still have the same error when I try to configure glibc.<\/p>\n\n<p>Here are my questions:<\/p>\n\n<p>How can I specify which version of binutils I want to use to build glibc ?<\/p>\n\n<p>EDIT: I should have RTFM, answer is use \u2018--with-binutils=directory\u2019 option<\/p>\n\n<p>How can I make tensorflow looking for the good version of glibc when it will be installed in \/opt\/glibc-2.18 ?<\/p>\n\n<p>FYI I use python 2.7.11 <\/p>\n\n<p>Thank you very much and best regards<\/p>\n","answers":[{"AnswerId":"35020825","CreationDate":"2016-01-26T18:09:15.087","ParentId":null,"OwnerUserId":"3507061","Title":null,"Body":"<p>I think tensorflow needs exactly GLIB 2.17\nand not 2.18. \nFor some reason it got hardcoded.<\/p>\n\n<p>See also:\n<a href=\"https:\/\/stackoverflow.com\/questions\/33655731\/error-while-importing-tensorflow-in-python2-7-in-ubuntu-12-04-glibc-2-17-not-f\/33658956#33658956\">Error while importing Tensorflow in python2.7 in Ubuntu 12.04. &#39;GLIBC_2.17 not found&#39;<\/a><\/p>\n"},{"AnswerId":"35081986","CreationDate":"2016-01-29T10:19:09.167","ParentId":null,"OwnerUserId":"2853544","Title":null,"Body":"<p>I have same problem. <\/p>\n\n<p>But I used CentOS 6.5 and my GLIBC(GNU libc) version is 2.12.<\/p>\n\n<p>I tried to changed my GLIBC version manually by removing version 2.12 and installing 2.17.<\/p>\n\n<p>It occured many problems for it's dependencies and cannot import TensorFlow.<\/p>\n\n<p>So, I used <strong>CentOS 7.0 and it has default GLIBC version 2.17.<\/strong><\/p>\n\n<p>I can install and run TensorFlow using CentOS 7.0 without any problem.<\/p>\n\n<p><em>Thank you.<\/em><\/p>\n"}]}
{"QuestionId":35020757,"AnswerCount":2,"Tags":"<linux><nvidia><caffe><pycaffe>","CreationDate":"2016-01-26T18:05:43.133","AcceptedAnswerId":null,"OwnerUserId":5105872.0,"Title":"fatal error: opencv2\/core\/core.hpp caffe framework installation","Body":"<p>I'm trying to install Caffe framework with this script: <a href=\"https:\/\/gist.github.com\/jetsonhacks\/acf63b993b44e1fb9528\" rel=\"nofollow\">https:\/\/gist.github.com\/jetsonhacks\/acf63b993b44e1fb9528<\/a> and I'm getting opencv2\/core\/core.hpp error.  <\/p>\n\n<pre><\/pre>\n\n<p>In the documentation, OpenCV is optional. But that error tells me not.\nDoes such  have nothing to do with OpenCV installation? <\/p>\n","answers":[{"AnswerId":"50324384","CreationDate":"2018-05-14T06:42:07.177","ParentId":null,"OwnerUserId":"4983420","Title":null,"Body":"<p>Installing opencv for python is not sufficient.\nYou also need to do<\/p>\n\n<pre><code>sudo apt-get install libopencv-dev\n<\/code><\/pre>\n"},{"AnswerId":"37406406","CreationDate":"2016-05-24T06:56:16.860","ParentId":null,"OwnerUserId":"3541400","Title":null,"Body":"<p>I faced the same problem and solved it just by installing OpenCV. You will find other similar errors, unless you have installed all prerequisites. <\/p>\n"}]}
{"QuestionId":35021018,"AnswerCount":2,"Tags":"<gradient><theano><tensorflow><backpropagation><autodiff>","CreationDate":"2016-01-26T18:21:49.733","AcceptedAnswerId":35021911.0,"OwnerUserId":1681590.0,"Title":"Can auto differentiation handle separate functions of array slices?","Body":"<p>Given a vector  of length  say , can auto differentiation tools in say theano or tensorflow be able to take the gradient of something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Would this even make sense? The way I picture it in my mind I would have to do some multiplications by identity vectors\/matrices with trailing 0's to convert <\/p>\n","answers":[{"AnswerId":"51766593","CreationDate":"2018-08-09T12:06:32.397","ParentId":null,"OwnerUserId":"7566962","Title":null,"Body":"<p>Let us take a look at the source code:<\/p>\n\n<pre><code>@ops.RegisterGradient(\"Reshape\")\ndef _ReshapeGrad(op, grad):\n  return [array_ops.reshape(grad, array_ops.shape(op.inputs[0])), None]\n<\/code><\/pre>\n\n<p>This is how tensorflow automatically differentiates.<\/p>\n"},{"AnswerId":"35021911","CreationDate":"2016-01-26T19:11:30.297","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Slice and reshape operations fit into standard reverse mode AD framework in the same way as any other op. Below is a simple TensorFlow program that is similar to the example you gave (I had to change a couple of things to make dimensions match), and the resulting computation graph for the gradient<\/p>\n\n<pre><code>def f(x, z):\n  \"\"\"Adds values together, reshapes into vector\"\"\"\n  return tf.reshape(x+z, (5,))\n\nx = tf.Variable(np.random.rand(5, 1))\nz = tf.Variable(np.random.rand(5, 1))\nv = f(x, z)\nw = tf.slice(v, 0, 5)\nw = tf.reshape(v, (5, 1))\ny = tf.matmul(tf.reshape(w, (5, 1)), tf.transpose(x)) + tf.slice(v, 0, 5)\ncost = tf.square(tf.reduce_sum(y-x))\nprint tf.gradients(cost, [x, z])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TjuOW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TjuOW.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35021462,"AnswerCount":2,"Tags":"<python><machine-learning><theano>","CreationDate":"2016-01-26T18:46:34.253","AcceptedAnswerId":null,"OwnerUserId":5194223.0,"Title":"Error Importing Theano on Windows 10: Failed to compile cuda_ndarray.cu","Body":"<p>When I try to import Theano in Python 2.7, I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I noticed that in the  --compiler-bindir  section, the path has no slashes (C:Program Files (x86)Microsoft Visual Studio 12.0VCbinamd64). I have checked my PATH, THEANO_FLAGS, and .theanorc.txt file and none have this path without the slashes. Is there a way to resolve this?<\/p>\n\n<p>When I try to paste this command, inserting the slashes, into the command prompt, I get the following results:<\/p>\n\n<pre><\/pre>\n\n<p>This makes no sense to me, as the first entry in my PATH is:<\/p>\n\n<pre><\/pre>\n\n<p>My .theanorc.txt is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Any help would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"42455474","CreationDate":"2017-02-25T11:29:33.727","ParentId":null,"OwnerUserId":"6086782","Title":null,"Body":"<p>Just check if <code>cl.exe<\/code> actually exists in the directory:<\/p>\n\n<p><strong><code>C:\\Program Files (x86)\\Microsoft Visual Studio 12.0\\VC\\bin\\amd64<\/code><\/strong><\/p>\n\n<p><strong><em>If it doesn't<\/em><\/strong>, then install it(along with all the other MS runtime files for C++, which doesn't come by default) by running the VS setup, select Modify option, then expand and select:<\/p>\n\n<p><strong><code>Programming Language -&gt; C++<\/code><\/strong><\/p>\n\n<p>and install.<\/p>\n\n<p>Hopefully, now it might work. Please be patient as that download might take quite a while...<\/p>\n\n<blockquote>\n  <p><strong>PS: Iff possible, migrate your Theano workbase to UNIX based OS like (Ubuntu\/Kali Linux\/Manjaro) + Anaconda for SciPy and NumPy, as it poses much less of a problem than in Windows<\/strong><\/p>\n<\/blockquote>\n"},{"AnswerId":"42455276","CreationDate":"2017-02-25T11:10:50.997","ParentId":null,"OwnerUserId":"2652482","Title":null,"Body":"<p>I could solve this problem by setting windows environment variable, instead of directly mentioning the path in compiler-bindir. <\/p>\n\n<p>You can set windows environment variable by right click on <\/p>\n\n<p>\"start->computer->properties->advanced system settings\"<\/p>\n"}]}
{"QuestionId":35022095,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-01-26T19:21:26.970","AcceptedAnswerId":null,"OwnerUserId":4671049.0,"Title":"pyCaffe not giving same results as command line interface","Body":"<p>I'm trying to use a python script to extract a feature value using a pre-trained network. Here's my script:<\/p>\n\n<pre><\/pre>\n\n<p>Now the image D_4051 should return a score somewhere between 0.002 - 0.006, but the python script returns [[ 2.99469399]] for ANY image from my test set. It's clear my code isn't doing what I want it to do, but from the documentation and examples I have found, I'm unsure what is wrong.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"35027234","CreationDate":"2016-01-27T01:26:22.203","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>When you use python interface, you must pay attention to that the format of test image should be the same as the images you use for training. For example, pixel values are [0..1] or [0..256], RGB order or BGR order, mean subtraction and so on. Only if you preprocess the test image exact the same as the training images, you can get reasonable results.<\/p>\n\n<p>Some useful examples: <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a><\/p>\n"}]}
{"QuestionId":35027152,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe>","CreationDate":"2016-01-27T01:18:14.083","AcceptedAnswerId":null,"OwnerUserId":3130081.0,"Title":"how to write caffe python layer with trainable parameters?","Body":"<p>I want to learn how to write caffe python layers.<br>\nBut I only find examples about very simple layers like <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py\" rel=\"nofollow\"><\/a>.<\/p>\n\n<p>How to write python caffe with trainable parameters?<\/p>\n\n<p>For example, how to write a fully connected python layer?<\/p>\n","answers":[{"AnswerId":"35030680","CreationDate":"2016-01-27T06:59:44.240","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe stores the layer's trainable parameters as a vector of <code>blobs<\/code>. By default this vector is empty and it is up to you to add parameters blobs to it in the <code>setup<\/code> of the layer. There is a simple example for a layer with parameters in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_python_layer.py#L31\" rel=\"nofollow noreferrer\"><code>test_python_layer.py<\/code><\/a>.<\/p>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/a\/41481539\/1714410\">this post<\/a> for more information about <code>\"Python\"<\/code> layers in caffe.<\/p>\n"}]}
{"QuestionId":35028173,"AnswerCount":2,"Tags":"<python><deep-learning><tensorflow>","CreationDate":"2016-01-27T03:12:38.380","AcceptedAnswerId":null,"OwnerUserId":5825952.0,"Title":"How to read images with different size in a TFRecord file","Body":"<p>I have created a dataset and saved it into a TFRecord file. The thing is the pictures have different size, so I want to save the size as well with the images. So I used the TFRecordWriter and defined the features like:<\/p>\n\n<pre><\/pre>\n\n<p>I expected that I can read and decode the image using TFRecordReader, but the thing is I cannot get the value of rows and cols from the file because they are tensors. So how am I supposed to do to read the size dynamically and reshape the image accordingly. Thanks guys<\/p>\n","answers":[{"AnswerId":"35028637","CreationDate":"2016-01-27T03:59:09.847","ParentId":null,"OwnerUserId":"3632556","Title":null,"Body":"<p>You can call <code>tf.reshape<\/code> with a dynamic <code>shape<\/code> parameter.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>image_rows = tf.cast(features['rows'], tf.int32)\nimage_cols = tf.cast(features['cols'], tf.int32)\nimage_data = tf.decode_raw(features['image_raw'], tf.uint8)\nimage = tf.reshape(image_data, tf.pack([image_rows, image_cols, 3]))\n<\/code><\/pre>\n"},{"AnswerId":"35050814","CreationDate":"2016-01-28T00:28:06.947","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>I suggest a workflow like:<\/p>\n\n<pre><code>TARGET_HEIGHT = 500\nTARGET_WIDTH = 500\n\nimage = tf.image.decode_jpeg(image_buffer, channels=3)\nimage = tf.image.convert_image_dtype(image, dtype=tf.float32)\n\n# Choose your bbox here.\nbbox_begin = ...  (should be (h_start, w_start, 0))\nbbox_size = tf.constant((TARGET_HEIGHT, TARGET_WIDTH, 3), dtype=tf.int32)\n\ncropped_image = tf.slice(image, bbox_begin, bbox_size)\n<\/code><\/pre>\n\n<p><code>cropped_image<\/code> has a constant tensor size, and can then be thrown into a shuffle batch.<\/p>\n\n<p>You can dynamically access the size of the decoded image using <code>tf.shape(image)<\/code>.  You can do computations on the resulting sub-elements and then stitch them back together using something like <code>bbox_begin = tf.pack([bbox_h_start, bbox_y_start, 0])<\/code>.  Just need to insert your own logic in there for determining the start points of the crop, and what you want to do if the image starts out smaller than you want for your pipeline.<\/p>\n\n<p>If you want to upsize <em>only<\/em> if the image is smaller than your target dimensions, you'll need to use <code>tf.control_flow_ops.cond<\/code> or equivalent.  But you <em>could<\/em> use min and max operations to set the size of your crop window so that you're returning the full image iff it's smaller than the requested dimensions, and then unconditionally resize up to 500x500.  The cropped image will already be at 500x500, so the resize should become an effective no-op.<\/p>\n"}]}
{"QuestionId":35029683,"AnswerCount":1,"Tags":"<caffe><pycaffe><cudnn>","CreationDate":"2016-01-27T05:42:49.100","AcceptedAnswerId":null,"OwnerUserId":5845019.0,"Title":"Caffe build failed linking cuDNN","Body":"<p>I am new to Ubuntu and caffe. I am trying to build an image recognition system using Caffe on Ubuntu on AWS, but I'm stuck on an error regarding an incompatible cuDNN.<\/p>\n\n<p>The command I executed : <\/p>\n\n<pre><\/pre>\n\n<p>The error I get : <\/p>\n\n<blockquote>\n  <p>LD -o .build_release\/lib\/libcaffe.so.1.0.0-rc3 \/usr\/bin\/ld: skipping\n  incompatible \/usr\/local\/cuda\/lib64\/libcudnn.so when searching for\n  -lcudnn \/usr\/bin\/ld: cannot find -lcudnn collect2: error: ld returned 1 exit status make:  [.build_release\/lib\/libcaffe.so.1.0.0-rc3]\n  Error 1<\/p>\n<\/blockquote>\n\n<p>I appreciate your help.<\/p>\n","answers":[{"AnswerId":"35100315","CreationDate":"2016-01-30T09:55:26.183","ParentId":null,"OwnerUserId":"3810057","Title":null,"Body":"<p>I wrote a <a href=\"http:\/\/cloudcv.wordpress.com\" rel=\"nofollow\">step-by-step blog post<\/a> for installation of caffe on Mac OSX and Ubuntu. Hope it helps.<\/p>\n"}]}
{"QuestionId":35031053,"AnswerCount":2,"Tags":"<converter><photos><tensorflow>","CreationDate":"2016-01-27T07:21:11.807","AcceptedAnswerId":null,"OwnerUserId":5845360.0,"Title":"how do i convert photographs to tensors","Body":"<p>I am a neophyte neural network user trying to get to grips with TensorFlow. I have used the MNIST dataset as a test, and would now like to use real world data.\nCan anyone point me to a \"Howto\" or paper or source which tells me how to go about converting digital photographs in files, (jpeg, png, gif, wmf), into a tensors ready for import into TensorFlow please?<\/p>\n\n<p>Cheers!<\/p>\n","answers":[{"AnswerId":"42993909","CreationDate":"2017-03-24T07:38:02.477","ParentId":null,"OwnerUserId":"6885324","Title":null,"Body":"<p>The standard way to load data into Tensorflow is to use a TFRecords file.<\/p>\n\n<blockquote>\n  <p>Another approach is to convert whatever data you have into a supported format. This approach makes it easier to mix and match data sets and network architectures. The recommended format for TensorFlow is a TFRecords file containing tf.train.Example protocol buffers.<\/p>\n<\/blockquote>\n\n<p>-<a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">Tensorflow Documentation<\/a><\/p>\n\n<p>Basically TFRecord is a binary representation of your data or images along with its labels, file names, and other information. Its main advantages are to allow you to stream data into the model efficiently by using Tensorflow's threading and to increase flexibility between different models.<\/p>\n\n<p>You can use this <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data\/build_image_data.py\" rel=\"nofollow noreferrer\">script<\/a> to generate your own TFRecord files.<\/p>\n\n<p>Additionally, you can read on how to use the script <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception#how-to-construct-a-new-dataset-for-retraining\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"42989129","CreationDate":"2017-03-24T00:02:24.367","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/image\" rel=\"nofollow noreferrer\">TensorFlow image functions<\/a> to load images and convert them into tensors. After loading the images, you will likely want to look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_bilinear\" rel=\"nofollow noreferrer\"><code>tf.image.resize_bilinear<\/code><\/a> to resize the images to standard sizes. <\/p>\n"}]}
{"QuestionId":35032675,"AnswerCount":3,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-01-27T08:53:54.137","AcceptedAnswerId":35034287.0,"OwnerUserId":5473359.0,"Title":"How to create dataset similar to cifar-10","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":"37822297","CreationDate":"2016-06-14T21:18:40.273","ParentId":null,"OwnerUserId":"5120590","Title":null,"Body":"<p>I didn't find any of the answers to do what I wanted to I made my own solution. It can be found on my github here: <a href=\"https:\/\/github.com\/jdeepee\/machine_learning\/tree\/master\" rel=\"nofollow\">https:\/\/github.com\/jdeepee\/machine_learning\/tree\/master<\/a><\/p>\n\n<p>This script will convert and amount of images to training and test data where the arrays are the same shape as the cifar10 dataset. <\/p>\n\n<p>The code is commented so should be easy enough to follow. I should note it iterated through a master directory containing multiple folders which contain the images.<\/p>\n"},{"AnswerId":"35034287","CreationDate":"2016-01-27T10:06:48.837","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>I have successfully run the CIFAR10 code on a few of my own data sets. I believe I should be able to give you a good answer.<\/p>\n\n<p>Before we do that though, we need to understand the format in which the CIFAR10 data set is in. If we refer to: <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\">https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a>, and specifically, the Binary Version section, we see this:<\/p>\n\n<blockquote>\n  <p>In other words, the first byte is the label of the first image, which\n  is a number in the range 0-9. The next 3072 bytes are the values of\n  the pixels of the image. The first 1024 bytes are the red channel\n  values, the next 1024 the green, and the final 1024 the blue. The\n  values are stored in row-major order, so the first 32 bytes are the\n  red channel values of the first row of the image.<\/p>\n<\/blockquote>\n\n<p>Intuitively, we need to store the data in this format. What you can do next as sort of a baseline experiment first, is to get images that are exactly the same size and same number of classes as CIFAR10 and put them in this format. This means that your images should have a size of 32x32x3 and have 10 classes. If you can successfully run this, then you can go further on to factor cases like single channels, different size inputs, and different classes. Doing so would mean that you have to change many variables in the other parts of the code. You have to slowly work your way through.<\/p>\n\n<p>I'm in the midst of working out a general module. My code for this is in <a href=\"https:\/\/github.com\/jkschin\/svhn\">https:\/\/github.com\/jkschin\/svhn<\/a>. If you refer to the svhn_flags.py code, you will see many flags there that can be changed to accommodate your needs. I admit it's cryptic now, as I haven't cleaned it up such that it is readable, but it works. If you are willing to spend some time taking a rough look, you will figure something out.<\/p>\n\n<p>This is probably the easy way to run your own data set on CIFAR10. You could of course just copy the neural network definition and implement your own reader, input format, batching, etc, but if you want it up and running fast, just tune your inputs to fit CIFAR10.<\/p>\n\n<p>I hope this helps.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Some really really basic code that I hope would help.<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>from PIL import Image\nimport numpy as np\n\nim = Image.open('images.jpeg')\nim = (np.array(im))\n\nr = im[:,:,0].flatten()\ng = im[:,:,1].flatten()\nb = im[:,:,2].flatten()\nlabel = [1]\n\nout = np.array(list(label) + list(r) + list(g) + list(b),np.uint8)\nout.tofile(\"out.bin\")\n<\/code><\/pre>\n\n<p>This would convert an image into a byte file that is ready for use in CIFAR10. For multiple images, just keep concatenating the arrays, as stated in the format above. To check if your format is correct, specifically for the Asker's use case, you should get a file size of 427*427*3 + 1 = 546988 bytes. Assuming your pictures are RGB and values range from 0-255. Once you verify that, you're all set to run in TensorFlow. Do use TensorBoard to perhaps visualize one image, just to guarantee correctness.<\/p>\n\n<p>EDIT 2:<\/p>\n\n<p>As per Asker's question in comments, <\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>if not eval_data:\n    filenames = [os.path.join(data_dir, 'data_batch_%d.bin' % i)\n                 for i in xrange(1, 6)]\n<\/code><\/pre>\n\n<p>If you really wanna it to work as it is, you need to study the function calls of CIFAR10 code. In cifar10_input, the batches are hardcoded. So you have to edit this line of code to fit the name of the bin file. Or, just distribute your images into 6 bin files evenly.<\/p>\n"},{"AnswerId":"37615133","CreationDate":"2016-06-03T13:01:26.870","ParentId":null,"OwnerUserId":"5651936","Title":null,"Body":"<p>for SVHN dataset\nYou can try like this for multiple input images:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.io \n\nmat = scipy.io.loadmat('train_32x32.mat')\ndata = mat['X']\nlabel = mat['y']\n\nR_data = data[:,:,0,:]\nG_data = data[:,:,1,:]\nB_data = data[:,:,2,:]\n\nR_data = np.transpose(R_data, (2,0,1))\nG_data = np.transpose(G_data, (2,0,1))\nB_data = np.transpose(B_data, (2,0,1))\n\nR_data = np.reshape(R_data,(73257,32*32))\nG_data = np.reshape(G_data,(73257,32*32))\nB_data = np.reshape(B_data,(73257,32*32))\n\noutdata = np.concatenate((label,R_data,G_data,B_data), axis = 1)\nstep = 10000\nfor i in range(1,6):\n    temp = outdata[i*step:(i+1)*step,:]\n    temp.tofile('SVHN_train_data_batch%d.bin' % i)\n    print('save data %d' % i)\n<\/code><\/pre>\n"}]}
{"QuestionId":35034120,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-27T10:00:23.130","AcceptedAnswerId":35034484.0,"OwnerUserId":309830.0,"Title":"How to understand the \"Densely Connected Layer\" section in tensorflow tutorial","Body":"<p>In the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/pros\/index.html#convolution-and-pooling\" rel=\"noreferrer\">Densely Connected Layer<\/a> section of the tensorflow tutorial, it says the image size is <strong>7 x 7<\/strong>, after it is been processed. I tried the code, and it seem these parameters works.<\/p>\n\n<p>But I do not know how to get this <strong>7 x 7<\/strong> dimension. I understand that:<\/p>\n\n<ul>\n<li>the original image is 28 x 28,<\/li>\n<li>in the 1st conv layer, the  function will reduce both of the image dimension by a factor of 4, so after the first pooling operation, the image size is <strong>7 x 7<\/strong><\/li>\n<li><p>HERE'S WHAT I DO NOT UNDERSTAND<\/p>\n\n<p>in the 2nd conv layer, there another  function call, so I think the image size should be reduce by a factor of 4 <em>again<\/em>. But actually did not.<\/p><\/li>\n<\/ul>\n\n<p>Which step I got wrong?<\/p>\n","answers":[{"AnswerId":"35034484","CreationDate":"2016-01-27T10:15:34.963","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>You also need to know the stride of the max pool and convolution.<\/p>\n\n<pre><code>def conv2d(x, W):\n  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')\n\ndef max_pool_2x2(x):\n  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],\n                        strides=[1, 2, 2, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>Here, we can see that convolution has a stride of 1 and max pool has a stride of 2. How you can look at max pool, is that it takes a 2x2 box, and slides it over the image, each time taking the maximum value over 4 pixels. If you have a stride of 2, it takes 2 steps each time it moves! The image size should reduce by a factor of 2, instead of 4.<\/p>\n\n<p>In other words, a 28x28 picture with max pool 2x2 and stride 2, will become 14x14. Another max pool 2x2 and stride 2 will reduce it to 7x7.<\/p>\n\n<p>To further illustrate my point, let's take the case of max pool 2x2 and stride 1. If we don't pad the image, it will become a 27x27 image after max pool.<\/p>\n\n<p>Here's an image for a more complete answer:\n<a href=\"https:\/\/i.stack.imgur.com\/8PQKD.jpg\"><img src=\"https:\/\/i.stack.imgur.com\/8PQKD.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"35035584","CreationDate":"2016-01-27T11:06:07.893","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>Take a look at <a href=\"http:\/\/googleresearch.blogspot.co.uk\/2016\/01\/teach-yourself-deep-learning-with.html\" rel=\"nofollow noreferrer\">Teach Yourself Deep Learning with TensorFlow<\/a> and <a href=\"https:\/\/www.udacity.com\/me#!\/\" rel=\"nofollow noreferrer\">Udacity<\/a>\n with Vincent Vanhoucke<\/p>\n\n<p>This is covered in the course. I am currently working through it.<\/p>\n\n<p>The course is free, however you do have to sign up. It is a series of videos, quizzes and coding projects all self paced and self graded. I am learning a lot and enjoy it.<\/p>\n\n<p>Here is one of the quizzes.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RYig4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RYig4.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35037360,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2016-01-27T12:26:59.950","AcceptedAnswerId":35440530.0,"OwnerUserId":1527879.0,"Title":"How to deploy and serve prediction using TensorFlow from API?","Body":"<p>From google tutorial we know how to train a model in TensorFlow. But what is the best way to save a trained model, then serve the prediction using a basic minimal python api in production server.<\/p>\n\n<p>My question is basically for TensorFlow best practices to save the model and serve prediction on live server without compromising speed and memory issue. Since the API server will be running on the background for forever.<\/p>\n\n<p>A small snippet of python code will be appreciated. <\/p>\n","answers":[{"AnswerId":"35440530","CreationDate":"2016-02-16T18:38:30.483","ParentId":null,"OwnerUserId":"1397927","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/serving\">TensorFlow Serving<\/a> is a high performance, open source serving system for machine learning models, designed for production environments and optimized for TensorFlow. The initial release contains C++ server and Python client examples based on <a href=\"http:\/\/www.grpc.io\/\">gRPC<\/a>. The basic architecture is shown in the diagram below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dfLtJ.png\"><img src=\"https:\/\/i.stack.imgur.com\/dfLtJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To get started quickly, check out the <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\">tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":35037642,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-27T12:39:50.633","AcceptedAnswerId":35044662.0,"OwnerUserId":258483.0,"Title":"How can read_cifar10() routine return anything other than first object in TensorFlow tutorial?","Body":"<p>TensorFlow has CIFAR-10 tutorial, which <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks\" rel=\"nofollow\">is discussed here<\/a>. Source code in Python <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/0.6.0\/tensorflow\/models\/image\/cifar10\" rel=\"nofollow\">is here<\/a>.<\/p>\n\n<p>It has <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0.6.0\/tensorflow\/models\/image\/cifar10\/cifar10_input.py\" rel=\"nofollow\">read_cifar10() routine here<\/a>, which is intended to read samples from binary file.<\/p>\n\n<p>I am failing to understand, how it works. Suspect this is somehow related with TensorFlow deferred nature, but can't figure out how.<\/p>\n\n<p>At some point routine does the following:<\/p>\n\n<pre><\/pre>\n\n<p>I see here, that a new reader is created from scratch, and then this reader is pointed to filename queue. <\/p>\n\n<p>How many samples is returned by  call? <\/p>\n\n<p>Later, inside  method a code does following:<\/p>\n\n<pre><\/pre>\n\n<p>Here  is normal Python call, not deferred, so comment assume that fetching of 20000 records will occur immediately.<\/p>\n\n<p>How can it happen? Everywhere I see only per-one-record logic. How it multiplies on many records?<\/p>\n","answers":[{"AnswerId":"35044662","CreationDate":"2016-01-27T17:55:42.503","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TLDR; <code>reader.read<\/code> only adds a <code>read<\/code> operation to computation graph, the actual execution happens during <code>session.run<\/code> which is done by a separate thread in a <code>while(True): session.run(...)<\/code> kind of loop initiated by <code>start_queue_runners<\/code><\/p>\n\n<p>Long version:\nThis is part of an \"input pipeline\" which is complicated by the fact that reading\/prefetching needs to happen asynchronously to avoid blocking. Official how-to describing input pipelines is <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/how_tos\/reading_data\/index.html\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>To be more specific, <code>reader.read<\/code> adds an operation to read a single record to computation graph. This operation then feeds into <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_input.py#L120\" rel=\"nofollow\"><code>shuffle_batch<\/code><\/a> created inside <code>_generate_image_and_label_batch<\/code>. To this point no reading has taken place. The <code>shuffle_batch<\/code> operation creates a queue which decouples the input flow, in a sense that evaluation of the part of the graph before the queue and after the queue\u00a0can be done asynchronously using different <code>session.run<\/code> calls, with queue providing buffering in the middle. Additionally <code>shuffle_batch<\/code> operation registers operations feeding into the queue as part of <code>GraphKeys.QUEUE_RUNNERS<\/code> collection.<\/p>\n\n<p>Inside <code>train()<\/code>, operation <code>tf.start_queue_runners<\/code> will create several threads corresponding to enqueue operations registered in <code>GraphKeys.QUEUE_RUNNERS<\/code> collection and start evaluating them in a loop. Results of <code>reader.read<\/code> will flow through other ops until reaching <code>shuffle_batch<\/code> queue and getting saved in its memory buffer.<\/p>\n\n<p>The part of the graph after <code>shuffle_batch<\/code> will get driven by the main Python thread, initiated by <code>sess.run([train_op, loss])<\/code> command. The thread will get a collect a batch of examples saved on the <code>shuffle_batch<\/code> queue and propagate it forward.<\/p>\n\n<p>Here's an example of feeding your input queue manually instead of using queue runners.<\/p>\n\n<pre><code>queue_dtype = np.int32\nqueue_capacity = 2\nvalues_queue = tf.FIFOQueue(capacity=queue_capacity, dtypes=queue_dtype)\nsize_op = values_queue.size()\nvalue_placeholder = tf.placeholder(dtype=queue_dtype)\nenqueue_op = values_queue.enqueue(value_placeholder)\ndequeue_op = values_queue.dequeue()\nclose_op = values_queue.close()\n\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\n\n# add two elements onto the queue\nsess.run([enqueue_op], {value_placeholder:2})\nsess.run([enqueue_op], {value_placeholder:3})\n# if you uncomment the next line, you'll hang because queue is full\n# sess.run([enqueue_op], {value_placeholder:4})\n\n# close the queue. This means 3rd read will throw OutOfRangeError instead of\n# hanging until queue is replenished\nsess.run([close_op])\nprint('queue has %d\/%d entries' % (sess.run([size_op])[0], queue_capacity))\n\n# take two elements off the queue\nfancy_computation = tf.square(dequeue_op)\nprint('Computation result %d' %(sess.run([fancy_computation])[0]))\nprint('queue has %d\/%d entries' % (sess.run([size_op])[0], queue_capacity))\nprint('Computation result %d' %(sess.run([fancy_computation])[0]))\nprint('queue has %d\/%d entries' % (sess.run([size_op])[0], queue_capacity))\nprint('Computation result %d' %(sess.run([fancy_computation])[0]))\nprint('queue has %d\/%d entries' % (sess.run([size_op])[0], queue_capacity))\n<\/code><\/pre>\n\n<p>What you should see if you run it<\/p>\n\n<pre><code>queue has 2\/2 entries\nComputation result 4\nqueue has 1\/2 entries\nComputation result 9\nqueue has 0\/2 entries\n---------------------------------------------------------------------------\nOutOfRangeError     \n<\/code><\/pre>\n"}]}
{"QuestionId":35041894,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-27T15:48:12.460","AcceptedAnswerId":35043509.0,"OwnerUserId":2835670.0,"Title":"Add the elements in a row, for each row, in a matrix, using theano","Body":"<p>I have a vector v and a matrix z, say<\/p>\n\n<pre><\/pre>\n\n<p>I want to create a new vector y given by v + sum of elements in each row of z.\nBasically: y[i] = v[i] + T.sum(z[:,i]) that I can do, for each entry, by:<\/p>\n\n<pre><\/pre>\n\n<p>My question is: is there a way, without doing a loop, to write y = v + T.sum(rows of z) in one single line?<\/p>\n","answers":[{"AnswerId":"35043509","CreationDate":"2016-01-27T16:58:33.663","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>You can obtain it like this<\/p>\n\n<pre><code>y = v + z.sum(axis=1)\n<\/code><\/pre>\n\n<p>In numpy, and thus in theano, many aggregator functions, such as <code>sum<\/code>, <code>mean<\/code>, <code>var<\/code>, <code>std<\/code>, <code>any<\/code>, <code>all<\/code>, ... have an <code>axis<\/code> keyword argument, and sometimes even an <code>axes<\/code> keyword argument, with which you can specify exactly in which direction you array is to be traversed.<\/p>\n"}]}
{"QuestionId":35049379,"AnswerCount":4,"Tags":"<machine-learning><neural-network><deep-learning><tensorflow><perceptron>","CreationDate":"2016-01-27T22:30:20.657","AcceptedAnswerId":38742267.0,"OwnerUserId":5849136.0,"Title":"Training on imbalanced data using TensorFlow","Body":"<p><strong>The Situation:<\/strong><\/p>\n\n<p>I am wondering how to use TensorFlow optimally when my training data is imbalanced in label distribution between 2 labels. For instance, suppose the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/beginners\/index.html\" rel=\"noreferrer\">MNIST tutorial<\/a> is simplified to only distinguish between 1's and 0's, where all images available to us are either 1's or 0's. This is straightforward to train using the provided TensorFlow tutorials when we have roughly 50% of each type of image to train and test on. But what about the case where 90% of the images available in our data are 0's and only 10% are 1's? I observe that in this case, TensorFlow routinely predicts my entire test set to be 0's, achieving an accuracy of a meaningless 90%.<\/p>\n\n<p>One strategy I have used to some success is to pick random batches for training that do have an even distribution of 0's and 1's. This approach ensures that I can still use all of my training data and produced decent results, with less than 90% accuracy, but a much more useful classifier. Since accuracy is somewhat useless to me in this case, my metric of choice is typically area under the ROC curve (AUROC), and this produces a result respectably higher than .50.<\/p>\n\n<p><strong>Questions:<\/strong><\/p>\n\n<p>(1) Is the strategy I have described an accepted or optimal way of training on imbalanced data, or is there one that might work better?<\/p>\n\n<p>(2) Since the accuracy metric is not as useful in the case of imbalanced data, is there another metric that can be maximized by altering the cost function? I can certainly calculate AUROC post-training, but can I train in such a way as to maximize AUROC?<\/p>\n\n<p>(3) Is there some other alteration I can make to my cost function to improve my results for imbalanced data? Currently, I am using a default suggestion given in TensorFlow tutorials:<\/p>\n\n<pre><\/pre>\n\n<p>I have heard this may be possible by up-weighting the cost of miscategorizing the smaller label class, but I am unsure of how to do this.<\/p>\n","answers":[{"AnswerId":"35050158","CreationDate":"2016-01-27T23:28:08.893","ParentId":null,"OwnerUserId":"1886138","Title":null,"Body":"<p>1) Yes. This is well received strategy to counter imbalanced data.  But this strategy is good in Neural Nets only if you using SGD.<\/p>\n\n<p>Another  easy way to balance the training data is using weighted examples. Just amplify the per-instance loss by a larger weight\/smaller when seeing imbalanced examples. If you use online gradient descent, it can be as simple as using a larger\/smaller learning rate when seeing imbalanced examples.<\/p>\n\n<p>Not sure about 2.<\/p>\n"},{"AnswerId":"37239111","CreationDate":"2016-05-15T13:53:19.100","ParentId":null,"OwnerUserId":"6337206","Title":null,"Body":"<p>I'm one who struggling with imbalanced data. What my strategy to counter imbalanced data are as below.<\/p>\n\n<p>1) Use cost function calculating 0 and 1 labels at the same time like below.<\/p>\n\n<pre><code>cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(_pred) + (1-y)*tf.log(1-_pred), reduction_indices=1))\n<\/code><\/pre>\n\n<p>2) Use SMOTE, oversampling method making number of 0 and 1 labels similar. Refer to here, <a href=\"http:\/\/comments.gmane.org\/gmane.comp.python.scikit-learn\/5278\" rel=\"nofollow\">http:\/\/comments.gmane.org\/gmane.comp.python.scikit-learn\/5278<\/a><\/p>\n\n<p>Both strategy worked when I tried to make credit rating model.<\/p>\n\n<p>Logistic regression is typical method to handle imbalanced data and binary classification such as predicting default rate. AUROC is one of the best metric to counter imbalanced data.<\/p>\n"},{"AnswerId":"38742267","CreationDate":"2016-08-03T11:27:28.750","ParentId":null,"OwnerUserId":"2649901","Title":null,"Body":"<p>(1)It's ok to use your strategy. I'm working with imbalanced data as well, which I try to use down-sampling and up-sampling methods first to make the training set even distributed. Or using ensemble method to train each classifier with an even distributed subset.<\/p>\n\n<p>(2)I haven't seen any method to maximise the AUROC. My thought is that AUROC is based on ture positive rate and false positive rate, which doesn't tell how well it works on each instance. Thus, it may not necessarily maximise the capability to separate the classes.<\/p>\n\n<p>(3)Regarding weighting the cost by the ratio of class instances, it similar to <a href=\"https:\/\/stackoverflow.com\/questions\/35155655\/loss-function-for-class-imbalanced-binary-classifier-in-tensor-flow?answertab=votes#tab-top\">Loss function for class imbalanced binary classifier in Tensor flow<\/a>\nand the answer.<\/p>\n"},{"AnswerId":"48269668","CreationDate":"2018-01-15T19:43:44.590","ParentId":null,"OwnerUserId":"2895528","Title":null,"Body":"<p>Regarding imbalanced datasets, the first two methods that come to mind are (upweighting positive samples, sampling to achieve balanced batch distributions). <\/p>\n\n<p><strong>Upweighting positive samples<\/strong>\nThis refers to increasing the losses of misclassified positive samples when training on datasets that have much fewer positive samples. This incentivizes the ML algorithm to learn parameters that are better for positive samples. For binary classification, there is a simple API in tensorflow that achieves this. See (weighted_cross_entropy) referenced below <\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/weighted_cross_entropy_with_logits\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/weighted_cross_entropy_with_logits<\/a><\/li>\n<\/ul>\n\n<p><strong>Batch Sampling<\/strong>\nThis involves sampling the dataset so that each batch of training data has an even distribution positive samples to negative samples. This can be done using the rejections sampling API provided from tensorflow. <\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/training\/rejection_sample\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/training\/rejection_sample<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":35049563,"AnswerCount":0,"Tags":"<git><curl><proxy><lua><torch>","CreationDate":"2016-01-27T22:42:49.723","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"Luarock fails to install new pacakages?","Body":"<p>I'm running the torch7 on a ubuntu 14.04 LTS pc behind a proxy.\nI have already made environment variables Http_proxy and proxy.<\/p>\n\n<p>The default options for  and  and  has been set to run via this proxy. here are the steps I used to change the settings of the packages of <a href=\"https:\/\/stackoverflow.com\/questions\/11211705\/setting-proxy-in-wget\">Wget proxy<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/7559103\/how-to-setup-curl-to-permanently-use-a-proxy\">Curl proxy<\/a> &amp; <a href=\"https:\/\/stackoverflow.com\/questions\/783811\/getting-git-to-work-with-a-proxy-server\">git proxy<\/a><\/p>\n\n<p>The env variables are available in my  and files.<\/p>\n\n<p>I have also added the proxy in the config.lua file as per <a href=\"https:\/\/github.com\/keplerproject\/luarocks\/wiki\/LuaRocks-through-a-proxy\" rel=\"nofollow noreferrer\">this<\/a> post as well.<\/p>\n\n<p>The error I get when i try to install this is <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35050753,"AnswerCount":5,"Tags":"<python><machine-learning><keras><data-science>","CreationDate":"2016-01-28T00:21:39.103","AcceptedAnswerId":null,"OwnerUserId":4984897.0,"Title":"How big should batch size and number of epochs be when fitting a model in Keras?","Body":"<p>I am training on 970 samples and validating on 243 samples. <\/p>\n\n<p>How big should batch size and number of epochs be when fitting a model in Keras to optimize the val_acc? Is there any sort of rule of thumb to use based on data input size?<\/p>\n","answers":[{"AnswerId":"52231546","CreationDate":"2018-09-08T02:45:30.560","ParentId":null,"OwnerUserId":"8138555","Title":null,"Body":"<p>Great answers above. Everyone gave good inputs.<\/p>\n\n<p>Ideally, this is the sequence of the batch sizes that should be used: <\/p>\n\n<pre><code>{1, 2, 4, 8, 16} - slow \n\n{ [32, 64],[ 128, 256] }- Good starters\n\n[32, 64] - CPU\n\n[128, 256] - GPU for more boost\n<\/code><\/pre>\n"},{"AnswerId":"55264099","CreationDate":"2019-03-20T15:08:54.007","ParentId":null,"OwnerUserId":"5660092","Title":null,"Body":"<p>Epochs is up to your wish, depending upon when validation loss stops improving further. This much should be batch size:<\/p>\n\n<pre><code>\n# To define function to find batch size for training the model\n# use this function to find out the batch size\n\n    def FindBatchSize(model):\n        \"\"\"#model: model architecture, that is yet to be trained\"\"\"\n        import os, sys, psutil, gc, tensorflow, keras\n        import numpy as np\n        from keras import backend as K\n        BatchFound= 16\n\n        try:\n            total_params= int(model.count_params());    GCPU= \"CPU\"\n            #find whether gpu is available\n            try:\n                if K.tensorflow_backend._get_available_gpus()== []:\n                    GCPU= \"CPU\";    #CPU and Cuda9GPU\n                else:\n                    GCPU= \"GPU\"\n            except:\n                from tensorflow.python.client import device_lib;    #Cuda8GPU\n                def get_available_gpus():\n                    local_device_protos= device_lib.list_local_devices()\n                    return [x.name for x in local_device_protos if x.device_type == 'GPU']\n                if \"gpu\" not in str(get_available_gpus()).lower():\n                    GCPU= \"CPU\"\n                else:\n                    GCPU= \"GPU\"\n\n            #decide batch size on the basis of GPU availability and model complexity\n            if (GCPU== \"GPU\") and (os.cpu_count() &gt;15) and (total_params &lt;1000000):\n                BatchFound= 64    \n            if (os.cpu_count() &lt;16) and (total_params &lt;500000):\n                BatchFound= 64  \n            if (GCPU== \"GPU\") and (os.cpu_count() &gt;15) and (total_params &lt;2000000) and (total_params &gt;=1000000):\n                BatchFound= 32      \n            if (GCPU== \"GPU\") and (os.cpu_count() &gt;15) and (total_params &gt;=2000000) and (total_params &lt;10000000):\n                BatchFound= 16  \n            if (GCPU== \"GPU\") and (os.cpu_count() &gt;15) and (total_params &gt;=10000000):\n                BatchFound= 8       \n            if (os.cpu_count() &lt;16) and (total_params &gt;5000000):\n                BatchFound= 8    \n            if total_params &gt;100000000:\n                BatchFound= 1\n\n        except:\n            pass\n        try:\n\n            #find percentage of memory used\n            memoryused= psutil.virtual_memory()\n            memoryused= float(str(memoryused).replace(\" \", \"\").split(\"percent=\")[1].split(\",\")[0])\n            if memoryused &gt;75.0:\n                BatchFound= 8\n            if memoryused &gt;85.0:\n                BatchFound= 4\n            if memoryused &gt;90.0:\n                BatchFound= 2\n            if total_params &gt;100000000:\n                BatchFound= 1\n            print(\"Batch Size:  \"+ str(BatchFound));    gc.collect()\n        except:\n            pass\n\n        memoryused= [];    total_params= [];    GCPU= \"\";\n        del memoryused, total_params, GCPU;    gc.collect()\n        return BatchFound\n<\/code><\/pre>\n"},{"AnswerId":"38405970","CreationDate":"2016-07-15T22:43:33.510","ParentId":null,"OwnerUserId":"4727781","Title":null,"Body":"<p>Since you have a pretty small dataset (~ 1000 samples), you would probably be safe using a batch size of 32, which is pretty standard. It won't make a huge difference for your problem unless you're training on hundreds of thousands or millions of observations. <\/p>\n\n<p><strong>To answer your questions on Batch Size and Epochs:<\/strong><\/p>\n\n<p><em>In general<\/em>: Larger batch sizes result in faster progress in training, but don't always converge as fast. Smaller batch sizes train slower, but <em>can<\/em> converge faster. It's definitely problem dependent. <\/p>\n\n<p><em>In general<\/em>, the models improve with more epochs of training, to a point. They'll start to plateau in accuracy as they converge. Try something like 50 and plot number of epochs (x axis) vs. accuracy (y axis). You'll see where it levels out. <\/p>\n\n<p>What is the type and\/or shape of your data? Are these images, or just tabular data? This is an important detail. <\/p>\n"},{"AnswerId":"38457655","CreationDate":"2016-07-19T11:46:59.373","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>I use Keras to perform non-linear regression on speech data. Each of my speech files gives me features that are 25000 rows in a text file, with each row containing 257 real valued numbers. I use a batch size of 100, epoch 50 to train  <code>Sequential<\/code> model in Keras with 1 hidden layer. After 50 epochs of training, it converges quite well to a low <code>val_loss<\/code>. <\/p>\n"},{"AnswerId":"44901953","CreationDate":"2017-07-04T09:20:48.003","ParentId":null,"OwnerUserId":"5829106","Title":null,"Body":"<p>I used Keras to perform non linear regression for market mix modelling. I got best results with a batch size of 32 and epochs = 100 while training a Sequential model in Keras with 3 hidden layers. Generally batch size of 32 or 25 is good, with epochs = 100 unless you have large dataset. in case of large dataset you can go with batch size of 10 with epochs b\/w 50 to 100.\nAgain the above mentioned figures have worked fine for me.<\/p>\n"}]}
{"QuestionId":35050812,"AnswerCount":1,"Tags":"<python><machine-learning><keras><data-science>","CreationDate":"2016-01-28T00:27:47.400","AcceptedAnswerId":null,"OwnerUserId":4984897.0,"Title":"How to log data about model expected performance in Keras?","Body":"<p>I built a script that saves my best model to the filesystem and I want it to also log data about the expected performance of the model to a different text file.<\/p>\n\n<p>Current code:<\/p>\n\n<pre><\/pre>\n\n<p>I currently save my best model as an HDF5 file. Is there a way to log the performance of the best model which is saved in the HDF5 file using Keras in a different text file?<\/p>\n\n<p>EDIT: Is there a way to use this here <a href=\"http:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">http:\/\/keras.io\/callbacks\/<\/a>?<\/p>\n","answers":[{"AnswerId":"35069019","CreationDate":"2016-01-28T18:17:42.593","ParentId":null,"OwnerUserId":"1735756","Title":null,"Body":"<p>You can use python <code>logging<\/code> module to log your data.<a href=\"https:\/\/docs.python.org\/2\/library\/logging.html\" rel=\"nofollow\">python logging<\/a><\/p>\n\n<p>Use <code>FileHandler<\/code> to write all your data into a file and also you can use <code>ConsoleHandler<\/code> to display all data on console.<\/p>\n\n<p><code>TimedRotatingFileHandler<\/code> can help you rotate your log file in certain time intervals. For instance: every hour, every day or every week.<\/p>\n"}]}
{"QuestionId":35050911,"AnswerCount":1,"Tags":"<python><multiprocessing><caffe><pycaffe>","CreationDate":"2016-01-28T00:39:23.917","AcceptedAnswerId":null,"OwnerUserId":3543300.0,"Title":"Multiprocessing with Python and Caffe","Body":"<p>I have a pretrained caffe model and I'm trying to make a prediction with it. However, I am trying to perform this prediction in a separate process using python's multiprocessing module so as to not slow down the main program. However, the process seems to freeze when trying to make the prediction. Is this possible with caffe? If not, are there other methods to do efficient parallel processing with caffe in python?<\/p>\n\n<p>Here's a simplified version of my code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35056059","CreationDate":"2016-01-28T08:16:11.453","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>You can process images in batch mode using pycaffe. This is much faster than processing image one by one.\nYou can examine the code of pycaffe classifier class\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L89\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L89<\/a>\nfor details.<\/p>\n"}]}
{"QuestionId":35056608,"AnswerCount":0,"Tags":"<python><macos><caffe>","CreationDate":"2016-01-28T08:44:15.617","AcceptedAnswerId":null,"OwnerUserId":5845063.0,"Title":"No rule to make target `\/usr\/lib\/libpython2.7.dylib', needed by `lib\/libcaffe.1.0.0-rc3.dylib'. Stop","Body":"<p>I am installing caffe on my mac 10.10 with cpu only model.<\/p>\n\n<blockquote>\n  <p>[ 72%] Building CXX object src\/caffe\/CMakeFiles\/caffe.dir\/util\/upgrade_proto.cpp.o<br>\n  make[2]: ** No rule to make target '\/usr\/lib\/libpython2.7.dylib', needed by 'lib\/libcaffe.1.0.0-rc3.dylib'.**  Stop.<br>\n  make[1]: ** [src\/caffe\/CMakeFiles\/caffe.dir\/all] Error 2<br>\n  make: ** [all] Error 2 **<\/p>\n<\/blockquote>\n\n<p>First,I check the file  and found the  exits but was useless then I copied a  from     that is my python's file<\/p>\n\n<p>Finally,<\/p>\n\n<pre><\/pre>\n\n<p>This error came out again. I checked the makefile in caffe\/bulid\/src\/caffe\/ but cannot found some url or libxxxxx. How can I fix this error?<\/p>\n","answers":[]}
{"QuestionId":35056909,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-01-28T08:59:47.617","AcceptedAnswerId":35163341.0,"OwnerUserId":1211587.0,"Title":"Input to LSTM network tensorflow","Body":"<p>I have a time series of length t (x0, ...,xt) each of the xi is a d-dimension vector i.e. xi=(x0i, x1i, ...., xdi). Thus my input X is of shape [batch_size, d]<\/p>\n\n<p>The input for the tensorflow LSTM should be of size [batchSize, hidden_size]. \nMy question is how should i input my time series to the LSTM. One possible solution that i thought of is to have additional weight matrix, W, of size [d,hidden_size] and to input the LSTM with X*W + B. <\/p>\n\n<p>Is this correct or should i input something else to the netwo\u05e8k?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35163341","CreationDate":"2016-02-02T20:37:32.920","ParentId":null,"OwnerUserId":"5693571","Title":null,"Body":"<p>Your intuition is correct; what you need (and what you have described) is an embedding to translate your input vector to the dimension of your LSTM's input. There are three primary ways that I know of to accomplish that.<\/p>\n\n<ul>\n<li>You could do this manually with an additional weight matrix <code>W<\/code> and bias vector <code>b<\/code> as you described.<\/li>\n<li>You could create the weight matrix and bias vectors automatically using the <code>linear()<\/code> function <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn_cell.py#L668\" rel=\"noreferrer\">from TensorFlow's rnn_cell.py library<\/a>. Then pass the output of that linear layer as the input of your LSTM when you create your LSTM via the <code>rnn_decoder()<\/code> function <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py#L36\" rel=\"noreferrer\">in Tensorflow's seq2seq.py library<\/a> or otherwise.<\/li>\n<li>Or you could have Tensorflow create this embedding and hook it up to the inputs of your LSTM  automatically, by creating the LSTM via the <code>embedding_rnn_decoder()<\/code> function at line 141 of the same seq2seq library. (If you trace through the code for this function without any optional arguments, you'll see that it is simply creating a linear embedding layer for the input as well as the LSTM and hooking them together.)<\/li>\n<\/ul>\n\n<p>Unless you need access to the individual components that you're creating for some reason, I would recommend the third option to keep your code at a high level.<\/p>\n"}]}
{"QuestionId":35064747,"AnswerCount":2,"Tags":"<python><caffe><pycaffe>","CreationDate":"2016-01-28T14:54:39.163","AcceptedAnswerId":35065272.0,"OwnerUserId":3106889.0,"Title":"Send two argument using map and lambda functions and return two values","Body":"<p>I would like to send two arguments within a function and receive two values as well <\/p>\n\n<p>I want to modify this code so that I code be able to send two arguments and receive two as well<\/p>\n\n<pre><\/pre>\n\n<p>function_f is a function return one number either 1 or 0.<\/p>\n\n<p>The above code can do the job when sending one argument but I want to send two instead of.\nthe first argument of the required function is list_value and the second argument is a model \"net model train it using caffe\"<\/p>\n\n<p>so I would like to write  a function that can do the same job of the previous function but returning two arguments one is [0 or 1] and the other is a modified model which this function has modified.<\/p>\n","answers":[{"AnswerId":"35065272","CreationDate":"2016-01-28T15:19:00.763","ParentId":null,"OwnerUserId":"364696","Title":null,"Body":"<p>Since you say you only want a single model, not a new model for each value, this is fairly simple. Change:<\/p>\n\n<pre><code>hnd = map(lambda (valua): function_f(valua), list_value)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>model = ... initialize a model that will be passed to every call ...\nhnd = map(lambda valua: function_f(valua, model), list_value)\n<\/code><\/pre>\n\n<p>Just make sure <code>function_f<\/code> returns both the new value and model, e.g. if it previously did:<\/p>\n\n<pre><code>def function_f(val, model):\n    ... calculate newval and make newmodel ...\n    return newval\n<\/code><\/pre>\n\n<p>just change it to:<\/p>\n\n<pre><code>def function_f(val, model):\n    ... calculate newval and make newmodel ...\n    return newval, newmodel\n<\/code><\/pre>\n\n<p>Note: If need to use <code>lambda<\/code>s to use <code>map<\/code>, don't use <code>map<\/code>; it gains you nothing (a generator expression or list comprehension is going to run with the same speed or even faster in most cases where the mapping function isn't a CPython built-in). That said, in your particular case you don't need a <code>lambda<\/code>, for example, you can do:<\/p>\n\n<pre><code>from future_builtins import map  # Only on Py2; Py3 map is good\n\nfrom itertools import repeat\n\nmodel = ... initialize a model that will be passed to every call ...\nhnd = map(function_f, list_value, repeat(model))\n<\/code><\/pre>\n\n<p>or just use a generator expression (unless the function is a Python built-in implemented in C, <code>map<\/code> basically never gains you performance; if you don't want to think about whether <code>map<\/code> is appropriate, always using list comprehensions\/generator expresssions instead of <code>map<\/code> is a good idea):<\/p>\n\n<pre><code># Change outside parens to brackets, [], for list comp\nhnd = (function_f(x, model) for x in list_value)\n<\/code><\/pre>\n"},{"AnswerId":"35065503","CreationDate":"2016-01-28T15:27:41.310","ParentId":null,"OwnerUserId":"5788912","Title":null,"Body":"<p>By using zip, you can use map with such function. I am providing a sample code as you have asked to demonstrate<\/p>\n\n<pre><code>ls1 = [1,2,3,4,5]\nls2 = [1,2,3,4,5]\n&gt;&gt;&gt; def func(x,y):\n    return x+1,y+1\n\n&gt;&gt;&gt; map(lambda (v1, v2): func(v1, v2), zip(ls1, ls2))\n[(2, 2), (3, 3), (4, 4), (5, 5), (6, 6)]\n<\/code><\/pre>\n"}]}
{"QuestionId":35065524,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><tensorflow>","CreationDate":"2016-01-28T15:28:55.563","AcceptedAnswerId":null,"OwnerUserId":5852644.0,"Title":"Accuracy remains identical between Tensorflow runs","Body":"<p>I've been trying to utilise Tensorflow to gauge it's suitability for classifying data that I'm studying on a Huntington Disease project (not relevant to the problem, just providing context). Previously, I was using Support Vector Machines to classify my data, which were 'ok'. I'm hoping NNetworks are better.<\/p>\n\n<p>Loading the data is fine, no issues there. After reading the documentation for Tensorflow, and following some tutorials and examples online, I wrote the following to do a very simple network example with CSV data. The data I use in this provided example is the standard MNIST image database, but in CSV format.<\/p>\n\n<pre><\/pre>\n\n<p>All functional. Train, test and onehot arrays are all the correct size and populated with the correct values.<\/p>\n\n<p>The actual tensorflow code is where I am most likely going wrong (?).<\/p>\n\n<pre><\/pre>\n\n<p>Each output of this code always produces the exact same accuracy, and I can't figure out why.<\/p>\n\n<pre><\/pre>\n\n<p>I have went back and looked at the tutorials, and examples I learned from. The iris dataset (loaded in the same manner) produced a proper output that predicted accurately. Yet, this code with the MNIST CSV data does not.<\/p>\n\n<p>Any insight would be appreciated.<\/p>\n\n<p><strong>EDIT 1:<\/strong><\/p>\n\n<p>So I had a few minutes and tried some of your suggestions, to no avail. I also decided to go back and test things with the Iris CSV dataset, for comparisons sake. Output is slightly different after use sess.run(train_step, feed=dict={...}:<\/p>\n\n<pre><\/pre>\n\n<p>Values generally hover around this range, until it reaches Run 64, where it is locked at:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35074474","CreationDate":"2016-01-28T23:57:31.203","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>I think the problem might be that your train_step is not inside a sess.run try this. Also consider using mini-batching for the training like so:<\/p>\n\n<pre><code>for i in range(100):\n    for start, end in zip(range(0, len(x_train), 20), range(20, len(x_train), 20)):\n        sess.run(train_step, feed_dict={x: x_train[start:end], y_: y_train_onehot[start:end]})\n<\/code><\/pre>\n"}]}
{"QuestionId":35066828,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-28T16:30:26.077","AcceptedAnswerId":35067476.0,"OwnerUserId":2280020.0,"Title":"tensorflow exception from session initializing one variable?","Body":"<p>I am getting an exception from tensorflow with this example code:<\/p>\n\n<pre><\/pre>\n\n<p>It gives<\/p>\n\n<p>UnboundLocalError: local variable 'opts' referenced before assignment<\/p>\n\n<p>Is this a bug, or am I doing something wrong? Maybe the example is too simple and Tensorflow expected more computations to be set up?<\/p>\n","answers":[{"AnswerId":"35067476","CreationDate":"2016-01-28T16:59:52.347","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I think the error is in your code, but that's an unhelpful error message from the <code>tf.Session<\/code> constructor.<\/p>\n\n<p>The correct way to initialize the variable is:<\/p>\n\n<pre><code>sess = tf.Session()  # Note no argument.\nsess.run(init_op)\n<\/code><\/pre>\n"}]}
{"QuestionId":35072754,"AnswerCount":1,"Tags":"<python><macos><tensorflow>","CreationDate":"2016-01-28T21:49:44.557","AcceptedAnswerId":35077111.0,"OwnerUserId":4079577.0,"Title":"ImportError: cannot import name 'compile'","Body":"<p>I'm trying to install python 3.4 and TensorFlow on my Mac (OSX El Capitan version 10.11.1).<br>\nI am following the instructions here  <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/get_started\/os_setup.html#pip_install\" rel=\"nofollow\">Pip Installation<\/a> <\/p>\n\n<p>I can do <\/p>\n\n<pre><\/pre>\n\n<p>but then this part gets difficult:<\/p>\n\n<pre><\/pre>\n\n<p>I put in my password;\nit gives me a traceback:<\/p>\n\n<pre><\/pre>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<pre><\/pre>\n\n<p>Did I miss any important information?<\/p>\n","answers":[{"AnswerId":"35077111","CreationDate":"2016-01-29T04:50:43.750","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The main reason for this error is that the package you are trying to install is a <em>Python 2<\/em> wheel file. (The <code>py2<\/code> in the URL is the clue to this Moreover, as Guy Coder points out in the comments, you're installing version 0.5.0 of TensorFlow, which did not support Python 3.)<\/p>\n\n<p>The best solution would be to install the <a href=\"http:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py3-none-any.whl\" rel=\"nofollow\">Python 3\u2013compatible package<\/a>:<\/p>\n\n<pre><code>$ sudo pip install --upgrade http:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.6.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p>(It appears that these links aren't easy to find on the website, so we'll fix that shortly.)<\/p>\n"}]}
{"QuestionId":35074549,"AnswerCount":5,"Tags":"<python><machine-learning><keras><data-science>","CreationDate":"2016-01-29T00:03:28.007","AcceptedAnswerId":35196851.0,"OwnerUserId":4984897.0,"Title":"How to load a model from an HDF5 file in Keras?","Body":"<p>How to load a model from an HDF5 file in Keras?<\/p>\n\n<p>What I tried:<\/p>\n\n<pre><\/pre>\n\n<p>The above code successfully saves the best model to a file named weights.hdf5. What I want to do is then load that model. The below code shows how I tried to do so:<\/p>\n\n<pre><\/pre>\n\n<p>This is the error I get:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43263973","CreationDate":"2017-04-06T19:17:07.037","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>If you stored the complete model, not only the weights, in the HDF5 file, then it is as simple as<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('model.h5')\n<\/code><\/pre>\n"},{"AnswerId":"53798402","CreationDate":"2018-12-16T00:23:13.280","ParentId":null,"OwnerUserId":"1323067","Title":null,"Body":"<p>According to official documentation\n <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-install-hdf5-or-h5py-to-save-my-models-in-keras\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-install-hdf5-or-h5py-to-save-my-models-in-keras<\/a><\/p>\n\n<p>you can do :<\/p>\n\n<p>first test if you have h5py installed by running the <\/p>\n\n<p>import h5py<\/p>\n\n<p>if you dont have errors while importing h5py you are good to save:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'\ndel model  # deletes the existing model\n\n# returns a compiled model\n# identical to the previous one\nmodel = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>If you need to install h5py <a href=\"http:\/\/docs.h5py.org\/en\/latest\/build.html\" rel=\"nofollow noreferrer\">http:\/\/docs.h5py.org\/en\/latest\/build.html<\/a><\/p>\n"},{"AnswerId":"41660061","CreationDate":"2017-01-15T10:23:14.463","ParentId":null,"OwnerUserId":"6469175","Title":null,"Body":"<p>See the following sample code on how to Build a basic Keras Neural Net Model, save Model (JSON) &amp; Weights (HDF5) and load them:<\/p>\n\n<pre><code># create model\nmodel = Sequential()\nmodel.add(Dense(X.shape[1], input_dim=X.shape[1], activation='relu')) #Input Layer\nmodel.add(Dense(X.shape[1], activation='relu')) #Hidden Layer\nmodel.add(Dense(output_dim, activation='softmax')) #Output Layer\n\n# Compile &amp; Fit model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit(X,Y,nb_epoch=5,batch_size=100,verbose=1)    \n\n# serialize model to JSON\nmodel_json = model.to_json()\nwith open(\"Data\/model.json\", \"w\") as json_file:\n    json_file.write(simplejson.dumps(simplejson.loads(model_json), indent=4))\n\n# serialize weights to HDF5\nmodel.save_weights(\"Data\/model.h5\")\nprint(\"Saved model to disk\")\n\n# load json and create model\njson_file = open('Data\/model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n\n# load weights into new model\nloaded_model.load_weights(\"Data\/model.h5\")\nprint(\"Loaded model from disk\")\n\n# evaluate loaded model on test data \n# Define X_test &amp; Y_test data first\nloaded_model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nscore = loaded_model.evaluate(X_test, Y_test, verbose=0)\nprint (\"%s: %.2f%%\" % (loaded_model.metrics_names[1], score[1]*100))\n<\/code><\/pre>\n"},{"AnswerId":"35196851","CreationDate":"2016-02-04T09:13:21.403","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p><code>load_weights<\/code> only sets the weights of your network. You still need to define its architecture before calling <code>load_weights<\/code>:<\/p>\n\n<pre><code>def create_model():\n   model = Sequential()\n   model.add(Dense(64, input_dim=14, init='uniform'))\n   model.add(LeakyReLU(alpha=0.3))\n   model.add(BatchNormalization(epsilon=1e-06, mode=0, momentum=0.9, weights=None))\n   model.add(Dropout(0.5)) \n   model.add(Dense(64, init='uniform'))\n   model.add(LeakyReLU(alpha=0.3))\n   model.add(BatchNormalization(epsilon=1e-06, mode=0, momentum=0.9, weights=None))\n   model.add(Dropout(0.5))\n   model.add(Dense(2, init='uniform'))\n   model.add(Activation('softmax'))\n   return model\n\ndef train():\n   model = create_model()\n   sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\n   model.compile(loss='binary_crossentropy', optimizer=sgd)\n\n   checkpointer = ModelCheckpoint(filepath=\"\/tmp\/weights.hdf5\", verbose=1, save_best_only=True)\n   model.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose=2, callbacks=[checkpointer])\n\ndef load_trained_model(weights_path):\n   model = create_model()\n   model.load_weights(weights_path)\n<\/code><\/pre>\n"},{"AnswerId":"54978213","CreationDate":"2019-03-04T07:09:34.483","ParentId":null,"OwnerUserId":"11122745","Title":null,"Body":"<p>I done in this way<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras_contrib.losses import import crf_loss\nfrom keras_contrib.metrics import crf_viterbi_accuracy\n\n# To save model\nmodel.save('my_model_01.hdf5')\n\n# To load the model\ncustom_objects={'CRF': CRF,'crf_loss': crf_loss,'crf_viterbi_accuracy':crf_viterbi_accuracy}\n\n# To load a persisted model that uses the CRF layer \nmodel1 = load_model(\"\/home\/abc\/my_model_01.hdf5\", custom_objects = custom_objects)\n<\/code><\/pre>\n"}]}
{"QuestionId":35078027,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-01-29T06:14:35.073","AcceptedAnswerId":36386287.0,"OwnerUserId":309830.0,"Title":"Implement MLP in tensorflow","Body":"<p>I want to implement the MLP model taught in <a href=\"https:\/\/www.coursera.org\/learn\/machine-learning\" rel=\"nofollow noreferrer\">https:\/\/www.coursera.org\/learn\/machine-learning<\/a>, using <em>tensorflow<\/em>. Here's implementation.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, it does not work. I think the definition for the layers are correct, but the problem is in the <em>cross_entropy<\/em>. If I use the first one, the one got commented out, the model converges quickly; but if I use the 2nd one, which I think\/hope is the translation of the previous equation, the model won't converge.<\/p>\n\n<p>If you want to take a look at the cost equation, you can find it at <a href=\"https:\/\/datascience.stackexchange.com\/questions\/10015\/implement-mlp-in-tensorflow?noredirect=1#comment9892_10015\">here<\/a>.<\/p>\n\n<h1>Update<\/h1>\n\n<p>I have implemented this same MLP model using <em>numpy<\/em> and <em>scipy<\/em>, and it works.<\/p>\n\n<p>In the tensorflow code, I added a <em>print<\/em> line in the training loop, and I found out that all the elements in  are <strong>nan<\/strong>...I think it is caused by arithmetic overflow or something alike.<\/p>\n","answers":[{"AnswerId":"36386287","CreationDate":"2016-04-03T13:56:46.900","ParentId":null,"OwnerUserId":"2220465","Title":null,"Body":"<p>It is likely 0*log(0) issue.<\/p>\n\n<p>Replacing <\/p>\n\n<pre><code>cross_entropy = tf.reduce_sum(- y * tf.log(y_) - (1 - y) * tf.log(1 - y_), 1)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>cross_entropy = tf.reduce_sum(- y * tf.log(tf.clip_by_value(y_, 1e-10, 1.0)) - (1 - y) * tf.log(tf.clip_by_value(1 - y_, 1e-10, 1.0)), 1)\n<\/code><\/pre>\n\n<p>Please see <a href=\"https:\/\/stackoverflow.com\/questions\/33712178\/tensorflow-nan-bug\">Tensorflow NaN bug?<\/a>.<\/p>\n"},{"AnswerId":"35082213","CreationDate":"2016-01-29T10:29:33.747","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>The problem I think is that nn.sigmoid_cross_entropy_with_logits expects unormalized results, where as the function you replace it with <code>cross_entropy = tf.reduce_sum(- y * tf.log(y_) - (1 - y) * tf.log(1 - y_), 1)<\/code><\/p>\n\n<p>Expects y_ to be normalized (by the sigmoid) between 0 and 1<\/p>\n\n<p>try replacing<\/p>\n\n<pre><code>y_ = tf.matmul(h1, W_out)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>y_ = tf.nn.sigmoid(tf.matmul(h1, W_out))\n<\/code><\/pre>\n"}]}
{"QuestionId":35081722,"AnswerCount":1,"Tags":"<python><macos><deep-learning><caffe>","CreationDate":"2016-01-29T10:05:31.683","AcceptedAnswerId":null,"OwnerUserId":5724595.0,"Title":"Trouble installing Caffe","Body":"<p>I have been having hard time installing Caffe on my Mac OS X 10.11.3.\nI have managed to run \"make all\", \"make test\" successfully, but \"make runtest\" has been causing the following error for some reason.<\/p>\n\n<pre><\/pre>\n\n<p>I don't know how to install that library separately, Please help...<\/p>\n","answers":[{"AnswerId":"35405419","CreationDate":"2016-02-15T09:17:46.607","ParentId":null,"OwnerUserId":"5929065","Title":null,"Body":"<p>I got this problem and solved by this:<\/p>\n\n<p>After make test in <code>$CAFFE_ROOT<\/code>:<\/p>\n\n<pre><code>cp -a .build_release\/lib\/. \/usr\/local\/lib\/\n<\/code><\/pre>\n\n<p>The answer comes from: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/3628\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/3628<\/a><\/p>\n"}]}
{"QuestionId":35088440,"AnswerCount":2,"Tags":"<ios><swift><deep-learning><caffe>","CreationDate":"2016-01-29T15:45:44.977","AcceptedAnswerId":35682171.0,"OwnerUserId":654005.0,"Title":"Load UIImage to float array for DeepLearningKit usage","Body":"<p>I'd like to use the <a href=\"https:\/\/github.com\/DeepLearningKit\/DeepLearningKit\" rel=\"nofollow\">DeepLearningKit<\/a> for iOS. I want to use UIImage objects to be classified. The sample application only uses a float array loaded from a json file. Thus I have to create the bitmap representation of the UIImage as a float array and use this for the classify-method. <\/p>\n\n<p>Can anybody help me on that? Is there a way to create a bitmap representation for UIImage? Moreover I have to swap the channels from RGB to BGR. <\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"35682171","CreationDate":"2016-02-28T11:53:02.630","ParentId":null,"OwnerUserId":"239834","Title":null,"Body":"<p>Have added an extension to UIImage that allows setting and getting RGB(A) pixels directly - key methods: <\/p>\n\n<pre><code>public func setPixelColorAtPoint(point:CGPoint, color: RawColorType) -&gt; UIImage? \nfunc getPixelColorAtLocation(point:CGPoint)-&gt;UIColor?\n<\/code><\/pre>\n\n<p>where RawColorType is defined as<\/p>\n\n<pre><code>public typealias RawColorType = (newRedColor:UInt8, newgreenColor:UInt8, newblueColor:UInt8,  newalphaValue:UInt8)\n<\/code><\/pre>\n\n<p>This way you should be able to convert back and forth between bitmap representation and UIImage. Wrote a blog post that gives some more context: <a href=\"http:\/\/deeplearningkit.org\/tutorials-for-ios-os-x-and-tvos\/tutorial-image-handling-in-deeplearningkit\/\" rel=\"nofollow\">http:\/\/deeplearningkit.org\/tutorials-for-ios-os-x-and-tvos\/tutorial-image-handling-in-deeplearningkit\/<\/a><\/p>\n"},{"AnswerId":"42094423","CreationDate":"2017-02-07T16:00:23.567","ParentId":null,"OwnerUserId":"2572084","Title":null,"Body":"<p>I've write a function to convert a image file to Caffe blob on iOS platform. You can find it <a href=\"https:\/\/github.com\/solrex\/caffe-mobile\/blob\/master\/examples\/ios\/simple\/CaffeSimple\/ImageReader.mm\" rel=\"nofollow noreferrer\">here<\/a>. I hope it will help you.<\/p>\n\n<p>Code snippet:<\/p>\n\n<pre><code>\/\/ Convert Bitmap (channels*width*height) to Matrix (width*height*channels)\n\/\/ Remove alpha channel\nint input_channels = input_layer-&gt;channels();\nLOG(INFO) &lt;&lt; \"image_channels:\" &lt;&lt; image_channels &lt;&lt; \" input_channels:\" &lt;&lt; input_channels;\nif (input_channels == 3 &amp;&amp; image_channels != 4) {\n    LOG(ERROR) &lt;&lt; \"image_channels input_channels not match.\";\n    return false;\n} else if (input_channels == 1 &amp;&amp; image_channels != 1) {\n    LOG(ERROR) &lt;&lt; \"image_channels input_channels not match.\";\n    return false;\n}\n\nfloat *input_data = input_layer-&gt;mutable_cpu_data();\n\nfor (size_t h = 0; h &lt; height; h++) {\n    for (size_t w = 0; w &lt; width; w++) {\n        for (size_t c = 0; c &lt; input_channels; c++) {\n            \/\/ OpenCV use BGR instead of RGB\n            size_t cc = c;\n            if (input_channels == 3) {\n                cc = 2 - c;\n            }\n            \/\/ Convert uint8_t to float\n            input_data[c*width*height + h*width + w] = \n                static_cast&lt;float&gt;(result[h*width*image_channels + w*image_channels + cc]);\n            if (mean.size() == input_channels) {\n                input_data[c*width*height + h*width + w] -= mean[c];\n            }\n        }\n    }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":35089956,"AnswerCount":2,"Tags":"<neural-network><embedding><keras><recurrent-neural-network>","CreationDate":"2016-01-29T16:57:09.507","AcceptedAnswerId":null,"OwnerUserId":963921.0,"Title":"How to use the Embedding Layer for Recurrent Neural Network (RNN) in Keras","Body":"<p>I'm rather new to Neural Networks and the Keras Library and I'm wondering how I can use the Embedding Layer as described <a href=\"http:\/\/keras.io\/layers\/embeddings\/\">here<\/a> to mask my input data from a 2D tensor to a 3D tensor for a <a href=\"http:\/\/keras.io\/layers\/recurrent\/\">RNN<\/a>.<\/p>\n\n<p>Say my timeseries data looking as follows (with an increasing time):<\/p>\n\n<pre><\/pre>\n\n<p>Now, say I would want to give the RNN the last 2 feature vectors in order to predict the feature vector for time t+1.<\/p>\n\n<p>Currently (without the Embedding Layer), I am creating the required 3D tensor with shape (nb_samples, timesteps, input_dim) myself (as in this example <a href=\"http:\/\/danielhnyk.cz\/predicting-sequences-vectors-keras-using-rnn-lstm\/\">here<\/a>).<\/p>\n\n<p>Related to my example, the final 3D Tensor would then look as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and Y_train:<\/p>\n\n<pre><\/pre>\n\n<p>My model looks as follows (adapted to the simplified example above):<\/p>\n\n<pre><\/pre>\n\n<p>And finally, my question would be, how can I avoid doing those 2D tensor to 3D tensor reshaping myself and use the Embedding layer instead? I guess after model = sequential() I would have to add something like:<\/p>\n\n<pre><\/pre>\n\n<p>Probably the answer is rather simple, I'm simply confused by the documentation of the embedding layer.<\/p>\n","answers":[{"AnswerId":"35305724","CreationDate":"2016-02-10T02:14:24.737","ParentId":null,"OwnerUserId":"5457186","Title":null,"Body":"<p>You can you it as follows:<\/p>\n\n<p>Note:<\/p>\n\n<ol>\n<li><p>I generated some X and y as 0s just to give you some idea of the input structure.<\/p><\/li>\n<li><p>If you are having a multi class y_train, you will need to binarize.<\/p><\/li>\n<li><p>You might need to add padding if you have data of various length.<\/p><\/li>\n<li><p>If I understood correctly about predicting at time t+1, you might want to look at Sequence to Sequence learning.<\/p><\/li>\n<\/ol>\n\n<p>Try something like:<\/p>\n\n<pre><code>hidden_neurons = 4\nnb_classes =3\nembedding_size =10\n\nX = np.zeros((128, hidden_neurons), dtype=np.float32)\ny = np.zeros((128, nb_classes), dtype=np.int8)\n\n\nmodel = Sequential()\nmodel.add(Embedding(hidden_neurons, embedding_size))\nmodel.add(SimpleRNN(hidden_neurons, return_sequences=False)) \nmodel.add(Dense(nb_classes))\nmodel.add(Activation(\"softmax\"))\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop', class_mode=\"categorical\")\nmodel.fit(X, y, batch_size=1, nb_epoch=1)\n<\/code><\/pre>\n"},{"AnswerId":"40666389","CreationDate":"2016-11-17T22:42:30.447","ParentId":null,"OwnerUserId":"1835885","Title":null,"Body":"<p>From what I know so far, the Embedding layer seems to be more or less for dimensionality reduction like <a href=\"https:\/\/en.wikipedia.org\/wiki\/Word_embedding\" rel=\"nofollow noreferrer\">word embedding<\/a>. So in this sense it does not seem applicable as general reshaping tool.<\/p>\n\n<p>Basicaly if you have a mapping of words to integers like {car: 1, mouse: 2 ... zebra: 9999}, your input text would be vector of words represented by they integer id, like [1, 2, 9999 ...], which would mean [car, mouse, zebra ...]. But it seems to be efficient to map words to vectors of real numbers with length of the vocabular, so if your text has 1000 unique words, you would map each word to vector of real numbers with length of 1000. I am not sure but I think that it mostly represents a weigh of how similar a meaning of a word is to all other words, but I am not sure that is right and wheher there are other ways to embed words.<\/p>\n"}]}
{"QuestionId":35090591,"AnswerCount":1,"Tags":"<gpu><tensorflow><bazel>","CreationDate":"2016-01-29T17:34:32.153","AcceptedAnswerId":35104015.0,"OwnerUserId":2116766.0,"Title":"Is there a way to force Bazel to run tests serially","Body":"<p>By default, Bazel runs tests in a parallel fashion to speed things up. However, I have a resource (GPU) that can't handle parallel jobs due to the GPU memory limit. Is there a way to force Bazel to run tests in a serial, i.e., non-parallel way? <\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"35104015","CreationDate":"2016-01-30T16:17:00.630","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p><code>--jobs 1<\/code> will limit the number of parallel jobs Bazel runs to 1.<\/p>\n\n<p>You can also modify the test targets and add <code>tags = [\"exclusive\"]<\/code> to prevent specific test to run in parallel (see <a href=\"http:\/\/bazel.io\/docs\/test-encyclopedia.html\" rel=\"noreferrer\">http:\/\/bazel.io\/docs\/test-encyclopedia.html<\/a>).<\/p>\n"}]}
{"QuestionId":35091074,"AnswerCount":0,"Tags":"<python><multithreading><theano><deep-learning>","CreationDate":"2016-01-29T18:02:08.580","AcceptedAnswerId":null,"OwnerUserId":3302611.0,"Title":"Python threading Queue is very slow while passing queue reference","Body":"<p>I'm building a Deeplearning model which optimizes using SGD. I have a python threading queue and I hold at a time 10 minibatches inside it. When I call the gradient descent step function, I pop out one minibatch and send it to it. But it takes a long time, and I'm not sure why. I'm trying to pass the reference of the queue to the function. The following are the small snippets of code which shows how I do it. The overhead being almost 100 times slower than other operations.<\/p>\n\n<pre><\/pre>\n\n<p>The  print 'time outside set_sharedval in step :', (time.time() - time_step) statement from the step function outputs 200 seconds required to do this part of the code, which really has nothing except loading the queue and popping a value from it makes me think there's something wrong with how I construct and pass the data loader. ALL the necessary code flow in between functions are in bolds for easy reading.<\/p>\n","answers":[]}
{"QuestionId":35091806,"AnswerCount":1,"Tags":"<lua><machine-learning><torch><conv-neural-network>","CreationDate":"2016-01-29T18:45:19.437","AcceptedAnswerId":35095100.0,"OwnerUserId":2452617.0,"Title":"torch7: Error in fine-tuning a CNN","Body":"<p>I am trying to fine-tune a pre-trained <strong><a href=\"https:\/\/github.com\/jhjin\/overfeat-torch\" rel=\"nofollow\">OverFeat<\/a><\/strong> conv-net (CNN) on a dataset of images of multi-class faces. My training-testing lua script is based on the tutorial given <a href=\"https:\/\/github.com\/torch\/tutorials\/tree\/master\/2_supervised\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>I first created a script and tried to test my script by training-testing on a subset of the <em>ImageNet<\/em> dataset. After resolving some issues, it was working as expected without any errors. But, then I made slight changes to the script, like adding few layers in the neural net, changing the input files and labels, to adapt it to the new dataset. My updated model to be fine-tuned is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I am using <strong>nn.ClassNLLCriterion()<\/strong> to train my network. But upon training, I am facing the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am unable to resolve it as I'm finding this error non-informative, and thus, difficult to trace and debug. Can someone please help me resolve this issue?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"35095100","CreationDate":"2016-01-29T22:21:27.273","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Here when you do a <a href=\"https:\/\/github.com\/torch\/nn\/blob\/3585e82\/ClassNLLCriterion.lua#L35-L42\" rel=\"nofollow\">forward on your criterion<\/a> the second argument (= target) has not the right type, hence this error.<\/p>\n\n<p>Since you apparently work with mini-batches you are supposed to pass a 1D torch long tensor of size N = mini-batch size (in non batch mode it could be a <code>number<\/code> or a single element 1D long tensor).<\/p>\n\n<p><em>Note: at training time, right before the <code>nn.ClassNLLCriterion<\/code>, you should use a <code>nn.LogSoftMax()<\/code> and not a <code>nn.SoftMax()<\/code> layer. As an alternative there is a built-in layer that combines both: <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/criterion.md#nn.CrossEntropyCriterion\" rel=\"nofollow\"><code>nn.CrossEntropyCriterion<\/code><\/a>.<\/em><\/p>\n"}]}
{"QuestionId":35093588,"AnswerCount":1,"Tags":"<python-2.7><numpy><theano><deep-learning><conv-neural-network>","CreationDate":"2016-01-29T20:36:14.893","AcceptedAnswerId":35132068.0,"OwnerUserId":1775763.0,"Title":"Visualize output of each layer in theano Convolutional MLP","Body":"<p>I am reading <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#convolutional-neural-networks-lenet\" rel=\"nofollow\">Convolutional Neural Networks tutorial<\/a>. I want to visualize output of each layer after model is trained. For example in function \"evaluate_lenet5\" I want to pass a instance (which is an image) to the network and see the output of each layer and the class that trained Neural Network set for the input. I thought it may be easy as doing a dot product on an image and Weight vector of each layer, but it did not work at all.<\/p>\n\n<p>I have objects of each layer as :<\/p>\n\n<pre><\/pre>\n\n<p>So can you suggest a way to visualize a sample of processing an image step by step after the neural network is trained?<\/p>\n","answers":[{"AnswerId":"35132068","CreationDate":"2016-02-01T13:43:39.687","ParentId":null,"OwnerUserId":"5181464","Title":null,"Body":"<p>This isn't so hard.\nIf you are using the same class definition of <strong>LeNetConvPoolLayer<\/strong> from the theano deep-learning tutorial, then you just need to compile a function with <code>x<\/code> as input and <code>[LayerObject].output<\/code> as output (where <em>LayerObject<\/em> can be any layer object like <em>layer0<\/em>, <em>layer1<\/em> etc. whichever layer you want to visualize.<\/p>\n\n<blockquote>\n  <p>vis_layer1 = function([x], [layer1.output])<\/p>\n<\/blockquote>\n\n<p>Pass a (or many) sample (exactly how you fed the input tensor while training) and you'll get the output of that particular layer for which your function is compiled for.<\/p>\n\n<p><em>Note:<\/em> In this way you will get the outputs in the exact same <em>shape<\/em> the model has used in calculation. However you can <em>reshape<\/em> it as you want by reshaping the output variable like <code>layer1.output.flatten(n)<\/code>.<\/p>\n"}]}
{"QuestionId":35094899,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-29T22:06:41.173","AcceptedAnswerId":35095052.0,"OwnerUserId":5540330.0,"Title":"TensorFlow operator overloading","Body":"<p>What is the difference between <\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>in TensorFlow? What would be different in your computation graph when you construct your graph with  instead of ? <\/p>\n\n<p>More generally, are   or other operations overloaded for tensors?<\/p>\n","answers":[{"AnswerId":"35095052","CreationDate":"2016-01-29T22:17:44.330","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If at least one of <code>x<\/code> or <code>y<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a> object, the expressions <code>tf.add(x, y)<\/code> and <code>x + y<\/code> are equivalent. The main reason you might use <code>tf.add()<\/code> is to specify an explicit <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#add\" rel=\"noreferrer\"><code>name<\/code><\/a> keyword argument for the created op, which is not possible with the overloaded operator version.<\/p>\n\n<p>Note that if neither <code>x<\/code> nor <code>y<\/code> is a <code>tf.Tensor<\/code>&mdash;for example if they are NumPy arrays&mdash;then <code>x + y<\/code> will not create a TensorFlow op. <code>tf.add()<\/code> always creates a TensorFlow op and converts its arguments to <code>tf.Tensor<\/code> objects. Therefore, if you are writing a library function that might accept both tensors and NumPy arrays, you might prefer to use <code>tf.add()<\/code>.<\/p>\n\n<p>The following operators are overloaded in the TensorFlow Python API:<\/p>\n\n<ul>\n<li><code>__neg__<\/code> (unary <code>-<\/code>)<\/li>\n<li><code>__abs__<\/code> (<code>abs()<\/code>)<\/li>\n<li><code>__invert__<\/code> (unary <code>~<\/code>)<\/li>\n<li><code>__add__<\/code> (binary <code>+<\/code>)<\/li>\n<li><code>__sub__<\/code> (binary <code>-<\/code>)<\/li>\n<li><code>__mul__<\/code> (binary elementwise <code>*<\/code>)<\/li>\n<li><code>__div__<\/code> (binary <code>\/<\/code> in Python 2)<\/li>\n<li><code>__floordiv__<\/code> (binary <code>\/\/<\/code> in Python 3)<\/li>\n<li><code>__truediv__<\/code> (binary <code>\/<\/code> in Python 3)<\/li>\n<li><code>__mod__<\/code> (binary <code>%<\/code>)<\/li>\n<li><code>__pow__<\/code> (binary <code>**<\/code>)<\/li>\n<li><code>__and__<\/code> (binary <code>&amp;<\/code>)<\/li>\n<li><code>__or__<\/code> (binary <code>|<\/code>)<\/li>\n<li><code>__xor__<\/code> (binary <code>^<\/code>)<\/li>\n<li><code>__lt__<\/code> (binary <code>&lt;<\/code>)<\/li>\n<li><code>__le__<\/code> (binary <code>&lt;=<\/code>)<\/li>\n<li><code>__gt__<\/code> (binary <code>&gt;<\/code>)<\/li>\n<li><code>__ge__<\/code> (binary <code>&gt;=<\/code>)<\/li>\n<\/ul>\n\n<p>Please note, <code>__eq__<\/code> ( binary <code>==<\/code> ) is <strong>not<\/strong> overloaded. <code>x == y<\/code> will simply return a Python boolean whether <code>x<\/code> and <code>y<\/code> refer to the same tensor. You need to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/equal\" rel=\"noreferrer\"><code>tf.equal()<\/code><\/a> explicitly to check for element-wise equality. Same goes for not equal, <code>__ne__<\/code> ( binary <code>!=<\/code> ).<\/p>\n"},{"AnswerId":"52961139","CreationDate":"2018-10-24T04:31:46.723","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Mrry nicely explained that there is no real difference. I will just add when using <code>tf.add<\/code> is beneficial.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/add\" rel=\"nofollow noreferrer\">tf.add<\/a> has one important parameter which is <code>name<\/code>. It allows you to name the operation in a graph which will be visible in tensorboard. So my rule of thumb, if it will be beneficial to name an operation in tensorboard, I use <code>tf.<\/code> equivalent, otherwise I go for brevity and use overloaded version.<\/p>\n"}]}
{"QuestionId":35095403,"AnswerCount":3,"Tags":"<python><git><theano>","CreationDate":"2016-01-29T22:43:58.297","AcceptedAnswerId":35096172.0,"OwnerUserId":3413239.0,"Title":"Uninstalling python library installed from github clone","Body":"<p>Can you tell me how should I uninstall Theano which I developed by cloning the Theano Git Repository?<\/p>\n\n<p>I have:<\/p>\n\n<pre><\/pre>\n\n<p> returns <\/p>\n\n<p> gives:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35096172","CreationDate":"2016-01-29T23:51:38.123","ParentId":null,"OwnerUserId":"346478","Title":null,"Body":"<p>You are probably not in the directory, where you were installing Theano.<\/p>\n\n<p>I thinky, you have following options::<\/p>\n\n<pre><code>$ pip uninstall Theano\/\n<\/code><\/pre>\n\n<p>This works for me (tested with <code>moto<\/code> as installing Theano requires some compiled packages which are\nfailing on my machine).<\/p>\n\n<p>Another option is to <code>cd<\/code> to the Theano dir and uninstal from here::<\/p>\n\n<pre><code>$ cd Theano\n$ pip uninstall theano\n<\/code><\/pre>\n\n<p>If none of this works, last option is to manually remove it from system by deleting the\nTheano.egg-link file.<\/p>\n\n<p>Assuming you are using virtualenv named <code>testtheano<\/code>::<\/p>\n\n<pre><code>$ cd ~\/.virtualenvs\/testtheano\/lib\/python2.7\/site-packages\n$ rm Theano.egg-link\n<\/code><\/pre>\n\n<p>If it installed some scripts, delete them too:<\/p>\n\n<pre><code>$ cd ~\/.virtualenvs\/testtheano\/bin\n$ rm &lt;all-the-theano-scripts&gt;\n<\/code><\/pre>\n"},{"AnswerId":"49892101","CreationDate":"2018-04-18T05:49:41.523","ParentId":null,"OwnerUserId":"6179398","Title":null,"Body":"<p>You can simply remove the Theano directory and egg file from \/opt\/anaconda\/lib\/pythonXX\/site-packages.<\/p>\n"},{"AnswerId":"35095998","CreationDate":"2016-01-29T23:32:27.280","ParentId":null,"OwnerUserId":"405303","Title":null,"Body":"<p>Find out where it's installed:<\/p>\n\n<pre><code>$ which Theano\n<\/code><\/pre>\n\n<p>and remove it. <\/p>\n\n<p>The dependencies shouldn't cause you any harm and can be left in place. If you want to remove them, you can, but be careful, since they may well be required by other packages.<\/p>\n"}]}
{"QuestionId":35096557,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-01-30T00:35:37.770","AcceptedAnswerId":54247965.0,"OwnerUserId":1978504.0,"Title":"Transfer learning with inception model in Tensorflow (python)","Body":"<p>How can I load a .pb protobuf model and then tweak the network as needed (specially the outer layers) in order to train a new model for completely different classes? Effectively doing transfer learning?<\/p>\n\n<p>I want to do something like <a href=\"https:\/\/stackoverflow.com\/questions\/34945554\/how-to-set-layer-wise-learning-rate-in-tensorflow\">these<\/a> (i.e. train the outer layers with a bigger learning rate than the inner layers) among other things, so I need a way to not only load the graph with the variables, but to alter the network's structure and hyperparameters too.<\/p>\n\n<p>If anyone has an example to follow with the inception model, it would be amazing!<\/p>\n\n<p>My question is very similar to <a href=\"https:\/\/stackoverflow.com\/questions\/34091929\/how-to-fine-tune-a-model-in-tensorflow#\">this one<\/a>.<\/p>\n\n<p>I've searched all over the internet (TF docs, Github, StackOverflow, Google...) but I can't seam to find something useful for a novice.<\/p>\n\n<p>Thanks a lot!<\/p>\n","answers":[{"AnswerId":"54247965","CreationDate":"2019-01-18T04:56:26.327","ParentId":null,"OwnerUserId":"4524241","Title":null,"Body":"<p>This is the updated tutorial from official Tensorflow website <a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining<\/a><\/p>\n\n<p>They use the pre trained Inception V3 model and everything works fine. You can change the dataset folder to your own dataset<\/p>\n"},{"AnswerId":"35139305","CreationDate":"2016-02-01T20:03:34.323","ParentId":null,"OwnerUserId":"5530615","Title":null,"Body":"<p>tf.import_graph_def() is the function for loading a GraphDef:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#import_graph_def\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#import_graph_def<\/a><\/p>\n\n<p>Hopefully once imported, you can make the modifications to the graph you need. It would be easier, though, to modify the Python code that generated the graph in the first place, if you have access to that.<\/p>\n"}]}
{"QuestionId":35100458,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-01-30T10:10:58.863","AcceptedAnswerId":null,"OwnerUserId":3130081.0,"Title":"caffe LayerSetUp and Reshape?","Body":"<p>I am reading caffe source code now.<br>\nI am confused about  and  methods.<br>\nSome layers have both of these methods, others have one or none...\nWhy? Can anyone explain this to me?<\/p>\n","answers":[{"AnswerId":"35104552","CreationDate":"2016-01-30T17:09:34.793","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><code>LayerSetUp<\/code> is called once when loading the net. It's aim is to<br>\n(a) Verify the layer has exactly the right number of input\/output blobs<br>\n(b) Read the parameters of the layer from the prototxt<br>\n(c) Initialize internal parameters<\/p>\n\n<p>On the other hand, <code>Reshape<\/code> is used to allocate memory for parameters and output blobs and can be called even after the net was setup. For instance, it is common for detection networks to change the input shape, thus <code>Reshape<\/code>ing all consequent blobs.<\/p>\n"}]}
{"QuestionId":35104026,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-01-30T16:17:42.460","AcceptedAnswerId":35104485.0,"OwnerUserId":5540330.0,"Title":"Adding two Tensors with different ranks on TensorFlow","Body":"<p>I tried following on TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>I expected an error like \"Two Tensors must have the same rank.\" However, output was a 10-by-10 zero matrix. Why do you think this happened?<\/p>\n","answers":[{"AnswerId":"35104485","CreationDate":"2016-01-30T17:02:49.160","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Because of <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#add\" rel=\"nofollow\">broadcasting<\/a>. You have one horizontal and one vertical vectors, once added they create a 10x10 matrix. The same applies for<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.Variable(tf.zeros([10, 1]))\nb = tf.Variable(tf.zeros([1, 10]) )\nc = a + b\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    print sess.run(c)\n<\/code><\/pre>\n"}]}
{"QuestionId":35104616,"AnswerCount":1,"Tags":"<theano><theano-cuda>","CreationDate":"2016-01-30T17:15:35.017","AcceptedAnswerId":35111904.0,"OwnerUserId":5861624.0,"Title":"Running Theano with GPU on windows, python 2.7, cuda 7.5","Body":"<p>I have been running Theano without GPU on py 2.7 for quite some time now. Then decided to check out GPU support. So followed all steps given in here:\n<a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-getting-started-guide-for-microsoft-windows\/index.html#abstract\" rel=\"nofollow noreferrer\">http:\/\/docs.nvidia.com\/cuda\/cuda-getting-started-guide-for-microsoft-windows\/index.html#abstract<\/a><\/p>\n\n<p>Then opened the device query sln in msvc 2013 and built it (as mentioned in the above link to check if installation was successful). On run it gives me an error (attached screen shot).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/b7m73.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/b7m73.jpg\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>A similar error comes up when I run the starter program with gpu in py 2.7 with Theanorc file contents:<\/p>\n\n<pre><\/pre>\n\n<p>received no errors for nvcc -V, and cl.exe.\nOS - win7 SP1\nNVIDIA GeForce GT650M<\/p>\n\n<p>Can anyone please help?<\/p>\n","answers":[{"AnswerId":"35111904","CreationDate":"2016-01-31T08:40:27.057","ParentId":null,"OwnerUserId":"5860627","Title":null,"Body":"<p>I got similar error when my video card driver mismatched with the CUDA Toolkit. <\/p>\n\n<p>Make sure you are using video driver included in CUDA installer. If you are not sure, reinstall latest CUDA Toolkit installer and update your video driver with it.<\/p>\n"}]}
{"QuestionId":35106101,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><conv-neural-network>","CreationDate":"2016-01-30T19:28:17.240","AcceptedAnswerId":35110972.0,"OwnerUserId":5507905.0,"Title":"Why does my TensorFlow Convnet (attempted) training result in NaN gradients?","Body":"<p>I have modified the MNIST (28x28) Convnet tutorial code to accept larger images (150x150). But when I try to train I receive this error (see end for full stack trace):<\/p>\n\n<pre><\/pre>\n\n<p>Here's my code. Worryingly, I get the same error when using image data from disk as when generating noisy red\/blue\/green squares and trying to classify them by colour. The code that generates RGB data is different to the code that scans directories for JPG image data. So either there is some systematically wrong way I am loading in my own data, or there is something wrong with my proposed architecture. (I could include these modules but I worry it might make this post unreadably long.)<\/p>\n\n<p>EDIT: I have tried this same code with moderately larger images (30x30) and it DOES work. So perhaps the error is related to the very high dimensionality of a (150x150) problem?<\/p>\n\n<pre><\/pre>\n\n<p>Error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35110972","CreationDate":"2016-01-31T06:15:19.370","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>One likely source of trouble is the <code>tf.log(y_conv)<\/code>, which will emit NaN values for any zeroes in <code>y_conv<\/code>. The <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits()<\/code><\/a> operator offers a numerically stable (and more efficient) version of your loss calculation. The following should work better:<\/p>\n\n<pre><code>logits = tf.matmul(h_fc1_drop, W_fc2) + b_fc2\ny_conv = tf.nn.softmax(logits)\n\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, y_)\n\ncorrect_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))\n<\/code><\/pre>\n"},{"AnswerId":"35106509","CreationDate":"2016-01-30T20:04:57.120","ParentId":null,"OwnerUserId":"5540330","Title":null,"Body":"<p>If you are getting NaN values, it is probably because your learning rate is high relative to your loss function. I would suggest you to try lower learning rates like <code>1e-5<\/code> or <code>1e-6<\/code>. This used to solve problem in my case.<\/p>\n"}]}
{"QuestionId":35108813,"AnswerCount":5,"Tags":"<caffe>","CreationDate":"2016-01-31T00:12:40.260","AcceptedAnswerId":null,"OwnerUserId":5452008.0,"Title":"Installing Caffe without CUDA: fatal error: cublas_v2.h No such file (Fedora23)","Body":"<p>I try to install caffe under Fedora23 following these instructions <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow\">enter link description here<\/a>. And I uncommented the  in the Makefile.config file, to install caffe without CUDA. I installed the dependencies including protobuf. However, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>As I understand this is a CUDA related error, but I want to install Caffe without CUDA. Do I misunderstand something?<\/p>\n","answers":[{"AnswerId":"53854641","CreationDate":"2018-12-19T15:42:45.350","ParentId":null,"OwnerUserId":"6543500","Title":null,"Body":"<p>In contrast to Anoop's answer, this is what worked for me:<\/p>\n\n<ul>\n<li><p>MacOS 10.12 Sierra, no CUDA, CPU_ONLY<\/p><\/li>\n<li><p>add <code>#define CPU_ONLY<\/code> to top of <code>device_alternate.hpp<\/code><\/p><\/li>\n<li><p>This will enforce the program to go with <code>#ifdef CPU_ONLY<\/code> and thus will not raise the <code>cublas_v2.h<\/code> error.<\/p><\/li>\n<\/ul>\n\n<p>That said, when I did this, the next error I got was <\/p>\n\n<pre><code>caffe\/build$ make\n.\n.\n.\nbuild\/install\/include\/caffe\/util\/mkl_alternate.hpp:14:10: fatal error: 'cblas.h' file not found\n#include &lt;cblas.h&gt;\n<\/code><\/pre>\n\n<p>which I fixed with adding <code>-DCMAKE_CXX_FLAGS=-I\/usr\/local\/opt\/openblas\/include<\/code> to my <code>cmake<\/code> command.<\/p>\n\n<p>But then the new error I got was <\/p>\n\n<pre><code>make[2]: *** No rule to make target ..\/caffe\/build\/lib\/libcaffe.so', needed by `libdeeplcd.dylib'.\n<\/code><\/pre>\n\n<p>And so the adventure continues...!<\/p>\n"},{"AnswerId":"57921978","CreationDate":"2019-09-13T10:33:48.483","ParentId":null,"OwnerUserId":"4949887","Title":null,"Body":"<p>You might have to uncomment <code>CPU_ONLY := 1<\/code> in your Makefile.config file before you compile the make all.\nhope this helps<\/p>\n"},{"AnswerId":"42672564","CreationDate":"2017-03-08T13:25:00.007","ParentId":null,"OwnerUserId":"768183","Title":null,"Body":"<p>I use Nvidia TX1 and met this problem.\nI solve by this instruction:<\/p>\n\n<pre><code>  export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:\/usr\/local\/cuda-7.0\/targets\/armv7-linux-gnueabihf\/include\/\n<\/code><\/pre>\n\n<p>You can change the directory to try it.<\/p>\n\n<p>Hope it works.<\/p>\n\n<p>Sam<\/p>\n"},{"AnswerId":"38107217","CreationDate":"2016-06-29T18:14:45.573","ParentId":null,"OwnerUserId":"2892909","Title":null,"Body":"<p>I faced the same problem and searched for hours only to realize I had to be more careful during installation!\nFirst thing to uncomment CPU_ONLY := 1 in Makefile.config and not Makefile.config.example. This file is in the directory where you git cloned caffe.\nSecond is to change cmake options. After adding OpenBLAS libraries to .bashrc and sourcing it, execute the following -<\/p>\n\n<pre><code>cd \/..\/..\/src\/caffe \/\/where you have git cloned\nmkdir build &amp;&amp; cd build\ncmake -D CPU_ONLY=ON ..\ncd ..\nsudo make clean\nsudo make all -j2\nsudo make install -j2\nsudo make runtest -j2\n<\/code><\/pre>\n\n<p>Third step is to define CPU_ONLY=1 in your header folder. I use Qt (v5.7) and I add this in the .pro file<\/p>\n\n<pre><code>DEFINES += CPU_ONLY=1\n<\/code><\/pre>\n\n<p>Caffe libraries and your code may not know if CPU_ONLY flag is defined or not when your code is compiled. This should solve your problem.<\/p>\n"},{"AnswerId":"35286011","CreationDate":"2016-02-09T07:15:39.857","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>In .\/include\/caffe\/util\/device_alternate.hpp<\/p>\n\n<pre><code>line  4: #ifdef CPU_ONLY\nline 32: #else\nline 34: #include &lt;cublas_v2.h&gt;\nline 99: #endif\n<\/code><\/pre>\n\n<p>If CPU_ONLY flag was enabled, the else condition shouldn't have worked. <\/p>\n\n<p>Make sure that you are editing the Makefile.config instead of Makefile.config.example file.<\/p>\n\n<p>Try debugging by editing this part of the code to get an idea on what is happening during compilation.<\/p>\n"}]}
{"QuestionId":35111611,"AnswerCount":1,"Tags":"<deep-learning><tensorflow>","CreationDate":"2016-01-31T07:59:06.583","AcceptedAnswerId":null,"OwnerUserId":1363774.0,"Title":"Need hint for the Exercise posed in the Tensorflow Convolution Neural Networks Tutorial","Body":"<p>Below is the exercise question posed on this page <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html<\/a><\/p>\n\n<blockquote>\n  <p>EXERCISE: The output of inference are un-normalized logits. Try\n  editing the network architecture to return normalized predictions\n  using tf.softmax().<\/p>\n<\/blockquote>\n\n<p>In the spirit of the exercise, I want to know if I'm on the right-track (not looking for the coded-up answer).<\/p>\n\n<p>Here's my proposed solution. <\/p>\n\n<p><strong>Step 1<\/strong>: The last layer (of the inference) in the example is a \"softmax_linear\", i.e., it simply does the <em>unnormalized<\/em> WX+b transformation. As stipulated, we apply the  operation with  as input. This normalizes the output as probabilities on the range [0, 1]. <\/p>\n\n<p><strong>Step 2:<\/strong> The next step is to modify the cross-entropy calculation in the loss-function. Since we already have normalized output, we need to replace the  operation with a plain  function (that does not further normalize the output before calculating the loss). I believe this function is <strong>not<\/strong> available in the tensorflow library; it needs to be written.<\/p>\n\n<p>That's it. Feedback is kindly solicited.<\/p>\n","answers":[{"AnswerId":"35141161","CreationDate":"2016-02-01T21:57:34.757","ParentId":null,"OwnerUserId":"5869947","Title":null,"Body":"<p>Step 1 is more then sufficient if you insert the <code>tf.nn.softmax()<\/code> in <code>cifar10_eval.py<\/code> (and not in <code>cifar10.py<\/code>). For example:<\/p>\n\n<pre><code>logits = cifar10.inference(images)\nnormalized_logits = tf.nn.softmax(logits)\ntop_k_op = tf.nn.in_top_k(normalized_logits, labels, 1)\n<\/code><\/pre>\n"}]}
{"QuestionId":35112212,"AnswerCount":1,"Tags":"<python><c++><xml><deep-learning><caffe>","CreationDate":"2016-01-31T09:22:52.063","AcceptedAnswerId":null,"OwnerUserId":1525479.0,"Title":"Conver xml file to LMDB file for caffe","Body":"<p>I want to use feature vectors which have been extracted from training image dataset as an input for caffe convolutional neural network. The feature vectors have been saved in xml file. \nI'm really new in caffe I want know how can I convert xml file to LMDB as an caffe input and how to mention in the model that I'm using feature vectors instead of images.\nthe xml file is :\n<img src=\"https:\/\/i.stack.imgur.com\/FO7r4.jpg\" alt=\"enter image description here\"><\/p>\n","answers":[{"AnswerId":"35113047","CreationDate":"2016-01-31T10:58:36.117","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you are willing to use python, then it should be quite easy for you.<\/p>\n\n<p>First, you need to parse the XML file and read the data into a numpy array.<br>\nYou can do this using <a href=\"https:\/\/docs.python.org\/2\/library\/xml.etree.elementtree.html\" rel=\"nofollow\">xml python package<\/a>. <\/p>\n\n<p>The second stage is to write the data you have in numpy array into lmdb.<br>\nYou can use <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow\">this solution<\/a> to write LMDB using python.<\/p>\n\n<p>Regarding caffe, as long as your input data is in order, caffe does not care if it works on images, feature vectors or any other input. <\/p>\n"}]}
{"QuestionId":35114024,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-01-31T12:49:07.337","AcceptedAnswerId":35122970.0,"OwnerUserId":2065691.0,"Title":"Loss functions in tensorflow (with an if - else)","Body":"<p>I am trying a different loss functions in tensorflow.<\/p>\n\n<p>The loss function I want is  a kind of an epsilon insensitive function (this is componentwise):<\/p>\n\n<pre><\/pre>\n\n<p>I tried this solution:<\/p>\n\n<pre><\/pre>\n\n<p>I also used<\/p>\n\n<pre><\/pre>\n\n<p>I do not find any error in the implementation. However, it does not converge, while the loss = tf.square(yData-yModel) converges and loss=tf.maximum(tf.square(yData-yModel)-epsilonTensor,tf.zeros_like(yData)) also converges.<\/p>\n\n<p>So, I also tried something simpler loss=tf.abs(yData-yModel) and it also does not converge. Am I making some mistake, or having problems with the non-differentiability of the abs at zero or something else? What is happenning with the abs function?<\/p>\n","answers":[{"AnswerId":"35122970","CreationDate":"2016-02-01T04:24:04.593","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>When your loss is something like <code>Loss(x)=abs(x-y)<\/code>, then solution is an unstable fixed point of SGD -- start your minimization with a point arbitrarily close to the solution, and the next step will increase the loss. <\/p>\n\n<p>Having a stable fixed point is a requirement for convergence of an iterative procedure like SGD. In practice this means your optimization will move towards a local minimum, but after getting close enough, will jump around the solution with steps proportional to the learning rate. Here's a toy TensorFlow program that illustrates the problem<\/p>\n\n<pre><code>x = tf.Variable(0.)\nloss_op = tf.abs(x-1.05)\nopt = tf.train.GradientDescentOptimizer(0.1)\ntrain_op = opt.minimize(loss_op)\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\nxvals = []\nfor i in range(20):\n  unused, loss, xval = sess.run([train_op, loss_op, x])\n  xvals.append(xval)\npyplot.plot(xvals)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hrw4i.png\"><img src=\"https:\/\/i.stack.imgur.com\/hrw4i.png\" alt=\"Graph of x estimate\"><\/a><\/p>\n\n<p>Some solutions to the problem:<\/p>\n\n<ol>\n<li>Use a more robust solver such as the Proximal Gradient Method<\/li>\n<li>Use more SGD friendly loss function such as Huber Loss<\/li>\n<li>Use learning rate schedule to gradually decrease learning rate<\/li>\n<\/ol>\n\n<p>Here's a way to implement (3) on the toy problem above<\/p>\n\n<pre><code>x = tf.Variable(0.)\nloss_op = tf.abs(x-1.05)\n\nstep = tf.Variable(0)\nlearning_rate = tf.train.exponential_decay(\n      0.2,   # Base learning rate.\n      step,  # Current index into the dataset.\n      1,     # Decay step.\n      0.9    # Decay rate\n)\n\nopt = tf.train.GradientDescentOptimizer(learning_rate)\ntrain_op = opt.minimize(loss_op, global_step=step)\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\nxvals = []\nfor i in range(40):\n  unused, loss, xval = sess.run([train_op, loss_op, x])\n  xvals.append(xval)\npyplot.plot(xvals)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/t0mLm.png\"><img src=\"https:\/\/i.stack.imgur.com\/t0mLm.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35114376,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-01-31T13:30:45.643","AcceptedAnswerId":null,"OwnerUserId":3972931.0,"Title":"Error when computing summaries in TensorFlow","Body":"<p>I am trying to use TensorFlow to produce summaries and visualize them using TensorBoard. However, I am getting an error () that I do not understand.<\/p>\n\n<p>This is the full source of my program:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I attempt to run the above code, the following error is raised:<\/p>\n\n<pre><\/pre>\n\n<p>It seems from the error that a placeholder in my source has not been fed an appropriate value. As far as I can tell, I am feeding values for all of the placeholders ( and ).<\/p>\n\n<p>If you need I'll add the full log to this issue.<\/p>\n\n<p>I also found that when I first fetch mnist, it does work (with the following output) but still no TensorBoard visualization is produced:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"49436133","CreationDate":"2018-03-22T18:38:19.093","ParentId":null,"OwnerUserId":"521776","Title":null,"Body":"<p><strong>for keras users<\/strong><\/p>\n\n<p>you might encounter this error when using the TensorBoard callback and fitting a <em>new<\/em> model from scratch. In this case, the solution is to call<\/p>\n\n<pre><code>from keras import backend as K\nK.clear_session()\n<\/code><\/pre>\n\n<p>before creating the new model. verified with keras 2.1.5 and tensorflow 1.6.0<\/p>\n"},{"AnswerId":"35117760","CreationDate":"2016-01-31T18:40:24.870","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From your error message, it looks like you are using IPython. One pitfall when using IPython to build a TensorFlow model is that functions like <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#merge_all_summaries\"><code>tf.merge_all_summaries()<\/code><\/a> will remember every summary created in the current session, including cells that failed with an error. This is a result of TensorFlow using a <strong>default graph<\/strong> to collect all of the operations, summaries, etc. that are created in a process, unless you specify the graph explicitly. I suspect that your call to <code>tf.merge_all_summaries()<\/code> is returning more than the three histogram summaries that you created in your code, and the older ones will have a dependency on a <em>previously created<\/em> placeholder.<\/p>\n\n<p>There are two main ways that you could fix that. The simplest is to explicitly merge the summaries, rather than using <code>tf.merge_all_summaries()<\/code>:<\/p>\n\n<pre><code>weights_summary = tf.histogram_summary(\"weights\", W)\nbiases_summary = tf.histogram_summary(\"biases\", b)\ny_summary = tf.histogram_summary(\"y\", y)\n\nmerged = tf.merge_summary([weights_summary, biases_summary, y_summary])\n<\/code><\/pre>\n\n<p>An alternative would be to set an <strong>explicit default graph<\/strong> before constructing your model. This is awkward if you want to split your model across multiple IPython cells, but should also work:<\/p>\n\n<pre><code># Sets a new default graph, and stores it in `g`.\nwith tf.Graph().as_default() as g:  \n\n    x = tf.placeholder(tf.float32, [None, 784])\n    # ...\n\n    # Summaries are added to `g`.\n    _ = tf.histogram_summary(\"weights\", W)\n    _ = tf.histogram_summary(\"biases\", b)\n    _ = tf.histogram_summary(\"y\", y)\n\n    # ...\n\n    # `merged` contains only summaries from `g`.\n    merged = tf.merge_all_summaries()\n\n    # ...\n<\/code><\/pre>\n"}]}
{"QuestionId":35114668,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-01-31T14:00:40.577","AcceptedAnswerId":35114941.0,"OwnerUserId":2452617.0,"Title":"torch7: How to get the dimensionality of a tensor?","Body":"<p>Given any torch7 tensor, how can one find its dimensionality?<\/p>\n\n<p>I want to apply it to check if a particular tensor is one-dimensional or not, but I am curious about any tensor in general.<\/p>\n","answers":[{"AnswerId":"35114941","CreationDate":"2016-01-31T14:28:54.893","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<pre><code>t:dim()\n<\/code><\/pre>\n\n<p>I suggest that you read the documentation here: <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md\" rel=\"nofollow\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md<\/a><\/p>\n"}]}
{"QuestionId":35116566,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-01-31T17:00:49.010","AcceptedAnswerId":null,"OwnerUserId":3972931.0,"Title":"How can I code feed_dict","Body":"<blockquote>\n  <p>Codes which produces AE<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>What I got as an error message<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>in log<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>I don't know what I need to do. It seems to me that the codes are valid.<\/p>\n","answers":[{"AnswerId":"35345905","CreationDate":"2016-02-11T17:19:21.020","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem here is almost certainly the call to <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/train.html#merge_all_summaries\" rel=\"nofollow noreferrer\"><code>tf.merge_all_summaries()<\/code><\/a>, which doesn't play well with interactive (e.g. IPython) usage. Perhaps unintuitively, <code>tf.merge_all_summaries()<\/code> adds a new operation to the graph that depends on <em>all<\/em> summaries that have been created in the same graph. When you are using IPython, the graph will remember all of the summaries that you created, even in previous executions of the same cell. This can lead to <code>tf.merge_all_summaries()<\/code> creating a tensor that depends on a <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> created in a previous execution of the cell.<\/p>\n\n<p>There are two main workarounds for this, as I described in <a href=\"https:\/\/stackoverflow.com\/a\/35117760\/3574081\">another answer<\/a>:<\/p>\n\n<ol>\n<li><p>Explicitly merge the summaries that you want to compute, using <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/train.html#merge_summary\" rel=\"nofollow noreferrer\"><code>tf.merge_summary([summary_1, summary_2, summary_3])<\/code><\/a>.<\/p><\/li>\n<li><p>Create your graph in an explicit <code>with tf.Graph().as_default():<\/code> block, which will ensure that only the summaries created within that block are included in the merge.<\/p><\/li>\n<\/ol>\n\n<p>Note also that <code>tf.merge_all_summaries()<\/code> <em>adds<\/em> a new op to the graph each time you run it. This means that your training loop has a subtle memory leak, and will add 200 copies of the same op to the graph. To avoid this, you should call <code>tf.merge_all_summaries()<\/code> outside your training loop and cache the resulting tensor. You can then use that tensor within the training loop to get the same result.<\/p>\n"}]}
{"QuestionId":35116689,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-01-31T17:11:25.163","AcceptedAnswerId":null,"OwnerUserId":5706386.0,"Title":"TensorFlow exited abnormally with code 137","Body":"<p>I'm trying convolutional neural net using TensorFlow.<\/p>\n\n<p>Though I could success some training, the script failed with<\/p>\n\n<pre><\/pre>\n\n<p>when I just changed the training data. \nThe data sizes of the first and second data are the same, and \nI could train CNN using the very second data on another platform (lasagne).\n(The first data is slice of photo images, and the second is artificially generated geological patterns).<\/p>\n\n<p>Can someone tell me the meaning of the error or the way to avoid this problem?<\/p>\n\n<p>My platform is Mac OS X 10.10.5, Python 2.7.10 |Anaconda 2.2.0 (x86_64),\nand tensorflow-0.5.0.<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"35117550","CreationDate":"2016-01-31T18:22:27.357","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Exit code 137 means that your Python process was killed by the SIGKILL signal. It's hard to say for certain, but one possibility is that your process was killed by the OOM (out-of-memory) killer. Check <code>\/var\/log\/messages<\/code> to see if there is any information about why your process was killed.<\/p>\n\n<p><a href=\"http:\/\/backdrift.org\/oom-killer-how-to-create-oom-exclusions-in-linux\" rel=\"noreferrer\">This post<\/a> has advice on how to disable the OOM killer for an individual process. However, this ignores the root cause, which is that your process (or TensorFlow itself) is using too much memory.<\/p>\n"}]}
{"QuestionId":35116857,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-01-31T17:26:05.320","AcceptedAnswerId":null,"OwnerUserId":3612009.0,"Title":"How to derive the shapes of tensors given the networks architecture, number of outputs and number of samples","Body":"<p>How can we calculate the shapes of various tensors involved in the computation graph once we know the network architecture (number of hidden layers, number of units in each layer), number of outputs, number of inputs and number of samples in the training set. (Assuming the network is fully connected).<\/p>\n\n<p>For example, lets say there are 100 features, 10000 samples, 2 hidden layers (H0, H1) and 10 outputs. H1 has 500 units and H2 has 5000 units. Assume RELU activation is used in H0\/H1 and softmax is used for output layer.<\/p>\n\n<p>In this case, although the sequence of calculations that need to happen is very clear, finding correct shape for each constant\/variable\/place holder  is difficult. \nI am trying to understand if there is a standard method which we can follow to do this.<\/p>\n","answers":[]}
{"QuestionId":35117842,"AnswerCount":1,"Tags":"<image><lua><bmp><torch>","CreationDate":"2016-01-31T18:46:45.473","AcceptedAnswerId":null,"OwnerUserId":2452617.0,"Title":"torch7: Unable to load BMP images","Body":"<p>I am trying to load some images to train my conv-net. I am using '<strong>image<\/strong>' package of torch7 for that purpose. But, during loading images as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error on trying to load a BMP file:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone please suggest some way to load BMP images for further processing in torch7?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"40497573","CreationDate":"2016-11-08T22:33:02.753","ParentId":null,"OwnerUserId":"118803","Title":null,"Body":"<p>Just for completeness,  I'll throw some code here:<\/p>\n\n<p>First, install graphics magic<\/p>\n\n<p><code>$ luarocks install graphicsmagick<\/code><\/p>\n\n<p>Then<\/p>\n\n<pre><code>require 'image'\ngm = require 'graphicsmagick'\n\n-- change the string\nlocal jpg_filename = string.gsub(train_files[i],\".[Bb][Mm][Pp]$\",\".jpg\")\n-- get the info so we can build the size string\nlocal image_info = gm.info(train_path .. train_files[i])\n-- convert takes a size string h x w\nlocal size_string = image_info[\"height\"] .. \"x\" .. image_info[\"width\"]\n\ngm.convert{\n  input = train_path .. train_files[i],\n  output = train_path .. jpg_filename,\n  size = size_string,\n  quality = 95,\n  verbose = true\n}\nlocal img_raw = image.load(train_path .. jpg_filename):mul(255)\n<\/code><\/pre>\n"}]}
{"QuestionId":35118576,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-01-31T19:46:37.523","AcceptedAnswerId":35121364.0,"OwnerUserId":2255757.0,"Title":"Theano Tensor remove element?","Body":"<p>I am currently trying to learn theano.<\/p>\n\n<p>Is there a way to, for example, delete\/add a row\/column from an NxN tensor? The <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.set_subtensor\" rel=\"nofollow\">subtensor<\/a> functionality highlighted in the documentation only modifies elements, not remove them.<\/p>\n","answers":[{"AnswerId":"35121364","CreationDate":"2016-02-01T00:36:16.527","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>Sutensor allow to take part of tensor. It is set_subtensor and inc_subtensor that allow to modify part of them.<\/p>\n\n<p>Theano support Python and NumPy indexing and advanced indexing. You can do what you want in many ways. Here is a simple one:<\/p>\n\n<pre><code>import theano, numpy\nT = theano.tensor.matrix()\nf = theano.function([T], T[[1, 2, 4]])\nf(numpy.arange(25).reshape(5, 5))\n<\/code><\/pre>\n\n<p>So mostly, you can just pass a list with the index of rows that you want to keep. For colums, just use:<\/p>\n\n<pre><code>import theano, numpy\nT = theano.tensor.matrix()\nf = theano.function([T], T[:, [1, 2, 4]])\nf(numpy.arange(25).reshape(5, 5))\n<\/code><\/pre>\n\n<p>To add rows, we support the same interface as numpy, so mostly, you can just build a new tensor by concatenating the part you want:<\/p>\n\n<pre><code>import theano, numpy\nT = theano.tensor.matrix()\no = theano.tensor.concatenate([T[2], T[4], [100, 101, 102, 103, 104]])\nf = theano.function([T], o)\nf(numpy.arange(25).reshape(5, 5))\n<\/code><\/pre>\n"}]}
{"QuestionId":35119109,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-01-31T20:29:52.503","AcceptedAnswerId":35119525.0,"OwnerUserId":2280020.0,"Title":"tensorflow scalar summary tags name exception","Body":"<p>I am trying to learn how to work with tensorflow summary writers by following the HowTo mnist tutorial. That tutorial adds a scalar summary for the loss function. I wrote a loss function in an unusual by building up a regularization term, and I get this exception:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The loss function and adding the summary look like <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>and if I build up the regularizationTerm like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>were regA and regB are tf.Variables previously defined, I get the exception, whereas is I build it up like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>then it works. So I guess I am not setting the name correctly, when I do the += I create a new tensor that is unamed? But why can't I add that into the loss, and then name the loss? That is the only thing I am trying to summarize?  <\/p>\n\n<p>Is there something like += where I can name the output, or preserve the name of the tensor I am modifying?<\/p>\n\n<p>In case the issue is related to something else, here is my simple example where I identified the problem:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"48032992","CreationDate":"2017-12-30T10:21:30.370","ParentId":null,"OwnerUserId":"5806421","Title":null,"Body":"<p>Since this is an issue related to shape of a tensor with dimension of size 1, you could try using <code>tf.squeeze<\/code> which specifically removes dimensions of size 1 from the shape of a tensor.<\/p>\n\n<p>In your case, you could replace<\/p>\n\n<p><code>tf.scalar_summary(loss.op.name, loss)<\/code><\/p>\n\n<p>with <\/p>\n\n<p><code>tf.scalar_summary(loss.op.name, tf.squeeze(loss))<\/code><\/p>\n"},{"AnswerId":"35119525","CreationDate":"2016-01-31T21:12:36.190","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> The problem is the shape of the argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#scalar_summary\" rel=\"noreferrer\"><code>tf.scalar_summary()<\/code><\/a>, not the names.<\/p>\n\n<p>I think the problem is a shape-related issue, stemming from this line:<\/p>\n\n<pre><code>regularizationTerm = tf.Variable(tf.zeros([1], dtype=np.float32), name='regterm')\n<\/code><\/pre>\n\n<p>This defines a variable whose shape is a vector of length 1. The subsequent <code>+=<\/code> operators (which are syntactic sugar for <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#add\" rel=\"noreferrer\"><code>tf.add()<\/code><\/a>) and the <code>tf.add()<\/code> to compute <code>loss<\/code> will produce vector-shaped results, because <code>tf.add()<\/code> <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"noreferrer\"><em>broadcasts<\/em><\/a> the scalar argument to become a vector. Finally, <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#scalar_summary\" rel=\"noreferrer\"><code>tf.scalar_summary()<\/code><\/a> expects its two arguments to have the same shape&mdash;unlike the broadcasting <code>add<\/code>, <code>tf.scalar_summary()<\/code> is not permissive about the shapes of its inputs. The <code>tags<\/code> input is a scalar string (the name of the <code>loss<\/code> op) whereas the <code>values<\/code> input is a vector of length one (the value of the <code>loss<\/code> tensor). Therefore you get the error that you reported.<\/p>\n\n<p>Fortunately, the solution is simple! Either define the <code>regularizationTerm<\/code> variable as a scalar, like so:<\/p>\n\n<pre><code># Note that `[]` is the scalar shape.\nregularizationTerm = tf.Variable(tf.zeros([], dtype=np.float32), name='regterm')\n<\/code><\/pre>\n\n<p>...or pass a vector (of length 1) of strings to <code>tf.scalar_summary()<\/code>:<\/p>\n\n<pre><code># Wrap `loss.op.name` in a list to make it a vector.\ntf.scalar_summary([loss.op.name], loss)\n<\/code><\/pre>\n"}]}
{"QuestionId":35120102,"AnswerCount":2,"Tags":"<python><c++><machine-learning><tensorflow><skflow>","CreationDate":"2016-01-31T22:04:39.073","AcceptedAnswerId":35141328.0,"OwnerUserId":4123073.0,"Title":"Executing a model trained in skflow using the Tensorflow C++ API","Body":"<p>Is it possible to execute a graph using the Tensorflow C++ API that does not have labeled input (or output) nodes? As far as I know, when training my model (using skflow in python, which I then later saved as a binary protobuf), I did not have labeled input\/output nodes, yet I was able to restore the model and do predictions without difficulty in Python. When using the C++ API to execute a graph, the input Vectors are pairs of Strings and Tensors, where I'm assuming the String refers to the label of an input node. From the docs:<\/p>\n\n<blockquote>\n<pre><\/pre>\n  \n  <p>Runs the graph with the provided input tensors and fills outputs for\n  the endpoints specified in output_tensor_names. Runs to but does not\n  return Tensors for the nodes in target_node_names.<\/p>\n<\/blockquote>\n\n<p>Is there any way I can execute the graph without knowing the labels of my input\/output nodes? Perhaps there's a way I can load the graph in Python, give the nodes labels, and then save it out as a protobuf again? Ideally I'd like to just pass in a vector which is applied to the input nodes and not have to worry about any labels.<\/p>\n","answers":[{"AnswerId":"35141328","CreationDate":"2016-02-01T22:07:58.047","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>In skflow all nodes already have labels and it just takes care of restoring them for you.<\/p>\n\n<p>Default names are <code>input:0<\/code> and <code>output:0<\/code> for <code>X<\/code> and <code>y<\/code> respectively and then some custom names for predictions and loss depending on the model you have used.<\/p>\n\n<p>The way to find out the names of predictions and probability nodes is to take a look at <code>endpoints<\/code> file in the directory you saved the model (if you used <code>estimator.save(path)<\/code> for saving).<\/p>\n\n<p>It should look like this:<\/p>\n\n<blockquote>\n  <p>input:0<\/p>\n  \n  <p>output:0<\/p>\n  \n  <p>logistic_regression\/softmax_classifier\/Softmax<\/p>\n  \n  <p>logistic_regression\/softmax_classifier\/xent:0<\/p>\n<\/blockquote>\n\n<p>Where first two are names of the input\/output nodes and second two are prediction and loss nodes.<\/p>\n"},{"AnswerId":"35139892","CreationDate":"2016-02-01T20:40:27.900","ParentId":null,"OwnerUserId":"5530615","Title":null,"Body":"<p>If you can \"restore the model and do predictions without difficulty in Python\", then you can find out the names\/labels of the input nodes or tensors using their \"name\" property, search for \".name\" in either:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Operation\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Operation<\/a>\nor:\n  <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Tensor\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/framework.html#Tensor<\/a><\/p>\n\n<p>All nodes have names\/labels whether you have explicitly named them or not.<\/p>\n"}]}
{"QuestionId":35120298,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><tic-tac-toe><torch><q-learning>","CreationDate":"2016-01-31T22:26:04.117","AcceptedAnswerId":null,"OwnerUserId":681005.0,"Title":"Tic tac toe machine learning - valid moves","Body":"<p>I am toying around with Machine learning. Especially Q-Learning where you have a state and actions and give rewards depending on how well the network did. <\/p>\n\n<p>Now for starters I set myself a simple goal: Train a network so it emits valid moves for tic-tac-toe (vs a random opponent) as actions. My problem is that the network does not learn at all or even gets worse over time.<\/p>\n\n<p>The first thing I did was getting in touch with torch and a deep q learning module for it: <a href=\"https:\/\/github.com\/blakeMilner\/DeepQLearning\" rel=\"nofollow\">https:\/\/github.com\/blakeMilner\/DeepQLearning<\/a> .<\/p>\n\n<p>Then I wrote a simple tic-tac-toe game where a random player competes with the neural net and plugged this into the code from this sample <a href=\"https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/test.lua\" rel=\"nofollow\">https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/test.lua<\/a> . The output of the network consists of 9 nodes for setting the respective cell.<\/p>\n\n<p>A move is valid if the network chooses an empty cell (no X or O in it). According to this I give positive reward (if network chooses empty cell) and negative rewards (if network chooses an occupied cell).<\/p>\n\n<p>The problem is it never seems to learn. I tried lots of variations:<\/p>\n\n<ul>\n<li>mapping the tic-tac-toe field as 9 inputs (0 = cell empty, 1 = player 1, 2 = player 2) or as 27 inputs (e.g. for an empty cell 0 [empty = 1, player1 = 0, player2 = 0])<\/li>\n<li>vary the hidden nodes count between 10 and 60<\/li>\n<li>tried up to 60k iterations<\/li>\n<li>varying learning rate between 0.001 and 0.1<\/li>\n<li>giving negative rewards for fails or only rewards for success, different reward values<\/li>\n<\/ul>\n\n<p>Nothing works :(<\/p>\n\n<p>Now I have a couple of questions:<\/p>\n\n<ol>\n<li>Since this is my very first attempt at Q-Learning is there anything I am fundamentally doing wrong?<\/li>\n<li>What parameters are worth changing? The \"Brain\" thing has a lot: <a href=\"https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/deepqlearn.lua#L57\" rel=\"nofollow\">https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/deepqlearn.lua#L57<\/a> .<\/li>\n<li>What would a good count for the number of hidden nodes be?<\/li>\n<li>Is the simple network structure as defined at <a href=\"https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/deepqlearn.lua#L116\" rel=\"nofollow\">https:\/\/github.com\/blakeMilner\/DeepQLearning\/blob\/master\/deepqlearn.lua#L116<\/a> too simple for this problem?<\/li>\n<li>Am I just too impatient and have to train much more iterations? <\/li>\n<\/ol>\n\n<p>Thank you,<\/p>\n\n<p>-Matthias<\/p>\n","answers":[{"AnswerId":"35352727","CreationDate":"2016-02-12T00:28:03.693","ParentId":null,"OwnerUserId":"3403018","Title":null,"Body":"<p>Matthias,<\/p>\n\n<p>It seems you are using one output node? \"The output of the network in the forward step is a number between 1 and 9\". If so, then I believe this is the problem. Instead of having one output node I would treat this as a classification problem and have nine output nodes corresponding to each board position. Then take the argmax of these nodes as the predicted move. This is how networks that play the game of Go are setup (there are 361 output nodes each representing an intersection on the board).  <\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":35125214,"AnswerCount":0,"Tags":"<python-2.7><theano>","CreationDate":"2016-02-01T07:39:51.290","AcceptedAnswerId":null,"OwnerUserId":5866269.0,"Title":"Calculate the cross entropy loss between two heatmaps in Theano","Body":"<p>I have a problem about calculating the cross entropy loss in Theano.<\/p>\n\n<p>Concretely, for my net:<\/p>\n\n<pre><\/pre>\n\n<p>And I want to calculate cross entropy for each output:<\/p>\n\n<pre><\/pre>\n\n<p>In Theano, ,the input must be a matrix.\nHowever, when the program call the cost,the  is , the groundtruth is . <\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n","answers":[]}
{"QuestionId":35125239,"AnswerCount":1,"Tags":"<python><python-2.7><neural-network><keras>","CreationDate":"2016-02-01T07:41:16.833","AcceptedAnswerId":42281505.0,"OwnerUserId":4437954.0,"Title":"How setup Keras model correctly?","Body":"<p>I am struggling with building model in Keras. I want to pass all my features as , and . However, once I run the code below <\/p>\n\n<pre><\/pre>\n\n<p>I got an error:<\/p>\n\n<pre><\/pre>\n\n<p>What are the correct parameters for model initialization in my case?<\/p>\n","answers":[{"AnswerId":"42281505","CreationDate":"2017-02-16T18:03:47.820","ParentId":null,"OwnerUserId":"7523984","Title":null,"Body":"<p>I copied your code and executed it without errors.(just commented the print as I am using Python 3)<\/p>\n\n<p>I set input_dim = 10 and output_classes = 3, there was no problems.<\/p>\n\n<p>May be your are passing a non-integer value for output_classes, if that is the case a range of errors are possible depending on the type of object passed.<\/p>\n"}]}
{"QuestionId":35126954,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2016-02-01T09:27:44.627","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"How can I know if the epoch point is reached in seq2seq model?","Body":"<p>I am training a seq2seq model since many days on a custom parallel corpus of about a million sentences with default settings for the seq2seq model.\nFollowing is the output log which has crossed 350k steps as mentioned in the tutorial. I saw that the bucket perplexity have suddenly increased significantly the overall train perplexity is constant at 1.02 since a long time now , also the learning rate was initialized at 0.5 but now it shows about 0.007 , so the learning rate has also significantly decreased, Also the output of the system is not close to satisfactory.\nHow can I know if the epoch point is reached and should I stop and reconfigure settings like parameter tuning and optimizer improvements?<\/p>\n\n<pre>\nglobal step 372800 learning rate 0.0071 step-time 1.71 perplexity 1.02\n  eval: bucket 0 perplexity 91819.49\n  eval: bucket 1 perplexity 21392511.38\n  eval: bucket 2 perplexity 16595488.15\n  eval: bucket 3 perplexity 7632624.78\nglobal step 373000 learning rate 0.0071 step-time 1.73 perplexity 1.02\n  eval: bucket 0 perplexity 140295.51\n  eval: bucket 1 perplexity 13456390.43\n  eval: bucket 2 perplexity 7234450.24\n  eval: bucket 3 perplexity 3700941.57\nglobal step 373200 learning rate 0.0071 step-time 1.69 perplexity 1.02\n  eval: bucket 0 perplexity 42996.45\n  eval: bucket 1 perplexity 37690535.99\n  eval: bucket 2 perplexity 12128765.09\n  eval: bucket 3 perplexity 5631090.67\nglobal step 373400 learning rate 0.0071 step-time 1.82 perplexity 1.02\n  eval: bucket 0 perplexity 119885.35\n  eval: bucket 1 perplexity 11166383.51\n  eval: bucket 2 perplexity 27781188.86\n  eval: bucket 3 perplexity 3885654.40\nglobal step 373600 learning rate 0.0071 step-time 1.69 perplexity 1.02\n  eval: bucket 0 perplexity 215824.91\n  eval: bucket 1 perplexity 12709769.99\n  eval: bucket 2 perplexity 6865776.55\n  eval: bucket 3 perplexity 5932146.75\nglobal step 373800 learning rate 0.0071 step-time 1.78 perplexity 1.02\n  eval: bucket 0 perplexity 400927.92\n  eval: bucket 1 perplexity 13383517.28\n  eval: bucket 2 perplexity 19885776.58\n  eval: bucket 3 perplexity 7053727.87\nglobal step 374000 learning rate 0.0071 step-time 1.85 perplexity 1.02\n  eval: bucket 0 perplexity 46706.22\n  eval: bucket 1 perplexity 35772455.34\n  eval: bucket 2 perplexity 8198331.56\n  eval: bucket 3 perplexity 7518406.42\nglobal step 374200 learning rate 0.0070 step-time 1.98 perplexity 1.03\n  eval: bucket 0 perplexity 73865.49\n  eval: bucket 1 perplexity 22784461.66\n  eval: bucket 2 perplexity 6340268.76\n  eval: bucket 3 perplexity 4086899.28\nglobal step 374400 learning rate 0.0069 step-time 1.89 perplexity 1.02\n  eval: bucket 0 perplexity 270132.56\n  eval: bucket 1 perplexity 17088126.51\n  eval: bucket 2 perplexity 15129051.30\n  eval: bucket 3 perplexity 4505976.67\nglobal step 374600 learning rate 0.0069 step-time 1.92 perplexity 1.02\n  eval: bucket 0 perplexity 137268.32\n  eval: bucket 1 perplexity 21451921.25\n  eval: bucket 2 perplexity 13817998.56\n  eval: bucket 3 perplexity 4826017.20\n\n<\/pre>\n\n<p>And when will this stop ?<\/p>\n","answers":[{"AnswerId":"35147923","CreationDate":"2016-02-02T07:59:54.410","ParentId":null,"OwnerUserId":"5776136","Title":null,"Body":"<p>It looks like there is a difference between your dev and train data:<\/p>\n\n<blockquote>\n  <p>global step 374600 learning rate 0.0069 step-time 1.92 perplexity 1.02<\/p>\n  \n  <p>eval: bucket 0 perplexity 137268.32<\/p>\n<\/blockquote>\n\n<p>Your training perplexity is 1.02 -- the model is basically perfect on the data it receives for training. But your dev perplexity is enormous, the model does not work at all for the dev set. How did it look in earlier epochs? I would suspect that there is some mismatch. Maybe the tokenization is different for train and dev? Maybe you loaded the wrong file? Maybe the sizes of the buckets from the original translation model are not appropriate for your dev data? It's hard to say without knowing more details.<\/p>\n\n<p>As to when to stop: the original translation model has an infinite training loop because it has a large data-set and capacity and could continue improving for many weeks of training. But it also lowers the learning rate when it's not improving any more, so if your learning rate is very low (as it seems to be in your case), it's a clear signal you can stop.<\/p>\n"}]}
{"QuestionId":35129424,"AnswerCount":1,"Tags":"<python><nlp><theano><recurrent-neural-network><language-model>","CreationDate":"2016-02-01T11:29:39.727","AcceptedAnswerId":null,"OwnerUserId":200340.0,"Title":"RNNLM using theano","Body":"<p>I asked the same question on theano user list, but got no reply, just wondering if anyone can help me here.<\/p>\n\n<p>I am trying to re-implement the RNNLM of <a href=\"http:\/\/www.fit.vutbr.cz\/research\/groups\/speech\/publi\/2010\/mikolov_interspeech2010_IS100722.pdf\" rel=\"nofollow\">http:\/\/www.fit.vutbr.cz\/research\/groups\/speech\/publi\/2010\/mikolov_interspeech2010_IS100722.pdf<\/a> based on this nice post.\nI tried a toy test case which training data is the first 100 sentences of PTB training data (downloaded from <a href=\"http:\/\/rnnlm.org\/\" rel=\"nofollow\">http:\/\/rnnlm.org\/<\/a> ), the same data also used for evaluation.<\/p>\n\n<p><strong>Baseline:<\/strong><\/p>\n\n<p>I trained the LM with 25 iterations, using rnnlm-0.4b from <a href=\"http:\/\/rnnlm.org\/\" rel=\"nofollow\">http:\/\/rnnlm.org\/<\/a>, I got\ntest log probability: -4107.323481\nPPL net: 85.496622<\/p>\n\n<p>The command lines that to produce the baseline are:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Using my implementation<\/strong>, after 25 iterations, there is a large difference in PPL:<\/p>\n\n<blockquote>\n  <p>epoch=24: log probability=-5374.255371 ppl=337.187731<\/p>\n<\/blockquote>\n\n<p>I am still learning Theano, did i miss something in my implementation?<\/p>\n\n<p>Thanks<\/p>\n\n<p>My implementation can be found at <a href=\"https:\/\/11350770138305416713.googlegroups.com\/attach\/179c4942cbf157\/rnnlm.py?part=0.1&amp;view=1&amp;vt=ANaJVrGZjtk4NfhfR2gY8Xk3x5yDxT5FUczkzSB6PRIGVu7lN_Dr9XHUZcYNE3WSrEc6SCL7vf8O9Xg5oCuzST4Z5oxo0WgCAIjBz7X0SC_4ZnvpsKf4HII\" rel=\"nofollow\">here<\/a>:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40195172","CreationDate":"2016-10-22T17:44:32.747","ParentId":null,"OwnerUserId":"2498151","Title":null,"Body":"<p>I'm working on the same problem. I'm not sure I've found all the problems but if you print out the V weights you'll find them much larger than the others. I divided by the vocabulary size instead of number of hidden weights. It's helping but the net still doesn't converge all the way.<\/p>\n\n<p>You'll want to print the values for all the weights, this code will get you started. <\/p>\n\n<p>I tried adding L1 and L2 regularization, it didn't help.<\/p>\n\n<pre><code>V = np.random.uniform(-np.sqrt(1.\/n_unique_words), np.sqrt(1.\/n_unique_words), (n_unique_words, n_hidden))\n\ndef sum_weights(self):\n    v = self.V.sum()\n    return v.eval()\n\ndef abs_sum_weights(self):\n    v = (np.abs(self.V)).sum()\n    return v.eval()\n<\/code><\/pre>\n\n<p>This is still a work in progress. I don't think you need more than the snippets I gave you. If you do the rest is here. <a href=\"https:\/\/github.com\/timestocome\/AliceInWonderland\" rel=\"nofollow\">https:\/\/github.com\/timestocome\/AliceInWonderland<\/a> If you walk back a few updates you'll find the L1\/L2 code.<\/p>\n"}]}
{"QuestionId":35133320,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-02-01T14:41:32.420","AcceptedAnswerId":null,"OwnerUserId":3271237.0,"Title":"Tensor variable indexing","Body":"<p>I'm trying to make some symbolic calculations using indexing of symbolic variable.<\/p>\n\n<pre><\/pre>\n\n<p>And I'm getting<\/p>\n\n<pre><\/pre>\n\n<p>How to use symbolic variable indexing correctly?<\/p>\n","answers":[{"AnswerId":"35134290","CreationDate":"2016-02-01T15:31:15.253","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You declare both <code>y<\/code> and <code>X<\/code> as matrices but your inputs to the compiled Theano function are an object array and a vector.<\/p>\n\n<p><code>np.array([np.array([[3],[5]]),np.array([[4,1]])])<\/code> is an object array because it is constructed as an array of numpy arrays. Note that <code>np.array([np.array([[3],[5]]),np.array([[4,1]])]).dtype == object<\/code>.<\/p>\n\n<p>To create a matrix, just use a multi-dimensional array in the numpy array construction. You don't even need to create numpy arrays, just pass vanilla Python lists. Since your second argument (for <code>X<\/code>) is a vector I've assumed the input value is correct and the symbolic variable declaration is incorrect. With these changes, the following code runs:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\nX = T.vector('X')\ny = T.matrix('y')\nz = T.dot(T.dot(X,y[0]),y[1]).norm(L=2)\ncallable_function = theano.function([y,X], z)\nprint callable_function([[3,5],[4,1]], [1,2])\n<\/code><\/pre>\n"}]}
{"QuestionId":35135149,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-02-01T16:14:06.250","AcceptedAnswerId":35140326.0,"OwnerUserId":1879926.0,"Title":"Subtract a vector across all columns in a tensor for Torch","Body":"<p>Given a tensor of dimensions NxM and a vector Nx1, how do I subtract the vector in each column of the tensor in Torch?<\/p>\n","answers":[{"AnswerId":"35140326","CreationDate":"2016-02-01T21:05:53.297","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>One possibility is to use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#expandingreplicatingsqueezing-tensors\" rel=\"nofollow\">expand<\/a>. Example:<\/p>\n\n<pre><code>local A = torch.Tensor{{1, 2},{3, 4},{5,6}}\nlocal B = torch.ones(3)\nlocal C = A - B:view(3, 1):expandAs(A) -- make a 3x1 tensor before expand\nprint(C)\n--  0  1\n--  2  3\n--  4  5\n-- [torch.DoubleTensor of size 3x2]\n<\/code><\/pre>\n"}]}
{"QuestionId":35135331,"AnswerCount":0,"Tags":"<python><neural-network><theano><lasagne><nolearn>","CreationDate":"2016-02-01T16:22:38.003","AcceptedAnswerId":null,"OwnerUserId":310242.0,"Title":"How can you train multiple neural networks simultaneously in nolearn\/lasagne\/theano on Python?","Body":"<p>I am writing a calibration pipeline to learn the hyperparameters for neural networks to detect properties of DNA sequences*. This therefore requires training a large number of models on the same dataset with different hyperparameters.<\/p>\n\n<p>I am trying to optimise this to run on GPU. DNA sequence datasets are quite small compared to image datasets (typically 10s or 100s of base-pairs in 4 'channels' to represent the 4 DNA bases, A, C, G and T, compared to 10,000s of pixels in 3 RGB channels), and consequently cannot make full use of the parallelisation on a GPU unless multiple models are trained at the same time.<\/p>\n\n<p>Is there a way to do this in nolearn, lasagne or, at worst, Theano?<\/p>\n\n<p>* It's based on the <a href=\"http:\/\/www.nature.com\/nbt\/journal\/v33\/n8\/full\/nbt.3300.html\" rel=\"nofollow\">DeepBind model<\/a> for detecting where transcription factors bind to DNA, if you're interested.<\/p>\n","answers":[]}
{"QuestionId":35136334,"AnswerCount":1,"Tags":"<lua><machine-learning><neural-network><torch><luajit>","CreationDate":"2016-02-01T17:11:03.577","AcceptedAnswerId":35137128.0,"OwnerUserId":4341948.0,"Title":"Unexpected behavior at torch7","Body":"<p>I was working on generating an XOR gate dataset with torch7. But when i printed the dataset i saw that the data was wrong, but i could not find the bug. There seems to be nothing wrong with the code. But I'm new to torch, so mistakes can happen.<\/p>\n\n<p>So, here is my code<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, I printed the values that are being inserted into the  list and for double checking i read from  again. <\/p>\n\n<p>And finally i checked from  after full insertion. The dataset was different somehow. I ran couple of times. Every time it was different. Like it was stuck on 1 or 0.<\/p>\n\n<p>So here is my output<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the format is like this <\/p>\n\n<p><\/p>\n\n<p>I printed original when i read from input[i] and output.<\/p>\n\n<p>I printed dataset when i read from dataset, after being inserted.<\/p>\n\n<p>Also you can see that the first set of values are different when i printed. It should be 1 0 1. But it is  1 0 0.<\/p>\n\n<p>I could not find the bug in my code. Can anyone help? If the question is not clear please let me know. <\/p>\n","answers":[{"AnswerId":"35137128","CreationDate":"2016-02-01T17:53:23.477","ParentId":null,"OwnerUserId":"5675002","Title":null,"Body":"<p>Problem is here: <code>dataset[i] = {input[i], output}<\/code><br>\nYou're not saving calculated result, you're saving reference to value that is changed with subsequent calculations for 'xor' function.<br>Naturally, when you read result, you're always getting the same number - last result written to <code>output[1]<\/code><br>\nTo fix it, either change <code>output<\/code> variable to store actual temporary value (not table), or at least read actual value from output table when saving to dataset[i], do not just save link to table, you won't get deep copy that way.<\/p>\n"}]}
{"QuestionId":35138131,"AnswerCount":1,"Tags":"<python><tensorflow><networking><deep-learning>","CreationDate":"2016-02-01T18:56:30.490","AcceptedAnswerId":35140852.0,"OwnerUserId":2769316.0,"Title":"Invalid argument error (incompatible shapes) with TensorFlow","Body":"<p>I'm trying to train a simple network with tensorflow for the MNIST dataset.\nAt the moment though it is not working.\nIt is basically a modified version of the example given on the TensorFlow website. I just changed a couple lines and removed a layer to see what happened.\nHere is my code:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to execute it it crashes giving me an ArgumentError:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35140852","CreationDate":"2016-02-01T21:38:47.217","ParentId":null,"OwnerUserId":"1575163","Title":null,"Body":"<p>You need for your stride size to reduce your outputs to the right shape - this should fix it (note the strides compared to yours):<\/p>\n\n<pre><code>def conv2d(x, W):\n  return tf.nn.conv2d(x, W, strides=[1, 2, 2, 1], padding='SAME')\n<\/code><\/pre>\n\n<p>To troubleshoot this kind of issue, try printing <code>.get_shape()<\/code> for all of your variables.  Both Tensor and Variable have this function - it will give you a better sense of what's going on and will help immensely with troubleshooting.<\/p>\n\n<p>Here's some code that will help - put this after your declaration of <code>h_pool2<\/code>, it will print the name and shape of each of your vars:<\/p>\n\n<pre><code>from tensorflow.python.ops.variables import Variable\n\nfor k, v in locals().items():\n    if type(v) is Variable or type(v) is tf.Tensor:\n        print(\"{0}: {1}\".format(k, v))\n<\/code><\/pre>\n"}]}
{"QuestionId":35142323,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-02-01T23:20:57.787","AcceptedAnswerId":null,"OwnerUserId":5652211.0,"Title":"caffe always predicts the same result","Body":"<p>I have trained my network with my own data, which is not a image but a vector(20\u00d7128\u00d7128).\nNow I want to use the trained network to forecast.\nHowever, when I try the code,\nI got the same output no matter what the input vector is.\nI will attach the file that I have used .\nAny advice would be appreciated.<\/p>\n\n<pre><\/pre>\n\n<p>This is my deploy file<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39393130","CreationDate":"2016-09-08T14:07:39.930","ParentId":null,"OwnerUserId":"1911043","Title":null,"Body":"<p>When you call <code>net.forward()<\/code> it replaces all layers with the data layer specified by the prototxt, rendering your call to <code>net.blobs['data'].data[...] = X<\/code> ineffective. To mend this run:<\/p>\n\n<p><code>net.forward(data=np.asarray(X))<\/code><\/p>\n\n<p>This should give you the correct predictions.<\/p>\n"}]}
{"QuestionId":35142709,"AnswerCount":1,"Tags":"<python><neural-network><theano><keras>","CreationDate":"2016-02-01T23:58:55.100","AcceptedAnswerId":null,"OwnerUserId":2030396.0,"Title":"Theano input and output sample number error","Body":"<p>I am working on some project and I need to use neural network in python. I am trying to train neural network but I always get error for FIT() function. This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting error that my input array and output array haven't same samples number.<\/p>\n\n<p>my console output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35907357","CreationDate":"2016-03-10T04:04:09.093","ParentId":null,"OwnerUserId":"6039008","Title":null,"Body":"<p>I was implementing Autoencoders when I got this problem saying \"Exception: All input arrays and the target array must have the same number of samples.\".\nmodel.fit expects numpy array as input.<\/p>\n\n<pre><code>X: data, as a numpy array.\ny: labels, as a numpy array.\n<\/code><\/pre>\n\n<p>converting list to numpy array solved my problem.<\/p>\n\n<p>If we see the documentation of model.fit it says:<\/p>\n\n<pre><code>if type(X) == list:\n            if len(set([len(a) for a in X] + [len(y)])) != 1:\n                raise Exception('All input arrays and the target array must ''have the same number of samples.')\n<\/code><\/pre>\n\n<p>so if X is of type list: length of all the items in the list X and length of list y should be same else the exception will be raised.<\/p>\n"}]}
{"QuestionId":35143046,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-02-02T00:33:39.910","AcceptedAnswerId":null,"OwnerUserId":3517912.0,"Title":"Tensorflow seq2seq get sequence hidden state","Body":"<p>I just started to work on tensorflow not long ago. I'm working on the seq2seq model and somehow got the tutorial to work, but I'm stuck at getting the states of each sentence. <\/p>\n\n<p>As far as I understand, the seq2seq model takes an input sequence and generates a hidden state for the sequence through RNN. Later, the model uses the sequence's hidden state to generate a new sequence of data. <\/p>\n\n<p>My problem is what should I do if I want to use the hidden state of input sequence directly?  Say for example if I have a trained model, how should I get the final hidden state of input sequence [token1, token2,....,token N] ?<\/p>\n\n<p>I've been stuck at this for 2 days, I tried many different methods but none of them work.<\/p>\n","answers":[{"AnswerId":"39979469","CreationDate":"2016-10-11T14:22:16.157","ParentId":null,"OwnerUserId":"4070452","Title":null,"Body":"<p>bearsteak's answer above is great, but it is based on tensorflow-0.6 which is quite out of date. So I update his answer in tensorflow-0.8 which is also similar to that in the newest version.<\/p>\n\n<p>(*represent where to modify)<\/p>\n\n<pre><code>losses = []\noutputs = []\n*states = []\nwith ops.op_scope(all_inputs, name, \"model_with_buckets\"):\n    for j, bucket in enumerate(buckets):\n        with variable_scope.variable_scope(variable_scope.get_variable_scope(),\n                                                                             reuse=True if j &gt; 0 else None):\n            *bucket_outputs, _ ,bucket_states= seq2seq(encoder_inputs[:bucket[0]],\n                                                                    decoder_inputs[:bucket[1]])\n            outputs.append(bucket_outputs)\n            if per_example_loss:\n                losses.append(sequence_loss_by_example(\n                        outputs[-1], targets[:bucket[1]], weights[:bucket[1]],\n                        softmax_loss_function=softmax_loss_function))\n            else:\n                losses.append(sequence_loss(\n                    outputs[-1], targets[:bucket[1]], weights[:bucket[1]],\n                    softmax_loss_function=softmax_loss_function))\n\nreturn outputs, losses, *states\n<\/code><\/pre>\n\n<p>at python\/ops\/seq2seq, modify embedding_attention_seq2seq()<\/p>\n\n<pre><code>if isinstance(feed_previous, bool):\n    *outputs, states = embedding_attention_decoder(\n                decoder_inputs, encoder_state, attention_states, cell,\n                num_decoder_symbols, embedding_size, num_heads=num_heads,\n                output_size=output_size, output_projection=output_projection,\n                feed_previous=feed_previous,\n                initial_state_attention=initial_state_attention)\n    *return outputs, states, encoder_state\n\n    # If feed_previous is a Tensor, we construct 2 graphs and use cond.\ndef decoder(feed_previous_bool):\n    reuse = None if feed_previous_bool else True\n    with variable_scope.variable_scope(variable_scope.get_variable_scope(),reuse=reuse):\n        outputs, state = embedding_attention_decoder(\n                decoder_inputs, encoder_state, attention_states, cell,\n                num_decoder_symbols, embedding_size, num_heads=num_heads,\n                output_size=output_size, output_projection=output_projection,\n                feed_previous=feed_previous_bool,\n                update_embedding_for_previous=False,\n                initial_state_attention=initial_state_attention)\n        return outputs + [state]\n\n    outputs_and_state = control_flow_ops.cond(feed_previous, lambda: decoder(True), lambda: decoder(False))                                                                                                                                                           \n    *return outputs_and_state[:-1], outputs_and_state[-1], encoder_state\n<\/code><\/pre>\n\n<p>at model\/rnn\/translate\/seq2seq_model.py modify init()<\/p>\n\n<pre><code>if forward_only:\n    *self.outputs, self.losses, self.states= tf.nn.seq2seq.model_with_buckets(\n            self.encoder_inputs, self.decoder_inputs, targets,\n            self.target_weights, buckets, lambda x, y: seq2seq_f(x, y, True),\n            softmax_loss_function=softmax_loss_function)\n    # If we use output projection, we need to project outputs for decoding.\n    if output_projection is not None:\n        for b in xrange(len(buckets)):\n            self.outputs[b] = [\n                    tf.matmul(output, output_projection[0]) + output_projection[1]\n                    for output in self.outputs[b]\n            ]\nelse:\n    *self.outputs, self.losses, _ = tf.nn.seq2seq.model_with_buckets(\n            self.encoder_inputs, self.decoder_inputs, targets,\n            self.target_weights, buckets,\n            lambda x, y: seq2seq_f(x, y, False),\n            softmax_loss_function=softmax_loss_function)\n<\/code><\/pre>\n\n<p>at model\/rnn\/translate\/seq2seq_model.py modify step()<\/p>\n\n<pre><code>if not forward_only:\n    return outputs[1], outputs[2], None    # Gradient norm, loss, no outputs.\nelse:\n    *return None, outputs[0], outputs[1:], outputs[-1]    # No gradient norm, loss, outputs.\n<\/code><\/pre>\n\n<p>with all these done, we can get the encoded states by calling :<\/p>\n\n<pre><code>_, _, output_logits, states = model.step(sess, encoder_inputs, decoder_inputs,\n                                                                     target_weights, bucket_id, True)\nprint (states)\n<\/code><\/pre>\n\n<p>in the translate.py.<\/p>\n"},{"AnswerId":"35166925","CreationDate":"2016-02-03T01:28:13.507","ParentId":null,"OwnerUserId":"5776136","Title":null,"Body":"<p>In the seq2seq model the encoder is always an RNN, called through rnn.rnn.<\/p>\n\n<p>The call to rnn.rnn returns outputs and state, so to get just the state you can do this:<\/p>\n\n<blockquote>\n  <p>_, encoder_state = rnn.rnn(encoder_cell, encoder_inputs, dtype=dtype)<\/p>\n<\/blockquote>\n\n<p>It's done in the same way in the seq2seq module. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py#L103\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py#L103<\/a><\/p>\n"},{"AnswerId":"35303116","CreationDate":"2016-02-09T22:04:33.397","ParentId":null,"OwnerUserId":"3517912","Title":null,"Body":"<p>Alright, I guess my issue is I don't really know how to code in tensorflow style, so I kind of brute forced it. <\/p>\n\n<p>(*represent where to modify)<\/p>\n\n<p>at python\/ops\/seq2seq, modify      model_with_buckets()  <\/p>\n\n<pre><code>outputs = []\n*states = []\n  with ops.op_scope(all_inputs, name, \"model_with_buckets\"):\n    for j in xrange(len(buckets)):\n      if j &gt; 0:\n        vs.get_variable_scope().reuse_variables()\n      bucket_encoder_inputs = [encoder_inputs[i]\n                               for i in xrange(buckets[j][0])]\n      bucket_decoder_inputs = [decoder_inputs[i]\n                               for i in xrange(buckets[j][1])]\n      *bucket_outputs, _ ,bucket_states= seq2seq(bucket_encoder_inputs,\n                                  bucket_decoder_inputs)\n      outputs.append(bucket_outputs)\n      states.append(bucket_states)\n      bucket_targets = [targets[i] for i in xrange(buckets[j][1])]\n      bucket_weights = [weights[i] for i in xrange(buckets[j][1])]\n      losses.append(sequence_loss(\n          outputs[-1], bucket_targets, bucket_weights, num_decoder_symbols,\n          softmax_loss_function=softmax_loss_function))\n\n  return outputs, losses,*states\n<\/code><\/pre>\n\n<p>at python\/ops\/seq2seq, modify   embedding_attention_seq2seq()<\/p>\n\n<pre><code>if isinstance(feed_previous, bool):\n     * outputs, states =  embedding_attention_decoder(\n          decoder_inputs, encoder_states[-1], attention_states, cell,\n          num_decoder_symbols, num_heads, output_size, output_projection,\n          feed_previous)\n      * return outputs, states, tf.constant(encoder_states[-1])\n    else:  # If feed_previous is a Tensor, we construct 2 graphs and use cond.\n      outputs1, states1 = embedding_attention_decoder(\n          decoder_inputs, encoder_states[-1], attention_states, cell,\n          num_decoder_symbols, num_heads, output_size, output_projection, True)\n      vs.get_variable_scope().reuse_variables()\n      outputs2, states2 = embedding_attention_decoder(\n          decoder_inputs, encoder_states[-1], attention_states, cell,\n          num_decoder_symbols, num_heads, output_size, output_projection, False)\n\n      outputs = control_flow_ops.cond(feed_previous,\n                                      lambda: outputs1, lambda: outputs2)\n      states = control_flow_ops.cond(feed_previous,\n                                     lambda: states1, lambda: states2)\n\n      *return outputs, states, tf.constant(encoder_states[-1])\n<\/code><\/pre>\n\n<p>at model\/rnn\/translate\/seq2seq_model.py modify <strong>init<\/strong>()<\/p>\n\n<pre><code>if forward_only:\n     * self.outputs, self.losses, self.states = seq2seq.model_with_buckets(\n          self.encoder_inputs, self.decoder_inputs, targets,\n          self.target_weights, buckets, self.target_vocab_size,\n          lambda x, y: seq2seq_f(x, y, True),\n          softmax_loss_function=softmax_loss_function)\n      # If we use output projection, we need to project outputs for decoding.\n      if output_projection is not None:\n        for b in xrange(len(buckets)):\n          self.outputs[b] = [tf.nn.xw_plus_b(output, output_projection[0],\n                                             output_projection[1])\n                             for output in self.outputs[b]]\n    else:\n  *    self.outputs, self.losses,_  = seq2seq.model_with_buckets(\n          self.encoder_inputs, self.decoder_inputs, targets,\n          self.target_weights, buckets, self.target_vocab_size,\n          lambda x, y: seq2seq_f(x, y, False),\n          softmax_loss_function=softmax_loss_function)\n<\/code><\/pre>\n\n<p>at model\/rnn\/translate\/seq2seq_model.py modify step()<\/p>\n\n<pre><code>if not forward_only:\n      return outputs[1], outputs[2], None  # Gradient norm, loss, no outputs.\nelse:\n      *return None, outputs[0], outputs[1:-1], outputs[-1]\n<\/code><\/pre>\n\n<p>with all these done, we can get the encoded states by calling :<\/p>\n\n<p><code>_, _, _,states = model.step(all_other_arguements, forward_only = True)<\/code><\/p>\n"}]}
{"QuestionId":35143058,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2016-02-02T00:35:17.170","AcceptedAnswerId":null,"OwnerUserId":5870277.0,"Title":"Importing theano.shared","Body":"<p>These three imports work:<\/p>\n\n<pre><\/pre>\n\n<p>But this one doesn't: <\/p>\n\n<pre><\/pre>\n\n<p>instead, there's this error message: <\/p>\n\n<blockquote>\n  <p>ImportError: No module named shared <\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":35145645,"AnswerCount":3,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2016-02-02T05:22:54.797","AcceptedAnswerId":null,"OwnerUserId":5808490.0,"Title":"How can I feed last output y(t-1) as input for generating y(t) in tensorflow RNN?","Body":"<p>I want to design a single layer RNN in Tensorflow such that last output  is participated in updating the hidden state.<\/p>\n\n<pre><\/pre>\n\n<p>How can I feed last input  as input for updating the hidden state?<\/p>\n","answers":[{"AnswerId":"35166980","CreationDate":"2016-02-03T01:34:02.590","ParentId":null,"OwnerUserId":"5776136","Title":null,"Body":"<p>Is y(t-1) the last input or output? In both cases it is not a straight fit with the TensorFlow RNN cell abstraction. If your RNN is simple you can just write the loop on your own, then you have full control. Another way that I would use is to pre-process your RNN input, e.g., do something like:<\/p>\n\n<blockquote>\n  <p>processed_input[t] = tf.concat(input[t], input[t-1])<\/p>\n<\/blockquote>\n\n<p>Then call the RNN cell with processed_input and split there.<\/p>\n"},{"AnswerId":"54509711","CreationDate":"2019-02-04T03:04:56.813","ParentId":null,"OwnerUserId":"1645784","Title":null,"Body":"<p>I would call what you described an \"autoregressive RNN\". Here's an (incomplete) code snippet that shows how you can create one using <code>tf.nn.raw_rnn<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nLSTM_SIZE = 128\nBATCH_SIZE = 64\nHORIZON = 10\n\nlstm_cell = tf.nn.rnn_cell.LSTMCell(LSTM_SIZE, use_peepholes=True)\n\n\nclass RnnLoop:\n    def __init__(self, initial_state, cell):\n        self.initial_state = initial_state\n        self.cell = cell\n\n    def __call__(self, time, cell_output, cell_state, loop_state):\n        emit_output = cell_output  # == None for time == 0\n        if cell_output is None:  # time == 0\n            initial_input = tf.fill([BATCH_SIZE, LSTM_SIZE], 0.0)\n            next_input = initial_input\n            next_cell_state = self.initial_state\n        else:\n            next_input = cell_output\n            next_cell_state = cell_state\n\n        elements_finished = (time &gt;= HORIZON)\n        next_loop_state = None\n        return elements_finished, next_input, next_cell_state, emit_output, next_loop_state\n\n\nrnn_loop = RnnLoop(initial_state=initial_state_tensor, cell=lstm_cell)\nrnn_outputs_tensor_array, _, _ = tf.nn.raw_rnn(lstm_cell, rnn_loop)\nrnn_outputs_tensor = rnn_outputs_tensor_array.stack()\n<\/code><\/pre>\n\n<p>Here we initialize internal state of LSTM with some vector <code>initial_state_tensor<\/code>, and feed zero array as input at <code>t=0<\/code>. After that, the output of the current timestep is the input for the next timestep.<\/p>\n"},{"AnswerId":"49302825","CreationDate":"2018-03-15T15:00:38.707","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>One possibility is to use <code>tf.nn.raw_rnn<\/code> which I found in this <a href=\"https:\/\/hanxiao.github.io\/2017\/08\/16\/Why-I-use-raw-rnn-Instead-of-dynamic-rnn-in-Tensorflow-So-Should-You-0\/\" rel=\"nofollow noreferrer\">article<\/a>. Check my answer to this <a href=\"https:\/\/stackoverflow.com\/questions\/39681026\/tensorflow-how-to-pass-output-from-previous-time-step-as-input-to-next-timestep\/49274234#49274234\">related post<\/a>. <\/p>\n"}]}
{"QuestionId":35146444,"AnswerCount":3,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-02-02T06:21:56.363","AcceptedAnswerId":null,"OwnerUserId":5239688.0,"Title":"Tensorflow python : Accessing individual elements in a tensor","Body":"<p>This question is with respect to accessing individual elements in a tensor, say [[1,2,3]]. I need to access the inner element [1,2,3] (This can be performed using .eval() or sess.run()) but it takes longer when the size of the tensor is huge) <\/p>\n\n<p>Is there any method to do the same faster?<\/p>\n\n<p>Thanks in Advance.<\/p>\n","answers":[{"AnswerId":"35172568","CreationDate":"2016-02-03T09:05:06.133","ParentId":null,"OwnerUserId":"3301357","Title":null,"Body":"<p>You simply can't get <strong>value<\/strong> of 0th element of [[1,2,3]] without run()-ning or eval()-ing an operation which would be getting it. Because before you 'run' or 'eval', you have only a description how to get this inner element(because TF uses symbolic graphs\/calculations). So even if you would use tf.gather\/tf.slice, you still would have to get <strong>values<\/strong> of these operations via eval\/run. See @mrry's answer.<\/p>\n"},{"AnswerId":"35158370","CreationDate":"2016-02-02T16:12:35.037","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are two main ways to access subsets of the elements in a tensor, either of which should work for your example.<\/p>\n\n<ol>\n<li><p>Use the indexing operator (based on <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#slice\"><code>tf.slice()<\/code><\/a>) to extract a contiguous slice from the tensor.<\/p>\n\n<pre><code>input = tf.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9]])\n\noutput = input[0, :]\nprint sess.run(output)  # ==&gt; [1 2 3]\n<\/code><\/pre>\n\n<p>The indexing operator supports many of the same slice specifications as NumPy does.<\/p><\/li>\n<li><p>Use the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#gather\"><code>tf.gather()<\/code><\/a> op to select a non-contiguous slice from the tensor.<\/p>\n\n<pre><code>input = tf.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9]])\n\noutput = tf.gather(input, 0)\nprint sess.run(output)  # ==&gt; [1 2 3]\n\noutput = tf.gather(input, [0, 2])\nprint sess.run(output)  # ==&gt; [[1 2 3] [7 8 9]]\n<\/code><\/pre>\n\n<p>Note that <code>tf.gather()<\/code> only allows you to select whole slices in the 0th dimension (whole rows in the example of a matrix), so you may need to <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#reshape\"><code>tf.reshape()<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#transpose\"><code>tf.transpose()<\/code><\/a> your input to obtain the appropriate elements.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"35148137","CreationDate":"2016-02-02T08:13:21.157","ParentId":null,"OwnerUserId":"241013","Title":null,"Body":"<p>I suspect it's the rest of the computation that takes time, rather than accessing one element.<\/p>\n\n<p>Also the result might require a copy from whatever memory is stored in, so if it's on the graphics card it will need to be copied back to RAM first and then you get access to your element. If this is the case you might skip it by adding an tensorflow operation to take the first element, and only return that.<\/p>\n"}]}
{"QuestionId":35147630,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2016-02-02T07:40:50.053","AcceptedAnswerId":null,"OwnerUserId":5808490.0,"Title":"How can I stop RNN after generating a special output word in Tensorflow?","Body":"<p>I want to implement an encoder-decoder model for sequence to sequence learning.<\/p>\n\n<p>Encoder reads the input sequence word by word and update its hidden state.<\/p>\n\n<p>Decoder uses the hidden state of encoder to initializing its hidden state. and then generating output with respect to last generated output (y(t-1)) and its hidden state. I want to stop this procedure when a special output () is generated. In fact, I want to able to generate different length outputs. How can I do that in Tensorflow?<\/p>\n","answers":[{"AnswerId":"35930531","CreationDate":"2016-03-11T01:17:20.387","ParentId":null,"OwnerUserId":"348412","Title":null,"Body":"<p>I suppose you want something like the <code>sequence_length<\/code> of <code>tf.nn.rnn<\/code>. I want it too, but it seems TensorFlow doesn't have it. <\/p>\n\n<p>What I've being doing so far and found a good way around this limitation is to pad the decoder labels at train time with the EOS symbol. Usually, you would need only one of these, but filling a lot of them doesn't do any harm. <\/p>\n\n<p>At execution time, you either control each iteration manually to stop when the first EOS is produced or just run a predefined number of time steps and then remove the extra EOS symbols from your output. The decoder quickly learns that after the first EOS, only more EOS may follow.<\/p>\n"}]}
{"QuestionId":35148121,"AnswerCount":2,"Tags":"<python><assign><tensorflow>","CreationDate":"2016-02-02T08:12:33.400","AcceptedAnswerId":35156269.0,"OwnerUserId":258483.0,"Title":"Assign op in TensorFlow: what is the return value?","Body":"<p>I was trying to build an autoincrementing graph in TensorFlow. I thought that the  op might be suitable for that, but found no documentation for it.<\/p>\n\n<p>I assumed that this op returns its value&mdash;like in C-like languages&mdash;and wrote the following code:<\/p>\n\n<pre><\/pre>\n\n<p>and this code works.<\/p>\n\n<p>The question is: is this the expected behavior? Why is the assign op not documented here: <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/index.html\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/index.html<\/a><\/p>\n\n<p>Is it a non-recommended op?<\/p>\n","answers":[{"AnswerId":"43510633","CreationDate":"2017-04-20T05:05:34.403","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><code>tf.assign()<\/code> is nicely <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\">documented in the latest versions<\/a> and it is used frequently in the projects.<\/p>\n\n<blockquote>\n  <p>This operation outputs \"ref\" after the assignment is done. This makes\n  it easier to chain operations that need to use the reset value.<\/p>\n<\/blockquote>\n\n<p>In easier words it takes your original tensor and a new tensor. It updates original value of your tensor with a new value and returns the reference of your original tensor.<\/p>\n"},{"AnswerId":"35156269","CreationDate":"2016-02-02T14:36:16.603","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.assign()<\/code> operator is the underlying mechanism that implements the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/state_ops.html#Variable.assign\" rel=\"noreferrer\"><code>Variable.assign()<\/code><\/a> method. It takes a <em>mutable tensor<\/em> (with <code>tf.*_ref<\/code> type) and a new value, and returns a <em>mutable tensor<\/em> that has been updated with the new value. The return value is provided to make it easier to order an assignment before a subsequent read, but this feature is not well documented. An example will hopefully illustrate:<\/p>\n\n<pre><code>v = tf.Variable(0)\nnew_v = v.assign(10)\noutput = v + 5  # `v` is evaluated before or after the assignment.\n\nsess.run(v.initializer)\n\nresult, _ = sess.run([output, new_v.op])\nprint result  # ==&gt; 10 or 15, depending on the order of execution.\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>v = tf.Variable(0)\nnew_v = v.assign(10)\noutput = new_v + 5  # `new_v` is evaluated after the assignment.\n\nsess.run(v.initializer)\n\nresult = sess.run([output])\nprint result  # ==&gt; 15\n<\/code><\/pre>\n\n<p>In your code example the dataflow dependencies enforce the order of execution <code>[read counter] -&gt; new_counter = tf.add(...) -&gt; tf.assign(...) -&gt; [read output of assign] -&gt; result = tf.add(...)<\/code>, which means that the semantics are unambiguous. <strong>However<\/strong>, the read-modify-write steps to update the counter are somewhat inefficient, and can have unexpected behavior when there are multiple steps running concurrently. For example, multiple threads accessing the same variable could observe the counter moving backwards (in the case that an older value was written back after a newer value).<\/p>\n\n<p>I would recommend that you use <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/state_ops.html#Variable.assign_add\" rel=\"noreferrer\"><code>Variable.assign_add()<\/code><\/a> to update the counter, as follows:<\/p>\n\n<pre><code>counter = tf.Variable(0, name=\"counter\")\n\none = tf.constant(1)\nten = tf.constant(10)\n\n# assign_add ensures that the counter always moves forward.\nupdated_counter = counter.assign_add(one, use_locking=True)\n\nresult = tf.add(updated_counter, ten)\n# ...\n<\/code><\/pre>\n"}]}
{"QuestionId":35151207,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-02-02T10:42:45.367","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"How to interpret Poolallocator messages in tensorflow?","Body":"<p>While training a tensorflow seq2seq model I see the following messages :<\/p>\n\n<pre>\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 27282 get requests, put_count=9311 evicted_count=1000 eviction_rate=0.1074 and unsatisfied allocation rate=0.699032\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 100 to 110\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 13715 get requests, put_count=14458 evicted_count=10000 eviction_rate=0.691659 and unsatisfied allocation rate=0.675684\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 110 to 121\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 6965 get requests, put_count=6813 evicted_count=5000 eviction_rate=0.733891 and unsatisfied allocation rate=0.741421\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 133 to 146\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 44 get requests, put_count=9058 evicted_count=9000 eviction_rate=0.993597 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 46 get requests, put_count=9062 evicted_count=9000 eviction_rate=0.993158 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 4 get requests, put_count=1029 evicted_count=1000 eviction_rate=0.971817 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 2 get requests, put_count=1030 evicted_count=1000 eviction_rate=0.970874 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 44 get requests, put_count=6074 evicted_count=6000 eviction_rate=0.987817 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 12 get requests, put_count=6045 evicted_count=6000 eviction_rate=0.992556 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 2 get requests, put_count=1042 evicted_count=1000 eviction_rate=0.959693 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 44 get requests, put_count=6093 evicted_count=6000 eviction_rate=0.984737 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 4 get requests, put_count=1069 evicted_count=1000 eviction_rate=0.935454 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 17722 get requests, put_count=9036 evicted_count=1000 eviction_rate=0.110668 and unsatisfied allocation rate=0.550615\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 792 to 871\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 6 get requests, put_count=1093 evicted_count=1000 eviction_rate=0.914913 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 6 get requests, put_count=1101 evicted_count=1000 eviction_rate=0.908265 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 3224 get requests, put_count=4684 evicted_count=2000 eviction_rate=0.426985 and unsatisfied allocation rate=0.200062\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 1158 to 1273\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 17794 get requests, put_count=17842 evicted_count=9000 eviction_rate=0.504428 and unsatisfied allocation rate=0.510228\nI tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:239] Raising pool_size_limit_ from 1400 to 1540\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 31 get requests, put_count=1185 evicted_count=1000 eviction_rate=0.843882 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 40 get requests, put_count=8209 evicted_count=8000 eviction_rate=0.97454 and unsatisfied allocation rate=0\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 0 get requests, put_count=2272 evicted_count=2000 eviction_rate=0.880282 and unsatisfied allocation rate=-nan\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 0 get requests, put_count=2362 evicted_count=2000 eviction_rate=0.84674 and unsatisfied allocation rate=-nan\nW tensorflow\/core\/common_runtime\/gpu\/pool_allocator.cc:227] PoolAllocator: After 38 get requests, put_count=5436 evicted_count=5000 eviction_rate=0.919794 and unsatisfied allocation rate=0\n\n<\/pre>\n\n<p>What does it mean , does it mean I am having some resource allocation issues? Am running on Titan X 3500+ CUDA ,12 GB GPU<\/p>\n","answers":[{"AnswerId":"50326911","CreationDate":"2018-05-14T09:18:39.207","ParentId":null,"OwnerUserId":"5031051","Title":null,"Body":"<p>tl;dr<\/p>\n\n<p>You are OOM.. \ntry reducing the size of your network or training fewer models simultaneously or play around with batch size etc.<\/p>\n"},{"AnswerId":"35166985","CreationDate":"2016-02-03T01:34:34.570","ParentId":null,"OwnerUserId":"5875585","Title":null,"Body":"<p>TensorFlow has multiple memory allocators, for memory that will be used in different ways.  Their behavior has some adaptive aspects.<\/p>\n\n<p>In your particular case, since you're using a GPU, there is a PoolAllocator for CPU memory that is pre-registered with the GPU for fast DMA.  A tensor that is expected to be transferred from CPU to GPU, e.g., will be allocated from this pool.<\/p>\n\n<p>The PoolAllocators attempt to amortize the cost of calling a more expensive underlying allocator by keeping around a pool of allocated then freed chunks that are eligible for immediate reuse.  Their default behavior is to grow slowly until the eviction rate drops below some constant.  (The eviction rate is the proportion of free calls where we return an unused chunk from the pool to the underlying pool in order not to exceed the size limit.)  In the log messages above, you see \"Raising pool_size_limit_\" lines that show the pool size growing.  Assuming that your program actually has a steady state behavior with a maximum size collection of chunks it needs, the pool will grow to accommodate it, and then grow no more.  It behaves this way rather than simply retaining all chunks ever allocated so that sizes needed only rarely, or only during program startup, are less likely to be retained in the pool.<\/p>\n\n<p>These messages should only be a cause for concern if you run out of memory.  In such a case the log messages may help diagnose the problem.  Note also that peak execution speed may only be attained after the memory pools have grown to the proper size.<\/p>\n"}]}
{"QuestionId":35151739,"AnswerCount":1,"Tags":"<python-2.7><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-02-02T11:06:20.687","AcceptedAnswerId":null,"OwnerUserId":3767230.0,"Title":"How to train and test LeNet using caffe using python","Body":"<p>I am new to caffe and Machine learning algorithms. Are there any tutorials to train and <strong>TEST<\/strong> LeNet on the MNIST data using caffe and python 2.7(preferably). Thanks <\/p>\n","answers":[{"AnswerId":"35182417","CreationDate":"2016-02-03T16:26:54.043","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>There's caffe's ipython notebook example <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/01-learning-lenet.ipynb\" rel=\"nofollow\">Learning LeNet<\/a>.<\/p>\n\n<p>The code is not limited to ipython notebook (although easier to navigate). You can copy them into the python interpreter and run them there.<\/p>\n"}]}
{"QuestionId":35152052,"AnswerCount":2,"Tags":"<python-2.7><theano>","CreationDate":"2016-02-02T11:21:02.917","AcceptedAnswerId":null,"OwnerUserId":5866269.0,"Title":"Take input of arbitrary size in theano","Body":"<p>I built a cnn network in Theano. Input are many images, but the size of them are different. The elements of numpy.array have the same size.<br>\nHow can I make them as the input?\nThanks a lot.<\/p>\n","answers":[{"AnswerId":"35155459","CreationDate":"2016-02-02T13:58:32.203","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>When training your network, for architectural and performance reasons, all your input images should be of the same shape, usually square. You can achieve square input e.g. by doing<\/p>\n\n<ol>\n<li>reshape the image such that the smaller side is equal to the square sidelength, while preserving aspect ratio<\/li>\n<li>crop the image along the other axis, either in the middle or in a random location, to obtain the square<\/li>\n<\/ol>\n\n<p>This makes sure that the image is not distorted.<\/p>\n\n<p>That said, if you have already trained your network, and would like to apply it to an image bigger than its normal input size, then you can convolve the convnet over the image. This is explained in Pierre Sermanet's OverFeat paper. It is also implemented for the estimators in sklearn-theano at <a href=\"http:\/\/sklearn-theano.github.io\" rel=\"nofollow\">http:\/\/sklearn-theano.github.io<\/a><\/p>\n"},{"AnswerId":"35153245","CreationDate":"2016-02-02T12:14:42.477","ParentId":null,"OwnerUserId":"2255757","Title":null,"Body":"<p>You need to do some data formatting. The input size of a NN is constant, so if the images for your CNN have different sizes you need to resize them to your input size before feeding them in. Its like a person being too close or far away from a painting, your field of view is contant, in order to see everything clearly you need to adjust your distance from the image.<\/p>\n"}]}
{"QuestionId":35155655,"AnswerCount":6,"Tags":"<classification><tensorflow>","CreationDate":"2016-02-02T14:07:36.470","AcceptedAnswerId":null,"OwnerUserId":5048076.0,"Title":"Loss function for class imbalanced binary classifier in Tensor flow","Body":"<p>I am trying to apply deep learning for a binary classification problem with high class imbalance between target classes (500k, 31K). I want to write a custom loss function which should be like: \nminimize(100-((predicted_smallerclass)\/(total_smallerclass))*100)<\/p>\n\n<p>Appreciate any pointers on how I can build this logic. <\/p>\n","answers":[{"AnswerId":"47597502","CreationDate":"2017-12-01T16:27:02.700","ParentId":null,"OwnerUserId":"2542699","Title":null,"Body":"<p>I had to work with a similar unbalanced dataset of multiple classes and this is how I worked through it, hope it will help somebody looking for a similar solution:<\/p>\n\n<p>This goes inside your training module:  <\/p>\n\n<pre><code>from sklearn.utils.class_weight import compute_sample_weight\n#use class weights for handling unbalanced dataset\nif mode == 'INFER' #test\/dev mode, not weighing loss in test mode\n   sample_weights = np.ones(labels.shape)\nelse:\n   sample_weights = compute_sample_weight(class_weight='balanced', y=labels)\n<\/code><\/pre>\n\n<p>This goes inside your model class definition:<\/p>\n\n<pre><code>#an extra placeholder for sample weights\n#assuming you already have batch_size tensor\nself.sample_weight = tf.placeholder(dtype=tf.float32, shape=[None],\n                       name='sample_weights')\ncross_entropy_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(\n                       labels=self.label, logits=logits, \n                       name='cross_entropy_loss')\ncross_entropy_loss = tf.reduce_sum(cross_entropy_loss*self.sample_weight) \/ batch_size\n<\/code><\/pre>\n"},{"AnswerId":"42163122","CreationDate":"2017-02-10T15:34:17.950","ParentId":null,"OwnerUserId":"4747756","Title":null,"Body":"<p>Did ops <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#weighted_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">tf.nn.weighted_cross_entropy_with_logits()<\/a> for two classes:<\/p>\n\n<pre><code>classes_weights = tf.constant([0.1, 1.0])\ncross_entropy = tf.nn.weighted_cross_entropy_with_logits(logits=logits, targets=labels, pos_weight=classes_weights)\n<\/code><\/pre>\n"},{"AnswerId":"35168022","CreationDate":"2016-02-03T03:35:30.013","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>You can add class weights to the loss function, by multiplying logits. \nRegular cross entropy loss is this:<\/p>\n\n<pre><code>loss(x, class) = -log(exp(x[class]) \/ (\\sum_j exp(x[j])))\n               = -x[class] + log(\\sum_j exp(x[j]))\n<\/code><\/pre>\n\n<p>in weighted case:<\/p>\n\n<pre><code>loss(x, class) = weights[class] * -x[class] + log(\\sum_j exp(weights[class] * x[j]))\n<\/code><\/pre>\n\n<p>So by multiplying logits, you are re-scaling predictions of each class by its class weight.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>ratio = 31.0 \/ (500.0 + 31.0)\nclass_weight = tf.constant([ratio, 1.0 - ratio])\nlogits = ... # shape [batch_size, 2]\nweighted_logits = tf.mul(logits, class_weight) # shape [batch_size, 2]\nxent = tf.nn.softmax_cross_entropy_with_logits(\n  weighted_logits, labels, name=\"xent_raw\")\n<\/code><\/pre>\n\n<p>There is a standard losses function now that supports weights per batch:<\/p>\n\n<pre><code>tf.losses.sparse_softmax_cross_entropy(labels=label, logits=logits, weights=weights)\n<\/code><\/pre>\n\n<p>Where weights should be transformed from class weights to a weight per example (with shape [batch_size]). See <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/losses\/softmax_cross_entropy\" rel=\"noreferrer\">documentation here<\/a>.<\/p>\n"},{"AnswerId":"38912982","CreationDate":"2016-08-12T08:07:02.707","ParentId":null,"OwnerUserId":"6708365","Title":null,"Body":"<p>The code you proposed seems wrong to me. \nThe loss should be multiplied by the weight, I agree.<\/p>\n\n<p>But if you multiply the logit by the class weights, you end with:<\/p>\n\n<pre><code>weights[class] * -x[class] + log( \\sum_j exp(x[j] * weights[class]) )\n<\/code><\/pre>\n\n<p>The <strong>second term<\/strong> is not equal to: <\/p>\n\n<pre><code>weights[class] * log(\\sum_j exp(x[j]))\n<\/code><\/pre>\n\n<p>To show this, we can be rewrite the latter as: <\/p>\n\n<pre><code>log( (\\sum_j exp(x[j]) ^ weights[class] )\n<\/code><\/pre>\n\n<p>So here is the code I'm proposing:<\/p>\n\n<pre><code>ratio = 31.0 \/ (500.0 + 31.0)\nclass_weight = tf.constant([[ratio, 1.0 - ratio]])\nlogits = ... # shape [batch_size, 2]\n\nweight_per_label = tf.transpose( tf.matmul(labels\n                           , tf.transpose(class_weight)) ) #shape [1, batch_size]\n# this is the weight for each datapoint, depending on its label\n\nxent = tf.mul(weight_per_label\n         , tf.nn.softmax_cross_entropy_with_logits(logits, labels, name=\"xent_raw\") #shape [1, batch_size]\nloss = tf.reduce_mean(xent) #shape 1\n<\/code><\/pre>\n"},{"AnswerId":"44371783","CreationDate":"2017-06-05T14:56:50.927","ParentId":null,"OwnerUserId":"1853062","Title":null,"Body":"<p>You can check the guides at tensorflow <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.losses\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.losses<\/a> <\/p>\n\n<p>...<\/p>\n\n<p>While specifying a scalar loss rescales the loss over the entire batch, we sometimes want to rescale the loss per batch sample. For example, if we have certain examples that matter more to us to get correctly, we might want to have a higher loss that other samples whose mistakes matter less. In this case, we can provide a weight vector of length batch_size which results in the loss for each sample in the batch being scaled by the corresponding weight element. For example, consider the case of a classification problem where we want to maximize our accuracy but we especially interested in obtaining high accuracy for a specific class:<\/p>\n\n<pre><code>inputs, labels = LoadData(batch_size=3)\nlogits = MyModelPredictions(inputs)\n\n# Ensures that the loss for examples whose ground truth class is `3` is 5x\n# higher than the loss for all other examples.\nweight = tf.multiply(4, tf.cast(tf.equal(labels, 3), tf.float32)) + 1\n\nonehot_labels = tf.one_hot(labels, num_classes=5)\ntf.contrib.losses.softmax_cross_entropy(logits, onehot_labels, weight=weight)\n<\/code><\/pre>\n"},{"AnswerId":"41254526","CreationDate":"2016-12-21T03:14:59.623","ParentId":null,"OwnerUserId":"7120165","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/weighted_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.weighted_cross_entropy_with_logits()<\/code><\/a> and set <code>pos_weight<\/code> to 1 \/ (expected ratio of positives).<\/p>\n"}]}
{"QuestionId":35156197,"AnswerCount":1,"Tags":"<python><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2016-02-02T14:32:55.463","AcceptedAnswerId":35157455.0,"OwnerUserId":4201966.0,"Title":"How to add external features to input images when training GoogLeNet in Caffe?","Body":"<p>I have been using the Caffe framework for fine-tuning the GoogLeNet model using images from my own dataset. I think that I could improve the results if I could add some externally extracted features (like SIFT based or HoG...) to the training images in the training process.<\/p>\n\n<p>Is there a way I can expand the training dataset to IMAGES + ADDITIONAL FEATURES and train the same model in Caffe using Python?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"35157455","CreationDate":"2016-02-02T15:28:58.410","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe is not limited to images only. In fact, it can accept (almost) any type of input. You can store your training data in a <a href=\"https:\/\/www.hdfgroup.org\/HDF5\/\" rel=\"nofollow\"><code>hdf5<\/code><\/a> binary file and read it in caffe using <code>\"HDF5Data\"<\/code> layer. The hdf5 file can store several variables, e.g., <code>\"image\"<\/code>, <code>\"label\"<\/code> and addtional <code>\"img_features\"<\/code>. Thus, your input layer to caffe will look something like:<\/p>\n\n<pre><code>layer {\n  name: \"input\"\n  type: \"HDF5Data\"\n  top: \"image\"\n  top: \"img_features\"\n  top: \"label\"\n  ...\n}\n<\/code><\/pre>\n\n<p>You can use python's <a href=\"http:\/\/www.h5py.org\/\" rel=\"nofollow\"><code>h5py<\/code><\/a> python package to create the input <code>hdf5<\/code> files. <\/p>\n"}]}
{"QuestionId":35157696,"AnswerCount":0,"Tags":"<neural-network><theano><conv-neural-network><keras>","CreationDate":"2016-02-02T15:40:45.090","AcceptedAnswerId":null,"OwnerUserId":5843689.0,"Title":"Regularisers in Keras vs. Caffe","Body":"<p>I've edited an AlexNet in KERAS, trying to learn a single class (i.e. dogs\/faces or not dog\/faces).\nI have training samples of <strong>correct<\/strong> images so the Ytrain is <strong>[1, 0]<\/strong>, and <strong>incorrect<\/strong> images so the Ytrain is <strong>[0,1]<\/strong>.\nThe goal is to know if the image is a dog (for example).<\/p>\n\n<p>while training the loss was very big ~ 100,000,000<\/p>\n\n<p>so i deleted the regularisers lines and it worked...<\/p>\n\n<pre><\/pre>\n\n<p>In Caffe the model works fine with these lines in the convolution layers:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the model in Keras(image size 3,68,56):<\/p>\n\n<pre><\/pre>\n\n<p>Y is that?? <\/p>\n","answers":[]}
{"QuestionId":35158365,"AnswerCount":2,"Tags":"<machine-learning><neural-network><conv-neural-network><torch><gradient-descent>","CreationDate":"2016-02-02T16:12:17.060","AcceptedAnswerId":35162884.0,"OwnerUserId":4299527.0,"Title":"Will larger batch size make computation time less in machine learning?","Body":"<p>I am trying to tune the hyper parameter i.e <strong>batch size<\/strong> in CNN.I have a computer of corei7,RAM 12GB and i am training a CNN network with CIFAR-10 dataset which can be found in this <a href=\"http:\/\/torch.ch\/blog\/2015\/07\/30\/cifar.html\" rel=\"nofollow noreferrer\">blog<\/a>.<br><br><em>Now At first what i have read and learnt about batch size in machine learning:<\/em><\/p>\n\n<blockquote>\n  <p>let's first suppose that we're doing online learning, i.e. that we're\n  using a mini\u00adbatch size of 1. The obvious worry about online learning\n  is that using mini\u00adbatches which contain just a single training\n  example will cause significant errors in our estimate of the gradient.\n  In fact, though, the errors turn out to not be such a problem. The\n  reason is that the individual gradient estimates don't need to be\n  super\u00adaccurate. All we need is an estimate accurate enough that our\n  cost function tends to keep decreasing. It's as though you are trying\n  to get to the North Magnetic Pole, but have a wonky compass that's\n  10\u00ad-20 degrees off each time you look at it. Provided you stop to\n  check the compass frequently, and the compass gets the direction right\n  on average, you'll end up at the North Magnetic Pole just\n  fine.<br><br><\/p>\n  \n  <p>Based on this argument, it sounds as though we should use online\n  learning. In fact, the situation turns out to be more complicated than\n  that.As we know we can use matrix techniques to compute the gradient\n  update for all examples in a mini\u00adbatch simultaneously, rather than\n  looping over them. Depending on the details of our hardware and linear\n  algebra library this can make it quite a bit faster to compute the\n  gradient estimate for a mini\u00adbatch of (for example) size 100 , rather\n  than computing the mini\u00adbatch gradient estimate by looping over the\n  100 training examples separately. It might take (say) only 50 times as\n  long, rather than 100 times as long.Now, at first it seems as though\n  this doesn't help us that much.<br><br><\/p>\n  \n  <p>With our mini\u00adbatch of size 100 the learning rule for the weights\n  looks like:<a href=\"https:\/\/i.stack.imgur.com\/QDhgB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QDhgB.png\" alt=\"enter image description here\"><\/a><br><\/p>\n  \n  <p>where the sum is over training examples in the mini\u00adbatch. This is\n  versus<a href=\"https:\/\/i.stack.imgur.com\/X4CkK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/X4CkK.png\" alt=\"enter image description here\"><\/a><br> for online learning.\n  Even if it only takes 50 times as long to do the mini\u00adbatch update, it\n  still seems likely to be better to do online learning, because we'd be\n  updating so much more frequently. Suppose, however, that in the\n  mini\u00adbatch case we increase the learning rate by a factor 100, so the\n  update rule becomes<br><a href=\"https:\/\/i.stack.imgur.com\/KMnF1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KMnF1.png\" alt=\"enter image description here\"><\/a><br>\n  That's a lot like doing separate instances of online learning with a\n  learning rate of . But it only takes 50 times as long as doing a\n  single instance of online learning. Still, it seems distinctly\n  possible that using the larger mini\u00adbatch would speed things up.<\/p>\n<\/blockquote>\n\n<p><br><br><\/p>\n\n<p>Now i tried with  and ran a sample program and set the batch size  at first.I noted down the training time needed for the full dataset.Then i increased the batch size and i noticed that it became faster.\n<br>\nBut in case of training with this  <a href=\"https:\/\/github.com\/szagoruyko\/cifar.torch\/blob\/master\/models\/vgg_bn_drop.lua\" rel=\"nofollow noreferrer\">code<\/a> and <a href=\"https:\/\/github.com\/szagoruyko\/cifar.torch\" rel=\"nofollow noreferrer\">github link<\/a> changing the batch size doesn't decrease the training time.It remained same if i use 30 or 128 or 64.They are saying that they got  accuracy.After two or three epoch they have got above  accuracy.But when i ran the code in my computer without changing anything other than the batch size i got worse result after 10 epoch like only 28% and test accuracy stuck there in the next epochs.Then i thought since they have used batch size of 128 i need to use that.Then i used the same but it became more worse only give 11% after 10 epoch and stuck in there.<strong>Why is that??<\/strong><\/p>\n","answers":[{"AnswerId":"35162884","CreationDate":"2016-02-02T20:09:49.897","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>Neural networks learn by gradient descent an error function in the weight space which is parametrized by the training examples. This means the variables are the weights of the neural network. The function is \"generic\" and becomes specific when you use training examples. The \"correct\" way would be to use all training examples to make the specific function. This is called \"batch gradient descent\" and is usually not done for two reasons:<\/p>\n\n<ol>\n<li>It might not fit in your RAM (usually GPU, as for neural networks you get a huge boost when you use the GPU).<\/li>\n<li>It is actually not necessary to use all examples.<\/li>\n<\/ol>\n\n<p>In machine learning problems, you usually have several thousands of training examples. But the error surface might look similar when you only look at a few (e.g. 64, 128 or 256) examples.<\/p>\n\n<p>Think of it as a photo: To get an idea of what the photo is about, you usually don't need a 2500x1800px resolution. A 256x256px image will give you a good idea what the photo is about. However, you miss details.<\/p>\n\n<p>So imagine gradient descent to be a walk on the error surface: You start on one point and you want to find the lowest point. To do so, you walk down. Then you check your height again, check in which direction it goes down and make a \"step\" (of which the size is determined by the learning rate and a couple of other factors) in that direction. When you have mini-batch training instead of batch-training, you walk down on a different error surface. In the low-resolution error surface. It might actually go up in the \"real\" error surface. But overall, you will go in the right direction. And you can make single steps much faster!<\/p>\n\n<p>Now, what happens when you make the resolution lower (the batch size smaller)?<\/p>\n\n<p>Right, your image of what the error surface looks like gets less accurate. How much this affects you depends on factors like:<\/p>\n\n<ul>\n<li>Your hardware\/implementation<\/li>\n<li>Dataset: How complex is the error surface and how good it is approximated by only a small portion?<\/li>\n<li>Learning: How exactly are you learning (momentum? newbob? rprop?)<\/li>\n<\/ul>\n"},{"AnswerId":"46590360","CreationDate":"2017-10-05T16:16:19.790","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<p>I'd like to add to what's been already said here that larger batch size <em>is not always<\/em> good for generalization. I've seen these cases myself, when an increase in batch size hurt validation accuracy, particularly for CNN working with CIFAR-10 dataset.<\/p>\n\n<p>From <a href=\"https:\/\/arxiv.org\/abs\/1609.04836\" rel=\"nofollow noreferrer\">\"On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima\"<\/a>:<\/p>\n\n<blockquote>\n  <p>The stochastic gradient descent (SGD) method and its variants are\n  algorithms of choice for many Deep Learning tasks. These methods\n  operate in a small-batch regime wherein a fraction of the training\n  data, say 32\u2013512 data points, is sampled to compute an approximation\n  to the gradient. <strong>It has been observed in practice that when using a\n  larger batch there is a degradation in the quality of the model, as\n  measured by its ability to generalize<\/strong>. We investigate the cause for\n  this generalization drop in the large-batch regime and present\n  numerical evidence that supports the view that large-batch methods\n  tend to converge to sharp minimizers of the training and testing\n  functions\u2014and as is well known, sharp minima lead to poorer\n  generalization. In contrast, small-batch methods consistently converge\n  to flat minimizers, and our experiments support a commonly held view\n  that this is due to the inherent noise in the gradient estimation. We\n  discuss several strategies to attempt to help large-batch methods\n  eliminate this generalization gap.<\/p>\n<\/blockquote>\n\n<p>Bottom-line: you should tune the batch size, just like <a href=\"https:\/\/stackoverflow.com\/questions\/41860817\/hyperparameter-optimization-for-deep-learning-structures-using-bayesian-optimiza\/46318446\">any other hyperparameter<\/a>, to find an optimal value.<\/p>\n"}]}
{"QuestionId":35158604,"AnswerCount":0,"Tags":"<python><opencv><caffe>","CreationDate":"2016-02-02T16:23:21.710","AcceptedAnswerId":null,"OwnerUserId":5873708.0,"Title":"Import Caffe: One of the two will be used. Which one is undefined","Body":"<p>I got this when trying to import caffe into Spyder(python2.7).\nCaffe was successfully installed and it went through test.\nAny idea how to solve it?<\/p>\n\n<blockquote>\n  <p>objc[3163]: Class  is implemented in both\n  \/Users\/xxx\/anaconda\/lib\/libopencv_highgui.2.4.8.dylib and\n  \/usr\/local\/opt\/opencv\/lib\/libopencv_highgui.2.4.dylib.<br>\n  One of the two will be used. Which one is undefined.<br>\n      objc[3163]: Class  is implemented in both \/Users\/xxx\/anaconda\/lib\/libopencv_highgui.2.4.8.dylib and\n  \/usr\/local\/opt\/opencv\/lib\/libopencv_highgui.2.4.dylib.<br>\n    One of the two will be used. Which one is undefined.<br>\n      objc[3163]: Class  is implemented in both \/Users\/xxx\/anaconda\/lib\/libopencv_highgui.2.4.8.dylib and\n  \/usr\/local\/opt\/opencv\/lib\/libopencv_highgui.2.4.dylib.<br>\n  One of the two will be used. Which one is undefined.<br>\n      objc[3163]: Class  is implemented in both \/Users\/xxx\/anaconda\/lib\/libopencv_highgui.2.4.8.dylib and\n  \/usr\/local\/opt\/opencv\/lib\/libopencv_highgui.2.4.dylib.<br>\n  One of the two will be used. Which one is undefined.<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":35159960,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2016-02-02T17:26:37.107","AcceptedAnswerId":35160145.0,"OwnerUserId":988029.0,"Title":"Tensorflow csv reader error: \"Quote inside a string has to be escaped by another quote\"","Body":"<p>I have a csv file, which I'm pretty sure doesn't have \" in, which I'm trying to reading using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run it I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Which I think I could debug, but I can't find where it references which entry in the csv file it's having problems with. It's quite a large csv file, and the first 100 or so entries don't have this problem. As I said, I can't find any \", and ' seems to parse fine in tests. Is there some way I can find the troublesome entry? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"35160145","CreationDate":"2016-02-02T17:35:08.207","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>One way to find the troublesome entry would be to add a <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/control_flow_ops.html#Print\" rel=\"nofollow\"><code>tf.Print()<\/code><\/a> operator before the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/io_ops.html#decode_csv\" rel=\"nofollow\"><code>tf.decode_csv()<\/code><\/a>:<\/p>\n\n<pre><code># ...\n\n# Prints out the contents of `key` and `value` every time the op executes.\nvalue = tf.Print(value, [key, value])\n\ncol1, col2, col3 = tf.decode_csv(\n    value, record_defaults=record_defaults, field_delim=\" \")\n\n# ...\n<\/code><\/pre>\n\n<p>The last logged entry before the failure should indicate which input was invalid. Hopefully the root cause becomes apparent when you make this modification.<\/p>\n"}]}
{"QuestionId":35161771,"AnswerCount":1,"Tags":"<neural-network><torch>","CreationDate":"2016-02-02T19:03:31.660","AcceptedAnswerId":35253780.0,"OwnerUserId":45843.0,"Title":"Packaged version of Torch for neural networks","Body":"<p>I tried installing Torch and the  package with a view to carrying out some experiments on neural networks, but the recommended procedure failed with a compile error while trying to install the latter. The procedure involves taking some git snapshots, so there is some degree of non-determinism.<\/p>\n\n<p>Does there exist a package of Torch and  snapshotted from known compatible versions, such that you could apply the package to Ubuntu X.Y and it would always work (or always fail, as the case may be)?<\/p>\n","answers":[{"AnswerId":"35253780","CreationDate":"2016-02-07T12:58:21.540","ParentId":null,"OwnerUserId":"3167182","Title":null,"Body":"<p>From the Torch website:<\/p>\n\n<p><em>Torch can be installed to your home folder in ~\/torch by running these three commands:<\/em><\/p>\n\n<pre><code># in a terminal, run the commands\ncurl -s https:\/\/raw.githubusercontent.com\/torch\/ezinstall\/master\/install-deps | bash\ngit clone https:\/\/github.com\/torch\/distro.git ~\/torch --recursive\ncd ~\/torch; .\/install.sh\n<\/code><\/pre>\n\n<p><em>The first script installs the basic package dependencies that LuaJIT and Torch require. The second script installs LuaJIT, LuaRocks, and then uses LuaRocks (the lua package manager) to install core packages like torch, nn and paths, as well as a few other packages.<\/em><\/p>\n\n<p>I've installed Torch a few times now and found this most successful - no compilation required.<\/p>\n"}]}
{"QuestionId":35161956,"AnswerCount":1,"Tags":"<python><arrays><numpy><theano>","CreationDate":"2016-02-02T19:14:24.023","AcceptedAnswerId":35178274.0,"OwnerUserId":3778384.0,"Title":"How to debug shape mismatch of theano tensor?","Body":"<p>I defined a Theano tensor  and used it as an argument of a theano function . <\/p>\n\n<p>When I now call  with a numpy array  of shape (100,3), I'd expect that  would have the shape (100,). <\/p>\n\n<p>However, the error message shows that the shape is (1,100). How can I examine the function mismatch step by step?<\/p>\n","answers":[{"AnswerId":"35178274","CreationDate":"2016-02-03T13:22:27.123","ParentId":null,"OwnerUserId":"3778384","Title":null,"Body":"<p>Thanks to the useful hints in the comments I was able to debug the shape mismatch. I set up another theano debug function with the same inputs and a custom output, which I could examine with the debugger, e.g.:<\/p>\n\n<pre><code># define a function ...\ninputs = T.matrix('inputs')\ndebug_out = T.sum(fancy_expression(inputs)) # expression to debug\ndebug_fn = theano.function(\n        inputs=[inputs], \n        outputs=debug_out,\n        on_unused_input='ignore' # to suppress unused input exeptions\n)\n\n# ... and debug it here\nresult = debug_fn(np.empty((100,3)))\n<\/code><\/pre>\n\n<p>Thanks again to @ali_m<\/p>\n"}]}
{"QuestionId":35163371,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><tensorflow>","CreationDate":"2016-02-02T20:39:56.193","AcceptedAnswerId":35166790.0,"OwnerUserId":3915457.0,"Title":"Tensorflow: Extracting the features of a trained model","Body":"<p>I have an implementation of the AlexNet. I'm interested in extracting the vector of features of a trained model <em>before the fully-connected classification layers<\/em><\/p>\n\n<ol>\n<li><p>I want to first train the model (below I included the evaluation methods for training and testing).<\/p><\/li>\n<li><p>How do I get a list of final output feature vectors (during the forward pass) for all the images in the training\/test set before they get classified?<\/p><\/li>\n<\/ol>\n\n<p>Here is the code (full version available <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/alexnet.py\" rel=\"nofollow\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/alexnet.py<\/a>) :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35166790","CreationDate":"2016-02-03T01:12:33.477","ParentId":null,"OwnerUserId":"5530615","Title":null,"Body":"<p>It sounds like you want the value of dense2 from alex_net()?  If so, you will need to return that from alex_net() in addition to out, so<\/p>\n\n<pre><code>return out\n<\/code><\/pre>\n\n<p>becomes<\/p>\n\n<pre><code>return dense2, out\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>pred = alex_net(x, weights, biases, keep_prob)\n<\/code><\/pre>\n\n<p>becomes<\/p>\n\n<pre><code>before_classification_layer, pred = alex_net(...)\n<\/code><\/pre>\n\n<p>Then you can fetch before_classification_layer when calling <code>sess.run()<\/code> whenever you want that value.  See <code>tf.Session.run<\/code> in <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/client.html#Session.run<\/a>.  Note that the fetches may be a list, so to avoid evaluating your graph twice in your example code, you can do<\/p>\n\n<pre><code># Calculate batch accuracy and loss\nacc, loss = sess.run([accuracy, cost], feed_dict={...})\n<\/code><\/pre>\n\n<p>instead of<\/p>\n\n<pre><code># Calculate batch accuracy\nacc = sess.run(accuracy, feed_dict={...})\n# Calculate batch loss\nloss = sess.run(cost, feed_dict={...})\n<\/code><\/pre>\n\n<p>(Adding <code>before_classification_layer<\/code> to that list when desired.)<\/p>\n"}]}
{"QuestionId":35163744,"AnswerCount":1,"Tags":"<python><theano><hessian-matrix>","CreationDate":"2016-02-02T21:02:20.350","AcceptedAnswerId":35165493.0,"OwnerUserId":2255757.0,"Title":"theano T.grad() error: not part of computational graph","Body":"<p>code: <\/p>\n\n<pre><\/pre>\n\n<p>output:<\/p>\n\n<pre><\/pre>\n\n<p>how is a derivative of the graph not part of the graph? Is this why scan is used to compute the hessian?<\/p>\n","answers":[{"AnswerId":"35165493","CreationDate":"2016-02-02T23:03:31.167","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>Here:<\/p>\n\n<pre><code>d2y = T.grad(loss, dy)\n<\/code><\/pre>\n\n<p>you are attempting to compute the gradient of the loss with respect to <code>dy<\/code>. However the loss depends only on the values of <code>a<\/code> and <code>b<\/code> and not <code>dy<\/code>, hence the error. It only makes sense to compute partial derivatives of the loss with respect to parameters that actually affect its value.<\/p>\n\n<p>The easiest way to compute the Hessian in Theano is to use the <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/gradient.html#theano.gradient.hessian\" rel=\"nofollow\"><code>theano.gradient.hessian<\/code><\/a> convenience function:<\/p>\n\n<pre><code>d2y = theano.gradient.hessian(loss, a)\n<\/code><\/pre>\n\n<p>See the documentation <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html#computing-the-hessian\" rel=\"nofollow\">here<\/a> for an alternative manual method that uses  a combination of <code>theano.grad<\/code> and <code>theano.scan<\/code>.<\/p>\n\n<p>In your example the Hessian will be a 3x3 matrix of zeros, since the partial derivative of the loss w.r.t. <code>a<\/code> is independent of <code>a<\/code> (it's just a vector of ones).<\/p>\n"}]}
{"QuestionId":35163789,"AnswerCount":4,"Tags":"<python><numpy><theano><tensorflow>","CreationDate":"2016-02-02T21:04:51.817","AcceptedAnswerId":35164907.0,"OwnerUserId":5874853.0,"Title":"Theano Dimshuffle equivalent in Google's TensorFlow?","Body":"<p>I have seen that transpose and reshape together can help but I don't know how to use.<\/p>\n\n<p>Eg. dimshuffle(0, 'x')<\/p>\n\n<p>What is its equivalent by using transpose and reshape? or is there a better way?\nThank you.<\/p>\n","answers":[{"AnswerId":"46177952","CreationDate":"2017-09-12T13:38:10.037","ParentId":null,"OwnerUserId":"7684762","Title":null,"Body":"<p>If tensorflow is your backend<\/p>\n\n<pre><code>from keras import baskend as K\nK.permute_dimension should do\n<\/code><\/pre>\n"},{"AnswerId":"35164907","CreationDate":"2016-02-02T22:20:24.420","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are three relevant ops for implementing Theano's <code>dimshuffle<\/code> in TensorFlow:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#transpose\"><code>tf.transpose()<\/code><\/a> is used to permute the dimensions of a tensor. If the pattern specified in the arguments to <code>dimshuffle<\/code> is a permutation of the input tensor's dimensions (i.e. there is no <code>'x'<\/code> or missing dimension) you can use <code>tf.transpose()<\/code> to implement <code>dimshuffle()<\/code>.<\/p><\/li>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#expand_dims\"><code>tf.expand_dims()<\/code><\/a> is used to add one or more size-1 dimensions to a tensor. This handles the case where <code>'x'<\/code> is specified as part of the <code>dimshuffle()<\/code> pattern, but does not reorder the existing dimensions.<\/p><\/li>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/array_ops.html#squeeze\"><code>tf.squeeze()<\/code><\/a> is used to remove one or more size-1 dimensions from a tensor. This handles the case where a dimension is omitted from a <code>dimshuffle()<\/code> pattern, but it does not reorder the existing dimensions.<\/p><\/li>\n<\/ul>\n\n<p>Assuming that the input is a vector, your example (<code>dimshuffle(0, 'x')<\/code>) can be expressed using <code>tf.expand_dims()<\/code> only:<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32, [None])  # Defines an arbitrary-sized vector.\nresult = tf.expand_dims(input, 1)\n\nprint result.get_shape()  # ==&gt; TensorShape([Dimension(None), Dimension(1)])\n<\/code><\/pre>\n\n<p>Taking a more complicated example, <code>dimshuffle(1, 'x', 0)<\/code> applied to a matrix would be:<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32, [128, 32])  # Defines a matrix.\noutput = tf.expand_dims(tf.transpose(input, [1, 0]), 1)\n\nprint output.get_shape()\n# ==&gt; TensorShape([Dimension(32), Dimension(1), Dimension(128)])\n<\/code><\/pre>\n"},{"AnswerId":"43605483","CreationDate":"2017-04-25T08:32:10.690","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>I implemented <code>dimshuffle<\/code> for TensorFlow in <a href=\"https:\/\/github.com\/rwth-i6\/returnn\/\" rel=\"nofollow noreferrer\">our framework Returnn<\/a> (<a href=\"https:\/\/github.com\/rwth-i6\/returnn\/blob\/master\/TFUtil.py\" rel=\"nofollow noreferrer\">here<\/a>). The code is this:<\/p>\n\n<pre><code>def expand_multiple_dims(x, axes, name=\"expand_multiple_dims\"):\n  \"\"\"\n  :param tf.Tensor x:\n  :param list[int]|tuple[int] axes: after completion, tf.shape(y)[axis] == 1 for axis in axes\n  :param str name: scope name\n  :return: y where we have a new broadcast axis for each axis in axes\n  :rtype: tf.Tensor\n  \"\"\"\n  with tf.name_scope(name):\n    for i in sorted(axes):\n      x = tf.expand_dims(x, axis=i, name=\"expand_axis_%i\" % i)\n    return x\n\n\ndef dimshuffle(x, axes, name=\"dimshuffle\"):\n  \"\"\"\n  Like Theanos dimshuffle.\n  Combines tf.transpose, tf.expand_dims and tf.squeeze.\n\n  :param tf.Tensor x:\n  :param list[int|str]|tuple[int|str] axes:\n  :param str name: scope name\n  :rtype: tf.Tensor\n  \"\"\"\n  with tf.name_scope(name):\n    assert all([i == \"x\" or isinstance(i, int) for i in axes])\n    real_axes = [i for i in axes if isinstance(i, int)]\n    bc_axes = [i for (i, j) in enumerate(axes) if j == \"x\"]\n    if x.get_shape().ndims is None:\n      x_shape = tf.shape(x)\n      x = tf.reshape(x, [x_shape[i] for i in range(max(real_axes) + 1)])  # will have static ndims\n    assert x.get_shape().ndims is not None\n\n    # First squeeze missing axes.\n    i = 0\n    while i &lt; x.get_shape().ndims:\n      if i not in real_axes:\n        x = tf.squeeze(x, axis=i)\n        real_axes = [(j if (j &lt; i) else (j - 1)) for j in real_axes]\n      else:\n        i += 1\n\n    # Now permute.\n    assert list(sorted(real_axes)) == list(range(x.get_shape().ndims))\n    if real_axes != list(range(x.get_shape().ndims)):\n      x = tf.transpose(x, real_axes)\n\n    # Now add broadcast dimensions.\n    if bc_axes:\n      x = expand_multiple_dims(x, bc_axes)\n    assert len(axes) == x.get_shape().ndims\n    return x\n<\/code><\/pre>\n"},{"AnswerId":"50763868","CreationDate":"2018-06-08T15:25:55.173","ParentId":null,"OwnerUserId":"6108816","Title":null,"Body":"<p>tf.transpose is probably what you are looking for.   it takes an arbitrary permutation.<\/p>\n"}]}
{"QuestionId":35163818,"AnswerCount":1,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2016-02-02T21:06:59.137","AcceptedAnswerId":37738355.0,"OwnerUserId":3368154.0,"Title":"Theano matrix multiplication of 2d matrix to give 3d matrix","Body":"<p>I want to do exactly what the following question is trying to do in numpy:<br>\n<a href=\"https:\/\/stackoverflow.com\/questions\/35162197\/numpy-matrix-multiplication-of-2d-matrix-to-give-3d-matrix\/35162306?noredirect=1#comment58044934_35162306\">Numpy matrix multiplication of 2d matrix to give 3d matrix<\/a><\/p>\n\n<p>I have written the following code (as per the suggestion for numpy case), assuming that broadcast works in theano as well: <\/p>\n\n<pre><\/pre>\n\n<p>But it is throwing following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37738355","CreationDate":"2016-06-10T00:46:34.110","ParentId":null,"OwnerUserId":"3150181","Title":null,"Body":"<p>As per @Divakar s comment, change it to<\/p>\n\n<pre><code>x[:,:,None] * y[:,None,:]\n<\/code><\/pre>\n\n<p>Just posting it as an answer so that people know there's an answer to this.<\/p>\n"}]}
{"QuestionId":35164529,"AnswerCount":7,"Tags":"<python><tensorflow>","CreationDate":"2016-02-02T21:55:18.603","AcceptedAnswerId":35164992.0,"OwnerUserId":2245776.0,"Title":"In TensorFlow is there any way to just initialize uninitialised variables?","Body":"<p>The standard way of initializing variables in TensorFlow is<\/p>\n\n<pre><\/pre>\n\n<p>After running some learning for a while I create a new set of variables but once I initialize them it resets all my existing variables. At the moment my way around this is to save all the variable I need and then reapply them after the tf.initalize_all_variables call. This works but is a bit ugly and clunky. I cannot find anything like this in the docs...<\/p>\n\n<p>Does anyone know of any good way to just initialize the uninitialized variables?   <\/p>\n","answers":[{"AnswerId":"43324104","CreationDate":"2017-04-10T13:10:26.843","ParentId":null,"OwnerUserId":"4776477","Title":null,"Body":"<p>I think that the easiest way is to create all the training operators first and initialize variables afterwards.<\/p>\n\n<p>For example, I solved the problem of layer-wise pretraining with Adam Optimizer in the following way:<\/p>\n\n<pre><code> # create an optimizer\n pretrain_optimizer =  tf.train.AdamOptimizer(learning_rate=learning_rate)\n\n # Make an array of the trainers for all the layers\n trainers=[pretrain_optimizer.minimize(loss_reconstruction(ae.run_less_layers(ae._input_, i+1), ae.run_less_layers(ae._input_, i+1, is_target=True)), global_step=tf.contrib.framework.get_or_create_global_step(), name='Layer_wise_optimizer_'+str(i)) for i in xrange(len(ae_shape) - 2)]\n\n # Initialize all the variables\n sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n"},{"AnswerId":"43601894","CreationDate":"2017-04-25T04:54:18.033","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>TF <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/state_ops\" rel=\"noreferrer\">does not have a function that does exactly what you want<\/a>, but you can easily write one:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef initialize_uninitialized(sess):\n    global_vars          = tf.global_variables()\n    is_not_initialized   = sess.run([tf.is_variable_initialized(var) for var in global_vars])\n    not_initialized_vars = [v for (v, f) in zip(global_vars, is_not_initialized) if not f]\n\n    print [str(i.name) for i in not_initialized_vars] # only for testing\n    if len(not_initialized_vars):\n        sess.run(tf.variables_initializer(not_initialized_vars))\n<\/code><\/pre>\n\n<p>Here I extract all <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/global_variables\" rel=\"noreferrer\">global variables<\/a>, iterate all of them and check whether they <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/is_variable_initialized\" rel=\"noreferrer\">are already initialized<\/a>. After this I get a list of uninitialized variables which I <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variables_initializer\" rel=\"noreferrer\">initialize<\/a>. I also print variables that I am going to initialize for debugging purposes.<\/p>\n\n<hr>\n\n<p>You can easily verify that it works as expected:<\/p>\n\n<pre><code>a = tf.Variable(3, name='my_var_a')\nb = tf.Variable(4, name='my_var_b')\n\nsess = tf.Session()\ninitialize_uninitialized(sess)\ninitialize_uninitialized(sess)\n\nc = tf.Variable(5, name='my_var_a') # the same name, will be resolved to different name\nd = tf.Variable(6, name='my_var_d')\ninitialize_uninitialized(sess)\n\nprint '\\n\\n', sess.run([a, b, c, d])\n<\/code><\/pre>\n\n<p>This will print all the unitialized variables before initializing them and the last sess.run will make sure persuade you that all variables are initialized.<\/p>\n\n<hr>\n\n<p>You can also use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/report_uninitialized_variables\" rel=\"noreferrer\"><code>tf.report_uninitialized_variables()<\/code><\/a> to write a similar function. A sketch of it is <a href=\"https:\/\/stackoverflow.com\/a\/44276281\/1090562\">here<\/a>.<\/p>\n"},{"AnswerId":"47271489","CreationDate":"2017-11-13T18:50:50.600","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Btw, if you want to initialize only a single tensor (for e.g.  <code>tf.Variable<\/code>) that hasn't been initialized using <code>tf.global_variables_initializer()<\/code>, then you can use <code>your_tensor.initializer<\/code> in the <code>sess.run()<\/code> as in the following example:<\/p>\n\n<pre><code>In [196]: weights = tf.Variable(tf.zeros(shape=(3, 4)), name='weights')\n\nIn [197]: with tf.Session() as sess:\n     ...:     sess.run(weights.initializer)\n     ...:     print(weights.eval())\n     ...:     \n\n# the result\n[[ 0.  0.  0.  0.]\n [ 0.  0.  0.  0.]\n [ 0.  0.  0.  0.]]\n<\/code><\/pre>\n"},{"AnswerId":"40676934","CreationDate":"2016-11-18T12:17:02.643","ParentId":null,"OwnerUserId":"4841393","Title":null,"Body":"<p>I've come up with a method for TensorFlow r0.11:<\/p>\n\n<pre><code>def get_uninitialized_variables(variables=None):\n    \"\"\"Get uninitialized variables as a list.\n\n    Parameters\n    ----------\n    variables : collections.Iterable[tf.Variable]\n        Return only uninitialized variables within this collection.\n        If not specified, will return all uninitialized variables.\n\n    Returns\n    -------\n    list[tf.Variable]\n    \"\"\"\n    sess = tf.get_default_session()\n    if variables is None:\n        variables = tf.all_variables()\n    else:\n        variables = list(variables)\n    init_flag = sess.run(\n        tf.pack([tf.is_variable_initialized(v) for v in variables]))\n    return [v for v, f in zip(variables, init_flag) if not f]\n<\/code><\/pre>\n"},{"AnswerId":"37291254","CreationDate":"2016-05-18T05:47:52.413","ParentId":null,"OwnerUserId":"4794308","Title":null,"Body":"<p>For the case @Poik mentioned, when variables are created by optimizers so that they cannot be accessed directly, a cleaner solution is to use <code>tf.train.Optimizer.get_slot<\/code>. <\/p>\n\n<p>Some optimizer subclasses, such as <code>MomentumOptimizer<\/code> and <code>AdagradOptimizer<\/code> allocate and manage additional variables associated with the variables to train. These are called Slots. You can use <code>tf.train.Optimizer.get_slot_names()<\/code> to get all slot names a optimizer has and then use <code>tf.train.Optimizer.get_slot<\/code> to get the variable allocated for these slots.<\/p>\n"},{"AnswerId":"35164992","CreationDate":"2016-02-02T22:26:12.940","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is no elegant* way to enumerate the uninitialized variables in a graph. However, if you have access to the new variable objects&mdash;let's call them <code>v_6<\/code>, <code>v_7<\/code>, and <code>v_8<\/code>&mdash;you can selectively initialize them using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/state_ops.html#initialize_variables\" rel=\"noreferrer\"><code>tf.initialize_variables()<\/code><\/a>:<\/p>\n\n<pre><code>init_new_vars_op = tf.initialize_variables([v_6, v_7, v_8])\nsess.run(init_new_vars_op)\n<\/code><\/pre>\n\n<hr>\n\n<p>*&nbsp;A process of trial and error could be used to identify the uninitialized variables, as follows:<\/p>\n\n<pre><code>uninitialized_vars = []\nfor var in tf.all_variables():\n    try:\n        sess.run(var)\n    except tf.errors.FailedPreconditionError:\n        uninitialized_vars.append(var)\n\ninit_new_vars_op = tf.initialize_variables(uninitialized_vars)\n# ...\n<\/code><\/pre>\n\n<p>...however, I would not condone such behavior :-).<\/p>\n"},{"AnswerId":"35618160","CreationDate":"2016-02-25T04:10:38.023","ParentId":null,"OwnerUserId":"786020","Title":null,"Body":"<p><strong>UPDATE:<\/strong> TensorFlow 0.9 has a new method that \"fixes\" all this <strong>but only if you're using a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/state_ops.html#VariableScope\" rel=\"noreferrer\">VariableScope<\/a> with <code>reuse<\/code> set to <code>True<\/code><\/strong>. <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/state_ops.html#report_uninitialized_variables\" rel=\"noreferrer\">tf.report_uninitialized_variables<\/a> which can be used in one line with <code>sess.run( tf.initialize_variables( list( tf.get_variable(name) for name in sess.run( tf.report_uninitialized_variables( tf.all_variables( ) ) ) ) ) )<\/code><\/p>\n\n<p>or more intelligently through the ability to specify the variables you expect to be initialized:<\/p>\n\n<pre><code>def guarantee_initialized_variables(session, list_of_variables = None):\n    if list_of_variables is None:\n        list_of_variables = tf.all_variables()\n    uninitialized_variables = list(tf.get_variable(name) for name in\n                                   session.run(tf.report_uninitialized_variables(list_of_variables)))\n    session.run(tf.initialize_variables(uninitialized_variables))\n    return unintialized_variables\n<\/code><\/pre>\n\n<p>This is still less ideal than actually knowing which variables are and are not initialized and taking care of that properly, but in the case of misdirection like the <code>optim<\/code> classes (see below) it may be hard to avoid.<\/p>\n\n<p>Also note, tf.initialize_variables cannot evaluate tf.report_uninitialized_variables, so both of them have to be run within the context of the session to work.<\/p>\n\n<hr>\n\n<p>There is an inelegant but concise way to do it. Before introducing your new variables run <code>temp = set(tf.all_variables())<\/code> and afterwards run <code>sess.run(tf.initialize_variables(set(tf.all_variables()) - temp))<\/code>. These together will only initialize any variables created after the temp value is assigned.<\/p>\n\n<p>I've been playing with transfer learning, so I wanted a quick way to do it too, but this is the best way I could find. Especially when using things like AdamOptimizer, which doesn't give you easy (or any, I'm not sure) access to the variables it uses. So the following actually shows up in my code. (I initialize the new layer's variables explicitly, and run it once to show the initial error before transfer learning. Just for a sanity check.)<\/p>\n\n<pre><code>temp = set(tf.all_variables())\ntrain_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)\n#I honestly don't know how else to initialize ADAM in TensorFlow.\nsess.run(tf.initialize_variables(set(tf.all_variables()) - temp))\n<\/code><\/pre>\n\n<p>And it solves all my problems.<\/p>\n\n<p><strong>EDIT:<\/strong> <a href=\"https:\/\/stackoverflow.com\/a\/37291254\/786020\">@Lifu_Huang's answer<\/a> states the proper way to fix my problem. Theoretically, you should use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/train.html#Optimizer.get_slot_names\" rel=\"noreferrer\">tf.train.Optimizer.get_slot_names<\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/train.html#Optimizer.get_slot\" rel=\"noreferrer\">tf.train.Optimizer.get_slot<\/a>:<\/p>\n\n<pre><code>optim = tf.train.AdadeltaOptimizer(1e-4)\nloss = cross_entropy(y,yhat)\ntrain_step = optim.minimize(loss)\nsess.run(tf.initialize_variables([optim.get_slot(loss, name)\n                                  for name in optim.get_slot_names()])\n<\/code><\/pre>\n\n<p>This however gives me <code>AttributeError: 'NoneType' object has no attribute 'initializer'<\/code>. I'll make edits when I figure out what I did wrong, so you don't make my mistakes.<\/p>\n"}]}
{"QuestionId":35166493,"AnswerCount":1,"Tags":"<python><windows><dll><32bit-64bit><theano>","CreationDate":"2016-02-03T00:38:05.057","AcceptedAnswerId":35166733.0,"OwnerUserId":667804.0,"Title":"Trouble installing Theano (64 bit DLL error)","Body":"<p>I'm working on a Windows 7 64 bit Laptop with an NVidia graphic card. I wanted to use Lasagne which uses Theano. I never would have thought this would be that complicated :\/<\/p>\n\n<p>After failing miserable on my own, I tried to follow this tutorial: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a> - it uses python2.7 but I wanted to use the newer 3.5 version (and I don't think that is the problem). I did everything and at the end I was able to call:<\/p>\n\n<pre><\/pre>\n\n<p>I think this looks promising, I even checked and the python.exe is a 64 bit version (I did the size() check from some other thread). So next I downloaded Theano and I was supposed to call \"python setup.py develop\" but it said \"unknown option develop\" so I used \"build\" and \"install\" instead. I then tried to run an example but when I type \"import theano\" I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>So this seems like a 64bit - 32bit DLL problem. Do you guys have any advice? Thank you very much!<\/p>\n","answers":[{"AnswerId":"35166733","CreationDate":"2016-02-03T01:05:32.197","ParentId":null,"OwnerUserId":"667804","Title":null,"Body":"<p>so yes kichik you were right. I somehow put the 35-32 folder in my path when installing python. Then I always assumed it was 64 bit because I used the WinPython 64 bit version. If you look at the output of \"where python\" you can see that both files were in my path. This caused the problem. Changing the path fixed my problem.<\/p>\n\n<p>Thank you very much!<\/p>\n"}]}
{"QuestionId":35169491,"AnswerCount":4,"Tags":"<deep-learning><keras><lstm>","CreationDate":"2016-02-03T05:54:02.873","AcceptedAnswerId":35180313.0,"OwnerUserId":3609599.0,"Title":"How to implement a deep bidirectional LSTM with Keras?","Body":"<p>I am trying to implement a LSTM based speech recognizer. So far I could set up bidirectional LSTM (i think it is working as a bidirectional LSTM) by following the example in Merge layer. Now I want to try it with another bidirectional LSTM layer, which make it a deep bidirectional LSTM. But I am unable to figure out how to connect the output of the previously merged two layers into a second set of LSTM layers. I don't know whether it is possible with Keras. Hope someone can help me with this.<\/p>\n\n<p>Code for my single layer bidirectional LSTM is as follows<\/p>\n\n<pre><\/pre>\n\n<p>Dimensions of my x and y values are as follows.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43795411","CreationDate":"2017-05-05T02:13:51.790","ParentId":null,"OwnerUserId":"7093436","Title":null,"Body":"<p>You can use <code>keras.layers.wrappers.Bidirectional<\/code>.\nOfficial manual can be referenced here, <a href=\"https:\/\/keras.io\/layers\/wrappers\/#bidirectional\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/wrappers\/#bidirectional<\/a><\/p>\n"},{"AnswerId":"35180313","CreationDate":"2016-02-03T14:54:49.827","ParentId":null,"OwnerUserId":"3609599","Title":null,"Body":"<p>Well, I got the answer for the issue posted on the Keras issues. Hope this would be useful to anyone who look for this kind of approach.\n<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1629\">How to implement deep bidirectional -LSTM<\/a><\/p>\n"},{"AnswerId":"55358921","CreationDate":"2019-03-26T13:55:54.087","ParentId":null,"OwnerUserId":"9276708","Title":null,"Body":"<pre><code>model.add(Bidirectional(LSTM(64)))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/ad578c4c19444af9d1f0e0d51a8283eb0db1a264\/examples\/imdb_bidirectional_lstm.py#L38\" rel=\"nofollow noreferrer\">Keras example<\/a><\/p>\n"},{"AnswerId":"53664646","CreationDate":"2018-12-07T06:53:16.147","ParentId":null,"OwnerUserId":"7056704","Title":null,"Body":"<p>Now designing BiLSTM is easier. A new class <code>Bidirectional<\/code> is added as per official doc here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Bidirectional\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Bidirectional<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/43035827\/whats-the-difference-between-a-bidirectional-lstm-and-an-lstm\/53664580#53664580\"><strong>For training result &amp; full code<\/strong>  <\/a><\/p>\n"}]}
{"QuestionId":35170791,"AnswerCount":3,"Tags":"<scikit-learn><tensorflow>","CreationDate":"2016-02-03T07:20:20.940","AcceptedAnswerId":35171132.0,"OwnerUserId":3695715.0,"Title":"sklearn.linear_model not found in TensorFlow Udacity course","Body":"<p>I'm following the instructions of the Deep Learning course by Google with TensorFlow. Unfortunately I'm stuck <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/1_notmnist.ipynb\" rel=\"nofollow noreferrer\" title=\"here\">with this workbook<\/a> right now.\nI work in the docker vm with all the assignment code loaded as described <a href=\"https:\/\/www.udacity.com\/course\/viewer#!\/c-ud730\/l-6452084188\/m-6560586345\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>When I do all the imports everything works except for following line:<\/p>\n\n<pre><\/pre>\n\n<p>it throws the following error:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/17348144\/3695715\">This SO answer<\/a> sounds promising, but I did not find the source directory of sklearn. <\/p>\n\n<p>Any help greatly aprreciated.<\/p>\n","answers":[{"AnswerId":"35171132","CreationDate":"2016-02-03T07:41:17.607","ParentId":null,"OwnerUserId":"5870826","Title":null,"Body":"<p>You can install and upgrade sklearn from the shell with pip. That may or may not be the problem - but at least you'll know its installed.<\/p>\n\n<pre><code>sudo pip install --upgrade scikit-learn\n<\/code><\/pre>\n"},{"AnswerId":"53603187","CreationDate":"2018-12-03T22:55:01.873","ParentId":null,"OwnerUserId":"10226602","Title":null,"Body":"<p>Please take note that when you are writing your code, you'll import the sklearn package using <code>import sklearn<\/code>, but when <strong>installing<\/strong> the package it with, say, conda, you should do the following: <\/p>\n\n<p><code>conda install scikit-learn<\/code><\/p>\n"},{"AnswerId":"36223958","CreationDate":"2016-03-25T16:30:51.930","ParentId":null,"OwnerUserId":"6114790","Title":null,"Body":"<p>In your Jupyter notebook:<\/p>\n\n<pre><code>import pip\n\npip.main(['install', 'sklearn'])\n<\/code><\/pre>\n"}]}
{"QuestionId":35171405,"AnswerCount":3,"Tags":"<machine-learning><tensorflow><gpu-programming><recurrent-neural-network>","CreationDate":"2016-02-03T07:57:57.170","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"How to determine maximum batch size for a seq2seq tensorflow RNN training model","Body":"<p>Currently, I am using the default 64 as the batch size for the seq2seq tensorflow model. What is the maximum batch size , layer size etc I can go with a single Titan X GPU with 12 GB RAM with Haswell-E xeon 128GB RAM. The input data is converted to embeddings. Following are some helpful parameters I am using , it seems the cell input size is 1024:<\/p>\n\n<pre><\/pre>\n\n<p>So based on my hardware what is the maximum batch size , layers, input size I can go? Currently the GPU shows that 99% memory is occupied.<\/p>\n","answers":[{"AnswerId":"42832364","CreationDate":"2017-03-16T11:08:16.093","ParentId":null,"OwnerUserId":"7499201","Title":null,"Body":"<p>By default, Tensorflow occupies all GPU memory available. However, there is a way to change this. In my model, I do this:<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\n<\/code><\/pre>\n\n<p>Then you can use this config when you start your session:<\/p>\n\n<pre><code>with tf.Session(config=config) as sess:\n<\/code><\/pre>\n\n<p>Now, the model will only use as much memory as it needs, and then you can try with different batch sizes and see when it runs out of memory.<\/p>\n"},{"AnswerId":"35205848","CreationDate":"2016-02-04T16:05:56.707","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>The memory usage when running a TensorFlow model depends on how many variables you have in your model, as well as the intermediate tensors that the TensorFlow run time uses to compute activations, gradients, etc. For instance, in your model, if the <code>input_size<\/code> is <code>1024<\/code>, the memory used for variables per layer would be <code>4MB + 4KB<\/code> (weights and biases). The  memory used for intermediate tensors would grow linearly with the batch size, but the exact amount is hard to estimate, as it depends on how the run time decides to schedule the operations. <code>12GB<\/code> should be able to fit quite a large model, though.<\/p>\n"},{"AnswerId":"35277541","CreationDate":"2016-02-08T19:18:11.003","ParentId":null,"OwnerUserId":"5875585","Title":null,"Body":"<p>Elaborating a bit on the prior answer, it is difficult to analytically forecast the exact max RAM consumption of a model because the TF runtime has some freedom to schedule independent operations simultaneously, and doing so can result in higher max RAM use than executing the same ops sequentially.  Op scheduling is dynamic, hence the maximum amount of RAM used in a training step can vary non-deterministically from step to step.  In practice, for non-trivial models it seems necessary to experiment to find the largest batch size that will consistently work.  <\/p>\n"}]}
{"QuestionId":35174423,"AnswerCount":2,"Tags":"<reshape><tensorflow>","CreationDate":"2016-02-03T10:27:03.513","AcceptedAnswerId":null,"OwnerUserId":85514.0,"Title":"Tensorflow transformations for frame sliding (numpy stride_tricks)","Body":"<p>My input in a tensorflow graph comes as a vector which contains multiple overlapping windows. How can I create this array using only tensorflow operations?<\/p>\n\n<pre><\/pre>\n\n<p>In numpy I would use stride_tricks, but something similar isn't available in tensorflow. How should I approach this?<\/p>\n","answers":[{"AnswerId":"54805683","CreationDate":"2019-02-21T11:10:35.970","ParentId":null,"OwnerUserId":"8342910","Title":null,"Body":"<p>You can use tf.map_fn() to accomplish this:<\/p>\n\n<pre><code>input = [1,2,3,4,5,6,7,8]\nshift = 2\nwindow_width = 4\nlimit = len(input) - window_width + 1\n\ninput_tensor = tf.placeholder(tf.int32, shape=(8,))\noutput_tensor = tf.map_fn(lambda i: input_tensor[i:i+window_width], elems=tf.range(start=0, limit=limit, delta=shift))\n\nwith tf.Session() as sess:\n    answer_test = sess.run(output_tensor, feed_dict = {input_tensor:input})\n    print(answer_test)\n\n:    [[1 2 3 4]\n     [3 4 5 6]\n     [5 6 7 8]]\n<\/code><\/pre>\n"},{"AnswerId":"35182849","CreationDate":"2016-02-03T16:45:32.893","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow doesn't have stride_tricks. The following would work for your particular use case.<\/p>\n\n<pre><code>&gt;&gt;&gt; b=tf.concat(0, [tf.reshape(input[i:i+4], [1, window_width]) for i in range(0, len(input) - window_width + 1, shift)])\n&gt;&gt;&gt; with tf.Session(\"\"): b.eval()\n... \narray([[1, 2, 3, 4],\n       [3, 4, 5, 6],\n       [5, 6, 7, 8]], dtype=int32)\n<\/code><\/pre>\n\n<p>If your input is large, you may also want to look at <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#slice_input_producer\" rel=\"nofollow\">slice_input_producer<\/a>.<\/p>\n"}]}
{"QuestionId":35175032,"AnswerCount":1,"Tags":"<python><reference><tensorflow><dereference>","CreationDate":"2016-02-03T10:52:13.673","AcceptedAnswerId":35181279.0,"OwnerUserId":258483.0,"Title":"How to dereference _ref tensor type in TensorFlow?","Body":"<p>How to convert reference tensor type to value tensor type?<\/p>\n\n<p>The only way I found is to add a zero to a tensor. Is there any convenient way?<\/p>\n\n<p>Below  is a tensor of reference type. How to get rid of ?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35181279","CreationDate":"2016-02-03T15:37:15.633","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In general, you should be able to use a <code>tf.foo_ref<\/code>-type tensor anywhere a <code>tf.foo<\/code>-type tensor is expected. TensorFlow ops will implicitly dereference their input arguments (unless a reference tensor is explicitly expected, e.g. in <code>tf.assign()<\/code>).<\/p>\n\n<p>The simplest way to dereference a tensor is to use <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/control_flow_ops.html#identity\" rel=\"nofollow\"><code>tf.identity()<\/code><\/a>, as follows:<\/p>\n\n<pre><code>counter = tf.Variable(0)\nassert counter.dtype == tf.int32_ref\n\ncounter_val = tf.identity(counter)\nassert counter_val.dtype == tf.int32\n<\/code><\/pre>\n\n<p>Note that this answers your question, but can have surprising semantics, because <code>tf.identity()<\/code> <strong>does not copy<\/strong> the underlying buffer. Therefore, <code>counter<\/code> and <code>counter_val<\/code> in the above example share the same buffer, and a modification to <code>counter<\/code> will be reflected in <code>counter_val<\/code>:<\/p>\n\n<pre><code>counter = tf.Variable(0)\ncounter_val = tf.identity(counter)  # Take alias before the `assign_add` happens.\ncounter_update = counter.assign_add(1)\n\nwith tf.control_dependencies([counter_update]):\n  # Force a copy after the `assign_add` happens.\n  result = counter_val + 0\n\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\n\nprint sess.run(result)  # ==&gt; 1  (result has effect of `assign_add`)\n<\/code><\/pre>\n"}]}
{"QuestionId":35175607,"AnswerCount":1,"Tags":"<machine-learning><torch><confusion-matrix><criterion>","CreationDate":"2016-02-03T11:18:56.983","AcceptedAnswerId":35234273.0,"OwnerUserId":4672217.0,"Title":"Torch Confusion matrix for binary criterion","Body":"<p>How do you use the optim packages' confusion matrix with the BCECriterion, which produces binary output? The last layer of my model is nn.logsigmoid. Thanks.<\/p>\n","answers":[{"AnswerId":"35234273","CreationDate":"2016-02-05T21:59:45.263","ParentId":null,"OwnerUserId":"4672217","Title":null,"Body":"<p>Confusion matrix cannot accept 0s, it wants the class indices of the target and prediction. <\/p>\n"}]}
{"QuestionId":35176812,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-03T12:15:27.870","AcceptedAnswerId":null,"OwnerUserId":3972931.0,"Title":"Cannot make array in using tensorflow","Body":"<p>In feeding batch_xs to x, I reshaped batch_xs, for BATCH_SIZE is 1.\nHere is my source.\nI'm not sure what is making the ValueError.<\/p>\n\n<pre><\/pre>\n\n<p>This gives me the Value error, which doesn't make sense to me.<\/p>\n\n<p>ValueError: Argument must be a dense tensor<\/p>\n","answers":[{"AnswerId":"35181828","CreationDate":"2016-02-03T16:02:22.517","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>From the documentation <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/client.html#Session\" rel=\"nofollow\">here<\/a> :<\/p>\n\n<blockquote>\n  <p>Each key in feed_dict can be one of the following types:<\/p>\n  \n  <ul>\n  <li><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><\/li>\n  <li><p>If the key is a SparseTensor, the value should be a SparseTensorValue.<\/p><\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>The types that you can use as the \"value\" for a key in <code>feed_dict<\/code> should be Python primitive types or numpy arrays. You are using the result of <code>tf.reshape<\/code>, which is a TensorFlow Tensor type. You can simply use <code>np.reshape<\/code> if you want to feed a reshaped array.<\/p>\n"}]}
{"QuestionId":35179452,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow>","CreationDate":"2016-02-03T14:15:05.523","AcceptedAnswerId":null,"OwnerUserId":258483.0,"Title":"How to pass parameters to the thread, operating on tensorflow graph?","Body":"<p>I would like to run a thread, which will populate a queue on the computation graph from normal python operations. The thread will use  object, some of tensors and some data. <\/p>\n\n<p>How to pass data to the thread?<\/p>\n\n<p>Here is seen some sample: <a href=\"https:\/\/stackoverflow.com\/a\/34596212\/258483\">https:\/\/stackoverflow.com\/a\/34596212\/258483<\/a><\/p>\n\n<p> function should be ran in a separate thread. It is started with <\/p>\n\n<pre><\/pre>\n\n<p>but how  and  reach inside of a thread? These declared as function parameters, so it is not a closure. So, is this author's error or Python allows this way of sending parameters?<\/p>\n","answers":[{"AnswerId":"35182605","CreationDate":"2016-02-03T16:34:34.563","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you have a function <code>load_and_enqueue(foo, bar, baz)<\/code> and you want to run it in a thread with particular values for the arguments, you can start it with those arguments as follows:<\/p>\n\n<pre><code>def load_and_enqueue(foo, bar, baz):\n  # ...\n\nfoo_arg = ...\nbar_arg = ...\nbaz_arg = ...\n\nt = threading.Thread(target=load_and_enqueue, args=(foo_arg, bar_arg, baz_arg))\nt.start()\n<\/code><\/pre>\n\n<p><strong>N.B.<\/strong> The <a href=\"https:\/\/stackoverflow.com\/a\/34596212\/3574081\">original SO answer<\/a> has been modified because it was intended to simply capture the variables from the enclosing scope.<\/p>\n"}]}
{"QuestionId":35180888,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-03T15:18:51.760","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"How to use pre-trained model as non trainable sub network in tensorflow?","Body":"<p>I'd like to train a network that contains a sub network that I need to stay fix during the training. The basic idea is to prepend and append some layers the the pre-trained network (inceptionV3) <\/p>\n\n<pre><\/pre>\n\n<p>and run the training process for the task I have without changing the pre-trained one. \nI also need to branch directly on some layer of the pre-trained network. For example, with the inceptionV3 I like to uses it from the conv 299x299 to the last pool layer or from the conv 79x79 to the last pool layer.<\/p>\n","answers":[{"AnswerId":"35182226","CreationDate":"2016-02-03T16:19:13.797","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>Whether or not a \"layer\" is trained is determined by whether the variables used in that layer get updated with gradients. If you are using the <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/train.html#optimizers\">Optimizer<\/a> interface to optimize your network, then you can simply not pass the variables used in the layers that you want to keep fixed to the <code>minimize<\/code> function, i.e.,<\/p>\n\n<pre><code>opt.minimize(loss, &lt;subset of variables you want to train&gt;)\n<\/code><\/pre>\n\n<p>If you are using <code>tf.gradients<\/code> function directly, then remove the variables that you want to keep fixed from the second argument to <code>tf.gradients<\/code>.<\/p>\n\n<p>Now, how you \"branch directly\" to a layer of a pre-trained network depends on how that network is implemented. I would simply locate the <code>tf.Conv2D<\/code> call to the 299x299 layer you are talking about, and pass as its input, the output of your new layer, and on the output side, locate the 79x79 layer, use its output as the input to your new layer.<\/p>\n"}]}
{"QuestionId":35184999,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><word2vec>","CreationDate":"2016-02-03T18:36:19.277","AcceptedAnswerId":35186508.0,"OwnerUserId":4190832.0,"Title":"Word2vec + Tensorflow and the shape of everything","Body":"<p>i'm looking for a solution to a simple Text classification issue with tensorflow. I made a model with the IMDB dataset to know wether a comment is positive or negative. The data was processed through word2vec so now i have a bunch of vector to classify. I think my problem here are due to the bad shape of the y_labels since they are one dimensionnal and i want to classify them through tensorflow with a two classes output, or maybe i am wrong. Final info, the model is working well, with an accuracy of 1.0, maybe too well! Thanks for the help !<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35186508","CreationDate":"2016-02-03T19:59:18.523","ParentId":null,"OwnerUserId":"5530615","Title":null,"Body":"<p>You are using softmax in your example.  Softmax assigns a probability to N different classes, where the probability adds up to 1. Basically, the model is choosing exactly one of the N choices. For this to make sense you need N to be at least 2.  With N == 1, the probability of that class will always be 1.  You have two possible fixes:<\/p>\n\n<ol>\n<li>Create two classes, one for \"positive sentiment\" and one for \"negative sentiment\", setting N to 2.<\/li>\n<li>Use logistic regression instead of Softmax. In logistic regression each class is independent. That means you have N questions, each of which gets its own \"yes\" or \"no\" answer, which makes sense with N == 1.<\/li>\n<\/ol>\n"}]}
{"QuestionId":35189999,"AnswerCount":1,"Tags":"<lua><deep-learning><torch><mnist>","CreationDate":"2016-02-03T23:47:47.893","AcceptedAnswerId":35266837.0,"OwnerUserId":4572570.0,"Title":"Lua\/Torch: Split MNIST dataset into training and test with equal number of labels in each set","Body":"<p>I am trying to split the data into training(80%) and test(20%) set, but I need to shuffle the data first and then assign equal number of samples for each label(y, 10 classes) in each dataset.<\/p>\n\n<p>How can I do this in lua\/torch? Thanks!<\/p>\n\n<p>This is my code so far...<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35266837","CreationDate":"2016-02-08T10:04:16.407","ParentId":null,"OwnerUserId":"5407700","Title":null,"Body":"<p>You could do something like below to shuffle;<\/p>\n\n<pre><code>    dataSize = Data.data:size()[1]\n    shuffleIdx = torch.randperm(dataSize)\n    Data.data = Data.data:index(1,shuffleIdx:long())\n    Data.labels = Data.labels:index(1,shuffleIdx:long())\n<\/code><\/pre>\n\n<p>but I am not sure about the second part of your question.<\/p>\n"}]}
{"QuestionId":35190898,"AnswerCount":1,"Tags":"<python><theano><gradient-descent>","CreationDate":"2016-02-04T01:25:05.097","AcceptedAnswerId":35433335.0,"OwnerUserId":2255757.0,"Title":"theano hard_sigmoid() breaks gradient descent","Body":"<p>for intents of highlighting the issue lets follow this <a href=\"http:\/\/outlace.com\/Beginner-Tutorial-Theano\/\" rel=\"nofollow\">tutorial<\/a>.<\/p>\n\n<hr>\n\n<p>theano has 3 ways to compute the sigmoid of a tensor, namely <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#tensor.nnet.sigmoid\" rel=\"nofollow\">sigmoid<\/a>, <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#tensor.nnet.ultra_fast_sigmoid\" rel=\"nofollow\">ultra_fast_sigmoid<\/a> and <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#tensor.nnet.hard_sigmoid\" rel=\"nofollow\">hard_sidmoid<\/a>. It seems using the latter two breaks the gradient descent algorithm.<\/p>\n\n<p>The conventional sigmoid converges as it should, but the the others have strange inconsistent behaviours. ultra_fast_sigmoid, just throws a straight error when trying to compute the gradient 'Method not defined ('grad', ultra_fast_sigmoid)', whilst hard_sigmoid compiles fine, but fails to converge on the solution. <\/p>\n\n<hr>\n\n<p>Does anyone know the source of this behaviour? It s not highlighted in the documentation that this should happen and it seems counter intuitive.<\/p>\n\n<hr>\n\n<p>code: <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>i changed the following lines from the code to make the output shorter for this post (they differ from the tutorial, but are already contained in the code above):<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>sigmoid<\/p>\n\n<pre><\/pre>\n\n<p>ultra_fast_sigmoid<\/p>\n\n<pre><\/pre>\n\n<p>hard_sigmoid<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35433335","CreationDate":"2016-02-16T13:00:56.580","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Here's the source code of <code>hard_sigmoid<\/code>:<\/p>\n\n<pre><code>def hard_sigmoid(x):\n    \"\"\"An approximation of sigmoid.\n    More approximate and faster than ultra_fast_sigmoid.\n    Approx in 3 parts: 0, scaled linear, 1\n    Removing the slope and shift does not make it faster.\n    \"\"\"\n    # Use the same dtype as determined by \"upgrade_to_float\",\n    # and perform computation in that dtype.\n    out_dtype = scalar.upgrade_to_float(scalar.Scalar(dtype=x.dtype))[0].dtype\n    slope = tensor.constant(0.2, dtype=out_dtype)\n    shift = tensor.constant(0.5, dtype=out_dtype)\n    x = (x * slope) + shift\n    x = tensor.clip(x, 0, 1)\n    return x\n<\/code><\/pre>\n\n<p>So it is just implemented as a piecewise linear function, whose gradient is 0.2 within the range of (-2.5, 2.5) and 0 elsewhere. Which means if the input falls outside the region (-2.5, 2.5), its gradient will be zero, and no learning will happen.  <\/p>\n\n<p>So it might not be appropriate for training, but can be use for approximating the prediction result.<\/p>\n\n<hr>\n\n<p>Edit:<br>\nTo evaluate the gradient of the network parameters, normally we use <a href=\"https:\/\/en.wikipedia.org\/wiki\/Backpropagation\" rel=\"nofollow\">backpropagation<\/a>.<br>\nHere's a very simple example.<\/p>\n\n<pre><code>x = theano.tensor.scalar()\nw = theano.shared(numpy.float32(1))\ny = theano.tensor.nnet.hard_sigmoid(w*x)  # y=w*x, w is initialized to 1.\n\ndw = theano.grad(y, w)  # gradient wrt w, which is equal to slope*x in this case\nnet = theano.function([x], [y, dw])\n\nprint net(-3)\nprint net(-1)\nprint net(0)\nprint net(1)\nprint net(3)\n\nOutput:\n[array(0.0), array(-0.0)]  # zero gradient because the slope is zero\n[array(0.3), array(-0.2)]\n[array(0.5), array(0.0)]  # zero gradient because x is zero\n[array(0.7), array(0.2)]\n[array(1.0), array(0.0)]  # zero gradient because the slope is zero\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT BY OP:<br>\n<code>ultra_hard_sigmoid<\/code> fails, if you look at the source code implementation, because it is hard-coded in python and not handled by tensor expressions.<\/p>\n"}]}
{"QuestionId":35192022,"AnswerCount":1,"Tags":"<image><image-processing><machine-learning><neural-network><torch>","CreationDate":"2016-02-04T03:34:22.047","AcceptedAnswerId":35211881.0,"OwnerUserId":4013781.0,"Title":"Visualizing CNN weights from Torch Tensor","Body":"<p>I have a Torch Cuda Tensor of size 64x64x3x3 and I want to visualise its weights for a given layer as follows: <\/p>\n\n<pre><\/pre>\n\n<p>This is producing an error: <\/p>\n\n<pre><\/pre>\n\n<p>How should I resize the tensor in order to use toDisplayTensor? Thanks.<\/p>\n","answers":[{"AnswerId":"35211881","CreationDate":"2016-02-04T21:22:56.813","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Error seems to be very straight forward - you should provide data in the form of either:<\/p>\n\n<ul>\n<li>HxW - height x weight, 2-dim tensor, one per kernel\/channel pair<\/li>\n<li>KxHxW - kernels x height x weight, 3-dim tensor, one per channel (for example if data is grayscale)<\/li>\n<li>Kx3xHxW - kernels x channels x height x weight, 4-dim tensor, one per network layer (for 3-channel colour data, as the second dim will be mapped to RGB scale)<\/li>\n<\/ul>\n\n<p>In your case I assume that your image is 64x64 (or after current convolution is)? Then you need 3x3x64x64 Tensor, where first dimension is the one iterating over kernels, second - over colors, third and forth over  the image itself.<\/p>\n"}]}
{"QuestionId":35196307,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><caffe>","CreationDate":"2016-02-04T08:47:12.140","AcceptedAnswerId":35204696.0,"OwnerUserId":708366.0,"Title":"Caffe fine-tuning vs. starting from scratch","Body":"<p><strong>Context<\/strong>: let's say I have trained a CNN on  and I've obtained .<\/p>\n\n<p><strong>Current situation<\/strong>: new pictures arrive so I can build a new dataset,   <\/p>\n\n<p><strong>Question<\/strong>: would these two situations lead to same caffemodel?<\/p>\n\n<ol>\n<li>merge  and  and train the net from scratch.<\/li>\n<li>perform some <em>fine-tuning<\/em> on existing  by training it only on  (as explained here: <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html<\/a>) <\/li>\n<\/ol>\n\n<p>It might seem a dumb question, but I'm not really sure about its answer. And it's really important because if the two approximations lead to same result I can save time by performing number 2.<\/p>\n\n<p><em>Note: bear in mind that it's the same problem, so no need to change architecture here, I just plan to add new images to the training.<\/em><\/p>\n","answers":[{"AnswerId":"35204696","CreationDate":"2016-02-04T15:14:29.817","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<p>In the Flicker-style example the situation is a bit more generic. They use the weights of first layers from a model trained for a different classification task and employ it for a new task, training only a new last layer and fine-tuning the first layers a bit (by setting a low learning rate for those pretrained layers). Your case is similar but more specific, you want to use the pretrained model to train the exact architecture for the exact same task but with an extension of your data.<\/p>\n\n<p>If your question if whether Option 1. will produce <strong>exactly<\/strong> the same model (all resulting weights are equal) as Option 2. Then no, most probably not.<\/p>\n\n<p>In Option 2. the network is trained for iterations of <code>dataset A<\/code> then for <code>dataset B<\/code> then <code>dataset A<\/code> again..and so on (assuming both were just concatenated together).\nWhile in Option 1. will have the network trained for some iterations\/epochs on <code>dataset A<\/code>, then later continue learning for iterations\/epochs on only <code>dataset B<\/code> and that's it. So the solver will see a different sequence of gradients in both options resulting in two different models. That's from a strict theoretical perspective.<\/p>\n\n<p>If you ask from a practical perspective, the two options will probably end up with very similar models. How many <strong>epochs<\/strong> (not iterations) did you train on <code>dataset A<\/code> ? say N <strong>epochs<\/strong>, then you can safely go with Option 2. and train your existing model further on <code>dataset B<\/code> for the same number of <strong>epochs<\/strong> and same learning rate and batch size.<\/p>\n"}]}
{"QuestionId":35198528,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-02-04T10:28:33.283","AcceptedAnswerId":35205193.0,"OwnerUserId":4190832.0,"Title":"Reshape y_train for binary text classification in Tensorflow","Body":"<p>i have a classical y_train which is composed of 0 (negative) and 1 (positive) in a one dimensionnal shape. I wanted to train a tensorflow model but i have to initialize the y placeholder with the number of classes i want. So in this text classification case, i want the model to check negative or positive value so 2 classes ? But how did i convert my y_train to fit the output that i'm looking for. Thanks for your time!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35205193","CreationDate":"2016-02-04T15:38:13.977","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>It appears your <code>y_train<\/code> contains the label values themselves, whereas the <code>y_train<\/code> needed by the model requires label probabilities. In your case, since there are only two labels, you can convert that to label probabilities as follows :<\/p>\n\n<pre><code>y_train = tf.concat(1, [1 - y_train, y_train])\n<\/code><\/pre>\n\n<p>If you have more labels, have a look at <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/sparse_ops.html#sparse_to_dense\" rel=\"nofollow\"><code>sparse_to_dense<\/code><\/a> to convert them to probabilities.<\/p>\n"}]}
{"QuestionId":35201889,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2016-02-04T13:10:26.340","AcceptedAnswerId":35202403.0,"OwnerUserId":957997.0,"Title":"How can I retrieve the output from both fc and softmax layers from Inception-v3 in a single run?","Body":"<p>I would like to extract the output of both  and  layers. I could run the model twice and, for each run, extract the output of a single layer, but it's a bit wasteful. Is it possible to do it running the model only once?<\/p>\n\n<p>I'm using the example provided by <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py\" rel=\"nofollow\"><\/a>. Here is the relevant snippet:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"35202403","CreationDate":"2016-02-04T13:32:53.703","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can pass a list of tensors to <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/client.html#Session.run\"><code>Session.run()<\/code><\/a> and TensorFlow will share the work done to compute them:<\/p>\n\n<pre><code>softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')\npool_3 = sess.graph.get_tensor_by_name('pool_3:0')\npredictions, pool3_val = sess.run([softmax_tensor, pool_3],\n                                  {'DecodeJpeg:0': image_data})\n<\/code><\/pre>\n"}]}
{"QuestionId":35206445,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-04T16:32:11.903","AcceptedAnswerId":35208939.0,"OwnerUserId":2615676.0,"Title":"Add control dependency after operations are created?","Body":"<p>Is it possible to create a control dependency between two ops after they've both been created? I realize that with  it's possible to make one op wait on another before executing, but the dependent op has to be created within the  context. I'd like to construct the two ops independently first and then add the dependency.<\/p>\n","answers":[{"AnswerId":"44394908","CreationDate":"2017-06-06T16:06:26.690","ParentId":null,"OwnerUserId":"2546106","Title":null,"Body":"<p>While not possible in general, you can slightly modify an existing <code>op<\/code> to enforce dependencies. For example, you can add zero to a tensor:\n<code>\ndef first_before_second_num(op_first, op_second):\n   with tf.control_dependencies([op_first]):\n       op_second += 0\n   return op_second\n<\/code><\/p>\n\n<p>If you have a <code>NoOp<\/code> like a <code>train_step<\/code>, then you might want to group it with another <code>NoOp<\/code>:\n<code>\ndef first_before_second_noop(op_first, op_second):\n   with tf.control_dependencies([op_first]):\n       op_second tf.group(op_second, tf.no_op())\n   return op_second\n<\/code><\/p>\n"},{"AnswerId":"35208939","CreationDate":"2016-02-04T18:36:08.943","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This might be a disappointing answer, but it is not possible to add a control dependency (or any other input) to a TensorFlow <code>Operation<\/code> after it has been created. Tensors and operations are <em>immutable<\/em> once they have been created.<\/p>\n\n<p>One possibility would be to <strong>clone<\/strong> the op that should run second, in the appropriate control dependencies context. Let's say that you have two <code>Operation<\/code> objects, <code>op_first<\/code> and <code>op_second<\/code>, and you want <code>op_first<\/code> to run before <code>op_second<\/code>:<\/p>\n\n<pre><code>def first_before_second(op_first, op_second):\n    \"\"\"Sequence `op_first` before `op_second`.\n\n    Given two operations, returns a pair of operations with the same behavior\n    except that the first returned operation will execute before the second\n    returned operation.\n    \"\"\"\n    with tf.control_dependencies([op_first]):\n        g = tf.get_default_graph()\n        op_second_clone = g.create_op(op_second.type,\n                                      [in_t for in_t in op_second.inputs],\n                                      [out_t.dtype for out_t in op_second.outputs],\n                                      attrs=op_second.node_def.attr,\n                                      op_def=op_second.op_def)\n\n    return op_first, op_second_clone\n<\/code><\/pre>\n\n<p>A similar modification could make this deal with <code>Tensor<\/code> objects as well, but that's left as an exercise for the reader.<\/p>\n"}]}
{"QuestionId":35208832,"AnswerCount":1,"Tags":"<python><image><shape><tensorflow>","CreationDate":"2016-02-04T18:29:47.673","AcceptedAnswerId":35209389.0,"OwnerUserId":2206976.0,"Title":"TensorFlow Resize image tensor to dynamic shape","Body":"<p>I am trying to read some image input for an image classification problem with TensorFlow.<\/p>\n\n<p>Of course, I am doing this with . My images have variable size and hence I am not able to specify a fixed shape for the image tensor.<\/p>\n\n<p>But I need to scale the images depending on their actual size. Specifically, I want to scale the shorter side to a fixed value and the longer side in a way that the aspect ratio is preserved.<\/p>\n\n<p>I can get the actual shape of a certain image by . I am also able to do the computation for the new longer edge like <\/p>\n\n<pre><\/pre>\n\n<p>My problem now is that I cannot pass  and  to  because one of them is a tensor and  expects integers as height and width inputs.<\/p>\n\n<p>Is there a way to \"pull out\" the integer of the tensor or is there any other way to do my task with TensorFlow?<\/p>\n\n<p>Thanks in advance.<\/p>\n\n<hr>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>Since I also had <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1029\" rel=\"noreferrer\">some other issues<\/a> with , here's the code that worked for me:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35209389","CreationDate":"2016-02-04T19:00:45.447","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The way to do this is to use the (currently experimental, but available in the next release) <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#cond\" rel=\"noreferrer\"><code>tf.cond()<\/code><\/a>* operator. This operator is able to test a value computed at runtime, and execute one of two branches based on that value.<\/p>\n\n<pre><code>shape = tf.shape(image)\nheight = shape[0]\nwidth = shape[1]\nnew_shorter_edge = 400\nheight_smaller_than_width = tf.less_equal(height, width)\n\nnew_shorter_edge = tf.constant(400)\nnew_height, new_width = tf.cond(\n    height_smaller_than_width,\n    lambda: new_shorter_edge, (width \/ height) * new_shorter_edge,\n    lambda: new_shorter_edge, (height \/ width) * new_shorter_edge)\n<\/code><\/pre>\n\n<p>Now you have <code>Tensor<\/code> values for <code>new_height<\/code> and <code>new_width<\/code> that will take the appropriate value at runtime.<\/p>\n\n<hr>\n\n<p>*&nbsp; To access the operator in the current released version, you'll need to import the following:<\/p>\n\n<pre><code>from tensorflow.python.ops import control_flow_ops\n<\/code><\/pre>\n\n<p>...and then use <code>control_flow_ops.cond()<\/code> instead of <code>tf.cond()<\/code>.<\/p>\n"}]}
{"QuestionId":35211377,"AnswerCount":1,"Tags":"<python><theano><gradient-descent><lstm><recurrent-neural-network>","CreationDate":"2016-02-04T20:51:21.377","AcceptedAnswerId":null,"OwnerUserId":3222797.0,"Title":"Compute updates in Theano after N number of loss calculations","Body":"<p>I've constructed a LSTM recurrent NNet using <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\" rel=\"nofollow\">lasagne<\/a> that is loosely based on the architecture in this <a href=\"http:\/\/colinraffel.com\/talks\/hammer2015recurrent.pdf\" rel=\"nofollow\">blog post<\/a>.  My input is a text file that has around 1,000,000 sentences and a vocabulary of 2,000 word tokens.  Normally, when I construct networks for image recognition my input layer will look something like the following:<\/p>\n\n<pre><\/pre>\n\n<p>(where the dimensions are batch size, channel, height and width) which is convenient because all the images are the same size so I can process them in batches.  Since each instance in my LSTM network has a varying sentence length, I have an input layer that looks like the following:<\/p>\n\n<pre><\/pre>\n\n<p>As described in above referenced blog post, <\/p>\n\n<blockquote>\n  <p><strong>Masks<\/strong>:<br>\n  Because not all sequences in each minibatch will always have the same length, all recurrent layers in \n  lasagne\n   accept a separate mask input which has shape \n  (batch_size, n_time_steps)\n  , which is populated such that \n  mask[i, j] = 1\n   when \n  j &lt;= (length of sequence i)\n   and \n  mask[i, j] = 0\n   when \n  j > (length\n  of sequence i)\n  . \n  When no mask is provided, it is assumed that all sequences in the minibatch are of length \n  n_time_steps.<\/p>\n<\/blockquote>\n\n<p>My question is:  Is there a way to process this type of network in mini-batches <strong><em>without<\/em><\/strong> using a mask?<\/p>\n\n<hr>\n\n<p>Here is a simplified version if my network.<\/p>\n\n<pre><\/pre>\n\n<p>From there I have generator that spits out  pairs and I am computing  and updating the gradient with each iteration.  What I want to do is do an N number of training steps and <strong><em>then<\/em><\/strong> update the parameters with the average gradient.<\/p>\n\n<p>To do this, I tried creating a  function:<\/p>\n\n<pre><\/pre>\n\n<p>and then looping over several training instances to create a \"batch\" and collect the gradient calculations to a list:<\/p>\n\n<pre><\/pre>\n\n<p>this produces a list of lists<\/p>\n\n<pre><\/pre>\n\n<p>From here I would need to take the mean of the gradient at each layer, and then update the model parameters.  This is possible to do in pieces like this does does the gradient calc\/parameter update need to happen all in one theano function?<\/p>\n\n<p>Thanks.  <\/p>\n","answers":[{"AnswerId":"35229771","CreationDate":"2016-02-05T17:11:29.170","ParentId":null,"OwnerUserId":"2255757","Title":null,"Body":"<p><em>NOTE: this is a solution, but by no means do i have enough experience to verify its best and the code is just a sloppy example<\/em><\/p>\n\n<p>You need 2 theano functions. The first being the grad one you seem to have already judging from the information provided in your question.<\/p>\n\n<p>So after computing the batched gradients you want to immediately feed them as an input argument back into another theano function dedicated to updating the shared variables. For this you need to specify the expected batch size at the compile time of your neural network. so you could do something like this: (for simplicity i will assume you have a global list variable where all your params are stored)<\/p>\n\n<pre><code>params #list of params you wish to update\nBATCH_SIZE = 1024 #size of the expected training batch\nG = [T.matrix() for i in range(BATCH_SIZE) for param in params] #placeholder for grads result flattened so they can be fed into a theano function\n\nupdates = [G[i] for i in range(len(params))] #starting with list of  param updates from first batch\n\nfor i in range(len(params)): #summing the gradients for each individual param\n     for j in range(1, len(G)\/len(params)):\n         updates[i] += G[i*BATCH_SIZE + j]\n\nfor i in range(len(params)): #making a list of tuples for theano.function updates argument\n     updates[i] = (params[i], updates[i]\/BATCH_SIZE) \nupdate = theano.function([G], 0, updates=updates)\n<\/code><\/pre>\n\n<p>Like this theano will be taking the mean of the gradients and updating the params as usual<\/p>\n\n<p><em>dont know if you need to flatten the inputs as I did, but probably<\/em><\/p>\n\n<p>EDIT: gathering from how you edited your question it seems important that the batch size can vary in that case you could add 2 theano functions to your existing one:<\/p>\n\n<ol>\n<li>the first theano function takes a batch of size 2 of your params and returns the sum. you could apply this theano function using python's reduce() and get the sum of the over the whole batch of gradients<\/li>\n<li>the second theano function takes those summed param gradients and a scaler (the batch size) as input and hence is able to update the NN params over the mean of the summed gradients.<\/li>\n<\/ol>\n"}]}
{"QuestionId":35213172,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-04T22:47:16.923","AcceptedAnswerId":35213412.0,"OwnerUserId":2628128.0,"Title":"Add a new device in TensorFlow","Body":"<p>In TensorFlow, a <em>Session<\/em> places the graph ops onto <em>Devices<\/em>, such as <em>CPU<\/em>s or <em>GPU<\/em>s, and provides methods to execute them. Is it possible to add an additional class of devices to TensorFlow? Let's call it XPU in addition to CPU and GPU?<\/p>\n","answers":[{"AnswerId":"35213412","CreationDate":"2016-02-04T23:06:30.083","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It is possible to add new device types to TensorFlow, although we don't have great documentation. Approximately, it would require adding a new implementation of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/device.h\" rel=\"noreferrer\"><code>tensorflow::Device<\/code><\/a>, adding a corresponding implementation of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/device_factory.h\" rel=\"noreferrer\"><code>tensorflow::DeviceFactory<\/code><\/a>, and invoking the following registration macro:<\/p>\n\n<pre><code>REGISTER_LOCAL_DEVICE_FACTORY(\"XPU\", MyXPUDeviceFactory);\n<\/code><\/pre>\n"}]}
{"QuestionId":35213787,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-04T23:37:59.570","AcceptedAnswerId":35213952.0,"OwnerUserId":1938262.0,"Title":"TensorFlow Batch Outer Product","Body":"<p>I have the following two tensors:<\/p>\n\n<pre><\/pre>\n\n<p>I want to perform a batch outer product: I'd like to multiply each element in the first column of  by each element in the first row of  to get a tensor of shape , then the second column of  by the second row of , and so on. The shape of the final tensor should be , where  is the batch size.<\/p>\n\n<p>Is there any simple way to achieve this in TensorFlow? I've tried to use  without success. <\/p>\n","answers":[{"AnswerId":"54273648","CreationDate":"2019-01-20T04:44:32.923","ParentId":null,"OwnerUserId":"4007718","Title":null,"Body":"<p>Perhaps, there is a more elegant solution using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/einsum\" rel=\"nofollow noreferrer\"><code>tf.einsum<\/code><\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>result = tf.einsum(\"un,nv-&gt;nuv\", x, y)\n<\/code><\/pre>\n"},{"AnswerId":"35213952","CreationDate":"2016-02-04T23:54:23.287","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Would the following work, using <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/math_ops.html#batch_matmul\" rel=\"noreferrer\"><code>tf.batch_matmul()<\/code><\/a>?<\/p>\n\n<pre><code>print x.get_shape()  # ==&gt; [U, N]\nprint y.get_shape()  # ==&gt; [N, V]\n\nx_transposed = tf.transpose(x)\nprint x_transposed.get_shape()  # ==&gt; [N, U]\n\nx_transposed_as_matrix_batch = tf.expand_dims(x_transposed, 2)\nprint x_transposed_as_matrix_batch.get_shape()  # ==&gt; [N, U, 1]\n\ny_as_matrix_batch = tf.expand_dims(y, 1)\nprint y_as_matrix_batch.get_shape()  # ==&gt; [N, 1, V]\n\nresult = tf.batch_matmul(x_transposed_as_matrix_batch, y_as_matrix_batch)\nprint result.get_shape()  # ==&gt; [N, U, V]\n<\/code><\/pre>\n"}]}
{"QuestionId":35214115,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-02-05T00:10:04.433","AcceptedAnswerId":35214135.0,"OwnerUserId":551538.0,"Title":"Column sum in torch","Body":"<p>How do i sum along a column in torch? I have a 128*1024 tensor, and I want to get a 1*1024 tensor by summing all the rows.<\/p>\n\n<p>For example:\na:<\/p>\n\n<p><\/p>\n\n<p>I want b<\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"35214135","CreationDate":"2016-02-05T00:12:35.023","ParentId":null,"OwnerUserId":"2857133","Title":null,"Body":"<p>To do this, you can use the sum method. <\/p>\n\n<p><code>torch.sum(a,1)<\/code><\/p>\n\n<p>In general, you can specify any axis you want to sum over.<\/p>\n\n<p><code>torch.sum(a,axis)<\/code><\/p>\n\n<p>(To sum over rows, you can use axis=2)<\/p>\n"}]}
{"QuestionId":35214262,"AnswerCount":1,"Tags":"<deep-learning><conv-neural-network><keras>","CreationDate":"2016-02-05T00:26:05.120","AcceptedAnswerId":35229643.0,"OwnerUserId":233964.0,"Title":"Convolutional2D Siamese Network in Keras","Body":"<p>I'm trying to use Keras's <a href=\"http:\/\/keras.io\/layers\/core\/#siamese\" rel=\"noreferrer\">Siamese<\/a> layer in conjunction with a shared  layer.\nI don't need the input to pass through any other layers before the  layer but the  layer requires that input layers be specified. I can't figure out how to create the input layers to match the input of the conv layer. The only concrete example of the  layer being used I could find is in the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/tests\/keras\/test_models.py#L445\" rel=\"noreferrer\">tests<\/a> where  layers (with vector inputs) are used as input. Basically, I want an input layer that allows me to specify the image dimensions as input so they can be passed on to the shared conv layer.<\/p>\n\n<p>In code I have something like the following:<\/p>\n\n<pre><\/pre>\n\n<p>What should  be? Or is my appraoch in general incorrect?<\/p>\n","answers":[{"AnswerId":"35229643","CreationDate":"2016-02-05T17:04:16.170","ParentId":null,"OwnerUserId":"233964","Title":null,"Body":"<p>Okay, I figured it out. The \"abstract\" <code>Layer<\/code> class is basically a pass through layer which is just what I need. I was also making a mistake where I thought <code>Siamese<\/code> could take an entire model (i.e. multiple layers) but it in fact only takes a single layer. To make the creation of these Siamese layers less painful there is a <code>add_shared_layer<\/code> helper function.<\/p>\n\n<p>I should also point out this <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/1421\" rel=\"nofollow\">pull request<\/a> that would allow a shared layer to the first layer in a model, exactly what I am trying to do.<\/p>\n"}]}
{"QuestionId":35221766,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><convolution><softmax>","CreationDate":"2016-02-05T10:28:07.580","AcceptedAnswerId":35229610.0,"OwnerUserId":4844184.0,"Title":"Getting rid of softmax saturation in DeepMNIST-like net for colour-images classification in TensorFlow","Body":"<p>I have a dataset for classification which is composed of a training of size 8000x(32x32x3 images) and of a test of size 2000x(same size images).<br>\nI am doing a very simple task of distinguishing vehicules and background. I am using the cross_entropy as a cost function.   <\/p>\n\n<p>The net I am using is almost the same as the one used in <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/pros\/index.html#second-convolutional-layer\" rel=\"nofollow\">DeepMNIST<\/a> except the first filter has size 3x... instead of 1x... because it is a colour  image and the output has size two because there are only two classes : vehicules or not vehicules. \nSeeing the results of this relatively straight forward task has led me to ask myself several interrogations :  <\/p>\n\n<p>-First if I do not use a large enough batch size (>200) I get stuck almost every time at accuracy 62% (in a local optima) over the two sets which is not sufficient for my need  <\/p>\n\n<p>-Secondly whenever I use the right optimizer Adam with the right batch size and learning rate I go up to 92% however the outputs are always very disturbingly good like [0.999999999 0.000000000001].    <\/p>\n\n<p>This should not happen as the task is difficult.<br>\nTherefore when I go fully convolutional to create a heatmap I got 1.000001 almost everywhere due to the saturation.<br>\nWhat am I doing wrong ? Do you think whitening would solve the problem ? Batch normalization ? Something else ? What am I facing ?<\/p>\n","answers":[{"AnswerId":"35229610","CreationDate":"2016-02-05T17:02:28.117","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>That's a sign of overfitting. If you train on small dataset long enough with a large enough model, eventually your confidences get saturated to 0's and 1's. Hence, same techniques that prevent overfitting (regularization penalties, dropout, early stopping, data augmentation) will help there.<\/p>\n\n<p>My first step for a tiny dataset like this, would be to augment the dataset with noise corrupted examples. IE, for your example I would add 800k noise corrupted examples with original labels, and train on those. <\/p>\n"}]}
{"QuestionId":35226198,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-05T14:08:45.253","AcceptedAnswerId":35227732.0,"OwnerUserId":1441121.0,"Title":"Is this one-hot encoding in TensorFlow fast? Or flawed for any reason?","Body":"<p>There are a few stack overflow questions about computing one-hot embeddings with TensorFlow, and here is the accepted solution:<\/p>\n\n<pre><\/pre>\n\n<p>This is almost identical to the code in an official tutorial: <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/tf\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/tf\/index.html<\/a><\/p>\n\n<p>To me it seems that since  exists, it's probably more efficient. Here's a version that uses this, and it supports arbitrarily-shaped inputs:<\/p>\n\n<pre><\/pre>\n\n<p>Do you expect this implementation to be faster? And is it flawed for any other reason?<\/p>\n","answers":[{"AnswerId":"35227732","CreationDate":"2016-02-05T15:28:09.023","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>one_hot()<\/code> function in your question looks correct. However, the reason that we do not recommend writing code this way is that it is <strong>very memory inefficient<\/strong>. To understand why, let's say you have a batch size of 32, and 1,000,000 classes.<\/p>\n\n<ul>\n<li><p>In the version suggested in the tutorial, the largest tensor will be the result of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_to_dense\" rel=\"noreferrer\"><code>tf.sparse_to_dense()<\/code><\/a>, which will be <code>32 x 1000000<\/code>.<\/p><\/li>\n<li><p>In the <code>one_hot()<\/code> function in the question, the largest tensor will be the result of <code>np.identity(1000000)<\/code>, which is 4 terabytes. Of course, allocating this tensor probably won't succeed. Even if the number of classes were much smaller, it would still waste memory to store all of those zeroes explicitly&mdash;TensorFlow does not automatically convert your data to a sparse representation even though it might be profitable to do so.<\/p><\/li>\n<\/ul>\n\n<p>Finally, I want to offer a plug for a new function that was recently added to the open-source repository, and will be available in the next release. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a> allows you to specify a vector of integers as the labels, and saves you from having to build the dense one-hot representation. It should be much more efficient that <em>either<\/em> solution for large numbers of classes.<\/p>\n"}]}
{"QuestionId":35226428,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-05T14:21:24.750","AcceptedAnswerId":35227842.0,"OwnerUserId":2245776.0,"Title":"How do I get the gradient of the loss at a TensorFlow variable?","Body":"<p>The feature I'm after is to be able to tell what the gradient of a given variable is with respect to my error function given some data.<\/p>\n\n<p>One way to do this would be to see how much the variable has changed after a call to train, but obviously that can vary massively based on the learning algorithm (for example it would be almost impossible to tell with something like RProp) and just isn't very clean.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"35227842","CreationDate":"2016-02-05T15:32:50.723","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gradients\" rel=\"noreferrer\"><code>tf.gradients()<\/code><\/a> function allows you to compute the symbolic gradient of one tensor with respect to one or more other tensors&mdash;including variables.  Consider the following simple example:<\/p>\n\n<pre><code>data = tf.placeholder(tf.float32)\nvar = tf.Variable(...)              # Must be a tf.float32 or tf.float64 variable.\nloss = some_function_of(var, data)  # some_function_of() returns a `Tensor`.\n\nvar_grad = tf.gradients(loss, [var])[0]\n<\/code><\/pre>\n\n<p>You can then use this symbolic gradient to evaluate the  gradient in some specific point (data):<\/p>\n\n<pre><code>sess = tf.Session()\n\nvar_grad_val = sess.run(var_grad, feed_dict={data: ...})\n<\/code><\/pre>\n"},{"AnswerId":"58314728","CreationDate":"2019-10-10T02:45:05.267","ParentId":null,"OwnerUserId":"1699075","Title":null,"Body":"<p>In TensorFlow 2.0 you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GradientTape\" rel=\"noreferrer\"><code>GradientTape<\/code><\/a> to achieve this. <code>GradientTape<\/code> records the gradients of any computation that happens in the context of that. Below is an example of how you might do that.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Here goes the neural network weights as tf.Variable\nx = tf.Variable(3.0)\n\nwith tf.GradientTape() as tape:\n  # Doing the computation in the context of the gradient tape\n  # For example computing loss\n  y = x**2 \n\n# Getting the gradient of weight w.r.t loss \ndy_dx = tape.gradient(y, x) \nprint(dy_dx) # Returns 6\n<\/code><\/pre>\n"}]}
{"QuestionId":35227224,"AnswerCount":4,"Tags":"<torch>","CreationDate":"2016-02-05T15:02:26.417","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch - repeat tensor like numpy repeat","Body":"<p>I am trying to repeat a tensor in torch in two ways. For example repeating the tensor  3 times both ways to yield;<\/p>\n\n<pre><\/pre>\n\n<p>There is a built in torch:repeatTensor function which will generate the first of the two (like ) but I can't find one for the latter (like ). I'm sure that I could call sort on the first to give the second but I think this might be computationally expensive for larger arrays?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"52259068","CreationDate":"2018-09-10T13:38:47.237","ParentId":null,"OwnerUserId":"4982729","Title":null,"Body":"<p>Quoting <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-tile-a-tensor\/13853\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-to-tile-a-tensor\/13853<\/a> - <\/p>\n\n<blockquote>\n<pre><code>z = torch.FloatTensor([[1,2,3],[4,5,6],[7,8,9]])\n1 2 3\n4 5 6\n7 8 9\nz.transpose(0,1).repeat(1,3).view(-1, 3).transpose(0,1)\n1 1 1 2 2 2 3 3 3\n4 4 4 5 5 5 6 6 6\n7 7 7 8 8 8 9 9 9\n<\/code><\/pre>\n<\/blockquote>\n\n<p>This will give you a intuitive feel of how it works.<\/p>\n"},{"AnswerId":"35387967","CreationDate":"2016-02-14T03:15:42.987","ParentId":null,"OwnerUserId":"1257954","Title":null,"Body":"<pre><code>a = torch.Tensor{1,2,3,4}\n<\/code><\/pre>\n\n<p>To get <code>{1,2,3,4,1,2,3,4,1,2,3,4}<\/code> we repeat the tensor thrice in the 1st dimension:<\/p>\n\n<pre><code>a:repeatTensor(3)\n<\/code><\/pre>\n\n<p>To get <code>{1,1,1,2,2,2,3,3,3,4,4,4}<\/code> we add a dimension to the tensor and repeat it thrice in the 2nd dimension to get a <code>4 x 3<\/code> tensor, which we can flatten.<\/p>\n\n<pre><code>b = a:reshape(4,1):repeatTensor(1,3)\nb:view(b:nElement())\n<\/code><\/pre>\n"},{"AnswerId":"55811350","CreationDate":"2019-04-23T12:34:27.077","ParentId":null,"OwnerUserId":"1155034","Title":null,"Body":"<p>Here's a generic function that repeats elements in tensors.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def repeat(tensor, dims):\n    if len(dims) != len(tensor.shape):\n        raise ValueError(\"The length of the second argument must equal the number of dimensions of the first.\")\n    for index, dim in enumerate(dims):\n        repetition_vector = [1]*(len(dims)+1)\n        repetition_vector[index+1] = dim\n        new_tensor_shape = list(tensor.shape)\n        new_tensor_shape[index] *= dim\n        tensor = tensor.unsqueeze(index+1).repeat(repetition_vector).reshape(new_tensor_shape)\n    return tensor\n<\/code><\/pre>\n\n<p>If you have <\/p>\n\n<pre><code>foo = tensor([[1, 2],\n              [3, 4]])\n<\/code><\/pre>\n\n<p>By calling <code>repeat(foo, [2,1])<\/code> you get<\/p>\n\n<pre><code>tensor([[1, 2],\n        [1, 2],\n        [3, 4],\n        [3, 4]])\n<\/code><\/pre>\n\n<p>So you duplicated every element along dimension 0 and left elements as they are on dimension 1.<\/p>\n"},{"AnswerId":"56161541","CreationDate":"2019-05-16T05:44:37.950","ParentId":null,"OwnerUserId":"6830722","Title":null,"Body":"<p><strong>Can you try something like:<\/strong><\/p>\n\n<pre><code>import torch as pt\n\n#1 work as numpy tile\n\nb = pt.arange(10)\nprint(b.repeat(3))\n\n#2 work as numpy tile\n\nb = pt.tensor(1).repeat(10).reshape(2,-1)\nprint(b)\n\n#3 work as numpy repeat\n\nt = pt.tensor([1,2,3])\nt.repeat(2).reshape(2,-1).transpose(1,0).reshape(-1)\n<\/code><\/pre>\n"}]}
{"QuestionId":35228095,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2016-02-05T15:44:40.177","AcceptedAnswerId":35256423.0,"OwnerUserId":5606352.0,"Title":"Python\/Keras\/Neural Networks - Basics to make it run","Body":"<p>I'm a beginner working with Keras to make predictions.\nI understand the concept and all the theory behind it. But I'm having a hard time to make it run. At this stage I'm not worried about the efficiency of it, I just want to run it and see an output so I can evolve later.<\/p>\n\n<p>I have this dummy Pandas DataFrame I'm using as predictor_train (X):<\/p>\n\n<pre><\/pre>\n\n<p>I have this other DataFrame which I'm using as target_train (Y):<\/p>\n\n<pre><\/pre>\n\n<p>I'm using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>And I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any insights on how to make this mammoth move?<\/p>\n","answers":[{"AnswerId":"35256423","CreationDate":"2016-02-07T17:07:41.853","ParentId":null,"OwnerUserId":"5699422","Title":null,"Body":"<p>So, you don't want to set your <code>input_dim<\/code> to 129.  That's just the number of samples.  Keras doesn't really care if your data has 1 row or 1 million rows; that's not the dimensionality of your data.  If your data were a numpy array, the input shape would be (1, 2, 3) = (2, 3).  The number of rows (1) doesn't make a difference.<\/p>\n\n<p>But, that's not the issue here.  The error that you're receiving doesn't have to do with your specified dimensionality.  If that were the case, then you'd be receiving a different error (<code>ValueError<\/code> and not <code>IndexError<\/code>) that traces to your backend (Theano or Tensorflow), if there was in fact a shape mismatch between your data and what you specified in the model.  For example:<\/p>\n\n<pre><code>ValueError: ('shapes (x,x) and (x,x) not aligned: x (dim 1) != x (dim 0)', (x, x), (x, x))\n<\/code><\/pre>\n\n<p>Looking at the error message, this is a problem with how you're preprocessing your data; i.e., the problem is with pandas and not with Keras.  You're using Keras' <code>slice_X<\/code> function to segment your data, but when Keras talks to pandas and tells it that you want to get the data from <code>dataframe[x:y]<\/code>, pandas is responding that the indices you've specified aren't valid given your dataframe.<\/p>\n\n<p>After that's sorted, though, you might run into different errors once your program gets to the Keras functions that you're calling in <code>model.fit<\/code>.  Specifically, you'll probably see a shape mismatch between your data and what your model is expecting.  How are your columns being encoded before being fed into your Keras model?  <a href=\"http:\/\/keras.io\/models\/#methods\" rel=\"nofollow\">Keras is expecting a numpy array<\/a> for <code>X<\/code> and <code>y<\/code> in <code>model.fit<\/code>.  Also, what kind of output are you expecting such that <code>output_dim = 10<\/code>?<\/p>\n\n<p>I recommend <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/addition_rnn.py\" rel=\"nofollow\">this Keras example<\/a> that shows sequence to sequence prediction with an LSTM model, in this case learning addition, since it does exactly what you want; i.e., it's concrete example that shows the model running and evolving at every epoch.<\/p>\n\n<pre><code>--------------------------------------------------\nIteration 1\nTrain on 45000 samples, validate on 5000 samples\nEpoch 1\/1\n45000\/45000 [==============================] - 30s - loss: 1.6745 - acc:0.3927- val_loss: 1.5373 - val_acc: 0.4320\nQ 58333+69862\nT 128195\n\u2612 13335 \n---\n.\n.\n.\n---\nQ 83911+65   \nT 83976 \n\u2612 11115 \n---\n<\/code><\/pre>\n\n<p>The model starts learning what's up around epoch 7:<\/p>\n\n<pre><code>Iteration 7\nTrain on 45000 samples, validate on 5000 samples\nEpoch 1\/1\n45000\/45000 [==============================] - 31s - loss: 0.9129 - acc: 0.6549 - val_loss: 0.9117 - val_acc: 0.6510\nQ 80+3375    \nT 3455  \n\u2612 3420  \n---\n.\n.\n.\n---\nQ 6+50853    \nT 50859 \n\u2611 50859    &lt;--------------------------\n---\n<\/code><\/pre>\n\n<p>You can add the following print line to the <code>decode<\/code> function in the <code>CharacterTable<\/code> class to the following in order to see what's happening.  In this case, the source addition questions are inputted in reverse, which improves performance in this type of model by introducing short term dependencies between the inputs and targets (<a href=\"http:\/\/papers.nips.cc\/paper\/5346-sequence-to-sequence-learning-with-neural-networks.pdf\" rel=\"nofollow\">Sutskever, Vinyals, and Le 2014<\/a>)<\/p>\n\n<pre><code>    def decode(self, X, calc_argmax=True):\n        if calc_argmax:\n            X = X.argmax(axis=-1)\n        print('{0} --&gt; {1}'.format(X, [self.indices_char[x] for x in X]))\n        return ''.join(self.indices_char[x] for x in X)\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>Iteration 3\nTrain on 45000 samples, validate on 5000 samples\nEpoch 1\/1\n45000\/45000 [==============================] - 21s - loss: 1.4865 - acc: 0.4491 - val_loss: 1.4132 - val_acc: 0.4676\n[ 0 11 12  2  4 11 12] --&gt; [' ', '7', '8', '+', '0', '7', '8']\n[13  9 11  0] --&gt; ['9', '5', '7', ' ']\n[12  4 11  0] --&gt; ['8', '0', '7', ' ']\nQ 870+87 \nT 957 \n\u2612 807 \n---\n[ 0  8 10  3  8  4 10] --&gt; [' ', '4', '6', '-', '4', '0', '6']\n[9 8 4 0] --&gt; ['5', '4', '0', ' ']\n[10  4 13  0] --&gt; ['6', '0', '9', ' ']\nQ 604-64 \nT 540 \n\u2612 609 \n---\n<\/code><\/pre>\n\n<p>This encoder\/decoder model can be changed quite easily to adapt to other kinds of sequence to sequence prediction problems.<\/p>\n\n<p>You'd probably want to revise the code to fit your data.  But, in your case, if you were going to feed the above data into the model as it is, I believe your <code>input_dim<\/code> would be <code>len('0123456789.- ')<\/code>, since that's the dimensionality of your data; i.e., the set of all characters in your inputs and targets.  In this linked example, sequences are converted to one-hot encodings of the individual characters.  <\/p>\n\n<p>So, you'd specify an input shape in the first layer of the model: <\/p>\n\n<pre><code>model.add(LSTM(HIDDEN_DIMENSIONS, input_shape=(MAX_LENGTH, INPUT_DIMENSIONALITY)))\n<\/code><\/pre>\n"}]}
{"QuestionId":35228133,"AnswerCount":0,"Tags":"<python><list><definition><theano>","CreationDate":"2016-02-05T15:46:46.830","AcceptedAnswerId":null,"OwnerUserId":594353.0,"Title":"skip-thought Theano model - list not declared","Body":"<p>I have defined a list  to take the contents of a file and feed the preprocessor of the skip-though model in Theano. The error says .<\/p>\n\n<p><a href=\"https:\/\/github.com\/ryankiros\/skip-thoughts\/blob\/master\/skipthoughts.\" rel=\"nofollow\">https:\/\/github.com\/ryankiros\/skip-thoughts\/blob\/master\/skipthoughts.py<\/a><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35235092,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-02-05T23:09:11.227","AcceptedAnswerId":null,"OwnerUserId":4040998.0,"Title":"Add Bias to a Matrix in Torch","Body":"<p>In Torch, how do I add a bias vector to each input to when i have a batch input? Suppose I have an input 3*2 matrix (where 2 = number of classes)<\/p>\n\n<pre><\/pre>\n\n<p>and I want to add the bias value to each element in the output matrix:<\/p>\n\n<pre><\/pre>\n\n<p>My final output should look like:<\/p>\n\n<pre><\/pre>\n\n<p>I am new to Torch and just figuring out the Syntax.<\/p>\n","answers":[{"AnswerId":"35235137","CreationDate":"2016-02-05T23:13:23.640","ParentId":null,"OwnerUserId":"2857133","Title":null,"Body":"<p>You can expand the bias to have the same dimensions as your input:<\/p>\n\n<p><code>expandedBias=torch.expand(BIAS,3,2)<\/code><\/p>\n\n<p>yields:<\/p>\n\n<p><code>th&gt; expandedBias\n 0.6588  0.6525\n 0.6588  0.6525\n 0.6588  0.6525<\/code><\/p>\n\n<p>After that you can simply add them:<\/p>\n\n<p><code>output=A+expandedBias<\/code><\/p>\n\n<p>to give:<\/p>\n\n<p><code>th&gt; A+expandedBias\n 1.4779  0.9155\n 1.1931  1.1063\n 1.3967  1.2410<\/code><\/p>\n"}]}
{"QuestionId":35235606,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-06T00:06:49.407","AcceptedAnswerId":35244793.0,"OwnerUserId":2615676.0,"Title":"Combining conditionals and control dependencies","Body":"<p>I'm trying to execute a conditional piece of code that is in turn dependent on another op executing first. Simple versions of this work, like the following:<\/p>\n\n<pre><\/pre>\n\n<p>Where evaluating  sets  to  as expected. However this more complex version doesn't work:<\/p>\n\n<pre><\/pre>\n\n<p>In particular  gets assigned  instead of . The logic I'm going for is for  to first be assigned , and <em>then<\/em> gets trimmed to . Incidentally this appears to have something to do with the size of  changing, since if in the initial  assignment  gets assigned  instead of , then evaluating  results in  being assigned , which is the correct behavior. I.e. it first gets updated to , and then trimmed to . <\/p>\n","answers":[{"AnswerId":"35244793","CreationDate":"2016-02-06T18:30:28.953","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>Note that the <code>with tf.control_dependencies<\/code> only applies to operations created inside the block. When you call <code>rest(x)<\/code> inside the block the <code>x<\/code> you are referring to is still the old <code>x<\/code> which was the return value of <code>tf.Variable<\/code> function, which simply is the <code>Tensor<\/code> holding the initial value of the variable. You can pass the new value by calling <code>rest(x_op)<\/code> instead. Here the complete working snippet :<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef rest(x): return tf.gather(x, tf.range(1, tf.size(x)))\n\nx = tf.Variable([0., 1.])\nx_op = tf.assign(x, [0., 1., 2.], validate_shape=False)\n\nwith tf.control_dependencies([x_op]):\n  true_fun  = lambda: tf.assign(x, rest(x_op), validate_shape=False)\n  false_fun = lambda: tf.constant([])\n  pred = tf.constant(True)\n  cond_op = tf.cond(pred, true_fun, false_fun)\n\nwith tf.Session(\"\"):\n  x.initializer.run()\n  print(cond_op.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":35235697,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-06T00:18:16.630","AcceptedAnswerId":35236642.0,"OwnerUserId":1085291.0,"Title":"Reading binary data in float32","Body":"<p>I want to train a network using Tensorflow, based on features from a time signal. Data is split up in  3 second epochs with  features for each epoch. Thus, the data has the form<\/p>\n\n<pre><\/pre>\n\n<p>Loading data to Tensorflow, I am trying to follow the cifar example and using . Thus, I have taken the data, and saved it to a binary file of type  with first label for the first epoch, followed by the  features for the first epoch, then the second, etc.<\/p>\n\n<p>Reading this into Tensorflow is a challenge for me, however. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Print output was:<\/p>\n\n<pre><\/pre>\n\n<p>Which surprises me, because since I have cast the values to float32, I expected the dimensions to be respectively 1 and 10, which are the actual numbers, but they are 4 and 40, which corresponds to the byte lengths.<\/p>\n\n<p>How come?<\/p>\n","answers":[{"AnswerId":"35236642","CreationDate":"2016-02-06T02:26:13.953","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I think the issue stems from the fact that <code>tf.decode_raw(value, tf.float32, little_endian=True)<\/code> returns a vector of type <code>tf.float32<\/code> rather than a vector of bytes. The slice size for extracting the features should be specified as a count of floating-point values (i.e. <code>NUM_FEATURES<\/code>) rather than a count of bytes (<code>features_bytes<\/code>).<\/p>\n\n<p>However, there's the slight wrinkle that your label is an integer, while the rest of the vector contains floating-point values. TensorFlow doesn't have many facilities for casting between binary representations (except <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#decode_raw\" rel=\"nofollow\"><code>tf.decode_raw()<\/code><\/a>), so you'll have to decode the string twice into different types:<\/p>\n\n<pre><code># Decode the result from bytes to int32\nvalue_as_ints = tf.decode_raw(value, tf.int32, little_endian=True)\nresult.label = value_as_ints[0]\n\n# Decode the result from bytes to float32\nvalue_as_floats = tf.decode_raw(value, tf.float32, little_endian=True)\nresult.features = value_as_floats[1:1+NUM_FEATURES]\n<\/code><\/pre>\n\n<p>Note that this only works because <code>sizeof(tf.int32) == sizeof(tf.float32)<\/code>, which wouldn't be true in general. Some more string manipulation tools would be useful for slicing out appropriate substrings of the raw <code>value<\/code> in the more general case. Hopefully this should be enough to get you going, though.<\/p>\n"}]}
{"QuestionId":35236851,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-02-06T03:00:44.430","AcceptedAnswerId":null,"OwnerUserId":5890827.0,"Title":"tensorflow no module named example.tutorials.mnist.input_data","Body":"<p>I've installed tensor flow on Mac OS X.\nSuccessfully ran simple command line test.\nNow trying the first tutorial.<\/p>\n\n<p>Fail on the first python line:<\/p>\n\n<p>[python prompt:] \nimport tensorflow.examples.tutorials.mnist.input_data<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \nImportError: No module named examples.tutorials.mnist.input_data<\/p>\n\n<p>But the file seems to be there:<\/p>\n\n<p>new-host-4:~ karlovitz$ ls \/Library\/Python\/2.7\/site-packages\/tensorflow\/examples\/tutorials\/mnist\/\nBUILD           fully_connected_feed.py mnist.py        mnist_with_summaries.py\n<strong>init<\/strong>.py     input_data.py       mnist_softmax.py<\/p>\n","answers":[{"AnswerId":"35237068","CreationDate":"2016-02-06T03:38:49.310","ParentId":null,"OwnerUserId":"5099417","Title":null,"Body":"<p>@mkarlovitz Looks like <code>\/Library\/Python\/2.7\/site-packages\/<\/code> is not in the list of paths python is looking for.   <\/p>\n\n<p>To see what paths are  python uses to find packages, do the below ( you can use command line for this ).<br>\n1. <code>import sys<\/code><br>\n2. <code>sys.path<\/code>  ( This tells the list of Paths )  <\/p>\n\n<p>If the  <code>\/Library\/Python\/2.7\/site-packages\/<\/code>  is not in the above list.<br>\nAdd it as follows:<br>\nIn the Python file\/script you are executing.<br>\n1. <code>import sys<\/code><br>\n2. <code>sys.path.append('\/Library\/Python\/2.7\/site-packages\/')<\/code>   <\/p>\n"}]}
{"QuestionId":35237401,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-06T04:42:55.780","AcceptedAnswerId":35244222.0,"OwnerUserId":2628128.0,"Title":"Check for broadcasting compatibility in Tensorflow - C++ API","Body":"<p>I am implementing element-wise operations in TensorFlow. Many TensorFlow operations, e.g. add, support broadcasting (<a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"nofollow\">from numpy<\/a>). Broadcasting is possible if the following rule is respected:<\/p>\n\n<blockquote>\n  <p>When operating on two tensors, their shapes should be compared element-wise. The procedure starts with the trailing dimensions, and works its way forward. Two dimensions are compatible when they are equal, or one of them is 1. If these conditions are not met, an exception is thrown, indicating that the tensors have incompatible shapes. The size of the resulting tensor is the maximum size along each dimension of the input arrays.<\/p>\n<\/blockquote>\n\n<p>Does TensorFlow <strong>C++ API<\/strong> provide any method for comparing the <em>compatibility<\/em> of two tensors? Or, which is the fastest way to do that?<\/p>\n","answers":[{"AnswerId":"35244222","CreationDate":"2016-02-06T17:39:54.273","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>All element-wise binary operations' kernel implementations in TensorFlow derive from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/cwise_ops_common.h#L36\" rel=\"nofollow\"><code>BinaryOpShared<\/code><\/a> class, that does the compatibility checking via the helper class <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/cwise_ops_common.cc#L46\" rel=\"nofollow\"><code>BinaryOpState<\/code><\/a>. Perhaps, you can simply derive your kernel class from <code>BinaryOpShared<\/code> and get the compatibility checking for free.<\/p>\n"}]}
{"QuestionId":35239461,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><theano><lstm>","CreationDate":"2016-02-06T09:34:57.510","AcceptedAnswerId":null,"OwnerUserId":797880.0,"Title":"LSTM implementation with peephole","Body":"<p>I have been reading papers about LSTM and checking its implementations. There is one point that is not clear to me.<br>\nIn most of the papers it is mentioned that the weight matrices from the cell to gate vectors should be diagonal(ex: <a href=\"http:\/\/arxiv.org\/pdf\/1308.0850v5.pdf\" rel=\"nofollow noreferrer\">Alex<\/a> page 5, 2013), but I haven't seen this in any implementation.<br>\nFor example this :<br>\n<a href=\"http:\/\/christianherta.de\/lehre\/dataScience\/machineLearning\/neuralNetworks\/LSTM.php\" rel=\"nofollow noreferrer\">1<\/a>\n<a href=\"https:\/\/github.com\/jhb86253817\/lstmlm\/blob\/master\/code\/lstm_ada_peep.py\" rel=\"nofollow noreferrer\">2<\/a>\nAnother example is from mila lab.\n<a href=\"https:\/\/github.com\/mila-udem\/blocks\/blob\/83c5d2a853f8057be03b910ac0f0e601a59a01a9\/blocks\/bricks\/recurrent.py#L334\" rel=\"nofollow noreferrer\">3<\/a><\/p>\n\n<p>Are these people implementing wrongly or am I missing something?   <\/p>\n","answers":[{"AnswerId":"35242036","CreationDate":"2016-02-06T14:16:18.233","ParentId":null,"OwnerUserId":"2615676","Title":null,"Body":"<p>The TensorFlow implementation does use a diagonal matrix, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/97f585d506cccc57dc98f234f4d5fcd824dd3c03\/tensorflow\/python\/ops\/rnn_cell.py#L353\" rel=\"noreferrer\">here<\/a>. Note that what this means in practice is that the peepholes only go from the cell to itself, and so you're doing elementwise vector multiplies.<\/p>\n"}]}
{"QuestionId":35239999,"AnswerCount":0,"Tags":"<ubuntu><caffe>","CreationDate":"2016-02-06T10:35:09.497","AcceptedAnswerId":null,"OwnerUserId":3174567.0,"Title":"undefined symbol when make runtest in Caffe installation","Body":"<p>There is a problem when I run the following command for the first time:<\/p>\n\n<pre><\/pre>\n\n<p>The output is as following:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I run   again:<\/p>\n\n<p>There error message is different:\n<\/p>\n\n<p>There is <strong>no problem<\/strong> when:<\/p>\n\n<pre><\/pre>\n\n<p>Additionally, when I run:<\/p>\n\n<pre><\/pre>\n\n<p>There is <strong>another error message<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>I am not sure what is going wrong. Last time install cafffe, I have everything going smooth with Ubuntu 14. Now, I am using Ubuntu 15 (I must use ubuntu 15 since I faced another problem with ubuntu 14).<\/p>\n\n<p>Is there anyone have encountered this issue?\nPlease help me to fix this issue. Any idea would be appreciated.<\/p>\n","answers":[]}
{"QuestionId":35241251,"AnswerCount":2,"Tags":"<python><machine-learning><deep-learning><tensorflow>","CreationDate":"2016-02-06T12:54:40.953","AcceptedAnswerId":35241735.0,"OwnerUserId":4473303.0,"Title":"In Tensorflow, what is the difference between sampled_softmax_loss and softmax_cross_entropy_with_logits","Body":"<p>In tensorflow, there are methods called <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/nn\/sampled_softmax_loss\" rel=\"nofollow noreferrer\"><\/a>.<\/p>\n\n<p>I read the tensorflow document and searched google for more information but I couldn't find the difference. It looks like to me both calculates the loss using softmax function.<\/p>\n\n<h3>Using  to calculate the loss<\/h3>\n\n<pre><\/pre>\n\n<h3>Using  to calculate the loss<\/h3>\n\n<pre><\/pre>\n\n<p>To me, calculating softmax loss is same as calculating softmaxed cross entropy (e.g. )<\/p>\n\n<p>Could somebody tell me the why there is two different methods and which method should I use in which case?<\/p>\n","answers":[{"AnswerId":"35241735","CreationDate":"2016-02-06T13:45:51.510","ParentId":null,"OwnerUserId":"1886138","Title":null,"Body":"<p>If your target vocabulary(or in other words amount of classes you want to predict) is really big, it is very hard to use regular softmax, because you have to calculate probability for every word in dictionary. By Using  <code>sampled_softmax_loss<\/code> you only take in account subset <strong>V<\/strong> of your vocabulary to calculate your loss.<\/p>\n\n<p>Sampled softmax only makes sense if we sample(our <strong>V<\/strong>) less than vocabulary size. If your vocabulary(amount of labels) is small, there is no point using <code>sampled_softmax_loss<\/code>.<\/p>\n\n<p>You can see implementation details in this paper:\n<a href=\"http:\/\/arxiv.org\/pdf\/1412.2007v2.pdf\" rel=\"noreferrer\">http:\/\/arxiv.org\/pdf\/1412.2007v2.pdf<\/a><\/p>\n\n<p>Also you can see example where it is used - Sequence to sequence translation in this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/translate\/seq2seq_model.py\" rel=\"noreferrer\">example<\/a><\/p>\n"},{"AnswerId":"44896285","CreationDate":"2017-07-04T02:54:14.843","ParentId":null,"OwnerUserId":"2476920","Title":null,"Body":"<h2>Sampled:<\/h2>\n\n<p>Sampled, in both cases, means you don't calculate it for all of what's possible as an output (e.g.: there are too many words in a dictionary to take all of them at each derivation, so we take just a few sample and learn on that for NLP problems). <\/p>\n\n<h3><code>softmax_cross_entropy_with_logits<\/code>:<\/h3>\n\n<p>This is the cross entropy and receives logits as inputs and yields what can be used as a loss.<\/p>\n\n<h3><code>sampled_softmax_loss<\/code>:<\/h3>\n\n<p>This is a sampled softmax_cross_entropy_with_logits, so it takes just a few samples before using the cross enthropy rather than using the full cross enthropy: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/ops\/nn_impl.py#L1269\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/ops\/nn_impl.py#L1269<\/a><\/p>\n"}]}
{"QuestionId":35241827,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-06T13:54:45.517","AcceptedAnswerId":35244828.0,"OwnerUserId":258483.0,"Title":"How to use \"FLAGS\" (command line switches) in TensorFlow?","Body":"<p>I am trying to set custom batch size in my application.<\/p>\n\n<p>If I put the following code into my app<\/p>\n\n<pre><\/pre>\n\n<p>it says the following error<\/p>\n\n<pre><\/pre>\n\n<p>and if I remove this statement, it swears:<\/p>\n\n<pre><\/pre>\n\n<p>at the line where  used. <\/p>\n\n<p> is the name of my script and I didn't write this message anywhere and don't expect these command line switches at all. Looks like  uses some Python switch parsing library and expecting these switches somehow. How to avoid this and expect custom switches?<\/p>\n\n<p>How to hardcode custom batch_size?<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>My command line is follows:<\/p>\n\n<pre><\/pre>\n\n<p>PNGs are images from CIFAR database I wish to recognize from command line. This is command line I wish it to be, I don't wish it contain options listed in \"usage\" output.<\/p>\n","answers":[{"AnswerId":"35244098","CreationDate":"2016-02-06T17:30:17.697","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>I suspect you are importing <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow\">cifar10.py<\/a> that already has the <code>batch_size<\/code> flag defined, and the error is due to you trying to re-define a flag with the same name. If you are importing <code>cifar10<\/code>, you can simply use <code>--batch_size<\/code> from the command line, and <code>FLAGS.batch_size<\/code> in your code.<\/p>\n"},{"AnswerId":"35244828","CreationDate":"2016-02-06T18:34:16.787","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From your update, it sounds like you don't want to use the <code>FLAGS<\/code> module at all. If you look at a program like <code>cifar10_train.py<\/code>, you'll see the following <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/e008ecf12607f691952a9bfe8971f406b30bc56e\/tensorflow\/models\/image\/cifar10\/cifar10_train.py#L129\">near the bottom of the script<\/a>:<\/p>\n\n<pre><code>def main(argv=None):  # pylint: disable=unused-argument\n  # ...\n\nif __name__ == '__main__':\n  tf.app.run()\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/e008ecf12607f691952a9bfe8971f406b30bc56e\/tensorflow\/python\/platform\/default\/_app.py\"><code>tf.app.run()<\/code><\/a> invocation is a bit of boilerplate that ensures that any flags are parsed, and then invokes the <code>main()<\/code> function in the same module. Notice that <code>main()<\/code> has an <code>argv<\/code> argument. This will be filled with the remaining arguments to your program: in your example, it will be a list <code>[\"image1.png\", \"image2.png\", \"image3.png\"]<\/code>. Therefore you can simply write your <code>main()<\/code> function to be something like:<\/p>\n\n<pre><code>def main(argv=None):\n  if argv:\n    for filename in argv:\n      run_inference_on_file(filename)\n<\/code><\/pre>\n"}]}
{"QuestionId":35247406,"AnswerCount":2,"Tags":"<android><python><tensorflow>","CreationDate":"2016-02-06T22:41:34.607","AcceptedAnswerId":35525092.0,"OwnerUserId":371408.0,"Title":"Converting trained Tensorflow model to protobuf","Body":"<p>The inception net in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\">Android Tensorflow Demo<\/a> stores the model as a protobuf file (<a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\">tensorflow_inception_graph.pb<\/a>). I'd like to replace this network with another one.<\/p>\n\n<p>Is there documentation or an example somewhere on how to convert a trained TensorFlow network in Python to ? I'm aware of TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/state_ops.html#Saver\">Saver<\/a> but this seems to be used to save intermediate training state. Not sure how it works if the model is already trained.<\/p>\n","answers":[{"AnswerId":"35248210","CreationDate":"2016-02-07T00:16:31.083","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's an example of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/00440e99ffb1ed1cfe4b4ea650e0c560838a6edc\/tensorflow\/python\/tools\/freeze_graph_test.py#L68\" rel=\"noreferrer\">saving<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/00440e99ffb1ed1cfe4b4ea650e0c560838a6edc\/tensorflow\/python\/tools\/freeze_graph_test.py#L79\" rel=\"noreferrer\">loading<\/a><\/p>\n"},{"AnswerId":"35525092","CreationDate":"2016-02-20T15:14:39.003","ParentId":null,"OwnerUserId":"5182129","Title":null,"Body":"<p><code>tf.Graph.as_graph_def()<\/code> retrieves the serialized graph. Then you just write it to a file.<\/p>\n"}]}
{"QuestionId":35248476,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-02-07T00:54:36.070","AcceptedAnswerId":null,"OwnerUserId":1289540.0,"Title":"Tensorflow compat modules issues?","Body":"<p>Getting the following error when working through the ipython notebooks on Google's tensorflow udacity course:    <\/p>\n\n<blockquote>\n  <p>AttributeError: 'module' object has no attribute 'compat'<\/p>\n<\/blockquote>\n\n<p>Trying to call:<\/p>\n\n<blockquote>\n  <p>tf.compat.as_str(f.read(name)).split()<\/p>\n<\/blockquote>\n\n<p>Running on Ubuntu 14.04 and wondering if this a tensorflow early bug issue or just me being stupid. :P<\/p>\n","answers":[{"AnswerId":"35256493","CreationDate":"2016-02-07T17:14:18.120","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You're most likely using an older version of TensorFlow.  I just noticed that some of our install docs still link to 0.5 -- try upgrading to 0.6 or to head.<\/p>\n\n<p>I'll fix the docs soon, but in the meantime, if you installed via pip, you can just change the 0.5 to 0.6 in the path.  If you're building from source, just check out the appropriate release tag (or head).<\/p>\n"},{"AnswerId":"35318403","CreationDate":"2016-02-10T14:51:36.003","ParentId":null,"OwnerUserId":"682227","Title":null,"Body":"<p>For me (on python 2.7) this seems to work:<\/p>\n\n<pre><code>return f.read(name).encode('utf-8').split()\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/python\/util\/compat.py?autodive=0%2F%2F\" rel=\"nofollow\">the module source<\/a> for what may help under other circumstances.<\/p>\n"},{"AnswerId":"35641400","CreationDate":"2016-02-26T00:55:11.860","ParentId":null,"OwnerUserId":"3147590","Title":null,"Body":"<p>I just solved the problem by reinstalling Tensorflow by using \"pip\" command (Ubuntu 14.04, 64-bit) which is suggested in the following link:\n   \"<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html<\/a>\"<\/p>\n\n<p>I guess you haven't used Docker or other virtualenv. So when another program is removed or installed it happens that it affects Tensorflows functionality. i guess you can solve the problem the way I did just a minute ago.<\/p>\n"},{"AnswerId":"35627095","CreationDate":"2016-02-25T12:18:13.183","ParentId":null,"OwnerUserId":"5107003","Title":null,"Body":"<p>For me(on python2.7), I copy the file: <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/python\/util\/compat.py?autodive=0%2F%2F\" rel=\"nofollow\">compat.py<\/a> to the build folder. \nThen, add the commands to the file:<\/p>\n\n<pre><code>import compat as cp\n<\/code><\/pre>\n\n<p>Next, replace the call as:<\/p>\n\n<pre><code>#tf.compat.as_str(f.read(name)).split()\ncp.as_str(f.read(name)).split()\n<\/code><\/pre>\n\n<p>I think this is the simplest solution.<\/p>\n"}]}
{"QuestionId":35251924,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-02-07T09:43:07.723","AcceptedAnswerId":null,"OwnerUserId":3654968.0,"Title":"Can tensors be concatenated without losing reference in torch (lua)?","Body":"<pre><\/pre>\n\n<p>Modifying z <strong>does not<\/strong> modify x. Is there any way of concatenating x and y such that modifying z does modify x?<\/p>\n","answers":[{"AnswerId":"35259902","CreationDate":"2016-02-07T22:25:00.690","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>You can achieve this kind of behaviour, but the other way around. You should start with a bigger tensor, your main \"storage\", and then you can create subtensors, which will share internal state.<\/p>\n\n<p>See in particular <code>:sub<\/code> method from torch (following code sample taken from <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md\" rel=\"nofollow\">Torch doc<\/a>)<\/p>\n\n<pre><code>x = torch.Tensor(5, 6):zero()\n&gt; x\n 0 0 0 0 0 0\n 0 0 0 0 0 0\n 0 0 0 0 0 0\n 0 0 0 0 0 0\n 0 0 0 0 0 0\n[torch.DoubleTensor of dimension 5x6]\n\ny = x:sub(2,4):fill(1) -- y is sub-tensor of x:\n&gt; y                    -- dimension 1 starts at index 2, ends at index 4\n 1  1  1  1  1  1\n 1  1  1  1  1  1\n 1  1  1  1  1  1\n[torch.DoubleTensor of dimension 3x6]\n\n&gt; x                    -- x has been modified!\n 0  0  0  0  0  0\n 1  1  1  1  1  1\n 1  1  1  1  1  1\n 1  1  1  1  1  1\n 0  0  0  0  0  0\n[torch.DoubleTensor of dimension 5x6]\n<\/code><\/pre>\n\n<p>As you can see you have <code>y<\/code> variable which is actually a part of <code>x<\/code>, and changing its values - change x's too. This is very generic way thus you can share multiple parts of tensor.<\/p>\n\n<p>Thus in your case it would be something like<\/p>\n\n<pre><code>z = torch.Tensor(4):zero()\nx = z:sub(1, 2)\ny = z:sub(3, 4)\nx[1] = 2\ny[2] = 8\nprint(z)\n<\/code><\/pre>\n\n<p>prints <\/p>\n\n<pre><code> 2\n 0\n 0\n 8\n[torch.DoubleTensor of size 4]\n<\/code><\/pre>\n\n<p>as desired.<\/p>\n"}]}
{"QuestionId":35252057,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-02-07T09:59:31.183","AcceptedAnswerId":35252155.0,"OwnerUserId":3698878.0,"Title":"Epoch size (in terms of iterations) with data augmentation caffe","Body":"<p>Suppose if one has training examples, and his batch size is 500, then it will take 2 iterations to complete 1 epoch. Now let's say I am using the caffe framework's on the fly data-augmentation, i.e; 10 crops per example.<br>\nMy question is will the epoch size will still be 2 iterations as in the above examples or become 2*10 = 20?<\/p>\n","answers":[{"AnswerId":"35252155","CreationDate":"2016-02-07T10:10:27.603","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>An epoch is the the number of iterations it takes to go over the training data once. Since you augment your data, it will take you 10 times more iterations to complete one pass over the training data. Hence <code>1 epoch = 2*10 iterations<\/code> now.<\/p>\n"}]}
{"QuestionId":35252888,"AnswerCount":12,"Tags":"<python><tensorflow>","CreationDate":"2016-02-07T11:27:38.060","AcceptedAnswerId":null,"OwnerUserId":1233675.0,"Title":"tensorflow installation problems","Body":"<p>I tried installing tensorflow on my ubuntu 14.04 64bit machine:<\/p>\n\n<pre><\/pre>\n\n<p>for python 2.7<\/p>\n\n<p>When I run  in the console I get 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>Google didn't provide a good answer to this, any thoughts?<\/p>\n","answers":[{"AnswerId":"44447193","CreationDate":"2017-06-08T23:46:43.107","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>I ran in to a similar issue and it turned out it was because the protobuf dependency wasn't met, or there was an issue with protobuf. I would uninstall protobuf:<\/p>\n\n<pre><code>&gt;&gt; sudo pip uninstall protobuf\n<\/code><\/pre>\n\n<p>Then re-install:<\/p>\n\n<pre><code>&gt;&gt; sudo pip install --upgrade protobuf\n<\/code><\/pre>\n\n<p>If that doesn't work, try uninstalling and re-installing both protobuf and tensorflow. If you are building from source, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1402\" rel=\"nofollow noreferrer\">check out this issue on GitHub<\/a><\/p>\n"},{"AnswerId":"40823297","CreationDate":"2016-11-26T21:22:33.833","ParentId":null,"OwnerUserId":"4674549","Title":null,"Body":"<p>Follow these steps:<\/p>\n\n<pre><code>sudo apt-get install python-pip python-dev\n\nexport TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.11.0-cp27-none-linux_x86_64.whl\n\nsudo pip install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>And then run python. It should work.<\/p>\n"},{"AnswerId":"48981925","CreationDate":"2018-02-26T05:25:42.197","ParentId":null,"OwnerUserId":"9411475","Title":null,"Body":"<p>It shows this kind of error because the python s\/w you installed is not supported by tensorflow.\nSo uninstall the previous version of python and go to the <code>tensorflow<\/code> installation doc then install the specified python software instructed there.   <\/p>\n"},{"AnswerId":"37076504","CreationDate":"2016-05-06T15:47:14.047","ParentId":null,"OwnerUserId":"1303432","Title":null,"Body":"<p>Before you type command <code>python<\/code> in terminal, your workspace shouldn't be the directory which contain the tensorflow-0.6.0-cp27-none-linux_x86_64.whl.\ni.e. your workspace should be different from the directory where tensorflow-0.6.0-cp27-none-linux_x86_64.whl is.<\/p>\n"},{"AnswerId":"51076520","CreationDate":"2018-06-28T06:58:16.297","ParentId":null,"OwnerUserId":"10003973","Title":null,"Body":"<p>it's very painful if you're not very familiar with python but you install tensorflow from source code. you cannot import tensorflow from your tensorflow directory. If you exit your TENSORFLOW_ROOT, it should work.<\/p>\n"},{"AnswerId":"45814760","CreationDate":"2017-08-22T10:08:37.787","ParentId":null,"OwnerUserId":"7034507","Title":null,"Body":"<p>The problem is with your tensorflow installation version and environmental configuration. I have also faced same issue, as I had installed tensorflow for Python version 3.5 while my system has 3.6. <\/p>\n\n<p>Select the correct wheel for your installation. \n<a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/pypi.python.org\/pypi\/tensorflow<\/a><\/p>\n\n<p>It should work.<\/p>\n"},{"AnswerId":"36575573","CreationDate":"2016-04-12T14:00:50.777","ParentId":null,"OwnerUserId":"2796145","Title":null,"Body":"<p>I had a similar problem. The error message also contained the line \"ImportError: No module named google.protobuf\". As mentioned here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1415\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1415<\/a> , a workaround for this problem is to create an empty __init_.py file in the tensorflow\/google directory, where you should also have the protobuf folder. Hope this helps.<\/p>\n"},{"AnswerId":"47150833","CreationDate":"2017-11-07T05:51:19.853","ParentId":null,"OwnerUserId":"243072","Title":null,"Body":"<p>I recommend you use <a href=\"https:\/\/anaconda.org\" rel=\"nofollow noreferrer\">Anaconda<\/a>, it's really helpful for ML and data science job.<\/p>\n"},{"AnswerId":"40031611","CreationDate":"2016-10-13T21:52:43.103","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>You cannot import tensorflow whilst you are in the tensorflow source directory. Change Directory i.e. <code>cd ..<\/code> first before running python. <\/p>\n\n<p>I've emphasized the key part of your error: \"Error importing tensorflow. Unless you are using bazel, <strong>you should not try to import tensorflow from its source directory<\/strong>; please exit the tensorflow source tree, and relaunch your python interpreter from there.\" <\/p>\n"},{"AnswerId":"38645861","CreationDate":"2016-07-28T20:10:38.200","ParentId":null,"OwnerUserId":"3220264","Title":null,"Body":"<p>I would recommend using <a href=\"https:\/\/www.continuum.io\/downloads\" rel=\"nofollow noreferrer\">Anaconda<\/a>. It makes it easy to install all dependencies needed (most are automatically installed) and there are multiple options for installing tensorflow.<\/p>\n\n<pre><code>conda install tensorflow\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>conda install tensorflow-gpu\n<\/code><\/pre>\n"},{"AnswerId":"41527012","CreationDate":"2017-01-07T21:55:20.527","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>You will see this issue if you start python when you are working inside the source tree of tensorflow. (wherever you cloned the git repository)\nPython gets confused about what to import when you run \"import tensorflow\" (the one in your directory or the system installed one).\nThat is what the error message is trying to say.<\/p>\n\n<p>Simply change to your home directory, or a new empty directory and it should work.<\/p>\n"},{"AnswerId":"38233712","CreationDate":"2016-07-06T21:01:44.670","ParentId":null,"OwnerUserId":"6100416","Title":null,"Body":"<p>Strongly suggest you to use virtualenv. I have been suffering in installing tensorflow as well as its dependencies for weeks until I learn about virtualenv. You can follow the official tutorial to install it. (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/get_started\/os_setup.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.9\/get_started\/os_setup.html<\/a>) And it comes with all the library that you are likely to use. Also, you can modify your environment in your Python IDE. Then you don't have to active and deactive virtualenv every time you use it.ery Very convenient\n!<\/p>\n"}]}
{"QuestionId":35254138,"AnswerCount":1,"Tags":"<python><machine-learning><convolution><keras><lstm>","CreationDate":"2016-02-07T13:34:41.450","AcceptedAnswerId":35255185.0,"OwnerUserId":3759736.0,"Title":"Python keras how to change the size of input after convolution layer into lstm layer","Body":"<p>I have a problem with the connection between convolution layer and lstm layer. The data is of shape(75,5) where there is 75 timesteps x 5 data points for each time step. What I want to do is do a convolution on (75x5), get new convolved (75x5) data and feed that data into lstm layer. However, it does not work because the shape of output of convolution layer has number of filters which I do not need. And therefore the shape of convolution layer output is (1,75,5) and input needed for lstm layer is (75,5). How do I just take the first filter.<\/p>\n\n<pre><\/pre>\n\n<p>And this is the error that comes up:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35255185","CreationDate":"2016-02-07T15:14:12.753","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>You can add Reshape() layer in between to make dimensions compatible.<\/p>\n\n<p><a href=\"http:\/\/keras.io\/layers\/core\/#reshape\" rel=\"noreferrer\">http:\/\/keras.io\/layers\/core\/#reshape<\/a><\/p>\n\n<blockquote>\n  <p><code>keras.layers.core.Reshape(dims)<\/code><\/p>\n  \n  <p>Reshape an output to a certain shape.<\/p>\n  \n  <p><strong>Input shape<\/strong><\/p>\n  \n  <p>Arbitrary, although all dimensions in the input shaped must be fixed. Use the keyword argument input_shape (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model.<\/p>\n  \n  <p><strong>Output shape<\/strong><\/p>\n  \n  <p><code>(batch_size,) + dims<\/code><\/p>\n  \n  <p><strong>Arguments<\/strong><\/p>\n  \n  <p><code>dims<\/code>: target shape. Tuple of integers, does not include the samples dimension (batch size).<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":35254982,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-02-07T14:54:11.003","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"what is the use of min_informtaive_str in theano","Body":"<p>What is the use of ? Some examples would help me to understand. I tried the theano documentation at <a href=\"http:\/\/deeplearning.net\/software\/theano\/theano.pdf\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/theano.pdf<\/a>. But I couldn't get it.<\/p>\n","answers":[{"AnswerId":"35256741","CreationDate":"2016-02-07T17:36:10.377","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If the documentation is missing from the PDF (it does get out of date) then you can find the details in the source code. Here, for example, is <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/printing.py#L1224\" rel=\"nofollow\">the source code for <code>min_informative_str<\/code><\/a>. It says:<\/p>\n\n<pre><code>Returns a string specifying to the user what obj is\nThe string will print out as much of the graph as is needed\nfor the whole thing to be specified in terms only of constants\nor named variables.\n\n\nParameters\n----------\nobj: the name to convert to a string\nindent_level: the number of tabs the tree should start printing at\n              (nested levels of the tree will get more tabs)\n_prev_obs: should only be used by min_informative_str\n                a dictionary mapping previously converted\n                objects to short tags\n\n\nBasic design philosophy\n-----------------------\n\nThe idea behind this function is that it can be used as parts of\ncommand line tools for debugging or for error messages. The\ninformation displayed is intended to be concise and easily read by\na human. In particular, it is intended to be informative when\nworking with large graphs composed of subgraphs from several\ndifferent people's code, as in pylearn2.\n\nStopping expanding subtrees when named variables are encountered\nmakes it easier to understand what is happening when a graph\nformed by composing several different graphs made by code written\nby different authors has a bug.\n\nAn example output is:\n\nA. Elemwise{add_no_inplace}\n    B. log_likelihood_v_given_h\n    C. log_likelihood_h\n\n\nIf the user is told they have a problem computing this value, it's\nobvious that either log_likelihood_h or log_likelihood_v_given_h\nhas the wrong dimensionality. The variable's str object would only\ntell you that there was a problem with an\nElemwise{add_no_inplace}. Since there are many such ops in a\ntypical graph, such an error message is considerably less\ninformative. Error messages based on this function should convey\nmuch more information about the location in the graph of the error\nwhile remaining succint.\n\nOne final note: the use of capital letters to uniquely identify\nnodes within the graph is motivated by legibility. I do not use\nnumbers or lower case letters since these are pretty common as\nparts of names of ops, etc. I also don't use the object's id like\nin debugprint because it gives such a long string that takes time\nto visually diff.\n<\/code><\/pre>\n"}]}
{"QuestionId":35255486,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-02-07T15:42:34.250","AcceptedAnswerId":null,"OwnerUserId":4040998.0,"Title":"How to do Element wise product torch?","Body":"<p>How do i find an element wise product of two tensors? Both my tensors are of the same dimension and i want to find their product?<\/p>\n\n<pre><\/pre>\n\n<p>It should yield:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35255502","CreationDate":"2016-02-07T15:44:23.890","ParentId":null,"OwnerUserId":"2857133","Title":null,"Body":"<p>You can do this by using cmul.<\/p>\n\n<pre><code>th&gt; torch.cmul(q,w)\n  1   4   9\n  4  16  36\n<\/code><\/pre>\n\n<p>As a side note:<\/p>\n\n<p><code>q:cmul(w)<\/code>: will multiply them and store the value back into q, <\/p>\n\n<p><code>z=torch.cmul(q,w)<\/code>: will multiply them and return a new tensor which will be stored in z.<\/p>\n"}]}
{"QuestionId":35256110,"AnswerCount":2,"Tags":"<tensorflow><bazel>","CreationDate":"2016-02-07T16:38:33.580","AcceptedAnswerId":35274720.0,"OwnerUserId":160015.0,"Title":"tensorflow build fails with \"missing dependency\" error","Body":"<p>I'm completely new to bazel and tensorflow so the solution to this may be obvious to someone with some experience. My bazel build of tensorflow fails with a \"missing dependency\" error message. Here is the relevant sequence of build commands and output:<\/p>\n\n<pre><\/pre>\n\n<p>The version of bazel I'm using is , I'm running on openSUSE 13.2. I confirmed that the header files do exist which is probably expected:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Note for anyone who finds this question<\/strong>:<\/p>\n\n<p>Use Damien's answer below except that you have to use  rather than . Also if you are building for GPU acceleration you will also need to modify the  file in the tensorflow repo like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42713416","CreationDate":"2017-03-10T08:13:30.373","ParentId":null,"OwnerUserId":"7670365","Title":null,"Body":"<p>Bazel complaints of system header files because compiler uses -MD (as opposed to -MMD) flag when generating dependences. While using -MD is reasonable for an environment that changes often, listing dependency on system header files causes the 'missing dependency declarations' errors. <\/p>\n\n<p>What helped me was converting the '-MD' flag into '-MMD' flag in the compiler wrapper files third_party\/gpus\/crosstool\/clang\/bin\/crosstool_wrapper_driver_is_not_gcc.tpl just before 'subprocess.call([CPU_COMPILER]...)':<\/p>\n\n<pre><code>cpu_compiler_flags = ['-MMD' if flag == '-MD' else flag for flag in cpu_compiler_flags]\n<\/code><\/pre>\n\n<p>and third_party\/sycl\/crosstool\/computecpp.tpl, similar place:<\/p>\n\n<pre><code>computecpp_device_compiler_flags = ['-MMD' if flag == '-MD' else flag for flag in computecpp_device_compiler_flags]\n<\/code><\/pre>\n"},{"AnswerId":"35274720","CreationDate":"2016-02-08T16:42:00.600","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p>You should tweak the C++ compiler.<\/p>\n\n<p>To do so, here's the best way to proceed:\nedit the file <code>tools\/cpp\/CROSSTOOL<\/code> (<a href=\"https:\/\/github.com\/bazelbuild\/bazel\/blob\/master\/tools\/cpp\/CROSSTOOL\" rel=\"nofollow\">https:\/\/github.com\/bazelbuild\/bazel\/blob\/master\/tools\/cpp\/CROSSTOOL<\/a>) from your package path directory (should be in <code>~\/.bazel\/base_workspace<\/code>, can be found with <code>bazel info package_path<\/code>) to add a line <code>cxx_builtin_include_directory: \/usr\/lib64\/gcc<\/code> around line 100 (see <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/blob\/master\/tools\/cpp\/CROSSTOOL#L101\" rel=\"nofollow\">https:\/\/github.com\/bazelbuild\/bazel\/blob\/master\/tools\/cpp\/CROSSTOOL#L101<\/a>).<\/p>\n\n<p>Then <code>echo \"build --crosstool=\/\/tools\/cpp:toolchain\" &gt;&gt;~\/.bazelrc<\/code> and then retries to build.<\/p>\n\n<p>Sorry for the mess, we are working on making C++ toolchain work better out of the box.<\/p>\n"}]}
{"QuestionId":35260280,"AnswerCount":2,"Tags":"<torch>","CreationDate":"2016-02-07T23:04:53.520","AcceptedAnswerId":null,"OwnerUserId":4040998.0,"Title":"Transpose a column vector in torch","Body":"<p>I have a colum vector which I want to tranpose into row vector, I get the following error while doing it. Is there a way to tranpose 1 dimensional vectors in torch<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"56067654","CreationDate":"2019-05-09T21:12:57.617","ParentId":null,"OwnerUserId":"3650983","Title":null,"Body":"<p>for Transpose of 1d tensor you can do something like this:<\/p>\n\n<p>let's say you have 1D tensor b:<\/p>\n\n<pre><code>import torch\n\na = torch.rand(1,10)\nb = a[0,:]\n<\/code><\/pre>\n\n<p>and .t() not working for 1D<\/p>\n\n<pre><code>print(b)\n#print(b.t()) # not going to work 1D\n<\/code><\/pre>\n\n<p>you can use one of the following option:<\/p>\n\n<pre><code>print(b.reshape(1,-1).t())\nprint(b.reshape(-1,1))\n<\/code><\/pre>\n\n<p>first line create 1*N matrix and then transpose it, and the second create N*1 matrix that transposed <\/p>\n"},{"AnswerId":"35260334","CreationDate":"2016-02-07T23:09:19.560","ParentId":null,"OwnerUserId":"2857133","Title":null,"Body":"<p>This is because your tensor has a dimension of 1. You can only take the transpose of a tensor with dimension 2.<\/p>\n\n<p>In order to do this, first resize your tensor as <\/p>\n\n<pre><code>bb:resize(2,1)\n<\/code><\/pre>\n\n<p>After that, it should work:<\/p>\n\n<pre><code>th&gt; bb:t()\n 1  2\n<\/code><\/pre>\n\n<p>More generally, for a tensor with any other size, you can simply use:<\/p>\n\n<pre><code>bb:resize(bb:size(1),1)\n<\/code><\/pre>\n"}]}
{"QuestionId":35261789,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-08T03:01:58.003","AcceptedAnswerId":35275088.0,"OwnerUserId":2138200.0,"Title":"Efficient way to extract pool_3 for large number of images?","Body":"<p>I'd like to use pool_3 features extracted from a set of images. Currently I have a loop over each image to extract the pool_3 features:<\/p>\n\n<pre><\/pre>\n\n<p>This is quite slow though. Is there a faster batch implementation to do this?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35275088","CreationDate":"2016-02-08T16:59:47.623","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>It doesn't - yet.  I've opened <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1021\" rel=\"nofollow\">a ticket for this feature request on github<\/a> in response to another question.<\/p>\n"}]}
{"QuestionId":35264563,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><lstm>","CreationDate":"2016-02-08T07:42:36.520","AcceptedAnswerId":null,"OwnerUserId":1211587.0,"Title":"Tensorflow LSTM for noisy sequence","Body":"<p>I tried to solve Experiment 3a described in the original LSTM paper here:  <a href=\"http:\/\/deeplearning.cs.cmu.edu\/pdfs\/Hochreiter97_lstm.pdf\" rel=\"nofollow\">http:\/\/deeplearning.cs.cmu.edu\/pdfs\/Hochreiter97_lstm.pdf<\/a>  with tensorflow LSTM and failed <\/p>\n\n<p>From the paper: The task is to observe and then classify input sequences. There are two classes, each occurring with probability 0.5. There is only one input line. Only the rst N real-valued sequence elements convey relevant information about the class. Sequence elements at positions t > N are generated by a Gaussian with mean zero and variance 0.2. <\/p>\n\n<p>The net architecture that he described in the paper:\n\"We use a 3-layer net with 1 input unit, 1 output unit, and 3 cell blocks of size 1. The output layer receives connections only from memory cells. Memory cells and gate units receive inputs from input units, memory cells and gate units, and have bias weights. Gate units and output unit are logistic sigmoid in [0; 1], h in [-1; 1], and g in [-2; 2]\"<\/p>\n\n<p>I tried to reproduce it with LSTM with 3 hidden units for T=100 and N=3 but failed.<\/p>\n\n<p>I used online training (i.e. update the weights after each sequence) as described in the original paper<\/p>\n\n<p>The core of my code was as follow:<\/p>\n\n<pre><\/pre>\n\n<h1>prepare the input<\/h1>\n\n<pre><\/pre>\n\n<h1>use the last output<\/h1>\n\n<pre><\/pre>\n\n<p>Any idea why it couldn't learn? <\/p>\n\n<p>My first instinct was to create 2 outputs one for each class but in the paper he specifically mentioned only one output unit. <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35294451","CreationDate":"2016-02-09T14:23:26.487","ParentId":null,"OwnerUserId":"1211587","Title":null,"Body":"<p>It seems that i needed forget_bias > 1.0. for long sequences the network couldn't work with default forget_bias for T=50 for example i needed forget_bias = 2.1<\/p>\n"}]}
{"QuestionId":35266217,"AnswerCount":1,"Tags":"<machine-learning><neural-network><torch><cross-entropy>","CreationDate":"2016-02-08T09:31:54.287","AcceptedAnswerId":35273151.0,"OwnerUserId":45843.0,"Title":"Torch CrossEntropyCriterion error","Body":"<p>I'm trying to train a simple test network on the XOR function in Torch. It works when I use MSECriterion, but when I try CrossEntropyCriterion it fails with the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I get the same error message when decomposing it into LogSoftMax and ClassNLLCriterion. Code is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35273151","CreationDate":"2016-02-08T15:23:48.187","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>MSE criterion was designed for regression problems. When it's used for classification tasks, the targets should be one-hot vectors. Cross entropy \/ Negative log likelihood criteria are used exclusively for classification; therefore, there's no need to explicitly represent the target class as a vector. In <code>torch<\/code> the target for such criteria is just an index of the assigned class (1 to the number of classes).<\/p>\n"}]}
{"QuestionId":35266370,"AnswerCount":1,"Tags":"<python><matrix><tensorflow>","CreationDate":"2016-02-08T09:40:48.847","AcceptedAnswerId":37666032.0,"OwnerUserId":1940138.0,"Title":"TensorFlow: Compute Hessian matrix (and higher order derivatives)","Body":"<p>I would like to be able to compute higher order derivatives for my loss function. At the very least I would like to be able to compute the Hessian matrix. At the moment I am computing a numerical approximation to the Hessian but this is more expensive, and more importantly, as far as I understand, inaccurate if the matrix is ill-conditioned (with very large condition number).<\/p>\n\n<p>Theano implements this through symbolic looping, see <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html\" rel=\"nofollow\">here<\/a>, but Tensorflow does not seem to support symbolic control flow yet, see <a href=\"https:\/\/github.com\/samjabrahams\/tensorflow-white-paper-notes\/commit\/156736904eb48a0a44de4c23b227ae938e37e654\" rel=\"nofollow\">here<\/a>. A similar issue has been raised on TF github page, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/675\" rel=\"nofollow\">here<\/a>, but it looks like nobody has followed up on the issue for a while.<\/p>\n\n<p>Is anyone aware of more recent developments or ways to compute higher order derivatives (symbolically) in TensorFlow?<\/p>\n","answers":[{"AnswerId":"37666032","CreationDate":"2016-06-06T20:10:10.720","ParentId":null,"OwnerUserId":"2452184","Title":null,"Body":"<p>Well, you can , with little effort, compute the hessian matrix!<\/p>\n\n<p>Suppose you have two variables : <\/p>\n\n<pre><code>x = tf.Variable(np.random.random_sample(), dtype=tf.float32)\ny = tf.Variable(np.random.random_sample(), dtype=tf.float32)\n<\/code><\/pre>\n\n<p>and a function defined using these 2 variables:<\/p>\n\n<pre><code>f = tf.pow(x, cons(2)) + cons(2) * x * y + cons(3) * tf.pow(y, cons(2)) + cons(4) * x + cons(5) * y + cons(6)\n<\/code><\/pre>\n\n<p>where:<\/p>\n\n<pre><code>def cons(x):\n    return tf.constant(x, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>So in algebraic terms, this function is <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VdkRN.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VdkRN.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now we define a method that compute the hessian:<\/p>\n\n<pre><code>def compute_hessian(fn, vars):\n    mat = []\n    for v1 in vars:\n        temp = []\n        for v2 in vars:\n            # computing derivative twice, first w.r.t v2 and then w.r.t v1\n            temp.append(tf.gradients(tf.gradients(f, v2)[0], v1)[0])\n        temp = [cons(0) if t == None else t for t in temp] # tensorflow returns None when there is no gradient, so we replace None with 0\n        temp = tf.pack(temp)\n        mat.append(temp)\n    mat = tf.pack(mat)\n    return mat\n<\/code><\/pre>\n\n<p>and call it with:<\/p>\n\n<pre><code># arg1: our defined function, arg2: list of tf variables associated with the function\nhessian = compute_hessian(f, [x, y])\n<\/code><\/pre>\n\n<p>Now we grab a tensorflow session, initialize the variables, and run <code>hessian<\/code>  :<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.initialize_all_variables())\nprint sess.run(hessian)\n<\/code><\/pre>\n\n<p>Note: Since the function we used is quadratic in nature (and we are differentiating twice), the hessian returned will have constant values irrespective of the variables. <\/p>\n\n<p>The output is : <\/p>\n\n<pre><code>[[ 2.  2.]\n[ 2.  6.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35266598,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-02-08T09:52:49.537","AcceptedAnswerId":35283932.0,"OwnerUserId":5407700.0,"Title":"Torch - Query matrix with another matrix","Body":"<p>I have a m x n tensor (Tensor 1) and another k x 2 tensor (Tensor 2) and I wish to extract all the values of Tensor 1 using indices based on Tensor 2. For example;<\/p>\n\n<pre><\/pre>\n\n<p>And the function would yield;<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35283932","CreationDate":"2016-02-09T04:31:23.503","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>The first solution that comes into mind is to simply loop through indexes and pick the correspoding values:   <\/p>\n\n<pre><code>function get_elems_simple(tensor, indices)\n    local res = torch.Tensor(indices:size(1)):typeAs(tensor)\n    local i = 0\n    res:apply(\n        function () \n            i = i + 1\n            return tensor[indices[i]:clone():storage()] \n        end)\n    return res\nend\n<\/code><\/pre>\n\n<p>Here <code>tensor[indices[i]:clone():storage()]<\/code> is just a generic way to pick an element from a multi-dimensional tensor. In k-dimensional case this is exactly analogous to <code>tensor[{indices[i][1], ... , indices[i][k]}]<\/code>.<\/p>\n\n<p>This method works fine if you don't have to extract lots of values (the bottleneck is <code>:apply<\/code> method which is not able to use many optimization techniques and SIMD instructions because the function it executes is a black box). The job can be done way more efficiently: the method <code>:index<\/code> does exactly what you need... with a one-dimensional tensor. Multi-dimensional target\/index tensors need to be flattened:<\/p>\n\n<pre><code>function flatten_indices(sp_indices, shape)\n    sp_indices = sp_indices - 1\n    local n_elem, n_dim = sp_indices:size(1), sp_indices:size(2)\n    local flat_ind = torch.LongTensor(n_elem):fill(1)\n\n    local mult = 1\n    for d = n_dim, 1, -1 do\n        flat_ind:add(sp_indices[{{}, d}] * mult)\n        mult = mult * shape[d]\n    end\n    return flat_ind\nend\n\nfunction get_elems_efficient(tensor, sp_indices)\n    local flat_indices = flatten_indices(sp_indices, tensor:size()) \n    local flat_tensor = tensor:view(-1)\n    return flat_tensor:index(1, flat_indices)\nend\n<\/code><\/pre>\n\n<p>The difference is drastic:<\/p>\n\n<pre><code>n = 500000\nk = 100\na = torch.rand(n, k)\nind = torch.LongTensor(n, 2)\nind[{{}, 1}]:random(1, n)\nind[{{}, 2}]:random(1, k)\n\nelems1 = get_elems_simple(a, ind)      # 4.53 sec\nelems2 = get_elems_efficient(a, ind)   # 0.05 sec\n\nprint(torch.all(elems1:eq(elems2)))    # true\n<\/code><\/pre>\n"}]}
{"QuestionId":35268167,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><tensorflow><image-recognition><object-detection>","CreationDate":"2016-02-08T11:09:06.533","AcceptedAnswerId":null,"OwnerUserId":829624.0,"Title":"Low accuracy with change to TensorFlow Cifar10 example","Body":"<p>I am trying to modify the network structure provided by Cifar10 in TensorFlow. Typically, I added another convolution layer (conv12) after the first convolution layer (conv1). No matter how I set the filter (I tried all 1x1, 3x3, 5x5) and whether using weight decay or not, having a new layer will decrease the accuracy to below than 10%. This is equivalent to a random guess in Cifar10 since there are 10 classes.<\/p>\n\n<p>The code structure is as following, I don't modify any other part of the cifar except setting the size of input image to be 48x48 (instead of 24x24). I guess the input size should not matter.<\/p>\n\n<p>Note that the conv12 is a depthwise convolution layer because I want to add just a linear layer after the conv1 layer in order to minimize the change to the original code. Doing that I expected that the accuracy should be similar to the original version, but it decreases to around 10%. (I also tried a normal convolution layer but it didn't work also.)<\/p>\n\n<pre><\/pre>\n\n<p>Could someone please tell me what wrong with the code?<\/p>\n","answers":[{"AnswerId":"35284165","CreationDate":"2016-02-09T04:53:55.987","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Your second convolution:<\/p>\n\n<pre><code>kernel12 = _variable_with_weight_decay('weights', shape=[1, 1, 64, 1]\n<\/code><\/pre>\n\n<p>is taking the depth-64 output of the previous layer and squeezing it down to a depth-1 output.  That doesn't seem like it will match with whichever code you have following this (if it's <code>conv2<\/code> from the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow\">cifar example from TensorFlow<\/a>, then it definitely isn't going to work well, because that one expects a depth-64 input.)<\/p>\n\n<p>Perhaps you really wanted <code>shape=[1, 1, 64, 64]<\/code>, which would simply add an extra \"inception-style\" 1x1 convolutional layer into your model?<\/p>\n"}]}
{"QuestionId":35270450,"AnswerCount":1,"Tags":"<android><ios><mobile><windows-phone-8.1><tensorflow>","CreationDate":"2016-02-08T13:08:24.070","AcceptedAnswerId":35273806.0,"OwnerUserId":5898518.0,"Title":"TensorFlow on Mobile Devices (Android, iOS, Windows Phone)","Body":"<p>I am currently looking on different deep learning frameworks out there, specifically to train and deploy convolutional neural networks. The requirements are, that it can be trained on a normal PC with a GPU, but then the trained model has to be deployed on the three main mobile operating systems, namely Android, iOS, and Windows Phone.<\/p>\n\n<p>TensorFlow caught my eye, because of its simplicity and great python interface. There is an example application for Android (<a href=\"https:\/\/jalammar.github.io\/Supercharging-android-apps-using-tensorflow\/\" rel=\"nofollow\">https:\/\/jalammar.github.io\/Supercharging-android-apps-using-tensorflow\/<\/a>), but I am not sure if it can be also deployed on iOS and Windows Phone? If not, can you recommend an alternative framework, which would meet these requirements? Ideally with a simple scripting interface for fast prototyping?<\/p>\n\n<p>Thank you very much for your answers!<\/p>\n\n<p><strong>EDIT:<\/strong> Currently I'm testing <a href=\"https:\/\/github.com\/Microsoft\/CNTK\" rel=\"nofollow\">Microsoft's CNTK<\/a>. Building on Windows and Linux from source works perfectly, it can be extended in a \"Lego blocks\" fashion, and the proprietary NDL (Network Description Language) is really easy to read and learn, and provides enough freedom to build a lot of different Neural Network architectures. The execution engine is only a small part of the framework, and it can read in the NN model defined by the NDL, as well as the trained parameters. I will keep the post updated, on how the porting process to ARM processors goes.<\/p>\n","answers":[{"AnswerId":"35273806","CreationDate":"2016-02-08T15:56:19.510","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow currently doesn't support iOS or Windows. Here are the open github issues tracking them :<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16\" rel=\"nofollow\">iOS support<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow\">Windows support<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":35270849,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-02-08T13:29:42.237","AcceptedAnswerId":35291223.0,"OwnerUserId":2147347.0,"Title":"Value Error: Dimension Mismatch .., in Theano","Body":"<p>I just followed convolutional neural network tutorial and try to re-write the  code from:\n<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/convolutional_mlp.py\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/convolutional_mlp.py<\/a><\/p>\n\n<p>and I try to change the code from line 108, from:<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>and then I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know why is this happen, because I think  and  have the same shape in the second dimension<\/p>\n\n<p>can someone help me to explain why is this happen?<\/p>\n","answers":[{"AnswerId":"35291223","CreationDate":"2016-02-09T11:44:53.883","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The problem is not inside the <code>LeNetConvPoolLayer<\/code> it's in the following layer which is expecting the output from the <code>LeNetConvPoolLayer<\/code> to be one size but is getting something else.<\/p>\n\n<p>To skip the pooling, you'll need to adjust every layer that follows a <code>LeNetConvPoolLayer<\/code> to expect its input to be the appropriately larger size.<\/p>\n"}]}
{"QuestionId":35274405,"AnswerCount":1,"Tags":"<python><multithreading><queue><tensorflow>","CreationDate":"2016-02-08T16:25:14.873","AcceptedAnswerId":35274785.0,"OwnerUserId":2206976.0,"Title":"Tensorflow Training using input queue gets stuck","Body":"<p>I am trying to build a NN training similar to the one in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#creating-threads-to-prefetch-using-queuerunner-objects\" rel=\"nofollow\">this<\/a> tutorial.<\/p>\n\n<p>My code looks as follows:<\/p>\n\n<pre><\/pre>\n\n<p>with<\/p>\n\n<pre><\/pre>\n\n<p>The tutorial says <\/p>\n\n<blockquote>\n  <p>These require that you call  before running any training or inference steps, or it will hang forever.<\/p>\n<\/blockquote>\n\n<p>. I am calling , but the execution of  still gets stuck at the first occurrence of .<\/p>\n\n<p>Does someone have an idea what I am doing wrong?<\/p>\n","answers":[{"AnswerId":"35274785","CreationDate":"2016-02-08T16:45:01.893","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You are redefining your network every time you try to run the training loop.<\/p>\n\n<p>Remember that TensorFlow defines an execution graph, and then executes it.  You want to call your train_op() outside of the run loop, and you need to define that graph BEFORE you call <code>initialize_all_variables<\/code> and <code>tf.train.start_queue_runners<\/code><\/p>\n"}]}
{"QuestionId":35274457,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-02-08T16:27:43.360","AcceptedAnswerId":35274630.0,"OwnerUserId":1259390.0,"Title":"InceptionV3 and transfer learning with tensorflow","Body":"<p>I would like to do a transfer learning from the given inceptionV3 in tensorflow example. Following the classify image example and the operator and tensor names given here <a href=\"https:\/\/github.com\/AKSHAYUBHAT\/VisualSearchServer\/blob\/master\/notebooks\/notebook_network.ipynb\" rel=\"nofollow\">https:\/\/github.com\/AKSHAYUBHAT\/VisualSearchServer\/blob\/master\/notebooks\/notebook_network.ipynb<\/a> I can create my graph. But when, I put a batch of images of size (100, 299, 299, 3) in the pre-computed inception graph, I get the following shape error at the pool_3 layer :<\/p>\n\n<pre><\/pre>\n\n<p>It seems that this inceptionV3 graph doesn't accept image batch as input. am I wrong ?<\/p>\n","answers":[{"AnswerId":"35319329","CreationDate":"2016-02-10T15:32:07.303","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>Actually it works for transfer learning if you extract the right thing. There is no problem feeding a batch of images in the shape of <code>[N, 299, 299, 3]<\/code> as <code>ResizeBilinear:0<\/code> and then using the <code>pool_3:0<\/code> tensor. It's the reshaping afterwards that breaks, but you can reshape yourself (you'll have your own layers afterwards anyway). If you wanted to use the original classifier with a batch, you could add your own reshaping on top of <code>pool_3:0<\/code> and then add the softmax layer, reusing the weights\/biases tensors of the original softmax.<\/p>\n\n<p>TLDR: With double_img being a stack of two images with shape (2, 299, 299, 3) this works:<\/p>\n\n<pre><code>pooled_2 = sess.graph.get_tensor_by_name(\"pool_3:0\").eval(session=sess, feed_dict={'ResizeBilinear:0':double_img})\npooled_2.shape\n# =&gt; (2, 1, 1, 2048)\n<\/code><\/pre>\n"},{"AnswerId":"35274630","CreationDate":"2016-02-08T16:37:34.407","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You're not wrong.  This seems like a very reasonable feature request, so I've opened <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1021\" rel=\"nofollow\">a ticket for it on github<\/a>.  Follow that for updates. <\/p>\n"},{"AnswerId":"57650506","CreationDate":"2019-08-25T23:49:03.593","ParentId":null,"OwnerUserId":"5858902","Title":null,"Body":"<p>etarion made a very good point. However, we don't have to reshape it ourselves; instead, we could change the value of <code>shape<\/code> that <code>reshape<\/code> takes as input. I.e.,<\/p>\n\n<pre><code>input_tensor_name = 'import\/input:0'\nshape_tensor_name = 'import\/InceptionV3\/Predictions\/Shape:0'\noutput_tensor_name= 'import\/InceptionV3\/Predictions\/Reshape_1:0'\n\noutput_tensor = tf.import_graph_def(\n    graph.as_graph_def(),\n    input_map={input_tensor_name: image_batch,\n               shape_tensor_name: [batch_size, num_class]},\n    return_elements=[output_tensor_name])\n<\/code><\/pre>\n\n<p>These tensor names are based on <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception_v3_2016_08_28_frozen.pb.tar.gz\" rel=\"nofollow noreferrer\"><code>inception_v3_2016_08_28_frozen.pb<\/code><\/a>.<\/p>\n"},{"AnswerId":"45414736","CreationDate":"2017-07-31T11:29:26.773","ParentId":null,"OwnerUserId":"5015616","Title":null,"Body":"<p>Something like this should do it:<\/p>\n\n<pre><code>    with g.as_default():\n     inputs = tf.placeholder(tf.float32, shape=[batch_size, 299, 299, 3],\n                                name='input')\n\n        with slim.arg_scope(inception.inception_v3_arg_scope()):\n\n            logits, end_points = inception.inception_v3( inputs, \n            num_classes=FLAGS.num_classes, is_training=False)\n            variables_to_restore = lim.get_variables_to_restore(exclude=exclude)\n        sess = tf.Session()\n\n        saver = tf_saver.Saver(variables_to_restore)\n<\/code><\/pre>\n\n<p>Then you should be able to call the operation:<\/p>\n\n<pre><code>        sess.run(\"pool_3:0\",feed_dict={'ResizeBilinear:0':images})\n<\/code><\/pre>\n"}]}
{"QuestionId":35276944,"AnswerCount":0,"Tags":"<deep-learning><gensim><keras>","CreationDate":"2016-02-08T18:43:09.703","AcceptedAnswerId":null,"OwnerUserId":2802468.0,"Title":"Index of Embedding layer with zero padding in Keras","Body":"<p>I am building an RNN model in Keras for sentences with word embeddings from gensim. I am initializing the embedding layer with GloVe vectors. Since this is a sequential model and sentences have variable lengths, vectors are zero-padded. e.g.<\/p>\n\n<pre><\/pre>\n\n<p>Let's say the GloVe vectors have dimensions . The zero index is masked (ignored) so to get the proper indexing of words, do we add an extra column to the GloVe matrix so the dimensions are: ?<\/p>\n\n<p>Seems like there is an unnecessary vector that the model will estimate unless there is a more elegant way.<\/p>\n\n<pre><\/pre>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":35277339,"AnswerCount":2,"Tags":"<python><pip><tensorflow>","CreationDate":"2016-02-08T19:07:16.103","AcceptedAnswerId":null,"OwnerUserId":2154591.0,"Title":"Error when importing TensorFlow in Python on Windows","Body":"<p>I am running Python 2.7.10 and I successfully installed TensorFlow, using .<\/p>\n\n<p>When I run the  interpreter and attempt to import the  module, I get the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41018561","CreationDate":"2016-12-07T13:18:23.827","ParentId":null,"OwnerUserId":"2579263","Title":null,"Body":"<p>you can use pip to install tensorflow<\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-0.12.0rc0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>If you are trying to install on windows , you will have to have python3.5 x64 bit and along with that you need to install  Visual C++ 2015 redistributable (x64 version) to be able to import tensorflow<\/p>\n"},{"AnswerId":"35277461","CreationDate":"2016-02-08T19:14:16.683","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Judging by the paths in the output, it looks like you are trying to run TensorFlow on Windows. The package that you have installed is a Mac OS X binary distribution of TensorFlow.<\/p>\n\n<p>Unfortunately, Windows is <a href=\"https:\/\/stackoverflow.com\/a\/33618580\/3574081\">not currently supported by TensorFlow<\/a>, and it is unlikely to be supported before its build system (Bazel) adds Windows support. There is some discussion of progress towards a Windows version on this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow noreferrer\">GitHub issue<\/a>, and an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/728\" rel=\"nofollow noreferrer\">unsupported contribution that builds TensorFlow using CMake on Windows<\/a>.<\/p>\n"}]}
{"QuestionId":35277898,"AnswerCount":2,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-02-08T19:40:36.193","AcceptedAnswerId":35279095.0,"OwnerUserId":2680707.0,"Title":"TensorFlow for binary classification","Body":"<p>I am trying to adapt <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"noreferrer\">this MNIST example<\/a> to binary classification.<\/p>\n\n<p>But when changing my  from  to , the loss function always returns 0 (and accuracy 1).<\/p>\n\n<pre><\/pre>\n\n<p>I have checked the dimensions of both  (fed into ) and  and they are both 1xN matrices when  so the problem seems to be prior to that. Maybe something to do with the matrix multiplication?<\/p>\n\n<p>I actually have got this same problem in a real project, so any help would be appreciated... Thanks!<\/p>\n","answers":[{"AnswerId":"57385551","CreationDate":"2019-08-07T00:34:01.437","ParentId":null,"OwnerUserId":"6410450","Title":null,"Body":"<p>I've been looking for good examples of how to implement binary classification in TensorFlow in a similar manner to the way it would be done in Keras.  I didn't find any, but after digging through the code a bit, I think I have it figured out.  I modified the problem here to implement a solution that uses sigmoid_cross_entropy_with_logits the way Keras does under the hood.  <\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nfrom tensorflow.examples.tutorials.mnist import input_data\nimport tensorflow as tf\n\n# Import data\nmnist = input_data.read_data_sets('data', one_hot=True)\nNLABELS = 1\n\nsess = tf.InteractiveSession()\n\n# Create the model\nx = tf.placeholder(tf.float32, [None, 784], name='x-input')\nW = tf.get_variable('weights', [784, NLABELS],\n                    initializer=tf.truncated_normal_initializer()) * 0.1\nb = tf.Variable(tf.zeros([NLABELS], name='bias'))\nlogits = tf.matmul(x, W) + b\n\n# Define loss and optimizer\ny_ = tf.placeholder(tf.float32, [None, NLABELS], name='y-input')\n\n# More name scopes will clean up the graph representation\nwith tf.name_scope('cross_entropy'):\n\n    #manual calculation : under the hood math, don't use this it will have gradient problems\n    # entropy = tf.multiply(tf.log(tf.sigmoid(logits)), y_) + tf.multiply((1 - y_), tf.log(1 - tf.sigmoid(logits)))\n    # loss = -tf.reduce_mean(entropy, name='loss')\n\n    entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=y_, logits=logits)\n    loss = tf.reduce_mean(entropy, name='loss')\n\nwith tf.name_scope('train'):\n    # Using Adam instead\n    # train_step = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss)\n    train_step = tf.train.AdamOptimizer(learning_rate=0.002).minimize(loss)\n\nwith tf.name_scope('test'):\n    preds = tf.cast((logits &gt; 0.5), tf.float32)\n    correct_prediction = tf.equal(preds, y_)\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\ntf.initialize_all_variables().run()\n\n# Train the model, and feed in test data and record summaries every 10 steps\n\nfor i in range(2000):\n    if i % 100 == 0:  # Record summary data and the accuracy\n        labels = mnist.test.labels[:, 0:NLABELS]\n        feed = {x: mnist.test.images, y_: labels}\n        result = sess.run([loss, accuracy], feed_dict=feed)\n        print('Accuracy at step %s: %s - loss: %f' % (i, result[1], result[0]))\n    else:\n        batch_xs, batch_ys = mnist.train.next_batch(100)\n        batch_ys = batch_ys[:, 0:NLABELS]\n        feed = {x: batch_xs, y_: batch_ys}\n    sess.run(train_step, feed_dict=feed)\n<\/code><\/pre>\n\n<p>Training:<\/p>\n\n<pre><code>Accuracy at step 0: 0.7373 - loss: 0.758670\nAccuracy at step 100: 0.9017 - loss: 0.423321\nAccuracy at step 200: 0.9031 - loss: 0.322541\nAccuracy at step 300: 0.9085 - loss: 0.255705\nAccuracy at step 400: 0.9188 - loss: 0.209892\nAccuracy at step 500: 0.9308 - loss: 0.178372\nAccuracy at step 600: 0.9453 - loss: 0.155927\nAccuracy at step 700: 0.9507 - loss: 0.139031\nAccuracy at step 800: 0.9556 - loss: 0.125855\nAccuracy at step 900: 0.9607 - loss: 0.115340\nAccuracy at step 1000: 0.9633 - loss: 0.106709\nAccuracy at step 1100: 0.9667 - loss: 0.099286\nAccuracy at step 1200: 0.971 - loss: 0.093048\nAccuracy at step 1300: 0.9714 - loss: 0.087915\nAccuracy at step 1400: 0.9745 - loss: 0.083300\nAccuracy at step 1500: 0.9745 - loss: 0.079019\nAccuracy at step 1600: 0.9761 - loss: 0.075164\nAccuracy at step 1700: 0.9768 - loss: 0.071803\nAccuracy at step 1800: 0.9777 - loss: 0.068825\nAccuracy at step 1900: 0.9788 - loss: 0.066270\n<\/code><\/pre>\n"},{"AnswerId":"35279095","CreationDate":"2016-02-08T20:53:13.573","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The original MNIST example uses a <a href=\"https:\/\/en.wikipedia.org\/wiki\/One-hot\" rel=\"nofollow noreferrer\">one-hot encoding<\/a> to represent the labels in the data: this means that if there are <code>NLABELS = 10<\/code> classes (as in MNIST), the target output is <code>[1 0 0 0 0 0 0 0 0 0]<\/code> for class 0, <code>[0 1 0 0 0 0 0 0 0 0]<\/code> for class 1, etc. The <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/nn.html#softmax\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax()<\/code><\/a> operator converts the logits computed by <code>tf.matmul(x, W) + b<\/code> into a probability distribution across the different output classes, which is then compared to the fed-in value for <code>y_<\/code>.<\/p>\n\n<p>If <code>NLABELS = 1<\/code>, this acts as if there were only a single class, and the <code>tf.nn.softmax()<\/code> op would compute a probability of <code>1.0<\/code> for that class, leading to a cross-entropy of <code>0.0<\/code>, since <code>tf.log(1.0)<\/code> is <code>0.0<\/code> for all of the examples.<\/p>\n\n<p>There are (at least) two approaches you could try for binary classification:<\/p>\n\n<ol>\n<li><p>The simplest would be to set <code>NLABELS = 2<\/code> for the two possible classes, and encode your training data as <code>[1 0]<\/code> for label 0 and <code>[0 1]<\/code> for label 1. <a href=\"https:\/\/stackoverflow.com\/a\/35205193\/3574081\">This answer<\/a> has a suggestion for how to do that.<\/p><\/li>\n<li><p>You could keep the labels as integers <code>0<\/code> and <code>1<\/code> and use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a>, as suggested in <a href=\"https:\/\/stackoverflow.com\/a\/35227732\/3574081\">this answer<\/a>.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35281225,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-08T23:22:52.903","AcceptedAnswerId":null,"OwnerUserId":1441121.0,"Title":"How can we replace a single element in an n-D Tensor in TensorFlow?","Body":"<p>If we have a 1-D Tensor, we can replace a single element with a 0-D Tensor by using  or by unpacking, replacing, packing.<\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n\n<p> will now produce .<\/p>\n\n<p>Is it possible to do this with n-D arrays?<\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n\n<p>I think I could get the desired result with a mixture of , , and , but it'd be verbose, and we'd be replacing a (potentially large) 3-D Tensor instead of just replacing a tiny 0-D Tensor. Is there a better way?<\/p>\n","answers":[{"AnswerId":"35304665","CreationDate":"2016-02-10T00:13:27.580","ParentId":null,"OwnerUserId":"5875585","Title":null,"Body":"<p>If you want to do this efficiently, I believe you will need to write a new Op or generalize scatter_update.<\/p>\n"}]}
{"QuestionId":35282146,"AnswerCount":2,"Tags":"<neural-network><theano><conv-neural-network><lasagne><bias-neuron>","CreationDate":"2016-02-09T00:54:42.710","AcceptedAnswerId":null,"OwnerUserId":5900989.0,"Title":"get_all_param_values() how to read lasagne.layer","Body":"<p>I am running Lasagne and Theano to create my Convolutional Neural Network. I currently consist of <\/p>\n\n<pre><\/pre>\n\n<p>My last layer is a denselayer which uses a softmax to output my classification. My ultimate goal is to retrieve the probability and not the classification (0 or 1). <\/p>\n\n<p>When I call get_all_param_values(), it provides me an extensive array. I only want the weights and bias for the last dense layer. How do you go about this? I have tried l_out.W and l_out.b  and get_values(). <\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"35289531","CreationDate":"2016-02-09T10:25:06.210","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>You can get the parameters for a single layer using <code>get_params<\/code>. This is explained in the <a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/modules\/layers\/base.html#lasagne.layers.Layer.get_params\" rel=\"nofollow\">documentation<\/a>.<\/p>\n"},{"AnswerId":"35380974","CreationDate":"2016-02-13T14:14:31.533","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>I modified your code because what you pasted references an l_in, but you don't include an l_in in your code. I defined the following network: <\/p>\n\n<pre><code>l_shape = lasagne.layers.InputLayer(shape = (None, 3, 130, 130))\nl_conv1 = lasagne.layers.Conv2DLayer(l_shape, num_filters=32, filter_size=3, pad=1)\nl_conv1_1 = lasagne.layers.Conv2DLayer(l_conv1, num_filters=32, filter_size=3, pad=1)\nl_pool1 = lasagne.layers.MaxPool2DLayer(l_conv1_1, 2)\nl_conv2 = lasagne.layers.Conv2DLayer(l_pool1, num_filters=64, filter_size=3, pad=1)\nl_conv2_2 = lasagne.layers.Conv2DLayer(l_conv2, num_filters=64, filter_size=3, pad=1)\nl_pool2 = lasagne.layers.MaxPool2DLayer(l_conv2_2, 2)\nl_conv3 = lasagne.layers.Conv2DLayer(l_pool2, num_filters=64, filter_size=3, pad=1)\nl_conv3_2 = lasagne.layers.Conv2DLayer(l_conv3, num_filters=64, filter_size=3, pad=1)\nl_pool3 = lasagne.layers.MaxPool2DLayer(l_conv3_2, 2)\nl_conv4 = lasagne.layers.Conv2DLayer(l_pool3, num_filters=64, filter_size=3, pad=1)\nl_conv4_2 = lasagne.layers.Conv2DLayer(l_conv4, num_filters=64, filter_size=3, pad=1)\nl_pool4 = lasagne.layers.MaxPool2DLayer(l_conv4_2, 2)\nl_conv5 = lasagne.layers.Conv2DLayer(l_pool4, num_filters=64, filter_size=3, pad=1)\nl_conv5_2 = lasagne.layers.Conv2DLayer(l_conv5, num_filters=64, filter_size=3, pad=1)\nl_pool5 = lasagne.layers.MaxPool2DLayer(l_conv5_2, 2)\nl_out = lasagne.layers.DenseLayer(l_pool5, num_units=2, nonlinearity=lasagne.nonlinearities.softmax)\n<\/code><\/pre>\n\n<p>Just to implement Daniel Renshaw's answer:<\/p>\n\n<pre><code>params = l_out.get_params()\nW = params[0].get_value()\n<\/code><\/pre>\n\n<p>When you print params, you will see all the parameters for l_out:<\/p>\n\n<pre><code>[W, b] \n<\/code><\/pre>\n\n<p>So each element of params, params[0] and params[1] is a Theano shared variable and you can get the numerical values by params[i].get_value().<\/p>\n"}]}
{"QuestionId":35285015,"AnswerCount":1,"Tags":"<tensorflow><attributeerror><mnist>","CreationDate":"2016-02-09T06:06:31.653","AcceptedAnswerId":null,"OwnerUserId":5871027.0,"Title":"import input_data MNIST tensorflow gives NoneType object (Attribute error)","Body":"<p>I am not able to import the mnist dataset properly. Can you please help me figuring out whats going wrong. \"input_data.py\" is correctly placed and called.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35353250","CreationDate":"2016-02-12T01:26:28.520","ParentId":null,"OwnerUserId":"5916397","Title":null,"Body":"<p>The method read_data_sets first downloads the data from <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a> and then extract it       <\/p>\n\n<p>local_file = maybe_download(TRAIN_IMAGES, train_dir)<br>\ntrain_images = extract_images(local_file)<\/p>\n\n<p>which for you is working fine. But after that the object which returns collection of DataSet is null. Since its working fine for me and I can not reproduce the error can you run the calls in method and provide where it is failing. Something like this...<\/p>\n\n<pre><code>&gt;&gt;&gt; local_file = input_data.maybe_download('train-labels-idx1-ubyte.gz', 'MNIST_data\/')\nSuccessfully downloaded train-labels-idx1-ubyte.gz 28881 bytes.\n&gt;&gt;&gt; train_labels = input_data.extract_labels(local_file, one_hot=True)\nExtracting MNIST_data\/train-labels-idx1-ubyte.gz\n&gt;&gt;&gt; local_file = input_data.maybe_download('train-images-idx3-ubyte.gz', 'MNIST_data\/')\n&gt;&gt;&gt; train_images = input_data.extract_images(local_file)\nExtracting MNIST_data\/train-images-idx3-ubyte.gz\n&gt;&gt;&gt; local_file = input_data.maybe_download('t10k-images-idx3-ubyte.gz', 'MNIST_data\/')\n&gt;&gt;&gt; test_images = input_data.extract_images(local_file)\nExtracting MNIST_data\/t10k-images-idx3-ubyte.gz\n&gt;&gt;&gt; local_file = input_data.maybe_download('t10k-labels-idx1-ubyte.gz', 'MNIST_data\/')\n&gt;&gt;&gt; test_labels = input_data.extract_labels(local_file,one_hot=True)\nExtracting MNIST_data\/t10k-labels-idx1-ubyte.gz\n&gt;&gt;&gt; VALIDATION_SIZE = 5000\n&gt;&gt;&gt; validation_images = train_images[:VALIDATION_SIZE]\n&gt;&gt;&gt; validation_labels = train_labels[:VALIDATION_SIZE]\n&gt;&gt;&gt; train_images = train_images[VALIDATION_SIZE:]\n&gt;&gt;&gt; train_labels = train_labels[VALIDATION_SIZE:]\n\n&gt;&gt;&gt; dtype = 'float32'\n&gt;&gt;&gt; data_set_train = input_data.DataSet(train_images, train_labels, dtype=dtype)\n&gt;&gt;&gt; data_set_validation = input_data.DataSet(validation_images, validation_labels, dtype=dtype)\n&gt;&gt;&gt; data_set_test = input_data.DataSet(test_images, test_labels, dtype=dtype)  \n&gt;&gt;&gt; trX = data_set_train.images\n&gt;&gt;&gt; print(data_set_train)\n&lt;tensorflow.examples.tutorials.mnist.input_data.DataSet object at 0x10508ff98&gt;\n&gt;&gt;&gt; print(trX)\n[[ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n ..., \n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35286989,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-09T08:17:44.937","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Sharing Queue between two graphs in tensorflow","Body":"<p>Is it possible to share a queue between two graphs in TensorFlow? I'd like to do a kind of bootstrapping to select \"hard negative\" examples during training.<\/p>\n\n<p>To speed up the process, I want separate threads for hard negative example selection, and for the training process. The hard negative selection is based on the evaluation of the current model, and it will load its graph from a checkpoint file. The training graph is run on another thread and writes the checkpoint file. The two graphs should share the same queue: the training graph will consume examples and the hard negative selection will produce them.<\/p>\n","answers":[{"AnswerId":"35297009","CreationDate":"2016-02-09T16:23:05.580","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Currently there's no support for sharing state between different graphs in the open-source version of TensorFlow: each graph runs in a separate session, and each session uses an isolated set of devices.<\/p>\n\n<p><strong>However<\/strong>, it seems like it would be possible to achieve your goal using a queue in single graph. Simply construct a queue (using e.g. <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/io_ops.html#FIFOQueue\" rel=\"noreferrer\"><code>tf.FIFOQueue<\/code><\/a>) and use <code>tf.import_graph_def()<\/code> to import the graph from the checkpoint file into the current graph. Using the <code>return_elements<\/code> argument to <code>tf.import_graph_def()<\/code> you can specify the name of the tensor that will contain the negative examples, and then add a <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/io_ops.html#QueueBase.enqueue_many\" rel=\"noreferrer\"><code>q.enqueue_many()<\/code><\/a> operation to add them to your queue. You would then fork a thread to run the <code>enqueue_many<\/code> operation in a loop. In your training graph, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\" rel=\"noreferrer\"><code>q.dequeue_many()<\/code><\/a> to get a batch of negative examples, and use them as the input to your training process.<\/p>\n"}]}
{"QuestionId":35289773,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-09T10:35:55.677","AcceptedAnswerId":35289851.0,"OwnerUserId":1851402.0,"Title":"Cannot convert a partially converted tensor in TensorFlow","Body":"<p>There are many methods in TensorFlow that requires specifying a shape, for example truncated_normal:<\/p>\n\n<pre><\/pre>\n\n<p>I have a placeholder for the input of shape [None, 784], where the first dimension is None because the batch size can vary. I could use a fixed batch size but it still would be different from the test\/validation set size.<\/p>\n\n<p>I cannot feed this placeholder to tf.truncated_normal because it requires a fully specified tensor shape. What is a simple way to having tf.truncated_normal accept different tensor shapes?<\/p>\n","answers":[{"AnswerId":"35289851","CreationDate":"2016-02-09T10:39:50.497","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>You just need to feed it in as a single example but in the batched shape. So that means adding an extra dimension to the shape e.g.<\/p>\n\n<pre><code>batch_size = 32 # set this to the actual size of your batch\ntf.truncated_normal((batch_size, 784), mean=0.0, stddev=1.0, dtype=tf.float32, seed=None, name=None)\n<\/code><\/pre>\n\n<p>This way it will \"fit\" into the placeholder.<\/p>\n\n<p>If you expect batch_size to change you can also use:<\/p>\n\n<pre><code>tf.truncated_normal(tf.shape(input_tensor), mean=0.0, stddev=1.0, dtype=tf.float32, seed=None, name=None)\n<\/code><\/pre>\n\n<p>Where input_tensor could be a placeholder or just whatever tensor is going to have this noise added to it.<\/p>\n"}]}
{"QuestionId":35290945,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-09T11:31:01.703","AcceptedAnswerId":null,"OwnerUserId":3559888.0,"Title":"Compiling tensorflow from source: Undefined reference to tensorflow::functor::FillFunctor","Body":"<p>When compiling tensorflow from source with GPU support I get:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried different bazel versions (0.1.1 and 0.1.4) but I get the same error. Any ideas?<\/p>\n","answers":[{"AnswerId":"35685183","CreationDate":"2016-02-28T16:34:57.860","ParentId":null,"OwnerUserId":"4925151","Title":null,"Body":"<p>It is most likely due to having an old version of the c compiler. By upgrading your Ubuntu, you've also also upgraded the gcc, which allows TensorFlow to build with bazel. <\/p>\n\n<p>For those who cannot upgrade the compiler, you can try with a local gcc, see here:\n<a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/649\" rel=\"nofollow\">https:\/\/github.com\/bazelbuild\/bazel\/issues\/649<\/a><\/p>\n"},{"AnswerId":"35314518","CreationDate":"2016-02-10T11:53:50.733","ParentId":null,"OwnerUserId":"3559888","Title":null,"Body":"<p>I am not sure what caused the problem, but upgrading from Ubuntu 12.04 to 14.04 removed it.<\/p>\n"}]}
{"QuestionId":35292311,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><tensorboard>","CreationDate":"2016-02-09T12:38:11.940","AcceptedAnswerId":null,"OwnerUserId":4618605.0,"Title":"ImportError: cannot import name tensorboard_server","Body":"<p>since I pulled the latest updates of the tensorflow repository, I figured out that the tensorboard is no longer functional. I installed from source using bazel-0.1.4. Afterwards, I built the tensorboard as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Then, when I'm running it with:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>What's wrong? Can someone help me?<\/p>\n","answers":[{"AnswerId":"35295754","CreationDate":"2016-02-09T15:25:19.310","ParentId":null,"OwnerUserId":"3622526","Title":null,"Body":"<p>It looks like a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/60765a4b210fc54ec32a7729f8532a60c5217544\" rel=\"nofollow\">recent commit<\/a> added the new module <code>tensorboard_server<\/code> without listing it as a dependency of tensorboard.<\/p>\n\n<p>I modified tensorflow\/tensorboard\/BUILD by adding the following section:<\/p>\n\n<pre><code>py_library(\n    name = \"tensorboard_server\",\n    srcs = [\"backend\/tensorboard_server.py\"],\n    deps = [\n        \":float_wrapper\",\n        \"\/\/tensorflow\/python:platform\",\n        \"\/\/tensorflow\/python:summary\",\n    ],\n    srcs_version = \"PY2AND3\",\n)\n<\/code><\/pre>\n\n<p>and added <code>\":tensorboard_server\"<\/code> to the dependencies of the tensorboard binary.<\/p>\n\n<p>Following this, I built and ran tensorboard as you specified, and it appears to work:<\/p>\n\n<pre><code>$ bazel build tensorflow\/tensorboard:tensorboard\nINFO: Found 1 target...\nTarget \/\/tensorflow\/tensorboard:tensorboard up-to-date:\n  bazel-bin\/tensorflow\/tensorboard\/tensorboard\nINFO: Elapsed time: 0.425s, Critical Path: 0.10s\n$ bazel-bin\/tensorflow\/tensorboard\/tensorboard --logdir .\nStarting TensorBoard 13 on port 6006\n(You can navigate to http:\/\/0.0.0.0:6006)\n<\/code><\/pre>\n"}]}
{"QuestionId":35294057,"AnswerCount":2,"Tags":"<tensorflow><bazel>","CreationDate":"2016-02-09T14:04:40.277","AcceptedAnswerId":null,"OwnerUserId":1002035.0,"Title":"How do I check Bazel version?","Body":"<p>I am trying to find out which version of Bazel I currently have on my computer to install TensorFlow from source (it requires version 0.1.4)<\/p>\n\n<pre><\/pre>\n\n<p>So where is the version actually?<\/p>\n","answers":[{"AnswerId":"35294802","CreationDate":"2016-02-09T14:39:45.257","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>See <a href=\"https:\/\/docs.bazel.build\/versions\/master\/user-manual.html\" rel=\"noreferrer\">Bazel users manual<\/a><\/p>\n\n<p>From the command line:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>$ bazel version \nBuild label: 0.1.1\n<\/code><\/pre>\n"},{"AnswerId":"57555765","CreationDate":"2019-08-19T11:23:50.753","ParentId":null,"OwnerUserId":"606897","Title":null,"Body":"<p>Since bazel version <code>0.27.0<\/code> there is <code>bazel --version<\/code> available, without the side-effect of starting the server daemon.<\/p>\n"}]}
{"QuestionId":35295191,"AnswerCount":2,"Tags":"<python><python-2.7><machine-learning><tensorflow><word-embedding>","CreationDate":"2016-02-09T14:57:35.167","AcceptedAnswerId":35296384.0,"OwnerUserId":4190832.0,"Title":"Tensorflow embedding_lookup","Body":"<p>I am trying to learn the word representation of the imdb dataset \"from scratch\" through the TensorFlow  function. If I understand it correctly, I have to set up an embedding layer before the other hidden layer, and then when I perform gradient descent, the layer will \"learn\" a word representation in the weights of this layer. However, when I try to do this, I get a shape error between my embedding layer and the first fully-connected layer of my network.<\/p>\n\n<pre><\/pre>\n\n<p>The error I get is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35296384","CreationDate":"2016-02-09T15:54:13.353","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The shape error arises because you are using a two-dimensional tensor, <code>x<\/code> to index into a two-dimensional embedding tensor <code>W<\/code>. Think of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/embedding_lookup\" rel=\"noreferrer\"><code>tf.nn.embedding_lookup()<\/code><\/a> (and its close cousin <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"noreferrer\"><code>tf.gather()<\/code><\/a>) as taking each integer value <code>i<\/code> in <code>x<\/code> and replacing it with the row <code>W[i, :]<\/code>. From the error message, one can infer that <code>n_input = 300<\/code> and <code>embedding_size = 128<\/code>. In general, the result of <code>tf.nn.embedding_lookup()<\/code> number of dimensions equal to <code>rank(x) + rank(W) - 1<\/code>&hellip; in this case, 3. The error arises when you try to multiply this result by <code>_weights['h1']<\/code>, which is a (two-dimensional) matrix.<\/p>\n\n<p>To fix this code, it depends on what you're trying to do, and why you are passing in a matrix of inputs to the embedding. One common thing to do is to <em>aggregate<\/em> the embedding vectors for each input example into a single row per example using an operation like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_sum\" rel=\"noreferrer\"><code>tf.reduce_sum()<\/code><\/a>. For example, you might do the following:<\/p>\n\n<pre><code>W = tf.Variable(\n    tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0) ,name=\"W\")\nembedding_layer = tf.nn.embedding_lookup(W, _X)\n\n# Reduce along dimension 1 (`n_input`) to get a single vector (row)\n# per input example.\nembedding_aggregated = tf.reduce_sum(embedding_layer, [1])\n\nlayer_1 = tf.nn.sigmoid(tf.add(tf.matmul(\n    embedding_aggregated, _weights['h1']), _biases['b1'])) \n<\/code><\/pre>\n"},{"AnswerId":"39383880","CreationDate":"2016-09-08T06:25:13.873","ParentId":null,"OwnerUserId":"3898714","Title":null,"Body":"<p>One another possible solution is : Instead of adding the embedding vectors, concatenate these vectors into a single vector and increase the number of neurons in the hidden layer. <br>\nI used : <br>\n    <code>embedding_aggregated = tf.reshape(embedding_layer, [-1, embedding_size *     sequence_length])<\/code> <br>\nAlso, i changed the number of neurons in hidden layer to <code>embedding_size * sequence_length<\/code>. \nObservation : Accuracy also improved on using concatenation rather than addition.<\/p>\n"}]}
{"QuestionId":35297798,"AnswerCount":1,"Tags":"<python><arrays><numpy><theano>","CreationDate":"2016-02-09T17:00:39.157","AcceptedAnswerId":35298402.0,"OwnerUserId":3661815.0,"Title":"Outer products using theano\/numpy.tensordot","Body":"<p>I have a NumPy array of  vectors of length , so . I am trying to compute the outer products of each vector using  such that the  and :<\/p>\n\n<pre><\/pre>\n\n<p>I'm pretty sure this is possible, but am having a hard time figuring out what arguments to use for axes. This can be easily done iterating over the array, but I am trying to complete this step using tensordot as it is implemented the same in theano and numpy.  This is important because I am using theano, but using numpy to debug currently. <\/p>\n","answers":[{"AnswerId":"35298402","CreationDate":"2016-02-09T17:30:53.980","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>You are not exactly using the sum reduction part of a dot product and just broadcasting elementwise multiiplication for each row with it's transposed version. This could be achieved in a vectorized manner for all rows again with <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow\"><code>broadcasting<\/code><\/a> by creating <code>singleton<\/code> dimensions in two versions of the input array - One with <code>singleton dim<\/code> at <code>axis=1<\/code> and another at <code>axis=2<\/code>. Since, the input array is already 3D shaped, so we already have the <code>version #2<\/code> with us. To get the first version, just swap axes along the last two dims. Thus, multiplying these two versions would give us the desired output, like so -<\/p>\n\n<pre><code>output = input*input.swapaxes(1,2)\n<\/code><\/pre>\n"}]}
{"QuestionId":35298142,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-02-09T17:17:20.180","AcceptedAnswerId":null,"OwnerUserId":5904419.0,"Title":"How to add broadcastable tensor with given values to another tensor in Theano?","Body":"<pre><\/pre>\n\n<p>I'm trying to create a custom layer in Theano\/Lasagne, which shifts the image matrix for some random bias. I want this bias to be a constant for a certain image, but different for every another image in a batch.<\/p>\n\n<p>'input' is a tensor with shape ('x',3,32,32),where 'x' is broadcastable dimension, representing image batch count. The code above adds a ('x',3,32,32) matrix of random numbers to 'input'.\nHowever, I want the number to be the same for every image (bias[i,:,:,:]=const)\nand random only among different images (bias[i,:]!=bias[j,:])\nThe problem is that I cannot access directly to the broadcastable dimension.\nAny ideas on how can I do this stuff?<\/p>\n","answers":[{"AnswerId":"35308077","CreationDate":"2016-02-10T06:10:40.953","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This code is untested but you can create a vector of random numbers whose size equals the batch size and then add broadcastable dimensions using <code>dimshuffle<\/code>:<\/p>\n\n<pre><code>bias=self._srng.uniform(input.shape[0]).dimshuffle(0, 'x', 'x', 'x')\n<\/code><\/pre>\n"}]}
{"QuestionId":35298299,"AnswerCount":2,"Tags":"<numpy><tensorflow>","CreationDate":"2016-02-09T17:24:55.163","AcceptedAnswerId":null,"OwnerUserId":1569461.0,"Title":"Print full value of tensor into console or write to file in tensorflow","Body":"<p>I need to print a large tensor ([32,32,3]) into the console, and I only get output like this:<\/p>\n\n<pre><\/pre>\n\n<p>How can I get tensorflow to print out the entire tensor, instead of truncating it with the ellipses?<\/p>\n","answers":[{"AnswerId":"35303645","CreationDate":"2016-02-09T22:42:38.983","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The value returned from a TensorFlow <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow\"><code>Session.run()<\/code><\/a> call is a NumPy ndarray, so this rendering is controlled by NumPy itself. One simple way to ensure that all elements are printed is to use <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/reference\/generated\/numpy.set_printoptions.html\" rel=\"nofollow\"><code>numpy.set_printoptions()<\/code><\/a>:<\/p>\n\n<pre><code>import numpy\nnumpy.set_printoptions(threshold=numpy.nan)\n<\/code><\/pre>\n"},{"AnswerId":"35298597","CreationDate":"2016-02-09T17:41:19.567","ParentId":null,"OwnerUserId":"3637404","Title":null,"Body":"<p>use <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.savetxt.html#numpy.savetxt\" rel=\"nofollow\"><code>numpy.savetxt<\/code><\/a>:<\/p>\n\n<pre><code>x = y = z = np.arange(0.0,5.0,1.0)\nnp.savetxt('test.out', x, delimiter=',')   # X is an array\nnp.savetxt('test.out', (x,y,z))   # x,y,z equal sized 1D arrays\nnp.savetxt('test.out', x, fmt='%1.4e')   # use exponential notation\n<\/code><\/pre>\n"}]}
{"QuestionId":35298326,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-02-09T17:26:36.937","AcceptedAnswerId":35304001.0,"OwnerUserId":1716733.0,"Title":"\"freeze\" some variables\/scopes in tensorflow: stop_gradient vs passing variables to minimize","Body":"<p>I am trying to implement <a href=\"http:\/\/arxiv.org\/abs\/1406.2661\">Adversarial NN<\/a>, which requires to 'freeze' one or the other part of the graph during alternating training minibatches. I.e. there two sub-networks: G and D.<\/p>\n\n<pre><\/pre>\n\n<p>where loss function of  depends on .<\/p>\n\n<p>First I need to train parameters in D with all G parameters fixed, and then parameters in G with parameters in D fixed. Loss function in first case will be negative loss function in the second case and the update will have to apply to the parameters of whether first or second subnetwork.<\/p>\n\n<p>I saw that tensorflow has  function. For purpose of training the D (downstream) subnetwork I can use this function to block the gradient flow to <\/p>\n\n<pre><\/pre>\n\n<p>The  is very succinctly annotated with no in-line example (and example  is too long and not that easy to read), but looks like it must be called during the graph creation. <strong>Does it imply that if I want to block\/unblock gradient flow in alternating batches, I need to re-create and re-initialize the graph model?<\/strong> <\/p>\n\n<p>Also it seems that <strong>one cannot block the gradient flowing through the G (upstream) network by means of , right?<\/strong><\/p>\n\n<p>As an alternative I saw that one can pass the list of variables to the optimizer call as , which would be an easy solution if one could get all variables in the scopes of each subnetwork. <strong>Can one get a  for a tf.scope?<\/strong><\/p>\n","answers":[{"AnswerId":"35298465","CreationDate":"2016-02-09T17:34:26.133","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>Another option you might want to consider is you can set trainable=False on a variable. Which means it will not be modified by training.<\/p>\n\n<pre><code>tf.Variable(my_weights, trainable=False)\n<\/code><\/pre>\n"},{"AnswerId":"39258326","CreationDate":"2016-08-31T20:18:13.387","ParentId":null,"OwnerUserId":"3307732","Title":null,"Body":"<p>I don't know if my approach has down sides, but I solved this issue for myself with this construct:<\/p>\n\n<pre><code>do_gradient = &lt;Tensor that evaluates to 0 or 1&gt;\nno_gradient = 1 - do_gradient\nwrapped_op = do_gradient * original + no_gradient * tf.stop_gradient(original)\n<\/code><\/pre>\n\n<p>So if <code>do_gradient = 1<\/code>, the values and gradients will flow through just fine, but if <code>do_gradient = 0<\/code>, then the values will only flow through the stop_gradient op, which will stop the gradients flowing back.<\/p>\n\n<p>For my scenario, hooking do_gradient up to an index of a random_shuffle tensor let me randomly train different pieces of my network.<\/p>\n"},{"AnswerId":"40327164","CreationDate":"2016-10-30T08:14:44.907","ParentId":null,"OwnerUserId":"1666349","Title":null,"Body":"<p>@mrry's answer is completely right and perhaps more general than what I'm about to suggest. But I think a simpler way to accomplish it is to just pass the python reference directly to <code>var_list<\/code>: <\/p>\n\n<pre><code>W = tf.Variable(...)\nC = tf.Variable(...)\nY_est = tf.matmul(W,C)\nloss = tf.reduce_sum((data-Y_est)**2)\noptimizer = tf.train.AdamOptimizer(0.001)\n\n# You can pass the python object directly\ntrain_W = optimizer.minimize(loss, var_list=[W])\ntrain_C = optimizer.minimize(loss, var_list=[C])\n<\/code><\/pre>\n\n<p>I have a self-contained example here: <a href=\"https:\/\/gist.github.com\/ahwillia\/8cedc710352eb919b684d8848bc2df3a\" rel=\"noreferrer\">https:\/\/gist.github.com\/ahwillia\/8cedc710352eb919b684d8848bc2df3a<\/a> <\/p>\n"},{"AnswerId":"35304001","CreationDate":"2016-02-09T23:08:58.603","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to achieve this, as you mention in your question, is to create two optimizer operations using separate calls to <code>opt.minimize(cost, ...)<\/code>. By default, the optimizer will use all of the variables in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#trainable_variables\"><code>tf.trainable_variables()<\/code><\/a>. If you want to filter the variables to a particular scope, you can use the optional <code>scope<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#get_collection\"><code>tf.get_collection()<\/code><\/a> as follows:<\/p>\n\n<pre><code>optimizer = tf.train.AdagradOptimzer(0.01)\n\nfirst_train_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\n                                     \"scope\/prefix\/for\/first\/vars\")\nfirst_train_op = optimizer.minimize(cost, var_list=first_train_vars)\n\nsecond_train_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\n                                      \"scope\/prefix\/for\/second\/vars\")                     \nsecond_train_op = optimizer.minimize(cost, var_list=second_train_vars)\n<\/code><\/pre>\n"}]}
{"QuestionId":35298823,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-02-09T17:53:11.820","AcceptedAnswerId":null,"OwnerUserId":628096.0,"Title":"How to write a custom pooling layer module in tensor flow?","Body":"<p>I have thought that adding a new module which will do the center pooling. <\/p>\n\n<p>I was looking in the tensorflow code and there is a file named  which internally calls function from another file by passing \"Maxpool\", \"AvgPool\", etc. parameters to do the required computation.<\/p>\n\n<p>I want to do centre pooling which selects the center element in the window. I have code ready for the matlab and c++ versions but need to know how to add a new module in TensorFlow for such computation. Also where to set the backpropogation code. <\/p>\n","answers":[{"AnswerId":"35303470","CreationDate":"2016-02-09T22:30:36.883","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>A custom pooling layer would probably be implemented in C++. To see what you'd need to do, let's see where the implementation of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#max_pool_with_argmax\" rel=\"nofollow noreferrer\"><code>tf.nn.max_pool_with_argmax()<\/code><\/a> lives:<\/p>\n\n<ol>\n<li><p>The Python wrapper function (<code>tf.nn.max_pool_with_argmax()<\/code> itself) is automatically generated, in <code>gen_nn_ops.py<\/code>. This is ultimately imported into <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/75937824b9834b04c9a371447da27f322dfd9882\/tensorflow\/python\/ops\/nn.py\" rel=\"nofollow noreferrer\"><code>nn.py<\/code><\/a>, so that it appears under <code>tf.nn<\/code> when you <code>import tensorflow as tf<\/code>.<\/p><\/li>\n<li><p>In C++, there is an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0c06d8d9aa88b0596b734c8feb6435d2cf359b7e\/tensorflow\/core\/ops\/nn_ops.cc#L356\" rel=\"nofollow noreferrer\">op registration in <code>ops\/nn_ops.cc<\/code><\/a>, and a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0c06d8d9aa88b0596b734c8feb6435d2cf359b7e\/tensorflow\/core\/kernels\/maxpooling_op.cc#L527\" rel=\"nofollow noreferrer\">kernel registration in <code>kernels\/maxpooling_op.cc<\/code><\/a>.<\/p><\/li>\n<li><p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0c06d8d9aa88b0596b734c8feb6435d2cf359b7e\/tensorflow\/core\/kernels\/maxpooling_op.cc#L393\" rel=\"nofollow noreferrer\">kernel itself is defined in <code>kernels\/maxpooling_op.cc<\/code><\/a>.<\/p><\/li>\n<li><p>The gradient is defined as a separate op&mdash;<code>\"MaxPoolWithArgmaxGrad\"<\/code>&mdash;in the same places.<\/p><\/li>\n<\/ol>\n\n<p>There's a fair amount of work to do to add a new op (see <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">this tutorial<\/a> for a more complete guide), but hopefully these pointers can help!<\/p>\n"}]}
{"QuestionId":35302446,"AnswerCount":1,"Tags":"<python><hdf5><deep-learning><caffe>","CreationDate":"2016-02-09T21:23:30.037","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Prepare data for multioutput regression in Caffe","Body":"<p>I need to write python script that prepare data to feed it to a caffe solver.\nMy input is images() and vector of ints() (I have multioutput regression problem not single  for each ) and I try to modify Lenet to my task.<\/p>\n\n<p><a href=\"http:\/\/vision.stanford.edu\/teaching\/cs231n\/slides\/caffe_tutorial.pdf\" rel=\"nofollow noreferrer\">Here<\/a> I found that hdf5 can be a good option - it can be used from python, but drawbacks is that we can't do data augmentation on-the-fly and input images must be float32\/float64. <\/p>\n\n<p>Also <a href=\"https:\/\/stackoverflow.com\/questions\/33140000\/how-to-feed-caffe-multi-label-data-in-hdf5-format\">here<\/a> I found an example, but in example there is only 1D data, so I'm curious what shape images should have?<\/p>\n\n<p>Also <a href=\"https:\/\/stackoverflow.com\/questions\/28057957\/multi-label-regression-in-caffe\">here<\/a> I found info about  and  layers. What layer type should I use for multioutput regression?<\/p>\n","answers":[{"AnswerId":"35308731","CreationDate":"2016-02-10T06:57:03.417","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe expects its input images to be 4-D <code>B<\/code>-by-<code>C<\/code>-by-<code>H<\/code>-by-<code>W<\/code>:<\/p>\n\n<ul>\n<li><code>B<\/code> is the \"batch size\", the number of images you process simultaneously  <\/li>\n<li><code>C<\/code> is the number of channels, usually 3 for BGR (most nets conforms to opencv BGR format, rather than RGB. go figure...)  <\/li>\n<li><code>H<\/code> and <code>W<\/code> are the height and width of the image, respectively.<\/li>\n<\/ul>\n\n<p>Therefore, you need a python script that read images (you can use <code>caffe.io.load_image<\/code>) then transpose, resize, rescale and finally stack them into <code>B<\/code>-by-<code>C<\/code>-by-<code>H<\/code>-by-<code>W<\/code> numpy arrays of type <code>float32<\/code>.<br>\nYou can do the augmentations in python and save all the data into <code>hdf5<\/code> files.<\/p>\n"}]}
{"QuestionId":35304393,"AnswerCount":1,"Tags":"<mathematical-optimization><torch><gradient-descent><softmax>","CreationDate":"2016-02-09T23:44:10.820","AcceptedAnswerId":35328031.0,"OwnerUserId":3113501.0,"Title":"Trying to understand code that computes the gradient wrt to the input for LogSoftMax in Torch","Body":"<p>Code comes from: <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/lib\/THNN\/generic\/LogSoftMax.c\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/lib\/THNN\/generic\/LogSoftMax.c<\/a><\/p>\n\n<p>I don't see how this code is computing the gradient w.r.t to the input for the module LogSoftMax. What I'm confused about is what the two for loops are doing.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35328031","CreationDate":"2016-02-10T23:27:58.067","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>At forward time we have (with x = input vector, y = output vector, f = logsoftmax, i = i-th component):<\/p>\n\n<pre><code>yi = f(xi)\n   = log( exp(xi) \/ sum_j(exp(xj)) )\n   = xi - log( sum_j(exp(xj)) )\n<\/code><\/pre>\n\n<p>When computing the jacobian Jf of f you have (i-th row):<\/p>\n\n<pre><code>dyi\/dxi = 1 - exp(xi) \/ sum_j(exp(xj))\n<\/code><\/pre>\n\n<p>And for k different than i:<\/p>\n\n<pre><code>dyi\/dxk = - exp(xk) \/ sum_j(exp(xj))\n<\/code><\/pre>\n\n<p>This gives for Jf:<\/p>\n\n<pre><code>1-E(x1)     -E(x2)     -E(x3)    ...\n -E(x1)    1-E(x2)     -E(x3)    ...\n -E(x1)     -E(x2)    1-E(x3)    ...\n...\n<\/code><\/pre>\n\n<p>With <code>E(xi) = exp(xi) \/ sum_j(exp(xj))<\/code><\/p>\n\n<p>If we name gradInput the gradient w.r.t input and gradOutput the gradient w.r.t output the backpropagation gives (chain rule):<\/p>\n\n<pre><code>gradInputi = sum_j( gradOutputj . dyj\/dxi )\n<\/code><\/pre>\n\n<p>This is equivalent to:<\/p>\n\n<pre><code>gradInput = transpose(Jf) . gradOutput\n<\/code><\/pre>\n\n<p>Which finally gives for the i-th component:<\/p>\n\n<pre><code>gradInputi = gradOutputi - E(xi) . sum_j( gradOutputj )\n<\/code><\/pre>\n\n<p>So the first loop pre-computes <code>sum_j( gradOutputj )<\/code> and the last one the above term, i.e. i-th component of grad. input - except there is a missing <code>1 \/ sum_j(exp(xj))<\/code> for the exponential term in the Torch implementation (<em>the above calculus should probably be double checked even though it sounds correct and explains the current implementation<\/em>).<\/p>\n\n<p><strong>UPDATE<\/strong>: there is no problem with the <em>missing<\/em> <code>1 \/ sum_j(exp(xj))<\/code> term. Since the the jacobian is computed on the <em>output<\/em> value, and since this formerly computed output is precisely a log-softmax distribution it comes that the sum-exp of this distribution is 1:<\/p>\n\n<pre><code>sum_j(exp(outputj)) = sum_j(exp( log(exp(inputj) \/ sum_k(exp(inputk) ))\n                    = sum_j(         exp(inputj) \/ sum_k(exp(inputk)  )\n                    = 1\n<\/code><\/pre>\n\n<p>So there is no need to explicit this term in the implementation, which gives (for x = output):<\/p>\n\n<pre><code>gradInputi = gradOutputi - exp(outputi) . sum_j( gradOutputj )\n<\/code><\/pre>\n"}]}
{"QuestionId":35306191,"AnswerCount":2,"Tags":"<tensorflow><skflow>","CreationDate":"2016-02-10T03:12:28.220","AcceptedAnswerId":35388289.0,"OwnerUserId":5889446.0,"Title":"How to see values inside tensor while using skflow","Body":"<p>I've checked <a href=\"https:\/\/stackoverflow.com\/questions\/33633370\/how-to-print-the-value-of-a-tensor-object-in-tensorflow\">How to print the value of a Tensor object in TensorFlow?<\/a>. However doesn't seem to be working out of the box with skflow.\nE.g. tried like this :<\/p>\n\n<pre><\/pre>\n\n<p>Also tried <\/p>\n\n<pre><\/pre>\n\n<p>before calling word_vectors.eval(). But all this leads to crash:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know relatively easy way to look at content of Tensors created by skflow ?<\/p>\n","answers":[{"AnswerId":"35446871","CreationDate":"2016-02-17T02:38:16.977","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p><code>categorical_variable<\/code> won't initialize all the variables for you. It's not like the estimators that initialize the session, variables, graph, etc. At this moment, if you use <code>tf.initialize_all_variables().run()<\/code>, you will be able to get those values. Also, note that you can access weights and biases via <code>bias_<\/code> and <code>weights_<\/code> properties of the estimators. <\/p>\n"},{"AnswerId":"35388289","CreationDate":"2016-02-14T04:15:23.260","ParentId":null,"OwnerUserId":"728291","Title":null,"Body":"<p>It is not clear what error you are throwing but I got the error \"Attempting to use uninitialized value words\/words_embeddings\".  Initializing the variables fixed the error.<\/p>\n\n<p>So the working code was:<\/p>\n\n<pre><code>word_vectors = skflow.ops.categorical_variable([1,2], n_classes=3,embedding_size=2, name='words')\nwith tf.Session() as sess:\n  tf.initialize_all_variables().run()\n  word_vectors.eval()\n<\/code><\/pre>\n\n<p>I have also moved the graph node <code>word_vectors<\/code> outside of the session context though that wasn't necessary to fix the error.<\/p>\n"}]}
{"QuestionId":35310200,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2016-02-10T08:35:30.573","AcceptedAnswerId":35310384.0,"OwnerUserId":3507061.0,"Title":"How to download Docker images without a direct internet connection","Body":"<p>I see we can install tensorflow (with GPU) using Docker - here <a href=\"https:\/\/stackoverflow.com\/questions\/34694701\/tensorflow-which-docker-image-to-use\">TensorFlow - which Docker image to use?<\/a><\/p>\n\n<p>But how do I do this on a machine that has no external internet connection?\nIs there a way to first download the tensor flow image <\/p>\n\n<pre><\/pre>\n\n<p>and then copy it to local file space and \"install\" it from there?<\/p>\n","answers":[{"AnswerId":"35310384","CreationDate":"2016-02-10T08:45:47.760","ParentId":null,"OwnerUserId":"434408","Title":null,"Body":"<p>You can pull the image on a computer that have access to the internet.<\/p>\n\n<pre><code>sudo docker pull b.gcr.io\/tensorflow\/tensorflow:latest-gpu\n<\/code><\/pre>\n\n<p>Then you can save this image to a file<\/p>\n\n<pre><code>sudo docker save -o tensorflow_image.docker b.gcr.io\/tensorflow\/tensorflow:latest-gpu\n<\/code><\/pre>\n\n<p>Transfer the file on the offline computer (USB\/CD\/whatever) and load the image from the file:<\/p>\n\n<pre><code>sudo docker load &lt; tensorflow_image.docker\n<\/code><\/pre>\n\n<p>Courtesy: <a href=\"https:\/\/serverfault.com\/questions\/701248\/downloading-docker-image-for-transfer-to-non-internet-connected-machine\">https:\/\/serverfault.com\/questions\/701248\/downloading-docker-image-for-transfer-to-non-internet-connected-machine<\/a><\/p>\n"}]}
{"QuestionId":35311039,"AnswerCount":4,"Tags":"<lua><torch>","CreationDate":"2016-02-10T09:18:35.517","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"lua - Read .raw file","Body":"<p>I am trying to read a .raw image file into lua but currently I am having little success converting the binary to integers. I am trying the following code;<\/p>\n\n<pre><\/pre>\n\n<p>which yields a big binary print including lines like the below;<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35314457","CreationDate":"2016-02-10T11:51:18.170","ParentId":null,"OwnerUserId":"5675002","Title":null,"Body":"<p>Here's your binary data read, assuming there's only shorts:<\/p>\n\n<pre><code>local f = io.open(\"data.bin\",\"rb\")\nlocal data = f:read(\"*all\")\nf:close()\n\nfor i=1,data:len(),2 do\n    print(data:byte(i)+256*data:byte(i+1))\nend\n<\/code><\/pre>\n\n<p>In short - you read it byte-by-byte and reconstruct numeric values.<br>You will need to now if your data is LSB or MSB. My sample assumes LSB.<\/p>\n"},{"AnswerId":"35315246","CreationDate":"2016-02-10T12:24:52.050","ParentId":null,"OwnerUserId":"5661707","Title":null,"Body":"<p>If the whole file is made of shorts, you can use the torch <a href=\"https:\/\/github.com\/andresy\/torch7-split\/blob\/master\/rocks\/torch\/doc\/storage.md#__torch.StorageMap\" rel=\"nofollow\">Storage<\/a> as follows<\/p>\n\n<p><code>local f = torch.ShortStorage(\"data.raw\")<\/code><\/p>\n"},{"AnswerId":"35335168","CreationDate":"2016-02-11T09:20:03.407","ParentId":null,"OwnerUserId":"5908892","Title":null,"Body":"<p>.raw files are in a other format. read it with binary or found out the correct format of the image. decode it.<\/p>\n"},{"AnswerId":"35314342","CreationDate":"2016-02-10T11:46:34.547","ParentId":null,"OwnerUserId":"5352026","Title":null,"Body":"<p>Maybe you should read this section in the manual about binary files:<\/p>\n\n<blockquote>\n  <p>The simple model functions io.input and io.output always open a file in text mode (the default). In Unix, there is no difference between binary files and text files. But in some systems, notably Windows, binary files must be opened with a special flag. To handle such binary files, you must use io.open, with the letter `b\u00b4 in the mode string.<\/p>\n<\/blockquote>\n\n<p>Source: <a href=\"http:\/\/www.lua.org\/pil\/21.2.2.html\" rel=\"nofollow\">http:\/\/www.lua.org\/pil\/21.2.2.html<\/a><\/p>\n"}]}
{"QuestionId":35314898,"AnswerCount":0,"Tags":"<tensorflow><stacked><mnist><autoencoder><softmax>","CreationDate":"2016-02-10T12:10:38.043","AcceptedAnswerId":null,"OwnerUserId":5686003.0,"Title":"how to extract (plot) the hidden units\/softmax (features) from a autoencoder using Tensorflow","Body":"<p>I'm new with ML, and I'm using tensorflow.<br>\nI want to see the features of my autoencoder, but I don't know how to extract(see) the hidden units.<br>\ncould someone help me?<\/p>\n\n<p>I made my own dataset, but the original code is with MNIST:\n<a href=\"https:\/\/jmetzen.github.io\/2015-11-27\/vae.html\" rel=\"nofollow\">https:\/\/jmetzen.github.io\/2015-11-27\/vae.html<\/a><\/p>\n\n<p>I added a layer of hidden units to transform it to Stacked AE.. (I don't know if it became stacked, but worked)<\/p>\n\n<p>my AE: <a href=\"https:\/\/github.com\/Snowedze\/Autoencoder.git\" rel=\"nofollow\">https:\/\/github.com\/Snowedze\/Autoencoder.git<\/a><\/p>\n","answers":[]}
{"QuestionId":35315723,"AnswerCount":1,"Tags":"<numpy><theano>","CreationDate":"2016-02-10T12:46:42.207","AcceptedAnswerId":35319787.0,"OwnerUserId":1150636.0,"Title":"Transform a matrix made of binomial vectors to ranges for consecutive zeros","Body":"<p>I am trying to figure out how to do this transformation symbolically in theano a matrix of undetermined size<\/p>\n\n<p>From:<\/p>\n\n<pre><\/pre>\n\n<p>To:<\/p>\n\n<pre><\/pre>\n\n<p>So for every consecutive 0 I want an increasing range and whenever I stumble on a 1 the range resets.<\/p>\n","answers":[{"AnswerId":"35319787","CreationDate":"2016-02-10T15:52:21.457","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>Here's one way to do it, using inefficient scans:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as tt\n\n\ndef inner_step(x_t_t, y_t_tm1):\n    return tt.switch(x_t_t, 0, y_t_tm1 + 1)\n\n\ndef outer_step(x_t):\n    return theano.scan(inner_step, sequences=[x_t], outputs_info=[0])[0]\n\n\ndef compile():\n    x = tt.bmatrix()\n    y = theano.scan(outer_step, sequences=[x])[0]\n    return theano.function([x], y)\n\n\ndef main():\n    f = compile()\n    data = [[0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 1],\n            [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1]]\n    print f(data)\n\nmain()\n<\/code><\/pre>\n\n<p>When run, this prints:<\/p>\n\n<pre><code>[[1 2 3 0 1 2 3 4 5 0 0 1 0 1 2 3 0]\n [1 2 3 4 5 6 7 8 0 1 2 0 0 0 0 0 0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35316250,"AnswerCount":4,"Tags":"<python><dictionary><lookup><tensorflow>","CreationDate":"2016-02-10T13:11:57.243","AcceptedAnswerId":null,"OwnerUserId":2206976.0,"Title":"Tensorflow Dictionary lookup with String tensor","Body":"<p>Is there any way to perform a dictionary lookup based on a String tensor in Tensorflow?<\/p>\n\n<p>In plain Python, I'd do something like<\/p>\n\n<pre><\/pre>\n\n<p>. Now I'd like to do the same thing at Tensorflow runtime, when I have my  as a String tensor. Something like<\/p>\n\n<pre><\/pre>\n\n<p>would be nice.<\/p>\n","answers":[{"AnswerId":"40640870","CreationDate":"2016-11-16T19:39:42.130","ParentId":null,"OwnerUserId":"3050716","Title":null,"Body":"<p>You might find <code>tensorflow.contrib.lookup<\/code> helpful:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lookup\/lookup_ops.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lookup\/lookup_ops.py<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/lookup\/HashTable\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/lookup\/HashTable<\/a><\/p>\n\n<p>In particular, you can do:<\/p>\n\n<pre><code>table = tf.contrib.lookup.HashTable(\n  tf.contrib.lookup.KeyValueTensorInitializer(keys, values), -1\n)\nout = table.lookup(input_tensor)\ntable.init.run()\nprint out.eval()\n<\/code><\/pre>\n"},{"AnswerId":"35321717","CreationDate":"2016-02-10T17:16:43.237","ParentId":null,"OwnerUserId":"5875585","Title":null,"Body":"<p>TensorFlow is a data flow language with no support for data structures other than tensors.  There is no map or dictionary type.  However, depending on what you need, when you're using the Python wrapper it is possible to maintain a dictionary in the driver process, which executes in Python, and use it to interact with the TensorFlow graph execution.  For example, you could execute one step of a TensorFlow graph within a session, return a string value to the Python driver, use it as a key into a dictionary in the driver, and use the retrieved value to determine the next computation to be requested from the session.  This is probably not a good solution if the speed of these dictionary lookups is performance critical.<\/p>\n"},{"AnswerId":"57123833","CreationDate":"2019-07-20T10:27:16.263","ParentId":null,"OwnerUserId":"11811625","Title":null,"Body":"<p>tf.gather can help you, but it only gets values of list. You can convert dictionary into key and value lists, and then apply tf.gather. Example:<\/p>\n\n<pre><code># Your dict\ndict_ = {'a': 1.12, 'b': 5.86, 'c': 68.}\n# concrete query\nquery_list = ['a', 'c']\n\n# unpack key and value lists\nkey, value = list(zip(*dict_.items()))\n# map query list to list -&gt; [0, 2]\nquery_list = [i for i, s in enumerate(key) if s in query_list]\n\n# query as tensor\nquery = tf.placeholder(tf.int32, shape=[None])\n# convert value list to tensor\nvl_tf = tf.constant(value)\n# get value\nmy_vl = tf.gather(vl_tf, query)\n\n# session run\nsess = tf.InteractiveSession()\nsess.run(my_vl, feed_dict={query:query_list})\n<\/code><\/pre>\n"},{"AnswerId":"55995756","CreationDate":"2019-05-05T19:44:44.003","ParentId":null,"OwnerUserId":"5681083","Title":null,"Body":"<p>If you want to run this with new TF 2.0 code with eager execution enabled by default. Below is the quick code snippet.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# build a lookup table\ntable = tf.lookup.StaticHashTable(\n    initializer=tf.lookup.KeyValueTensorInitializer(\n        keys=tf.constant([0, 1, 2, 3]),\n        values=tf.constant([10, 11, 12, 13]),\n    ),\n    default_value=tf.constant(-1),\n    name=\"class_weight\"\n)\n\n# now let us do a lookup\ninput_tensor = tf.constant([0, 0, 1, 1, 2, 2, 3, 3])\nout = table.lookup(input_tensor)\nprint(out)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tf.Tensor([10 10 11 11 12 12 13 13], shape=(8,), dtype=int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":35317029,"AnswerCount":1,"Tags":"<computer-vision><classification><tensorflow><scene><labeling>","CreationDate":"2016-02-10T13:49:53.983","AcceptedAnswerId":null,"OwnerUserId":2204511.0,"Title":"How to implement pixel-wise classification for scene labeling in TensorFlow?","Body":"<p>I am working on a deep learning model using <strong>Google's TensorFlow<\/strong>. The model should be used to <strong>segment and label scenes<\/strong>. <\/p>\n\n<ol>\n<li>I am using the <strong>SiftFlow dataset<\/strong> which has <em>33 semantic\nclasses<\/em> and <em>images with 256x256 pixels<\/em>.  <\/li>\n<li>As a result, at my final layer using convolution and deconvolution I arrive at the following tensor(array) <em>[256, 256, 33]<\/em>. <\/li>\n<li>Next I would like to\napply <em>Softmax<\/em> and compare the results to a semantic label of size\n<em>[256, 256]<\/em>.<\/li>\n<\/ol>\n\n<p><strong>Questions:<\/strong>\nShould I apply mean averaging or argmax to my final layer so its shape becomes <em>[256,256,1]<\/em> and then loop through each pixel and classify as if I were classying <em>256x256<\/em> instances? If the answer is yes, how, if not, what other options?<\/p>\n","answers":[{"AnswerId":"37294185","CreationDate":"2016-05-18T08:30:31.877","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>To apply softmax and use a <strong>cross entropy loss<\/strong>, you have to keep <strong>intact<\/strong> the final output of your network of size <em>batch_size x 256 x 256 x 33<\/em>. Therefore you <strong>cannot use<\/strong> mean averaging or argmax because it would destroy the output probabilities of your network.<\/p>\n\n<p>You have to loop through all the <em>batch_size x 256 x 256<\/em> pixels and apply a cross entropy loss to your prediction for this pixel. This is easy with the built-in function <code>tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels)<\/code>.<\/p>\n\n<p>Some warnings <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/nn.html#sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\">from the doc<\/a> before applying the code below:<\/p>\n\n<ul>\n<li>WARNING: This op expects <strong>unscaled logits<\/strong>, since it performs a softmax on logits internally for efficiency. Do not call this op with the output of softmax, as it will produce incorrect results.  <\/li>\n<li>logits and must have the shape [batch_size, num_classes] and the dtype (either float32 or float64).<\/li>\n<li>labels must have the shape [batch_size] and the dtype int64.<\/li>\n<\/ul>\n\n<p>The trick is to use <code>batch_size * 256 * 256<\/code> as the batch size required by the function. We will reshape <code>logits<\/code> and <code>labels<\/code> to this format.\nHere is the code I use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = tf.placeholder(tf.float32, [batch_size, 256, 256, 3])  # input images\nlogits = inference(inputs)  # your outputs of shape [batch_size, 256, 256, 33] (no final softmax !!)\nlabels = tf.placeholder(tf.float32, [batch_size, 256, 256])  # your labels of shape [batch_size, 256, 256] and type int64\n\nreshaped_logits = tf.reshape(logits, [-1, 33])  # shape [batch_size*256*256, 33]\nreshaped_labels = tf.reshape(labels, [-1])  # shape [batch_size*256*256]\nloss = sparse_softmax_cross_entropy_with_logits(reshaped_logits, reshaped_labels)\n<\/code><\/pre>\n\n<p>You can then apply your optimizer on that loss.<\/p>\n\n<hr>\n\n<h3>Update: v0.10<\/h3>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn.html#sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\">documentation<\/a> of <code>tf.sparse_softmax_cross_entropy_with_logits<\/code> shows that it now accepts any shape for <code>logits<\/code>, so there is no need to reshape the tensors (thanks @chillinger):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = tf.placeholder(tf.float32, [batch_size, 256, 256, 3])  # input images\nlogits = inference(inputs)  # your outputs of shape [batch_size, 256, 256, 33] (no final softmax !!)\nlabels = tf.placeholder(tf.float32, [batch_size, 256, 256])  # your labels of shape [batch_size, 256, 256] and type int64\n\nloss = sparse_softmax_cross_entropy_with_logits(logits, labels)\n<\/code><\/pre>\n"}]}
{"QuestionId":35320056,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-10T16:03:53.387","AcceptedAnswerId":null,"OwnerUserId":5894128.0,"Title":"notMNIST not downloadable in TensorFlow Udacity course","Body":"<p>I'm doing the Udacity TensorFlow course, first exercise: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/1_notmnist.ipynb\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/1_notmnist.ipynb<\/a><\/p>\n\n<p>OSX 10.11 (El Capitan)\nPython 2.7\nvirtualenv installation of TF<\/p>\n\n<p>I am getting an error:<\/p>\n\n<p>\"Exception: Failed to verifynotMNIST_large.tar.gz. Can you get to it with a browser?\"<\/p>\n\n<p>It finds the \u201csmall\u201d file, but not the \u201clarge.\u201d  Appreciate help.  Thanks.<\/p>\n\n<p>Here is the whole block of code:<\/p>\n\n<pre><\/pre>\n\n<p>Here is what gets returned:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40660303","CreationDate":"2016-11-17T16:37:07.567","ParentId":null,"OwnerUserId":"4004167","Title":null,"Body":"<p>Also face the same situation.<\/p>\n\n<p>That is a simple thing to handle it and continue.<\/p>\n\n<blockquote>\n  <p>There's a size mismatch.<\/p>\n  \n  <p>Just re-run the code with <strong>force=True<\/strong>, and\n  it works right now!<\/p>\n<\/blockquote>\n\n<p>If you try to download it manually it will also be Ok.<\/p>\n\n<p>As described in this thread: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1475\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1475<\/a><\/p>\n\n<p>Hope it helps.<\/p>\n"},{"AnswerId":"35967892","CreationDate":"2016-03-13T07:52:38.373","ParentId":null,"OwnerUserId":"6054156","Title":null,"Body":"<p>I found the solution (workaround?) of the issue in Udacity forum.<\/p>\n\n<p><a href=\"https:\/\/discussions.udacity.com\/t\/not-able-to-load-the-dataset-for-assingment-1\/160124\" rel=\"nofollow\">https:\/\/discussions.udacity.com\/t\/not-able-to-load-the-dataset-for-assingment-1\/160124<\/a><\/p>\n\n<p>In the page, two solutions are proposed.<\/p>\n\n<p>a. Download notMNIST_large.tar.gz locally via browser, and copy it using docker command.<\/p>\n\n<p>or<\/p>\n\n<p>b. Add \"User-Agent\" header in fetching script.<\/p>\n\n<p>I think both approaches works fine.<\/p>\n"}]}
{"QuestionId":35320365,"AnswerCount":2,"Tags":"<gpu><gpgpu><theano><tensorflow>","CreationDate":"2016-02-10T16:16:26.813","AcceptedAnswerId":35321198.0,"OwnerUserId":2680707.0,"Title":"Using a GPU both as video card and GPGPU","Body":"<p>Where I work, we do a lot of numerical computations and we are considering buying workstations with NVIDIA video cards because of CUDA (to work with TensorFlow and Theano).<\/p>\n\n<p>My question is: should these computers come with another video card to handle the display and free the NVIDIA for the GPGPU?<\/p>\n\n<p>I would appreciate if anyone knows of hard data on using a video card for display and GPGPU at the same time.<\/p>\n","answers":[{"AnswerId":"35344940","CreationDate":"2016-02-11T16:36:16.127","ParentId":null,"OwnerUserId":"5785176","Title":null,"Body":"<p>Having been through this, I'll add my two cents.<\/p>\n\n<p>It is helpful to have a dedicated card for computations, but it is definitely not necessary.<\/p>\n\n<p>I have used a development workstation with a single high-end GPU for both display and compute.  I have also used workstations with multiple GPUs, as well as headless compute servers.<\/p>\n\n<p>My experience is that doing compute on the display GPU is fine as long as demands on the display are typical for software engineering.  In a Linux setup with a couple monitors, web browsers, text editors, etc., I use about 200MB for display out of the 6GB of the card -- so only about 3% overhead.  You might see the display stutter a bit during a web page refresh or something like that, but the throughput demands of the display are very small.<\/p>\n\n<p>One technical issue worth noting for completeness is that the NVIDIA driver, GPU firmware, or OS may have a timeout for kernel completion on the display GPU (run NVIDIA's 'deviceQueryDrv' to see the driver's \"run time limit on kernels\" setting).  In my experience (on Linux), with machine learning, this has never been a problem since the timeout is several seconds and, even with custom kernels, synchronization across multiprocessors constrains how much you can stuff into a single kernel launch.  I would expect the typical runs of the pre-baked ops in TensorFlow to be two or more orders of magnitude below this limit.<\/p>\n\n<p>That said, there are some big advantages of having multiple compute-capable cards in a workstation (whether or not one is used for display).  Of course there is the potential for more throughput (if your software can use it).  However, the main advantage in my experience, is being able to run long experiments while concurrently developing new experiments.<\/p>\n\n<p>It is of course feasible to start with one card and then add one later, but make sure your motherboard has lots of room and your power supply can handle the load.  If you decide to have two cards, with one being a low-end card dedicated to display, I would specifically advise against having the low-end card be a CUDA-capable card lest it get selected as a default for computation.<\/p>\n\n<p>Hope that helps.<\/p>\n"},{"AnswerId":"35321198","CreationDate":"2016-02-10T16:51:57.263","ParentId":null,"OwnerUserId":"5875585","Title":null,"Body":"<p>In my experience it is awkward to share a GPU card between numerical computation tasks and driving a video monitor.  For example, there is limited memory available on any GPU, which is often the limiting factor in the size of a model you can train.  Unless you're doing gaming, a fairly modest GPU is probably adequate to drive the video.  But for serious ML work you will probably want a high-performance card.  Where I work (Google) we typically put two GPUs in desk-side machines when one is to be used for numerical computation.<\/p>\n"}]}
{"QuestionId":35322409,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2016-02-10T17:49:28.227","AcceptedAnswerId":null,"OwnerUserId":5905804.0,"Title":"Running Memnet model using caffe","Body":"<p>I try to use the dataset of mnist, which is an example provided in github, to run Memnet model on the interface of cmd.<\/p>\n\n<p>The model is downloaded from <a href=\"http:\/\/memorability.csail.mit.edu\/download.html\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>I modified its deploy.prototxt accordingly. Having no idea... Can someone help me with this?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/NRGgk.png\" alt=\"memnet\"><\/p>\n\n<p>But it keeps telling me something wrong going on, as the pic shows:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/0tmmT.png\" alt=\"cmd error\"><\/p>\n","answers":[{"AnswerId":"35332619","CreationDate":"2016-02-11T06:50:56.273","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The command line interface should get as a <code>-solver<\/code> a <code>solver.prototxt<\/code> file (which has <code>train_val.prototxt<\/code> as one of its parameters). You cannot supply <code>train_val.prototxt<\/code> directly to <code>caffe train<\/code>.<\/p>\n\n<p>You can look at the <code>examples<\/code> subfolder of caffe and find some examples of <code>solver.prototxt<\/code>. A simple one can be found in <code>examples\/mnist<\/code>, you can check out <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow\"><code>lenet_solver.prototxt<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":35323023,"AnswerCount":1,"Tags":"<python-2.7><neural-network><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-02-10T18:22:27.867","AcceptedAnswerId":null,"OwnerUserId":4975126.0,"Title":"State Normalization of RNNs","Body":"<p>Perhaps a question better posed to Computer Science or Cross Validated?<\/p>\n\n<hr>\n\n<p>I'm beginning some work with LSTM on sequences of arbitrary length and one problem I'm experiencing and that I haven't seen addressed, is that my network seems to have developed a couple parameters that grow linearly (perhaps as a measure of time?).<\/p>\n\n<p>The obvious issue with this is that the training data is bounded at a sequence of length  and so the network grows this parameter reasonably up until tilmestep . But after that, the network will eventually NAN because values are getting too extreme.<\/p>\n\n<p>Has anyone read anything about the normalization of stabilization of states over time?<\/p>\n\n<p>Any suggestions would be much appreciated.<\/p>\n","answers":[{"AnswerId":"39259772","CreationDate":"2016-08-31T22:09:51.013","ParentId":null,"OwnerUserId":"447599","Title":null,"Body":"<p>Idea #1: Gradient clipping is often applied in RNNs. Here is an example of implementation: <a href=\"https:\/\/stackoverflow.com\/questions\/36498127\/how-to-effectively-apply-gradient-clipping-in-tensor-flow\">How to effectively apply gradient clipping in tensor flow?<\/a><\/p>\n\n<p>Idea #2: Using <a href=\"https:\/\/arxiv.org\/pdf\/1603.09025.pdf\" rel=\"nofollow noreferrer\">Recurrent Batch Normalization (arXiv)<\/a> (<a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"nofollow noreferrer\">Batch Normalization<\/a>)<\/p>\n\n<p>Here is a Tensorflow implementation of a batch normalized LSTM cell: <a href=\"https:\/\/github.com\/OlavHN\/bnlstm\/blob\/master\/lstm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/OlavHN\/bnlstm\/blob\/master\/lstm.py<\/a><\/p>\n\n<p>This implementation is explained in the article here : <a href=\"http:\/\/olavnymoen.com\/2016\/07\/07\/rnn-batch-normalization\" rel=\"nofollow noreferrer\">Batch normalized LSTM for Tensorflow<\/a><\/p>\n"}]}
{"QuestionId":35324721,"AnswerCount":1,"Tags":"<android><tensorflow><bazel>","CreationDate":"2016-02-10T19:58:26.310","AcceptedAnswerId":null,"OwnerUserId":3920097.0,"Title":"Bazel generate_workspace not working?","Body":"<p>I'm trying to create my own Android app that uses TensorFlow.  I have previously built the TensorFlow Android example but that Android project is embedded inside the TensorFlow bazel project and I'd like to put my project in it's own place elsewhere on the filesystem.  So I've started down the path of using bazel to build my Android app and incorporate TensorFlow that way.<\/p>\n\n<p>I started by looking at the bazel docs on <a href=\"http:\/\/bazel.io\/docs\/external.html\" rel=\"nofollow\">external dependencies<\/a>.  It mentions running  to auto generate the WORKSPACE and BUILD files to include all the dependencies for all dependent bazel projects. So I ran this:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following output with tons of warnings about skipping transitive dependencies in tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>The generated WORKSPACE and BUILD files are essentially blank (only a few generated comments in each).  This obviously doesn't seem right but admittedly I don't know what I'm doing with incorporating external bazel projects as dependencies in my Android bazel project.<\/p>\n\n<p>What have I done wrong and has anyone managed to incorporate TensorFlow into a separate Android project without jamming the Android project down inside the TensorFlow project?<\/p>\n","answers":[{"AnswerId":"37974112","CreationDate":"2016-06-22T16:52:15.863","ParentId":null,"OwnerUserId":"436830","Title":null,"Body":"<p>I faced the same problem and had a look in the utility's source code. It supports only <code>maven_jar<\/code> WORKSPACE rule at the moment.<\/p>\n"}]}
{"QuestionId":35325480,"AnswerCount":1,"Tags":"<python><python-2.7><neural-network><tensorflow><recurrent-neural-network>","CreationDate":"2016-02-10T20:41:24.790","AcceptedAnswerId":null,"OwnerUserId":4975126.0,"Title":"TensorFlow: Performing this loss computation","Body":"<p>My question and problem is stated below the two blocks of code.<\/p>\n\n<hr>\n\n<h1>Loss Function<\/h1>\n\n<pre><\/pre>\n\n<hr>\n\n<h1>Being used in<\/h1>\n\n<p>I've tried to flatten my code so that everything is happening in one place. Let me know if there are typos\/points of confusion.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h1>My issue<\/h1>\n\n<p>The issue is that this is returning the error: <\/p>\n\n<pre><\/pre>\n\n<p>So I assume that this is TensorFlow complaining that it can't compute the gradients of my loss because the loss is performed by numpy, outside the scope of TF.<\/p>\n\n<p>So naturally to fix that I would try and implement this in TensorFlow. The issue is, my  and  are both Tensors, so when I try and access a single element, I'm returned a Tensor of shape []. This is an issue when I'm trying to use  which takes an  for its  parameter.<\/p>\n\n<p>Another issue with that is my  is a Placeholder and since my  value need to be defined before the  call, I also get an error that says a value needs to be passed for the placeholder.<\/p>\n\n<p>I'm just wondering how I could try and implement this loss function. Or if I'm missing something obvious.<\/p>\n\n<hr>\n\n<p><strong>Edit:<\/strong> After some <a href=\"http:\/\/arxiv.org\/pdf\/1511.06456v4.pdf\" rel=\"noreferrer\">further reading<\/a> I see that usually losses like the one I describe are used in validation and in training a surrogate loss that minimizes in the same place as the true loss is used. Does anyone know what surrogate loss is used for an edit distance based scenario like mine?<\/p>\n","answers":[{"AnswerId":"43889136","CreationDate":"2017-05-10T09:57:23.953","ParentId":null,"OwnerUserId":"277089","Title":null,"Body":"<p>The first thing I would do is to calculate loss using tensorflow instead of numpy. That will allow tensorflow to compute gradients for you, so you will be able to back-propagate, meaning you can minimize the loss.<\/p>\n\n<p>There is tf.edit_distance(<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/edit_distance\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/edit_distance<\/a>) function in the core library. <\/p>\n\n<blockquote>\n  <p>So naturally to fix that I would try and implement this in TensorFlow. The issue is, my logit_lengths and label_lengths are both Tensors, so when I try and access a single element, I'm returned a Tensor of shape []. This is an issue when I'm trying to use tf.nn.top_k() which takes an Int for its k parameter.<\/p>\n<\/blockquote>\n\n<p>Could you provide a little bit more details why it is an issue?<\/p>\n"}]}
{"QuestionId":35330084,"AnswerCount":1,"Tags":"<tensorflow><bazel><protocol-buffers>","CreationDate":"2016-02-11T03:05:55.903","AcceptedAnswerId":null,"OwnerUserId":1111929.0,"Title":"installing TensorFlow on the IBM power8","Body":"<p>I have access to a large IBM Power8 machine, and would like to install TensorFlow on it. Naturally, I tried the quick pip install, but it failed: <\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, pip.log cotains little useful info.<\/p>\n\n<pre><\/pre>\n\n<p>Next thing I tried was to build TensorFlow from source. To no avail, all my attempts ended with some  message, e.g.:<\/p>\n\n<pre><\/pre>\n\n<p>So then I tried to compile Bazel from source, which also resulted in a similar hard error. <\/p>\n\n<pre><\/pre>\n\n<p>I however found this link <a href=\"http:\/\/www.cnblogs.com\/rodenpark\/p\/5007744.html\" rel=\"nofollow\">http:\/\/www.cnblogs.com\/rodenpark\/p\/5007744.html<\/a> that explains how to build the Protobuf compiler from source on the Power8 machine. This worked and after the modifications described in his other topic <a href=\"http:\/\/www.cnblogs.com\/rodenpark\/p\/5007846.html\" rel=\"nofollow\">http:\/\/www.cnblogs.com\/rodenpark\/p\/5007846.html<\/a> I managed to at least get the compilation process started. But now it crashes with a ton of errors which each seem less severe on their own but the vast amount of them makes it look really hopeless, I posted them on <a href=\"http:\/\/pastebin.com\/KjkseaGx\" rel=\"nofollow\">http:\/\/pastebin.com\/KjkseaGx<\/a> for reference.<\/p>\n\n<p>So... I'm running out of inspiration. What can I do to make TensorFlow work on the Power8 machine? <\/p>\n","answers":[{"AnswerId":"39764135","CreationDate":"2016-09-29T07:34:33.587","ParentId":null,"OwnerUserId":"3404962","Title":null,"Body":"<h1>Install bazel 0.2.0-ppc<\/h1>\n\n<pre><code>tf@ubuntu16:~$ git clone https:\/\/github.com\/ibmsoe\/bazel\ntf@ubuntu16:~\/bazel$ git checkout v0.2.0-ppc\ntf@ubuntu16:~\/bazel$ .\/compile.sh\n<\/code><\/pre>\n\n<h1>Install tensorflow<\/h1>\n\n<pre><code>tf@ubuntu16:~$ git clone --recurse-submodules https:\/\/github.com\/tensorflow\/tensorflow\ntf@ubuntu16:~\/tensorflow$ git checkout  v0.10.0rc0\ntf@ubuntu16:~\/tensorflow$ git commit -m\"v0.10.0rc0\"\ntf@ubuntu16:~\/tensorflow$ git cherry-pick ce70f6cf842a46296119337247c24d307e279fa0\ntf@ubuntu16:~\/tensorflow$ git cherry-pick f1acb3bd828a73b15670fc8019f06a5cd51bd564\ntf@ubuntu16:~\/tensorflow$ git cherry-pick 9b6215a691a2eebaadb8253bd0cf706f2309a0b8\ntf@ubuntu16:~\/tensorflow$ .\/configure\ntf@ubuntu16:~\/tensorflow$ bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>Here you'll encounter an error, something like this<\/p>\n\n<blockquote>\n  <p>ERROR: \/home\/tf\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/external\/farmhash_archive\/BUILD:5:1: Executing genrule @farmhash_archive\/\/:configure failed: bash failed: error executing command \/bin\/bash -c ... (remaining 1 argument(s) skipped): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.\n  \/home\/tf\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/tensorflow\/external\/farmhash_archive\/farmhash-34c13ddfab0e35422f4c3979f360635a8c050260 \/home\/tf\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/tensorflow\n  \/tmp\/tmp.XdCPQefJyZ \/home\/tf\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/tensorflow\/external\/farmhash_archive\/farmhash-34c13ddfab0e35422f4c3979f360635a8c050260 \/home\/tf\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/tensorflow<\/p>\n<\/blockquote>\n\n<p>You'll have to edit config.guess as below to insert a stanza for ppc64le <\/p>\n\n<pre><code>tf@ubuntu16:~\/.cache\/bazel\/_bazel_tf\/b2f766da603b0bed56d4c1d0b178456a\/external\/farmhash_archive\/farmhash-34c13ddfab0e35422f4c3979f360635a8c050260$ vi config.guess\n<\/code><\/pre>\n\n\n\n<pre><code>    *:BSD\/OS:*:*)\n        echo ${UNAME_MACHINE}-unknown-bsdi${UNAME_RELEASE}\n        exit ;;\n+    ppc64le:Linux:*:*)\n+      echo powerpc64le-unknown-linux-gnu\n+      exit ;;\n    *:FreeBSD:*:*)\n        case ${UNAME_MACHINE} in\n<\/code><\/pre>\n\n\n\n<pre><code>tf@ubuntu16:~\/tensorflow$ bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\ntf@ubuntu16:~\/tensorflow$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\ntf@ubuntu16:~\/tensorflow$ sudo pip install \/tmp\/tensorflow_pkg\/tensorflow*.whl\ntf@ubuntu16:~\/tensorflow\/bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package.runfi\ntf@ubuntu16:~\/tensorflow$ mkdir _python_build\ntf@ubuntu16:~\/tensorflow$ cd _python_build\ntf@ubuntu16:~\/tensorflow\/_python_build$ ln -s ~\/tensorflow\/bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package.runfiles\/tensorflow\/* .\ntf@ubuntu16:~\/tensorflow\/_python_build$ ln -s ~\/tensorflow\/tools\/* .\ntf@ubuntu16:~\/tensorflow\/_python_build$ python  __init__.py develop\n<\/code><\/pre>\n"}]}
{"QuestionId":35330117,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2016-02-11T03:10:17.813","AcceptedAnswerId":null,"OwnerUserId":5808490.0,"Title":"How can I run a loop with a tensor as its range? (in tensorflow)","Body":"<p>I want to have a for loop that the number of its iterations is depend on a tensor value. For example:<\/p>\n\n<pre><\/pre>\n\n<p>However I get the following error:<\/p>\n\n<p><strong>\"TypeError: 'Tensor' object is not iterable\"<\/strong><\/p>\n\n<p>What should I do?<\/p>\n","answers":[{"AnswerId":"39010125","CreationDate":"2016-08-18T04:40:15.317","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>To do this you will need to use the tensorflow while loop (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"noreferrer\"><code>tf.while_loop<\/code><\/a>) as follows:<\/p>\n\n<pre><code>i = tf.constant(0)\nwhile_condition = lambda i: tf.less(i, input_placeholder[1, 1])\ndef body(i):\n    # do something here which you want to do in your loop\n    # increment i\n    return [tf.add(i, 1)]\n\n# do the loop:\nr = tf.while_loop(while_condition, body, [i])\n<\/code><\/pre>\n"},{"AnswerId":"35350571","CreationDate":"2016-02-11T21:38:33.323","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>The type of the return value of TensorFlow Python API functions, including <code>tf.range<\/code> is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L114\" rel=\"nofollow\"><code>Tensor<\/code><\/a>. A <code>Tensor<\/code> is a symbolic handle to node in a graph that represents computation. You perform the actual computation by calling the <code>eval<\/code> method on a <code>Tensor<\/code>, or by passing the object to <code>run<\/code> method of a <code>Session<\/code>. In your case, perhaps what you intended to do was simply iterate over <code>numpy<\/code>'s <code>range<\/code>.<\/p>\n\n<pre><code>for in in np.range(...):\n  # do something\n<\/code><\/pre>\n"}]}
{"QuestionId":35332495,"AnswerCount":3,"Tags":"<neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-02-11T06:42:52.160","AcceptedAnswerId":null,"OwnerUserId":1242108.0,"Title":"Tensorflow tf.train.Saver saves suspiciously large .ckpt files?","Body":"<p>I'm working with a reasonably sized net (1 convolutional layer, 2 fully connected layers). Every time I save variables using , the  files are half a gigabyte each of disk space (512 MB to be exact). Is this normal? I have a Caffe net with the same architecture that requires only a 7MB  file. Is there a particular reason why Tensorflow saves such large file sizes?<\/p>\n\n<p>Many thanks.<\/p>\n","answers":[{"AnswerId":"43569653","CreationDate":"2017-04-23T09:54:45.790","ParentId":null,"OwnerUserId":"6464677","Title":null,"Body":"<p>Maybe you can try (in Tensorflow 1.0):<\/p>\n\n<pre><code>saver.save(sess, filename, write_meta_graph=False)\n<\/code><\/pre>\n\n<p>which doesn't save meta Graph information.\nSee:\n <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/train\/Saver<\/a>\n <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph<\/a><\/p>\n"},{"AnswerId":"44741562","CreationDate":"2017-06-24T22:36:14.220","ParentId":null,"OwnerUserId":"852592","Title":null,"Body":"<p>Typically you only save <code>tf.global_variables()<\/code> (which is shorthand for <code>tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)<\/code>, i.e. the collection of global variables). This collection is meant to include variables which are necessary for restoring the state of the model, so things like current moving averages for batch normalization, the global step, the states of the optimizer(s) and, of course, the <code>tf.GraphKeys.TRAINABLE_VARIABLES<\/code> collection. Variables of more temporary nature, such as the gradients, are collected in <code>LOCAL_VARIABLES<\/code> and it is usually not necessary to store them and they might take up a lot of disk space.<\/p>\n"},{"AnswerId":"35375626","CreationDate":"2016-02-13T03:10:23.380","ParentId":null,"OwnerUserId":"5693571","Title":null,"Body":"<p>Hard to tell how large your net is from what you've described -- the number of connections between two fully connected layers scales up quadratically with the size of each layer, so perhaps your net is quite large depending on the size of your fully connected layers.<\/p>\n\n<p>If you'd like to save space in the checkpoint files, you could replace this line:<\/p>\n\n<p><code>saver = tf.train.Saver()<\/code><\/p>\n\n<p>with the following:<\/p>\n\n<p><code>saver = tf.train.Saver(tf.trainable_variables())<\/code><\/p>\n\n<p>By default, <code>tf.train.Saver()<\/code> saves all variables in your graph -- including the variables created by your optimizer to accumulate gradient information. Telling it to save only <em>trainable<\/em> variables means it will save only the weights and biases of your network, and discard the accumulated optimizer state. Your checkpoints will probably be a lot smaller, with the tradeoff that it you may experience slower training for the first few training batches after you resume training, while the optimizer re-accumulates gradient information. It doesn't take long at all to get back up to speed, in my experience, so personally, I think the tradeoff is worth it for the smaller checkpoints.<\/p>\n"}]}
{"QuestionId":35336648,"AnswerCount":6,"Tags":"<python><tensorflow>","CreationDate":"2016-02-11T10:24:57.010","AcceptedAnswerId":35337827.0,"OwnerUserId":2372357.0,"Title":"List of tensor names in graph in Tensorflow","Body":"<p>The graph object in Tensorflow has a method called \"get_tensor_by_name(name)\". Is there anyway to get a list of valid tensor names?<\/p>\n\n<p>If not, does anyone know the valid names for the pretrained model inception-v3 <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/tutorials\/image_recognition\/index.html\">from here<\/a>? From their example, pool_3, is one valid tensor but a list of all of them would be nice. I looked at <a href=\"http:\/\/arxiv.org\/abs\/1512.00567\">the paper referred to<\/a> and some of the layers seems to correspond to the sizes in table 1 but not all of them.<\/p>\n","answers":[{"AnswerId":"42037139","CreationDate":"2017-02-04T06:09:27.910","ParentId":null,"OwnerUserId":"5861256","Title":null,"Body":"<p>To see the operations in the graph (You will see many, so to cut short I have given here only the first string).<\/p>\n\n<pre><code>sess = tf.Session()\nop = sess.graph.get_operations()\n[m.values() for m in op][1]\n\nout:\n(&lt;tf.Tensor 'conv1\/weights:0' shape=(4, 4, 3, 32) dtype=float32_ref&gt;,)\n<\/code><\/pre>\n"},{"AnswerId":"50620593","CreationDate":"2018-05-31T09:08:48.203","ParentId":null,"OwnerUserId":"2091339","Title":null,"Body":"<p>The above answers are correct. I came across an easy to understand \/ simple code for the above task. So sharing it here :-<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef printTensors(pb_file):\n\n    # read pb into graph_def\n    with tf.gfile.GFile(pb_file, \"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    # import graph_def\n    with tf.Graph().as_default() as graph:\n        tf.import_graph_def(graph_def)\n\n    # print operations\n    for op in graph.get_operations():\n        print(op.name)\n\n\nprintTensors(\"path-to-my-pbfile.pb\")\n<\/code><\/pre>\n"},{"AnswerId":"58846046","CreationDate":"2019-11-13T21:40:34.473","ParentId":null,"OwnerUserId":"196732","Title":null,"Body":"<p><code>saved_model_cli<\/code> is An alternative command line tool comes with TF that <em>might<\/em> be useful if your dealing with the \"SavedModel\" format. From the <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#the_savedmodel_format_on_disk\" rel=\"nofollow noreferrer\">docs<\/a><\/p>\n\n<pre><code>!saved_model_cli show --dir \/tmp\/mobilenet\/1 --tag_set serve --all\n<\/code><\/pre>\n\n<p>This output might be useful, something like:<\/p>\n\n<pre><code>MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:\n\nsignature_def['__saved_model_init_op']:\n  The given SavedModel SignatureDef contains the following input(s):\n  The given SavedModel SignatureDef contains the following output(s):\n    outputs['__saved_model_init_op'] tensor_info:\n        dtype: DT_INVALID\n        shape: unknown_rank\n        name: NoOp\n  Method name is: \n\nsignature_def['serving_default']:\n  The given SavedModel SignatureDef contains the following input(s):\n    inputs['dense_input'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 1280)\n        name: serving_default_dense_input:0\n  The given SavedModel SignatureDef contains the following output(s):\n    outputs['dense_1'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 1)\n        name: StatefulPartitionedCall:0\n  Method name is: tensorflow\/serving\/predict\n<\/code><\/pre>\n"},{"AnswerId":"35337827","CreationDate":"2016-02-11T11:15:22.920","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>The paper is not accurately reflecting the model. If you download the source from arxiv it has an accurate model description as model.txt, and the names in there correlate strongly with the names in the released model.<\/p>\n\n<p>To answer your first question, <code>sess.graph.get_operations()<\/code> gives you a list of operations. For an op, <code>op.name<\/code> gives you the name and <code>op.values()<\/code> gives you a list of tensors it produces (in the inception-v3 model, all tensor names are the op name with a \":0\" appended to it, so <code>pool_3:0<\/code> is the tensor produced by the final pooling op.)<\/p>\n"},{"AnswerId":"49040588","CreationDate":"2018-02-28T23:58:04.763","ParentId":null,"OwnerUserId":"209246","Title":null,"Body":"<p>As a nested list comprehension:<\/p>\n\n<pre><code>tensor_names = [t.name for op in tf.get_default_graph().get_operations() for t in op.values()]\n<\/code><\/pre>\n\n<p>Function to get names of Tensors in a graph (defaults to default graph):<\/p>\n\n<pre><code>def get_names(graph=tf.get_default_graph()):\n    return [t.name for op in graph.get_operations() for t in op.values()]\n<\/code><\/pre>\n\n<p>Function to get Tensors in a graph (defaults to default graph):<\/p>\n\n<pre><code>def get_tensors(graph=tf.get_default_graph()):\n    return [t for op in graph.get_operations() for t in op.values()]\n<\/code><\/pre>\n"},{"AnswerId":"43703647","CreationDate":"2017-04-30T06:33:23.993","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You do not even have to create a session to see the names of all operation names in the graph. To do this you just need to grab a default graph <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_default_graph\" rel=\"nofollow noreferrer\"><code>tf.get_default_graph()<\/code><\/a> and extract all the operations: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/framework\/core_graph_data_structures#Graph\" rel=\"nofollow noreferrer\"><code>.get_operations<\/code><\/a>. Each operation has <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/framework\/core_graph_data_structures#Operation\" rel=\"nofollow noreferrer\">many fields<\/a>, the one you need is name.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.Variable(5)\nb = tf.Variable(6)\nc = tf.Variable(7)\nd = (a + b) * c\n\nfor i in tf.get_default_graph().get_operations():\n    print i.name\n<\/code><\/pre>\n"}]}
{"QuestionId":35339482,"AnswerCount":1,"Tags":"<slice><tensorflow>","CreationDate":"2016-02-11T12:34:41.500","AcceptedAnswerId":35343131.0,"OwnerUserId":3285421.0,"Title":"How to slice a rank 4 tensor in TensorFlow?","Body":"<p>I am trying to slice a four-dimensional tensor using the  operator, as follows:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I try to run this code, I get the following exception:<\/p>\n\n<pre><\/pre>\n\n<p>How can I slice this tensor?<\/p>\n","answers":[{"AnswerId":"35343131","CreationDate":"2016-02-11T15:18:01.103","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice\" rel=\"nofollow noreferrer\"><code>tf.slice(input, begin, size)<\/code><\/a> operator requires that the <code>begin<\/code> and <code>size<\/code> vectors&mdash;which define the subtensor to be sliced&mdash;have the same length as the number of dimensions in <code>input<\/code>. Therefore, to slice a 4-D tensor, you must pass a vector (or list) of four numbers as the second and third arguments of <code>tf.slice()<\/code>.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>x_image = tf.reshape(x, [-1, 28, 28, 1], name='Images_2D')\n\nslice_im = tf.slice(x_image, [0, 2, 2, 0], [1, 24, 24, 1])\n\n# Or, using the indexing operator:\nslice_im = x_image[0:1, 2:26, 2:26, :]\n<\/code><\/pre>\n\n<p>The indexing operator is slightly more powerful, as it can also reduce the rank of the output, if for a dimension you specify a single integer, rather than a range:<\/p>\n\n<pre><code>slice_im = x_image[0:1, 2:26, 2:26, :]\nprint slice_im_2d.get_shape()  # ==&gt; [1, 24, 24, 1]\n\nslice_im_2d = x_image[0, 2:26, 2:26, 0]\nprint slice_im_2d.get_shape()  # ==&gt; [24, 24]\n<\/code><\/pre>\n"}]}
{"QuestionId":35343002,"AnswerCount":1,"Tags":"<neural-network><theano><tensorflow>","CreationDate":"2016-02-11T15:11:37.453","AcceptedAnswerId":35344963.0,"OwnerUserId":2680707.0,"Title":"TensorFlow or Theano: how do they know the loss function derivative based on the neural network graph?","Body":"<p>In TensorFlow or Theano, you only tell the library how your neural network is, and how feed-forward should operate.<\/p>\n\n<p>For instance, in TensorFlow, you would write:<\/p>\n\n<pre><\/pre>\n\n<p>I am using L2-norm loss function, C=0.5*sum((y-yp)^2), and in the backpropagation step presumably the derivative will have to be computed, dC=sum(y-yp). See <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap2.html\" rel=\"noreferrer\">(30) in this book<\/a>.<\/p>\n\n<p>My question is: how can TensorFlow (or Theano) know the analytical derivative for backpropagation? Or do they do an approximation? Or somehow do not use the derivative?<\/p>\n\n<p>I have done the <a href=\"https:\/\/www.udacity.com\/course\/deep-learning--ud730\" rel=\"noreferrer\">deep learning udacity course<\/a> on TensorFlow, but I am still at odds at how to make sense on how these libraries work.<\/p>\n","answers":[{"AnswerId":"35344963","CreationDate":"2016-02-11T16:37:15.167","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The differentiation happens in the final line:<\/p>\n\n<pre><code>    optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)\n<\/code><\/pre>\n\n<p>When you execute the <code>minimize()<\/code> method, TensorFlow identifies the set of variables on which <code>loss<\/code> depends, and computes gradients for each of these. The differentiation is implemented in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/gradients.py\"><code>ops\/gradients.py<\/code><\/a>, and it uses <a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation#Reverse_accumulation\">\"reverse accumulation\"<\/a>. Essentially it searches backwards from the <code>loss<\/code> tensor to the variables, applying the chain rule at each operator in the dataflow graph. TensorFlow includes \"gradient functions\" for most (differentiable) operators, and you can see an example of how these are implemented in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/math_grad.py\"><code>ops\/math_grad.py<\/code><\/a>. A gradient function can use the original op (including its inputs, outputs, and attributes) and the gradients computed for each of its outputs to produce gradients for each of its inputs.<\/p>\n\n<p>Page 7 of <a href=\"http:\/\/www.cs.utoronto.ca\/~ilya\/pubs\/ilya_sutskever_phd_thesis.pdf\">Ilya Sutskever's PhD thesis<\/a> has a nice explanation of how this process works in general.<\/p>\n"}]}
{"QuestionId":35345283,"AnswerCount":1,"Tags":"<python><multithreading><queue><race-condition><tensorflow>","CreationDate":"2016-02-11T16:50:59.267","AcceptedAnswerId":35349023.0,"OwnerUserId":2206976.0,"Title":"Tensorflow Race conditions when chaining multiple queues","Body":"<p>I'd like to compute the mean of each of the RGB channels of a set of images in a multithreaded manner.<\/p>\n\n<p>My idea was to have a  that fills a  and then have a second  that is filled by  threads that load images from the filenames in , perform some ops on them and then enqueue the result.<\/p>\n\n<p>This second queue is then accessed by one single thread (the main thread) that sums up all the values from the queue.<\/p>\n\n<p>This is the code I have:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is each time I'm running this function I get different results, for example:<\/p>\n\n<pre><\/pre>\n\n<p>I blame this to race conditions. But where do those race conditions come from? Can someone help me out?<\/p>\n","answers":[{"AnswerId":"35349023","CreationDate":"2016-02-11T20:05:57.363","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Your <code>QueueRunner<\/code> will start <code>num_threads<\/code> threads which will race to access your <code>reader<\/code> and push the result onto the queue. The order of images on <code>queue<\/code> will vary depending on which thread finishes first.<\/p>\n\n<p><strong>Update Feb 12<\/strong><\/p>\n\n<p>A simple example of chaining two queues, and summing up values from the second queue. When using <code>num_threads<\/code> > 1, there's some non-determinism in the intermediate values, but the final value will always be <code>30<\/code>. When <code>num_threads=1<\/code>, everything is deterministic<\/p>\n\n<pre><code>tf.reset_default_graph()\n\nqueue_dtype = np.int32\n\n# values_queue is a queue that will be filled with 0,1,2,3,4\n# range_input_producer creates the queue and registers its queue_runner\nvalue_queue = tf.range_input_producer(limit=5, num_epochs=1, shuffle=False)\nvalue = value_queue.dequeue()\n\n# value_squared_queue will be filled with 0,1,4,9,16\nvalue_squared_queue = tf.FIFOQueue(capacity=50, dtypes=queue_dtype)\nvalue_squared_enqueue = value_squared_queue.enqueue(tf.square(value))\nvalue_squared = value_squared_queue.dequeue()\n\n# value_squared_sum keeps running sum of squares of values \nvalue_squared_sum = tf.Variable(0)\nvalue_squared_sum_update = value_squared_sum.assign_add(value_squared)\n\n# register queue_runner in the global queue runners collection\nnum_threads = 2\nqr = tf.train.QueueRunner(value_squared_queue, [value_squared_enqueue] * num_threads)\ntf.train.queue_runner.add_queue_runner(qr)\n\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\ntf.start_queue_runners()\n\nfor i in range(5):\n  sess.run([value_squared_sum_update])\n  print sess.run([value_squared_sum])\n<\/code><\/pre>\n\n<p>You should see:<\/p>\n\n<pre><code>[0]\n[1]\n[5]\n[14]\n[30]\n<\/code><\/pre>\n\n<p>Or sometimes (when the order of first 2 values is flipped)<\/p>\n\n<pre><code>[1]\n[1]\n[5]\n[14]\n[30]\n<\/code><\/pre>\n"}]}
{"QuestionId":35345875,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-02-11T17:17:40.573","AcceptedAnswerId":null,"OwnerUserId":3631541.0,"Title":"Running Tensorflow on GeForce 940M (Ubuntu)","Body":"<p>I'm running the CIFAR-10 classification from the Tensorflow for the very first time on my laptop with GeForce 940M. I'm running the training with the pre-defined parameters as follows:<\/p>\n\n<pre><\/pre>\n\n<p>after step 1800 I'm getting the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody have any idea?<\/p>\n\n<p>Thanks a lot in advance for your help! Any advice is kindly appreciated!<\/p>\n","answers":[]}
{"QuestionId":35346117,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2016-02-11T17:29:56.377","AcceptedAnswerId":null,"OwnerUserId":2576346.0,"Title":"TensorFlow C++, runtime issue","Body":"<p>I have created a model using TensorFlow python. Now I have loaded the model in C++ and run it using session->Run (feeding input tensors). The code is complied and linked well, however as soon as it reaches the session->Run during the run time, It doesn't proceed further, and it doesn't create any error message either! It looks like it runs forever! I also checked the CPU usage, but it does not also indicate any intensive computation!<\/p>\n","answers":[{"AnswerId":"35346656","CreationDate":"2016-02-11T17:56:44.017","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>When a TensorFlow program blocks forever, one common issue is that the step is blocked on an empty queue in a <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#QueueBase.dequeue\" rel=\"nofollow\"><code>q.dequeue()<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\" rel=\"nofollow\"><code>q.dequeue_many()<\/code><\/a> op.<\/p>\n\n<p>One possibility is that your Python model depends on prefetching threads (derived from <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/train.html#QueueRunner\" rel=\"nofollow\"><code>tf.train.QueueRunner<\/code><\/a> objects). Many of the input reading  (e.g. using <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#TFRecordReader\" rel=\"nofollow\"><code>tf.TFRecordReader<\/code><\/a>), and batching (e.g. using <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#batch\" rel=\"nofollow\"><code>tf.train.batch()<\/code><\/a>) pipelines implicitly create queues and queue runners.<\/p>\n\n<p>If you had to run <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/train.html#start_queue_runners\" rel=\"nofollow\"><code>tf.train.start_queue_runners()<\/code><\/a> in your Python program, then you will need to do the equivalent thing in your C++ code, by forking threads to run the appropriate <code>q.enqueue()<\/code> ops. Alternatively, you can prepare the input in your C++ program and feed the graph so that the operations you are trying to run do not depend on dequeuing elements for a queue.<\/p>\n"}]}
{"QuestionId":35346599,"AnswerCount":1,"Tags":"<tensorflow><deconvolution>","CreationDate":"2016-02-11T17:53:41.530","AcceptedAnswerId":null,"OwnerUserId":3285421.0,"Title":"How can I give a variable batch_dim in output_shape argument of deconv2d in tensorflow?","Body":"<p>I am trying to use the  op on a variable-sized batch of data. However, it appears that I need to set the  argument as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Why does  take a fixed ? Is there any way to specify a variable batch dimension? What happens if the input batch size varies?<\/p>\n","answers":[{"AnswerId":"35347928","CreationDate":"2016-02-11T19:07:27.633","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>N.B.<\/strong> <code>tf.nn.deconv2d()<\/code> will be called <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#conv2d_transpose\" rel=\"noreferrer\"><code>tf.nn.conv2d_transpose()<\/code><\/a> in the next release of TensorFlow (0.7.0).<\/p>\n\n<p>The <code>output_shape<\/code> argument to <code>tf.nn.deconv2d()<\/code> accepts a computed <code>Tensor<\/code> as its value, which enables you specify a dynamic shape. For example, let's say your input is defined as follows:<\/p>\n\n<pre><code># N.B. Other dimensions are chosen arbitrarily.\ninput = tf.placeholder(tf.float32, [None, 24, 24, 5])\n<\/code><\/pre>\n\n<p>...then the batch size for a particular step can be computed at runtime:<\/p>\n\n<pre><code>batch_size = tf.shape(input)[0]\n<\/code><\/pre>\n\n<p>With this value, you can then build the <code>output_shape<\/code> argument to <code>tf.nn.deconv2d()<\/code> using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#pack\" rel=\"noreferrer\"><code>tf.pack()<\/code><\/a>:<\/p>\n\n<pre><code>output_shape = tf.pack([batch_size, 24, 24, 5])\n\nresult = tf.nn.deconv2d(..., filter, output_shape=output_shape,\n                        strides=[1, 2, 2, 1], padding='SAME')\n<\/code><\/pre>\n"}]}
{"QuestionId":35348929,"AnswerCount":1,"Tags":"<python><slice><theano>","CreationDate":"2016-02-11T20:00:23.767","AcceptedAnswerId":35349401.0,"OwnerUserId":4858356.0,"Title":"Slicing in Theano converts a matrix into a vector","Body":"<p>Consider the following code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>If you run it, you get the following output: <\/p>\n\n<pre><\/pre>\n\n<p>Apparently indexing with an array of length 1 turns xsub2 into a col instead of a matrix. How can i make xsub2 be a matrix?<\/p>\n","answers":[{"AnswerId":"35349401","CreationDate":"2016-02-11T20:29:08.243","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>A <code>col<\/code> or \"column vector\" is the name Theano uses for a symbolic matrix that it knows contains only one column. It should be possible to use it just like a matrix.<\/p>\n\n<p>Theano often doesn't know the shape of a particular symbolic tensor, only its dimensionality. However, in some circumstances, such as that given in the question, Theano is able to infer that a tensor has a particular special case of shape and can sometimes use this information to optimize the computation. This is why <code>col<\/code> (and <code>row<\/code>) exist as special cases of <code>matrix<\/code>.<\/p>\n\n<p>If you think about the shape more than the type then you'll see that Theano is behaving just the same as numpy:<\/p>\n\n<pre><code>import theano\nimport theano.tensor\nimport numpy as np\n\n\ndef compute(minibatch):\n    xflat = minibatch.reshape((minibatch.shape[0], -1))\n    partition = np.array([1, 2, 3])\n    xsub1 = xflat[:, partition]\n    partition = np.array([1])\n    xsub2 = xflat[:, partition]\n    return xsub1, xsub2\n\n\ndef compile_theano_version():\n    minibatch = theano.tensor.tensor4(name='minibatch', dtype=theano.config.floatX)\n    xsub1, xsub2 = compute(minibatch)\n    print xsub1.type, xsub2.type\n    return theano.function([minibatch], [xsub1, xsub2])\n\n\ndef numpy_version(minibatch):\n    return compute(minibatch)\n\n\ndef main():\n    batch_shape = (50, 40, 30, 30)\n    minibatch = np.random.standard_normal(size=batch_shape).astype(theano.config.floatX)\n\n    xsub1, xsub2 = numpy_version(minibatch)\n    print xsub1.shape, xsub2.shape\n\n    theano_version = compile_theano_version()\n    xsub1, xsub2 = theano_version(minibatch)\n    print xsub1.shape, xsub2.shape\n\n\nmain()\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>(50L, 3L) (50L, 1L)\nTensorType(float64, matrix) TensorType(float64, col)\n(50L, 3L) (50L, 1L)\n<\/code><\/pre>\n\n<p>So a <code>col<\/code> is indeed a matrix with one column and <em>not<\/em> a vector.<\/p>\n"}]}
{"QuestionId":35350699,"AnswerCount":2,"Tags":"<floating-point><tensorflow>","CreationDate":"2016-02-11T21:46:14.140","AcceptedAnswerId":null,"OwnerUserId":16480.0,"Title":"Fast, portable conversion of 24 bit ints to float without losing a bit","Body":"<p>When generating random uniform floats in , TensorFlow currently uses <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/ed6f783690a35e54441881930d4291894bc03285\/tensorflow\/core\/lib\/random\/random_distributions.h#L402\" rel=\"nofollow\">bit twiddling<\/a> to convert 23 bit integers to floats in , then subtracts one:<\/p>\n\n<pre><\/pre>\n\n<p>This irritates me, since subtracting one means we only get 23 bits of precision, rather than the 24 bits available.  Unfortunately, the naive algorithm is about 9% slower on CPUs (it's the same speed on GPU):<\/p>\n\n<pre><\/pre>\n\n<p>I also tried explicitly truncating to 24 bits in case it would teach the compiler that rounding mode flags didn't matter; this didn't solve the problem:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to get the full 24 bits of available precision without sacrificing performance?  I'm pretty sure I could do it in assembly, but portability is required.<\/p>\n\n<p>Note that the remaining 8 bits of precision that one can get <em>sometimes<\/em> for small floats are not interesting: I only care about the one missing bit.<\/p>\n","answers":[{"AnswerId":"35351145","CreationDate":"2016-02-11T22:13:59.760","ParentId":null,"OwnerUserId":"139746","Title":null,"Body":"<p>You can try not doing the subtraction when the 24th bit is set:<\/p>\n\n<pre><code>  \u2026\n  const uint32 exp = static_cast&lt;uint32&gt;(126); \/\/ 0.5\n  \u2026\n  if ((x &amp; 0x800000) == 0) result -= 0.5f;\n  return result;\n}\n<\/code><\/pre>\n\n<p>However, just 9% penalty for the 24th bit is already pretty good, and this will not necessarily be faster. (Here you are sometimes avoiding the price of a subtraction, but always paying the price of a test and conditional branch. I will let you do the timings: the 0x800000 mask can be done in parallel with the rest but the cost of the conditional branch  entirely depends on the distribution of the values in practice.)<\/p>\n\n<p>This can easily be made branchless for GPUs by always doing the subtraction followed by a conditional move, but the compiler should do this automatically.<\/p>\n"},{"AnswerId":"35357803","CreationDate":"2016-02-12T08:21:23.720","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can use <code>__builtin_clz<\/code> to directly adjust the exponent and map the remainder of the number as the mantissa, avoiding the floating point subtraction and the loss of precision:<\/p>\n\n<pre><code>float Uint32ToFloat(uint32_t x) {\n  \/\/ IEEE754 floats are formatted as follows (MSB first):\n  \/\/    sign(1) exponent(8) mantissa(23)\n  \/\/ Conceptually construct the following:\n  \/\/    sign == 0\n  \/\/    exponent == 126  -- an excess 127 representation of a -1 exponent\n  \/\/    mantissa == 23 random bits\n  uint32_t exp = static_cast&lt;uint32_t&gt;(126);\n\n  auto lz = __builtin_clz(x);\n  exp -= lz;\n  x &lt;&lt;= (lz+1);  \/\/ +1 to chop off implicit 1 in FP representation.\n  const uint32_t man = x &gt;&gt; 9;  \/\/ 23 bit mantissa.\n  const uint32_t val = (exp &lt;&lt; 23) | man;\n\n  \/\/ Assumes that endian-ness is same for float and uint32.\n  float result;\n  memcpy(&amp;result, &amp;val, sizeof(val));\n  return result;\n}\n<\/code><\/pre>\n\n<p>Note that the CUDA equivalent of gcc's <code>__builtin_clz<\/code> is <code>__clz()<\/code>.<\/p>\n\n<p>Advantages:  Preserves as much precision as possible from the original random number.<\/p>\n\n<p>Drawback:  I think the original version vectorizes better and has a bit less instruction latency.<\/p>\n\n<p>A third alternative is to directly adjust the exponent after letting the FP hardware do the conversion from the integer:<\/p>\n\n<pre><code>inline float Uint32ToFloat_bit(uint32_t x) {\n  float f(x);\n  uint32_t f_as_int;\n  memcpy(&amp;f_as_int, &amp;f, sizeof(f_as_int));\n  f_as_int -= (32 &lt;&lt; 23);  \/\/ Subtract 32 from the exponent.\n  float result;\n  memcpy(&amp;result, &amp;f_as_int, sizeof(f_as_int));\n  return result;\n}\n<\/code><\/pre>\n\n<p>This was faster than the builtin_clz version for me, but slower than your basic one, but I again suspect it's very context-dependent.  This one might vectorize quite well - but so does your very basic one, since it's just a <code>vmulss<\/code>.<\/p>\n\n<p>After having gone through all of this, I think that the best step would be a timing harness that produces a batch of 8 random numbers and then bulk converts them, to let the compiler vectorize the conversions, and then see which is best.<\/p>\n"}]}
{"QuestionId":35351760,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-02-11T22:57:05.270","AcceptedAnswerId":35352156.0,"OwnerUserId":5766785.0,"Title":"TF save\/restore graph fails at tf.GraphDef.ParseFromString()","Body":"<p>Based on this <a href=\"https:\/\/stackoverflow.com\/questions\/35247406\/converting-trained-tensorflow-model-to-protobuf\">converting-trained-tensorflow-model-to-protobuf<\/a> I am trying to save\/restore TF graph without success. <\/p>\n\n<p>Here is saver:  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>which produces  with text graph description. Pretty much copy paste from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/00440e99ffb1ed1cfe4b4ea650e0c560838a6edc\/tensorflow\/python\/tools\/freeze_graph_test.py#L79\" rel=\"noreferrer\">freeze_graph_test.py<\/a>.<\/p>\n\n<p>Here is reader:  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>which fails at  with <\/p>\n\n<p>I am on docker container  in case it makes any difference. <\/p>\n","answers":[{"AnswerId":"43942741","CreationDate":"2017-05-12T16:44:58.010","ParentId":null,"OwnerUserId":"3387598","Title":null,"Body":"<p>I tried to load model via the java API which accepts only binary. But in the python where we uses contrib.Estimator produces textual model file format.\nI found a <a href=\"https:\/\/gist.github.com\/Arafatk\/c063bddb9b8d17a037695d748db4f592\" rel=\"nofollow noreferrer\">model file converter<\/a> online, seems it's working fine.\nThis might solve the original issue too (use the binary model loader) if you have an existing text format model file. <\/p>\n"},{"AnswerId":"35352131","CreationDate":"2016-02-11T23:28:55.603","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><code>ParseFromString<\/code> needs binary serialized protocol buffer, for human-readable representation you need to use <code>text_format.Merge<\/code> as used <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/test_util_test.py#L84\" rel=\"nofollow\">here<\/a><\/p>\n"},{"AnswerId":"35352156","CreationDate":"2016-02-11T23:31:10.427","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>GraphDef.ParseFromString()<\/code> method (and, in general, the <code>ParseFromString()<\/code> method on any Python protobuf wrapper) expects a string in the binary protocol buffer format. If you pass <code>as_text=False<\/code> to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#write_graph\"><code>tf.train.write_graph()<\/code><\/a>, then the file will be in the appropriate format.<\/p>\n\n<p>Otherwise you can do the following to read the text-based format:<\/p>\n\n<pre><code>from google.protobuf import text_format\n# ...\ngraph_def = tf.GraphDef()\ntext_format.Merge(proto_b, graph_def) \n<\/code><\/pre>\n"}]}
{"QuestionId":35351801,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-11T23:01:14.623","AcceptedAnswerId":35352250.0,"OwnerUserId":964199.0,"Title":"how to set test iteration variable in tensorflow","Body":"<p>Is test_iter variable in BVLC and Nvidia's caffe same as tensorflow's NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN in the cifar example code? If not, where can I set test_iter and test_interval (as used in Caffe)?<\/p>\n","answers":[{"AnswerId":"35352250","CreationDate":"2016-02-11T23:39:42.413","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7718d2a052042d6ee1d821c979661371e217f0d7\/tensorflow\/models\/image\/cifar10\/cifar10_input.py#L34\" rel=\"nofollow\"><code>NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN<\/code><\/a> constant represents the number of images in the <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow\">training set for CIFAR-10 training<\/a>.<\/p>\n\n<p>In Caffe, <code>test_iter<\/code> is the number of evaluation iterations to perform. \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7718d2a052042d6ee1d821c979661371e217f0d7\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py\" rel=\"nofollow\">TensorFlow's CIFAR-10 eval<\/a> performs evaluation on the complete set of 10000 images, does not have a quantity analogous to <code>test_iter<\/code>.<\/p>\n"}]}
{"QuestionId":35352787,"AnswerCount":1,"Tags":"<python><machine-learning><theano><pymc3>","CreationDate":"2016-02-12T00:34:20.137","AcceptedAnswerId":null,"OwnerUserId":1237998.0,"Title":"Theano Scan and Repeat","Body":"<p>In Theano, there is an option to use the repeat function  and supply a pair of vectors, such that each element of  is repeated  times.<\/p>\n\n<p>Unfortunately, this operation has no defined gradient (it throws a notimplemented exception) which is a problem, as I'm trying to use this with 's gradient based samplers. <\/p>\n\n<p>I think I can address this using the  function and calling repeat recursively for each element of the two vectors, however my code isn't working, probably because I'm calling  incorrectly. Can anyone help me understand why the below code isn't working?<\/p>\n\n<pre><\/pre>\n\n<p>I'd expect this to return [1,1,1,1] but instead it returns the below error. <\/p>\n\n<pre><\/pre>\n\n<p>I've raised an <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/issues\/960\" rel=\"nofollow\">issue<\/a> on the Pymc3 github to see if this is something that should be fixed more permanently, but I figure its a good opportunity to learn more about Theano for me anyway, and if I can resolve the problem maybe I can contribute back to the project.<\/p>\n","answers":[{"AnswerId":"36378957","CreationDate":"2016-04-02T21:31:41.920","ParentId":null,"OwnerUserId":"1165170","Title":null,"Body":"<p>I see here two things:<\/p>\n\n<ol>\n<li>Bad ordering in lambda expression: it should be A, B, prior_result (now B is treated as outputs_info)<\/li>\n<li>shape of A.repeat(B) is different from shape of prior_result (at this stage of compilation) <\/li>\n<\/ol>\n\n<p>quick fix: just remove outputs_info from scan's arguments (and prior_result from lambda) and you will get [1,1,1,1].<\/p>\n"}]}
{"QuestionId":35355528,"AnswerCount":1,"Tags":"<deep-learning><keras>","CreationDate":"2016-02-12T05:37:19.457","AcceptedAnswerId":35361418.0,"OwnerUserId":1472128.0,"Title":"How to implement a Mean Pooling layer in Keras","Body":"<p>There seems no built-in support for Mean Pooling layer for RNN in Keras. Anyone knows how to wrap one?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vApwo.png\"><img src=\"https:\/\/i.stack.imgur.com\/vApwo.png\" alt=\"enter image description here\"><\/a>\n<a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\">http:\/\/deeplearning.net\/tutorial\/lstm.html<\/a><\/p>\n","answers":[{"AnswerId":"35361418","CreationDate":"2016-02-12T11:25:37.737","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>Keras has a layer <code>AveragePooling1D<\/code> for that. If you use the graph API, you should be able to do something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add_node(AveragePooling1D(...), \n               inputs=['h0', 'h1', ..., 'hn'], \n               merge_mode='concat', ...)\n<\/code><\/pre>\n"}]}
{"QuestionId":35356514,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-12T06:52:51.163","AcceptedAnswerId":35947076.0,"OwnerUserId":964199.0,"Title":"Run Tensorflow with googlenet","Body":"<p>I would like to run the <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_googlenet\" rel=\"nofollow\">googlenet<\/a> with tensorflow. Is there any multi-GPU version available that can be run with tensorflow? <\/p>\n","answers":[{"AnswerId":"36708214","CreationDate":"2016-04-19T03:48:58.590","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Can you also please check out: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception<\/a>? (The image is from the URL.)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/svzPR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/svzPR.png\" alt=\"Img from https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\"><\/a><\/p>\n"},{"AnswerId":"35947076","CreationDate":"2016-03-11T18:10:32.577","ParentId":null,"OwnerUserId":"6051382","Title":null,"Body":"<p>You can train, evaluate and fine-tune an Inception v3 model. See links there-in for pointers.\n<a href=\"http:\/\/googleresearch.blogspot.com\/2016\/03\/train-your-own-image-classifier-with.html\" rel=\"nofollow\">http:\/\/googleresearch.blogspot.com\/2016\/03\/train-your-own-image-classifier-with.html<\/a><\/p>\n"}]}
{"QuestionId":35356608,"AnswerCount":1,"Tags":"<theano><dimension><mismatch><lasagne>","CreationDate":"2016-02-12T06:58:37.890","AcceptedAnswerId":35357456.0,"OwnerUserId":868972.0,"Title":"Input dimension mismatch binary crossentropy Lasagne and Theano","Body":"<p>I read all posts in the net adressing the issue where people forgot to change the target vector to a matrix, and as a problem remains after this change, I decided to ask my question here. Workarounds are mentioned below, but new problems show and I am thankful for suggestions!<\/p>\n\n<p>Using a convolution network setup and binary crossentropy with sigmoid activation function, I get a dimension mismatch problem, but not during the training data, only during validation \/ test data evaluation. For some strange reason, of of my validation set vectors get his dimension switched and I have no idea, why. Training, as mentioned above, works fine. Code follows below, thanks a lot for help (and sorry for hijacking the thread, but I saw no reason for creating a new one), most of it copied from the lasagne tutorial example.<\/p>\n\n<p>Workarounds and new problems:<\/p>\n\n<ol>\n<li>Removing \"axis=1\" in the valAcc definition helps, but validation accuracy remains zero and test classification always returns the same result, no matter how many nodes, layers, filters etc. I have. Even changing training set size (I have around 350 samples for each class with 48x64 grayscale images) does not change this. So something seems off<\/li>\n<\/ol>\n\n<p>Network creation:<\/p>\n\n<pre><\/pre>\n\n<p>Target matrices for all sets are created like this (training target vector as an example)<\/p>\n\n<pre><\/pre>\n\n<p>...and the theano variables as such<\/p>\n\n<pre><\/pre>\n\n<p>Finally, here the two loops, training and test. The first is fine, the second throws the error, excerpts below<\/p>\n\n<pre><\/pre>\n\n<p>Erorr (excerpts)<\/p>\n\n<pre><\/pre>\n\n<p>Again, thanks for help!<\/p>\n","answers":[{"AnswerId":"35357456","CreationDate":"2016-02-12T07:58:22.827","ParentId":null,"OwnerUserId":"5631237","Title":null,"Body":"<p>so it seems the error is in the evaluation of the validation accuracy.\nWhen you remove the \"axis=1\" in your calculation, the argmax goes on everything, returning only a number.\nThen, broadcasting steps in and this is why you would see the same value for the whole set.<\/p>\n\n<p>But from the error you have posted, the \"T.eq\" op throws the error because it has to compare a 52 x 1 with a 1 x 52 vector (matrix for theano\/numpy).\nSo, I suggest you try to replace the line with:<\/p>\n\n<pre><code>    valAcc = T.mean(T.eq(T.argmax(valPrediction, axis=1), targetVar.T))\n<\/code><\/pre>\n\n<p>I hope this should fix the error, but I haven't tested it myself.<\/p>\n\n<p>EDIT:\nThe error lies in the argmax op that is called.\nNormally, the argmax is there to determine which of the output units is activated the most.\nHowever, in your setting you only have one output neuron which means that the argmax over all output neurons will always return 0 (for first arg).<\/p>\n\n<p>This is why you have the impression your network gives you always 0 as output.<\/p>\n\n<p>By replacing:<\/p>\n\n<pre><code>    valAcc = T.mean(T.eq(T.argmax(valPrediction, axis=1), targetVar.T))\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>    binaryPrediction = valPrediction &gt; .5\n    valAcc = T.mean(T.eq(binaryPrediction, targetVar.T)\n<\/code><\/pre>\n\n<p>you should get the desired result.<\/p>\n\n<p>I'm just not sure, if the transpose is still necessary or not.<\/p>\n"}]}
{"QuestionId":35358642,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-02-12T09:10:12.873","AcceptedAnswerId":null,"OwnerUserId":4755260.0,"Title":"Error on the first convolution layer at the point of plus biases","Body":"<p>I'm stuck on the error regarding CNN.<\/p>\n\n<p>Trying to implement the following structure on tensorflow.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/afpCN.png\" alt=\"enter image description here\"><\/p>\n\n<p>I wrote my code as the tutorial (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow noreferrer\">Deep MNIST for Experts<\/a>) shows.Defined weight and bias initialization as below.<\/p>\n\n<pre><\/pre>\n\n<p>And first convolution layer as following. Almost the same as Tutorial.<\/p>\n\n<pre><\/pre>\n\n<p>However, I got error as following. How to solve this error?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35360791","CreationDate":"2016-02-12T10:54:58.303","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>That error is not from the code you showed. It looks like you are creating multiple graphs (maybe because of repeated invocations of an ipython notebook cell?) in  and try to use things from different graphs in the same computation.<\/p>\n"}]}
{"QuestionId":35358786,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2016-02-12T09:19:00.563","AcceptedAnswerId":null,"OwnerUserId":2576346.0,"Title":"Tensorflow c++, issue with tensors and feeding","Body":"<p>I am very new to TensorFlow C++ API and trying to just build a very simple graph in Python and load\/test it in C++ API. Here is the python code to create the graph:<\/p>\n\n<pre><\/pre>\n\n<p>and here is the C code to load and run the graph:<\/p>\n\n<pre><\/pre>\n\n<p>However I get this error message:<\/p>\n\n<blockquote>\n  <p>.\/tensorflow\/core\/framework\/tensor.h:500] Check failed: 1 ==\n  NumElements() (1 vs. 4)Must have a one element tensor<\/p>\n<\/blockquote>\n\n<p>What could be the problem? I noticed if I define my tensors in he python and C++ as [1,1], it goes through without any problem!<\/p>\n","answers":[{"AnswerId":"35937224","CreationDate":"2016-03-11T10:04:57.253","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/medium.com\/@hamedmp\/exporting-trained-tensorflow-models-to-c-the-right-way-cf24b609d183#.t9iq9asfs\" rel=\"nofollow\">here<\/a>, I tried to explain everything completely \n To summarize you should use 'freeze_graph' file to make trained variables constant and then import it in your C++ file. <\/p>\n"}]}
{"QuestionId":35361467,"AnswerCount":10,"Tags":"<tensorflow>","CreationDate":"2016-02-12T11:28:07.753","AcceptedAnswerId":35367161.0,"OwnerUserId":2680707.0,"Title":"TensorFlow: numpy.repeat() alternative","Body":"<p>I want to compare the predicted values  from my neural network in a pairwise fashion, and so I was using (back in my old numpy implementation):<\/p>\n\n<pre><\/pre>\n\n<p>This basically create a indexing mesh which I then use.  while . I do not know if there is a simpler manner of doing it...<\/p>\n\n<p>Anyhow, TensorFlow has a , but it seems to be lacking a .<\/p>\n\n<pre><\/pre>\n\n<p>And I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>It also does not work to use a TensorFlow constant for the indexing:<\/p>\n\n<pre><\/pre>\n\n<p>-<\/p>\n\n<pre><\/pre>\n\n<p>The idea is to convert my <a href=\"http:\/\/research.microsoft.com\/en-us\/um\/people\/cburges\/papers\/ICML_ranking.pdf\" rel=\"noreferrer\">RankNet<\/a> implementation to TensorFlow.<\/p>\n","answers":[{"AnswerId":"35367161","CreationDate":"2016-02-12T16:11:16.123","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can achieve the effect of <code>np.repeat()<\/code> using a combination of <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.2\/api_docs\/python\/tf\/tile\" rel=\"noreferrer\"><code>tf.tile()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.2\/api_docs\/python\/tf\/reshape\" rel=\"noreferrer\"><code>tf.reshape()<\/code><\/a>:<\/p>\n\n<pre><code>idx = tf.range(len(yp))\nidx = tf.reshape(idx, [-1, 1])    # Convert to a len(yp) x 1 matrix.\nidx = tf.tile(idx, [1, len(yp)])  # Create multiple columns.\nidx = tf.reshape(idx, [-1])       # Convert back to a vector.\n<\/code><\/pre>\n\n<p>You can simply compute <code>jdx<\/code> using <code>tf.tile()<\/code>:<\/p>\n\n<pre><code>jdx = tf.range(len(yp))\njdx = tf.tile(jdx, [len(yp)])\n<\/code><\/pre>\n\n<p>For the indexing, you could try using <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.2\/api_docs\/python\/tf\/gather\" rel=\"noreferrer\"><code>tf.gather()<\/code><\/a> to extract non-contiguous slices from the <code>yp<\/code> tensor:<\/p>\n\n<pre><code>s = tf.gather(yp, idx) - tf.gather(yp, jdx)\n<\/code><\/pre>\n"},{"AnswerId":"44527309","CreationDate":"2017-06-13T16:49:31.147","ParentId":null,"OwnerUserId":"3015202","Title":null,"Body":"<p>Just in case anybody is interested for a 2D method to copy the matrices. I think this could work:<\/p>\n\n<pre><code>TF_obj = tf.zeros([128, 128])\ntf.tile(tf.expand_dims(TF_obj, 2), [1, 1, 2])\n<\/code><\/pre>\n"},{"AnswerId":"44490169","CreationDate":"2017-06-12T01:15:41.853","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>It looks like your question is so popular that people <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8246\" rel=\"nofollow noreferrer\">refer it on TF tracker<\/a>. Sadly the same function is not still implemented in TF.<\/p>\n\n<p>You can implement it by combining <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tile\" rel=\"nofollow noreferrer\">tf.tile<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\">tf.reshape<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/squeeze\" rel=\"nofollow noreferrer\">tf.squeeze<\/a>. Here is a way to convert examples from <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.repeat.html\" rel=\"nofollow noreferrer\">np.repeat<\/a>:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nx = [[1,2],[3,4]]\nprint np.repeat(3, 4)\nprint np.repeat(x, 2)\nprint np.repeat(x, 3, axis=1)\n\nx = tf.constant([[1,2],[3,4]])\nwith tf.Session() as sess:\n    print sess.run(tf.tile([3], [4]))\n    print sess.run(tf.squeeze(tf.reshape(tf.tile(tf.reshape(x, (-1, 1)), (1, 2)), (1, -1))))\n    print sess.run(tf.reshape(tf.tile(tf.reshape(x, (-1, 1)), (1, 3)), (2, -1)))\n<\/code><\/pre>\n\n<p>In the last case where repeats are different for each element you most probably will need <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">loops<\/a>.<\/p>\n"},{"AnswerId":"55059587","CreationDate":"2019-03-08T08:49:18.913","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>A <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/ragged\/ragged_util.py#L113\" rel=\"nofollow noreferrer\">relatively fast<\/a> implementation was recently added with <code>RaggedTensor<\/code> utilities from 1.13, but it's not a part of the officially exported API. You can still use it, but there's a chance it might disappear.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python.ops.ragged.ragged_util import repeat\n<\/code><\/pre>\n\n<p>From the source code:<\/p>\n\n<pre><code># This op is intended to exactly match the semantics of numpy.repeat, with\n# one exception: numpy.repeat has special (and somewhat non-intuitive) behavior\n# when axis is not specified.  Rather than implement that special behavior, we\n# simply make `axis` be a required argument.\n<\/code><\/pre>\n"},{"AnswerId":"56388308","CreationDate":"2019-05-31T03:29:34.347","ParentId":null,"OwnerUserId":"4715661","Title":null,"Body":"<p>According to tf api <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/repeat_elements\" rel=\"nofollow noreferrer\">document<\/a>, <code>tf.keras.backend.repeat_elements()<\/code> does the same work with <code>np.repeat()<\/code> . For example,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant([1, 3, 3, 1], dtype=tf.float32)\nrep_x = tf.keras.backend.repeat_elements(x, 5, axis=0)\n# result: [1. 1. 1. 1. 1. 3. 3. 3. 3. 3. 3. 3. 3. 3. 3. 1. 1. 1. 1. 1.]\n<\/code><\/pre>\n"},{"AnswerId":"52450552","CreationDate":"2018-09-21T20:13:25.573","ParentId":null,"OwnerUserId":"2114757","Title":null,"Body":"<p>You can simulate missing <code>tf.repeat<\/code> by <code>tf.stack<\/code>ing the value with itself:<\/p>\n\n<pre><code>value = np.arange(len(yp))  # what to repeat\nrepeat_count = len(yp)      # how many times\nrepeated = tf.stack ([value for i in range(repeat_count)], axis=1)\n<\/code><\/pre>\n\n<p>I advice using this only on small repeat counts.<\/p>\n"},{"AnswerId":"57929586","CreationDate":"2019-09-13T19:49:04.150","ParentId":null,"OwnerUserId":"2082026","Title":null,"Body":"<p>So I have found that tensorflow has one such method to repeat the elements of an array. The method <code>tf.keras.backend.repeat_elements<\/code> is what you are looking for. Anyone who comes at a later point of time can save lot of their efforts.  This <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/repeat_elements\" rel=\"nofollow noreferrer\">link<\/a> offers an explanation to the method and specifically says<\/p>\n\n<blockquote>\n  <p>Repeats the elements of a tensor along an axis, like np.repeat<\/p>\n<\/blockquote>\n\n<p>I have included a very short example which proves that the elements are copied in the exact way as <code>np.repeat<\/code> would do.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\nx = np.random.rand(2,2)\n# print(x)  # uncomment this line to see the array's elements\n\ny = tf.convert_to_tensor(x)\ny = tf.keras.backend.repeat_elements(x, rep=3, axis=0)\n# print(y)  # uncomment this line to see the results\n<\/code><\/pre>\n"},{"AnswerId":"53298183","CreationDate":"2018-11-14T10:36:33.037","ParentId":null,"OwnerUserId":"10651406","Title":null,"Body":"<p>Just for 1-d tensors, I've made this function<\/p>\n\n<pre><code>def tf_repeat(y,repeat_num)   \n        return tf.reshape(tf.tile(tf.expand_dims(y,axis=-1),[1,repeat_num]),[-1]) \n<\/code><\/pre>\n"},{"AnswerId":"52460748","CreationDate":"2018-09-22T20:47:15.750","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<pre><code>import numpy as np\nimport tensorflow as tf\nimport itertools    \n\nx = np.arange(6).reshape(3,2)\nx = tf.convert_to_tensor(x)\nN = 3 # number of repetition\nK = x.shape[0] # for here 3\n\norder = list(range(0, N*K, K))\norder = [[x+i for x in order] for i in range(K)]\norder = list(itertools.chain.from_iterable(order))\nx_rep = tf.gather(tf.tile(x, [N, 1]), order)\n<\/code><\/pre>\n\n<p>Results from:<\/p>\n\n<pre><code>   [0, 1],\n   [2, 3],\n   [4, 5]]\n<\/code><\/pre>\n\n<p>To: <\/p>\n\n<pre><code>  [[0, 1],\n   [0, 1],\n   [0, 1],\n   [2, 3],\n   [2, 3],\n   [2, 3],\n   [4, 5],\n   [4, 5],\n   [4, 5]]\n<\/code><\/pre>\n\n<p>If you want:<\/p>\n\n<pre><code>  [[0, 1],\n   [2, 3],\n   [4, 5],\n   [0, 1],\n   [2, 3],\n   [4, 5],\n   [0, 1],\n   [2, 3],\n   [4, 5]]\n<\/code><\/pre>\n\n<p>Simply use <code>tf.tile(x, [N, 1])<\/code><\/p>\n"},{"AnswerId":"52941809","CreationDate":"2018-10-23T05:37:18.167","ParentId":null,"OwnerUserId":"6642334","Title":null,"Body":"<p>Though many clean and working solutions have been given, they seem to all be based on producing the set of indices from scratch each iteration.   <\/p>\n\n<p>While the cost to produce these node's isn't typically significant during training, it may be significant if using your model for inference.  <\/p>\n\n<p>Repeating tf.range (like your example) has come up a few times so I built the following function creator.  Given the maximum number of times something will be repeated and the maximum number of things that will need repeating, it returns a function which produces the same values as <code>np.repeat(np.arange(len(multiples)), multiples)<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n\ndef numpy_style_repeat_1d_creator(max_multiple=100, max_to_repeat=10000):\n    board_num_lookup_ary = np.repeat(\n        np.arange(max_to_repeat),\n        np.full([max_to_repeat], max_multiple))\n    board_num_lookup_ary = board_num_lookup_ary.reshape(max_to_repeat, max_multiple)\n\n    def fn_to_return(multiples):\n        board_num_lookup_tensor = tf.constant(board_num_lookup_ary, dtype=tf.int32)\n        casted_multiples = tf.cast(multiples, dtype=tf.int32)\n        padded_multiples = tf.pad(\n            casted_multiples,\n            [[0, max_to_repeat - tf.shape(multiples)[0]]])\n\n        return tf.boolean_mask(\n            board_num_lookup_tensor,\n            tf.sequence_mask(padded_multiples, maxlen=max_multiple))\n\n    return fn_to_return\n\n#Here's an example of how it can be used\nwith tf.Session() as sess:\n    repeater = numpy_style_repeat_1d_creator(5,4)\n    multiples = tf.constant([4,1,3])\n\n    repeated_values = repeater(multiples)\n    print(sess.run(repeated_values))\n<\/code><\/pre>\n\n<p>The general idea is to store a repeated tensor and then mask it, but it may help to see it visually (this is for the example given above):<\/p>\n\n<pre><code>In the example above the following Tensor is produced:\n[[0,0,0,0,0],\n [1,1,1,1,1],\n [2,2,2,2,2],\n [3,3,3,3,3]]\n\nFor multiples [4,1,3] it will collect the non-X values:\n[[0,0,0,0,X],\n [1,X,X,X,X],\n [2,2,2,X,X],\n [X,X,X,X,X]]\n\nresulting in:\n[0,0,0,0,1,2,2,2]\n<\/code><\/pre>\n\n<p>tl;dr: To avoid producing the indices each time (can be costly), pre-repeat everything and then mask that tensor each time<\/p>\n"}]}
{"QuestionId":35362184,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2016-02-12T12:03:14.323","AcceptedAnswerId":35362457.0,"OwnerUserId":5324819.0,"Title":"How to duplicate matrix in theano","Body":"<p>Hi I am new to theano and I need [N x M * 4] dot [M x P], so I want second matrix to be duplicated 4 times. <\/p>\n\n<p>Something like <\/p>\n\n<pre><\/pre>\n\n<p>But in my case I want<\/p>\n\n<pre><\/pre>\n\n<p>How could I do that?<\/p>\n","answers":[{"AnswerId":"35362397","CreationDate":"2016-02-12T12:12:59.783","ParentId":null,"OwnerUserId":"5856763","Title":null,"Body":"<p>You may do this using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/reference\/generated\/numpy.tile.html\" rel=\"nofollow\">numpy tile<\/a> function:<\/p>\n\n<pre><code>&gt;&gt;&gt; np.tile(np.array([[1, 2], [3, 4]]), (2, 1))\narray([[1, 2],\n       [3, 4],\n       [1, 2],\n       [3, 4]])\n<\/code><\/pre>\n\n<p>Where (2, 1) is the number of repetitions along each axis. It should be (4, 1) in your case<\/p>\n"},{"AnswerId":"35362457","CreationDate":"2016-02-12T12:16:21.753","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>You want <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.tile\" rel=\"nofollow\"><code>theano.tensor.tile<\/code><\/a>:<\/p>\n\n<pre><code>z2 = T.tile(x, (3, 1))  # repeat 3x in the first dimension, 1x in the second\nbar = theano.function([x], z2)\n\nprint(bar(a))\n# [[ 1.  2.]\n#  [ 3.  4.]\n#  [ 1.  2.]\n#  [ 3.  4.]\n#  [ 1.  2.]\n#  [ 3.  4.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35363698,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-12T13:21:37.030","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Why calling a dequeue_op in the following code throw tensorflow.python.framework.errors.OutOfRangeError","Body":"<p>In the following example I read a data from file, decode it and create a batch with tf.train.shuffle_batch. I start every queue runner threads and then fetch my batch. It throws outOfRange error as if my data queues are empty. I don't understand why since I start every queue runners. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35367258","CreationDate":"2016-02-12T16:15:30.540","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The most common reason for an <em>unexpected<\/em> <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#OutOfRangeError\" rel=\"nofollow\"><code>tf.errors.OutOfRangeError<\/code><\/a> is that one of the operations that enqueues elements to the queue has failed. Check whether one of the ops in your <code>decode_and_transform()<\/code> function is failing, by running the upstream code manually (and without calling <code>tf.train.start_queue_runners()<\/code>:<\/p>\n\n<pre><code># Runs the ops before `tf.train.shuffle_batch()`.\nexample_val, label_val = sess.run([example, label])\n<\/code><\/pre>\n"}]}
{"QuestionId":35365007,"AnswerCount":5,"Tags":"<python><machine-learning><scikit-learn><tensorflow>","CreationDate":"2016-02-12T14:26:17.743","AcceptedAnswerId":43960730.0,"OwnerUserId":4190832.0,"Title":"Tensorflow Precision \/ Recall \/ F1 score and Confusion matrix","Body":"<p>I would like to know if there is a way to implement the different score function from the scikit learn package like this one :<\/p>\n\n<pre><\/pre>\n\n<p>into a tensorflow model to get the different score.<\/p>\n\n<pre><\/pre>\n\n<p>Will i have to run the session again to get the prediction ?<\/p>\n","answers":[{"AnswerId":"45165180","CreationDate":"2017-07-18T11:13:42.603","ParentId":null,"OwnerUserId":"7825531","Title":null,"Body":"<p>Since i have not enough reputation to add a comment to Salvador Dalis answer this is the way to go:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/count_nonzero\" rel=\"nofollow noreferrer\"><code>tf.count_nonzero<\/code><\/a> casts your values into an <code>tf.int64<\/code> unless specified otherwise. Using:<\/p>\n\n<pre><code>argmax_prediction = tf.argmax(prediction, 1)\nargmax_y = tf.argmax(y, 1)\n\nTP = tf.count_nonzero(argmax_prediction * argmax_y, dtype=tf.float32)\nTN = tf.count_nonzero((argmax_prediction - 1) * (argmax_y - 1), dtype=tf.float32)\nFP = tf.count_nonzero(argmax_prediction * (argmax_y - 1), dtype=tf.float32)\nFN = tf.count_nonzero((argmax_prediction - 1) * argmax_y, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>is a realy good idea.<\/p>\n"},{"AnswerId":"50251763","CreationDate":"2018-05-09T11:00:10.843","ParentId":null,"OwnerUserId":"3867406","Title":null,"Body":"<h1>Multi-label case<\/h1>\n\n<p>Previous answers do not specify how to handle the multi-label case so here is such a version implementing <strong>three types of multi-label f1 score in tensorflow<\/strong>: micro, macro and weighted (as per scikit-learn)<\/p>\n\n<p><strong>Update (06\/06\/18):<\/strong> I wrote a <a href=\"http:\/\/vict0rsch.github.io\/2018\/06\/06\/tensorflow-streaming-multilabel-f1\/\" rel=\"noreferrer\">blog post<\/a> about how to compute the <strong>streaming multilabel f1 score<\/strong> in case it helps anyone (it's a longer process, don't want to overload this answer) <\/p>\n\n<pre><code>f1s = [0, 0, 0]\n\ny_true = tf.cast(y_true, tf.float64)\ny_pred = tf.cast(y_pred, tf.float64)\n\nfor i, axis in enumerate([None, 0]):\n    TP = tf.count_nonzero(y_pred * y_true, axis=axis)\n    FP = tf.count_nonzero(y_pred * (y_true - 1), axis=axis)\n    FN = tf.count_nonzero((y_pred - 1) * y_true, axis=axis)\n\n    precision = TP \/ (TP + FP)\n    recall = TP \/ (TP + FN)\n    f1 = 2 * precision * recall \/ (precision + recall)\n\n    f1s[i] = tf.reduce_mean(f1)\n\nweights = tf.reduce_sum(y_true, axis=0)\nweights \/= tf.reduce_sum(weights)\n\nf1s[2] = tf.reduce_sum(f1 * weights)\n\nmicro, macro, weighted = f1s\n<\/code><\/pre>\n\n<hr>\n\n<h3>Correctness<\/h3>\n\n<pre><code>def tf_f1_score(y_true, y_pred):\n    \"\"\"Computes 3 different f1 scores, micro macro\n    weighted.\n    micro: f1 score accross the classes, as 1\n    macro: mean of f1 scores per class\n    weighted: weighted average of f1 scores per class,\n            weighted from the support of each class\n\n\n    Args:\n        y_true (Tensor): labels, with shape (batch, num_classes)\n        y_pred (Tensor): model's predictions, same shape as y_true\n\n    Returns:\n        tuple(Tensor): (micro, macro, weighted)\n                    tuple of the computed f1 scores\n    \"\"\"\n\n    f1s = [0, 0, 0]\n\n    y_true = tf.cast(y_true, tf.float64)\n    y_pred = tf.cast(y_pred, tf.float64)\n\n    for i, axis in enumerate([None, 0]):\n        TP = tf.count_nonzero(y_pred * y_true, axis=axis)\n        FP = tf.count_nonzero(y_pred * (y_true - 1), axis=axis)\n        FN = tf.count_nonzero((y_pred - 1) * y_true, axis=axis)\n\n        precision = TP \/ (TP + FP)\n        recall = TP \/ (TP + FN)\n        f1 = 2 * precision * recall \/ (precision + recall)\n\n        f1s[i] = tf.reduce_mean(f1)\n\n    weights = tf.reduce_sum(y_true, axis=0)\n    weights \/= tf.reduce_sum(weights)\n\n    f1s[2] = tf.reduce_sum(f1 * weights)\n\n    micro, macro, weighted = f1s\n    return micro, macro, weighted\n\n\ndef compare(nb, dims):\n    labels = (np.random.randn(nb, dims) &gt; 0.5).astype(int)\n    predictions = (np.random.randn(nb, dims) &gt; 0.5).astype(int)\n\n    stime = time()\n    mic = f1_score(labels, predictions, average='micro')\n    mac = f1_score(labels, predictions, average='macro')\n    wei = f1_score(labels, predictions, average='weighted')\n\n    print('sklearn in {:.4f}:\\n    micro: {:.8f}\\n    macro: {:.8f}\\n    weighted: {:.8f}'.format(\n        time() - stime, mic, mac, wei\n    ))\n\n    gtime = time()\n    tf.reset_default_graph()\n    y_true = tf.Variable(labels)\n    y_pred = tf.Variable(predictions)\n    micro, macro, weighted = tf_f1_score(y_true, y_pred)\n    with tf.Session() as sess:\n        tf.global_variables_initializer().run(session=sess)\n        stime = time()\n        mic, mac, wei = sess.run([micro, macro, weighted])\n        print('tensorflow in {:.4f} ({:.4f} with graph time):\\n    micro: {:.8f}\\n    macro: {:.8f}\\n    weighted: {:.8f}'.format(\n            time() - stime, time()-gtime,  mic, mac, wei\n        ))\n\ncompare(10 ** 6, 10)\n<\/code><\/pre>\n\n<p><strong>outputs<\/strong>: <\/p>\n\n<pre><code>&gt;&gt; rows: 10^6 dimensions: 10\nsklearn in 2.3939:\n    micro: 0.30890287\n    macro: 0.30890275\n    weighted: 0.30890279\ntensorflow in 0.2465 (3.3246 with graph time):\n    micro: 0.30890287\n    macro: 0.30890275\n    weighted: 0.30890279\n<\/code><\/pre>\n"},{"AnswerId":"35795176","CreationDate":"2016-03-04T11:42:08.910","ParentId":null,"OwnerUserId":"4190832","Title":null,"Body":"<p>Maybe this example will speak to you :    <\/p>\n\n<pre><code>    pred = multilayer_perceptron(x, weights, biases)\n    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n\n    with tf.Session() as sess:\n    init = tf.initialize_all_variables()\n    sess.run(init)\n    for epoch in xrange(150):\n            for i in xrange(total_batch):\n                    train_step.run(feed_dict = {x: train_arrays, y: train_labels})\n                    avg_cost += sess.run(cost, feed_dict={x: train_arrays, y: train_labels})\/total_batch         \n            if epoch % display_step == 0:\n                    print \"Epoch:\", '%04d' % (epoch+1), \"cost=\", \"{:.9f}\".format(avg_cost)\n\n    #metrics\n    y_p = tf.argmax(pred, 1)\n    val_accuracy, y_pred = sess.run([accuracy, y_p], feed_dict={x:test_arrays, y:test_label})\n\n    print \"validation accuracy:\", val_accuracy\n    y_true = np.argmax(test_label,1)\n    print \"Precision\", sk.metrics.precision_score(y_true, y_pred)\n    print \"Recall\", sk.metrics.recall_score(y_true, y_pred)\n    print \"f1_score\", sk.metrics.f1_score(y_true, y_pred)\n    print \"confusion_matrix\"\n    print sk.metrics.confusion_matrix(y_true, y_pred)\n    fpr, tpr, tresholds = sk.metrics.roc_curve(y_true, y_pred)\n<\/code><\/pre>\n"},{"AnswerId":"43960730","CreationDate":"2017-05-14T05:39:00.433","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You do not really need sklearn to calculate precision\/recall\/f1 score. You can easily express them in TF-ish way by looking at the formulas:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/U0hjG.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/U0hjG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now if you have your <code>actual<\/code> and <code>predicted<\/code> values as vectors of 0\/1, you can calculate TP, TN, FP, FN using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/count_nonzero\" rel=\"noreferrer\">tf.count_nonzero<\/a>:<\/p>\n\n<pre><code>TP = tf.count_nonzero(predicted * actual)\nTN = tf.count_nonzero((predicted - 1) * (actual - 1))\nFP = tf.count_nonzero(predicted * (actual - 1))\nFN = tf.count_nonzero((predicted - 1) * actual)\n<\/code><\/pre>\n\n<p>Now your metrics are easy to calculate:<\/p>\n\n<pre><code>precision = TP \/ (TP + FP)\nrecall = TP \/ (TP + FN)\nf1 = 2 * precision * recall \/ (precision + recall)\n<\/code><\/pre>\n"},{"AnswerId":"44849578","CreationDate":"2017-06-30T15:06:48.407","ParentId":null,"OwnerUserId":"3323084","Title":null,"Body":"<p>Use the metrics APIs provided in tf.contrib.metrics, for example:<\/p>\n\n<pre><code>labels = ...\npredictions = ...\n\naccuracy, update_op_acc = tf.contrib.metrics.streaming_accuracy(labels, predictions)\nerror, update_op_error = tf.contrib.metrics.streaming_mean_absolute_error(labels, predictions)\n\nsess.run(tf.local_variables_initializer())\nfor batch in range(num_batches):\n  sess.run([update_op_acc, update_op_error])\naccuracy, mean_absolute_error = sess.run([accuracy, mean_absolute_error])\n<\/code><\/pre>\n"}]}
{"QuestionId":35366139,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-12T15:21:58.267","AcceptedAnswerId":null,"OwnerUserId":1993300.0,"Title":"Building a conversational model using TensorFlow","Body":"<p>I'd like to build a conversational modal that can predict a sentence using the previous sentences using TensorFlow LSTMs . The example provided in TensorFlow tutorial can be used to predict the next word in a sentence .<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/tutorials\/recurrent\/index.html\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/v0.6.0\/tutorials\/recurrent\/index.html<\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Can I use the same technique to predict the next sentence ? Is there any working example on how to do this ? <\/p>\n","answers":[{"AnswerId":"35368570","CreationDate":"2016-02-12T17:18:48.673","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>A sentence is composed words, so you can indeed predict the next sentence by predicting words sequentially. There are models, such as the one described in <a href=\"http:\/\/arxiv.org\/abs\/1507.07998\" rel=\"nofollow\">this<\/a> paper, that build embeddings for entire paragraphs, which can be useful for your purpose. Of course there is <a href=\"http:\/\/arxiv.org\/abs\/1506.05869\" rel=\"nofollow\">Neural Conversational Model<\/a> work that probably directly fits your need. TensorFlow doesn't ship with working examples of these models, but the recurrent models that come with TensorFlow should give you a good starting point for implementing them.<\/p>\n"},{"AnswerId":"37632382","CreationDate":"2016-06-04T16:11:57.377","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>You want to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/tutorials\/seq2seq\/index.html#sequence-to-sequence-models\" rel=\"nofollow\">Sequence-to-sequence<\/a> model. Instead of having it learn to translate sentences from a source language to a target language you have it learn responses to previous utterances in the conversation.<\/p>\n\n<p>You can adapt the example seq2seq model in tensorflow by using the analogy that the source language 'English' is your set of previous sentences and target language 'French' are your response sentences.<\/p>\n\n<p>In theory you could use the basic LSTM you were looking at by concatenating your training examples with a special symbol like this:<\/p>\n\n<p><code>hello there ! __RESPONSE hi , how can i help ?<\/code><\/p>\n\n<p>Then during testing you run it forward with a sequence up to and including the __RESPONSE symbol and the LSTM can carry it the rest of the way.<\/p>\n\n<p>However, the seq2seq model above should be much more accurate and powerful because it had a separate encoder \/ decoder and includes an attention mechanism.<\/p>\n"}]}
{"QuestionId":35366408,"AnswerCount":1,"Tags":"<python><gpu><theano>","CreationDate":"2016-02-12T15:33:32.617","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Theano GPU Install failed nvcc","Body":"<p>After installing Winpython, and launching <\/p>\n\n<pre><\/pre>\n\n<p>I got this error message:<\/p>\n\n<p>ValueError: Theano nvcc.flags support only parameter\/value pairs without space between them. e.g.: '--machine 64' is not supported, but '--machine=64' is supported. Please add the '=' symbol. nvcc.flags value is 'D:_devs\\Python01\\WinPython-64\\python-3.4.3.amd64\\libs'<\/p>\n\n<p>My theanorc. config file is:<\/p>\n\n<pre><\/pre>\n\n<p>Has anybody have an idea ?<\/p>\n","answers":[{"AnswerId":"35366622","CreationDate":"2016-02-12T15:44:21.947","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Find the issue and correct file syntax:<\/p>\n\n<pre><code>[nvcc]\nfastmath = True\n--flags=-LD:\\_devs\\Python01\\WinPython-64\\python-3.4.3.amd64\\libs\n--compiler_bindir=D:\\_app\\visualstudio13\\VC\\bin\\\n<\/code><\/pre>\n\n<p>Solved.<\/p>\n"}]}
{"QuestionId":35366970,"AnswerCount":2,"Tags":"<python><numpy><theano><tensorflow>","CreationDate":"2016-02-12T16:00:48.757","AcceptedAnswerId":35367166.0,"OwnerUserId":5874853.0,"Title":"Theano.function equivalent in Tensorflow","Body":"<p>I am wondering if there is any equivalent to <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>in TensorFlow<\/p>\n","answers":[{"AnswerId":"35369025","CreationDate":"2016-02-12T17:46:05.867","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Theano's <code>function<\/code> returns an object that acts like a Python function and executes the computational graph when called. In TensorFlow, you execute computational graph using session's <code>run<\/code> method. If you want to have a similar Theano-style function object that you can call, you could use <code>TensorFlowTheanoFunction<\/code> wrapper below as a drop-in replacement for theano's <code>function<\/code> <\/p>\n\n<pre><code>class TensorFlowTheanoFunction(object):   \n  def __init__(self, inputs, outputs):\n    self._inputs = inputs\n    self._outputs = outputs\n\n  def __call__(self, *args, **kwargs):\n    feeds = {}\n    for (argpos, arg) in enumerate(args):\n      feeds[self._inputs[argpos]] = arg\n    return tf.get_default_session().run(self._outputs, feeds)\n\na = tf.placeholder(dtype=tf.int32)\nb = tf.placeholder(dtype=tf.int32)\nc = a+b\nd = a-b\nsess = tf.InteractiveSession()\nf = TensorFlowTheanoFunction([a, b], [c, d])\nprint f(1, 2)\n<\/code><\/pre>\n\n<p>You will see<\/p>\n\n<pre><code>[3, -1]\n<\/code><\/pre>\n"},{"AnswerId":"35367166","CreationDate":"2016-02-12T16:11:25.440","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>The <code>run<\/code> method on the <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/client.html#Session\" rel=\"noreferrer\"><code>tf.Session<\/code><\/a> class is quite close to <code>theano.function<\/code>. Its <code>fetches<\/code> and <code>feed_dict<\/code> arguments are moral equivalents of <code>outputs<\/code> and <code>givens<\/code>.<\/p>\n"}]}
{"QuestionId":35372795,"AnswerCount":1,"Tags":"<python><computer-vision><caffe>","CreationDate":"2016-02-12T21:51:44.920","AcceptedAnswerId":35376659.0,"OwnerUserId":4308982.0,"Title":"Your own caffe model from first pricinples","Body":"<p>I am new to caffe. I've gotten so far that I've installed caffe, pycaffe and matcaffe and looked at some tutorials.<\/p>\n\n<p>I'd like to build my own (lm)db file to use to train a model.<\/p>\n\n<p>As far as I can tell, all the tutorials at caffe.berkeleyvision.com use a set of preprocessed binary files when they create their lmdb files.<\/p>\n\n<p>I'd like to start from basics: Say that I have a small set of images in just the right format (jpeg, tiff, whatever it happens to be). In this example that's all I've got. However the number of images is small so I could, in this hypothetical scenario, assign every picture a label myself, manually. Not good if you are actually going to train a model, if one is out to understand the process of creating a lmdb file, however - gold!<\/p>\n\n<p>Feel free to substitute lmdb for h5py or whatever suits the situation the best.<\/p>\n\n<p><strong>How would I go about creating a databasde from a set of images alone?<\/strong> <\/p>\n","answers":[{"AnswerId":"35376659","CreationDate":"2016-02-13T05:58:25.887","ParentId":null,"OwnerUserId":"5352059","Title":null,"Body":"<p>If you are okay with hand labeling, you can create a text document that has a line for each image with the file name\/location along with the label, e.g.<\/p>\n\n<pre><code>img1.jpg 0\nimg2.jpg 1\nimg3.jpb 3\nimg4.jpg 0\n[...]\n<\/code><\/pre>\n\n<p>Then, either use an <code>ImageData<\/code> layer or convert it to an lmdb or leveldb with the <code>convert_imageset<\/code> tool that comes with caffe.<\/p>\n"}]}
{"QuestionId":35374958,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-13T01:21:42.463","AcceptedAnswerId":35375029.0,"OwnerUserId":2272798.0,"Title":"Reshape tensor using placeholder value","Body":"<p>I want to reshape a tensor using the [int, -1] notation (to flatten an image, for example). But I don't know the first dimension ahead of time. One use case is train on a large batch, then evaluate on a smaller batch.<\/p>\n\n<p>Why does this give the following error: ?<\/p>\n\n<pre><\/pre>\n\n<p>Note: when I have the reshape outside of the function it seems to work, but I have very large models that I use multiple times, so I need to keep them in a function and pass the dim using an argument.<\/p>\n","answers":[{"AnswerId":"42603755","CreationDate":"2017-03-05T02:21:43.847","ParentId":null,"OwnerUserId":"7660139","Title":null,"Body":"<p>hi all the issue is due to Keras version. I tried above all without any success. Uninstall Keras and install via pip. It worked for me.<\/p>\n\n<p>I was facing this error with Keras 1.0.2 &amp; resolved with Keras 1.2.0<\/p>\n\n<p>Hope this will help. Thank you<\/p>\n"},{"AnswerId":"35375029","CreationDate":"2016-02-13T01:33:25.900","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To make this work, replace the function:<\/p>\n\n<pre><code>def reshape(_batch_size):\n    return tf.reshape(x, [_batch_size, -1])\n<\/code><\/pre>\n\n<p>\u2026with the function:<\/p>\n\n<pre><code>def reshape(_batch_size):\n    return tf.reshape(x, tf.pack([_batch_size, -1]))\n<\/code><\/pre>\n\n<p>The reason for the error is that <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#reshape\" rel=\"noreferrer\"><code>tf.reshape()<\/code><\/a> expects a value that is convertible to a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a> as its second argument. TensorFlow will automatically convert a list of Python numbers to a <code>tf.Tensor<\/code> but will not automatically convert a mixed list of numbers and tensors (such as a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a>)\u2014instead raising the somewhat unintuitive error message you saw.<\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#pack\" rel=\"noreferrer\"><code>tf.pack()<\/code><\/a> op takes a <em>list<\/em> of objects convertible to a tensor, and converts each element individually, so it can handle the combination of a placeholder and an integer.<\/p>\n"}]}
{"QuestionId":35376302,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-02-13T05:04:26.577","AcceptedAnswerId":35376423.0,"OwnerUserId":2562693.0,"Title":"How to write a net with a Memory Data Layer using pycaffe?","Body":"<p>It is possible to write a caffe prototxt in pycaffe with a hdf5 data layer from the following lines:<\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to use a similar method to write a prototxt that has a Memory Data Layer?<\/p>\n","answers":[{"AnswerId":"35376423","CreationDate":"2016-02-13T05:22:30.503","ParentId":null,"OwnerUserId":"5352059","Title":null,"Body":"<p>Try something like this:<\/p>\n\n<pre><code>import caffe\nfrom caffe import layers as L\n\ndef logreg(height, width, channels, batch_size):\n    n = caffe.NetSpec()\n    n.data = L.MemoryData(batch_size = batch_size, height = height, width = width, channels = channels)\n    n.ip1 = L.InnerProduct(n.data, num_output = 2, weight_filler = dict(type='xavier'))\n    return n.to_proto()\n\nwith open('models\/logreg_memdata.prototxt', 'w') as f:\n    f.write(str(logreg(128,128,3, chunck_size)))\n<\/code><\/pre>\n"}]}
{"QuestionId":35378592,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-02-13T10:13:51.237","AcceptedAnswerId":35386430.0,"OwnerUserId":2989201.0,"Title":"Measuring Training Error in Caffe","Body":"<p>I'm working with Caffe and am interested in comparing my training and test errors in order to determine if my network is overfitting or underfitting. However, I can't seem to figure out how to have Caffe report training error. It will show training loss (the value of the loss function computed over the batch), but this is not useful in determining if the network is overfitting\/underfitting. Is there a straightforward way to do this?<\/p>\n\n<p>I'm using the Python interface to Caffe (pycaffe). If I could get access to the raw training set somehow, I could just put batches through with forward passes and evaluate the results. But, I can't seem to figure out how to access more than the currently-processing batch of training data. Is this possible? My data is in a LMDB format.<\/p>\n","answers":[{"AnswerId":"35386430","CreationDate":"2016-02-13T23:10:13.657","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>In the <code>train_val.prototxt<\/code> file change the <code>source<\/code> in the <code>TEST<\/code> phase to point to the training LMDB database (by default it points to the validation LMDB database) and then run this command:<\/p>\n\n<pre><code>$ .\/build\/tools\/caffe test -solver models\/bvlc_reference_caffenet\/solver.prototxt -weights models\/bvlc_reference_caffenet\/&lt;caffenet_train_iter&gt;.caffemodel -gpu 0\n<\/code><\/pre>\n"}]}
{"QuestionId":35382120,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><conv-neural-network>","CreationDate":"2016-02-13T16:11:12.060","AcceptedAnswerId":null,"OwnerUserId":755557.0,"Title":"Implement rotation invariant convolution layer using tensorflow","Body":"<p>I am trying to implement rotation-invariant convolution layer using tensorflow as a part of udacity deep learning course. <\/p>\n\n<p>I tried the approach below but it doesn't because i get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Here I tried to implement the simplest approach by feeding an original and 90 degree rotated image to the same convolution kernel. After this i rotated back the output of the corresponding convolution layer.<\/p>\n\n<p>Do you have any ideas how to fix this approach or may be you know a better way to implement at least 90 degree rotation-invariant convolution layer?<\/p>\n\n<p>my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"35382665","CreationDate":"2016-02-13T16:59:22.447","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The latest release of TensorFlow (0.6.0) was missing a gradient definition for <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/array_ops.html#reverse\" rel=\"nofollow\"><code>tf.reverse()<\/code><\/a>. It was added in a subsequent <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/1e416cd23464aa77d23a2d63a9f7df12325d6491\" rel=\"nofollow\">commit<\/a>, and you can add the following code to the top level of your program to make use of it without upgrading:<\/p>\n\n<pre><code>@tf.RegisterGradient(\"Reverse\")\ndef _ReverseGrad(op, grad):\n  reverse_dims = op.inputs[1]\n  return tf.reverse(grad, reverse_dims), None\n<\/code><\/pre>\n"}]}
{"QuestionId":35382409,"AnswerCount":1,"Tags":"<neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-02-13T16:35:56.977","AcceptedAnswerId":35382604.0,"OwnerUserId":5318180.0,"Title":"Tensorflow error when calculating cross entropy loss","Body":"<p>I get following error <\/p>\n\n<pre><\/pre>\n\n<p>when I try to calculate cross entropy loss<\/p>\n\n<pre><\/pre>\n\n<p>I use Python 3.4.3. <\/p>\n\n<p>Any ideas why? <\/p>\n","answers":[{"AnswerId":"35382604","CreationDate":"2016-02-13T16:54:16.810","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It sounds like you have defined <code>input_y<\/code>\u2014which I am assuming is a <code>tf.placeholder()<\/code>\u2014as having type <code>tf.int32<\/code>. Either change this to <code>tf.float32<\/code> or add a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/array_ops.html#cast\">cast<\/a>: <code>tf.cast(input_y, tf.float32)<\/code> or <code>tf.to_float(input_y)<\/code>.<\/p>\n"}]}
{"QuestionId":35382977,"AnswerCount":2,"Tags":"<python><sorting><theano>","CreationDate":"2016-02-13T17:28:08.690","AcceptedAnswerId":null,"OwnerUserId":542156.0,"Title":"How to sort rows in Theano based on 1st column, then 2nd column, etc","Body":"<p>I have a Theano array that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I want to sort the rows based on the first column, and for rows which have the same value in the first column, sort based on the second column, etc... In this example there are 3 columns but ideally I'd like the method to work with an arbitrary number of columns.<\/p>\n\n<p>The expected result is:<\/p>\n\n<pre><\/pre>\n\n<p>Edit: the cells can have very large values (int64), so combining the columns together doesn't look like a viable solution to me.<\/p>\n","answers":[{"AnswerId":"35383056","CreationDate":"2016-02-13T17:35:07.293","ParentId":null,"OwnerUserId":"3629665","Title":null,"Body":"<p>You can simply do that with a lexical sort using <code>sorted<\/code>, assuming that your data is formatted consistently like that.<\/p>\n\n<p>Assuming that this is stored in a text file, you can just do:<\/p>\n\n<pre><code>with open('foo.txt', 'r') as inputfile:\n    lines = inputfile.read().splitlines()\n    with open('bar.txt', 'w') as outputfile:\n        outputfile.write('\\n'.join(sorted(lines)))\n<\/code><\/pre>\n\n<p>Why? Because space comes before numbers in a lexical sort, so <code>' 8'<\/code> is less than <code>'10'<\/code>, for instance. ASCII was specifically designed for sorts like this.<\/p>\n"},{"AnswerId":"38415701","CreationDate":"2016-07-16T21:15:03.680","ParentId":null,"OwnerUserId":"5424661","Title":null,"Body":"<p>It's better to compute the order list out of Theano in numpy, because one have to deal with tied scenario, which might need if statement.<\/p>\n\n<p>After you got the order list, convert it to a theano ivector or vector with dtype='int64' or 'int32' and use that as index on the original matrix.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>    0  43  1\nA = 4   8  5\n    0  41  8\n    4   8  4\n<\/code><\/pre>\n\n<p>compute order list in nump, which leads to <code>b = [2,0,3,1]<\/code><\/p>\n\n<pre><code>A = theano.matrix()\norder_ls = theano.tensor.ivector(name='orderlist')\nreordered_matrix = A[order_ls]\n<\/code><\/pre>\n\n<p>to get the value of the reordered matrix, one could create a theano function like<\/p>\n\n<pre><code>get_value = theano.function([A, order_ls], reordered_matrix)\n<\/code><\/pre>\n"}]}
{"QuestionId":35384624,"AnswerCount":1,"Tags":"<python><python-2.7><machine-learning><neural-network><tensorflow>","CreationDate":"2016-02-13T19:53:16.577","AcceptedAnswerId":35632838.0,"OwnerUserId":5096669.0,"Title":"Rank Mismatch error in TensorFlow concat","Body":"<p>I have been trying to make a simple 2 layered neural network. I studied the tensorflow api and official tutorials, I made the one layered model, but am having trouble in the neural network. Here is a section of my code that is causing the error:<\/p>\n\n<pre><\/pre>\n\n<p>the error is:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the full code: <a href=\"http:\/\/pastebin.com\/sX7RqbAf\" rel=\"nofollow\">http:\/\/pastebin.com\/sX7RqbAf<\/a><\/p>\n\n<p>I have used TensorFlow and Python 2.7. I am quite new to neural nets and machine learning so pardon me for any mistakes, thanks in advance.<\/p>\n","answers":[{"AnswerId":"35632838","CreationDate":"2016-02-25T16:24:57.947","ParentId":null,"OwnerUserId":"358696","Title":null,"Body":"<p>In your example:<\/p>\n\n<ul>\n<li><code>tf_train_labels<\/code> has shape <code>[batch_size, num_labels]<\/code><\/li>\n<li>Therefore <code>labels<\/code> has shape <code>[batch_size, 1, num_labels]<\/code><\/li>\n<li><code>indices<\/code> has shape <code>[batch_size, 1]<\/code><\/li>\n<\/ul>\n\n<p>Therefore, when you write:<\/p>\n\n<pre><code>concated = tf.concat(1, [indices, tf.cast(labels,tf.int32)])\n<\/code><\/pre>\n\n<p>it throws an error, because the third dimension of <code>labels<\/code> and <code>indices<\/code> is different.  <code>labels<\/code> has a third dimension of size <code>num_labels<\/code> (presumably 10), and <code>indices<\/code> has no third dimension.<\/p>\n"}]}
{"QuestionId":35387109,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-02-14T00:45:32.053","AcceptedAnswerId":35387337.0,"OwnerUserId":5256563.0,"Title":"Validation and Test with TensorFlow","Body":"<p>I have created a one hidden layer neural network with a pyramidal architecture using TensorFlow. Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>And now I train my NN:<\/p>\n\n<pre><\/pre>\n\n<p>Now I would like to validate and test my NN after training. But I don't know how to create the new feed_dict and use session.run in order to validate\/test.<\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"35387337","CreationDate":"2016-02-14T01:18:23.580","ParentId":null,"OwnerUserId":"1391220","Title":null,"Body":"<p>You will first have to create appropriate validation\/test tensor functions. For one-layer MPL, it involves nested multiply with weights and addition of biases (and also Relu's since you have them in your original model). Define these right below your train predictions<\/p>\n\n<pre><code>valid_prediction = tf.nn.softmax(\n                      tf.nn.relu(tf.matmul(\n                         tf.nn.relu(tf.matmul(tf_valid_dataset, Wh) + bh)), Wout) + bout)))\ntest_prediction = tf.nn.softmax(\n                      tf.nn.relu(tf.matmul(\n                         tf.nn.relu(tf.matmul(tf_test_dataset, Wh) + bh)), Wout) + bout)))\n<\/code><\/pre>\n\n<p>These expressions are in fact quite identical to <code>logit<\/code> variable defined in your code, only using <code>tf_valid_dataset<\/code> and <code>tf_test_dataset<\/code> respectively. You can create intermediate variables to simplify them.<\/p>\n\n<p>You will then have to create some validation\/test function to test accuracy. Simplest would be to test most likely predicted class (Misclassification error roughly). Define this outside your graph\/session.<\/p>\n\n<pre><code>def accuracy(predictions, labels):\n      pred_class = np.argmax(predictions, 1)\n      true_class = np.argmax(labels, 1)\n      return (100.0 * np.sum(pred_class == true_class) \/ predictions.shape[0])\n<\/code><\/pre>\n\n<p>After that, you can simply pass this accuracy function inside same session\/feed_dict to compute validation\/test score.<\/p>\n\n<pre><code>print 'Validation accuracy: %.1f%%' % accuracy(valid_prediction.eval(), valid_labels)\nprint 'Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels)\n<\/code><\/pre>\n"}]}
{"QuestionId":35393212,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-14T15:03:26.707","AcceptedAnswerId":35394687.0,"OwnerUserId":2946792.0,"Title":"Partial update on Tensor in TensorFlow","Body":"<p>I am trying to run multiple convolutions on image in TensorFlow and then concatenate results. Because  allocates new tensor I sometimes run into  (current solution is to change  to smaller value). <\/p>\n\n<p>So here is my question. Is there a way to create big tensor (I know all dimensions in advance) and then assign results of convolutions to it (part by part to avoid concatenating and memory allocation)? Or maybe there is other more efficient way of doing this?<\/p>\n\n<p>Something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35394687","CreationDate":"2016-02-14T17:18:51.550","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>There isn't a way to do this - TensorFlow objects are immutable by design.<\/p>\n\n<p>There might be another way to accomplish what you want (and it'd be interesting to know about cases that are running out of memory for future improvements).  <\/p>\n"}]}
{"QuestionId":35393742,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-02-14T15:54:40.603","AcceptedAnswerId":null,"OwnerUserId":2997380.0,"Title":"pruning tensorflow connections and weights (using cifar10 cnn)","Body":"<p>I'm using tensorflow to run a cnn for image classification.\nI use tensorflow cifar10 cnn implementation.(<a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks\" rel=\"nofollow\">tensorflow cifar10<\/a>)\nI want to decrease the number of connections, meaning I want to  prune the low-weight connections.\nHow can I create a new graph(subgraph) without some of the nuerones?<\/p>\n","answers":[{"AnswerId":"50504004","CreationDate":"2018-05-24T07:58:53.830","ParentId":null,"OwnerUserId":"9298396","Title":null,"Body":"<p>Another way is add an addition mask to control the connections.<\/p>\n\n<blockquote>\n  <p>The first step involves adding mask and threshold variables to the\n  layers that need to undergo pruning. The variable mask is the same\n  shape as the layer's weight tensor and determines which of the weights\n  participate in the forward execution of the graph.<\/p>\n<\/blockquote>\n\n<p>There is a pruning implementation under <strong>tensorflow\/contrib\/model_pruning<\/strong> to prune the model. Hope this can help you to prune model quickly.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning<\/a><\/p>\n"},{"AnswerId":"52704140","CreationDate":"2018-10-08T14:09:38.637","ParentId":null,"OwnerUserId":"2221231","Title":null,"Body":"<p>I think google has an updated answer here : <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning<\/a><\/p>\n\n<p><strong>Removing pruning nodes from the trained graph:<\/strong><\/p>\n\n<pre><code>$ bazel build -c opt contrib\/model_pruning:strip_pruning_vars\n$ bazel-bin\/contrib\/model_pruning\/strip_pruning_vars --checkpoint_path=\/tmp\/cifar10_train --output_node_names=softmax_linear\/softmax_linear_2 --filename=cifar_pruned.pb\n<\/code><\/pre>\n\n<p>I suppose that cifar_pruned.pb will be smaller, since the pruned \"or zero masked\" variables are removed.<\/p>\n"},{"AnswerId":"42557499","CreationDate":"2017-03-02T14:02:22.517","ParentId":null,"OwnerUserId":"6232024","Title":null,"Body":"<p>Tensorflow does not allow you lock\/freeze a particular kernel of a particular layer, that I have found. The only I've found to do this is to use the tf.assign() function as shown in<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42517926\/how-to-freeze-lock-weights-of-one-tensorflow-variable-e-g-one-cnn-kernel-of-o\">How to freeze\/lock weights of one Tensorflow variable (e.g., one CNN kernel of one layer<\/a><\/p>\n\n<p>It's fairly cave-man but I've seen no other solution that works. Essentially, you have to .assign() the values every so often as you iterate through the data. Since this approach is so inelegant and brute-force, it's very slow. I do the .assign() every 100 batches.<\/p>\n\n<p>Someone please post a better solution and soon!<\/p>\n"},{"AnswerId":"35399820","CreationDate":"2016-02-15T00:35:26.800","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>The <code>cifar10<\/code> model you point to, and for that matter, most models written in TensorFlow, do not model the weights (and hence, connections) of individual neurons directly in the computation graph. For instance, for fully connected layers, all the connections between the two layers, say, with <code>M<\/code> neurons in the layer below, and 'N' neurons in the layer above, are modeled by one <code>MxN<\/code> weight matrix. If you wanted to completely remove a neuron and all of its outgoing connections from the layer below, you can simply slice out a <code>(M-1)xN<\/code> matrix by removing the relevant row, and multiply it with the corresponding <code>M-1<\/code> activations of the neurons.<\/p>\n"}]}
{"QuestionId":35394103,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-02-14T16:28:11.500","AcceptedAnswerId":35394643.0,"OwnerUserId":1115491.0,"Title":"Initializing tensorflow Variable with an array larger than 2GB","Body":"<p>I am trying to initialize a tensorflow  with pre-trained  embeddings.<\/p>\n\n<p>I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>And I am receiving the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The array () I am trying to assign is of shape  and its size is 3.6GB. <\/p>\n\n<p>I am getting the same error if I try  as well.<\/p>\n\n<p>I know that it fails due to the fact that the array is larger than 2GB. However, I do not know how to assign an array larger than 2GB to a tensorflow <\/p>\n","answers":[{"AnswerId":"35394643","CreationDate":"2016-02-14T17:14:57.473","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The easiest solution is to feed_dict'ing it into a placeholder node that you use to tf.assign to the variable.<\/p>\n\n<pre><code>X = tf.Variable([0.0])\nplace = tf.placeholder(tf.float32, shape=(3000000, 300))\nset_x = X.assign(place)\n# set up your session here....\nsess.run(set_x, feed_dict={place: model.syn0})\n<\/code><\/pre>\n\n<p>As Joshua Little noted in a separate answer, you can also use it in the initializer:<\/p>\n\n<pre><code>X = tf.Variable(place)    # place as defined above\n...\ninit = tf.initialize_all_variables()\n... create sess ...\nsess.run(init, feed_dict={place: model.syn0})\n<\/code><\/pre>\n"},{"AnswerId":"35495519","CreationDate":"2016-02-19T01:05:27.713","ParentId":null,"OwnerUserId":"875389","Title":null,"Body":"<p>\nIt seems like the only option is to use a placeholder. The cleanest way I can find is to initialize to a placeholder directly:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X_init = tf.placeholder(tf.float32, shape=(3000000, 300))\nX = tf.Variable(X_init)\n# The rest of the setup...\nsess.run(tf.initialize_all_variables(), feed_dict={X_init: model.syn0})\n<\/code><\/pre>\n"},{"AnswerId":"51413353","CreationDate":"2018-07-19T02:06:39.823","ParentId":null,"OwnerUserId":"3395301","Title":null,"Body":"<p>try this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom gensim import models\n\nmodel = models.KeyedVectors.load_word2vec_format('.\/GoogleNews-vectors-negative300.bin', binary=True)\nX = model.syn0\n\nembeddings = tf.Variable(tf.random_uniform(X.shape, minval=-0.1, maxval=0.1), trainable=False)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nembeddings.load(model.syn0, sess)\n<\/code><\/pre>\n"}]}
{"QuestionId":35394446,"AnswerCount":3,"Tags":"<python><neural-network><tensorflow><deep-learning><q-learning>","CreationDate":"2016-02-14T16:56:47.157","AcceptedAnswerId":null,"OwnerUserId":1990276.0,"Title":"Why doesn't my Deep Q Network master a simple Gridworld (Tensorflow)? (How to evaluate a Deep-Q-Net)","Body":"<p>I try to familiarize myself with Q-learning and Deep Neural Networks, currently try to implement <a href=\"https:\/\/www.cs.toronto.edu\/~vmnih\/docs\/dqn.pdf\">Playing Atari with Deep Reinforcement Learning<\/a>. <\/p>\n\n<p>To test my implementation and play around with it, I tought I try a simple gridworld. Where i have a N x N Grid and start in the top left corner and finishes at the bottom right. The possible actions are: left, up, right, down. <\/p>\n\n<p>Even though my implementation has become very similar to <a href=\"https:\/\/github.com\/asrivat1\/DeepLearningVideoGames\/blob\/master\/deep_q_network.py\">this<\/a>(hope its a good one) it dosn't seem to learn anything. Looking at the total steps it needs to finish(I guess the average would be aroung 500 with a gridsize of 10x10, but there also very low and high values), it seams more random than anything else to me.<\/p>\n\n<p>I tried it with and without convolutional layers and played around with all the parameters but to be honest, I've no idea if something with my implementation is wrong or it needs to train longer(I let it train for quite a time) or what ever. But at least it seams to converge, here the plot of the loss value one training session:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JGSVK.png\"><img src=\"https:\/\/i.stack.imgur.com\/JGSVK.png\" alt=\"Loss image\"><\/a><\/p>\n\n<p>So what is the problem in this case?<\/p>\n\n<p>But also and maybe more importantly how can I \"debug\" this Deep-Q-Nets, in supervised training there are training, test and validation sets and for example with precision and recall it is possible to evaluate them. What options do I have for unsupervised learning with Deep-Q-Nets, so that the next time maybe I can fix it myself?<\/p>\n\n<p>Finally here is the code:<\/p>\n\n<p>This is the network:<\/p>\n\n<pre><\/pre>\n\n<p>And here the training:<\/p>\n\n<pre><\/pre>\n\n<p>I appreciate every help and ideas you may have!<\/p>\n","answers":[{"AnswerId":"44349845","CreationDate":"2017-06-04T00:54:08.180","ParentId":null,"OwnerUserId":"1990276","Title":null,"Body":"<p>So it's quite a time ago that i wrote this question but it seams that there is still quite some interest and request for the running code i finally decided to create a github repository<\/p>\n\n<p>Since it's quite a time ago i wrote it and so on it won't run out of the box, but it shouldn't be that hard to get it running. So here is the deep q network and example i wrote at the time which back then worked, hope you enjoy: <a href=\"https:\/\/github.com\/natschz\/deepq\" rel=\"nofollow noreferrer\">Link to deep q repository<\/a><\/p>\n\n<p>Would be nice to see some contribution and if you fix it and get it running make a pull request!<\/p>\n"},{"AnswerId":"35468407","CreationDate":"2016-02-17T21:44:27.027","ParentId":null,"OwnerUserId":"1990276","Title":null,"Body":"<p>For those interested, I ajusted the parameters and the model further but the biggest improvment was switching to a simple feed forward network with 3 Layers and about 50 neurons in the hidden layer. For me it then converged in a pretty decent time.<\/p>\n\n<p>Btw further tips for debuggin are appreciated!<\/p>\n"},{"AnswerId":"52124708","CreationDate":"2018-09-01T03:51:44.993","ParentId":null,"OwnerUserId":"9473627","Title":null,"Body":"<p>I have implemented a simple toy DQN without CNN layers, and it works. Here are some findings during my implementation , hope it will help.<\/p>\n\n<ol>\n<li><p>According to DeepMind's paper they didn't use max pooling layer, the reason is that the image will become position invariant which is not good for the game. The position of agent is crucial for the information of the game.\n<a href=\"https:\/\/i.stack.imgur.com\/meSbH.jpg\" rel=\"nofollow noreferrer\">DQN Architecture<\/a><\/p><\/li>\n<li><p>If you want to skip the CNN first use gym environment(Like what I have done for the toy implementation), during my development, here are couple of things I found:<\/p>\n\n<ul>\n<li>Encode your state of environment by one-hot encoding, it will increase the training efficiency.<\/li>\n<li>I only use a matrix of weights with [number of state, number of action] shape, to do the matrix multiplication with input one-hot encoded state. No bias, no activation function(It will increase training time, I assume, it never work after I add other layer or anything).<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>This are two things I found extremely crucial for my implementation to work, I am not fully understand the reason behind it, hope my answer can give you a little bit insight.<\/p>\n"}]}
{"QuestionId":35394514,"AnswerCount":1,"Tags":"<python><if-statement><theano>","CreationDate":"2016-02-14T17:03:20.880","AcceptedAnswerId":35397714.0,"OwnerUserId":2255757.0,"Title":"theano T.switch(): if tensor is empty strange behaviour","Body":"<p>i want to check if vector <\/p>\n\n<pre><\/pre>\n\n<p>this prints:<\/p>\n\n<pre><\/pre>\n\n<p>the conditional statement by itself works, it only returns 1 if A is empty.<\/p>\n","answers":[{"AnswerId":"35397714","CreationDate":"2016-02-14T20:39:15.477","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>This is because <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html\" rel=\"nofollow\"><code>theano.tensor.switch<\/code> operates differently to <code>theano.ifelse.ifelse<\/code><\/a>.<\/p>\n\n<p><code>theano.tensor.switch<\/code> operates element-wise. The three arguments need to have the same shape, or be broadcastable to the same shape. <code>T.eq(A.size, 0)<\/code> will always be a scalar and the true value, <code>[0]<\/code> is a vector with a single element so both will be broadcast to the shape of <code>A<\/code>. The case of <code>A == []<\/code> is undoubtedly odd and I don't know if it is by design; Theano appears to be \"broadcasting\" the scalar and single entry vector to the empty vector.<\/p>\n\n<p>The solution is to switch to <code>theano.ifelse.ifelse<\/code>:<\/p>\n\n<pre><code>A = tt.ivector()\nout = theano.ifelse.ifelse(\n    tt.eq(A.size, 0), tt.unbroadcast(tt.zeros((1,), dtype=A.dtype), 0), A)\nf = theano.function([A], out)\nprint f([])\nprint f([1])\nprint f([1, 2])\n<\/code><\/pre>\n\n<p>As desired, this prints<\/p>\n\n<pre><code>[0]\n[1]\n[1 2]\n<\/code><\/pre>\n\n<p>Note that the two possible <code>ifelse<\/code> output values (the value if the comparison is true and the value if the comparison is false) must have identical types, hence the convoluted method for constructing a zeros vector with a single entry.<\/p>\n"}]}
{"QuestionId":35398836,"AnswerCount":1,"Tags":"<lua><statistics><torch><confusion-matrix>","CreationDate":"2016-02-14T22:34:34.827","AcceptedAnswerId":35587923.0,"OwnerUserId":4013781.0,"Title":"Alternative to optim confusion matrix for multi label classification in torch","Body":"<p>I have a convnet that produces multi label output. Ie: each image can be labelled with 0 to 10 labels. The output is a tensor of size 10, containing 1s (indicating a label at that index) and -1s (no label at that index). Targets are in the same format. So:<\/p>\n\n<pre><\/pre>\n\n<p>Indicates labels 1-3 and 5-10. Multi label output cannot be used with the confusion matrix of the torch optim package. What are some good alternatives for evaluating prediction accuracy? <\/p>\n\n<p>Ideally, I would like to know how accurately <strong>each label<\/strong> is being predicted, in terms of false positives, true positives, etc. One idea would be to make 10 separate confusion matrices and feed each the data for one index. Is there a better solution though?<\/p>\n\n<p>Any input appreciated. <\/p>\n\n<hr>\n\n<p>Update: the above approach worked fine, but it's cumbersome when there are many labels. I just wondered if there was a better way. <\/p>\n","answers":[{"AnswerId":"35587923","CreationDate":"2016-02-23T20:52:27.993","ParentId":null,"OwnerUserId":"4013781","Title":null,"Body":"<p>I made separate confusion matrices for each label, split the output predictions, and added them to the appropriate matrix. By co-plotting classification accuracy for training\/validation set, it's pretty easy to interpret. <\/p>\n"}]}
{"QuestionId":35400065,"AnswerCount":2,"Tags":"<python><tensorflow><text-classification><multilabel-classification>","CreationDate":"2016-02-15T01:10:07.903","AcceptedAnswerId":39472895.0,"OwnerUserId":1336058.0,"Title":"Multilabel Text Classification using TensorFlow","Body":"<p>The text data is organized as vector with 20,000 elements, like [2, 1, 0, 0, 5, ...., 0]. \ni-th element indicates the frequency of the i-th word in a text. <\/p>\n\n<p>The ground truth label data is also represented as vector with 4,000 elements, like [0, 0, 1, 0, 1, ...., 0]. \ni-th element indicates whether the i-th label is a positive label for a text. \nThe number of labels for a text differs depending on texts. <\/p>\n\n<p>I have a code for single-label text classification. <\/p>\n\n<p>How can I edit the following code for multilabel text classification?<\/p>\n\n<p>Especially, I would like to know following points. <\/p>\n\n<ul>\n<li>How to compute accuracy using TensorFlow. <\/li>\n<li>How to set a threshold which judges whether a label is positive or negative. For instance, if the output is [0.80, 0.43, 0.21, 0.01, 0.32] and the ground truth is [1, 1, 0, 0, 1], the labels with scores over 0.25 should be judged as positive. <\/li>\n<\/ul>\n\n<p>Thank you. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39472895","CreationDate":"2016-09-13T14:47:17.330","ParentId":null,"OwnerUserId":"1756427","Title":null,"Body":"<p>Change relu to sigmoid of output layer.\nModify cross entropy loss to explicit mathematical formula of sigmoid cross entropy loss (explicit loss was working in my case\/version of tensorflow )<\/p>\n\n<pre><code>import tensorflow as tf\n\n# hidden Layer\nclass HiddenLayer(object):\n    def __init__(self, input, n_in, n_out):\n        self.input = input\n\n        w_h = tf.Variable(tf.random_normal([n_in, n_out],mean = 0.0,stddev = 0.05))\n        b_h = tf.Variable(tf.zeros([n_out]))\n\n        self.w = w_h\n        self.b = b_h\n        self.params = [self.w, self.b]\n\n    def output(self):\n        linarg = tf.matmul(self.input, self.w) + self.b\n        self.output = tf.nn.relu(linarg)\n\n        return self.output\n\n# output Layer\nclass OutputLayer(object):\n    def __init__(self, input, n_in, n_out):\n        self.input = input\n\n        w_o = tf.Variable(tf.random_normal([n_in, n_out], mean = 0.0, stddev = 0.05))\n        b_o = tf.Variable(tf.zeros([n_out]))\n\n        self.w = w_o\n        self.b = b_o\n        self.params = [self.w, self.b]\n\n    def output(self):\n        linarg = tf.matmul(self.input, self.w) + self.b\n        #changed relu to sigmoid\n        self.output = tf.nn.sigmoid(linarg)\n\n        return self.output\n\n# model\ndef model():\n    h_layer = HiddenLayer(input = x, n_in = 20000, n_out = 1000)\n    o_layer = OutputLayer(input = h_layer.output(), n_in = 1000, n_out = 4000)\n\n    # loss function\n    out = o_layer.output()\n    # modified cross entropy to explicit mathematical formula of sigmoid cross entropy loss\n    cross_entropy = -tf.reduce_sum( (  (y_*tf.log(out + 1e-9)) + ((1-y_) * tf.log(1 - out + 1e-9)) )  , name='xentropy' )    \n\n    # regularization\n    l2 = (tf.nn.l2_loss(h_layer.w) + tf.nn.l2_loss(o_layer.w))\n    lambda_2 = 0.01\n\n    # compute loss\n    loss = cross_entropy + lambda_2 * l2\n\n    # compute accuracy for single label classification task\n    correct_pred = tf.equal(tf.argmax(out, 1), tf.argmax(y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, \"float\"))\n\n    return loss, accuracy\n<\/code><\/pre>\n"},{"AnswerId":"37051795","CreationDate":"2016-05-05T13:22:48.517","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>You have to use variations of cross entropy function in other to support multilabel classification. In case you have less than one thousand of ouputs you should use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/nn.html#sigmoid_cross_entropy_with_logits\" rel=\"noreferrer\">sigmoid_cross_entropy_with_logits<\/a>, in your case that you have 4000 outputs you may consider <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/extras\/candidate_sampling.pdf\" rel=\"noreferrer\">candidate sampling<\/a> as it is faster than the previous.<\/p>\n\n<blockquote>\n  <p>How to compute accuracy using TensorFlow.<\/p>\n<\/blockquote>\n\n<p>This depends on your problem and what you want to achieve. If you don't want to miss any object in an image then if the classifier get all right but one, then you should consider the whole image an error. You can also consider that an object missed or missclassiffied is an error. The latter I think it supported by sigmoid_cross_entropy_with_logits.<\/p>\n\n<blockquote>\n  <p>How to set a threshold which judges whether a label is positive or\n  negative. For instance, if the output is [0.80, 0.43, 0.21, 0.01,\n  0.32] and the ground truth is [1, 1, 0, 0, 1], the labels with scores over 0.25 should be judged as positive.<\/p>\n<\/blockquote>\n\n<p>Threshold is one way to go, you have to decided which one. But that is some kind of hack, not real multilable classification. For that you need the previous functions I said before.<\/p>\n"}]}
{"QuestionId":35400134,"AnswerCount":1,"Tags":"<android><torch><camera2>","CreationDate":"2016-02-15T01:19:56.710","AcceptedAnswerId":35400204.0,"OwnerUserId":3668324.0,"Title":"android api 21+ camera2 api","Body":"<p>i try to run this code, but it say that there is some thing not found in framworke.jar <\/p>\n\n<pre><\/pre>\n\n<p>but i got this error, <\/p>\n\n<pre><\/pre>\n\n<p>plz i need some explanations<\/p>\n","answers":[{"AnswerId":"35400204","CreationDate":"2016-02-15T01:28:37.647","ParentId":null,"OwnerUserId":"1676363","Title":null,"Body":"<p><a href=\"http:\/\/developer.android.com\/reference\/android\/hardware\/camera2\/CameraManager.html#setTorchMode(java.lang.String,%20boolean)\" rel=\"nofollow\">setTorchMode()<\/a> was added in API 23 - API 21-22 devices will not have that method.<\/p>\n"}]}
{"QuestionId":35400936,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2016-02-15T03:13:31.017","AcceptedAnswerId":35417846.0,"OwnerUserId":4360343.0,"Title":"When I try to mount a location in my container I get a error: Invalid Flag for flag -v: bad mode specified","Body":"<p>I am new to Docker and TensorFlow.<\/p>\n\n<p>I am trying to get TensorFlow working with Windows using Docker. I am using the command:  <\/p>\n\n<pre><\/pre>\n\n<p>but I am getting the error:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lz7ho.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lz7ho.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"35417846","CreationDate":"2016-02-15T19:49:24.293","ParentId":null,"OwnerUserId":"4509510","Title":null,"Body":"<p>Could you check if connecting via ssh to your docker-machine and executing from there the <code>docker run<\/code> command the volumes are created? You can follow the below steps:<\/p>\n\n<pre><code>$ docker-machine ssh default\n\n$ docker run -dit -v \/c\/Users\/Aditya\/notebook:\/media\/disk -p 8888:8888 b.gcr.io\/tensorflow\/tensorflow:latest\n<\/code><\/pre>\n\n<p>If the issue persists, could you check if updating your Docker to the latest one the issue fixes?<\/p>\n"}]}
{"QuestionId":35402749,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-15T06:24:44.930","AcceptedAnswerId":35445083.0,"OwnerUserId":4747268.0,"Title":"While running TensorFlow's classify_image, getting AttributeError: 'module' object has no attribute 'gfile'","Body":"<p>I am getting this error while running the sample file given with TensorFlow, in the imagenet model,<\/p>\n\n<pre><\/pre>\n\n<p>AttributeError: 'module' object has no attribute 'gfile'<\/p>\n\n<p>I have tried installing using both, from pip as well as source, on virtualserver as well, still I get this error.<\/p>\n","answers":[{"AnswerId":"35445083","CreationDate":"2016-02-16T23:30:40.853","ParentId":null,"OwnerUserId":"625087","Title":null,"Body":"<p>There are two ways to go about solving this problem:<\/p>\n\n<h2>Option 1<\/h2>\n\n<p>This is an expansion on Yaroslav's comment above, and is easier than option 2.<\/p>\n\n<p>Modify <code>classify_image.py<\/code> as follows:<\/p>\n\n<p>Replace all instances of <code>tf.gfile.Exists<\/code> to <code>os.path.exists<\/code>, and\nreplace all instances of <code>tf.gfile.GFile<\/code> and <code>tf.gfile.FastGFile<\/code> to <code>open<\/code><\/p>\n\n<p>Then run the modified <code>classify_image.py<\/code>, and it should work.<\/p>\n\n<h2>Option 2<\/h2>\n\n<p>Update tensorflow to the latest version that includes <code>gfile<\/code> as described <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>However, after you do that, you might encounter the following error when you try to run <code>classify_image.py<\/code>:<\/p>\n\n<pre><code>$ python classify_image.py\n&gt;&gt; Downloading inception-2015-12-05.tgz 100.0%\nSuccesfully downloaded inception-2015-12-05.tgz 88931400 bytes.\n[libprotobuf ERROR google\/protobuf\/src\/google\/protobuf\/io\/coded_stream.cc:207] A protocol message was rejected because it was too big (more than 67108864 bytes).  To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google\/protobuf\/io\/coded_stream.h.\nTraceback (most recent call last):\n  File \"classify_image.py\", line 213, in &lt;module&gt;\n    tf.app.run()\n  File \"\/Users\/USERNAME\/.virtualenvs\/mlnd\/lib\/python2.7\/site-packages\/tensorflow\/python\/platform\/default\/_app.py\", line 30, in run\n    sys.exit(main(sys.argv))\n  File \"classify_image.py\", line 209, in main\n    run_inference_on_image(image)\n  File \"classify_image.py\", line 159, in run_inference_on_image\n    create_graph()\n  File \"classify_image.py\", line 141, in create_graph\n    graph_def.ParseFromString(f.read())\ngoogle.protobuf.message.DecodeError: Error parsing message\n<\/code><\/pre>\n\n<p>To fix this, you can change a line in the source code as described <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/582\" rel=\"noreferrer\">here<\/a> and then recompile tensorflow on your machine.<\/p>\n\n<p>Option 2 might be a bit of work, especially if you're on a Mac.<\/p>\n"},{"AnswerId":"37480605","CreationDate":"2016-05-27T10:04:35.743","ParentId":null,"OwnerUserId":"214647","Title":null,"Body":"<p>The following steps helped me resolve the issue<\/p>\n\n<ol>\n<li>update tensorflow  as described here <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html<\/a><\/li>\n<\/ol>\n\n<p>2.For me this resulted in protobuf error. To resolve I ran <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>$sudo pip uninstall protobuf <\/p>\n    \n    <p>$sudo pip install --upgrade\n    <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl<\/a><\/p>\n  <\/blockquote>\n<\/blockquote>\n"}]}
{"QuestionId":35403703,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2016-02-15T07:29:55.403","AcceptedAnswerId":null,"OwnerUserId":3755873.0,"Title":"Permission denied when running make_imagenet_mean.sh","Body":"<p>I'm making an ImageNet with my own data which I have successfully made into two database files in the caffe root directory under folders \"DB_train\" and \"DB_verify\".<\/p>\n\n<p>My  looks like:<\/p>\n\n<pre><\/pre>\n\n<p>When I run the command:<\/p>\n\n<pre><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Any reason as to why this is happening?<\/p>\n\n<p>When I run <\/p>\n\n<pre><\/pre>\n\n<p>I get <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n","answers":[]}
{"QuestionId":35404048,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne><recurrent-neural-network>","CreationDate":"2016-02-15T07:52:03.410","AcceptedAnswerId":35546855.0,"OwnerUserId":4565947.0,"Title":"Weird gradient results with recurrent layers","Body":"<p>I've been experimenting with very basic recurrent networks and have seen really strange behaviors. I've spent quite a bit of time trying to narrow down where it goes wrong and I ended up finding that the gradients computed by theano and by finite differentiation are radically different when using a recurrent layer. What is going on here?<\/p>\n\n<p>Here is the kind of problem I have:<\/p>\n\n<p>I have n_seq sequences of n_steps feature vectors of dimension n_feat, along with their labels among n_class classes. The labels are per time step, not per sequence (so I have n_seq*n_steps labels).\nMy goal is to train a model to correctly classify the feature vectors.<\/p>\n\n<p>Here's my minimal example:<\/p>\n\n<p>(In reality, there would be some sequential information in the data, hence recurrent networks should do better, but in this minimal example I generate purely random data, which is good enough to expose the bug.)<\/p>\n\n<p>I create 2 minimal networks:<\/p>\n\n<p>1) A regular feed-forward (not recurrent), with just the input layer and an output layer with softmax  (no hidden layer). I discard the sequential information by considering a \"batch\" of n_seq*n_steps \"independent\" feature vectors.<\/p>\n\n<p>2) An identical network but where the output layer is recurrent. The batch is now of size n_seq and each input is a full sequence of n_steps feature vectors. Finally I reshape the output back to a \"batch\" of size n_seq*n_steps.<\/p>\n\n<p>If the recurrent weights are set to 0, the 2 networks should be equivalent. Indeed I do see that the initial loss for both networks are the same in this case, no matter what random initialization for the feed-forward weights I have.\nIf I implement a finite differentiation, I also get that the (initial) gradients for the feed-forward weights are the same (as they should). However the gradients obtained from theano are radically different (but only for the recurrent network). <\/p>\n\n<p>Here is my code with sample result:<\/p>\n\n<p>NOTE: when run for the first time, I get this warning, I don't know what triggers it but I bet it is relevant to my problem.\nWarning: In the strict mode, all neccessary shared variables must be passed as a part of non_sequences\n  'must be passed as a part of non_sequences', Warning)<\/p>\n\n<p>Any insight would be much appreciated!<\/p>\n\n<p>CODE:<\/p>\n\n<pre><\/pre>\n\n<p>RESULTS:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35546855","CreationDate":"2016-02-22T06:10:44.500","ParentId":null,"OwnerUserId":"4565947","Title":null,"Body":"<p>The problem was coming from non-intuitive (maybe) effect of gradient_steps clipping in BPTT as explained here:\n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/QNge6fC6C4s\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/QNge6fC6C4s<\/a><\/p>\n"}]}
{"QuestionId":35409545,"AnswerCount":1,"Tags":"<python><numpy><theano><cosine-similarity><trigonometry>","CreationDate":"2016-02-15T12:34:55.173","AcceptedAnswerId":35422457.0,"OwnerUserId":5550721.0,"Title":"Cosine similarity in Theano","Body":"<p>What is the easiest way to compute cosine similarity with numpy and theano?\nVectors given as numpy arrays.<\/p>\n\n<p>I've tried to calculate cosine similarity matrix just using numpy, and it works maddeningly slow. However, I am absolutely new to theano, but suppose that this library may help me to build my cosine similarity matrix.<\/p>\n\n<p>Well, help! :)<\/p>\n","answers":[{"AnswerId":"35422457","CreationDate":"2016-02-16T02:10:46.593","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Here's a post about cosine similarity in Python: <a href=\"https:\/\/stackoverflow.com\/questions\/18424228\/cosine-similarity-between-2-number-lists\">Cosine Similarity between 2 Number Lists<\/a>.<\/p>\n\n<p>I rewrote <a href=\"https:\/\/stackoverflow.com\/a\/18424953\/3041068\">this answer<\/a> in Numpy and Theano:<\/p>\n\n<pre><code>def cos_sim_numpy(v1, v2):\n    numerator = sum(v1*v2)\n    denominator = math.sqrt(sum(v1**2)*sum(v2**2))\n    return numerator\/denominator\n\ndef compile_cos_sim_theano():\n    v1 = theano.tensor.vector(dtype=theano.config.floatX)\n    v2 = theano.tensor.vector(dtype=theano.config.floatX)\n    numerator = theano.tensor.sum(v1*v2)\n    denominator = theano.tensor.sqrt(theano.tensor.sum(v1**2)*theano.tensor.sum(v2**2))\n    return theano.function([v1, v2], numerator\/denominator)\n\ncos_sim_theano_fn = compile_cos_sim_theano()\n\nv1 = numpy.asarray([3,45,7,2], dtype=np.float32)\nv2 = numpy.asarray([2,54,13,15], dtype=np.float32)\n\nprint cos_sim_theano_fn(v1, v2), cos_sim_numpy(v1, v2)\n\nOutput: 0.972284251712 0.972284251712\n<\/code><\/pre>\n"}]}
{"QuestionId":35411194,"AnswerCount":4,"Tags":"<math><tensorflow><deep-learning><keras><theano>","CreationDate":"2016-02-15T13:54:11.357","AcceptedAnswerId":49030550.0,"OwnerUserId":4497662.0,"Title":"How is Hard Sigmoid defined","Body":"<p>I am working on Deep Nets using keras. There is an activation \"hard sigmoid\". Whats its mathematical definition ?<\/p>\n\n<p>I know what is Sigmoid. Someone asked similar question on Quora: <a href=\"https:\/\/www.quora.com\/What-is-hard-sigmoid-in-artificial-neural-networks-Why-is-it-faster-than-standard-sigmoid-Are-there-any-disadvantages-over-the-standard-sigmoid\" rel=\"nofollow\">https:\/\/www.quora.com\/What-is-hard-sigmoid-in-artificial-neural-networks-Why-is-it-faster-than-standard-sigmoid-Are-there-any-disadvantages-over-the-standard-sigmoid<\/a><\/p>\n\n<p>But I could not find the precise mathematical definition anywhere ?<\/p>\n","answers":[{"AnswerId":"49030550","CreationDate":"2018-02-28T13:18:31.587","ParentId":null,"OwnerUserId":"4497662","Title":null,"Body":"<p>it is <\/p>\n\n<pre><code>  clip((x + 1)\/2, 0, 1) \n<\/code><\/pre>\n\n<p>in coding parlance:<\/p>\n\n<pre><code>  max(0, min(1, (x + 1)\/2)) \n<\/code><\/pre>\n"},{"AnswerId":"39625795","CreationDate":"2016-09-21T20:14:11.457","ParentId":null,"OwnerUserId":"3277902","Title":null,"Body":"<p>For reference, the <code>hard sigmoid function<\/code> may be defined differently in different places. In Courbariaux et al. 2016 [1] it's defined as:<\/p>\n\n<blockquote>\n  <p>\u03c3 is the \u201chard sigmoid\u201d function: \u03c3(x) = clip((x + 1)\/2, 0, 1) =\n  max(0, min(1, (x + 1)\/2))<\/p>\n<\/blockquote>\n\n<p>The intent is to provide a probability value (hence constraining it to be between <code>0<\/code> and <code>1<\/code>) for use in stochastic binarization of neural network parameters (e.g. weight, activation, gradient). You use the probability <code>p = \u03c3(x)<\/code> returned from the hard sigmoid function to set the parameter <code>x<\/code> to <code>+1<\/code> with <code>p<\/code> probability, or <code>-1<\/code> with probability <code>1-p<\/code>.<\/p>\n\n<p>[1] <a href=\"https:\/\/arxiv.org\/abs\/1602.02830\" rel=\"nofollow\">https:\/\/arxiv.org\/abs\/1602.02830<\/a> - \"Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1\", Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, Yoshua Bengio, (Submitted on 9 Feb 2016 (v1), last revised 17 Mar 2016 (this version, v3))<\/p>\n"},{"AnswerId":"35580682","CreationDate":"2016-02-23T14:51:59.867","ParentId":null,"OwnerUserId":"2602240","Title":null,"Body":"<p>Since Keras supports both Tensorflow and Theano, the exact implementation might be different for each backend - I'll cover Theano only. For Theano backend Keras uses <code>T.nnet.hard_sigmoid<\/code>, which is in turn <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nnet\/sigm.py#L277\" rel=\"nofollow noreferrer\">linearly approximated standard sigmoid<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>slope = tensor.constant(0.2, dtype=out_dtype)\nshift = tensor.constant(0.5, dtype=out_dtype)\nx = (x * slope) + shift\nx = tensor.clip(x, 0, 1)\n<\/code><\/pre>\n\n<p>i.e. it is: <code>max(0, min(1, x*0.2 + 0.5))<\/code><\/p>\n"},{"AnswerId":"52314126","CreationDate":"2018-09-13T12:49:52.120","ParentId":null,"OwnerUserId":"4375377","Title":null,"Body":"<p>The hard sigmoid is normally a piecewise linear approximation of the logistic sigmoid function. Depending on what properties of the original sigmoid you want to keep, you can use a different approximation.<\/p>\n\n<p>I personally like to keep the function correct at zero, i.e. <code>\u03c3(0) = 0.5<\/code> (shift) and <code>\u03c3'(0) = 0.25<\/code> (slope). This could be coded as follows<\/p>\n\n<pre><code>def hard_sigmoid(x):\n    return np.maximum(0, np.minimum(1, (x + 2) \/ 4))\n<\/code><\/pre>\n"}]}
{"QuestionId":35412375,"AnswerCount":1,"Tags":"<python><macos><tensorflow>","CreationDate":"2016-02-15T14:50:28.560","AcceptedAnswerId":null,"OwnerUserId":3490622.0,"Title":"Unable to get tensorflow working on Mac OSX","Body":"<p>I am trying to get tensor flow on my Mac OS X Yosemite.  I tried it in a number of ways, all of which result in an error.<\/p>\n\n<ol>\n<li><p>Method #1: using virtualenv.  When I get to the <\/p>\n\n<pre><\/pre>\n\n<p>I get <\/p>\n\n<pre><\/pre><\/li>\n<li><p>Method #2: using Anaconda.  Again, when within the anaconda environment I do<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>I get an error, although a different one this time:<\/p>\n\n<pre><\/pre>\n\n<ol start=\"3\">\n<li><p>Method #3: using bazel.  Numerous dependency issues - e.g.,<\/p>\n\n<pre><\/pre>\n\n<p>etc.<\/p>\n\n<p>Has anyone been able to successfully build tensorflow on a mac (that's sitting behind a corporate firewall, btw - maybe that's part of the problem).  Any insight would be much appreciated.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Natalia<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"35415565","CreationDate":"2016-02-15T17:34:23.900","ParentId":null,"OwnerUserId":"3626961","Title":null,"Body":"<p>Did you try removing secure to <code>http<\/code>, something like <code>sudo pip install --upgrade http:\/\/...<\/code> instead (and also, you should try checking <code>pip<\/code> version too)<\/p>\n"}]}
{"QuestionId":35413618,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-02-15T15:52:35.397","AcceptedAnswerId":35424017.0,"OwnerUserId":2680707.0,"Title":"TensorFlow: PlaceHolder error when using tf.merge_all_summaries()","Body":"<p>I am getting a placeholder error.<\/p>\n\n<p>I do not know what it means, because I am mapping correctly on ... I provide here a fully functional MWE reproducing the error:<\/p>\n\n<pre><\/pre>\n\n<p>The error:<\/p>\n\n<pre><\/pre>\n\n<p>If I remove the  or remove  from  then it runs okay.<\/p>\n\n<p>This looks similar to this post:\n<a href=\"https:\/\/stackoverflow.com\/questions\/35114376\/error-when-computing-summaries-in-tensorflow\">Error when computing summaries in TensorFlow<\/a>\nHowever, I am not using iPython...<\/p>\n","answers":[{"AnswerId":"35424017","CreationDate":"2016-02-16T04:51:34.093","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.merge_all_summaries()<\/code> function is convenient, but also somewhat dangerous: it merges <strong>all summaries in the default graph<\/strong>, which includes any summaries from previous&mdash;apparently unconnected&mdash;invocations of code that also added summary nodes to the default graph. If old summary nodes depend on an old placeholder, you will get errors like the one you have shown in your question (and like <a href=\"https:\/\/stackoverflow.com\/q\/35114376\/3574081\">previous<\/a> <a href=\"https:\/\/stackoverflow.com\/q\/35116566\/3574081\">questions<\/a> as well).<\/p>\n\n<p>There are two independent workarounds:<\/p>\n\n<ol>\n<li><p>Ensure that you explicitly collect the summaries that you wish to compute. This is as simple as using the explicit <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/train.html#merge_summary\" rel=\"nofollow noreferrer\"><code>tf.merge_summary()<\/code><\/a> op in your example:<\/p>\n\n<pre><code>accuracy_summary = tf.scalar_summary(\"accuracy\", accuracy)\nloss_summary = tf.scalar_summary(\"loss\", C)\n\nmerged = tf.merge_summary([accuracy_summary, loss_summary])\n<\/code><\/pre><\/li>\n<li><p>Ensure that each time you create a new set of summaries, you do so in a new graph. The recommended style is to use an explicit default graph:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n  # Build model and create session in this scope.\n  #\n  # Only summary nodes created in this scope will be returned by a call to\n  # `tf.merge_all_summaries()`\n<\/code><\/pre>\n\n<p>Alternatively, if you are using the latest open-source version of TensorFlow (or the forthcoming 0.7.0 release), you can call <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#reset_default_graph\" rel=\"nofollow noreferrer\"><code>tf.reset_default_graph()<\/code><\/a> to reset the state of the graph and remove any old summary nodes.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35414009,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-15T16:12:36.647","AcceptedAnswerId":35420394.0,"OwnerUserId":2615676.0,"Title":"Multiple queues causing TF to lock up","Body":"<p>I'm trying to use multiple queues for reading and batching, but this is causing TF to occasionally lock up. Here's some sample code:<\/p>\n\n<pre><\/pre>\n\n<p>TensorFlow occasionally locks up at the last line, when I actually try to run something. This behavior is rather hard to reproduce using the above code however. In 1000+ runs, I could only get it to hang once. In my real code, the actual reader is more complicated, and it's using TFRecords, but otherwise everything is the same. There it hangs up 2\/3 of the time with 3 queues in total. With 5 queues it seemingly never runs, and with 1 queue it seemingly never hangs. This is on a Mac with 0.6. I have a different system running Ubuntu, also with 0.6, and I get the same problem (although the frequency of locking up is much higher on the Ubuntu system).<\/p>\n\n<p>UPDATE: A more accurate estimate of how often the above code locks up is 1 in 5,000 trials.<\/p>\n","answers":[{"AnswerId":"35420394","CreationDate":"2016-02-15T22:38:11.743","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This is probably caused by not having enough operation threads. If you have a queue runner 1 depending on work of queue runner 2, and you run them asynchronously, then you'll need at least two op threads, set through <code>inter_op_parallelism_threads<\/code>, to guarantee that progress is being made.<\/p>\n\n<p>In your case, you have queue runner that's filling <code>batch<\/code> thread depending on <code>string_input_producer<\/code> queue being not empty. If the queue runner associated with <code>string_input_producer<\/code> queue runs first, then everything is fine. But if <code>batch<\/code> queue runner is scheduled first, it will get stuck in <code>string_input_producer.dequeue<\/code> op waiting for <code>string_input_producer<\/code> queue to get some filenames. Since there's only 1 thread in TensorFlow op thread pool, the <code>enqueue<\/code> op of <code>string_input_producer<\/code> will never get allocated a thread to complete (ie, to execute its <code>Compute<\/code> method)<\/p>\n\n<p>Simplest solution is to have at least as many operation threads as you have simultaneous <code>run<\/code> calls (ie, number of queues + 1). If you really want to restrict yourself to one thread, you could preload filename queue file filenames synchronously using main thread. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>  coordinator = tf.train.Coordinator()\n\n  import glob\n  files = glob.glob('\/temp\/pipeline\/*')\n  if FLAGS.preload_filenames:\n    file_queue = tf.FIFOQueue(capacity=len(files), dtypes=tf.string)\n    enqueue_val = tf.placeholder(dtype=tf.string)\n    enqueue_op = file_queue.enqueue(enqueue_val)\n  else:\n    file_queue = tf.train.string_input_producer(files)\n\n  reader = tf.TextLineReader()\n  key, serialized_example = reader.read(file_queue)\n  keys, serialized_examples = tf.train.batch([key, serialized_example], 5,\n                                             capacity=10)\n\n  initializer = tf.initialize_all_variables()\n\n  session = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=1,\n                                             intra_op_parallelism_threads=1))\n  print 'running initializer'\n  session.run(initializer)\n\n  if FLAGS.preload_filenames:\n    print 'preloading filenames'\n    for fn in files:\n      session.run([enqueue_op], feed_dict={enqueue_val: fn})\n      print 'size - ', session.run([file_queue.size()])\n    session.run([file_queue.close()])\n\n  print 'starting queue runners'\n  threads = tf.train.start_queue_runners(sess=session, coord=coordinator)\n  print 'about to run session'\n  print session.run(keys)\n<\/code><\/pre>\n\n<p>Code above will need some encapsulation if you have more than one filenames queue. Alternatively here's a hacky work-around which should work if there's exactly <code>prebuffer_amount<\/code> filenames for all input_producer queues<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>queue_runners=tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)\nfilename_queue_runners=[qr for qr in queue_runners if 'input_producer' in qr.name]\nfor qr in filename_queue_runners:\n  for k in prebuffer_amount:\n    sess.run(qr._enqueue_ops[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":35414824,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><tensorflow-serving>","CreationDate":"2016-02-15T16:53:26.363","AcceptedAnswerId":35438859.0,"OwnerUserId":64106.0,"Title":"TensorFlow in production for real time predictions in high traffic app - how to use?","Body":"<p>What is the right way to use TensorFlow for real time predictions in a high traffic application.<\/p>\n\n<p>Ideally I would have a server\/cluster running tensorflow listening on a port(s) where I can connect from app servers and get predictions similar to the way databases are used.\nTraining should be done by cron jobs feeding the training data through the network to the same server\/cluster.<\/p>\n\n<p>How does one actually use tensorflow in production? Should I build a setup where the python is running as a server and use the python scripts to get predictions? I'm still new to this but I feel that such script will need to open sessions etc.. which is not scalable. (I'm talking about 100s of predictions\/sec).<\/p>\n\n<p>Any pointer to relevant information will be highly appreciated. I could not find any.<\/p>\n","answers":[{"AnswerId":"35438859","CreationDate":"2016-02-16T17:13:04.000","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This morning, our colleagues released <a href=\"http:\/\/tensorflow.github.io\/serving\/\">TensorFlow Serving<\/a> on GitHub, which addresses some of the use cases that you mentioned. It is a distributed wrapper for TensorFlow that is designed to support high-performance serving of multiple models. It supports both bulk processing and interactive requests from app servers.<\/p>\n\n<p>For more information, see the <a href=\"http:\/\/tensorflow.github.io\/serving\/serving_basic\">basic<\/a> and <a href=\"http:\/\/tensorflow.github.io\/serving\/serving_advanced\">advanced<\/a> tutorials.<\/p>\n"}]}
{"QuestionId":35414942,"AnswerCount":0,"Tags":"<python><numpy><theano>","CreationDate":"2016-02-15T16:59:06.930","AcceptedAnswerId":null,"OwnerUserId":5931014.0,"Title":"Looping through Tensor Theano","Body":"<p>I am a beginner in theano and I'm searching a way to loop through a Tensor.\nI can develop the functionality with numpy array's, but my code with Theano won't work. <\/p>\n\n<p>numpy: code<\/p>\n\n<pre><\/pre>\n\n<p>What I tried and doesn't work<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35418410,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-02-15T20:25:34.647","AcceptedAnswerId":null,"OwnerUserId":3314143.0,"Title":"Tensorflow: train two layer streams separately?","Body":"<p>Let's assume I have a convolutional neural network which should predict two different (semantically) things out of images which can be classified in a N-dimensional output each. So my network looks like this:<\/p>\n\n<pre><\/pre>\n\n<p> is a vector of dimension , therefore predicting over 1000 classes.  is a vector of dimension , therefore predicting over 500 classes.<\/p>\n\n<p>The classes in x and y are semantically related, so they are <em>not<\/em> one-hot encoded. Instead, a normal distribution around the real class is fitted in the training vectors.<\/p>\n\n<p>With every training step I am minimizing the cross entropy error like this:<\/p>\n\n<pre><\/pre>\n\n<p>So I'm calculating different cross entropy errors for  and .<\/p>\n\n<p>Up to now I am satisfied with the (training) results of the network. However, the network architecture is not final. It will get bigger and eventually deeper and the training data set will be several orders of magnitude bigger as it is by now.<\/p>\n\n<p><strong>My question is<\/strong>: Is this a reasonable network architecture? Or is it considered rather a <em>bad design<\/em>?<\/p>\n\n<p><strong>Is Tensorflow really training the two streams separately<\/strong>? Or are my two AdamOptimizers per learning step <em>bad design<\/em> and am I just getting \"magically\" good looking results although my architecture wouldn't be considered \"good\" for this problem?<\/p>\n\n<p>I thought two <em>different<\/em> fully-connected \"streams\" for x and y data\/classes would be reasonable for an architecture which should predict between two different \"sets\" of classes.<\/p>\n","answers":[]}
{"QuestionId":35420275,"AnswerCount":1,"Tags":"<python><colors><neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-02-15T22:29:11.667","AcceptedAnswerId":35420717.0,"OwnerUserId":3314143.0,"Title":"Tensorflow weights for kernels of convolution for colored images?","Body":"<p>At the moment I have some networks doing classification stuff with greyscaled images. I want to move on to colored (RGB) images.<\/p>\n\n<p>In the CIFAR-10 tutorial of Tensorflow I got confused by the weights for the convolution kernels. The first convolution there looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>So it is a  convolution with an input of 3 (one for each color channel: the red, green and blue image information) and it is generating 64 feature maps.<\/p>\n\n<p>However, the second convolution layer takes an input of 64 feature maps:<\/p>\n\n<pre><\/pre>\n\n<p>...so, how does this process the color information? Does this mean the different color channels are somehow <em>spread<\/em> on the 64 feature maps of convolution layer 1?<\/p>\n\n<p>I thought conv layer 1 produces 64 feature maps for each color channel, therefore ending up in 3 * 64 = 196 feature maps...but obviously I were wrong.<\/p>\n\n<p>How is the color information <em>mixed<\/em> there in conv layer 1?<\/p>\n","answers":[{"AnswerId":"35420717","CreationDate":"2016-02-15T23:02:17.193","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>See equation 3 in description of CuDNN <a href=\"http:\/\/arxiv.org\/pdf\/1410.0759.pdf\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p>Basically for a single example (<code>n<\/code>), single row (<code>p<\/code>) and single column (<code>q<\/code>), the result of spatial convolution will be a weighted sum of <code>5x5x3<\/code> values. So each activation will contain information from all 3 colors.<\/p>\n"}]}
{"QuestionId":35421385,"AnswerCount":0,"Tags":"<caffe><pycaffe>","CreationDate":"2016-02-16T00:04:09.390","AcceptedAnswerId":null,"OwnerUserId":4386025.0,"Title":"caffe - Input data, train and test set","Body":"<p><strong>RAW data dim :<\/strong><\/p>\n\n<p>1)  in  format... ---> Class1<\/p>\n\n<p> in  format... ---> labels <\/p>\n\n<p>2)  in  format... ---> Class2<\/p>\n\n<p> in  format... ---> labels <\/p>\n\n<p>3)  in  format... ---> Class3<\/p>\n\n<p> in  format... ---> labels <\/p>\n\n<p><strong>How to use\/convert it for caffe as a input... using python?<\/strong><\/p>\n","answers":[]}
{"QuestionId":35422047,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-02-16T01:19:37.100","AcceptedAnswerId":35432693.0,"OwnerUserId":5870277.0,"Title":"Matrix value not needed for Lop?","Body":"<p>In the theano derivatives tutorial here:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html#tutcomputinggrads\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/gradients.html#tutcomputinggrads<\/a><\/p>\n\n<p>the example of Lop works without an explicit value of the W matrix in the dot product. And, in fact, the partial derivatives in this case do remove the values of the components of W so they are not needed. <\/p>\n\n<p>But, attempting a similar thing with the Rop throws an error: <\/p>\n\n<p>theano.gof.fg.MissingInputError: (\"An input of the graph, used to compute dot(Elemwise{second,no_inplace}.0, ), was not provided and not given a value.<\/p>\n\n<p>How is this different? <\/p>\n","answers":[{"AnswerId":"35432693","CreationDate":"2016-02-16T12:31:18.583","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Theano will try to optimize the computation graph, but it does not always work.  <\/p>\n\n<p>In the Lop example, Theano can detect that we don't actually need that W, but when changed to the Rop it just can't.<\/p>\n\n<p>The Lop example:<\/p>\n\n<pre><code>W = T.dmatrix('W')\nv = T.dvector('v')\nx = T.dvector('x')\ny = T.dot(x, W)\nVJ = T.Lop(y, W, v)\nf = theano.function([v, x], VJ)\nf([2, 2], [0, 1])\n<\/code><\/pre>\n\n<p>If I just change <code>y = T.dot(x, W)<\/code> to <code>y = T.dot(x, W**1)<\/code>, Theano will fail to do the optimization and throw the same error message at me say that I did not provide enough parameters.  <\/p>\n\n<p>Actually in the Rop example, if we change the values given to W, it does not affect the result at all, because Theano failed to optimize that.<\/p>\n\n<p>p.s. I find the Theano documents very unclear sometimes.<\/p>\n"}]}
{"QuestionId":35422388,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-16T02:02:45.810","AcceptedAnswerId":35436452.0,"OwnerUserId":5703903.0,"Title":"What is Bazel in TensorFlow? When do I need to build again?","Body":"<p>I'm new to Bazel. I'm not sure how this thing works. On the <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/get_started\/os_setup.html\" rel=\"noreferrer\">TF website<\/a>, there's this section on \"Create the pip package and install\".<\/p>\n\n<pre><\/pre>\n\n<p>Here's the situation:<\/p>\n\n<ol>\n<li>There's a new commit on the master branch of TensorFlow and I merge it into my fork.<\/li>\n<li>I need to rebuild the wheel and do a pip install of the new wheel (correct me if I am wrong).<\/li>\n<li>I .\/configure first, then bazel build, then bazel-bin, then pip install.<\/li>\n<\/ol>\n\n<p>Is this the correct way to properly update changes from master? The bazel build step takes a really long time.<\/p>\n","answers":[{"AnswerId":"35436452","CreationDate":"2016-02-16T15:26:46.403","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p><a href=\"http:\/\/bazel.io\" rel=\"noreferrer\">Bazel<\/a> is a build tool just like other build tools like <a href=\"http:\/\/cmake.org\" rel=\"noreferrer\">cmake<\/a> and <a href=\"https:\/\/www.gnu.org\/software\/make\/\" rel=\"noreferrer\">make<\/a>. The steps you listed is the correct way to get updates from master. The build step could take long the first time you build TensorFlow. Later builds, after updates from master, should be faster, as Bazel, just like any other build tool, doesn't re-build targets whose dependencies have not been modified.<\/p>\n"}]}
{"QuestionId":35423309,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-02-16T03:44:31.790","AcceptedAnswerId":null,"OwnerUserId":1270543.0,"Title":"How reconstruct the caffe net by using pycaffe","Body":"<p>What I want is, After loading a net,  I will decompose some certain layers and save the new net. For example<\/p>\n\n<p>Orignial net: <\/p>\n\n<blockquote>\n  <p>data -> conv1 -> conv2 -> fc1 -> fc2 -> softmax;<\/p>\n<\/blockquote>\n\n<p>New net:<\/p>\n\n<blockquote>\n  <p>data -> conv1_1 -> conv1_2 -> conv2_1 -> conv2_2 -> fc1 -> fc2 -> softmax<\/p>\n<\/blockquote>\n\n<p>Therefore, during this process, I stuck in the following situation:<br>\n1. How to new a certain layer with specified layer parameters in ?<br>\n2. How to copy the layer parameters from existing layers(such as  and  above)?<\/p>\n\n<p>I know by using , we can define a new net manually. But  can not specify a layer from a existing one(e.g: ). <\/p>\n","answers":[{"AnswerId":"35469774","CreationDate":"2016-02-17T23:23:18.610","ParentId":null,"OwnerUserId":"5352059","Title":null,"Body":"<p>I didn't see how to load in previous nets with net_spec but you can always use the protobuf objects directly.\n(I use your network structure as an example)<\/p>\n\n<pre><code>import caffe.proto.caffe_pb2 as caffe_pb2\nimport google.protobuf as pb\nfrom caffe import layers as L\n\nnet = caffe_pb2.NetParameter()\nwith open('net.prototxt', 'r') as f:\n    pb.text_format.Merge(f.read(), net)\n\n#example of modifing the network:\nnet.layer[1].name = 'conv1_1'\nnet.layer[1].top[0] = 'conv1_1'\nnet.layer[2].name = 'conv1_2'\nnet.layer[2].top[0] = 'conv1_2'\nnet.layer[2].bottom[0] = 'conv1_1'\n\nnet.layer[3].bottom[0] = 'conv2_2'\n\n#example of adding new layers (using net_spec):\nconv2_1 = net.layer.add()\nconv2_1.CopyFrom(L.Convolution(kernel_size=7, stride=1, num_output=48, pad=0).to_proto().layer[0])\nconv2_1.name = 'conv2_1'\nconv2_1.top[0] = 'conv2_1'\nconv2_1.bottom.add('conv1_2')\n\nconv2_2 = net.layer.add()\nconv2_2.CopyFrom(L.Convolution(kernel_size=7, stride=1, num_output=48, pad=0).to_proto().layer[0])\nconv2_2.name = 'conv2_2'\nconv2_2.top[0] = 'conv2_2'\nconv2_2.bottom.add('conv2_1')\n\n# then write back out:\nwith open('net2.prototxt, 'w') as f:\n    f.write(pb.text_format.MessageToString(net))\n<\/code><\/pre>\n\n<p>Also see <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/reference\/python-generated\" rel=\"noreferrer\">here<\/a> as a guide on protocol buffers in python and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\" rel=\"noreferrer\">here<\/a> for the current caffe message formats.<\/p>\n"}]}
{"QuestionId":35425766,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-02-16T06:57:02.927","AcceptedAnswerId":null,"OwnerUserId":859840.0,"Title":"Implementing Coordinate Descent with mini-batches","Body":"<p>I am trying to implement the Learned Coordinate Descent Algorithm (LCoD) from <a href=\"http:\/\/www.cs.nyu.edu\/~kgregor\/gregor-icml-10.pdf\" rel=\"nofollow\">this paper<\/a> by Karol Gregor and Yann Lecun with TensorFlow. The algorithm shows how to perform the forward pass (and backward propagation) with mini-batch size=1. I would like to extend this to larger mini-batch sizes.<\/p>\n\n<p>In each step, (greedy) coordinate descent chooses the index of the largest component of , where  is the sparse code (), and  is the proximal operator, such as the soft-thresholding function.  is the number of components\/atoms in the dictionary. In the paper, , but I want to generalize this to .<\/p>\n\n<p>In the update step of algorithm 5:<\/p>\n\n<pre><\/pre>\n\n<p> is the index of the largest component from above. This step is not easy to  perform since index  is different for every example in the mini-batch (again, not a problem with , but problematic with ).<\/p>\n\n<p>A quick-fix would be to split the mini-batch into  rows, feed each row independently into the LCoD algorithm, then concatenate all  rows back into a full-sized matrix. However, I was wondering if there is a more elegant solution. On both CPU\/GPU, I find the forward\/back-propagation of this quick-fix to be very, very slow.<\/p>\n","answers":[]}
{"QuestionId":35426162,"AnswerCount":1,"Tags":"<theano><deep-learning><convolution><caffe><mnist>","CreationDate":"2016-02-16T07:19:24.997","AcceptedAnswerId":35434809.0,"OwnerUserId":3870344.0,"Title":"Caffe vs Theano MNIST example","Body":"<p>I'm trying to learn (and compare) different deep learning frameworks, by the time they are Caffe and Theano.<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html<\/a><\/p>\n\n<p>and <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html<\/a><\/p>\n\n<p>I follow the tutorial to run those frameworks on MNIST dataset. However, I notice a quite difference in term of accuracy and performance.<\/p>\n\n<p>For Caffe, it's extremely fast for the accuracy to build up to ~97%. In fact, it only takes 5 mins to finish the program (using GPU) which the final accuracy on test set of over 99%. How impressive!<\/p>\n\n<p>However, on Theano, it is much poorer. It took me more than 46 minutes (using same GPU), just to achieve 92% test performance.<\/p>\n\n<p>I'm confused as it should not have so much difference between the frameworks running relatively same architectures on same dataset.<\/p>\n\n<p>So my question is. Is the accuracy number reported by Caffe is the percentage of correct prediction on test set? If so, is there any explanation for the discrepancy?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"35434809","CreationDate":"2016-02-16T14:11:57.337","ParentId":null,"OwnerUserId":"2490541","Title":null,"Body":"<p>The examples for Theano and Caffe are not exactly the same network. Two key differences which I can think of are that the Theano example uses sigmoid\/tanh activation functions, while the Caffe tutorial uses the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Rectifier_(neural_networks)\">ReLU<\/a> activation function, and that the Theano code uses normal <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent\">minibatch gradient descent<\/a> while Caffe uses a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent#Momentum\">momentum optimiser<\/a>. Both differences will significantly affect the training time of your network. And using the ReLU unit will likely also affect the accuracy.  <\/p>\n\n<p>Note that Caffe is a deep learning framework which already has ready-to-use functions for many commonly used things like the momentum optimiser. Theano, on the other hand, is a symbolic maths library which can be used to build neural networks. However, it is not a deep learning framework. <\/p>\n\n<p>The Theano tutorial you mentioned is an excellent resource to understand how exactly convolutional and other neural networks work on a basic level. However, it will be cumbersome to implement all the state-of-the-art tweaks. If you want to get state-of-the-art results quickly you are better off using one of the existing deep learning frameworks. Apart from Caffe, there are a number of frameworks based on Theano. I know of <a href=\"http:\/\/keras.io\/\">keras<\/a>, <a href=\"https:\/\/blocks.readthedocs.org\/en\/latest\/\">blocks<\/a>, <a href=\"http:\/\/deeplearning.net\/software\/pylearn2\/\">pylearn2<\/a>, and my personal favourite <a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/\">lasagne<\/a>. <\/p>\n"}]}
{"QuestionId":35428297,"AnswerCount":1,"Tags":"<gpu><matrix-multiplication><tensorflow><invalid-argument>","CreationDate":"2016-02-16T09:15:54.950","AcceptedAnswerId":null,"OwnerUserId":5933970.0,"Title":"Tensorflow issue with GPU on matmul. GPU isn't recognized","Body":"<p>I installed tensorflow with gpu, cuda 7.0 and cudnn 6.5. When I import tensorflow it works well.<\/p>\n\n<p>I am trying to run a simple matrix multiplication on Tensorflow and it doesn't want to use my gpu though it seems to recognize it. I have this issue on my computer with a nvidia geforce 970m and on a cluster with two titan Z. <\/p>\n\n<p>My first code is :<\/p>\n\n<pre><\/pre>\n\n<p>This code works and the result is :<\/p>\n\n<pre><\/pre>\n\n<p>So in my way, tensorflow uses my gpu to create constant but not for matmul (that is weird). Then, I force the gpu like this :<\/p>\n\n<pre><\/pre>\n\n<p>And Tensorflow returns :<\/p>\n\n<pre><\/pre>\n\n<p>If someone have the same problem or an idea, I will be glad to read your answer !<\/p>\n","answers":[{"AnswerId":"35443287","CreationDate":"2016-02-16T21:24:54.080","ParentId":null,"OwnerUserId":"1940138","Title":null,"Body":"<p>I do not have enough reputation to comment, I have come across a similar issue, my question is here <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/35443080\/tensorflow-critical-graph-operations-assigned-to-cpu-rather-than-gpu\">TensorFlow: critical graph operations assigned to cpu rather than gpu<\/a><\/p>\n"}]}
{"QuestionId":35429468,"AnswerCount":1,"Tags":"<python><linux><amazon-web-services><caffe>","CreationDate":"2016-02-16T10:07:33.340","AcceptedAnswerId":null,"OwnerUserId":5082576.0,"Title":"Running Caffe creates high load over ksoftirqd\/0","Body":"<p>I'm running Caffe using python on AWS. <\/p>\n\n<p>The scripts uses the GPU, uploads an existing model, and checks the output of it per image URL. <\/p>\n\n<p>On the first few tries the script ran well. Than it stuck on a few different phases at each time. <\/p>\n\n<p>Using 'top' I could see that  gets about 93% of the CPU when the process is stuck. <\/p>\n\n<p>I don't think there is a bug in my script, because originally it ran well no the server. When I reboot the server, sometimes it helps. But later we get the same problem. <\/p>\n\n<p>Killing all python process on the server doesn't help. Only rebootting it. <\/p>\n\n<p>Any ideas what I can do here?<\/p>\n","answers":[{"AnswerId":"35432099","CreationDate":"2016-02-16T12:02:41.020","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you are experiencing a <a href=\"https:\/\/askubuntu.com\/a\/7919\">very high network load<\/a>.<br>\nWhat exactly are you trying to download from URLs?<br>\nAre there any other processes running at the same time on the machine?<\/p>\n\n<p>It is difficult to diagnose your problem without the specifics of your script.<\/p>\n"}]}
{"QuestionId":35431460,"AnswerCount":3,"Tags":"<python><linux><python-import><theano><keras>","CreationDate":"2016-02-16T11:33:10.307","AcceptedAnswerId":45907083.0,"OwnerUserId":3990607.0,"Title":"Import keras.datasets not working","Body":"<p>I have keras installed on my linux machine, but when I try to import a dataset from the keras.datasets I get an error that it cannot find it.<\/p>\n\n<p>So for example:<\/p>\n\n<pre><\/pre>\n\n<p>I get the error<\/p>\n\n<blockquote>\n  <p>ImportError: No module named keras.datasets<\/p>\n<\/blockquote>\n\n<p>I installed keras using  and it installed successfully.<\/p>\n","answers":[{"AnswerId":"35434208","CreationDate":"2016-02-16T13:42:17.093","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>Thanks to the comment from Selcuk which got me on the right track.<\/p>\n\n<p>Indeed the problem was that I had multiple versions of Python. \nI followed some online instructions to installing Keras which recommended installing MiniConda\/Conda\/Anaconda which is its own version of python. So I had two Python2.7 versions installed:<\/p>\n\n<ul>\n<li>Normal Linux Python 2.7<\/li>\n<li>Anaconda Python 2.7<\/li>\n<\/ul>\n\n<p>Removing Anaconda Python and installing all libraries using <code>pip<\/code> \/ <code>apt-get<\/code> instead of <code>conda<\/code> solved my problem. <\/p>\n"},{"AnswerId":"45907083","CreationDate":"2017-08-27T16:33:57.770","ParentId":null,"OwnerUserId":"8508686","Title":null,"Body":"<blockquote>\n  <p>Indeed the problem was that I had multiple versions of Python.<\/p>\n  \n  <p>Removing Anaconda Python and installing all libraries using pip \/ apt-get instead of conda solved my problem.<\/p>\n<\/blockquote>\n\n<p>I found this to be true but it is not necessary to delete anaconda.\nI had the same issue but with multiple python versions. However, i created an environment that only used the Anaconda version (while in that environment). In terminal (on mac and other suitable terminals), type\/copy<\/p>\n\n<p><code>conda create -n dataweekends python=2.7 pandas scikit-learn jupyter matplotlib<\/code><\/p>\n\n<p>dataweekends is simply the name of the environment you created. To access this, just use the command <\/p>\n\n<p><code>source activate dataweekends<\/code><\/p>\n\n<p>Be mindful that you might (probably) have to reinstall dependencies once in that new environment.<\/p>\n\n<p>I got this trick from here \"<a href=\"https:\/\/www.dataweekends.com\/blog\/2017\/03\/09\/set-up-your-mac-for-deep-learning-with-python-keras-and-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/www.dataweekends.com\/blog\/2017\/03\/09\/set-up-your-mac-for-deep-learning-with-python-keras-and-tensorflow<\/a>\"<\/p>\n\n<p>I would also recommend setting up different environments for each project you do in python.<\/p>\n"},{"AnswerId":"42609531","CreationDate":"2017-03-05T14:24:38.440","ParentId":null,"OwnerUserId":"6256888","Title":null,"Body":"<p>Do you have <code>keras.py<\/code> or <code>keras.pyc<\/code> in the current working directory? If so, this will mess up the imports. Try <strong>renaming the file<\/strong> and\/or <strong>deleting <code>keras.pyc<\/code><\/strong>.<\/p>\n"}]}
{"QuestionId":35432545,"AnswerCount":1,"Tags":"<python><server><tensorflow>","CreationDate":"2016-02-16T12:24:12.210","AcceptedAnswerId":35435955.0,"OwnerUserId":4919612.0,"Title":"Deploy Tensorflow on a server without installing using Slurm Manager","Body":"<p>I would like to install Tensorflow on a High Performance Computing Cluster that use Slurm Manager. I do not have The admin right so I want to deploy Tensorflow without installing it. However I cannot find any informations about this. Does anybody know how to do this? Thanks a lot!<\/p>\n","answers":[{"AnswerId":"35435955","CreationDate":"2016-02-16T15:05:22.077","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to install TensorFlow if you do not have admin (root) privileges is to install it in a virtualenv:<\/p>\n\n<pre><code>$ virtualenv --system-site-packages ~\/tensorflow\n$ source ~\/tensorflow\/bin\/activate\n(tensorflow)$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.6.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/get_started\/os_setup.html#virtualenv_install\" rel=\"nofollow\">Full instructions<\/a> are available on the TensorFlow website.<\/p>\n"}]}
{"QuestionId":35435852,"AnswerCount":0,"Tags":"<lua><torch><conv-neural-network>","CreationDate":"2016-02-16T15:00:35.913","AcceptedAnswerId":null,"OwnerUserId":5861164.0,"Title":"bad argument #1 to 'updateOutput' (2D or 3D(batch mode) tensor expected)","Body":"<p>I'm trying torch since last week and I have a problem with the temporal convolution.\nI'm working on the UCR datasets (see <a href=\"http:\/\/www.cs.ucr.edu\/~eamonn\/time_series_data\/\" rel=\"nofollow\">http:\/\/www.cs.ucr.edu\/~eamonn\/time_series_data\/<\/a>) and when I try a convolution, I have this error:<\/p>\n\n<pre><\/pre>\n\n<p>I have this model in lua:<\/p>\n\n<pre><\/pre>\n\n<p>I train my model with  (but manually, I have the same error).\nI understand how convolution works but I don't know how solve this.\nI suppose I have to resize my input but how ?\nIf you can help me, I will be thankful !!<\/p>\n\n<p>Thanks,\nArthur<\/p>\n\n<p>ps: sorry for my english<\/p>\n","answers":[]}
{"QuestionId":35436078,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-16T15:11:27.173","AcceptedAnswerId":35437050.0,"OwnerUserId":4077588.0,"Title":"Tensorflow read serialized data example undefined tf.FixedLenFeature","Body":"<p>I am trying to use serialized data using the proto interface as suggested here <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#reading-from-files\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#reading-from-files<\/a><\/p>\n\n<p>I try to use the example : \n<a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py<\/a><\/p>\n\n<p>But it fails because : <\/p>\n\n<pre><\/pre>\n\n<p>I guess I am missing something here... <\/p>\n","answers":[{"AnswerId":"35437050","CreationDate":"2016-02-16T15:52:38.687","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The version of <code>fully_connected_reader.py<\/code> includes some (breaking) <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/492e55ff2297dc13f5ed540056ad40603f2b1dd0\" rel=\"nofollow\">changes to the <code>tf.parse_example()<\/code> API<\/a> that were made after the TensorFlow 0.6.0 release. These changes included adding the <code>tf.FixedLenFeature<\/code> class as a helper for defining the schema to be used in <code>tf.parse_example()<\/code>. You should build TensorFlow from source or wait for the upcoming 0.7.0 release to use this version of the API.<\/p>\n\n<p>Alternatively, the old version of the example code is available <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/cd53f3c3302c9312c1840389a9988a879b8b9dd5\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow\">here<\/a>, and the documentation for <code>tf.parse_example()<\/code> is available <a href=\"https:\/\/www.tensorflow.org\/versions\/v0.6.0\/api_docs\/python\/io_ops.html#parse_example\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":35436970,"AnswerCount":2,"Tags":"<python><neural-network><gpu><deep-learning><caffe>","CreationDate":"2016-02-16T15:49:04.573","AcceptedAnswerId":null,"OwnerUserId":310242.0,"Title":"Can you train multiple networks on the same data using Caffe?","Body":"<p>I am writing a calibration pipeline to learn the hyperparameters for neural networks to detect properties of DNA sequences*. This therefore requires training a large number of models on the same dataset with different hyperparameters.<\/p>\n\n<p>I am trying to optimise this to run on GPU. DNA sequence datasets are quite small compared to image datasets (typically 10s or 100s of base-pairs in 4 'channels' to represent the 4 DNA bases, A, C, G and T, compared to 10,000s of pixels in 3 RGB channels), and consequently cannot make full use of the parallelisation on a GPU unless multiple models are trained at the same time.<\/p>\n\n<p>Is there a way to do this in Caffe in Python?<\/p>\n\n<p>(I previously <a href=\"https:\/\/stackoverflow.com\/questions\/35135331\/how-can-you-train-multiple-neural-networks-simultaneously-in-nolearn-lasagne-the\">asked this question<\/a> with reference to doing this in nolearn, lasagne or Theano, but I'm not sure it's possible so have moved on to Caffe.)<\/p>\n\n<p>* It's based on the <a href=\"http:\/\/www.nature.com\/nbt\/journal\/v33\/n8\/full\/nbt.3300.html\" rel=\"nofollow noreferrer\">DeepBind model<\/a> for detecting where transcription factors bind to DNA, if you're interested.<\/p>\n","answers":[{"AnswerId":"35802883","CreationDate":"2016-03-04T17:55:37.447","ParentId":null,"OwnerUserId":"2466336","Title":null,"Body":"<p>Yes. The following options are available for this.<\/p>\n\n<p><strong>Option 1: Parallel training processes<\/strong>. Write a network definition for each hyper-parameter configuration and run a training process for each.<\/p>\n\n<p>Pros: No need to manage resources yourself.\nCons: If all networks are working off of the same data, you'll be paying the cost of data copy from CPU to GPU RAM multiple times. This can get highly redundant. I've noticed this slowing down so much that I might as well have trained the different network sequentially.<\/p>\n\n<p><strong>Option 2: Merge different network definitions into a single prototxt<\/strong> and train with a single solver. The layer and blob names have to be unique for the sub-networks to be independent.<\/p>\n\n<p><a href=\"https:\/\/github.com\/kashefy\/caffe_sandbox\/blob\/master\/examples\/parallel_train\/verify_parallel_training.ipynb\" rel=\"nofollow\">This ipython notebook<\/a> provides a demo of this. The merging of networks into the prototxt for you. The notebook also demonstrates that by keeping the layer and blob names unique fo each sub network, the combined training process will reach the same results as if you trained each independently.<\/p>\n"},{"AnswerId":"35464461","CreationDate":"2016-02-17T18:11:36.900","ParentId":null,"OwnerUserId":"5893783","Title":null,"Body":"<p>Maybe I haven't understood the question but to train multiple networks on whatever framework you just have to run a different process for each training (if you have more than one GPU, set the flags in order to that each prices use a different one).<\/p>\n\n<p>With one GPU, in image classification the main problem we usually have is the lack of memory (a part of the obvious ones) but presuming that you won't have that, you should be able run both trainings over the same GPU. Because the context switching between processes in the GPU, you're going to get a slow-down in the training performance. From my experience, this slow-down was by a factor of x2-x4 so usually is not worth it (but check anyway).<\/p>\n"}]}
{"QuestionId":35438160,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-02-16T16:42:51.580","AcceptedAnswerId":35449401.0,"OwnerUserId":4229825.0,"Title":"Caffe does not make snapshots on SIGINT","Body":"<p>When I'm pressing <kbd>CTRL<\/kbd>+<kbd>C<\/kbd> in terminal, caffe stops training but does not make snapshots. How to fix it?\nMy solver:<\/p>\n\n<pre><\/pre>\n\n<p>Bash script:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42217315","CreationDate":"2017-02-14T03:21:46.130","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>A good way to log caffe output is <\/p>\n\n<pre><code>GLOG_log_dir=\/path\/to\/log\/dir $CAFFE_ROOT\/bin\/caffe.bin train \n\u2014solver=\/path\/to\/solver.prototxt\n<\/code><\/pre>\n\n<p>This does live logging of caffe output and <code>SIGINT<\/code> definitely reaches caffe.<\/p>\n"},{"AnswerId":"35449401","CreationDate":"2016-02-17T06:28:17.607","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Redirecting caffe's output through <code>tee<\/code> and pipes might alter the way the OS handles and transfers signals to processes. Try avoiding <code>| tee<\/code> to make sure <code>SIGINT<\/code> reaches caffe.<\/p>\n\n<p>Note that <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/caffe.cpp#L51-L56\" rel=\"nofollow noreferrer\"><code>caffe<\/code> tool<\/a> has two flags<\/p>\n\n<pre><code>DEFINE_string(sigint_effect, \"stop\",\n             \"Optional; action to take when a SIGINT signal is received: \"\n              \"snapshot, stop or none.\");\nDEFINE_string(sighup_effect, \"snapshot\",\n             \"Optional; action to take when a SIGHUP signal is received: \"\n             \"snapshot, stop or none.\");\n<\/code><\/pre>\n\n<p>These flags can help you define caffe's behavior on <a href=\"https:\/\/en.wikipedia.org\/wiki\/Unix_signal#SIGINT\" rel=\"nofollow noreferrer\">SIGINT<\/a> and <a href=\"https:\/\/en.wikipedia.org\/wiki\/Unix_signal#SIGHUP\" rel=\"nofollow noreferrer\">SIGHUP<\/a>.<\/p>\n"}]}
{"QuestionId":35438808,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-02-16T17:11:09.853","AcceptedAnswerId":35440281.0,"OwnerUserId":2515942.0,"Title":"Is there an arbitary limit on the size of vectors you can put into session.run?","Body":"<p>I recently started learning tensorflow, I was doing the first tutorial on tensorflow website and I encountered a problem.<\/p>\n\n<p>I am using tensorflow on Cloud 9.<\/p>\n\n<p>The starting code has no problems:<\/p>\n\n<pre><\/pre>\n\n<p>In the training part of the tutorial, it says it is faster to get batches of 100 images from the whole training set of 55000 images. The code for that is:<\/p>\n\n<pre><\/pre>\n\n<p>This works and I get the %91 accuracy as the tutorial says.<\/p>\n\n<p>Then I tried to send the whole set into training as the tutorial suggested was computationally expensive. I kept the iterations low to see what actually happens.<\/p>\n\n<pre><\/pre>\n\n<p>However, whatever the number of iterations I always got 0.098 accuracy.<\/p>\n\n<p>Then I tried another way to achieve the same result.<\/p>\n\n<pre><\/pre>\n\n<p>In this example, it iterates over the whole set 50 by 50. And this method works up to 200 by 200. After 200, whatever the number is I get 0.098 accuracy.<\/p>\n\n<p>So as I understand there is an arbitary limit on the number of items you can send into training. Is is true or am I missing something here?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"35440281","CreationDate":"2016-02-16T18:25:51.417","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>There's no limit on number of examples, but learning rate (<code>0.01<\/code> in your case) interacts with batch size. If you make your batch size 10x larger, your gradients are 10x larger and you take 10x longer steps. This can end up overshooting and increasing your objective at each step instead of decreasing it. You need to tune your learning rate (try diving by square root of batch size multiplier), or use an adaptive descent method (<code>AdamOptimizer<\/code> will automatically adjust learning rate over time) <\/p>\n"}]}
{"QuestionId":35438967,"AnswerCount":1,"Tags":"<python><theano><mcmc><pymc3><mixture-model>","CreationDate":"2016-02-16T17:18:30.003","AcceptedAnswerId":null,"OwnerUserId":2486302.0,"Title":"PyMC3 Multivariate Mixture Model: Constraining components to be non-empty","Body":"<p>I am implementing a Personalized Mixture of Multivariate Gaussian Regressions in  and running into an issue with empty components. After referring to the <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/examples\/gaussian_mixture_model.ipynb\" rel=\"nofollow noreferrer\">related PyMC3 mixture model example<\/a>, I tried implementing the model using univariate normals instead, but I've had <a href=\"https:\/\/stackoverflow.com\/q\/35371811\/2486302\">some issues there as well<\/a>.<\/p>\n\n<p>I've tried several strategies to constrain each component to be non-empty, but each has failed. These are shown in the code below. My specific question is: <strong>What is the best way to constrain all components to be non-empty in a mixture of multivariate Gaussians using ?<\/strong><\/p>\n\n<p>Note that attempt #1 in the code below comes from the <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/blob\/master\/pymc3\/examples\/gaussian_mixture_model.ipynb\" rel=\"nofollow noreferrer\">Mixture Model in PyMC3 Example<\/a> and does not work here.<\/p>\n\n<p>You can replicate the synthetic data I am using with the <a href=\"https:\/\/gist.github.com\/macks22\/7853e65d8fd136b40613\" rel=\"nofollow noreferrer\">function in this gist<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>The first two approaches fail to ensure non-empty clusters; attempting to sample results in a :<\/p>\n\n<pre><\/pre>\n\n<p>...which indicates the component is empty, since the precision matrix has shape .<\/p>\n\n<p>The third method actually resolves the empty component issue but gives very strange inference behavior. I selected a burn-in based on traceplots and thinned to every 10th sample. The samples are still highly autocorrelated but much better than without thinning. At this point, I summed the Z values across the samples, and this is what I get:<\/p>\n\n<pre><\/pre>\n\n<p>So for some reason, all of the users are being assigned to the last cluster for every sample.<\/p>\n","answers":[{"AnswerId":"44286908","CreationDate":"2017-05-31T14:01:59.873","ParentId":null,"OwnerUserId":"5828100","Title":null,"Body":"<p>I have run into a similar problem. Something like this worked for a mixture of multivariate gaussians model. As for whether it's the <em>best<\/em>, it's certainly the best solution I've found.<\/p>\n\n<pre><code>pm.Potential('pi_min_potential', T.switch(\n                                           T.all(\n                                                [pi[i, 0] &lt; 0.1 for i in range(K)]), -np.inf, 0))\n<\/code><\/pre>\n\n<p>The key here is that you need to account for <em>each<\/em> potential that is below your cutoff. Further, you should adjust the <code>shape<\/code> of your <code>pi<\/code> distribution, as mentioned in the comments. This will affect your indexing in the <code>T.switch<\/code> call (on the <code>pi[i,0]<\/code>).<\/p>\n"}]}
{"QuestionId":35443080,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-16T21:12:46.243","AcceptedAnswerId":35446265.0,"OwnerUserId":1940138.0,"Title":"TensorFlow: critical graph operations assigned to cpu rather than gpu","Body":"<p>I have implemented a TensorFlow DNN model (2 hidden layers with elu activation functions trained on MNIST) as a Python class in order to wrap TF calls within another library with its own optimization routines and tools. <\/p>\n\n<p>When running some tests on a TeslaK20 I noticed that the GPU was being used at 4% of the total capacity. Therefore I looked a bit more closely to the log-device-placement and figured that all critical operations like , , ,  etc were being assigned to the CPU. <\/p>\n\n<p>The first thing that came to mind was that it was because I was using , therefore I switched to . While a lot more operations were assigned to GPU, still a good number were assigned to the CPU, like , , . <\/p>\n\n<p>So here comes my first question (I'm linking a working code example at the end), <\/p>\n\n<p>1) why would that be? I have written different TF models that consist of simple tensor multiplications and reductions and they run fully on GPU as long as I use single precision. <\/p>\n\n<p>So here comes the second question,<\/p>\n\n<p>2) why does TF assign the graph to different devices depending on the data type? I understand that not all kernels are implemented for GPU but I would have thought that things like  could run on GPU for both single and double precision.<\/p>\n\n<p>3) Could the fact that the model is wrapped within a Python class have an effect? I do not think this is the case because as I said, it did not happen for other models wrapped similarly but that were simpler. <\/p>\n\n<p>4) What sort of steps can I take to run the model fully on a GPU?<\/p>\n\n<p>Here is a full working example of my code that I have isolated from my library <\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/smcantab\/8ecb679150a327738102\" rel=\"nofollow\">https:\/\/gist.github.com\/smcantab\/8ecb679150a327738102<\/a> . <\/p>\n\n<p>If you run it and look at the output you'll see how different parts of the graph have been assigned to different devices. To see how this changes with types and devices change  and  within  at the end of the example. Note that if I set  the graph fails to initialize.<\/p>\n\n<p>Any word of advice would be really appreciated.<\/p>\n","answers":[{"AnswerId":"35446265","CreationDate":"2016-02-17T01:24:22.967","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>As Yaroslav noted:  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/reduction_ops_mean.cc\" rel=\"nofollow noreferrer\">Mean, in particular, was not yet implemented for GPU<\/a>, but it is now available so these operations should run on the GPU with the latest TensorFlow.  (as per the DEVICE_GPU registration at that link)<\/p>\n\n<p>Prior to availability of mean, the status of this was:<\/p>\n\n<p>(a)  You can implement mean by hand, because <code>reduce_sum<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/reduction_ops_sum.cc\" rel=\"nofollow noreferrer\">is available on GPU<\/a>.<\/p>\n\n<p>(b)  I've re-pinged someone to see if there's an easy way to add the GPU support, but we'll see.<\/p>\n\n<p>Re <code>float64<\/code> on GPU, someone opened an issue three days ago with a patch for <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/1089\" rel=\"nofollow noreferrer\">supporting float64 reductions on GPU<\/a>.  Currently being reviewed and tested.<\/p>\n\n<p>No, it doesn't matter if it's wrapped in Python - it's really just about whether a kernel has been defined for it to execute on the GPU or not.  In many cases, the answer to \"why is X supported on GPU by Y not?\" comes down to whether or not there's been demand for Y to run on the GPU.  The answer for float64 is simpler:  float32 is a lot faster, so in most cases, people work to make their models work in float32 when possible because it gives all-around speed benefits.<\/p>\n"},{"AnswerId":"38490375","CreationDate":"2016-07-20T20:35:14.597","ParentId":null,"OwnerUserId":"6615666","Title":null,"Body":"<p>Most of the graphics cards like the GTX 980, 1080, etc are stripped of the double precision floating point hardware units. Since these are much cheaper and therefore more ubiquitous than the Newer Tesla Units (which have FP64 double precision hardware), doing double precision calculations on the graphics cards is very slow compared to single precision. FP64 calculations on a GPU seem to be about 32 X slower than FP32 on a GPU without the FP64 hardware. I believe this is why the FP32 calculations tend to be set up for the GPU while FP64 for the CPU (which is faster in most systems.) Hopefully in the future, the frameworks will test the GPU capabilities at runtime to decide where to assign the FP64 calculations.<\/p>\n"}]}
{"QuestionId":35445387,"AnswerCount":0,"Tags":"<centos><tensorflow><bazel>","CreationDate":"2016-02-17T00:00:13.493","AcceptedAnswerId":null,"OwnerUserId":5256563.0,"Title":"Bazel on CentOS for Tensorflow with CUDA","Body":"<p>We had to use Bazel on CentOS 6 (no choice) in order to install TensorFlow, and it worked.<\/p>\n\n<p>Unfortunately, we were not able to install TensorFlow with CUDA, thus so far it runs only on CPUs.<\/p>\n\n<p>We think that it is because the link to the CUDA compiler is wrong.\nHow can we modify\/tune Bazel in order to give the proper link?<\/p>\n","answers":[]}
{"QuestionId":35448177,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-17T04:50:20.440","AcceptedAnswerId":null,"OwnerUserId":4755260.0,"Title":"ValueError when training Tensorflow: setting an array element with a sequence","Body":"<p>I was stuck on training process with ValueError.\nThe following are the details:<\/p>\n\n<p>I made flags as below.<\/p>\n\n<pre><\/pre>\n\n<p>Moreover training process as below:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this code I faced following error. How to solve this problem?<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>I add code around train_image and train_label as following.\n\u3000\u3000\u3000<\/p>\n\n<p>\u3000<\/p>\n\n<pre><\/pre>\n\n<p>\u3000\u3000 <\/p>\n\n<pre><\/pre>\n\n<p>I constructing this model.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ODmCk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ODmCk.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"35448419","CreationDate":"2016-02-17T05:13:37.517","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This exception is raised when TensorFlow converts the values in the <code>feed_dict<\/code> into dense NumPy <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.ndarray.html\" rel=\"nofollow\">ndarrays<\/a>, and will depend on what's in your <code>train_image<\/code> and <code>train_label<\/code> objects.<\/p>\n\n<p>A common cause of these errors is when feed value is a <em>ragged<\/em> list: i.e. a list of lists, where the sublists have different sizes. For example:<\/p>\n\n<pre><code>&gt;&gt;&gt; train_image = [[1., 2., 3.], [4., 5.]]\n&gt;&gt;&gt; np.array(train_image, dtype=np.float32)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT:<\/strong> Thanks for sharing your code to build <code>train_image<\/code> and <code>train_label<\/code>. I suspect the problem is in the creation of <code>train_image<\/code>, if the elements of <code>flatten_data<\/code> can have different lengths for each example. Try making the following modification to confirm this: <code>train_image = np.asarray(flatten_data, dtype=np.float32)<\/code>. If you get the same <code>ValueError<\/code>, you will need to pad or crop the individual rows so that they have <code>IMAGE_PIXELS<\/code> elements.<\/p>\n"}]}
{"QuestionId":35448312,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-17T05:03:33.897","AcceptedAnswerId":null,"OwnerUserId":4747268.0,"Title":"How to visualize features classified by tensorflow?","Body":"<p>I'm running the default <strong>classify_image<\/strong> code of the imagenet model. Is there any way to visualize the features that it has extracted? If I use <strong>'pool_3:0'<\/strong>, that gives me the feature vector. Is there any way to overlay this on top of my image to see which features it has picked as important?<\/p>\n","answers":[{"AnswerId":"36141668","CreationDate":"2016-03-21T21:10:36.347","ParentId":null,"OwnerUserId":"1760830","Title":null,"Body":"<p>Ross Girshick described one way to visualize what a pooling layer has learned: <a href=\"https:\/\/www.cs.berkeley.edu\/~rbg\/papers\/r-cnn-cvpr.pdf\" rel=\"nofollow\">https:\/\/www.cs.berkeley.edu\/~rbg\/papers\/r-cnn-cvpr.pdf<\/a><\/p>\n\n<p>Essentially instead of visualizing features, you find a few images that a neuron fires most on. You repeat that for a few or all neurons from your feature vector. The algorithm needs lots of images to choose from of course, e.g. the test set.<\/p>\n\n<p>I wrote my implementation of this idea for <code>cifar10<\/code> model in Tensorflow today, which I want to share (uses OpenCV): <a href=\"https:\/\/gist.github.com\/kukuruza\/bb640cebefcc550f357c\" rel=\"nofollow\">https:\/\/gist.github.com\/kukuruza\/bb640cebefcc550f357c<\/a><\/p>\n\n<p>You could use it if you manage to provide the <code>images<\/code> tensor for reading images by batches, and the <code>pool_3:0<\/code> tensor.<\/p>\n"}]}
{"QuestionId":35448894,"AnswerCount":1,"Tags":"<sequence><tensorflow><recurrent-neural-network>","CreationDate":"2016-02-17T05:51:16.110","AcceptedAnswerId":null,"OwnerUserId":5676182.0,"Title":"Incremental continuous timeseries using TensorFlow RNN","Body":"<p>I want to feed timeseries data as it comes - one step at a time, to incrementally build the rnn ( after few initial steps ).<\/p>\n\n<p>currently the rnn() takes encoder_input and decoder_input as complete sequence. <\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to feed decoder_input one step at a time instead of whole sequence, because that's how the data would come in real time ?<\/p>\n","answers":[{"AnswerId":"35471177","CreationDate":"2016-02-18T01:36:06.697","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>Check out <a href=\"https:\/\/github.com\/tensorflow\/skflow\" rel=\"nofollow\">Scikit Flow<\/a>. The example folder includes a lot of examples dealing with RNNs and there are built-in RNN estimators that you can just plug into your existing code. <\/p>\n\n<p>Check this <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/estimators\/base.py#L171\" rel=\"nofollow\"><code>fit()<\/code><\/a> method in estimators and you'll find <code>partial_fit()<\/code> that allows continuous training, which suits your needs. Many of the examples use this to continue training in a <code>while<\/code> loop and save the checkpoint over time (you can also configure the frequency).  <\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":35449798,"AnswerCount":1,"Tags":"<python><numpy><scipy><sympy><theano>","CreationDate":"2016-02-17T06:51:43.067","AcceptedAnswerId":null,"OwnerUserId":2051572.0,"Title":"Gaussian filter of 2d image with location dependant sigma in Python","Body":"<p>I am trying to implement in Python a Gaussian filter in which the kernel is \n<a href=\"https:\/\/i.stack.imgur.com\/Pk7VJ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Pk7VJ.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where   is some function of the coordinates. Is it possible somehow to trick <a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.14.0\/reference\/generated\/scipy.ndimage.filters.gaussian_filter.html\" rel=\"nofollow noreferrer\">scipy.ndimage.filters.gaussian_filter<\/a> to do the job?<\/p>\n\n<p>Eventually I will have to perform the following integration over space:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UE0dZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UE0dZ.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And so I get $\\tilde G(x',y') $ as a coordinate dependent coefficient.<\/p>\n\n<p>Right now I am making it with tedious  loops:<\/p>\n\n<pre><\/pre>\n\n<p>** Can I do it somehow using Symbolic calculation using SymPy or Theano?<\/p>\n","answers":[{"AnswerId":"35517690","CreationDate":"2016-02-20T00:47:01.753","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Convolutions, from a certain signal and filter support size onwards are always faster to perform in Fourier space by pointwise multiplication of the Fourier transforms of signal and filter and inverse transforming.<\/p>\n\n<p>In that case one assumes the filtering to be linear and shift-invariant, so your context does not fit the classic context of convolution.<\/p>\n\n<p>So you will have to implement this yourself.<\/p>\n\n<p>If you have a lot of signals to apply this to, your best bet without resorting to writing low-level code is to write this transformation in the form of a <code>scipy.sparse<\/code> matrix. For this, you have to create all the Gaussian filters you need and place them appropriately in the image and stacking the ravelled versions of these filter images in a sparse matrix. You should truncate the support of you Gaussians so most of the pixels of your image are equal to 0. Once this matrix is created, you can apply it immediately to the pixel array of any image to get the desired result.<\/p>\n"}]}
{"QuestionId":35450239,"AnswerCount":0,"Tags":"<google-api><tensorflow>","CreationDate":"2016-02-17T07:18:51.043","AcceptedAnswerId":null,"OwnerUserId":5871027.0,"Title":"Tensorflow vs Google Prediction API","Body":"<p>Talking only about the machine learning libraries (and not about the platform), <strong>is the prediction API available on GCP (Google Cloud Platform) same as its newly released tensorflow library?<\/strong> Is there any official documentation which clarifies this thing?<\/p>\n","answers":[]}
{"QuestionId":35450747,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-02-17T07:50:31.287","AcceptedAnswerId":35450912.0,"OwnerUserId":4747268.0,"Title":"How to see the operations performed in a Tensorflow pb file?","Body":"<p>More specifically, I want to see the operations performed in the graph inside <strong>'classify_image_graph_def.pb'<\/strong> which is inside Tensorflow's imagenet Inception model.<\/p>\n","answers":[{"AnswerId":"35450912","CreationDate":"2016-02-17T08:00:15.407","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>This is related to <a href=\"https:\/\/stackoverflow.com\/questions\/35336648\/list-of-tensor-names-in-graph-in-tensorflow\/35337827#35337827\" title=\"this answer\">this question<\/a> - basically you can load it (like in the python image classification example) and then run <code>get_operations()<\/code> on the graph, which gives you a list of operations, each of which have a name and a type attribute. You also could in principle use tensorboard I believe, but for that specific graph I always ran into \"too big\" errors.<\/p>\n"}]}
{"QuestionId":35451948,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-17T08:57:43.397","AcceptedAnswerId":35464950.0,"OwnerUserId":5703903.0,"Title":"Clarification on tf.Tensor.set_shape()","Body":"<p>I have an image that is 478 x 717 x 3 = 1028178 pixels, with a rank of 1. I verified it by calling tf.shape and tf.rank.<\/p>\n\n<p>When I call image.set_shape([478, 717, 3]), it throws the following error.<\/p>\n\n<pre><\/pre>\n\n<p>I tested again by first casting to 1028178, but the error still exists.<\/p>\n\n<pre><\/pre>\n\n<p>Well, that does make sense because one is of rank 1 and the other is of rank 3. However, why is it necessary to throw an error, as the total number of pixels still match.<\/p>\n\n<p>I could of course use tf.reshape and it works, but I think that's not optimal.<\/p>\n\n<p>As stated on the TensorFlow FAQ<\/p>\n\n<blockquote>\n  <p>What is the difference between x.set_shape() and x = tf.reshape(x)?<\/p>\n  \n  <p>The tf.Tensor.set_shape() method updates the static shape of a Tensor\n  object, and it is typically used to provide additional shape\n  information when this cannot be inferred directly. It does not change\n  the dynamic shape of the tensor.<\/p>\n  \n  <p>The tf.reshape() operation creates a new tensor with a different dynamic shape.<\/p>\n<\/blockquote>\n\n<p>Creating a new tensor involves memory allocation and that could potentially be more costly when more training examples are involved. Is this by design, or am I  missing something here?<\/p>\n","answers":[{"AnswerId":"35464950","CreationDate":"2016-02-17T18:38:02.770","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As far as I know (and I wrote that code), there isn't a bug in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#set_shape\" rel=\"noreferrer\"><code>Tensor.set_shape()<\/code><\/a>. I think the misunderstanding stems from the confusing name of that method.<\/p>\n\n<p>To elaborate on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/resources\/faq.html#what_is_the_difference_between_xset_shape_and_x_tfreshapex\" rel=\"noreferrer\">FAQ entry you quoted<\/a>, <code>Tensor.set_shape()<\/code> is a pure-Python function that <em>improves<\/em> the shape information for a given <code>tf.Tensor<\/code> object. By \"improves\", I mean \"makes more specific\". <\/p>\n\n<p>Therefore, when you have a <code>Tensor<\/code> object <code>t<\/code> with shape <code>(?,)<\/code>, that is a one-dimensional tensor of unknown length. You can call <code>t.set_shape((1028178,))<\/code>, and then <code>t<\/code> will have shape <code>(1028178,)<\/code> when you call <code>t.get_shape()<\/code>. This doesn't affect the underlying storage, or indeed anything on the backend: it merely means that subsequent shape inference using <code>t<\/code> can rely on the assertion that it is a vector of length 1028178.<\/p>\n\n<p>If <code>t<\/code> has shape <code>(?,)<\/code>, a call to <code>t.set_shape((478, 717, 3))<\/code> will fail, because TensorFlow already knows that <code>t<\/code> is a vector, so it <em>cannot<\/em> have shape <code>(478, 717, 3)<\/code>. If you want to make a new Tensor with that shape from the contents of <code>t<\/code>, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#reshape\" rel=\"noreferrer\"><code>reshaped_t = tf.reshape(t, (478, 717, 3))<\/code><\/a>. This creates a new <code>tf.Tensor<\/code> object in Python; the actual <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/03bff43060229357cbe2cc1659e7d129c2799b06\/tensorflow\/core\/kernels\/reshape_op.h#L89\" rel=\"noreferrer\">implementation of <code>tf.reshape()<\/code><\/a> does this using a shallow copy of the tensor buffer, so it is inexpensive in practice.<\/p>\n\n<p>One analogy is that <code>Tensor.set_shape()<\/code> is like a run-time cast in an object-oriented language like Java. For example, if you have a pointer to an <code>Object<\/code> but know that, in fact, it is a <code>String<\/code>, you might do the cast <code>(String) obj<\/code> in order to pass <code>obj<\/code> to a method that expects a <code>String<\/code> argument. However, if you have a <code>String<\/code> <code>s<\/code> and try to cast it to a <code>java.util.Vector<\/code>, the compiler will give you an error, because these two types are unrelated.<\/p>\n"}]}
{"QuestionId":35453095,"AnswerCount":1,"Tags":"<python-2.7><theano>","CreationDate":"2016-02-17T09:48:22.200","AcceptedAnswerId":null,"OwnerUserId":5866269.0,"Title":"argument in theano.tensor.nnet.conv2d","Body":"<p>I want to add pad to make the output shape be equal to the input shape when using theano.tensor.nnet.con2d.<\/p>\n\n<pre><\/pre>\n\n<p>If border_mode = (pad_C,pad_R),there is error like \"not all arguments converted during string formatting\".\nIf border_mode = 'half', there is error like \"Mode half not implemented\".\nIf border_mode is equal to default, everything is OK.<\/p>\n\n<p>What's wrong?\nThanks a lot.<\/p>\n","answers":[{"AnswerId":"35518618","CreationDate":"2016-02-20T03:04:01.747","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>Tough luck, if you take a look at the source code of <code>theano.tensor.nnet.con2d<\/code> you'll see it does not support padding or the 'half' option at all, as opposed to what is claimed in the current documentation.  <\/p>\n\n<p>I use <code>theano.sandbox.cuda.dnn.dnn_conv<\/code> instead, it supports padding.<\/p>\n"}]}
{"QuestionId":35453273,"AnswerCount":0,"Tags":"<python><gpgpu><theano><deep-learning>","CreationDate":"2016-02-17T09:55:37.803","AcceptedAnswerId":null,"OwnerUserId":5939610.0,"Title":"Schematic link between Python, C, Theano, CUDA, Windows 7, Machine Code, CPU, GPU and many more?","Body":"<p>I do not have a thorough computer science background and I am trying to find my way trough deep learning. I was at the point that I wanted to make a decision on which framework I want to use to start coding some networks. Nevertheless, going through many discussions on fora and other websites raised another question that kept me busy for two days and which seems to be answered nowhere. <\/p>\n\n<p>At this point I still can't get a complete understanding on how my Theano code in Python is finally executed on the CPU and GPU (is this \"machine code\"?). <\/p>\n\n<p>As an answer I imagine a schematic view (of course can be in words) on the link between how all the parts play together in a Theano program, mainly:<\/p>\n\n<ul>\n<li>my code<\/li>\n<li>Theano<\/li>\n<li>Python interpreter<\/li>\n<li>C<\/li>\n<li>NVCC<\/li>\n<li>CPU<\/li>\n<li>GPU <\/li>\n<li>CUDA <\/li>\n<\/ul>\n\n<p>Also I'd like to understand the role of Python bytecode in the context of Theano, and how it differs from the code (machine code?) that is acutally executed on the CPU or the GPU.<\/p>\n\n<p>If it helps, I am working on a Windows 7 64 bit. <\/p>\n","answers":[]}
{"QuestionId":35456689,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><training-data>","CreationDate":"2016-02-17T12:25:28.277","AcceptedAnswerId":35457870.0,"OwnerUserId":4386025.0,"Title":"Caffe : train network accuracy = 1 constant ! Accuracy issue","Body":"<p>Right now, I am train network with with 2 class data... but accuracy is constant 1 after first iteration ! <\/p>\n\n<p>Input data is grayscale images. both class images are randomly selected when HDF5Data creation. <\/p>\n\n<p>Why is that happened ? What's wrong or where is mistake ! <\/p>\n\n<p>network.prototxt :<\/p>\n\n<pre><\/pre>\n\n<p>Output : <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n","answers":[{"AnswerId":"35457870","CreationDate":"2016-02-17T13:19:21.767","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Summarizing some information from the comments:<br>\n - You run test at intervals of <code>test_interval:100<\/code> iterations.<br>\n - Each test interval goes over <code>test_iter:5<\/code> * <code>batch_size:10<\/code> = 50 samples.<br>\n - Your train and test sets seems to be very nit: all the negative samples (label=0) are grouped together before all the positive samples.<\/p>\n\n<hr>\n\n<p>Consider your SGD iterative solver, you feed it batches of <code>batch_size:10<\/code> during training. Your training set has 14,746 negative samples (that is 1474 batches) before any positive sample. So, for the first 1474 iterations your solver only \"sees\" negative examples and no positive ones.<br>\nWhat do you expect this solver will learn?<\/p>\n\n<h3>The problem<\/h3>\n\n<p>Your solver only sees negative examples, thus learns that no matter what the input is it should output \"0\". Your test set is also ordered in the same fashion, so testing only 50 samples at each test_interval, you only test on the negative examples in the test set resulting with a perfect accuracy of 1.<br>\nBut as you noted, your net actually learned nothing.<\/p>\n\n<h3>Solution<\/h3>\n\n<p>I suppose you already guess what the solution should be by now. You need to shuffle your training set, and test your net on your <em>entire<\/em> test set.<\/p>\n"}]}
{"QuestionId":35457879,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-17T13:19:36.320","AcceptedAnswerId":35460770.0,"OwnerUserId":3733814.0,"Title":"Evaluation of loss with session.run([]) in TensorFlow","Body":"<p>For simple gradient descent, I am using this :<\/p>\n\n<p>\n where optimizer op minimizes loss. <\/p>\n\n<p>But I am feeding loss after optimizer, so will the loss, be evaluated two times at each .run(), one using initial weight, and then using updated weights ?<\/p>\n","answers":[{"AnswerId":"35460770","CreationDate":"2016-02-17T15:23:30.930","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>The order of elements in the list of fetched tensors doesn't matter (except for capturing the values on the Python side). They will all be executed during the same evaluation.<\/p>\n\n<p>In your example, the loss will be evaluated only once and any computations that can be shared between the <code>[optimizer, loss, train_prediction]<\/code> will be shared.<\/p>\n"}]}
{"QuestionId":35459334,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-17T14:24:33.747","AcceptedAnswerId":35474015.0,"OwnerUserId":5587428.0,"Title":"Reading Tensorflow Record Files, does not work after first run","Body":"<p>I have a small block of code that reads in data from some TFRecord files. If I run the code from an ipython notebook, it works fine the first time I execute the block. But the same block of code produces an error if I try to execute it a second time without restarting the kernel (Error: StatusNotOK: Not found: FetchOutputs node DecodeRaw_2:0: not found). The code is shown below. Do I need to close\/clear\/reinitialize something in order to have the code run properly multiple times?<\/p>\n\n<pre><\/pre>\n\n<p>Note that read_and_decode() is taken from <a href=\"https:\/\/www.tensorflow.org\/code\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow\">here<\/a><\/p>\n","answers":[{"AnswerId":"35468717","CreationDate":"2016-02-17T22:05:24.930","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>By default tf. commands append to default graph with new names. You can use tf.reset_default_graph() before running the snippet the second time to clear the default graph.<\/p>\n"},{"AnswerId":"35474015","CreationDate":"2016-02-18T06:04:23.233","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are a few issues in the code in the question.<\/p>\n\n<ol>\n<li><p>The first, <a href=\"https:\/\/stackoverflow.com\/a\/35468717\/3574081\">as Yaroslav pointed out<\/a> is that all of the ops are added to the same graph. The means that when you call <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#start_queue_runners\" rel=\"nofollow noreferrer\"><code>tf.train.start_queue_runners()<\/code><\/a> (or run the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#initialize_all_variables\" rel=\"nofollow noreferrer\"><code>tf.initialize_all_variables()<\/code><\/a> op) the session will be doing work that is proportional to the number of times that you've run this code snippet. You can call <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#reset_default_graph\" rel=\"nofollow noreferrer\"><code>tf.reset_default_graph()<\/code><\/a> in between invocations to this code, but a cleaner way to isolate things might be to declare a separate graph each time:<\/p>\n\n<pre><code>with tf.Graph().as_default():  # Declares a new graph for the life of the block.\n    filename_queue = tf.train.string_input_producer(filename_list)\n    init = tf.initialize_all_variables()\n    image = []\n    label = []\n    with tf.Session() as sess:\n        # ...\n        coord.join(threads)\n<\/code><\/pre><\/li>\n<li><p>The second issue is that the separate calls to <code>sess.run(tf_image)<\/code> and <code>sess.run(tf_label)<\/code> means that the association between image and label is lost. When you call <code>sess.run(tf_image)<\/code> you consume an image <em>and<\/em> a label from the reader, but discard the label (and vice versa for <code>sess.run(tf_label)<\/code>. The correct solution is to fetch them both in the same step:<\/p>\n\n<pre><code>image_val, label_val = sess.run([tf_image, tf_label])\nimage.append(image_val)\nlabel.append(label_val)\n<\/code><\/pre><\/li>\n<li><p>The final issue&mdash;which could cause problems even if you reset the graph&mdash;is that the code adds nodes to the graph after calling <code>tf.train.start_queue_runners()<\/code>. There is a possibility of data races on the TensorFlow graph, because <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/fdf2b0e9dc0017ab7fe38893b01e6adffa43c289\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py#L55\" rel=\"nofollow noreferrer\"><code>read_and_decode()<\/code><\/a> adds nodes to the graph, while the parallel queue runners concurrently read it, and the <code>tf.Graph<\/code> isn't thread safe.<\/p>\n\n<p>The best way to deal with this is to define all of your graph before starting the queue runners:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n    filename_queue = tf.train.string_input_producer(filename_list)\n    image = []\n    label = []\n    tf_image, tf_label = read_and_decode(filename_queue)\n\n    with tf.Session() as sess:\n       coord = tf.train.Coordinator()\n       threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n       for i in range(len(filename_list)):\n           image_val, label_val = sess.run([tf_image, tf_label])\n           image.append(image_val)\n           label.append(label_val)\n\n       coord.request_stop()\n       coord.join(threads)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":35460750,"AnswerCount":1,"Tags":"<mingw><openmp><anaconda><theano>","CreationDate":"2016-02-17T15:22:42.703","AcceptedAnswerId":36874644.0,"OwnerUserId":3271237.0,"Title":"Theano and MinGW from Anaconda. How to make openmp work?","Body":"<p>I have to work with large data sets using Theano, so I want to use parallelism. I've installed MinGW using conda, and set openmp flag in .theanorc True. But when I run my code I saw \n\"UserWarning: Your g++ compiler fails to compile OpenMP code. We know this happen with some version of the EPD mingw compiler and LLVM compiler on Mac OS X. We disable openmp everywhere in Theano. To remove this warning set the theano flags  to False\" How to make things work? \nI'm using Windows 7 x64, Anaconda 2.5.0, theano 0.7, mingw 4.7<\/p>\n","answers":[{"AnswerId":"36874644","CreationDate":"2016-04-26T20:00:22.930","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>In newest version of Anaconda you have INTEL MKL libraries installed by default. Configuring MKL properly automatically allows you to use multithreading mode in your numpy and theano computations.<\/p>\n"}]}
{"QuestionId":35461177,"AnswerCount":2,"Tags":"<caffe><pycaffe><matcaffe>","CreationDate":"2016-02-17T15:38:32.407","AcceptedAnswerId":35466130.0,"OwnerUserId":4308982.0,"Title":"How do you export .caffemodels to other applications?","Body":"<p>Is it possible to translate the info in a .caffemodel file such that it could be read by (for example) Matlab. That is, is there a way to write your model using something else that prototxt and import the weights trained using Caffe?<\/p>\n\n<p>If the answer is \"Nope, it's a binary file and will always remain that way\", is there some documentation regarding the structure of the file so that one could extract the important information somehow?<\/p>\n","answers":[{"AnswerId":"39159478","CreationDate":"2016-08-26T06:15:08.137","ParentId":null,"OwnerUserId":"4265775","Title":null,"Body":"<p>I'll take VGG as an example  <\/p>\n\n<pre><code>from caffe.proto import caffe_pb2\nnet = caffe_pb2.NetParameter()\ncaffemodel = sys.argv[1]\nwith open(caffemodel, 'rb') as f:\n    net.ParseFromString(f.read())\nfor i in net.layer:\n    print i.ListFields()[0][-1]\n#conv1\n#relu1\n#norm1\n#pool1\n#conv2\n#relu2\n#norm2\n#pool2\n#conv3\n#relu3\n#conv4\n#relu4\n#conv5\n#relu5\n#pool5\n#fc6\n#relu6\n#drop6\n#fc7\n#relu7\n#drop7\n#fc8\n#prob\n<\/code><\/pre>\n"},{"AnswerId":"35466130","CreationDate":"2016-02-17T19:38:58.307","ParentId":null,"OwnerUserId":"733941","Title":null,"Body":"<p>As you know, <code>.caffemodel<\/code> consists of weights and biases.\nA simple way to read weights and biases for a caffemodel given the prototxt would be to just load the network in Python and read the weights. <\/p>\n\n<p>You can use:<\/p>\n\n<pre><code>import caffe\nnet = caffe.Net(&lt;prototxt-file&gt;,&lt;model-file&gt;,&lt;phase&gt;);\n<\/code><\/pre>\n\n<p>and access the params from <code>net.params<\/code><\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/ES86FI-3vbY\/I13VVwIBAwAJ\" rel=\"nofollow\">source<\/a><\/p>\n"}]}
{"QuestionId":35461516,"AnswerCount":1,"Tags":"<python><theano><keras><lstm><recurrent-neural-network>","CreationDate":"2016-02-17T15:53:17.860","AcceptedAnswerId":null,"OwnerUserId":5155529.0,"Title":"Assertion Error when compiling LSTM with Keras","Body":"<p>I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>But everytime I run it, I get an error at the line<\/p>\n\n<pre><\/pre>\n\n<p>which states:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know what that means?<\/p>\n","answers":[{"AnswerId":"50171848","CreationDate":"2018-05-04T09:39:48.900","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p><em>Answer by OP:<\/em><\/p>\n\n<p>I have fixed this problem, turns out it has something to do with an outdated Theano version. So if you are experiencing this problem, update your theano module! <\/p>\n"}]}
{"QuestionId":35461536,"AnswerCount":3,"Tags":"<python><machine-learning><caffe><pycaffe>","CreationDate":"2016-02-17T15:53:52.190","AcceptedAnswerId":null,"OwnerUserId":4755260.0,"Title":"Error on import caffe","Body":"<p>For DeepDream or other deeplearning project, constructing environment of Caffe.<\/p>\n\n<p>I installed required packages for PyCaffe and made PYTHONPATH to caffe\/python.<\/p>\n\n<p>However when I import caffe on python:<\/p>\n\n<pre><\/pre>\n\n<p>Error occurred as below.How to solve this problem?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35464625","CreationDate":"2016-02-17T18:20:30.810","ParentId":null,"OwnerUserId":"5918371","Title":null,"Body":"<p>Try to manually set python path in your python script if you get no module named caffe error<\/p>\n\n<p>Ex.\nimport sys<\/p>\n\n<p>sys.path.insert(0, \"\/home\/nviso\/GitHub\/caffe\/distribute\/python\")<\/p>\n\n<p>import caffe<\/p>\n\n<p>This generally works for me.\nAdding the caffe or python path to .bashrc manually should probably solve the issue as well though not sure, don't have my Office PC now to try :)<\/p>\n"},{"AnswerId":"37617562","CreationDate":"2016-06-03T14:54:02.017","ParentId":null,"OwnerUserId":"6362908","Title":null,"Body":"<p>Are you using a mac? I had a very hard time making pycaffe on mac, until I realized that there is a native python installed on all macs and that I was using another version I had installed.  While compiling, caffe was using some of the stuff from the native python, and some of the stuff from the other python.  I had to make sure to change all the relevant paths in the makefile.config file and to change the python my bash was using.  I recommend working in a virtual environment as well.  <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/591\" rel=\"nofollow\">This<\/a> is a good link to help you out, good luck! <\/p>\n"},{"AnswerId":"47586414","CreationDate":"2017-12-01T03:54:35.797","ParentId":null,"OwnerUserId":"901973","Title":null,"Body":"<p>This has been discussed since 2015 in <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2677\" rel=\"nofollow noreferrer\">Github issue<\/a>.\nMostly the reason is the conflict of homebrew python and OS X system python.<\/p>\n\n<p>Homebrew has provided an <a href=\"http:\/\/docs.python-guide.org\/en\/latest\/starting\/install\/osx\/\" rel=\"nofollow noreferrer\">awesome solution for OS X<\/a>:<\/p>\n\n<pre><code>$ python -V   # system Python interpreter\n$ python2 -V  # Homebrew installed Python 2 interpreter\n$ python3 -V  # Homebrew installed Python 3 interpreter (if installed)\n<\/code><\/pre>\n\n<p>So the solution is changing all the python paths to python2. Bellowing is my Makefile.config related:<\/p>\n\n<pre><code># ...\n# NOTE: this is required only if you will compile the python interface.\n# We need to be able to find Python.h and numpy\/arrayobject.h.\n# PYTHON_INCLUDE := \/usr\/include\/python2.7 \\\n#       \/usr\/lib\/python2.7\/dist-packages\/numpy\/core\/include\n# ------ For Homebrew installed python. Numpy path is added using python commands. \nPYTHON_INCLUDE := \/usr\/local\/Cellar\/python\/2.7.14\/Frameworks\/Python.framework\/Versions\/2.7\/include\/python2.7\n\n# We need to be able to find libpythonX.X.so or .dylib. ------ (Update Homebrew path)\n# PYTHON_LIB := \/usr\/lib\n# PYTHON_LIB := $(ANACONDA_HOME)\/lib\nPYTHON_LIB := \/usr\/local\/Cellar\/python\/2.7.14\/Frameworks\/Python.framework\/Versions\/2.7\/lib\n\n# Homebrew installs numpy in a non standard path (keg only) ------ (python2 for brew instead of python for system)\nPYTHON_INCLUDE += $(dir $(shell python2 -c 'import numpy.core; print(numpy.core.__file__)'))\/include\nPYTHON_LIB += $(shell brew --prefix numpy)\/lib\n# ...\n<\/code><\/pre>\n"}]}
{"QuestionId":35464097,"AnswerCount":1,"Tags":"<neural-network><tensorflow>","CreationDate":"2016-02-17T17:51:14.323","AcceptedAnswerId":null,"OwnerUserId":455959.0,"Title":"How to build a simple RNN with a cycle in the graph in TensorFlow?","Body":"<p>I've just started playing with TensorFlow and I'm trying to implement a very simple RNN. The RNN has  as input,  as output and consists of just a single layer that takes  and it's previous output as input. Here's a picture of the sort of thing I have in mind:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wvNGy.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wvNGy.png\" alt=\"A simple RNN\"><\/a><\/p>\n\n<p>The problem is, I can't see any way through the TensorFlow API to construct a graph with a cycle in it. Whenever I define a Tensor I have to specify what it's inputs are, which means I have to have already have defined it's inputs. So there's a chicken-and-egg problem.<\/p>\n\n<p>I don't even know if it makes sense to want to define a graph with a cycle (What gets computed first? Would I have to define an initial value of the softmax node?). I played with the idea of using a variable to represent the previous output and then manually take the value of  and store it in the variable every time after feeding through a training sample. But that would be very slow unless there's a way to represent this procedure in the graph itself (?).<\/p>\n\n<p>I know the TensorFlow tutorials show example implementations of RNNs but they cheat and pull an LSTM module out of the library which already has the cycle in it. Overall the tutorials are good for stepping you through how to build certain things but they could do a better job of explaining how this beast really works.<\/p>\n\n<p>So, TensorFlow experts, is there a way to build this thing? How would I go about doing it?<\/p>\n","answers":[{"AnswerId":"35473787","CreationDate":"2016-02-18T05:47:01.177","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>As a matter of fact, both the forward and backward pass in all the machine learning frameworks assume that your network does not have cycles. A common way of implementing a recurrent network is unrolling it in time for several steps (say 50), and therefore converting a network that has loops into one that does not have any. <\/p>\n\n<p>For instance, in the docs you are referring to:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html<\/a><\/p>\n\n<p>They mention<\/p>\n\n<blockquote>\n  <p>In order to make the learning process tractable, it is a common practice to truncate the gradients for backpropagation to a fixed number (num_steps) of unrolled steps.<\/p>\n<\/blockquote>\n\n<p>What it effectively means is that they will create <code>num_steps<\/code> LSTM cells, where each takes as an input the value <code>x<\/code> for the current timestep, and the output of the <strong>previous<\/strong> LSTM module.<\/p>\n\n<p>The <code>BasicLSTMCell<\/code> that they use and that you think has a loop in fact does not have a loop. An LSTM cell is just an implementation of a single LSTM step (a block that has two inputs [input and memory] and two outputs [output and memory], and uses gates to compute outputs from inputs), not the entire LSTM network.<\/p>\n"}]}
{"QuestionId":35464652,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-17T18:21:17.667","AcceptedAnswerId":null,"OwnerUserId":604964.0,"Title":"How to create ensemble in tensorflow?","Body":"<p>I am trying to create an ensemble of many trained models. All models have the same graph and just differ by its weights. I am creating the model graph using . I have several different checkpoints (with different weights) for the same graph architecture and I want to make one instance model for each checkpoint. <\/p>\n\n<p>How can I load many checkpoints without overwriting the previous loaded weights? <\/p>\n\n<p>As I created my graphs with , the only way I can create multiple graph is by passing the argument . Now if I try changing the names of my graph variables enclosing the build method in a new scope (so they become non-sharable with other created graphs) before loading, then this is not going to work because the new names will differ from the saved weights and I will not be able to load it.<\/p>\n","answers":[{"AnswerId":"51018846","CreationDate":"2018-06-25T08:09:57.227","ParentId":null,"OwnerUserId":"7443104","Title":null,"Body":"<p>This requires a few hacks. Let us save a few simple models<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#! \/usr\/bin\/env python\n# -*- coding: utf-8 -*-\n\nimport argparse\nimport tensorflow as tf\n\n\ndef build_graph(init_val=0.0):\n    x = tf.placeholder(tf.float32)\n    w = tf.get_variable('w', initializer=init_val)\n    y = x + w\n    return x, y\n\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser()\n    parser.add_argument('--init', help='dummy string', type=float)\n    parser.add_argument('--path', help='dummy string', type=str)\n    args = parser.parse_args()\n\n    x1, y1 = build_graph(args.init)\n\n    saver = tf.train.Saver()\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        print(sess.run(y1, {x1: 10}))  # outputs: 10 + i\n\n        save_path = saver.save(sess, args.path)\n        print(\"Model saved in path: %s\" % save_path)\n\n# python ensemble.py --init 1 --path .\/models\/model1.chpt\n# python ensemble.py --init 2 --path .\/models\/model2.chpt\n# python ensemble.py --init 3 --path .\/models\/model3.chpt\n<\/code><\/pre>\n\n<p>These models produce outputs of \"10 + i\" where i=1, 2, 3.\nNote this script creates, runs and saves multiple times the <em>same<\/em> graph-structure. Loading these values and restoring each graph individually is folklore and can be done by<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#! \/usr\/bin\/env python\n# -*- coding: utf-8 -*-\n\nimport argparse\nimport tensorflow as tf\n\n\ndef build_graph(init_val=0.0):\n    x = tf.placeholder(tf.float32)\n    w = tf.get_variable('w', initializer=init_val)\n    y = x + w\n    return x, y\n\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser()\n    parser.add_argument('--path', help='dummy string', type=str)\n    args = parser.parse_args()\n\n    x1, y1 = build_graph(-5.)\n\n    saver = tf.train.Saver()\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n\n        saver.restore(sess, args.path)\n        print(\"Model loaded from path: %s\" % args.path)\n\n        print(sess.run(y1, {x1: 10}))\n\n# python ensemble_load.py --path .\/models\/model1.chpt  # gives 11\n# python ensemble_load.py --path .\/models\/model2.chpt  # gives 12\n# python ensemble_load.py --path .\/models\/model3.chpt  # gives 13\n<\/code><\/pre>\n\n<p>These produce again the outputs 11,12,13 like expected. Now the trick is to create for each model from the ensemble its own scope like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_graph(x, init_val=0.0):\n    w = tf.get_variable('w', initializer=init_val)\n    y = x + w\n    return x, y\n\n\nif __name__ == '__main__':\n    models = ['.\/models\/model1.chpt', '.\/models\/model2.chpt', '.\/models\/model3.chpt']\n    x = tf.placeholder(tf.float32)\n    outputs = []\n    for k, path in enumerate(models):\n        # THE VARIABLE SCOPE IS IMPORTANT\n        with tf.variable_scope('model_%03i' % (k + 1)):\n            outputs.append(build_graph(x, -100 * np.random.rand())[1])\n<\/code><\/pre>\n\n<p>Hence each model lives under a different <em>variable<\/em>-scope, ie. we have variables 'model_001\/w:0, model_002\/w:0, model_003\/w:0' although they have a similar (not the same) sub-graph, these variables are indeed different objects. Now, the trick is to manage two sets of variables (those of the graph under the current scope and those from the checkpoint):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def restore_collection(path, scopename, sess):\n    # retrieve all variables under scope\n    variables = {v.name: v for v in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scopename)}\n    # retrieves all variables in checkpoint\n    for var_name, _ in tf.contrib.framework.list_variables(path):\n        # get the value of the variable\n        var_value = tf.contrib.framework.load_variable(path, var_name)\n        # construct expected variablename under new scope\n        target_var_name = '%s\/%s:0' % (scopename, var_name)\n        # reference to variable-tensor\n        target_variable = variables[target_var_name]\n        # assign old value from checkpoint to new variable\n        sess.run(target_variable.assign(var_value))\n<\/code><\/pre>\n\n<p>The full solution would be<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#! \/usr\/bin\/env python\n# -*- coding: utf-8 -*-\n\nimport numpy as np\nimport tensorflow as tf\n\n\ndef restore_collection(path, scopename, sess):\n    # retrieve all variables under scope\n    variables = {v.name: v for v in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scopename)}\n    # retrieves all variables in checkpoint\n    for var_name, _ in tf.contrib.framework.list_variables(path):\n        # get the value of the variable\n        var_value = tf.contrib.framework.load_variable(path, var_name)\n        # construct expected variablename under new scope\n        target_var_name = '%s\/%s:0' % (scopename, var_name)\n        # reference to variable-tensor\n        target_variable = variables[target_var_name]\n        # assign old value from checkpoint to new variable\n        sess.run(target_variable.assign(var_value))\n\n\ndef build_graph(x, init_val=0.0):\n    w = tf.get_variable('w', initializer=init_val)\n    y = x + w\n    return x, y\n\n\nif __name__ == '__main__':\n    models = ['.\/models\/model1.chpt', '.\/models\/model2.chpt', '.\/models\/model3.chpt']\n    x = tf.placeholder(tf.float32)\n    outputs = []\n    for k, path in enumerate(models):\n        with tf.variable_scope('model_%03i' % (k + 1)):\n            outputs.append(build_graph(x, -100 * np.random.rand())[1])\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n\n        print(sess.run(outputs[0], {x: 10}))  # random output -82.4929\n        print(sess.run(outputs[1], {x: 10}))  # random output -63.65792\n        print(sess.run(outputs[2], {x: 10}))  # random output -19.888203\n\n        print(sess.run(W[0]))  # randomly initialize value -92.4929\n        print(sess.run(W[1]))  # randomly initialize value -73.65792\n        print(sess.run(W[2]))  # randomly initialize value -29.888203\n\n        restore_collection(models[0], 'model_001', sess)  # restore all variables from different checkpoints\n        restore_collection(models[1], 'model_002', sess)  # restore all variables from different checkpoints\n        restore_collection(models[2], 'model_003', sess)  # restore all variables from different checkpoints\n\n        print(sess.run(W[0]))  # old values from different checkpoints: 1.0\n        print(sess.run(W[1]))  # old values from different checkpoints: 2.0\n        print(sess.run(W[2]))  # old values from different checkpoints: 3.0\n\n        print(sess.run(outputs[0], {x: 10}))  # what we expect: 11.0\n        print(sess.run(outputs[1], {x: 10}))  # what we expect: 12.0\n        print(sess.run(outputs[2], {x: 10}))  # what we expect: 13.0\n\n# python ensemble_load_all.py\n<\/code><\/pre>\n\n<p>Now having a list of <em>outputs<\/em>, you can average these values <em>within<\/em> TensorFlow or do some other ensemble predictions.<\/p>\n\n<p><strong>edit<\/strong>: <\/p>\n\n<ul>\n<li>It is way easier to store the model as a numpy dictionary using NumPy (npz) and load these values, like in my answer here:\n<a href=\"https:\/\/stackoverflow.com\/a\/50181741\/7443104\">https:\/\/stackoverflow.com\/a\/50181741\/7443104<\/a><\/li>\n<li>The code above just illustrates a solution. It does not feature sanity checks (like does the variable really exists). A try-catch might help.<\/li>\n<\/ul>\n"},{"AnswerId":"56042339","CreationDate":"2019-05-08T13:48:46.193","ParentId":null,"OwnerUserId":"7807658","Title":null,"Body":"<p>There are several questions regarding this topic and a lot of possible answers\/ways to do it. Here, I'd like to show how I came up with the most elegant and cleanest way of making an ensemble of <code>N<\/code> models, where <code>N<\/code> is arbitrary. This solution was tested with tf 1.12.0, python 2.7<\/p>\n\n<p>The following code snippet is what you are looking for (comments below):<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nnum_of_ensembles = N\nsavers = list()\npalceholders = list()\ninference_ops = list()\n\nfor i in xrange(num_of_ensembles):\n    with tf.name_scope('model_{}'.format(i)):\n        savers.append(tf.train.import_meta_graph('saved_model.ckpt.meta'))\n\ngraph = tf.get_default_graph()\n\nfor i in xrange(num_of_ensembles):\n    placeholders.append(graph.get_operation_by_name('model_{}\/input_ph'.format(i)).outputs[0])\n    inference_ops.append(graph.get_operation_by_name('model_{}\/last_operation_in_the_network'.format(i)).outputs[0])\n\n\nwith tf.Session() as sess:\n    for i in xrange(num_of_ensembles):\n        savers[i].restore(sess, 'saved_model.ckpt')\n        prediction = sess.run(inference_ops[i], feed_dict={placeholders[i]: np.random.rand(your_input.shape)})\n<\/code><\/pre>\n\n<p>So, first thing to do is to import the meta graph of each model. As suggested in the comments above, the key is to create for each model from the ensemble its own scope, in order to add a prefix like <em>model_001\/, model_002\/ ...<\/em> to each variable scopes. This will allow you to restore <code>N<\/code> different models, with their own independent variables.<\/p>\n\n<p>All this graphs will live in the current default graph. Now, when you load a model you have to extract the inputs, outputs, and operations that you wish to use from the graph into new variables. To do so, you'll need to know the names of those tensors from the old model. You can inspect all saved operations using the command: <code>ops = graph.get_operations()<\/code>. In the example above the first operation is the placeholder assignment <em>\/input_ph<\/em> while the last operation was named <em>\/last_operation_in_the_network<\/em> (normally, if the author of the network doesn't specify the field <code>name<\/code> per each layer, you will find something like <em>\/dense_3, \/conv2d_1<\/em> etc.). Note that it must be the exact final operation of your model and, also, you must provide the <strong>tensor<\/strong> which is the value <code>.outputs[0]<\/code> of the operation itself.<\/p>\n\n<p>Finally, you can run the session with the correct inference operation and placeholder, getting the prediction as numpy array and doing whatever you want (averaging, majority voting, etc.)<\/p>\n\n<p>Useful links that you may want to check:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.reddit.com\/r\/learnmachinelearning\/comments\/8hw2bf\/how_to_restore_model_and_make_prediction_with_it\/?depth=2\" rel=\"nofollow noreferrer\">reddit post<\/a> <\/li>\n<li><a href=\"https:\/\/cv-tricks.com\/tensorflow-tutorial\/save-restore-tensorflow-models-quick-complete-tutorial\/\" rel=\"nofollow noreferrer\">A quick complete tutorial to save and restore Tensorflow models<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":35465651,"AnswerCount":2,"Tags":"<c++><opencv><caffe><mnist>","CreationDate":"2016-02-17T19:14:04.753","AcceptedAnswerId":null,"OwnerUserId":733941.0,"Title":"How to use opencv_dnn module for digit classification by using LeNet?","Body":"<p>I tried to use the same code with some modification <a href=\"http:\/\/docs.opencv.org\/master\/d5\/de7\/tutorial_dnn_googlenet.html\" rel=\"nofollow\">here<\/a> for digit recognitions using a trained model for mnist based on Caffe. \nMy modified source code and network is as follows. I also included the error message at the end. The image is used in the code is a sample image from mnist.\nThe line of code that it cannot go through is <\/p>\n\n<pre><\/pre>\n\n<p>And I'm pretty sure it is because of the dimension of input. <\/p>\n\n<p>Any ideas how to run the code for LeNet?<\/p>\n\n<p><strong>caffe_mnist.cpp<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>mnist.prototxt<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36215234","CreationDate":"2016-03-25T06:38:56.800","ParentId":null,"OwnerUserId":"6092010","Title":null,"Body":"<p>Perhaps you are applying an incorrect image. Correct line: \nMat img = imread(imageFile, CV_LOAD_IMAGE_GRAYSCALE); <\/p>\n"},{"AnswerId":"37287524","CreationDate":"2016-05-17T22:57:02.263","ParentId":null,"OwnerUserId":"3736842","Title":null,"Body":"<p>I faced the same problem. Upon changing it to CV_LOAD_IMAGE_GRAYSCALE, the code ran without errors. <\/p>\n"}]}
{"QuestionId":35469664,"AnswerCount":1,"Tags":"<tensorflow><skflow>","CreationDate":"2016-02-17T23:12:56.880","AcceptedAnswerId":null,"OwnerUserId":1810559.0,"Title":"RNN with numeric data in SKFLOW","Body":"<p>I want to try an SKFLOW recurrent neural network on some time sequence data with real values for a binary classification problem. Each row of my data contains 57 features (variables) and I would like to look at the previous 2 samples and the next 2 samples to make predictions on each row.<\/p>\n\n<p>My data looks like this:<\/p>\n\n<p>sample -2: f1, f2, f3, f4,...f57, sample -1: f1, f2, f3, f4,...f57, current sample: f1, f2, f3, f4,...f57, sample +1: f1, f2, f3, f4,...f57, sample +2: f1, f2, f3, f4,...f57<\/p>\n\n<p>I started with the <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/examples\/text_classification_builtin_rnn_model.py\" rel=\"nofollow\">SKFLOW example RNN for text classification<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>It looks as if I should be able to just modify the input_op_fn to make it work but I am not sure how to correctly convert my numpy array to a tensor for the skflow.TensorFlowRNNClassifier. This is what it looks like for the text classification example.<\/p>\n\n<pre><\/pre>\n\n<p>If I am interpreting the text problem correctly then for my problem it would be \nTensorShape([Dimension(# rows), Dimension(57), Dimension(3)])<\/p>\n","answers":[{"AnswerId":"35471052","CreationDate":"2016-02-18T01:21:20.480","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>Check out this <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/tests\/test_nonlinear.py#L61\" rel=\"nofollow\">unit test<\/a> for RNN. <\/p>\n\n<p>Suppose this is the numeric data:\n<code>\n        data = np.array(list([[2, 1, 2, 2, 3],\n                              [2, 2, 3, 4, 5],\n                              [3, 3, 1, 2, 1],\n                              [2, 4, 5, 4, 1]]), dtype=np.float32)\n        labels = np.array(list([1, 0, 1, 0]), dtype=np.float32)\n<\/code><\/p>\n\n<p><code>data<\/code> is of shape <code>(4, 5)<\/code> where 4 is batch_size and 5 is the sequence_length. Then you want to have <code>tf.split(1, 5, X)<\/code> in <code>input_op_fn()<\/code>. Hope this helps. You are welcomed to submit an PR for adding an example dealing with this. <\/p>\n"}]}
{"QuestionId":35469731,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-17T23:19:31.580","AcceptedAnswerId":35473661.0,"OwnerUserId":4747268.0,"Title":"Tensorflow GPU error: ImportError: libcudart.so.7.0: cannot open shared object file: No such file or directory","Body":"<p>This is the traceback,<\/p>\n\n<pre><\/pre>\n\n<p>This is my .bashrc file,<\/p>\n\n<pre><\/pre>\n\n<p>I'm using Ubuntu 15.04, cuda7.5 an cudnn 4.0.7, and I have installed tensorflow using pip inside a virtualenv.<\/p>\n","answers":[{"AnswerId":"35473661","CreationDate":"2016-02-18T05:37:05.213","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Your problem is cuda7.5. As far as I understand, TensorFlow is linked in a way that it only works with cuda7.0. You either need to build it from source, or degrade your local cuda version to 7.0 (as an option, create a docker container and install lower version of cuda there, if you need 7.5 for other purposes).<\/p>\n"}]}
{"QuestionId":35470006,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2016-02-17T23:42:49.747","AcceptedAnswerId":35701181.0,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, which neural network structure for mini-batch training?","Body":"<p>I'm still working on implementing the mini-batch gradient update on my siamese neural network. Previously I had an implementation problem, that was <a href=\"https:\/\/stackoverflow.com\/questions\/31324633\/torch-lua-how-to-correctly-implement-minibatch-training-in-a-siamese-neural-n\">correctly solved here<\/a>. <\/p>\n\n<p>Now I realized that there was also a mistake in the architecture of my neural network, that is related to my incomplete understanding of the correct implementation.<\/p>\n\n<p>So far, I've always used a non-minibatch gradient descent approach, in which I was passing the training elements one by one to the gradient update. Now, I want to implement a gradient update through mini-batch, starting say with minibatches made of N=2 elements.<\/p>\n\n<p><strong>My question is: how should I change the architecture of my siamese neural network to make it able to handle a mini-batch of N=2 elements instead of a single element?<\/strong><\/p>\n\n<p>This is the (simplified) architecture of my siamese neural network:<\/p>\n\n<pre><\/pre>\n\n<p>I have:<\/p>\n\n<ul>\n<li>2 identical siamese neural networks (upper and lower)<\/li>\n<li>6 input units<\/li>\n<li>3 hidden units<\/li>\n<li>2 output units<\/li>\n<li>cosine distance function that compares the output of the two parallel neural networks<\/li>\n<\/ul>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>This architecture works very well if I have a gradient update function that takes 1 element; how should modify it to let it manage a <strong>minibatch<\/strong>?<\/p>\n\n<p><strong>EDIT:<\/strong> I probably should use the <a href=\"https:\/\/github.com\/Element-Research\/rnn\/blob\/master\/Sequencer.lua\" rel=\"nofollow noreferrer\">nn.Sequencer() class<\/a>, by modifying the last two lines of my code in: <\/p>\n\n<pre><\/pre>\n\n<p>What do you guys think?<\/p>\n","answers":[{"AnswerId":"35701181","CreationDate":"2016-02-29T13:45:38.597","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>Every <code>nn<\/code> module can work with minibatches. Some work only with minibatches, e.g. <code>(Spatial)BatchNormalization<\/code>. A module knows how many dimensions its input must contain (let's say D) and if the module receives a D+1 dimensional tensor, it assumes the first dimension to be the batch dimension. For example, take a look at <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/simple.md#linear\" rel=\"nofollow\"><code>nn.Linear<\/code> module documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>The input tensor given in forward(input) must be either a vector (1D\n  tensor) or matrix (2D tensor). If the input is a matrix, then each row\n  is assumed to be an input sample of given batch.<\/p>\n<\/blockquote>\n\n<pre><code>function table_of_tensors_to_batch(tbl)\n    local batch = torch.Tensor(#tbl, unpack(tbl[1]:size():totable()))\n    for i = 1, #tbl do\n       batch[i] = tbl[i] \n    end\n    return batch\nend\n\ninputs = {\n    torch.Tensor(5):fill(1),\n    torch.Tensor(5):fill(2),\n    torch.Tensor(5):fill(3),\n}\ninput_batch = table_of_tensors_to_batch(inputs)\nlinear = nn.Linear(5, 2)\noutput_batch = linear:forward(input_batch)\n\nprint(input_batch)\n 1  1  1  1  1\n 2  2  2  2  2\n 3  3  3  3  3\n[torch.DoubleTensor of size 3x5]\n\nprint(output_batch)\n 0,3128 -1,1384\n 0,7382 -2,1815\n 1,1637 -3,2247\n[torch.DoubleTensor of size 3x2]\n<\/code><\/pre>\n\n<p>Ok, but what about containers (<code>nn.Sequential<\/code>, <code>nn.Paralel<\/code>, <code>nn.ParallelTable<\/code> and others)? Container itself does not deal with an input, it just sends the input (or its corresponding part) to the corresponding module it contains. <code>ParallelTable<\/code>, for example, simply applies the i-th member module to the i-th input table element. Thus, if you want it to handle a batch, each input[i] (input is a table) must be a tensor with the batch dimension as described above. <\/p>\n\n<pre><code>input_number = 5\noutput_number = 2\n\ninputs1 = {\n    torch.Tensor(5):fill(1),\n    torch.Tensor(5):fill(2),\n    torch.Tensor(5):fill(3),\n}\ninputs2 = {\n    torch.Tensor(5):fill(4),\n    torch.Tensor(5):fill(5),\n    torch.Tensor(5):fill(6),\n}\ninput1_batch = table_of_tensors_to_batch(inputs1)\ninput2_batch = table_of_tensors_to_batch(inputs2)\n\ninput_batch = {input1_batch, input2_batch}\noutput_batch = perceptron:forward(input_batch)\n\nprint(input_batch)\n{\n  1 : DoubleTensor - size: 3x5\n  2 : DoubleTensor - size: 3x5\n}\nprint(output_batch)\n 0,6490\n 0,9757\n 0,9947\n[torch.DoubleTensor of size 3]\n\n\ntarget_batch = torch.Tensor({1, 0, 1})\ncriterion = nn.MSECriterion()\nerr = criterion:forward(output_batch, target_batch)\ngradCriterion = criterion:backward(output_batch, target_batch)\nperceptron:zeroGradParameters()\nperceptron:backward(input_batch, gradCriterion)\n<\/code><\/pre>\n\n<p>Why is there <code>nn.Sequencer<\/code> then? Can one use it instead? Yes, but it's <strong>highly not recommended<\/strong>. Sequencer takes a sequence table and applies the module to each element in the table independently providing no speedup. Besides, it has to make copies of that module, so such \"batch mode\" is considerably less efficient than online (non-batch) training. Sequencer was designed to be a part of recurrent nets, no point to using it in your case.<\/p>\n"}]}
{"QuestionId":35471900,"AnswerCount":0,"Tags":"<python><graph><tensorflow><tensorboard>","CreationDate":"2016-02-18T02:49:51.763","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Tensorboard cannot show the graph","Body":"<p>I've already installed the latest version of Tensorflow (version_0.7), and I followed tutorial steps on training and testing mnist data, which was good. When I visualized the result using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html#tensorboard-visualizing-learning\" rel=\"nofollow\">Tensorboard<\/a>, I noticed that I could only see results for  and , but except ... I tried different browsers including Safari, Chrome and Firefox, but nothing was showed for . I am sure all code is correct, as I download the code from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"nofollow\">website<\/a>. and I also remembered that  on Tensorboard worked for Tensorflow version_0.6. I am wondering if there is wrong with the new version, or my setups. I really appreciate for any helps.<\/p>\n","answers":[]}
{"QuestionId":35473510,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2016-02-18T05:26:22.090","AcceptedAnswerId":null,"OwnerUserId":5405823.0,"Title":"Python Deep Learning Keras: Wrong number of dimensions: expected 3, got 2 with shape","Body":"<p>I am new to Keras, a deep learning library and need your help. The model is built without errors, but has the following issue when calling model.fit(X, y):<\/p>\n\n<pre><\/pre>\n\n<p>This is similar as this one <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/815\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/issues\/815<\/a>\nMy y train matrix is a matrix with multiple rows and one column.<\/p>\n\n<p>One solution mentioned about converting y to 3d tensor with binary one-hot coding. Is there an example of this?<\/p>\n","answers":[{"AnswerId":"35598867","CreationDate":"2016-02-24T10:01:28.697","ParentId":null,"OwnerUserId":"2602240","Title":null,"Body":"<p>You can use: <\/p>\n\n<pre><code>from keras.utils import np_utils\nnp_utils.to_categorical(y_train, n_classes)\n<\/code><\/pre>\n\n<p>for one hot encoding, where y_train is a train class vector and n_classes - total number of a classes, <\/p>\n\n<p>However, based on that the error description mentions (16, 40), rather than a (Nx1), I suspect you might have an issue with your X as well.<\/p>\n"},{"AnswerId":"35477191","CreationDate":"2016-02-18T09:14:16.490","ParentId":null,"OwnerUserId":"1586821","Title":null,"Body":"<p>Are you looking for something like this?<\/p>\n\n<pre><code>&gt;b=np.arange(640)\n&gt;b.reshape(16,40).shape\n (16, 40)\n&gt;c=b.reshape(1,16,40)\n&gt;c.shape\n (1, 16, 40)\n<\/code><\/pre>\n"}]}
{"QuestionId":35474974,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-18T07:08:11.620","AcceptedAnswerId":null,"OwnerUserId":5944204.0,"Title":"Not able to run fully_connected_feed.py in Tensorflow","Body":"<p>I am following the tutorial of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/tf\/index.html\" rel=\"nofollow\">TensorFlow Mechanics 101<\/a> (version 0.7.0). As per the document, I download the two files ( and ) and save them to the same directory on my local machine.<\/p>\n\n<p>When I run the following command:<\/p>\n\n<pre><\/pre>\n\n<p>...I get this error: . The full output and stack trace are below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35487956","CreationDate":"2016-02-18T17:07:19.127","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is a bug in the 0.7.0 release of TensorFlow, which was fixed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/42f06d870014dec70769cbdd8697821c92880be9\" rel=\"nofollow\">a recent commit<\/a> and will appear in a bugfix release shortly. The issue is caused when the <code>--train_dir<\/code> flag doesn't contain a directory name component.<\/p>\n\n<p>In the meantime, you can avoid this issue by passing the flag <code>--train_dir=.\/<\/code> when you run the example.<\/p>\n"},{"AnswerId":"35502872","CreationDate":"2016-02-19T10:22:50.690","ParentId":null,"OwnerUserId":"5814701","Title":null,"Body":"<p>This should be a comment to mrry's post (I'm missing reputation)<\/p>\n\n<p>Changing line #42 from fully_connected_feed.py to<\/p>\n\n<pre><code>flags.DEFINE_string('train_dir', '.\/data', 'Directory to put the training data.')\n<\/code><\/pre>\n\n<p>solved the problem for me. I'm also on 0.7.0 and was able to run all other mnist examples.<\/p>\n"}]}
{"QuestionId":35477548,"AnswerCount":1,"Tags":"<image-processing><neural-network><tensorflow>","CreationDate":"2016-02-18T09:31:21.823","AcceptedAnswerId":null,"OwnerUserId":5944790.0,"Title":"Interpretation of Neural Network (CNN) Result \/ Accuracy","Body":"<p>I'm kind of new to the subject and build a convolutional neural network based on google's tensorflow. I wanted to classify a test data set of pictures belonging to 10 categories. My CNN setup is aligned to the tensorflow tutorial with some amendmends to meet my images' size. \nI ran the trainig step repeatedly for 20 times over a random sample of 500 images and then repeated that step for 50 times on different samples of size 500. I used a sample of 200 as validation data set (kept this fixed for all runs). As a result I got an accuracy of about 35%, which isn't to bad in my eyes, since I didn't do any optimizations and the images are kind of hard to assign to a single category evan for humans.<\/p>\n\n<p>So here are my questions:<\/p>\n\n<ul>\n<li>Does it really make sense to run a step for 20 times over the same batch? (I did this becuase it's about what fits in the RAM and loading a new batch took quite a while - so I could get more runs in less time)<\/li>\n<li>In the training accuracy diagram (see below) there's a jump at some point around step 120-130. From there on the accuracy goes up close to 100% for each 20-run of the same random batch. What does that jump mean in terms of network structure \/ learning?<\/li>\n<\/ul>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Rkby5.png\" alt=\"Test and Training Accuracy Diagram\"><\/p>\n","answers":[{"AnswerId":"35479112","CreationDate":"2016-02-18T10:39:02.810","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>Your spikes are likely due to the network overfitting on the batch that you are repeatedly showing it, while not really learning something that is useful in general. This also answers your first question - in this case, it doesn't make sense. <\/p>\n"}]}
{"QuestionId":35481805,"AnswerCount":1,"Tags":"<c++><opencv><caffe>","CreationDate":"2016-02-18T12:42:35.317","AcceptedAnswerId":null,"OwnerUserId":3437275.0,"Title":"Using Caffe memory layer does NOT produce consistent and deterministic results","Body":"<p>I am using the Caffe framework for windows (downloaded from <a href=\"https:\/\/github.com\/happynear\/caffe-windows\" rel=\"nofollow\">here<\/a>) on a Windows 7 64-bit machine. I am using C++ within Visual Studio Community 2013. I use the pre-trained GoogLeNet model to extract the loss1-fc layer output to use as a feature vector for each image. So far so good.<\/p>\n\n<p>Recently i tried changing my software for use with video frames. So i changed the first layer from an ImageData layer to a Memory layer, so i can send to Caffe a vector of OpenCV mats instead of the naive approach of writing each frame to disk and sending a file list to caffe.<\/p>\n\n<p>Now, i noticed i won't get the same results for the same images! When using the ImageData layer, there's no such thing. <\/p>\n\n<p>I use CPU (no Cudnn, no GPU).<\/p>\n\n<p>The function i use for feature extraction is the following:<\/p>\n\n<pre><\/pre>\n\n<p>The  vector is a vector of mat. The  is a vector of int, all set to 0. I wrote all images to disk again once they were added to vector. I checked and there is no problem with that. So there is nothing wrong when loading the images. By the way i use OpenCV 3.1.<\/p>\n\n<p>The memory layer in GoogLeNet prototxt file is declared as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and is the first layer.<\/p>\n\n<p>I print the first 10 values for each image. Note that images 0, 1, 2, 3 are the EXACT same file copied and the same holds for 6, 7 and 8 images.<\/p>\n\n<pre><\/pre>\n\n<p>2nd run:<\/p>\n\n<pre><\/pre>\n\n<p>The layers output is different for the same images and different for different runs! When using the same procedure with the ImageData layer there is no such problem. Also, the problem holds for the output of other layers too, for example loss3\/classifier. So, i suspect there might be a bug within the MemoryLayer implementation.<\/p>\n\n<p>Has anyone noticed this strange behaviour? I read that cudnn might produce non-deterministic results but i ran my model on CPU. Any thoughts on this are welcome.<\/p>\n","answers":[{"AnswerId":"35500846","CreationDate":"2016-02-19T08:41:46.583","ParentId":null,"OwnerUserId":"3437275","Title":null,"Body":"<p>I found out what went wrong and i'll post here the answer to help others.<\/p>\n\n<p>It turns out that GoogLeNet requires input images to be 224x224x3 size, and you must NOT subtract the mean in TEST phase. So by changing the definition of the memory layer in .prototxt file to this:<\/p>\n\n<pre><code>name: \"GoogleNet\"\nlayer {\n  name: \"data\"\n  type: \"MemoryData\"\n  top: \"data\"\n  top: \"label\"\n  memory_data_param {\n    batch_size: 1\n    channels: 3\n    height: 224\n    width: 224\n  }\n}\n...\n<\/code><\/pre>\n\n<p>i got the results i expected. Many thanks to @Miki for pointing out the OpenCV tutorial on their dnn module, which helped me clarify this.<\/p>\n"}]}
{"QuestionId":35483367,"AnswerCount":1,"Tags":"<tensorflow><skflow>","CreationDate":"2016-02-18T13:54:43.603","AcceptedAnswerId":35485392.0,"OwnerUserId":2902280.0,"Title":"Limit GPU memory allocation in skflow","Body":"<p>I am training convnets with Tensorflow and skflow, on an EC2 instance I share with other people. For all of us to be able to work at the same time, I'd like to limit the fraction of available GPU memory which is allocated. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory?rq=1\">This question<\/a> does it with Tensorflow, but since I'm using sklfow I'm never using a .<\/p>\n\n<p>Is it possible to do the same thing through skflow ?<\/p>\n","answers":[{"AnswerId":"35485392","CreationDate":"2016-02-18T15:19:34.603","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>At this moment, you can only control the number of cores (<code>num_cores<\/code>) to be used in estimators by passing this parameter to estimator. <\/p>\n\n<p>One can add <code>gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)<\/code> to <code>tf.ConfigProto<\/code> as suggested by <a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory?rq=1\">this question<\/a> you linked to achieve what you need. <\/p>\n\n<p>Feel free to submit a PR to make changes <a href=\"https:\/\/github.com\/tensorflow\/skflow\/blob\/master\/skflow\/estimators\/base.py#L160\" rel=\"nofollow noreferrer\">here<\/a> as well as adding this additional parameters to all estimators. Otherwise, I'll make the changes some time this week. <\/p>\n\n<p>Edit: <\/p>\n\n<p>I have made the changes to allow those options. Please check \"Building A Model Using Different GPU Configurations\" example in examples folder. Let me know if there's any particular need or other options you want to add. Pull requests are always welcomed!<\/p>\n"}]}
{"QuestionId":35486961,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-18T16:25:05.527","AcceptedAnswerId":null,"OwnerUserId":5379551.0,"Title":"how to store trained weights and biases in model A for use in model B in tensorflow","Body":"<p>Suppose I have construct a neural network A with a dictionary of weights and biases variable, and I have get specific values for weights and biases after training in neural network A. And I want to use these specific value to replace the  in neural network B. How could I achieve this purpose? I have tried , however, I don't know how to restore the weights and biases in another network B (in another file). I have also tried storing them using , however, I met with another problem, i.e. when restoring  and  using , it said  type is not hashable. Could anyone help me solve this problem?<\/p>\n","answers":[{"AnswerId":"35487604","CreationDate":"2016-02-18T16:52:17.790","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow\"><code>tf.train.Saver<\/code><\/a> class should help with this, although you might need to use some of the optional arguments to get it to work.<\/p>\n\n<p>Let's say your model A looks like this, and you've trained it and saved it to a file called <code>\"\/tmp\/model_a_ckpt\"<\/code>:<\/p>\n\n<pre><code>weights_a = tf.Variable(..., name=\"weights_a\")\nbiases_a = tf.Variable(..., name=\"biases_a\")\n# ...\n\nsaver_a = tf.train.Saver()\n# ...\nsaver_a.save(sess, \"\/tmp\/model_a_ckpt\")\n<\/code><\/pre>\n\n<p>...and then let's say your model B looks like this:<\/p>\n\n<pre><code>weights_b = tf.Variable(..., name=\"weights_b\")\nbiases_b = tf.Variable(..., name=\"biases_b\")\n<\/code><\/pre>\n\n<p>To load the checkpoint into model B you have to create a saver that <em>maps<\/em> the names of variables in the checkpoint (i.e. <code>\"weights_a\"<\/code> and <code>\"biases_a\"<\/code> because they default to the <code>name<\/code> property of the corresponding <code>tf.Variable<\/code> objects in model A) to the variables in model B:<\/p>\n\n<pre><code>saver_b = tf.train.Saver({\"weights_a\": weights_b, \"biases_a\": biases_b})\n# ...\nsaver_b.restore(sess, \"\/tmp\/model_a_ckpt\")\n<\/code><\/pre>\n\n<p>After running <code>saver_b.restore()<\/code>, your variables in model B will have the values trained in model A.<\/p>\n"}]}
{"QuestionId":35487598,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-02-18T16:52:00.610","AcceptedAnswerId":44975517.0,"OwnerUserId":1994648.0,"Title":"Sampling Bernoulli random variables in TensorFlow","Body":"<p>Given a 1D tensor containing the means of a Bernoulli distribution, how do I sample a corresponding 1D tensor with the given means?<\/p>\n\n<p>TensorFlow only seems to have  and  functions implemented. I could use something complicated like:<\/p>\n\n<pre><\/pre>\n\n<p>but the  function has no gradient defined in TensorFlow.<\/p>\n","answers":[{"AnswerId":"50230661","CreationDate":"2018-05-08T09:46:43.340","ParentId":null,"OwnerUserId":"3986320","Title":null,"Body":"<p><a href=\"https:\/\/www.linkedin.com\/pulse\/deep-learning-model-rbmrestricted-boltzmann-machine-deepak-kumar\/\" rel=\"nofollow noreferrer\">I've seen<\/a> also the following trick as a way of sampling from the Bernoulli distribution:<\/p>\n\n<pre><code>tf.nn.relu(tf.sign(means - tf.random_uniform(tf.shape(means))))\n<\/code><\/pre>\n"},{"AnswerId":"44975517","CreationDate":"2017-07-07T16:17:36.693","ParentId":null,"OwnerUserId":"7303170","Title":null,"Body":"<p>Since TFr1.0, <code>tf.select<\/code> is deprecated in favor of <code>tf.where<\/code>. Furthermore, the answer given by @keveman should compare the uniform random sampling with &lt; 0, neither with > 0.5 nor with > 0:<\/p>\n\n<pre><code>    means = tf.constant([.3,.8])\n    sample = tf.where(tf.random_uniform([1, 2]) - means &lt; 0, \n      tf.ones([1,2]), tf.zeros([1,2]))\n    with tf.Session(''): sample.eval()\n<\/code><\/pre>\n"},{"AnswerId":"35492614","CreationDate":"2016-02-18T21:18:50.757","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>You can use <code>tf.select<\/code>, which is differentiable.<\/p>\n\n<pre><code>means = tf.constant([.3,.8])\na = tf.select(tf.random_uniform([1, 2])- means &gt; 0.5, tf.ones([1,2]), tf.zeros([1,2]))\nwith tf.Session(''): a.eval()\n<\/code><\/pre>\n"}]}
{"QuestionId":35488019,"AnswerCount":1,"Tags":"<tensorflow><gradient-descent><momentum>","CreationDate":"2016-02-18T17:10:11.757","AcceptedAnswerId":null,"OwnerUserId":5782844.0,"Title":"What's different about momentum gradient update in Tensorflow and Theano like this?","Body":"<p>I'm trying to use TensorFlow with my deep learning project.<br>\nHere I need implement my gradient update in this formula :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qJw0f.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qJw0f.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have also implement this part in Theano, and it came out the expected answer. But when I try to use TensorFlow's , the result is really bad. I don't know what is different between them.<\/p>\n\n<p>Theano:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>TensorFlow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"35683298","CreationDate":"2016-02-28T13:46:25.350","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>If you look at the implementation of momentum optimizer in TensorFlow [<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/training_ops.cc#L129\" rel=\"noreferrer\">link<\/a>], it is implemented as follows:<\/p>\n\n<pre><code>accum = accum * momentum() + grad;\nvar -= accum * lr();\n<\/code><\/pre>\n\n<p>As you see, the formulas are a bit different. Scaling momentum term by the learning rate should resolve your differences.<\/p>\n\n<p>It is also very easy to implement such optimizer by yourself. The resulting \ncode would look similar to the snippet in Theano that you included.<\/p>\n"}]}
{"QuestionId":35488717,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-18T17:46:09.110","AcceptedAnswerId":38059483.0,"OwnerUserId":307401.0,"Title":"Confused about conv2d_transpose","Body":"<p>I'm getting this error message when using :<\/p>\n\n<pre><\/pre>\n\n<p>However, it occurs after the graph is built while compiling the loss function (Adam). Any ideas on what would cause this? I suspect it's related to the input dimensions but I'm not sure exactly why.<\/p>\n\n<p>Full error: <a href=\"https:\/\/gist.github.com\/jimfleming\/75d88e888044615dd6e3\">https:\/\/gist.github.com\/jimfleming\/75d88e888044615dd6e3<\/a><\/p>\n\n<p>Relevant code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38059483","CreationDate":"2016-06-27T17:05:03.037","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>Thanks for the question! You're exactly right---the problem is that the input and output dimensions being passed to tf.nn.conv2d_transpose don't agree. (The error may be detected when computing gradients, but the gradient computation isn't the problem.)<\/p>\n\n<p>Let's look at just the first part of your code, and simplify it a little bit:<\/p>\n\n<pre><code>sess = tf.Session()\nbatch_size = 3\noutput_shape = [batch_size, 8, 8, 128]\nstrides = [1, 2, 2, 1]\n\nl = tf.constant(0.1, shape=[batch_size, 32, 32, 4])\nw = tf.constant(0.1, shape=[7, 7, 128, 4])\n\nh1 = tf.nn.conv2d_transpose(l, w, output_shape=output_shape, strides=strides, padding='SAME')\nprint sess.run(h1)\n<\/code><\/pre>\n\n<p>I replaced the variables with constants --- it's easier to see what's going on.<\/p>\n\n<p>If you try to run this code, you get a similar error:<\/p>\n\n<pre><code>InvalidArgumentError: Conv2DCustomBackpropInput: Size of out_backprop doesn't match computed: actual = 32, computed = 4\n  [[Node: conv2d_transpose_6 = Conv2DBackpropInput[T=DT_FLOAT, data_format=\"NHWC\", padding=\"SAME\", strides=[1, 2, 2, 1], use_cudnn_on_gpu=true, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](conv2d_transpose_6\/output_shape, Const_25, Const_24)]]\n<\/code><\/pre>\n\n<p>Now, the error is a little misleading --- it talks about the 'out_backprop' argument to 'Conv2DCustomBackpropInput'. The key is that tf.nn.conv2d_transpose is actually just the gradient of tf.nn.conv2d, so Tensorflow uses the same code internally (Conv2DCustomBackpropInput) to compute the gradient of tf.nn.conv2d and to compute tf.nn.conv2d_transpose.<\/p>\n\n<p>The error means that the 'output_shape' you requested is not possible, given the shapes of 'l' and 'w'.<\/p>\n\n<p>Since tf.nn.conv2d_transpose is the backward (gradient) counterpart of tf.nn.conv2d, one way to see what the correct shapes should be is to use the corresponding forward operation:<\/p>\n\n<pre><code>output = tf.constant(0.1, shape=output_shape)\nexpected_l = tf.nn.conv2d(output, w, strides=strides, padding='SAME')\nprint expected_l.get_shape()\n# Prints (3, 4, 4, 4)\n<\/code><\/pre>\n\n<p>That is, in the forward direction, if you provided a tensor of shape 'output_shape', you would get out a tensor of shape (3, 4, 4, 4). \nSo one way to fix the problem is to change the shape of 'l' to (3, 4, 4, 4); if you change the code above to:<\/p>\n\n<pre><code>l = tf.constant(0.1, shape=[batch_size, 4, 4, 4])\n<\/code><\/pre>\n\n<p>everything works fine.<\/p>\n\n<p>In general, try using tf.nn.conv2d to get a feel for what the relationship between the tensor shapes is. Since tf.nn.conv2d_transpose is its backward counterpart, it has the same relationship between input, output and filter shapes (but with the roles of the input and output reversed.) <\/p>\n\n<p>Hope that helps!<\/p>\n"},{"AnswerId":"40485785","CreationDate":"2016-11-08T11:30:09.880","ParentId":null,"OwnerUserId":"6457746","Title":null,"Body":"<p>Using <code>padding='SAME'<\/code> in <code>tf.nn.conv2d_transpose()<\/code> function may works too<\/p>\n"}]}
{"QuestionId":35491633,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-02-18T20:19:42.797","AcceptedAnswerId":null,"OwnerUserId":3320467.0,"Title":"At what stage is a tensorflow graph set up?","Body":"<p>An optimizer typically run the same computation graph for many steps until convergence. Does tensorflow setup the graph at the beginning and reuse it for every step? What if I change the batch size during training? What if I make some minus change to the graph like changing the loss function? What if I made some major change to the graph? Does tensorflow pre-generate all possible graphs? Does tensorflow know how to optimize the entire computation when the graph changes?<\/p>\n","answers":[{"AnswerId":"35492929","CreationDate":"2016-02-18T21:38:49.043","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow exposes only one graph that is visible to the user, namely the one specified by the user. The user can run the graph with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>Session.run()<\/code><\/a> or by calling <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Tensor.eval\" rel=\"noreferrer\"><code>Tensor.eval()<\/code><\/a> on some tensor. A <code>Session.run()<\/code> call can specify some tensors to be fed and others to be fetched. Depending on what needs to be fetched, the TensorFlow runtime could be internally constructing and optimizing various data structures, including a pruned version of the user visible graph. However, this internal graph is not visible to the user in anyway. No, TensorFlow doesn't 'pre-generate' all possible graphs. Yes, TensorFlow does perform extensive optimizations on the computation graph. And finally, changing the batch size of a tensor that is fed doesn't change the structure of the graph.<\/p>\n"},{"AnswerId":"35518402","CreationDate":"2016-02-20T02:27:25.147","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As <a href=\"https:\/\/stackoverflow.com\/a\/35492929\/3574081\">keveman says<\/a>, from the client's perspective there is a single TensorFlow graph. In the runtime, there can be multiple <em>pruned<\/em> subgraphs that contain just the nodes that are necessary to compute the values <code>t1<\/code>, <code>t2<\/code> etc. that you fetch when calling <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow noreferrer\"><code>sess.run([t1, t2, ...])<\/code><\/a>.<\/p>\n\n<p>If you call <code>sess.run([t1, t2])<\/code> will prune the overall graph (<code>sess.graph<\/code>) down to the subgraph required to compute those values: i.e. the operations that produce <code>t1<\/code> and <code>t2<\/code> and all of their antecedents. If you subsequently call <code>sess.run([t3, t4])<\/code>, the runtime will prune the graph down to the subgraph required to compute <code>t3<\/code> and <code>t4<\/code>. Each time you pass a new combination of values to fetch, TensorFlow will compute a new pruned graph and cache it&mdash;this is why the first <code>sess.run()<\/code> can be somewhat slower than subsequent ones.<\/p>\n\n<p>If the pruned graphs overlap, TensorFlow will reuse the \"kernel\" for the ops that are shared. This is relevant because some ops (e.g. <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow noreferrer\"><code>tf.Variable<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#FIFOQueue\" rel=\"nofollow noreferrer\"><code>tf.FIFOQueue<\/code><\/a>) are <em>stateful<\/em>, and their contents can be used in both pruned graphs. This allows you, for example, to initialize your variables with one subgraph (e.g. <code>sess.run(tf.initialize_all_variables())<\/code>), train them with another (e.g. <code>sess.run(train_op)<\/code>), and evaluate your model with a third (e.g. <code>sess.run(loss, feed_dict={x: ...})<\/code>). It also lets you enqueue elements to a queue with one subgraph, and dequeue them with another, which is the foundation of the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/threading_and_queues\/index.html\" rel=\"nofollow noreferrer\">input pipelines<\/a>.<\/p>\n"}]}
{"QuestionId":35491844,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-02-18T20:32:32.217","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Multiple Event files on Tensorboard","Body":"<p>I am using Tensorboard to show the training results of the code using Tensorflow (0.7). The previous Tensorflow version had <a href=\"https:\/\/stackoverflow.com\/questions\/34233767\/event-files-in-google-tensorflow\">issue<\/a> for multiple event files: when I run my local server using , it shows up error if there are more than 1 event files. It seems that the latest version (0.7) does not show up the same error for multiple event files, but it still shows the overlapped curves for multiple event files on Tensorboard. I wonder how to solve this problem. Thanks! <\/p>\n","answers":[{"AnswerId":"37156919","CreationDate":"2016-05-11T08:22:53.350","ParentId":null,"OwnerUserId":"6297189","Title":null,"Body":"<p>Training my own networks, I am writing summaries in different subfolders like <code>\/tmp\/project\/train<\/code>  and <code>\/tmp\/project\/eval<\/code>. If you start TensorBoard using <\/p>\n\n<pre><code>tensorboard --logdir=\/tmp\/project\/\n<\/code><\/pre>\n\n<p>you will still receive multiple graphs from each event file in the subfolders at once, as you mentioned. To see seperate graphs you can start TensorBoard from the desired subfolder: <\/p>\n\n<pre><code>tensorboard --logdir=\/tmp\/project\/train\/\n<\/code><\/pre>\n"}]}
{"QuestionId":35493764,"AnswerCount":1,"Tags":"<python><arrays><image><numpy><tensorflow>","CreationDate":"2016-02-18T22:31:01.100","AcceptedAnswerId":35494199.0,"OwnerUserId":5948325.0,"Title":"After calculating a tensor, how can I show it as a image?","Body":"<p>I have one dimensional numpy array. After performing a calculation in TensorFlow, I get a  as output. I am trying to reshape it into a 2-dimensional array and show it as an image.<\/p>\n\n<p>If it were a numpy ndarray, I would know how to plot it as an image. But it is a tensor now!<\/p>\n\n<p>Although I tried  to convert it into numpy array, I got an error saying \"No default session\".<\/p>\n\n<p>Can anyone teach me how to show a tensor as an image?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35494199","CreationDate":"2016-02-18T23:00:20.100","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The immediate error you're seeing is because <code>Tensor.eval()<\/code> only works when there is a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#get_default_session\" rel=\"nofollow\">\"default <code>Session<\/code>\"<\/a>. This requires that either (i) you're executing in a <code>with tf.Session():<\/code> block, (ii) you're executing in a <code>with sess.as_default():<\/code> block, or (iii) you're using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#InteractiveSession\" rel=\"nofollow\"><code>tf.InteractiveSession<\/code><\/a>. <\/p>\n\n<p>There are two simple workarounds to make your case work:<\/p>\n\n<pre><code># Pass the session to eval().\nax.imshow(img_res.eval(session=sess))\n\n# Use sess.run().\nax.imshow(sess.run(img_res))\n<\/code><\/pre>\n\n<p>Note that, as a large point about visualizing your image, you might consider using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#image_summary\" rel=\"nofollow\"><code>tf.image_summary()<\/code><\/a> op along with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">TensorBoard<\/a> to visualize tensors produced by a larger training pipeline.<\/p>\n"}]}
{"QuestionId":35493910,"AnswerCount":2,"Tags":"<lua><torch>","CreationDate":"2016-02-18T22:39:03.477","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Torch \/ Lua, how to edit an original abstract file of a known package?","Body":"<p>I tried to add a function to the  <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/Module.lua\" rel=\"nofollow\">abstract file<\/a> in the Torch package called , but my main program does not find it.<\/p>\n\n<p>Suppose my function is simply:<\/p>\n\n<pre><\/pre>\n\n<p>I added this function at the end of the  file and I was supposed to use it in my Torch terminal:<\/p>\n\n<pre><\/pre>\n\n<p>But the system generates:<\/p>\n\n<pre><\/pre>\n\n<p>Probably the system is not aware of this new function that I added... what should I do to use this new method?<\/p>\n","answers":[{"AnswerId":"35494230","CreationDate":"2016-02-18T23:02:06.500","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>You simply defined a global function <code>printTry<\/code>, but you call it as a method on <code>perceptron<\/code>. You need to define it as a field of <code>Module<\/code> (in that <code>Module.lua<\/code> file, assuming this is what <code>nn.Sequential<\/code> returns):<\/p>\n\n<pre><code>function Module:printTry()\n  print(\"printTry()\")\nend\n<\/code><\/pre>\n"},{"AnswerId":"35495926","CreationDate":"2016-02-19T01:50:45.980","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Use torch.getmetatable:<\/p>\n\n<pre><code>require 'nn'\ntorch.getmetatable('nn.Module').printTry = function() print('PrintTry') end\n\nperceptron = nn.Sequential()\nperceptron:printTry()\n<\/code><\/pre>\n"}]}
{"QuestionId":35495673,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-19T01:24:10.593","AcceptedAnswerId":null,"OwnerUserId":604964.0,"Title":"How to randomly choose and apply an op from a list of ops in tensorflow?","Body":"<p>For data augmentation purposes I want to have n different transformations to the data and want to randomly choose and apply one of them for each image in a batch. Something like:<\/p>\n\n<pre><\/pre>\n\n<p>Is that possible?<\/p>\n\n<p>Obs: I want the op to be chosen randomly while the session is being executed.<\/p>\n","answers":[{"AnswerId":"35509400","CreationDate":"2016-02-19T15:53:45.897","ParentId":null,"OwnerUserId":"604964","Title":null,"Body":"<p>I think I came up with a solution:<\/p>\n\n<pre><code>def apply_random_op(tensor, ops):\n  n = len(ops)\n  rand_idx = tf.floor(tf.random_uniform([], 0, n, dtype=tf.float32))\n  op_idx = tf.constant(0.0, dtype=tf.float32)\n  chain = tf.cond(tf.equal(op_idx, rand_idx), lambda: ops[0](tensor), lambda: ops[1](tensor))\n  for i in xrange(2, n):\n    op_idx = tf.constant(float(i), dtype=tf.float32)\n    chain = tf.cond(tf.equal(op_idx, rand_idx), lambda: ops[i](tensor), lambda: chain)\n  return chain\n<\/code><\/pre>\n"}]}
{"QuestionId":35495851,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-02-19T01:42:40.640","AcceptedAnswerId":35496224.0,"OwnerUserId":4524141.0,"Title":"In TensorFlow, how to feed 2 parameters into session.run()","Body":"<p>I'm trying to get into TensorFlow, and trying to make some changes of beginner example.<\/p>\n\n<p>I am trying to combine <a href=\"https:\/\/github.com\/dennybritz\/nn-from-scratch\/blob\/master\/nn-from-scratch.ipynb\" rel=\"nofollow\">Implementing a Neural Network from Scratch <\/a> with <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow\">Deep MNIST for Experts<\/a><\/p>\n\n<p>I getting data by using . Basically, this line gives 2D X (,) and 2 class Y (0\/1)<\/p>\n\n<pre><\/pre>\n\n<p>The structure of network is the same as <em>Deep MNIST for Experts<\/em>. The difference is the session running function.<\/p>\n\n<pre><\/pre>\n\n<p>But that gives an<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone give me some hints on this issue? Here is the code.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Here is the correct code after struggling with TensorFlow:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35496224","CreationDate":"2016-02-19T02:23:29.383","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem arises because you redefine <code>y<\/code> on the following line:<\/p>\n\n<pre><code>y = tf.nn.softmax(tf.matmul(x,W) + b)\n<\/code><\/pre>\n\n<p>TensorFlow then gives an error because feeding <code>y_: y<\/code> in the <code>feed_dict<\/code> would be feeding a tensor with another tensor, which isn't possible (and\u2014even if it were\u2014this particular feed would create a circular dependency!).<\/p>\n\n<p>The solution is to rewrite your softmax and cross-entropy ops:<\/p>\n\n<pre><code>y_softmax = tf.nn.softmax(tf.matmul(x,W) + b)\ncross_entropy = -tf.reduce_sum(y_*tf.log(y_softmax))\n<\/code><\/pre>\n"}]}
{"QuestionId":35497710,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-02-19T05:06:46.740","AcceptedAnswerId":null,"OwnerUserId":2853544.0,"Title":"Convolution and Pooling in TensorFlow Deep MNIST","Body":"<p>When I study <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/pros\/index.html#deep-mnist-for-experts\" rel=\"nofollow noreferrer\">Deep MNIST for Experts tutorial<\/a>, I have many difficulties.\nI'd to know why they used <strong>Convolution and Pooling<\/strong> in a Multilayer Convolutional Network.<\/p>\n\n<p>And I don't understand the following two functions.<\/p>\n\n<pre><\/pre>\n\n<p>I'd to know the meaning of <strong>strides=[1,1,1,1]<\/strong> in conv2d function.<\/p>\n\n<p>Should we always use <strong>ksize=[1, 2, 2, 1]<\/strong> and <strong>strides=[1, 2, 2, 1]<\/strong> in max_pool_2x2 function.<\/p>\n\n<p>What is the difference between <strong>padding='SAME'<\/strong> and <strong>padding='VALID'<\/strong><\/p>\n","answers":[{"AnswerId":"40986314","CreationDate":"2016-12-06T01:26:54.020","ParentId":null,"OwnerUserId":"6080123","Title":null,"Body":"<p>I would say check the following <a href=\"https:\/\/stackoverflow.com\/questions\/34619177\/what-does-tf-nn-conv2d-do-in-tensorflow\/34698115#34698115\">answer<\/a>. It has a wonderful explanation for the whole convolution operation. This should cover your query for conv2d .<\/p>\n\n<p>for max pooling, <\/p>\n\n<p><strong>ksize<\/strong>: is basically the kernal size. Its the size of the window for each dimension of the input tensor. you can change it according to your need. Like in the paper <a href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow noreferrer\">AlexNet<\/a> they have used ksize=[1, 3, 3, 1] and <\/p>\n\n<p><strong>stride<\/strong>: The filter is applied to image patches of the same size as the filter and strided according to the strides argument. strides = [1, 2, 2, 1] applies the filter to every other image patch in each dimension, etc.<\/p>\n\n<p>The difference of padding is explained well in this <a href=\"https:\/\/stackoverflow.com\/questions\/37674306\/what-is-the-difference-between-same-and-valid-padding-in-tf-nn-max-pool-of-t\">post<\/a>.<\/p>\n"}]}
{"QuestionId":35497979,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2016-02-19T05:32:30.560","AcceptedAnswerId":35501042.0,"OwnerUserId":3755873.0,"Title":"KeyError: 'data' in Caffe","Body":"<p>While I was following the deepdream iPython notebook which is here: <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"nofollow\">https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb<\/a>, I successfully ran the code and initialized the network until i get this error:<\/p>\n\n<pre><\/pre>\n\n<p>This is my code for the python file:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know what's happening? I am using my own data that I successfully trained a model with.<\/p>\n","answers":[{"AnswerId":"35501042","CreationDate":"2016-02-19T08:51:36.450","ParentId":null,"OwnerUserId":"5352059","Title":null,"Body":"<p>From the deepdream iPython notebook:<\/p>\n\n<pre><code>net = caffe.Classifier('tmp.prototxt', param_fn,\n                       mean = np.float32([104.0, 116.0, 122.0]), # ImageNet mean, training set dependent\n                       channel_swap = (2,1,0)) # the reference model has channels in BGR order instead of RGB\n<\/code><\/pre>\n\n<p>vs your:<\/p>\n\n<pre><code>net = caffe.Classifier('\/home\/andrew\/caffe\/models\/bvlc_reference_caffenet\/deploy.prototxt', '\/home\/andrew\/caffe\/models\/bvlc_reference_caffenet\/caffenet_train_iter_500.caffemodel', caffe.TEST)\n<\/code><\/pre>\n\n<p>You do not seem to include a mean when you create a <code>caffe.Classifier<\/code>.<\/p>\n\n<p>See the definition of <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py\" rel=\"nofollow\"><code>caffe.Classifier<\/code><\/a>.<\/p>\n\n<p>If you don't have a mean, you could probably just remove the mention of mean from preprocess\/deprocess:<\/p>\n\n<pre><code>def preprocess(net, img):\n    return np.float32(np.rollaxis(img, 2)[::-1])\ndef deprocess(net, img):\n    return np.dstack((img)[::-1])\n<\/code><\/pre>\n"}]}
{"QuestionId":35498064,"AnswerCount":1,"Tags":"<neural-network><theano><tensorflow><deep-learning>","CreationDate":"2016-02-19T05:40:13.573","AcceptedAnswerId":null,"OwnerUserId":2886263.0,"Title":"DNN: using framework like Theano or TensorFlow, how to make sure the gradient or calculation is correct when the NN has more than several layers?","Body":"<p>I already write some python or matlab code for neural network, but not using any framework or auto differentiation, but as we know, Theano and TensorFlow using auto differentiation, you build a calculation graph, they do calcultaion(back propagation) for you, but some times written program can run but definitely not run as I wish, So i wonder have some methods to make sure my program is correct?\nprinting the constructed calculation graph? but seems complicated when the number of NN layers is big like the winner of Imagenet adopted 152 layers\nOr write another program using simple matlab or python code, then compare this output with program using framework?<\/p>\n","answers":[{"AnswerId":"35499073","CreationDate":"2016-02-19T06:51:18.740","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Standard solution is numerical gradient checking. You can inefficiently compute gradient by doing forward propagation at two nearby values.<\/p>\n\n<p>See section on numerical gradient checking here:\n<a href=\"https:\/\/web.stanford.edu\/class\/cs294a\/sparseAutoencoder_2011new.pdf\" rel=\"nofollow\">https:\/\/web.stanford.edu\/class\/cs294a\/sparseAutoencoder_2011new.pdf<\/a><\/p>\n\n<p>In TensorFlow this is implemented using compute_numeric_jacobian  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/kernel_tests\/gradient_checker.py#L93\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":35498852,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><eigenvalue>","CreationDate":"2016-02-19T06:38:22.763","AcceptedAnswerId":35499462.0,"OwnerUserId":4727011.0,"Title":"Why the eigenvalue output is different for tensorflow and numpy?","Body":"<p>I tried to calculate Eigenvalues for matrix using both numpy and tensorflow but I am getting different eigenvalues for each implementation. Below are the details<\/p>\n\n<pre><\/pre>\n\n<p>EigenValues of A with numpy are  <\/p>\n\n<p>EigenValues of A with tensorflow are <\/p>\n\n<p>I used  for tensorflow implementation and  for numpy implementation.<\/p>\n","answers":[{"AnswerId":"35499462","CreationDate":"2016-02-19T07:14:02.223","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>From description of the function:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#self_adjoint_eig\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#self_adjoint_eig<\/a><\/p>\n\n<blockquote>\n  <p>Calculates the Eigen Decomposition of a square Self-Adjoint matrix.<\/p>\n  \n  <p>Only the lower-triangular part of the input will be used in this case.\n  The upper-triangular part will not be read.<\/p>\n<\/blockquote>\n\n<p>Therefore TensorFlow's <code>self_adjoint_eig<\/code> on your matrix is equivalent to numpy's <code>eig<\/code> of the following matrix <\/p>\n\n<blockquote>\n  <p>({1,1,1},{1,2,2},{1,2,3})<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":35500242,"AnswerCount":1,"Tags":"<tensorflow><logistic-regression>","CreationDate":"2016-02-19T08:01:28.193","AcceptedAnswerId":null,"OwnerUserId":5949009.0,"Title":"Implement Logistic Regression(LR) in Tensorflow,But Result looks strange when using bias but good without bias","Body":"<p>I tried the LR for sparse dataset using the tensorflow package(0.7.0)<br>\nThe following is part of my procedure:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>but I find the results is wired when using the bias,using the bias I find the abs(cost) became larger and larger and even the abs(bias) also became larger and larger.  <\/p>\n\n<p>IN the early iterations the results like this :  <\/p>\n\n<pre><\/pre>\n\n<p><strong>but with more iteration the results like this :<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<p>However if I donot use the bias the results looks good, what maybe the problems and I wonder the reasons very much.<\/p>\n\n<p>Can anynone help me, Thank you!<\/p>\n","answers":[{"AnswerId":"41725842","CreationDate":"2017-01-18T17:31:41.783","ParentId":null,"OwnerUserId":"7372608","Title":null,"Body":"<p>I think the use of \u201cembedding_lookup_sparse\u201d is wrong.\nLook at this<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/cb17d1b0e2b581b5da1d9597b7929ba764749d38\/tensorflow\/python\/ops\/embedding_ops.py#L280\" rel=\"nofollow noreferrer\">source code of embedding_lookup_sparse<\/a><\/p>\n\n<p>the default value of combiner is \"mean\", but if you want to get the w * x, which x is a sparse data, you should set the combiner is \"sum\", not the \"mean\".\nSo I think this is your main problem.<\/p>\n"}]}
{"QuestionId":35504528,"AnswerCount":2,"Tags":"<python><amazon-web-services><amazon-ec2><tensorflow><skflow>","CreationDate":"2016-02-19T11:45:22.177","AcceptedAnswerId":null,"OwnerUserId":2902280.0,"Title":"skflow allocates memory in gpu0 even when other gpu is specified","Body":"<p>I'm running into this problem on a 4 GPU Amazon instance, using a simple example script:<\/p>\n\n<pre><\/pre>\n\n<p>The result of  before launching the script is:<\/p>\n\n<pre><\/pre>\n\n<p>and while the script is running:<\/p>\n\n<pre><\/pre>\n\n<p>so memory is allocated to GPU0, even though no part in the code mentions it. Do you know where this behavior comes from? This causes an issue because we are multiple users on this instance, and GPU0 gets saturated even if nobody means to use it.<\/p>\n","answers":[{"AnswerId":"35506516","CreationDate":"2016-02-19T13:29:08.750","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>If you are interested in only using GPU1, I'd consider wrapping the script in something that sets <code>CUDA_VISIBLE_DEVICES<\/code> (see <a href=\"https:\/\/devblogs.nvidia.com\/parallelforall\/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices\/\" rel=\"nofollow\">https:\/\/devblogs.nvidia.com\/parallelforall\/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices\/<\/a>) to <code>1<\/code>. That way, only one GPU will be visible to the script (and it will look like its id is <code>0<\/code>). If you'd set it to <code>2,3<\/code> you would get those GPUs with ids <code>0,1<\/code> respectively.<\/p>\n"},{"AnswerId":"35507394","CreationDate":"2016-02-19T14:14:28.450","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>A workaround we have found is to modify <code>skflow.TensorFlowEstimator<\/code><\/p>\n\n<p>The culprit is <\/p>\n\n<pre><code>with self._graph.as_default():\n    tf.set_random_seed(self.tf_random_seed)\n    self._global_step = tf.Variable(\n        0, name=\"global_step\", trainable=False)\n<\/code><\/pre>\n\n<p>in <code>skflow.TensorFlowEstimator.setup_training()<\/code>, which we've modified as<\/p>\n\n<pre><code>with self._graph.as_default(), tf.device(\"\/gpu:{0}\".format(self.gpu_number)):\n    tf.set_random_seed(self.tf_random_seed)\n    self._global_step = tf.get_variable('global_step', [],\n                                      initializer=tf.constant_initializer(0), trainable=False)\n<\/code><\/pre>\n\n<p>adding an attribute <code>gpu_number<\/code> to the class, and initiliazing <code>session<\/code> with <code>allow_soft_placement=True<\/code> in <code>skflow.TensorFlowEstimator._setup_training()<\/code><\/p>\n"}]}
{"QuestionId":35504948,"AnswerCount":1,"Tags":"<matlab><caffe><matcaffe>","CreationDate":"2016-02-19T12:06:00.093","AcceptedAnswerId":35548823.0,"OwnerUserId":2191652.0,"Title":"Extract Caffe features with variable batch size in Matlab","Body":"<p>I know how to extract Caffe feats \/ scores using the  that is provided along with Caffe. However when using this file one has to provide a prototxt file that determines not only the network architecture but also the input dimensions including the .<\/p>\n\n<p>Since I'm processing the frames of videos of variable sequence lenght I need a way to use  along with a variable batch size.<\/p>\n\n<p>Does anyone know how to do that?<\/p>\n\n<p>It would probably involve changing this line from <\/p>\n\n<pre><\/pre>\n\n<p>to something that dynamically passes the current batch size needed dynamically to caffe<\/p>\n","answers":[{"AnswerId":"35548823","CreationDate":"2016-02-22T08:21:15.393","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>I ended up using the <code>reshape<\/code> function:<\/p>\n\n<pre><code>net = caffe.Net(net_model, net_weights, phase);\nnet.blobs('data').reshape([dim1 dim2 numChannels numFrames]);\nscores = net.forward(inputData);\ncaffe.reset_all();\n<\/code><\/pre>\n"}]}
{"QuestionId":35506904,"AnswerCount":1,"Tags":"<theano><recurrent-neural-network><language-model>","CreationDate":"2016-02-19T13:48:36.270","AcceptedAnswerId":35509785.0,"OwnerUserId":557478.0,"Title":"Wrong number of dimensions: expected 0, got 1 with shape (1,)","Body":"<p>I am doing word-level language modelling with a vanilla rnn, I am able to train the model but for some weird reasons I am not able to get any samples\/predictions from the model; here is the relevant part of the code:<\/p>\n\n<pre><\/pre>\n\n<p>I get the error: <strong>\"TypeError: ('Bad input argument to theano function with name \"train.py:94\"  at index 0(0-based)', 'Wrong number of dimensions: expected 0, got 1 with shape (1,).')\"<\/strong><\/p>\n\n<p>Now this corresponds to the following line (in the predict_model):<\/p>\n\n<pre><\/pre>\n\n<p>Even after spending hours I am not able to comprehend how could there be a dimension mis-match when:<\/p>\n\n<pre><\/pre>\n\n<p>And when I do train_set_x[index], I am getting <em>(4, 109)<\/em> which '<strong>x<\/strong>' Tensor of type fmatrix can hold (this is what happens in <strong>train_model<\/strong>) but when I do voc[index], I am getting <em>(1, 109)<\/em>, which is also a matrix but '<strong>x<\/strong>' cannot hold this, why ? !<\/p>\n\n<p>Any help will be much appreciated.<\/p>\n\n<p>Thanks !<\/p>\n","answers":[{"AnswerId":"35509785","CreationDate":"2016-02-19T16:12:35.677","ParentId":null,"OwnerUserId":"1985353","Title":null,"Body":"<p>The error message refers to the definition of the whole Theano function named <code>predict_model<\/code>, not the specific line where the substitution with <code>givens<\/code> occurs.<\/p>\n\n<p>The issue seems to be that <code>predict_model<\/code> gets called with an argument that is a <em>vector of length 1<\/em> instead of a <em>scalar<\/em>. The initial <code>seed<\/code> sampled from <code>randint<\/code> is actually a scalar, but I would guess that the output <code>p<\/code> of <code>predict_model(seed)<\/code> is a vector and not a scalar.<\/p>\n\n<p>In that case, you could either return <code>rnn.y[0]<\/code> in <code>predict_model<\/code>, or replace <code>seed = p<\/code> with <code>seed = p[0]<\/code> in the loop over <code>j<\/code>.<\/p>\n"}]}
{"QuestionId":35508356,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-19T15:03:09.263","AcceptedAnswerId":40981698.0,"OwnerUserId":5951780.0,"Title":"translate.py doesnt use embedding?","Body":"<p>I am trying to get my head around the lstm-with-attention script in tensorflow:\n<a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/translate\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/translate<\/a><\/p>\n\n<p>This does not use embeddings, even though it calls embedding_rnn_seq2seq, but fed into this call are only the scalar word-IDs, not any kind of embedding vectors.<\/p>\n\n<p>At least I cannot find the embedding parameters, nor the matrix, nor a possibility to feed pretrained embeddings. Any idea? Thx a bunch<\/p>\n","answers":[{"AnswerId":"40981698","CreationDate":"2016-12-05T19:18:01.933","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The code in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d9c173fc7adf9443d58310dd08a84b9eb89d4c8f\/tensorflow\/models\/rnn\/translate\/translate.py\" rel=\"nofollow noreferrer\"><code>translate.py<\/code><\/a> <em>does<\/em> use an embedding, but its creation is buried deep in the implementation. The model uses the (undocumented) <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4172ca2cd7ac34be67bda2600944d284e8907b95\/tensorflow\/python\/ops\/seq2seq.py#L754\" rel=\"nofollow noreferrer\"><code>tf.nn.seq2seq.embedding_attention_seq2seq()<\/code><\/a> function, which <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4172ca2cd7ac34be67bda2600944d284e8907b95\/tensorflow\/python\/ops\/seq2seq.py#L814\" rel=\"nofollow noreferrer\">creates<\/a> a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/rnn_cell.html#EmbeddingWrapper\" rel=\"nofollow noreferrer\"><code>tf.nn.rnn_cell.EmbeddingWrapper<\/code><\/a> that internally creates the embedding matrix.<\/p>\n"}]}
{"QuestionId":35508866,"AnswerCount":2,"Tags":"<python><c++><tensorflow>","CreationDate":"2016-02-19T15:26:39.400","AcceptedAnswerId":null,"OwnerUserId":1973820.0,"Title":"Tensorflow Different ways to Export and Run graph in C++","Body":"<p>For importing your trained network to the C++ you need to export your network to be able to do so. After searching a lot and finding almost no information about it, it was clarified that we should use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"noreferrer\">freeze_graph()<\/a> to be able to do it.<\/p>\n\n<p>Thanks to the new 0.7 version of Tensorflow, they added <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#export_meta_graph\" rel=\"noreferrer\">documentation<\/a> of it. <\/p>\n\n<p>After looking into documentations, I found that there are few similar methods, can you tell what is the difference between  and:\n  as it has similar parameters, but it seems it can also be used for importing models to C++ (I just guess the difference is that for using the file output by this method you can only use  or it's something else?)<\/p>\n\n<p>Also one question about how to use :\nIn documentations the  is given by  but in examples in  it is . What is the difference between these two? <\/p>\n\n<p>This question is related to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/615\" rel=\"noreferrer\">this issue.<\/a><\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"35663177","CreationDate":"2016-02-26T22:50:18.690","ParentId":null,"OwnerUserId":"5936741","Title":null,"Body":"<p>For predicting (and every other operations) you can do something like this:<\/p>\n\n<p>First of all in python you should <strong>name<\/strong> your variables or operation for the future use<\/p>\n\n<pre><code>self.init = tf.initialize_variables(tf.all_variables(), name=\"nInit\")\n<\/code><\/pre>\n\n<p>After training, calculations of so.. when you have your variables assigned go through all of them and save as constants to your graph. (almost the same can be done with that freeze tool, but i usually do it by myself, check \"name=nWeights\" in py and cpp below)<\/p>\n\n<pre><code>def save(self, filename):\n    for variable in tf.trainable_variables():\n        tensor = tf.constant(variable.eval())\n        tf.assign(variable, tensor, name=\"nWeights\")\n\n    tf.train.write_graph(self.sess.graph_def, 'graph\/', 'my_graph.pb', as_text=False)\n<\/code><\/pre>\n\n<p>Now go c++ and load our graph and load variables from saved constants:<\/p>\n\n<pre><code>void load(std::string my_model) {\n        auto load_graph_status =\n                ReadBinaryProto(tensorflow::Env::Default(), my_model, &amp;graph_def);\n\n        auto session_status = session-&gt;Create(graph_def);\n\n        std::vector&lt;tensorflow::Tensor&gt; out;\n        std::vector&lt;string&gt; vNames;\n\n        int node_count = graph_def.node_size();\n        for (int i = 0; i &lt; node_count; i++) {\n            auto n = graph_def.node(i);\n\n            if (n.name().find(\"nWeights\") != std::string::npos) {\n                vNames.push_back(n.name());\n            }\n        }\n\n        session-&gt;Run({}, vNames, {}, &amp;out);\n<\/code><\/pre>\n\n<p>Now you have all of your neural net weights or other variables loaded. <\/p>\n\n<p>Similarly, you can perform other operations (remember about names?); make input and output tensors of proper size, fill input tensor with data and run session like so:<\/p>\n\n<pre><code>auto operationStatus = session-&gt;Run(input, {\"put_your_operation_here\"}, {}, &amp;out);\n<\/code><\/pre>\n"},{"AnswerId":"43639305","CreationDate":"2017-04-26T16:03:29.657","ParentId":null,"OwnerUserId":"1076564","Title":null,"Body":"<p>Here's my solution utilizing the V2 checkpoints introduced in TF 0.12.<\/p>\n\n<p>There's no need to convert all variables to constants or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/615#issuecomment-184784850\" rel=\"noreferrer\">freeze the graph<\/a>.<\/p>\n\n<p>Just for clarity, a V2 checkpoint looks like this in my directory <code>models<\/code>:<\/p>\n\n<pre><code>checkpoint  # some information on the name of the files in the checkpoint\nmy-model.data-00000-of-00001  # the saved weights\nmy-model.index  # probably definition of data layout in the previous file\nmy-model.meta  # protobuf of the graph (nodes and topology info)\n<\/code><\/pre>\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 <a href=\"https:\/\/stackoverflow.com\/a\/43526252\/1076564\">different problems<\/a>.<\/p>\n\n<p><strong>Python part (inference)<\/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\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"}]}
{"QuestionId":35508894,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-19T15:28:02.550","AcceptedAnswerId":35525974.0,"OwnerUserId":2615676.0,"Title":"sparse_to_dense requires indices to be lexicographically sorted in 0.7","Body":"<p>There was a change in behavior between versions 0.7 and 0.6 in  that now requires the indices to be lexicographically sorted. This is a somewhat onerous requirement for some of my uses because the way I generate the indices is matched to the way I generate the elements (third arg of ) and sorting both is a bit of a hassle. I see that the  option can be set to  to ignore this. Is it safe to do so? Why the requirement in the first place?<\/p>\n","answers":[{"AnswerId":"35525974","CreationDate":"2016-02-20T16:36:45.693","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Looking at the current implementation for <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/875a67f3d298ab7f7438a3004e4597c71d92b529\/tensorflow\/core\/util\/sparse\/sparse_tensor.h#L310\" rel=\"noreferrer\"><code>tensorflow::sparse::SparseTensor::ToDense&lt;T&gt;()<\/code><\/a>, it seems that the conversion does not depend on the order of indices. Therefore, you can run <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/sparse_ops.html#sparse_to_dense\" rel=\"noreferrer\"><code>tf.sparse_to_dense(sparse_indices, ..., validate_indices=False)<\/code><\/a> on unsorted <code>sparse_indices<\/code>. A consistent result will be produced as long as there are no duplicate indices (otherwise it looks like the current implementation is last writer wins, but that is not guaranteed in the API).<\/p>\n\n<p>It looks like the requirement for lexicographically sorted indices arises from a desire to check that there are no duplicates in <code>sparse_indices<\/code>. You can check the order property <em>and<\/em> absence of duplicates with a simple scan over the indices tensor (only considering rows <code>i<\/code> and <code>i - 1<\/code>), whereas checking for duplicates in an order independent fashion is more expensive. Since there's limited upside to the more conservative check, and a potentially large performance downside, we chose not to implement it, but the documentation could be clearer on this point!<\/p>\n"}]}
{"QuestionId":35512046,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-02-19T18:12:34.060","AcceptedAnswerId":35513102.0,"OwnerUserId":2272798.0,"Title":"Requesting multiple values from graph at same time","Body":"<p>In the code below l2 surprisingly returns the same value as l1, but since the optimizer is being requested in the list before l2, I expected the loss to be the new loss after training. Can I not request multiple values at the same time from the graph and expect consistent output?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"35513654","CreationDate":"2016-02-19T19:43:28.600","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As Dave <a href=\"https:\/\/stackoverflow.com\/a\/35513102\/3574081\">points out<\/a>, the order of arguments to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow noreferrer\"><code>Session.run()<\/code><\/a> has no effect on the order of evaluation, and the <code>loss<\/code> tensor in your example does not have a dependency on the <code>optimizer<\/code> op. To add a dependency, you could use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#control_dependencies\" rel=\"nofollow noreferrer\"><code>tf.control_dependencies()<\/code><\/a> to add an explicit dependency on the optimizer running before fetching the loss:<\/p>\n\n<pre><code>with tf.control_dependencies([optimizer]):\n    loss_after_optimizer = tf.identity(loss)\n\n_, l2 = sess.run([optimizer, loss_after_optimizer], feed_dict={x: X, y: Y})\n<\/code><\/pre>\n"},{"AnswerId":"44896587","CreationDate":"2017-07-04T03:33:05.623","ParentId":null,"OwnerUserId":"8251322","Title":null,"Body":"<p>I've tested logistic regression implemented in tensorflow with three ways of session.run:<\/p>\n\n<ol>\n<li><p>all together<\/p>\n\n<blockquote>\n  <p>res1, res2, res3 = sess.run([op1, op2, op3])<\/p>\n<\/blockquote><\/li>\n<li><p>separately<\/p>\n\n<blockquote>\n  <p>res1 = sess.run(op1)<\/p>\n  \n  <p>res2 = sess.run(op2)<\/p>\n  \n  <p>res3 = sess.run(op3)<\/p>\n<\/blockquote><\/li>\n<li><p>with dependencies<\/p>\n\n<blockquote>\n  <p>with tf.control_dependencies([op1]):<\/p>\n  \n  <p>op2_after = tf.identity(op1)<\/p>\n  \n  <p>op3_after = tf.identity(op1)<\/p>\n  \n  <p>res1,res2,res3 = session.run([op1, op2_after, op3_after])<\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>set batch size as 10000, the result is: <\/p>\n\n<pre><code>1: 0.05+ secs &lt; 2: 0.11+ secs &lt; 3: 0.25+ secs\n<\/code><\/pre>\n\n<p>The main difference between 1 and 3 is only one mini-batch. It may not worth it to use 3 instead of 1.<\/p>\n\n<p>Here is the test code (it is an LR example written by someone else...).<\/p>\n\n<p>Here is the <a href=\"https:\/\/github.com\/StryxZilla\/noob\/blob\/master\/notMNIST.pickle\" rel=\"nofollow noreferrer\">data<\/a><\/p>\n\n<pre><code>#!\/usr\/bin\/env python2\n# -*- coding: utf-8 -*-\n\"\"\"\nCreated on Fri Jun  2 13:38:14 2017\n\n@author: inse7en\n\"\"\"\n\nfrom __future__ import print_function\nimport numpy as np\nimport tensorflow as tf\nfrom six.moves import cPickle as pickle\nimport time\n\npickle_file = '\/Users\/inse7en\/Downloads\/notMNIST.pickle'\nwith open(pickle_file, 'rb') as f:\n  save = pickle.load(f)\n  train_dataset = save['train_dataset']\n  train_labels = save['train_labels']\n  valid_dataset = save['valid_dataset']\n  valid_labels = save['valid_labels']\n  test_dataset = save['test_dataset']\n  test_labels = save['test_labels']\n  del save  # hint to help gc free up memory\n  print('Training set', train_dataset.shape, train_labels.shape)\n  print('Validation set', valid_dataset.shape, valid_labels.shape)\n  print('Test set', test_dataset.shape, test_labels.shape)\n\n\nimage_size = 28\nnum_labels = 10\n\ndef reformat(dataset, labels):\n  dataset = dataset.reshape((-1, image_size * image_size)).astype(np.float32)\n  # Map 2 to [0.0, 1.0, 0.0 ...], 3 to [0.0, 0.0, 1.0 ...]\n  labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)\n  return dataset, labels\ntrain_dataset, train_labels = reformat(train_dataset, train_labels)\nvalid_dataset, valid_labels = reformat(valid_dataset, valid_labels)\ntest_dataset, test_labels = reformat(test_dataset, test_labels)\nprint('Training set', train_dataset.shape, train_labels.shape)\nprint('Validation set', valid_dataset.shape, valid_labels.shape)\nprint('Test set', test_dataset.shape, test_labels.shape)\n\n# This is to expedite the process\ntrain_subset = 10000\n# This is a good beta value to start with\nbeta = 0.01\n\ngraph = tf.Graph()\nwith graph.as_default():\n    # Input data.\n    # They're all constants.\n    tf_train_dataset = tf.constant(train_dataset[:train_subset, :])\n    tf_train_labels = tf.constant(train_labels[:train_subset])\n    tf_valid_dataset = tf.constant(valid_dataset)\n    tf_test_dataset = tf.constant(test_dataset)\n\n    # Variables\n    # They are variables we want to update and optimize.\n    weights = tf.Variable(tf.truncated_normal([image_size * image_size, num_labels]))\n    biases = tf.Variable(tf.zeros([num_labels]))\n\n    # Training computation.\n    logits = tf.matmul(tf_train_dataset, weights) + biases\n    # Original loss function\n    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))\n    # Loss function using L2 Regularization\n    regularizer = tf.nn.l2_loss(weights)\n    loss = tf.reduce_mean(loss + beta * regularizer)\n\n    # Optimizer.\n    optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)\n\n    # Predictions for the training, validation, and test data.\n    train_prediction = tf.nn.softmax(logits)\n    valid_prediction = tf.nn.softmax(tf.matmul(tf_valid_dataset, weights) + biases)\n    test_prediction = tf.nn.softmax(tf.matmul(tf_test_dataset, weights) + biases)\n\n    num_steps = 50\n\n\n    def accuracy(predictions, labels):\n        return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))\n                \/ predictions.shape[0])\n\n\n    with tf.Session(graph=graph) as session:\n        # This is a one-time operation which ensures the parameters get initialized as\n        # we described in the graph: random weights for the matrix, zeros for the\n        # biases.\n        tf.initialize_all_variables().run()\n        print('Initialized')\n        for step in range(num_steps):\n            # Run the computations. We tell .run() that we want to run the optimizer,\n            # and get the loss value and the training predictions returned as numpy\n            # arrays.\n            #_, l, predictions = session.run([optimizer, loss, train_prediction])\n\n            start_time = time.time()\n            with tf.control_dependencies([optimizer]):\n                loss_after_optimizer = tf.identity(loss)\n                predictions_after = tf.identity(train_prediction)\n                regularizers_after = tf.identity(regularizer)\n\n\n            _, l, predictions,regularizers = session.run([optimizer, loss_after_optimizer, predictions_after, regularizers_after])\n\n            print(\"--- with dependencies: %s seconds ---\" % (time.time() - start_time))\n            #start_time = time.time()\n            #opt = session.run(optimizer)\n            #l = session.run(loss)\n            #predictions = session.run(train_prediction)\n            #regularizers = session.run(regularizer)\n\n            #print(\"--- run separately: %s seconds ---\" % (time.time() - start_time))\n\n            #start_time = time.time()\n            #_, l, predictions,regularizers = session.run([optimizer, loss, train_prediction, regularizer])\n\n            #print(\"--- all together: %s seconds ---\" % (time.time() - start_time))\n\n            #if (step % 100 == 0):\n                #print('Loss at step {}: {}'.format(step, l))\n                #print('Training accuracy: {:.1f}'.format(accuracy(predictions,\n                                                                  #train_labels[:train_subset, :])))\n                # Calling .eval() on valid_prediction is basically like calling run(), but\n                # just to get that one numpy array. Note that it recomputes all its graph\n                # dependencies.\n\n                # You don't have to do .eval above because we already ran the session for the\n                # train_prediction\n                #print('Validation accuracy: {:.1f}'.format(accuracy(valid_prediction.eval(),\n                                                                    #valid_labels)))\n        #print('Test accuracy: {:.1f}'.format(accuracy(test_prediction.eval(), test_labels)))\n        #print(regularizer)\n<\/code><\/pre>\n"},{"AnswerId":"35513102","CreationDate":"2016-02-19T19:10:19.570","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>No - the order in which you request them in the list has no effect on the evaluation order.  For side-effect-having operations such as the optimizer, if you want to guarantee a specific ordering, you need to enforce it using <code>with_dependencies<\/code> or similar control-flow constructs.  In general, ignoring side-effects, TensorFlow will return results to you by grabbing the node from the graph as soon as it's computed - and, obviously, the loss is computed <em>before<\/em> the optimizer, since the optimizer requires the loss as one of its input.  (Remember that 'loss' is not a variable;  it's a tensor; so it's not actually affected by the optimizer step.)<\/p>\n\n<pre><code>sess.run([loss, optimizer], ...)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>sess.run([optimizer, loss], ...)\n<\/code><\/pre>\n\n<p>are equivalent.<\/p>\n"}]}
{"QuestionId":35513370,"AnswerCount":2,"Tags":"<python><tensorflow><python-unicode>","CreationDate":"2016-02-19T19:25:01.137","AcceptedAnswerId":35518224.0,"OwnerUserId":1739403.0,"Title":"'utf-8' decode error in tensorflow tutorial","Body":"<p>I'm running into this bizarre problem where when I run <\/p>\n\n<pre><\/pre>\n\n<p>I get:<\/p>\n\n<pre><\/pre>\n\n<p>However, if I just run the code in input_data.py directly, everything appears to be fine:<\/p>\n\n<pre><\/pre>\n\n<p>Anyone has any idea what's going on?<\/p>\n\n<p>My system: Ubuntu 15.10 x64 python 3.5.0.<\/p>\n","answers":[{"AnswerId":"41363103","CreationDate":"2016-12-28T13:27:40.967","ParentId":null,"OwnerUserId":"394594","Title":null,"Body":"<p>In my case, the problem was in the encoding of the data file.<\/p>\n\n<p>Open the file using <code>vim<\/code> and execute:<\/p>\n\n<pre><code>:set fileencoding=utf-8\n<\/code><\/pre>\n\n<p>That solved the issue in my case.<\/p>\n"},{"AnswerId":"35518224","CreationDate":"2016-02-20T01:56:27.497","ParentId":null,"OwnerUserId":"1739403","Title":null,"Body":"<p>The bug has been addressed by a recent change <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/555e73da8f171992085c68614f74b23b8180292c\" rel=\"noreferrer\">555e73d<\/a>. MNIST files need to be opened with binary 'rb' mode instead of just text 'r'.<\/p>\n"}]}
{"QuestionId":35513574,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><skflow>","CreationDate":"2016-02-19T19:38:34.497","AcceptedAnswerId":null,"OwnerUserId":786637.0,"Title":"Multiple Categorical Input Variables in Tensorflow","Body":"<p>I have a data-set in which each feature vector has 50 features, 45 of which are categorical. I am having trouble sending the categorical variables into tensorflow. I have found an example <a href=\"https:\/\/medium.com\/@ilblackdragon\/tensorflow-tutorial-part-3-c5fc0662bc08\" rel=\"nofollow\">tutorial for tensorflow with categorical variables<\/a>, but do not understand how to adapt this to work with a set which has both types of data, and multiple features. My first attempt is below, but this does not encode the majority of variables.<\/p>\n\n<pre><\/pre>\n\n<p>I think the real problem, is I don't really understand how to handle the input 'tensor' X to the my_model function in the above code.<\/p>\n","answers":[{"AnswerId":"35944998","CreationDate":"2016-03-11T16:15:49.180","ParentId":null,"OwnerUserId":"3865298","Title":null,"Body":"<p>Use a categorical processor to map your categories into integers before inputting, like so<\/p>\n\n<pre><code>cat_processor = skflow.preprocessing.CategoricalProcessor()\nX_train = np.array(list(cat_processor.fit_transform(X_train)))\nX_test = np.array(list(cat_processor.transform(X_test)))\nn_classes = len(cat_processor.vocabularies_[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":35514898,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-02-19T20:57:56.443","AcceptedAnswerId":null,"OwnerUserId":5952845.0,"Title":"When should I define a new TensorFlow op?","Body":"<p>For my application, I was able to create a new function using only predefined ops. Is there any need to define a new op in this case?<\/p>\n\n<p>The pseudocode for my function is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35515600","CreationDate":"2016-02-19T21:42:11.523","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you can achieve what you set out to do with a composition of existing ops, then that's great! You <strong>don't<\/strong> need to create a new op.<\/p>\n\n<p>There are circumstances when we've found it necessary to create a new op, however:<\/p>\n\n<ul>\n<li><p>Sometimes you can gain performance by <em>fusing<\/em> ops together into a single op. For example many of the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/training_ops.cc\" rel=\"nofollow\">\"training\" ops<\/a> have fused implementations, even though they were initially implemented using simple ops.<\/p><\/li>\n<li><p>Another example is when you want to define a gradient for a composition of ops (because it's more efficient or stable to consider the expression as a whole). This is the rationale for ops like <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"nofollow\"><code>tf.nn.softmax_cross_entropy_with_logits()<\/code><\/a>.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":35518072,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-20T01:32:29.417","AcceptedAnswerId":35536511.0,"OwnerUserId":2125302.0,"Title":"Feeding tensors for training vs validation data","Body":"<p>In the tensorflow examples, feed_dict is used to send either training or validation input into the same model graph. Unfortunately, you cannot feed tensors:<\/p>\n\n<pre><\/pre>\n\n<p>I've been using an input pipeline and , so my data never really enters python. Having to call  to get the data into python just to feed it back to tensorflow seems silly and is definitely slow.<\/p>\n\n<p>Does anyone have a good solution for this?<\/p>\n\n<p>Currently, I just create two identical copies of the model subgraph that use the same parameters. This works, but forces me to organize my code in an odd way.<\/p>\n\n<p>EDIT<\/p>\n\n<p>For example, I'm currently doing something like:<\/p>\n\n<pre><\/pre>\n\n<p>so that the test model uses the parameters learned by training. The nice thing about  is that I only need to define the model once and I do not have to separate the model's parameters from its structure.<\/p>\n","answers":[{"AnswerId":"35536511","CreationDate":"2016-02-21T13:06:12.327","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<h2>Warning:<\/h2>\n\n<p>This solution can cause significant problems when input queues are involved. See: <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!msg\/discuss\/mLrt5qc9_uU\/sGNbC7GpAwAJ\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!msg\/discuss\/mLrt5qc9_uU\/sGNbC7GpAwAJ<\/a><\/p>\n\n<p>Thanks to @fwalch for pointing this out in the comments<\/p>\n\n<hr>\n\n<p>There's no way to do exactly what you're asking, see the answer to my question <a href=\"https:\/\/stackoverflow.com\/questions\/33748552\/tensorflow-how-to-replace-a-node-in-a-calculation-graph?lq=1\">here<\/a>.<\/p>\n\n<p>But the newly public \"<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/control_flow_ops.html#cond\" rel=\"nofollow noreferrer\">cond<\/a>\" from version 0.7 can fill your use case:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Here are the two data streams.\ntrain_data = tf.Variable(999)\ntest_data = tf.Variable(1000)\n\n# This selects which stream to use.\nselect_test = tf.placeholder(dtype=bool,shape=[],name='select_test')\ndata = tf.cond(\n    select_test,\n    lambda:test_data,\n    lambda:train_data\n)\n\n# Here is the model.\nmodel = data-500;\n\ninit = tf.initialize_all_variables()\nwith tf.Session():\n    init.run()\n\n    # You just have to feed `select_test` when you evaluate the model.\n    print(model.eval({select_test:False})) #499\n    print(model.eval({select_test:True})) #500\n<\/code><\/pre>\n\n<p>You can use the same sort of trick for <a href=\"https:\/\/stackoverflow.com\/a\/34634291\/997378\">switching Batch Normalization to use a moving average during test<\/a>.<\/p>\n"},{"AnswerId":"35529705","CreationDate":"2016-02-20T22:10:48.707","ParentId":null,"OwnerUserId":"1589147","Title":null,"Body":"<p><strong>TL;DR<\/strong><\/p>\n\n<p>You can use your input pipeline with the train, test and validation step instead of using a placeholder. When you already have a Tensor then I don't see a reason not to use it.<\/p>\n\n<p><strong>Relevant Documentation<\/strong><\/p>\n\n<p>Further information can be found on TF's site.<\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#feeding\" rel=\"nofollow\">Feeding Data from Python<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#placeholder\" rel=\"nofollow\">Placeholders<\/a><\/li>\n<\/ol>\n\n<p><strong>Example<\/strong><\/p>\n\n<p>In this example, we'll send the model and expected output to train and test functions. The only difference is that for some we use a placeholder and others we're using Tensors.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\n\ndef train(model, T):\n    \"\"\"Example train operation which returns sum of parameters.\"\"\"\n    return model + T\n\n\ndef test(model, T):\n    \"\"\"Example test operation which returns parameters multiplied.\"\"\"\n    return model * T\n\n\n# Placeholders which will be required in the feed_dict once we execute a TF_Run.\nx_placeholder_model = tf.placeholder(tf.int32)\nt_placeholder = tf.placeholder(tf.int32)\n\n# Tensors, using constants for brevity but these could be from an input pipeline\n#  or some other set of operations.\nx_tensor_model = tf.constant([1, 2, 3])\nt_tensor = tf.constant([1, 2, 3])\n\nusing_placeholder_train = train(x_placeholder_model, t_placeholder)\nusing_tensor_train = train(x_tensor_model, t_tensor)\n\nusing_placeholder_test = test(x_placeholder_model, t_placeholder)\nusing_tensor_test = test(x_tensor_model, t_tensor)\n\nwith tf.Session() as sess:\n    print(sess.run(\n        using_placeholder_train, feed_dict={\n            x_placeholder_model: [1, 2, 3],\n            t_placeholder: [1, 2, 3]}))\n\n    print(sess.run(\n        using_placeholder_test, feed_dict={\n            x_placeholder_model: [1, 2, 3],\n            t_placeholder: [1, 2, 3]}))\n\n    print(sess.run(using_tensor_train))\n    print(sess.run(using_tensor_test))\n<\/code><\/pre>\n\n<p>Executing this code will result in the following output:<\/p>\n\n<pre><code>[2 4 6]\n[1 4 9]\n[2 4 6]\n[1 4 9]\n<\/code><\/pre>\n\n<p>Illustrating that calls to:<\/p>\n\n<pre><code>print(sess.run(\n    using_placeholder_train, feed_dict={\n        x_placeholder_model: [1, 2, 3],\n        t_placeholder: [1, 2, 3]}))\n<\/code><\/pre>\n\n<p>Return the same output as:<\/p>\n\n<pre><code>print(sess.run(using_tensor_train))\n<\/code><\/pre>\n\n<p>Without any additions to the <code>feed_dict<\/code> because we already have the tensors available on the graph.<\/p>\n"}]}
{"QuestionId":35518155,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-02-20T01:43:23.750","AcceptedAnswerId":null,"OwnerUserId":5870277.0,"Title":"Why does scan upcast?","Body":"<p>This code to calculate the trace of a matrix (based on an example in the Theano \"loop\" tutorial) works fine:<\/p>\n\n<pre><\/pre>\n\n<p>But if I remove the cast operation from the lambda function like this:<\/p>\n\n<pre><\/pre>\n\n<p>The following error message is produced: <\/p>\n\n<p>ValueError: When compiling the inner function of scan the following error has been encountered: The initial state ( in scan nomenclature) of variable IncSubtensor{Set;:int64:}.0 (argument number 2) has dtype float32, while the result of the inner function () has dtype float64. This can happen if the inner function of scan results in an upcast or downcast.<\/p>\n\n<p>Why so?  and  are explicitly set to float32. How does the result of adding them get to be float64?<\/p>\n","answers":[]}
{"QuestionId":35518413,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-02-20T02:28:47.263","AcceptedAnswerId":null,"OwnerUserId":2138200.0,"Title":"Python crashes running assemble_data.py for Flickr Style dataset","Body":"<p>I'm trying to download the Flickr Style dataset using  provided in the examples folder. However, whenever I run this python crashes with error 'python quit unexpectedly'. <\/p>\n\n<p>It seems to be related to multiprocessing and urllib. When I replace  with a single threaded loop it works but is very slow. Also, if I run with multiprocessing but remove  it seems to work too.<\/p>\n","answers":[{"AnswerId":"35526214","CreationDate":"2016-02-20T16:57:00.620","ParentId":null,"OwnerUserId":"2138200","Title":null,"Body":"<p>Answering my own question here... I resolved this by using urllib3 instead.<\/p>\n\n<pre><code>http = urllib3.PoolManager(10)\ndef download_image(args_tuple):\n    \"For use with multiprocessing map. Returns filename on fail.\"\n    url, filename = args_tuple\n    try:\n        if not os.path.exists(filename):\n            print url + ' -&gt; ' + filename\n            # Dont redirect.\n            response = http.request('GET', url, redirect=False)\n            with open(filename, 'wb') as f:\n                f.write(response.data)\n        with open(filename) as f:\n            assert hashlib.sha1(f.read()).hexdigest() != MISSING_IMAGE_SHA1\n        test_read_image = io.imread(filename)\n        return True\n    except KeyboardInterrupt:\n        raise Exception()  # multiprocessing doesn't catch keyboard exceptions\n    except:\n        os.remove(filename)\n        return False\n<\/code><\/pre>\n\n<p>Gist <a href=\"https:\/\/gist.github.com\/sthomp\/48d35095a414580c055b\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":35524246,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-20T13:56:10.613","AcceptedAnswerId":null,"OwnerUserId":2954187.0,"Title":"Train a tensorflow model minimizing the loss of several batches","Body":"<p>I would like to train the weights of a model based on the sum of the loss value of several batches. However it seems that once you run the graph for each of the individual batches, the object that is returned is just a regular numpy array. So when you try and use an optimizer like GradientDescentOptimizer, it no longer has information about the variables that were used to calculate the sum of the losses, so it can't find the gradients of the weights that what help minimize the loss. Here's an example tensorflow script to illustrate what I'm talking about:<\/p>\n\n<pre><\/pre>\n\n<p>The  is a numpy array and thus cannot be used in the optimizer. Does anyone know a way around the problem, where I want to use information across many batches but still need the graph intact in order to preserve the fact that the  is a function of the weights?<\/p>\n","answers":[{"AnswerId":"35526431","CreationDate":"2016-02-20T17:15:04.447","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>The thing you optimize in any of the trainers must be a part of the graph, here what you train on is the actual realized result, so it won't work.<\/p>\n\n<p>I think the way you should probably do this is to construct your input as a batch of batches e.g.<\/p>\n\n<pre><code>intput = tf.placeholder(\"float\", (number_of_batches, batch_size, input_size)\n<\/code><\/pre>\n\n<p>Then have your target also be a 3d tensor which can be trained on.<\/p>\n"}]}
{"QuestionId":35525694,"AnswerCount":2,"Tags":"<csv><pandas><hdf5><caffe><np>","CreationDate":"2016-02-20T16:13:31.023","AcceptedAnswerId":null,"OwnerUserId":5030342.0,"Title":"Memory error when using read_csv","Body":"<p>I'd like to convert the csv files into hdf5 format,which are used for caffe training.Because the csv files is 80G,it will report memory error.The machine memory is 128G.So can it possbile to improve my code?handle it one by one?Below is my code,it reported memory error when run in np.array<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43386154","CreationDate":"2017-04-13T07:32:39.643","ParentId":null,"OwnerUserId":"3051715","Title":null,"Body":"<p>Actually, since you treat chunks of size 100000 separately, there is no need to load the whole CSV at one. The <code>chunksize<\/code> option in the <code>read_csv<\/code> is exactly for this case.<\/p>\n\n<p>When specifying <code>chunksize<\/code>, <code>read_csv<\/code> will become an iterator, returning <code>DataFrames<\/code>'s of size <code>chunksize<\/code>. You can iterate over instead of slicing the arrays each time.<\/p>\n\n<p>Minus all the lines setting the different variables, your code should look more like this:<\/p>\n\n<pre><code>chuncks = pd.read_csv(file, header=None, chunksize=100000)\n\nfor chunk_number, data in enumerate(chunks):\n    y = np.array(data.iloc[:,0], np.float32)\n    x = np.array(data.iloc[:,1:], np.float32)\n\n    file_w = open(filename, 'w')\n    with h5py.File(modelname + str(idx) + '.h5', 'w') as f:\n        f['data'] = x\n        f['label'] = y\n        file_w.write(modelname + str(chunk_number) + '.h5\\n')\n    file_w.close()\n<\/code><\/pre>\n"},{"AnswerId":"35527596","CreationDate":"2016-02-20T18:54:19.433","ParentId":null,"OwnerUserId":"3964931","Title":null,"Body":"<p>The best approach would be to read n lines and then write these to the HDF5 file, extending it be n elements each time.  This way the amount of memory needed is not dependent on the size of the csv file.  You could read a line at a time as well, but that would be slightly less efficient.<\/p>\n\n<p>Here's code that applies this process for reading weather station data:\n<a href=\"https:\/\/github.com\/HDFGroup\/datacontainer\/blob\/master\/util\/ghcn\/convert_ghcn.py\" rel=\"nofollow\">https:\/\/github.com\/HDFGroup\/datacontainer\/blob\/master\/util\/ghcn\/convert_ghcn.py<\/a>.<\/p>\n"}]}
{"QuestionId":35527360,"AnswerCount":1,"Tags":"<lua><deep-learning><torch><softmax><resnet>","CreationDate":"2016-02-20T18:34:30.647","AcceptedAnswerId":35527507.0,"OwnerUserId":28802.0,"Title":"Adding Softmax to ResNet model in Torch","Body":"<p><strong>Context<\/strong>: I'm trying to modify <a href=\"https:\/\/github.com\/facebook\/fb.resnet.torch\/blob\/master\/pretrained\/extract-features.lua\" rel=\"nofollow\">this Facebook's ResNet feature extractor script<\/a> to classify an image and print the ImageNet class label.\nLet's say I have the model in torch:<\/p>\n\n<pre><\/pre>\n\n<p>That gives me the scores for each class. I want to obtain the top 5 classes and their probabilities. I think that to transform scores to probabilities I should use a SoftMax layer first.<\/p>\n\n<p>So I do: <\/p>\n\n<pre><\/pre>\n\n<p>But when I run it I get:<\/p>\n\n<blockquote>\n  <p>\/SpatialSoftMax.lua:38: bad argument #1 to 'resizeAs'\n  (torch.DoubleTensor expected, got torch.CudaTensor)<\/p>\n<\/blockquote>\n\n<p>The model looks good to me: (showing last layers only)<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on what might be wrong?<\/p>\n","answers":[{"AnswerId":"35527507","CreationDate":"2016-02-20T18:46:56.147","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Layers have types associated with them. By default you are given a double type<\/p>\n\n<pre><code>local softMaxLayer = cudnn.LogSoftMax():cuda()\nmodel:add(softMaxLayer)\n<\/code><\/pre>\n"}]}
{"QuestionId":35528905,"AnswerCount":1,"Tags":"<label><torch><data-science><resnet><imagenet>","CreationDate":"2016-02-20T20:56:13.223","AcceptedAnswerId":null,"OwnerUserId":28802.0,"Title":"Get ImageNet label for a specific index in the 1000-dimensional output tensor in torch","Body":"<p>I have the output Tensor of a forward pass for <a href=\"https:\/\/github.com\/facebook\/fb.resnet.torch\" rel=\"nofollow\">a Facebook implementation of the ResNet model<\/a> with a cat image. That is a 1000-dimensional Tensor with the classification probabilities. Using <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/03c04c6\/doc\/maths.md#torchtopkresval-resind-x-k-dim-dir-sort\" rel=\"nofollow\">torch.topk<\/a> I can obtain the top-5 probabilities and their indexes in the output tensor. Now I want to see the human-readable labels for those most-probable indexes.<\/p>\n\n<p>I searched online for the list of labels (which apparently are also called sysnets) and only found this: \n<a href=\"http:\/\/image-net.org\/challenges\/LSVRC\/2015\/browse-synsets\" rel=\"nofollow\">http:\/\/image-net.org\/challenges\/LSVRC\/2015\/browse-synsets<\/a><\/p>\n\n<p>I put those labels in a file using line numbers as the label index and when I run the network with two different cat images, I get \"screwdriver\" as the top guess for both. If I sort the label file alphabetically, I get \"cinema\" for both.<\/p>\n\n<p>This appears to be a problem with converting index to label, right?\nSo...the question is:\n<strong>How can I properly map index in network output tensor to Imagenet label?<\/strong> <\/p>\n","answers":[{"AnswerId":"35529631","CreationDate":"2016-02-20T22:03:39.960","ParentId":null,"OwnerUserId":"28802","Title":null,"Body":"<p>Found <a href=\"https:\/\/dato.com\/learn\/gallery\/notebooks\/build_imagenet_deeplearning.html\" rel=\"nofollow\">this tutorial on training ConvNets on ImageNet by Dato<\/a> and at the end it contains the correct mapping. Reporting it here for the record:<\/p>\n\n<pre><code>{\n 0: 'tench, Tinca tinca',\n 1: 'goldfish, Carassius auratus',\n 2: 'great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias',\n 3: 'tiger shark, Galeocerdo cuvieri',\n 4: 'hammerhead, hammerhead shark',\n 5: 'electric ray, crampfish, numbfish, torpedo',\n 6: 'stingray',\n 7: 'cock',\n 8: 'hen',\n 9: 'ostrich, Struthio camelus',\n 10: 'brambling, Fringilla montifringilla',\n ... [truncated for space]\n 990: 'buckeye, horse chestnut, conker',\n 991: 'coral fungus',\n 992: 'agaric',\n 993: 'gyromitra',\n 994: 'stinkhorn, carrion fungus',\n 995: 'earthstar',\n 996: 'hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa',\n 997: 'bolete',\n 998: 'ear, spike, capitulum',\n 999: 'toilet tissue, toilet paper, bathroom tissue'\n}\n<\/code><\/pre>\n\n<p>Full mapping here: <a href=\"https:\/\/gist.github.com\/maraoz\/388eddec39d60c6d52d4\" rel=\"nofollow\">https:\/\/gist.github.com\/maraoz\/388eddec39d60c6d52d4<\/a><\/p>\n"}]}
{"QuestionId":35529078,"AnswerCount":1,"Tags":"<neural-network><caffe>","CreationDate":"2016-02-20T21:11:23.800","AcceptedAnswerId":35572495.0,"OwnerUserId":4229825.0,"Title":"How to predict in pycaffe?","Body":"<p>I have a <a href=\"https:\/\/gist.github.com\/yutkin\/bf36e098a4eef7d30c54\" rel=\"nofollow\">model<\/a> that has been trained on CIFAR-10, but I don't realise how can I make a prediction in pycaffe.<\/p>\n\n<p>I got an image from lmdb but I don't know how to load it in a net and get a predicted class.<\/p>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35572495","CreationDate":"2016-02-23T08:34:07.157","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Use this python script<\/p>\n\n<pre><code># Run the script with anaconda-python\n# $ \/home\/&lt;path to anaconda directory&gt;\/anaconda\/bin\/python LmdbClassification.py\nimport sys\nimport numpy as np\nimport lmdb\nimport caffe\nfrom collections import defaultdict\ncaffe.set_mode_gpu()\n\n# Modify the paths given below\ndeploy_prototxt_file_path = '\/home\/&lt;username&gt;\/caffe\/examples\/cifar10\/cifar10_deploy.prototxt' # Network definition file\ncaffe_model_file_path = '\/home\/&lt;username&gt;\/caffe\/examples\/cifar10\/cifar10_iter_5000.caffemodel' # Trained Caffe model file\ntest_lmdb_path = '\/home\/&lt;username&gt;\/caffe\/examples\/cifar10\/cifar10_test_lmdb\/' # Test LMDB database path\nmean_file_binaryproto = '\/home\/&lt;username&gt;\/caffe\/examples\/cifar10\/mean.binaryproto' # Mean image file\n\n# Extract mean from the mean image file\nmean_blobproto_new = caffe.proto.caffe_pb2.BlobProto()\nf = open(mean_file_binaryproto, 'rb')\nmean_blobproto_new.ParseFromString(f.read())\nmean_image = caffe.io.blobproto_to_array(mean_blobproto_new)\nf.close()\n\n# CNN reconstruction and loading the trained weights\nnet = caffe.Net(deploy_prototxt_file_path, caffe_model_file_path, caffe.TEST)\n\ncount = 0\ncorrect = 0\nmatrix = defaultdict(int) # (real,pred) -&gt; int\nlabels_set = set()\n\nlmdb_env = lmdb.open(test_lmdb_path)\nlmdb_txn = lmdb_env.begin()\nlmdb_cursor = lmdb_txn.cursor()\n\nfor key, value in lmdb_cursor:\n    datum = caffe.proto.caffe_pb2.Datum()\n        datum.ParseFromString(value)\n        label = int(datum.label)\n        image = caffe.io.datum_to_array(datum)\n        image = image.astype(np.uint8)\n    out = net.forward_all(data=np.asarray([image]) - mean_image)\n    plabel = int(out['prob'][0].argmax(axis=0))\n    count += 1\n    iscorrect = label == plabel\n    correct += (1 if iscorrect else 0)\n    matrix[(label, plabel)] += 1\n    labels_set.update([label, plabel])\n\n    if not iscorrect:\n            print(\"\\rError: key = %s, expected %i but predicted %i\" % (key, label, plabel))\n        sys.stdout.write(\"\\rAccuracy: %.1f%%\" % (100.*correct\/count))\n        sys.stdout.flush()\n\nprint(\"\\n\" + str(correct) + \" out of \" + str(count) + \" were classified correctly\")\nprint \"\"\nprint \"Confusion matrix:\"\nprint \"(r , p) | count\"\nfor l in labels_set:\n    for pl in labels_set:\n        print \"(%i , %i) | %i\" % (l, pl, matrix[(l,pl)])\n<\/code><\/pre>\n"}]}
{"QuestionId":35529626,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-20T22:02:51.273","AcceptedAnswerId":35530400.0,"OwnerUserId":2138200.0,"Title":"tf.decode_csv() error: \"Unquoted fields cannot have quotes\/CRLFs inside\"","Body":"<p>I have a csv file  that looks like:<\/p>\n\n<pre><\/pre>\n\n<p>I can read the csv as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Which outputs:<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to decode the csv as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Now when I run  I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I decode this csv?<\/p>\n","answers":[{"AnswerId":"35530400","CreationDate":"2016-02-20T23:30:38.177","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#read_file\" rel=\"nofollow\"><code>tf.read_file()<\/code><\/a> op reads the <em>entire<\/em> contents of the given file into a single string, whereas <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#decode_csv\" rel=\"nofollow\"><code>tf.decode_csv()<\/code><\/a> op expects each element of its input to be a single <em>record<\/em> (i.e. one line). Therefore you need something that reads one line at a time, which the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#TextLineReader\" rel=\"nofollow\"><code>tf.TextLineReader<\/code><\/a> supports.<\/p>\n\n<p>Using a reader is slightly more complicated than using a simple op, because it's designed for reading large multi-file datasets with a lot of flexibility in how the files are chosen. You can see the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#reading-from-files\" rel=\"nofollow\">tutorial<\/a> \nfor a complete explanation, but the following example code should help to get you started:<\/p>\n\n<pre><code># Read the file once.\nfilenames = tf.train.string_input_producer([\"blah.txt\"], num_epochs=1)\n\nreader = tf.TextLineReader()\n\n_, line = reader.read(filenames)\n\ncol1, col2 = tf.decode_csv(line,\n                           record_defaults=[tf.constant([],dtype=tf.int32),\n                                            tf.constant([],dtype=tf.int32)])\n<\/code><\/pre>\n\n<p>Now <code>col1<\/code> and <code>col2<\/code> represent single values. If you evaluate them, you'll get the contents of the next line:<\/p>\n\n<pre><code># N.B. These must be called before evaluating the inputs.\nsess.run(tf.initialize_all_variables()\ntf.train.start_queue_runners(sess)\n\nprint sess.run([col1, col2])  # ==&gt; 1, 2\nprint sess.run([col1, col2])  # ==&gt; 3, 4\n<\/code><\/pre>\n\n<p>If instead you want to batch the columns, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#batch\" rel=\"nofollow\"><code>tf.train.batch()<\/code><\/a>:<\/p>\n\n<pre><code>_, line = reader.read(filenames)\n\n# N.B. Batch size is 2, to match the size of your file.\nline_batch, = tf.train.batch([line], batch_size=2)\n\ncol1, col2 = tf.decode_csv(line_batch,\n                           record_defaults=[tf.constant([],dtype=tf.int32),\n                                            tf.constant([],dtype=tf.int32)])\n\nsess.run(tf.initialize_all_variables()\ntf.train.start_queue_runners(sess)\n\nprint sess.run([col1, col2])  # ==&gt; [1, 3], [2, 4]\n<\/code><\/pre>\n"}]}
{"QuestionId":35529944,"AnswerCount":0,"Tags":"<caffe><multi-targeting>","CreationDate":"2016-02-20T22:37:10.223","AcceptedAnswerId":null,"OwnerUserId":5765030.0,"Title":"Implementing a Multi Target Regression Net with Caffe with Non-image data input","Body":"<p>I was trying to implement a simple net in Caffe for coping with multi-target regression. <\/p>\n\n<p>The dimension of my input is 5.\nThe dimension of my output is 3.<\/p>\n\n<p>I have this data on a mat file.<\/p>\n\n<p>I have used  and  to create both  and  files.<\/p>\n\n<pre><\/pre>\n\n<p>I designed a network in Caffe as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately It does not work! I got the following error<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I think that the problem is in the way I'm organizing the data. And also in the use of the Accuracy layer.<\/p>\n\n<p>However, any idea about how to solve it?<\/p>\n","answers":[]}
{"QuestionId":35530307,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><torch>","CreationDate":"2016-02-20T23:18:47.193","AcceptedAnswerId":35530414.0,"OwnerUserId":5954650.0,"Title":"Transfer function of neural network","Body":"<p>I used torch7 building a  3 hidden layers neural network to handle a classification problem (3 classes.). But I am confused about what transfer function to use and how to use them. Below is the structure of my network:<\/p>\n\n<pre><\/pre>\n\n<p>As above, I used all Tanh() transfer functions, is that correct? Can I use other transfer fucntion (like Sigmoid().. )? And do I have to insert transfer function between each layer?<\/p>\n\n<p>Many thanks ahead.<\/p>\n","answers":[{"AnswerId":"35530414","CreationDate":"2016-02-20T23:33:46.673","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<blockquote>\n  <p>As above, I used all Tanh() transfer functions, is that correct? <\/p>\n<\/blockquote>\n\n<p>It is correct, as well as using any other transfer function would be correct.<\/p>\n\n<blockquote>\n  <p>Can I use other transfer fucntion (like Sigmoid().. )? <\/p>\n<\/blockquote>\n\n<p>Yes, you can use any transfer function, each has its own properties, which are way to long to express in SO form, however nowadays you might find that ReLU is one of the most commonly used ones, especially in deeper networks.<\/p>\n\n<blockquote>\n  <p>And do I have to insert transfer function between each layer?<\/p>\n<\/blockquote>\n\n<p>Yes, if you do not - mathematically speaking your layers collapse (consequtive linear layers behave as a single linear layer at least in the sense of final behaviour - training might be different).<\/p>\n"}]}
{"QuestionId":35531290,"AnswerCount":1,"Tags":"<machine-learning><neural-network><computer-vision><deep-learning><caffe>","CreationDate":"2016-02-21T01:40:47.063","AcceptedAnswerId":35535145.0,"OwnerUserId":4308982.0,"Title":"Caffe only trains for one label","Body":"<p>This is really weird. I'm implementing this model:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ARwF2.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ARwF2.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Except that I read data from a text file using an ImageData blob, batch_size: 1. There are only two labels and the text file is organized as usual<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>Still, Caffe only trains and tests label 0!<\/p>\n\n<p>I run caffe using the regular tool.<\/p>\n\n<p><\/p>\n\n<p>When I open the net in pycaffe I get this message for the first time ever:<\/p>\n\n<p><\/p>\n\n<p>and the size of the\n\nis now 1, when it should be 2!<\/p>\n\n<p>Not only that but that label seems to be a float rather than an integer.<\/p>\n\n<pre><\/pre>\n\n<p>I know that this has worked before, I just can't get my head around what I'm doing wrong or where to begin troubleshoot.<\/p>\n","answers":[{"AnswerId":"35535145","CreationDate":"2016-02-21T10:49:27.487","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The output shape of your network depends on the input <code>batch_size<\/code>: if you define <code>batch_size: 1<\/code> than your net processes a <strong>single<\/strong> example each time, thus it only reads a single <code>label<\/code>. If you change <code>batch_size<\/code> to 2, caffe will read two samples and consequently the shape of <code>label<\/code> will become 2.<br>\nOne exception to this \"shape rule\" is the loss output: the loss defines a <strong>scalar<\/strong> function with respect to which gradients are computed. Thus, the loss output will always be a scalar regardless of the input shape.<\/p>\n\n<p>Regarding the data type of <code>label<\/code>: Caffe stores all variables in \"Blobs\" of type <code>float32<\/code>.<\/p>\n"}]}
{"QuestionId":35532457,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-21T04:50:09.877","AcceptedAnswerId":null,"OwnerUserId":3067467.0,"Title":"TensorFlow placeholder variable for integer or boolean isn't working","Body":"<p>I am using the following code snippet in TensorFlow to conditionally pull data from one source or another:<\/p>\n\n<pre><\/pre>\n\n<p>I keep obtaining error vomit as threads spazz out, but the recurring theme is the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any help would very much be appreciated!<\/p>\n","answers":[{"AnswerId":"35749825","CreationDate":"2016-03-02T14:22:30.033","ParentId":null,"OwnerUserId":"5499453","Title":null,"Body":"<p>A solution is mentionned here:\n<a href=\"http:\/\/andyljones.tumblr.com\/\" rel=\"nofollow\">http:\/\/andyljones.tumblr.com\/<\/a>.<\/p>\n\n<p>You have to change your placeholder <code>is_train = tf.placeholder(tf.int32)<\/code> for a tf.Variable:<code>is_train =  tf.Variable(True,  name='training')<\/code>.\nSo you will be able to initialize through your call at sess.run(tf.initialize_all_variables()) before making a call too <code>tf.train.start_queue_runners(sess=sess)<\/code>.<\/p>\n"}]}
{"QuestionId":35533395,"AnswerCount":3,"Tags":"<machine-learning><neural-network><keras><theano><conv-neural-network>","CreationDate":"2016-02-21T07:18:16.517","AcceptedAnswerId":35599246.0,"OwnerUserId":5957862.0,"Title":"Keras: How to feed input directly into other hidden layers of the neural net than the first?","Body":"<p>I have a question about using Keras to which I'm rather new. I'm using a convolutional neural net that feeds its results into a standard perceptron layer, which generates my output. This CNN is fed with a series of images. This is so far quite normal.<\/p>\n\n<p>Now I like to pass a short non-image input vector directly into the last perceptron layer without sending it through all the CNN layers. How can this be done in Keras?<\/p>\n\n<p>My code looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Any answers are greatly appreciated, thanks!<\/p>\n","answers":[{"AnswerId":"35599246","CreationDate":"2016-02-24T10:17:18.213","ParentId":null,"OwnerUserId":"2602240","Title":null,"Body":"<p>Provided your Keras's backend is Theano, you can do the following:\n<\/p>\n\n<pre><code>import theano\nimport numpy as np\n\nd = Dense(1500, W_regularizer=l2(1e-3), activation='relu') # I've joined activation and dense layers, based on assumption you might be interested in post-activation values\nmodel.add(d)\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\n\nc = theano.function([d.get_input(train=False)], d.get_output(train=False))\nlayer_input_data = np.random.random((1,20000)).astype('float32') # refer to d.input_shape to get proper dimensions of layer's input, in my case it was (None, 20000)\no = c(layer_input_data)\n<\/code><\/pre>\n"},{"AnswerId":"35640316","CreationDate":"2016-02-25T23:11:52.233","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>You didn't show which kind of model you us, but I assume that you initialized your model as <code>Sequential<\/code>. In a sequential model you can only stack one layer after another - so adding a \"short-cut\" connection is not possible.<\/p>\n\n<p>For this reason authors of Keras added option of building \"graph\" models. In this case you can build a graph (DAG) of your computations. It's a more complicated than designing a stack of layers, but still quite easy. <\/p>\n\n<p>Check the documentation <a href=\"http:\/\/keras.io\/models\/#using-the-graph-model\" rel=\"nofollow noreferrer\">site<\/a> to look for more details.<\/p>\n"},{"AnswerId":"48911206","CreationDate":"2018-02-21T16:46:55.637","ParentId":null,"OwnerUserId":"898213","Title":null,"Body":"<p>The answer <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2774\" rel=\"nofollow noreferrer\">here<\/a> works is more high level and works also for <code>tensorflow<\/code> backend:<\/p>\n\n<pre><code>input_1 = Input(input_shape)\ninput_2 = Input(input_shape)\n\nmerge = merge([input_1, input_2], mode=\"concat\")  # could also to \"sum\", \"dot\", etc.\nhidden = Dense(hidden_dims)(merge)\nclassify = Dense(output_dims, activation=\"softmax\")(hidden)\n\nmodel = Model(input=[input_1, input_2], output=hidden)\n<\/code><\/pre>\n"}]}
{"QuestionId":35533703,"AnswerCount":2,"Tags":"<deep-learning><convolution><caffe><mnist>","CreationDate":"2016-02-21T07:58:09.830","AcceptedAnswerId":35553300.0,"OwnerUserId":3870344.0,"Title":"Hard to understand Caffe MNIST example","Body":"<p>After going through the Caffe tutorial here: <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html<\/a><\/p>\n\n<p>I am really confused about the different (and efficient) model using in this tutorial, which is defined here:     <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt<\/a><\/p>\n\n<p>As I understand, Convolutional layer in Caffe simply calculate the sum of Wx+b for each input, without applying any activation function. If we would like to add the activation function, we should add another layer immediately below that convolutional layer, like Sigmoid, Tanh, or Relu layer. Any paper\/tutorial I read on the internet applies the activation function to the neuron units.<\/p>\n\n<p>It leaves me a big question mark as we only can see the Convolutional layers and Pooling layers interleaving in the model. I hope someone can give me an explanation.<\/p>\n\n<p>As a site note, another doubt for me is the max_iter in this solver:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver.prototxt<\/a><\/p>\n\n<p>We have 60.000 images for training, 10.000 images for testing. So why does the max_iter here only 10.000 (and it still can get > 99% accuracy rate)? What does Caffe do in each iteration?\nActually, I'm not so sure if the accuracy rate is the total correct prediction\/test size.<\/p>\n\n<p>I'm very amazed of this example, as I haven't found any example, framework that can achieve this high accuracy rate in that very short time (only 5 mins to get >99% accuracy rate). Hence, I doubt there should be something I misunderstood.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"43812557","CreationDate":"2017-05-05T19:34:41.893","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>For your question about the missing activation layers, you are correct. The model in the tutorial is missing activation layers. This seems to be an oversight of the tutorial. For the real LeNet-5 model, there should be activation functions following the convolution layers. For MNIST, the model still works surprisingly well without the additional activation layers.<\/p>\n\n<p>For reference, in Le Cun's 2001 paper, it states:<\/p>\n\n<blockquote>\n  <p>As in classical neural networks, units in layers up to F6 compute a dot product between their input vector and their weight vector, to which a bias is added. This weighted sum, denoted <em>a_i<\/em>, for unit <em>i<\/em>, is then passed through a sigmoid squashing function to produce the state of unit <em>i<\/em> ...<\/p>\n<\/blockquote>\n\n<p>F6 is the \"blob\" between the two fully connected layers. Hence the first fully connected layers should have an activation function applied (the tutorial uses ReLU activation functions instead of sigmoid).<\/p>\n\n<p>MNIST is the hello world example for neural networks. It is very simple to today's standard. A single fully connected layer can solve the problem with accuracy of about 92%. Lenet-5 is a big improvement over this example.<\/p>\n"},{"AnswerId":"35553300","CreationDate":"2016-02-22T12:10:56.723","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Caffe uses batch processing. The <code>max_iter<\/code> is 10,000 because the <code>batch_size<\/code> is 64. <code>No of epochs = (batch_size x max_iter)\/No of train samples<\/code>. So the number of <code>epochs<\/code> is nearly 10. The accuracy is calculated on the <code>test data<\/code>. And yes, the accuracy of the model is indeed >99% as the dataset is not very complicated.<\/p>\n"}]}
{"QuestionId":35534179,"AnswerCount":3,"Tags":"<python><scipy><theano>","CreationDate":"2016-02-21T08:57:31.110","AcceptedAnswerId":36312607.0,"OwnerUserId":2051572.0,"Title":"Solving reaction-diffusion system with Theano","Body":"<p>I am new to Theano, and I try to implement a numerical integrator of a reaction-diffusion system - <a href=\"https:\/\/en.wikipedia.org\/wiki\/FitzHugh%E2%80%93Nagumo_model\" rel=\"noreferrer\">FitzHugh\u2013Nagumo model<\/a> of this version:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OuQSG.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OuQSG.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KTKyV.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KTKyV.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>For now my expressions are:<\/p>\n\n<pre><\/pre>\n\n<p>So I haven't implemented the finite-differences <a href=\"https:\/\/en.wikipedia.org\/wiki\/Discrete_Laplace_operator#Finite_Differences\" rel=\"noreferrer\">laplacian operator<\/a> yet. My question is whether there is a smart way of doing it in Theano?<\/p>\n","answers":[{"AnswerId":"35534439","CreationDate":"2016-02-21T09:29:54.333","ParentId":null,"OwnerUserId":"37213","Title":null,"Body":"<p>I see two coupled, first order, non-linear, ordinary differential equations here.<\/p>\n\n<p>Update: Now your equations are clear - Laplacians are there; two coupled nonlinear PDEs.  <em>Much<\/em> better.<\/p>\n\n<p>You need a finite difference or finite element approach for your spatial discretization.  Your choice, of course, but I'd prefer a finite element approach over finite differences.  <\/p>\n\n<p>You also need some kind of numerical integration in time.  An implicit error correcting scheme would be best.<\/p>\n\n<p>I looked quickly at the Theano docs.  I didn't see anything to help you with your spatial discretization problem.  Once you accomplish that you'll have matrix equations that you can solve, but I don't believe Theano will help you formulate the problem.  <\/p>\n\n<p>I'll admit that I'm not a Theano maven.<\/p>\n"},{"AnswerId":"36313100","CreationDate":"2016-03-30T15:20:19.250","ParentId":null,"OwnerUserId":"5609523","Title":null,"Body":"<p>An interesting example of a similar, but simpler problem, solved using convolutional networks on Google's tensorflow can be found here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/pdes\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/pdes\/index.html<\/a><\/p>\n\n<p>In particular they use the following definition of the diffusion kernel:<\/p>\n\n<pre><code>  laplace_k = make_kernel([[0.5, 1.0, 0.5],\n                           [1.0, -6., 1.0],\n                           [0.5, 1.0, 0.5]])\n<\/code><\/pre>\n"},{"AnswerId":"36312607","CreationDate":"2016-03-30T14:58:16.677","ParentId":null,"OwnerUserId":"3667142","Title":null,"Body":"<p>Is there any reason for using Theano? There are other ways in Python to solve a system of coupled non-linear ODEs.<\/p>\n\n<p>The reaction-diffusion system definition from Google seems to suggest that u(x,y,t), v(x,y,t).<\/p>\n\n<p>I am not a user of Theano, but it looks like casting the problem in the form of an equation like b = Ax is the way to go. <\/p>\n\n<p>Some resources that I came across on Google for using Theano and generally solving PDEs are below.<\/p>\n\n<p><a href=\"http:\/\/www.ericlwilkinson.com\/blog\/2015\/1\/11\/numerical-relaxation-on-the-gpu-with-theano\" rel=\"nofollow\">Expressing the Laplacian using Theano<\/a><\/p>\n\n<p><a href=\"http:\/\/ipython-books.github.io\/featured-05\/\" rel=\"nofollow\">Solving a reaction-diffusion problem using numpy<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/mrocklin\/ShallowWater\" rel=\"nofollow\">Github project using Theano to solve the shallow water PDE<\/a><\/p>\n"}]}
{"QuestionId":35535246,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-02-21T11:01:16.113","AcceptedAnswerId":35540010.0,"OwnerUserId":5958323.0,"Title":"tensorflow restore only variables","Body":"<p>I do some training in Tensorflow and save the whole session using a saver:<\/p>\n\n<pre><\/pre>\n\n<p>It works fine, and I can successfully restore the whole session by using the exact same model and calling:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to modify only the optimizer while keeping the rest of the model constant (the computation graph stays the same apart from the optimizer):<\/p>\n\n<pre><\/pre>\n\n<p>I also want to continue the training from the last graph state I saved (i.e., I want to restore the state of my variables and continue with another training algorithm). Of course I cannot use:<\/p>\n\n<pre><\/pre>\n\n<p>any longer, because the graph has changed.<\/p>\n\n<p>So my question is: is there a way to restore only variables using the saver.restore command (or even, maybe for later use, only a subset of variables), when the whole session has been saved? I looked for such feature in the API documentation and online, but could not find any example \/ detailed enough explanations that could help me get it to work.<\/p>\n","answers":[{"AnswerId":"39007632","CreationDate":"2016-08-17T23:15:20.567","ParentId":null,"OwnerUserId":"1121123","Title":null,"Body":"<p>you could recover the original <code>Saver<\/code> from a <code>MetaGraph<\/code> protobuf first and then use that saver to restore <strong>all<\/strong> old variables safely. For a concrete example, you can take a look at the <code>eval.py<\/code> script: <a href=\"https:\/\/stackoverflow.com\/questions\/36462229\/tensorflow-how-do-i-release-a-model-without-source-code\/38921538#38921538\">TensorFlow: How do I release a model without source code?<\/a><\/p>\n"},{"AnswerId":"45539128","CreationDate":"2017-08-07T03:47:33.257","ParentId":null,"OwnerUserId":"7949687","Title":null,"Body":"<p>I see the same problem. Inspired by keveman' s answer. My solution is:<\/p>\n\n<ol>\n<li><p>Define your new graph, (here only new optimizer related variables are different from the old graph).<\/p><\/li>\n<li><p>Get all variables using tf.global_variables(). This return a var list I called g_vars.<\/p><\/li>\n<li><p>Get all optimizer related variables using tf.contrib.framework.get_variables_by_suffix('some variable filter'). The filter may be RMSProp or RMSPRrop_*. This function returns a var list I called exclude_vars.<\/p><\/li>\n<li><p>Get the variables in g_vars but not in exclude_vars. Simply use<\/p>\n\n<p>vars = [item for item in g_vars if item not in exclude_vars]<\/p><\/li>\n<\/ol>\n\n<p>these vars are common vars in both new and old graph, which you can restore from old model now.<\/p>\n"},{"AnswerId":"35540010","CreationDate":"2016-02-21T18:18:51.700","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>It is possible to restore a subset of variables by passing the list of variables as the <code>var_list<\/code> argument to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow\"><code>Saver<\/code><\/a> constructor. However, when you change the optimizer, additional variables may have been created (momentum accumulators, for instance) and variable associated with the previous optimizer, if any, would have been removed from the model. So simply using the old <code>Saver<\/code> object to restore will not work, especially if you had constructed it with the default constructor, which uses <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/saver.py#L775\" rel=\"nofollow\"><code>tf.all_variables<\/code><\/a> as the argument to <code>var_list<\/code> parameter. You have to construct the <code>Saver<\/code> object on the subset of variables that you created in your model and then <code>restore<\/code> would work. Note that, this would leave the new variables created by the new optimizer uninitialized, so you have to explicitly initialize them.<\/p>\n"}]}
{"QuestionId":35536983,"AnswerCount":1,"Tags":"<macos><python-3.x><tensorflow><pip><anaconda>","CreationDate":"2016-02-21T13:53:54.003","AcceptedAnswerId":35553696.0,"OwnerUserId":1833854.0,"Title":"Installing Tensorflow with Python 3.5 and Anaconda","Body":"<p>I have Anaconda 3.19.3 with Python 3.5 on a Mac. Now, I would like to install TensorFlow via . The installation instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"nofollow noreferrer\">here<\/a> also include setup instructions for installing TensorFlow under Python 3.3+.<\/p>\n\n<p>So, I have executed , no problem. When I execute , after downloading the necessary packages, I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>I have left out some of the error message to save space.<\/p>\n\n<p>What is this and, more importantly, what can I do about it?<\/p>\n","answers":[{"AnswerId":"35553696","CreationDate":"2016-02-22T12:28:27.077","ParentId":null,"OwnerUserId":"4372162","Title":null,"Body":"<p>See the answer here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/34264512\/error-setuptools-when-installing-tensorflow\">Error setuptools when installing tensorflow<\/a><\/p>\n\n<p>In particular, you could try adding the <code>--ignore-installed<\/code> tag to the end of your pip command.<\/p>\n"}]}
{"QuestionId":35539359,"AnswerCount":1,"Tags":"<lua><torch><luaj>","CreationDate":"2016-02-21T17:25:09.007","AcceptedAnswerId":null,"OwnerUserId":5959245.0,"Title":"luaj doesn't find .so modules","Body":"<p>I've written a GUI in Java and use a lua-script to calculate some values for images  with a neural network. Therefore the lua-script requires some modules from torch7. \nI got so far that it finds the modules which have a init.lua file. However,  it fails when a module only has a .so file. The module is required in one of the init.lua files.\nBefore I require the modules in the script i set the new package.path and the package.cpath to LUA_PATH and LUA_CPATH because luaj only used the default path which didn't work. <\/p>\n\n<p>I think this a LuaJ problem because when I run the script in the terminal with lua script_name.lua it works fine. <\/p>\n\n<pre><\/pre>\n\n<p>The error message is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I call the script from my java program: <\/p>\n\n<pre><\/pre>\n\n<p>I use lua 5.1 because some of the problems where solved by changing from lua 5.2 to lua 5.1. I have Ubuntu 14.04 LTS and luaj 3.0.1.<\/p>\n\n<p>I really appreciate any help!<\/p>\n","answers":[{"AnswerId":"35585449","CreationDate":"2016-02-23T18:34:16.967","ParentId":null,"OwnerUserId":"1686442","Title":null,"Body":"<p>It seems that LuaJ does not have the ability to load dll files and so files.\nYou may want to look into jnlua:\n<a href=\"https:\/\/code.google.com\/archive\/p\/jnlua\/\" rel=\"nofollow\">https:\/\/code.google.com\/archive\/p\/jnlua\/<\/a><\/p>\n"}]}
{"QuestionId":35540269,"AnswerCount":3,"Tags":"<conv-neural-network><keras>","CreationDate":"2016-02-21T18:39:42.293","AcceptedAnswerId":49602859.0,"OwnerUserId":5808464.0,"Title":"keras giving same loss on every epoch","Body":"<p>I am newbie to keras.<\/p>\n\n<p>I ran it on a dataset where my objective was to reduce the logloss.\nFor every epoch it is giving me the same loss value. I am confused whether i am on the right track or not.<\/p>\n\n<p>For example:<\/p>\n\n<pre><\/pre>\n\n<p>Here 3.8019 is same in every epoch. It is supposed to be less.<\/p>\n","answers":[{"AnswerId":"49602859","CreationDate":"2018-04-01T21:36:56.990","ParentId":null,"OwnerUserId":"6318002","Title":null,"Body":"<p>I ran into this issue as well. After much deliberation, I figured out that it was my <strong>activation function on my output layer<\/strong>. <\/p>\n\n<p>I had this model to predict a binary outcome:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(16,input_shape=(8,),activation='relu'))\nmodel.add(Dense(32,activation='relu'))\nmodel.add(Dense(32,activation='relu'))\nmodel.add(Dense(1, activation='softmax'))\n<\/code><\/pre>\n\n<p>and I needed this for binary cross entropy<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(16,input_shape=(8,),activation='relu'))\nmodel.add(Dense(32,activation='relu'))\nmodel.add(Dense(32,activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>I would look towards the problem you are trying to solve and the output needed to ensure that your activation functions are what they need to be.<\/p>\n"},{"AnswerId":"49603149","CreationDate":"2018-04-01T22:19:28.277","ParentId":null,"OwnerUserId":"6806531","Title":null,"Body":"<p>Try decreasing your learning rate to 0.0001 and use Adam. What is your learning rate?<\/p>\n\n<p>It's actually not clear to see if its the problem of learning rate or model complexity, could you explain a bit further with these instructions:<\/p>\n\n<ol>\n<li>What is your data size, what is it about?<\/li>\n<li>What is your model's complexity? We can compare your complexity with analysing your data. If your data is too big, you need more complex model.<\/li>\n<li>Did you normalize your outputs? For inputs, it couldn't be a big deal since not-normalization gives results, but if your outputs are some numbers bigger than 1, you need to normalize your data. If you check for your last layer's activation function of model, they're usually sigmoid, softmax, tanh that frequently squeezes your output to 0-1 and -1 - 1. You need to normalize your data according to your last activation function, and then reverse multiplying them for getting real-life result.<\/li>\n<\/ol>\n\n<p>Since you're new to deep learning, these advices are enough to check if there's a problem on your model. Can you please check them and reply?<\/p>\n"},{"AnswerId":"47614504","CreationDate":"2017-12-03T01:49:47.143","ParentId":null,"OwnerUserId":"907770","Title":null,"Body":"<p>Usually this problem occurs when the model you are training does not have enough capacity (or the cost function is not appropriate). Or in some cases it happens that by mistake the data which we are feeding into the model is not prepared correctly and therefore the labels for each sample might not be  correct which makes the model helpless and it will not able to decrease the loss.<\/p>\n"}]}
{"QuestionId":35543428,"AnswerCount":2,"Tags":"<neural-network><theano><convolution>","CreationDate":"2016-02-21T23:20:54.400","AcceptedAnswerId":35550406.0,"OwnerUserId":2147347.0,"Title":"Activation function after pooling layer or convolutional layer?","Body":"<p>The theory from these links show that the order of Convolutional Network is: .<\/p>\n\n<ol>\n<li><a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html\" rel=\"noreferrer\">Neural networks and deep learning (equation (125)<\/a><\/li>\n<li><a href=\"http:\/\/www.deeplearningbook.org\/contents\/convnets.html\" rel=\"noreferrer\">Deep learning book (page 304, 1st paragraph)<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#details-and-notation\" rel=\"noreferrer\">Lenet (the equation)<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#the-convolution-operator\" rel=\"noreferrer\">The source in this headline<\/a><\/li>\n<\/ol>\n\n<p>But, in the last implementation from those sites, it said that the order is: <\/p>\n\n<ol>\n<li><a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap6.html#the_code_for_our_convolutional_networks\" rel=\"noreferrer\">network3.py<\/a><\/li>\n<li><a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html#putting-it-all-together\" rel=\"noreferrer\">The sourcecode, LeNetConvPoolLayer class<\/a><\/li>\n<\/ol>\n\n<p>I've tried too to explore a Conv2D operation syntax, but there is no activation function, it's only convolution with flipped kernel. Can someone help me to explain why is this happen?<\/p>\n","answers":[{"AnswerId":"44250343","CreationDate":"2017-05-29T21:12:12.687","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>In many papers people use <code>conv -&gt; pooling -&gt; non-linearity<\/code>. It does not mean that you can't use another order and get reasonable results. In case of max-pooling layer and ReLU the order does not matter (both calculate the same thing):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DggnU.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DggnU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You can proof that this is the case by remembering that ReLU is an element-wise operation and a non-decreasing function so <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LHbZl.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LHbZl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The same thing happens for almost every activation function (most of them are non-decreasing). But does not work for a general pooling layer (average-pooling).<\/p>\n\n<hr>\n\n<p>Nonetheless both orders produce the same result, <code>Activation(MaxPool(x))<\/code> does it significantly faster by doing less amount of operations. For a pooling layer of size <code>k<\/code>, it uses <code>k^2<\/code> times less calls to activation function.<\/p>\n\n<p>Sadly this optimization is negligible for CNN, because majority of the time is used in convolutional layers.<\/p>\n"},{"AnswerId":"35550406","CreationDate":"2016-02-22T09:48:59.147","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Well, max-pooling and monotonely increasing non-linearities commute. This means that MaxPool(Relu(x)) = Relu(MaxPool(x)) for any input. So the result is the same in that case. So it is technically better to first subsample through max-pooling and then apply the non-linearity (if it is costly, such as the sigmoid). In practice it is often done the other way round - it doesn't seem to change much in performance.<\/p>\n\n<p>As for conv2D, it does <em>not<\/em> flip the kernel. It implements exactly the definition of convolution. This is a linear operation, so you have to add the non-linearity yourself in the next step, e.g. <code>theano.tensor.nnet.relu<\/code>.<\/p>\n"}]}
{"QuestionId":35544356,"AnswerCount":6,"Tags":"<macos><caffe>","CreationDate":"2016-02-22T01:20:28.103","AcceptedAnswerId":35638559.0,"OwnerUserId":5136799.0,"Title":"Install Caffe on Mac error: \"fatal error: 'cblas.h' file not found\"","Body":"<p>I've been following <a href=\"http:\/\/playittodeath.ru\/how-to-install-caffe-on-mac-os-x-yosemite-10-10-4\/\" rel=\"noreferrer\">this guide<\/a> to install on my El Capitan macbook pro, using CMake to build the makefile. However, when I attempt to make, I run into the error of missing the cblas.h file. I successfully reinstalled openblas with: <\/p>\n\n<p> <\/p>\n\n<p>and adding the two build flags:<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>However I still get the error. Then I reinstalled CUDA and xcode-select install tools as recommended <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/eHcmYJvYBHs\" rel=\"noreferrer\">here<\/a>. <\/p>\n\n<p>Any recommendations\/further speculation about why make doesn't see my cblas.h file is appreciated!<\/p>\n","answers":[{"AnswerId":"46268162","CreationDate":"2017-09-17T19:39:24.463","ParentId":null,"OwnerUserId":"4848859","Title":null,"Body":"<p>Uncomment this lines in Makefile.config:<\/p>\n\n<pre><code># Homebrew puts openblas in a directory that is not on the standard search path \nBLAS_INCLUDE := $(shell brew --prefix openblas)\/include \nBLAS_LIB := $(shell brew --prefix openblas)\/lib\n<\/code><\/pre>\n"},{"AnswerId":"35638559","CreationDate":"2016-02-25T21:18:57.923","ParentId":null,"OwnerUserId":"2915032","Title":null,"Body":"<p>In your question I see <code>cmake<\/code>, but in provided link <code>Makefile.config<\/code>.<\/p>\n\n<p>If you're still want to proceed with <strong>cmake<\/strong>:<\/p>\n\n<p>first of all, build script is NOT broken - I tried it on Yosemite with the latest <a href=\"https:\/\/github.com\/BVLC\/caffe\/commit\/fe0f44112a153377ff4c418adefc8c690b872c37\">master<\/a> branch (2016-02-25), but you should slightly adjust it if you already made<\/p>\n\n<p><code>brew uninstall openblas; brew install --fresh -vd openblas<\/code><\/p>\n\n<p>Just run your cmake with additional arguments:<\/p>\n\n<pre><code>cmake -DCMAKE_CXX_FLAGS=-I\/usr\/local\/opt\/openblas\/include &lt;path to your sources&gt;\n<\/code><\/pre>\n\n<p>and then usual make or xcode build (depends on what generator you use).<\/p>\n\n<p>After this it was successfully built on my system.<\/p>\n\n<p><strong>UPD:<\/strong> you can also use libraries in <code>Accelerate.framework<\/code> (<code>vecLib<\/code>) instead of <strong>openblas<\/strong>.<\/p>\n"},{"AnswerId":"37557631","CreationDate":"2016-06-01T00:25:23.380","ParentId":null,"OwnerUserId":"448695","Title":null,"Body":"<p>I ran into this issue myself today and submitted a pull request to mainline Caffe:<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/4247\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/4247<\/a><\/p>\n\n<p>For me, the issue was that Caffe was finding the deprecated vecLib framework (which doesn't ship cblas.h) instead of the new Accelerate framework (which does ship cblas.h).<\/p>\n"},{"AnswerId":"53899739","CreationDate":"2018-12-22T22:25:34.303","ParentId":null,"OwnerUserId":"6543500","Title":null,"Body":"<p>For my Caffe installation, this was my setup:<\/p>\n\n<ul>\n<li>MacOS 10.12 Sierra, CPU_ONLY=ON, no CUDA, no Anaconda, with Python interface<\/li>\n<li>Compilation via Make (instead of Cmake) <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow noreferrer\">official instructions here<\/a><\/li>\n<\/ul>\n\n<p>The directory of my <code>cblas.h<\/code> file (via Finder search):\n<code>\/System\/Library\/Frameworks\/Accelerate.framework\/Versions\/A\/Frameworks\/vecLib.framework\/Headers<\/code><\/p>\n\n<p>Add that cblas.h path to <code>.\/caffe\/Makefile.config<\/code>:<\/p>\n\n<pre><code># Whatever else you find you need goes here.\nINCLUDE_DIRS := $(PYTHON_INCLUDE) \/usr\/local\/include \/System\/Library\/Frameworks\/Accelerate.framework\/Versions\/A\/Frameworks\/vecLib.framework\/Headers\/\nLIBRARY_DIRS := $(PYTHON_LIB) \/usr\/local\/lib \/usr\/lib\n<\/code><\/pre>\n\n<p><strong>Additional notes<\/strong><\/p>\n\n<p>Also for me, I had OpenBlas. Might impact your situation.<\/p>\n\n<pre><code># BLAS choice:\n# atlas for ATLAS (default)\n# mkl for MKL\n# open for OpenBlas\nBLAS := open\n<\/code><\/pre>\n\n<p>I also modified <code>.\/caffe\/Makefile<\/code> with the cblas.h path, but not sure if it was necessary \/ made a difference.<\/p>\n\n<pre><code>else ifeq ($(BLAS), open)\n    # OpenBLAS\n    LIBRARIES += openblas\n    BLAS_INCLUDE ?= \/System\/Library\/Frameworks\/Accelerate.framework\/Versions\/A\/Frameworks\/vecLib.framework\/Headers\/\n<\/code><\/pre>\n\n<p><strong>Resulting outcome<\/strong><\/p>\n\n<p>The original error disappeared:<\/p>\n\n<pre><code>Tams-MacBook-Pro:caffe tamtam$ make all\nCXX src\/caffe\/blob.cpp\nIn file included from src\/caffe\/blob.cpp:7:\nIn file included from .\/include\/caffe\/util\/math_functions.hpp:11:\n.\/include\/caffe\/util\/mkl_alternate.hpp:14:10: fatal error: 'cblas.h' file not found\n#include &lt;cblas.h&gt;\n         ^\n1 error generated.\nmake: *** [.build_release\/src\/caffe\/blob.o] Error 1\n<\/code><\/pre>\n\n<p>But the next error unraveled was (the tail end of output):<\/p>\n\n<pre><code>CXX src\/caffe\/util\/upgrade_proto.cpp\nAR -o .build_release\/lib\/libcaffe.a\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_conv_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_lcn_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_lrn_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_pooling_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_relu_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_sigmoid_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_softmax_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn_tanh_layer.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(parallel.o) has no symbols\n\/Applications\/Xcode833.app\/Contents\/Developer\/Toolchains\/XcodeDefault.xctoolchain\/usr\/bin\/ranlib: file: .build_release\/lib\/libcaffe.a(cudnn.o) has no symbols\nLD -o .build_release\/lib\/libcaffe.so.1.0.0-rc5\nclang: warning: argument unused during compilation: '-pthread' [-Wunused-command-line-argument]\nld: warning: directory not found for option '-L\/lib'\nld: library not found for -lopenblas\nclang: error: linker command failed with exit code 1 (use -v to see invocation)\nmake: *** [.build_release\/lib\/libcaffe.so.1.0.0-rc5] Error 1\n<\/code><\/pre>\n\n<p>Which I fix\/resolve\/modify via this in <code>Makefile.config<\/code>:<\/p>\n\n<pre><code># Homebrew puts openblas in a directory that is not on the standard search path\nBLAS_INCLUDE := $(brew --prefix openblas)\/include \/usr\/local\/Cellar\/openblas\/0.3.3\/include\nBLAS_LIB := $(brew --prefix openblas)\/lib \/usr\/local\/Cellar\/openblas\/0.3.3\/lib\n<\/code><\/pre>\n\n<p>I also resolve the <code>-pthread<\/code> warning via this in <code>Makefile<\/code>:<\/p>\n\n<pre><code>CXXFLAGS += -fPIC $(COMMON_FLAGS) $(WARNINGS) -std=c++11 #rm'd -pthread\nLINKFLAGS += -fPIC $(COMMON_FLAGS) $(WARNINGS) -std=c++11 #rm'd -pthread\n<\/code><\/pre>\n\n<p>With another <code>make clean<\/code>, <code>cp Makefile.config.example Makefile.config<\/code>, and <code>make all<\/code>, this is the tail end of the output and similarly the entire gist of the output:<\/p>\n\n<pre><code>In file included from .\/include\/caffe\/caffe.hpp:12:\n.\/include\/caffe\/net.hpp:41:5: warning: unused typedef 'INVALID_REQUESTED_LOG_SEVERITY' [-Wunused-local-typedef]\n    LOG_EVERY_N(WARNING, 1000) &lt;&lt; \"DEPRECATED: ForwardPrefilled() \"\n    ^\n\/usr\/local\/include\/glog\/logging.h:943:30: note: expanded from macro 'LOG_EVERY_N'\n                             INVALID_REQUESTED_LOG_SEVERITY);           \\\n                             ^\n1 warning generated.\nCXX\/LD -o .build_release\/examples\/cpp_classification\/classification.bin\nld: warning: directory not found for option '-L\/lib'\nCXX examples\/mnist\/convert_mnist_data.cpp\nCXX\/LD -o .build_release\/examples\/mnist\/convert_mnist_data.bin\nld: warning: directory not found for option '-L\/lib'\nCXX examples\/siamese\/convert_mnist_siamese_data.cpp\nCXX\/LD -o .build_release\/examples\/siamese\/convert_mnist_siamese_data.bin\nld: warning: directory not found for option '-L\/lib'\n<\/code><\/pre>\n\n<p>So it seems that <code>make all<\/code> built successfully, assuming that this <code>-L\/lib<\/code> error will not cause further problems down the line. Note: <code>make all<\/code>, <code>make test<\/code>, and <code>make runtest<\/code> run fine with just warnings produced. <code>make pycaffe<\/code> however produced further errors; I might link to my solutions later. Also, yes this answer is indeed long and beyond what the OP asked, but hopefully my answer might help someone else in this dreadful installation process.<\/p>\n"},{"AnswerId":"42081523","CreationDate":"2017-02-07T04:03:43.470","ParentId":null,"OwnerUserId":"7405440","Title":null,"Body":"<p>Go to the <code>Makefile.config<\/code> and edit the following lines with the path where your openblas is located, in my case the version is <code>0.2.18_2<\/code><\/p>\n\n<p>Be sure to uncomment these lines:<\/p>\n\n<pre><code>BLAS_INCLUDE := \/usr\/local\/Cellar\/openblas\/0.2.18_2\/include\nBLAS_LIB := \/usr\/local\/Cellar\/openblas\/0.2.18_2\/lib\n<\/code><\/pre>\n"},{"AnswerId":"53391190","CreationDate":"2018-11-20T10:40:29.967","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>In <code>Makefile.config<\/code>:<\/p>\n\n<pre><code>INCLUDE_DIRS := $(PYTHON_INCLUDE) \/usr\/local\/include \/usr\/local\/opt\/openblas\/include\/\n\nLIBRARY_DIRS := $(PYTHON_LIB) \/usr\/local\/lib \/usr\/lib \/usr\/local\/opt\/openblas\/lib\/\n<\/code><\/pre>\n\n<p>Not sure if it's the most elegant solution, but it works.<\/p>\n"}]}
{"QuestionId":35545798,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-22T04:24:11.157","AcceptedAnswerId":35546179.0,"OwnerUserId":2853544.0,"Title":"keep_prob in TensorFlow MNIST tutorial","Body":"<p>I can't understand the following code in <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/pros\/index.html#variables\" rel=\"noreferrer\">the Deep MNIST for Experts tutorial<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>What is the purpose of <strong><\/strong> when running ?<\/p>\n","answers":[{"AnswerId":"35546179","CreationDate":"2016-02-22T05:08:11.270","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>keep_prob<\/code> value is used to control the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/tutorials\/mnist\/pros\/index.html#dropout\" rel=\"noreferrer\">dropout rate<\/a> used when training the neural network. Essentially, it means that each connection between layers (in this case between the last densely connected layer and the readout layer) will only be used with probability <code>0.5<\/code> when training. This reduces overfitting. For more information on the theory of dropout, you can see the original <a href=\"https:\/\/www.cs.toronto.edu\/~hinton\/absps\/JMLRdropout.pdf\" rel=\"noreferrer\">paper by Srivastava et al<\/a>. To see how to use it in TensorFlow, see the documentation on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/nn.html#dropout\" rel=\"noreferrer\"><code>tf.nn.dropout()<\/code><\/a> operator.<\/p>\n\n<p>The <code>keep_prob<\/code> value is fed in via a placeholder so that the same graph can be used for training (with <code>keep_prob = 0.5<\/code>) and evaluation (with <code>keep_prob = 1.0<\/code>). An alternative way to handle these cases is to build different graphs for training and evaluation: look at the use of dropout in the current <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7e3b8b23835ab0ac55d390aed2349af6e05dbe3b\/tensorflow\/models\/image\/mnist\/convolutional.py#L216\" rel=\"noreferrer\"><code>convolutional.py<\/code><\/a> model for an example.<\/p>\n"}]}
{"QuestionId":35547241,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow><convolution>","CreationDate":"2016-02-22T06:39:31.300","AcceptedAnswerId":null,"OwnerUserId":5551505.0,"Title":"Tensorflow weight matrix rank error","Body":"<pre><\/pre>\n\n<p>When I run the code, I get the error,<\/p>\n\n<pre><\/pre>\n\n<p>But the Weights have been initialized and even then, it shows them as empty tensors.\nThe files and labels are being read and transfered properly.\nThe first convolutional layer has a 5x5 window with a depth of 3 and I want to have 32 such 5X5 filters. Hence a shape of [5,5,3,32] for W_conv1.<\/p>\n","answers":[{"AnswerId":"35549196","CreationDate":"2016-02-22T08:44:49.460","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>It looks like inputs is returning a 3d tensor while conv2d expects a 4d one (first dimension is batch idx) - if you just want to run one image, you'd reshape it to [1,180,180,3] first. <\/p>\n"},{"AnswerId":"35559980","CreationDate":"2016-02-22T17:25:59.860","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>inputs()<\/code> function returns a tensor of shape <code>180 x 180 x 3<\/code>, but <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.conv2d()<\/code><\/a> expects a 4-D tensor of shape <code>batch_size x height x width x num_channels<\/code>.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/35549196\/3574081\">As etarion suggests<\/a>, you can make this work by reshaping the <code>image<\/code> tensor (e.g. using <code>image = tf.expand_dims(image, 0)<\/code>). However, if you are training a neural network you will probably want to <em>batch<\/em> the inputs. One way to do this is using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#batch\" rel=\"nofollow noreferrer\"><code>tf.train.batch()<\/code><\/a>:<\/p>\n\n<pre><code>image, label = inputs()\n\n# Set batch_size to the largest value that works for your configuration.\nimage_batch, label_batch = tf.train.batch([image, label], batch_size=32)\n<\/code><\/pre>\n\n<p>...then use <code>image_batch<\/code> or <code>label_batch<\/code> where you've used <code>image<\/code> and <code>label<\/code> respectively.<\/p>\n"}]}
{"QuestionId":35550219,"AnswerCount":3,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-02-22T09:40:20.823","AcceptedAnswerId":35559023.0,"OwnerUserId":5870826.0,"Title":"Tensorflow pip built from source fails on 'import tensorflow'","Body":"<p>Built tensorflow 0.7 from source on Ubuntu 15.10, Cuda 7.5, cudnn 4.0, python 2.7.  Bazel build and pip package generation OK.  pip install ok.  Python script \"import tensorflow\" fails:<\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts?\nThanks<\/p>\n","answers":[{"AnswerId":"45349511","CreationDate":"2017-07-27T11:28:55.200","ParentId":null,"OwnerUserId":"4294554","Title":null,"Body":"<p>This is because of protobuf version conflict:\nfor me i had python-protobuf v2.6.1 installed from apt-get package manager\nand while installing tensorflow it installed protobuf v3.3.0 from pip manager<\/p>\n\n<p>so i uninstalled both:<\/p>\n\n<pre><code>sudo pip uninstall protobuf\nsudo apt-get remove python-protobuf\n<\/code><\/pre>\n\n<p>Reinstalled using pip and it solved the issue!!!<\/p>\n\n<pre><code>sudo pip install protobuf\n<\/code><\/pre>\n\n<p>Hope this helps..<\/p>\n"},{"AnswerId":"35972477","CreationDate":"2016-03-13T16:03:28.263","ParentId":null,"OwnerUserId":"601314","Title":null,"Body":"<p>I succeed in running TensorFlow 0.7 on a manjaro linux laptop using cuda acceleration by running python with optirun:<\/p>\n\n<pre><code>$optirun python\n<\/code><\/pre>\n\n<p>However, I had problem with protobuf and I had to uninstall\/reinstall previous version of protobuf and tensorflow with pip. <a href=\"http:\/\/bit.ly\/1RHApU5\" rel=\"nofollow\">http:\/\/bit.ly\/1RHApU5<\/a><\/p>\n"},{"AnswerId":"35559023","CreationDate":"2016-02-22T16:40:57.377","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>This does look like a protobuf library versioning error. There's a long thread on similar issues here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11<\/a>\nTowards the end there are some suggested solutions that might help.<\/p>\n"}]}
{"QuestionId":35550451,"AnswerCount":2,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-02-22T09:50:56.660","AcceptedAnswerId":35559836.0,"OwnerUserId":5962133.0,"Title":"Tensorflow, py_func, or custom function","Body":"<p>I'm currently working on a quaternionic Neural Network using Tensorflow (I want to use GPUs). TensorFlow doesn't have support for quaternions, <em>but<\/em> you can represent than as a 4x4 real matrix, so it might be possible to build such a neural network in TensorFlow.<\/p>\n\n<p>Is there a simple way to add a custom operation or to do a custom operation on tensors?<\/p>\n\n<p>For example, I can write: <\/p>\n\n<pre><\/pre>\n\n<p>...and that's pretty cool! All you have to do is add a loss function and then do backpropagation. However, I want to do the same thing but with quaternions, for example:<\/p>\n\n<pre><\/pre>\n\n<p>However, I need to transform the quaternions to and from tensors to optimize the GPU calculation. So I need to create a function that gets some tensors as parameters and returns the transformed tensors. \nI've looked at , but it seems that you can't return tensors.<\/p>\n\n<p>I tried the following, but it failed:<\/p>\n\n<pre><\/pre>\n\n<p>...and in :<\/p>\n\n<pre><\/pre>\n\n<p>Error : Unimplemented: Unsupported object type Tensor<\/p>\n\n<p>Ideally I could use this  in the standard backprop algorithm of TensorFlow but I don't know if it's possible.<\/p>\n","answers":[{"AnswerId":"35557811","CreationDate":"2016-02-22T15:41:18.010","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>If you want to run your custom operations on GPUs, you have to provide GPU implementation (kernels) in C++. Look at the documentation <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/adding_an_op\/index.html#adding-a-new-op\" rel=\"nofollow\">here<\/a> for how to extend TensorFlow with custom operations, and especially the section on <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/adding_an_op\/index.html#gpu-support\" rel=\"nofollow\">GPU support<\/a>.<\/p>\n"},{"AnswerId":"35559836","CreationDate":"2016-02-22T17:18:24.620","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Depending on the functionality required, you might be able to implement your operation as a composition of existing TensorFlow ops, without needing to use <code>tf.py_func()<\/code>.<\/p>\n\n<p>For example, the following works and will run on a GPU:<\/p>\n\n<pre><code>def layer_activation(input_tensor, weight_tensor):\n    return tf.matmul(input_tensor, weight_tensor)\n\n# ...\nx = tf.placeholder(...)\nW_to_hidden = tf.Variable(...)\ntest = layer_activation(input_tensor, weight_tensor)\n# ...\n<\/code><\/pre>\n\n<p>The main reason to use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/script_ops.html#py_func\" rel=\"nofollow noreferrer\"><code>tf.py_func()<\/code><\/a> is if your operations cannot be implemented using TensorFlow operations, and you want to inject some Python code (e.g. using NumPy) that works on the actual values of your tensor. <\/p>\n\n<p>However, if your <code>mySigmoid()<\/code> or <code>myFunction()<\/code> operations cannot be implemented in terms of existing TensorFlow operations, and you want to implement them on GPU, then&mdash;<a href=\"https:\/\/stackoverflow.com\/a\/35557811\/3574081\">as keveman says<\/a>&mdash;you will need to add a new op.<\/p>\n"}]}
{"QuestionId":35550628,"AnswerCount":1,"Tags":"<machine-learning><neural-network><tensorflow><recurrent-neural-network>","CreationDate":"2016-02-22T09:59:16.103","AcceptedAnswerId":null,"OwnerUserId":2305993.0,"Title":"How to construct a character based seq2seq model in tensorflow","Body":"<p>What changes are required to the existing seq2seq model in tensorflow so that I can use character units rather then the existing word units for the seq2seq task? And will this be a good configuration for a predictive ext application?<\/p>\n\n<p>The following function signatures may need modification for this task:<\/p>\n\n<pre><\/pre>\n\n<p>Apart from reduced input output vocabulary what other parameter changes would be be required to implement such a character level seq2seq model ?<\/p>\n","answers":[{"AnswerId":"36091600","CreationDate":"2016-03-18T18:17:31.160","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>I think you could use the existing seq2seq model in tensorflow without any code changes for character-based units if you prepare your input data files by whitespace separating your training examples like this:<\/p>\n\n<pre><code>The quick brown fox.\n<\/code><\/pre>\n\n<p>Becomes:<\/p>\n\n<pre><code>T h e _SPACE_ q u i c k _SPACE_ b r o w n _SPACE_ f o x .\n<\/code><\/pre>\n\n<p>Then your vocabulary naturally becomes characters not words.<\/p>\n\n<p>You can experiment vocab sizes, with embedding size, eliminate embedding layer, etc. to see what works best for your data.<\/p>\n"}]}
{"QuestionId":35551326,"AnswerCount":3,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-02-22T10:31:45.690","AcceptedAnswerId":null,"OwnerUserId":4190832.0,"Title":"Tensorflow Tensorboard default port","Body":"<p>Is there a way to change the default port (\"6006\") on tensorboard so we could open multiple tensorboard? Maybe an option like --port=\"8008\" ?<\/p>\n","answers":[{"AnswerId":"43568782","CreationDate":"2017-04-23T08:04:31.473","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>You should provide a port flag (<code>--port=6007<\/code>). <\/p>\n\n<p>But I am here to explain how you can find it and other flags without any documentation. Almost all command line tools have a flag <code>-h<\/code> or <code>--help<\/code> which shows all possible flags this tool allows.<\/p>\n\n<p>By running it you will see information about a port flag and that <code>--logdir<\/code> allows you to <\/p>\n\n<blockquote>\n  <p>also pass a comma separated list of log directories<\/p>\n<\/blockquote>\n\n<p>and you can also inspect separate event-files and tags with <code>--event_file<\/code> and <code>--tag<\/code> flags<\/p>\n"},{"AnswerId":"50582874","CreationDate":"2018-05-29T11:08:58.253","ParentId":null,"OwnerUserId":"7428659","Title":null,"Body":"<p>If you are already in the directory where TensorFlow writes its logs, you should specify the port first:<\/p>\n\n<pre><code>tensorboard --port=6007 --logdir runs\n<\/code><\/pre>\n\n<p>If you are feeding a directory to <code>logdir<\/code>, then the order doesn't matter. (I am using TensorBaord 1.8)<\/p>\n"},{"AnswerId":"35556485","CreationDate":"2016-02-22T14:42:43.483","ParentId":null,"OwnerUserId":"4190832","Title":null,"Body":"<p>In fact there is an option to change the default port ... <\/p>\n\n<pre><code>tensorboard --logdir=\/tmp  --port=8008\n<\/code><\/pre>\n"}]}
{"QuestionId":35552674,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><tensorflow><deep-learning>","CreationDate":"2016-02-22T11:40:13.147","AcceptedAnswerId":35559609.0,"OwnerUserId":3510526.0,"Title":"How to use event-by-event weights in Tensorflow?","Body":"<p>In my data set I have for every entry (event) a weight. This weight consist of several quantities but basically represent how important this event for the data and must be accounted for.<\/p>\n\n<p>How can I use this weights when training in Tensorflow? I don't want to simply use this as another feature.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35559609","CreationDate":"2016-02-22T17:07:14.233","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>One simple solution is to multiply the computed cost for each example by its weight, before computing the overall cost for a mini-batch.<\/p>\n\n<p>Let's say you have the following:<\/p>\n\n<pre><code># Vector of features per example.\nx = tf.placeholder(tf.float32, shape=[batch_size, num_features])\n\n# Scalar weight per example.\nx_weights = tf.placeholder(tf.float32, shape=[batch_size])\n\n# Vector of outputs per example.\ny = tf.placeholder(tf.float32, shape=[batch_size, num_outputs])\n\n# ...\nlogits = ...\n\n# Insert appropriate cost function here.\ncost = tf.nn.softmax_cross_entropy_with_logits(logits, y)\n<\/code><\/pre>\n\n<p>The computed <code>cost<\/code> tensor is a vector of length <code>batch_size<\/code>. You can simply perform an element-wise multiplication with <code>x_weights<\/code> to get a weighted cost.<\/p>\n\n<pre><code>overall_cost = tf.mul(cost, x_weights) \/ batch_size\n<\/code><\/pre>\n\n<p>Finally you can use <code>overall_cost<\/code> as the value to minimize in your <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#optimizers\" rel=\"nofollow\">optimizer<\/a>.<\/p>\n"}]}
{"QuestionId":35560123,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-02-22T17:32:14.400","AcceptedAnswerId":null,"OwnerUserId":5499453.0,"Title":"Stuck in an infinite loop when using while operation in tensorflow","Body":"<p>To crop a batch of images in tensorflow in parallel, I used those lines of code:<\/p>\n\n<pre><\/pre>\n\n<p>My question is this one: if I don't put the instruction  the program get stuck in what seems to be an infinite loop, Why?<\/p>\n","answers":[]}
{"QuestionId":35561665,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-22T18:57:59.373","AcceptedAnswerId":null,"OwnerUserId":5964720.0,"Title":"Tensorflow sequence2sequence model padding","Body":"<p>In the seq2seq models, paddings are applied to make all sequences in a bucket have the same lengths. And apart from this, it looks like no special handling is applied to the paddings:<\/p>\n\n<p>the encoder encodes the paddings as well\nthe basic decoder w\/o attention decodes using the last encoding which encodes the paddings\nthe decoder with attention attends to the hidden states of the padding inputs too\nIt would be really helpful if this could be clarified: is it true that, basically the paddings are just a special id\/embedding, and the current seq2seq implementation treats them just like other embeddings? And no special mechanism is needed to ignore these padding, for example when encoding a sequence containing paddings; or to decode a sequence containing paddings using the attention-based decoder? So after padding, nothing special is done to the paddings, we can just pretend a padding is just another embedding (apart from maybe when doing the weighted x-entropy using target_weights)?<\/p>\n\n<p>If the above is true, then when testing a trained model, is padding needed at all (since at test time, each sentence is decoded separately and not in a batch)? --- It looks like from the code, at test time an input sentence is still bucketed first and then padded?<\/p>\n","answers":[{"AnswerId":"37626818","CreationDate":"2016-06-04T05:47:09.027","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>I think your basic premise is correct: the model does not treat the padding symbol differently than any other symbol. However, when packing the data tensors the padding always shows up at the end of decoder training examples after the 'EOS' symbol, and at the beginning of encoder training examples (because the encoder sequences get reversed.) <\/p>\n\n<p>Presumably the model will learn that padding on the encoder side carries no real semantic information as it wouldn't correlate with anything of the other words...I suppose it could convey something about the length of the sequence when attention is factored in, but that wouldn't be a bad thing. The bucket approach is an attempt to limit the amount of excess padding.<\/p>\n\n<p>On the decoder side the model will quickly learn that padding always occurs after the 'EOS' symbol so you ignore everything after the 'EOS' symbol is emitted anyway.  <\/p>\n\n<p>Padding is mainly useful because when running tensors in batches, the sequences all have to be the same size...so when testing one-at-a-time you really don't need to pad. However, when testing large validation sets it is still useful to run in batches...with padding.<\/p>\n\n<p>(I have more questions and concerns about the <em>UNK<\/em> symbol myself.)<\/p>\n"}]}
{"QuestionId":35562272,"AnswerCount":1,"Tags":"<lua><torch><luaj>","CreationDate":"2016-02-22T19:33:47.623","AcceptedAnswerId":null,"OwnerUserId":5959245.0,"Title":"Require doesn't find shared libraries","Body":"<p>I'm trying to run a neural network from torch inside Java. I'm using luaj as a wrapper.<\/p>\n\n<p>The problem is that when I require for example the  module which depends on the  module which is a shared library it throws the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Before I require any module I set the  and  to the folders where the libraries are because before it only looked in the default path which was just the project folder. However I have the feeling that 'require' only looks in the package.path for modules and not in the cpath because then it would also find libpaths.so. Also because this output includes only the directories from path and not cpath:<\/p>\n\n<pre><\/pre>\n\n<p>I tried package.searchpath() with the path from the cpath and 'libpath' and it did find the libpaths.so. I tried a lot of different things but nothing seems to work so I would really appreciate some help!<\/p>\n\n<p>I use Lua 5.1.5, my OS is Ubuntu 14.04 LTS.<\/p>\n","answers":[{"AnswerId":"35587366","CreationDate":"2016-02-23T20:20:51.047","ParentId":null,"OwnerUserId":"2779972","Title":null,"Body":"<p>When you're using luarocks to install packages locally, you should setup your environment before starting lua\/luajit\/torch. The easiest way is to execute the output of <code>luarocks-5.1 path<\/code><\/p>\n\n<pre><code>$ luarocks-5.1 path\nexport LUA_PATH='\/home\/markus\/.luarocks\/share\/lua\/5.1\/?.lua;\/home\/markus\/.luarocks\/share\/lua\/5.1\/?\/init.lua;\/usr\/share\/lua\/5.1\/?.lua;\/usr\/share\/lua\/5.1\/?\/init.lua;.\/?.lua;\/usr\/lib\/lua\/5.1\/?.lua;\/usr\/lib\/lua\/5.1\/?\/init.lua'\nexport LUA_CPATH='\/home\/markus\/.luarocks\/lib\/lua\/5.1\/?.so;\/usr\/lib\/lua\/5.1\/?.so;.\/?.so;\/usr\/lib\/lua\/5.1\/loadall.so'\n<\/code><\/pre>\n\n<p>A more comfortable way is to put the output into your <code>~\/.bashrc<\/code><\/p>\n\n<pre><code>$ luarocks-5.1 path &gt;&gt; ~\/.bashrc\n<\/code><\/pre>\n\n<p>To reload your <code>~\/.bashrc<\/code> without logout and login, do <code>source ~\/.bashrc<\/code><\/p>\n"}]}
{"QuestionId":35562576,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-02-22T19:52:08.943","AcceptedAnswerId":35562954.0,"OwnerUserId":1112654.0,"Title":"Tensorflow equivalent for Numpy indexed assignment","Body":"<p>What is the pseudo(*)-equivalent in Tensorflow for this?<\/p>\n\n<pre><\/pre>\n\n<p>I guess it should be something like:<\/p>\n\n<pre><\/pre>\n\n<p>(*) I do not pretend to keep the array mutable neither it to be executed at the moment as I would be a tensor.<\/p>\n\n<p>Maybe another way to ask this is: How would be the code for applying an array of conditional tensors of tf.cond() depending on tf.less() to an array of numbers?<\/p>\n","answers":[{"AnswerId":"35562954","CreationDate":"2016-02-22T20:15:26.580","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>You can do<\/p>\n\n<pre><code>tf.select(array &lt; 50, tf.zeros_like(array), array)\n<\/code><\/pre>\n\n<p>which will return an expression equivalent to what <code>array<\/code> will contain after <code>array[array &lt; 50] = 0<\/code>. If <code>array<\/code> was a TensorFlow variable, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#assign\" rel=\"noreferrer\"><code>tf.assign<\/code><\/a> to assign the above expression to <code>array<\/code>.<\/p>\n"}]}
{"QuestionId":35564253,"AnswerCount":1,"Tags":"<matrix><matrix-multiplication><tensorflow>","CreationDate":"2016-02-22T21:29:15.477","AcceptedAnswerId":35564506.0,"OwnerUserId":2666216.0,"Title":"tensorflow element-wise matrix multiplication","Body":"<p>Say I have two tensors in tensorflow, with the first dimension representing the index of a training example in a batch, and the others representing some vectors of matrices of data.  Eg<\/p>\n\n<pre><\/pre>\n\n<p>I'm curious what the most idiomatic way to perform a vector*matrix multiply, for each of the pairs of vectors, matrices that share an index along the first dimension.<\/p>\n\n<p>Aka a the most idiomatic way to write:<\/p>\n\n<pre><\/pre>\n\n<p>What would be the best way to organize the shape of the input vectors to make this process as simple\/clean as possible?<\/p>\n","answers":[{"AnswerId":"35564506","CreationDate":"2016-02-22T21:42:55.423","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Probably the most idiomatic way to do this is using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#batch_matmul\" rel=\"noreferrer\"><code>tf.batch_matmul()<\/code><\/a> operator (in conjunction with <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#expand_dims\" rel=\"noreferrer\"><code>tf.expand_dims()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#squeeze\" rel=\"noreferrer\"><code>tf.squeeze()<\/code><\/a>:<\/p>\n\n<pre><code>vector_batch = tf.placeholder(tf.float32, shape=[64, 50])\nmatrix_batch = tf.placeholder(tf.float32, shape=[64, 50, 50])\n\nvector_batch_as_matrices = tf.expand_dims(vector_batch, 1)\n# vector_batch_as_matrices.get_shape() ==&gt; [64, 1, 50]\n\nresult = tf.batch_matmul(vector_batch_as_matrices, matrix_batch)\n# result.get_shape() ==&gt; [64, 1, 50]\n\nresult = tf.squeeze(result, [1])\n# result.get_shape() ==&gt; [64, 50]\n<\/code><\/pre>\n"}]}
{"QuestionId":35565040,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-02-22T22:17:03.260","AcceptedAnswerId":35607545.0,"OwnerUserId":2477654.0,"Title":"Caffe : cifar 10_quick testing issue","Body":"<p>I have trained CIFAR QUICK using caffe, but when I test the cifar10_quick_iter_5000.caffemodel.h5 using a python wrapper I get an accuracy around 52-54% whereas it should be 75%. I do not understand why I am geting such a low accuracy, because when I test Lenet MNIST I get the expected accuracy as per the MNIST example in caffe website. To verify if my method is right or wrong I have tried the cifar trained model file from <a href=\"https:\/\/github.com\/IsaacRocos\/Clasificador_Cifar-10.git\" rel=\"nofollow\">Clasificador_Cifar-10<\/a> and I get and accuracy of 68%. <\/p>\n\n<p>Please let me know if I am missing something when I test the model.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35607545","CreationDate":"2016-02-24T16:25:10.770","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>See my answer <a href=\"https:\/\/stackoverflow.com\/a\/35572495\/3579977\">here<\/a>. You are not subtracting the mean which results in low accuracy. The link to the code, posted above, takes care of that. Apart from this there's nothing wrong with your approach.<\/p>\n"}]}
{"QuestionId":35565312,"AnswerCount":2,"Tags":"<python><tensorflow><convolution><sobel>","CreationDate":"2016-02-22T22:35:24.650","AcceptedAnswerId":35569977.0,"OwnerUserId":1112654.0,"Title":"Is there a convolution function in Tensorflow to apply a Sobel filter?","Body":"<p>Is there any  method in Tensorflow to apply a Sobel filter to an image  (tensor of type  and rank 2)?<\/p>\n\n<pre><\/pre>\n\n<p>I've already seen  but I can't see how to use it for this operation. Is there some way to use  to solve my problem?<\/p>\n","answers":[{"AnswerId":"50430612","CreationDate":"2018-05-20T00:18:20.050","ParentId":null,"OwnerUserId":"292896","Title":null,"Body":"<p>Tensorflow 1.8 has added tf.image.sobel_edges() so that is the easiest and probably most robust way todo this now.<\/p>\n"},{"AnswerId":"35569977","CreationDate":"2016-02-23T06:00:40.783","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Perhaps I'm missing a subtlety here, but it appears that you could apply a Sobel filter to an image using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#expand_dims\" rel=\"noreferrer\"><code>tf.expand_dims()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#conv2d\" rel=\"noreferrer\"><code>tf.nn.conv2d()<\/code><\/a>, as follows:<\/p>\n\n<pre><code>sobel_x = tf.constant([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], tf.float32)\nsobel_x_filter = tf.reshape(sobel_x, [3, 3, 1, 1])\nsobel_y_filter = tf.transpose(sobel_x_filter, [1, 0, 2, 3])\n\n# Shape = height x width.\nimage = tf.placeholder(tf.float32, shape=[None, None])\n\n# Shape = 1 x height x width x 1.\nimage_resized = tf.expand_dims(tf.expand_dims(image, 0), 3)\n\nfiltered_x = tf.nn.conv2d(image_resized, sobel_x_filter,\n                          strides=[1, 1, 1, 1], padding='SAME')\nfiltered_y = tf.nn.conv2d(image_resized, sobel_y_filter,\n                          strides=[1, 1, 1, 1], padding='SAME')\n<\/code><\/pre>\n"}]}
{"QuestionId":35567132,"AnswerCount":4,"Tags":"<histogram><tensorflow><tensorboard>","CreationDate":"2016-02-23T01:26:22.733","AcceptedAnswerId":null,"OwnerUserId":4480756.0,"Title":"Meaning of Histogram on Tensorboard","Body":"<p>I am working on Google Tensorboard, and I'm feeling confused about the meaning of Histogram Plot. I read the tutorial, but it seems unclear to me. I really appreciate if anyone could help me figure out the meaning of each axis for Tensorboard Histogram Plot.<\/p>\n\n<h3>Sample histogram from TensorBoard<\/h3>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yqmOA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yqmOA.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"35583107","CreationDate":"2016-02-23T16:38:43.910","ParentId":null,"OwnerUserId":"5952845","Title":null,"Body":"<p>Roufan,<\/p>\n\n<p>The histogram plot allows you to plot variables from your graph. <\/p>\n\n<pre><code>w1 = tf.Variable(tf.zeros([1]),name=\"a\",trainable=True)\n    tf.histogram_summary(\"firstLayerWeight\",w1)\n<\/code><\/pre>\n\n<p>For the example above the vertical axis would have the units of my w1 variable. The horizontal axis would have units of the step which I think is captured here:<\/p>\n\n<pre><code>summary_str = sess.run(summary_op, feed_dict=feed_dict)\n    summary_writer.add_summary(summary_str, **step**)\n<\/code><\/pre>\n\n<p>It may be useful to see this on how to make <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train.html#summary-operations\" rel=\"nofollow\">summaries<\/a> for the tensorboard.<\/p>\n\n<p>Don<\/p>\n"},{"AnswerId":"36229943","CreationDate":"2016-03-25T23:54:40.063","ParentId":null,"OwnerUserId":"2488162","Title":null,"Body":"<p>@marc_alain, you're a star for making such a simple script for TB, which are hard to find.<\/p>\n\n<p>To add to what he said the histograms showing 1,2,3 sigma of the <em>distribution<\/em> of weights. which is equivalent to the 68th,95th, and 98th percentiles. So think if you're model has 784 weights, the histogram shows how the values of those weights change with training.<\/p>\n\n<p>These histograms are probably not that interesting for shallow models, you could imagine that with deep networks, weights in high layers might take a while to grow because of the logistic function being saturated. Of course I'm just mindlessly parroting <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf\" rel=\"nofollow\">this paper by Glorot and Bengio<\/a>, in which they study the weights distribution through training and show how the logistic function is saturated for the higher layers for quite a while.<\/p>\n"},{"AnswerId":"35966354","CreationDate":"2016-03-13T04:08:40.343","ParentId":null,"OwnerUserId":"3937068","Title":null,"Body":"<p>I came across this question earlier, while also seeking information on how to interpret the histogram plots in TensorBoard. For me, the answer came from experiments of plotting known distributions.\nSo, the conventional normal distribution with mean = 0 and sigma = 1 can be produced in TensorFlow with the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\ncwd = \"test_logs\"\n\nW1 = tf.Variable(tf.random_normal([200, 10], stddev=1.0))\nW2 = tf.Variable(tf.random_normal([200, 10], stddev=0.13))\n\nw1_hist = tf.summary.histogram(\"weights-stdev_1.0\", W1)\nw2_hist = tf.summary.histogram(\"weights-stdev_0.13\", W2)\n\nsummary_op = tf.summary.merge_all()\n\ninit = tf.initialize_all_variables()\nsess = tf.Session()\n\nwriter = tf.summary.FileWriter(cwd, session.graph)\n\nsess.run(init)\n\nfor i in range(2):\n    writer.add_summary(sess.run(summary_op),i)\n\nwriter.flush()\nwriter.close()\nsess.close()\n<\/code><\/pre>\n\n<p>Here is what the result looks like:\n<a href=\"https:\/\/i.stack.imgur.com\/HCJSG.png\" rel=\"nofollow noreferrer\">histogram of normal distribution with 1.0 standard deviation<\/a>.\nThe horizontal axis represents time steps. \nThe plot is a contour plot and has contour lines at the vertical axis values of -1.5, -1.0, -0.5, 0.0, 0.5, 1.0, and 1.5.<\/p>\n\n<p>Since the plot represents a normal distribution with mean = 0 and sigma = 1 (and remember that sigma means standard deviation), the contour line at 0 represents the mean value of the samples.<\/p>\n\n<p>The area between the contour lines at -0.5 and +0.5 represent the area under a normal distribution curve captured within +\/- 0.5 standard deviations from the mean, suggesting that it is 38.3% of the sampling.<\/p>\n\n<p>The area between the contour lines at -1.0 and +1.0 represent the area under a normal distribution curve captured within +\/- 1.0 standard deviations from the mean, suggesting that it is 68.3% of the sampling.<\/p>\n\n<p>The area between the contour lines at -1.5 and +1-.5 represent the area under a normal distribution curve captured within +\/- 1.5 standard deviations from the mean, suggesting that it is 86.6% of the sampling.<\/p>\n\n<p>The palest region extends a little beyond +\/- 4.0 standard deviations from the mean, and only about 60 per 1,000,000 samples will be outside of this range.<\/p>\n\n<p>While Wikipedia has a very thorough explanation, you can get the most relevant nuggets <a href=\"http:\/\/www.mathsisfun.com\/data\/standard-normal-distribution.html\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Actual histogram plots will show several things. The plot regions will grow and shrink in vertical width as the variation of the monitored values increases or decreases. The plots may also shift up or down as the mean of the monitored values increases or decreases.<\/p>\n\n<p>(You may have noted that the code actually produces a second histogram with a standard deviation of 0.13. I did this to clear up any confusion between the plot contour lines and the vertical axis tick marks.)<\/p>\n"},{"AnswerId":"38481873","CreationDate":"2016-07-20T12:52:41.917","ParentId":null,"OwnerUserId":"1845941","Title":null,"Body":"<p>Each line on the chart represents a percentile in the distribution over the data: for example, the bottom line shows how the minimum value has changed over time, and the line in the middle shows how the median has changed. Reading from top to bottom, the lines have the following meaning: <code>[maximum, 93%, 84%, 69%, 50%, 31%, 16%, 7%, minimum]<\/code><\/p>\n\n<p>These percentiles can also be viewed as standard deviation boundaries on a normal distribution: <code>[maximum, \u03bc+1.5\u03c3, \u03bc+\u03c3, \u03bc+0.5\u03c3, \u03bc, \u03bc-0.5\u03c3, \u03bc-\u03c3, \u03bc-1.5\u03c3, minimum]<\/code> so that the colored regions, read from inside to outside, have widths <code>[\u03c3, 2\u03c3, 3\u03c3]<\/code> respectively.<\/p>\n"}]}
{"QuestionId":35569253,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-02-23T05:06:07.193","AcceptedAnswerId":null,"OwnerUserId":1658445.0,"Title":"Theano missing packages","Body":"<p>I recently installed Theano on my macbook pro, and tried running an example code. I get the following error<\/p>\n\n<p>ImportError: No module named theano.tensor<\/p>\n\n<p>I upgraded Theano to it's latest version, using <\/p>\n\n<p>sudo pip install --upgrade --no-deps theano<\/p>\n\n<p>and I get a message saying everything is up-to-date:<\/p>\n\n<p>Requirement already up-to-date: theano in \/anaconda\/anaconda\/lib\/python2.7\/site-packages<\/p>\n\n<p>Any idea how to proceed and what is causing the issue?<\/p>\n","answers":[]}
{"QuestionId":35575982,"AnswerCount":1,"Tags":"<concatenation><slice><tensorflow>","CreationDate":"2016-02-23T11:12:21.877","AcceptedAnswerId":null,"OwnerUserId":5967574.0,"Title":"How to slice a batch and apply an operation on each slice in TensorFlow","Body":"<p>I am a beginner with TensorFlow, I am trying to implement a function that takes a batch as input. It has to slice this batch into several ones, apply some operations on them, then concatenate them to build a new tensor to return. Through my readings, I found there are some implemented function like input_slice_producer and batch_join but I didn't get to work with them.  I attached what I've found as solution below, but it's kinda slow, not proper and incapable of detecting the current size of batch. Does any know a better way of doing this?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40981831","CreationDate":"2016-12-05T19:28:48.440","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow function <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/functional_ops.html#map_fn\" rel=\"nofollow noreferrer\"><code>tf.map_fn(fn, elems)<\/code><\/a> allows you to apply a function (<code>fn<\/code>) to each slice of a tensor (<code>elems<\/code>). For example, you could express your program as follows:<\/p>\n\n<pre><code>def model(x):\n    W_1 = tf.Variable(tf.random_normal([6, 1]), name=\"W_1\")\n\n    def fn(x_slice):\n        return tf.reduce_sum(x_slice, W_1)\n\n    return tf.map_fn(fn, x)\n<\/code><\/pre>\n\n<p>It may also be possible to implement your operation more concisely using broadcasting on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/math_ops.html#mul\" rel=\"nofollow noreferrer\"><code>tf.mul()<\/code><\/a> operator, which uses <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">NumPy broadcasting semantics<\/a>, and the <code>axis<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/math_ops.html#reduce_sum\" rel=\"nofollow noreferrer\"><code>tf.reduce_sum()<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":35576364,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-23T11:29:20.790","AcceptedAnswerId":35577521.0,"OwnerUserId":681714.0,"Title":"Avoid cluttering the tensorflow graph with assign operations","Body":"<p>I have to run something like the following code<\/p>\n\n<pre><\/pre>\n\n<p>several times. The actual code is much more complicated and uses more variables.\nThe problem is that the TensorFlow graph grows with each instantiated assign op, which makes the graph grow, eventually slowing down the computation.<\/p>\n\n<p>I could use  to set the value, but I would like to keep my state in the graph, so that I can easily query it in other places.<\/p>\n\n<p>Is there some way of avoiding cluttering the current graph in this case?<\/p>\n","answers":[{"AnswerId":"37027233","CreationDate":"2016-05-04T12:02:38.127","ParentId":null,"OwnerUserId":"6290635","Title":null,"Body":"<p>Each time you call <code>sess.run(x.assign(42.))<\/code>\ntwo things happen: (i) a new <code>assign<\/code> operation is added to the computational graph <code>sess.graph<\/code>, (ii) the newly added operation executes. No wonder the graph gets pretty large if loop repeats many times. If you define assignment operation before execution (<code>asgnmnt_operation<\/code> in example below), just a single operation is added to the graph so the performance is great: <\/p>\n\n<pre><code>import tensorflow as tf\n\n\nx = tf.Variable(42.)\nc = tf.constant(42.)\nasgnmnt_operation = x.assign(c) \nsess = tf.Session()\n\nfor i in range(10000):\n    sess.run(asgnmnt_operation) \n    sess.run(x)\n    print(i)\n<\/code><\/pre>\n"},{"AnswerId":"35577521","CreationDate":"2016-02-23T12:24:51.823","ParentId":null,"OwnerUserId":"681714","Title":null,"Body":"<p>I think I've found a good solution for this:\nI define a placeholder <code>y<\/code> and create an op that assigns the value of <code>y<\/code> to <code>x<\/code>.\nI can then use that op repeatedly, using <code>feed_dict={y: value}<\/code> to assign a new value to x.\nThis doesn't add another op to the graph.\nIt turns out that the loop runs <em>much<\/em> more quickly than before as well.<\/p>\n\n<pre><code>import tensorflow as tf\n\nsess = tf.Session()\n\nx = tf.Variable(42.)\ny = tf.placeholder(dtype=tf.float32)\n\nassign = x.assign(y)\n\nsess.run(tf.initialize_all_variables())\n\nfor i in range(10000):\n    sess.run(assign, feed_dict={y: i})\n    print(i, sess.run(x))\n<\/code><\/pre>\n"}]}
{"QuestionId":35578253,"AnswerCount":1,"Tags":"<lua><training-data><torch><conv-neural-network>","CreationDate":"2016-02-23T12:57:50.273","AcceptedAnswerId":null,"OwnerUserId":2452617.0,"Title":"torch7: Setting Variable Learning Rates for Different Conv-net Layers","Body":"<p>I am trying to fine-tune a conv-net. It has the following structure (adapted from <strong>OverFeat<\/strong>):<\/p>\n\n<pre><\/pre>\n\n<p>And I'm using <strong>SGD<\/strong> as the optimization method with the following parameters:<\/p>\n\n<pre><\/pre>\n\n<p>I am training it as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where:<\/p>\n\n<pre><\/pre>\n\n<p>From my references, I have learned that while fine-tuning a pre-trained network, it is recommended that the lower (convolutional) layers should have lower learning rates and the higher layers should have relatively higher learning rates.<\/p>\n\n<p>I referred to torch7's documentation of <a href=\"https:\/\/github.com\/torch\/optim\/blob\/master\/sgd.lua\" rel=\"nofollow\">optim\/sgd<\/a> to set different learning rates for each layer. From there, I get that setting <strong>config.learningRates<\/strong> i.e. a vector of individual learning rates, I can achieve what I want. I am new to Torch, so, please pardon me if this seems as a silly question, but it would be really helpful if someone can please explain me how and where to create\/use this vector to serve my purpose?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42403166","CreationDate":"2017-02-22T22:04:49.763","ParentId":null,"OwnerUserId":"1851402","Title":null,"Body":"<p>I don't know if you still need an answer, as you posted this question one year ago.<\/p>\n\n<p>Anyway, just in case someone sees this, I've written a <a href=\"http:\/\/www.thedataware.com\/post\/the-torch-adventures-setting-layer-wise-training-parameters\" rel=\"nofollow noreferrer\">post here<\/a> about how to set different learning rates for different layers in torch.<\/p>\n\n<p>The solution is to use <code>net:parameters()<\/code> instead of <code>net:getParameters()<\/code>. Instead of returning two tensors, it returns two tables of tensors, containing the parameters (and the gradParameters) for each layer in separate tensors.<\/p>\n\n<p>In this way, you can run an <code>sgd()<\/code> step (with a different learning rate) for each layer. You can find the full code by clicking the above link.<\/p>\n"}]}
{"QuestionId":35582746,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard><tensorflow-serving>","CreationDate":"2016-02-23T16:23:25.210","AcceptedAnswerId":null,"OwnerUserId":5952845.0,"Title":"Attribute Error from tensorboard.py following v0.7.1 upgrade","Body":"<p>I just ran at my home directory:<\/p>\n\n<pre><\/pre>\n\n<p>Now when I try to run tensorboard using<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Has anyone else had v0.7.1 mess up tensorboard? Also am I missing a step on the install?<\/p>\n","answers":[]}
{"QuestionId":35582858,"AnswerCount":1,"Tags":"<tensorflow><word2vec>","CreationDate":"2016-02-23T16:28:26.343","AcceptedAnswerId":null,"OwnerUserId":5969670.0,"Title":"how can I get final embeddings from word2vec.py?","Body":"<p>I am using word2vec.py from Google, see the link below:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/models\/embedding\/word2vec.py\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/models\/embedding\/word2vec.py<\/a><\/p>\n\n<p>I want to extract the final embeddings of words and save it to a txt file. How can I find the final embeddings?<\/p>\n","answers":[{"AnswerId":"35611003","CreationDate":"2016-02-24T19:13:28.053","ParentId":null,"OwnerUserId":"5969670","Title":null,"Body":"<p>I figured out how to do it. <\/p>\n\n<p>I created a self._nemb_final in build_eval_graph model to transfer the nemb out.\nAfter the model is trained, we can then do sth. like below:<\/p>\n\n<pre><code>final_embedding = self._nemb_final.eval()     \nnp.savetxt('\/...\/final_embedding_dic.txt',final_embedding)\n<\/code><\/pre>\n"}]}
{"QuestionId":35582875,"AnswerCount":2,"Tags":"<docker><tensorflow>","CreationDate":"2016-02-23T16:28:59.210","AcceptedAnswerId":35605106.0,"OwnerUserId":5786837.0,"Title":"Unable to start TensorFlow within Docker, on Windows","Body":"<p>Hope I didn't miss anything.<br>\nI've installed docker on my win 7 using this guide:<br>\n<a href=\"https:\/\/docs.docker.com\/engine\/installation\/\">https:\/\/docs.docker.com\/engine\/installation\/<\/a><br>\nI opened a new terminal and entered the following command:  <\/p>\n\n<pre><\/pre>\n\n<p>All donwloaded and extracted and then I get the following massages:<\/p>\n\n<pre><\/pre>\n\n<p>And then it just gets stuck like this, there's no command line and I can't enter anything... what am I missing?<\/p>\n","answers":[{"AnswerId":"36873007","CreationDate":"2016-04-26T18:30:32.613","ParentId":null,"OwnerUserId":"2993768","Title":null,"Body":"<p>I had the same problem and was able to get it working by following these steps:<\/p>\n\n<pre><code>$ docker-machine ip default\n<\/code><\/pre>\n\n<p>Remember this <code>DOCKER_IP<\/code> value (copy to clipboard) in my case <\/p>\n\n<pre><code>192.168.99.100\n<\/code><\/pre>\n\n<p>Now start your TensorFlow docker container (with port forwarding):<\/p>\n\n<pre><code>$ docker run -it -p 8888:8888 gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p>Now open the web browser:<\/p>\n\n<pre><code>$ open http:\/\/localhost:8888\n<\/code><\/pre>\n\n<p>You should now see your browser with the jupyter home page<\/p>\n\n<p>I'm working on more notes on <a href=\"https:\/\/github.com\/stujo\/tensorflow-what-is-it\" rel=\"nofollow\">Getting started wtih TensorFlow here<\/a> some of that is OSX specific though<\/p>\n\n<p>Update: I have a better understanding now so I'm updating the answer - A docker image makes certain ports available (EXPORTable) for mapping, but by default does not map them to the host machines ports when the container is run. <\/p>\n\n<p>We can map them to the host by using the -p option. We specify which host port (if any) the already EXPORTed port should be mapped to on the host. <\/p>\n\n<pre><code>$ docker run -p $HOSTPORT:$CONTAINERPORT someimage\n<\/code><\/pre>\n"},{"AnswerId":"35605106","CreationDate":"2016-02-24T14:39:36.723","ParentId":null,"OwnerUserId":"5786837","Title":null,"Body":"<p>Ok, So i found a sort of an answer,<br>\nThere are two ways to solve it:<br>\n1) Install tensorflow with source code instead, this seems to solve the problem.\nThis is done by writing:\n<code>docker run -it b.gcr.io\/tensorflow\/tensorflow:latest-devel<\/code><br>\n2) Or, if you use the regular install, before installing check\nthe default VM IP with:<\/p>\n\n<pre><code>docker-machine ip default\n<\/code><\/pre>\n\n<p>And then, after installtion go in the brwoser to <a href=\"http:\/\/(default_ip):8888\/\" rel=\"noreferrer\">http:\/\/(default_ip):8888\/<\/a><\/p>\n"}]}
{"QuestionId":35585053,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-23T18:10:11.387","AcceptedAnswerId":null,"OwnerUserId":2615676.0,"Title":"Restoring from checkpoint when variable shape changes","Body":"<p>I'm unable to restore checkpointed models that include variables that change shape. For example with this simple model:<\/p>\n\n<pre><\/pre>\n\n<p>if I run  and then save the model, when I try to restore it I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>which seems to have to do with the changing shape and  trying to validate shape. If I set  to  when constructing the , which according to the docs should solve this problem, I instead get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm inclined to think that this is a bug.<\/p>\n","answers":[{"AnswerId":"47386431","CreationDate":"2017-11-20T06:49:27.373","ParentId":null,"OwnerUserId":"4862950","Title":null,"Body":"<p>Set <code>validate_shape<\/code> to False when loading meta graph by adding code<\/p>\n\n<pre><code>if graph.node[-1].attr.get(\"validate_shape\"):\n    graph.node[-1].attr[\"validate_shape\"].b = False\n<\/code><\/pre>\n\n<p>to tensorflow\/python\/framework\/ops.py#2318<\/p>\n\n<pre><code>with self._lock:\n  graph = graph_pb2.GraphDef()\n  graph.versions.CopyFrom(self._graph_def_versions)\n  bytesize = 0\n  for op_id in sorted(self._nodes_by_id):\n    op = self._nodes_by_id[op_id]\n    if from_version is None or op_id &gt; from_version:\n      graph.node.extend([op.node_def])\n      if graph.node[-1].attr.get(\"validate_shape\"):\n        graph.node[-1].attr[\"validate_shape\"].b = False\n      if op.outputs and add_shapes:\n        assert \"_output_shapes\" not in graph.node[-1].attr\n        graph.node[-1].attr[\"_output_shapes\"].list.shape.extend([\n            output.get_shape().as_proto() for output in op.outputs])\n      bytesize += op.node_def.ByteSize()\n      if bytesize &gt;= (1 &lt;&lt; 31) or bytesize &lt; 0:\n        raise ValueError(\"GraphDef cannot be larger than 2GB.\")\n<\/code><\/pre>\n"},{"AnswerId":"35875511","CreationDate":"2016-03-08T18:47:10.520","ParentId":null,"OwnerUserId":"16480","Title":null,"Body":"<p>The reshape option to <code>Saver<\/code> only works if the shapes have the same total number of elements.  For example, it will let you load a variable with shape <code>[1]<\/code> from data with shape <code>[]<\/code>, or a variable with shape <code>[15, 7]<\/code> from data with shape <code>[5, 21]<\/code>.  If the shapes aren't compatible in this way, you'll have to build a new graph.<\/p>\n"}]}
{"QuestionId":35585915,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-23T18:59:21.883","AcceptedAnswerId":null,"OwnerUserId":5175089.0,"Title":"Does Tensorflow have optimizers for compute graphs involving complex64 tensors?","Body":"<p>The tf.float32 version of the following code works. However, when we try to run the following, we get an exception  (on the line where we try to define an optimizer).<\/p>\n\n<pre><\/pre>\n\n<p>Here is the error message from the top of the stack: <\/p>\n\n<pre><\/pre>\n\n<p>Does this mean that complex numbers are not supported as intermediate steps when trying to do optimization? Short of manually encoding complex number operations, are there any workarounds?<\/p>\n","answers":[{"AnswerId":"35593874","CreationDate":"2016-02-24T05:21:16.010","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Yes, you hit intermediate op that didn't have complex support. See this <a href=\"https:\/\/stackoverflow.com\/questions\/35443080\/tensorflow-critical-graph-operations-assigned-to-cpu-rather-than-gpu\">question<\/a> for some pointers to code where you can see which data types are registered for which devices. In general I think complex support is fairly slim, most ops don't support it.<\/p>\n"}]}
{"QuestionId":35586752,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2016-02-23T19:46:00.180","AcceptedAnswerId":35611879.0,"OwnerUserId":2628128.0,"Title":"Passing compilation flags to bazel (TensorFlow)","Body":"<p>I am compiling TensorFlow from sources on an ARM7-equipped SoC, thus it is a non-standard compilation. Bazel fails in many cases but it is easy to fix if I could pass it some additional compilation flags: for example<\/p>\n\n<pre><\/pre>\n\n<p>How can I pass those extra flags to Bazel?<\/p>\n","answers":[{"AnswerId":"35611879","CreationDate":"2016-02-24T20:00:15.590","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p><a href=\"http:\/\/bazel.io\/docs\/bazel-user-manual.html#flag--copt\" rel=\"nofollow\">--copt<\/a> is the way to set C flags from the command line but those flag should be set already by bazel when building tensorflow.<\/p>\n\n<p>What is the output of <code>bazel build -s<\/code>? Can you see the includes of gcc? (Maybe we should continue the discussion on the github issue tracker).<\/p>\n"}]}
{"QuestionId":35591081,"AnswerCount":1,"Tags":"<tensorflow><gradient-descent>","CreationDate":"2016-02-24T00:33:22.340","AcceptedAnswerId":null,"OwnerUserId":5813490.0,"Title":"FailedPreconditionError while trying to use RMSPropOptimizer on tensorflow","Body":"<p>I am trying to use the RMSPropOptimizer for minimizing loss. Here's the part of the code that is relevant:<\/p>\n\n<pre><\/pre>\n\n<p>Problem is as soon as I run this I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Note that I don't get this error If am using the usual GradientDescentOptimizer. I am initializing my variables as you can see above but I don't know what 'train\/output\/bias\/RMSProp' is because I don't create any such variable. I only have 'train\/output\/bias\/' which does get initialized above.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36294634","CreationDate":"2016-03-29T20:38:30.113","ParentId":null,"OwnerUserId":"5813490","Title":null,"Body":"<p>So for people from the future running into similar trouble, I found this post helpful:\n<a href=\"https:\/\/stackoverflow.com\/questions\/33788989\/tensorflow-using-adam-optimizer\">Tensorflow: Using Adam optimizer<\/a><\/p>\n\n<p>Basically, I was running <\/p>\n\n<pre><code>sess.run(tf.initialize_all_variables()) \n<\/code><\/pre>\n\n<p>before I had defined my loss minimization op<\/p>\n\n<pre><code>loss = tf.square(targets)\n#create the gradient descent op\ngrads_and_vars = opt.compute_gradients(loss)\ncapped_grads_and_vars = [(tf.clip_by_value(g, -self.clip_delta, self.clip_delta), v) for g, v in grads_and_vars]    #gradient capping\nself.opt_op = self.opt.apply_gradients(capped_grads_and_vars)\n<\/code><\/pre>\n\n<p>This needs to be done <em>before<\/em> running the initialization op!<\/p>\n"}]}
{"QuestionId":35592867,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><smartphone>","CreationDate":"2016-02-24T03:46:47.163","AcceptedAnswerId":35601102.0,"OwnerUserId":4480180.0,"Title":"Possible use case\/real application for mobile distributed version of Tensorflow?","Body":"<p>I'm developing this project where I'm trying to create a distributed version of <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"nofollow\">Tensorflow<\/a> (the actual open source version is single node) and where the cluster is entirely composed by mobile devices (e.g. smartphones).<\/p>\n\n<p>In your opinion, what is a possible application or use case where this could be useful? Can you give me some example please?<\/p>\n\n<p>I know that this is not a \"standard\" Stack Overflow question, but I didn't know where to post it (if you know a better place where to post it, please let me know it). Thanks so much for your help!<\/p>\n","answers":[{"AnswerId":"35601102","CreationDate":"2016-02-24T11:36:40.827","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p><a href=\"http:\/\/www.google.com.hk\/search?q=teonsoflow+android\" rel=\"nofollow\">http:\/\/www.google.com.hk\/search?q=teonsoflow+android<\/a><\/p>\n\n<p>TensorFlow can be used for image identification and there is an example using the camera for Android.<\/p>\n\n<p>There could be many distributed uses for this. Face recognition, 3D space construction from 2D images.<\/p>\n\n<p>TensorFlow can be used for a chat bot. I am working towards using it for a personal assistant. The Ai on one phone could communicate with the Ai on other phones.<\/p>\n\n<p>It could use vision and GPS to 'reserve' a lane for you on the road. Intelligent crowd planned roads and intersections would be safer.<\/p>\n\n<p>I am also interested in using it for distributed mobile. Please contact me with my user name at gmail or Skype.<\/p>\n"},{"AnswerId":"35601706","CreationDate":"2016-02-24T12:04:29.137","ParentId":null,"OwnerUserId":"594353","Title":null,"Body":"<p><a href=\"https:\/\/boinc.berkeley.edu\" rel=\"nofollow\">https:\/\/boinc.berkeley.edu<\/a><\/p>\n\n<p>I think all my answers could run on individual phones with communication between them. If you want them to act like a cluster as @Yaroslav pointed out there is Seti@home and other projects running in the BOINC client.<\/p>\n\n<p>TensorFlow could be combined with a game engine. You could have a proceduraly generated Ai learning augumented reality game generating the story as multiple players interact with it. I have seen research papers for each of these components.<\/p>\n"}]}
{"QuestionId":35596136,"AnswerCount":1,"Tags":"<python><named-pipes><theano>","CreationDate":"2016-02-24T07:45:21.930","AcceptedAnswerId":null,"OwnerUserId":2539074.0,"Title":"Is it possible to use named-pipes in Theano?","Body":"<p>Does  support  [for both reading\\writing]? What about reading\\writing files in general?<\/p>\n","answers":[{"AnswerId":"35601031","CreationDate":"2016-02-24T11:33:07.647","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>If you're familiar with how to to do the I\/O using Python and\/or numpy then that's all you need. Just pass the numpy data to the compiled Theano function (or first store the numpy array in a shared variable if that's what the Theano computation expects).<\/p>\n\n<p>Theano itself doesn't do any I\/O. Inputs to compiled Theano functions should be numpy arrays (or castable to numpy arrays) and the outputs of a compiled Theano function will also be numpy arrays. How you read the input arrays into memory, or write the output arrays elsewhere, is none of Theano's concern, just use Python and\/or numpy as normal.<\/p>\n\n<p>There's an example using named pipes in Python <a href=\"http:\/\/www.roman10.net\/named-pipe-in-linux-with-a-python-example\/\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":35596468,"AnswerCount":1,"Tags":"<cmake><caffe>","CreationDate":"2016-02-24T08:04:46.443","AcceptedAnswerId":null,"OwnerUserId":1309735.0,"Title":"Linking Shared library (which has dependency on other shared library) in CMake","Body":"<p>I am trying to build an executable which links to a shared library (named 'caffe'). The shared library is dependent on another shared library (named 'cblas'). When I try to link to caffe in my CMake file it shows the following error:<\/p>\n\n<blockquote>\n  <p>libcblas.so.3, needed by libcaffe.so, not found (try using -rpath or\n  -rpath-link)<\/p>\n<\/blockquote>\n\n<p>I am using the following statements in my CMakeLists.txt:<\/p>\n\n<pre><\/pre>\n\n<p>Both 'cblas' and 'caffe' libraries are present in  folder.<\/p>\n\n<p>Do I need to add cblas.so to  also? Also, i am not building caffe.so so building it via CMake and keeping it as a dependency is not an option<\/p>\n\n<p>Is there any other feasible solution for the same problem where dependency tree of shared library needs to be resolved while linking?<\/p>\n","answers":[{"AnswerId":"35596820","CreationDate":"2016-02-24T08:23:24.090","ParentId":null,"OwnerUserId":"1782465","Title":null,"Body":"<p>Browsing through the library's GitHub tree, it seems to me that it provides a package config file. Therefore, if you have installed it in the normal way, you should be able to find it as a package, instead of defining the imported target yourself:<\/p>\n\n<pre><code>find_package(Caffe)\ninclude_directories(${Caffe_INCLUDE_DIRS})\nadd_definitions(${Caffe_DEFINITIONS})    # ex. -DCPU_ONLY\n\nadd_executable(caffeinated_application main.cpp)\ntarget_link_libraries(caffeinated_application ${Caffe_LIBRARIES})\n<\/code><\/pre>\n\n<p>The example above comes from the <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1667\" rel=\"nofollow\">Caffe documentation<\/a> on the topic.<\/p>\n"}]}
{"QuestionId":35596629,"AnswerCount":4,"Tags":"<tensorflow><int64>","CreationDate":"2016-02-24T08:13:17.963","AcceptedAnswerId":35596680.0,"OwnerUserId":5948325.0,"Title":"How to convert tf.int64 to tf.float32?","Body":"<p>I tried:<\/p>\n\n<pre><\/pre>\n\n<p>Then following error appears:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35596680","CreationDate":"2016-02-24T08:15:46.683","ParentId":null,"OwnerUserId":"5948325","Title":null,"Body":"<p>Oops, I find the function in the API...<\/p>\n\n<pre><code> tf.to_float(x, name='ToFloat')\n<\/code><\/pre>\n"},{"AnswerId":"39812549","CreationDate":"2016-10-01T23:44:58.033","ParentId":null,"OwnerUserId":"4561314","Title":null,"Body":"<p>You can use either <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#cast\" rel=\"noreferrer\"><code>tf.cast(x, tf.float32)<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops.html#to_float\" rel=\"noreferrer\"><code>tf.to_float(x)<\/code><\/a>, both of which cast to float32.<\/p>\n\n<p>Example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>sess = tf.Session()\n\n# Create an integer tensor.\ntensor = tf.convert_to_tensor(np.array([0, 1, 2, 3, 4]), dtype=tf.int64)\nsess.run(tensor)\n# array([0, 1, 2, 3, 4])\n\n# Use tf.cast()\ntensor_float = tf.cast(tensor, tf.float32)\nsess.run(tensor_float)\n# array([ 0.,  1.,  2.,  3.,  4.], dtype=float32)\n\n# Use tf.to_float() to cast to float32\ntensor_float = tf.to_float(tensor)\nsess.run(tensor_float)\n# array([ 0.,  1.,  2.,  3.,  4.], dtype=float32)\n<\/code><\/pre>\n"},{"AnswerId":"37809635","CreationDate":"2016-06-14T10:38:23.333","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>You can cast generally using:<\/p>\n\n<pre><code>tf.cast(my_tensor, tf.float32)\n<\/code><\/pre>\n\n<p>Replace tf.float32 with your desired type.<\/p>\n\n<hr>\n\n<p><strong>Edit<\/strong>: It seems at the moment at least, that <code>tf.cast<\/code> won't cast to an unsigned dtype (e.g. <code>tf.uint8<\/code>). To work around this, you can cast to the signed equivalent and used <code>tf.bitcast<\/code> to get all the way. e.g.<\/p>\n\n<pre><code>tf.bitcast(tf.cast(my_tensor, tf.int8), tf.uint8)\n<\/code><\/pre>\n"},{"AnswerId":"51605200","CreationDate":"2018-07-31T03:50:29.920","ParentId":null,"OwnerUserId":"3331276","Title":null,"Body":"<p><code>image<\/code>type cast you can use <code>tf.image.convert_image_dtype()<\/code> which convert image range <code>[0 255]<\/code> to <code>[0 1]<\/code>:<\/p>\n\n<pre><code>img_uint8 = tf.constant([1,2,3], dtype=tf.uint8)\nimg_float = tf.image.convert_image_dtype(img_uint8, dtype=tf.float32)\nwith tf.Session() as sess:\n    _img= sess.run([img_float])\n    print(_img, _img.dtype)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[0.00392157 0.00784314 0.01176471] float32\n<\/code><\/pre>\n\n<p>if you only want to cast type and keep value range use <code>tf.cast<\/code> or <code>tf.to_float<\/code> as @stackoverflowuser2010 and @Mark McDonald answered<\/p>\n"}]}
{"QuestionId":35596795,"AnswerCount":1,"Tags":"<machine-learning><neural-network><opencl><theano>","CreationDate":"2016-02-24T08:22:03.950","AcceptedAnswerId":35597443.0,"OwnerUserId":1256369.0,"Title":"OpenCL Theano - How to forcefully disable CUDA?","Body":"<p>After a series of pains, I have installed Theano on a machine with AMD graphics card - .<\/p>\n\n<p>Now, consider a following code. <\/p>\n\n<pre><\/pre>\n\n<p>I think this code should work just fine. But I get a series of errors: <\/p>\n\n<pre><\/pre>\n\n<p>And I get the same set of messages multiple times. Then at the end:<\/p>\n\n<pre><\/pre>\n\n<p>Does this mean I cannot use this GPU or I have done something wrong in my code. Moreover, from the errors, it seems there is been a search for . But I do not have CUDA, I have opencl. <\/p>\n\n<pre><\/pre>\n\n<p>also:<\/p>\n\n<pre><\/pre>\n\n<p>So how do I go from here? Is there way to make sure  is searched instead of .<\/p>\n\n<p>PS_1: hello world works.\nPS_2: System = 14.04 64 bit<\/p>\n","answers":[{"AnswerId":"35597443","CreationDate":"2016-02-24T08:55:50.710","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>OpenCL is not yet supported by Theano. As a result, only NVIDIA GPUs are supported.<\/p>\n\n<p>The status of OpenCL is <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/2936\" rel=\"noreferrer\">recorded on GitHub<\/a>.<\/p>\n\n<p>You need to disable GPU operation by setting <code>device=cpu<\/code> in your Theano config. There are multiple ways to do this (i.e. via <code>THEANO_FLAGS<\/code> environment variable or via a <code>.theanorc<\/code> file; <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"noreferrer\">see documentation<\/a>).<\/p>\n\n<p>Before running the script, try setting<\/p>\n\n<pre><code>export THEANO_FLAGS=device=cpu,floatX=float64\n<\/code><\/pre>\n\n<p>Your situation may need additional configuration options. See <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"noreferrer\">the documentation<\/a> for more.<\/p>\n"}]}
{"QuestionId":35599179,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-24T10:14:12.823","AcceptedAnswerId":35606830.0,"OwnerUserId":234167.0,"Title":"How to perform subtraction on a single element of a tensor","Body":"<p>I have a tensor that consists of 4 floats, called .<\/p>\n\n<p>How do I with a 50% chance execute ?<\/p>\n\n<p>Right now I have:<\/p>\n\n<pre><\/pre>\n\n<p>which works, but looks extremely ugly. Isn't there a smarter\/simpler way of doing this?<\/p>\n\n<p>Related question: How do I apply a certain op (e.g. sqrt) just to two elements? I'm guessing I have to remove these two elements, perform the op and then concat them back to the original vector?<\/p>\n","answers":[{"AnswerId":"35606830","CreationDate":"2016-02-24T15:55:41.173","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/control_flow_ops.html#select\" rel=\"nofollow\"><code>tf.select<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/control_flow_ops.html#cond\" rel=\"nofollow\"><code>tf.cond<\/code><\/a> come in handy for situations where you have to perform computations conditionally on elements of a tensor. For your example, the following would work :<\/p>\n\n<pre><code>label = tf.constant([0.35, 0.5, 0.17, 0.14])\ninv = tf.pack([1.0, 0.0, 0.0, 0.0])\nmask = tf.pack([1.0, -1.0, -1.0, -1.0])\noutput = tf.cond(tf.random_uniform([], 0, 1.0) &gt; 0.5,\n                 lambda: label,\n                 lambda: (inv - label) * mask)\nwith tf.Session(''):\n  print(output.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":35599328,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2016-02-24T10:21:16.493","AcceptedAnswerId":null,"OwnerUserId":3165370.0,"Title":"Keras - unable to reduce loss between epochs","Body":"<p>I am training a VGG-like convnet (like in the example <a href=\"http:\/\/keras.io\/examples\/\" rel=\"noreferrer\">http:\/\/keras.io\/examples\/<\/a>) with a set of images. I convert images to arrays and resize them using scipy:<\/p>\n\n<pre><\/pre>\n\n<p>In the last dense layer I have 2 neurons and I activate with softmax. Here are the last lines:<\/p>\n\n<pre><\/pre>\n\n<p>I am not able to improve reduce the loss and every epoch has the same loss and the same precision as the one before. The loss actually goes up between 1st and 2nd epoch:<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"37482373","CreationDate":"2016-05-27T11:30:34.090","ParentId":null,"OwnerUserId":"5056575","Title":null,"Body":"<p>My suggestions would be to reduce the learning rate, try data augmentation.<\/p>\n\n<p>Data augmentation code: <\/p>\n\n<pre><code>print('Using real-time data augmentation.')\n\n    # this will do preprocessing and realtime data augmentation\n     datagen = ImageDataGenerator(\n        featurewise_center=False,  # set input mean to 0 over the dataset\n        samplewise_center=False,  # set each sample mean to 0\n        featurewise_std_normalization=False,  # divide inputs by std of the dataset\n        samplewise_std_normalization=False,  # divide each input by its std\n        zca_whitening=True,  # apply ZCA whitening\n        rotation_range=90,  # randomly rotate images in the range (degrees, 0 to 180)\n        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)\n        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)\n        horizontal_flip=True,  # randomly flip images\n        vertical_flip=False)  # randomly flip images\n\n    # compute quantities required for featurewise normalization\n    # (std, mean, and principal components if ZCA whitening is applied)\n     datagen.fit(X_train)\n\n    # fit the model on the batches generated by datagen.flow()\n     model.fit_generator(datagen.flow(X_train, Y_train,\n                         batch_size=batch_size),\n                         samples_per_epoch=X_train.shape[0],\n                         nb_epoch=nb_epoch)\n<\/code><\/pre>\n"},{"AnswerId":"37476487","CreationDate":"2016-05-27T06:39:27.913","ParentId":null,"OwnerUserId":"3875760","Title":null,"Body":"<p>From my experience, this often happens when the learning rate is too high.\nThe optimization will be unable to find a minima and just \"turn around\".<\/p>\n\n<p>The ideal rate will depend on your data and on the architecture of your network. <\/p>\n\n<p>(As a reference, I'm at the moment running a convnet with 8 layers, on a sample size similar to yours, and the same lack of convergence could be observed until I reduced the learning rate to 0.001)<\/p>\n"}]}
{"QuestionId":35605551,"AnswerCount":2,"Tags":"<python><centos7><caffe><pycaffe>","CreationDate":"2016-02-24T14:57:42.253","AcceptedAnswerId":null,"OwnerUserId":1922589.0,"Title":"relocation R_X86_64_32S against `_Py_NotImplementedStruct' can not be used when making a shared object; recompile with -fPIC","Body":"<p>I got this error message after run it on Centos 7 (512MB).<\/p>\n\n<pre><\/pre>\n\n<p>or <\/p>\n\n<pre><\/pre>\n\n<p>Error message<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42292073","CreationDate":"2017-02-17T07:37:46.680","ParentId":null,"OwnerUserId":"5242412","Title":null,"Body":"<p>I had recently encountered a similar error with another software package. <\/p>\n\n<p>Try passing <code>CFLAGS=\"-fPIC\"<\/code> or <code>CXXFLAGS=\"-fPIC\"<\/code> while configuring or building. That worked for me.<\/p>\n\n<p>You might also want to go through <a href=\"https:\/\/web.archive.org\/web\/20150319085301\/http:\/\/www.gentoo.org\/proj\/en\/base\/amd64\/howtos\/index.xml?part=1&amp;chap=3\" rel=\"nofollow noreferrer\">this page on <code>-fPIC<\/code> errors<\/a>:<\/p>\n"},{"AnswerId":"42582759","CreationDate":"2017-03-03T15:29:46.830","ParentId":null,"OwnerUserId":"784210","Title":null,"Body":"<p>I had the same problem with trying to install dlib. The solution was to build Python with the <code>--enable-shared<\/code> flag. When using <code>pyenv<\/code> it can be done like this:<\/p>\n\n<p><code>PYTHON_CONFIGURE_OPTS=\"--enable-shared\" pyenv install --force 2.7.11<\/code><\/p>\n"}]}
{"QuestionId":35607270,"AnswerCount":0,"Tags":"<python><csv><tensorflow>","CreationDate":"2016-02-24T16:14:00.683","AcceptedAnswerId":null,"OwnerUserId":5975539.0,"Title":"TensorFlow CSV issues: Unquoted fields cannot have quotes\/CRLFs inside","Body":"<p>I'm trying to use the TensorFlow Python API to train a multilayer perceptron on a numeric dataset with 119 features.\nLooking at the documentation, it seems more geared towards reading text files: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#file-formats\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#file-formats<\/a><\/p>\n\n<p>I get the following error:<\/p>\n\n<p>InvalidArgumentError: Unquoted fields cannot have quotes\/CRLFs inside\n     [[Node: DecodeCSV_15 = DecodeCSV[OUT_TYPE=[DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32], field_delim=\",\", _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](ReaderRead_11:1, DecodeCSV_15\/record_defaults_0, DecodeCSV_15\/record_defaults_1, DecodeCSV_15\/record_defaults_2, DecodeCSV_15\/record_defaults_3, DecodeCSV_15\/record_defaults_4, DecodeCSV_15\/record_defaults_5, DecodeCSV_15\/record_defaults_6)]]<\/p>\n\n<p>The relevant part of my code is:<\/p>\n\n<pre><\/pre>\n\n<p>I'm also conscious that even if I get this working, I'll have to do the same for 119 columns, which is a bit impractical, and I'm wondering if it's overkill to use TensorFlow for this task in the first place.<\/p>\n","answers":[]}
{"QuestionId":35610868,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-02-24T19:06:09.553","AcceptedAnswerId":null,"OwnerUserId":5976355.0,"Title":"Tensorflow code breaking","Body":"<p>I am modifying cifar multi GPU tensorflow code to read the Imagenet dataset.<\/p>\n\n<p>The edits that I made are:<\/p>\n\n<p><strong>Cifar10.py:<\/strong><\/p>\n\n<p>1) Changed tf.app.flags.DEFINE_string('data_dir',...)<\/p>\n\n<p>2) Removed the later part in data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin')<\/p>\n\n<p>3) Removed the download part from maybe_download_and_extract()<\/p>\n\n<p><strong>cifar10_input.py:<\/strong><\/p>\n\n<p>1) IMAGE SIZE = 227<\/p>\n\n<p>2) result.height = 256 and result.width = 256<\/p>\n\n<p>3) Changed<\/p>\n\n<pre><\/pre>\n\n<p>to <\/p>\n\n<pre><\/pre>\n\n<p>But this is throwing an ugly error:\ntensorflow.python.framework.errors.OutOfRangeError: RandomShuffleQueue '_1_tower_0\/shuffle_batch\/random_shuffle_queue' is closed and has insufficient elements (requested 128, current size 0)<\/p>\n\n<p>[[Node: tower_0\/shuffle_batch = QueueDequeueMany[component_types=[DT_FLOAT, DT_INT32], timeout_ms=-1, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](tower_0\/shuffle_batch\/random_shuffle_queue, tower_0\/shuffle_batch\/n\/_775)]]<\/p>\n\n<p>[[Node: tower_1\/shuffle_batch\/n\/_664 = _HostSendT=DT_INT32, client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/gpu:1\", send_device_incarnation=1, tensor_name=\"edge_170_tower_1\/shuffle_batch\/n\", _device=\"\/job:localhost\/replica:0\/task:0\/gpu:1\"]]\nCaused by op u'tower_0\/shuffle_batch', defined at:<\/p>\n\n<p>File \"lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_multi-gpu_train.py\", line 224, in <\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/platform\/default\/_app.py\", line 30, in run<\/p>\n\n<pre><\/pre>\n\n<p>File \"lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_multi-gpu_train.py\", line 222, in main<\/p>\n\n<pre><\/pre>\n\n<p>File \"lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_multi-gpu_train.py\", line 150, in train<\/p>\n\n<pre><\/pre>\n\n<p>File \"lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_multi-gpu_train.py\", line 65, in tower_loss<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10.py\", line 119, in distorted_inputs<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_input.py\", line 153, in distorted_inputs<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/models\/image\/cifar10\/cifar10_input.py\", line 104, in _generate_image_and_label_batch<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/input.py\", line 496, in shuffle_batch\n    return queue.dequeue_many(batch_size, name=name)<\/p>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/data_flow_ops.py\", line 287, in dequeue_many<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_data_flow_ops.py\", line 319, in _queue_dequeue_many<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/op_def_library.py\", line 664, in apply_op\n    op_def=op_def)<\/p>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1834, in create_op<\/p>\n\n<pre><\/pre>\n\n<p>File \"\/home\/saoni.m\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1043, in <strong>init<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>When I traced back to the line where shuffle_batch() is called:<\/p>\n\n<pre><\/pre>\n\n<p>The values that are passed to it are:\nbatch size 128, num_threads 16, capacity 20384, min_after_deque 20000<\/p>\n","answers":[{"AnswerId":"45800065","CreationDate":"2017-08-21T14:47:32.670","ParentId":null,"OwnerUserId":"7964238","Title":null,"Body":"<p>I meet a similar problem, and i try to change python list,like\n        [os.path.join(data_dir, i) for i in os.listdir(data_dir)]\nto\n        files=tf.train.match_filenames_once(\"\/path\/to\/data.tfrecords-*\"),\n        filequeue = tf.train.input_string_input_producer(files).\nit works for me and you can try it.<\/p>\n"},{"AnswerId":"35614740","CreationDate":"2016-02-24T22:45:07.843","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Looks like you're not getting any data input from your readers.<\/p>\n\n<p>You changed:<\/p>\n\n<pre><code>[os.path.join(data_dir, i) for i in os.listdir(data_dir)]\n<\/code><\/pre>\n\n<p>What's actually in data_dir\/ ?  (Are you sure the right dirname is being used, etc.?)<\/p>\n\n<p>My suggestion would be to <code>print filenames<\/code> at the start of your execution -- that's not doing anything in tensorflow, just python, so you'll get an instant easy-to-read answer.  If it looks valid, we'll work from there. :)<\/p>\n\n<p>The second concern is that your changes aren't enough to start working on imagenet.  The <code>read_cifar10<\/code> function is specialized for the cifar input format, but the ImageNet data is (mostly) JPEGs, with a separate file specifying the labels.  You can decode the jpegs with <code>tf.image.decode_jpeg<\/code>, but you also need to merge the synset labels in.<\/p>\n"}]}
{"QuestionId":35611970,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-02-24T20:04:36.370","AcceptedAnswerId":null,"OwnerUserId":1217329.0,"Title":"Requiring a Torch command line script without parsing Command line arguments","Body":"<p>I have two Torch scripts. One is some code for training a model, and the other is for testing it.<\/p>\n\n<p>The training script accepts command line arguments<\/p>\n\n<pre><\/pre>\n\n<p>The testing script also accepts command line arguments. It requires the training script because it needs some functions and constants the training script defines<\/p>\n\n<pre><\/pre>\n\n<p>When I run my testing script, it runs the  call from training, and it fails because  isn't a valid option to pass in. Is there a convention for how to make the cmd:parse call not happen unless it's run from the command line?<\/p>\n","answers":[{"AnswerId":"35641447","CreationDate":"2016-02-26T00:59:04.107","ParentId":null,"OwnerUserId":"1217329","Title":null,"Body":"<p>As said in the comment, <code>debug.getinfo(3)<\/code> gives different behavior when called from an include.<\/p>\n\n<p>When called from the command line, <code>debug.getinfo(3).name<\/code> is <code>nil<\/code>. When called from <code>require<\/code> on that file, <code>debug.getinfo(3).name<\/code> is <code>pcall<\/code>. So I wrapped all the command line function into an if block that checks <code>debug.getinfo(3)<\/code> first. (It's analogous to <code>__name__ = '__main__'<\/code> from Python.)<\/p>\n\n<p>This is with the Torch interpreter, I don't know if this works with other Lua interpreters.<\/p>\n"}]}
{"QuestionId":35614993,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-24T23:00:42.370","AcceptedAnswerId":35617218.0,"OwnerUserId":5977269.0,"Title":"Tensorflow: cifar10.py or cifar10_input.py does not work","Body":"<p>If I try to use cifar10.py and cifar10_input.py in cifar10_train.py file,\nit give below errors:<\/p>\n\n<pre><\/pre>\n\n<p>and then terminal hangs...<\/p>\n\n<p>I have only changed 2 things:<\/p>\n\n<ol>\n<li><p>I replaced \"from tensorflow.models.image.cifar10 import cifar10\" by \"import cifar10\" and same for cifar10_input.py<\/p><\/li>\n<li><p>Set max_steps to 5000, to check if the code runs without error.<\/p><\/li>\n<\/ol>\n\n<p>Any help would be much appreciated. Thanks<\/p>\n","answers":[{"AnswerId":"35617218","CreationDate":"2016-02-25T02:40:07.097","ParentId":null,"OwnerUserId":"5977269","Title":null,"Body":"<p>Solved the problem.<\/p>\n\n<p>There should be at least 2 input(.bin) files. I tried with one file before. E.g. There should be data_batch_1.bin and data_batch_2.bin. I tried with data_batch_1.bin before.<\/p>\n"}]}
{"QuestionId":35618951,"AnswerCount":1,"Tags":"<neural-network><torch><recurrent-neural-network>","CreationDate":"2016-02-25T05:17:04.700","AcceptedAnswerId":36135833.0,"OwnerUserId":2491303.0,"Title":"In torch, how can I fix the pre-trained embedding in training time?","Body":"<p>I am trying to train an RNN on pre-trained word embeddings.\nSuppose these pre-trained embedding are kept in a matrix E, which I can use to initialize a LookupTable:<\/p>\n\n<pre><\/pre>\n\n<p>How can I force the model to keep these embeddings fixed during training?<\/p>\n","answers":[{"AnswerId":"36135833","CreationDate":"2016-03-21T15:56:04.157","ParentId":null,"OwnerUserId":"2011802","Title":null,"Body":"<p>Maybe two possibilities :<\/p>\n\n<ol>\n<li><p>Force the weights for this layer to stay at 1 at each mini batch iteration<\/p><\/li>\n<li><p>Implement your own lookuptable by extending the nn.LookupTable and override updateOutput so that the weights are not updated.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35619962,"AnswerCount":4,"Tags":"<python><amazon-web-services><cuda><gpu><theano>","CreationDate":"2016-02-25T06:30:24.027","AcceptedAnswerId":42512445.0,"OwnerUserId":5435514.0,"Title":"Failed to compile cuda_ndarray.cu: libcublas.so.7.5: cannot open shared object file","Body":"<p>I am trying to import theano library in an aws instance to use GPU. I have written a python script using boto to automate aws setup which will essentially do an ssh to the instance from my local machine and then start a bash script where I do \"<strong>python -c 'import theano'\"<\/strong> to start the GPU. But I get the following error: <br><\/p>\n\n<blockquote>\n  <p><strong><em>ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: libcublas.so.7.5: cannot open shared object file: No such file or\n  directory<\/em><\/strong><\/p>\n<\/blockquote>\n\n<p>When I tried to import theano module directly in the instance command shell it automatically starts using GPU.<\/p>\n\n<blockquote>\n  <p>Using gpu device 0: GRID K520 (CNMeM is disabled)<\/p>\n<\/blockquote>\n\n<p>I guess I am missing some other import that has to made while importing through my automation python script. What could possibly be the solution?<\/p>\n","answers":[{"AnswerId":"42512445","CreationDate":"2017-02-28T15:15:48.380","ParentId":null,"OwnerUserId":"6035289","Title":null,"Body":"<p>I will try to solve this problem clearly and concise, as I found not really good answer for people which are starting using unix or are not familiar with compilation and linking.<\/p>\n\n<p>The problem has to do with dynamic linking and it can be solved in two ways. First one is by setting LD_LIBRARY_PATH enviroment variable. Assuming cuda is installed in \/usr\/local\/cuda\/, just add in your enviroment file \/etc\/enviroment:<\/p>\n\n<pre><code>LD_LIBRARY_PATH=\/usr\/local\/cuda\/\n<\/code><\/pre>\n\n<p>Or simply in your bashrc:<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\/\n<\/code><\/pre>\n\n<p>This solution is not recommended by unix gurus (i am not one i have just read that on the internet and i follow linux gurus). So the solution I found is simple, modify the path where the linux ld software search for libraries by default. To do that just do (you have to do it as root):<\/p>\n\n<pre><code>cd \/etc\/ld.so.conf.d\/\n<\/code><\/pre>\n\n<p>Then pick for example and edit:<\/p>\n\n<pre><code>vi libc.conf \n<\/code><\/pre>\n\n<p>Inside this file just add the path to the lib64 root like:<\/p>\n\n<pre><code>\/usr\/local\/cuda\/lib64\/\n<\/code><\/pre>\n\n<p>You would get something like this in the file:<\/p>\n\n<pre><code>\\# libc default configuration\n\n\/usr\/local\/lib\n\n\/usr\/local\/cuda\/lib64\/\n<\/code><\/pre>\n\n<p>And then just run:<\/p>\n\n<pre><code>sudo ldconfig\n<\/code><\/pre>\n\n<p>Hope this answer helps people which are starting seen programming, or using high level languages such as python that uses C code below (like theano does) and are not familiar with compilation, linkig...<\/p>\n"},{"AnswerId":"39025647","CreationDate":"2016-08-18T18:58:55.097","ParentId":null,"OwnerUserId":"4562921","Title":null,"Body":"<p>I faced the same error on Ubuntu 16.04 with cuda 7.5 and found the solution <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu-gpu\" rel=\"nofollow\">here<\/a>:<\/p>\n\n<ol>\n<li><p>cuda 7.5 don't support the default g++ version. Install an supported version and make it the default:<\/p>\n\n<pre><code>sudo apt-get install g++-4.9\n\nsudo update-alternatives --install \/usr\/bin\/gcc gcc \/usr\/bin\/gcc-4.9 20\nsudo update-alternatives --install \/usr\/bin\/gcc gcc \/usr\/bin\/gcc-5 10\n\nsudo update-alternatives --install \/usr\/bin\/g++ g++ \/usr\/bin\/g++-4.9 20\nsudo update-alternatives --install \/usr\/bin\/g++ g++ \/usr\/bin\/g++-5 10\n\nsudo update-alternatives --install \/usr\/bin\/cc cc \/usr\/bin\/gcc 30\nsudo update-alternatives --set cc \/usr\/bin\/gcc\n\nsudo update-alternatives --install \/usr\/bin\/c++ c++ \/usr\/bin\/g++ 30\nsudo update-alternatives --set c++ \/usr\/bin\/g++\n<\/code><\/pre><\/li>\n<li><p>Work around a glibc bug - create .theanorc in the home directory with the following settings:<\/p>\n\n<pre><code>[global]\ndevice=gpu\nfloatX=float32\n\n[nvcc]\nflags=-D_FORCE_INLINES\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>And don't forget to check environment variables:\nPATH should contain your cuda bin folder location and CUDA_HOME should contain cuda home location<\/p>\n\n<p>I've added it to mine .bashrc file this way:<\/p>\n\n<pre><code>export PATH=\"\/usr\/local\/cuda\/bin:$PATH\"\nexport CUDA_HOME=\"\/usr\/local\/cuda:$CUDA_HOME\"\n<\/code><\/pre>\n"},{"AnswerId":"45378608","CreationDate":"2017-07-28T16:44:35.730","ParentId":null,"OwnerUserId":"851699","Title":null,"Body":"<p>On the suggestion of Kumar <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/buSFF7dsrPQ\" rel=\"nofollow noreferrer\">here<\/a>, I did<\/p>\n\n<pre><code>sudo ldconfig \/usr\/local\/cuda\/lib64\n<\/code><\/pre>\n\n<p>And it magically started working.  Thanks Kumar!<\/p>\n"},{"AnswerId":"35896996","CreationDate":"2016-03-09T16:20:04.237","ParentId":null,"OwnerUserId":"4841248","Title":null,"Body":"<p>I had a similar problem recently and spend ages figuring out what was going wrong (to the point I corrupted my Linux install and had to do a fresh install).<\/p>\n\n<p>A <strong>potential solution<\/strong> for this error is to <strong>delete the <code>.theano\/<\/code> directory<\/strong> that is (possibly) located in your home directory:<\/p>\n\n<pre><code>sudo rm -rf ~\/.theano\n<\/code><\/pre>\n\n<p>To prevent this error from happening again, do not run your scripts as root user (i.e. without <code>sudo<\/code>).<\/p>\n\n<p>Running a script as root will create the hidden directory with root permissions, making it inaccessible to other processes.<\/p>\n"}]}
{"QuestionId":35620157,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-02-25T06:41:45.820","AcceptedAnswerId":35630798.0,"OwnerUserId":5703903.0,"Title":"How can I use values read from TFRecords as arguments to tf.reshape?","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'm using a TFRecord to store all my data. The function read_and_decode is from the TFRecords example provided by TensorFlow. Currently I reshape by having predefined values:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, the data that I will be using now is of different dimensions. For example, I could have an image that is [40, 30, 3] (scaling this is not an option as I don't want it to be warped). I would like to read in data of different dimensions and use random_crop in the data augmentation stage to circumvent this problem. What I need is something like the following.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, I can't seem to find a way to do this. Thanks for your help!<\/p>\n\n<p>EDIT:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is definitely with these 2 lines. The hard coded variables work, but not the one with tf.pack().<\/p>\n","answers":[{"AnswerId":"42874109","CreationDate":"2017-03-18T11:57:13.543","ParentId":null,"OwnerUserId":"1574779","Title":null,"Body":"<p>I have faced same issue. According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/shuffle_batch\" rel=\"nofollow noreferrer\">Tensorflow documentation<\/a>, you will be encountered this situation if you are trying to use shuffle_batch \noperation after reading required data.<\/p>\n\n<p>Like in this example, if you don't use shuffle_batch processing, you can load dynamic dimensional files.<\/p>\n\n<pre><code>    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(\n        serialized_example,\n        features={\n             'clip_height': tf.FixedLenFeature([], tf.int64),\n             'clip_width': tf.FixedLenFeature([], tf.int64),\n             'clip_raw': tf.FixedLenFeature([], tf.string),\n             'clip_label_raw': tf.FixedLenFeature([], tf.int64)\n        })\n    image = tf.decode_raw(features['clip_raw'], tf.float64)\n    label = tf.cast(features['clip_label_raw'], tf.int32)\n    height = tf.cast(features['clip_height'], tf.int32)\n    width = tf.cast(features['clip_width'], tf.int32)\n    im_shape = tf.stack([height, width, -1])\n    new_image = tf.reshape(image, im_shape )\n<\/code><\/pre>\n\n<p>But if you are to use shuffle batch processing, you can't use <code>tf.stack<\/code>. You have to define dimensions statically similar to this. <\/p>\n\n<pre><code>    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(\n        serialized_example,\n        features={\n             'clip_height': tf.FixedLenFeature([], tf.int64),\n             'clip_width': tf.FixedLenFeature([], tf.int64),\n             'clip_raw': tf.FixedLenFeature([], tf.string),\n             'clip_label_raw': tf.FixedLenFeature([], tf.int64)\n        })\n    image = tf.decode_raw(features['clip_raw'], tf.float64)\n    label = tf.cast(features['clip_label_raw'], tf.int32)\n    height = tf.cast(features['clip_height'], tf.int32)\n    width = tf.cast(features['clip_width'], tf.int32)\n    image = tf.reshape(image, [1, 512, 1])\n    images, sparse_labels = tf.train.shuffle_batch(\n            [image, label], batch_size=batch_size, num_threads=2,\n            capacity=1000 + 3 * batch_size,\n            min_after_dequeue=100)\n<\/code><\/pre>\n\n<p>@mrry please correct me if I am wrong.<\/p>\n"},{"AnswerId":"35630798","CreationDate":"2016-02-25T14:57:13.210","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You're very close to having a working solution! Right now there's no <em>automatic<\/em> way to give TensorFlow a list made up of tensors and numbers and make a tensor from it, which <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\"><code>tf.reshape()<\/code><\/a> is expecting. The answer is to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"nofollow noreferrer\"><code>tf.stack()<\/code><\/a>, which explicitly takes a list of N-dimensional tensors (or things convertible to tensors) and packs them into an (N+1)-dimensional tensor.<\/p>\n\n<p>This means you can write:<\/p>\n\n<pre><code>features = ...  # Parse from an example proto.\nheight = tf.cast(features['height'], tf.int32)\nwidth = tf.cast(features['width'], tf.int32)\n\nimage = tf.reshape(image, tf.stack([height, width, 3]))\n<\/code><\/pre>\n"}]}
{"QuestionId":35621745,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-25T08:15:31.223","AcceptedAnswerId":40981970.0,"OwnerUserId":2853544.0,"Title":"Importing tf.train.Saver from another python file","Body":"<p>I use <strong>tf.train.Saver()<\/strong> in <strong>one.py<\/strong> file with the following code.<\/p>\n\n<pre><\/pre>\n\n<p>How can I restore <strong>checkpoint.data<\/strong> in another python file?<\/p>\n\n<p>I used the following code, but it didn't work.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40981970","CreationDate":"2016-12-05T19:36:29.273","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The checkpoint file (i.e. <code>'checkpoint.data'<\/code>) does not provide TensorFlow with enough information to reconstruct your model structure. In your second program, you need to reconstruct the same TensorFlow graph that was used in the first program. There are a few options for doing this:<\/p>\n\n<ul>\n<li>Extract the model building code into a Python function, and call it before creating the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/state_ops.html#Saver\" rel=\"nofollow noreferrer\"><code>tf.train.Saver<\/code><\/a> in each program.<\/li>\n<li>Use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/state_ops.html#Saver.export_meta_graph\" rel=\"nofollow noreferrer\"><code>saver.export_meta_graph()<\/code><\/a> to write out the graph structure along with a checkpoint in your first program, and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/state_ops.html#import_meta_graph\" rel=\"nofollow noreferrer\"><code>tf.train.import_meta_graph()<\/code><\/a> to import the graph structure (and create an appropriately configure <code>tf.train.Saver<\/code> instance) in your second program.<\/li>\n<\/ul>\n"}]}
{"QuestionId":35622434,"AnswerCount":0,"Tags":"<theano><tensorflow><deep-learning><rbm><dbn>","CreationDate":"2016-02-25T08:53:16.673","AcceptedAnswerId":null,"OwnerUserId":1749667.0,"Title":"Custom operation implementation for RBM\/DBN with tensorflow?","Body":"<p>Since Google released out tensorflow, it becomes kind of trend in the current deep learning selections.<\/p>\n\n<p>I'd like to do some experiments about RBM\/DBN (Restricted Boltzmann Machine\/Deep Belief Network), I've made some attempt by myself and kind of implement it well through the combination of available APIs from tensorflow. See <a href=\"https:\/\/github.com\/myme5261314\/dbn_tf\" rel=\"nofollow noreferrer\">code<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/34760981\/rbm-implemention-with-tensorflow\/35446666#35446666\">previous answer<\/a>.<\/p>\n\n<p>So, if doesn't bother the code running performance, here's the gift for RBM\/DBN implementation with tensorflow.<\/p>\n\n<p>But, the running performance must be considered for the future. Because of the special progress of CD (Contrastive Divergence) algorithm, I think it just works against the framework (data flow graph) used by tensorflow. That's why my <strong>code<\/strong> seems weired.<\/p>\n\n<p>So, the custom operation should be implemented for acceleration. I've followed the current <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/adding_an_op\/index.html#implement-the-kernel-for-the-op\" rel=\"nofollow noreferrer\">documentation<\/a> about adding custom ops.<\/p>\n\n<pre><\/pre>\n\n<p>In my design,  should is an operation that takes ,,, as input, but output by only <strong>first<\/strong> 3  ( simply sigmoid(X*W+hb) ), its gradient should return at least gradients for <strong>last<\/strong> 3 .<\/p>\n\n<p>Imagine example psuedo code like this:<\/p>\n\n<pre><\/pre>\n\n<p><strong>But<\/strong> the tensorflow library is too complex for me. And after too much time seeking for how to implement these existing operations (, , , , ) in custom operation, no solution is found by myself.<\/p>\n\n<p>So, I'd like to ask if someone could help me achieve the remain works.<\/p>\n\n<p>PS: before getting some ideas, I'd like to dive into  since it implements RBM\/DBN already. Just in my opinion,  is kind of not suitable for RBM\/DBN because of its framework.<\/p>\n\n<p><strong>Update<\/strong>: After scratch through the tutorials from Theano, I found the key reason for Theano implemented the RBM\/DBN while the tensorflow haven't is the  technology. So, there might wait tensorflow to implement  technology to prepare for RBM\/DBN implementation.<\/p>\n","answers":[]}
{"QuestionId":35622784,"AnswerCount":1,"Tags":"<python><theano><deep-learning><autoencoder>","CreationDate":"2016-02-25T09:09:34.787","AcceptedAnswerId":null,"OwnerUserId":5979241.0,"Title":"What is the right way to pass inputs parameters to a Theano function?","Body":"<p>I'm using Python 2.7 with Theano library installed (updated version) and I've got a problem with the inputs parameters, defining a Theano function. <\/p>\n\n<p>The code is:<\/p>\n\n<pre><\/pre>\n\n<p>It's taken from here:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/tutorial\/code\/SdA.py\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/code\/SdA.py<\/a><\/p>\n\n<p>and it gives me this error, with Eclipse:<\/p>\n\n<pre><\/pre>\n\n<p>So, if I change the code in this way:<\/p>\n\n<pre><\/pre>\n\n<p>it works but I can't pass the value of corruption_level and learning_rate.<\/p>\n\n<p>Anyone could help? Thanks!<\/p>\n\n<p>Luca<\/p>\n","answers":[{"AnswerId":"36635839","CreationDate":"2016-04-14T23:35:11.597","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>In was officially deprecated and there was a planed replacement. During a few days, it was removed from Theano development version. But then we realized that it was better to keep it and change it and get rid our planned replacement.<\/p>\n\n<p>During that time, there was also some inconsistency between the what Theano wanted and the Deep Learning tutorial.<\/p>\n\n<p>This was fixed. So now, update to Theano 0.8 or use the current development version of Theano and it should work correctly.<\/p>\n\n<p>Maybe other people having related problems could need to update there Deep Learning tutorial code as for a few days, it was using the planned replacement that we removed.<\/p>\n\n<p>theano.In() now work as in your question.<\/p>\n"}]}
{"QuestionId":35624462,"AnswerCount":1,"Tags":"<neural-network><theano><lasagne>","CreationDate":"2016-02-25T10:20:33.643","AcceptedAnswerId":null,"OwnerUserId":1869634.0,"Title":"Max pooling indices","Body":"<p>I am trying to find the indices 2d max pooling in lasagne <\/p>\n\n<pre><\/pre>\n\n<p>When I try to build the model it raises a error<\/p>\n\n<pre><\/pre>\n\n<p>What is the right way of coding functions on lasagne layers intermediate outputs.<\/p>\n","answers":[{"AnswerId":"35720655","CreationDate":"2016-03-01T10:38:28.790","ParentId":null,"OwnerUserId":"5115006","Title":null,"Body":"<p>I had a similar problem a while ago, check out my solution for 2d and 3d max pooling indices: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/32989064\/theano-max-pool-3d\/35720432#35720432\">Theano max_pool_3d<\/a><\/p>\n\n<p>(Its based on the same Google-groups post, i guess)<\/p>\n"}]}
{"QuestionId":35625027,"AnswerCount":1,"Tags":"<apache-spark><caffe-on-spark>","CreationDate":"2016-02-25T10:44:43.733","AcceptedAnswerId":null,"OwnerUserId":4291805.0,"Title":"Yahoo - caffe on Spark library dependency?","Body":"<p>Yahoo just released a version of caffe that uses the latest version of Apache-Spark yesterday, the git repo is not well documented yet: <a href=\"https:\/\/github.com\/yahoo\/CaffeOnSpark\" rel=\"nofollow\">git link<\/a><\/p>\n\n<p>There is a scala test file which is suppose to run an example: <a href=\"https:\/\/github.com\/yahoo\/CaffeOnSpark\/blob\/master\/caffe-grid\/src\/main\/scala\/com\/yahoo\/ml\/caffe\/examples\/MyMLPipeline.scala\" rel=\"nofollow\">Scala Example<\/a><\/p>\n\n<p>but it requires the dependency com.yahoo.ml.caffe.{Config, CaffeOnSpark, DataSource} which I assume contains basically the data, the config and the API. Has this been made into a library yet? How could I build this using sbt?<\/p>\n","answers":[{"AnswerId":"35656573","CreationDate":"2016-02-26T16:13:29.163","ParentId":null,"OwnerUserId":"5987035","Title":null,"Body":"<p>Our CaffeOnSpark release contains all the code that you need to run. com.yahoo.ml.caffe.* are collection of Scala classes in caffe-grid folder. Please follow our guides at CaffeOnSpark wiki page, and ask questions at your mailing list.  <\/p>\n"}]}
{"QuestionId":35625140,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-25T10:49:23.297","AcceptedAnswerId":null,"OwnerUserId":1441121.0,"Title":"Change default configuration for TensorFlow Session?","Body":"<p>Is it possible to change the default Session configuration, either within Python or by setting environment variables, etc.?<\/p>\n\n<p>Specifically I'd like<\/p>\n\n<pre><\/pre>\n\n<p>to use up much less memory when I'm running small side tests in parallel to other jobs. So I'd the above to behave the same as<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35636976","CreationDate":"2016-02-25T19:52:12.837","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Don't think there's a way to set process-wide default, but here's the pattern I use.<\/p>\n\n<pre><code>def create_session():\n  config = tf.ConfigProto(log_device_placement=True)\n  config.gpu_options.per_process_gpu_memory_fraction=0.3 # don't hog all vRAM\n  sess = tf.InteractiveSession(\"\", config=config)\n  return sess\n\nsess=create_session()\na=tf.constant(1)\nb=tf.constant(2)\nsess.run([a+b])\n<\/code><\/pre>\n"}]}
{"QuestionId":35625635,"AnswerCount":6,"Tags":"<caffe><mnist>","CreationDate":"2016-02-25T11:12:40.417","AcceptedAnswerId":35889485.0,"OwnerUserId":2446538.0,"Title":"mnist dataset Lenet training","Body":"<p>I am new to caffe. To train Lenet with mnist dataset,I downloaded the data to the data\/mnist in caffe. \nWhen trying to convert to lmdb, using  ,  I get error as :<\/p>\n\n<pre><\/pre>\n\n<p>Can somebody help me with this? What exactly could be the issue?<\/p>\n","answers":[{"AnswerId":"40779290","CreationDate":"2016-11-24T06:21:12.697","ParentId":null,"OwnerUserId":"7203600","Title":null,"Body":"<p>rename your mnist dataset,because you downloaded mnist dataset maybe was \"t10k-images.idx3-ubyte\" but it is the \"t10k-images-idx3-ubyte\" in create_mnist.sh,so you have to use '-'instead of '.'<\/p>\n"},{"AnswerId":"45284547","CreationDate":"2017-07-24T15:29:40.180","ParentId":null,"OwnerUserId":"5599843","Title":null,"Body":"<p>Download the mnist dataset from another location (<a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a>), place them in $CAFFE_ROOT\/data\/ and unzip them. After that .\/examples\/mnist\/create_mnist.sh works when run from $CAFFE_ROOT. At least for me it worked.<\/p>\n"},{"AnswerId":"51403010","CreationDate":"2018-07-18T13:15:31.030","ParentId":null,"OwnerUserId":"8702330","Title":null,"Body":"<p>You should run your script from the $CAFFE_ROOT, such as the following: <\/p>\n\n<pre><code>sudo sh \/examples\/mnist\/create_mnist.sh    \n<\/code><\/pre>\n"},{"AnswerId":"42917826","CreationDate":"2017-03-21T03:19:28.713","ParentId":null,"OwnerUserId":"4728909","Title":null,"Body":"<p>I met the same problem. In <strong>$CAFFE_ROOT\/data\/minist\/get_mnist.sh<\/strong>, line 15: <strong>gunzip ${fname}.gz<\/strong> I use ubuntu 14.04 and I don't know why it failed when gunzip train-images-idx3-ubyte.tz and t10k-images-idx3-ubyte.tz. <\/p>\n\n<p>As a result, there are no train-images-idx3-ubyte and t10k-images-idx3-ubyte. In $CAFFE_ROOT\/examples\/mnist\/create_mnist.sh,  it failed to convert them to lmdb format.<\/p>\n\n<p>So I just gunzip them in OSX and then conversion worked.<\/p>\n"},{"AnswerId":"35889485","CreationDate":"2016-03-09T10:51:44.437","ParentId":null,"OwnerUserId":"2446538","Title":null,"Body":"<p>The manual compilation worked, though the script failed for unknown reason.<\/p>\n"},{"AnswerId":"35920988","CreationDate":"2016-03-10T15:41:40.483","ParentId":null,"OwnerUserId":"5837113","Title":null,"Body":"<p>You can use absolutely path not like <code>data\/mnist\/t10k-images-idx3-ubyte<\/code>.caffe sometimes can't get it.<\/p>\n"}]}
{"QuestionId":35628221,"AnswerCount":4,"Tags":"<python><anaconda><tensorflow>","CreationDate":"2016-02-25T13:08:15.410","AcceptedAnswerId":null,"OwnerUserId":2065691.0,"Title":"Anaconda with Tensorflow: No module named any_pb2","Body":"<p>I have just installed Anaconda in a Ubuntu 14.04. Then I installed tensorflow using<\/p>\n\n<pre><\/pre>\n\n<p>Then I installed protobuf<\/p>\n\n<pre><\/pre>\n\n<p>However, I am not able to solve this issue:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35642964","CreationDate":"2016-02-26T03:36:17.960","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Here's my comment on the issue from github, hope this helps.<\/p>\n\n<p>\"For anyone else who is running into the same issue, I was also having similar problems when trying to import tensorflow on notebook via jupyter. The modules were installed using conda. After spending two days on it without any success (kept getting the same, no module any_pb2\/contrib), it was obvious it was a notebook problem and not tensorflow related since I could import tensorflow from python shell. So I decided to switch back to vanilla python and installed everything manually (tensorflow via pip &amp; got rid of jupyter &amp; installed ipython instead) and now I am able to import tensorflow from my ipython notebook.\"<\/p>\n\n<p>Original comment link: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1161#issuecomment-188664103\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1161#issuecomment-188664103<\/a><\/p>\n"},{"AnswerId":"35631850","CreationDate":"2016-02-25T15:42:05.257","ParentId":null,"OwnerUserId":"908231","Title":null,"Body":"<p>The following worked for me:<\/p>\n\n<p>conda install -c <a href=\"https:\/\/conda.anaconda.org\/jjhelmus\" rel=\"nofollow\">https:\/\/conda.anaconda.org\/jjhelmus<\/a> tensorflow<\/p>\n"},{"AnswerId":"35630537","CreationDate":"2016-02-25T14:48:05.097","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>Use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#pip-installation\" rel=\"nofollow\">instructions on the TensorFlow site<\/a>.<\/p>\n\n<p>TLDR;<\/p>\n\n<p>I am posting this as an answer because people don't look for answers in the comments.<\/p>\n"},{"AnswerId":"35666634","CreationDate":"2016-02-27T06:38:28.743","ParentId":null,"OwnerUserId":"4172036","Title":null,"Body":"<p>Posting this in case it helps anyone else. I was having the same issue when trying to import TensorFlow in the IPython console in Spyder using the conda TensorFlow package. I was able to fix the issue by uninstalling the conda package:<\/p>\n\n<pre><code>conda uninstall tensorflow\n<\/code><\/pre>\n\n<p>And then installing using <em>pip<\/em>: <\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl --ignore-installed\n<\/code><\/pre>\n\n<p>Note the final flag, without the flag I was getting an error message about <em>easy-install.pth<\/em> and the installation would fail. Hope this helps someone else! <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/622\" rel=\"nofollow\">Credit to frankcarey for the fix.<\/a><\/p>\n"}]}
{"QuestionId":35629854,"AnswerCount":1,"Tags":"<reshape><tensorflow>","CreationDate":"2016-02-25T14:18:41.577","AcceptedAnswerId":35631180.0,"OwnerUserId":4650495.0,"Title":"reshape output or tf.gather(), tensorflow","Body":"<p>I have a tensor in  from  and I want to reshape it to \nbut using  shows me error of  because of dynamic aspect of the graph. \nis there any way to first reshape the inside of  as far as the dimensions are known in graph and then using the  having the ?<\/p>\n","answers":[{"AnswerId":"35631180","CreationDate":"2016-02-25T15:12:51.470","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#reshape\" rel=\"nofollow\"><code>tf.reshape()<\/code><\/a> op doesn't understand partial shapes (i.e. those with a <code>None<\/code> for one or more dimensions), because they can be ambiguous: e.g. there could be many possible concrete shapes for a partial shape <code>[None, None, 50]<\/code>.<\/p>\n\n<p>However, <code>tf.reshape()<\/code> also allows you to specify <em>one<\/em> of the dimensions as a wildcard, which will be chosen automatically, so you can use this in your case. To specify the wildcard, use <code>-1<\/code> as one of the dimensions:<\/p>\n\n<pre><code>input = ...\nprint input.get_shape()  ==&gt; [None, 100, 5, 50]\n\nreshaped = tf.reshape(input, [-1, 100, 250])\n<\/code><\/pre>\n"}]}
{"QuestionId":35630020,"AnswerCount":1,"Tags":"<python><install><theano>","CreationDate":"2016-02-25T14:25:41.430","AcceptedAnswerId":null,"OwnerUserId":1649095.0,"Title":"Installing theano","Body":"<p>I installed theano with Spyder 2.3.8 on Windows 7 64-bit by running \"pip install theano\". It worked well. But then, when I tried to run \"import theano\", I got the following error:<\/p>\n\n<pre>\nTraceback (most recent call last):\n\n  File \"\", line 1, in \n    import theano\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\__init__.py\", line 55, in \n    from theano.compile import \\\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\compile\\__init__.py\", line 9, in \n    from theano.compile.function_module import *\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\compile\\function_module.py\", line 18, in \n    import theano.compile.mode\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\compile\\mode.py\", line 11, in \n    import theano.gof.vm\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\gof\\vm.py\", line 25, in \n    in_c_key=False)\n\n  File \"D:\\Anaconda3\\lib\\site-packages\\theano\\configparser.py\", line 231, in AddConfigVar\n    configparam.fullname)\n\nAttributeError: ('This name is already taken', 'profile')\n<\/pre>\n\n<p>What does that mean?<\/p>\n","answers":[{"AnswerId":"37780769","CreationDate":"2016-06-13T01:35:27.950","ParentId":null,"OwnerUserId":"2280645","Title":null,"Body":"<p>Is there a reason you are using Spyder?<\/p>\n\n<p>On windows 7 x64, I would advise to either use Canopy from Enthought, which, if you are a student , you can get the complete package for free.<\/p>\n\n<p>Or, i also used WinPython before, it worked fine.\nAlthough you have to follow all steps, including :<\/p>\n\n<p><strong>Installing Windows Software Development Kit version 7.1 or you can install the newest Visual Studio which comes with it and it comes with an option to install Microsoft Visual C++ Compiler for Python 2.7<\/strong><\/p>\n\n<p><strong>Installing TDM GCC<\/strong> <\/p>\n\n<p><strong><em>But be sure to check<\/em><\/strong> <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a><\/p>\n\n<p><em>You might wanna check if you missed any steps.<\/em><\/p>\n\n<p>I had problems with Anaconda and Theano before. So i gave up on using it.\nHowever a friend had success with it and Theano as well.<\/p>\n\n<p>Just be sure you have your environment and python setup installed properly prior to theano install.<\/p>\n\n<p>Doing <\/p>\n\n<pre><code>where gcc\nwhere gendef\nwhere cl\nwhere nvcc\n<\/code><\/pre>\n\n<p><strong><em>on the Command Prompt Might reveal something...<\/em><\/strong><\/p>\n\n<p>Im not a <strong>GCC<\/strong> specialist, but according from the link from the user above, \ni read a friend saying its something related to GCC.\nSo, yes, i suggest you re-check as i said .<\/p>\n\n<p>Edit the question with that info and i might be able to help more and so will others.<\/p>\n\n<p>@CantFindName<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/25423173\/name-conflicting-in-theano\">Name conflicting in Theano<\/a><\/p>\n"}]}
{"QuestionId":35633094,"AnswerCount":2,"Tags":"<multithreading><lua><torch>","CreationDate":"2016-02-25T16:36:01.407","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Torch - Multithreading to load tensors into a queue for training purposes","Body":"<p>I would like to use the library threads (or perhaps parallel) for loading\/preprocessing data into a queue but I am not entirely sure how it works. In summary;<\/p>\n\n<ol>\n<li>Load data (tensors), pre-process tensors (this takes time, hence why I am here) and put them in a queue. I would like to have as many threads as possible doing this so that the model is not waiting or not waiting for long.<\/li>\n<li>For the tensor at the top of the queue, extract it and forward it through the model and remove it from the queue.<\/li>\n<\/ol>\n\n<p>I don't really understand the example in <a href=\"https:\/\/github.com\/torch\/threads\" rel=\"nofollow\">https:\/\/github.com\/torch\/threads<\/a> enough. A hint or example as to where I would load data into the queue and train would be great.<\/p>\n\n<p>EDIT 14\/03\/2016<\/p>\n\n<p>In this example \"<a href=\"https:\/\/github.com\/torch\/threads\/blob\/master\/test\/test-low-level.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/threads\/blob\/master\/test\/test-low-level.lua<\/a>\" using a low level thread, does anyone know how I can extract data from these threads into the main thread? <\/p>\n","answers":[{"AnswerId":"35634194","CreationDate":"2016-02-25T17:28:27.943","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Look at this multi-threaded data provider:\n<a href=\"https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua\" rel=\"nofollow\">https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua<\/a><\/p>\n\n<p>It runs this file in the thread:\n<a href=\"https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L18\" rel=\"nofollow\">https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L18<\/a><\/p>\n\n<p>by calling it here:\n<a href=\"https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L30-L43\" rel=\"nofollow\">https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L30-L43<\/a><\/p>\n\n<p>And afterwards, if you want to queue a job into the thread, you provide two functions:\n<a href=\"https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L84\" rel=\"nofollow\">https:\/\/github.com\/soumith\/dcgan.torch\/blob\/master\/data\/data.lua#L84<\/a>\nThe first one runs inside the thread, and the second one runs in the main thread after the first one completes.<\/p>\n\n<p>Hopefully that makes it a bit more clear.<\/p>\n"},{"AnswerId":"38245869","CreationDate":"2016-07-07T12:46:56.023","ParentId":null,"OwnerUserId":"2104596","Title":null,"Body":"<p>If Soumith's examples in the previous answer are not very easy to use, I suggest you build your own pipeline from scratch. I provide here an example of two synchronized threads : one for writing data and one for reading data:<\/p>\n\n<pre><code>local t = require 'threads'\nt.Threads.serialization('threads.sharedserialize')\nlocal tds = require 'tds'\nlocal dict = tds.Hash()  -- only local variables work here, and only tables or tds.Hash()\ndict[1] = torch.zeros(4)\n\nlocal m1 = t.Mutex()\nlocal m2 = t.Mutex()\nlocal m1id  = m1:id()\nlocal m2id  = m2:id()\n\nm1:lock()\n\nlocal pool = t.Threads(\n  1,\n  function(threadIdx)\n  end\n)\n\npool:addjob(\n  function()\n    local t = require 'threads'\n    local m1 = t.Mutex(m1id)\n    local m2 = t.Mutex(m2id)\n\n    while true do\n      m2:lock()\n      dict[1] = torch.randn(4)\n      m1:unlock()\n\n      print ('W ===&gt; ')\n      print(dict[1])\n      collectgarbage()\n      collectgarbage()      \n    end\n\n    return __threadid\n  end,\n  function(id)\n  end\n)\n\n-- Code executing on master:\nlocal a = 1\nwhile true do\n  m1:lock()\n  a = dict[1]\n  m2:unlock()\n\n  print('R --&gt; ')\n  print(a)\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":35635131,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-02-25T18:17:55.440","AcceptedAnswerId":null,"OwnerUserId":1410871.0,"Title":"Unable to visualize Inception v3 model in TensorBoard with TensorFlow 0.7.1","Body":"<p>I'm attempting to visualize Google's Inception v3 model using TensorBoard in TensorFlow 0.7.1 and am unable to do so. The TensorBoard Graph tab stalls with the statement<\/p>\n\n<blockquote>\n  <p>Data : Reading graph.pbtxt<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5G0oo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5G0oo.png\" alt=\"Data : Reading graph.pbtxt\"><\/a><\/p>\n\n<p>I downloaded an un-Tarred the <a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz\" rel=\"nofollow noreferrer\">inception v3 model<\/a>. The graph protobuffer is in .<\/p>\n\n<p>Here's my code to dump the model:<\/p>\n\n<pre><\/pre>\n\n<p>This dumps a 91 MB file called  (same size as the graph protobuffer) so it seems that the graph is in there somewhere.<\/p>\n\n<p>I ran TensorBoard as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Which results in the aforementioned hung loading bar on the Graph page.<\/p>\n\n<p>The Chrome JavaScript console yields this error:<\/p>\n\n<blockquote>\n  <p>Uncaught TypeError: Cannot read property '0' of undefined<\/p>\n<\/blockquote>\n\n<p>which I assume is related to the fact that the graph is missing.<\/p>\n\n<p>I've tried this with Chrome 48.0.2564.116 (64-bit) on OS X 10.11.3 both with TensorFlow 0.7.1 for Python 3 built using Bazel and TensorFlow 0.7.1 for Python 2 installed via pip with the exact same results.<\/p>\n\n<p>I also verified that I <em>can<\/em> visualize the graph generated with the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"nofollow noreferrer\">mnist_with_summaries example<\/a>, so this is an issue specifically with the Inception model.<\/p>\n","answers":[{"AnswerId":"35636567","CreationDate":"2016-02-25T19:30:20.777","ParentId":null,"OwnerUserId":"5952845","Title":null,"Body":"<p>Daniel,<\/p>\n\n<p>I don't know about the protobuffer stuff you mention, but I think you may want to uninstall the protobuf and reinstall tensorflow.<\/p>\n\n<p>After my upgrade (on Ubuntu) from  tensorflow v0.6 to v0.7.1 I found the following post when I had trouble launching the tensorboard.<\/p>\n\n<p>I think there was a problem with protobuf. Talked about here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1134#issuecomment-185279000\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1134#issuecomment-185279000<\/a><\/p>\n\n<p>I ended up uninstalling protobuf and reinstalling tensorboard. Now I can launch the board and view my graph. Good luck! :)<\/p>\n\n<p>Don <\/p>\n"},{"AnswerId":"51897808","CreationDate":"2018-08-17T14:42:40.277","ParentId":null,"OwnerUserId":"7769653","Title":null,"Body":"<p>Use this code:\nThis works for me:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.platform import gfile\nwith tf.Session() as sess:\n    model_filename ='YouGraphNameWithPath.pb'\n    with gfile.FastGFile(model_filename, 'rb') as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n        g_in = tf.import_graph_def(graph_def)\nLOGDIR='PathWhereSummaryWillBeSaved'\ntrain_writer = tf.summary.FileWriter(LOGDIR)\ntrain_writer.add_graph(sess.graph)\n<\/code><\/pre>\n"}]}
{"QuestionId":35636868,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-25T19:46:25.003","AcceptedAnswerId":null,"OwnerUserId":1585638.0,"Title":"looping through dataset once at test time in tensorflow","Body":"<p>What's the best way to take a single pass over a dataset, in order to evaluate on test data?  I'd like to avoid scripting the data loading in python and using a . Instead, I'd like to use all the nice TF infrastructure for queueing, batching, etc. <\/p>\n\n<p>In the cifar <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py#L92\" rel=\"noreferrer\">example<\/a>, the number of test examples is hard-coded and the code takes  steps in order to do evaluation. It seems like there should be a better way to do this, using the batching infrastructure. <\/p>\n\n<p>It seems that the standard pattern is to stop running when you catch some exception thrown by the queue. I've tried some things, such that the queue complains when there are no more examples to populate the queue (ie the Producer can't produce any more). This isn't the exception you want to catch. You want to catch when the consumer has nothing left to consume, ie the queue is empty. How do I do this?<\/p>\n\n<p>Also, what do you do if the number of test examples isn't divisible by a batch size (eg. the number of test examples is prime). <\/p>\n\n<p>Additional information:<\/p>\n\n<p>In practice, we typically evaluate on test data multiple times during learning, by calling a  function. Yaroslav's answer is useful if you only want process the test data once. Ideally, each call to do_evaluation would run over every example in the test dataset exactly once. We need some mechanism for resetting the batcher so that you can take a single pass over it one more time. Here's some code for that. Don't use the  command. It takes a batcher that doesn't shuffle and specify the number of batches in the test set (this doesn't work if the number of examples set isn't divisible by minibatchsize). The function returns a new op for grabbing data that will throw a  when you've run over the whole set. The second return value is an op that should be called to reset the counter. This should be the first call inside a  function. <\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"52841385","CreationDate":"2018-10-16T17:56:35.797","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>You can use the tf.Data API for this. Like so. <\/p>\n\n<pre><code>import tensorflow as tf\n\ngraph = tf.Graph()\nsess = tf.Session(graph=graph)\n\n\ndef build_dataset(train_or_test):\n    if train_or_test == 'test':\n        dataset = tf.data.Dataset.from_tensor_slices(tf.zeros([4, 2]))\n    elif train_or_test == 'train':\n        dataset = tf.data.Dataset.from_tensor_slices(tf.ones([10, 2]))\n    else:\n        raise ValueError('wrong name')\n    batch_size = 3\n    dataset = dataset.batch(batch_size)\n    return dataset\n\n\ndef build_inputs():\n    train_dataset = build_dataset('train')\n    test_dataset = build_dataset('test')\n    iterator = tf.data.Iterator.from_structure(\n        train_dataset.output_types,\n        train_dataset.output_shapes,)\n    data = iterator.get_next()\n    tf.identity(data, name='data')\n    iterator.make_initializer(train_dataset, name='train_init')\n    iterator.make_initializer(test_dataset, name='test_init')\n\n\ndef model(inputs):\n    return tf.add(inputs, 1, name='output')\n\n\ndef build_graph():\n    with graph.as_default():\n        build_inputs()\n        data = graph.get_tensor_by_name('data:0')\n        model(data)\n\n\ndef train():\n    train_init = graph.get_operation_by_name('train_init')\n    sess.run(train_init)\n    out = graph.get_tensor_by_name('output:0')\n    while True:\n        try:\n            network_out = sess.run(out)\n            print(network_out.shape)\n            print(network_out)\n        except tf.errors.OutOfRangeError:\n            break\n\n\ndef test():\n    test_init = graph.get_operation_by_name('test_init')\n    sess.run(test_init)\n    out = graph.get_tensor_by_name('output:0')\n    while True:\n        try:\n            network_out = sess.run(out)\n            print(network_out.shape)\n            print(network_out)\n        except tf.errors.OutOfRangeError:\n            break\n\n\ndef train_loop():\n    cur_epoch = 0\n    while cur_epoch &lt; 1:\n        print('Test epoch')\n        test()\n        print('Train epoch')\n        train()\n        cur_epoch += 1\n\n\ndef initialise_graph():\n    with graph.as_default():\n        sess.run(tf.global_variables_initializer())\n\n\nbuild_graph()\ninitialise_graph()\ntrain_loop()\n<\/code><\/pre>\n\n<p>This will output:<\/p>\n\n<pre><code>Test epoch\n(3, 2)\n[[1. 1.]\n [1. 1.]\n [1. 1.]]\n(1, 2)\n[[1. 1.]]\nTrain epoch\n(3, 2)\n[[2. 2.]\n [2. 2.]\n [2. 2.]]\n(3, 2)\n[[2. 2.]\n [2. 2.]\n [2. 2.]]\n(3, 2)\n[[2. 2.]\n [2. 2.]\n [2. 2.]]\n(1, 2)\n[[2. 2.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35639131,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-02-25T21:53:57.497","AcceptedAnswerId":35644548.0,"OwnerUserId":5982966.0,"Title":"Efficiently apply function to neuron outputs then sum, not apply function to sum","Body":"<p>I want to apply a nonlinearity to a neuron's individual inputs and sum the result, not sum the individual inputs and apply a nonlinearity. My current implementation is slow.<\/p>\n\n<p>I have an input layer of size D, and the batch size is B. The next layer is  neurons wide, so the weight matrix . Call my nonlinearity .<\/p>\n\n<p>In symbols, for neuron n, I want  instead of, as we normally have, .<\/p>\n\n<p>My vectorization of this process is slow. Currently, it exploits broadcasting, and I am wondering if there is a faster way to get what I want.<\/p>\n\n<pre><\/pre>\n\n<p>I then proceed to feed this to the next layer. Doing this several times is  slow. I am wondering if there is some way I can make this more efficient or compute this differently.<\/p>\n","answers":[{"AnswerId":"35644548","CreationDate":"2016-02-26T05:56:46.953","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>(Just to sanity check that we're on the same page:  A way of restating what you're asking for is to be able to apply a function f(x) to each scalar element*element product inside a matrix multiply, <em>before<\/em> those products are summed to produce the final output value.)<\/p>\n\n<p>An unsatisfying - and perhaps someone can prove me wrong - answer is that I think it's the best you can do without writing a kernel to implement it.  The core challenge is that without a custom kernel (or implementing a block-decomposed matrix multiply out of raw ops, which sounds unpleasant), I can't think of a way to get the cache locality that you need to make this operation fast.<\/p>\n"}]}
{"QuestionId":35640267,"AnswerCount":1,"Tags":"<nlp><theano><recurrent-neural-network>","CreationDate":"2016-02-25T23:08:47.970","AcceptedAnswerId":null,"OwnerUserId":4833664.0,"Title":"Why we need fine tune word embedding in recurrent neural networks?","Body":"<p>In the theano's tutorial about <a href=\"http:\/\/deeplearning.net\/tutorial\/rnnslu.html\" rel=\"nofollow\">RNN<\/a>, at the final part it mentioned that <\/p>\n\n<blockquote>\n  <p>We keep the word embeddings on the unit sphere by normalizing them after each update:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>I searched online and only find the paper <a href=\"http:\/\/www.iro.umontreal.ca\/~lisa\/pointeurs\/RNNSpokenLanguage2013.pdf\" rel=\"nofollow\">Investigation of Recurrent-Neural-Network Architectures and Learning Methods for Spoken Language Understanding<\/a> mentioned it is fine-tuning word embedding in Chapter 3.5.1.<\/p>\n\n<p>I don't understand why we can fine tune word embedding.      is the imput of this RNN model, right? How could we change the input value?<\/p>\n","answers":[{"AnswerId":"42068757","CreationDate":"2017-02-06T13:16:53.760","ParentId":null,"OwnerUserId":"4323148","Title":null,"Body":"<p>I encountered this problem, and this is my understanding:<\/p>\n\n<p>Actually <code>self.emb<\/code> is not purely the input of the RNN, it also likes a part of parameters(or vectors in word2vec) which is trained to fit the words in <a href=\"http:\/\/deeplearning.net\/tutorial\/rnnslu.html#dataset\" rel=\"nofollow noreferrer\">ATIS<\/a>.<\/p>\n\n<p>When we treat it as the parameters, the fine-tune is reasonable.<\/p>\n\n<p>But when we treat it as the input, normalization becomes reasonable.<\/p>\n\n<p>In addtion, I read the paper you mentioned and I think fine-tuning doesn't contain the normalization because fine-tuning is a type of supervised learning, but the normalization is just a processing of data.<\/p>\n"}]}
{"QuestionId":35641536,"AnswerCount":2,"Tags":"<lua><jpeg><binary-data><torch><luasocket>","CreationDate":"2016-02-26T01:06:28.250","AcceptedAnswerId":35646907.0,"OwnerUserId":4917583.0,"Title":"Convert Binary Data to Torch Tensor in Lua","Body":"<p>I have Lua code that downloads an image from a url using a luasocket:<\/p>\n\n<pre><\/pre>\n\n<p>In order to read this image into a Torch tensor, I save it in a jpg file and read it using :<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to convert the binary jpg data to a torch tensor directly without doing a write-to-and-read-from the hard-drive? Something like:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"35646907","CreationDate":"2016-02-26T08:30:27.870","ParentId":null,"OwnerUserId":"2546626","Title":null,"Body":"<p>See <a href=\"https:\/\/github.com\/torch\/image\/blob\/master\/doc\/saveload.md#res-imagedecompressjpgtensor-depth-tensortype\" rel=\"nofollow\">image.decompressJPG<\/a>.<\/p>\n\n<p>You just have to pack your <code>body<\/code> string inside a ByteTensor first. This can be done by constructing this tensor with a <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/storage.md#constructors-and-access-methods\" rel=\"nofollow\">storage<\/a> which can set his contents with <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/storage.md#self-stringstr\" rel=\"nofollow\">string(str)<\/a>.<\/p>\n"},{"AnswerId":"37491887","CreationDate":"2016-05-27T20:38:04.273","ParentId":null,"OwnerUserId":"3859312","Title":null,"Body":"<p>One potential solution is to use graphicsmagick.<\/p>\n\n<pre><code>local gm = require 'graphicsmagick'\nlocal img = gm.Image()\nlocal ok = pcall(img.fromString, img, body)\nimg = img:toTensor('float', 'RGB', 'DHW')\n<\/code><\/pre>\n\n<p>I found this example in <a href=\"https:\/\/github.com\/clementfarabet\/graphicsmagick\/blob\/master\/test\/corrupt.lua\" rel=\"nofollow\">https:\/\/github.com\/clementfarabet\/graphicsmagick\/blob\/master\/test\/corrupt.lua<\/a> and I know that <\/p>\n\n<pre><code>local body, code = http.request(image_url)\n<\/code><\/pre>\n\n<p>will return <code>body<\/code> as a string. And, obviously if <code>pcall<\/code> returns false, the image was corrupt.<\/p>\n"}]}
{"QuestionId":35641732,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-02-26T01:27:42.323","AcceptedAnswerId":null,"OwnerUserId":5703903.0,"Title":"How can I sequentially read and then evaluate from a test set and output the predictions to a file in TensorFlow?","Body":"<p>I noticed that CIFAR10 in TensorFlow does batch evaluation. The batches are created with many threads. As a result of this, the image sequences are pseudo random in the batch.<\/p>\n\n<p>To circumvent this, my current implementation of doing evaluation is to <\/p>\n\n<ol>\n<li>Set num_epochs in tf.train.string_input_producer to 1, thus reading the TFRecords file only once.<\/li>\n<li>Set num_threads in tf.train.batch to 1, thus only using a single thread and preventing multiple reads from the same TFRecords file.<\/li>\n<\/ol>\n\n<p>Following that, I print the predicted results on the console.<\/p>\n\n<p>I have 2 questions:<\/p>\n\n<ol>\n<li>Is this the best way to evaluate a test set sequentially?<\/li>\n<li>Is simply opening a file and writing the predicted results out the best way to do this?<\/li>\n<\/ol>\n\n<p>Thank you for your help.<\/p>\n","answers":[]}
{"QuestionId":35642033,"AnswerCount":0,"Tags":"<python><loading><theano>","CreationDate":"2016-02-26T01:58:59.870","AcceptedAnswerId":null,"OwnerUserId":1362215.0,"Title":"How do I update my training data by loading a new chunk from a file with theano?","Body":"<p>I am loading image files for a neural network, and their size becomes nontrivial when they are read into numpy. Currently I have all the files as .png files on my hard drive.  I can safely load the file names and the response variable, since they are several orders of magnitude less than the images themselves.<\/p>\n\n<p>I would like to read them like this:<\/p>\n\n<pre><\/pre>\n\n<p>I have my givens defined as follows<\/p>\n\n<pre><\/pre>\n\n<p>Obviously, this doesn't work as I get an IndexError from an invalid slice, but I am not sure how to get theano to update something that isn't already in memory.  It should be possible, as part of <a href=\"http:\/\/deeplearning.net\/tutorial\/gettingstarted.html\" rel=\"nofollow\">the tutorial I am reading<\/a> is as follows:<\/p>\n\n<blockquote>\n  <p>If you are running your code on the GPU and the dataset you are using is too large to fit in memory the code will crash. In such a case you should store the data in a shared variable. You can however store a sufficiently small chunk of your data (several minibatches) in a shared variable and use that during training. Once you got through the chunk, update the values it stores. This way you minimize the number of data transfers between CPU memory and GPU memory.<\/p>\n<\/blockquote>\n\n<p>How do I do this?<\/p>\n","answers":[]}
{"QuestionId":35644264,"AnswerCount":4,"Tags":"<python><mongodb><csv><tensorflow>","CreationDate":"2016-02-26T05:36:18.560","AcceptedAnswerId":37302871.0,"OwnerUserId":3723683.0,"Title":"How to read data into Tensorflow?","Body":"<p>I'm trying to read data from CSV files to tensorflow,<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#filenames-shuffling-and-epoch-limits\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#filenames-shuffling-and-epoch-limits<\/a><\/p>\n\n<p>The sample code in official document is like this:<\/p>\n\n<pre><\/pre>\n\n<p>To read the file, I need to know how many columns and lines in the file beforehand, and if there are 1000 columns, I need to define 1000 variables like  this doesn't look like an efficient way to read data.<\/p>\n\n<p>My questions<\/p>\n\n<ol>\n<li><p>What is the best way to read CSV files into Tensorflow ?<\/p><\/li>\n<li><p>Is there any way to read Database (such as mongoDB) in Tensorflow ?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"48716741","CreationDate":"2018-02-10T02:47:51.887","ParentId":null,"OwnerUserId":"8012473","Title":null,"Body":"<p>of course you can implement to directly read batch random sort trained data from mongo to feed to tensorflow. below is my way: <\/p>\n\n<pre><code>        for step in range(self.steps):\n\n\n            pageNum=1;\n            while(True):\n                trainArray,trainLabelsArray = loadBatchTrainDataFromMongo(****)\n                if len(trainArray)==0:\n                    logging.info(\"train datas consume up!\")\n                    break;\n                logging.info(\"started to train\")\n                sess.run([model.train_op],\n                         feed_dict={self.input: trainArray,\n                                    self.output: np.asarray(trainLabelsArray),\n                                    self.keep_prob: params['dropout_rate']})\n\n                pageNum=pageNum+1;\n<\/code><\/pre>\n\n<p>and also you need preprocess trained data in mongodb, such like: assign each trained data in mongodb a random sort value...<\/p>\n"},{"AnswerId":"37302871","CreationDate":"2016-05-18T14:40:03.037","ParentId":null,"OwnerUserId":"4794308","Title":null,"Body":"<ol>\n<li><p>You definitely don't need to define col1, col2, to col1000...<\/p>\n\n<p>generally, you might do things like this:<\/p>\n\n<pre><code>\ncolumns = tf.decode_csv(value, record_defaults=record_defaults)\nfeatures = tf.pack(columns)\ndo_whatever_you_want_to_play_with_features(features)\n<\/code><\/pre><\/li>\n<li><p>I do not know any off-the-shelf way to directly read data from MongoDB. Maybe you can just write a short script to convert data from MongoDB in a format that Tensorflow supports, I would recommend binary form <code>TFRecord<\/code>, which is much faster to read than csv record. <a href=\"https:\/\/indico.io\/blog\/tensorflow-data-inputs-part1-placeholders-protobufs-queues\/\" rel=\"noreferrer\">This<\/a> is a good blog post about this topic. Or you can choose to implement a customized data reader by yourself, see <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/new_data_formats\/index.html#custom-data-readers\" rel=\"noreferrer\">the official doc<\/a> here.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"55526796","CreationDate":"2019-04-05T00:48:04.770","ParentId":null,"OwnerUserId":"2934579","Title":null,"Body":"<blockquote>\n  <p>Is there any way to read Database (such as mongoDB) in Tensorflow ?<\/p>\n<\/blockquote>\n\n<p>Try <a href=\"https:\/\/github.com\/svenboesiger\/tfmongodb\" rel=\"nofollow noreferrer\">TFMongoDB<\/a>, a C++ implemented dataset op for TensorFlow that allows you to connect to your MongoDB:<\/p>\n\n<pre><code>pip install tfmongodb\n<\/code><\/pre>\n\n<p>There's an example on the GitHub page on how to read data. See also <a href=\"https:\/\/pypi.org\/project\/TFMongoDB\/#description\" rel=\"nofollow noreferrer\">pypi: TFMongoDB<\/a><\/p>\n"},{"AnswerId":"37269345","CreationDate":"2016-05-17T07:06:48.390","ParentId":null,"OwnerUserId":"1363836","Title":null,"Body":"<pre><code>def func()\n    return 1,2,3,4\n\nb = func() \n\nprint b #(1, 2, 3, 4)\n\nprint [num for num in b] # [1, 2, 3, 4]\n<\/code><\/pre>\n\n<p>Hi its nothing to do with tensorflow its simple python need not define 1000 variable. tf.decode_csv returns a tuple.<\/p>\n\n<p>No idea on database handling, I think u can use python and just input the data in the form of array to the tensorflow.<\/p>\n\n<p>Hope this is helpful<\/p>\n"}]}
{"QuestionId":35645262,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-02-26T06:47:48.223","AcceptedAnswerId":35652213.0,"OwnerUserId":5107003.0,"Title":"cifar10: TypeError: range() takes at least 2 arguments (1 given)","Body":"<p>When I build the file: <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10_train.py\" rel=\"nofollow\"><\/a>, it occurs that:<\/p>\n\n<pre><\/pre>\n\n<p>The trouble occurs in the file <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow\"><\/a>.<\/p>\n","answers":[{"AnswerId":"35652213","CreationDate":"2016-02-26T12:50:37.140","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It seems that you are using TensorFlow version 0.5 and a more recent version of the <code>cifar10_train.py<\/code> script. The signature of <code>tf.range()<\/code> was changed to accept a single argument (like the Python <code>range()<\/code> built-in function) after version 0.5 was released.<\/p>\n\n<p>I'd always recommend <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html\" rel=\"nofollow\">upgrading to the latest version<\/a> of TensorFlow, since the runtime has had many performance and stability improvements since the initial release.<\/p>\n\n<p>If that doesn't work, the following equivalent code is taken <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f41959ccb2d9d4c722fe8fc3351401d53bcf4900\/tensorflow\/models\/image\/cifar10\/cifar10.py#L364\" rel=\"nofollow\">from the original release<\/a>:<\/p>\n\n<pre><code>indices = tf.reshape(tf.range(0, FLAGS.batch_size, 1), [FLAGS.batch_size, 1])\n<\/code><\/pre>\n"},{"AnswerId":"35645263","CreationDate":"2016-02-26T06:47:48.223","ParentId":null,"OwnerUserId":"5107003","Title":null,"Body":"<p>I alter the command as follows:  <\/p>\n\n<pre><code>#indices = tf.reshape(tf.range(FLAGS.batch_size), [FLAGS.batch_size, 1])\nindices = tf.reshape(range(FLAGS.batch_size), [FLAGS.batch_size, 1])\n<\/code><\/pre>\n\n<p>Then, the code works well.<\/p>\n"}]}
{"QuestionId":35646577,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><cntk>","CreationDate":"2016-02-26T08:10:33.400","AcceptedAnswerId":35674999.0,"OwnerUserId":263228.0,"Title":"What is the usefulness of the mean file with AlexNet neural network?","Body":"<p>When using an AlexNet neural network, be it with caffe or CNTK, it needs a mean file as input. What is this mean file for ? How does it affect the training ? How is it generated, only from training sample ?<\/p>\n","answers":[{"AnswerId":"35674999","CreationDate":"2016-02-27T20:22:16.137","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Mean subtraction removes the DC component from images. It has the geometric interpretation of centering the cloud of data around the origin along every dimension. It reduces the correlation between images which improves training. From my experience I can say that it improves the training accuracy significantly. It is computed from the training data. Computing mean from the testing data makes no sense.<\/p>\n"}]}
{"QuestionId":35647512,"AnswerCount":1,"Tags":"<theano><cudnn>","CreationDate":"2016-02-26T09:05:56.517","AcceptedAnswerId":35975660.0,"OwnerUserId":5256563.0,"Title":"Errors after Theano upgrade from 0.7 to bleeding edge","Body":"<p>I installed and used Theano 0.7 and everything was working perfectly. But now for the purpose of my future works, I need the bleeding edge version, and the installation went fine.<\/p>\n\n<p>But when I run this <a href=\"http:\/\/www.thibault.biz\/StackOverflow\/Theano\/tests.py\" rel=\"nofollow\">little test<\/a> (found into the Theano documentation), it generates many errors (<a href=\"http:\/\/www.thibault.biz\/StackOverflow\/Theano\/errors.txt\" rel=\"nofollow\">see here for the full list<\/a>).<\/p>\n\n<ul>\n<li>We can observe that the GPU is detected and used, but cuDNN is not found anymore:<\/li>\n<\/ul>\n\n<p><\/p>\n\n<ul>\n<li>And then I have an import error, I think it is also about cuDNN:<\/li>\n<\/ul>\n\n<p><\/p>\n\n<p>I've checked and cudnn.h is still in , in  we still find libcudnn.dylib which is a symbolic link to libcudnn.4.dylib, and everything in  points to <\/p>\n\n<p>Any idea?<\/p>\n\n<p>[EDIT] In my .profile we find:<\/p>\n\n<pre><\/pre>\n\n<p>In  there is a symbolic link to the cudnn library that is actually in .<\/p>\n\n<p>Here is the result from the command :<\/p>\n\n<pre><\/pre>\n\n<p>And here is the link between: , and in  I have <\/p>\n\n<p>[EDIT 2]<\/p>\n\n<pre><\/pre>\n\n<p>[EDIT 3]  <a href=\"http:\/\/www.thibault.biz\/StackOverflow\/Error.txt\" rel=\"nofollow\">Here is the last error<\/a> displayed. At least one part, because this error appears at each epoch.<\/p>\n\n<p>With :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35975660","CreationDate":"2016-03-13T20:49:20.020","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>This looks like a bug in Theano. It probably would work if they added <code>[\"-Wl,-rpath,%s\" % l for l in c_lib_dirs()]<\/code> to the compile args. You should report that upstream <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>It might work as a workaround if you add the path of <code>libcudnn.4.dylib<\/code> to your <code>LD_LIBRARY_PATH<\/code> (or maybe <code>DYLD_LIBRARY_PATH<\/code>) environment variable, because that is where <code>@rpath<\/code> will also look at, so that the path <code>@rpath\/libcudnn.4.dylib<\/code> can be resolved.<\/p>\n"}]}
{"QuestionId":35649011,"AnswerCount":1,"Tags":"<shell><curl><permissions><torch>","CreationDate":"2016-02-26T10:16:59.490","AcceptedAnswerId":35652430.0,"OwnerUserId":639973.0,"Title":"Torch install without curl","Body":"<p>I need to install Torch on server,<\/p>\n\n<p>but its installation guide requires the following line as the first step:<\/p>\n\n<pre><\/pre>\n\n<p>curl is not installed on my server, and I'm not the root, so I don't have permission to install it.<\/p>\n\n<p>I downloaded the script file in the URL, and chmod-ed it, <\/p>\n\n<p>but can't execute it, again due to permission.<\/p>\n\n<p>Is there a way to install torch without usage of curl?<\/p>\n\n<pre><\/pre>\n\n<p>resulted in <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35652430","CreationDate":"2016-02-26T13:01:08.363","ParentId":null,"OwnerUserId":"3685419","Title":null,"Body":"<p>Torch dependencies needs to be installed -  if you install it via this script, that'll require root privileges.<\/p>\n\n<p>If you saved the script, you can start it (without chmod) with<\/p>\n\n<pre><code>bash install-deps\n<\/code><\/pre>\n"}]}
{"QuestionId":35652757,"AnswerCount":1,"Tags":"<theano><deep-learning><lasagne>","CreationDate":"2016-02-26T13:16:51.387","AcceptedAnswerId":null,"OwnerUserId":3130081.0,"Title":"Best practice for training on large scale datasets like ImageNet using Theano\/Lasagne?","Body":"<p>I found that all of the examples of Theano\/Lasagne deal with small data set like mnist and cifar10 which can be loaded into memory completely.<\/p>\n\n<p>My question is how to write efficient code for training on large scale datasets?\nSpecifically, what is the best way to prepare mini-batches (including real time data augmentation) in order to keep the GPU busy? <\/p>\n\n<p>Maybe like using CAFFE's ImageDataLayer?\nFor example, I have a big txt file which contains all the image paths and labels.\nIt would be appreciated to show some code.<\/p>\n\n<p>Thank you very much!<\/p>\n","answers":[{"AnswerId":"35675198","CreationDate":"2016-02-27T20:41:27.350","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>In case the data doesn't fit into memory, a good way is to prepare the minibatches and store them into an HDF5 file, which is then used at training time.<\/p>\n\n<p>However, this does suffice when doing data augmentation as this is done on the fly. Because of Pythons global interpreter lock, images cannot already be loaded and preprocesed while the GPU is busy. \nThe best way around this, that I know of, is the Fuel library.\nFuel loads and preprocesses the minibatches in a different python process and then streams them to the training process over a TCP socket:\n<a href=\"http:\/\/fuel.readthedocs.org\/en\/latest\/server.html#data-processing-server\" rel=\"nofollow\">http:\/\/fuel.readthedocs.org\/en\/latest\/server.html#data-processing-server<\/a><\/p>\n\n<p>It additionally provides some functions to preprocess the data, such as scaling and mean subtraction:\n<a href=\"http:\/\/fuel.readthedocs.org\/en\/latest\/overview.html#transformers-apply-some-transformation-on-the-fly\" rel=\"nofollow\">http:\/\/fuel.readthedocs.org\/en\/latest\/overview.html#transformers-apply-some-transformation-on-the-fly<\/a><\/p>\n\n<p>Hope this helps.\nMichael<\/p>\n"}]}
{"QuestionId":35653017,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-02-26T13:29:15.350","AcceptedAnswerId":35659064.0,"OwnerUserId":2243015.0,"Title":"What are the learnable parameters in batch normalization?","Body":"<p>There are four variables in question: gamma, beta, moving average of mean, moving average of variance.<\/p>\n\n<p>Is it necessary to snapshot the moving averages, and load them in the testing time?<\/p>\n\n<p>A better question: <\/p>\n\n<p>For <a href=\"https:\/\/stackoverflow.com\/a\/34634291\/2243015\">this implementation<\/a> of batch normalization in tensorflow, do I need to transfer the batch-mean and batch-var from training time to testing time? If so, how can I achieve that in tensorflow?<\/p>\n","answers":[{"AnswerId":"35659064","CreationDate":"2016-02-26T18:27:09.667","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Yes - for any use of batch normalization, you <em>train<\/em> by normalizing based upon the statistics of the single batch, but then you run inference by using the long-term average of the statistics.<\/p>\n\n<p>You should save a copy of your mean and variance-holding-variables and restore it when you're doing tests.<\/p>\n\n<p>There shouldn't be any magic required:  They're just variables that will be saved and restored when you <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variables\/index.html\" rel=\"nofollow\">use the Saver<\/a>.<\/p>\n\n<p>In the specific implementation you reference, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#ExponentialMovingAverage\" rel=\"nofollow\">documentation for <code>tf.train.ExponentialMovingAverage<\/code><\/a> has a specific example of how to save and restore the moving averages for training and inference, respectively.<\/p>\n"}]}
{"QuestionId":35656643,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-02-26T16:16:39.553","AcceptedAnswerId":35657724.0,"OwnerUserId":4618605.0,"Title":"tf.initialize_variables() inconvenience - FailedPreconditionError tensorflow","Body":"<p>I want to be able to initialize tensorflow variables with a specific list using . So, to test the case I am trying this but it fails:<\/p>\n\n<pre><\/pre>\n\n<p>with the error:<\/p>\n\n<pre><\/pre>\n\n<p>What is really weird to me, is that it works this way:<\/p>\n\n<pre><\/pre>\n\n<p>So, the same happens if I create my own list of variables that I want to initialize. Probably I am misunderstanding something of the tensorflow mechanisms. Any ideas? I'd be glad if I don't have to do that using a loop for each variable.<\/p>\n","answers":[{"AnswerId":"35657724","CreationDate":"2016-02-26T17:09:54.543","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It's hard to tell without seeing the code, but my guess is that you're initializing one variable from another variable, which is a little tricky in TensorFlow. For example, I imagine you're doing something like this:<\/p>\n\n<pre><code>v_1 = tf.Variable(...)\nv_2 = tf.Variable(v_1)\n<\/code><\/pre>\n\n<p>This will fail if you try to do <code>sess.run(tf.initalize_all_variables())<\/code> because it will attempt to initialize <code>v_1<\/code> and <code>v_2<\/code> in parallel, but the initialization of <code>v_2<\/code> will fail (because its input is currently uninitialized).<\/p>\n\n<p>If you iterate over the <code>tf.all_variables()<\/code> list in increasing order, it will work, because <code>v_1<\/code> is added to the list before <code>v_2<\/code>, so running things sequentially will cause <code>v_1<\/code> to be initialized before <code>v_2<\/code>. (At a guess, if you iterated over <code>tf.all_variables()<\/code> in <em>reverse<\/em> order, it would fail with a similar error.)<\/p>\n\n<p>The fix is a bit subtle: any time you initialize a variable <code>v_2<\/code> from another variable <code>v_1<\/code>, you should use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.initialized_value\" rel=\"nofollow\"><code>v_1.initialized_value()<\/code><\/a> as the argument to <code>v_2<\/code>'s constructor:<\/p>\n\n<pre><code>v_1 = tf.Variable(...)\nv_2 = tf.Variable(v_1.initialized_value())\n<\/code><\/pre>\n\n<p>Using <code>v_1.initialized_value<\/code> adds a control dependency between the two initializers, so that <code>v_1<\/code> is initialized before <code>v_2<\/code>, even if you try to initialize them in parallel.<\/p>\n"}]}
{"QuestionId":35657003,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-26T16:33:47.697","AcceptedAnswerId":36182018.0,"OwnerUserId":4650495.0,"Title":"Aggregate each element of tensor in tensorflow","Body":"<p>I have a tensor 3-D as follow:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to convert it to 2-D tensor in tensorflow in graph where if \nI have a True in any element the aggregate result should be True, otherwise mush be False. Like follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36182018","CreationDate":"2016-03-23T15:17:03.320","ParentId":null,"OwnerUserId":"1990276","Title":null,"Body":"<p>This should do exactly what you want (directly from the documentation):<\/p>\n\n<pre><code># 'x' is [[True,  True]\n#         [False, False]]\ntf.reduce_any(x) ==&gt; True\ntf.reduce_any(x, 0) ==&gt; [True, True]\ntf.reduce_any(x, 1) ==&gt; [True, False]   &lt;== this is what you need\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#reduce_any\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#reduce_any<\/a><\/p>\n"}]}
{"QuestionId":35657015,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-26T16:34:17.950","AcceptedAnswerId":null,"OwnerUserId":5948619.0,"Title":"TFRecords and record shuffling","Body":"<p>My understanding is that it is good practice to shuffle training samples for each epoch so that each mini-batch contains a nice random sample of the entire dataset. If I convert my entire data-set into a single file containing TFRecords then how is this shuffling to be achieved short of loading the entire data-set? My understanding is that there is no efficient random access to TFRecord files. So, to be specific, I am looking for guidance as to how TFRecord files are used in this scenario.<\/p>\n","answers":[{"AnswerId":"47323386","CreationDate":"2017-11-16T06:53:31.393","ParentId":null,"OwnerUserId":"4259486","Title":null,"Body":"<p>Actually now you don't have to worry about shuffling before saving to TFRecords. It's because (currently) recommended method to read TFRecords uses <code>tf.data.TFRecordDataset<\/code> which implements <code>.shuffle()<\/code> method.<\/p>\n"},{"AnswerId":"35658968","CreationDate":"2016-02-26T18:21:34.297","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>It's not - you can improve the mixing somewhat by sharding your input into multiple input data files, and then treating them <a href=\"https:\/\/stackoverflow.com\/questions\/34258043\/getting-good-mixing-with-many-input-datafiles-in-tensorflow\">as explained in this answer<\/a>.<\/p>\n\n<p>If you need anything close to \"perfect\" shuffling, you would need to read it into memory, but in practice for most things, you'll probably get \"good enough\" shuffling by just splitting into 100 or 1000 files and then using a shuffle queue that's big enough to hold 8-16 files worth of data.<\/p>\n\n<p>I have an itch in the back of my head to write an external random shuffle queue that can spill to disk, but it's very low on my priority list -- if someone wanted to contribute one, I'm volunteering to review it. :)<\/p>\n"}]}
{"QuestionId":35658519,"AnswerCount":1,"Tags":"<python><numpy><theano><keras>","CreationDate":"2016-02-26T17:53:32.133","AcceptedAnswerId":35658764.0,"OwnerUserId":5606352.0,"Title":"Python\/Keras\/Theano - ValueError: Dimension mismatch; shapes are (98, 10), (98, 1)","Body":"<p>I've been trying to run a neural network with the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I haven't been able to find out what this error means:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I understand, the shapes should be equal, like  or  and that's what causing the problem. Is that right?<\/p>\n\n<p>If yes, does anyone know where can I fix this in the code or in the dataset? What does that 10 and 1 mean?<\/p>\n\n<p>If no, can anyone explain to me what is happening?<\/p>\n\n<p><br><br><\/p>\n\n<hr>\n\n<hr>\n\n<p><b>EXTRA INFORMATION:<\/b><\/p>\n\n<p>Variable predictor_train:<\/p>\n\n<pre><\/pre>\n\n<p>Variable target_train:<\/p>\n\n<pre><\/pre>\n\n<p>Full Traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35658764","CreationDate":"2016-02-26T18:09:00.307","ParentId":null,"OwnerUserId":"3076438","Title":null,"Body":"<p>The mismatch is between the expected output dimension (98,10) and the dimension of the data you're using (98,1)<\/p>\n\n<p>That is because you're using a sample code that is supposed to do classification on a 10-class database. If you want to do prediction then change your last layer to<\/p>\n\n<pre><code>model.add(Dense(output_dim=1, init=\"glorot_uniform\"))\n<\/code><\/pre>\n\n<hr>\n\n<p>Also, I think you will have problems with your cost function. You're not supposed to use absolute percentage error if you have continuous data. Change this<\/p>\n\n<pre><code>model.compile(loss=\"mean_absolute_percentage_error\", optimizer=\"rmsprop\")\n<\/code><\/pre>\n\n<p>to this maybe<\/p>\n\n<pre><code>model.compile(loss=\"mean_squared_error\", optimizer=\"rmsprop\")\n<\/code><\/pre>\n"}]}
{"QuestionId":35659055,"AnswerCount":1,"Tags":"<lua><torch><luarocks>","CreationDate":"2016-02-26T18:26:40.337","AcceptedAnswerId":35659797.0,"OwnerUserId":2886575.0,"Title":"luarocks install local source","Body":"<p>I am making edits to a file in the nn Library.  It has a rockspec file.  When I luarocks install that rockspec file, it downloads the git source. I want it to use my local source.  The rockspec file has all sorts of compile time command line gibberish that I don't want to invoke manually. Something like  would be great.<\/p>\n","answers":[{"AnswerId":"35659797","CreationDate":"2016-02-26T19:09:18.513","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>For local source, you need to go into the directory and issue:<\/p>\n\n<pre><code>luarocks make rocks\/nn-scm-1.rockspec\n<\/code><\/pre>\n"}]}
{"QuestionId":35659531,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-02-26T18:53:55.200","AcceptedAnswerId":35666153.0,"OwnerUserId":4723732.0,"Title":"Could I install Caffe framework on my laptop?","Body":"<p>My graphics card<\/p>\n\n<pre><\/pre>\n\n<p>So i do not have NVIDIA card,does this mean I can not install it?<\/p>\n","answers":[{"AnswerId":"35666153","CreationDate":"2016-02-27T05:36:06.687","ParentId":null,"OwnerUserId":"2915032","Title":null,"Body":"<p>No, it does not. You can install it with <code>CPU_ONLY=ON<\/code> variable in cmake. Of course, <strong>CUDA<\/strong> will be unavailable for you, but you still can use <strong>caffe<\/strong> on CPU.<\/p>\n\n<p>Moreover, you can try to checkout <em>opencl<\/em> caffe branch to utilize your AMD hardware (but I didn't try it yet).<\/p>\n"}]}
{"QuestionId":35659897,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><softmax>","CreationDate":"2016-02-26T19:15:29.453","AcceptedAnswerId":null,"OwnerUserId":5987713.0,"Title":"Using the softmax layer within the objective function itself","Body":"<p>This is going to be long and hard to describe so apologies in advance.<\/p>\n\n<p>I have a regular CNN like network with standard MLP layers on top of it. On top of the MLP, I have a softmax layer too, however, unlike conventional networks, this is NOT fully connected to the MLP below and it consists of subgroups.<\/p>\n\n<p>To further describe the softmax, it looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>There are many more groups. Each group has a softmax that is independent from the other groups. So it is in a way, several independent classifications (even though it actually is not).<\/p>\n\n<p>What I need is for the index of the activated neuron to be monotonically increasing between groups. For example, if I have Neuron5 in Group A activated, I want the activated neuron in group B to be >=5.  Same with Group B and Group C and so on..<\/p>\n\n<p>This softmax layer containing all the neurons for all groups is actually NOT my last layer and it is interestingly an intermediate one.<\/p>\n\n<p>To achieve this monotonicity,  I add another term to my loss function that penalizes non monotonic activated neuron indices. Here is some of the code:<\/p>\n\n<p>The code for softmax layer and its output:<\/p>\n\n<pre><\/pre>\n\n<p>The estimated_yps are passed onto a function that quantifies monotonicity:<\/p>\n\n<pre><\/pre>\n\n<p>and the loss function is:<\/p>\n\n<pre><\/pre>\n\n<p>Now my problem is, when I do not use SOME_OTHER_THINGS that are conventional metrics,  I get  for the monotonocity metric.<\/p>\n\n<p>Seems like gradients are not defined when the softmax layer outputs are used like this.<\/p>\n\n<p>Am I doing something wrong?   Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"35663283","CreationDate":"2016-02-26T22:59:22.923","ParentId":null,"OwnerUserId":"5987713","Title":null,"Body":"<p>Apologies.. I realized that the problem is that tf.argmax function obviously does not have a gradient defined.<\/p>\n"}]}
{"QuestionId":35660713,"AnswerCount":1,"Tags":"<python><theano><theano-cuda>","CreationDate":"2016-02-26T20:02:55.567","AcceptedAnswerId":null,"OwnerUserId":5987916.0,"Title":"theano throwing redefinition of argument 'cl-version' when importing","Body":"<p>I'm new to theano.\nAfter following this guide: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a> I managed to work with theano properly. Later I upgraded to theano dev version and now I keep getting this exception:<\/p>\n\n<pre><\/pre>\n\n<p>The .theanorc file:<\/p>\n\n<pre><\/pre>\n\n<p>the nvcc.profile:<\/p>\n\n<pre><\/pre>\n\n<p>I defenitely found out that for some reason FLAGS from .theanorc is being called twice, if I put it twice in .theanorc, I get it 4 times in exception.\nBut I have no clue where could be those twice-mentioned flags.\nHow does this theano-nvcc-cuda-?? chain work? Who is calling nvcc and causing this issue?<\/p>\n\n<p>BTW if I don't specify cl-version, I get <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40638668","CreationDate":"2016-11-16T17:36:59.890","ParentId":null,"OwnerUserId":"5987916","Title":null,"Body":"<p>seems to be issue in dev version of theano, as stated here: github.com\/Theano\/Theano\/issues\/4091 <\/p>\n"}]}
{"QuestionId":35661677,"AnswerCount":1,"Tags":"<python><numpy><initialization><tensorflow><interactive-session>","CreationDate":"2016-02-26T21:01:25.703","AcceptedAnswerId":35662013.0,"OwnerUserId":3314143.0,"Title":"TensorFlow: generating a random constant","Body":"<p>In ipython I imported  and  and created an TensorFlow .\nWhen I am running or initializing some normal distribution with numpy input, everything runs fine:<\/p>\n\n<pre><\/pre>\n\n<p>Returns:<\/p>\n\n<pre><\/pre>\n\n<p>Just as expected.<\/p>\n\n<p>...but when I use the Tensorflow normal distribution function:<\/p>\n\n<pre><\/pre>\n\n<p>...it raises a Type error saying:<\/p>\n\n<pre><\/pre>\n\n<p><strong>What am I missing here?<\/strong><\/p>\n\n<p>The output of:<\/p>\n\n<pre><\/pre>\n\n<p>alone returns the exact same thing which  generates -> a matrix of shape  with values taken from a normal distribution.<\/p>\n","answers":[{"AnswerId":"35662013","CreationDate":"2016-02-26T21:24:57.540","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> op takes a numpy array (or something implicitly convertible to a numpy array), and returns a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a> whose value is the same as that array. It does <em>not<\/em> accept a <code>tf.Tensor<\/code> as its argument.<\/p>\n\n<p>On the other hand, the <code>tf.random_normal()<\/code> op returns a <code>tf.Tensor<\/code> whose value is generated randomly according to the given distribution each time it runs. Since it returns a <code>tf.Tensor<\/code>, it cannot be used as the argument to <code>tf.constant()<\/code>. This explains the <code>TypeError<\/code> (which is unrelated to the use of <code>tf.InteractiveSession<\/code>, since it occurs when you build the graph).<\/p>\n\n<p>I'm assuming you want your graph to include a tensor that (i) is randomly generated on its first use, and (ii) constant thereafter. There are two ways to do this:<\/p>\n\n<ol>\n<li><p>Use NumPy to generate the random value and put it in a <code>tf.constant()<\/code>, as you did in your question:<\/p>\n\n<pre><code>some_test = tf.constant(\n    np.random.normal(loc=0.0, scale=1.0, size=(2, 2)).astype(np.float32))\n<\/code><\/pre><\/li>\n<li><p>(Potentially faster, as it can use the GPU to generate the random numbers) Use TensorFlow to generate the random value and put it in a <code>tf.Variable<\/code>:<\/p>\n\n<pre><code>some_test = tf.Variable(\n    tf.random_normal([2, 2], mean=0.0, stddev=1.0, dtype=tf.float32)\nsess.run(some_test.initializer)  # Must run this before using `some_test`\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":35662220,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-26T21:37:18.813","AcceptedAnswerId":null,"OwnerUserId":1002035.0,"Title":"tensorflow copy a variable in rnn","Body":"<p>related to this: <a href=\"https:\/\/stackoverflow.com\/questions\/33717772\/how-can-i-copy-a-variable-in-tensorflow\">How can I copy a variable in tensorflow<\/a><\/p>\n\n<p>I am trying to copy the values of a lstm decoding units to use it elsewhere for beamsearch. in pseudo code, I would like something like this:<\/p>\n\n<pre><\/pre>\n\n<p>The issue is that in tensorflow, the lstm variables are not generated during the call \"tf.nn.rnn_cell(...)\", but it is actually generated during the unrolling of the function call to rnn_decoder.<\/p>\n\n<p>I could set the scope to the \"tf.nn.seq2seq.rnn_decoder\" function call, but the actual initialization of the lstm weights are not transparent to me. How might I capture these values and re-use them to make an lstm cell with the same weights as the ones learned?<\/p>\n\n<p>thanks! <\/p>\n","answers":[{"AnswerId":"35663417","CreationDate":"2016-02-26T23:11:44.710","ParentId":null,"OwnerUserId":"1002035","Title":null,"Body":"<p>I think I figured it out.<\/p>\n\n<p>What you want is to set the scope for the decoder call to a particular value, say \"decoding\", in this line:<\/p>\n\n<pre><code>training_output = tf.nn.seq2seq.rnn_decoder(...scope=\"decoding\")\n<\/code><\/pre>\n\n<p>And later on when you want to use the lstm units you learned during the decoding, you set the variable scope to \"decoding\" again, and use scope.reuse_variables() to allow re-using of the variables of the decoding. Then simply use the \"lstm_decode\" as you would otherwise, which will be bound to the same value as before.<\/p>\n\n<pre><code>with tf.variable_scope(\"decoding\") as scope:\n  scope.reuse_variables()\n  ... use lstm_decode as usual ...\n<\/code><\/pre>\n\n<p>this way all the weights in the lstm_decode will be shared in these two sub-graphs, and whichever value it learned during the training will be present in the 2nd part as well.<\/p>\n"}]}
{"QuestionId":35666835,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-27T07:03:48.663","AcceptedAnswerId":null,"OwnerUserId":4557009.0,"Title":"TensorFlow: Plotting a graph that describes data points and decision boundary","Body":"<p>I would like to achieve something similar: <\/p>\n\n<p><a href=\"https:\/\/rootpy.github.io\/root_numpy\/_images\/plot_multiclass_1.png\" rel=\"nofollow\">https:\/\/rootpy.github.io\/root_numpy\/_images\/plot_multiclass_1.png<\/a><\/p>\n\n<p>What would be the most elegant solution? Get the weights, bias, function and data and plot it with some other tool or does TensorFlow have support for that?<\/p>\n","answers":[{"AnswerId":"38445742","CreationDate":"2016-07-18T20:47:47.707","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>As far as I know, Tensorflow does not directly support plotting decision boundaries.<\/p>\n\n<p>It is certainly not the most elegant solution, but you can create a grid. Classify each point of the grid and then plot it. For example:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n\n\"\"\"\nSolve the XOR problem with Tensorflow.\n\nThe XOR problem is a two-class classification problem. You only have four\ndatapoints, all of which are given during training time. Each datapoint has\ntwo features:\n\n      x    o\n\n      o    x\n\nAs you can see, the classifier has to learn a non-linear transformation of\nthe features to find a propper decision boundary.\n\"\"\"\n\n__author__ = \"Martin Thoma\"\n\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\nimport numpy as np\n\n# The training data\nXOR_X = [[0, 0], [0, 1], [1, 0], [1, 1]]  # Features\nXOR_Y = [[0], [1], [1], [0]]  # Class labels\nXOR_Y = [[1, 0], [0, 1], [0, 1], [1, 0]]  # Target values\n\nassert len(XOR_X) == len(XOR_Y)  # sanity check\n\n# The network\nnb_classes = 2\ninput_ = tf.placeholder(tf.float32,\n                        shape=[None, len(XOR_X[0])],\n                        name=\"input\")\ntarget = tf.placeholder(tf.float32,\n                        shape=[None, nb_classes],\n                        name=\"output\")\nnb_hidden_nodes = 2\n# enc = tf.one_hot([0, 1], 2)\nw1 = tf.Variable(tf.random_uniform([2, nb_hidden_nodes], -1, 1),\n                 name=\"Weights1\")\nw2 = tf.Variable(tf.random_uniform([nb_hidden_nodes, nb_classes], -1, 1),\n                 name=\"Weights2\")\nb1 = tf.Variable(tf.zeros([nb_hidden_nodes]), name=\"Biases1\")\nb2 = tf.Variable(tf.zeros([nb_classes]), name=\"Biases2\")\nactivation2 = tf.sigmoid(tf.matmul(input_, w1) + b1)\nhypothesis = tf.nn.softmax(tf.matmul(activation2, w2) + b2)\ncross_entropy = -tf.reduce_sum(target * tf.log(hypothesis))\ntrain_step = tf.train.GradientDescentOptimizer(0.1).minimize(cross_entropy)\n\n# Start training\ninit = tf.initialize_all_variables()\nwith tf.Session() as sess:\n    sess.run(init)\n\n    for i in range(100000):\n        sess.run(train_step, feed_dict={input_: XOR_X, target: XOR_Y})\n\n        if i % 10000 == 0:\n            print('Epoch ', i)\n            print('Hypothesis ', sess.run(hypothesis,\n                                          feed_dict={input_: XOR_X,\n                                                     target: XOR_Y}))\n            print('w1 ', sess.run(w1))\n            print('b1 ', sess.run(b1))\n            print('w2 ', sess.run(w2))\n            print('b2 ', sess.run(b2))\n            print('cost (ce)', sess.run(cross_entropy,\n                                        feed_dict={input_: XOR_X,\n                                                   target: XOR_Y}))\n\n            # Visualize classification boundary\n            xs = np.linspace(-5, 5)\n            ys = np.linspace(-5, 5)\n            pred_classes = []\n            for x in xs:\n                for y in ys:\n                    pred_class = sess.run(hypothesis,\n                                          feed_dict={input_: [[x, y]]})\n                    pred_classes.append((x, y, pred_class.argmax()))\n            xs_p, ys_p = [], []\n            xs_n, ys_n = [], []\n            for x, y, c in pred_classes:\n                if c == 0:\n                    xs_n.append(x)\n                    ys_n.append(y)\n                else:\n                    xs_p.append(x)\n                    ys_p.append(y)\n            plt.plot(xs_p, ys_p, 'ro', xs_n, ys_n, 'bo')\n            plt.show()\n<\/code><\/pre>\n\n<p>which gives<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Vi8xX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Vi8xX.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35667330,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><classification><theano>","CreationDate":"2016-02-27T08:09:31.043","AcceptedAnswerId":35667690.0,"OwnerUserId":2147347.0,"Title":"log-likelihood cost function: mean or sum?","Body":"<p>In <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\/blob\/d2764f288b4e58e12bd492953d1c1a0b43d92e21\/code\/logistic_sgd.py#L118-L135\" rel=\"nofollow\">this code<\/a> for computing the negative log-likelihood they say:<\/p>\n\n<blockquote>\n  <p>Note: we use the mean instead of the sum so that the learning rate is\n  less dependent on the batch size<\/p>\n<\/blockquote>\n\n<p>and this is how they get the negative log-likelihood:<\/p>\n\n<pre><\/pre>\n\n<p>This is true in many textbooks (e.g. <em>Pattern Recognition and Machine Learning<\/em> by Bishop), the negative log-likelihood is calculated by using the <strong>sum<\/strong> of each individual sample error rather than the <strong>mean<\/strong>. I still don't understand the note from the author. Every time when we calculate cost function should we use the mean rather than the sum? Even when we are not using batch?<\/p>\n","answers":[{"AnswerId":"35667690","CreationDate":"2016-02-27T08:51:08.410","ParentId":null,"OwnerUserId":"157726","Title":null,"Body":"<p>The difference between the mean and the sum is just the multiplication by 1\/N. <\/p>\n\n<p>The problem with using the sum is that the batch size (N) will influence your gradients. The learning rate indicates how much in the direction of the gradient you want to adjust your parameters. <\/p>\n\n<p>If your gradient is larger for larger batch sizes (N), it means that you will need to adjust the learning rate as you increase the batch size (N).<\/p>\n\n<p>In practice, in order to keep these two (learning rate and batch size) independent, it is common to use the mean instead of the sum. This makes the gradient magnitude independent of N.<\/p>\n\n<p>If you are not using a batch, then N=1 and the mean is the same as the sum.<\/p>\n"},{"AnswerId":"35667701","CreationDate":"2016-02-27T08:52:45.520","ParentId":null,"OwnerUserId":"3076438","Title":null,"Body":"<p>Consider the two situations when you want to change the batch size of your algorithm lets say from 10 to 100:<\/p>\n\n<ul>\n<li><p>If you are using the sum of the log-likelihood cost function the average step that you're taking will be multiplied by 10 (because the sum will have 10x more terms) and then you will have to readjust the learning rate<\/p><\/li>\n<li><p>If you are using the mean, then that result will be divided by 100 instead of 10, and your steps will be naturally of the same order of magnitude<\/p><\/li>\n<\/ul>\n\n<p>So, if you're not using mini-batches or adjusting the batch size, don't worry about this<\/p>\n"}]}
{"QuestionId":35668236,"AnswerCount":2,"Tags":"<neural-network><caffe>","CreationDate":"2016-02-27T09:47:03.390","AcceptedAnswerId":35686577.0,"OwnerUserId":4229825.0,"Title":"Accuracy in Caffe keeps on 0.1 and does not change","Body":"<p>Through all training process, accuracy is 0.1. What am I doing wrong?<\/p>\n\n<p>Model, solver and part of log here:\n<a href=\"https:\/\/gist.github.com\/yutkin\/3a147ebbb9b293697010\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/yutkin\/3a147ebbb9b293697010<\/a><\/p>\n\n<p>Topology in png format: \n<img src=\"https:\/\/i.imgur.com\/xTK1fxS.png\" alt=\"\"><\/p>\n\n<p>P.S. I am using the latest version of Caffe and g2.2xlarge instance on AWS. <\/p>\n","answers":[{"AnswerId":"35686577","CreationDate":"2016-02-28T18:34:49.170","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>You're working on CIFAR-10 dataset which has 10 classes. When the training of a network commences, the first guess is usually random due to which your accuracy is <code>1\/N<\/code>, where <code>N<\/code> is the number of classes. In your case it is <code>1\/10<\/code>, i.e., <code>0.1<\/code>. If your accuracy stays the same over time it implies that your network isn't learning anything. This may happen due to a large learning rate. The basic idea of training a network is that you calculate the loss and propagate it back. The gradients are multiplied with the learning rate and added to the current weights and biases. If the learning rate is too big you may overshoot the local minima every time. If it is too small, the convergence will be slow. I see that your <code>base_lr<\/code> here is 0.01. As far as my experience goes, this is somewhat large. You may want to keep it at 0.001 in the beginning and then go on reducing it by a factor of 10 whenever you observe that the accuracy is not improving. But then anything below 0.00001 usually doesn't make much of a difference. The trick is to observe the progress of the training and make parameter changes as and when required.<\/p>\n"},{"AnswerId":"39713073","CreationDate":"2016-09-26T21:59:24.833","ParentId":null,"OwnerUserId":"5084931","Title":null,"Body":"<p>I know the thread is quite old but maybe my answer helps somebody. I experienced the same problem with an accuracy like a random guess. <\/p>\n\n<p>What helped was to set the number of outputs of the last layer before the accuracy layer to the number of labels. <\/p>\n\n<p>In your case that should be the <code>ip2<\/code> layer. Open the model definition of your net and set <code>num_outputs<\/code> to the number of labels. <\/p>\n\n<p>See Section 4.4 for more information: <a href=\"http:\/\/adilmoujahid.com\/posts\/2016\/06\/introduction-deep-learning-python-caffe\/\" rel=\"nofollow\">A Practical Introduction to Deep Learning with Caffe and Python<\/a><\/p>\n"}]}
{"QuestionId":35669032,"AnswerCount":0,"Tags":"<python><anaconda><theano>","CreationDate":"2016-02-27T11:16:22.620","AcceptedAnswerId":null,"OwnerUserId":5310324.0,"Title":"Installing Theano on Windows 10 - error when importing theano","Body":"<p>I've followed numerous installation tutorials for windows over the past few days. I use anaconda on windows 10. But I always get the following error when importing theano:<\/p>\n\n<pre><\/pre>\n\n<p>Could anyone help me out on this? <\/p>\n\n<p>I have run <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35671516,"AnswerCount":0,"Tags":"<neural-network><time-series><keras>","CreationDate":"2016-02-27T15:22:23.700","AcceptedAnswerId":null,"OwnerUserId":5276948.0,"Title":"Temporal weights, unequal length of time series, how to align?","Body":"<p>I have time series of unequal length that I feed into a RNN-LSTM created in keras. My understanding is that if I choose sample_weight_mode='temporal' the network will learn a different set of weights for every time step. Please, correct me if I am wrong.<\/p>\n\n<p>I guess that the first time step of every time series will be handled by one set of weights, the second step of every series by another set and so on. Accordingly, the last step of the longest time series will get a different set of weights than the last step of the shortest time series. How can I get the same set of weights for the last step of every series?<\/p>\n\n<p>Viewed differently: I believe that all the time series are left-aligned when using sample_weight_mode='temporal' whereas I would like to have them right-aligned. How could I achieve that?<\/p>\n","answers":[]}
{"QuestionId":35673874,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-27T18:42:50.520","AcceptedAnswerId":35675045.0,"OwnerUserId":5551505.0,"Title":"Tensorflow error \"shape Tensorshape() must have rank 1\"","Body":"<pre><\/pre>\n\n<p>I am trying to run the tensorflow convolutional model over my own images of dimension of[32x32x3] images. The images are being read properly and are being assigned to the placeholder, during training. The problem arises during running the train_step op. When I execute the graph I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>But when I see the example <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/5_convolutional_net.py\">here<\/a>, the images are in the form of [batch_size,height,width,depth] tensors only. And this example works fine. \nAm I missing anything?<\/p>\n","answers":[{"AnswerId":"35675045","CreationDate":"2016-02-27T20:26:27.817","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I think the error is coming from this line:<\/p>\n\n<pre><code>cross_entropy= -tf.reduce_sum(tf.cast(image_batch[1],tf.float32)*tf.log(y_conv))\n<\/code><\/pre>\n\n<p><code>image_batch<\/code> is a 5-D tensor, with shape <code>[2, 1, 32, 32, 3]<\/code>, where 2 is the <code>batch_size<\/code> argument to <code>tf.train.batch()<\/code>, and the 1 was added by the earlier <code>image = tf.reshape(image, [-1, 32, 32, 3])<\/code>. (<strong>N.B.<\/strong> This reshape is unnecessary, since <code>tf.train.batch()<\/code> already adds a batch dimension, and you end up having to <em>undo<\/em> the effect of the reshape when you later construct <code>image_4d<\/code>).<\/p>\n\n<p>In TensorFlow, the slicing operation (i.e. <code>image_batch[1]<\/code>) is slightly less flexible than in NumPy. The number of dimensions specified in the slice must be equal to the rank of the tensor: i.e. you must specify all five dimensions for this to work. You could specify <code>image_batch[1, :, :, :, :]<\/code> to get a 4-D slice of <code>image_batch<\/code>.<\/p>\n\n<p>I noticed a few other issues in your program, however:<\/p>\n\n<ol>\n<li><p>The <code>cross_entropy<\/code> calculation seems strange. Typically this uses a predicted label and compares it to the known <em>label<\/em>, as opposed to the image data.<\/p><\/li>\n<li><p>The feed on the training step seems to have no effect, because placeholders <code>x<\/code> and <code>y_<\/code> are unused in your program. Furthermore, you appear to be feeding a <code>tf.Tensor<\/code> (indeed, an illegal slice of <code>image_batch<\/code>), so this will fail when you execute that statement. If you intend to use feeding, you should feed in NumPy arrays holding the input data.<\/p><\/li>\n<li><p>If you're not using feeding&mdash;i.e. using the <code>tf.WholeFileReader<\/code> as shown in your program&mdash;you will need to call <code>tf.train.start_queue_runners()<\/code> to get started. Otherwise your program will hang, waiting for input.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35674073,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-27T18:59:38.750","AcceptedAnswerId":35675700.0,"OwnerUserId":5703903.0,"Title":"Compute status: Not found: Tensor name \"input_producer\/limit_epochs\/epochs\" not found in checkpoint files","Body":"<p>I'm using the CIFAR10 example. I trained the net as it is with the code provided. The training was done successfully. As I wanted to evaluate each example only once on my data set, I have modified inputs in cifar10_input.py to the following.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have isolated the problem to the following:<\/p>\n\n<blockquote>\n  <p>tf.train_string_input_producer([filename], num_epochs = 1)<\/p>\n<\/blockquote>\n\n<p>If I don't set num_epochs = 1, everything works fine as it is. If I do, I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>Thank you for your help!<\/p>\n\n<p>EDIT 3 @mrry:<\/p>\n\n<p>It still fails. Here's the trace.<\/p>\n\n<pre><\/pre>\n\n<p>EDIT 4 @mrry:<\/p>\n\n<p>softmax_linear\/biases\/ExponentialMovingAverage<\/p>\n\n<pre><\/pre>\n\n<p>EDIT 5 @mrry:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38853530","CreationDate":"2016-08-09T14:27:37.097","ParentId":null,"OwnerUserId":"6132001","Title":null,"Body":"<p>Another way of eliminating the implicit variable <code>\"input_producer\/limit_epochs\/epochs\"<\/code> is to only load the trainable variables:<\/p>\n\n<pre><code>saver = tf.train.Saver(tf.trainable_variables())\n<\/code><\/pre>\n"},{"AnswerId":"35675700","CreationDate":"2016-02-27T21:31:20.007","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> In <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/77c2042e77a11ee442ecc7e369cd91d91e4a98c3\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py#L134\" rel=\"nofollow\"><code>cifar10_eval.py<\/code><\/a>, change the saver constructor so that it is:<\/p>\n\n<pre><code>saver = tf.train.Saver([v for v in variables_to_restore\n                        if v != \"input_producer\/limit_epochs\/epochs\"])\n<\/code><\/pre>\n\n<p>This problem arises because <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#string_input_producer\" rel=\"nofollow\"><code>tf.train.string_input_producer()<\/code><\/a> internally creates a variable (called <code>\"input_producer\/limit_epochs\/epochs\"<\/code>) when its <code>num_epochs<\/code> argument is not <code>None<\/code>. When, in <code>cifar10_eval.py<\/code> a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/77c2042e77a11ee442ecc7e369cd91d91e4a98c3\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py#L134\" rel=\"nofollow\"><code>tf.train.Saver<\/code> is created<\/a>, it uses <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#all_variables\" rel=\"nofollow\"><code>tf.all_variables()<\/code><\/a>, which includes the implicitly-created variable from the <code>tf.nn.string_input_producer()<\/code>. This list of variables determines the set of names that TensorFlow looks up in the checkpoint file.<\/p>\n\n<p>Currently there isn't a great way to refer to implicitly created variables, other than by their name. Therefore, the best fix is to exclude the variable from the <code>Saver<\/code> constructor by name.<\/p>\n"}]}
{"QuestionId":35675185,"AnswerCount":0,"Tags":"<matlab><computer-vision><caffe><conv-neural-network><lmdb>","CreationDate":"2016-02-27T20:40:44.280","AcceptedAnswerId":null,"OwnerUserId":5991420.0,"Title":"Segmented Images to LMDB for Caffe","Body":"<p>I'm using a dataset of achitectural images to try to finetune a caffe model. This dataset consists of images in one folder, and segmentation ground-truth in another folder, stored as a png image with each colour representing a label (eg. window, wall, sky.)<\/p>\n\n<p>I'm not sure how to convert this to the lmdb format, since most tutorials don't focus on pixel-wise labeling or labels stored as colours in a png image.<\/p>\n\n<p>Any hints would be great!<\/p>\n","answers":[]}
{"QuestionId":35676138,"AnswerCount":1,"Tags":"<lua><neural-network><deep-learning><torch><conv-neural-network>","CreationDate":"2016-02-27T22:17:07.453","AcceptedAnswerId":35688733.0,"OwnerUserId":700023.0,"Title":"torch nn : how to get the dimensions of the outputs (feature maps)?","Body":"<p>In a convolutional module many parameters may alter the dimensionality of the input. Is there any way to get the dimensionality of the output of the convolutional module, or in general of any module?<\/p>\n\n<p>I tried looking at the methods exposed in  but there are none that suggest giving me that information.<\/p>\n\n<p>Moreover, the output tensor seems to have these dimensions:<\/p>\n\n<pre><\/pre>\n\n<p>any ideas?<\/p>\n\n<p>I realize that I didn't even set the input dimensionality, except for the number of channels (that would be ). How to set it? Why is it not mandatory?<\/p>\n","answers":[{"AnswerId":"35688733","CreationDate":"2016-02-28T21:51:42.937","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>The output size depends on your input size. You can try something like this:<\/p>\n\n<pre><code>th&gt; require 'nngraph'\nth&gt; conv1 = nn.SpatialConvolution(3, 96, 5, 5, 1, 1, 2, 2)()\nth&gt; model = nn.gModule({conv1},{conv1})\nth&gt; x = torch.rand(3,20,20)\nth&gt; y = model:forward(x)\nth&gt; y:size()\n\n96\n20\n20\n[torch.LongStorage of size 3]\n<\/code><\/pre>\n"}]}
{"QuestionId":35677724,"AnswerCount":4,"Tags":"<python><c++><machine-learning><tensorflow>","CreationDate":"2016-02-28T01:40:08.017","AcceptedAnswerId":35678837.0,"OwnerUserId":2473294.0,"Title":"TensorFlow, why was python the chosen language?","Body":"<p>I recently started studying deep learning and other ML techniques, and I started searching for frameworks that simplify the process of build a net and training it, then I found TensorFlow, having little experience in the field, for me, it seems that speed is a big factor for making a big ML system even more if working with deep learning, so why python was chosen by Google to make TensorFlow? Wouldn't it be better to make it over an language that can be compiled and not interpreted?<\/p>\n\n<p>What are the advantages of using Python over a language like C++ for machine learning?<\/p>\n","answers":[{"AnswerId":"45124752","CreationDate":"2017-07-16T03:44:42.857","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>TF is not written in python. It is written in C++ (and uses high-performant numerical <a href=\"https:\/\/en.wikipedia.org\/wiki\/Eigen_(C%2B%2B_library)\" rel=\"noreferrer\">libraries<\/a> and <a href=\"https:\/\/www.geforce.com\/hardware\/technology\/cuda\" rel=\"noreferrer\">CUDA<\/a> code) and you can check this by looking at their <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"noreferrer\">github<\/a>. So <a href=\"https:\/\/youtu.be\/t64ortpgS-E?t=4m38s\" rel=\"noreferrer\">the core is written not in python<\/a> but TF provide an interface to many other languages (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/\" rel=\"noreferrer\">python, C++, Java, Go<\/a>)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yAI6m.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yAI6m.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If you come from a data analysis world, you can think about it like numpy (not written in python, but provides an interface to Python) or if you are a web-developer - think about it as a database (PostgreSQL, MySQL, which can be invoked from Java, Python, PHP)<\/p>\n\n<hr>\n\n<p>Python frontend (the language in which people write models in TF) is the most popular due to <a href=\"https:\/\/www.quora.com\/Why-is-Python-so-popular-in-machine-learning\" rel=\"noreferrer\">many<\/a> <a href=\"https:\/\/www.quora.com\/Why-is-Python-considered-a-good-language-for-AI-and-Machine-Learning\" rel=\"noreferrer\">reasons<\/a>. In my opinion the main reason is historical: majority of ML users already use it (another popular choice is R) so if you will not provide an interface to python, your library is most probably doomed to obscurity.<\/p>\n\n<hr>\n\n<p>But being written in python does not mean that your model is executed in python. On the contrary, if you written your model in the right way Python is never executed during the evaluation of the TF graph (except of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"noreferrer\">tf.py_func()<\/a>, which exists for debugging and should be avoided in real model exactly because it is executed on Python's side).<\/p>\n\n<p>This is different from for example numpy. For example if you do <code>np.linalg.eig(np.matmul(A, np.transpose(A))<\/code> (which is <code>eig(AA')<\/code>), the operation will compute transpose in some fast language (C++ or fortran), return it to python, take it from python together with A, and compute a multiplication in some fast language and return it to python, then compute eigenvalues and return it to python. So nonetheless expensive operations like matmul and eig are calculated efficiently, you still lose time by moving the results to python back and force. <strong>TF does not do it<\/strong>, once you defined the graph your tensors flow not in python but in C++\/CUDA\/something else.<\/p>\n"},{"AnswerId":"35677761","CreationDate":"2016-02-28T01:46:32.893","ParentId":null,"OwnerUserId":"1530134","Title":null,"Body":"<p>Python allows you to create extension modules using C and C++, interfacing with native code, and still getting the advantages that Python gives you.<\/p>\n\n<p>TensorFlow uses Python, yes, but it also contains large amounts of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=&amp;type=Code\" rel=\"nofollow\">C++<\/a>.<\/p>\n\n<p>This allows a simpler interface for experimentation with less human-thought overhead with Python, and add performance by programming the most important parts in C++.<\/p>\n"},{"AnswerId":"53675411","CreationDate":"2018-12-07T19:00:55.177","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>The latest ratio you can check from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">here<\/a> shows inside TensorFlow C++ takes ~50% of code, and Python takes ~40% of code.<\/p>\n\n<p>Both C++ and Python are the official languages at Google so there is no wonder why this is so. If I would have to provide fast regression where C++ and Python are present... <\/p>\n\n<p>C++ is inside the computational algebra, and Python is used for everything else including for the testing. Knowing how ubiquitous the testing is today it is no wonder why Python code contributes that much to TF.<\/p>\n"},{"AnswerId":"35678837","CreationDate":"2016-02-28T04:52:37.567","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>The most important thing to realize about TensorFlow is that, for the most part, <em>the core is not written in Python<\/em>:  It's written in a combination of highly-optimized C++ and CUDA (Nvidia's language for programming GPUs).  Much of that happens, in turn, by using <a href=\"http:\/\/eigen.tuxfamily.org\/index.php?title=Main_Page\" rel=\"noreferrer\">Eigen<\/a> (a high-performance C++ and CUDA numerical library) and <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"noreferrer\">NVidia's cuDNN<\/a> (a very optimized DNN library for <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"noreferrer\">NVidia GPUs<\/a>, for functions such as <a href=\"https:\/\/en.wikipedia.org\/wiki\/Convolutional_neural_network\" rel=\"noreferrer\">convolutions<\/a>).<\/p>\n\n<p>The model for TensorFlow is that the programmer uses \"some language\" (most likely Python!) to express the model.  This model, written in the TensorFlow constructs such as:<\/p>\n\n<pre><code>h1 = tf.nn.relu(tf.matmul(l1, W1) + b1)\nh2 = ...\n<\/code><\/pre>\n\n<p>is not actually executed when the Python is run.  Instead, what's actually created is a <a href=\"https:\/\/www.tensorflow.org\/get_started\/graph_viz\" rel=\"noreferrer\">dataflow graph<\/a> that says to take particular inputs, apply particular operations, supply the results as the inputs to other operations, and so on.  <em>This model is executed by fast C++ code, and for the most part, the data going between operations is never copied back to the Python code<\/em>.<\/p>\n\n<p>Then the programmer \"drives\" the execution of this model by pulling on nodes -- for training, usually in Python, and for serving, sometimes in Python and sometimes in raw C++:<\/p>\n\n<pre><code>sess.run(eval_results)\n<\/code><\/pre>\n\n<p>This one Python (or C++ function call) uses either an in-process call to C++ or an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Remote_procedure_call\" rel=\"noreferrer\">RPC<\/a> for the distributed version to call into the C++ TensorFlow server to tell it to execute, and then copies back the results.<\/p>\n\n<p><strong>So, with that said, let's re-phrase the question:  Why did TensorFlow choose  Python as the first well-supported language for expressing and controlling the training of models?<\/strong><\/p>\n\n<p>The answer to that is simple:  Python is probably <em>the<\/em> most comfortable language for a large range of data scientists and machine learning experts that's also that easy to integrate and have control a C++ backend, while also being general, widely-used both inside and outside of Google, and open source.  Given that with the basic model of TensorFlow, the performance of Python isn't that important, it was a natural fit.  It's also a huge plus that <a href=\"http:\/\/www.numpy.org\/\" rel=\"noreferrer\">NumPy<\/a> makes it easy to do pre-processing in Python -- also with high performance -- before feeding it in to TensorFlow for the truly CPU-heavy things.<\/p>\n\n<p>There's also a bunch of complexity in expressing the model that isn't used when executing it -- shape inference (e.g., if you do matmul(A, B), what is the shape of the resulting data?) and automatic <a href=\"https:\/\/en.wikipedia.org\/wiki\/Gradient\" rel=\"noreferrer\">gradient<\/a> computation.  It turns out to have been nice to be able to express those in Python, though I think in the long term they'll probably move to the C++ backend to make adding other languages easier.<\/p>\n\n<p>(The hope, of course, is to support other languages in the future for creating and expressing models.  It's already quite straightforward to run inference using several other languages -- C++ works now, someone from Facebook contributed <a href=\"https:\/\/golang.org\/\" rel=\"noreferrer\">Go<\/a> bindings that we're reviewing now, etc.)<\/p>\n"}]}
{"QuestionId":35678812,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-28T04:49:53.460","AcceptedAnswerId":35678932.0,"OwnerUserId":1044011.0,"Title":"reduce_sum by certain dimension","Body":"<p>I have two embeddings tensor  and , which looks like<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>what I want to do is calculate the L2 distance  element-wise. <\/p>\n\n<p>First I did a  to get<\/p>\n\n<pre><\/pre>\n\n<p>and then I want to do an element-wise reduce which returns <\/p>\n\n<pre><\/pre>\n\n<p>but  does not allow my to reduce by row. Any inputs would be appreciated. Thanks.<\/p>\n","answers":[{"AnswerId":"35678932","CreationDate":"2016-02-28T05:04:03.187","ParentId":null,"OwnerUserId":"5862726","Title":null,"Body":"<p>Add the <code>reduction_indices<\/code> argument with a value of 1, eg.:<\/p>\n\n<pre><code>tf.reduce_sum( tf.square( tf.sub( lhs, rhs) ), 1 )\n<\/code><\/pre>\n\n<p>That should produce the result you're looking for. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_sum\" rel=\"nofollow noreferrer\">Here is the documentation<\/a> on <code>reduce_sum()<\/code>.<\/p>\n"},{"AnswerId":"42756879","CreationDate":"2017-03-13T04:54:28.310","ParentId":null,"OwnerUserId":"1574779","Title":null,"Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops\/reduction#reduce_sum\" rel=\"noreferrer\">TensorFlow documentation<\/a>, <code>reduce_sum<\/code> function which takes four arguments.<\/p>\n\n<pre><code>tf.reduce_sum(input_tensor, axis=None, keep_dims=False, name=None, reduction_indices=None).\n<\/code><\/pre>\n\n<p>But <code>reduction_indices<\/code> has been deprecated. Better to use axis instead of. If the axis is not set, reduces all its dimensions.<\/p>\n\n<p>As an example,this is taken from the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops\/reduction#reduce_sum\" rel=\"noreferrer\">documentation<\/a>, <\/p>\n\n<pre><code># 'x' is [[1, 1, 1]\n#         [1, 1, 1]]\ntf.reduce_sum(x) ==&gt; 6\ntf.reduce_sum(x, 0) ==&gt; [2, 2, 2]\ntf.reduce_sum(x, 1) ==&gt; [3, 3]\ntf.reduce_sum(x, 1, keep_dims=True) ==&gt; [[3], [3]]\ntf.reduce_sum(x, [0, 1]) ==&gt; 6\n<\/code><\/pre>\n\n<p>Above requirement can be written in this manner,<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\na = np.array([[1,7,1],[1,1,1]])\nb = np.array([[0,0,0],[1,1,1]])\n\nxtr = tf.placeholder(\"float\", [None, 3])\nxte = tf.placeholder(\"float\", [None, 3])\n\npred = tf.reduce_sum(tf.square(tf.subtract(xtr, xte)),1)\n\n# Initializing the variables\ninit = tf.global_variables_initializer()\n\n# Launch the graph\nwith tf.Session() as sess:\n    sess.run(init)\n    nn_index = sess.run(pred, feed_dict={xtr: a, xte: b})\n    print nn_index\n<\/code><\/pre>\n"}]}
{"QuestionId":35678883,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-28T04:57:43.750","AcceptedAnswerId":null,"OwnerUserId":3528623.0,"Title":"TensorFlow: getting variable by name","Body":"<p>When using the TensorFlow Python API, I created a variable (without specifying its  in the constructor), and its  property had the value .  When I try to select this variable using , a new variable called  is created instead. How do I correctly select  instead of creating a new one? <\/p>\n\n<p>What I want to do is select the variable by name, and reinitialize it so I can finetune weights.<\/p>\n","answers":[{"AnswerId":"39027917","CreationDate":"2016-08-18T21:38:59.707","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>The <code>get_variable()<\/code> function creates a new variable or returns one created earlier by <code>get_variable()<\/code>. It won't return a variable created using <code>tf.Variable()<\/code>. Here's a quick example:<\/p>\n\n<pre><code>&gt;&gt;&gt; with tf.variable_scope(\"foo\"):\n...   bar1 = tf.get_variable(\"bar\", (2,3)) # create\n... \n&gt;&gt;&gt; with tf.variable_scope(\"foo\", reuse=True):\n...   bar2 = tf.get_variable(\"bar\")  # reuse\n... \n\n&gt;&gt;&gt; with tf.variable_scope(\"\", reuse=True): # root variable scope\n...   bar3 = tf.get_variable(\"foo\/bar\") # reuse (equivalent to the above)\n... \n&gt;&gt;&gt; (bar1 is bar2) and (bar2 is bar3)\nTrue\n<\/code><\/pre>\n\n<p>If you did not create the variable using <code>tf.get_variable()<\/code>, you have a couple options. First, you can use <code>tf.global_variables()<\/code> (as @mrry suggests):<\/p>\n\n<pre><code>&gt;&gt;&gt; bar1 = tf.Variable(0.0, name=\"bar\")\n&gt;&gt;&gt; bar2 = [var for var in tf.global_variables() if var.op.name==\"bar\"][0]\n&gt;&gt;&gt; bar1 is bar2\nTrue\n<\/code><\/pre>\n\n<p>Or you can use <code>tf.get_collection()<\/code> like so:<\/p>\n\n<pre><code>&gt;&gt;&gt; bar1 = tf.Variable(0.0, name=\"bar\")\n&gt;&gt;&gt; bar2 = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=\"bar\")[0]\n&gt;&gt;&gt; bar1 is bar2\nTrue\n<\/code><\/pre>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>You can also use <code>get_tensor_by_name()<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; bar1 = tf.Variable(0.0, name=\"bar\")\n&gt;&gt;&gt; graph = tf.get_default_graph()\n&gt;&gt;&gt; bar2 = graph.get_tensor_by_name(\"bar:0\")\n&gt;&gt;&gt; bar1 is bar2\nFalse, bar2 is a Tensor througn convert_to_tensor on bar1. but bar1 equal \nbar2 in value.\n<\/code><\/pre>\n\n<p>Recall that a tensor is the output of an operation. It has the same name as the operation, plus <code>:0<\/code>. If the operation has multiple outputs, they have the same name as the operation plus <code>:0<\/code>, <code>:1<\/code>, <code>:2<\/code>, and so on.<\/p>\n"},{"AnswerId":"35686754","CreationDate":"2016-02-28T18:48:46.577","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to get a variable by name is to search for it in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/global_variables\" rel=\"noreferrer\"><code>tf.global_variables()<\/code><\/a> collection:<\/p>\n\n<pre><code>var_23 = [v for v in tf.global_variables() if v.name == \"Variable_23:0\"][0]\n<\/code><\/pre>\n\n<p>This works well for ad hoc reuse of existing variables. A more structured approach&mdash;for when you want to share variables between multiple parts of a model&mdash;is covered in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variable_scope\/index.html#sharing-variables\" rel=\"noreferrer\">Sharing Variables tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":35679160,"AnswerCount":3,"Tags":"<json><lua><torch>","CreationDate":"2016-02-28T05:39:49.550","AcceptedAnswerId":null,"OwnerUserId":734524.0,"Title":"Read large JSON file in Lua","Body":"<p>I am trying to read some data in a JSON file in order to use it from lua. The data are sound files that have been preprocessed in python and stored in JSON for easier access.<\/p>\n\n<p>The file is roughly 800Mb. When I try to read the read the entire file with , I get back a  response. The libraries I have looked at are <a href=\"https:\/\/github.com\/clementfarabet\/lua---json\/\" rel=\"nofollow\">lua-json<\/a>, <a href=\"http:\/\/www.kyne.com.au\/~mark\/software\/lua-cjson.php\" rel=\"nofollow\">lua-cjson<\/a> and <a href=\"https:\/\/github.com\/harningt\/luajson\" rel=\"nofollow\">luajson<\/a>. The first two don't provide a method to access files directly, the third one does, however is just a wrapper that calls .<\/p>\n\n<p>My ultimate goal is to use <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow\">torch<\/a> to train some models on some audio data, but I want to keep the processing of the raw signals in python. I chose JSON over other formats for convenience, so if you think there is a format that would work better, I am open for ideas.<\/p>\n","answers":[{"AnswerId":"37087612","CreationDate":"2016-05-07T10:50:31.543","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>Instead of using json, you could also try <a href=\"https:\/\/github.com\/htwaijry\/npy4th\" rel=\"nofollow\">npy4th<\/a>, and you can save data as \"npz\" file.<\/p>\n\n<p>Another option is to use <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a>, a library which allows you run lua\/torch in python and provide convenient utilities for converting between numpy array and torch tensor, the advantage is memory copy or disk copy is not necessary, they share the underlying memory, so it's very fast. Check <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">the website<\/a> for more information.<\/p>\n\n<p>A basic example:<\/p>\n\n<pre><code>import lutorpy as lua\nimport numpy as np\n\n## use require(\"MODULE\") to import lua modules\nrequire(\"nn\")\n\n## run lua code in python with minimal modification:  replace \":\" to \"._\"\nt = torch.DoubleTensor(10,3)\nprint(t._size()) # the corresponding lua version is t:size()\n\n## or, you can use numpy array\nxn = np.random.randn(100)\n## convert the numpy array into torch tensor\nxt = torch.fromNumpyArray(xn)\n\n## convert torch tensor to numpy array\n### Note: the underlying object are sharing the same memory, so the conversion is instant\narr = xt.asNumpyArray()\nprint(arr.shape)\n<\/code><\/pre>\n"},{"AnswerId":"35679518","CreationDate":"2016-02-28T06:28:45.707","ParentId":null,"OwnerUserId":"1442917","Title":null,"Body":"<p>I'm not sure json is the best format for storing audio data, but it seems like in this situation you'll need to write your own json parser that will read the file, parse the data, and pass them through your training process without storing the entire data set in memory.<\/p>\n\n<p>Since the json format is fairly simple and you can limit the processing to just handle your format, it should be relatively straightforward to write SAX-like parser that will generate events you need. <a href=\"https:\/\/stackoverflow.com\/questions\/444380\/is-there-a-streaming-api-for-json\">This SO answer<\/a> may be a good starting point (or at least give you ideas on what keywords to search for).<\/p>\n"},{"AnswerId":"35687217","CreationDate":"2016-02-28T19:29:13.380","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>You have two options:<\/p>\n\n<p>Option 1: Install torch with Lua52 instead of LuaJIT. Nothing changes, everything works as expected, and you can now load your json file and decode it without memory issues. To do this:<\/p>\n\n<pre><code>cd ~\/torch\n.\/clean.sh\nTORCH_LUA_VERSION=LUA52 .\/install.sh\n<\/code><\/pre>\n\n<p>Option 2:\nUse HDF5 to save your python pre-processed files, and use torch-hdf5 to load them. HDF5 is much more suited for your data than JSON anyways.<\/p>\n"}]}
{"QuestionId":35680702,"AnswerCount":2,"Tags":"<deep-learning><caffe>","CreationDate":"2016-02-28T09:17:41.450","AcceptedAnswerId":35715139.0,"OwnerUserId":3130081.0,"Title":"The speed between ImageDataLayer and LMDB data layer","Body":"<p>Caffe support LMDB data layer and ImageDataLayer. \nCreate LMDB database from some dataset require some time and a lot of space.\nIn contrast, ImageDataLayer only use a txt file which is very convenient.\nMy question is, is there big speed difference between these two kinds of layers?\nThank you very much!<\/p>\n","answers":[{"AnswerId":"35713558","CreationDate":"2016-03-01T02:22:03.967","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Yes, the speed difference is indeed big. LMDB is optimized for high speed batch processing.<\/p>\n"},{"AnswerId":"35715139","CreationDate":"2016-03-01T05:12:59.813","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p><strong>LMDB<\/strong> is designed for faster fetching of data from a given <code>key value<\/code>. Also the data is stored in uncompressed format, which makes it easy for the machine to just read the data and directly pass them to the GPU for processing. <\/p>\n\n<p>In <strong>ImageDataLayer<\/strong>, we have to read the image details from the text file, and use OpenCV to read the image to memory. This uncompressing of image is computationally expensive.<\/p>\n\n<p>But the best performance may not always be for the LMDB layer, it depends heavily on the configuration of the machine. Consider an example of 256 image batch size and the images of size 227x227x3. Also consider than you are using a very good GPU and a high end i8 processor machine. Here single image in LMDB format may occupy 151KB. A whole batch may occupy 37MB. If the GPU is able to perform 10 batches a second, the harddisk should have a speed of reading 370MB\/s. If you are using a normal SATA or external harddisk, there will be bottlenecks on reading such large chunks of data due to the limits of the hard disk.<\/p>\n\n<p>If caffe could not fetch data in the required speed, the bottleneck slows the whole training process even worse. At the same time, if you were reading 256 images and use multi-core version of OpenCV, the data prefetching may be handled more effectively than reading an LMDB.<\/p>\n\n<p>The above case will not occur if you have stored the LMDB data on a SSD though!<\/p>\n"}]}
{"QuestionId":35682758,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-28T12:50:38.053","AcceptedAnswerId":35683145.0,"OwnerUserId":5964720.0,"Title":"PTB rnn model one PTBModel object instead of three","Body":"<p>in the PTB rnn model, three PTBModel objects are created, namely m, mvalid and mtest:      <\/p>\n\n<pre><\/pre>\n\n<p>my questions are:<\/p>\n\n<ul>\n<li><p>do all these three objects live in the same graph? (It looks like they all live under the default graph.)<\/p><\/li>\n<li><p>do these three objects share the same placeholders, e.g., _input_data? Or is it the case that different sets of placeholders are created with each PTBModel object, so that for example there are three _input_data placeholders within the same graph (one _input_data used for feeding training data, another for validation and yet another for testing)?<\/p><\/li>\n<li><p>suppose I only create one PTBModel object, would it be possible to reuse  the _input_data placeholder used for training and change its shape and use it for testing as well (where the 1st dimension, num_steps, is set to 1 at test time)?<\/p><\/li>\n<\/ul>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"35683145","CreationDate":"2016-02-28T13:31:44.607","ParentId":null,"OwnerUserId":"5543198","Title":null,"Body":"<p>Yes, these three objects live in the same graph.<\/p>\n\n<p>The placeholders are different and you need to use the correct one if you want to evaluate particular part of the graph.<\/p>\n\n<p>It would in theory be possible but it is not as trivial. E.g. you could have a training graph unrolled for 20 steps but use only a subset of the steps for evaluation. The other possibility might be to use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn.py#L332\" rel=\"nofollow\">dynamic_rnn<\/a> functionality.<\/p>\n\n<p>In general, building a few copies of the graph is not very expensive and it might not be worth spending a lot of time on optimizing the number of allocated nodes.<\/p>\n"}]}
{"QuestionId":35687678,"AnswerCount":4,"Tags":"<python><numpy><tensorflow><deep-learning>","CreationDate":"2016-02-28T20:11:43.207","AcceptedAnswerId":35688187.0,"OwnerUserId":3147590.0,"Title":"Using a pre-trained word embedding (word2vec or Glove) in TensorFlow","Body":"<p>I've recently reviewed an interesting implementation for <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\">convolutional text classification<\/a>. However all TensorFlow code I've reviewed uses a random (not pre-trained) embedding vectors like the following:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody know how to use the results of Word2vec or a GloVe pre-trained word embedding instead of a random one?<\/p>\n","answers":[{"AnswerId":"49779653","CreationDate":"2018-04-11T15:59:13.663","ParentId":null,"OwnerUserId":"5904928","Title":null,"Body":"<p>I was also facing embedding issue, So i wrote detailed tutorial with dataset. \nHere I would like to add what I tried You can also try this method, <\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.reset_default_graph()\n\ninput_x=tf.placeholder(tf.int32,shape=[None,None])\n\n#you have to edit shape according to your embedding size\n\n\nWord_embedding = tf.get_variable(name=\"W\", shape=[400000,100], initializer=tf.constant_initializer(np.array(word_embedding)), trainable=False)\nembedding_loopup= tf.nn.embedding_lookup(Word_embedding,input_x)\n\nwith tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        for ii in final_:\n            print(sess.run(embedding_loopup,feed_dict={input_x:[ii]}))\n<\/code><\/pre>\n\n<p>Here is working detailed Tutorial <a href=\"https:\/\/github.com\/monk1337\/word_embedding-in-tensorflow\/blob\/master\/Use%20Pre-trained%20word_embedding%20in%20Tensorflow.ipynb\" rel=\"nofollow noreferrer\">Ipython example<\/a> if you want to understand from scratch , take a look .<\/p>\n"},{"AnswerId":"36880018","CreationDate":"2016-04-27T03:58:22.547","ParentId":null,"OwnerUserId":"5772548","Title":null,"Body":"<p>I use this method to load and share embedding.<\/p>\n\n<pre><code>W = tf.get_variable(name=\"W\", shape=embedding.shape, initializer=tf.constant_initializer(embedding), trainable=False)\n<\/code><\/pre>\n"},{"AnswerId":"35688187","CreationDate":"2016-02-28T20:59:12.407","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are a few ways that you can use a pre-trained embedding in TensorFlow. Let's say that you have the embedding in a NumPy array called <code>embedding<\/code>, with <code>vocab_size<\/code> rows and <code>embedding_dim<\/code> columns and you want to create a tensor <code>W<\/code> that can be used in a call to <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/nn#Embeddings\" rel=\"noreferrer\"><code>tf.nn.embedding_lookup()<\/code><\/a>.<\/p>\n\n<ol>\n<li><p>Simply create <code>W<\/code> as a <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/constant_op#constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> that takes <code>embedding<\/code> as its value:<\/p>\n\n<pre><code>W = tf.constant(embedding, name=\"W\")\n<\/code><\/pre>\n\n<p>This is the easiest approach, but it is not memory efficient because the value of a <code>tf.constant()<\/code> is stored multiple times in memory. Since <code>embedding<\/code> can be very large, you should only use this approach for toy examples.<\/p><\/li>\n<li><p>Create <code>W<\/code> as a <code>tf.Variable<\/code> and initialize it from the NumPy array via a <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/io_ops#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a>:<\/p>\n\n<pre><code>W = tf.Variable(tf.constant(0.0, shape=[vocab_size, embedding_dim]),\n                trainable=False, name=\"W\")\n\nembedding_placeholder = tf.placeholder(tf.float32, [vocab_size, embedding_dim])\nembedding_init = W.assign(embedding_placeholder)\n\n# ...\nsess = tf.Session()\n\nsess.run(embedding_init, feed_dict={embedding_placeholder: embedding})\n<\/code><\/pre>\n\n<p>This avoid storing a copy of <code>embedding<\/code> in the graph, but it does require enough memory to keep two copies of the matrix in memory at once (one for the NumPy array, and one for the <code>tf.Variable<\/code>). Note that I've assumed that you want to hold the embedding matrix constant during training, so <code>W<\/code> is created with <code>trainable=False<\/code>.<\/p><\/li>\n<li><p>If the embedding was trained as part of another TensorFlow model, you can use a <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/state_ops#Saver\" rel=\"noreferrer\"><code>tf.train.Saver<\/code><\/a> to load the value from the other model's checkpoint file. This means that the embedding matrix can bypass Python altogether. Create <code>W<\/code> as in option 2, then do the following:<\/p>\n\n<pre><code>W = tf.Variable(...)\n\nembedding_saver = tf.train.Saver({\"name_of_variable_in_other_model\": W})\n\n# ...\nsess = tf.Session()\nembedding_saver.restore(sess, \"checkpoint_filename.ckpt\")\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"40214916","CreationDate":"2016-10-24T09:25:01.397","ParentId":null,"OwnerUserId":"2984547","Title":null,"Body":"<p>The answer of @mrry is not right because it provoques the overwriting of the embeddings weights each the network is run, so if you are following a minibatch approach to train your network, you are overwriting the weights of the embeddings. So, on my point of view the right way to pre-trained embeddings is:<\/p>\n\n<pre><code>embeddings = tf.get_variable(\"embeddings\", shape=[dim1, dim2], initializer=tf.constant_initializer(np.array(embeddings_matrix))\n<\/code><\/pre>\n"}]}
{"QuestionId":35689248,"AnswerCount":1,"Tags":"<python><argparse><tensorflow><gflags>","CreationDate":"2016-02-28T22:44:21.860","AcceptedAnswerId":35689315.0,"OwnerUserId":4674418.0,"Title":"Train Tensorflow model without using command line","Body":"<p>I want to call <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/embedding\/word2vec_optimized.py#L408\" rel=\"nofollow\">this  function<\/a> from another Python script without spawning a new process (so that it's easier to debug). However, that function is written to work with command line arguments. What would be the cleanest way call that function directly from another function?<\/p>\n","answers":[{"AnswerId":"35689315","CreationDate":"2016-02-28T22:53:32.167","ParentId":null,"OwnerUserId":"5699422","Title":null,"Body":"<p>You can import <code>FLAGS<\/code> and then define the required args (train_data, eval_data, save_path).<\/p>\n\n<pre><code>In [13]: from tensorflow.models.embedding.word2vec_optimized import FLAGS\nIn [14]: from tensorflow.models.embedding.word2vec_optimized import main\nIn [16]: main(_)\n--train_data --eval_data and --save_path must be specified.\nAn exception has occurred, use %tb to see the full traceback.\nIn [17]: FLAGS.train_data = \"this\"\nIn [18]: FLAGS.eval_data = \"that\"\nIn [19]: FLAGS.save_path = \"some_path\"\nIn [20]: main(_)\nI tensorflow\/core\/common_runtime\/local_device.cc:40] Local device intra op parallelism threads: 8\n<\/code><\/pre>\n"}]}
{"QuestionId":35689547,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-28T23:19:26.860","AcceptedAnswerId":35689728.0,"OwnerUserId":2138200.0,"Title":"How to process single training file in parallel","Body":"<p>I have a file  that contains paths to images and their labels. ie:<\/p>\n\n<pre><\/pre>\n\n<p>After going through the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html\" rel=\"nofollow\">reading data tutorial<\/a> I came up with some code to go through each image, resize it and apply distortions:<\/p>\n\n<pre><\/pre>\n\n<p>The problem now is that I'm confused how to do these operations across the file in parallel. The documentation suggests either using  or  with num_threads=N.<\/p>\n\n<p>I first tried following the example code using  but this seems to be intended for processing multiple files in parallel. In my case however I just have 1 file. <\/p>\n\n<pre><\/pre>\n\n<p>I also tried setting  but its not clear to me if this is doing the right thing (although I can see more cpu cores in use)<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code for executing the graph:<\/p>\n\n<pre><\/pre>\n\n<p>Whats the best way to process this file in parallel?<\/p>\n","answers":[{"AnswerId":"35689728","CreationDate":"2016-02-28T23:39:18.287","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Both seem like viable approaches. Using <code>batch<\/code> with <code>threads=N<\/code> will create <code>N<\/code> copies of your reader op connected to your queue so that they can run in parallel, while <code>batch_join<\/code> you have to create the copies manually.<\/p>\n\n<p>In your usage with <code>batch_join<\/code> you are creating several copies of <code>TextLineReader<\/code> which (as you noticed) will only parallelize across files. To have several threads reading a single file, you could instead create one <code>TextLineReader<\/code> and have several <code>line_reader.read<\/code> ops using the same reader.<\/p>\n\n<p>Here's an example with some textfiles containing numbers<\/p>\n\n<p>Generate numbers:<\/p>\n\n<pre><code>num_files=10\nnum_entries_per_file=10\nfile_root=\"\/temp\/pipeline\"\nos.system('mkdir -p '+file_root)\nfor fi in range(num_files):\n  fname = file_root+\"\/\"+str(fi)\n  dump_numbers_to_file(fname, fi*num_entries_per_file, (fi+1)*num_entries_per_file)\n<\/code><\/pre>\n\n<p>Read those numbers in batches of size 2, with parallelism of 2<\/p>\n\n<pre><code>ops.reset_default_graph()\nfilename_queue = tf.train.string_input_producer([\"\/temp\/pipeline\/0\",\n                                                 \"\/temp\/pipeline\/1\"],\n                                                shuffle=False)\nreader = tf.TextLineReader()\nkey, value = reader.read(filename_queue)\nnumeric_val1, = tf.decode_csv(value, record_defaults=[[-1]])\nnumeric_val2, = tf.decode_csv(value, record_defaults=[[-1]])\nnumeric_batch = tf.batch_join([[numeric_val1,], [numeric_val2]], 2)\n# have to create session before queue runners because they use default session\nsess = create_session()\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(coord=coord)\n\nprint '\\n'.join([t.name for t in threads])\nfor i in range(20):\n  print sess.run([numeric_batch])\n\ncoord.request_stop()\ncoord.join(threads)\n<\/code><\/pre>\n\n<p>You may see something like this:<\/p>\n\n<pre><code>QueueRunner(input_producer:input_producer\/input_producer_EnqueueMany)\nQueueRunner(input_producer:input_producer\/input_producer_Close_1)\nQueueRunner(batch_join\/fifo_queue:batch_join\/fifo_queue_enqueue)\nQueueRunner(batch_join\/fifo_queue:batch_join\/fifo_queue_enqueue_1)\nQueueRunner(batch_join\/fifo_queue:batch_join\/fifo_queue_Close_1)\n[array([0, 1], dtype=int32)]\n[array([2, 3], dtype=int32)]\n[array([4, 5], dtype=int32)]\n[array([6, 7], dtype=int32)]\n[array([8, 9], dtype=int32)]\n[array([10, 11], dtype=int32)]\n[array([12, 13], dtype=int32)]\n[array([14, 15], dtype=int32)]\n[array([16, 17], dtype=int32)]\n[array([18, 19], dtype=int32)]\n<\/code><\/pre>\n\n<p>From the list of threads, you can see that there are 2 threads corresponding to read operations (<code>fifo_queue_enqueue<\/code> and <code>fifo_queue_enqueue_1<\/code> so you can do 2 reads in parallel)<\/p>\n"}]}
{"QuestionId":35689613,"AnswerCount":1,"Tags":"<performance><profiling><gpu><gpgpu><theano>","CreationDate":"2016-02-28T23:27:04.897","AcceptedAnswerId":null,"OwnerUserId":403875.0,"Title":"Is my Theano program actually using the GPU?","Body":"<p>Theano claims it's using the GPU; it says what device when it starts up, etc.  Furthermore nvidia-smi says it's being used.<\/p>\n\n<p>But the running time seems to be exactly the same regardless of whether or not I use it.<\/p>\n\n<p>Could it have something to do with integer arithmetic?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35918512","CreationDate":"2016-03-10T13:56:21.833","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>The answer is yes. And no. If you profile your code in a GPU enabled Theano installation using <a href=\"http:\/\/docs.nvidia.com\/cuda\/profiler-users-guide\/#abstract\" rel=\"nofollow noreferrer\">nvprof<\/a>, you will see something like this:<\/p>\n\n<pre><code>==16540== Profiling application: python .\/theano_test.py\n==16540== Profiling result:\nTime(%)      Time     Calls       Avg       Min       Max  Name\n 49.22%  12.096us         1  12.096us  12.096us  12.096us  kernel_reduce_ccontig_node_c8d7bd33dfef61705c2854dd1f0cb7ce_0(unsigned int, float const *, float*)\n 30.60%  7.5200us         3  2.5060us     832ns  5.7600us  [CUDA memcpy HtoD]\n 13.93%  3.4240us         1  3.4240us  3.4240us  3.4240us  [CUDA memset]\n  6.25%  1.5350us         1  1.5350us  1.5350us  1.5350us  [CUDA memcpy DtoH]\n<\/code><\/pre>\n\n<p>i.e. There is a least a reduce operation being performed on your GPU. However, if you modify your main like this:<\/p>\n\n<pre><code>def main():\n  n = 100000000 \n  print(primes_to(n).size.eval())\n\nif __name__ == '__main__':\n    import cProfile, pstats\n    cProfile.run(\"main()\", \"{}.profile\".format(__file__))\n    s = pstats.Stats(\"{}.profile\".format(__file__))\n    s.strip_dirs()\n    s.sort_stats(\"time\").print_stats(10)\n<\/code><\/pre>\n\n<p>and use <a href=\"https:\/\/docs.python.org\/2\/library\/profile.html#module-cProfile\" rel=\"nofollow noreferrer\">cProfile<\/a> to profile your code, you will see something like this:<\/p>\n\n<pre><code>Thu Mar 10 14:35:24 2016    .\/theano_test.py.profile\n\n         486743 function calls (480590 primitive calls) in 17.444 seconds\n\n   Ordered by: internal time\n   List reduced from 1138 to 10 due to restriction &lt;10&gt;\n\n   ncalls  tottime  percall  cumtime  percall filename:lineno(function)\n        1    6.376    6.376   16.655   16.655 {theano.scan_module.scan_perform.perform}\n       13    6.168    0.474    6.168    0.474 subtensor.py:2084(perform)\n       27    2.910    0.108    2.910    0.108 {method 'nonzero' of 'numpy.ndarray' objects}\n       30    0.852    0.028    0.852    0.028 {numpy.core.multiarray.concatenate}\n       27    0.711    0.026    0.711    0.026 {method 'astype' of 'numpy.ndarray' objects}\n       13    0.072    0.006    0.072    0.006 {numpy.core.multiarray.arange}\n        1    0.034    0.034   17.142   17.142 function_module.py:482(__call__)\n      387    0.020    0.000    0.052    0.000 graph.py:486(stack_search)\n       77    0.016    0.000   10.731    0.139 op.py:767(rval)\n      316    0.013    0.000    0.066    0.000 graph.py:715(general_toposort)\n<\/code><\/pre>\n\n<p>The slowest operation (just) is the scan call, and looking at the source for scan, you can see that presently, <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/b7e68d2241d28db48a1a8dadf949ba1a0d97c1c6\/theano\/scan_module\/scan.py#L1012\" rel=\"nofollow noreferrer\">GPU execution of scan is disabled<\/a>.<\/p>\n\n<p>So then answer is, yes, the GPU is being used for something in your code, but no, the most time consuming operation(s) are being run on the CPU because GPU execution appears to be hard disabled in the code at present.<\/p>\n"}]}
{"QuestionId":35689671,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-02-28T23:34:15.527","AcceptedAnswerId":35689756.0,"OwnerUserId":440221.0,"Title":"How to do a column sum in Tensorflow?","Body":"<p>What is the equivalent of the following in Tensorflow?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35689756","CreationDate":"2016-02-28T23:42:13.853","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>There is tf.reduce_sum which is a bit more powerfull tool for doing so.<\/p>\n\n<pre><code># 'x' is [[1, 1, 1]\n#         [1, 1, 1]]\ntf.reduce_sum(x) ==&gt; 6\ntf.reduce_sum(x, 0) ==&gt; [2, 2, 2]\ntf.reduce_sum(x, 1) ==&gt; [3, 3]\ntf.reduce_sum(x, 1, keep_dims=True) ==&gt; [[3], [3]]\ntf.reduce_sum(x, [0, 1]) ==&gt; 6\n<\/code><\/pre>\n"}]}
{"QuestionId":35690931,"AnswerCount":1,"Tags":"<neural-network><deep-learning><torch>","CreationDate":"2016-02-29T02:19:50.300","AcceptedAnswerId":35773752.0,"OwnerUserId":3853033.0,"Title":"Criteria Behind Structuring a Neural Network","Body":"<p>I'm just starting with Torch and neural networks and just glancing at a lot of sample code and tutorials, I see a lot of variety in the how people structure their neural networks. There are layers like Linear(), Tanh(), Sigmoid() as well as criterions like MSE, ClassNLL, MultiMargin, etc. <\/p>\n\n<p>I'm wondering what kind of factors people keep in mind when creating the structure of their network? For example, I know that in a ClassNLLCriterion, you want to have the last layer of your network be a LogSoftMax() layer so that you can input the right log probabilities. <\/p>\n\n<p>Are there any other general rules or guidelines when it comes to creating these networks?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35773752","CreationDate":"2016-03-03T13:38:37.367","ParentId":null,"OwnerUserId":"5407700","Title":null,"Body":"<p>Here is a good webpage which contains the pros and cons of some of the main activation functions;<\/p>\n\n<p><a href=\"http:\/\/cs231n.github.io\/neural-networks-1\/#actfun\" rel=\"nofollow\">http:\/\/cs231n.github.io\/neural-networks-1\/#actfun<\/a><\/p>\n\n<p>It can boil down to the problem at hand and knowing what to do when something goes wrong. As an example, if you have a huge dataset and you can't churn through it terribly quickly then a ReLU might be better in order to quickly get to a local minimum. However you could find that some of the ReLU units \"die\" so you might want to keep a track on the proportion of activated neurons in that particular layer to make sure this hasn't happened. <\/p>\n\n<p>In terms of criterions, they are also problem specific but a bit less ambiguous. For example, binary cross entropy for binary classification, MSE for regression etc. It really depends on the objective of the whole project.<\/p>\n\n<p>For the overall network architecture, I personally find it can be a case of trying out different architectures and seeing which ones work and which don't on your test set. If you think that the problem at hand is terribly complex and you need a complex network to solve the problem then you will probably want to try making a very deep network to begin with, then add\/remove a few layers at a time to see if you have under\/overfitted. As another example, if you are using  convolutional network and the input is relatively small then you might try and use a smaller set of convolutional filters to begin with.<\/p>\n"}]}
{"QuestionId":35691099,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><tensorflow><deep-learning>","CreationDate":"2016-02-29T02:40:01.360","AcceptedAnswerId":null,"OwnerUserId":5551505.0,"Title":"Tensorflow - Convolutonal neural network with custom data set DOES NOT LEARN","Body":"<p>I am trying to make a convolutional neural network for custom data set. The classifier has only two classes. I am able to read the input images properly and have also assigned them the batch_labels for the two corresponding classes . The code executes without error, but the output is anomalous.\nFor some reason, the accuracy is always 50%. <\/p>\n\n<pre><\/pre>\n\n<p>When I print the  tensor, I get the following output no matter what.<\/p>\n\n<pre><\/pre>\n\n<p>The accuracy is always 0.5, as if the weights are not being updated at all. When I print weights after each training step, they remain unchanged. I think I have some coding error.   Could it be that the network is training on the same image again and again? But even so, the weights must update.   I have 150 training examples, with 75 belonging to each class.\nCould someone please point me in the right direction ? <\/p>\n\n<p><strong>EDIT<\/strong>:\nThis is how I initialize weights <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35711648","CreationDate":"2016-02-29T23:14:06.060","ParentId":null,"OwnerUserId":"3314143","Title":null,"Body":"<p>Your network has some design flaws. Because of mathematical issues it is not a good idea to calculate the cross entropy yourself and apply a softmax on the output layer. If you are interested in the mathematics I could add this, if not stick to the Tensorflow explanation and method: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"nofollow\">tf.nn.softmax_cross_entropy_with_logits<\/a>.<\/p>\n\n<p>Have you already tried many different configurations? Depending on the complexity of your images a higher or lower kernel size and amount of feature maps might be a good idea. Generally, if your images are relatively homogeneous, a lot of rather similar information is added up and therefore the network has it harder to converge if you have many feature maps. Since you have only two output neurons I assume are images are not very complex?<\/p>\n\n<p>The next thing is your dropout. You are always using a dropout of 0.5 but normally, for test\/validation (like your accuracy prediction) you do not use dropout. In most cases you only use it for training. You can create a placeholder specifying your dropout rate and feed this <code>sess.run<\/code>.<\/p>\n\n<p>Here is some example of my own:<\/p>\n\n<pre><code>h_fc_drop = tf.nn.dropout(h_fc, keep_prob)\n\n(...)\n\naccu, top1, top3, top5 = sess.run([accuracy, te_top1, te_top3, te_top5],\n                            feed_dict={\n                                x: teX[i: i + batch_size],\n                                y: teY[i: i + batch_size]\n                                keep_prob: 1.0\n                            }\n                         )\n<\/code><\/pre>\n\n<p>This lets Tensorflow calculate my equations for <code>accuracy<\/code> and <code>topX<\/code> error rate, while I feed in the test data input <code>teX<\/code> and the real labels <code>teY<\/code> for the output with a keep probability <code>keep_prob<\/code> of 1.0 for the dropout.<\/p>\n\n<p>Despite this the initialization of your weights is really important in deep neural networks. Even if your design is sufficient for your kind of problem (this also has to be investigated) your network could refuse to learn, diverge or converge to 0 if your weights are not initialized properly. You did not add details to your initialization so you maybe want to look up <em>Xavier initialization<\/em>. <a href=\"http:\/\/andyljones.tumblr.com\/post\/110998971763\/an-explanation-of-xavier-initialization\" rel=\"nofollow\">This<\/a> is an easy beginning for Xavier initialization.<\/p>\n\n<p>Conclusively I can just encourage you to plot some weights, feature maps, the output over time and so on to get an idea of what your network is doing. Normally this helps a lot.<\/p>\n"}]}
{"QuestionId":35691102,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-02-29T02:40:19.783","AcceptedAnswerId":35692452.0,"OwnerUserId":5703903.0,"Title":"ValueError: All shapes must be fully defined. Issue due to commenting out tf.random_crop","Body":"<p>This question is somewhat an extension of <a href=\"https:\/\/stackoverflow.com\/questions\/35620157\/how-can-i-use-values-read-from-tfrecords-as-arguments-to-tf-reshape\">How can I use values read from TFRecords as arguments to tf.reshape?<\/a><\/p>\n\n<p>I cast my images into a certain shape with the following code:<\/p>\n\n<pre><\/pre>\n\n<p>In cifar10_input code, the image is then distorted with the following, where IMAGE_SIZE = 32:<\/p>\n\n<pre><\/pre>\n\n<p>However, for my purposes, I don't need to do a random crop now. As such, I replaced that line with:<\/p>\n\n<pre><\/pre>\n\n<p>When I do this, it throws the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have 2 questions:<\/p>\n\n<ol>\n<li>Why does it throw an error when I don't do a tf.random_crop? It seems to me as if tf.random_crop returns something that is totally diffrent from image.<\/li>\n<li>Is simply setting IMAGE_SIZE to the size that I want a good solution? For example, if the images are 32 x 32 and I want to crop it to 24 x 24, I will set IMAGE_SIZE = 24. Now, since I want it to be 32 x 32, should I simply set IMAGE_SIZE = 32?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"35692452","CreationDate":"2016-02-29T05:19:33.197","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Because you're generating your image dynamically, including pulling out the height &amp; width dynamically from the tf record file, TensorFlow doesn't know the shape of the resulting image.  Many of the later ops in the pipeline need to be able to determine the shape at the time Python executes.<\/p>\n\n<p>The <code>tf.random_crop<\/code> has the incidental effect of setting the image size to a known, fixed size, and leaving its shape exposed for subsequent processing.<\/p>\n\n<p>You can just slice the image to the size you want instead of doing a random_crop, but you need to perform <em>some<\/em> operation to turn the image into a fixed-size thing.  If you want it to be 32x32 <em>and you know that your input height and width are 32x32<\/em>, then you can just do set_shape on it (but you'd better be right).  Otherwise, you can crop and\/or resize to the size you want.<\/p>\n"}]}
{"QuestionId":35691947,"AnswerCount":1,"Tags":"<python><numpy><cython><theano><conv-neural-network>","CreationDate":"2016-02-29T04:23:31.917","AcceptedAnswerId":null,"OwnerUserId":868736.0,"Title":"3d sliding window operation in Theano?","Body":"<p>TL.DR.  Is there a 3-dimensional friendly implementation of <a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/tensor\/nnet\/neighbours.py\" rel=\"nofollow noreferrer\"><\/a>?<\/p>\n\n<p>I would like to perform voxel-wise classification of a volume (NxNxN) using a neural network that takes in a nxnxn image, where N>n.  To classify each voxel in the volume, I have to iterate through each voxel.  For each iterration, I obtain and pass the neighborhood voxels as the input to the neural network.  This is simply a sliding window operation, which the operation is the neural network.<\/p>\n\n<p>While my neural network is implemented in Theano, the sliding window implementation is in python\/numpy.  Since this is not a pure Theano operation, the classification takes forever (> 3 hours) to classify all voxels in one volume.  For 2d sliding window operation, Theano has a helper method, , is there a similar implementation for 3-dimensional images?<\/p>\n\n<p>Edit:\nThere are existing numpy solutions (<a href=\"http:\/\/github.com\/Theano\/Theano\/issues\/2166\" rel=\"nofollow noreferrer\">1<\/a> and <a href=\"http:\/\/gist.github.com\/seberg\/3866040%20johnvinyard.com\/blog\/?p=268\" rel=\"nofollow noreferrer\">2<\/a>) for n-d sliding window, both use  to provide \"views of the sliding window\", thus preventing memory issues.  In my implementation, the sliding window arrays are being passed from numpy (Cython) to Python and then to Theano.  To boost performance, it's likely I have to bypass Python.<\/p>\n","answers":[{"AnswerId":"48053225","CreationDate":"2018-01-01T21:05:51.623","ParentId":null,"OwnerUserId":"8431111","Title":null,"Body":"<p>Eickenberg and Kastner's OverfeatTransformer utility in sklearn_theano.feature_extraction.overfeat would be a good match for this operation, as mentioned by OP.<\/p>\n"}]}
{"QuestionId":35693687,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-02-29T07:04:37.540","AcceptedAnswerId":null,"OwnerUserId":5156504.0,"Title":"Forcing dependence on a variable update","Body":"<p>Say I have some function  of variable :<\/p>\n\n<pre><\/pre>\n\n<p>and an op which updates :<\/p>\n\n<pre><\/pre>\n\n<p>and I want to get the value of  resulting from the updated . I had thought that <\/p>\n\n<pre><\/pre>\n\n<p>would force  to depend on the update , but this doesn't seem to be the case:<\/p>\n\n<pre><\/pre>\n\n<p>Is there some other way to define the updated value of ?<\/p>\n\n<p>Obviously I could create a new independent copy by writing something like  but this blows up the graph size, and also requires access to the definition of , which I'd prefer to treat as a black box. <\/p>\n\n<p><strong>Edit<\/strong>: to motivate this, here's a sketch of the code I want to write:<\/p>\n\n<pre><\/pre>\n\n<p>It feels really inelegant to define a new copy of logpdf and rederive gradients each time through the loop: it requires access to  define_logpdf() and blows up the graph size by a factor of 50. Is there no better way to do this (barring some equivalent of theano.scan)?<\/p>\n","answers":[{"AnswerId":"35704144","CreationDate":"2016-02-29T16:08:45.543","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>The <code>with tf.control_dependencies([op])<\/code> block enforces control dependency on <code>op<\/code> to other ops <em>created<\/em> within with block. In your case, <code>x*x<\/code> is created outside, and the <code>tf.identity<\/code> just gets the old value. Here is what you want :<\/p>\n\n<pre><code>with tf.control_dependencies([new_x,]):\n  new_fx = x*x\n<\/code><\/pre>\n"}]}
{"QuestionId":35695183,"AnswerCount":1,"Tags":"<memory-management><memory-leaks><neural-network><tensorflow>","CreationDate":"2016-02-29T08:47:22.857","AcceptedAnswerId":35705890.0,"OwnerUserId":5962133.0,"Title":"Tensorflow : Memory leak even while closing Session?","Body":"<p>I was just trying some stuff for a quaternionic neural network when I realized that, even if I close my current Session in a for loop, my program slows down massively and I get a memory leak caused by ops being constructed. This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>The problem seems to come from . I need to declare these ops after executing  once, because  can't be a placeholder (I need to iterate through it). Without this line, the program runs very well, fast and without a memory leak. I can't understand why it seems like TensorFlow is trying to \"save\"  even if I close the session ... <\/p>\n","answers":[{"AnswerId":"35705890","CreationDate":"2016-02-29T17:38:26.027","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> Closing a session does not free the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Graph\" rel=\"noreferrer\"><code>tf.Graph<\/code><\/a> data structure in your Python program, and if each iteration of the loop adds nodes to the graph, you'll have a leak.<\/p>\n\n<p>Since your function <code>feedForwardStep<\/code> creates new TensorFlow operations, and you call it within the <code>for<\/code> loop, then there <em>is<\/em> a leak in your code&mdash;albeit a subtle one.<\/p>\n\n<p>Unless you specify otherwise (using a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Graph.as_default\" rel=\"noreferrer\"><code>with tf.Graph().as_default():<\/code><\/a> block), all TensorFlow operations are added to a global default graph. This means that every call to <code>tf.constant()<\/code>, <code>tf.matmul()<\/code>, <code>tf.Variable()<\/code> etc. adds objects to a global data structure. There are two ways to avoid this:<\/p>\n\n<ol>\n<li><p>Structure your program so that you build the graph once, then use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> ops to feed in different values in each iteration. You mention in your question that this might not be possible.<\/p><\/li>\n<li><p>Explicitly create a new graph in each for loop. This might be necessary if the structure of the graph depends on the data available in the current iteration. You would do this as follows:<\/p>\n\n<pre><code>for step in xrange(200):\n    with tf.Graph().as_default(), tf.Session() as sess:\n        # Remainder of loop body goes here.\n<\/code><\/pre>\n\n<p>Note that in this version, you cannot use <code>Tensor<\/code> or <code>Operation<\/code> objects from a previous iteration. (For example, it's not clear from your code snippet where <code>test<\/code> comes from.)<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35699851,"AnswerCount":2,"Tags":"<python><windows><ide><pycharm><tensorflow>","CreationDate":"2016-02-29T12:41:27.187","AcceptedAnswerId":40801992.0,"OwnerUserId":3345257.0,"Title":"Using tensorflow in IDE on windows","Body":"<p>I have used the docker installation as suggested at <a href=\"https:\/\/bishnusarker.wordpress.com\/2015\/11\/11\/how-to-install-google-tensorflow-in-your-windows-machine-using-docker-image\/\" rel=\"noreferrer\">https:\/\/bishnusarker.wordpress.com\/2015\/11\/11\/how-to-install-google-tensorflow-in-your-windows-machine-using-docker-image\/<\/a> to run tensorflow on my windows 7 machine. I have it up and running on the docker terminal. I would like to know how to import the tensorflow library from the PyCharm IDE. Any help or pointers will be highly appreciated as information on the matter is rather sparse.<\/p>\n","answers":[{"AnswerId":"40743395","CreationDate":"2016-11-22T13:41:27.570","ParentId":null,"OwnerUserId":"348412","Title":null,"Body":"<p>There is no way to make it work. Your tensorflow installation is inside the docker container, while PyCharm running on Windows searches for your available Python libraries in Windows; it has no access to docker.<\/p>\n\n<p>You could, instead, try to install and run PyCharm in docker as well (I'm not sure if it is viable though).<\/p>\n"},{"AnswerId":"40801992","CreationDate":"2016-11-25T09:49:12.507","ParentId":null,"OwnerUserId":"7208993","Title":null,"Body":"<p>Since I am currently struggling with a similar issue, I came across your question. You can run PyCharm on your Windows machine and configure a remote interpreter to mount your python code to the docker image, in which your libraries reside and run it there as described here: <a href=\"https:\/\/blog.jetbrains.com\/pycharm\/2015\/12\/using-docker-in-pycharm\/\" rel=\"nofollow noreferrer\">https:\/\/blog.jetbrains.com\/pycharm\/2015\/12\/using-docker-in-pycharm\/<\/a>\nFor tensorflow in particular, this might be interesting as well: <a href=\"http:\/\/killianlevacher.github.io\/blog\/posts\/post-2016-07-22\/post.html\" rel=\"nofollow noreferrer\">http:\/\/killianlevacher.github.io\/blog\/posts\/post-2016-07-22\/post.html<\/a><\/p>\n"}]}
{"QuestionId":35700964,"AnswerCount":0,"Tags":"<android><python><c++><tensorflow>","CreationDate":"2016-02-29T13:35:55.967","AcceptedAnswerId":null,"OwnerUserId":2371579.0,"Title":"Loading Tensorflow graph from binary file in Android application","Body":"<p>I'm trying to integrate TensorFlow into an Android application. Since I'm new to TensorFlow, I'm proceeding starting with very simple operations.<\/p>\n\n<p>As a first step I created just the following model:<\/p>\n\n<pre><\/pre>\n\n<p>This works fine. I'm able to properly get output from my C++ code (and then from Android as well).<\/p>\n\n<p>I tried then to generate  by using . It seems this is not feasible by just replacing  with . That's because  returs a costant, while  not. In particular it seems I must handle it as a variable.<\/p>\n\n<p>Idea I followed is the same proposed in other examples I've found on GitHub... so, evaluating  before writing the graph, as reported in code below:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately it seems it doesn't work due to an error occurring in reading binary file:<\/p>\n\n<blockquote>\n  <p>Out of range: Read less bytes than requested<\/p>\n<\/blockquote>\n\n<p>This is C++ code that I'm currently using for loading graph from file:<\/p>\n\n<pre><\/pre>\n\n<p>Hope someone could help me with this problem.<\/p>\n","answers":[]}
{"QuestionId":35702403,"AnswerCount":5,"Tags":"<python><linux><tensorflow>","CreationDate":"2016-02-29T14:43:17.177","AcceptedAnswerId":null,"OwnerUserId":5997869.0,"Title":"Tensorflow 0.7.1 with Cuda Toolkit 7.5 and cuDNN 7.0","Body":"<p>I recently tried to upgrade my Tensorflow installation from 0.6 to 0.7.1 (Ubuntu 15.10, Python 2.7) because it is described to be compatible with more up-to-date Cuda libraries. Everything works well including the simple test from the Tensorflow getting started page. However I'm not able to use cuDNN. When running a program using cuDNN, I first get a warning<\/p>\n\n<p>\"Unable to load cuDNN DSO\"<\/p>\n\n<p>and later the program crashes with<\/p>\n\n<pre><\/pre>\n\n<p>The files I downloaded for the Cuda Installation were<\/p>\n\n<ul>\n<li>cuda-repo-ubuntu1504-7-5-local_7.5-18_amd64.deb\nand<\/li>\n<li>cudnn-7.0-linux-x64-v4.0-prod.tgz<\/li>\n<\/ul>\n\n<p>I followed the instructions on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#optional-install-cuda-gpus-on-linux\">Tensorflow getting started page<\/a> with the exception of using cuDNN 7.0 instead of 6.5. $LD_LIBRARY_PATH is\n\"\/usr\/local\/cuda\/lib64\"<\/p>\n\n<p>I have no clue why cudnnCreate is not found. Is there somebody who has successfully installed this configuration and can give me advice?<\/p>\n","answers":[{"AnswerId":"39797729","CreationDate":"2016-09-30T18:20:53.987","ParentId":null,"OwnerUserId":"6906047","Title":null,"Body":"<p>Ubuntu 14.04  &amp;&amp;  cudnnV5.0 &amp;&amp; Cuda7.5<\/p>\n\n<p>I got the some error and solve it in another way.\nFollow the official get-started page, I install the cudnn with these commands below, which is basically just copy those files into our cuda directory<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/get_started\/os_setup.html#optional-install-cuda-gpus-on-linux\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/get_started\/os_setup.html#optional-install-cuda-gpus-on-linux<\/a><\/p>\n\n<blockquote>\n  <p>tar xvzf cudnn-7.5-linux-x64-v5.1-ga.tgz<\/p>\n  \n  <p>sudo cp cuda\/include\/cudnn.h \/usr\/local\/cuda\/include<\/p>\n  \n  <p>sudo cp cuda\/lib64\/libcudnn* \/usr\/local\/cuda\/lib64<\/p>\n  \n  <p>sudo chmod a+r \/usr\/local\/cuda\/include\/cudnn.h \/usr\/local\/cuda\/lib64\/libcudnn*<\/p>\n<\/blockquote>\n\n<p>But after doing this ,if we use ll command to show all the file in \"\/usr\/local\/cuda\/lib64\" and  compare with the origin files <\/p>\n\n<blockquote>\n  <p>ll<\/p>\n<\/blockquote>\n\n<p>it seems that those soft links has broken after copy.\nso I delete them and create manually, like this:<\/p>\n\n<blockquote>\n  <p>sudo rm libcudnn.so.5 libcudnn.so<\/p>\n  \n  <p>sudo ln -sf libcudnn.so.5 libcudnn.so<\/p>\n  \n  <p>sudo ln -sf libcudnn.so.5.1.3 libcudnn.so.5    <\/p>\n<\/blockquote>\n\n<p>after that, execute<\/p>\n\n<blockquote>\n  <p>sudo ldconfig \/usr\/local\/cuda\/lib64<\/p>\n<\/blockquote>\n\n<p>and it finally worked!<\/p>\n"},{"AnswerId":"36965792","CreationDate":"2016-05-01T10:51:39.587","ParentId":null,"OwnerUserId":"5634178","Title":null,"Body":"<p>I get the same error when I forgot to set the <code>LD_LIBRARY_PATH<\/code> and <code>CUDA_HOME<\/code> environment variables:<\/p>\n\n<p>export <code>LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:\/usr\/local\/cuda\/lib64\"<\/code>\nexport <code>CUDA_HOME=\/usr\/local\/cuda<\/code><\/p>\n"},{"AnswerId":"41859792","CreationDate":"2017-01-25T19:14:40.053","ParentId":null,"OwnerUserId":"3759231","Title":null,"Body":"<h2>Errorsolving for windows 10 users:<\/h2>\n\n<ul>\n<li><p>Download cuDNN v5.1 Library for Windows 10 from the <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">cuda site<\/a>,\nregister if necessary.<\/p><\/li>\n<li><p><strong>Copy<\/strong> the <strong>cudnn64_5.dll<\/strong> (cuda\\bin\\cudnn64_5.dll) from that zip\narchive into <\/p><\/li>\n<\/ul>\n\n<blockquote>\n  <p>C:\\Program Files\\NVIDIA GPU Computing\n            Toolkit\\CUDA\\v8.0\\bin\\;<\/p>\n<\/blockquote>\n\n<p>If C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0 is your install PATH\n        for the CUDA toolkit.<\/p>\n"},{"AnswerId":"35760994","CreationDate":"2016-03-03T00:36:24.570","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>I am following this instructions to install TensorFlow in archlinux:\n<a href=\"https:\/\/github.com\/ddigiorg\/AI-TensorFlow\/blob\/master\/install\/install-TF_2016-02-27.md\" rel=\"nofollow\">https:\/\/github.com\/ddigiorg\/AI-TensorFlow\/blob\/master\/install\/install-TF_2016-02-27.md<\/a><\/p>\n\n<p>It seems you need cuDNN v2 or above, which you can get by register for their Accelerated Computing Developer Program, which usually takes 2 days:\n<a href=\"https:\/\/developer.nvidia.com\/accelerated-computing-developer\" rel=\"nofollow\">https:\/\/developer.nvidia.com\/accelerated-computing-developer<\/a><\/p>\n\n<p><strong>UPDATE<\/strong>: It seems you already have cuDNNv2<\/p>\n"},{"AnswerId":"35793510","CreationDate":"2016-03-04T10:21:15.887","ParentId":null,"OwnerUserId":"5997869","Title":null,"Body":"<p>The link sent by jorgemf (thank you) describes a Python 3.5 installation and I almost switched to Python 3.5. \nMy last attempt with my present installation was to again copy the cuDNN libraries to \/usr\/local\/cuda\/lib64. <\/p>\n\n<p>And it worked! So the problem is solved, although I still don't know why I had it.<\/p>\n"}]}
{"QuestionId":35703201,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-02-29T15:24:02.250","AcceptedAnswerId":null,"OwnerUserId":1585638.0,"Title":"speed benchmark for testing tensorflow install","Body":"<p>I'm doubting whether tensorflow is correctly configured on my gpu box, since it's about 100x slower per iteration to train a simple linear regression model (batchsize = 32, 1500 input features, 150 output variables) on my fancy gpu machine than on my laptop. <\/p>\n\n<p>I'm using a Titan X, with a modern cpu, etc. nvidia-smi says that I'm only at 10% gpu utilization, but I expect that's because of the small batchsizes. I'm not using a feed_dict to move data into the computation graph. Everything is coming via a tf.decode_csv and tf.train.shuffle_batch.<\/p>\n\n<p>Does anyone have any recommendations for how to easily test whether my install is correct?  Are there any simple speed benchmarks? The speed difference between my laptop and the gpu machine is so dramatic that I'm expecting that things aren't configured properly. <\/p>\n","answers":[{"AnswerId":"35710218","CreationDate":"2016-02-29T21:40:42.827","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Try <code>tensorflow\/tensorflow\/models\/image\/mnist\/convolutional.py<\/code>, that'll print per-step timing.<\/p>\n\n<p>On Tesla K40c that should get about <code>16 ms<\/code> per step, while about <code>120 ms<\/code> for CPU-only on my 3 year old machine<\/p>\n\n<hr>\n\n<p><em>Edit:<\/em> This moved to the <code>models<\/code> repositories: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/mnist\/convolutional.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/mnist\/convolutional.py<\/a>.<\/p>\n\n<p>The <code>convolutional.py<\/code> file is now at <code>models\/tutorials\/image\/mnist\/convolutional.py<\/code><\/p>\n"},{"AnswerId":"44327563","CreationDate":"2017-06-02T11:14:19.333","ParentId":null,"OwnerUserId":"7572354","Title":null,"Body":"<p>Extending Yaroslavs answer: \nHere is how to do the entire testing process (CUDA and cudNN installed already)<\/p>\n\n<pre><code>git clone https:\/\/github.com\/tensorflow\/models.git\n<\/code><\/pre>\n\n<p>Create a Virtual Environment for tensorflow and install tensorflow<\/p>\n\n<pre><code>virtualenv --system-site-packages -p python3 tf-venv3\nsource tf-venv3\/bin\/activate\npip install --upgrade pip\npip install --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p>Run the model within your Virtual Environment<\/p>\n\n<pre><code>python models\/tutorials\/image\/mnist\/convolutional.py \n<\/code><\/pre>\n\n<p>My GTX 1070 needs ~5ms per step<\/p>\n\n<p><strong>Note:<\/strong> On Geforce 1050 Ti it takes ~10ms per step\n<a href=\"https:\/\/i.stack.imgur.com\/6do73.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6do73.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35704078,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-02-29T16:05:03.887","AcceptedAnswerId":null,"OwnerUserId":5467869.0,"Title":"theano tutorial copy past pieces of code to try it?? where are py code files?","Body":"<p>theano tutorial is great!!\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/<\/a><\/p>\n\n<p>but how to find .py files, since meantime I need copy past pieces of code to try it?? <\/p>\n","answers":[{"AnswerId":"35704229","CreationDate":"2016-02-29T16:13:12.027","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>The tutorial code can all be found in <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/code\/<\/a>.<\/p>\n\n<p>These files are linked to from the tutorial pages that make use of them. For example, the <a href=\"http:\/\/deeplearning.net\/tutorial\/code\/lstm.py\" rel=\"nofollow\">LSTM code<\/a> is linked to from the <a href=\"http:\/\/deeplearning.net\/tutorial\/lstm.html\" rel=\"nofollow\">LSTM tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":35710938,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-02-29T22:24:22.663","AcceptedAnswerId":35711410.0,"OwnerUserId":859453.0,"Title":"Are these functions equivalent in TensorFlow?","Body":"<p>I'm a newbie wit TensorFlow and I've studying it for the last few days.\nI want to understand if the following two functions are equivalent or not:<\/p>\n\n<p>1.<\/p>\n\n<pre><\/pre>\n\n<p>2.<\/p>\n\n<pre><\/pre>\n\n<p>If they are in fact different, what is the main difference?<\/p>\n","answers":[{"AnswerId":"35711410","CreationDate":"2016-02-29T22:55:52.090","ParentId":null,"OwnerUserId":"3314143","Title":null,"Body":"<pre><code>softmax1 = tf.add(tf.matmul(x, weights), biases, name=scope.name)\n<\/code><\/pre>\n\n<p><strong>is not equal to<\/strong><\/p>\n\n<pre><code>softmax2 = tf.nn.softmax(tf.matmul(x, weights) + biases, name=scope.name)\n<\/code><\/pre>\n\n<p>since <code>softmax1<\/code> has no softmax calculation at all while <code>softmax2<\/code> does. See the Tensorflow API for <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax\" rel=\"nofollow\">tf.nn.softmax<\/a>. The <a href=\"https:\/\/en.wikipedia.org\/wiki\/Softmax_function\" rel=\"nofollow\">general idea of a softmax<\/a> is that it <em>normalizes<\/em> the input by rescaling the whole data sequence ensuring their entries are in the interval (0, 1) and the sum is 1.<\/p>\n\n<p>The only thing that is equal between the two statements is the basic calculation. <code>+<\/code> does the same thing <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#add\" rel=\"nofollow\">tf.add<\/a> does so <code>tf.add(tf.matmul(x, weights), biases)<\/code> <strong>is equal<\/strong> to <code>tf.matmul(x, weights) + biases<\/code>.<\/p>\n\n<hr>\n\n<p><strong>EDIT<\/strong>: To add some clarification (I think you do not know really know what softmax is doing?):<\/p>\n\n<pre><code>tf.matmul(x, W) + bias\n<\/code><\/pre>\n\n<p>Calculates a matrix multiplication between <code>x<\/code> (your input vector) and <code>W<\/code> the weights for the current layer. Afterwards the <code>bias<\/code> is added.<\/p>\n\n<p>This calculation models the <em>activation<\/em> of one layer. Additionally you have an activation function, like the <code>sigmoid<\/code> function which transforms your activation. So for one layer you normally do something like this:<\/p>\n\n<pre><code>h1 = tf.sigmoid(tf.matmul(x, W) + bias)\n<\/code><\/pre>\n\n<p>Here <code>h1<\/code> would be the activation of this layer.<\/p>\n\n<p>The <code>softmax<\/code> operation simply rescales your input. E.g., if you got this activation on your output layer:<\/p>\n\n<pre><code>output = [[1.0, 2.0, 3.0, 5.0, 0.5, 0.2]]\n<\/code><\/pre>\n\n<p>The softmax rescales this input for fitting the values in the interval (0, 1) and being the sum equal to 1:<\/p>\n\n<pre><code>tf.nn.softmax(output)\n&gt; [[ 0.01497873,  0.0407164 ,  0.11067866,  0.81781083,  0.00908506,\n     0.00673038]]\n\ntf.reduce_sum(tf.nn.softmax(output))\n&gt; 1.0\n<\/code><\/pre>\n"}]}
{"QuestionId":35714540,"AnswerCount":1,"Tags":"<macos><numpy><nan><caffe><blas>","CreationDate":"2016-03-01T04:15:07.543","AcceptedAnswerId":null,"OwnerUserId":1270543.0,"Title":"Matrix multiplication result in NaN by using Mac osx Accelerate framework","Body":"<p>When I use  for , a strange phenomenon happened. Here is the info of the  installed on my MacOSX:<\/p>\n\n<pre><\/pre>\n\n<p>Assume there are two matrices , , and , I found there are some positions in  is . For example: . And then I do the following operation:<\/p>\n\n<pre><\/pre>\n\n<p>And I found  is not a ! Besides, I also found that the positions of  in y are difference between each experiment.<\/p>\n\n<p>I used the same  and  and do that operation in another PC with , the info of the  in that PC is:<\/p>\n\n<pre><\/pre>\n\n<p> is normal and there are no any  in it.<\/p>\n\n<p>I know how  is implemented and  is called for the double-precision matrix-times-matrix case. <strong>So Is there any bug in the MacOSX Accelerate framework??<\/strong><\/p>\n\n<hr>\n\n<p>The matrix  is the weight blobs in the , and  is the input to this layer. I also used  on my MacOSX, which used the  as its matrix multiplication library.That strange things also happened.<\/p>\n","answers":[{"AnswerId":"36547807","CreationDate":"2016-04-11T11:46:07.523","ParentId":null,"OwnerUserId":"923709","Title":null,"Body":"<p>If the matrices contain zero and inf\/-inf then the product zero*inf seems to be defined as nan. There seem to be inconsistencies in the BLAS libraries, mine (libblas3) sometimes omits the NaN values and gives different results for differently sized matrices. Compare the positions which differ by AxB == trans(trans(A)*trans(B)) where A and B are matrices and trans() is the transpose of a matrix.<\/p>\n"}]}
{"QuestionId":35714832,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-01T04:43:02.323","AcceptedAnswerId":35729020.0,"OwnerUserId":6000679.0,"Title":"TensorFlow shuffle_batch not working","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Hi,<\/p>\n\n<p>I am trying to build a simple neural network in TensorFlow. I am trying to load my input images in batches. Now I am testing the code with 11 images and batch_size = 5. Eventually I will be working with 100000 images.<\/p>\n\n<p>This bit of code is modified from TensorFlow's cifar10.py example. For some reason my code stops (doesn't terminate, it just hangs there) at <\/p>\n\n<p>I've tried different combinations of batch_size, capacity, min_after_dequeue etc. but I still can't figure out what is going wrong.<\/p>\n\n<p>Any help will be much appreciated! Thank you!<\/p>\n","answers":[{"AnswerId":"35729020","CreationDate":"2016-03-01T17:07:21.493","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like the problem arises because the statement<\/p>\n\n<pre><code>tf.train.start_queue_runners(sess=sess)\n<\/code><\/pre>\n\n<p>...executes before any queue runners have been created. If you move this line after <code>images = get_batch()<\/code>, your program should work.<\/p>\n\n<p>What is the problem here? The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#shuffle_batch\"><code>tf.train.shuffle_batch()<\/code><\/a> function internally uses a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#RandomShuffleQueue\"><code>tf.RandomShuffleQueue<\/code><\/a> to produce a randomized batch. Currently, the only way to get elements into that queue is to run a step that invokes the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.enqueue\"><code>q.enqueue()<\/code><\/a> op. To make this easier, TensorFlow has a notion of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/threading_and_queues\/index.html#queuerunner\">\"queue runners\"<\/a> that are implicitly collected as you build your graph, and then started with a call to <code>tf.train.start_queue_runners()<\/code>. However, calling <code>tf.train.start_queue_runners()<\/code> only starts the queue runners that have been defined at that point in time, so it must come <strong>after<\/strong> the code that creates the queue runners.<\/p>\n"}]}
{"QuestionId":35714995,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-01T05:00:07.607","AcceptedAnswerId":35715513.0,"OwnerUserId":2125302.0,"Title":"Computing exact moving average over multiple batches in tensorflow","Body":"<p>During training, I would like to write the average loss over the last N mini-batches to  as a way of smoothing the very noisy batch loss. It's easy to compute this in python and print it, but I would like to add this to a summary so that I can see it in tensorboard. Here's an overly simplified example of what I'm doing now.<\/p>\n\n<pre><\/pre>\n\n<p>I'm aware that I could use  with a decay of 1.0, but I would still need some way to reset this every N batches. Really, if all I care about is visualizing loss in tensorboard, the reset probably isn't necessary, but I'm still curious how one would go about aggregating across batches for other reasons (e.g. computing total accuracy over a test dataset that is too big to run in a single batch).<\/p>\n","answers":[{"AnswerId":"35715513","CreationDate":"2016-03-01T05:43:30.570","ParentId":null,"OwnerUserId":"728291","Title":null,"Body":"<p>Passing data from python to a graph function like <code>tf.scalar_summary<\/code> can be done using a placeholder and feed_dict.<\/p>\n\n<pre><code>average_pl = tf.placeholder(tf.float32)\naverage_summary = tf.summary.scalar(\"average_loss\", average_pl)\nwriter = tf.summary.FileWriter(\"\/tmp\/mnist_logs\", sess.graph_def)\n\nlosses = []\nfor i in range(10000):\n  _, loss = session.run([train_op, loss_op])\n  losses.append(loss)\n  if i % 100 == 0:\n    # How to produce a scalar_summary here?\n    feed = {average_pl: sum(losses)\/len(losses)}\n    summary_str = sess.run(average_summary, feed_dict=feed)\n    writer.add_summary(summary_str, i)\n    losses = []\n<\/code><\/pre>\n\n<p>I haven't tried it and this was hastily copied from the visualizing data how to but I expect something like this would work.<\/p>\n"},{"AnswerId":"36230796","CreationDate":"2016-03-26T02:01:34.953","ParentId":null,"OwnerUserId":"6116225","Title":null,"Body":"<p>You can manually construct the Summary object, like this:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from tensorflow.core.framework import summary_pb2\n\ndef make_summary(name, val):\n    return summary_pb2.Summary(value=[summary_pb2.Summary.Value(tag=name, \n                                                                simple_value=val)])\n\nsummary_writer.add_summary(make_summary('myvalue', myvalue), step)\n<\/code><\/pre>\n"}]}
{"QuestionId":35715494,"AnswerCount":1,"Tags":"<ipython><theano><jupyter-notebook>","CreationDate":"2016-03-01T05:41:43.747","AcceptedAnswerId":null,"OwnerUserId":2436776.0,"Title":"libcublas.so cannot open shared object file: No such file or directory in ipython and notebook","Body":"<p>The missing  problem has been around for some time. The most common problem is that the  and  environment variable is not set properly. And solutions for command line scenarios have been posted in the <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/845363\/libcublas-so-7-0-cannot-open-shared-object-file\/\" rel=\"nofollow noreferrer\">NVIDIA forum<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/20026127\/oserror-libcublas-so-3-cannot-open-shared-object-file-no-such-file-or-directo\">here<\/a>.<\/p>\n\n<p>But no specific solution has been out for similar symptoms in ipython or notebook. Here is my own work around.<\/p>\n","answers":[{"AnswerId":"35715495","CreationDate":"2016-03-01T05:41:43.747","ParentId":null,"OwnerUserId":"2436776","Title":null,"Body":"<p>The problem is still due to environment variables: ipython and notebook cannot propagate the settled <code>$PATH<\/code> and <code>$LD_LIBRARY_PATH<\/code>. So when this happens, the first thing to check is<\/p>\n\n<pre><code>import os; print(os.environ['PATH']); print(os.environ['LD_LIBRARY_PATH'])\n<\/code><\/pre>\n\n<p>Most probably the <code>bin<\/code> and <code>lib<\/code> paths are not in these environs.<\/p>\n\n<p>To solve this for ipython, use <code>sudo PATH=$PATH LD_LIBRARY_PATH=$LD_LIBRARY_PATH ipython<\/code> when starting ipython.<\/p>\n\n<p>And for notebook, add these lines to the end of <code>jupyter_notebook_config.py<\/code>:<\/p>\n\n<pre><code>import os\nos.environ['PATH'] += ':\/usr\/local\/cuda\/bin'\nos.environ['LD_LIBRARY_PATH'] = '\/usr\/local\/cuda\/lib64'\n<\/code><\/pre>\n"}]}
{"QuestionId":35716400,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-01T06:47:59.307","AcceptedAnswerId":null,"OwnerUserId":5944204.0,"Title":"How to get Tensor Value which is define inside a function or variable scope?","Body":"<p>Suppose I define some tensor as:<\/p>\n\n<pre><\/pre>\n\n<p>I can get the value of RESULT by using .<\/p>\n\n<p>But if I define  that <em>RESULT<\/em> tensor in some function eg.<\/p>\n\n<pre><\/pre>\n\n<p>then how can I get the value of RESULT? Same way if I define within a Variable Scope then What is the way to get value of that tensor? <\/p>\n","answers":[{"AnswerId":"46745046","CreationDate":"2017-10-14T13:21:55.950","ParentId":null,"OwnerUserId":"8744105","Title":null,"Body":"<p>For me the following worked:<\/p>\n\n<pre><code>    with tf.Session() as sess:\n        ...\n        op = sess.graph.get_operations()\n        test = tf.get_default_graph().get_tensor_by_name(\"tensor_name\")\n<\/code><\/pre>\n\n<ul>\n<li><code>op = sess.graph.get_operations()<\/code> gives you the list of all tensors used in computations. You can print it <code>print([p.values() for p in op])<\/code> and find the name of the tensor you're interested in.<\/li>\n<li><code>test = tf.get_default_graph().get_tensor_by_name(\"tensor_name\")<\/code> gives you the tensor with a specified name.<\/li>\n<\/ul>\n"},{"AnswerId":"35724094","CreationDate":"2016-03-01T13:20:59.703","ParentId":null,"OwnerUserId":"5499453","Title":null,"Body":"<p>you can use collection I think:<\/p>\n\n<pre><code>def fun1():\n    RESULT = tf.add(tf.matmul(Var_1, Var_2))\n    tf.add_to_collection(\"FV\",RESULT )\ndef fun2():\n    RESULT =tf.getCollection(\"FV\")[0]\n    sess.run(RESULT)\n<\/code><\/pre>\n"}]}
{"QuestionId":35717535,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-01T08:00:52.310","AcceptedAnswerId":35723703.0,"OwnerUserId":762295.0,"Title":"Getting variables values from Tensorflow","Body":"<p>After a module is built, and checkpoint is stored in model_ckpt,\nhow can I get the values(weights, biases) of each node ?<\/p>\n","answers":[{"AnswerId":"35723703","CreationDate":"2016-03-01T13:02:55.473","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>One way to do it is to load a checkpoint then call <\/p>\n\n<pre><code>node_value = session.run(variable_name_here)\n<\/code><\/pre>\n\n<p>Where variable_name_here should be there weight tensor<\/p>\n"}]}
{"QuestionId":35718042,"AnswerCount":0,"Tags":"<theano><deep-learning><activation-function>","CreationDate":"2016-03-01T08:34:45.187","AcceptedAnswerId":null,"OwnerUserId":988125.0,"Title":"Implementing Maxout Activation in Theano","Body":"<p>The only example of maxout implementation in Theano is on this <a href=\"http:\/\/hallvardnydal.github.io\/2015\/07\/10\/maxout\/\" rel=\"nofollow\">link<\/a>. My understanding is that I use any activation function and then maxout is just a post processing of the hidden layer outputs. <\/p>\n\n<p>I tried to apply this to my own  class. Below is the class before maxout:<\/p>\n\n<pre><\/pre>\n\n<p>If I understood the <a href=\"http:\/\/hallvardnydal.github.io\/2015\/07\/10\/maxout\/\" rel=\"nofollow\">link<\/a> correctly, the class after maxout implementation should look as below. <strong>Is this correct?<\/strong> If not, could you point out which part I misunderstood? <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35718824,"AnswerCount":1,"Tags":"<mathematical-optimization><tensorflow>","CreationDate":"2016-03-01T09:18:38.653","AcceptedAnswerId":null,"OwnerUserId":5602836.0,"Title":"beginner - solve a simple convex optimisation","Body":"<p>I am a beginner at Tensorflow. I took the line-fit example from the \"getting started\" page, \nand made what I thought is a nearly trivial modification to it, but it fails completely.\nI do not understand.<\/p>\n\n<p>In the modified version, an array b_data is a sum of two known Gaussians, with unknown weights.\nTry to solve for those weights.  This is a quadratic problem that can be solved\nas a linear system.<\/p>\n\n<p>Although the true weights are 0.4, 0.2,  the gradient descent gives w[0] negative,\nand w[1] positive.<\/p>\n\n<p>That is the issue: <strong>although the problem is convex (quadratic even), tensorflow does not find the right answer.<\/strong><\/p>\n\n<p>I think I must be doing something wrong with the loss function?\nIn fact I think using a loss <\/p>\n\n<p><\/p>\n\n<p>is what I want (corresponding to the squared 2-norm || b - b_data ||^2),\nhowever trying this is even worse, it results in NaNs. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"35953313","CreationDate":"2016-03-12T03:13:46.183","ParentId":null,"OwnerUserId":"2642845","Title":null,"Body":"<p>You are not optimizing the right cost function but a larger one as the substraction broadcast your arrays <code>b_data<\/code> and <code>b<\/code> :  <\/p>\n\n<pre><code>&gt; print(tf.square(b - b_data))\nTensor(\"Square:0\", shape=(200, 200), dtype=float32)\n\n&gt;print(tf.square(b[:, 0] - b_data))\nTensor(\"Square:0\", shape=(200,), dtype=float32)\n<\/code><\/pre>\n\n<p>This is caused by a counter intuitive broadcast in the tensorflow implementation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1406\" rel=\"nofollow\">(see for instance this issue)<\/a>. <\/p>\n\n<p>If you replace your loss with <code>loss = tf.reduce_mean(tf.square(b[:, 0] - b_data))<\/code>, the optimization succeeds and gives back the result.<\/p>\n"}]}
{"QuestionId":35719371,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-03-01T09:41:38.493","AcceptedAnswerId":null,"OwnerUserId":1616037.0,"Title":"Array of functions in theano","Body":"<p>Wanted to create a mapping function 'key-to-function' in theano (provides a function given some number from a predefined finite set); this <em>minimal crashing example<\/em> would not work.<\/p>\n\n<pre><\/pre>\n\n<p>Here's stacktrace:<\/p>\n\n<pre><\/pre>\n\n<p>Is it possible in theano at all?<\/p>\n","answers":[]}
{"QuestionId":35720595,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-01T10:35:43.723","AcceptedAnswerId":35726305.0,"OwnerUserId":3733814.0,"Title":"Tensorflow: issue with placeholder and summaries","Body":"<p>I have modified existing cifar10 example to work as a siamese network.\nBut I am facing some difficulties in training it.<\/p>\n\n<p>Changes Made : <\/p>\n\n<ul>\n<li>placeholder instead of queue<\/li>\n<li>custom loss function<\/li>\n<\/ul>\n\n<p>Here is my modified cifar10_train.py :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Modified cifar10.py<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Error I am getting : <\/p>\n\n<pre><\/pre>\n\n<p>Also, when I comment out merge_all_summaries(), the model diverges with loss= NaN<\/p>\n","answers":[{"AnswerId":"35726305","CreationDate":"2016-03-01T15:03:13.307","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem here is that some of the summaries in your graph&mdash;collected by <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#merge_all_summaries\" rel=\"nofollow\"><code>tf.merge_all_summaries()<\/code><\/a>&mdash; <em>depend<\/em> on your placeholders. For example, the code in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"nofollow\"><code>cifar10.py<\/code><\/a> creates summaries for various activations at each step, which depend on the training example used.<\/p>\n\n<p>The solution is to feed the same training batch when you evaluate <code>summary_op<\/code>:<\/p>\n\n<pre><code>if step % 100 == 0:\n  summary_str = sess.run(summary_op, feed_dict={\n      images: image[offset:(offset + batch_size)],\n      images2: image_p[offset:(offset + batch_size)],\n      labels: 1.0 * label[offset:(offset + batch_size)]})\n<\/code><\/pre>\n\n<p>While this gives the smallest modification to your original code, it is slightly inefficient, because it will re-execute the training step every 100 steps. The best way to address this (although it will require some restructuring of your training loop) is to fetch the summaries in the same call to <code>sess.run()<\/code> that performs a training step:<\/p>\n\n<pre><code>if step % 100 == 0:\n  _, loss_value, summary_str = sess.run([train_op, loss, summary_op], feed_dict={\n      images: image[offset:(offset + batch_size)],\n      images2: image_p[offset:(offset + batch_size)],\n      labels: 1.0 * label[offset:(offset + batch_size)]})\n<\/code><\/pre>\n"}]}
{"QuestionId":35723309,"AnswerCount":1,"Tags":"<linux><amazon-ec2><ubuntu-14.04><torch><jupyter-notebook>","CreationDate":"2016-03-01T12:44:53.930","AcceptedAnswerId":35729389.0,"OwnerUserId":5443023.0,"Title":"Jupyter notebook can't find kernel when run through \/etc\/rc.local","Body":"<p>I have a Jupyter Notebook server on an EC2 instance with python2 (default) and ITorch kernels. Everything works fine when run through SSH command line.<\/p>\n\n<p>The problem is that when I try to make this run at startup, the notebook server can't find ITorch kernel. I does work fine for python notebooks though. The error I get is a window saying<\/p>\n\n<pre><\/pre>\n\n<p>What I did was adding this line to <\/p>\n\n<pre><\/pre>\n\n<p>The notebook server apparently runs fine. This is from the system log<\/p>\n\n<pre><\/pre>\n\n<p>These are the only lines I modified in my config file<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35729389","CreationDate":"2016-03-01T17:26:34.047","ParentId":null,"OwnerUserId":"5443023","Title":null,"Body":"<p>@Thomas K comment solved it:<\/p>\n\n<p>\"The ITorch kernel is probably set up for your user, and when the system runs it, it's a different user. Run <code>jupyter kernelspec list<\/code> to see where the kernels are, and if ITorch is under your home directory, copy it to <code>\/usr\/local\/share\/jupyter\/kernels\/<\/code>\".<\/p>\n"}]}
{"QuestionId":35724469,"AnswerCount":1,"Tags":"<tensorflow><gradient-descent>","CreationDate":"2016-03-01T13:39:39.830","AcceptedAnswerId":35728022.0,"OwnerUserId":574187.0,"Title":"Where can I have a look at TensorFlow gradient descent main loop?","Body":"<p>(Sorry if this sounds a bit naive) \nI want to have a look at the meat of the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/deff101b27e2d3d2d0f68417991e9183bdcd3256\/tensorflow\/python\/training\/gradient_descent.py\" rel=\"nofollow\">TensorFlow<\/a> implementation for GradientDescent - and see for myself how are they handling termination condition, step-size adaptiveness, etc. I traced the code down for  but I can't find the implementation :( <\/p>\n","answers":[{"AnswerId":"35728022","CreationDate":"2016-03-01T16:20:26.100","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TensorFlow <code>Optimizer<\/code> interface, (which <code>GradientDescentOptimizer<\/code> implements) defines a a single step of minimization. Termination conditions or adjusting step size is implemented by the user. In <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">MNIST for Beginners<\/a> tutorial, the termination conditions is \"stop after 1000\" steps which you can see in <code>for i in range(1000)<\/code> loop<\/p>\n\n<p><code>apply_gradient_descent(a,b,c)<\/code> is a fused op that multiplies <code>c<\/code> by <code>b<\/code> and adds it to <code>a<\/code>. There are some extra levels of indirection to go from Python wrapper to C++ implementation detailed in <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/g3doc\/how_tos\/adding_an_op\/index.md\" rel=\"nofollow\">Adding a new op HowTo<\/a>, but as a shortcut you can usually find C++ implementation by converting from snake-case and searching for that, so <code>ApplyGradientDescent<\/code> in this case. That leads to implementation in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/training_ops.cc#L32\" rel=\"nofollow\">tensorflow\/core\/kernels\/training_ops.cc<\/a><\/p>\n"}]}
{"QuestionId":35725513,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-03-01T14:27:14.657","AcceptedAnswerId":35727708.0,"OwnerUserId":2857133.0,"Title":"TensorFlow: cast a float64 tensor to float32","Body":"<p>I am trying to use:  but the standard optimizers do not work with . Therefore I want to truncate my  from  to only .  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35727708","CreationDate":"2016-03-01T16:05:27.990","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The short answer is that you can convert a tensor from <code>tf.float64<\/code> to <code>tf.float32<\/code> using the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#cast\" rel=\"noreferrer\"><code>tf.cast()<\/code><\/a> op:<\/p>\n\n<pre><code>loss = tf.cast(loss, tf.float32)\n<\/code><\/pre>\n\n<p>The longer answer is that this will not solve all of your problems with the optimizers. (The lack of support for <code>tf.float64<\/code> is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/761\" rel=\"noreferrer\">known issue<\/a>.) The optimizers require that all of the <strong><code>tf.Variable<\/code><\/strong> objects that you are trying to optimize must also have type <code>tf.float32<\/code>.<\/p>\n"}]}
{"QuestionId":35726182,"AnswerCount":1,"Tags":"<python><classification><theano><deep-learning><conv-neural-network>","CreationDate":"2016-03-01T14:57:24.593","AcceptedAnswerId":36371926.0,"OwnerUserId":null,"Title":"Predicting the class of images using parameters of trained ConvNet","Body":"<p>I am trying to use Convolutional Neural Networks for classification of images. I have gone through this <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/\" rel=\"nofollow\">tutorial<\/a> on Deep Learning and implemented the given <a href=\"https:\/\/github.com\/mnielsen\/neural-networks-and-deep-learning\/blob\/master\/src\/network3.py\" rel=\"nofollow\">code<\/a> with many modifications. I added more layers of convolution and max-pooling and altered the inputs for accepting an input of 166x166. To save the parameters after training, we used cPickle.dump() in the function save() which is defined for each layer (ConvPool, FullyConnected and Softmax) separately.  This function is called after the training is completed, in sgd() method, for all layers. In another program, the parameters for the softmax, fully connected layer and convolution layers are loaded back from the .p pickled files in another program which is the same, except we are not calling the SGD method. The problem is that, I want to print y_out of the Softmax layer (y_out is used to calculate the accuracy of our network), to obtain a prediction of the class of the image. But after trying <\/p>\n\n<pre><\/pre>\n\n<p>I am still getting 'argmax' as the value of Tensor Variable,' y_out or otherwise an error of Missing Input Error when I used the eval() function to get the value of the Tensor Variable.<\/p>\n\n<p>Thus need help to print the prediction for a single test image.    <\/p>\n\n<p>This is the code for network3.py (renamed as net3.py) after our modification: <\/p>\n\n<pre><\/pre>\n\n<p>The main code is:<\/p>\n\n<pre><\/pre>\n\n<p>The code to load the single test image in .mat form is as follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36371926","CreationDate":"2016-04-02T10:12:47.917","ParentId":null,"OwnerUserId":"2446478","Title":null,"Body":"<p>You may use the test_mb_predictions function which is already defined in your code . You just need to load a single image, in your testing code, with a batch-size of 1, (and not 10) and print the class as returned by test_mb_predictions.<\/p>\n"}]}
{"QuestionId":35726909,"AnswerCount":1,"Tags":"<gnuplot><torch>","CreationDate":"2016-03-01T15:30:43.183","AcceptedAnswerId":35764922.0,"OwnerUserId":5489845.0,"Title":"gnuplot: multiplots using multiple lines in each plot and in-line data","Body":"<p>I am trying to call gnuplot from torch in order to plot several things. I am trying to use multiplots (like subplots in matlab), and at the same time plot different curves in the same plots. Furthermore, I'm defining the data in-line, i.e., avoiding writing the data in external files (plot '-').<\/p>\n\n<p>I've tried plotting several curves with '-' in several ways, with multiplot previous and next, with replot... but any combination messes up the layout in one way or another. Does anybody knows how to do this or can give some hints? <\/p>\n\n<p>Thanks!<\/p>\n\n<p><strong>Update<\/strong>: Added a small example. Blue and red lines should be plotted in the same plot (top one), while green one should be plotted alone in the bottom one. <\/p>\n\n<p>Example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35764922","CreationDate":"2016-03-03T06:41:41.733","ParentId":null,"OwnerUserId":"4927616","Title":null,"Body":"<p>I'm not familiar with torch, but I think your problem is with trying to do three separate plot statements.  In gnuplot, to do multiple curves in the same plot (two in your top plot), you specify them in the same command separated by a comma.  Feeding inline data, this means that you will need to provide one set of data, end it with e, provide another, and then end it with e.<\/p>\n\n<p>Additionally, you can use <code>lc<\/code> (linecolor) instead of <code>lt<\/code> to set your colors.<\/p>\n\n<p>In straight gnuplot, you will do this like so:<\/p>\n\n<pre><code>set terminal x11 0 position 1200,20 persist\nset multiplot layout 2,1\nplot '-' lc rgb 'blue', '-' lc rgb 'red'\n0 0\n100 30\ne\n0 30\n100 60\ne\nplot '-' lc rgb 'green'\n0 60\n100 90\ne\nunset multiplot\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Lc5HJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Lc5HJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If you want lines, just add <code>with lines<\/code> to all three plot specifications like<\/p>\n\n<pre><code>plot '-' lc rgb 'blue' with lines, '-' lc rgb 'red' with lines\nplot '-' lc rgb 'green' with lines\n<\/code><\/pre>\n\n<p>or change the default data style with <code>set style data lines<\/code>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/X7UJ1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/X7UJ1.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35730161,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-01T18:05:25.893","AcceptedAnswerId":null,"OwnerUserId":460760.0,"Title":"How to convert a list of tensors of dim N to a tensor of dim N+1","Body":"<p>I need to convert a list of tensors of dimensionality N to a new tensor with dimensionality N+1 so that the new dimension would be the right most dimension.<\/p>\n\n<p>For example if x and y would be tensors of shape (4,3) both then I am trying to create a new tensor z of shape (4,3,2) by forming z and setting tensor x as the 0th element along the third dimension and setting tensor y as the 1st element along the third dimension. In pseudocode:<\/p>\n\n<pre><\/pre>\n\n<p>What's the best way to do that in Tensorflow. I was unable to figure it out from the documentation of TF 0.7.1.<\/p>\n","answers":[{"AnswerId":"43435988","CreationDate":"2017-04-16T10:00:45.207","ParentId":null,"OwnerUserId":"2006462","Title":null,"Body":"<p>dga's method works, but <code>tf.pack()<\/code> has been removed from TensorFlow V1.0 onwards. <\/p>\n\n<p>You can use <code>tf.stack()<\/code> to achieve the same. <\/p>\n\n<p>Docs: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack<\/a><\/p>\n"},{"AnswerId":"35730627","CreationDate":"2016-03-01T18:29:55.150","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>If I'm reading you correctly, you want to <em>interleave<\/em> the data of the two tensors.<\/p>\n\n<p>You want to <code>tf.pack()<\/code> them together, which would form a tensor of shape <code>[2, 4, 3]<\/code> and then <code>tf.transpose([1, 2, 0])<\/code> that resulting tensor to get to the interleaving you want.<\/p>\n"}]}
{"QuestionId":35731506,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-03-01T19:16:22.593","AcceptedAnswerId":35736757.0,"OwnerUserId":246567.0,"Title":"Unaggregated gradients \/ gradients per example in tensorflow","Body":"<p>Given a simple mini-batch gradient descent problem on mnist in tensorflow (like in this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"noreferrer\">tutorial<\/a>), how can I retrieve the gradients for each example in the batch individually.<\/p>\n\n<p> seems to return gradients averaged over all examples in the batch. Is there a way to retrieve gradients before aggregation?<\/p>\n\n<p>Edit: A first step towards this answer is figuring out at which point tensorflow averages the gradients over the examples in the batch. I thought this happened in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/8048088bfb82846078f8023bc6199e6424926624\/tensorflow\/python\/ops\/gradients.py#L437\" rel=\"noreferrer\">_AggregatedGrads<\/a>, but that doesn't appear to be the case. Any ideas?<\/p>\n","answers":[{"AnswerId":"35741903","CreationDate":"2016-03-02T08:26:16.883","ParentId":null,"OwnerUserId":"246567","Title":null,"Body":"<p>To partly answer my own question after tinkering with this for a while. It appears that it is possible to manipulate gradients per example while still working in batch by doing the following: <\/p>\n\n<ul>\n<li>Create a copy of tf.gradients() that accepts an extra tensor\/placeholder with example-specific factors<\/li>\n<li>Create a copy of _AggregatedGrads() and add a custom aggregation method that uses the example-specific factors<\/li>\n<li>Call your custom tf.gradients function and give your loss as a list of slices:<\/li>\n<\/ul>\n\n<p><code>\n    custagg_gradients(\n          ys=[cross_entropy[i] for i in xrange(batch_size)],<br>\n          xs=variables.trainable_variables(), \n          aggregation_method=CUSTOM, \n          gradient_factors=gradient_factors\n    )\n<\/code><\/p>\n\n<p>But this will probably have the same complexity as doing individual passes per example, and I need to check if the gradients are correct :-).<\/p>\n"},{"AnswerId":"35736757","CreationDate":"2016-03-02T01:21:29.963","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><code>tf.gradients<\/code> returns the gradient with respect to the loss. This means that if your loss is a sum of per-example losses, then the gradient is also the sum of per-example loss gradients. <\/p>\n\n<p>The summing up is implicit. For instance if you want to minimize the sum of squared norms of <code>Wx-y<\/code> errors, the gradient with respect to <code>W<\/code> is <code>2(WX-Y)X'<\/code> where <code>X<\/code> is the batch of observations and <code>Y<\/code> is the batch of labels. You never explicitly form \"per-example\" gradients that you later sum up, so it's not a simple matter of removing some stage in the gradient pipeline.<\/p>\n\n<p>A simple way to get <code>k<\/code> per-example loss gradients is to use batches of size 1 and do <code>k<\/code> passes. Ian Goodfellow <a href=\"http:\/\/arxiv.org\/pdf\/1510.01799v2.pdf\" rel=\"noreferrer\">wrote up<\/a> how to get all <code>k<\/code> gradients in a single pass, for this you would need to specify gradients explicitly and not rely on <code>tf.gradients<\/code> method<\/p>\n"},{"AnswerId":"48584037","CreationDate":"2018-02-02T13:46:06.173","ParentId":null,"OwnerUserId":"3309610","Title":null,"Body":"<p>One way of retrieving gradients before aggregation is to use the <code>grads_ys<\/code> parameter. A good discussion is found here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42399401\/use-of-grads-ys-parameter-in-tf-gradients-tensorflow\">Use of grads_ys parameter in tf.gradients - TensorFlow<\/a><\/p>\n\n<p>EDIT: <\/p>\n\n<p>I haven't been working with Tensorflow a lot lately, but here is an open issue tracking the best way to compute unaggregated gradients:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/675\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/675<\/a><\/p>\n\n<p>There is a lot of sample code solutions provided by users (including myself) that you can try based on your needs.  <\/p>\n"}]}
{"QuestionId":35732281,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-01T19:58:40.937","AcceptedAnswerId":null,"OwnerUserId":5964720.0,"Title":"variable-length rnn padding and mask out padding gradients","Body":"<p>I'm building an rnn and using the sequene_length parameter to supply a list of lengths for sequences in a batch, and all of sequences in a batch are padded to the same length. <\/p>\n\n<p>However, when doing backprop, is it possible to mask out the gradients corresponding to the padded steps, so these steps would have 0 contribution to the weight updates? I'm already masking out their corresponding costs like this (where batch_weights is a vector of 0's and 1's, where the elements corresponding to the padding steps are 0's): <\/p>\n\n<pre><\/pre>\n\n<p>the problem is I'm not sure by doing the above whether the gradients from the padding steps are zeroed out or not? <\/p>\n","answers":[{"AnswerId":"40675844","CreationDate":"2016-11-18T11:21:44.937","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>For all framewise \/ feed-forward (non-recurrent) operations, masking the loss\/cost is enough.<\/p>\n\n<p>For all sequence \/ recurrent operations (e.g. <code>dynamic_rnn<\/code>), there is always a <code>sequence_length<\/code> parameter which you need to set to the corresponding sequence lengths. Then there wont be a gradient for the zero-padded steps, or in other terms, it will have 0 contribution.<\/p>\n"}]}
{"QuestionId":35733302,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2016-03-01T20:55:32.160","AcceptedAnswerId":null,"OwnerUserId":4975126.0,"Title":"Back-propagation in Tensorflow","Body":"<p>I am constructing a time delayed recurrent model and I need to know about how and when TensorFlow computes its backwards step.<\/p>\n\n<p>Consider the following model and pseudo-code:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/h6qlg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/h6qlg.png\" alt=\"A time-delayed recurrent network\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/gist.github.com\/aidangomez\/7657ff6c053d0ccc2f4d\" rel=\"nofollow noreferrer\">Here is a gist to a usable version of this code, based on tensorflow\/python\/ops\/rnn.py<\/a><\/p>\n\n<h1>My Question:<\/h1>\n\n<p>For time steps where cells are not called (i.e. at T=1, unit_0 is called, while all the rest are not) are their weights updated? I'm torn as to whether it would be a good idea to have them update at every state or not. The cells themselves haven't been exposed to any new data in the steps, so I afraid that back propagating may result in over-correction. Would appreciate other's insights into this.<\/p>\n\n<p>Let me know if any clarification is necessary.<\/p>\n","answers":[]}
{"QuestionId":35733917,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-01T21:29:49.333","AcceptedAnswerId":35734473.0,"OwnerUserId":1002035.0,"Title":"TensorFlow: Restoring variables from from multiple checkpoints","Body":"<p>I have the following situation:<\/p>\n\n<ul>\n<li><p>I have 2 models written in 2 separate scripts:<\/p><\/li>\n<li><p>Model A consists of variables , , and , and is written in <\/p><\/li>\n<li><p>Model B consists of variables , , and  , and is written in B.py<\/p><\/li>\n<\/ul>\n\n<p>In each of  and ,  I have a  that saves the checkpoint of all the local variables, and let's call the checkpoint files  and  respectively.<\/p>\n\n<p>I now want to make a model C that uses  and . I can make it so that the exact same variable name for  is used in both A and C by using the var_scope (and the same for ).<\/p>\n\n<p>The question is how might I load  and  from  and  into model C? For example, would the following work?<\/p>\n\n<pre><\/pre>\n\n<p>Would an error be raised if you are try to restore the same session twice? Would it complain that there are no allocated \"slots\" for the extra variables (, , , ), or would it simply restore the variables it can, and only complain if there are some other variables in C that are uninitialized?<\/p>\n\n<p>I'm trying to write some code to test this now but I would love to see a canonical approach to this problem, because one encounters this often when trying to re-use some pre-trained weights.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"35734473","CreationDate":"2016-03-01T22:02:21.903","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You would get a <code>tf.errors.NotFoundError<\/code> if you tried to use a saver (by default representing all six variables) to restore from a checkpoint that does not contain all of the variables that the saver represents. (Note however that you are free to call <code>Saver.restore()<\/code> multiple times in the same session, for any subset of the variables, as long as all of the requested variables are present in the corresponding file.)<\/p>\n\n<p>The canonical approach is to define <strong>two separate <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Saver\" rel=\"noreferrer\"><code>tf.train.Saver<\/code><\/a> instances<\/strong> covering each subset of variables that is entirely contained in a single checkpoint. For example:<\/p>\n\n<pre><code>saver_a = tf.train.Saver([a1])\nsaver_b = tf.train.Saver([b1])\n\nsaver_a.restore(session, ckptA_location)\nsaver_b.restore(session, ckptB_location)\n<\/code><\/pre>\n\n<p>Depending on how your code is built, if you have pointers to <code>tf.Variable<\/code> objects called <code>a1<\/code> and <code>b1<\/code> in the local scope, you can stop reading here.<\/p>\n\n<p>On the other hand, if variables <code>a1<\/code> and <code>b1<\/code> are defined in separate files, you might need to do something creative to retrieve pointers to those variables. Although it's not ideal, what people typically do is to use a common prefix, for example as follows (assuming the variable names are <code>\"a1:0\"<\/code> and <code>\"b1:0\"<\/code> respectively):<\/p>\n\n<pre><code>saver_a = tf.train.Saver([v for v in tf.all_variables() if v.name == \"a1:0\"])\nsaver_b = tf.train.Saver([v for v in tf.all_variables() if v.name == \"b1:0\"])\n<\/code><\/pre>\n\n<p>One final note: you don't have to make heroic efforts to ensure that the variables have the same names in A and C. You can pass a name-to-<code>Variable<\/code> dictionary as the first argument to the <code>tf.train.Saver<\/code> constructor, and thereby remap names in the checkpoint file to <code>Variable<\/code> objects in your code. This helps if <code>A.py<\/code> and <code>B.py<\/code> have similarly-named variables, or if in <code>C.py<\/code> you want to organize the model code from those files in a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#name_scope\" rel=\"noreferrer\"><code>tf.name_scope()<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":35733929,"AnswerCount":0,"Tags":"<scikit-learn><tensorflow><gradient-descent><vowpalwabbit>","CreationDate":"2016-03-01T21:30:49.907","AcceptedAnswerId":null,"OwnerUserId":574187.0,"Title":"Why no automatic termination for stochastic gradient descent in the frameworks?","Body":"<p>I checked out some notable open-source frameworks with SGD implementations - <a href=\"https:\/\/github.com\/scikit-learn\/scikit-learn\/blob\/51a765acfa4c5d1ec05fc4b406968ad233c75162\/sklearn\/linear_model\/stochastic_gradient.py\" rel=\"nofollow noreferrer\">scikit-learn<\/a>, <a href=\"https:\/\/github.com\/JohnLangford\/vowpal_wabbit\/blob\/d8cc7312b16fcd44655266e348cebfa3a68458fd\/vowpalwabbit\/gd.cc\" rel=\"nofollow noreferrer\">vowpal-wabbit<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/35724469\/where-can-i-have-a-look-at-tensorflow-gradient-descent-main-loop\/35728022#35728022\">tensor-flow<\/a>. <\/p>\n\n<p>All of them leave the task of deciding how many iterations to the user! scikit requires the user to specify it explicitly, vowpal assumes by default 1 epoch (pass thru all examples) but allows changing to any number of epochs, and tensor implements just a single step for a single example, leaving the entire iteration loop to the user.<\/p>\n\n<p>Why is that? The task of deciding on termination isn't trivial at all- should it be decided when the loss doesn't get any better? the average loss for the last N iterations? Should the user use a validation\/hold-out examples for measuring loss? Or maybe it's not the loss at all and we should check if the optimized weights aren't changing by much? Should we check after every example for termination, or once in a while?<\/p>\n\n<p>Would be happy if someone sheds a light on this design decision, am I missing something and it can't be done internally? The theory for this area is heavy and I was hoping for some support from the frameworks..<\/p>\n","answers":[]}
{"QuestionId":35735118,"AnswerCount":1,"Tags":"<linux><tensorflow><deep-learning>","CreationDate":"2016-03-01T22:45:18.963","AcceptedAnswerId":35738250.0,"OwnerUserId":5977269.0,"Title":"Solving SVHN using Tensorflow Error: \"Resource exhausted: OOM when allocating tensor..\"","Body":"<p>I was trying to solve \"SVHN\" dataset classification problem using Convolutional Neural Network provided here <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html#convolutional-neural-networks<\/a> <\/p>\n\n<p>I read the data and formatted them in this way:<\/p>\n\n<pre><\/pre>\n\n<p>In the  function, I modified the code a little bit as input images in  and  were already in 4D format.<\/p>\n\n<pre><\/pre>\n\n<p>I ended up with these errors:<\/p>\n\n<pre><\/pre>\n\n<p>Please let me know if I did any mistake in any logic.<\/p>\n\n<p>Thanks<\/p>\n\n<p>Sarah<\/p>\n","answers":[{"AnswerId":"35738250","CreationDate":"2016-03-02T03:58:11.993","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Note that your data contains 2*32*3*73257 entries, that's 900 MB for float or 1800MB for double. So you have 1800MB allocated at <code>read_input['X']<\/code>, then TF converts it to a tensor to feed into <code>cast<\/code> which is another 900MB. Output of <code>tf.cast<\/code>  is another 900MB tensor, and output of <code>transpose<\/code> is another 900MB tensor.<\/p>\n\n<p>So you may need 4.5GB of RAM for this to work.<\/p>\n\n<p>In general this approach (converting to a <code>Constant<\/code> node) is only recommended for \"small\" problems. There's a hard limit of 2GB you can put into a constant, but even much smaller values (ie >100MB) may cause problems if you ever move to GPU (example <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/838\" rel=\"nofollow\">here<\/a>)<\/p>\n\n<p>An alternative, scalable way, is to use an input pipeline like in Cifar example<\/p>\n"}]}
{"QuestionId":35735162,"AnswerCount":3,"Tags":"<ubuntu><tensorflow>","CreationDate":"2016-03-01T22:48:26.010","AcceptedAnswerId":35787103.0,"OwnerUserId":1850692.0,"Title":"TensorFlow (GPU) libcudart.so.7.5 error -- cannot open shared object file","Body":"<p>Installing TensorFlow with GPU on Ubuntu 14.04. Fairly new to the Ubuntu\/UNIX environment so lots of things I don't get. When searching this issue, I came across instances of being unable to import CUDA 7.0 when they have CUDA 7.5 installed (since TensorFlow doesn't support 7.5). I have CUDA 7.0 installed, but it seems to be looking for 7.5 - why is this? I have in my .bashrc file the following lines:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm still seeing the issue below... Anyone have any advice?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35781268","CreationDate":"2016-03-03T19:28:49.783","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>If you're using the latest r0.7 binaries, in my experience they are now built to support CUDA 7.5 by default...whereas r0.6 binaries worked with CUDA 7.0<\/p>\n\n<p>You can build tensorflow from source to work with CUDA 7.0 instead, see the documentation here <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#optional-install-cuda-gpus-on-linux\" rel=\"nofollow\">Configure TensorFlow's canonical view of Cuda libraries<\/a><\/p>\n\n<pre><code>$ .\/configure\nPlease specify the location of python. [Default is \/usr\/bin\/python]:\nDo you wish to build TensorFlow with GPU support? [y\/N] y\nGPU support will be enabled for TensorFlow\n\nPlease specify the Cuda SDK version you want to use, e.g. 7.0. [Leave\nempty to use system default]: 7.0\n\nPlease specify the location where CUDA 7.0 toolkit is installed. Refer to\nREADME.md for more details. [default is: \/usr\/local\/cuda]: \/usr\/local\/cuda\n\nPlease specify the Cudnn version you want to use. [Leave empty to use system\ndefault]: 4.0.4\n\nPlease specify the location where the cuDNN 4.0.4 library is installed. Refer to\nREADME.md for more details. [default is: \/usr\/local\/cuda]: \/usr\/local\/cudnn-r4-rc\/\n\nPlease specify a list of comma-separated Cuda compute capabilities you want to\nbuild with. You can find the compute capability of your device at: \nhttps:\/\/developer.nvidia.com\/cuda-gpus.\nPlease note that each additional compute capability significantly increases your\nbuild time and binary size. [Default is: \\\"3.5,5.2\\\"]: 3.5\n\nSetting up Cuda include\nSetting up Cuda lib64\nSetting up Cuda bin\nSetting up Cuda nvvm\nConfiguration finished\n<\/code><\/pre>\n"},{"AnswerId":"35787103","CreationDate":"2016-03-04T02:50:47.690","ParentId":null,"OwnerUserId":"6006609","Title":null,"Body":"<p>You need to install tensorflow from <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">SOURCE<\/a> to configure the settings. Installing by pip or easy_install will not work. <\/p>\n\n<p>I have the same problem after upgrading tensorflow from 0.6.0 to 0.7.1 using pip.<\/p>\n"},{"AnswerId":"35760743","CreationDate":"2016-03-03T00:10:33.067","ParentId":null,"OwnerUserId":"6010587","Title":null,"Body":"<p>I had the same issue. Have you tried installing CUDA 7.5 alongside 7? That worked for me, and I changed my PATH variables to include 7.5 instead of 7.0<\/p>\n"}]}
{"QuestionId":35736229,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-02T00:23:11.873","AcceptedAnswerId":35737543.0,"OwnerUserId":2138200.0,"Title":"How to use tf.train.batch with enqueue_many=true","Body":"<p>I'm looking for an example of using  with .<\/p>\n\n<p>In my case, I have an image tensor of shape [299,299,3] and when I call a function  it will return a new tensor of shape [10,299,299,3] (in this example, it will apply 10 distortions to the image and return them all as a new tensor). I'd then like to enqueue all these by calling .<\/p>\n\n<p>I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>But when I  I get back a list of length 10 (I was expecting a batch of size 5). <\/p>\n\n<p>Also, how would I include the label to  in this case? The label is the same for all 10 distortions.<\/p>\n","answers":[{"AnswerId":"35737543","CreationDate":"2016-03-02T02:47:25.313","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Don't unpack <code>distortions<\/code>.  The semantics of <code>enqueue_many<\/code> is that you feed it a tensor with first dimension being the batching dimension, so a <code>[10, 299, 299, 3]<\/code> tensor with enqueue_many will result in ten separate items, each of shape <code>299, 299, 3<\/code> being enqueued -- which is what you want.<\/p>\n"},{"AnswerId":"44868654","CreationDate":"2017-07-02T08:02:35.100","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/batch\" rel=\"nofollow noreferrer\"><code>tf.train.batch<\/code><\/a> tells you:<\/p>\n\n<blockquote>\n  <p>If enqueue_many is True, tensors is assumed to represent a batch of\n  examples, where the first dimension is indexed by example, and all\n  members of tensors should have the same size in the first dimension.\n  If an input tensor has shape [*, x, y, z], the output will have shape\n  [batch_size, x, y, z]. The capacity argument controls the how long the\n  prefetching is allowed to grow the queues.<\/p>\n<\/blockquote>\n\n<p>Which is exactly what happens in your case: <code>[10, 299, 299, 3]<\/code>, where 10 is the batch size. So you do not need to do any unpacking and <code>tf.train.batch(distortions, 5, enqueue_many=True)<\/code> will do the job.<\/p>\n"}]}
{"QuestionId":35736722,"AnswerCount":1,"Tags":"<python><numpy><matrix><tensorflow><exponential>","CreationDate":"2016-03-02T01:18:01.110","AcceptedAnswerId":35736934.0,"OwnerUserId":678572.0,"Title":"How to compute matrix operation with TensorFlow?","Body":"<p>I have a pandas dataframe containing floats from 0 to 1.<br>\n<strong>I want to exponentiate this matrix to a certain power (e.g. 6).<\/strong><\/p>\n\n<p>I started using  but the operation was taking really, really long for my 7000x7000 matrix so I thought this would be an excellent opportunity to test out <\/p>\n\n<p>My apologies if the notation is about trippy, I thought I was inputting everything correctly.  I want o use a  and .\nMy function  inputs a pandas dataframe object and then exponentiates the matrix to the power of a certain integer.<\/p>\n\n<p><strong>How do I use the placeholder with the feed_dict?<\/strong><\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>Dataframe looks like:<\/p>\n\n<pre><\/pre>\n\n<p>Matrix exponential function I tried:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35736934","CreationDate":"2016-03-02T01:45:07.110","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>EDIT:<\/strong> The question has been updated to remove the error message. You are very close to being able to feed the matrix into your program. The following version of your <code>exp_corr()<\/code> function should do the trick:<\/p>\n\n<pre><code>def exp_corr(DF_var,exp=6):\n    T_feed = tf.placeholder(tf.float32, DF_var.shape)\n    T_exp = tf.pow(T_feed, exp)\n\n    sess = tf.Session()\n\n    # Use the `feed_dict` argument to specify feeds.\n    DF_exp = pd.DataFrame(sess.run(T_exp, feed_dict={T_feed: DF_var.as_matrix()}))\n    DF_exp.columns = DF_var.column; DF_exp.index = DF_var.index\n\n    sess.close() \n\n    return DF_exp\n<\/code><\/pre>\n\n<hr>\n\n<p>The original issue with your program was in the error message:<\/p>\n\n<pre><code>Node 'Input Dataframe': Node name contains invalid characters\n<\/code><\/pre>\n\n<p>In particular, the <code>name<\/code> argument to TensorFlow op constructors (like <code>tf.constant()<\/code> and <code>tf.pow()<\/code>) must be a string that <strong>does not contain spaces<\/strong>.<\/p>\n\n<p>The syntax for node names is defined <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b8ff42a4b3c0bbc4eadaa5ee5cd1b10adb58de7e\/tensorflow\/core\/framework\/graph.proto#L61\" rel=\"nofollow\">here<\/a>. Node names must match the following regular expression (essentially alpha-numeric, plus <code>.<\/code>, <code>_<\/code>, and <code>\/<\/code>, but not starting with <code>_<\/code> or <code>\/<\/code>):<\/p>\n\n<pre><code>[A-Za-z0-9.][A-Za-z0-9_.\/]*\n<\/code><\/pre>\n"}]}
{"QuestionId":35737488,"AnswerCount":1,"Tags":"<python><memory-management><gpu><caffe>","CreationDate":"2016-03-02T02:42:34.720","AcceptedAnswerId":null,"OwnerUserId":1844346.0,"Title":"Memory not getting reclaimed on restart of a process","Body":"<p>I have a python job that runs a caffe net for image processing on nvidia GPUs. The job takes images from a rabbitmq queue, processes it and then writes the result in another queue. When I restart this job, the processes are getting killed but memory is not getting reclaimed.<\/p>\n\n<p>So after certain number of restarts the machine crashes. Once I kill the job there is no python process running in ps or top but the CPU memory is not getting reclaimed. <\/p>\n\n<p>How can I debug this issue?<\/p>\n\n<p>EDIT: CPU memory<\/p>\n","answers":[{"AnswerId":"35748621","CreationDate":"2016-03-02T13:30:20.057","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>It's your GPU memory which is not getting freed. Get the process id<\/p>\n\n<pre><code>$ nvidia-smi\n<\/code><\/pre>\n\n<p>and then<\/p>\n\n<pre><code>$ kill -9 &lt;process id&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":35738217,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2016-03-02T03:55:46.180","AcceptedAnswerId":null,"OwnerUserId":6005719.0,"Title":"TypeError: slice indices must be integers","Body":"<p>It gives me the following error when test_model is called:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm calling test_model with an integer (a particular batch). Inputtest is a list of list of floats and labeltes is a vector of ints. I'm not sure what the problem is.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35739635","CreationDate":"2016-03-02T06:02:02.887","ParentId":null,"OwnerUserId":"3875230","Title":null,"Body":"<p>The <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">Theano documentation<\/a> mentions that doing this<\/p>\n\n<pre><code>index = T.lscalar('index')  \n<\/code><\/pre>\n\n<p>will return the index to a [mini]batch <\/p>\n\n<p>I think the replacement:<\/p>\n\n<pre><code>x --&gt; inputstest[index * batch_size:(index + 1) * batch_size]\ny --&gt; labeltest[index * batch_size:(index + 1) * batch_size]\n<\/code><\/pre>\n\n<p>in your graph is not working because you may need to use batch start and batch stop and you're not passing your learning rate to your function.  <\/p>\n"},{"AnswerId":"35738504","CreationDate":"2016-03-02T04:27:10.820","ParentId":null,"OwnerUserId":"5927858","Title":null,"Body":"<p>Sometimes you will get a value looks like integer. I make two experiment to help you understand.<\/p>\n\n<p>1.<\/p>\n\n<pre><code>list1 = [2,3]\n\nindex = '1'\n\nprint(list1[index])\n<\/code><\/pre>\n\n<p>output: TypeError: list indices must be integers or slices, not str<\/p>\n\n<p>2.<\/p>\n\n<pre><code>list1 = [2,3]\n\nindex = '1'\n\nprint(list1[int(index)])\n<\/code><\/pre>\n\n<p>output: 3<\/p>\n\n<p>So you have to make sure the index you input to list or other data structure should be integer. But it will not always be integer. If you use dictionary, you should input string. It depends on the data structure you use. <\/p>\n\n<p>Hope it helps. :)<\/p>\n"}]}
{"QuestionId":35738297,"AnswerCount":1,"Tags":"<python><ubuntu><hdf5><caffe><pycaffe>","CreationDate":"2016-03-02T04:04:01.847","AcceptedAnswerId":35788692.0,"OwnerUserId":1056563.0,"Title":"Unable to find hdf5 libraries when building pycaffe","Body":"<p>When running <\/p>\n\n<pre><\/pre>\n\n<p>We are ending up with missing  libraries.<\/p>\n\n<pre><\/pre>\n\n<p>Note that every conceivable combination of these libraries has been installed on this  system.<\/p>\n\n<pre><\/pre>\n\n<p>Now I realize that is overkill .. but they were added one at a time - and then we retried the  before adding another..  We just wanted to be sure <em>missing hdf5 libs<\/em> were not the core issue.<\/p>\n\n<p><strong>Update<\/strong> In response to the supplied answer from @jon<\/p>\n\n<pre><\/pre>\n\n<p>And in the :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35788692","CreationDate":"2016-03-04T05:31:39.733","ParentId":null,"OwnerUserId":"2297751","Title":null,"Body":"<p>run <code>dpkg -L libhdf5-dev | grep so<\/code> should output:<\/p>\n\n<pre><code>\/usr\/lib\/x86_64-linux-gnu\/libhdf5_hl.so\n\/usr\/lib\/x86_64-linux-gnu\/libhdf5hl_fortran.so\n\/usr\/lib\/x86_64-linux-gnu\/libhdf5_fortran.so\n\/usr\/lib\/x86_64-linux-gnu\/libhdf5_hl_cpp.so\n\/usr\/lib\/x86_64-linux-gnu\/libhdf5_cpp.so\n\/usr\/lib\/x86_64-linux-gnu\/libhdf5.so\n<\/code><\/pre>\n\n<p>Make sure the path <code>\/usr\/lib\/x86_64-linux-gnu<\/code> (which contains <code>libhdf5.so<\/code> and <code>libhdf5_hl.so<\/code>) is in <code>Makefile.config<\/code>'s <code>LIBRARY_DIRS<\/code><\/p>\n"}]}
{"QuestionId":35738854,"AnswerCount":1,"Tags":"<python><performance><numpy><matrix><tensorflow>","CreationDate":"2016-03-02T04:59:09.170","AcceptedAnswerId":43433271.0,"OwnerUserId":678572.0,"Title":"Time comparison for TensorFlow operation and Numpy multiplication","Body":"<p>I've been trying to optimize my computations; and for most operations that I've tried,  is much faster.  I'm trying to do a fairly simple operation...Transform a matrix (multiply each value by 1\/2 and then add 1\/2 to that value).  <\/p>\n\n<p>With the help of @mrry , I was able to do these operations in .  However to my surprise, the  method was significantly faster?!<\/p>\n\n<p> seems like an extremely useful tool for data scientists and I think this could help clarify it's use and advantages. <\/p>\n\n<p><strong>Am I not using  data structures and operations in the most efficient way?<\/strong>  I'm not sure how non-tensorflow methods would be faster.  I'm using a Mid-2012 Macbook Air 4GB RAM<\/p>\n\n<p><em>trans1 is the tensorflow version while trans2 is numpy. DF_var is a pandas dataframe object<\/em><\/p>\n\n<pre><\/pre>\n\n<p>My timing method was:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43433271","CreationDate":"2017-04-16T02:59:17.300","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Your results are compatible with the <a href=\"https:\/\/simplyml.com\/linear-algebra-shootout-numpy-vs-theano-vs-tensorflow-2\/\" rel=\"nofollow noreferrer\">benchmarks from another guy<\/a>.<\/p>\n\n<p>In his benchmark he compared NumPy, Theano and Tensorflow on <\/p>\n\n<blockquote>\n  <p>an Intel core i5-4460 CPU with 16GiB RAM and a Nvidia GTX 970 with 4\n  GiB RAM using Theano 0.8.2, Tensorflow 0.11.0, CUDA 8.0 on Linux Mint\n  18<\/p>\n<\/blockquote>\n\n<p>His results for addition shows that:\n<a href=\"https:\/\/i.stack.imgur.com\/6YKGg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6YKGg.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>He also tested a few other functions such as matrix multiplication:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mvT9T.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mvT9T.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The results are:<\/p>\n\n<blockquote>\n  <p>It is clear that the main strengths of Theano and TensorFlow are very\n  fast dot products and matrix exponents. The dot product is\n  approximately 8 and 7 times faster respectively with Theano\/Tensorflow\n  compared to NumPy for the largest matrices. Strangely, matrix addition\n  is slow with the GPU libraries and NumPy is the fastest in these\n  tests.<\/p>\n  \n  <p>The minimum and mean of matrices are slow in Theano and quick in\n  Tensorflow. It is not clear why Theano is as slow (worse than NumPy)\n  for these operations.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":35739367,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-02T05:42:46.533","AcceptedAnswerId":null,"OwnerUserId":748240.0,"Title":"Tensorflow tf.cond evaluating both pedicate","Body":"<pre><\/pre>\n\n<p>This code snippet should calculate the mean of user_input across training stage and output mean during testing stage. <\/p>\n\n<p>This is the output result :<\/p>\n\n<pre><\/pre>\n\n<p>However,  always get evaluated when calling  even if . <\/p>\n\n<p>Is there any mistake in the code ?  tensorflow version is 0.7.1 <\/p>\n","answers":[{"AnswerId":"44355372","CreationDate":"2017-06-04T14:52:30.150","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>I think that is the same as <a href=\"https:\/\/stackoverflow.com\/a\/37064128\/133374\">answered here<\/a>. The <code>tf.control_dependencies<\/code> inside the conditionals will add the dependencies to the <code>tf.cond<\/code> itself.<\/p>\n\n<p>So try to create the <code>ema_apply_op<\/code> inside the <code>mean_var_with_update<\/code> function.<\/p>\n"},{"AnswerId":"35754389","CreationDate":"2016-03-02T17:40:46.663","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I've added some logging statements and <code>ema_mean<\/code> only seems to get evaluated when <code>isTrain<\/code> is false<\/p>\n\n<pre><code>tf.reset_default_graph()\n\nisTrain = tf.placeholder(tf.bool)\nuser_input = tf.placeholder(tf.float32)\n\n# ema = tf.train.ExponentialMovingAverage(decay=.5)\n\nwith tf.device('\/cpu:0'):\n    beta = tf.Variable(tf.ones([1]))\n\n    batch_mean = beta.assign(user_input)\n    ema = tf.train.ExponentialMovingAverage(decay=0.5)\n    ema_apply_op = ema.apply([batch_mean])\n    ema_mean = ema.average(batch_mean)\n\n    def mean_var_with_update():\n        with tf.control_dependencies([ema_apply_op]):\n            return tf.Print(tf.identity(batch_mean), [\"mean_var_with_update\"])\n            #return tf.identity(batch_mean)\n\n    mean = tf.Print(tf.cond(isTrain,\n        mean_var_with_update,\n        lambda: (tf.Print(ema_mean, [\"ema_mean\"]))),\n                    [\"evaluating mean\", isTrain])\n\n# ======= End Here ==========\nsaver = tf.train.Saver()\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\nu_input = [[2], [3], [4] ]\nfor u in u_input:\n    aa = sess.run([mean], feed_dict={user_input:u, isTrain: True })\n    print(\"Train\", aa)\n\nfor u in u_input:\n    aa = sess.run([ema_mean], feed_dict={user_input:u, isTrain: False })\n    print(\"Test correct\", aa)\n\nfor u in u_input:\n    aa = sess.run([mean], feed_dict={user_input:u, isTrain: False })\n    print(\"Test\", aa)\n<\/code><\/pre>\n\n<p>You see <\/p>\n\n<pre><code>[mean_var_with_update]\n[evaluating mean][True]\n[mean_var_with_update]\n[evaluating mean][True]\n[mean_var_with_update]\n[evaluating mean][True]\n[ema_mean]\n[evaluating mean][False]\n[ema_mean]\n[evaluating mean][False]\n[ema_mean]\n[evaluating mean][False]\n<\/code><\/pre>\n\n<p>Note that the <code>Print<\/code> statement is evaluated after all the inputs have been evaluated so outer print statement is printed last<\/p>\n"}]}
{"QuestionId":35740594,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><tensorflow><conv-neural-network>","CreationDate":"2016-03-02T07:06:12.223","AcceptedAnswerId":35743023.0,"OwnerUserId":3301357.0,"Title":"Layer names for pretrained inception v3 model (tensorflow)","Body":"<p>The task is to get per-layer output of a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow noreferrer\">pretrained cnn inceptionv3<\/a> model.\nFor example I feed an image to this network, and I want to get not only its output, but output of each layer(layer-wise).<\/p>\n\n<p>In order to do that, I have to know names of each layer output. It's quite easy to do for last and pre-last layer:<\/p>\n\n<pre><\/pre>\n\n<p>(names pool_3:0 and softmax:0 are listed in sources related to the model)<\/p>\n\n<p>But what if I want to get output of not only last and pre-last layer, but also pre-pre-last, pre-pre-pre-last, etc, output of any layer of this model?<\/p>\n\n<p>Then I have to know layer names. A similar question <a href=\"https:\/\/stackoverflow.com\/questions\/34175174\/extract-features-using-pre-trained-tensorflow-cnn\">has been asked<\/a>\n on stackoverflow, but there are only names for pre-last and last layers, which are also available in comments in the sources('pool_3:0', 'softmax:0')<\/p>\n\n<p>The obvious way to get names is to draw a graph of the model. But it's not the case <a href=\"https:\/\/stackoverflow.com\/questions\/35635131\/unable-to-visualize-inception-v3-model-in-tensorboard-with-tensorflow-0-7-1\">because graph-drawing tool simply crashes<\/a>.\nAlso there is an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1287\" rel=\"nofollow noreferrer\">issue<\/a> related to this.<\/p>\n\n<p>I also tried:<\/p>\n\n<ol>\n<li>To print all graph node names and construct graph using information only from node names. But node names are quite cryptic. Also connections between nodes are not so obvious, because of <a href=\"http:\/\/arxiv.org\/pdf\/1512.00567v3.pdf\" rel=\"nofollow noreferrer\">inception v3 architecture<\/a> So without knowing the connections between nodes, this approach is like to break the code of enigma :)<\/li>\n<li>Then I tried to draw a graph of this model using own-written tool(get all inputs of given node(say, 'pool_3:0'), then get all inputs of each of inputs of 'pool_3:0', etc. This approach works for simple models. But when I tried to use it for inceptionv3 implementation from THE image recognition tutorial, I've got 16 GB of my RAM used, then 10 GB of swap, then I simply stopped looking under the hood. Maybe there are some cycles in graph(but how could it be if the network is feed-forward?)<\/li>\n<\/ol>\n\n<p>So, maybe somebody who's participated in this pre-trained model implementation &amp; 'shipping' is reading stackoverflow.<\/p>\n\n<p>Or maybe somebody has already got layer-wise names for this model.\nPlease share it with me and ml community. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"35743023","CreationDate":"2016-03-02T09:22:55.727","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>The base question has already been asked (and answered) in <a href=\"https:\/\/stackoverflow.com\/questions\/35336648\/list-of-tensor-names-in-graph-in-tensorflow\/\">List of tensor names in graph in Tensorflow<\/a><\/p>\n\n<p>To answer your concern with your approach (2) - that does work (i've done it). If I had to guess, you probably don't use a dictionary-like structure to store the nodes, so if multiple nodes get their input from the same node (like it happens a lot in the inception model), you insert that input node multiple times in your list of things to store \/ visit. If you take care to visit each node only once, that should run basically instantly without memory problems.<\/p>\n"}]}
{"QuestionId":35741311,"AnswerCount":2,"Tags":"<python><windows><install><tensorflow>","CreationDate":"2016-03-02T07:48:41.983","AcceptedAnswerId":36788158.0,"OwnerUserId":4805357.0,"Title":"Issues with installing TensorFlow\/Python onto my Windows computer","Body":"<p>Currently I am trying to follow the <a href=\"https:\/\/www.udacity.com\/course\/deep-learning--ud730\" rel=\"nofollow noreferrer\">Google Deep Learning Course<\/a>, and am facing the issue of installing and running <strong>TensorFlow<\/strong> on my Windows PC. <\/p>\n\n<p><a href=\"http:\/\/www.netinstructions.com\/how-to-install-and-run-tensorflow-on-a-windows-pc\/\" rel=\"nofollow noreferrer\">These<\/a> are the instructions of how to install <strong>TensorFlow<\/strong> onto my computer. However these fail at this point: <\/p>\n\n<pre><\/pre>\n\n<p>with the error shown below: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yMdUd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yMdUd.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now after looking at <a href=\"https:\/\/caffinc.github.io\/2015\/11\/tensorflow-windows\/\" rel=\"nofollow noreferrer\">this<\/a> article, and trying out the line , I seem to get the same error: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AIgJn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AIgJn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Has anyone succesfully installed  and  onto their computers? Thank you for your answers. Please do not hesitate to ask me about anything that would clarify this question. <\/p>\n","answers":[{"AnswerId":"36788158","CreationDate":"2016-04-22T07:59:21.800","ParentId":null,"OwnerUserId":"3274693","Title":null,"Body":"<p>It seems that you don't have a running VM. The easiest way is to launch \"Docker Quick Start Terminal\". It will (create and) run a VM.<\/p>\n\n<p>Since you want to do Udacity assignments, you can try the following command in the opened Docker Terminal:<\/p>\n\n<pre><code>docker run -it -p 8888:8888 -v b.gcr.io\/tensorflow-udacity\/assignments:0.5.0\n<\/code><\/pre>\n\n<p>I personally use this command:<\/p>\n\n<pre><code>docker run -d -p 8888:8888 -v \/$(pwd)\/tensorflow:\/notebooks --memory=8g --name tf b.gcr.io\/tensorflow-udacity\/assignments:0.5.0\n<\/code><\/pre>\n\n<p>where:<\/p>\n\n<ul>\n<li><code>--memory=8g<\/code> is to increase memory<\/li>\n<li><code>--name tf<\/code> is to give the container a name<\/li>\n<li><code>-v \/$(pwd)\/tensorflow:\/notebooks<\/code> mounts a folder from host to the container. This folder MUST be inside your windows user folder. As you can see I created a folder named <code>tensorflow<\/code> inside my windows user folder and put all the assignment notebooks in there.<\/li>\n<\/ul>\n\n<p>To access the Jupyter notebook on your browser go to \"the Virtual Machine IP\":8888 which is usually 192.168.99.100:8888.<\/p>\n\n<p>If you named the container as I did, you can use <code>docker start tf<\/code> to start it next time.<\/p>\n"},{"AnswerId":"35753193","CreationDate":"2016-03-02T16:43:04.037","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>To build and run a CPU only tensorflow on a windows system that has docker installed you can <\/p>\n\n<pre><code>cd tensorflow\/tools\/docker\ndocker build -t tensorflow \ndocker run -it -p 8888:8888 tensorflow \n<\/code><\/pre>\n\n<p>That will give you a bash console with a built tensorflow <\/p>\n"}]}
{"QuestionId":35742706,"AnswerCount":1,"Tags":"<machine-learning><neural-network><image-recognition><conv-neural-network><keras>","CreationDate":"2016-03-02T09:08:30.200","AcceptedAnswerId":null,"OwnerUserId":5876383.0,"Title":"keras(cnn+nn) predicting only one class among 4 classes","Body":"<p>I am using two different datasets having 1200 images each. First dataset has 4 classes and second dataset has 6 classes. <\/p>\n\n<p>This is simple image classification problem. But while training, on each epoch I am getting same value for validation accuracy for both the datasets. <\/p>\n\n<p>I have resized all the images of both datasets to 100x100 using imagemagick.<\/p>\n\n<p>I don't know where I am making mistake.\nThanks in advance<\/p>\n\n<p>terminal output :<\/p>\n\n<pre><\/pre>\n\n<p>code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"56182550","CreationDate":"2019-05-17T08:29:46.123","ParentId":null,"OwnerUserId":"8893595","Title":null,"Body":"<p>The problem is with the optimizer. It is seen that you are using SDG as an optimizer, which usually performs poorly on CNNs. Please use adam\/nadam\/tanh activations.<\/p>\n"}]}
{"QuestionId":35744473,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2016-03-02T10:26:07.307","AcceptedAnswerId":null,"OwnerUserId":4571518.0,"Title":"ImportError: No module named vis (Python )?","Body":"<p>Hello l tried to run my python code and the errors is: <\/p>\n\n<pre><\/pre>\n\n<p>knowing that l have correctly installed foxhound.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38345709","CreationDate":"2016-07-13T07:54:05.370","ParentId":null,"OwnerUserId":"6583234","Title":null,"Body":"<p>I have met with the same problem. from foxhound vis import grayscale_grid_vis is enough. And I just comment out the other part. And the program works well. It seems that the module is not used.<\/p>\n"},{"AnswerId":"35745539","CreationDate":"2016-03-02T11:11:10.130","ParentId":null,"OwnerUserId":"4887136","Title":null,"Body":"<p>is Foxhound a scikit? if yes, it doe's not contains utils module. You can clone it on github <a href=\"https:\/\/github.com\/IndicoDataSolutions\/Foxhound\" rel=\"nofollow\">https:\/\/github.com\/IndicoDataSolutions\/Foxhound<\/a><\/p>\n"}]}
{"QuestionId":35747276,"AnswerCount":1,"Tags":"<deep-learning><convolution><conv-neural-network><keras>","CreationDate":"2016-03-02T12:28:16.300","AcceptedAnswerId":null,"OwnerUserId":6007683.0,"Title":"Implementing joint learning in keras","Body":"<p>i am trying to implement a model that is composed of two layers to segment object candidates in keras\nSo basically this model has the following architecture<\/p>\n\n<p>Image(channel,width,height) -> multiple convolution and pooling layers- > output('n' feature maps , height width ) <\/p>\n\n<p>Now this <strong><em>single output<\/em><\/strong> is being used by two layers\nwhich are as follows\n1) convolution (1*1) - > dense layer with m units (output = n * 1*1 )  - > pixel classifier using fully connected layers of h*w dimesion -> upsmapling to (H,N) - > output<\/p>\n\n<p>2) convolution -> maxpooling->dense layer - > score<\/p>\n\n<p>Cost function uses outputs of both these layers which is sum of binary logistic regression of each output<\/p>\n\n<p>Now I have two questions\n1) how to implement dense connection over  convoluted output in layer 1 to produce h*w pixel classifier as mentioned above\n2) How to merge  the two layers to calculate the single cost function and then train both the layers jointly using back-propagation <\/p>\n\n<p>Can anyone tell me how to create the model for above mentioned network architecture.i am new to deep learning so if there something which i misunderstood i ll appreciate if anyone can explain me the errors in my understanding\nThanks<\/p>\n","answers":[{"AnswerId":"35757129","CreationDate":"2016-03-02T20:09:46.603","ParentId":null,"OwnerUserId":"5256563","Title":null,"Body":"<p>It's easier when you share the code you already have.<\/p>\n\n<p>For the transition convolution to dense, you have to use <code>model.add(Flatten())<\/code>, like in the <a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow\">examples here<\/a>.<\/p>\n\n<p>Unfortunately, I don't know for the second question, but according to what I just read in the Keras Models, you have to use the graph model.<\/p>\n"}]}
{"QuestionId":35747388,"AnswerCount":1,"Tags":"<python-2.7><theano>","CreationDate":"2016-03-02T12:33:39.257","AcceptedAnswerId":35752669.0,"OwnerUserId":1334336.0,"Title":"Using a neural network for regression (with Theano)","Body":"<p>I'm new to theano, and I'm having troubles.\nI'm trying to use theano to create a neural network that can be used for a regression task (instead of a classification task)\nAfter reading a lot of Tutorials, I came to the conclusion that I could do that by creating an output layer which just handles the regression, and prepand a \"normal\" neural net with a few hidden layers. (But that still lies in the future). <\/p>\n\n<p>So this is my \"model\":<\/p>\n\n<pre><\/pre>\n\n<p>to train the model as in the theano tutorials I tried the following:<\/p>\n\n<pre><\/pre>\n\n<p>I feel like I did exactly the same (only with the math I need) like it was done in theanos logistic regression tutorial (<a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a>) but it doesn't work. So why cant I create the gradients?<\/p>\n","answers":[{"AnswerId":"35752669","CreationDate":"2016-03-02T16:18:57.387","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>Your cost function should probably be a sum of squares. At the moment it is a vector of squares, but you need to condense it down to one value in order to be able to the the gradient of the then scalar function. This is usually done like this:<\/p>\n\n<pre><code>def cost_function(self, y):\n    return ((y - self.regression) ** 2).mean()\n<\/code><\/pre>\n"}]}
{"QuestionId":35747651,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-03-02T12:46:47.633","AcceptedAnswerId":null,"OwnerUserId":6007803.0,"Title":"the type of variable returned by ''theano.tensor.cast''","Body":"<p>What is the exact type of a variable returned by  when the input is a shared variable?<\/p>\n\n<p>For example, execute the following code:<\/p>\n\n<pre><\/pre>\n\n<p>So.... what is the type of variable ?<\/p>\n","answers":[]}
{"QuestionId":35749431,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><caffe>","CreationDate":"2016-03-02T14:05:05.857","AcceptedAnswerId":null,"OwnerUserId":3743384.0,"Title":"In Caffe, how to finetune on two or more models?","Body":"<p>Now I have trained GoogleNet and VGG16 respectively for the same task, i.e. googlenet.caffemodel and vgg16.caffemodel. Now I want to merge the two networks to form an ensemble model. That is, I want to define the two network in a single prototxt. This means I need to fintune on two models. How can I do that?<\/p>\n","answers":[]}
{"QuestionId":35752358,"AnswerCount":1,"Tags":"<python><apache-spark><pyspark><rdd><pycaffe>","CreationDate":"2016-03-02T16:06:45.853","AcceptedAnswerId":null,"OwnerUserId":2085454.0,"Title":"How to classify images using Spark and Caffe","Body":"<p>I am using Caffe to do image classification, can I am using MAC OS X, Pyhton.<\/p>\n\n<p>Right now I know how to classify a list of images using Caffe with Spark python, but if I want to make it faster, I want to use Spark.<\/p>\n\n<p>Therefore, I tried to apply the image classification on each element of an RDD, the RDD created from a list of image_path. However, Spark does not allow me to do so.<\/p>\n\n<p>Here is my code:<\/p>\n\n<p>This is the code for image classification:<\/p>\n\n<pre><\/pre>\n\n<p>This this the code generates Caffe parameters and apply the classify_image method on each element of the RDD:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, here I tried to broadcast the caffe parameters, , \nThe error is:<\/p>\n\n<blockquote>\n  <p>RuntimeError: Pickling of \"caffe._caffe.Net\" instances is not enabled<\/p>\n<\/blockquote>\n\n<p>Before I am doing the broadcast, the error was :<\/p>\n\n<blockquote>\n  <p>Driver stacktrace.... Caused by: org.apache.spark.api.python.PythonException: Traceback (most recent call last):....<\/p>\n<\/blockquote>\n\n<p>So, do you know, is there any way I can classify images using Caffe and Spark but also take advantage of Spark?<\/p>\n","answers":[{"AnswerId":"35756115","CreationDate":"2016-03-02T19:13:40.380","ParentId":null,"OwnerUserId":"1560062","Title":null,"Body":"<p>When you work with complex, non-native objects initialization has to moved directly to the workers for example with singleton module:<\/p>\n\n<p><code>net_builder.py<\/code>:<\/p>\n\n<pre><code>import cafe \n\nnet = None\n\ndef build_net(*args, **kwargs):\n     ...  # Initialize net here\n     return net       \n\ndef get_net(*args, **kwargs):\n    global net\n    if net is None:\n        net = build_net(*args, **kwargs)\n    return net\n<\/code><\/pre>\n\n<p><code>main.py<\/code>:<\/p>\n\n<pre><code>import net_builder\n\nsc.addPyFile(\"net_builder.py\")\n\ndef classify_image(image_path, transformer, *args, **kwargs):\n    net = net_builder.get_net(*args, **kwargs)\n<\/code><\/pre>\n\n<p>It means you'll have to distribute all required files as well. It can be done either manually or <a href=\"https:\/\/stackoverflow.com\/a\/31764684\/1560062\">using <code>SparkFiles<\/code><\/a> mechanism. <\/p>\n\n<p>On a side note you should take a look at the <a href=\"https:\/\/github.com\/amplab\/SparkNet\" rel=\"nofollow noreferrer\">SparkNet<\/a> package.<\/p>\n"}]}
{"QuestionId":35753414,"AnswerCount":1,"Tags":"<python><deep-learning><keras>","CreationDate":"2016-03-02T16:53:09.833","AcceptedAnswerId":null,"OwnerUserId":5654554.0,"Title":"Don't know whats going wrong while training dataset. Keras Model Fit","Body":"<p>I have just started using Keras and was trying to train a model using Keras deep learning kit. Works till the epochs are runned but crashes just after it.<\/p>\n\n<pre><\/pre>\n\n<p>Error Recieved:\nI have pasted it over here. Please have a look at it.\n<a href=\"http:\/\/pastebin.com\/dSw9ckkk\" rel=\"nofollow\">http:\/\/pastebin.com\/dSw9ckkk<\/a><\/p>\n","answers":[{"AnswerId":"36970497","CreationDate":"2016-05-01T18:41:31.323","ParentId":null,"OwnerUserId":"1397061","Title":null,"Body":"<p>It looks like you have two classes, a positive class and a negative class, so that the positive class labels are 1 minus the negative class labels.  In that case, you can discard the negative class labels and make it a single-class problem:<\/p>\n\n<pre><code>model.add(Dense(1), activation='sigmoid')  # instead of Dense(nb_classes) and Activation('softmax')\n<\/code><\/pre>\n\n<p>Alternatively, you can still train the model on both classes and just use the positive class in the AUC calculation:<\/p>\n\n<pre><code>roc_auc_score(y_test[:, 1],y_pre[:, 1])\n<\/code><\/pre>\n"}]}
{"QuestionId":35754790,"AnswerCount":3,"Tags":"<python><classification><caffe>","CreationDate":"2016-03-02T18:02:31.677","AcceptedAnswerId":null,"OwnerUserId":5956107.0,"Title":"Caffe Classifier","Body":"<p>I looked around but I don't think the already raised questions can treat my problem. <\/p>\n\n<p>I am having a problem with a function in caffe : Classifier. The code is bugging in this line :<\/p>\n\n<pre><\/pre>\n\n<p>and I get the following (extract from the last lines) :<\/p>\n\n<pre><\/pre>\n\n<p>Anyone can help ?<\/p>\n\n<p>Thanks <\/p>\n","answers":[{"AnswerId":"35755099","CreationDate":"2016-03-02T18:18:45.210","ParentId":null,"OwnerUserId":"2289713","Title":null,"Body":"<blockquote>\n  <p>Process finished with exit code 139<\/p>\n<\/blockquote>\n\n<p>It means that process received signal 11 which means <code>SIGSEGV<\/code>\nI had some experience with Caffe. It has no any normal error reporting and segfaults on any error.\nYou can use some debugger like <code>gdb<\/code> to find out where error occurs<\/p>\n"},{"AnswerId":"43286728","CreationDate":"2017-04-07T20:36:24.747","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>The problem lies in the model specification prototype file. You are trying to import a caffe model that used deprecated prototype specifications for training. You can modify the net.cpp file in src folder to remove these warnings. It's just telling you that the upgrading from the old caffe model to the current caffe model you are using, failed. You should be good to go with small modification in the net.cpp file.<\/p>\n"},{"AnswerId":"40237537","CreationDate":"2016-10-25T10:27:59.517","ParentId":null,"OwnerUserId":"4726173","Title":null,"Body":"<p>This is not a real solution, it just worked in my case. And it only works in my development env. If anyone has some ideas why this works, please don't hesitate to comment\/answer.<\/p>\n\n<p>It could be a memory problem. (<a href=\"https:\/\/stackoverflow.com\/questions\/23704040\/exit-code-139-when-perfroming-image-subtraction\/34146379#34146379\">This<\/a> got me on that track.) Caffe is very memory intensive, naturally. But it could be something else as well, e.g. in my case a problem with the data layer I'm using.<\/p>\n\n<p>The \"solution\": <\/p>\n\n<ol>\n<li>run in pycharm (or other dev env?), set a breakpoint at the net creation line (it does not work with pdb breakpoints). It also seemed to help to switch to Console view before the call to caffe.Net().<\/li>\n<li>get rid of as many layers as possible (for me, removing one unnecessary data layer did the trick - BUT I believe I had bigger nets running already).<\/li>\n<\/ol>\n\n<p>In my case, it only failed sometimes at first. When running it in pycharm and stopping at and after net creation, it worked most of the time. (sleep()'s didn't seem to help.) Then I changed the net, and it always segfaulted (139, as vitalii said). <\/p>\n\n<p>Maybe using valgrind to track down the problem would be best (if you know how).<\/p>\n"}]}
{"QuestionId":35755973,"AnswerCount":1,"Tags":"<python><theano><tensorflow>","CreationDate":"2016-03-02T19:05:56.360","AcceptedAnswerId":35756827.0,"OwnerUserId":5874853.0,"Title":"TensorFlow error: TensorShape() must have the same rank","Body":"<pre><\/pre>\n\n<p>I am running an activation function which computes  and passes it into a sigmoid activation.\nWhen I try to execute the above function, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The theano equivalent for computing  is working just fine:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35756827","CreationDate":"2016-03-02T19:53:45.757","ParentId":null,"OwnerUserId":"5952845","Title":null,"Body":"<p>Mihir,<\/p>\n\n<p><strong>When I encountered this problem it was because my placeholders were the wrong size in my feed dictionary.<\/strong> Also you should know how to run the graph in a session. <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\" rel=\"nofollow noreferrer\">tf.Session.run(fetches, feed_dict=None)<\/a><\/p>\n\n<p>Here's my code to make the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#placeholder\" rel=\"nofollow noreferrer\">placeholders<\/a><\/p>\n\n<pre><code># Note this place holder is for the input data feed-dict definition\ninput_placeholder = tf.placeholder(tf.float32, shape=(batch_size, FLAGS.InputLayer))\n# Not sure yet what this will be used for. \ndesired_output_placeholder = tf.placeholder(tf.float32, shape=(batch_size, FLAGS.OutputLayer))\n<\/code><\/pre>\n\n<p>Here's my fill feed dictionary function:<\/p>\n\n<pre><code>def feel_feed_funct(data_sets_train, input_pl, output_pl):\n  ti_feed, dto_feed = data_sets_train.next_batch(FLAGS.batch_size)\n\n  feed_dict = {\n    input_pl: ti_feed,\n    output_pl: dto_feed\n  }\n  return feed_dict\n<\/code><\/pre>\n\n<p>Later I do this:<\/p>\n\n<pre><code># Fill a feed dictionary with the actual set of images and labels\n# for this particular training step.\nfeed_dict = fill_feed_dict(data_sets.train, input_placeholder, desired_output_placeholder)\n<\/code><\/pre>\n\n<p>Then to run the session and fetch the outputs I have this line<\/p>\n\n<pre><code>_, l = sess.run([train_op, loss], feed_dict=feed_dict)\n<\/code><\/pre>\n"}]}
{"QuestionId":35756236,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-03-02T19:20:40.877","AcceptedAnswerId":null,"OwnerUserId":5977269.0,"Title":"Tensorflow: Tried SVHN by editing label 0-9, still not working","Body":"<p>I am trying to classify SVHN data set by following this tutorial: \n    <a href=\"https:\/\/stackoverflow.com\/questions\/34759227\/tensorflow-cifar10-example\/34763642\">https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/deep_cnn\/index.html<\/a><\/p>\n\n<p>I am using train_32x32.mat file. In order to use it with CNN code (mentioned above), I converted this .mat file to several .bin files using this simple code:<\/p>\n\n<pre><\/pre>\n\n<p>But when I tried to classify SVHN using these customized .bin files I'm getting stuck with error \"Invalid argument: Indices are not valid (out of bounds)\" listed below:<\/p>\n\n<pre><\/pre>\n\n<p>I found <a href=\"https:\/\/stackoverflow.com\/questions\/34759227\/tensorflow-cifar10-example\/34763642\">TensorFlow CIFAR10 Example<\/a> , similar issue in stackoverflow. But even if I change the label, it's still not working.<\/p>\n\n<p>Please let me know if I did something wrong or not understanding any logic.<\/p>\n\n<p>Thanks<\/p>\n\n<p>Sarah<\/p>\n","answers":[{"AnswerId":"35781919","CreationDate":"2016-03-03T20:03:45.233","ParentId":null,"OwnerUserId":"5977269","Title":null,"Body":"<p>Something was wrong with my installed version of Tensorflow (might be a bug). Upgrading to new version solved the issue.<\/p>\n\n<p>Thanks<\/p>\n"}]}
{"QuestionId":35756710,"AnswerCount":3,"Tags":"<python><tensorflow><confusion-matrix>","CreationDate":"2016-03-02T19:47:03.457","AcceptedAnswerId":35876136.0,"OwnerUserId":2235817.0,"Title":"How do i create Confusion matrix of predicted and ground truth labels with Tensorflow?","Body":"<p>I have implemented a Nueral Network model for a classification  with the help of using TensorFlow. But, i don't know how can i able to draw confusion matrix by using predicted scores (accuracy). I am not an expert of TensorFlow and still in learning phase. Here i pasted my code below please tell me how can i write a code for making confusion from the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Up till now, i am able to print predicted scores but failed to implement confusion matrix please help. \n<strong>Note:<em><\/strong>(I am using one hot vectors for representing my labels)<\/em><\/p>\n","answers":[{"AnswerId":"49516367","CreationDate":"2018-03-27T15:01:52.793","ParentId":null,"OwnerUserId":"730650","Title":null,"Body":"<p>For the moment, I use this solution to obtain the confusion matrix :<\/p>\n\n<pre><code># load the data\n(train_x, train_y), (dev_x, dev_y), (test_x, test_y) = dataLoader.load()\n\n# build the classifier\nclassifier = tf.estimator.DNNClassifier(...)\n\n# train the classifier\nclassifier.train(input_fn=lambda:train_input_fn(), steps=1000)\n\n# evaluate and prediction on the test set\ntest_evaluate = classifier.evaluate(input_fn=lambda:eval_input_fn())\ntest_predict = classifier.predict(input_fn = lambda:eval_input_fn())\n\n# parse the prediction to retrieve the predicted labels\npredictions = []\n\nfor i in list(test_predict):\n    predictions.append(i['class_ids'][0])\n\n# build the prediction matrix\nmatrix = tf.confusion_matrix(test_y, predictions)\n\n#display the prediction matrix\nwith tf.Session():\n    print(str(tf.Tensor.eval(matrix)))\n<\/code><\/pre>\n\n<p>But I am not convince by my loop to retrieve the predicted labels...there should be a better Python way to do this...(or a TensorFlow way...)<\/p>\n"},{"AnswerId":"38349755","CreationDate":"2016-07-13T10:59:16.653","ParentId":null,"OwnerUserId":"6256833","Title":null,"Body":"<p>If you want to produce a confusion matrix, and then later precision and recall, you first need to get your counts of true positives, true negatives, false positives and false negatives. Here is how:<\/p>\n\n<p>For better readibility, I wrote the code very verbose.<\/p>\n\n<pre><code>def evaluation(logits,labels):\n\"Returns correct predictions, and 4 values needed for precision, recall and F1 score\"\n\n\n    # Step 1:\n    # Let's create 2 vectors that will contain boolean values, and will describe our labels\n\n    is_label_one = tf.cast(labels, dtype=tf.bool)\n    is_label_zero = tf.logical_not(is_label_one)\n    # Imagine that labels = [0,1]\n    # Then\n    # is_label_one = [False,True]\n    # is_label_zero = [True,False]\n\n    # Step 2:\n    # get the prediction and false prediction vectors. correct_prediction is something that you choose within your model.\n    correct_prediction = tf.nn.in_top_k(logits, labels, 1, name=\"correct_answers\")\n    false_prediction = tf.logical_not(correct_prediction)\n\n    # Step 3:\n    # get the 4 metrics by comparing boolean vectors\n    # TRUE POSITIVES\n    true_positives = tf.reduce_sum(tf.to_int32(tf.logical_and(correct_prediction,is_label_one)))\n\n    # FALSE POSITIVES\n    false_positives = tf.reduce_sum(tf.to_int32(tf.logical_and(false_prediction, is_label_zero)))\n\n    # TRUE NEGATIVES\n    true_negatives = tf.reduce_sum(tf.to_int32(tf.logical_and(correct_prediction, is_label_zero)))\n\n    # FALSE NEGATIVES\n    false_negatives = tf.reduce_sum(tf.to_int32(tf.logical_and(false_prediction, is_label_one)))\n\n\nreturn true_positives, false_positives, true_negatives, false_negatives\n\n# Now you can do something like this in your session:\n\ntrue_positives, \\\nfalse_positives, \\\ntrue_negatives, \\\nfalse_negatives = sess.run(evaluation(logits,labels), feed_dict=feed_dict)\n\n# you can print the confusion matrix using the 4 values from above, or get precision and recall:\nprecision = float(true_positives) \/ float(true_positives+false_positives)\nrecall = float(true_positives) \/ float(true_positives+false_negatives)\n\n# or F1 score:\nF1_score = 2 * ( precision * recall ) \/ ( precision+recall )\n<\/code><\/pre>\n"},{"AnswerId":"35876136","CreationDate":"2016-03-08T19:19:15.920","ParentId":null,"OwnerUserId":"2235817","Title":null,"Body":"<p>This code worked for me. I sort it out myself :)<\/p>\n\n<pre><code>from sklearn.metrics import precision_recall_fscore_support as score\nfrom sklearn.metrics import classification_report\n\ndef print_confusion_matrix(plabels,tlabels):\n\"\"\"\n    functions print the confusion matrix for the different classes\n    to find the error...\n\n    Input:\n    -----------\n    plabels: predicted labels for the classes...\n    tlabels: true labels for the classes\n\n    code from: http:\/\/stackoverflow.com\/questions\/2148543\/how-to-write-a-confusion-matrix-in-python\n\"\"\"\nimport pandas as pd\nplabels = pd.Series(plabels)\ntlabels = pd.Series(tlabels)\n\n# draw a cross tabulation...\ndf_confusion = pd.crosstab(tlabels,plabels, rownames=['Actual'], colnames=['Predicted'], margins=True)\n\n#print df_confusion\nreturn df_confusion\n\ndef confusionMatrix(text,Labels,y_pred, not_partial):\n    y_actu = np.where(Labels[:]==1)[1]\n    df = print_confusion_matrix(y_pred,y_actu)\n    print \"\\n\",df\n    #print plt.imshow(df.as_matrix())\n    if not_partial:\n       print \"\\n\",classification_report(y_actu, y_pred)\n    print \"\\n\\t------------------------------------------------------\\n\"\n\ndef do_eval(message, sess, correct_prediction, accuracy, pred, X_, y_,x,y):\n    predictions = sess.run([correct_prediction], feed_dict={x: X_, y: y_})\n    prediction  = tf.argmax(pred,1)\n    labels = prediction.eval(feed_dict={x: X_, y: y_}, session=sess)\n    print message, accuracy.eval({x: X_, y: y_}),\"\\n\"\n    confusionMatrix(\"Partial Confusion matrix\",y_,predictions[0], False)#Partial confusion Matrix\n    confusionMatrix(\"Complete Confusion matrix\",y_,labels, True) #complete confusion Matrix\n\n# Launch the graph\nwith tf.Session() as sess:\nsess.run(init)\ndata = zip(X_train,y_train)\ndata = np.array(data)\ndata_size = len(data)\nnum_batches_per_epoch = int(len(data)\/batch_size) + 1\nfor epoch in range(training_epochs):\n    avg_cost = 0.\n    # Shuffle the data at each epoch\n    shuffle_indices = np.random.permutation(np.arange(data_size))\n    shuffled_data = data[shuffle_indices]\n    for batch_num in range(num_batches_per_epoch):\n        start_index = batch_num * batch_size\n        end_index = min((batch_num + 1) * batch_size, data_size)\n        sample = zip(*shuffled_data[start_index:end_index])\n        #picking up random batches from training set of specific size\n        batch_xs, batch_ys = sample[0],sample[1]\n        # Fit training using batch data\n        sess.run(optimizer, feed_dict={x: batch_xs, y: batch_ys})\n        # Compute average loss\n        avg_cost += sess.run(cost, feed_dict={x: batch_xs, y: batch_ys})\/num_batches_per_epoch\n    #append loss\n    loss_history.append(avg_cost)\n\n    # Display logs per epoch step\n    if (epoch % display_step == 0):           \n        correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))            \n        # Calculate training  accuracy\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n        trainAccuracy = accuracy.eval({x: X_train, y: y_train})\n        train_acc_history.append(trainAccuracy)           \n        # Calculate validation  accuracy\n        valAccuracy = accuracy.eval({x: X_val, y: y_val})\n        val_acc_history.append(valAccuracy) \n        print \"Epoch:\", '%04d' % (epoch+1), \"cost=\", \"{:.9f}\".format(avg_cost), \"train=\",trainAccuracy,\"val=\", valAccuracy\n\nprint \"Optimization Finished!\\n\"\n\n# Evaluation of  model\ncorrect_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1)) \n# Calculate accuracy\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n\ndo_eval(\"Accuracy of Gold Test set Results: \", sess, correct_prediction, accuracy, pred, X_gold, y_gold, x, y)\n<\/code><\/pre>\n\n<p>and here is the sample output:<\/p>\n\n<pre><code>Accuracy of Gold Test set Results:  0.642608 \n\n\nPredicted  False  True  All\nActual                     \n0             20    46   66\n1              3     1    4\n2             21     1   22\n3              8     4   12\n4             16     7   23\n5             54   259  313\n6             41    14   55\n7             11     2   13\n8             48    94  142\n9             29     4   33\n10            17     4   21\n11            39   116  155\nAll          307   552  859\n\nPredicted   0  1  2   3   4    5   6   7    8   9  10   11  All\nActual                                                         \n0          46  0  0   0   0    8   0   2    2   2   0    6   66\n1           0  1  0   1   0    2   0   0    0   0   0    0    4\n2           3  0  1   3   0   12   0   0    1   0   0    2   22\n3           2  0  0   4   1    3   1   1    0   0   0    0   12\n4           1  0  0   0   7   12   0   0    1   0   0    2   23\n5           8  0  0   1   5  259   9   0    9   3   1   18  313\n6           1  0  0   1   6   30  14   1    2   0   0    0   55\n7           3  0  0   0   0    2   0   2    4   0   1    1   13\n8           6  0  0   1   1   18   0   3   94   8   1   10  142\n9           9  0  0   0   0    1   1   1    9   4   0    8   33\n10          1  0  0   0   3    6   0   1    1   0   4    5   21\n11          5  1  0   1   0   18   1   0    6   5   2  116  155\nAll        85  2  1  12  23  371  26  11  129  22   9  168  859\n\n         precision    recall  f1-score   support\n\n      0       0.54      0.70      0.61        66\n      1       0.50      0.25      0.33         4\n      2       1.00      0.05      0.09        22\n      3       0.33      0.33      0.33        12\n      4       0.30      0.30      0.30        23\n      5       0.70      0.83      0.76       313\n      6       0.54      0.25      0.35        55\n      7       0.18      0.15      0.17        13\n      8       0.73      0.66      0.69       142\n      9       0.18      0.12      0.15        33\n     10       0.44      0.19      0.27        21\n     11       0.69      0.75      0.72       155\n\n     avg \/ total       0.64      0.64      0.62       859\n<\/code><\/pre>\n"}]}
{"QuestionId":35757151,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2016-03-02T20:10:45.847","AcceptedAnswerId":35799201.0,"OwnerUserId":1464282.0,"Title":"Why does this Keras model require over 6GB of memory?","Body":"<p>This Keras model seems to require 6GB+ of RAM using the Tensorflow backend. My back-of-the-envelope math suggests that storing the weights shouldn't require more than 500MB. What's going on?<\/p>\n\n<pre><\/pre>\n\n<p>It's a convolution layer (16 3x3 filters) connected to a single neuron, and then that single neuron is connected to ~50k neurons. <\/p>\n\n<p>I'm pretty new to Keras, so I imagine my misunderstanding is pretty fundamental, but I can't seem to figure it out.<\/p>\n","answers":[{"AnswerId":"35799201","CreationDate":"2016-03-04T15:02:51.313","ParentId":null,"OwnerUserId":"1464282","Title":null,"Body":"<p>Turns out, my issue was including a path to CUDA 7.5 in my <code>LD_CONFIG_PATH<\/code>, but including a path to CUDA 7.0 in <code>PATH<\/code>. Apparently this awkward combination spawns some undefined behavior, which in my case produced a memory leak.<\/p>\n\n<p>After examining the code with a <code>valgrind<\/code>, I found that the <code>nvcc<\/code> from 7.0 was essentially jumping into nonsense areas of the CUDA (7.5) library, which is not unexpected. It's actually pretty amazing it leaked memory instead of just crashing, and that Theano had the same error.<\/p>\n\n<p>Hopefully no one else will be plagued by this particular issue in the future, but if you are, double check your version paths!<\/p>\n\n<p>On my local machine, without a GPU'd Tensorflow installed, I still got the memory leak, which appeared to a bug in the previous (0.7.0) version that has been resolved with the (0.7.1) release. Again, I haven't figured out why my non-GPU Theano backend also produced the leak, <em>but after upgrading Tensorflow, the Theano backend doesn't leak either<\/em>. It's a very strange thing, but I believe the general solution to this problem is \"upgrade\" and \"double-check your env\".<\/p>\n"}]}
{"QuestionId":35757439,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-03-02T20:28:50.093","AcceptedAnswerId":null,"OwnerUserId":5590385.0,"Title":"would it be straight forward to implement a spatial transformer network in tensorflow?","Body":"<p>i am interested in trying things out with a spatial transformer network and I can't find any implementation of it in caffe or tensorflow, which are the only two libraries I'm interested in using.  I have a pretty good grasp of tensorflow but was wondering if it would be straight forward to implement with the existing building blocks that tensorflow offers without having to do something too complicated like write a custom c++ module <\/p>\n","answers":[{"AnswerId":"45226970","CreationDate":"2017-07-21T00:03:27.097","ParentId":null,"OwnerUserId":"433310","Title":null,"Body":"<p>Tensorflow has a implementation of Spatial Transformer Network in the models repository - <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/transformer\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/transformer<\/a><\/p>\n"},{"AnswerId":"35837716","CreationDate":"2016-03-07T06:45:40.540","ParentId":null,"OwnerUserId":"2495287","Title":null,"Body":"<p>There is an implementation in caffe here. <a href=\"https:\/\/github.com\/daerduoCarey\/SpatialTransformerLayer\" rel=\"nofollow\">https:\/\/github.com\/daerduoCarey\/SpatialTransformerLayer<\/a><\/p>\n"},{"AnswerId":"36244259","CreationDate":"2016-03-27T05:46:08.810","ParentId":null,"OwnerUserId":"4490576","Title":null,"Body":"<p>Yes, it is very straight forward to setup the Tensorflow graph for a spatial transformer network with the existing API.<\/p>\n\n<p>You can find an example implementation in Tensorflow here [1].<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/daviddao\/spatial-transformer-tensorflow\" rel=\"noreferrer\">https:\/\/github.com\/daviddao\/spatial-transformer-tensorflow<\/a><\/p>\n"}]}
{"QuestionId":35757460,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2016-03-02T20:30:04.943","AcceptedAnswerId":37364615.0,"OwnerUserId":1637894.0,"Title":"Bayes factors in pymc3","Body":"<p>I'm interested in computing Bayes factors to compare two models in PyMC 3. According to <a href=\"http:\/\/stronginference.com\/bayes-factors-pymc.html\" rel=\"nofollow\">this website<\/a>, in PyMC 2, the procedure seems relatively straightforward: include a Bernoulli random variable and a custom likelihood function, which returns the likelihood of the first model when the value of the Bernoulli variable is 0, and the likelihood of the second model when the value is 1. In PyMC 3, however, things get more complicated, because the stochastic nodes need to be Theano variables.<\/p>\n\n<p>My two likelihood functions are Binomials, so I guess I need to re-write this class:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions on where to start?<\/p>\n","answers":[{"AnswerId":"37364615","CreationDate":"2016-05-21T15:10:29.417","ParentId":null,"OwnerUserId":"815542","Title":null,"Body":"<p>You could try something like this:<\/p>\n\n<pre><code>with pm.Model() as model:\n    p = np.array([0.5, 0.5])\n    model_index = pm.Categorical('model_index', p=p)\n    model0 # define one model here\n    model1 # define the other model here\n    m = pm.switch(pm.math.eq(model_index, 0), model0, model1)\n    # pass M to a prior or likelihood, for example\n    y = pm.SomeDistribution('y', m, observed=data)\n\n    step0 = pm.ElemwiseCategorical(vars=[model_index], values=[0,1])\n    step1 = pm.NUTS()\n    trace = pm.sample(5000, step=[step0, step1], start=start)\n<\/code><\/pre>\n\n<p>Then you compute the Bayes factors as: (add burnin if necessary)<\/p>\n\n<pre><code>pm.traceplot(trace);\npM1 = trace['model_index'].mean()\npM0 = 1 - pM1\npM0, pM1, (pM0\/pM1)*(p[1]\/p[0])\n<\/code><\/pre>\n\n<p>You may also want to check how yo use Information Criteria to compare models, see an example <a href=\"https:\/\/github.com\/pymc-devs\/pymc3\/blob\/6bbb7911787fa664373e88ae9bd0def2e5b4f7ed\/pymc3\/examples\/notebooks\/Model%20Comparison.ipynb\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":35757586,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-03-02T20:37:53.083","AcceptedAnswerId":null,"OwnerUserId":3528623.0,"Title":"Tensorflow: All variables become nan upon first training run","Body":"<p>I initialized all my variables, but upon the first training run, all variables become nan. I've checked that all variables were initialized before starting training.<\/p>\n\n<h1>EDIT (Solved)<\/h1>\n\n<p>The image array fed into the net was filled with NaN's causing the problem.<\/p>\n","answers":[]}
{"QuestionId":35759220,"AnswerCount":1,"Tags":"<filter><visualization><tensorflow>","CreationDate":"2016-03-02T22:13:24.453","AcceptedAnswerId":35858950.0,"OwnerUserId":5577321.0,"Title":"How to visualize learned filters on tensorflow","Body":"<p>Similarly to the Caffe framework, where it is possible to watch the learned filters during CNNs training and it's resulting convolution with input images, I wonder if is it possible to do the same with TensorFlow?<\/p>\n\n<p>A Caffe example can be viewed in this link: <\/p>\n\n<blockquote>\n  <p><a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"noreferrer\">http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb<\/a><\/p>\n<\/blockquote>\n\n<p>Grateful for your help!<\/p>\n","answers":[{"AnswerId":"35858950","CreationDate":"2016-03-08T04:27:07.050","ParentId":null,"OwnerUserId":"1760830","Title":null,"Body":"<p>To see just a few conv1 filters in Tensorboard, you can use this code (it works for cifar10)<\/p>\n\n<pre><code># this should be a part of the inference(images) function in cifar10.py file\n\n# conv1\nwith tf.variable_scope('conv1') as scope:\n  kernel = _variable_with_weight_decay('weights', shape=[5, 5, 3, 64],\n                                       stddev=1e-4, wd=0.0)\n  conv = tf.nn.conv2d(images, kernel, [1, 1, 1, 1], padding='SAME')\n  biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.0))\n  bias = tf.nn.bias_add(conv, biases)\n  conv1 = tf.nn.relu(bias, name=scope.name)\n  _activation_summary(conv1)\n\n  with tf.variable_scope('visualization'):\n    # scale weights to [0 1], type is still float\n    x_min = tf.reduce_min(kernel)\n    x_max = tf.reduce_max(kernel)\n    kernel_0_to_1 = (kernel - x_min) \/ (x_max - x_min)\n\n    # to tf.image_summary format [batch_size, height, width, channels]\n    kernel_transposed = tf.transpose (kernel_0_to_1, [3, 0, 1, 2])\n\n    # this will display random 3 filters from the 64 in conv1\n    tf.image_summary('conv1\/filters', kernel_transposed, max_images=3)\n<\/code><\/pre>\n\n<p>I also wrote a simple <a href=\"https:\/\/gist.github.com\/kukuruza\/03731dc494603ceab0c5\" rel=\"noreferrer\">gist<\/a> to display all 64 conv1 filters in a grid.<\/p>\n"}]}
{"QuestionId":35759950,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-02T23:02:21.757","AcceptedAnswerId":35761203.0,"OwnerUserId":2138200.0,"Title":"Incompatible GraphDef versions in Extend","Body":"<p>I have some code which creates a graph to process some images and then iterates  in a loop to fetch batches of image tensors of shape [*, 299, 299, 3]. I'd like to then feed these images into the inception model.<\/p>\n\n<p>So, I added some code to load the inception model:<\/p>\n\n<pre><\/pre>\n\n<p>Now Im getting this error when running :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35761203","CreationDate":"2016-03-03T01:00:55.253","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>This is most likely from using too old a version of TensorFlow to read in and run the graph -- the graph was created using a newer version of the GraphDef.  Try upgrading to 0.7 or to HEAD and then run your code again.<\/p>\n"}]}
{"QuestionId":35760996,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-03T00:36:29.293","AcceptedAnswerId":null,"OwnerUserId":5948619.0,"Title":"Does histogram_summary respect name_scope","Body":"<p>I am getting a Duplicate tag error when I try to write out histogram summaries for a multi-layer network that I generate procedurally. I think that the problem might be related to naming. Imagine code like the following:<\/p>\n\n<pre><\/pre>\n\n<p>I'd naively assumed that 'weights' would be scoped to some_unique_name but I'm suspecting that it is not. Are summary names independent of name_scope?<\/p>\n","answers":[{"AnswerId":"35761189","CreationDate":"2016-03-03T00:59:32.060","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>They are.<\/p>\n\n<p>I'm with you in thinking this is a bug, but I haven't run it past the designers of the op yet.  Go ahead and open an issue for it on GitHub!<\/p>\n\n<p>(I've run into this also and found it terribly annoying -- it prevents reuse of the model without deliberately parameterizing the summary op invocations.)<\/p>\n"},{"AnswerId":"35761337","CreationDate":"2016-03-03T01:16:35.147","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As <a href=\"https:\/\/stackoverflow.com\/a\/35761189\/3574081\">Dave points out,<\/a> the <code>tag<\/code> argument to <code>tf.histogram_summary(tag, ...)<\/code> is indeed independent of the current name scope. Part of the reason for this is that the <code>tag<\/code> may be a string <code>Tensor<\/code> (i.e. computed by part of your graph), whereas name scopes are a purely client-side construct (i.e. Python-only), so there's no good way to make the scoping work consistently across the two modes of use.<\/p>\n\n<p>However, if you're using TensorFlow build from source (and should be available in the next release, 0.8.0), you can use the following recipe to scope your tags (using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph.unique_name\" rel=\"nofollow noreferrer\"><code>Graph.unique_name(..., mark_as_used=False)<\/code><\/a>):<\/p>\n\n<pre><code>with tf.name_scope(some_unique_name):\n  # ...\n  tf.histogram_summary(\n      tf.get_default_graph().unique_name('weights', mark_as_used=False),\n      kernel_weights)\n<\/code><\/pre>\n\n<p>Alternatively, you can do the following in the current version:<\/p>\n\n<pre><code>with tf.name_scope(some_unique_name) as scope:\n  # ...\n  tf.histogram_summary(scope + 'weights', kernel_weights)\n<\/code><\/pre>\n"}]}
{"QuestionId":35761030,"AnswerCount":1,"Tags":"<macos><opencl><amd><torch>","CreationDate":"2016-03-03T00:40:22.200","AcceptedAnswerId":37628967.0,"OwnerUserId":3355076.0,"Title":"Set OpenCL device on OSX to dedicated AMD GPU","Body":"<p>I'm in the process of setting up Torch on OSX on a 2015 macbook pro with a Radeon GPU using <a href=\"https:\/\/github.com\/hughperkins\/cltorch\/\" rel=\"nofollow\">this library (cltorch)<\/a> for OpenCL support<\/p>\n\n<p>I can successfully run Torch scripts now, but running <a href=\"https:\/\/github.com\/hughperkins\/cltorch\/#custom-user-kernels\" rel=\"nofollow\">this test script<\/a> which outputs the device and platform being used I get:<\/p>\n\n<pre><\/pre>\n\n<p>Obviously, I want torch to run on Radeon instead of the integrated Iris, but I have no idea how to do that.<\/p>\n","answers":[{"AnswerId":"37628967","CreationDate":"2016-06-04T10:11:08.717","ParentId":null,"OwnerUserId":"212731","Title":null,"Body":"<p>You can use <code>cltorch.setDevice<\/code> to choose the device, like:<\/p>\n\n<pre><code>cltorch.setDevice(2)\n<\/code><\/pre>\n"}]}
{"QuestionId":35761661,"AnswerCount":1,"Tags":"<lua><deep-learning><torch>","CreationDate":"2016-03-03T01:50:36.523","AcceptedAnswerId":35762524.0,"OwnerUserId":1094908.0,"Title":"Where is the main code of convolutional nets?","Body":"<p>Although I want to modify code of convolutional nets, I could not find the main routine of the convolutional nets. The main routine is convolution and pooling.<\/p>\n\n<p> had the code below.<\/p>\n\n<pre><\/pre>\n\n<p>So I thought the routine was in .\nHowever,  had the code below.<\/p>\n\n<pre><\/pre>\n\n<p>So does anyone know where is ?<\/p>\n","answers":[{"AnswerId":"35762524","CreationDate":"2016-03-03T03:15:05.633","ParentId":null,"OwnerUserId":"1190388","Title":null,"Body":"<p>There's an open issue <a href=\"https:\/\/github.com\/chuanli11\/CNNMRF\/issues\/9\" rel=\"nofollow\">on GitHub<\/a> for the same. One of <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/torch7\/CnkUoI6Dh1E\/wZrXJaXTAwAJ\" rel=\"nofollow\">the solution<\/a> mentions upgrading <code>nn<\/code> and <code>cunn<\/code>:<\/p>\n\n<pre><code>luarocks install nn\nluarocks install cunn\n<\/code><\/pre>\n\n<p>Also, take a look at <a href=\"https:\/\/github.com\/chuanli11\/CNNMRF\/issues\/9#issuecomment-188605005\" rel=\"nofollow\">this reply<\/a>.<\/p>\n"}]}
{"QuestionId":35761791,"AnswerCount":0,"Tags":"<theano><keras><lstm><language-model>","CreationDate":"2016-03-03T02:02:29.330","AcceptedAnswerId":null,"OwnerUserId":5995117.0,"Title":"Testing accuracy always more than 99%","Body":"<p>I am trying to implement a language model using LSTMs in theano\/keras. My network runs fine and I also see that the training loss decreases but the testing accuracy is always above 99% even if I don not train my network for long. \nI have used word2vec vectors and embedded the weights in the Embedding layer. My network looks like:<\/p>\n\n<pre><\/pre>\n\n<p>My training and testing array shapes are:\nX_train shape: (100000, 18)\nX_test shape: (10000, 18)\ny_train shape: (100000, 998)\ny_test shape: (10000, 998)<\/p>\n\n<p>where there are 100000 training and 10000 testing sentences and each sentence contains 18 words. The number of classes for output is 998. <\/p>\n\n<p>Can anyone suggest why I might not be getting a genuine testing error?<\/p>\n","answers":[]}
{"QuestionId":35762491,"AnswerCount":1,"Tags":"<slice><torch>","CreationDate":"2016-03-03T03:12:38.430","AcceptedAnswerId":null,"OwnerUserId":6010932.0,"Title":"How to slice a Torch tensor by a set of indices?","Body":"<p>I am wondering if there is an efficient way to slice a tensor by a set of indices? In Matlab, I can do the following:<\/p>\n\n<pre><\/pre>\n\n<p>Then B is the 1st, 3rd, and 5th slice of A.<\/p>\n\n<p>In Torch, it seems you can only slice with a continuous range. Is that true?<\/p>\n\n<p>A more general question will be if I can get a subset by arbitrary indices such as <\/p>\n\n<pre><\/pre>\n\n<p>in Matlab.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"35767283","CreationDate":"2016-03-03T08:52:08.817","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Use the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-indexdim-index\" rel=\"nofollow\">index<\/a> operator, e.g.:<\/p>\n\n<pre><code>t = torch.rand(2, 3, 6)\n-- (1,.,.) = \n--   0.1790  0.7774  0.5343  0.0628  0.3077  0.7203\n--   0.0677  0.5847  0.2401  0.6885  0.8724  0.4413\n--   0.1849  0.2704  0.2745  0.5508  0.4634  0.6340\n-- \n-- (2,.,.) = \n--   0.2523  0.6135  0.6037  0.0194  0.6456  0.0229\n--   0.9966  0.8688  0.2078  0.7169  0.1528  0.5708\n--   0.8671  0.7731  0.4596  0.0636  0.8873  0.2205\n-- [torch.DoubleTensor of size 2x3x6]\n\n\nt:index(3, torch.LongTensor{1, 3, 5})\n-- (1,.,.) = \n--   0.1790  0.5343  0.3077\n--   0.0677  0.2401  0.8724\n--   0.1849  0.2745  0.4634\n-- \n-- (2,.,.) = \n--   0.2523  0.6037  0.6456\n--   0.9966  0.2078  0.1528\n--   0.8671  0.4596  0.8873\n-- [torch.DoubleTensor of size 2x3x3]\n<\/code><\/pre>\n\n<p>You can do as well <code>t:index(3, torch.LongTensor{1, 2, 6})<\/code>.<\/p>\n"}]}
{"QuestionId":35763011,"AnswerCount":2,"Tags":"<neural-network><deep-learning><torch>","CreationDate":"2016-03-03T04:08:21.220","AcceptedAnswerId":null,"OwnerUserId":3853033.0,"Title":"Using Torch's ClassNLLCriterion","Body":"<p>I am currently using Torch and just trying to get a simple neural network program running. Each one of my inputs has 3 attributes and the output is supposed to be a classification between the numbers 1 and 7. I've extracted my data from a CSV file and have put it into 2 Tensors (1 with the inputs and 1 with the outputs). The data is in this format. <\/p>\n\n<pre><\/pre>\n\n<p>For the model I'm using to train the network, I simply have <\/p>\n\n<pre><\/pre>\n\n<p>And this is the code I have to train the network<\/p>\n\n<pre><\/pre>\n\n<p>In my test data tensor, I have formatted it in the same way as I formatted the test data (Tensor of 99999x3). I want the program to give me a prediction of what the classification would be when I run this line. <\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting negative numbers (which shouldn't happen with ClassNLLCriterion?) and the sums of the probabilities are not adding to 0. My doubt is that I have either not formatted the data correctly or that I wasn't able to perform the training process correctly. If anyone could help me figure out what the issue is, I'd be very grateful.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"40149660","CreationDate":"2016-10-20T08:49:13.260","ParentId":null,"OwnerUserId":"7046851","Title":null,"Body":"<p>Here is what I currently use which maybe the wrong way of using classnllcriterion but atleast it will get you somewhere to understanding it.<\/p>\n\n<p>Make the targets to be either <\/p>\n\n<pre><code>(7,1,1,1,1,1,1) &lt;--First class representation\n.......\n(1,1,1,1,1,1,7) &lt;--Last class representation\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code> (1,1,1,1,1,1,1) &lt;--First class representation\n .......\n (7,7,7,7,7,7,7) &lt;--Last class representation\n<\/code><\/pre>\n\n<p>I figured it's much easier to train the last representation as targets, but I have a feeling we should use the first one instead.<\/p>\n\n<p><strong>EDIT: I've just found out that classnllcriterion only accepts a scalars as targets, hence using the above is wrong!<\/strong><\/p>\n\n<p><strong>You should instead use either<\/strong> \n<code>1 .. 7<\/code>  as target values, either just 1 or just 7.<\/p>\n\n<p>That's <\/p>\n"},{"AnswerId":"40174308","CreationDate":"2016-10-21T10:30:51.843","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<p>The reason why you cannot see the prediction yields on the layer <code>model:add(nn.LogSoftMax())<\/code> which implements the <code>log<\/code> function, that's why you have negative values (they are not probabilities). As an example, to get the probabilities back you should do:<\/p>\n\n<pre><code>model = nn.Sequential()\nmodel:add(nn.Linear(3, 7));\nmodel:add(nn.LogSoftMax());\ncriterion = nn.ClassNLLCriterion();\ndata = torch.Tensor{1914,  1993 , 2386}\nprint (model:forward(data):exp())\n\n&gt;&gt;  0.0000\n  0.0000\n  1.0000\n  0.0000\n  0.0000\n  0.0000\n  0.0000 [torch.DoubleTensor of size 7]\n<\/code><\/pre>\n\n<p>Sorry for the late answer.<\/p>\n"}]}
{"QuestionId":35768436,"AnswerCount":1,"Tags":"<amazon-web-services><theano><deep-learning><keras>","CreationDate":"2016-03-03T09:43:17.593","AcceptedAnswerId":null,"OwnerUserId":4497662.0,"Title":"Memory error using Theano on AWS GPU","Body":"<p>I am building a simple a FeedForward Network for a classification task.\nI am using keras on top of theano on AWS g2.2xlarge instance. My network config is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>When I compile the model, I get MemoryError<\/p>\n\n<pre><\/pre>\n\n<p>From AWS this is a 4GB video memory, 16 GB ram machine.\n<strong>If I half the value of 'dimof_middle' to 8196. everything runs fine.<\/strong><\/p>\n\n<p>1073741824 bytes ~ 1 GB which is well within GPU memory<\/p>\n\n<p>What wrong am I doing ? Any help would be much appreciated! <\/p>\n","answers":[{"AnswerId":"35792662","CreationDate":"2016-03-04T09:42:59.037","ParentId":null,"OwnerUserId":"3086290","Title":null,"Body":"<p>The error message is not about the total amount of allocated memory. It is about the amount that it was trying to allocate when failed. Theano allocates and deallocates memory during computation. <\/p>\n\n<p>You can disable CNMeM and use nvidia-smi to find out real allocation.<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANORC\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html#envvar-THEANORC<\/a><\/p>\n"}]}
{"QuestionId":35769944,"AnswerCount":1,"Tags":"<matrix><indexing><element><variable-assignment><tensorflow>","CreationDate":"2016-03-03T10:44:59.387","AcceptedAnswerId":35780087.0,"OwnerUserId":3860525.0,"Title":"Manipulating matrix elements in tensorflow","Body":"<p>How can I do the following in tensorflow? <\/p>\n\n<pre><\/pre>\n\n<p>I can't use the [] brackets because it only works on constants and not on\nvariables. I cant use the slice function either because that returns a tensor and you can't assign to a tensor.<\/p>\n\n<pre><\/pre>\n\n<h1>Will print<\/h1>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35780087","CreationDate":"2016-03-03T18:27:23.360","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>You can't change a tensor - but, as you noted, you can change a variable.<\/p>\n\n<p>There are three patterns you could use to accomplish what you want:<\/p>\n\n<p>(a)  Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_update\" rel=\"noreferrer\"><code>tf.scatter_update<\/code><\/a> to directly poke to the part of the variable you want to change.<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.Variable(initial_value=[2, 5, -4, 0])\nb = tf.scatter_update(a, [1], [9])\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as s:\n  s.run(init)\n  print s.run(a)\n  print s.run(b)\n  print s.run(a)\n<\/code><\/pre>\n\n<blockquote>\n  <p>[ 2  5 -4  0]<\/p>\n  \n  <p>[ 2  9 -4  0]<\/p>\n  \n  <p>[ 2  9 -4  0]<\/p>\n<\/blockquote>\n\n<p>(b)  Create two <code>tf.slice()<\/code>s of the tensor, excluding the item you want to change, and then <code>tf.concat(0, [a, 0, b])<\/code> them back together.<\/p>\n\n<p>(c)  Create <code>b = tf.zeros_like(a)<\/code>, and then use <code>tf.select()<\/code> to choose which items from <code>a<\/code> you want, and which zeros from <code>b<\/code> that you want.<\/p>\n\n<p>I've included (b) and (c) because they work with normal tensors, not just variables.<\/p>\n"}]}
{"QuestionId":35770360,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-03-03T11:01:06.317","AcceptedAnswerId":null,"OwnerUserId":4755260.0,"Title":"Tensorflow The same training accuracy continues","Body":"<p>I'm stuck on CNN model on Tensorflow.\nMy code as below.<\/p>\n\n<h1>Libraries<\/h1>\n\n<pre><\/pre>\n\n<h1>Model<\/h1>\n\n<pre><\/pre>\n\n<h1>Loss<\/h1>\n\n<pre><\/pre>\n\n<h1>Training<\/h1>\n\n<pre><\/pre>\n\n<h1>Accuracy<\/h1>\n\n<pre><\/pre>\n\n<h1>Main<\/h1>\n\n<pre><\/pre>\n\n<p>However, I got the same training accuracy. How to fix this problem?<\/p>\n\n<pre><\/pre>\n\n<p>I referred the following model and my tensorboard is as below.\n<a href=\"https:\/\/i.stack.imgur.com\/yAggg.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yAggg.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/mzoyh.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mzoyh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kYZu8.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kYZu8.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"36023486","CreationDate":"2016-03-15T22:30:14.137","ParentId":null,"OwnerUserId":"6068783","Title":null,"Body":"<p>Could it be that you are not minimizing the right tensor?\nYou are minimizing cross_entropy, but should be cross_entropy_mean (accuracy in your code).<\/p>\n\n<p>Basically with the following logic:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>cross_entropy = tf.nn.softmax_cross_entropy_with_logits(\n      logits, ground_truth_placeholder)\n\ncross_entropy_mean = tf.reduce_mean(cross_entropy)\n\ntrain_step = tf.train.GradientDescentOptimizer(FLAGS.learning_rate).minimize(\n      cross_entropy_mean)\n<\/code><\/pre>\n"}]}
{"QuestionId":35770617,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2016-03-03T11:13:31.227","AcceptedAnswerId":35816617.0,"OwnerUserId":4928066.0,"Title":"Error trying to run custom caffenet with new data","Body":"<p>I just tried to train the provided Caffenet network with my own lmdb file.\nI changed the fully connected to convolutional layer of depth 4096 and custom frame_size. <\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>When compiling i get this error: <\/p>\n\n<pre><\/pre>\n\n<p>My images are 336*254. <\/p>\n\n<p>What does this error mean ?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"35816617","CreationDate":"2016-03-05T16:26:33.470","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>You forgot to set the resize flag to <code>true<\/code> in your shell script , create_imagenet.sh, for creating the LMDB. Set <code>RESIZE = true<\/code> and set the <code>RESIZE_HEIGHT=221<\/code> and <code>RESIZE_WIDTH=221<\/code>. And make sure that the <code>crop_size<\/code> in the <code>data<\/code> layer of your <code>train_val.prototxt<\/code> is also set to 221.<\/p>\n"}]}
{"QuestionId":35771285,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><jupyter>","CreationDate":"2016-03-03T11:44:03.463","AcceptedAnswerId":null,"OwnerUserId":4611375.0,"Title":"Using TensorFlow through Jupyter (Python 3)","Body":"<p><em>Apologies in advance, I think the issue is quite perplexing!<\/em><\/p>\n\n<p><strong>I would like to use TensorFlow through Jupyter, with a Python3 kernel.<\/strong><\/p>\n\n<p>However the command  returns the error:  when either Python2 or Python3 is specified as the Jupyter kernel.<\/p>\n\n<ul>\n<li>I have Python 2 and Python 3 installed on my Mac and can access both\nversions through Terminal.<\/li>\n<li>I installed <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#download-and-setup\" rel=\"noreferrer\">TensorFlow for Python 3<\/a>, however I can only access it via Python 2 on the Terminal.<\/li>\n<\/ul>\n\n<p>As such, this question is really two-fold:<\/p>\n\n<ol>\n<li>I want to get TensorFlow working with Python3<\/li>\n<li>...which should lead to TensorFlow working with Jupyter on the Python3 terminal.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"37081342","CreationDate":"2016-05-06T20:59:55.837","ParentId":null,"OwnerUserId":"6301373","Title":null,"Body":"<p>I had the same problem and solved it using the tutorial <a href=\"http:\/\/help.pythonanywhere.com\/pages\/IPythonNotebookVirtualenvs\">Using a virtualenv in an IPython notebook<\/a>. I'll walk you through the steps I took. <\/p>\n\n<p>I am using Anaconda, and I installed a new environment <code>tensorflow<\/code> using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#anaconda-installation\">these instructions<\/a> at <code>tensorflow.org<\/code>. After that, here is how I got <code>tensorflow<\/code> to work in a Jupyter notebook:<\/p>\n\n<ol>\n<li>Open Terminal<\/li>\n<li>Run <code>source activate tensorflow<\/code>. You should now see <code>(tensorflow)<\/code> at the beginning of the prompt.<\/li>\n<li><p>Now that we are in the <code>tensorflow<\/code> environment, we want to install <code>ipython<\/code> and <code>jupyter<\/code> in this environment: Run <\/p>\n\n<pre><code>conda install ipython \n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>conda install jupyter\n<\/code><\/pre><\/li>\n<li><p>Now follow the instructions in the tutorial linked above. I'll repeat them here with a bit more information added. First run <\/p>\n\n<pre><code>ipython kernelspec install-self --user \n<\/code><\/pre>\n\n<p>The result for me was <code>Installed kernelspec python3 in \/Users\/charliebrummitt\/Library\/Jupyter\/kernels\/python3<\/code><\/p><\/li>\n<li><p>Run the following: <\/p>\n\n<pre><code>mkdir -p ~\/.ipython\/kernels\n<\/code><\/pre>\n\n<p>Then run the following with <code>&lt;kernel_name&gt;<\/code> replaced by a name of your choice (I chose <code>tfkernel<\/code>) and replace the first path (i.e., <code>~\/.local\/share\/jupyter\/kernels\/pythonX<\/code>) by the path generated in step 4: <\/p>\n\n<pre><code>mv ~\/.local\/share\/jupyter\/kernels\/pythonX ~\/.ipython\/kernels\/&lt;kernel_name&gt;\n<\/code><\/pre><\/li>\n<li><p>Now you'll see a new kernel if you open a Jupyter notebook and select <code>Kernel -&gt; Change kernel<\/code> from the menu. But the new kernel will have the same name as your previous kernel (for me it was called <code>Python 3<\/code>). To give your new kernel a unique name, run in Terminal<\/p>\n\n<pre><code>cd ~\/.ipython\/kernels\/tfkernel\/\n<\/code><\/pre>\n\n<p>and then run <code>vim kernel.json<\/code> to edit the file <code>kernel.json<\/code> so that you replace the value of <code>\"display_name\"<\/code> from the default (<code>Python 3<\/code>) to a new name (I chose to call it <code>\"tfkernel\"<\/code>). Save and exit <code>vim<\/code> by typing <code>:wq<\/code> while in command mode. <\/p><\/li>\n<li>Open a new Jupyter notebook and type <code>import tensorflow as tf<\/code>. If you didn't get <code>ImportError<\/code> then you are ready to go!<\/li>\n<\/ol>\n"}]}
{"QuestionId":35772068,"AnswerCount":0,"Tags":"<lua><neural-network><linear-regression><torch><mean-square-error>","CreationDate":"2016-03-03T12:20:55.147","AcceptedAnswerId":null,"OwnerUserId":5562717.0,"Title":"Doing linear regression in Torch gives NaN as error","Body":"<p>I am new for torch. Recently, I am trying to use torch to do multi-linear regression. But the error is always being infinity and nan. <\/p>\n\n<p>For the first two error, it is obviously increasing.\nHere is my code. <\/p>\n\n<pre><\/pre>\n\n<p>What should I do for this problem? And I use the same way for train logistic regression. The result seems better than this. But still not good enough. \nIs there something wrong? Thanks a lot.<\/p>\n","answers":[]}
{"QuestionId":35772292,"AnswerCount":2,"Tags":"<csv><tensorflow>","CreationDate":"2016-03-03T12:32:20.453","AcceptedAnswerId":null,"OwnerUserId":6012841.0,"Title":"TensorFlow: Reading a CSV file hangs forever","Body":"<p>I've followed the tutorial and the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/reading_data\/index.html#csv-files\" rel=\"nofollow noreferrer\">csv example<\/a> shown there doesn't seem to work. It gets stuck forever...<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'm using Tensorflow 0.7.1 and Python3.<\/p>\n\n<p>What am I doing wrong?<\/p>\n\n<p>My files have only this row:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39276664","CreationDate":"2016-09-01T16:32:57.783","ParentId":null,"OwnerUserId":"945522","Title":null,"Body":"<p>It may save some people grief to note in TensorFlow 0.10 the both the amendment from <a href=\"https:\/\/stackoverflow.com\/users\/6012841\/bruno-oliveira\" title=\"Bruno Oliveira\">Bruno Oliveira<\/a> and formatting is required in Python 3.0, not seen this working in Python 2.x even with correct 2.x print syntax.<\/p>\n\n<pre><code>import tensorflow as tf\n\nfilename_queue = tf.train.string_input_producer([\".\/iris.data\"])\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 = [tf.constant([], dtype=tf.float32),   # Column 0\n               tf.constant([], dtype=tf.float32),   # Column 1\n               tf.constant([], dtype=tf.float32),   # Column 2\n               tf.constant([], dtype=tf.float32),\n              tf.constant([], dtype=tf.string)] \ncol1, col2, col3, col4, col5 = tf.decode_csv(\n    value, record_defaults=record_defaults)\nfeatures = tf.pack([col1, col2, col3, col4])\n\nconfig = tf.ConfigProto(inter_op_parallelism_threads=2)\nwith tf.Session(config=config) 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        print(example, label)\n\n   coord.request_stop()\n   coord.join(threads)\n<\/code><\/pre>\n"},{"AnswerId":"35799498","CreationDate":"2016-03-04T15:15:04.673","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Thanks for your persistence on trying to debug this. It turns out that you were running into a bug that was fixed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/841656c9fd84902733e2ee13ac4dd9020a5b6361\" rel=\"nofollow noreferrer\">a recent commit<\/a>, but the fix hasn't made it into a release yet. There are two possible fixes (other than <a href=\"https:\/\/stackoverflow.com\/questions\/35772292\/cant-read-csv-with-tensorflow?noredirect=1#comment59255028_35772292\">acquiring more processors<\/a>):<\/p>\n\n<ol>\n<li>Upgrade to the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#installation\" rel=\"nofollow noreferrer\">nightly binary release<\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow noreferrer\">install from source<\/a>, to get the fix.<\/li>\n<li><p>In your Python program, add the following to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.__init__\" rel=\"nofollow noreferrer\">session creation<\/a>:<\/p>\n\n<pre><code>config = tf.ConfigProto(inter_op_parallelism_threads=2)\nwith tf.Session(config=config) as sess:\n  # ...\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>The reason for the issue is that TensorFlow uses a bounded threadpool for dispatching ops, and (until the fix) the reader op could block, which would lead to deadlock if another op had to run before the reader could complete (for example because of a producer-consumer relationship). The fix addresses this by running the reader asynchronously.<\/p>\n"}]}
{"QuestionId":35772854,"AnswerCount":1,"Tags":"<machine-learning><torch>","CreationDate":"2016-03-03T12:58:56.003","AcceptedAnswerId":null,"OwnerUserId":4718136.0,"Title":"Torch, testing issue - imagenet-multiGPU.torch","Body":"<p>I am a newbies to torch. I have trainined my model according to 2 classes via this instruction <a href=\"https:\/\/github.com\/soumith\/imagenet-multiGPU.torch\" rel=\"nofollow\">https:\/\/github.com\/soumith\/imagenet-multiGPU.torch<\/a> (number of classes is modifed) then, I wantted to test my model. These lines of codes are writen in instruction for testing: <\/p>\n\n<pre><\/pre>\n\n<p>I got errors at the initial line of codes while I was trying to write.\nWhen I try to write ; <\/p>\n\n<pre><\/pre>\n\n<p>I get these errors;<\/p>\n\n<pre><\/pre>\n\n<p>I don't know what to do. Thanks for helping.<\/p>\n","answers":[{"AnswerId":"35878559","CreationDate":"2016-03-08T21:40:21.813","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>You need to run <code>opts.lua<\/code> before running <code>donkey.lua<\/code> using<\/p>\n\n<pre><code>th&gt; dofile(\"opts.lua\")\n<\/code><\/pre>\n\n<p>Check <code>main.lua<\/code> and make sure, you are not missing any other dependencies.<\/p>\n"}]}
{"QuestionId":35773898,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-03T13:44:44.153","AcceptedAnswerId":35777637.0,"OwnerUserId":5703903.0,"Title":"How can I use values read from TFRecords as arguments to tf.set_shape?","Body":"<p>I would like to do something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone point me in the right direction? Thank you.<\/p>\n","answers":[{"AnswerId":"35777637","CreationDate":"2016-03-03T16:25:48.553","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You cannot pass a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Tensor\" rel=\"nofollow\"><code>tf.Tensor<\/code><\/a> to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Tensor.set_shape\" rel=\"nofollow\"><code>Tensor.set_shape()<\/code><\/a>. The only way to use a value read from a TFRecord as the argument to <code>set_shape()<\/code> is to evaluate it (e.g. using <code>sess.run()<\/code>) and pass the resulting NumPy array to <code>set_shape()<\/code>.<\/p>\n\n<p>The reason for this limitation is somewhat subtle. <code>Tensor.set_shape()<\/code> is used to make assertions about the shapes of <em>all<\/em> values that a particular <code>Tensor<\/code> (i.e. a symbolic value) can hold. In particular, it is used to fill the gaps left by shape inference, when the programmer knows more about that data than the library can infer. However, these assertions can only use information that is available at graph construction time. Dynamically-varying shapes are represented using the special values <code>Dimension(None)<\/code> (for an unknown dimension) and <code>TensorShape(None)<\/code> (for a tensor of unknown rank).<\/p>\n\n<p>The intent throughout TensorFlow is that most operations should be able to work without statically known shapes, so that tensors of dynamically varying shape can be used. However, there are a few exceptions among the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#cropping\" rel=\"nofollow\">image processing ops used for cropping\/padding<\/a>:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#crop_to_bounding_box\" rel=\"nofollow\"><code>tf.image.crop_to_bounding_box()<\/code><\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#pad_to_bounding_box\" rel=\"nofollow\"><code>tf.image.pad_to_bounding_box()<\/code><\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#resize_image_with_crop_or_pad\" rel=\"nofollow\"><code>tf.image.resize_image_with_crop_or_pad()<\/code><\/a><\/li>\n<\/ul>\n\n<p>These three ops require that the shape is statically known. However, they are simply convenience wrappers around <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#pad\" rel=\"nofollow\"><code>tf.pad()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#slice\" rel=\"nofollow\"><code>tf.slice()<\/code><\/a> that take advantage of the static shape to avoid doing needless computation in every step. We've been slowly removing this dependency from the image ops, but feel free to raise a GitHub issue about this. In the meantime, you can use the lower-level ops directly to work around the lack of shape inference for your dynamic tensors.<\/p>\n"}]}
{"QuestionId":35774379,"AnswerCount":0,"Tags":"<dll><linker><anaconda><deep-learning><caffe>","CreationDate":"2016-03-03T14:06:21.207","AcceptedAnswerId":null,"OwnerUserId":3130081.0,"Title":"anaconda\/lib\/libtiff.so.5: no version information available","Body":"<p>I am running caffe on Ubuntu 14.04 with anaconda.<\/p>\n\n<p>When I run <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I don't know what's wrong. \nThis seems like some warnings, since the training process is OK.<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n","answers":[]}
{"QuestionId":35775359,"AnswerCount":2,"Tags":"<gpgpu><tensorflow><multi-gpu><lstm><recurrent-neural-network>","CreationDate":"2016-03-03T14:49:13.920","AcceptedAnswerId":40097325.0,"OwnerUserId":3052875.0,"Title":"Distributed (multi-device) implementation of sequence-to-sequence models in TensorFlow?","Body":"<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/seq2seq\/index.html\" rel=\"nofollow\">Here<\/a> is a very good tutorial on training a <strong>sequence-to-sequence model<\/strong> in TensorFlow. I just interested to know if there is a <strong>distributed version<\/strong> which leverage <strong>a set of GPUs on a single machine<\/strong> for better performance?<\/p>\n\n<p><a href=\"http:\/\/download.tensorflow.org\/paper\/whitepaper2015.pdf\" rel=\"nofollow\">TensorFlow white paper<\/a> has been mentioned that it is possible to train a large multilayer recurrent neural network (See <strong>Figure 8<\/strong> and \"<strong>model parallel training<\/strong>\" section) as used in <a href=\"http:\/\/papers.nips.cc\/paper\/5346-sequence-to-sequence-learning-with-neural\" rel=\"nofollow\">Sequence to Sequence Learning with Neural Networks<\/a>. Anybody know if the current tutorial cover model parallel training? \nIf no, how to improve the original tutorial to leverage a set of GPUs?<\/p>\n","answers":[{"AnswerId":"45949688","CreationDate":"2017-08-29T23:23:26.363","ParentId":null,"OwnerUserId":"1780428","Title":null,"Body":"<p>Check this out: \n<a href=\"https:\/\/github.com\/shixing\/xing_rnn\/tree\/master\/Seq2Seq\" rel=\"nofollow noreferrer\">https:\/\/github.com\/shixing\/xing_rnn\/tree\/master\/Seq2Seq<\/a><\/p>\n\n<p>it implements a attention-based seq2seq model which can put each layer on different GPU. <\/p>\n"},{"AnswerId":"40097325","CreationDate":"2016-10-17T23:41:37.157","ParentId":null,"OwnerUserId":"2081611","Title":null,"Body":"<p>This example covers multi-gpu training <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/deep_cnn\/index.html#training-a-model-using-multiple-gpu-cards\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/deep_cnn\/index.html#training-a-model-using-multiple-gpu-cards<\/a><\/p>\n\n<p>Basically to utilize multiple GPU's, you have to use tensorflow's built in scoping to explicitly tell it which device to run on. You can have a bunch of GPUs run in parallel, and then use some device to sum up the results. This involves a lot of tensorflow scoping, which must be explicitly done. So to use gpu 0, you could have a call similar to this: <\/p>\n\n<pre><code>        with tf.device(\"\/gpu:0\"):\n            #do your calculations\n<\/code><\/pre>\n\n<p>and then somewhere else in your code you have a synchronization step:<\/p>\n\n<pre><code>        with tf.device(aggregation_device):\n            #make your updates\n<\/code><\/pre>\n\n<p>The link explains this pretty well, but hopefully this can get you started<\/p>\n"}]}
{"QuestionId":35775468,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow><deep-learning>","CreationDate":"2016-03-03T14:53:53.113","AcceptedAnswerId":35778541.0,"OwnerUserId":4584951.0,"Title":"the missing example | pre-fetch and pre-process data using threads","Body":"<p>seems there many open questions about the usage of TensorFlow out there and some developer of tensorflow here active on stackoverflow. Here is another question. I want to generate training data on-the-fly in other thread(s) using numpy or something which does not belongs to TensorFlow. But, I do not want to go through re-compiling the entire TensorFlow source again and again. I simply waiting for another way. \"tf.py_func\" seems to be a workaround. But the <\/p>\n\n<p>This is related to [how-to-prefetch-data-using-a-custom-python-function-in-tensorflow][1]<\/p>\n\n<p>Here is my MnWE (minmal-not-working-example):<\/p>\n\n<p><strong>Update (now there is an output but a race-condition, too):<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I basically have three question:<\/p>\n\n<ul>\n<li>What's wrong with this code? It runs into an endless loss (which is not debug-able). See Line \"HERE THE SILENCE BEGIN ...\"<\/li>\n<li>How to extend this code to use more threads?<\/li>\n<li>Is it worth to convert to tf.Record large datasets or data which can be generated on the fly?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"35778541","CreationDate":"2016-03-03T17:08:09.700","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You have a mistake on this line:<\/p>\n\n<pre><code>t = threading.Thread(target=populate_queue, args=(sess, thread_pool, qData))\n<\/code><\/pre>\n\n<p>It should be <code>qData_enqueue_op<\/code> instead of <code>qData<\/code>. Otherwise your enqueue operations fail, and you get stuck trying to dequeue from queue of size 0. I saw this when trying to run your code and getting<\/p>\n\n<pre><code>TypeError: Fetch argument &lt;google3.third_party.tensorflow.python.ops.data_flow_ops.FIFOQueue object at 0x4bc1f10&gt; of &lt;google3.third_party.tensorflow.python.ops.data_flow_ops.FIFOQueue object at 0x4bc1f10&gt; has invalid type &lt;class 'google3.third_party.tensorflow.python.ops.data_flow_ops.FIFOQueue'&gt;, must be a string or Tensor. (Can not convert a FIFOQueue into a Tensor or Operation.)\n<\/code><\/pre>\n\n<p>Regarding other questions:<\/p>\n\n<ul>\n<li>You don't need to start queue runners in this example because you don't have any. Queue runners are created by input producers like string_input_producer which is essentially FIFO queue + logic to launch threads. You are replicating 50% of queue runner functionality by launching your own threads that do enqueue ops. (the other 50% is closing the queue)<\/li>\n<li>RE: converting to tf.record -- Python has this thing called Global Interpreter Lock which means that two bits of Python code can't execute concurrently. In practice that's mitigated by the fact that a lot of the time is spent in numpy C++ code or IO ops (which release GIL). So I think it's a matter of checking if you are able to achieve required parallelism using Python pre-processing pipelines. <\/li>\n<\/ul>\n"}]}
{"QuestionId":35775853,"AnswerCount":1,"Tags":"<python><machine-learning><sentiment-analysis><keras><lstm>","CreationDate":"2016-03-03T15:11:13.677","AcceptedAnswerId":35779034.0,"OwnerUserId":5930493.0,"Title":"How to specify the output labels to keras lstm","Body":"<p>I am new to Keras and I want to predict the polarity of each word in a sentence using LSTM. I am representing each sentence, by the corresponding pre trained word vectors. So my input representation is (maxlen,input_dimensions).\nBut I am not being able to understand how to give the labels. For each sentence, the word can be in 3 classes(pos\/neg\/neutral). So it will something like [0,2,0,0,1.....]. How do I give this output to a sequential model?<\/p>\n","answers":[{"AnswerId":"35779034","CreationDate":"2016-03-03T17:32:06.657","ParentId":null,"OwnerUserId":"1056563","Title":null,"Body":"<p>When calling<\/p>\n\n<pre><code>model.fit\n<\/code><\/pre>\n\n<p>you will supply the inputs and <em>optionally<\/em> the expected output . Usually the input is called 'X' and the output 'y'.<\/p>\n\n<p>The input will include a dimension representing the sentence\/phrase: you need to decide how long this will be for training. Note: another related consideration is the mini-batch size.<\/p>\n\n<p>The output will have one less dimension than the input. You want to put the <strong>next<\/strong> word <strong>after<\/strong> the sentence in the same input array spot (by ordinal): this is the <strong><em>expected output for sentence K<\/em><\/strong> where k is the ordinal within the input array and the corresponding ordinal in the output array.<\/p>\n"}]}
{"QuestionId":35777774,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-03T16:32:02.360","AcceptedAnswerId":null,"OwnerUserId":2615676.0,"Title":"What happens to events written after a checkpoint?","Body":"<p>I have a bunch of summary nodes (scalars, histograms, etc) that are constantly writing to the log. Checkpointing is not as frequent, and so I often have situations in which I'm recovering from a checkpoint that is earlier than the events that have been written to the log. When I resume from the checkpoint and start writing to the log again, what exactly happens? Do the old events get overwritten? The documentation is not very clear on this. Looking in TensorBoard, it appears as if the \"future\" events are still there. Ideally I'd like to flush everything ahead of the current  and just start over.<\/p>\n","answers":[{"AnswerId":"35853122","CreationDate":"2016-03-07T20:15:01.090","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>TensorBoard does have logic to handle this case - it looks for restart events, and tries to purge everything with a global_step greater than the restart step. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/python\/summary\/event_accumulator.py#L310\" rel=\"nofollow\">See this code<\/a>. If you are still seeing the orphaned events, that means something isn't working - maybe the SessionLog.START event isn't being written when your job restarts from checkpoint?<\/p>\n\n<p>Can you create a simple repro of this and file an issue on GitHub?<\/p>\n"}]}
{"QuestionId":35783948,"AnswerCount":0,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-03-03T22:02:18.483","AcceptedAnswerId":null,"OwnerUserId":4250174.0,"Title":"Implement a simple NN using Tensorflow, want values as outputs","Body":"<p>I am following the link:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners<\/a><\/p>\n\n<p>But in my case, I do not want one-hot vectors as the results ([True, False, True,..]). I want numeric values as outputs. \nSo I want to write my own evaluation parts, to measure the accuracy. For example, I will have maybe a vector of values ([1,2,3,....]) as an output. I am thinking maybe cosine similarity could be a good way. Then I will normalise the similarity to range of [0,1]. (That's where I am)\nCould anyone give better ideas on how I can measure the accuracy?<\/p>\n","answers":[]}
{"QuestionId":35784044,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-03T22:08:33.747","AcceptedAnswerId":35821850.0,"OwnerUserId":234167.0,"Title":"Tensorflow how to check if a tensor row is only zeroes?","Body":"<p>I'm training a simple network for predicting bounding box coordinates of a single object. There are however pictures where there is no object to be found. Since the network always makes a prediction, it also predicts a confidence value between 0 and 1 which should indicate the probability that there is an object in the picture. My tensor with the predictions is called  and it's a  tensor (confidence, x, y, width and height). Similarly the  tensor is also .<\/p>\n\n<p>Previously I was training only with images that always had an object, so I could basically do <\/p>\n\n<p><\/p>\n\n<p>I want to start training also with pictures with no objects and when there is no object in the picture, I don't want the network to be penalized for whichever coordinates it predicted. In this case all that matters is the confidence value, which should be close to 0 (no object).<\/p>\n\n<p>How should I structure my labels and loss function to accomplish this? I can set the label of images with  no objects to all zeroes, but how do I check that a particular row is only zeroes? And in that case, the corresponding row in the logits needs also to be set to zeroes (except the confidence value!) so that the loss incurred because of the coordinates is also zero.<\/p>\n","answers":[{"AnswerId":"50542084","CreationDate":"2018-05-26T11:05:23.903","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>You can use <code>tf.count_nonzero()<\/code> to check whether the tensor has all zeros or not. You can check the guide <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/count_nonzero\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"35821850","CreationDate":"2016-03-06T01:02:50.897","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>An approach to find if a row of a tensor is all zero:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimage = tf.fill([8,8], 0)\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\nimage_row = tf.slice(image, [1,0], [1, -1])\ntotal = tf.reduce_sum(tf.abs(image_row))\nis_all_zero = tf.equal(total, 0)\nprint sess.run([total, is_all_zero, image_row])\n<\/code><\/pre>\n\n<p>Result: <\/p>\n\n<pre><code>[0, True, array([[0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":35784696,"AnswerCount":1,"Tags":"<machine-learning><cluster-computing><tensorflow><cudnn>","CreationDate":"2016-03-03T22:53:26.650","AcceptedAnswerId":35787766.0,"OwnerUserId":3314143.0,"Title":"TensorFlow: Multi-GPU configuration (performance)","Body":"<p>I would like to know what is considered \"best practice\" for multi-GPU systems when training networks with TensorFlow.<\/p>\n\n<p>E.g., one of my networks looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Does TensorFlow allocate multiple GPUs efficiently? Or should I specify myself which GPU TensorFlow should use for a specific operation?<\/p>\n\n<p>I have not benchmarked it by now, just started some GPU experiments today. However, at the moment I have <strong>not<\/strong> specified which device to use on the convolutional layers but I did specify it for the fully-connected layers:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a good idea? Or should leave resource allocation open to TensorFlow?<\/p>\n\n<p>I guess one single \"stream\" of convolutional layers can not be computed in parallel?! So it does not matter which device does the convolution-, pooling-, ... part?!<\/p>\n\n<p>Any tips to get the best performance?<\/p>\n\n<p>Currently I am training on one node of a Slurm cluster with 2 GPUs but potentially I could train on more nodes, so 4, 6 or even 8 GPUs. However, I guess there would be much overhead with more than 2 GPUs?<\/p>\n\n<hr>\n\n<p><strong>EDIT (slow multi-GPU performance):<\/strong> After some tests I am quite astonished...if I let TensorFlow decide what to allocate and remove the device-specific statements the network trains <strong>considerably<\/strong> faster. This was really surprising to me...what could be more effective than having each output stream on one GPU when there are two GPUs total? Additionally it seems (according to the output) that Tensorflow is only using one GPU?!<\/p>\n\n<hr>\n\n<p><strong>EDIT2 (NaN values):<\/strong> After some more tests I experienced that my manual setup of  for stream 1 and  for stream 2 is not only slower than letting TensorFlow decide what to use (and according to the piped script output TensorFlow just uses <strong>one<\/strong> GPU) but also sometimes my (I do not know why) my \" for stream 1 and  for stream 2\"-solution just generates NaN values. Like directly or short after the init. Very weird.<\/p>\n\n<p>Does TensorFlow need some kind of thread locking or manual copy of input data for multiple GPUs?<\/p>\n","answers":[{"AnswerId":"35787766","CreationDate":"2016-03-04T04:02:03.463","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>The logic for default placement of devices lies in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/simple_placer.cc\" rel=\"noreferrer\">simple_placer.cc<\/a><\/p>\n\n<p>I may be missing something in the logic, but from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/simple_placer.cc#L349\" rel=\"noreferrer\">this line<\/a> it seems that it will put all GPU ops on gpu:0<\/p>\n\n<p>You can see from implementation that placement strategy doesn't take into account data transfer or computation costs, so manual placement is often better than automatic. For instance, if you are doing some kind of input pipeline, default placement usually places some data processing ops on GPU which makes things slower overall.<\/p>\n\n<p>As far as your implementation being slow...perhaps there's gpu0->gpu1 copy happening somewhere?<\/p>\n\n<p>Getting multi-GPU setups to work is very much an open area, let us know what you find!<\/p>\n"}]}
{"QuestionId":35786289,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-04T01:19:07.717","AcceptedAnswerId":null,"OwnerUserId":2292800.0,"Title":"Is it possible to feed a tensor to a placeholder?","Body":"<p>Is there a way to feed a tensor to a placeholder, or a similar mechanism of connecting tensors after they have been defined? <\/p>\n\n<p>I have been defining my networks in layers, but now I find myself needing to feed an early layer with a tensor from a later layer. This is pseudo code, I apologize if there are errors.<\/p>\n\n<pre><\/pre>\n\n<p>It would be great if inputs[1:] could be placeholders and I could feed preds[:-1] to them.<\/p>\n\n<p>If there is not a way, the messy approach I intend to use is to loop over time, from the rnn onward, defining all layers for that back propagation through time time step. Any help is greatly appreciated.<\/p>\n","answers":[{"AnswerId":"37771481","CreationDate":"2016-06-12T06:45:34.183","ParentId":null,"OwnerUserId":"6455403","Title":null,"Body":"<h3>using  tensorflow.placeholder_with_default ( ...  )<\/h3>\n\n<pre><code>#!\/usr\/bin\/python\n\nimport tensorflow as tf\n\ns = tf.InteractiveSession()\n\nx = tf.placeholder_with_default( [ [1,2,3,4,5,6] ] , [None , 6])\n\nprint s.run ( x )\n<\/code><\/pre>\n"}]}
{"QuestionId":35786317,"AnswerCount":6,"Tags":"<tensorflow>","CreationDate":"2016-03-04T01:21:58.790","AcceptedAnswerId":null,"OwnerUserId":6016036.0,"Title":"Installing tensorflow on Linux","Body":"<p>I am trying to install Tensorflow on Linux. Unfortunately, I can barely use Ubuntu. <\/p>\n\n<p>I followed the page \"<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#create-the-pip-package-and-install\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#create-the-pip-package-and-install<\/a>\" <\/p>\n\n<p>and I am stuck at \"Create the pip package and install\" part. \nWhen I typed \"\" the terminal gives me below<\/p>\n\n<pre><\/pre>\n\n<p>What's wrong with it?<\/p>\n","answers":[{"AnswerId":"37136447","CreationDate":"2016-05-10T10:54:26.177","ParentId":null,"OwnerUserId":"3930957","Title":null,"Body":"<p>I have faced similar problem. Then I realized that I forgot to set up the environment as follows:<\/p>\n\n<pre><code>export PATH=\"$PATH:$HOME\/bin\"\n<\/code><\/pre>\n"},{"AnswerId":"37076423","CreationDate":"2016-05-06T15:42:57.407","ParentId":null,"OwnerUserId":"1303432","Title":null,"Body":"<p>You can  install tensorflow from \"pip install\" directly or you can install it from source. If you choose to install from source code. First you should install the bazel, it is a tool to build tensorflow from source. <\/p>\n"},{"AnswerId":"36781403","CreationDate":"2016-04-21T22:16:46.160","ParentId":null,"OwnerUserId":"5627842","Title":null,"Body":"<p>I wrote a small bash script that sets up Ubuntu to be ready to run TensorFlow. It may be of use to you:<\/p>\n\n<p><a href=\"https:\/\/github.com\/jasonmayes\/Tensor-Flow-on-Google-Compute-Engine\" rel=\"nofollow\">https:\/\/github.com\/jasonmayes\/Tensor-Flow-on-Google-Compute-Engine<\/a><\/p>\n\n<p>From the script here you can see how I install bazel, however that also has a number of dependencies :-)<\/p>\n\n<pre><code>wget https:\/\/github.com\/bazelbuild\/bazel\/releases\/download\/0.1.5\/bazel-0.1.5-installer-linux-x86_64.sh -O bazel-0.1.5-installer-linux-x86_64.sh\nchmod +x bazel-0.1.5-installer-linux-x86_64.sh\nsudo .\/bazel-0.1.5-installer-linux-x86_64.sh --user\nrm bazel-0.1.5-installer-linux-x86_64.sh\nsudo chown $USER:$USER ~\/.cache\/bazel\/\n<\/code><\/pre>\n"},{"AnswerId":"42608028","CreationDate":"2017-03-05T11:49:12.267","ParentId":null,"OwnerUserId":"4266234","Title":null,"Body":"<p>If your system has pip installed, it is an easy task. If not, install pip using:<\/p>\n\n<pre><code>sudo apt-get install python-pip python-dev\n<\/code><\/pre>\n\n<p>You can run Tensorflow using the system GPU or CPU. Also note that, as of now, tensorflow only supports Nvidia GPU. If you have an Nvidia GPU, <a href=\"http:\/\/www.nvidia.com\/object\/gpu-accelerated-applications-tensorflow-installation.html\" rel=\"nofollow noreferrer\">reading this link<\/a> would be better as it gives good level of clarity.<\/p>\n\n<p>For CPU installation run-<\/p>\n\n<ol>\n<li><p><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.10.0-cp27-none-linux_x86_64.whl<\/code><\/p>\n\n<p>This sets the url where the binaries reside. It changes from time to time, so make sure you use the appropriate url. Next, run-<\/p><\/li>\n<li><p><code>sudo pip install\u200a--upgrade $TF_BINARY_URL<\/code><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"38594644","CreationDate":"2016-07-26T15:55:12.173","ParentId":null,"OwnerUserId":"3256363","Title":null,"Body":"<p>Just an update:<\/p>\n\n<pre class=\"lang-bash prettyprint-override\"><code># Ubuntu\/Linux 64-bit, CPU only, Python 2.7\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.9.0-cp27-none-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, GPU enabled, Python 2.7\n# Requires CUDA toolkit 7.5 and CuDNN v4. For other versions, see \"Install from sources\" below.\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.9.0-cp27-none-linux_x86_64.whl\n\n# Mac OS X, CPU only, Python 2.7:\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.9.0-py2-none-any.whl\n\n# Ubuntu\/Linux 64-bit, CPU only, Python 3.4\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.9.0-cp34-cp34m-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, GPU enabled, Python 3.4\n# Requires CUDA toolkit 7.5 and CuDNN v4. For other versions, see \"Install from sources\" below.\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.9.0-cp34-cp34m-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, CPU only, Python 3.5\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.9.0-cp35-cp35m-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, GPU enabled, Python 3.5\n# Requires CUDA toolkit 7.5 and CuDNN v4. For other versions, see \"Install from sources\" below.\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.9.0-cp35-cp35m-linux_x86_64.whl\n\n# Mac OS X, CPU only, Python 3.4 or 3.5:\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.9.0-py3-none-any.whl\nInstall TensorFlow:\n\n# Python 2\n$ sudo pip install --upgrade $TF_BINARY_URL\n\n# Python 3\n$ sudo pip3 install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>More details on this:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md<\/a><\/p>\n"},{"AnswerId":"35786405","CreationDate":"2016-03-04T01:31:56.773","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>You need to install \"Bazel\", google's homegrown build tool for building the sources. You can get it here: <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"noreferrer\">Installing Bazel<\/a><\/p>\n\n<p>But if you are just getting started with this do you even need to bother with building from the source code?  <\/p>\n\n<p>Maybe you should try just installing the binaries using pip with a one line command as described here: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#pip-installation\" rel=\"noreferrer\">pip installation<\/a><\/p>\n\n<pre><code># Ubuntu\/Linux 64-bit, CPU only:\n$ sudo pip install --upgrade   https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl\n\n# Ubuntu\/Linux 64-bit, GPU enabled:\n$ sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":35788873,"AnswerCount":1,"Tags":"<neural-network><caffe><conv-neural-network>","CreationDate":"2016-03-04T05:48:41.387","AcceptedAnswerId":35789367.0,"OwnerUserId":1279459.0,"Title":"What is the output of fully connected layer in CNN?","Body":"<p>For example, in Caffe, one should define num_output in an Inner Product (Fully Connected) layer. What is the meaning of this output number?<\/p>\n","answers":[{"AnswerId":"35789367","CreationDate":"2016-03-04T06:25:48.273","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>Consider fully connect layer as a simple matrix-matrix multiplication of <code>1xN<\/code> and <code>NxM<\/code> to produce a result of dimension <code>1xM<\/code>.<\/p>\n\n<p>Let us consider that we pass a data of dimension say <code>56x56x3<\/code> as the input of a fully connected layer. Let the dimension of the weight be unknown <code>NxM<\/code>. Consider, we set <code>num_ouput = 4096<\/code>.<\/p>\n\n<p>For computing these data, the fully connected layer reshapes the input data of dimension <code>56x56x3<\/code> as <code>1xN<\/code>, <code>1x(56x56x3) = 1x9408<\/code>.<\/p>\n\n<p>Thus,<\/p>\n\n<blockquote>\n  <p>N = 9408<\/p>\n  \n  <p>M=num_output=4096<\/p>\n<\/blockquote>\n\n<p>In effect we end up doing a <code>(1x9408)matrix - (9408x4096) matrix<\/code> multiplication.<\/p>\n\n<p>If the num_output value was changed to say <code>100<\/code>, it would end up doing <code>(1x9408)matrix - (9408x100) matrix<\/code> multiplication. <\/p>\n\n<p>Thus increasing the <code>num_ouput<\/code> value will increase the number of weight parameters that the model has to learn.<\/p>\n"}]}
{"QuestionId":35789500,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-03-04T06:34:27.023","AcceptedAnswerId":null,"OwnerUserId":5638068.0,"Title":"memory issue in tensorflow","Body":"<p>I am trying to build a gaussian RBM model with Tensorflow. But the program will use too much memory.<\/p>\n\n<p>gaussian_rbm.py<\/p>\n\n<pre><\/pre>\n\n<p>test_mnist.py<\/p>\n\n<pre><\/pre>\n\n<p>the output is<\/p>\n\n<blockquote>\n  <p>run test_mnist.py Extracting MNIST_data\/train-images-idx3-ubyte.gz\n  Extracting MNIST_data\/train-labels-idx1-ubyte.gz Extracting\n  MNIST_data\/t10k-images-idx3-ubyte.gz Extracting\n  MNIST_data\/t10k-labels-idx1-ubyte.gz I\n  tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:900] 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_init.cc:102] Found device 0\n  with properties:  name: GeForce GTX 560 major: 2 minor: 1\n  memoryClockRate (GHz) 1.62 pciBusID 0000:01:00.0 Total memory:\n  1018.69MiB Free memory: 916.73MiB I tensorflow\/core\/common_runtime\/gpu\/gpu_init.cc:126] DMA: 0  I\n  tensorflow\/core\/common_runtime\/gpu\/gpu_init.cc:136] 0:   Y  I\n  tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:684] Ignoring gpu\n  device (device: 0, name: GeForce GTX 560, pci bus id: 0000:01:00.0)\n  with Cuda compute capability 2.1. The minimum required Cuda capability\n  is 3.5. step: 0\n  0.0911714\n  0.0781856\n  0.0773076\n  0.0770751\n  0.0776582\n  0.0764748\n  0.0755164\n  0.0741131\n  0.0726497\n  0.0712237\n  0.0701839\n  0.0686315\n  0.0664856\n  0.0658309\n  0.0646239\n  0.0626652\n  0.0616178\n  0.0610061\n  0.0598332\n  0.0588843\n  0.0587477\n  0.0572056\n  0.0561556\n  0.0554848 Killed<\/p>\n<\/blockquote>\n\n<p>Is there some way to monitor the memory?\nCan someone help me?<\/p>\n","answers":[{"AnswerId":"47206453","CreationDate":"2017-11-09T16:05:45.890","ParentId":null,"OwnerUserId":"2592317","Title":null,"Body":"<p>Make sure there are no memory leaks by making your graph read-only before training with<\/p>\n\n<pre><code>tf.get_default_graph().finalize()\n<\/code><\/pre>\n\n<p>TensorFlow will raise an exception each time you will try to add a new node.<\/p>\n"},{"AnswerId":"44066729","CreationDate":"2017-05-19T09:47:45.287","ParentId":null,"OwnerUserId":"1738214","Title":null,"Body":"<p>There may be a problem with the training loop, that causes your computer to run out of memory.<\/p>\n\n<p>For each iteration of your loop, you are calling:<\/p>\n\n<pre><code>sess.run(rbm_modle.cd1(), feed_dict={rbm_modle.input : trX[start : end]})\n<\/code><\/pre>\n\n<p>Inside this <code>rbm_modle.cd1()<\/code> function, you are creating several new operations, such as <code>tf.matmul()<\/code>, so every time you call <code>rbm_modle.cd1()<\/code> you will create new operations, which will cause the used memory to increase after each iteration.<\/p>\n\n<p>You should define all the operations before the loop, and then during run the operation with <code>sess.run()<\/code>, without creating new ones.<\/p>\n"},{"AnswerId":"35803706","CreationDate":"2016-03-04T18:43:55.383","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>You can monitory GPU memory with the command  <a href=\"https:\/\/developer.nvidia.com\/nvidia-system-management-interface\" rel=\"nofollow\">nvidia-smi<\/a><\/p>\n\n<p>It looks like your GPU does not support the later versions of CUDA required to run tensorflow. You can check <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow\">CUDA-Enabled GeForce Products<\/a><\/p>\n\n<p>From your output it looks like tensorflow is smart enough to not use the GPU, so either your model \/ batch size is too big for your RAM or you have a memory leak.<\/p>\n\n<p>Try running running session with log_device_placement=True to see what tensorflow is doing step by step, while running 'top' to monitor memory?  <\/p>\n\n<pre><code>    with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:\n<\/code><\/pre>\n"},{"AnswerId":"43278030","CreationDate":"2017-04-07T12:21:21.653","ParentId":null,"OwnerUserId":"7414362","Title":null,"Body":"<p>The answer seems correct, (Insufficient compute capabilities to run latest version of CUDA \/ Tensorflow<\/p>\n\n<p>However, the minimal requirement seems to be \"Compute Capabilites = 3.0\", as my GTX_770M is able to run Tensorflow 1.0 \/ CUDA 8.0 (see below) <\/p>\n\n<p>And \/ Or trying to recompile tensorflow from the sources, and include the 2.0 target during generation (it is proposing 3.5-5.5 by default)<\/p>\n\n<p>Have a good day!!<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 375.51                 Driver Version: 375.51                    |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  GeForce GTX 770M    Off  | 0000:01:00.0     N\/A |                  N\/A |\n|100%   48C    P0    N\/A \/  N\/A |   2819MiB \/  3017MiB |     N\/A      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID  Type  Process name                               Usage      |\n|=============================================================================|\n|    0                  Not Supported                                         |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n"}]}
{"QuestionId":35789666,"AnswerCount":2,"Tags":"<python><tensorflow><skflow>","CreationDate":"2016-03-04T06:45:47.960","AcceptedAnswerId":null,"OwnerUserId":5395642.0,"Title":"TensorFlow with skflow: AttributeError: 'module' object has no attribute 'saver_pb2'","Body":"<p>I tried to run the  example from skflow, without changing the code at all.<\/p>\n\n<p>However, I got an error, and I am unsure how to proceed.\nIt seems that it happened when it restores the model from a saved file.<\/p>\n\n<pre><\/pre>\n\n<p>I use the following versions:<\/p>\n\n<ul>\n<li>anaconda2-2.5.0<\/li>\n<li>tensorflow 0.7 CPU-only for Mac OS X<\/li>\n<li>skflow 0.1.0<\/li>\n<\/ul>\n\n<p>How can I fix it?<\/p>\n","answers":[{"AnswerId":"35789961","CreationDate":"2016-03-04T07:02:57.713","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This looks like a bug in skflow: the current version of skflow is not compatible with the latest version of TensorFlow (0.7.0 and up). I submitted a <a href=\"https:\/\/github.com\/tensorflow\/skflow\/pull\/127\" rel=\"nofollow\">pull request<\/a> to fix it. (<strong>EDIT:<\/strong> The pull request has now been merged, so upgrading to the latest version of skflow will also fix things.)<\/p>\n\n<p>In the meantime, you can fix it by editing line 401 of <code>\"skflow\/estimators\/base.py\"<\/code> (full path in your stacktrace), and replacing:<\/p>\n\n<pre><code>saver_def = tf.python.training.saver_pb2.SaverDef()\n<\/code><\/pre>\n\n<p>...with:<\/p>\n\n<pre><code>saver_def = tf.train.SaverDef()\n<\/code><\/pre>\n"},{"AnswerId":"35946221","CreationDate":"2016-03-11T17:20:01.000","ParentId":null,"OwnerUserId":"3865298","Title":null,"Body":"<p>The author just fixed this issue today: <a href=\"https:\/\/github.com\/tensorflow\/skflow\/issues\/144\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/skflow\/issues\/144<\/a>. You'd have to remove and reinstall skflow from github (the most up-to-date release).<\/p>\n"}]}
{"QuestionId":35790125,"AnswerCount":1,"Tags":"<python><import><module><caffe>","CreationDate":"2016-03-04T07:13:53.757","AcceptedAnswerId":null,"OwnerUserId":1157401.0,"Title":"python from . import fails","Body":"<p>Trying caffe python examples from: <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html<\/a> gives me error:<\/p>\n\n<pre><\/pre>\n\n<p>Apparently the sigtools import fails, but I can't figure out why. The \/usr\/lib\/python2.7\/site-packages\/scipy\/signal contains all files:<\/p>\n\n<pre><\/pre>\n\n<p>In general, how python process directives like this, specifically what dot is resolved to if my working directory was completely different from the location where sigtools package is located?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35791314","CreationDate":"2016-03-04T08:31:47.443","ParentId":null,"OwnerUserId":"1351609","Title":null,"Body":"<p>As stated here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/9439480\/from-import-vs-import\">`from ... import` vs `import .`<\/a><\/p>\n\n<p>\"<strong>from . import sigtools<\/strong>\" imports the main module \".\" (which is \"signal\") than imports the object\/module sigtools. If \".\" has been already imported it rely on that structure.<\/p>\n\n<p>I think that <strong>this can be tricky in case you have 2 modules with the same name<\/strong> in the python import path: the interpreter imports the first one found and never imports the second one. If the second one has more modules than the first one, this can lead to something similar to your problem.<\/p>\n"}]}
{"QuestionId":35790877,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><lstm>","CreationDate":"2016-03-04T08:01:22.557","AcceptedAnswerId":null,"OwnerUserId":6017152.0,"Title":"Tensorflow: Attempting to use uninitialized value RNN\/GRUCell\/Gates\/Linear\/Bias","Body":"<h3>Environment info<\/h3>\n\n<p>Operating System:\nRocks OS (Centos 6.5)<\/p>\n\n<p>I installed from sources, and here is my version:\n<a href=\"https:\/\/github.com\/shiyemin\/tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/shiyemin\/tensorflow\/<\/a>\nNothing changed but to make it compile successfully on our server.<\/p>\n\n<h3>ERROR<\/h3>\n\n<p>I use caffe-tensorflow to convert caffe model to tensorflow and the GoogLeNet is selected to construct our network.<\/p>\n\n<p>I add  a LSTM layer to this code as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When i add this LSTM layer to GoogLeNet, the following error occurs.<\/p>\n\n<pre><\/pre>\n\n<p>But when i using the code from: <\/p>\n\n<p><a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/recurrent_network.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/recurrent_network.py<\/a>, <\/p>\n\n<p>everything works well.<\/p>\n\n<p>Anyone knows what happened? I don't know how to debug this error.<\/p>\n","answers":[]}
{"QuestionId":35792278,"AnswerCount":3,"Tags":"<deep-learning><keras>","CreationDate":"2016-03-04T09:25:12.327","AcceptedAnswerId":35827171.0,"OwnerUserId":4497662.0,"Title":"How to find Number of parameters of a keras model?","Body":"<p>For a Feedforward Network (FFN), it is easy to compute the number of parameters. Given a CNN, LSTM etc is there a quick way to find the number of parameters in a keras model?<\/p>\n","answers":[{"AnswerId":"35827171","CreationDate":"2016-03-06T13:07:21.063","ParentId":null,"OwnerUserId":"2602240","Title":null,"Body":"<p>Models and layers have special method for that purpose:<\/p>\n\n<pre><code>model.count_params()\n<\/code><\/pre>\n\n<p>Also, to get a short summary of each layer dimensions and parameters, you might find useful the following method <\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n"},{"AnswerId":"50934285","CreationDate":"2018-06-19T17:50:42.410","ParentId":null,"OwnerUserId":"4367801","Title":null,"Body":"<p>Tracing back the <code>print_summary()<\/code> function, Keras developers compute the number of trainable and non_trainable parameters of a given <code>model<\/code> as follows:<\/p>\n\n<pre><code>import keras.backend as K\nimport numpy as np\n\ntrainable_count = int(np.sum([K.count_params(p) for p in set(model.trainable_weights)]))\n\nnon_trainable_count = int(np.sum([K.count_params(p) for p in set(model.non_trainable_weights)]))\n<\/code><\/pre>\n\n<p>Given that <code>K.count_params()<\/code> is defined as <code>np.prod(int_shape(x))<\/code>, this solution is quite similar to the one of Anuj Gupta, except for the use of <code>set()<\/code> and the way the shape of the tensors are retrieved.<\/p>\n"},{"AnswerId":"35792801","CreationDate":"2016-03-04T09:49:27.380","ParentId":null,"OwnerUserId":"4497662","Title":null,"Body":"<pre><code>import keras.backend as K\n\ndef size(model): # Compute number of params in a model (the actual number of floats)\n    return sum([np.prod(K.get_value(w).shape) for w in model.trainable_weights])\n<\/code><\/pre>\n"}]}
{"QuestionId":35792910,"AnswerCount":4,"Tags":"<python><theano><deep-learning>","CreationDate":"2016-03-04T09:54:05.150","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"logistic_sgd module, where to find it?","Body":"<p>doing a deep learning tutorial and my python cannot find that module.<\/p>\n\n<pre><\/pre>\n\n<p>How can i install it ? <\/p>\n","answers":[{"AnswerId":"42459689","CreationDate":"2017-02-25T18:05:03.220","ParentId":null,"OwnerUserId":"7622531","Title":null,"Body":"<p>are you using the Theano library, if you are, that means your project cannot find the Theano library, use <\/p>\n\n<pre><code>sys.path.append(r'~\/TheanoDL')\n<\/code><\/pre>\n\n<p>to declare it at the beginning of your project, change the path of it with your Theano is in.<\/p>\n\n<p>Note that it works for a project with a customed Theano library.(my situation)<\/p>\n\n<p>Best luck!<\/p>\n"},{"AnswerId":"41695109","CreationDate":"2017-01-17T10:48:43.967","ParentId":null,"OwnerUserId":"6936471","Title":null,"Body":"<p>Actually, you should download all the code from the following link, it contains all the dependencies you need. \nHere is the source code:\n<a href=\"http:\/\/deeplearning.net\/tutorial\/code\/\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/code\/<\/a><\/p>\n"},{"AnswerId":"45426320","CreationDate":"2017-07-31T22:34:16.827","ParentId":null,"OwnerUserId":"3954988","Title":null,"Body":"<p>To download all the code necessary, you can go to GitHub.<\/p>\n\n<p>Citing from: <a href=\"http:\/\/deeplearning.net\/tutorial\/\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/<\/a><\/p>\n\n<blockquote>\n  <p>The code is available on the <a href=\"https:\/\/github.com\/lisa-lab\/DeepLearningTutorials\" rel=\"nofollow noreferrer\" title=\"Deep Learning Tutorial repositories\">Deep Learning Tutorial repositories<\/a>.<\/p>\n<\/blockquote>\n\n<p>If you aren't familiar with Git, you can download the source code in one zip archive file from there.<\/p>\n"},{"AnswerId":"35937736","CreationDate":"2016-03-11T10:27:13.627","ParentId":null,"OwnerUserId":"4032607","Title":null,"Body":"<p>Download and Save logistic_sgd.py from the following link:<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py\" rel=\"noreferrer\">http:\/\/deeplearning.net\/tutorial\/code\/logistic_sgd.py<\/a><\/p>\n\n<p>Store it in the working directory. That's it!<\/p>\n"}]}
{"QuestionId":35796227,"AnswerCount":1,"Tags":"<linux><ubuntu><caffe>","CreationDate":"2016-03-04T12:35:06.500","AcceptedAnswerId":null,"OwnerUserId":6018264.0,"Title":"\/usr\/bin\/ld: cannot find -lxxx","Body":"<p>I'm trying to install caffe in Ubuntu 15.04, but get stuck in the make execution. Here is my command and the results. Can anybody help me out?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35803065","CreationDate":"2016-03-04T18:06:21.960","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>It seems like cuDNN is not installed in your system.<\/p>\n\n<p>Try commenting out the line <code>USE_CUDNN := 1<\/code> in Makefile.config. Do a <code>make clean<\/code> and build the project again.<\/p>\n"}]}
{"QuestionId":35797653,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2016-03-04T13:44:46.307","AcceptedAnswerId":35805728.0,"OwnerUserId":2213825.0,"Title":"Theano matrix composed of theano scalars","Body":"<p>How can I create a  matrix composed of  scalars?\nThe following code creates a  array composed of  scalars. But I want to have a  matrix instead.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35805728","CreationDate":"2016-03-04T20:47:21.380","ParentId":null,"OwnerUserId":"1461210","Title":null,"Body":"<p>You can use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.stacklists\" rel=\"nofollow\"><code>theano.tensor.stacklists<\/code><\/a> in much the same way as you would use <code>np.array<\/code> to construct a normal numpy array:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nfrom theano import tensor as te\n\na = te.fscalar(\"a\")\nb = te.fscalar(\"b\")\nM = te.stacklists([[a, b], [b, a]])\n\nf = theano.function([a, b], M)\n\nprint(f(1.0, 2.0))\n# [[ 1.  2.]\n#  [ 2.  1.]]\n<\/code><\/pre>\n\n<p>You could achieve the same result by using <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.stack\" rel=\"nofollow\"><code>theano.tensor.stack<\/code><\/a> or <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor.concatenate\" rel=\"nofollow\"><code>theano.tensor.concatenate<\/code><\/a> to construct a 1D vector from your scalars, then using its <code>reshape<\/code> method to reshape it into a matrix\/tensor with your desired dimensions.<\/p>\n"}]}
{"QuestionId":35799198,"AnswerCount":1,"Tags":"<python><tensorflow><skflow>","CreationDate":"2016-03-04T15:02:44.860","AcceptedAnswerId":35800085.0,"OwnerUserId":4291805.0,"Title":"skflow example throws errors","Body":"<p>I'm quite new to python and skflow, but I am experiencing a really basic problem and hope someone could help me. <\/p>\n\n<p>I've installed skflow and downloaded the dbpedia_csv which is used in the examples. However upon running the 'text_classification_character_rnn.py' example, I am hit with an error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on what I am doing wrong? Also, the  is working, but the  no accuracy is printed. I'm using PyCharm Edu 2.0.3 to run the code.<\/p>\n","answers":[{"AnswerId":"35800085","CreationDate":"2016-03-04T15:42:02.363","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This looks like a version mismatch between skflow and TensorFlow. From the error message, it looks like you have an old version (0.5 or 0.6) of TensorFlow. However, skflow was <a href=\"https:\/\/github.com\/tensorflow\/skflow\/commit\/d99245f1ccd6be547702c3e4d4e8f8e54597c569\" rel=\"nofollow\">recently updated<\/a> to use a new interface in TensorFlow, and it looks like you are using this latest version of skflow.<\/p>\n\n<p>The simplest solution would be to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"nofollow\">upgrade TensorFlow<\/a> to version 0.7.1<\/p>\n"}]}
{"QuestionId":35800780,"AnswerCount":1,"Tags":"<ubuntu><makefile><caffe>","CreationDate":"2016-03-04T16:13:01.987","AcceptedAnswerId":null,"OwnerUserId":6018264.0,"Title":"Caffe make runtest error","Body":"<p>I just used make and make test command successfully without errors. However, when I used make runtest command, it failed several times for the same problem.<\/p>\n\n<p>Though my installed boost is boost-1.55, it doesn't work. The result is as following. What should I do? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38673485","CreationDate":"2016-07-30T11:34:24.323","ParentId":null,"OwnerUserId":"2465534","Title":null,"Body":"<p>If you use Intel MKL, the cause of this issue maybe: Intel MKL's float point operations (such as matrix multiplication) are non-deterministic by default.<\/p>\n\n<pre><code>    export MKL_CBWR=AUTO\n<\/code><\/pre>\n"}]}
{"QuestionId":35800893,"AnswerCount":2,"Tags":"<python><deep-learning><face-recognition><torch>","CreationDate":"2016-03-04T16:17:43.990","AcceptedAnswerId":35881615.0,"OwnerUserId":5024858.0,"Title":"Trouble Installing OpenFace in Python","Body":"<p>I am new to deeplearning and face recognition. After searching, I found this python package about deeplearning applied to face recognition called OpenFace. From its documentation, I think it is build on top of Torch for the neural nets computation.<\/p>\n\n<p>I want to install the package in a virtual environment, so basically these are steps I made:<\/p>\n\n<ol>\n<li>brew install the necessary system requirements: bash,coreutils,curl,findutils,opencv, python and boost-python<\/li>\n<li>Make a virtual environment and install dlib, numpy, scipy, pandas, scikit-learn, scikit-image<\/li>\n<li>Cloned the openface github repository <\/li>\n<li><p>Install Torch <\/p>\n\n<p><\/p><\/li>\n<li>cd to cloned openface repo and run \n<\/li>\n<\/ol>\n\n<p>However when I run python:<\/p>\n\n<p><\/p>\n\n<p>I get:<\/p>\n\n<p><\/p>\n\n<p>How do I fix this? Also, are there any other tutorials for using openface?\nHow to properly install OpenFace? <\/p>\n","answers":[{"AnswerId":"35881615","CreationDate":"2016-03-09T02:03:54.340","ParentId":null,"OwnerUserId":"1381755","Title":null,"Body":"<p>As I posted in the comments, this segfault was caused by compiling dlib with one Python version and running it with another. This was resolved by manually installing dlib rather than using their pip package.<\/p>\n"},{"AnswerId":"47253270","CreationDate":"2017-11-12T19:55:00.057","ParentId":null,"OwnerUserId":"2414957","Title":null,"Body":"<p>I installed dlib with pip for openface and had no problem importing it for Python 2.7:<\/p>\n\n<pre><code>[jalal@goku scratch]$ sudo \/usr\/bin\/pip install dlib\n[sudo] password for jalal:\nCollecting dlib\n  Downloading dlib-19.7.0.tar.gz (4.0MB)\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| 4.0MB 347kB\/s\nInstalling collected packages: dlib\n\n\u25bd\n  Running setup.py install for dlib ... done\nSuccessfully installed dlib-19.7.0\nYou are using pip version 8.1.2, however version 9.0.1 is available.\nYou should consider upgrading via the 'pip install --upgrade pip' command.\n<\/code><\/pre>\n"}]}
{"QuestionId":35801397,"AnswerCount":2,"Tags":"<python><pickle><theano>","CreationDate":"2016-03-04T16:40:47.523","AcceptedAnswerId":35806750.0,"OwnerUserId":2213825.0,"Title":"Saving and Loading theano variables","Body":"<p>I am using a  variable to define a function. Then I save the variable. When I load the variable  does not recognize it as being the same variable used before.<\/p>\n\n<p>This code:<\/p>\n\n<pre><\/pre>\n\n<p>Generates the error:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried to Save\/Load using  with the same results.<\/p>\n","answers":[{"AnswerId":"35802055","CreationDate":"2016-03-04T17:14:14.717","ParentId":null,"OwnerUserId":"5837113","Title":null,"Body":"<p>I think your input is right,but <em>theano<\/em> claim your function must be used or set another parameter <em>'warn'<\/em> or <em>'ignore'<\/em> in your function.<\/p>\n"},{"AnswerId":"35806750","CreationDate":"2016-03-04T21:51:41.400","ParentId":null,"OwnerUserId":"1913824","Title":null,"Body":"<p>When you load <code>x<\/code> from pickle, you get a <em>copy<\/em> of <code>x<\/code>, not the original <code>x<\/code>. Pickle is a serialization protocol. For its nature, the identity of objects cannot be guaranteed to be the same.<\/p>\n\n<p>You must also consider that pickle is meant to be used to dump\/load data across different Python processes, not within the same process (I do not see any benefit in using pickle in your code).<\/p>\n\n<p>What you should do is dumping and loading <code>x<\/code> together with <code>y<\/code>:<\/p>\n\n<pre><code>dump((x, y), f)\nx, y = load(f)\n<\/code><\/pre>\n\n<p>Pickle can guarantee that the dumped <code>x<\/code> is \"linked\" to the dumped <code>y<\/code> (and, consequently, the loaded <code>x<\/code> is \"linked\" to the loaded <code>y<\/code>).<\/p>\n"}]}
{"QuestionId":35803425,"AnswerCount":2,"Tags":"<tensorflow><word-embedding>","CreationDate":"2016-03-04T18:26:53.133","AcceptedAnswerId":null,"OwnerUserId":1718064.0,"Title":"Update only part of the word embedding matrix in Tensorflow","Body":"<p>Assuming that I want to update a pre-trained word-embedding matrix during training, is there a way to update only a subset of the word embedding matrix?<\/p>\n\n<p>I have looked into the Tensorflow API page and found this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However how do I apply that to the word-embedding matrix. Suppose I do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>How do I then use  and  to update the original embeddign matrix? (Also, tensorflow throws an error if the second argument of  is not a )<\/p>\n","answers":[{"AnswerId":"35837251","CreationDate":"2016-03-07T06:09:56.477","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> The default implementation of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.minimize\"><code>opt.minimize(loss)<\/code><\/a>, TensorFlow will generate a <em>sparse<\/em> update for <code>word_emb<\/code> that modifies only the rows of <code>word_emb<\/code> that participated in the forward pass.<\/p>\n\n<p>The gradient of the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#gather\"><code>tf.gather(word_emb, indices)<\/code><\/a> op with respect to <code>word_emb<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#IndexedSlices\"><code>tf.IndexedSlices<\/code><\/a> object (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c1a40c7b9085719cb120ae92f991f86d0757af7e\/tensorflow\/python\/ops\/array_grad.py#L176\">see the implementation for more details<\/a>). This object represents a sparse tensor that is zero everywhere, except for the rows selected by <code>indices<\/code>. A call to <code>opt.minimize(loss)<\/code> calls <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c1a40c7b9085719cb120ae92f991f86d0757af7e\/tensorflow\/python\/training\/adam.py#L131\"><code>AdamOptimizer._apply_sparse(word_emb_grad, word_emb)<\/code><\/a>, which makes a call to <code>tf.scatter_sub(word_emb, ...)<\/code>* that updates only the rows of <code>word_emb<\/code> that were selected by <code>indices<\/code>.<\/p>\n\n<p>If on the other hand you want to modify the <code>tf.IndexedSlices<\/code> that is returned by <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.compute_gradients\"><code>opt.compute_gradients(loss, word_emb)<\/code><\/a>, you can perform arbitrary TensorFlow operations on its <code>indices<\/code> and <code>values<\/code> properties, and create a new <code>tf.IndexedSlices<\/code> that can be passed to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.apply_gradients\"><code>opt.apply_gradients([(word_emb, ...)])<\/code><\/a>. For example, you could cap the gradients using <code>MyCapper()<\/code> (as in the example) using the following calls:<\/p>\n\n<pre><code>grad, = opt.compute_gradients(loss, word_emb)\ntrain_op = opt.apply_gradients(\n    [tf.IndexedSlices(MyCapper(grad.values), grad.indices)])\n<\/code><\/pre>\n\n<p>Similarly, you could change the set of indices that will be modified by creating a new <code>tf.IndexedSlices<\/code> with a different indices.<\/p>\n\n<hr>\n\n<p>*&nbsp;In general, if you want to update only part of a variable in TensorFlow, you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#sparse-variable-updates\"><code>tf.scatter_update()<\/code>, <code>tf.scatter_add()<\/code>, or <code>tf.scatter_sub()<\/code> operators<\/a>, which respectively set, add to (<code>+=<\/code>) or subtract from (<code>-=<\/code>) the value previously stored in a variable.<\/p>\n"},{"AnswerId":"45831707","CreationDate":"2017-08-23T05:59:22.723","ParentId":null,"OwnerUserId":"6718080","Title":null,"Body":"<p>Since you just want to select the elements to be updated (and not to change the gradients), you can do as follows.<\/p>\n\n<p>Let <code>indices_to_update<\/code> be a boolean tensor that indicates the indices you wish to update, and <code>entry_stop_gradients<\/code> is defined in the link, Then:<\/p>\n\n<pre><code>gather_emb = entry_stop_gradients(gather_emb, indices_to_update)\n<\/code><\/pre>\n\n<p>(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9162\" rel=\"noreferrer\">Source<\/a>)<\/p>\n"}]}
{"QuestionId":35805595,"AnswerCount":1,"Tags":"<gpu><tensorflow><cudnn>","CreationDate":"2016-03-04T20:38:23.233","AcceptedAnswerId":null,"OwnerUserId":1104696.0,"Title":"Tensorflow failed to run on k20m and k40m GPUs","Body":"<p>I was trying to run Tensorflow on k40m GPU and got an error:<\/p>\n\n<pre><\/pre>\n\n<p>The issued can be reproduced on k40m and k20m GPU: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/526\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/526<\/a><\/p>\n\n<p>Anyone knows a fix?<\/p>\n\n<p>Detailed log:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37486522","CreationDate":"2016-05-27T14:51:43.600","ParentId":null,"OwnerUserId":"432021","Title":null,"Body":"<p>This is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/526\" rel=\"nofollow\">bug<\/a> in tensorflow.<\/p>\n\n<p>It helps to set driver in EXCLUSIVE_PROCESS mode instead of EXCLUSIVE_THREAD:<\/p>\n\n<pre><code> sudo nvidia-smi -c 3\n<\/code><\/pre>\n"}]}
{"QuestionId":35806105,"AnswerCount":2,"Tags":"<machine-learning><deep-learning><caffe><pycaffe>","CreationDate":"2016-03-04T21:11:50.997","AcceptedAnswerId":35967309.0,"OwnerUserId":3495236.0,"Title":"how can i upgrade my caffe model so it doesn't 'upgrade' every time I use it","Body":"<p>I am using caffe with python(pycaffe).  I am using the prebuilt <a href=\"http:\/\/dl.caffe.berkeleyvision.org\/bvlc_alexnet.caffemodel\" rel=\"nofollow\">alexnet model from model zoo<\/a>. <\/p>\n\n<p>from this page:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet<\/a><\/p>\n\n<p>Every time I use the model, with this code:<\/p>\n\n<pre><\/pre>\n\n<p>caffe tells me the file format is old and it needs to upgrade the file.  Shouldn't this happen only once?<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>how can I properly upgrade the file so that this doesn't happen every single time. <\/p>\n","answers":[{"AnswerId":"46782941","CreationDate":"2017-10-17T05:16:50.077","ParentId":null,"OwnerUserId":"1533823","Title":null,"Body":"<p>Thank you for Shai for your <a href=\"https:\/\/stackoverflow.com\/a\/35967309\/1714410\">help<\/a>.<br>\nHowever, if you are in <strong>Windows<\/strong> <code>upgrade_net_proto_binary<\/code> and <code>upgrade_net_proto_text<\/code> .exe files are in <code>path-to-caffe-master\/caffe\/build\/tools\/Release<\/code>. <\/p>\n\n<p>Hope this will help Windows users<\/p>\n"},{"AnswerId":"35967309","CreationDate":"2016-03-13T06:34:39.393","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When you load the model caffe upgrades your prototxt and binary proto, but does not override the original files you are using. This is why you keep getting this message.  <\/p>\n\n<p>Upgrading is very straight forward. In <code>$CAFFE_ROOT\/build\/tools<\/code> you'll find two binaries: <code>upgrade_net_proto_binary<\/code> and <code>upgrade_net_proto_text<\/code>. Simply apply them to your <code>deploy.prototxt<\/code> and <code>bvlc_alexnet.caffemodel<\/code> and save the results:<\/p>\n\n<pre><code>~$ mv deploy.prototxt deploy_old.prototxt\n~$ mv bvlc_alexnet.caffemodel bvlc_alexnet_old.caffemodel\n~$ $CAFFE_ROOT\/build\/tools\/upgrade_net_proto_text deploy_old.prototx deploy.prototxt\n~$ $CAFFE_ROOT\/build\/tools\/upgrade_net_proto_binary bvlc_alexnet_old.caffemodel bvlc_alexnet.caffemodel\n<\/code><\/pre>\n\n<p>And that's it!<\/p>\n"}]}
{"QuestionId":35810396,"AnswerCount":1,"Tags":"<python><gcc><install><macports><theano>","CreationDate":"2016-03-05T05:19:24.220","AcceptedAnswerId":null,"OwnerUserId":2362907.0,"Title":"Used MacPorts version of gcc and MacPorts version of python","Body":"<p>I am installing theano on Mavericks using macports.\nOne of the installation step is to use macport version of python  and gcc.\nI am not sure where is the path of the macport python.<\/p>\n\n<pre><\/pre>\n\n<p>usually macport save package on under <\/p>\n\n<p>in the .bash_profile file I have these lines<\/p>\n\n<pre><\/pre>\n\n<p>and under , there  is binary file            <\/p>\n\n<p>So what it the version corresponding to Macport?<\/p>\n\n<p><strong>for gcc<\/strong>\nI have the same confusion:<\/p>\n\n<p>Available versions for gcc:\nsudo port select gcc \n    mp-gcc47\n    mp-gcc49 (active)\n    mp-gcc5\n    none\nand<\/p>\n\n<pre><\/pre>\n\n<p>and under  I have a binary file <\/p>\n\n<p>Any explanation?<\/p>\n","answers":[{"AnswerId":"35852619","CreationDate":"2016-03-07T19:45:48.513","ParentId":null,"OwnerUserId":"5162389","Title":null,"Body":"<p>MacPorts adds <code>export PATH=\"\/opt\/local\/bin:\/opt\/local\/sbin:$PATH\"<\/code> to the end of <code>~\/.profile<\/code>, so that <code>\/opt\/local\/bin<\/code> is searched first.<\/p>\n\n<p>Clearly, if you override this by prepending for instance <code>\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/bin<\/code> to the <code>PATH<\/code>, then <code>python<\/code> will be found in that directory first. You can run <code>which -a python<\/code> by the way to show all available python executables.<\/p>\n\n<p>So, one way to \"fix\" this issue is to move the path setting for python into <code>~\/.profile<\/code> above the entry for MacPorts -- or to simply comment it out.<\/p>\n\n<p>I assume there is some similar path issue with <code>gcc<\/code>.<\/p>\n"}]}
{"QuestionId":35810574,"AnswerCount":2,"Tags":"<python><ruby><image-processing><tensorflow><deep-learning>","CreationDate":"2016-03-05T05:43:42.760","AcceptedAnswerId":35811160.0,"OwnerUserId":1031775.0,"Title":"Converting uint8 array to image","Body":"<p>I am trying to convert a  of a 48x48 image to its corresponding 48x48 image file (jpg\/jpeg\/gif). I tried converting the array contents to first binary and then wrote ('wb' mode) it to a file, but that did not work out. <\/p>\n\n<p>Is there a way I can accomplish this?<\/p>\n","answers":[{"AnswerId":"35810800","CreationDate":"2016-03-05T06:13:52.517","ParentId":null,"OwnerUserId":"2861137","Title":null,"Body":"<p>I was able to the same very easily using Octave.\nHere i have tried to generate a random matrix of 48x48 and then i saved the image as jpg format.<\/p>\n\n<pre><code>img = rand(48,48);\nimwrite(img, \"test.jpg\")\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RzAGY.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RzAGY.jpg\" alt=\"Image of 48x48\"><\/a><\/p>\n\n<p>You can save any type of image with this approach.<\/p>\n\n<p>If you can give some more details about what you want to achieve. Do u need to do it just once or u need it as part of program. <\/p>\n\n<p>Hope that helped.<\/p>\n"},{"AnswerId":"35811160","CreationDate":"2016-03-05T06:58:37.830","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you are producing the image in TensorFlow (as I'm inferring from your tag), you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#encode_jpeg\" rel=\"nofollow\"><code>tf.image.encode_jpeg()<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#encode_png\" rel=\"nofollow\"><code>tf.image.encode_png()<\/code><\/a> ops to encode a <code>uint8<\/code> tensor as an image:<\/p>\n\n<pre><code>uint8_data = ...\nimage_data = tf.image.encode_png(uint8_data)\n<\/code><\/pre>\n\n<p>The result of either op is a <code>tf.string<\/code> tensor that you can evaluate and write out to a file.<\/p>\n"}]}
{"QuestionId":35811245,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-03-05T07:09:22.113","AcceptedAnswerId":35816130.0,"OwnerUserId":4127806.0,"Title":"Error using grad in theano","Body":"<p>I am following a tutorial that shows how to implement logistic regression using Theano. The listed line is giving me an error. I don't know how to fix it.<\/p>\n\n<pre><\/pre>\n\n<p>This is the error I am getting:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35816130","CreationDate":"2016-03-05T15:46:02.857","ParentId":null,"OwnerUserId":"557478","Title":null,"Body":"<p>In line <code>grad_err = function([W,TS,E],grad(e,W))<\/code> you want to compute gradient of error 'def_err' w.r.t 'W', but you are passing a function 'e' to grad(..) without the list of inputs, this will never work.\nAlso please note that TS, W, E, O etc are tensor\/symbolic variables which are general expressions and need to be provided with extra input to determine their value.<\/p>\n\n<p>I would recommend going through the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">following tutorial<\/a> for logistic regression, If you have just started Theano then <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/index.html\" rel=\"nofollow\">these tutorials<\/a> will definitely help you to get started.<\/p>\n\n<p>This should work:<\/p>\n\n<pre><code>from theano import tensor, function, grad\n\nTS = tensor.matrix('training-set')\nW = tensor.matrix('weights')\nE = tensor.matrix('expected')\nO = tensor.dot(TS,W)\ndef_err = ((E-O)**2).sum()\ne = function([W,TS,E],def_err)\ngrad_err = function([W,TS,E],grad(def_err,W))\n<\/code><\/pre>\n"}]}
{"QuestionId":35811941,"AnswerCount":1,"Tags":"<python-2.7><anaconda><theano>","CreationDate":"2016-03-05T08:34:44.980","AcceptedAnswerId":null,"OwnerUserId":5961628.0,"Title":"Running Theano on Python 2.7","Body":"<p>New to Theano when I tried to use the package I keep getting the following error: <\/p>\n\n<p>ImportError: ('The following error happened while compiling the node', Dot22(, ), '\\n', 'dlopen(\/Userdir\/.theano\/compiledir_Darwin-14.3.0-x86_64-i386-64bit-i386-2.7.11-64\/tmpEBdQ_0\/eb163660e6e45b373cd7909e14efd44a.so, 2): Library not loaded: libmkl_intel_lp64.dylib\\n  Referenced from: \/Userdir\/.theano\/compiledir_Darwin-14.3.0-x86_64-i386-64bit-i386-2.7.11-64\/tmpEBdQ_0\/eb163660e6e45b373cd7909e14efd44a.so\\n  Reason: image not found', '[Dot22(, )]')<\/p>\n\n<p>Can someone tell me how to fix this issue?  Thanks.<\/p>\n","answers":[{"AnswerId":"42725755","CreationDate":"2017-03-10T18:53:32.567","ParentId":null,"OwnerUserId":"5544720","Title":null,"Body":"<p>get rid of theano and reinstall. If that doesn't work, reinstall all of python<\/p>\n"}]}
{"QuestionId":35812643,"AnswerCount":2,"Tags":"<git><github><caffe><git-fork><remote-branch>","CreationDate":"2016-03-05T09:53:52.993","AcceptedAnswerId":35812709.0,"OwnerUserId":1179925.0,"Title":"How to add fork as branch?","Body":"<p>There are a lot of forks of Caffe, for example MS have their own Caffe <a href=\"https:\/\/github.com\/Microsoft\/caffe\" rel=\"nofollow noreferrer\">fork<\/a>.<\/p>\n\n<p>What I want to do is to compare some fork to master branch of original <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">Caffe<\/a>.<\/p>\n\n<p>I know that using git I can <a href=\"https:\/\/stackoverflow.com\/questions\/9834689\/comparing-two-branches-in-git\">compare 2 branches<\/a>, but my question is how to add fork as branch.<\/p>\n\n<p>I tried <a href=\"https:\/\/stackoverflow.com\/questions\/5884784\/how-to-pull-remote-branch-from-somebody-elses-repo\">this<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>What I'm doing wrong, whick branch name I should specify?<\/p>\n\n<p><strong>Update:<\/strong>\nResult of solution by @hek2mgl <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35812707","CreationDate":"2016-03-05T10:01:13.690","ParentId":null,"OwnerUserId":"171318","Title":null,"Body":"<p><strong>Update<\/strong><\/p>\n\n<p>You are using an ancient version of git. Update to <code>git &gt;= 2<\/code> to make the below commands work. I'm using <code>2.4.10<\/code>.<\/p>\n\n<hr>\n\n<p>The following commands will do the trick:<\/p>\n\n\n\n<pre><code># Clone repo\ngit clone https:\/\/github.com\/BVLC\/caffe.git\ncd caffe\n\n# Add remote\ngit remote add caffe_MS https:\/\/github.com\/Microsoft\/caffe.git\n\n# Fetch remote master from fork\ngit fetch caffe_MS master\n\n# Compare to local master\n# We are currently working in the local master that's why it can be omitted\ngit diff caffe_MS\/master\n<\/code><\/pre>\n"},{"AnswerId":"35812709","CreationDate":"2016-03-05T10:01:36.887","ParentId":null,"OwnerUserId":"641955","Title":null,"Body":"<blockquote>\n  <p>I know that using git I can compare 2 branches, but my question is how to add fork as branch.<\/p>\n<\/blockquote>\n\n<p>When you say you <em>know<\/em> how to compare two branches, I think you really meant you know how to compare <em>local branches<\/em>, and what you're missing is how to compare <em>remote branches<\/em>.<\/p>\n\n<p>First of all, you can see a list of remote branches with:<\/p>\n\n<pre><code>git branch -r\n<\/code><\/pre>\n\n<p>And you can compare remote branches by prefixing with the remote names, for example:<\/p>\n\n<pre><code>git diff origin\/master caffe_MS\/master\n<\/code><\/pre>\n\n<p>As for the last error, it's unclear what you're trying to do with <code>git checkout --track caffe_MS\/MS<\/code>. Your explanation is about comparing branches and adding branches, you never mention tracking. If you want to checkout a branch from a fork, you can use the same syntax as when checkout a branch of origin:<\/p>\n\n<pre><code>git checkout -b caffe_MS_master caffe_MS\/master\n<\/code><\/pre>\n"}]}
{"QuestionId":35814637,"AnswerCount":2,"Tags":"<python><apache><cgi><torch>","CreationDate":"2016-03-05T13:17:25.743","AcceptedAnswerId":35814651.0,"OwnerUserId":6001121.0,"Title":"call external function in python cgi-bin","Body":"<p>I am trying to call a Torch 7 program from within a Python cgi-bin script.<\/p>\n\n<p>If I run the following Python script from the command line:<\/p>\n\n<pre><\/pre>\n\n<p>This works fine and I get the following output<\/p>\n\n<blockquote>\n  <p>ubuntu@ip-172-31-45-110:\/usr\/lib\/cgi-bin$ python test2.py\n  creating an lstm...<\/p>\n  \n  <h2>seeding with \"\u05d0\u05de\u05e8 \u05d4\u05d2\u05d0\u05d5\u05df \u05d4\u05d2\u05e8\u05e4\u05e7\u05d0 \u05d4\u05de\u05df \u05d0\u05d9\u05e9 \u05d8\u05d5\u05d1 \u05d4\u05d9\u05d4 \u05e9\u05e0\u05d0\u05de\u05e8\"<\/h2>\n  \n  <p>\"\u05d0\u05de\u05e8 \u05d4\u05d2\u05d0\u05d5\u05df \u05d4\u05d2\u05e8\u05e4\u05e7\u05d0 \u05d4\u05de\u05df \u05d0\u05d9\u05e9 \u05d8\u05d5\u05d1 \u05d4\u05d9\u05d4 \u05e9\u05e0\u05d0\u05de\u05e8\"\u05d6\u05d4 \u05d9\u05e7\u05e8\u05e7 \u05d9\u05d3\u05d5 \u05de\u05d9 \u05e9\u05db\u05df \u05d1\u05dc\u05d1\u05d3 \u05d5\u05de\u05d7\u05dc\u05dc\u05d9\u05df \u05e2\u05dc\u05d9\u05d5 \u05d0\u05d5\u05ea\u05d5\u05de\u05de\u05e0\u05d5 \u05d5\u05d6\u05db\u05d4 \u05d6\u05d4 \u05dc\u05d0 \u05e1\u05d5\u05dc\u05d3 \u05d3\u05d9\u05e0\u05e8 \u05d4\u05e7\u05d3\u05e9\u05d5 \u05de\u05d7\u05d5\u05e4\u05d4 \u05d0\u05ea \u05d0\u05e9\u05ea\u05d5 \u05de\u05d1\u05d9\u05ea \u05d4\u05e8\u05df \u05d5\u05d1\u05d9\u05e0\u05d9\u05d4\u05df\u05d0\u05d1\u05dc \u05d9\u05d5\u05dd \u05d0\u05d7\u05d3 \u05d5\u05dc\u05de\u05e0\u05d9\u05df \u05e9\u05e0\u05d9\u05d4 \u05d9\u05de\u05d9\u05d1\u05ea\u05df \u05db\u05db\u05dc\u05d5\u05dd \u05d6\u05db\u05d5\u05ea\u05d3\u05db\u05ea\u05d9\u05d1 \u05d0\u05e9\u05e8 \u05d1\u05d7\u05e8\u05dd \u05d0\u05dd \u05d0\u05e9\u05e8 \u05e2\u05e9\u05d5 \u05dc\u05d0 \u05db\u05dc \u05d7\u05d8\u05d0\u05ea \u05e7\u05de\u05d9\u05d4\u05d1\u05d3\u05d9\u05df \u05e0\u05e7\u05e8\u05d0 \u05d5\u05dc\u05d0 \u05de\u05d4\u05d5\u05e6\u05d9\u05d0\u05d4 \u05d5\u05dc\u05d0 \u05e0\u05d9\u05e7\u05d1\u05d4 \u05dc\u05d0 \u05d0\u05de\u05e8 \u05dc\u05d9\u05d4 \u05d0\u05dd \u05ea\u05d9\u05de\u05e6\u05d9\u05dc\u05d5\u05de\u05e8 \u05e2\u05d3 \u05d3\u05db\u05dd \u05e8\u05d0\u05e9\u05d5\u05df \u05d1\u05d2\u05d5\u05d1\u05d4\u05d5 \u05d4\u05e8\u05d9 \u05d6] \u05e0\u05d9\u05d3\u05d5\u05df\u05d8\u05d5\u05de\u05d0\u05d4 \u05de\u05d9\u05d3\u05e7\u05d9\u05e1\u05d0 \u05d0\u05d9 \u05ea\u05dc\u05de\u05d5\u05d3\u05d4 \u05d5\u05de\u05d9 \u05d0\u05d9\u05db\u05d0 \u05dc\u05de\u05d9\u05e4\u05e8\u05da \u05de\u05d4\u05dc\u05e9\u05e8\u05e2\u05ea\u05d5 \u05dc\u05d0\u05d4\u05ea\u05e0\u05d0\u05d5 \u05dc\u05d0 \u05d0\"\u05dc \u05d4\u05db\u05d9 \u05e7\u05d0\u05de\u05e8 \u05d0\u05e4\u05d9\u05dc\u05d5\u05e9\u05d7\u05d9 \u05d5\u05e9\u05d5\u05d1\u05e8\u05d5 \u05e9\u05dc \u05de\u05e7\u05d5\u05dd \u05d4\u05e8\u05d0\u05e9\u05d5\u05df\u05db\u05e1\u05d9\u05e4\u05d5 \u05d5\u05d4\u05e8\u05d9 \u05d7\u05d9\u05dc\u05d9\u05df \u05d1\u05d1\u05d9\u05ea \u05d4\u05de\u05e7\u05d3\u05e9 \u05d5\u05d4\u05d9\u05db\u05d9\u05de\u05de\u05e2\u05d9\u05d8\u05e8\u05d9\u05dd \u05d0\u05d7\u05e8\u05d9\u05e0\u05d0 \u05d5\u05e2\u05d9\u05d1\u05d5\u05e8 \u05d1\u05d9\u05e9\u05e8\u05d0\u05dc \u05de\u05d9 \u05d0\u05d9\u05db\u05d0 \u05de\u05d9\u05d3\u05d9\u05d0\u05d9\u05e9\u05ea\u05d9 \u05de\u05d9\u05e8\u05d7\u05d5 \u05d3\u05e9\u05e7\u05d9\u05dc \u05d0\u05de\u05e8 \u05dc\u05d4\u05d5 \u05d0\u05d9 \u05e9\u05d1\u05e2\u05d9\u05dd \u05e9\u05e0\u05d4 \u05e9\u05e2\u05d5\u05e8\u05d9\u05dd\u05e2\u05dc\u05d9\u05df \u05d0\"\u05db \u05e8\u05d1\u05d9 \u05d9\u05e0\u05d0\u05d9 \u05d1\u05e8 \u05de\u05d9\u05e9\u05e8\u05d0\u05dc\u05d0\u05de\u05e8 \u05dc\u05d5 \u05e0\u05e4\u05e9 \u05d2\u05d6\u05e8 \u05e7\u05d5\u05e0\u05e9\u05d9 \u05d1 \u05ea\u05e8\u05e6\u05d5\u05ea<\/p>\n<\/blockquote>\n\n<p>However if I try calling Torch 7 from a cgi-bin script<\/p>\n\n<pre><\/pre>\n\n<p>Then I get no output at all to the webpage, there is also no error in the apache error log.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"36896419","CreationDate":"2016-04-27T17:07:16.037","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>May be you could try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a> then you can run the torch code directly with python. You can use require(\"sample\") to import your sample.lua module, and then run the inner function just like you run a python function.<\/p>\n"},{"AnswerId":"35814651","CreationDate":"2016-03-05T13:18:56.097","ParentId":null,"OwnerUserId":"1832539","Title":null,"Body":"<p>You want to use <a href=\"https:\/\/docs.python.org\/3\/library\/subprocess.html#subprocess.check_output\" rel=\"nofollow\">check_output<\/a> to store the output of the command you are executing. <a href=\"https:\/\/docs.python.org\/3\/library\/subprocess.html#subprocess.call\" rel=\"nofollow\">call<\/a> will not do this. Call will only give you the return code of what you are executing.  <\/p>\n\n<pre><code>subprocess.check_output(['th', 'sample.lua', 'cv\/lm_lstm_epoch3.54_0.9324.t7', '-gpuid', '-1', '-primetext', '\"\u05d0\u05de\u05e8 \u05d4\u05d2\u05d0\u05d5\u05df \u05d4\u05d2\u05e8\u05e4\u05e7\u05d0 \u05d4\u05de\u05df \u05d0\u05d9\u05e9 \u05d8\u05d5\u05d1 \u05d4\u05d9\u05d4 \u05e9\u05e0\u05d0\u05de\u05e8\"', '-temperature', '1.0', '-length', '1000'])\n<\/code><\/pre>\n"}]}
{"QuestionId":35815145,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-03-05T14:12:25.700","AcceptedAnswerId":35831824.0,"OwnerUserId":1245262.0,"Title":"When is it advantageous to use theano's scan function","Body":"<p>I've been reading the Theano documentation on scan and find myself confused by two, seemingly, contradictory statements. <\/p>\n\n<p>On <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/loop.html#scan\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/loop.html#scan<\/a>, one of the advantages of scan is listed as:<\/p>\n\n<pre><\/pre>\n\n<p>But, on <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#lib-scan\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#lib-scan<\/a>, in a section on optimizing the use of scan, it says:<\/p>\n\n<pre><\/pre>\n\n<p>My reading of 'performance', here, is as a synonym for speed. So, I'm left confused as to when\/if scan will lead to shorter runtimes, once compilation has been completed.<\/p>\n","answers":[{"AnswerId":"35831824","CreationDate":"2016-03-06T19:59:30.280","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>If your expression intrinsically needs a for-loop, then you sometimes have two options:<\/p>\n\n<ol>\n<li>Build the expression using a python for loop<\/li>\n<li>Build the expression using scan<\/li>\n<\/ol>\n\n<p>Option 1 only works if you know in advance the length of your for-loop. It can happen that the length of your for-loop depends on a symbolic variable that is not available at script-writing time. In that case you need to use scan. Although oftentimes you can formulate the problem either way (see the absence of scan in tensorflow).<\/p>\n\n<p>As for time performance, there have been a number of results showing that it really depends on the problem which one is faster.<\/p>\n"}]}
{"QuestionId":35815170,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-03-05T14:15:17.417","AcceptedAnswerId":null,"OwnerUserId":5473359.0,"Title":"Get label prediction from Cifar-10 model","Body":"<p>I'm currently working on the Cifar-10 tutorial of tensorflow. I'd like to change the evaluation such that I can see for each image what the prediction of my model was, and whether it was true\/false. I struggle with the first part: if I print the predictions (sess.run([top_k_op])) I get true\/false values which I assume are whether the prediction was correct or not. However, if I try to print the actual prediction (I tried so far to print the logits, and print the top_k_op tensor), I get some numbers or values, but nothing that looks like the labels. What do I have to change about my code to actually see the labels that my model predicted?<\/p>\n","answers":[{"AnswerId":"35821483","CreationDate":"2016-03-06T00:07:20.297","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You want to evaluate first the <code>logits<\/code>. This is a probability distribution over your classes out of your network. The index of the tensor with the higher value will give you the most likely class for your label.<\/p>\n\n<p>you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#argmax\" rel=\"nofollow\">tf.argmax<\/a> to get the index and then use the index in your labels to print it out<\/p>\n\n<pre><code>print labels[index]\n<\/code><\/pre>\n"},{"AnswerId":"35947850","CreationDate":"2016-03-11T18:56:17.023","ParentId":null,"OwnerUserId":"5577321","Title":null,"Body":"<p>You can figure out an answer by looking <a href=\"https:\/\/github.com\/jkschin\/svhn\" rel=\"nofollow noreferrer\" title=\"here\">here<\/a><\/p>\n\n<p>In svhn.py, at line 116 the predicted label is printed: <code>print (step, int(test_labels[0]))<\/code><\/p>\n\n<p>I did it in a clear way by using: <\/p>\n\n<pre><code>classification = sess.run(top_k_predict_op)\nprint (step, int(test_labels[0]))\nprint \"network predicted:\", classification[0], \"for real label:\", test_labels\n<\/code><\/pre>\n\n<p>Be sure that you are predicting on 24*24 images, in case you trained your model with the original version of the TensorFlow CIFAR-10 model.<\/p>\n"}]}
{"QuestionId":35819313,"AnswerCount":1,"Tags":"<c++><macos><osx-yosemite><deep-learning><caffe>","CreationDate":"2016-03-05T20:13:43.977","AcceptedAnswerId":null,"OwnerUserId":3610694.0,"Title":"Error compiling caffe classification","Body":"<p>I'm trying to compile a caffe classification example on OS X 10.10.5, with the following command:<\/p>\n\n<pre><\/pre>\n\n<p>But I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on what the problem is?<\/p>\n","answers":[{"AnswerId":"35820690","CreationDate":"2016-03-05T22:25:44.280","ParentId":null,"OwnerUserId":"3610694","Title":null,"Body":"<p>Oh my goodness I forgot to link Caffe... Well that problem is solved.<\/p>\n\n<p>Edit: Here is the command I used in case anyone needs it<\/p>\n\n<pre><code>g++ -ggdb `pkg-config --cflags --libs opencv` -I \/Users\/SullyChen\/caffe\/include -I \/usr\/local\/cuda\/include -I \/usr\/local\/Cellar\/openblas\/0.2.15\/include -lboost_system -lglog -lprotobuf -L \/Users\/SullyChen\/caffe\/.build_release\/lib -lcaffe classification.cpp -o classification\n<\/code><\/pre>\n"}]}
{"QuestionId":35819407,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2016-03-05T20:21:01.577","AcceptedAnswerId":35821373.0,"OwnerUserId":5540330.0,"Title":"A Simple Network on TensorFlow","Body":"<p>I was trying to train a very simple model on TensorFlow. Model takes a single float as input and returns the probability of input being greater than 0. I used 1 hidden layer with 10 hidden units. Full code is shown below:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>The problem is, I am getting terribly unrelated results. Model does not learn anything and returns irrelevant probabilities. I tried to adjust learning rate and change variable initialization, but I did not get anything useful. Do you have any suggestions?<\/p>\n","answers":[{"AnswerId":"35821373","CreationDate":"2016-03-05T23:52:16.767","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You are computing only one probability what you want is to have two classes:  <\/p>\n\n<ul>\n<li>greater\/equal than zero.<\/li>\n<li>less than zero. <\/li>\n<\/ul>\n\n<p>So the output of the network will be a tensor of shape two that will contain the probabilities of each class. I renamed y_ in your example to <code>labels<\/code>:<\/p>\n\n<pre><code>labels = tf.placeholder(tf.float32, shape = [None,2])\n<\/code><\/pre>\n\n<p>Next we compute the cross entropy between the result of the network and the expected classification. The classes for positive numbers would be <code>[1.0, 0]<\/code> and for negative numbers would be <code>[0.0, 1.0]<\/code>.\nThe loss function becomes:<\/p>\n\n<pre><code>cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, labels)\nloss = tf.reduce_mean(cross_entropy)\n<\/code><\/pre>\n\n<p>I renamed the <code>y<\/code> to <code>logits<\/code> as that is a more descriptive name. <\/p>\n\n<p>Training this network for 10000 steps gives:<\/p>\n\n<pre><code>Give an x\n3.0\n[[ 0.96353203  0.03686807]]\nGive an x\n200\n[[ 0.97816485  0.02264325]]\nGive an x\n-20\n[[ 0.12095013  0.87537241]]\n<\/code><\/pre>\n\n<p>Full code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport random\n\n# Graph construction\n\nx = tf.placeholder(tf.float32, shape = [None,1])\nlabels = tf.placeholder(tf.float32, shape = [None,2])\n\nW = tf.Variable(tf.random_uniform([1,10],0.,0.1))\nb = tf.Variable(tf.random_uniform([10],0.,0.1))\n\nlayer1 = tf.nn.sigmoid( tf.add(tf.matmul(x,W), b) )\n\nW1 = tf.Variable(tf.random_uniform([10, 2],0.,0.1))\nb1 = tf.Variable(tf.random_uniform([1],0.,0.1))\n\nlogits = tf.nn.sigmoid( tf.add( tf.matmul(layer1,W1),b1) )\n\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, labels)\n\nloss = tf.reduce_mean(cross_entropy)\n\ntrain_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\n\n# Training\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    N = 1000\n    while N != 0:\n        batch = ([],[])\n        u = random.uniform(-10.0,+10.0)\n        if u &gt;= 0.:\n            batch[0].append([u])\n            batch[1].append([1.0, 0.0])\n        if  u &lt; 0.:\n            batch[0].append([u])\n            batch[1].append([0.0, 1.0])\n\n        sess.run(train_step, feed_dict = {x : batch[0] , labels : batch[1]} )\n\n        N -= 1\n\n    while(True):\n        u = raw_input(\"Give an x\\n\")\n        print sess.run(logits, feed_dict = {x : [[u]]})\n<\/code><\/pre>\n"}]}
{"QuestionId":35819923,"AnswerCount":0,"Tags":"<nvidia><caffe><cudnn>","CreationDate":"2016-03-05T21:08:25.070","AcceptedAnswerId":null,"OwnerUserId":4935271.0,"Title":"Caffe Performance with cuDNN v2 on the Jetson TK1","Body":"<p>I am using Caffe with cuDNN v2 on the Nvidia Jetson TK1.  However, I am not seeing a hardly any performance improvement with linking to cuDNN v2 over just using CUDA 6.5 without cuDNN.  The Nvidia website states that a 1.5x improvement is typical for cuDNN.  I am seeing perhaps a 1.05x improvement in forward propagation time.  Has anyone seen a better performance improvement using cuDNN and Caffe on the TK1 ? <\/p>\n","answers":[]}
{"QuestionId":35820878,"AnswerCount":1,"Tags":"<macos><opencl><caffe><amd-processor>","CreationDate":"2016-03-05T22:49:34.917","AcceptedAnswerId":35821346.0,"OwnerUserId":5500368.0,"Title":"Mac OSX 10.10 AMD APP SDK Location","Body":"<p>I have spent the past hour looking for the AMD APP SDK location for Mac OSX 10.10 and cannot find it for the life of me. I don't know the file name even so I can't just search for it and searching SDK doesn't work. <\/p>\n\n<p>I know I have it because I'm on a Mac Pro with AMD cards. <\/p>\n\n<p>I'm trying to set up an OpenCL version of Caffe and need a reference to my SDK for it to compile (<a href=\"https:\/\/github.com\/amd\/OpenCL-caffe\/wiki\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amd\/OpenCL-caffe\/wiki<\/a>). Because I'm not able to set the  I'm getting a compile error <\/p>\n\n<p>I've looked in my usr and found nothing. I looked for this comment <a href=\"https:\/\/stackoverflow.com\/a\/5795305\/5500368\">on where the openCL SDK would be<\/a> and found nothing there, I don't even have a Developer folder as far as I can tell... much less a GPU Computing folder.<\/p>\n\n<p>Help?<\/p>\n","answers":[{"AnswerId":"35821346","CreationDate":"2016-03-05T23:49:45.973","ParentId":null,"OwnerUserId":"3100482","Title":null,"Body":"<blockquote>\n  <p>I know I have it because I'm on a Mac Pro with AMD cards.<\/p>\n<\/blockquote>\n\n<p>If you are running OS X, you neither have nor need the AMD APP SDK.<\/p>\n\n<p>I've just built <code>OpenCL-Caffe<\/code> from the GitHub repository on my own MacBook (10.11). The only configuration options I had to pass to CMake were to tell it where <code>clBLAS<\/code> was installed.<\/p>\n\n<p>The only changes I needed to make to the source code were to replace <code>#include &lt;CL\/...&gt;<\/code> with <code>#include &lt;OpenCL\/...&gt;<\/code> in these files:<\/p>\n\n<pre><code>include\/caffe\/common.hpp\ninclude\/caffe\/util\/math_functions.hpp\n<\/code><\/pre>\n\n<p>and to remove <code>#include &lt;malloc.h&gt;<\/code> from <code>src\/caffe\/device.cpp<\/code>.<\/p>\n"}]}
{"QuestionId":35820929,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2016-03-05T22:54:40.313","AcceptedAnswerId":null,"OwnerUserId":234167.0,"Title":"Why no weight decay on the convolutional layers in the cifar10 example of tensorflow?","Body":"\n\n<p>There seems to be no weight decay on convolutional layers in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/deep_cnn\/index.html\" rel=\"nofollow noreferrer\">cifar10<\/a> example on tensorflow. Actually there is no weight decay on any layers except for the two fully connected layers. Is this a common practice? I thought weight decay was applied to all weights (except biases).<\/p>\n\n<p>For reference, here's the relevant code ( is the weight decay factor):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"46939842","CreationDate":"2017-10-25T18:40:15.787","ParentId":null,"OwnerUserId":"4880003","Title":null,"Body":"<p>Weight decay doesn't necessarily improve performance. In my own experience, I've found reasonably often that my models perform worse (as measured by some metric on a held-out set) with any significant amount of weight decay. It is a useful form of regularization to be aware of, but you don't need to add it to every model without considering if it seems needed or comparing the performance with and without.<\/p>\n\n<p>As for whether weight decay on only part of a model can be good compared with weight decay on the entire model, it does seem less common to only regularize some of the weights this way. I don't know that there's a theoretical reason for this, however. In general, neural networks already have too many hyperparameters to configure. Whether to use weight decay or not is already a question, and how strongly to regularize the weights if you do. If you also wonder, which layers should I regularize this way, you'll quickly run out of time to test the performance of all of the different ways you could turn it on and off for each layer.<\/p>\n\n<p>I imagine that there are models that would benefit from weight decay on only part of the model; I don't think it's done often because it's difficult to test all of the possibilities and find out which one works best.<\/p>\n"}]}
{"QuestionId":35822036,"AnswerCount":1,"Tags":"<github><caffe><cudnn>","CreationDate":"2016-03-06T01:31:27.230","AcceptedAnswerId":35823669.0,"OwnerUserId":4935271.0,"Title":"Previous version of Caffe compatible with cuDNN v2","Body":"<p>I am trying to locate the last version of Caffe that was still compatible with cuDNN v2.  The current master requires cuDNN v3.  I've been searching through the github repository for Caffe trying to locate what I need but there doesn't seem to be a clear way of doing this.  Is there a relatively painless way of finding this on github ?<\/p>\n","answers":[{"AnswerId":"35823669","CreationDate":"2016-03-06T05:55:53.260","ParentId":null,"OwnerUserId":"2297751","Title":null,"Body":"<p>Here is a list of <a href=\"https:\/\/github.com\/BVLC\/caffe\/releases\" rel=\"nofollow\">Caffe release<\/a>, and <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/1849\" rel=\"nofollow\">rc2<\/a> is the one you are looking for.<\/p>\n"}]}
{"QuestionId":35822251,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><tensorflow>","CreationDate":"2016-03-06T02:07:38.590","AcceptedAnswerId":null,"OwnerUserId":3314143.0,"Title":"TensorFlow: one network, two GPUs?","Body":"<p>I have a convolutional neural network with two different output streams:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to compute stream 1 on  and stream 2 on . Unfortunately I were not able to set it up properly.<\/p>\n\n<p>This attempt:<\/p>\n\n<pre><\/pre>\n\n<p>Runs dead slow (slower than training on 1 GPU solely) and sometimes produces NaN values in the output. I thought this might be because the  statements may not be synchronized properly. So I added  and placed the conv layers on  explicitly:<\/p>\n\n<pre><\/pre>\n\n<p>...but with this approach the network isn't even running. No matter what I've tried, it complained about the input not being initialized:<\/p>\n\n<pre><\/pre>\n\n<p>Without the  statements the network is training on  only and runs fine - trains reasonable stuff, no errors.<\/p>\n\n<p>What am I doing wrong? Is TensorFlow not able to split different streams of layers in one network to different GPUs? Do I always <em>have to<\/em> split the <em>complete<\/em> network in different towers?<\/p>\n","answers":[{"AnswerId":"35824398","CreationDate":"2016-03-06T07:44:12.993","ParentId":null,"OwnerUserId":"6024628","Title":null,"Body":"<p>There is an example of how to use many gpus on one network \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py<\/a>\nprobably you can copy the code.\nAlso can get something like this<\/p>\n\n<pre><code># Creates a graph.\nc = []\nfor d in ['\/gpu:2', '\/gpu:3']:\nwith tf.device(d):\n   a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3])\n   b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2])\n   c.append(tf.matmul(a, b))\nwith tf.device('\/cpu:0'):\nsum = tf.add_n(c)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint sess.run(sum)\n<\/code><\/pre>\n\n<p>Looking at: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/using_gpu\/index.html#using-multiple-gpus\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/using_gpu\/index.html#using-multiple-gpus<\/a><\/p>\n\n<p>Best Regards<\/p>\n"}]}
{"QuestionId":35823021,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2016-03-06T04:12:59.533","AcceptedAnswerId":null,"OwnerUserId":3217504.0,"Title":"Asking for the installation of Caffe","Body":"<p>I was installing the Caffe library on Mac OS, but when I type 'make run test', I encountered the following problem.  What should I do? Thanks in advance. My macbook doesn't contain Cudas, does this affect the installation?<\/p>\n\n<hr>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35823300","CreationDate":"2016-03-06T04:56:43.220","ParentId":null,"OwnerUserId":"3217504","Title":null,"Body":"<p>Finally, I find a solution by setting CPU_ONLY := 1 in Makefile.config (uncomment the original one by removing the '#' in line \"CPU_ONLY := 1 in Makefile.config\") and rerun the command\"make clean\", \"make all\", then \"make test\", then \"make runtest\" referring to this link - <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/736\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/issues\/736<\/a><\/p>\n"},{"AnswerId":"36049569","CreationDate":"2016-03-17T00:33:51.747","ParentId":null,"OwnerUserId":"824979","Title":null,"Body":"<p>I had the same issue. But i have a Graphic card specifically to run Caffe on it, so CPU_ONLY was not an option ;-)<\/p>\n\n<p>To check if it's the same cause that mine, try to run CUDA Samples <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-getting-started-guide-for-linux\/index.html#running-binaries\" rel=\"noreferrer\">deviceQuery example<\/a><\/p>\n\n<p>I fixed using <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/index.html#runfile-verifications\" rel=\"noreferrer\">CUDA Guide runfile verifications<\/a><\/p>\n\n<pre><code>sudo chmod 0666 \/dev\/nvidia*\n<\/code><\/pre>\n"}]}
{"QuestionId":35823165,"AnswerCount":2,"Tags":"<neural-network><theano><deep-learning>","CreationDate":"2016-03-06T04:37:19.637","AcceptedAnswerId":35859464.0,"OwnerUserId":4571055.0,"Title":"What is purpose of ifelse and switch in theano?","Body":"<p>I am not able to understand anything about them, I read it from <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/conditions.html<\/a>. I suppose they function like our ifelse but not able to understand how to use them. Please some guide on this. Thanks<\/p>\n","answers":[{"AnswerId":"37260894","CreationDate":"2016-05-16T18:35:34.630","ParentId":null,"OwnerUserId":"4557429","Title":null,"Body":"<p>The code is copied from the link to be explained easier:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/graphstructures.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/extending\/graphstructures.html<\/a><\/p>\n\n<pre><code>from theano import tensor as T\nfrom theano.ifelse import ifelse\nimport theano, time, numpy\n\na,b = T.scalars('a','b')\nx,y = T.matrices('x','y')\n\nz_switch = T.switch(T.lt(a,b), T.mean(x), T.mean(y))\nz_lazy = ifelse(T.lt(a,b), T.mean(x), T.mean(y))\n\nf_switch = theano.function([a,b,x,y], z_switch,\n                           mode=theano.Mode(linker='vm'))\nf_lazyifelse = theano.function([a,b,x,y], z_lazy,\n                               mode=theano.Mode(linker='vm'))\n\nval1 = 0.\nval2 = 1.\nbig_mat1 = numpy.ones((10000,1000))\nbig_mat2 = numpy.ones((10000,1000))\n\nn_times = 10\n\ntic = time.clock()\nfor i in range(n_times):\n    f_switch(val1, val2, big_mat1, big_mat2)\nprint('time spent evaluating both values %f sec' % (time.clock()-tic))\n\ntic = time.clock()\nfor i in range(n_times):\n    f_lazyifelse(val1, val2, big_mat1, big_mat2)\nprint('time spent evaluating one value %f sec' % (time.clock()-tic))\n<\/code><\/pre>\n\n<p>In the example, both f_switch(val1, val2, big_mat1, big_mat2) and f_lazyifelse(val1, val2, big_mat1, big_mat2) are doing same thing. If val1 is \"Less Than\" val2, it returns big_mat1 otherwise big_mat2.<\/p>\n\n<p>However there is a difference between \"switch\" and \"ifelse\" in performance. In \"ifelse\" depending to the condition, only one of the possible outputs will be evaluated. In this example, in f_lazyifelse(val1, val2, big_mat1, big_mat2) , if val1 is less than val2 then only big_mat1 will be evaluated which is performing mean function on all elements of a big matrix, and big_mat2 will not be evaluated. <\/p>\n\n<p>On the other hand, in case of \"switch\" regardless of the condition both the possible outputs will be evaluated, and in the example, in f_switch(val1, val2, big_mat1, big_mat2), regardless of val1 and val2 both big_mat1 and big_mat2 will be evaluated (here, performing mean function on all elements of a big matrix). Therefore \"switch\" is almost takes two times of \"ifelse\" to perform.<\/p>\n"},{"AnswerId":"35859464","CreationDate":"2016-03-08T05:10:06.740","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>I would think of it as just another operator that acts on three symbolic variables, if the first is true, return the second, else return the third.  <\/p>\n\n<p>But for many operators (like <code>-<\/code> and <code>+<\/code>) theano has overloaded them for symbolic variables, so probably you don't feel the difference.  <\/p>\n\n<p>For example, if <code>a<\/code> and <code>b<\/code> are numbers, then <code>c=a+b<\/code> creates a variable <code>c<\/code> with the value of <code>a+b<\/code>. If <code>a<\/code> and <code>b<\/code> are symbolic variables, then <code>c=a+b<\/code> creates another symbolic variable <code>c<\/code>, that will apply (element-wise) addition to <code>a<\/code> and <code>b<\/code> when the corresponded function gets called\/evaluated.<\/p>\n\n<p>Here's an introduction on theano operators and graphs.\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/graphstructures.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/extending\/graphstructures.html<\/a><\/p>\n\n<p>The <a href=\"http:\/\/northisup.com\/blog\/the-ternary-operator-in-python\/\" rel=\"nofollow noreferrer\">ternary operator<\/a> is python's native equivalent to theano's <code>ifelse<\/code> or <code>switch<\/code> for symbolic variables in some sense.<\/p>\n"}]}
{"QuestionId":35823311,"AnswerCount":0,"Tags":"<python><numpy><theano>","CreationDate":"2016-03-06T04:58:04.153","AcceptedAnswerId":null,"OwnerUserId":2877599.0,"Title":"Theano shared variable vs numpy array speed in theano function","Body":"<p>I'm learning about theano, and trying to implement a modular ML tool with it.<\/p>\n\n<p>I've a simple linear combination with:<\/p>\n\n<pre><\/pre>\n\n<p>but now I'm a little confused about the best way to describe the  as a theano function. I came up with two functions:<\/p>\n\n<pre><\/pre>\n\n<p>where  is a shared variable. <\/p>\n\n<p>My goal is to run it on GPUs, so I'd like to use shared variables as much as possible.<\/p>\n\n<p>The problem is that, if I use the  function, I cannot use a shared variable in the input. And when using the , the only way I got it working is reading as numpy arrays and constantly updating the  with . (<a href=\"https:\/\/stackoverflow.com\/a\/17296524\/2877599\">https:\/\/stackoverflow.com\/a\/17296524\/2877599<\/a>)<\/p>\n\n<p>My question is: What is the best,  or  to run on gpus ?<\/p>\n","answers":[]}
{"QuestionId":35823709,"AnswerCount":2,"Tags":"<caffe><pycaffe>","CreationDate":"2016-03-06T05:59:51.203","AcceptedAnswerId":null,"OwnerUserId":4159177.0,"Title":"I \"import caffe\" from ipython, but I got \"RuntimeWarning\". How to resolve it?","Body":"<p>I have read the article \"Ubuntu Installation --Guide for Ubuntu 14.04 with a 64 bit processor.\" from Github website (<a href=\"https:\/\/github.com\/tiangolo\/caffe\/blob\/ubuntu-tutorial-b\/docs\/install_apt2.md\" rel=\"nofollow\">https:\/\/github.com\/tiangolo\/caffe\/blob\/ubuntu-tutorial-b\/docs\/install_apt2.md<\/a>).\nAnd now, I open IPython to test that PyCaffe is working. I input \"ipython\" command, and enter to the ipython page.<\/p>\n\n<p>Then, I input the command \"import caffe\", but I got below warnings: <\/p>\n\n<blockquote>\n  <p>\/root\/code\/caffe-master\/python\/caffe\/pycaffe.py:13:RuntimeWarning: to-Python converter for boost::shared_ptr > already registered; second conversion method ignored. from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\<\/p>\n  \n  <p>\/root\/code\/caffe-master\/python\/caffe\/pycaffe.py:13: RuntimeWarning: to-Python converter for boost::shared_ptr > already registered; second conversion method ignored. from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\<\/p>\n  \n  <p>\/root\/code\/caffe-master\/python\/caffe\/pycaffe.py:13: RuntimeWarning: to-Python converter for boost::shared_ptr > already registered; second conversion method ignored. from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\<\/p>\n<\/blockquote>\n\n<p>Would you please help to give me some suggestions to resolve it?<\/p>\n","answers":[{"AnswerId":"36224242","CreationDate":"2016-03-25T16:49:59.947","ParentId":null,"OwnerUserId":"5837113","Title":null,"Body":"<p>You can edit \/caffe\/python\/caffe\/_caffe.cpp . There are four places need to change,like this<\/p>\n\n<pre><code>bp::register_ptr_to_python&lt;shared_ptr&lt;Layer&lt;Dtype&gt; &gt; &gt;();\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>const boost::python::type_info cinfo = boost::python::type_id&lt;shared_ptr&lt;Blob&lt;Dtype&gt; &gt; &gt;();\nconst boost::python::converter::registration* creg = boost::python::converter::registry::query(cinfo);\n      if (creg == NULL)\n      {\n         bp::register_ptr_to_python&lt;shared_ptr&lt;Layer&lt;Dtype&gt; &gt; &gt;();\n      }\n      else if ((*creg).m_to_python == NULL) {\n         bp::register_ptr_to_python&lt;shared_ptr&lt;Layer&lt;Dtype&gt; &gt; &gt;();\n      }\n<\/code><\/pre>\n\n<p>then compile use 'make pycaffe' ,it will be ok.<\/p>\n"},{"AnswerId":"35834148","CreationDate":"2016-03-06T23:58:27.090","ParentId":null,"OwnerUserId":"1168319","Title":null,"Body":"<p>I found this:<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/C_air48cISU\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/C_air48cISU<\/a><\/p>\n\n<p>Claiming that this is a non-error, cause by mis-matched versions of Boost.  You can safely ignore it.  They've promised to clean up the warning (at some point not yet specified)<\/p>\n"}]}
{"QuestionId":35824742,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-06T08:32:41.257","AcceptedAnswerId":35909109.0,"OwnerUserId":6024628.0,"Title":"Tensorflow distributed passing devices","Body":"<p>I recently installed the version of tensorflow for distributed processing.\nFrom the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/distributed_runtime\" rel=\"nofollow\">trend<\/a>, I tried to implement with multiple gpus on multiple computers, and also found a <a href=\"https:\/\/stackoverflow.com\/questions\/34439045\/tensorflow-setup-for-distributed-computing\">white paper<\/a> for some additional specifications.\nI can run the server and a worker on 2 different computers with 2 and 1 gpus respectively, and using the session grpc, allocate and run the program on remote or local mode.<\/p>\n\n<p>I ran locally tensorflow in the remote computer with: <\/p>\n\n<pre><\/pre>\n\n<p>and for using on the server <\/p>\n\n<pre><\/pre>\n\n<p>However, when i try to specify the device for running on 2 computers at the same time the python show me the error:<\/p>\n\n<pre><\/pre>\n\n<p>when i use<\/p>\n\n<pre><\/pre>\n\n<p>or even changin the name of job. So I am wondering if it is required to <a href=\"https:\/\/stackoverflow.com\/questions\/35213172\/add-a-new-device-in-tensorflow\">Add a New Device<\/a> or probably I am doing something wrong with the initialization of the cluster.<\/p>\n","answers":[{"AnswerId":"35909109","CreationDate":"2016-03-10T06:24:42.497","ParentId":null,"OwnerUserId":"5212207","Title":null,"Body":"<p>The worker is really the name of the cluster. <\/p>\n\n<p>Your first bazel call should be like this:<\/p>\n\n<p><code>bazel-bin\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_tensorflow_server \\\n--cluster_spec='worker|192.168.170.193:2500;192.168.170.226:2501' --job_name=worker --task_id=0 &amp;<\/code><\/p>\n\n<p>Which you run on the first node, 192.168.170.193<\/p>\n\n<p>Your cluster name is worker, which includes the IP addresses of the two nodes.  The task then refers to the the two running nodes.  You must start the protocol on both nodes, specifying the different task ID for each node, ie. then run:<\/p>\n\n<pre><code>bazel-bin\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_tensorflow_server \\\n--cluster_spec='worker|192.168.170.193:2500;192.168.170.226:2501' --job_name=worker --task_id=1 &amp;`\n<\/code><\/pre>\n\n<p>on your second node, 192.168.170.226<\/p>\n\n<p>then run:<\/p>\n\n<pre><code>with tf.device(\"\/job:worker\/task:0\/device:gpu:0\"):\n  x = tf.placeholder(tf.float32, [None, 784], name='x-input')\n  W = tf.Variable(tf.zeros([784, 10]), name='weights')\nwith tf.device(\"\/job:worker\/task:0\/device:gpu:1\"):\n  b = tf.Variable(tf.zeros([10], name='bias'))\n# Use a name scope to organize nodes in the graph visualizer\nwith tf.device(\"\/job:worker\/task:1\/device:gpu:0\"):\n  with tf.name_scope('Wx_b'):\n    y = tf.nn.softmax(tf.matmul(x, W) + b)\n<\/code><\/pre>\n"}]}
{"QuestionId":35824798,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-03-06T08:39:05.443","AcceptedAnswerId":35833270.0,"OwnerUserId":6024751.0,"Title":"Tensorflow error using tf.image.random : 'numpy.ndarray' object has no attribute 'get_shape'","Body":"<p><strong>Intro<\/strong><\/p>\n\n<p>I am using a modified version of the Tensorflow tutorial \"Deep MNIST for experts\" with the Python API for a medical images classification project using convolutionnal networks.<\/p>\n\n<p>I want to artificially increase the size of my training set by applying random modifications on the images of my training set.<\/p>\n\n<p><strong>Problem<\/strong><\/p>\n\n<p>When I run the line : <\/p>\n\n<pre><\/pre>\n\n<p>I get de following error :<\/p>\n\n<blockquote>\n  <p>AttributeError: 'numpy.ndarray' object has no attribute 'get_shape'<\/p>\n<\/blockquote>\n\n<p>My Tensor \"images\" is an ndarray  of \"batch\" ndarrays .<\/p>\n\n<p>Just to check if my input data was packed in the right shape and type, I have tried to apply this simple function in the (not modified) tutorial \"Tensorflow Mechanics 101\" and I get the same error.<\/p>\n\n<p>Finally, I still get the same error trying to use the following functions :<\/p>\n\n<ul>\n<li> <\/li>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<p><strong>Questions<\/strong><\/p>\n\n<p>As input data is usually carried in Tensorflow as ndarrays, I would like to know :   <\/p>\n\n<ol>\n<li>Is it a bug of Tensorflow Python API or is it my \"fault\" because\nof the type\/shape of my input data?<\/li>\n<li>How could I get it to work and be able to apply     to my training set?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"35833270","CreationDate":"2016-03-06T22:13:15.323","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This seems like an inconsistency in the TensorFlow API, since almost all other op functions accept NumPy arrays wherever a <code>tf.Tensor<\/code> is expected. I've filed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1409\" rel=\"noreferrer\">an issue<\/a> to track the fix.<\/p>\n\n<p>Fortunately, there is a simple workaround, using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#convert_to_tensor\" rel=\"noreferrer\"><code>tf.convert_to_tensor()<\/code><\/a>. Replace your code with the following:<\/p>\n\n<pre><code>flipped_images = tf.image.random_flip_left_right(tf.convert_to_tensor(images))\n<\/code><\/pre>\n"}]}
{"QuestionId":35826648,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-06T12:06:23.310","AcceptedAnswerId":null,"OwnerUserId":2886263.0,"Title":"About tensorflow graph: what am I wrong with this program?","Body":"<pre><\/pre>\n\n<p>output error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41391105","CreationDate":"2016-12-30T05:29:56.913","ParentId":null,"OwnerUserId":"4259486","Title":null,"Body":"<p>Another option is to initialize graph variable first:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n<\/code><\/pre>\n\n<p>and than pass it to your session:<\/p>\n\n<pre><code>with tf.Session(graph=graph) as session:\n<\/code><\/pre>\n"},{"AnswerId":"35832034","CreationDate":"2016-03-06T20:19:34.127","ParentId":null,"OwnerUserId":"234167","Title":null,"Body":"<p>The error is actually giving you a tip of why it's failing. When you started the session, you are actually using a different graph than the one referenced by <code>res<\/code>. The easiest solution is to simply move everything inside <code>with tf.Graph().as_default():<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef activation(e, f, g):\n  return e + f + g\n\nwith tf.Graph().as_default():\n  a = tf.constant([5, 4, 5], name='a')\n  b = tf.constant([0, 1, 2], name='b')\n  c = tf.constant([5, 0, 5], name='c')\n  res = activation(a, b, c)\n  init = tf.initialize_all_variables()\n\n  with tf.Session() as sess:\n    # Start running operations on the Graph.\n    merged = tf.merge_all_summaries()\n    sess.run(init)\n    hi = sess.run(res)\n    print hi\n    writer = tf.train.SummaryWriter(\"\/tmp\/basic\", sess.graph_def)\n<\/code><\/pre>\n\n<p>Alternatively you can also just remove the line <code>with tf.Graph().as_default():<\/code> and then it will also worked as expected.<\/p>\n"}]}
{"QuestionId":35828037,"AnswerCount":1,"Tags":"<nlp><neural-network><tensorflow>","CreationDate":"2016-03-06T14:25:26.467","AcceptedAnswerId":35898025.0,"OwnerUserId":524283.0,"Title":"Training a CNN with pre-trained word embeddings is very slow (TensorFlow)","Body":"<p>I'm using TensorFlow (0.6) to train a CNN on text data. I'm using a method similar to the second option specified in <a href=\"https:\/\/stackoverflow.com\/questions\/35687678\/using-a-pre-trained-word-embedding-word2vec-or-glove-in-tensorflow\">this SO thread<\/a> (with the exception that the embeddings are trainable). My dataset is pretty small and the vocabulary is around 12,000 words. When I train using random word embeddings everything works nicely. However, when I switch to <a href=\"https:\/\/code.google.com\/archive\/p\/word2vec\/\" rel=\"nofollow noreferrer\">the pre-trained embeddings from the word2vec site<\/a>, the vocabulary grows to over 3,000,000 words and training iterations become over 100 times slower. I'm also seeing this warning:<\/p>\n\n<blockquote>\n  <p>UserWarning: Converting sparse IndexedSlices to a dense Tensor with\n  900482700 elements<\/p>\n<\/blockquote>\n\n<p>I saw the discussion on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/464\" rel=\"nofollow noreferrer\">this TensorFlow issue<\/a>, but I'm still not sure if the slowdown I'm experiencing is expected or if it's a bug. I'm using the Adam optimizer but it's pretty much the same thing with Adagrad.<\/p>\n\n<p>One workaround I guess I could try is to train using a minimal embedding matrix with only the ~12,000 words in my dataset, serialize the resulting embeddings and at runtime merge them with the remaining words from the pre-trained embeddings. I think this should work but it sounds hacky. <\/p>\n\n<p>Is that currently the best solution or am I missing something?<\/p>\n","answers":[{"AnswerId":"35898025","CreationDate":"2016-03-09T17:06:25.537","ParentId":null,"OwnerUserId":"524283","Title":null,"Body":"<p>So there were two issues here:<\/p>\n\n<ol>\n<li>As mrry pointed out in his comment to the question, the warning was not a result of a conversion during the updates. Rather, I was calculating summary statistics (sparsity and histogram) on the embeddings gradient and that caused the conversion.<\/li>\n<li>Interestingly, removing the summaries made the message go away, but the code remained slow. Per <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/464\" rel=\"nofollow\">the TensorFlow issue referenced in the question<\/a>, I had to also replace the <code>AdamOptimizer<\/code> with the <code>AdagradOptimizer<\/code> and once I did that the runtime was back on par with the one obtained from a small vocabulary.<\/li>\n<\/ol>\n"}]}
{"QuestionId":35828366,"AnswerCount":1,"Tags":"<lua><neural-network><torch><loss>","CreationDate":"2016-03-06T14:56:03.837","AcceptedAnswerId":35878481.0,"OwnerUserId":2011802.0,"Title":"Pairwise similarity criterion in torch","Body":"<p>I want to implement a new Criterion in torch7.<\/p>\n\n<p>Basically, I have pair of samples for which I have a normalised similarity real value (let's denote it 'd').<\/p>\n\n<p>Among existing Criterion, the closest I can have is the <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/criterion.md#nn.CosineEmbeddingCriterion\" rel=\"nofollow noreferrer\">CosineEmbeddingCriterion<\/a> which provides the following loss : <\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n\n<p>Obviously this is designed for pairs, but this is for classification problems.<\/p>\n\n<p>In my problem, I want to implement a Criterion that provide the following loss function : \n <\/p>\n\n<p>Unfortunately, unlike <a href=\"https:\/\/stackoverflow.com\/questions\/33648796\/add-my-custom-loss-function-to-torch\">in this question<\/a>, it does not appear to me that I cannot simply combine existing criterion to do that.<\/p>\n\n<p>So I plan to go on the painful way to create a new module from <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/CosineEmbeddingCriterion.lua\" rel=\"nofollow noreferrer\">torch7 source file.<\/a><\/p>\n\n<p>My questions are as follows :<\/p>\n\n<ul>\n<li>Am I correct that I don't any other straightforward solutions ?<\/li>\n<li>If I implement my own Criterion in torch7 :\n\n<ul>\n<li>is there a skeleton to write criterion or a guide ?<\/li>\n<li>how can I check that it is valid ? (how to unit test it ?)<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>(it seems to me very easy to write a criterion that seems correct to me -- but not in practise)<\/p>\n\n<p>Thank you in advance for any clue !!<\/p>\n","answers":[{"AnswerId":"35878481","CreationDate":"2016-03-08T21:35:16.727","ParentId":null,"OwnerUserId":"2669077","Title":null,"Body":"<p>You just need to write the <code>updateOutput<\/code> and <code>updateGradInput<\/code> functions for your criterion. Then, you can use it like any other criterion. Here is a skeleton.<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/BCECriterion.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/nn\/blob\/master\/BCECriterion.lua<\/a><\/p>\n"}]}
{"QuestionId":35829247,"AnswerCount":2,"Tags":"<caffe>","CreationDate":"2016-03-06T16:15:08.193","AcceptedAnswerId":null,"OwnerUserId":3735010.0,"Title":"Caffe compilation fails: Undefined symbols for architecture x86_64?","Body":"<p>If I try building the newest version of Caffe, it leads to this error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm building on osx, OpenBLAS, and CPU_ONLY. I found a <a href=\"https:\/\/stackoverflow.com\/questions\/27890137\/undefined-symbols-for-architecture-x86-64-for-caffe-build\">kind of similar issue<\/a> on here but it appears to have been a resolved issue, and I'm not getting the exact same error, though perhaps it's related? I can also build and run an older version of Caffe from a month ago, so I think something has changed very recently.<\/p>\n\n<p>Any ideas on how to overcome this error?<\/p>\n","answers":[{"AnswerId":"39249413","CreationDate":"2016-08-31T12:16:47.160","ParentId":null,"OwnerUserId":"6778675","Title":null,"Body":"<p>I had exactly the same problem. Its now resolved.\nDo check if you already have a libcaffe.so in your system library paths (maybe \/usr\/local\/lib). If so, delete the existing libcaffe.so and build again.<\/p>\n"},{"AnswerId":"35936292","CreationDate":"2016-03-11T09:19:43.097","ParentId":null,"OwnerUserId":"5837113","Title":null,"Body":"<p>It shows a link problem when compile.On osx this problem will happen usually.I guess the problem happened in <code>Makefile.config<\/code>.You can change to<\/p>\n\n<pre><code># To customize your choice of compiler, uncomment and set the following.\n# N.B. the default for Linux is g++ and the default for OSX is clang++\nCUSTOM_CXX := g++\n<\/code><\/pre>\n\n<p>and confirm some path is correct.<\/p>\n"}]}
{"QuestionId":35829730,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><convolution><caffe><pycaffe>","CreationDate":"2016-03-06T17:00:42.560","AcceptedAnswerId":35830797.0,"OwnerUserId":4308982.0,"Title":"Why does the output of this convolution have these dimensions?","Body":"<p>I am trying to replicate the output of a convolution in Caffe.<\/p>\n\n<p>As far as I understand, Caffe uses the im2col algorithm to cast nD arrays into matrices and multiply them together. However, the dimensions of the output in Caffe confuse me. <\/p>\n\n<p>Using the ImageData layer, I input 4 images of dimension 150x149 with a batch size of 4. Caffe creates a 4D array with dimensions 4x3x149x150. <\/p>\n\n<p>I convolve these with a convolution layer with a filter of size 7 and a stride of 1 (num_output = 1 &amp; bias = the zero vector). This means that the weights are of dimensions 1x3x7x7. As far as I understand, if the stride is 1, the filter should be applied to every element and the output should have the same dimensions as the input. What I get, however, is an output of the following dimensionality: 4x1x143x144.     <\/p>\n\n<p>I don't see how this is possible. How would one carry out the same operations in Matlab (or whatever)?<\/p>\n\n<p><strong>How do you get from the input to the output?<\/strong><\/p>\n","answers":[{"AnswerId":"35830797","CreationDate":"2016-03-06T18:30:30.780","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>Your convolution filter has width 7. Looking at it in 1D, its first application will be to the pixels<\/p>\n\n<pre><code>1, 2, 3, 4, 5, 6, 7,\n<\/code><\/pre>\n\n<p>the second one to the pixels<\/p>\n\n<pre><code>2, 3, 4, 5, 6, 7, 8,\n<\/code><\/pre>\n\n<p>and so on until the last one, which operates on pixels<\/p>\n\n<pre><code>143, 144, 145, 146, 147, 148, 149.\n<\/code><\/pre>\n\n<p>So, as you can see, there are 143 different applications of the filter, and each one produces one output pixel. Hence, the output dimension is 143. It works analogously for the other coordinate direction.<\/p>\n\n<p>In short, the output width for a filter with stride 1 will always be<\/p>\n\n<pre><code>output width = image width - filter width + 1.\n<\/code><\/pre>\n"}]}
{"QuestionId":35831803,"AnswerCount":1,"Tags":"<deep-learning><keras><lasagne>","CreationDate":"2016-03-06T19:57:57.833","AcceptedAnswerId":35832178.0,"OwnerUserId":6026364.0,"Title":"Highway networks in keras and lasagne - significant perfomance difference","Body":"<p>I implemented highway networks with keras and with lasagne, and the keras version consistently underperforms to the lasagne version. I am using the same dataset and metaparameters in both of them. Here is the keras version's code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>And here is the lasagne version's code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Now the keras version barely outperforms logistic regression, while the lasagne version is the best scoring algorithm so far. Any ideas as to why?<\/p>\n","answers":[{"AnswerId":"35832178","CreationDate":"2016-03-06T20:32:50.850","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>Here are some suggestions (I'm not sure if they will actually close the performance gap you are observing):<\/p>\n\n<p>According to the <a href=\"http:\/\/keras.io\/layers\/core\/#highway\" rel=\"nofollow\">Keras documentation<\/a> the Highway layer is initialized using Glorot Uniform weights while in your Lasagne code you are using Orthogonal weight initialization. Unless you have another part of your code where you set the weight initialization to Orthogonal for the Keras Highway layer, this could be a source of the performance gap.<\/p>\n\n<p>It also seems like you are using Adagrad for your Keras model, but you are using Adadelta for your Lasagne model. <\/p>\n\n<p>Also I am not 100% sure about this, but you may also want to verify that your transform bias terms are initialized the same way. <\/p>\n"}]}
{"QuestionId":35833011,"AnswerCount":2,"Tags":"<python><if-statement><tensorflow>","CreationDate":"2016-03-06T21:47:46.527","AcceptedAnswerId":35833133.0,"OwnerUserId":5948325.0,"Title":"How to add if condition in a TensorFlow graph?","Body":"<p>Let's say I have following code:<\/p>\n\n<pre><\/pre>\n\n<p>Would the  statement work in the calculation (I do not think so)? If not, how can I add an  statement into the TensorFlow calculation graph? <\/p>\n","answers":[{"AnswerId":"56616372","CreationDate":"2019-06-16T05:36:41.813","ParentId":null,"OwnerUserId":"4909087","Title":null,"Body":"<h2>TensorFlow 2.0<\/h2>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/autograph\" rel=\"nofollow noreferrer\">TF 2.0 introduces a feature called AutoGraph<\/a> which lets you JIT compile python code into Graph executions. This means you can use python control flow statements (yes, this includes <code>if<\/code> statements). From the docs,<\/p>\n\n<blockquote>\n  <p>AutoGraph supports common Python statements like <code>while<\/code>, <code>for<\/code>, <code>if<\/code>,\n  <code>break<\/code>, <code>continue<\/code> and <code>return<\/code>, with support for nesting. That means you\n  can use Tensor expressions in the condition of <code>while<\/code> and <code>if<\/code>\n  statements, or iterate over a Tensor in a <code>for<\/code> loop.<\/p>\n<\/blockquote>\n\n<p>You will need to define a function implementing your logic and annotate it with <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/function\" rel=\"nofollow noreferrer\"><strong><code>tf.function<\/code><\/strong><\/a>. Here is a modified example from the documentation:<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.function\ndef sum_even(items):\n  s = 0\n  for c in items:\n    if tf.equal(c % 2, 0): \n        s += c\n  return s\n\nsum_even(tf.constant([10, 12, 15, 20]))\n#  &lt;tf.Tensor: id=1146, shape=(), dtype=int32, numpy=42&gt;\n<\/code><\/pre>\n"},{"AnswerId":"35833133","CreationDate":"2016-03-06T22:00:44.600","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You're correct that the <code>if<\/code> statement doesn't work here, because the condition is evaluated at graph construction time, whereas presumably you want the condition to depend on the value fed to the placeholder at runtime. (In fact, it will always take the first branch, because <code>condition &gt; 0<\/code> evaluates to a <code>Tensor<\/code>, which is <a href=\"https:\/\/www.udacity.com\/wiki\/cs258\/truthiness-in-python\" rel=\"noreferrer\">\"truthy\" in Python<\/a>.)<\/p>\n\n<p>To support conditional control flow, TensorFlow provides the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/control_flow_ops.html#cond\" rel=\"noreferrer\"><code>tf.cond()<\/code><\/a> operator, which evaluates one of two branches, depending on a boolean condition. To show you how to use it, I'll rewrite your program so that <code>condition<\/code> is a scalar <code>tf.int32<\/code> value for simplicity:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.placeholder(tf.float32, shape=[None, ins_size**2*3], name=\"x_input\")\ncondition = tf.placeholder(tf.int32, shape=[], name=\"condition\")\nW = tf.Variable(tf.zeros([ins_size**2 * 3, label_option]), name=\"weights\")\nb = tf.Variable(tf.zeros([label_option]), name=\"bias\")\n\ny = tf.cond(condition &gt; 0, lambda: tf.matmul(x, W) + b, lambda: tf.matmul(x, W) - b)\n<\/code><\/pre>\n"}]}
{"QuestionId":35833953,"AnswerCount":1,"Tags":"<indexing><tensorflow>","CreationDate":"2016-03-06T23:32:56.387","AcceptedAnswerId":35836991.0,"OwnerUserId":5948325.0,"Title":"Index operation in TensorFlow","Body":"<p>When I was doing batch labeling for some data, I have a variable for recording all calculation results:<\/p>\n\n<pre><\/pre>\n\n<p>In the calculation, I have a loop to deal with each batch:<\/p>\n\n<pre><\/pre>\n\n<p>How can I copy value of  into  ?<\/p>\n\n<p>To be more clear, I want something like:<\/p>\n\n<pre><\/pre>\n\n<p>How can I make those code actually work ?<\/p>\n","answers":[{"AnswerId":"35836991","CreationDate":"2016-03-07T05:47:29.507","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Since <code>p_all<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow\"><code>tf.Variable<\/code><\/a>, you can use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#scatter_update\" rel=\"nofollow\"><code>tf.scatter_update()<\/code><\/a> op to update individual rows in each batch:<\/p>\n\n<pre><code># Equivalent to `p_all[batch_index, :] = p_each_batch`\nupdate_op = tf.scatter_update(p_all,\n                              tf.expand_dims(batch_index, 0),\n                              tf.expand_dims(p_each_batch, 0)) \n\nfor i in range(batch_num):\n    feed = {x: testDS.test.next_batch(batch_size), batch_index: i}\n    sess.run(update_op, feed_dict=feed)\n<\/code><\/pre>\n"}]}
{"QuestionId":35835989,"AnswerCount":3,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-03-07T04:02:34.900","AcceptedAnswerId":37001392.0,"OwnerUserId":348412.0,"Title":"How to pick the last valid output values from tensorflow RNN","Body":"<p>I'm training an LSTM cell on batches of sequences that have different lengths. The  has the very convenient parameter , but after calling it, I don't know how to pick the output rows corresponding the last time step of each item in the batch.<\/p>\n\n<p>My code is basically as follows:<\/p>\n\n<pre><\/pre>\n\n<p> is a list with the LSTM output at each time step. However, each item in my batch has a different length, and so I would like to create a tensor containing the last LSTM output valid for each item in my batch.<\/p>\n\n<p>If I could use numpy indexing, I would just do something like this:<\/p>\n\n<pre><\/pre>\n\n<p>But it turns out that for the time begin tensorflow doesn't support it (I'm aware of the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow\">feature request<\/a>).<\/p>\n\n<p>So, how could I get these values?<\/p>\n","answers":[{"AnswerId":"37001392","CreationDate":"2016-05-03T10:13:49.707","ParentId":null,"OwnerUserId":"348412","Title":null,"Body":"<p>A more acceptable workaround was published by danijar on the feature request page I linked in the question. It doesn't need to evaluate the tensors, which is a big plus. <\/p>\n\n<p>I got it to work with tensorflow 0.8. Here is the code:<\/p>\n\n<pre><code>def extract_last_relevant(outputs, length):\n    \"\"\"\n    Args:\n        outputs: [Tensor(batch_size, output_neurons)]: A list containing the output\n            activations of each in the batch for each time step as returned by\n            tensorflow.models.rnn.rnn.\n        length: Tensor(batch_size): The used sequence length of each example in the\n            batch with all later time steps being zeros. Should be of type tf.int32.\n\n    Returns:\n        Tensor(batch_size, output_neurons): The last relevant output activation for\n            each example in the batch.\n    \"\"\"\n    output = tf.transpose(tf.pack(outputs), perm=[1, 0, 2])\n    # Query shape.\n    batch_size = tf.shape(output)[0]\n    max_length = int(output.get_shape()[1])\n    num_neurons = int(output.get_shape()[2])\n    # Index into flattened array as a workaround.\n    index = tf.range(0, batch_size) * max_length + (length - 1)\n    flat = tf.reshape(output, [-1, num_neurons])\n    relevant = tf.gather(flat, index)\n    return relevant\n<\/code><\/pre>\n"},{"AnswerId":"35846025","CreationDate":"2016-03-07T14:17:17.710","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>It's not be the nicest solution but you could evaluate your outputs then just use numpy indexing to get the results and create a tensor variable from that? It might work as a stop gap until tensorflow gets this feature. e.g.<\/p>\n\n<pre><code>all_outputs = session.run(lstm_outputs, feed_dict={'your inputs'})\nlast_outputs = all_outputs[sequence_lengths, tf.range(batch_size), :]\nuse_this_as_an_input_to_new_tensorflow_op = tf.constant(last_outputs)\n<\/code><\/pre>\n"},{"AnswerId":"41661612","CreationDate":"2017-01-15T13:28:31.497","ParentId":null,"OwnerUserId":"5068645","Title":null,"Body":"<p>if you're only interested in the last valid output you can retrieve it through the state returned by <code>tf.nn.rnn()<\/code> considering that it's always a tuple (c, h) where c is the last state and h is the last output. When the state is a <code>LSTMStateTuple<\/code> you can use the following snippet (working in tensorflow 0.12):<\/p>\n\n<pre><code>lstm_cell = tf.nn.rnn_cell.LSTMCell(num_lstm_units, input_size)\nlstm_outputs, state = tf.nn.rnn(lstm_cell, input_list, dtype=tf.float32, sequence_length=sequence_lengths)\nlast_output = state[1]\n<\/code><\/pre>\n"}]}
{"QuestionId":35836326,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-07T04:46:11.057","AcceptedAnswerId":35857902.0,"OwnerUserId":2853544.0,"Title":"Convert RGBA-image to black and white image in Tensorflow","Body":"<p><strong>tf.image.decode_png()<\/strong> can output grayscale, RGB and RGBA image.<\/p>\n\n<p>But I'd like to convert RGBA to pure black and white in <strong>Tensorflow<\/strong> (without using <strong>pillow<\/strong>).<\/p>\n\n<p>Please give me some advice.<\/p>\n","answers":[{"AnswerId":"35857902","CreationDate":"2016-03-08T02:37:34.620","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Use <code>tf.select<\/code> to do the thresholding<\/p>\n\n<pre><code>pil_image = PilImage.open('\/temp\/panda.png')\nshow_pil_image(pil_image)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/p662z.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p662z.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>pil_buf = open('\/temp\/panda.png').read()\ncontents = tf.placeholder(dtype=tf.string)\ndecode_op = tf.image.decode_png(contents, channels=1)\ngray_image = tf.squeeze(decode_op) # shape (127,127,1) -&gt; shape (127,127)\nsess = create_session()\n[decoded] = sess.run([gray_image], feed_dict={contents: pil_buf})\nshow_pil_image(PilImage.fromarray(decoded))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sHHm9.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sHHm9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>contents = tf.placeholder(dtype=tf.string)\ndecode_op = tf.image.decode_png(contents, channels=1)\ngray_image = tf.squeeze(decode_op) # shape (127,127,1) -&gt; shape (127,127)\nselect_op = tf.select(gray_image&gt;127, 255*tf.ones_like(gray_image), tf.zeros_like(gray_image))\nsess = create_session()\n[decoded] = sess.run([select_op], feed_dict={contents: pil_buf})\nshow_pil_image(PilImage.fromarray(decoded))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Qfj3B.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Qfj3B.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":35837644,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-03-07T06:39:42.477","AcceptedAnswerId":35857765.0,"OwnerUserId":5355865.0,"Title":"Update Option of Theano Function","Body":"<p>I am using the theano tutorials of deep learning and I have a doubt regarding how the <strong>theano function's update<\/strong> works. Whatever parameters the update is defined with will it take the new value if any of the parameters changed?<\/p>\n\n<pre><\/pre>\n\n<p>Updates is defined as <\/p>\n\n<pre><\/pre>\n\n<p>Here learnRate is not a theano variable. My question is if I change the learnRate at some point of time  then will the theano function  take the changed value of learnRate or will continue with the old value?  <\/p>\n\n<p>LearnRate is changed as<\/p>\n\n<pre><\/pre>\n\n<p>Initially the LearnRate was 0.05<\/p>\n","answers":[{"AnswerId":"35857765","CreationDate":"2016-03-08T02:22:08.087","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>It will continue with the old value.  <\/p>\n\n<p>Numbers are immutable in python so changing the value of <code>LearnRate<\/code> won't affect the value being used in the <code>updates<\/code> list, and I guess once a theano function is compiled with the provided <code>updates<\/code> it can not be changed unless you compile it again with the new value.<\/p>\n\n<p>As you probably have known, you can make the learning rate a theano variable and then change its value when you want to.  <\/p>\n\n<p>Here's another question that might be useful, <a href=\"https:\/\/stackoverflow.com\/questions\/34146656\/is-there-a-way-to-change-a-functions-update-list-without-re-compiling-it-in-the\">Is there a way to change a function&#39;s update list without re-compiling it in Theano?<\/a><\/p>\n"}]}
{"QuestionId":35840231,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2016-03-07T09:31:15.743","AcceptedAnswerId":null,"OwnerUserId":988125.0,"Title":"Windows Keras Install error - Import error occurred. No module named queue","Body":"<p>I am running the below Anaconda distribution:<\/p>\n\n<pre><\/pre>\n\n<p>I have got Theano (0.7.0) installed, which works fine. I have downloaded and unzipped the Keras distribution (as I cannot do pip due to firewall restrictions).\nI run <\/p>\n\n<pre><\/pre>\n\n<p>which installs Keras. When it comes to importing it, the below line does not throw an error:<\/p>\n\n<pre><\/pre>\n\n<p>However,  throws the error:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to run  as below, this time it cannot find .<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know what can be going wrong?<\/p>\n","answers":[{"AnswerId":"36639961","CreationDate":"2016-04-15T06:43:01.457","ParentId":null,"OwnerUserId":"5142261","Title":null,"Body":"<p>The amount of your information is too little,I could only specify my suspect.<\/p>\n\n<p>1\u3001\"Environment Path\"\nI suppose you had a old python version in addition to anaconda python,and the old version python isn't removed from the environment path,so when you exec the command \"python setup.py install\" ,the keras was installed in old python site-package<\/p>\n\n<p>2\u3001\"install failed\"\nthe core problem is this sentence \"from keras.models import Sequential\"\nwhen you can't compile this, this next error \"'module' object has no attribute 'models'\" must be thrown.the first error pointed out the module queue missing, you install it in addition.<\/p>\n\n<p>otherwise,reinstall all the module,you can try this procedure:<\/p>\n\n<ul>\n<li>Install Anaconda x64.<\/li>\n<li>run conda update conda<\/li>\n<li>run conda update --all<\/li>\n<li>run conda install mingw libpython<\/li>\n<li><p>pip install git+git:\/\/github.com\/Theano\/Theano.git<\/p>\n\n<p>you can git clone then python install<\/p><\/li>\n<li>pip install git+git:\/\/github.com\/fchollet\/keras.git <\/li>\n<\/ul>\n"}]}
{"QuestionId":35842274,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-03-07T11:15:10.580","AcceptedAnswerId":null,"OwnerUserId":5881101.0,"Title":"Convert own image to MNIST's image","Body":"<p>I am newbie of tensorflow.\nI trained the digit prediction model using MNIST's train data.\nAnd then I test the model using my own image.\nIt cannot predict the actual result.<\/p>\n\n<p>The problems are :<\/p>\n\n<ol>\n<li>MNIST's images are needed black and white   <\/li>\n<li>The images are size normalized to fit in a 20x20 pixel box and there are centered in a 28x28 image using the center of mass.<\/li>\n<li>I don't want to use  <\/li>\n<\/ol>\n\n<p>The question is <strong>How to shift my own handwritten digit image to the center of 28x28 image. Own image can be any color and that image to change Black and White MNIST's image<\/strong><\/p>\n","answers":[{"AnswerId":"55968472","CreationDate":"2019-05-03T11:00:51.007","ParentId":null,"OwnerUserId":"10967437","Title":null,"Body":"<pre><code>from PIL import Image, ImageFilter\n\n\ndef imageprepare(argv):\n    \"\"\"\n    This function returns the pixel values.\n    The imput is a png file location.\n    \"\"\"\n    im = Image.open(argv).convert('L')\n    width = float(im.size[0])\n    height = float(im.size[1])\n    newImage = Image.new('L', (28, 28), (255))  # creates white canvas of 28x28 pixels\n\n    if width &gt; height:  # check which dimension is bigger\n        # Width is bigger. Width becomes 20 pixels.\n        nheight = int(round((20.0 \/ width * height), 0))  # resize height according to ratio width\n        if (nheight == 0):  # rare case but minimum is 1 pixel\n            nheight = 1\n            # resize and sharpen\n        img = im.resize((20, nheight), Image.ANTIALIAS).filter(ImageFilter.SHARPEN)\n        wtop = int(round(((28 - nheight) \/ 2), 0))  # calculate horizontal position\n        newImage.paste(img, (4, wtop))  # paste resized image on white canvas\n    else:\n        # Height is bigger. Heigth becomes 20 pixels.\n        nwidth = int(round((20.0 \/ height * width), 0))  # resize width according to ratio height\n        if (nwidth == 0):  # rare case but minimum is 1 pixel\n            nwidth = 1\n            # resize and sharpen\n        img = im.resize((nwidth, 20), Image.ANTIALIAS).filter(ImageFilter.SHARPEN)\n        wleft = int(round(((28 - nwidth) \/ 2), 0))  # caculate vertical pozition\n        newImage.paste(img, (wleft, 4))  # paste resized image on white canvas\n\n    # newImage.save(\"sample.png\n\n    tv = list(newImage.getdata())  # get pixel values\n\n    # normalize pixels to 0 and 1. 0 is pure white, 1 is pure black.\n    tva = [(255 - x) * 1.0 \/ 255.0 for x in tv]\n    print(tva)\n    return tva\n\nx=imageprepare('.\/image.png')#file path here\nprint(len(x))# mnist IMAGES are 28x28=784 pixels\n<\/code><\/pre>\n"},{"AnswerId":"35857604","CreationDate":"2016-03-08T02:04:02.477","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I would use numpy recipe like this one --\n<a href=\"https:\/\/www.kaggle.com\/c\/digit-recognizer\/forums\/t\/6366\/normalization-and-centering-of-images-in-mnist\" rel=\"nofollow\">https:\/\/www.kaggle.com\/c\/digit-recognizer\/forums\/t\/6366\/normalization-and-centering-of-images-in-mnist<\/a><\/p>\n\n<p>You could probably remap this to pure TensorFlow pipeline, but I'm not sure it's necessary given that it's tiny images.<\/p>\n\n<p>Also you would get better accuracy if you went the other way -- instead of normalizing your input data, make your network robust to lack of normalization by training on a larger dataset of randomly shifted\/rescaled MNIST digits.<\/p>\n"}]}
{"QuestionId":35842598,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-03-07T11:31:52.560","AcceptedAnswerId":35847836.0,"OwnerUserId":200340.0,"Title":"TensorFlow: using a tensor to index another tensor","Body":"<p>I have a basic question about how to do indexing in TensorFlow.<\/p>\n\n<p>In numpy:<\/p>\n\n<pre><\/pre>\n\n<p>I can get<\/p>\n\n<pre><\/pre>\n\n<p>How can I do this in TensorFlow?<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"35847836","CreationDate":"2016-03-07T15:40:42.670","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Fortunately, the exact case you're asking about is supported in TensorFlow by <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"noreferrer\"><code>tf.gather()<\/code><\/a>:<\/p>\n\n<pre><code>result = x_t * tf.gather(e_t, x_t)\n\nwith tf.Session() as sess:\n    print sess.run(result)  # ==&gt; 'array([1, 0, 3, 3, 0, 5, 0, 7, 1, 3])'\n<\/code><\/pre>\n\n<p>The <code>tf.gather()<\/code> op is less powerful than <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html\" rel=\"noreferrer\">NumPy's advanced indexing<\/a>: it only supports extracting full slices of a tensor on its 0th dimension. Support for more general indexing has been requested, and is being tracked in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"noreferrer\">this GitHub issue<\/a>.<\/p>\n"}]}
{"QuestionId":35845448,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2016-03-07T13:48:18.880","AcceptedAnswerId":48579234.0,"OwnerUserId":5324819.0,"Title":"Simple XOR example on theano","Body":"<p>I am a newbie for theano and I am failed to make simple XOR example with theano. I tried plenty things to make it work but it is seems that I just doing shamanism. Look at code it is pretty simple, but with it I got random results. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48579234","CreationDate":"2018-02-02T09:09:19.830","ParentId":null,"OwnerUserId":"4851239","Title":null,"Body":"<p>The problem is in your update section. I renamed the variable to 'updates' as 'list' and 'dict' are reserved words; not a good choice. Also, I do not know why you want to reduce your learning rate so fast, I removed that.\nThe updates should be an array like here<\/p>\n\n<pre><code>updates = [(w1, w1 - lr * gw),\n           (b1, b1 - lr * gb),\n           (w2, w2 - lr * gv),\n           (b2, b2 - lr * gc)]\n\n# compile a fast training function\ntrain = theano.function([sx, sy], err, updates=updates)\n<\/code><\/pre>\n\n<p>I ran the changed example and got the result below. It needs more iterations to reduce the loss, but apart from that, it is ok.<\/p>\n\n<pre><code>Processing...\nError: 1.4456279556\n...\nError: 0.0767515052046\nWeights:\n[[ 0.52955082 -1.26936557]\n [-1.05887804  0.04998216]]\n[ 0.29209577 -0.22703456]\n[[-0.89983822]\n [-0.88619565]]\n[-0.86047891]\n[[-5. -5.]\n [-5.  5.]\n [ 5. -5.]\n [ 5.  5.]]\n[[-0.98989634]\n [-0.68941057]\n [-0.7034631 ]\n [ 0.72087948]]\n[[-0.89999998]\n [-0.89999998]\n [-0.89999998]\n [ 0.89999998]]\n<\/code><\/pre>\n"}]}
{"QuestionId":35849107,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-07T16:38:00.130","AcceptedAnswerId":null,"OwnerUserId":4383691.0,"Title":"Tensorflow: Split a placeholder BEFORE input has been fed?","Body":"<p>My model receives an input sequence via feed and then has to split this sequence by a delimiter within the model. However as Tensorflow's graph requires that we put placeholders before hand, I don't know how to make this work. The issue is that I don't know how many pieces the input sequence will be split into before I actually feed the input. <\/p>\n\n<p>Is there a way to create a model that can handle a changing number and shape of tensors within the model?<\/p>\n","answers":[{"AnswerId":"35857158","CreationDate":"2016-03-08T01:10:27.570","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>Are you trying to use an RNN?  If so, use parse_single_sequence_example to read data, in conjunction with a PaddingFIFOQueue, and dynamic_rnn to run the RNN on it.  You'll have to store your data as tensorflow.SequenceExample protos.  Alternatively you need a TF op that reads a string, splits on commas, and outputs a vector Tensor of values.<\/p>\n"}]}
{"QuestionId":35851036,"AnswerCount":2,"Tags":"<lua><torch>","CreationDate":"2016-03-07T18:21:06.597","AcceptedAnswerId":null,"OwnerUserId":5378521.0,"Title":"attempt to index local 'f' (a nil value) Lua\/Torch7","Body":"<p>I'm trying to run an example from torch7 only to come across this error.<\/p>\n\n<pre>\n    sandesh@sandesh-H87M-D3H:~\/Downloads\/tutorials-master\/2_supervised$ luajit doall.lua \n    ==> processing options\n    ==> executing all\n    ==> downloading dataset\n    ==> using regular, full training data\n    ==> loading dataset\n    ==> preprocessing data\n    ==> preprocessing data: colorspace RGB -> YUV\n    ==> preprocessing data: normalize each feature (channel) globally\n    ==> preprocessing data: normalize all three channels locally\n    ==> verify statistics\n    training data, y-channel, mean: 0.00067706172257129\n    training data, y-channel, standard deviation: 0.39473240322794\n    test data, y-channel, mean: -0.0010822884348063\n    test data, y-channel, standard deviation: 0.38091408093043\n    training data, u-channel, mean: -0.0048219975630079\n    training data, u-channel, standard deviation: 0.29768662619471\n    test data, u-channel, mean: -0.0030795217110624\n    test data, u-channel, standard deviation: 0.22289780235542\n    training data, v-channel, mean: 0.0036312269637064\n    training data, v-channel, standard deviation: 0.25405592463897\n    test data, v-channel, mean: 0.0033847450016769\n    test data, v-channel, standard deviation: 0.20362829592977\n    ==> visualizing data\n    ==> define parameters\n    ==> construct model\n    ==> here is the model:\n    nn.Sequential {\n      [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> output]\n      (1): nn.SpatialConvolutionMM(3 -> 64, 5x5)\n      (2): nn.Tanh\n      (3): nn.Sequential {\n        [input -> (1) -> (2) -> (3) -> (4) -> output]\n        (1): nn.Square\n        (2): nn.SpatialAveragePooling(2,2,2,2)\n        (3): nn.MulConstant\n        (4): nn.Sqrt\n      }\n      (4): nn.SpatialSubtractiveNormalization\n      (5): nn.SpatialConvolutionMM(64 -> 64, 5x5)\n      (6): nn.Tanh\n      (7): nn.Sequential {\n        [input -> (1) -> (2) -> (3) -> (4) -> output]\n        (1): nn.Square\n        (2): nn.SpatialAveragePooling(2,2,2,2)\n        (3): nn.MulConstant\n        (4): nn.Sqrt\n      }\n      (8): nn.SpatialSubtractiveNormalization\n      (9): nn.Reshape(1600)\n      (10): nn.Linear(1600 -> 128)\n      (11): nn.Tanh\n      (12): nn.Linear(128 -> 10)\n    }\n    ==> define loss\n    ==> here is the loss function:\n    nn.ClassNLLCriterion\n    ==> defining some tools\n    luajit: \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:38: attempt to index local 'f' (a nil value)\n    stack traceback:\n        \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:38: in function 'execute'\n        \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:71: in function 'uname'\n        \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/optim\/Logger.lua:38: in function '__init'\n        \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/torch\/init.lua:91: in function \n        [C]: in function 'Logger'\n        4_train.lua:60: in main chunk\n        [C]: in function 'dofile'\n        doall.lua:70: in main chunk\n        [C]: at 0x00406670\n\n    <\/pre>\n\n<p><br>\nI did not change any code in any of the lua files ...\n<br>\n<br>\nThis is the 4_train.lua file \n<br><\/p>\n\n<pre> <\/pre>\n\n<p><br>\n<br>\nThis is doall.lua<\/p>\n\n<pre><\/pre>\n\n<p><br>\n<br>\nThe git link is <a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/4_train.lua\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/4_train.lua<\/a>\n<br>\n<br>\nAlso I'm not using cuda as I dont have a GPU<\/p>\n","answers":[{"AnswerId":"55791856","CreationDate":"2019-04-22T09:04:41.530","ParentId":null,"OwnerUserId":"11394019","Title":null,"Body":"<p>I am currently playing with torch-rnn and stumbled accross the issue repeatedly. Check your input and see if the incorrect \/ no file was inputted. If the code worked before and no attempts were made to change it, it should be intact and it is your input that is problematic. <\/p>\n"},{"AnswerId":"35867637","CreationDate":"2016-03-08T12:38:51.917","ParentId":null,"OwnerUserId":"2858170","Title":null,"Body":"<p>I won't tell you what is wrong as you did not show any efforts to solve the problem yourself. But I will tell you how to proceed.<\/p>\n\n<blockquote>\n  <p>luajit: \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:38:\n  attempt to index local 'f' (a nil value)\n      stack traceback:\n          \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:38: in function 'execute'\n          \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/sys\/init.lua:71: in function 'uname'\n          \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/optim\/Logger.lua:38: in function '__init'\n          \/home\/sandesh\/torch\/install\/share\/lua\/5.1\/torch\/init.lua:91: in function \n          [C]: in function 'Logger'<\/p>\n<\/blockquote>\n\n<p>This tells you that some local f in init.lua line 38 is nil, which causes a problem. So open that file and find out where that f's value should come from and why it is nil. Then fix that. Also see if there is a more recent version of Torch that handles f being nil correctly. If not, change the code yourself, if possible. Otherwise try to stop this from happening by validating your inputs to torch.<\/p>\n"}]}
{"QuestionId":35851921,"AnswerCount":1,"Tags":"<matlab><computer-vision><deep-learning><convolution><caffe>","CreationDate":"2016-03-07T19:07:56.510","AcceptedAnswerId":35866899.0,"OwnerUserId":4308982.0,"Title":"Implementing convolution as a matrix multiplication","Body":"<p><em>Pithy: Help with Matlab script that takes ImageData array and Convolution weights  from Caffe and returns convolution. Please.<\/em><\/p>\n\n<hr>\n\n<p>I am trying to recreate a convolution generated by Caffe in Matlab.<\/p>\n\n<p>Let's make the following definitions <\/p>\n\n<pre><\/pre>\n\n<p>The following text describes how to generalize the convolution as a matrix multiplication:<\/p>\n\n<blockquote>\n  <p>The local regions in the input image are stretched out into columns in an operation commonly called im2col. For example, if the input is [227x227x3] and it is to be convolved with 11x11x3 filters at stride 4, then we would take [11x11x3] blocks of pixels in the input and stretch each block into a column vector of size 11*11*3 = 363. Iterating this process in the input at stride of 4 gives (227-11)\/4+1 = 55 locations along both width and height, leading to an output matrix X_col of im2col of size [363 x 3025], where every column is a stretched out receptive field and there are 55*55 = 3025 of them in total. Note that since the receptive fields overlap, every number in the input volume may be duplicated in multiple distinct columns.<\/p>\n<\/blockquote>\n\n<p>From this, one could draw the conclusion that the im2col function call would look something like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, if I use the following parameter-values<\/p>\n\n<pre><\/pre>\n\n<p>I get the following dimensions<\/p>\n\n<pre><\/pre>\n\n<p>And if I use the im2col function call from above, I end up with an empty matrix. <\/p>\n\n<p>If I change the stride to 1 in the above example, the size of the input, the output and the filter remains the same. If I use Matlab's 'convn' command, the size is not the same as the actual output from Caffe.<\/p>\n\n<pre><\/pre>\n\n<p><strong>What would be the general way to resize your array for matrix multiplication?<\/strong><\/p>\n","answers":[{"AnswerId":"35866899","CreationDate":"2016-03-08T12:04:34.823","ParentId":null,"OwnerUserId":"2929337","Title":null,"Body":"<p>You are using the second argument to <code>im2col<\/code> wrong, see <a href=\"http:\/\/de.mathworks.com\/help\/images\/ref\/im2col.html\" rel=\"nofollow\">the documentation<\/a>.<\/p>\n\n<p>You should give it the size of the filter window that you are trying to slide over the image, i.e.:<\/p>\n\n<pre><code>cols = im2col( input, [F, F])\n<\/code><\/pre>\n"}]}
{"QuestionId":35853483,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-07T20:35:23.670","AcceptedAnswerId":35855219.0,"OwnerUserId":1002035.0,"Title":"tensorflow constant with variable size","Body":"<p>I have a variable batch size, so all of my inputs are of the form<\/p>\n\n<pre><\/pre>\n\n<p>to accept the variable batch sizes. However, how might you create a constant value with variable batch size? The issue is with this line:<\/p>\n\n<pre><\/pre>\n\n<p>It is giving me an error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm sure it is possible to initialize a constant tensor with variable batch size, how might I do so?<\/p>\n\n<p>I've also tried the following:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"51533312","CreationDate":"2018-07-26T07:30:41.603","ParentId":null,"OwnerUserId":"7309971","Title":null,"Body":"<p>Suppose you want to do something using <code>log_probs<\/code>. For example, you want to do power operation on a tensor <code>v<\/code> and a constant <code>log_probs<\/code>. And you want the shape of log_probs to vary with the shape of v. <\/p>\n\n<pre><code>v = tf.placeholder(tf.float32, shape=(None, 1)\nlog_probs = tf.constant(0.0, dtype=tf.float32, shape=[None, 1])\nresult = tf.pow(v, log_probs)\n<\/code><\/pre>\n\n<p>However, you cannot construct the constant <code>log_probs<\/code>. While, firstly, you can construct tf.constant just with shape =[1] <code>log_prob = tf.constant(0.0, dtype=tf.float32, shape=[None, 1])<\/code>. Then use <code>tf.map_fn()<\/code> to do pow operation for each element of v. <\/p>\n\n<pre><code>v = tf.placeholder(tf.float32, shape=(None, 1)\nlog_prob = tf.constant(0.0, dtype=tf.float32, shape=[1])\nresult = tf.map_fn(lambda ele : tf.pow(ele, log_prob), v)\n<\/code><\/pre>\n"},{"AnswerId":"35855219","CreationDate":"2016-03-07T22:17:27.960","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>A <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> has fixed size and value at graph construction time, so it probably isn't the right op for your application.<\/p>\n\n<p>If you are trying to create a tensor with a dynamic size and the same (constant) value for every element, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#fill\" rel=\"noreferrer\"><code>tf.fill()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#shape\" rel=\"noreferrer\"><code>tf.shape()<\/code><\/a> to create an appropriately-shaped tensor. For example, to create a tensor <code>t<\/code> that has the same shape as <code>input<\/code> and the value <code>0.5<\/code> everywhere:<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32, shape=(None, ...))\n\n# `tf.shape(input)` takes the dynamic shape of `input`.\nt = tf.fill(tf.shape(input), 0.5)\n<\/code><\/pre>\n\n<p>As <a href=\"https:\/\/stackoverflow.com\/questions\/35853483\/tensorflow-constant-with-variable-size#comment59374756_35853483\">Yaroslav mentions in his comment<\/a>, you may also be able to use <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"noreferrer\">(NumPy-style) broadcasting<\/a> to avoid materializing a tensor with dynamic shape. For example, if <code>input<\/code> has shape <code>(None, 32)<\/code> and <code>t<\/code> has shape <code>(1, 32)<\/code> then computing <code>tf.mul(input, t)<\/code> will broadcast <code>t<\/code> on the first dimension to match the shape of <code>input<\/code>.<\/p>\n"}]}
{"QuestionId":35863073,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-08T09:04:12.840","AcceptedAnswerId":null,"OwnerUserId":5886358.0,"Title":"How do I get my images into Tensorflow?","Body":"<p>I am new to Tensorflow, and I am following the Mnist simple tutorial. Now I would like to do something similar with my own images. I cannot figure out how to do that. The tutorial does this:<\/p>\n\n<pre><\/pre>\n\n<p>What precisely should batch_xs, batch_ys be when I create it from my images?<\/p>\n\n<p>I see there is a library ImageFlow, and this seems to do exactly what I want, but I cannot figure out how to use that either. The description says I should call <\/p>\n\n<pre><\/pre>\n\n<p>but that does not even contain a path to my images.<\/p>\n\n<p>Many thanks for your attention.<\/p>\n","answers":[{"AnswerId":"35871985","CreationDate":"2016-03-08T15:55:45.073","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Looking at the implementation of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c54b03ad9c6bbb40d6a3adb2ca6a43333cde87e0\/tensorflow\/examples\/tutorials\/mnist\/input_data.py#L151\" rel=\"nofollow\"><code>mnist.train.next_batch()<\/code><\/a>, it appears that <code>batch_xs<\/code> should be a matrix of size <code>batch_size<\/code> x <code>num_pixels<\/code>, and <code>batch_ys<\/code> should be either a matrix of size <code>batch_size<\/code> x <code>num_classes<\/code> (if <code>one_hot<\/code> is True) or a vector of length <code>batch_size<\/code> (otherwise).<\/p>\n\n<p>When you call <code>mnist.train.next_batch(100)<\/code>, <code>batch_xs<\/code> will be <code>100<\/code> x <code>784<\/code>, and <code>batch_ys<\/code> will be <code>100<\/code> x <code>10<\/code>. For your own application you will probably want to change both the number of pixels and the number of classes.<\/p>\n"}]}
{"QuestionId":35864198,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-08T09:57:48.730","AcceptedAnswerId":35871863.0,"OwnerUserId":234167.0,"Title":"Why doesn't `tf.image.resize_images` set the image shape?","Body":"<p>My objective is to introduce random scaling and translation for data augmentation purposes.<\/p>\n\n<pre><\/pre>\n\n<p>This fails with  Swapping the lines works, but doesn't do what I really need.<\/p>\n\n<pre><\/pre>\n\n<p>So it seems like resize_images loses the shape of the image tensor and then  fails. Is this on purpose, am I missing something? How come  works after a resize but  doesn't?<\/p>\n","answers":[{"AnswerId":"35871863","CreationDate":"2016-03-08T15:49:44.350","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.image.resize_images()<\/code> op <em>does<\/em> set the image shape, on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/34880c4ca1a83b443b44dc415b74deb217c42257\/tensorflow\/python\/ops\/image_ops.py#L675\" rel=\"nofollow noreferrer\">this line<\/a> of the implementation. (This was added in TensorFlow 0.7.) <\/p>\n\n<p>However, if either of the <code>new_height<\/code> or <code>new_width<\/code> arguments is a dynamic value, then TensorFlow cannot infer a single shape for that dimension, and so uses <code>None<\/code> for that dimension. I notice in your code that the new height and width values are called <code>random_scale<\/code>: if a new random value is drawn on each step, then the shape will have <code>None<\/code> for the height and width dimensions.<\/p>\n\n<p>Note that in this case, the <code>tf.image.crop_to_bounding_box()<\/code> op will not work because&mdash;as the error message indicates&mdash;the current implementation requires that the shape of the input be fully defined. As I noted in <a href=\"https:\/\/stackoverflow.com\/a\/35777637\/3574081\">a recent answer<\/a>, the best workaround might be to use the lower level ops from which <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#crop_to_bounding_box\" rel=\"nofollow noreferrer\"><code>tf.image.crop_to_bounding_box()<\/code><\/a> is implemented (in particular <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#slice\" rel=\"nofollow noreferrer\"><code>tf.slice()<\/code><\/a> with computed indices).<\/p>\n"}]}
{"QuestionId":35869137,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-08T13:47:42.113","AcceptedAnswerId":null,"OwnerUserId":3441991.0,"Title":"Avoid tensorflow print on standard error","Body":"<p>anyone knows if there is a method to prevent tensorflow from polluting standard error with gpus' memory allocation log?.\nI noted that when the following command is executed:<\/p>\n\n<pre><\/pre>\n\n<p>tensorflow prints on standard error a log about memory and gpu resources allocation. Something like:<\/p>\n\n<pre><\/pre>\n\n<p>For important reasons, I wanna avoid this printing.<\/p>\n","answers":[{"AnswerId":"40982782","CreationDate":"2016-12-05T20:30:58.290","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This was recently fixed, and should be available if you upgrade to TensorFlow 0.12 or later.<\/p>\n\n<p>To disable all logging output from TensorFlow, set the following environment  variable before launching Python:<\/p>\n\n<pre><code>$ export TF_CPP_MIN_LOG_LEVEL=3\n$ python ...\n<\/code><\/pre>\n\n<p>You can also adjust the verbosity by changing the value of <code>TF_CPP_MIN_LOG_LEVEL<\/code>:<\/p>\n\n<ul>\n<li>0 = all messages are logged (default behavior)<\/li>\n<li>1 = <code>INFO<\/code> messages are not printed<\/li>\n<li>2 = <code>INFO<\/code> and <code>WARNING<\/code> messages are not printed<\/li>\n<li>3 = <code>INFO<\/code>, <code>WARNING<\/code>, and <code>ERROR<\/code> messages are not printed<\/li>\n<\/ul>\n"},{"AnswerId":"47305903","CreationDate":"2017-11-15T11:13:31.917","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<blockquote>\n  <p>Defaults to 0, so all logs are shown. Set TF_CPP_MIN_LOG_LEVEL to 1 to filter out INFO logs, 2 to additionall filter out WARNING, 3 to additionally filter out ERROR.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":35875652,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><lstm>","CreationDate":"2016-03-08T18:54:48.127","AcceptedAnswerId":null,"OwnerUserId":3921875.0,"Title":"LSTM inputs for Tensorflow","Body":"<p>I'm trying to create an LSTM network in Tensorflow and I'm lost in terminology\/basics. I have <em>n<\/em> time series examples so <strong>X<\/strong>=<strong>x<\/strong><sub>n<\/sub>, where \n<strong>x<\/strong><sub>i<\/sub>=[[x<sub>1<\/sub><sup>1<\/sup>x<sub>1<\/sub><sup>2<\/sup>,x<sub>1<\/sub><sup>3<\/sup>],...,[x<sub>m<\/sub><sup>1<\/sup>x<sub>m<\/sub><sup>2<\/sup>,x<sub>m<\/sub><sup>3<\/sup>]] and where x<sub>i<\/sub><sup>i<\/sup> is a float. First of all I want to train a model that given the start of a sequence ([x<sub>1<\/sub><sup>1<\/sup>x<sub>1<\/sub><sup>2<\/sup>,x<sub>1<\/sub><sup>3<\/sup>]) I can predict the rest of the sequence. Then later I hope to include a classifier to predict which binary class each <strong>x<\/strong><sub>i<\/sub> belongs to.<\/p>\n\n<p>So my problem is what do I feed in to the start and pull out the end of my model? So far I have something that looks like the below<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT<\/strong>:\nSpecifically, how to I finish the  function so that it is compatible with my data?<\/p>\n","answers":[{"AnswerId":"35926781","CreationDate":"2016-03-10T20:38:10.920","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>An RNN predicts the value of N+1 given the values from 1 to N so far. (LSTM is just one way to implement an RNN cell.)<\/p>\n\n<p>The short answer is:<\/p>\n\n<ul>\n<li>train your model using back propagation on your complete sequences [[x<sub>1<\/sub><sup>1<\/sup>x<sub>1<\/sub><sup>2<\/sup>,x<sub>1<\/sub><sup>3<\/sup>],...,[x<sub>m<\/sub><sup>1<\/sup>x<sub>m<\/sub><sup>2<\/sup>,x<sub>m<\/sub><sup>3<\/sup>]]<\/li>\n<li>run your trained model forward on the start of your sequence [x<sub>1<\/sub><sup>1<\/sup>x<sub>1<\/sub><sup>2<\/sup>,x<sub>1<\/sub><sup>3<\/sup>,...] then sample from the model to predict the rest of your sequence [x<sub>m<\/sub><sup>1<\/sup>x<sub>m<\/sub><sup>2<\/sup>,x<sub>m<\/sub><sup>3<\/sup>,...]. <\/li>\n<\/ul>\n\n<p>The longer answer is:<\/p>\n\n<p>Your example just shows the initialization of the model. You also need to implement a training function to run back propagation as well as a sample function that predicts the results.<\/p>\n\n<p>The following code snippets are mix &amp; match and are for illustration purposes only...<\/p>\n\n<p>For training just feed in your complete sequences with start + rest in your data iterator.<\/p>\n\n<p>For example in the sample code tensorflow\/models\/rnn\/ptb_word_lm.py the training loop computes a cost function for batches of input_data against targets (which are the input_data shifted by one timestep) <\/p>\n\n<pre><code>        # compute a learning rate decay\n        session.run(tf.assign(self.learning_rate_variable, learning_rate))\n\n        logger.info(\"Epoch: %d Learning rate: %.3f\" % (i + 1, session.run(self.learning_rate_variable)))\n\n\n        \"\"\"Runs the model on the given data.\"\"\"\n        epoch_size = ((len(training_data) \/\/ self.batch_size) - 1) \/\/ self.num_steps\n        costs = 0.0\n        iters = 0\n        state = self.initial_state.eval()\n        for step, (x, y) in enumerate(self.data_iterator(training_data, self.batch_size, self.num_steps)):\n\n            # x and y should have shape [batch_size, num_steps]\n            cost, state, _ = session.run([self.cost_function, self.final_state, self.train_op],\n                                     {self.input_data: x,\n                                      self.targets: y,\n                                      self.initial_state: state})\n            costs += cost\n            iters += self.num_steps\n<\/code><\/pre>\n\n<p>Note the data iterator in tensorflow\/models\/rnn\/reader.py returns the input data as 'x' and the targets as 'y' which are just shifted one step forward from x. (You would need to create a data iterator like this that packages your set of training sequences.) <\/p>\n\n<pre><code>def ptb_iterator(raw_data, batch_size, num_steps):\n  raw_data = np.array(raw_data, dtype=np.int32)\n\n  data_len = len(raw_data)\n  batch_len = data_len \/\/ batch_size\n  data = np.zeros([batch_size, batch_len], dtype=np.int32)\n  for i in range(batch_size):\n    data[i] = raw_data[batch_len * i:batch_len * (i + 1)]\n\n  epoch_size = (batch_len - 1) \/\/ num_steps\n\n  if epoch_size == 0:\n    raise ValueError(\"epoch_size == 0, decrease batch_size or num_steps\")\n\n  for i in range(epoch_size):\n    x = data[:, i*num_steps:(i+1)*num_steps]\n    y = data[:, i*num_steps+1:(i+1)*num_steps+1]\n    yield (x, y)\n<\/code><\/pre>\n\n<p>After training, you run the model forward to make predictions for sequences by feeding in the start of your sequence start_x=[X1, X2, X3,...]...this snippets assumes binary values representing classes, you'd have to adjust the sampling function for float values.<\/p>\n\n<pre><code>def sample(self, sess, num=25, start_x):\n\n    # return state tensor with batch size 1 set to zeros, eval\n    state = self.rnn_layers.zero_state(1, tf.float32).eval()\n\n    # run model forward through the start of the sequence\n    for char in start_x:\n\n        # create a 1,1 tensor\/scalar set to zero\n        x = np.zeros((1, 1))\n\n        # set to the vocab index\n        x[0, 0] = char\n\n\n        # fetch: final_state\n        # input_data = x, initial_state = state\n        [state] = sess.run([self.final_state], {self.input_data: x, self.initial_state:state})\n\n    def weighted_pick(weights):\n\n        # an array of cummulative sum of weights\n        t = np.cumsum(weights)\n\n        # scalar sum of tensor\n        s = np.sum(weights)\n\n        # randomly selects a value from the probability distribution\n        return(int(np.searchsorted(t, np.random.rand(1)*s)))\n\n    # PREDICT REST OF SEQUENCE\n    rest_x = []\n\n    # get last character in init\n    char = start_x[-1]\n\n    # sample next num chars in the sequence after init\n    score = 0.0\n\n    for n in xrange(num):\n\n        # init input to zeros\n        x = np.zeros((1, 1))\n\n        # lookup character index\n        x[0, 0] = char\n\n        # probs = tf.nn.softmax(self.logits)\n        # fetch: probs, final_state\n        # input_data = x, initial_state = state\n        [probs, state] = sess.run([self.output_layer, self.final_state], {self.input_data: x, self.initial_state:state})\n\n        p = probs[0]\n        logger.info(\"output=%s\" % np.shape(p))\n        # sample = int(np.random.choice(len(p), p=p))\n\n        # select a random value from the probability distribution\n        sample = weighted_pick(p)\n        score += p[sample]\n        # look up the key with the index\n        logger.debug(\"sample[%d]=%d\" % (n, sample))\n        pred = self.vocabulary[sample]\n        logger.debug(\"pred=%s\" % pred)\n\n        # add the car to the output\n        rest_x.append(pred) \n\n        # set the next input character\n        char = pred\n    return rest_x, score\n<\/code><\/pre>\n"}]}
{"QuestionId":35876365,"AnswerCount":1,"Tags":"<python><command-line><centos><theano><keras>","CreationDate":"2016-03-08T19:30:03.240","AcceptedAnswerId":35925267.0,"OwnerUserId":5256563.0,"Title":"The neural networks is compiled but the training does not start","Body":"<p>I wanted to train a neural network using keras, but the training does not start, and it happens in really weird situations.<\/p>\n\n<p>Here is a portion of my code:<\/p>\n\n<pre><\/pre>\n\n<p>If I run this code in the terminal using , everything goes fine and the training starts.<\/p>\n\n<p>But if I run this code in the terminal using , , or if I redirect the outputs in my code , then the program stall for the  method, so the learning does not start. The is displayed, but nothing more, no training.<\/p>\n\n<p>I took a look and it seems that the program keeps calling this function <\/p>\n\n<p>Any ideas why it does not start when I redirect the outputs?<\/p>\n","answers":[{"AnswerId":"35925267","CreationDate":"2016-03-10T19:14:50.393","ParentId":null,"OwnerUserId":"5256563","Title":null,"Body":"<p>Finally the process starts but nothing is displayed into the terminal or written into the output file until the training is completely over (all the epoch performed).<\/p>\n\n<p>I found out by starting many different small trainings with just few epochs, and everything was displayed only when all the trainings were over.<\/p>\n"}]}
{"QuestionId":35878229,"AnswerCount":1,"Tags":"<neural-network><gpu><theano>","CreationDate":"2016-03-08T21:19:55.310","AcceptedAnswerId":null,"OwnerUserId":6036456.0,"Title":"Theano Rarely Using GPU","Body":"<p>Just getting started with Theano and my GTX 960, but when training a network (specifically recurrent regression) and GPU usage spikes up for less than a second, maybe two or three times throughout the several minutes it takes. Is there a way to force it to actually use the GPU the entire time? <\/p>\n\n<p>Seems to take an obscenely long time for not much data.<\/p>\n","answers":[{"AnswerId":"35878412","CreationDate":"2016-03-08T21:29:54.687","ParentId":null,"OwnerUserId":"2773607","Title":null,"Body":"<p>What is actually happening in your code is that much of the time is spent getting data onto and off of the GPU, but any GPU monitor will show that the GPU is \"not being used\" during this time, even though what it is doing is loading data. You can't avoid doing this, other than being smart about how and when you load data onto the GPU. The spikes you are describing are completely normal - that is when the GPU actually does the computation. Nothing you can do!<\/p>\n"}]}
{"QuestionId":35878951,"AnswerCount":0,"Tags":"<python-3.x><theano>","CreationDate":"2016-03-08T22:04:33.203","AcceptedAnswerId":null,"OwnerUserId":6036622.0,"Title":"Theano doesn't load on Windows","Body":"<p>I have an issue when importing Theano into a python script, I cloned the latest dev version as the 0.7.0 version didn't work either:<\/p>\n\n<p><strong>error message:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35879593,"AnswerCount":0,"Tags":"<shared-ptr><caffe>","CreationDate":"2016-03-08T22:49:22.587","AcceptedAnswerId":null,"OwnerUserId":5348390.0,"Title":"Caffe Boost Shared_Ptr SegFault in Net::Init?","Body":"<p>Caffe crashes using Boost shared_ptr inside Net::Init<\/p>\n\n<pre><\/pre>\n\n<p>Note the BAD 'this' Pointer for RNG (0x1):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qwIcy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qwIcy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It actually dies inside BaseConvulutionLayer::Setup --> XavierFiller::Fill --> Caffe::RNG (corrupt this pointer)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Yrbg2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Yrbg2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h2>Compiler\/Linker Flags<\/h2>\n\n<p>g++ mytest.cpp -o testcaffe -lcaffe -lglog -lopencv_core -lopencv_highgui -lopencv_imgproc -lboost_system -std=c++11 -g -O0<\/p>\n\n<p>The same version of libcaffe.so works with a previously-compiled application using the same model files (prototxt, etc.)<\/p>\n\n<p>Could this be the result of a mismatch of g++ build\/link flags and the version of boost?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":35881260,"AnswerCount":1,"Tags":"<python><tensorflow><sage>","CreationDate":"2016-03-09T01:24:24.057","AcceptedAnswerId":35895336.0,"OwnerUserId":5049786.0,"Title":"Using TensorFlow with Sage","Body":"<p>I've written something in TensorFlow that makes use of some nice group theory functions that work very easily in Sage (and seem prohibitively difficult to code from scratch). The Sage part works on its own, and the TensorFlow part works on its own, but I can't figure out how to get them working together.<\/p>\n\n<p>Specifically: I can make a file test.py using Sage functions and run it from the command line using:<\/p>\n\n<pre><\/pre>\n\n<p>with no problem. But calling a function defined in test.py from a .py file using TensorFlow fails (\"Import error, no module named Sage\"), presumably because Sage (6.x) uses Python 2.6.x, while TensorFlow uses Python 2.7 or 3.3+.<\/p>\n\n<p>Is there a way around this?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>EDIT: I'm not sure if this is relevant, but if I fire up normal Python (the kind TensorFlow uses), I get this:<\/p>\n\n<pre><\/pre>\n\n<p>which outputs .<\/p>\n\n<p>However if I fire up Sage first I get this:<\/p>\n\n<pre><\/pre>\n\n<p>which outputs '.<\/p>\n\n<p>I just upgraded to Sage 7.0 if that matters (this didn't work in 6.10 either, though).<\/p>\n","answers":[{"AnswerId":"35895336","CreationDate":"2016-03-09T15:10:02.427","ParentId":null,"OwnerUserId":"782821","Title":null,"Body":"<p>Here's something <strong>NOT<\/strong> to do (yet); don't just take whatever Sage install you happen to have and do:<\/p>\n\n<pre><code>$ sage -pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\n<\/code><\/pre>\n\n<p>Even though this \"works\", it also had several worrying messages about upgrading numpy and six, which completely broke the numpy part of my Sage installation.  This was with Sage-6.9.<\/p>\n\n<p>Which means you have to make sure you have a Sage that has the right versions of Numpy and six.  With the latest development version, we do, apparently:<\/p>\n\n<pre><code>$ cd \/path\/to\/my\/bleeding\/edge\/sage\/directory\n$ .\/sage -pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\nCollecting tensorflow==0.7.1 from https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\n  Using cached https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\nRequirement already satisfied (use --upgrade to upgrade): six&gt;=1.10.0 in .\/local\/lib\/python2.7\/site-packages\/six-1.10.0-py2.7.egg (from tensorflow==0.7.1)\nCollecting protobuf==3.0.0b2 (from tensorflow==0.7.1)\n  Using cached protobuf-3.0.0b2-py2.py3-none-any.whl\nCollecting wheel (from tensorflow==0.7.1)\n  Using cached wheel-0.29.0-py2.py3-none-any.whl\nRequirement already satisfied (use --upgrade to upgrade): numpy&gt;=1.10.1 in .\/local\/lib\/python2.7\/site-packages (from tensorflow==0.7.1)\nRequirement already satisfied (use --upgrade to upgrade): setuptools in .\/local\/lib\/python2.7\/site-packages\/setuptools-20.1.1-py2.7.egg (from protobuf==3.0.0b2-&gt;tensorflow==0.7.1)\nInstalling collected packages: protobuf, wheel, tensorflow\nSuccessfully installed protobuf-3.0.0b2 tensorflow-0.7.1 wheel-0.29.0\nYou are using pip version 8.0.2, however version 8.1.0 is available.\nYou should consider upgrading via the 'pip install --upgrade pip' command.\n<\/code><\/pre>\n\n<p>And then I don't get any failures.  <\/p>\n\n<p>So one has to be careful, but apparently it's possible.  However, you definitely have to use it from \"within\" Sage; Sage-as-distribution wouldn't allow you to use your other tensorflow with it.<\/p>\n"}]}
{"QuestionId":35881370,"AnswerCount":1,"Tags":"<numpy><lua><torch>","CreationDate":"2016-03-09T01:36:09.633","AcceptedAnswerId":35886539.0,"OwnerUserId":3204676.0,"Title":"Lua Torch equivalent of np.where()?","Body":"<p>I have a  and want to grab the indices where there is a . In numpy, I could do something like<\/p>\n\n<pre><\/pre>\n\n<p>which would return . Is this functionality defined in Torch?<\/p>\n\n<p>(In my particular case, I want to use these indices to index into several different Tensor objects, but it'd be nice to know how to do this in general.)<\/p>\n","answers":[{"AnswerId":"35886539","CreationDate":"2016-03-09T08:33:11.500","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>You can use <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#longtensor-nonzerotensor\" rel=\"nofollow\">torch.nonzero<\/a>, e.g.:<\/p>\n\n<pre><code>&gt; a = torch.ByteTensor{1,0,1,0,1}\n&gt; print(torch.nonzero(a))                                                                                         \n 1                                                                                                                  \n 3                                                                                                                  \n 5                                                                                                                  \n[torch.LongTensor of size 3x1]\n<\/code><\/pre>\n\n<p>If you really need to find the 1-s only you can chain a logical operator:<\/p>\n\n<pre><code>&gt; a = torch.ByteTensor{1,2,1,6,1}\n&gt; a:eq(1):nonzero()\n<\/code><\/pre>\n"}]}
{"QuestionId":35881949,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2016-03-09T02:45:55.217","AcceptedAnswerId":null,"OwnerUserId":4625712.0,"Title":"the input shape of array about Keras on Tensorflow","Body":"<p>I have a question about the 4D tensor on keras about Convolution2D Layers.<\/p>\n\n<p>The Keras doc says:\n4D tensor with shape: (samples, channels, rows, cols) if dim_ordering='th' or 4D tensor with shape: (samples, rows, cols, channels) if dim_ordering='tf'.<\/p>\n\n<p>I use 'tf', how about my input? When I use , it is ok, but when I use  as input, it has some problems.<\/p>\n","answers":[{"AnswerId":"43664585","CreationDate":"2017-04-27T17:41:39.020","ParentId":null,"OwnerUserId":"1453258","Title":null,"Body":"<p><code>Keras<\/code> assumes that if you are using tensorflow, you are going with <code>(samples, channels, rows, cols)<\/code><\/p>\n"}]}
{"QuestionId":35884045,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><tensorflow>","CreationDate":"2016-03-09T05:58:20.447","AcceptedAnswerId":35884344.0,"OwnerUserId":1257953.0,"Title":"Creating a float64 Variable in tensorflow","Body":"<p>I am trying to do a logistic regression and my training dataset is derived from a a numpy float64 array. My code looks like, <\/p>\n\n<pre><\/pre>\n\n<p>This throws an exception<\/p>\n\n<pre><\/pre>\n\n<p>This probably because W and b are float32 and not float64. Is there a way to convert W and b OR create it as a float64<\/p>\n","answers":[{"AnswerId":"35884344","CreationDate":"2016-03-09T06:17:43.733","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To make this work, you should define the <code>W<\/code> and <code>b<\/code> variables with <code>tf.float64<\/code> initial values. The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#truncated_normal\" rel=\"noreferrer\"><code>tf.truncated_normal()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#zeros\" rel=\"noreferrer\"><code>tf.zeros()<\/code><\/a> ops each take an optional <code>dtype<\/code> argument that can be set to <code>tf.float64<\/code> as follows:<\/p>\n\n<pre><code>    W = tf.Variable(tf.truncated_normal([115713, 2], dtype=tf.float64))\n    b = tf.Variable(tf.zeros([2], dtype=tf.float64))\n<\/code><\/pre>\n"}]}
{"QuestionId":35885648,"AnswerCount":1,"Tags":"<indexing><gpu><theano>","CreationDate":"2016-03-09T07:38:36.370","AcceptedAnswerId":null,"OwnerUserId":6037919.0,"Title":"Shared variable indexing error in theano","Body":"<p>There's problem on shared variable indexing.<\/p>\n\n<p>The shape of train_set_x is (n, 3, 50176) and n is larger than index<\/p>\n\n<p>I can get value from shared variable train_set_x with following code<\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n\n<p>but I can't get value with following code<\/p>\n\n<pre><\/pre>\n\n<p>It shows an error message<\/p>\n\n<pre><\/pre>\n\n<p>What is the difference between them? and What would be the way to use the the value in train_set_x with theano.function?<\/p>\n","answers":[{"AnswerId":"35918495","CreationDate":"2016-03-10T13:55:29.270","ParentId":null,"OwnerUserId":"3086290","Title":null,"Body":"<p>The <code>Inputs values: ['not shown', 1039]<\/code> shows value passed as index argument (it is 1039) and the <code>Inputs shapes: [(1039, 3, 50176), ()]<\/code> line shows that your <code>n<\/code> is 1039. <\/p>\n\n<p>So it looks like you are really out of bounds as theano says.\nThe following sample works for me:<\/p>\n\n<pre><code>train_set_x = theano.shared(np.random.rand(10, 3, 3))\nx = theano.tensor.matrix()\nindex = theano.tensor.iscalar()\ncheck = theano.function([index], x, givens = {x : train_set_x[index]}\ncheck(3)\n<\/code><\/pre>\n\n<p>While <code>check(10)<\/code> displays very similar error:<\/p>\n\n<pre><code>IndexError: index out of bounds\nApply node that caused the error: Subtensor{int32}(&lt;TensorType(float64, 3D)&gt;, ScalarFromTensor.0)\nToposort index: 1\nInputs types: [TensorType(float64, 3D), Scalar(int32)]\nInputs shapes: [(10, 3, 3), ()]\nInputs strides: [(72, 24, 8), ()]\nInputs values: ['not shown', 10]\nOutputs clients: [[DeepCopyOp(Subtensor{int32}.0)]]\n\nHINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node\n<\/code><\/pre>\n"}]}
{"QuestionId":35886943,"AnswerCount":4,"Tags":"<linux><caffe>","CreationDate":"2016-03-09T08:54:32.993","AcceptedAnswerId":null,"OwnerUserId":5837113.0,"Title":"caffe:\/usr\/bin\/ld: cannot find -lcblas","Body":"<p>I have installed BLAS in my CentOS7(64 bit).But when my use  in my <\/p>\n\n<p>'caffe'.It reports an eror:<\/p>\n\n<pre><\/pre>\n\n<p>I dont know why this happened and how to resolve it.<\/p>\n","answers":[{"AnswerId":"41715841","CreationDate":"2017-01-18T09:36:52.463","ParentId":null,"OwnerUserId":"2455061","Title":null,"Body":"<p>1.Download atlas from <a href=\"https:\/\/sourceforge.net\/projects\/math-atlas\/files\/latest\/download?source=files\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>2.Go to the root directory of the unzipped atlas. Install it using the instruction:<\/p>\n\n<pre><code>mkdir build\ncd build \n..\/configure --shared\nmake\nmake install\n<\/code><\/pre>\n\n<p>3.Edit the <code>Makefile.config<\/code> file of caffe:<\/p>\n\n<pre><code>BLAS_LIB := \/usr\/local\/atlas\/lib\/\n<\/code><\/pre>\n\n<p><code>\/usr\/local\/atlas\/lib\/<\/code> contains <code>libtatlas.so<\/code><\/p>\n"},{"AnswerId":"35893432","CreationDate":"2016-03-09T13:48:31.223","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p><code>sudo yum install atlas-devel<\/code> to install the ATLAS library and try compiling Caffe again.<\/p>\n"},{"AnswerId":"48260138","CreationDate":"2018-01-15T09:29:28.513","ParentId":null,"OwnerUserId":"9218673","Title":null,"Body":"<p>I ended up using OpenBLAS to work around this problem.<\/p>\n\n<pre><code>yum install openblas-devel\nsed -i 's\/BLAS := atlas\/BLAS := open\/' Makefile.config\nsed -i '\/BLAS := open\/aBLAS_INCLUDE := \\\/usr\\\/include\\\/openblas\/' Makefile.config\n<\/code><\/pre>\n\n<p>which means install OpenBLAS and change <code>BLAS := altas<\/code> to <code>BLAS := open<\/code>\nand add <code>BLAS_INCLUDE := \/usr\/include\/openblas\/<\/code> in <code>Makefile.config<\/code>. Once you have made these changes, rebuild.<\/p>\n"},{"AnswerId":"40099055","CreationDate":"2016-10-18T03:32:52.200","ParentId":null,"OwnerUserId":"3526875","Title":null,"Body":"<p>First issue<\/p>\n\n<pre><code>sudo yum install atlas-devel\n<\/code><\/pre>\n\n<p>and then add the following lines to your Makefile.config<\/p>\n\n<pre><code>BLAS_INCLUDE := \/usr\/include\nBLAS_LIB := \/usr\/lib64\/atlas\n<\/code><\/pre>\n\n<p>in the target shared library section. The section after adding the change looks like the following<\/p>\n\n<pre><code> 36 # The target shared library name\n 37 LIBRARY_NAME := $(PROJECT)\n 38 LIB_BUILD_DIR := $(BUILD_DIR)\/lib\n 39 STATIC_NAME := $(LIB_BUILD_DIR)\/lib$(LIBRARY_NAME).a\n 40 DYNAMIC_VERSION_MAJOR       := 1\n 41 DYNAMIC_VERSION_MINOR       := 0\n 42 DYNAMIC_VERSION_REVISION    := 0-rc3\n 43 DYNAMIC_NAME_SHORT := lib$(LIBRARY_NAME).so\n 44 #DYNAMIC_SONAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR)\n 45 DYNAMIC_VERSIONED_NAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR).$(DYNAMIC_VERSION_    MINOR).$(DYNAMIC_VERSION_REVISION)\n 46 DYNAMIC_NAME := $(LIB_BUILD_DIR)\/$(DYNAMIC_VERSIONED_NAME_SHORT)\n 47 COMMON_FLAGS += -DCAFFE_VERSION=$(DYNAMIC_VERSION_MAJOR).$(DYNAMIC_VERSION_MINOR).$(DYNAMIC_VERSI    ON_REVISION)\n 48 \n 49 BLAS_INCLUDE := \/usr\/include\n 50 BLAS_LIB := \/usr\/lib64\/atlas\n<\/code><\/pre>\n\n<p>I also faced the same problem when building caffe on the amazon deep learning ami and found the solution here: \n<a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/Pyfp9eQoIMQ\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/Pyfp9eQoIMQ<\/a><\/p>\n"}]}
{"QuestionId":35890051,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-09T11:16:09.243","AcceptedAnswerId":35897221.0,"OwnerUserId":2864407.0,"Title":"How does a static member in a TensorFlow OpKernel behave?","Body":"<p>Specifically, if I have an Op inheriting from OpKernel, and inside this I have some private member declared . Am I safe to assume that all instantiations of this kernel will access the same static member? Are all kernels placed on the same machine\/worker (possibly different CPUs) in the same address space \/ process? Obviously, this would not apply to kernels mapped to separate machines in a distributed setting. E.g.<\/p>\n\n<pre><\/pre>\n\n<p>Finally, when implementing an OpKernel, is there a better way than  to share a chunk of read-only data in memory between OpKernels of the same type? The data is too large to house multiple copies in memory. <\/p>\n","answers":[{"AnswerId":"35897221","CreationDate":"2016-03-09T16:29:42.990","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>A <code>static<\/code> member of a <code>tensorflow::OpKernel<\/code> is shared between all instances of that kernel in the same process. If you are running a single TensorFlow process, then all instances (including instances assigned to CPU or GPU, and instances from different sessions) will share the same <code>static<\/code> member.<\/p>\n\n<p>Using a <code>static<\/code> member does not handle destruction when the session is closed and doesn't permit different objects in different sessions, so I wouldn't recommend this style; however the alternative is a bit more complicated. For shared state, we typically use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/resource_mgr.h\" rel=\"nofollow\"><code>ResourceMgr<\/code><\/a> (via <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7022c01ac7282eed90db6b2c52195819264bc146\/tensorflow\/core\/framework\/op_kernel.h#L854\" rel=\"nofollow\"><code>OpKernelContext::resource_manager()<\/code><\/a>) to store objects. The typical implementation uses \"constructor\" op that instantiates the object on first use, owns it, and outputs a (string) handle. Users of the shared state take a data dependency on the handle, and look it up in the <code>ResourceMgr<\/code> to access the object. The object is then deleted when the \"constructor\" op is deleted. This is the approach used for stateful ops like <code>tf.Variable<\/code>, <code>tf.FIFOQueue<\/code>, etc.<\/p>\n"}]}
{"QuestionId":35892412,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-03-09T13:03:35.170","AcceptedAnswerId":null,"OwnerUserId":2666216.0,"Title":"Tensorflow dense gradient explanation?","Body":"<p>I recently implemented a model and when I ran it I received this warning:<\/p>\n\n<pre><\/pre>\n\n<p>With some similar parameter settings (embedding dimensionalities) suddenly the model is ridiculously slow.<\/p>\n\n<ol>\n<li>What does this warning imply? It appears that something I've done has caused all of the gradients to be dense and so backprop is doing dense matrix computations<\/li>\n<li>If it's that there is an issue with the model that's causing this, how can I identify it and fix it?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"35896823","CreationDate":"2016-03-09T16:13:17.843","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This warning is printed when a sparse <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/IndexedSlices\" rel=\"noreferrer\"><code>tf.IndexedSlices<\/code><\/a> object is implicitly converted to a dense <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a>. This typically happens when one op (usually <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"noreferrer\"><code>tf.gather()<\/code><\/a>) backpropagates a sparse gradient, but the op that receives it does not have a specialized gradient function that can handle sparse gradients. As a result, TensorFlow automatically densifies the <code>tf.IndexedSlices<\/code>, which can have a devastating effect on performance if the tensor is large. <\/p>\n\n<p>To fix this problem, you should try to ensure that the <code>params<\/code> input to <code>tf.gather()<\/code> (or the <code>params<\/code> inputs to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#embedding_lookup\" rel=\"noreferrer\"><code>tf.nn.embedding_lookup()<\/code><\/a>) is a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a>. Variables can receive the sparse updates directly, so no conversion is needed. Although <code>tf.gather()<\/code> (and <code>tf.nn.embedding_lookup()<\/code>) accept arbitrary tensors as inputs, this may lead to a more complicated backpropagation graph, resulting in implicit conversion.<\/p>\n"},{"AnswerId":"35893467","CreationDate":"2016-03-09T13:50:13.920","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>A dense Tensor can be thought of like a standard python array. A sparse one can be thought of as a collection of indices and values e.g.<\/p>\n\n<pre><code># dense\narray = ['a', None, None, 'c']\n\n# sparse\narray = [(0, 'a'), (3, 'c')]\n<\/code><\/pre>\n\n<p>So as you can see if you have a lot of empty entries a sparse array will be much more efficient than a dense one. But if all entries are filled in, dense is far more efficient. In your case somewhere in the tensor flow graph a sparse array is being converted to a dense one of indeterminate size. The warning is just saying it is possible that you can waste a lot of memory like this. But it might not be a problem at all if the sparse array is not too big\/already quite dense.<\/p>\n\n<p>If you want to diagnose it I would advise naming your various tensor objects then it will print exactly which ones are being used in this conversion and you can work out what you might be able to adjust to remove it.<\/p>\n"},{"AnswerId":"41542814","CreationDate":"2017-01-09T07:16:08.040","ParentId":null,"OwnerUserId":"5963300","Title":null,"Body":"<p>Totally agree with the answer of <code>mrry<\/code>.<\/p>\n\n<p>Actually I will post another solution for this problem.<\/p>\n\n<p>You could use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/dynamic_partition\" rel=\"nofollow noreferrer\"><code>tf.dynamic_partition()<\/code><\/a> instead of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather()<\/code><\/a> to eliminate the warning.<\/p>\n\n<p>The example code is below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Create the cells for the RNN network\nlstm = tf.nn.rnn_cell.BasicLSTMCell(128)\n\n# Get the output and state from dynamic rnn\noutput, state = tf.nn.dynamic_rnn(lstm, sequence, dtype=tf.float32, sequence_length = seqlen)\n\n# Convert output to a tessor and reshape it\noutputs = tf.reshape(tf.pack(output), [-1, lstm.output_size])\n\n# Set partions to 2\nnum_partitions = 2\n\n# The partitions argument is a tensor which is already fed to a placeholder.\n# It is a 1-D tensor with the length of batch_size * max_sequence_length.\n# In this partitions tensor, you need to set the last output idx for each seq to 1 and \n# others remain 0, so that the result could be separated to two parts,\n# one is the last outputs and the other one is the non-last outputs.\nres_out = tf.dynamic_partition(outputs, partitions, num_partitions)\n\n# prediction\npreds = tf.matmul(res_out[1], weights) + bias\n<\/code><\/pre>\n\n<p>Hope this could help you.<\/p>\n"}]}
{"QuestionId":35893278,"AnswerCount":1,"Tags":"<python><visual-studio><visual-studio-2015><tensorflow><canopy>","CreationDate":"2016-03-09T13:42:37.910","AcceptedAnswerId":35908016.0,"OwnerUserId":3559686.0,"Title":"TensorFlow in Visual Studio 2015, using Canopy as the Python environment","Body":"<p>I followed the instructions on <a href=\"https:\/\/discussions.udacity.com\/t\/windows-tensorflow-and-visual-studio-2015\/45636\" rel=\"nofollow\">this page<\/a> and managed to install <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"nofollow\">TensorFlow<\/a> via pip within Visual Studio 2015. My Python environment runs in <a href=\"https:\/\/www.enthought.com\/products\/canopy\/\" rel=\"nofollow\">Enthought Canopy<\/a>. The installation was successful. I refreshed the DB, and auto-completion via IntelliSense was available when I wrote the following line:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I run it in the REPL, I see the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone shed some light on why I get the message  even though the installation of  took place without a hiccup? Any suggestions on how I can resolve this issue?<\/p>\n","answers":[{"AnswerId":"35908016","CreationDate":"2016-03-10T05:06:38.740","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The current version of TensorFlow does not support running on Windows. In particular, the PIP packages available <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#pip-installation\" rel=\"nofollow noreferrer\">here<\/a> are for Linux and Mac OS X only: they include a native compiled library called <code>_pywrap_tensorflow<\/code> that has not been compiled for Windows.<\/p>\n\n<p>As <a href=\"https:\/\/stackoverflow.com\/a\/33618580\/3574081\">suggested in another answer<\/a>, you can run TensorFlow on Windows using a Docker container. There is an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow noreferrer\">open issue on GitHub<\/a> concerning better Windows support, but this will depend on Bazel adding support for Windows build (currently <a href=\"http:\/\/bazel.io\/roadmap.html\" rel=\"nofollow noreferrer\">targeted<\/a> for later this year).<\/p>\n"}]}
{"QuestionId":35895257,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-09T15:05:44.410","AcceptedAnswerId":null,"OwnerUserId":6040027.0,"Title":"Error in Tensorflow when run \"mnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\"","Body":"<p>I would like to download and import MNIST data according to the Tensorflow tutorial using following commands in Ubuntu:<\/p>\n\n<pre><\/pre>\n\n<p>But, I get the below error when I run the second command:<\/p>\n\n<pre><\/pre>\n\n<p>Could you please advise me how I can fix this error? I need detailed instruction to run these commands. Do I need to run Python or import Tensorflow before running these commands?<\/p>\n\n<p>Thanks <\/p>\n","answers":[{"AnswerId":"43801973","CreationDate":"2017-05-05T09:57:59.057","ParentId":null,"OwnerUserId":"2944428","Title":null,"Body":"<p>Maybe you forgot to put '#!\/usr\/bin\/python' :)<\/p>\n"},{"AnswerId":"35898123","CreationDate":"2016-03-09T17:10:20.300","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>Haha, yes you need to run those commands in Python. Were you running them on the sh command line? I admit the tensorflow MNIST tutorial does not explicitly state you need to run Python.<\/p>\n\n<p>Try this:<\/p>\n\n<pre><code>$ python\nPython 2.7.5 (default, Jun 24 2015, 00:41:19) \n[GCC 4.8.3 20140911 (Red Hat 4.8.3-9)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; from tensorflow.examples.tutorials.mnist import input_data\n&gt;&gt;&gt; mnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":35896369,"AnswerCount":0,"Tags":"<python><classification><tensorflow>","CreationDate":"2016-03-09T15:53:49.533","AcceptedAnswerId":null,"OwnerUserId":5577321.0,"Title":"How to classify images using tensorflow cifar10 model","Body":"<p>I am trying to use a cifar10 trained model to classify images (that's the purpose of the whole thing, right?!) but untill now I had no success.<\/p>\n\n<p>In cifar10_eval.py, I made the following adaptions:<\/p>\n\n<p>1) in evaluate(), I call a custom made function \"test_image(saver, summary_writer, logits, summary_op)\":<\/p>\n\n<pre><\/pre>\n\n<p>Which is a similar approach to the same problem with the mnist model tutorial. In the mnist tutorial I could successfully  classify digits, as explained in my <a href=\"https:\/\/stackoverflow.com\/questions\/33784214\/how-to-test-tensorflow-cifar10-cnn-tutorial-model\/33785193#33785193\">other question:<\/a>.<\/p>\n\n<p>Has anyone tried something similar? Thanks for your help!<\/p>\n","answers":[]}
{"QuestionId":35897218,"AnswerCount":1,"Tags":"<ubuntu><machine-learning><neural-network><nvidia><theano>","CreationDate":"2016-03-09T16:29:37.740","AcceptedAnswerId":null,"OwnerUserId":4561404.0,"Title":"Problems getting THEANO to work with GPU ubuntu 14.04","Body":"<p>I have the following error when running:<\/p>\n\n<pre><\/pre>\n\n<p>My bashrc file has \":\/usr\/local\/cuda-7.5\" at the end along with export LD_LIBRARY_PATH=\"\/usr\/local\/cuda-7.5\/lib64\" at the end of it. <\/p>\n\n<p>I suspect it has something to do with the LD_LIBRARY_PATH but not sure how to fix this.<\/p>\n","answers":[{"AnswerId":"35897693","CreationDate":"2016-03-09T16:51:49.737","ParentId":null,"OwnerUserId":"157726","Title":null,"Body":"<p>Your problem is not about finding the library. The issue is that you have not upgraded your driver. Maybe it is not installed at all. Have you rebooted?<\/p>\n\n<p>Check your driver version by using <code>nvidia-smi<\/code> or you can use <code>cat \/proc\/driver\/nvidia\/version<\/code><\/p>\n"}]}
{"QuestionId":35900315,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-09T18:59:41.477","AcceptedAnswerId":50406596.0,"OwnerUserId":1760830.0,"Title":"How to simplify Tensorboard graph with shared variables?","Body":"<p>I'm working with  model of Tensorflow. I use the  function for both training and evaluation pipelines and share all the variables inside. My Tensorboard graph visualization looks like in the pic.<\/p>\n\n<p>Arrows between two main columns (e.g. ) inside the  scope hopefully reflect that all the  variables are shared.<\/p>\n\n<p>At the same time, node  doesn't exist in the code. It looks like it's replicated from  to separate the pipeline for my two inputs. <\/p>\n\n<p>My question is what all of  nodes in the graph exactly mean? Do scopes and functions in Tensorflow get replicated when I use them with different inputs? \nThe main question: is there a way to hide  nodes in the graph visualization because they are confusing and messy.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3uxwq.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3uxwq.png\" alt=\"tensorboard graph\"><\/a><\/p>\n","answers":[{"AnswerId":"50406596","CreationDate":"2018-05-18T08:04:16.997","ParentId":null,"OwnerUserId":"9763766","Title":null,"Body":"<p>There is an issue (still open) on GitHub with a workaround <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9545\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9545<\/a><\/p>\n\n<pre><code>from tensorflow.python.ops import variable_scope as var_scope\n\ndef simple_variable_scope(name_or_scope, reuse=None):\n  \"\"\"Creates a variable scope without also creating a name scope.\"\"\"\n  return var_scope.variable_scope(name_or_scope, reuse=reuse,\n                                  auxiliary_name_scope=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":35901400,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-03-09T19:57:29.900","AcceptedAnswerId":null,"OwnerUserId":5786837.0,"Title":"ReluGrad input is not finite on multi-layer network in TensorFlow","Body":"<p>I'm doing the udacity course of TensotFlow, I'm trying to train a neural network on the notMNIST set.<br>\nWhen using a 1-hidden layer network all works fine, but when I try to add another layer, after ~150 steps I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>This is the network model:<\/p>\n\n<pre><\/pre>\n\n<p>And the error is pointing at a specific line:<\/p>\n\n<pre><\/pre>\n\n<p>I guess the with two-layer network the w_h are becoming very small so the matmul product go to zero, but I don't understand how I can solve it\nNotice that I'm using this optimizer:<\/p>\n\n<pre><\/pre>\n\n<p>The net is 784->1024->512->10<\/p>\n\n<p>Any help would be appreciated...<\/p>\n","answers":[{"AnswerId":"45388645","CreationDate":"2017-07-29T11:17:32.343","ParentId":null,"OwnerUserId":"832621","Title":null,"Body":"<p>I was having the same problem when my weights were being initialized randomly and by biases with zeros. Using <a href=\"http:\/\/proceedings.mlr.press\/v9\/glorot10a\/glorot10a.pdf\" rel=\"nofollow noreferrer\">Xavier and Yoshua<\/a>'s initialization solved the problem, and here is my full example:<\/p>\n\n<pre><code>hidden_size = 1024\nbatch_size = 256\n\ndef multilayer(x, w, b):\n    for i, (wi, bi) in enumerate(zip(w, b)):\n        if i == 0:\n            out = tf.nn.relu(tf.matmul(x, wi) + bi)\n        elif i == len(w) - 1:\n            out = tf.matmul(out, wi) + bi\n        else:\n            out = tf.nn.relu(tf.matmul(out, wi) + bi)\n    print(out.shape, x.shape)\n    return out\n\ngraph = tf.Graph()\nwith graph.as_default():\n\n    # Input data. For the training data, we use a placeholder that will be fed\n    # at run time with a training minibatch.\n    tf_train_dataset = tf.placeholder(tf.float32, shape=(batch_size, image_size * image_size))\n    tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))\n    tf_valid_dataset = tf.constant(valid_dataset)\n    tf_test_dataset = tf.constant(test_dataset)\n\n    # Defining Xavier and Yoshua's initializer\n    initializer = tf.contrib.layers.xavier_initializer()\n\n    # Variables\n    W1 = tf.Variable(initializer([image_size * image_size, hidden_size]))\n    b1 = tf.Variable(initializer([hidden_size]))\n    W2 = tf.Variable(initializer([hidden_size, hidden_size]))\n    b2 = tf.Variable(initializer([hidden_size]))\n    W3 = tf.Variable(initializer([hidden_size, hidden_size]))\n    b3 = tf.Variable(initializer([hidden_size]))\n    W4 = tf.Variable(initializer([hidden_size, hidden_size]))\n    b4 = tf.Variable(initializer([hidden_size]))\n    W5 = tf.Variable(initializer([hidden_size, num_labels]))\n    b5 = tf.Variable(initializer([num_labels]))\n\n    Ws = [W1, W2, W3, W4, W5]\n    bs = [b1, b2, b3, b4, b5]\n\n    # Training computation\n    logits = multilayer(tf_train_dataset, Ws, bs)\n    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels))\n    #NOTE loss is actually a scalar value that represents the effectiveness of the\n    #     current prediction. A minimized loss means that the weights and biases\n    #     are adjusted at their best for the training data.\n\n    # Optimizer\n    optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)\n\n    # Predictions for the training, validation, and test data.\n    train_prediction = tf.nn.softmax(logits)\n    valid_prediction = tf.nn.softmax(multilayer(tf_valid_dataset, Ws, bs))\n    test_prediction = tf.nn.softmax(multilayer(tf_test_dataset, Ws, bs))\n<\/code><\/pre>\n"}]}
{"QuestionId":35903910,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-09T22:24:36.677","AcceptedAnswerId":35904439.0,"OwnerUserId":2125302.0,"Title":"Tensorflow: reading the entire contents of a file exactly once into a single tensor","Body":"<p>I'm working with about ~300MB of word embedding data (currently an .npz but I'm willing to do the work to translate into any format) and I'd like to know if there is a way to get that data into tensorflow that doesn't involve initializing it in python (i.e. initializing a  from a numpy array).<\/p>\n\n<p>My reason for wanting to avoid this is that doing so causes tensorflow to dump my embeddings along with the graph definition when writing summaries. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1444\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1444<\/a>.<\/p>\n\n<p>For my training data, I use the normal tensorflow reader pipeline (TFRecordReader, filename queues, ). That's very good at reading fixed sized batches of examples for a predefined number of epochs. What I have no idea how to do is read the entire contents of a file into a <strong>single tensor<\/strong>.  I could solve this pretty easily by just reading a single batch that is the full size of my embeddings, but I'd like a more general solution that doesn't rely on knowing the number of records, just the individual record format.<\/p>\n","answers":[{"AnswerId":"35904439","CreationDate":"2016-03-09T23:03:25.610","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The easiest way to achieve this would be to create a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow\"><code>tf.Variable<\/code><\/a> of the appropriate type and shape by initializing it from a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#placeholder\" rel=\"nofollow\"><code>tf.placeholder()<\/code><\/a>, then use the feed mechanism to pass in the value. As a result, the actual value will never appear in the graph itself.<\/p>\n\n<p>Let's say your embedding matrix is 1000 x 100:<\/p>\n\n<pre><code>embedding_init = tf.placeholder(tf.float32, shape=[1000, 100])\nembedding = tf.Variable(embedding_init)\n<\/code><\/pre>\n\n<p>You can then initialize the variable with the value from your <code>.npz<\/code> file:<\/p>\n\n<pre><code>datafile = numpy.load(\"data.npz\")\nembedding_value = datafile[\"embedding\"]\n\nsess = tf.Session()\nsess.run(tf.initialize_all_variables(),\n         feed_dict={embedding_init: embedding_value})\n\ndatafile.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":35905264,"AnswerCount":4,"Tags":"<python><python-2.7><python-3.x><import><keras>","CreationDate":"2016-03-10T00:17:14.530","AcceptedAnswerId":35911254.0,"OwnerUserId":4802797.0,"Title":"Keras: ImportError: No module named data_utils","Body":"<p>I am trying to import the module  but its not working. However, I can find this module <a href=\"https:\/\/github.com\/fluency03\/keras\/blob\/master\/keras\/utils\/data_utils.py\" rel=\"nofollow noreferrer\">here<\/a>. It is indeed existing. Why I cannot import it while I can import some other modules like and ?<\/p>\n\n<pre><\/pre>\n\n<p><strong>Edit:<\/strong> <\/p>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/a\/35911254\/4802797\">here<\/a> for the answer. <\/p>\n","answers":[{"AnswerId":"36171836","CreationDate":"2016-03-23T07:11:35.323","ParentId":null,"OwnerUserId":"4497662","Title":null,"Body":"<p>I hope your issue is resolved by now, if not:<\/p>\n\n<pre><code>sudo pip install git+git:\/\/github.com\/fchollet\/keras.git --upgrade\n<\/code><\/pre>\n\n<p>should do the job for you. <\/p>\n"},{"AnswerId":"35911254","CreationDate":"2016-03-10T08:33:01.643","ParentId":null,"OwnerUserId":"4802797","Title":null,"Body":"<p>Even though <a href=\"https:\/\/stackoverflow.com\/a\/35905570\/4802797\">this<\/a> answer is correct but is not complete. Thanks to <a href=\"https:\/\/stackoverflow.com\/users\/6005227\/ben-j\">Ben J.<\/a>'s answer but <a href=\"https:\/\/stackoverflow.com\/users\/5827215\/tadhg-mcdonald-jensen\">Tadhg McDonald-Jensen<\/a> is the first one offering me the answers <a href=\"https:\/\/stackoverflow.com\/questions\/35905264\/keras-importerror-no-module-named-data-utils?noredirect=1#comment59471285_35905264\">here<\/a>. <\/p>\n\n<p><strong>Summarize it<\/strong><\/p>\n\n<p>I was using <code>pip install keras<\/code> to install <code>keras<\/code>, but it did not install the latest version of <code>keras<\/code> according to <a href=\"https:\/\/pypi.python.org\/pypi\/Keras\/0.1.3\" rel=\"nofollow noreferrer\">this<\/a>. That is why I could do things like <code>from keras.models import Sequential<\/code>, <code>from keras.layers.core import Dense, Activation, Dropout<\/code>, and <code>from keras.layers.recurrent import LSTM<\/code> but not <code>from keras.utils.data_utils import get_file<\/code>. Because it is not in the previous versions. <\/p>\n\n<p>SO, just <code>clone<\/code> the <code>keras<\/code> from their <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"nofollow noreferrer\">github<\/a>, <code>cd<\/code> into it, and run <code>sudo python setup.py install<\/code> will solve this problem. <\/p>\n\n<p>REMEMBER, if you already did this <code>pip install keras<\/code>, you have to make sure clear all <code>keras<\/code> versions you have installed by doing this <code>pip uninstall keras<\/code> many time until no <code>keras<\/code> existing, then do this <code>sudo python setup.py install<\/code>.  <\/p>\n"},{"AnswerId":"41529784","CreationDate":"2017-01-08T05:28:09.130","ParentId":null,"OwnerUserId":"3588374","Title":null,"Body":"<p>this might solve your problem.<\/p>\n\n<pre><code>sudo -H  pip3 install git+https:\/\/github.com\/fchollet\/keras.git --upgrade\n<\/code><\/pre>\n"},{"AnswerId":"35905570","CreationDate":"2016-03-10T00:47:26.017","ParentId":null,"OwnerUserId":"6005227","Title":null,"Body":"<p>I looked at the tar.gz from <a href=\"https:\/\/pypi.python.org\/pypi\/Keras\/0.1.3\" rel=\"nofollow\">https:\/\/pypi.python.org\/pypi\/Keras\/0.1.3<\/a> and there is no file named data_utils.py in the utils folder. However the github does include this file: <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras<\/a>.<\/p>\n"}]}
{"QuestionId":35905830,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-10T01:14:20.060","AcceptedAnswerId":null,"OwnerUserId":3728184.0,"Title":"Can TensorFlow cache (sub-)graph computations?","Body":"<p>Can TensorFlow automatically cache computations if they involve\nmultiple calls to the same computation (sub-)graph?<\/p>\n\n<p>For example, I have a matrix  in which each entry represents a\ncomputation based on trainable variables . My objective function\nmultiplies this matrix several times with different vectors (each\ntime with unchanged W).<\/p>\n\n<p>Will TensorFlow recompute, for example,  whenever I access\nit, or will it cache that value?<\/p>\n\n<p>In theory, one could precompute the matrix  given a fixed ,\nsuch that each entry in  is a . But that would\nprevent the correct computation of the gradients of .<\/p>\n","answers":[{"AnswerId":"35907755","CreationDate":"2016-03-10T04:43:42.597","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow performs a limited amount of caching, but it probably doesn't cover the case that you describe.<\/p>\n\n<p>If you create a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\" rel=\"nofollow\"><code>tf.Session<\/code><\/a> with the following options, constant folding will be enabled:<\/p>\n\n<pre><code>config = tf.ConfigProto(graph_options=tf.GraphOptions(\n    optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L2)))\nsess = tf.Session(config=config)\n<\/code><\/pre>\n\n<p>When you call <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow\"><code>sess.run()<\/code><\/a> with this configuration, TensorFlow will evaluate the appropriate nodes to run, then identify the subgraph of those nodes whose outputs are constant, evaluate them, and cache the results. Therefore, it will avoid re-executing redundant computation.<\/p>\n\n<p>However, in your question you mention that <code>F<\/code> is a function of some trainable variables. From TensorFlow's point of view, these variables are volatile&mdash;they may change at any time&mdash;so it does not cache values that are derived from these variables. If you want to reuse the same value for <code>F<\/code> multiple times, you could consider storing it in a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#constant\" rel=\"nofollow\"><code>tf.constant()<\/code><\/a> so that the constant folding optimization is more useful.<\/p>\n"}]}
{"QuestionId":35908126,"AnswerCount":1,"Tags":"<neural-network><tensorflow><skflow>","CreationDate":"2016-03-10T05:15:06.727","AcceptedAnswerId":36089530.0,"OwnerUserId":1810559.0,"Title":"Is it desirable to scale data for skflow.TensorFlowDNNClassifier?","Body":"<p>My colleagues and this question on Cross Validated say you should transform data to zero mean and unit variance for neural networks. However, my performance was slightly worse with scaling than without.<\/p>\n\n<p>I tried using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Did I do something wrong or is scaling not necessary?<\/p>\n","answers":[{"AnswerId":"36089530","CreationDate":"2016-03-18T16:27:46.090","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>Usually scaling benefits most for models that don't have regularization and linear models. For example simple mean squared error loss (like in <code>TensorFlowLinearRegressor<\/code>) without regularization won't work very well on not scaled data.<\/p>\n\n<p>In your case you are using classifier that runs softmax regularization and you are using DNN, so scaling is not needed. DNNs themselve can model rescaling (via bias and weight on the feature in the first layer) if that's a useful thing to do.<\/p>\n"}]}
{"QuestionId":35909427,"AnswerCount":1,"Tags":"<python><runtime><distributed><tensorflow>","CreationDate":"2016-03-10T06:44:44.980","AcceptedAnswerId":null,"OwnerUserId":5212207.0,"Title":"TensorFlow Distributed Runtime Model Parallel CIFAR-10","Body":"<p>I have tried to modify the CIFAR-10 example to run on the new TensorFlow distributed runtime. However, I get the following error when trying to run the program:<\/p>\n\n<pre><\/pre>\n\n<p>I start the cluster using the following commands. On the first node I run:<\/p>\n\n<pre><\/pre>\n\n<p>...and on the second node I run:<\/p>\n\n<pre><\/pre>\n\n<p>For the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow\">CIFAR-10 multi-GPU code<\/a>, I make the simple modifications, replacing two lines in the  function. The following line:<\/p>\n\n<pre><\/pre>\n\n<p>...is replaced with:<\/p>\n\n<pre><\/pre>\n\n<p>and the following line:<\/p>\n\n<pre><\/pre>\n\n<p>...is replaced with:<\/p>\n\n<pre><\/pre>\n\n<p>In my understanding, the second substitution should take care of the model substitution.  Running a simpler example, like the code below, works fine:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36087083","CreationDate":"2016-03-18T14:35:05.010","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I can't tell from your program, but my guess is that you also have to modify the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/263d00d2710779d5c4ac66e335b2ba07d8385b6b\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py#L234\" rel=\"nofollow\">line that creates the session<\/a> to specify the address of one of your worker tasks. For example, given your configuration above, you might write:<\/p>\n\n<pre><code>sess = tf.Session(\n    \"grpc:\/\/10.31.101.101:7777\",\n    config=tf.ConfigProto(\n        allow_soft_placement=True,\n        log_device_placement=FLAGS.log_device_placement))\n<\/code><\/pre>\n\n<p>As it happens, we've been trying to improve that error message to make it less confusing. If you update to the latest version in GitHub and run the same code, you should see an error message that explains <em>why<\/em> the device specification could not be satisfied.<\/p>\n"}]}
{"QuestionId":35911252,"AnswerCount":9,"Tags":"<python><tensorflow>","CreationDate":"2016-03-10T08:32:58.807","AcceptedAnswerId":null,"OwnerUserId":1545917.0,"Title":"Disable Tensorflow debugging information","Body":"<p>By debugging information I mean what TensorFlow shows in my terminal about loaded libraries and found devices etc. not Python errors.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"51734310","CreationDate":"2018-08-07T19:41:10.277","ParentId":null,"OwnerUserId":"3561","Title":null,"Body":"<p>To add some flexibility here, you can achieve more fine-grained control over the level of logging by writing a function that filters out messages however you like:<\/p>\n\n<pre><code>logging.getLogger('tensorflow').addFilter(my_filter_func)\n<\/code><\/pre>\n\n<p>where <code>my_filter_func<\/code> accepts a <code>LogRecord<\/code> object as input [<a href=\"https:\/\/docs.python.org\/3\/library\/logging.html#logrecord-objects\" rel=\"nofollow noreferrer\"><code>LogRecord<\/code> docs<\/a>] and\nreturns zero if you want the message thrown out; nonzero otherwise.<\/p>\n\n<p>Here's an example filter that only keeps every nth info message (Python 3 due\nto the use of <code>nonlocal<\/code> here):<\/p>\n\n<pre><code>def keep_every_nth_info(n):\n    i = -1\n    def filter_record(record):\n        nonlocal i\n        i += 1\n        return int(record.levelname != 'INFO' or i % n == 0)\n    return filter_record\n\n# Example usage for TensorFlow:\nlogging.getLogger('tensorflow').addFilter(keep_every_nth_info(5))\n<\/code><\/pre>\n\n<p>All of the above has assumed that TensorFlow has set up its logging state already. You can ensure this without side effects by calling <code>tf.logging.get_verbosity()<\/code> before adding a filter.<\/p>\n"},{"AnswerId":"51327615","CreationDate":"2018-07-13T14:38:03.433","ParentId":null,"OwnerUserId":"2501747","Title":null,"Body":"<p>As <code>TF_CPP_MIN_LOG_LEVEL<\/code> didn't work for me you can try:<\/p>\n\n<pre><code>tf.logging.set_verbosity(tf.logging.WARN)\n<\/code><\/pre>\n\n<p>Worked for me in tensorflow v1.6.0<\/p>\n"},{"AnswerId":"38645250","CreationDate":"2016-07-28T19:33:12.127","ParentId":null,"OwnerUserId":"6557588","Title":null,"Body":"<p><strong>2.0 Update (10\/8\/19)<\/strong>\nSetting <code>TF_CPP_MIN_LOG_LEVEL<\/code> should still work (see below in v0.12+ update), but there is currently an issue open (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31870\" rel=\"noreferrer\">issue #31870<\/a>). If setting <code>TF_CPP_MIN_LOG_LEVEL<\/code> does not work for you (again, see below), try doing the following to set the log level:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ntf.get_logger().setLevel('INFO')\n<\/code><\/pre>\n\n<p>In addition, please see the documentation on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/autograph\/set_verbosity\" rel=\"noreferrer\"><code>tf.autograph.set_verbosity<\/code><\/a> which sets the verbosity of autograph log messages - for example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Can also be set using the AUTOGRAPH_VERBOSITY environment variable\ntf.autograph.set_verbosity(1)\n<\/code><\/pre>\n\n<p><strong>v0.12+ Update (5\/20\/17), Working through TF 2.0+:<\/strong><\/p>\n\n<p>In TensorFlow 0.12+, per this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1258\" rel=\"noreferrer\">issue<\/a>, you can now control logging via the environmental variable called <code>TF_CPP_MIN_LOG_LEVEL<\/code>; it defaults to 0 (all logs shown) but can be set to one of the following values under the <code>Level<\/code> column.<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>  Level | Level for Humans | Level Description                  \n -------|------------------|------------------------------------ \n  0     | DEBUG            | [Default] Print all messages       \n  1     | INFO             | Filter out INFO messages           \n  2     | WARNING          | Filter out INFO &amp; WARNING messages \n  3     | ERROR            | Filter out all messages      \n<\/code><\/pre>\n\n<p>See the following generic OS example using Python:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'  # or any {'0', '1', '2'}\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>To be thorough, you call also set the level for the Python <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/logging\" rel=\"noreferrer\"><code>tf_logging<\/code><\/a> module, which is used in e.g. summary ops, tensorboard, various estimators, etc.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># append to lines above\ntf.logging.set_verbosity(tf.logging.ERROR)  # or any {DEBUG, INFO, WARN, ERROR, FATAL}\n<\/code><\/pre>\n\n<p>For 1.14 you will receive warnings if you do not change to use the v1 API as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># append to lines above\ntf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)  # or any {DEBUG, INFO, WARN, ERROR, FATAL}\n<\/code><\/pre>\n\n<p><hr>\n<strong>For Prior Versions of TensorFlow or TF-Learn Logging (v0.11.x or lower):<\/strong><\/p>\n\n<p>View the page below for information on TensorFlow logging; with the new update, you're able to set the logging verbosity to either <code>DEBUG<\/code>, <code>INFO<\/code>, <code>WARN<\/code>, <code>ERROR<\/code>, or <code>FATAL<\/code>. For example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.logging.set_verbosity(tf.logging.ERROR)\n<\/code><\/pre>\n\n<p>The page additionally goes over monitors which can be used with TF-Learn models. <a href=\"https:\/\/www.tensorflow.org\/get_started\/monitors\" rel=\"noreferrer\">Here is the page<\/a>.<\/p>\n\n<p>This <strong>doesn't<\/strong> block all logging, though (only TF-Learn). I have two solutions; one is a 'technically correct' solution (Linux) and the other involves rebuilding TensorFlow.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>script -c 'python [FILENAME].py' | grep -v 'I tensorflow\/'\n<\/code><\/pre>\n\n<p>For the other, please see <a href=\"https:\/\/stackoverflow.com\/questions\/36331419\/tensorflow-how-to-measure-how-much-gpu-memory-each-tensor-takes\/36505898#36505898\">this answer<\/a> which involves modifying source and rebuilding TensorFlow.<\/p>\n"},{"AnswerId":"57294815","CreationDate":"2019-07-31T15:58:54.257","ParentId":null,"OwnerUserId":"11864958","Title":null,"Body":"<p>I solved with this post <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/27045#issuecomment-477385197\" rel=\"nofollow noreferrer\"> Cannot remove all warnings #27045 <\/a>, and the solution was:<\/p>\n\n<pre><code>import logging\nlogging.getLogger('tensorflow').disabled = True\n<\/code><\/pre>\n"},{"AnswerId":"55142079","CreationDate":"2019-03-13T12:35:05.350","ParentId":null,"OwnerUserId":"1587329","Title":null,"Body":"<p>For <strong>compatibility with Tensorflow 2.0<\/strong>, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/get_logger\" rel=\"noreferrer\"><code>tf.get_logger<\/code><\/a> <\/p>\n\n<pre><code>import logging\ntf.get_logger().setLevel(logging.ERROR)\n<\/code><\/pre>\n"},{"AnswerId":"52973625","CreationDate":"2018-10-24T16:12:06.290","ParentId":null,"OwnerUserId":"3375874","Title":null,"Body":"<p>Usual python3 log manager works for me with tensorflow==1.11.0:<\/p>\n\n<pre><code>import logging\nlogging.getLogger('tensorflow').setLevel(logging.INFO)\n<\/code><\/pre>\n"},{"AnswerId":"39708493","CreationDate":"2016-09-26T17:04:27.150","ParentId":null,"OwnerUserId":"1765893","Title":null,"Body":"<p>I have had this problem as well (on <code>tensorflow-0.10.0rc0<\/code>), but could not fix the excessive nose tests logging problem via the suggested answers.<\/p>\n\n<p>I managed to solve this by probing directly into the tensorflow logger. Not the most correct of fixes, but works great and only pollutes the test files which directly or indirectly import tensorflow:<\/p>\n\n<pre><code># Place this before directly or indirectly importing tensorflow\nimport logging\nlogging.getLogger(\"tensorflow\").setLevel(logging.WARNING)\n<\/code><\/pre>\n"},{"AnswerId":"42121886","CreationDate":"2017-02-08T19:23:32.660","ParentId":null,"OwnerUserId":"2244946","Title":null,"Body":"<p>You can disable all debugging logs using <code>os.environ<\/code> :<\/p>\n\n<pre><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' \nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>Tested on tf 0.12 and 1.0<\/p>\n\n<p>In details, <\/p>\n\n<pre><code>0 = all messages are logged (default behavior)\n1 = INFO messages are not printed\n2 = INFO and WARNING messages are not printed\n3 = INFO, WARNING, and ERROR messages are not printed\n<\/code><\/pre>\n"},{"AnswerId":"57466495","CreationDate":"2019-08-12T18:35:52.987","ParentId":null,"OwnerUserId":"1329739","Title":null,"Body":"<p>Yeah, I'm using tf 2.0-beta and want to enable\/disable the default logging.  The environment variable and methods in tf1.X don't seem to exist anymore.<\/p>\n\n<p>I stepped around in PDB and found this to work:<\/p>\n\n<pre><code># close the TF2 logger\ntf2logger = tf.get_logger()\ntf2logger.error('Close TF2 logger handlers')\ntf2logger.root.removeHandler(tf2logger.root.handlers[0])\n<\/code><\/pre>\n\n<p>I then add my own logger API (in this case file-based)<\/p>\n\n<pre><code>logtf = logging.getLogger('DST')\nlogtf.setLevel(logging.DEBUG)\n\n# file handler\nlogfile='\/tmp\/tf_s.log'\nfh = logging.FileHandler(logfile)\nfh.setFormatter( logging.Formatter('fh %(asctime)s %(name)s %(filename)s:%(lineno)d :%(message)s') )\nlogtf.addHandler(fh)\nlogtf.info('writing to %s', logfile)\n<\/code><\/pre>\n"}]}
{"QuestionId":35916535,"AnswerCount":1,"Tags":"<image-processing><matlab-figure><tensorflow><pixels><opticalflow>","CreationDate":"2016-03-10T12:29:39.543","AcceptedAnswerId":null,"OwnerUserId":5523820.0,"Title":"What is the difference between the tensor flow and the optical flow?","Body":"<p>Can any one give me a accurate response to this question:<\/p>\n\n<p><strong>What is the difference between the  tensor flow and the optical  flow?<\/strong><\/p>\n\n<p>THANKS<\/p>\n","answers":[{"AnswerId":"35918648","CreationDate":"2016-03-10T14:03:20.120","ParentId":null,"OwnerUserId":"3507061","Title":null,"Body":"<p>Tensorflow is the code released by Google for implementing Deep Learning - in the field of Machine Intelligence\/Neural Networks.\n<a href=\"https:\/\/www.tensorflow.org\/\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/<\/a><\/p>\n\n<p>Optical flow is way of analysing successive frames of a video such that features are tracked as they move from pixel to pixel - in the field of Computer Vision\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Optical_flow\" rel=\"nofollow\">https:\/\/en.wikipedia.org\/wiki\/Optical_flow<\/a><\/p>\n"}]}
{"QuestionId":35917701,"AnswerCount":1,"Tags":"<machine-learning><theano><deep-learning><keras>","CreationDate":"2016-03-10T13:19:25.817","AcceptedAnswerId":null,"OwnerUserId":6007683.0,"Title":"All input arrays and target arrays must have the same number of samples.\"- Training on single image to check if the model works in keras","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Hi ,above is my code where i am trying to segment object candidate through graph model,<\/p>\n\n<p>i want to check for one input if the code works or not so i am giving it one input image and the output image,<\/p>\n\n<p>But keras gives me an error - \"All input arrays and target arrays must have the same number of samples.\"<\/p>\n\n<p>Can anyone tell me what do i do to see if my model runs .i am training  on one input so that i can verify that my model is correct and start training ,is there any other way to do it?<\/p>\n","answers":[{"AnswerId":"38612739","CreationDate":"2016-07-27T12:17:38.437","ParentId":null,"OwnerUserId":"2534758","Title":null,"Body":"<p>In the part where you do this - <code>history = graph.fit({'input1':inputvgg, 'output1':outputmodel}, nb_epoch=1)<\/code> inputvgg and outputmodel should have same number of dimensions.<\/p>\n"}]}
{"QuestionId":35918811,"AnswerCount":1,"Tags":"<gpgpu><theano><theano-cuda>","CreationDate":"2016-03-10T14:10:57.030","AcceptedAnswerId":35920796.0,"OwnerUserId":133374.0,"Title":"How to get AdvancedSubtensor on GPU","Body":"<p>I have some subtensor and for some reason, Theano cannot transfer it to the GPU.<\/p>\n\n<p>Some sample code:<\/p>\n\n<pre><\/pre>\n\n<p>I added some  to make the problem more clear because in the output, you see the transfers  and , which it would avoid if it could do the AdvancedSubtensor on GPU. Output.<\/p>\n\n<pre><\/pre>\n\n<p>So, why is Theano not able to transform this into a GPU op?<\/p>\n\n<p>Also, how can I rewrite the code that Theano will do the calculation on GPU?<\/p>\n\n<hr>\n\n<p>Related question in Google Groups: <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/theano-users\/renshaw%7Csort:date\/theano-users\/SKipmABM6aY\/7mPEanvGxwAJ\" rel=\"nofollow\">here<\/a> and <a href=\"https:\/\/groups.google.com\/forum\/#!searchin\/theano-users\/Advanced$20index$20on$20GPU\/theano-users\/pukE6_tGnd8\/-mnM2Q1NZ5cJ\" rel=\"nofollow\">here<\/a>\nand <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/y8XxgYouYSs\" rel=\"nofollow\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"35920796","CreationDate":"2016-03-10T15:33:27.570","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>Ok, so in the Google Groups posts which I linked, it's pretty good explained why it doesn't work. AdvancedSubtensor is the most generic form which works with all crazy types of indexing variants. Then there is AdvancedSubtensor1, which only works for a certain kind of subset. There only exists a GPU version for AdvancedSubtensor1, not for AdvancedSubtensor. I didn't fully understand the reason but there is an ongoing discussion about it.<\/p>\n\n<p>AdvancedSubtensor1 can be used when there is a single list of indices. However, in my example, that is not the case. The common workaround you see, also in some other example in those Google Groups post, is to flatten the array first and calculate the indices for the flattened array.<\/p>\n\n<p>Most examples work with some kind of <code>nonzero()<\/code> or so, where you also would flatten the base arguments in the same and then you get the indices for the flattened version.<\/p>\n\n<p>So, the question is, how to apply this to my code?<\/p>\n\n<p>Actually, there is a simpler solution where it will use AdvancedSubtensor1 which I didn't realized initially:<\/p>\n\n<pre><code>meminkeyP = meminkey[:, P]  # (batch,n_copies,n_cells)\n<\/code><\/pre>\n\n<p>However, before I realized that, I came up with a generic solution which also works for other cases. I transform my indices tuple <code>(batches, P_bc)<\/code> into indices for the flattened version. This is done with this function:<\/p>\n\n<pre><code>def indices_in_flatten_array(ndim, shape, *args):\n  \"\"\"\n  We expect that all args can be broadcasted together.\n  So, if we have some array A with ndim&amp;shape as given,\n  A[args] would give us a subtensor.\n  We return the indices so that A[args].flatten()\n  and A.flatten()[indices] are the same.\n  \"\"\"\n  assert ndim &gt; 0\n  assert len(args) == ndim\n  indices_per_axis = [args[i] for i in range(ndim)]\n  for i in range(ndim):\n    for j in range(i + 1, ndim):\n      indices_per_axis[i] *= shape[j]\n  indices = indices_per_axis[0]\n  for i in range(1, ndim):\n    indices += indices_per_axis[i]\n  return indices\n<\/code><\/pre>\n\n<p>Then, I use it like this:<\/p>\n\n<pre><code>meminkeyP = meminkey.flatten()[indices_in_flatten_array(meminkey.ndim, meminkey.shape, batches, P_bc)]\n<\/code><\/pre>\n\n<p>This seems to work.<\/p>\n\n<p>And I get this output:<\/p>\n\n<pre><code>Using gpu device 0: GeForce GTX TITAN (CNMeM is disabled, CuDNN not available)\nGpuReshape{3} [id A] ''   11\n |GpuAdvancedSubtensor1 [id B] ''   10\n | |GpuReshape{1} [id C] ''   2\n | | |&lt;CudaNdarrayType(float32, matrix)&gt; [id D]\n | | |TensorConstant{(1,) of -1} [id E]\n | |Reshape{1} [id F] ''   9\n |   |Elemwise{second,no_inplace} [id G] ''   8\n |   | |TensorConstant{(1, 5, 10) of 0} [id H]\n |   | |Elemwise{Mul}[(0, 0)] [id I] ''   7\n |   |   |InplaceDimShuffle{0,x,x} [id J] ''   6\n |   |   | |ARange{dtype='int64'} [id K] ''   4\n |   |   |   |TensorConstant{0} [id L]\n |   |   |   |Shape_i{0} [id M] ''   0\n |   |   |   | |&lt;CudaNdarrayType(float32, matrix)&gt; [id D]\n |   |   |   |TensorConstant{1} [id N]\n |   |   |InplaceDimShuffle{x,x,x} [id O] ''   5\n |   |     |Shape_i{1} [id P] ''   1\n |   |       |&lt;CudaNdarrayType(float32, matrix)&gt; [id D]\n |   |TensorConstant{(1,) of -1} [id E]\n |MakeVector{dtype='int64'} [id Q] ''   3\n   |Shape_i{0} [id M] ''   0\n   |TensorConstant{5} [id R]\n   |TensorConstant{10} [id S]\n<\/code><\/pre>\n\n<p>Small test case:<\/p>\n\n<pre><code>def test_indices_in_flatten_array():\n  n_copies, n_cells = 5, 4\n  n_complex_cells = n_cells \/ 2\n  n_batch = 3\n  static_rng = numpy.random.RandomState(1234)\n  def make_permut():\n    p = numpy.zeros((n_copies, n_cells), dtype=\"int32\")\n    for i in range(n_copies):\n      p[i, :n_complex_cells] = static_rng.permutation(n_complex_cells)\n      # Same permutation for imaginary part.\n      p[i, n_complex_cells:] = p[i, :n_complex_cells] + n_complex_cells\n    return T.constant(p)\n  P = make_permut()  # (n_copies,n_cells) -&gt; list of indices\n\n  meminkey = T.as_tensor_variable(static_rng.rand(n_batch, n_cells).astype(\"float32\"))\n  i_t = T.ones((meminkey.shape[0],))  # (batch,)\n  n_batch = i_t.shape[0]\n  batches = T.arange(0, n_batch).dimshuffle(0, 'x', 'x')  # (batch,n_copies,n_cells)\n  P_bc = P.dimshuffle('x', 0, 1)  # (batch,n_copies,n_cells)\n  meminkeyP1 = meminkey[batches, P_bc]  # (batch,n_copies,n_cells)\n  meminkeyP2 = meminkey.flatten()[indices_in_flatten_array(meminkey.ndim, meminkey.shape, batches, P_bc)]\n\n  numpy.testing.assert_allclose(meminkeyP1.eval(), meminkeyP2.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":35919020,"AnswerCount":7,"Tags":"<tensorflow>","CreationDate":"2016-03-10T14:19:55.273","AcceptedAnswerId":null,"OwnerUserId":2886263.0,"Title":"What's the difference of name scope and a variable scope in tensorflow?","Body":"<p>What's the differences between these functions?<\/p>\n\n<blockquote>\n  <p><\/p>\n  \n  <p>Returns a context manager for defining an op that creates variables.\n  This context manager validates that the given values are from the same graph, ensures that that graph is the default graph, and pushes a name scope and a variable scope.<\/p>\n<\/blockquote>\n\n<hr>\n\n<blockquote>\n  <p><\/p>\n  \n  <p>Returns a context manager for use when defining a Python op.\n  This context manager validates that the given values are from the same graph, ensures that that graph is the default graph, and pushes a name scope.            <\/p>\n<\/blockquote>\n\n<hr>\n\n<blockquote>\n  <p><\/p>\n  \n  <p>Wrapper for  using the default graph.\n  See  for more details.<\/p>\n<\/blockquote>\n\n<hr>\n\n<blockquote>\n  <p><\/p>\n  \n  <p>Returns a context for variable scope.\n  Variable scope allows to create new variables and to share already created ones while providing checks to not create or share by accident. For details, see the Variable Scope How To, here we present only a few basic examples.  <\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"41682701","CreationDate":"2017-01-16T18:20:21.717","ParentId":null,"OwnerUserId":"2658138","Title":null,"Body":"<p>\nAs for API r0.11, <code>op_scope<\/code> and <code>variable_op_scope<\/code> are both <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/state_ops\/sharing_variables#variable_op_scope\" rel=\"nofollow noreferrer\">deprecated<\/a>. \n<code>name_scope<\/code> and <code>variable_scope<\/code> can be nested: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.name_scope('ns'):\n    with tf.variable_scope('vs'): #scope creation\n        v1 = tf.get_variable(\"v1\",[1.0])   #v1.name = 'vs\/v1:0'\n        v2 = tf.Variable([2.0],name = 'v2')  #v2.name= 'ns\/vs\/v2:0'\n        v3 = v1 + v2       #v3.name = 'ns\/vs\/add:0'\n<\/code><\/pre>\n"},{"AnswerId":"43581502","CreationDate":"2017-04-24T06:48:26.333","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Both <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_op_scope\" rel=\"nofollow noreferrer\">variable_op_scope<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/op_scope\" rel=\"nofollow noreferrer\">op_scope<\/a> are now deprecated and should not be used at all. <\/p>\n\n<p>Regarding the other two, I also had problems understanding the difference between <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"nofollow noreferrer\">variable_scope<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/name_scope\" rel=\"nofollow noreferrer\">name_scope<\/a> (they looked almost the same) before I tried to visualize everything by creating a simple example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\n\n\ndef scoping(fn, scope1, scope2, vals):\n    with fn(scope1):\n        a = tf.Variable(vals[0], name='a')\n        b = tf.get_variable('b', initializer=vals[1])\n        c = tf.constant(vals[2], name='c')\n\n        with fn(scope2):\n            d = tf.add(a * b, c, name='res')\n\n        print '\\n  '.join([scope1, a.name, b.name, c.name, d.name]), '\\n'\n    return d\n\nd1 = scoping(tf.variable_scope, 'scope_vars', 'res', [1, 2, 3])\nd2 = scoping(tf.name_scope,     'scope_name', 'res', [1, 2, 3])\n\nwith tf.Session() as sess:\n    writer = tf.summary.FileWriter('logs', sess.graph)\n    sess.run(tf.global_variables_initializer())\n    print sess.run([d1, d2])\n    writer.close()\n<\/code><\/pre>\n\n<p>Here I create a function that creates some variables and constants and groups them in scopes (depending on the type I provided). In this function, I also print the names of all the variables. After that, I executes the graph to get values of the resulting values and save event-files to investigate them in TensorBoard. If you run this, you will get the following:<\/p>\n\n<pre><code>scope_vars\n  scope_vars\/a:0\n  scope_vars\/b:0\n  scope_vars\/c:0\n  scope_vars\/res\/res:0 \n\nscope_name\n  scope_name\/a:0\n  b:0\n  scope_name\/c:0\n  scope_name\/res\/res:0 \n<\/code><\/pre>\n\n<p>You see the similar pattern if you open TensorBoard (as you see <code>b<\/code> is outside of <code>scope_name<\/code> rectangular):<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/MN3S3.png\" width=\"550\"><\/p>\n\n<hr>\n\n<p><strong>This gives you the answer<\/strong>:<\/p>\n\n<p>Now you see that <code>tf.variable_scope()<\/code> adds a prefix to the names of all variables (no matter how you create them), ops, constants. On the other hand <code>tf.name_scope()<\/code> ignores variables created with <code>tf.get_variable()<\/code> because it assumes that you know which variable and in which scope you wanted to use.<\/p>\n\n<p>A good documentation on <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"nofollow noreferrer\">Sharing variables<\/a> tells you that <\/p>\n\n<blockquote>\n  <p><code>tf.variable_scope()<\/code>: Manages namespaces for names passed to <code>tf.get_variable()<\/code>.<\/p>\n<\/blockquote>\n\n<p>The same documentation provides a more details how does Variable Scope work and when it is useful.<\/p>\n"},{"AnswerId":"37534656","CreationDate":"2016-05-30T23:09:19.793","ParentId":null,"OwnerUserId":"1576602","Title":null,"Body":"<p>Let's begin by a short introduction to variable sharing. It is a mechanism in <code>TensorFlow<\/code> that allows for sharing variables accessed in different parts of the code without passing references to the variable around. <\/p>\n\n<p>The method <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_variable\" rel=\"noreferrer\"><code>tf.get_variable<\/code><\/a> can be used with the name of the variable as the argument to either create a new variable with such name or retrieve the one that was created before. This is different from using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> constructor which will create a new variable every time it is called (and potentially add a suffix to the variable name if a variable with such name already exists). <\/p>\n\n<p>It is for the purpose of the variable sharing mechanism that a separate type of scope (variable scope) was introduced.<\/p>\n\n<p>As a result, we end up having two different types of scopes:<\/p>\n\n<ul>\n<li><em>name scope<\/em>, created using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/name_scope\" rel=\"noreferrer\"><code>tf.name_scope<\/code><\/a><\/li>\n<li><em>variable scope<\/em>, created using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"noreferrer\"><code>tf.variable_scope<\/code><\/a><\/li>\n<\/ul>\n\n<p>Both scopes have the same effect on all operations as well as variables created using <code>tf.Variable<\/code>, i.e., the scope will be added as a prefix to the operation or variable name. <\/p>\n\n<p>However, name scope is ignored by <code>tf.get_variable<\/code>. We can see that in the following example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.name_scope(\"my_scope\"):\n    v1 = tf.get_variable(\"var1\", [1], dtype=tf.float32)\n    v2 = tf.Variable(1, name=\"var2\", dtype=tf.float32)\n    a = tf.add(v1, v2)\n\nprint(v1.name)  # var1:0\nprint(v2.name)  # my_scope\/var2:0\nprint(a.name)   # my_scope\/Add:0\n<\/code><\/pre>\n\n<p>The only way to place a variable accessed using <code>tf.get_variable<\/code> in a scope is to use a variable scope, as in the following example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.variable_scope(\"my_scope\"):\n    v1 = tf.get_variable(\"var1\", [1], dtype=tf.float32)\n    v2 = tf.Variable(1, name=\"var2\", dtype=tf.float32)\n    a = tf.add(v1, v2)\n\nprint(v1.name)  # my_scope\/var1:0\nprint(v2.name)  # my_scope\/var2:0\nprint(a.name)   # my_scope\/Add:0\n<\/code><\/pre>\n\n<p>This allows us to easily share variables across different parts of the program, even within different name scopes:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.name_scope(\"foo\"):\n    with tf.variable_scope(\"var_scope\"):\n        v = tf.get_variable(\"var\", [1])\nwith tf.name_scope(\"bar\"):\n    with tf.variable_scope(\"var_scope\", reuse=True):\n        v1 = tf.get_variable(\"var\", [1])\nassert v1 == v\nprint(v.name)   # var_scope\/var:0\nprint(v1.name)  # var_scope\/var:0\n<\/code><\/pre>\n\n<hr>\n\n<h2>UPDATE<\/h2>\n\n<p><strong>As of version r0.11, <code>op_scope<\/code> and <code>variable_op_scope<\/code> are both <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/state_ops\/sharing_variables#variable_op_scope\" rel=\"noreferrer\">deprecated<\/a> and replaced by <code>name_scope<\/code> and <code>variable_scope<\/code>.<\/strong> <\/p>\n"},{"AnswerId":"50387030","CreationDate":"2018-05-17T08:43:24.727","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>Let's make it simple: just use <code>tf.variable_scope<\/code>. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6007#issuecomment-264535954\" rel=\"noreferrer\">Quoting a TF developer,<\/a>:<\/p>\n\n<blockquote>\n  <p>Currently, we recommend everyone to use <code>variable_scope<\/code> and not use <code>name_scope<\/code> except for internal code and libraries.<\/p>\n<\/blockquote>\n\n<p>Besides the fact that <code>variable_scope<\/code>'s functionality basically extends those of <code>name_scope<\/code>, consider how they do not play so nice together:<\/p>\n\n<pre><code>with tf.name_scope('foo'):\n  with tf.variable_scope('bar'):\n    x = tf.get_variable('x', shape=())\n    x2 = tf.square(x**2, name='x2')\nprint(x.name)\n# bar\/x:0\nprint(x2.name)\n# foo\/bar\/x2:0\n<\/code><\/pre>\n\n<p>By sticking to <code>variable_scope<\/code> only you avoid some headaches due to this kind of incompatibility.<\/p>\n"},{"AnswerId":"42659005","CreationDate":"2017-03-07T21:49:51.477","ParentId":null,"OwnerUserId":"4974791","Title":null,"Body":"<p>From the last section of this page of the tensorflow documentation: <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"nofollow noreferrer\">Names of ops in <code>tf.variable_scope()<\/code><\/a><\/p>\n\n<blockquote>\n  <p>[...] when we do <code>with tf.variable_scope(\"name\")<\/code>, this implicitly opens a <code>tf.name_scope(\"name\")<\/code>. For example:<\/p>\n<\/blockquote>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.variable_scope(\"foo\"):\n  x = 1.0 + tf.get_variable(\"v\", [1])\nassert x.op.name == \"foo\/add\"\n<\/code><\/pre>\n\n<blockquote>\n  <p>Name scopes can be opened in addition to a variable scope, and then they will only affect the names of the ops, but not of variables.<\/p>\n<\/blockquote>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.variable_scope(\"foo\"):\n    with tf.name_scope(\"bar\"):\n        v = tf.get_variable(\"v\", [1])\n        x = 1.0 + v\nassert v.name == \"foo\/v:0\"\nassert x.op.name == \"foo\/bar\/add\"\n<\/code><\/pre>\n\n<blockquote>\n  <p>When opening a variable scope using a captured object instead of a string, we do not alter the current name scope for ops.<\/p>\n<\/blockquote>\n"},{"AnswerId":"35929502","CreationDate":"2016-03-10T23:37:42.513","ParentId":null,"OwnerUserId":"743658","Title":null,"Body":"<p>Namespaces is a way to organize names for variables and operators in hierarchical manner (e.g. \"scopeA\/scopeB\/scopeC\/op1\")<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13ea3ca91ba5aecab6f21acc14b9cb6a9afa8630\/tensorflow\/python\/framework\/ops.py#L2395\" rel=\"noreferrer\"><code>tf.name_scope<\/code><\/a> creates namespace for operators in the default graph.<\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13ea3ca91ba5aecab6f21acc14b9cb6a9afa8630\/tensorflow\/python\/ops\/variable_scope.py#L403\" rel=\"noreferrer\"><code>tf.variable_scope<\/code><\/a> creates namespace for both variables and operators in the default graph.<\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13ea3ca91ba5aecab6f21acc14b9cb6a9afa8630\/tensorflow\/python\/framework\/ops.py#L3511\" rel=\"noreferrer\"><code>tf.op_scope<\/code><\/a> same as <code>tf.name_scope<\/code>, but for the graph in which specified variables were created.<\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13ea3ca91ba5aecab6f21acc14b9cb6a9afa8630\/tensorflow\/python\/ops\/variable_scope.py#L501\" rel=\"noreferrer\"><code>tf.variable_op_scope<\/code><\/a> same as <code>tf.variable_scope<\/code>, but for the graph in which specified variables were created.<\/p><\/li>\n<\/ul>\n\n<p>Links to the sources above help to disambiguate this documentation issue.<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/alexgorban\/fa5755c7f7fa49b47961\" rel=\"noreferrer\">This example<\/a> shows that all types of scopes define namespaces for both variables and operators with following differences:<\/p>\n\n<ol>\n<li>scopes defined by <code>tf.variable_op_scope<\/code> or <code>tf.variable_scope<\/code> are compatible with <code>tf.get_variable<\/code> (it ignores two other scopes)<\/li>\n<li><code>tf.op_scope<\/code> and <code>tf.variable_op_scope<\/code> just select a graph from a list of specified variables to create a scope for. Other than than their behavior equal to <code>tf.name_scope<\/code> and <code>tf.variable_scope<\/code> accordingly<\/li>\n<li><code>tf.variable_scope<\/code> and <code>variable_op_scope<\/code> add specified or default initializer. <\/li>\n<\/ol>\n"},{"AnswerId":"35927880","CreationDate":"2016-03-10T21:42:10.883","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You can think them as two groups: <code>variable_op_scope<\/code> and <code>op_scope<\/code> take a set of variables as input and are designed to create operations. The difference is in how they affect the creation of variables with <code>tf.get_variable<\/code>:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>def mysum(a,b,name=None):\n    with tf.op_scope([a,b],name,\"mysum\") as scope:\n        v = tf.get_variable(\"v\", 1)\n        v2 = tf.Variable([0], name=\"v2\")\n        assert v.name == \"v:0\", v.name\n        assert v2.name == \"mysum\/v2:0\", v2.name\n        return tf.add(a,b)\n\ndef mysum2(a,b,name=None):\n    with tf.variable_op_scope([a,b],name,\"mysum2\") as scope:\n        v = tf.get_variable(\"v\", 1)\n        v2 = tf.Variable([0], name=\"v2\")\n        assert v.name == \"mysum2\/v:0\", v.name\n        assert v2.name == \"mysum2\/v2:0\", v2.name\n        return tf.add(a,b)\n\nwith tf.Graph().as_default():\n    op = mysum(tf.Variable(1), tf.Variable(2))\n    op2 = mysum2(tf.Variable(1), tf.Variable(2))\n    assert op.name == 'mysum\/Add:0', op.name\n    assert op2.name == 'mysum2\/Add:0', op2.name\n<\/code><\/pre>\n\n<p>notice the name of the variable <code>v<\/code> in the two examples.<\/p>\n\n<p>same for <code>tf.name_scope<\/code> and <code>tf.variable_scope<\/code>:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.Graph().as_default():\n    with tf.name_scope(\"name_scope\") as scope:\n        v = tf.get_variable(\"v\", [1])\n        op = tf.add(v, v)\n        v2 = tf.Variable([0], name=\"v2\")\n        assert v.name == \"v:0\", v.name\n        assert op.name == \"name_scope\/Add:0\", op.name\n        assert v2.name == \"name_scope\/v2:0\", v2.name\n\nwith tf.Graph().as_default():\n    with tf.variable_scope(\"name_scope\") as scope:\n        v = tf.get_variable(\"v\", [1])\n        op = tf.add(v, v)\n        v2 = tf.Variable([0], name=\"v2\")\n        assert v.name == \"name_scope\/v:0\", v.name\n        assert op.name == \"name_scope\/Add:0\", op.name\n        assert v2.name == \"name_scope\/v2:0\", v2.name\n<\/code><\/pre>\n\n<p>You can read more about variable scope in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variable_scope\/index.html\" rel=\"nofollow noreferrer\">tutorial<\/a>.\nA similar question was <a href=\"https:\/\/stackoverflow.com\/questions\/34215746\/difference-between-variable-scope-and-name-scope-in-tensorflow\">asked before<\/a> on Stack Overflow.<\/p>\n"}]}
{"QuestionId":35919715,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2016-03-10T14:47:42.543","AcceptedAnswerId":41203925.0,"OwnerUserId":6045284.0,"Title":"understanding dtype in theano","Body":"<p>I am new to theano , an am trying to reproduce result for <a href=\"https:\/\/github.com\/mbeissinger\/intro_deep\/blob\/master\/MLP_theano.ipynb\" rel=\"nofollow\">this<\/a> code on my local machine. When I run the code as it is I get following error when I run the theano functions(In [48])  and .<\/p>\n\n<blockquote>\n  <p>TypeError: ('An update must have the same type as the original shared variable (shared_var=b_y, shared_var.type=TensorType(float32, vector), update_val=Elemwise{sub,no_inplace}.0, update_val.type=TensorType(float64, vector)).', 'If the difference is related to the broadcast pattern, you can call the tensor.unbroadcast(var, axis_to_unbroadcast[, ...]) function to remove broadcastable dimensions.')<\/p>\n<\/blockquote>\n\n<p>When I change the line:<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>then it runs fine. Now in my  I have . I am not able to understand why this is happening.<\/p>\n\n<p>Moreover if I change all  to , then also I get the same error.<\/p>\n\n<p>Please help me understand what is going on here.<\/p>\n","answers":[{"AnswerId":"41203925","CreationDate":"2016-12-17T23:08:14.493","ParentId":null,"OwnerUserId":"1693354","Title":null,"Body":"<p>Seems your environment is somehow messed up.  Without going into your environment details, I cannt give you a better solution; but whenever I have such issues, I always go for that code environment's command line flags.  You can set the flags for Theano when you run the code on the command line like this: <\/p>\n\n<pre><code>THEANO_FLAGS='floatX=float32,device=gpu0'  python &lt;myscript&gt;.py\n<\/code><\/pre>\n\n<p>More here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a><\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
{"QuestionId":35926504,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-10T20:23:06.263","AcceptedAnswerId":35928570.0,"OwnerUserId":563762.0,"Title":"Where to find Street View House Number example for CNN","Body":"<p>At the end of the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/deep_cnn\/index.html\" rel=\"nofollow\">Convolutional Neural Network example<\/a> there is an exercise to modify the program to use the Street View House Numbers (SVHN) data set. I have been working on this for the past few days and am finding it quite difficult to modify, especially since the data formats are different and the SVHN data set has variably sized images. <\/p>\n\n<p>This example would be extremely valuable to build an image classifier for use in a production environment. Any ideas where I could get this completed exercise?<\/p>\n","answers":[{"AnswerId":"35928570","CreationDate":"2016-03-10T22:25:05.703","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>This was answered in the tensorflow mailing list but in case others end up with this question the data can be fetched from the <a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"nofollow\">standford website<\/a>.<\/p>\n\n<p>The format is in matlab but can be read in python with scipy with <a href=\"http:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.io.loadmat.html\" rel=\"nofollow\">scipy.io.loadmat<\/a><\/p>\n"}]}
{"QuestionId":35930950,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-11T02:05:05.437","AcceptedAnswerId":35931354.0,"OwnerUserId":200340.0,"Title":"Trainable weight for TensorFlow sequence_loss_by_example()","Body":"<p>I want to have a trainable weight in , e.g.<\/p>\n\n<pre><\/pre>\n\n<p>However, running this code gives me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>According to the docstring for this function in : <\/p>\n\n<pre><\/pre>\n\n<p>It requires a \"Tensor\", but I want to pass a . Is there a way to have trainable weights in this function?<\/p>\n","answers":[{"AnswerId":"35931354","CreationDate":"2016-03-11T02:49:37.763","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"nofollow\"><code>tf.Variable<\/code><\/a> can be used anywhere a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Tensor\" rel=\"nofollow\"><code>tf.Tensor<\/code><\/a> (of the same element type and shape) is expected.<\/p>\n\n<p>Therefore, it you want to define a trainable weight, you can pass a list of <code>tf.Variable<\/code> objects as the <code>weights<\/code> argument to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13ea3ca91ba5aecab6f21acc14b9cb6a9afa8630\/tensorflow\/python\/ops\/seq2seq.py#L814\" rel=\"nofollow\"><code>seq2seq.sequence_loss_by_example()<\/code><\/a>. For example, you could do the following:<\/p>\n\n<pre><code># Defines a list of `num_steps` variables, each 1-D with length `batch_size`.\nweights = [tf.get_variable(\"w\", [batch_size]) for _ in range(num_steps)]\n\nloss = seq2seq.sequence_loss_by_example([logits_1, ..., logits_n],\n                                        [targets_1, ..., targets_n],\n                                        weights,\n                                        vocab_size_all)\n<\/code><\/pre>\n"}]}
{"QuestionId":35935414,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-03-11T08:26:06.100","AcceptedAnswerId":null,"OwnerUserId":6032878.0,"Title":"How to make a dataset similar to CIFAR10 with several images","Body":"<p>I try to make a dataset that is similar to CIFAR10. I found this tutorial:\n<a href=\"https:\/\/stackoverflow.com\/questions\/35032675\/how-to-create-dataset-similar-to-cifar-10\">How to create dataset similar to cifar-10<\/a><\/p>\n\n<p>I already can make a dataset with 1 image, but when I try to use several images I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone help me to solve this problem?<\/p>\n\n<p>This is my code: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37614682","CreationDate":"2016-06-03T12:41:37.457","ParentId":null,"OwnerUserId":"5651936","Title":null,"Body":"<p>I do it like this: <\/p>\n\n<pre><code>import numpy as np\nimport scipy.io \n\nmat = scipy.io.loadmat('train_32x32.mat')\ndata = mat['X']\nlabel = mat['y']\n\nR_data = data[:,:,0,:]\nG_data = data[:,:,1,:]\nB_data = data[:,:,2,:]\n\nR_data = np.transpose(R_data, (2,0,1))\nG_data = np.transpose(G_data, (2,0,1))\nB_data = np.transpose(B_data, (2,0,1))\n\nR_data = np.reshape(R_data,(73257,32*32))\nG_data = np.reshape(G_data,(73257,32*32))\nB_data = np.reshape(B_data,(73257,32*32))\n\noutdata = np.concatenate((label,R_data,G_data,B_data), axis = 1)\nstep = 10000\nfor i in range(1,6):\n    temp = outdata[i*step:(i+1)*step,:]\n    temp.tofile('SVHN_train_data_batch%d.bin' % i)\n    print('save data %d' % i)\n<\/code><\/pre>\n\n<p>Then, just put it directly in the train code of Cifar10 tensorflow example.<\/p>\n"},{"AnswerId":"37822224","CreationDate":"2016-06-14T21:14:54.533","ParentId":null,"OwnerUserId":"5120590","Title":null,"Body":"<p>I too tried to follow the tutorial you posted in the question however I couldn't get it to work so I made my own solution. It can be found on my github here: <a href=\"https:\/\/github.com\/jdeepee\/machine_learning\/tree\/master\" rel=\"nofollow\">https:\/\/github.com\/jdeepee\/machine_learning\/tree\/master<\/a><\/p>\n\n<p>The code is commented so should be easy enough to follow. I should note it iterated through a master directory containing multiple folders which contain the images.<\/p>\n"},{"AnswerId":"41082549","CreationDate":"2016-12-11T03:00:02.813","ParentId":null,"OwnerUserId":"435089","Title":null,"Body":"<p>The below snippet is what I did to adapt CIFAR-10 to GTSRB. More details here. <a href=\"https:\/\/github.com\/hashkanna\/traffic-signs\/blob\/master\/Traffic_Signs_Recognition_binFiles.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hashkanna\/traffic-signs\/blob\/master\/Traffic_Signs_Recognition_binFiles.ipynb<\/a><\/p>\n\n<pre><code>out = {}\n\nfor i in range(5):\n    bin_val = (i%5) + 1\n    #im = Image.open(X_train[i])\n    #im = np.array(im)\n    im = X_train[i]\n    r = im[:,:,0].flatten()\n    g = im[:,:,1].flatten()\n    b = im[:,:,2].flatten()\n    label = [y_train[i]]\n    out[bin_val] = np.array(list(label) + list(r) + list(g) + list(b),np.uint8)\n\nfor i in range(5,len(X_train)):\n    bin_val = (i%5) + 1\n    #im = Image.open(X_train[i])\n    #im = np.array(im)\n    im = X_train[i]\n    r = im[:,:,0].flatten()\n    g = im[:,:,1].flatten()\n    b = im[:,:,2].flatten()\n    label = [y_train[i]]\n    new_array = np.array(list(label) + list(r) + list(g) + list(b),np.uint8)\n    out[bin_val] = np.append(out[bin_val], new_array, 0)\n\nfor bin_val in range(1,6):\n    out[bin_val].tofile(\"\/Users\/kanna\/Downloads\/data_batch_%s.bin\"%bin_val)\n<\/code><\/pre>\n"}]}
{"QuestionId":35935462,"AnswerCount":3,"Tags":"<python><bash><pip><theano>","CreationDate":"2016-03-11T08:29:51.887","AcceptedAnswerId":35965549.0,"OwnerUserId":3302563.0,"Title":"Theano AttributeError: 'module' object has no attribute 'relu'","Body":"<p>I want to use , but I keep getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>I have already updated theano via  command as described in <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#updating-theano\" rel=\"nofollow\">theano's documentation<\/a>, also i've tried . Neither worked, I still get the same error. <\/p>\n\n<p>I was trying to do  to confirm that I have installed the latest version but then I get the following error: <\/p>\n\n<p>So my two questions here are:<\/p>\n\n<ul>\n<li>How can I see theano's version?<\/li>\n<li>Am I doing something wrong when updating theano? How can I solve the error first mentioned?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"35965549","CreationDate":"2016-03-13T01:51:53.317","ParentId":null,"OwnerUserId":"2067976","Title":null,"Body":"<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html?highlight=relu#theano.tensor.nnet.relu\" rel=\"nofollow\">relu<\/a> is available for <strong>theano >=  0.7.1<\/strong>. My guess is that pip linked to <code>theano==0.7<\/code>.<\/p>\n\n<p>You can check theano version with <code>pip freeze<\/code>:<\/p>\n\n<pre><code>pip freeze | grep Theano\n<\/code><\/pre>\n\n<p>So you have to install latest <code>theano<\/code> with pointing pip to theano git repo :<\/p>\n\n<pre><code>pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n<\/code><\/pre>\n\n<p>Also note that relu is function and <strong>not module<\/strong>, so to access it you have to use one of imports below:<\/p>\n\n<pre><code>from theano.tensor.nnet import relu # access `relu` as is ..\nimport theano.tensor.nnet as theano_nnet #access `relu` as `theano_nnet.relu`\n<\/code><\/pre>\n"},{"AnswerId":"35940254","CreationDate":"2016-03-11T12:25:14.380","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The <code>theano.tensor.nnet<\/code> module only supports <code>relu<\/code> in the most recent versions. In order to use it, you need to install the bleeding edge version from github or wait until the next release.<\/p>\n\n<p>Alternatively, you can implement it like this:<\/p>\n\n<pre><code>def relu(x):\n    return T.maximum(x, 0.)\n<\/code><\/pre>\n\n<p>This may not be inplace but it gives you the result and the gradient.<\/p>\n"},{"AnswerId":"35965514","CreationDate":"2016-03-13T01:44:54.327","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<ol>\n<li><p>to see theano version you can run the following code:<\/p>\n\n<pre><code>import theano\nprint theano.__version__\n<\/code><\/pre><\/li>\n<li><p>You should follow the instructions from <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\" rel=\"nofollow\">here<\/a> to get the bleeding edge version<\/p><\/li>\n<\/ol>\n\n<p>Actually relu function is simple to code, you can create your own relu function, like eickenberg's answer, or like <code>theano.tensor.nnet<\/code> style:<\/p>\n\n<pre><code>    def relu(x):   \n       return 0.5 * (x + abs(x))\n<\/code><\/pre>\n"}]}
{"QuestionId":35946810,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-03-11T17:54:44.987","AcceptedAnswerId":35948352.0,"OwnerUserId":5551505.0,"Title":"Compare tf.string and python string","Body":"<p>How can I compare a Tensorflow string tensor to a python string type.\nI tried something like this<\/p>\n\n<pre><\/pre>\n\n<p>key here, is a tf.string. I want to see if 'neg' is in the key tensor.<\/p>\n","answers":[{"AnswerId":"35948352","CreationDate":"2016-03-11T19:27:05.197","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Unfortunately, currently you cannot. Please ask this on TF github in issues section.<\/p>\n"}]}
{"QuestionId":35948961,"AnswerCount":0,"Tags":"<python><boost><caffe><pycaffe>","CreationDate":"2016-03-11T20:06:10.667","AcceptedAnswerId":null,"OwnerUserId":3919542.0,"Title":"Import caffe giving Runtimewarning","Body":"<p>I installed caffe on my system following this link <a href=\"http:\/\/hanzratech.in\/2015\/07\/27\/installing-caffe-on-ubuntu.html\" rel=\"nofollow\">http:\/\/hanzratech.in\/2015\/07\/27\/installing-caffe-on-ubuntu.html<\/a> . I was able to install it without any errors but when i did import caffe in the python terminal it is giving me this warning. Any reason as why is this happening and how it can be removed .<\/p>\n\n<pre><\/pre>\n\n<p>I am using a virtualbox in which i have installed caffe having Ubuntu14.04<\/p>\n","answers":[]}
{"QuestionId":35950032,"AnswerCount":1,"Tags":"<lua><deep-learning><torch>","CreationDate":"2016-03-11T21:14:37.420","AcceptedAnswerId":35956629.0,"OwnerUserId":846970.0,"Title":"torch extract sub tensor using indices","Body":"<p>I would like to extract a subtensor from the original tensor which contains elements at indices defined by another tensor.<\/p>\n\n<p>Say <\/p>\n\n<pre><\/pre>\n\n<p>How can I query ls using the indices in i?<\/p>\n","answers":[{"AnswerId":"35956629","CreationDate":"2016-03-12T10:43:40.397","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<pre><code>sub = ls:index(1, torch.LongTensor{2, 7, 9})\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-indexdim-index\" rel=\"nofollow\"><code>[Tensor] index(dim, index)<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":35951487,"AnswerCount":1,"Tags":"<python><gpu><gpgpu><tensorflow>","CreationDate":"2016-03-11T23:05:19.700","AcceptedAnswerId":null,"OwnerUserId":2267681.0,"Title":"TensorFlow only works with GPU 0","Body":"<p>I am having a hard time trying to run a tensorflow program in the GPU 1. It doesn't matter if I use  or inside the program I use the , all the time I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>This doesn't happen when running with GPU0, but the problem is that someone else is using that GPU and GPU1 is idle<\/p>\n","answers":[{"AnswerId":"35959188","CreationDate":"2016-03-12T14:55:37.380","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>If you are using CUDA_VISIBLE_DEVICES=1 make sure that \"someone else\" is running with CUDA_VISIBLE_DEVICES=0. <\/p>\n\n<p>By default, if no CUDA_VISIBLE_DEVICES is specified, a tensorflow process hogs all the GPUs even if it is not actively using them. They may appear idle, but will not accessible to subsequent tensorflow processes. However, you can use CUDA_VISIBLE_DEVICES on each process to allocate different GPUs to each. <\/p>\n\n<p>Remember that within each process, the tensorflow devices \/gpu:0, \/gpu:1, etc. are relative to the process, not global for your machine...so, they will refer to whatever cuda visible devices are available to that process...put another way, if you want each process to use 1 GPU you would refer to \/gpu:0 in the the code for all of them while using CUDA_VISIBLE_DEVICES to assign specific GPUs to each process.     <\/p>\n\n<p>Use command <code>nvidia-smi<\/code> to see which processes are using which GPUs.<\/p>\n"}]}
{"QuestionId":35951503,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2016-03-11T23:06:21.917","AcceptedAnswerId":null,"OwnerUserId":5606352.0,"Title":"Keras\/Python - If a RNN is stateful, a complete input_shape must be provided (including batch size)","Body":"<p>I'm trying to implement a stateful RNN, however it keeps asking me for a \"complete input_shape (including batch size)\". So I tried different things for the  and  arguments, neither of which seems to work. Can anyone shine some light on it?<\/p>\n\n<p>Code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35991984","CreationDate":"2016-03-14T15:44:32.907","ParentId":null,"OwnerUserId":"627042","Title":null,"Body":"<p>You need to provide only the batch_input_shape= parameter, and <em>not<\/em> the input_shape parameter.  Also, to avoid input shape errors, make sure the training data size is a multiple of batch_size.   And finally, if you are using validation splits, you have to be sure that both splits are also multiples of the batch_size.<\/p>\n\n<pre><code># ensure data size is a multiple of batch_size\ndata_size=data_size-data_size%batch_size\n# ensure validation splits are multiples of batch_size\nincrement=float(batch_size)\/len(data_size)\nval_split=float(int(val_split\/(increment))) * increment\n<\/code><\/pre>\n"},{"AnswerId":"38118457","CreationDate":"2016-06-30T09:02:08.047","ParentId":null,"OwnerUserId":"6253353","Title":null,"Body":"<p>In your definition of <code>SimpleRNN<\/code>, remove <code>input_dim<\/code> and <code>input_shape<\/code>, set:<\/p>\n\n<pre><code>batch_input_shape = (Number_Of_sequences, Size_Of_Each_Sequence,\n                     Shape_Of_Element_In_Each_Sequence) \n<\/code><\/pre>\n\n<p><code>batch_input_shape<\/code> should be a tuple of length at least 3.<\/p>\n\n<p>If you passes your sequences one by one, set:<\/p>\n\n<pre><code>Number_Of_sequences = 1\n<\/code><\/pre>\n\n<p>If the size of your sequences is not fixed, set:<\/p>\n\n<pre><code>Size_Of_Each_Sequence = None\n<\/code><\/pre>\n"}]}
{"QuestionId":35953210,"AnswerCount":10,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2016-03-12T02:55:20.680","AcceptedAnswerId":35963479.0,"OwnerUserId":1056563.0,"Title":"Error running basic tensorflow example","Body":"<p>I have just reinstalled latest tensorflow on ubuntu:<\/p>\n\n<pre><\/pre>\n\n<p>When following the directions to test it fails with <strong>cannot import name pywrap_tensorflow<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>Is there an additional change needed to my python or ubuntu\/bash environment?<\/p>\n","answers":[{"AnswerId":"35963479","CreationDate":"2016-03-12T21:22:44.010","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From the path in your stack trace (<code>\/git\/tensorflow\/tensorflow\/\u2026<\/code>), it looks like your Python path may be loading the tensorflow libraries from the source directory, rather than the version that you have installed. As a result, it is unable to find the (compiled) <code>pywrap_tensorflow<\/code> library, which is installed in a different directory.<\/p>\n\n<p>A common solution is to <code>cd<\/code> out of the <code>\/git\/tensorflow<\/code> directory before starting <code>python<\/code> or <code>ipython<\/code>.<\/p>\n"},{"AnswerId":"38536906","CreationDate":"2016-07-23T00:16:09.407","ParentId":null,"OwnerUserId":"1904943","Title":null,"Body":"<p>I compiled, installed TensorFlow from source (GitHub: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow<\/a>) in a Python 2.7 virtual environment (venv).  It worked fine, but I needed (as others have mentioned, e.g. user \"mrry\" at <a href=\"https:\/\/stackoverflow.com\/questions\/35953210\/error-running-basic-tensorflow-example\">Error running basic tensorflow example<\/a>) to cd out of the partition in which I had compiled TensorFlow, to another partition, in order to be able to import tensorflow into Python. Otherwise, I got various errors, depending on which (source partition) directory I was in.  To clarify:<\/p>\n\n<pre><code>         source: \/mnt\/Vancouver\/apps\/tensorflow\ncan't import tf: Python launched in any of \/mnt\/...\n  can import tf: Python launched in \/home\/victoria\/...\n<\/code><\/pre>\n\n<p>I later just followed the instructions here,<\/p>\n\n<pre><code>https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#anaconda-installation\n<\/code><\/pre>\n\n<p>and everything just works, fine.<\/p>\n\n<p>For reference, I am operating on<\/p>\n\n<ul>\n<li>Arch Linux [4.6.3-1-ARCH] x86_64<\/li>\n<li>Intel i7-4790<\/li>\n<li>xfce 4.12 desktop environment<\/li>\n<\/ul>\n\n<p>INSTALL STEPS:<\/p>\n\n<p>Modify paths, venv names as per your preferences.<\/p>\n\n<ol>\n<li><p>Create tf-env:<\/p>\n\n<pre><code>cd \/home\/victoria\/anaconda3\/envs\n\nconda create -n tf-env python=2.7 anaconda\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Note: appending the 'anaconda' metapackage installs ALL Anaconda packages (NumPy; ...).<\/p>\n\n<ol start=\"2\">\n<li><p>Source activate that venv (tf-env)<\/p>\n\n<pre><code>source activate tf-env\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Note: added to ~\/.bashrc as an alias:<\/p>\n\n<pre><code>alias tf='echo \"  [TensorFlow in Anaconda Python 2.7 venv (source activate tf]\" &amp;&amp; source activate tf-env'\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li><p>Install TensorFlow in the tf-env venv:<\/p>\n\n<pre><code>(tf-env)$ conda install -c conda-forge tensorflow\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>THIS obviates the use of pip (including the *.whl install script), an alternative approach to installing TensorFlow that is described here: <\/p>\n\n<pre><code>https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md).\n<\/code><\/pre>\n\n<p>WORKS!<\/p>\n\n<pre><code>(tf-env)[victoria@victoria ~]$ P\n\n  [P: python]\nPython 2.7.12 |Anaconda 4.1.1 (64-bit)| (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\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; print tensorflow.__version__\n0.9.0\n&gt;&gt;&gt; [Ctrl-D]\n\n(tf-env)[victoria@victoria ~]$\n<\/code><\/pre>\n\n<p>You may then work with TensorFlow in the tf-env; e.g., in Jupyter notebooks launched in this (tf-env) venv.<\/p>\n"},{"AnswerId":"38253674","CreationDate":"2016-07-07T19:20:35.787","ParentId":null,"OwnerUserId":"6562358","Title":null,"Body":"<p>I resolve the problem.\nTry the follow command:<\/p>\n\n<p><code>pip install --upgrade pip<\/code><\/p>\n"},{"AnswerId":"50939234","CreationDate":"2018-06-20T02:13:25.233","ParentId":null,"OwnerUserId":"819417","Title":null,"Body":"<p>In my <code>python -m venv environment<\/code> it appeared to be fixed by:<\/p>\n\n<pre><code>pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>Change <code>requirements.txt<\/code> to say <code>tensorflow==1.5.0<\/code> instead of <code>tensorflow==1.8.0<\/code>.<\/p>\n\n<pre><code>pip install -r requirements.txt\n<\/code><\/pre>\n"},{"AnswerId":"54319278","CreationDate":"2019-01-23T02:39:32.873","ParentId":null,"OwnerUserId":"9352979","Title":null,"Body":"<p>I figured out that TensorFlow 1.12.0 only works with Python version 3.5.2. I had Python 3.7 but that didn't work. So, I had to downgrade Python and then I could install TensorFlow it works.<\/p>\n\n<p>To downgrade your python version from 3.7 to 3.6<\/p>\n\n<pre><code>conda install python=3.6.8\n<\/code><\/pre>\n"},{"AnswerId":"51903798","CreationDate":"2018-08-17T23:04:19.800","ParentId":null,"OwnerUserId":"10241453","Title":null,"Body":"<p>Installing via the following command solves the problem: <\/p>\n\n<p><code>pip install --upgrade<\/code><\/p>\n\n<p>Here's the <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">link<\/a> to download<\/p>\n"},{"AnswerId":"56410901","CreationDate":"2019-06-01T22:31:00.473","ParentId":null,"OwnerUserId":"8293976","Title":null,"Body":"<p>I'm using tensorflow 1.13.1. Just reinstalling tensorflow worked for me. <\/p>\n"},{"AnswerId":"57068872","CreationDate":"2019-07-17T05:10:20.640","ParentId":null,"OwnerUserId":"8361934","Title":null,"Body":"<p>I also have this issue and have been looking into it for a long time. It seems that there is no such problem on python 3+. On py2.7- The problem is actually on google.protobuf<\/p>\n\n<h2><strong>Solution 1:<\/strong><\/h2>\n\n<pre><code>pip uninstall protobuf\npip uninstall google\npip install google\npip install protobuf\npip install google-cloud\n<\/code><\/pre>\n\n<h2><strong>Solution 2:<\/strong><\/h2>\n\n<p><strong>create an <code>__init__.py<\/code> in \"google\" folder.<\/strong><\/p>\n\n<pre><code>cd \/path\/to\/your\/env\/lib\/python2.7\/site-packages\/google\ntouch __init__.py\n<\/code><\/pre>\n\n<p>Hopefully it will work.<\/p>\n"},{"AnswerId":"51259760","CreationDate":"2018-07-10T07:46:20.953","ParentId":null,"OwnerUserId":"7800702","Title":null,"Body":"<p>if you using CPU manufactured before 2011 then  downgraded the tensorflow version from 1.8.0 to  1.5.0 or   1.2.0 and try to import the module it worked for me.<\/p>\n"},{"AnswerId":"48896566","CreationDate":"2018-02-21T00:49:11.047","ParentId":null,"OwnerUserId":"2876310","Title":null,"Body":"<p>Command below helped me.<\/p>\n\n<pre><code> pip install tensorflow --upgrade --force-reinstall\n<\/code><\/pre>\n"}]}
{"QuestionId":35955144,"AnswerCount":6,"Tags":"<tensorflow>","CreationDate":"2016-03-12T07:50:29.693","AcceptedAnswerId":null,"OwnerUserId":1430078.0,"Title":"Working with multiple graphs in TensorFlow","Body":"<p>Can someone explain to me how  works in TensorFlow?  <\/p>\n\n<p>Suppose I have the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I run this code I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I agree \"g2\/MatMul\" is not an element of graph , but why is it selecting \"g2\/MatMul\" when the session graph is set to ? Why doesn't it select \"g1\/MatMul\"?<\/p>\n\n<hr>\n\n<h3>Edit<\/h3>\n\n<p>The following code seems to work:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>By flipping the switch , graph  or  will run in the session. Is this the way name scoping was meant to work?<\/p>\n","answers":[{"AnswerId":"43637553","CreationDate":"2017-04-26T14:48:49.257","ParentId":null,"OwnerUserId":"6784405","Title":null,"Body":"<p>This is quite the necro, but it's still a top search result for questions related to this and I thought it might help to make very explicit something that the prior answers (which are correct) note in passing:<\/p>\n\n<p>The Q's variable <code>product<\/code> is a <em>python<\/em> variable. As such, it points to an object: When defined, it points to the <code>tf.Tensor<\/code> output of the matmul <code>tf.Operation<\/code> defined in the name_scope 'g1'. It is later redefined to point to a different object, the <code>tf.Tensor<\/code> output of 'g2'. This python variable has never heard of <code>tf.name_scope<\/code>s and doesn't care.<\/p>\n\n<p><em>That<\/em> is why you need to do lookups by the name attributes of the tensorflow objects... Those, by use of name_scopes, are unique and accessible. Or generate distinct python variables--that are unique and accessible according to python scoping rules--to point to each <code>tf.Tensor<\/code> object you want to reference.<\/p>\n\n<p>Dunno if this is helpful for anybody else, but if I ever forget, I'll thank my past self for this.<\/p>\n"},{"AnswerId":"43654576","CreationDate":"2017-04-27T10:01:59.247","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>Your problem is that you are calling the latest variable <code>product<\/code> which points to a tensor created on g2 - you overwrote it in your second scope. Just relabel all your variables and you should be good to go. The working code is below.<\/p>\n\n<pre><code>import tensorflow as tf\n\ng1 = tf.Graph() with g1.as_default() as g:\n    with g.name_scope( \"g1\" ) as scope:\n        matrix11 = tf.constant([[3., 3.]])\n        matrix21 = tf.constant([[2.],[2.]])\n        product1 = tf.matmul(matrix11, matrix21)\n\ntf.reset_default_graph()\n\ng2 = tf.Graph() with g2.as_default() as g:\n    with g.name_scope( \"g2\" ) as scope:\n        matrix12 = tf.constant([[4., 4.]])\n        matrix22 = tf.constant([[5.],[5.]])\n        product2 = tf.matmul(matrix12, matrix22)\n\ntf.reset_default_graph()\n\nwith tf.Session( graph = g1 ) as sess:\n    result = sess.run( product1 )\n    print( result )\n<\/code><\/pre>\n"},{"AnswerId":"39838433","CreationDate":"2016-10-03T18:50:05.100","ParentId":null,"OwnerUserId":"6917313","Title":null,"Body":"<p>I was having similar difficulties dealing with multiple graphs on a IPython notebook. What works for my purposes is to encapsulate each graph and its session in a function. I realize this is more of a hack I guess, I don't know anything about namespaces and I know OP wanted something along those lines. Maybe it will help someone I dunno, you can also pass results between computations.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef Graph1():\n    g1 = tf.Graph()\n    with g1.as_default() as g:\n        matrix1 = tf.constant([[3., 3.]])\n        matrix2 = tf.constant([[2.],[2.]])\n        product = tf.matmul( matrix1, matrix2, name = \"product\")\n\n    with tf.Session( graph = g ) as sess:\n        tf.initialize_all_variables().run()\n        return product\n\n\ndef Graph2(incoming):\n    i = incoming\n    g2 = tf.Graph()\n    with g2.as_default() as g:\n        matrix1 = tf.constant([[4., 4.]])\n        matrix2 = tf.constant([[5.],[5.]])\n        product = tf.matmul( matrix1, matrix2, name = \"product\" )\n\n    with tf.Session( graph = g ) as sess:\n        tf.initialize_all_variables().run()\n        print product\n        print i\n\nprint Graph1()\n\nGraph2(Graph1())\n<\/code><\/pre>\n"},{"AnswerId":"35962721","CreationDate":"2016-03-12T20:08:39.370","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Your <code>product<\/code> is a global variable, and you've set it to point to \"g2\/MatMul\".<\/p>\n\n<p>In particular<\/p>\n\n<p>Try<\/p>\n\n<pre><code>print product\n<\/code><\/pre>\n\n<p>and you'll see<\/p>\n\n<pre><code>Tensor(\"g2\/MatMul:0\", shape=(1, 1), dtype=float32)\n<\/code><\/pre>\n\n<p>So the system takes <code>\"g2\/MatMul:0\"<\/code> since that's the Tensor's name, and tries to find it in the graph <code>g1<\/code> since that's the graph you set for the session. Incidentally you can see all nodes in the graph <code>print [n.name for n in g1.as_graph_def().node]<\/code><\/p>\n\n<p>Generally, using more than one graph is rarely useful. You can't merge them and can't pass tensors between them. I'd recommend just doing<\/p>\n\n<pre><code>tf.reset_default_graph()\na = tf.Constant(2)\nsess = tf.InteractiveSession()\n....\n<\/code><\/pre>\n\n<p>This way you'll have one default graph and one default session and you can omit specifying graph or session in most cases. If you ever need to refer to them explicitly, you can get them from <code>tf.get_default_graph()<\/code> or <code>tf.get_default_session()<\/code><\/p>\n"},{"AnswerId":"47753619","CreationDate":"2017-12-11T13:02:47.303","ParentId":null,"OwnerUserId":"8022263","Title":null,"Body":"<p>I have written two snippets to do what you are trying to do.<\/p>\n\n<p>The first switches between graphs and does NOT use name_scope. The second uses the default_graph for both ops and uses name_scope to switch between them...<\/p>\n\n<pre><code>import tensorflow as tf\n\ng1 = tf.Graph()\nwith g1.as_default() as g:\n  matrix1 = tf.constant([[3., 3.]])\n  matrix2 = tf.constant([[2.],[2.]])\n  tf.matmul(matrix1, matrix2, name=\"productX\")\n\ng2 = tf.Graph()\nwith g2.as_default() as g:\n  matrix1 = tf.constant([[4., 4.]])\n  matrix2 = tf.constant([[5.],[5.]])\n  tf.matmul(matrix1, matrix2, name=\"productX\")\n\nproduct_op = g1.get_tensor_by_name(  \"productX:0\" )\nwith tf.Session( graph = g1 ) as sess:\n    result = sess.run( product_op )\n    print( result )\n\nproduct_op = g2.get_tensor_by_name(  \"productX:0\" )\nwith tf.Session( graph = g2 ) as sess:\n  result = sess.run( product_op )\n  print( result )\n<\/code><\/pre>\n\n<p>NOTE A) I have taken out the resetting of the default graph (the default graph is never used) and B) I have gotten rid of the 'product' variable and given the operation a name, instead<\/p>\n\n<p>The important switching code is...<\/p>\n\n<pre><code>product_op = g2.get_tensor_by_name(  \"productX:0\" )\n<\/code><\/pre>\n\n<p>where we are using the variable name of the two graphs to distinguish between the two product ops.<\/p>\n\n<p>NOTE the annoying ':0' that you have to put at the end of the variable name.<\/p>\n\n<p>And now using name_scope...<\/p>\n\n<pre><code>import tensorflow as tf\n\ng = tf.get_default_graph()\n\nmatrix1 = tf.constant([[3., 3.]])\nmatrix2 = tf.constant([[2.],[2.]])\nwith g.name_scope(\"prodA\"):\n  tf.matmul(matrix1, matrix2, name=\"productX\")\n\nmatrix1 = tf.constant([[4., 4.]])\nmatrix2 = tf.constant([[5.],[5.]])\nwith g.name_scope(\"prodB\"):\n  tf.matmul(matrix1, matrix2, name=\"productX\")\n\nwith tf.Session() as sess:\n\n  product_op = g.get_tensor_by_name(  \"prodA\/productX:0\" )\n  result = sess.run( product_op )\n  print( result )\n\n  product_op = g.get_tensor_by_name(  \"prodB\/productX:0\" )\n  result = sess.run( product_op )\n  print( result )\n<\/code><\/pre>\n\n<p>Now the switching line of code is this...<\/p>\n\n<pre><code>product_op = g.get_tensor_by_name(  \"prodB\/productX:0\" )\n<\/code><\/pre>\n\n<p>NOTE A) There is no switching between graphs or sessions. B) name_scope give you a kind of directory structure looking name hierarchy.<\/p>\n\n<p>What are the pros and cons of graph vs name_scope switching? If someone wrote a blog on that, I would read it!<\/p>\n\n<p>I don't imagine that switching between graphs and sessions can be very fast, however.<\/p>\n"},{"AnswerId":"41872248","CreationDate":"2017-01-26T11:14:00.237","ParentId":null,"OwnerUserId":"6455403","Title":null,"Body":"<p>You have 3 graphs , not 2 graphs ( g1 or g2 ) .<\/p>\n\n<p>You can see 3 graphs by id() <\/p>\n\n<pre><code>...\n\nprint 'g1           ', id(g1)\nprint 'g2           ', id(g2)\nprint 'current graph', id ( tf.get_default_graph() )\n\nwith tf.Session( graph = g1 ) as sess:\n    result = sess.run( product )\n    print( result )\n<\/code><\/pre>\n\n<p>Using \"with g1.as_default(): \" make the same error<\/p>\n\n<pre><code>...\n\nwith g1.as_default() :\n    with tf.Session( graph = g1 ) as sess:\n        result = sess.run( product )\n        print( result )\n<\/code><\/pre>\n\n<p>Because, you have two 'product' , not one.<\/p>\n\n<pre><code>g1 = tf.Graph()\nwith g1.as_default() as g:\n    ...\n    print 'g1 product', id ( product )\n\n...\n\ng2 = tf.Graph()\nwith g2.as_default() as g:\n    ...\n    print 'g2 product', id ( product )\n\nwith tf.Session( graph = g1 ) as sess:\n    print 'last product', id(product)\n    ...\n<\/code><\/pre>\n\n<p>last product == g2 product<\/p>\n\n<pre><code>...\n    product = tf.matmul(matrix1, matrix2, name='g1_product')\n...\nwith g1.as_default() as g:\n    with tf.Session() as sess:\n        product = g.get_tensor_by_name(  \"g1_product:0\" )\n        result = sess.run( product )\n        print( result )\n<\/code><\/pre>\n\n<p>Above code work.<\/p>\n\n<p>But, two variables with the same name ( product )<\/p>\n\n<p>Encapsulating with class is good?<\/p>\n"}]}
{"QuestionId":35955989,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-12T09:38:13.170","AcceptedAnswerId":35963259.0,"OwnerUserId":3052875.0,"Title":"How to add summaries to seq2seq tutorial?","Body":"<p>I'm working with <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/seq2seq\/index.html\" rel=\"nofollow\">Seq2Seq example in TensorFlow<\/a>. I can run the training and see the outputs of perplexity on the development set. It's great!<\/p>\n\n<p>I just want to add summaries (especially  such as perplexity on dev set) to the event file and monitor them in TensorBoard. After <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">reading the documentation<\/a>, I don't understand how to annotate <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/translate\/translate.py#L27\" rel=\"nofollow\"><\/a> with summary ops.<\/p>\n\n<p>Can anybody can help me with simple pseudo-code?<\/p>\n","answers":[{"AnswerId":"35963259","CreationDate":"2016-03-12T21:00:39.667","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like <code>translate.py<\/code> doesn't create a TensorBoard summary log at all. (Part of the reason may be that much of the evaluation happens in Python, rather than in the TensorFlow graph.) Let's see how to add one.<\/p>\n\n<ol>\n<li><p>You'll need to create a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#SummaryWriter\" rel=\"nofollow\"><code>tf.train.SummaryWriter<\/code><\/a>. Add the following before entering the training loop (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3a955eb32d919bde8580ff53fb4e753a9579b8f5\/tensorflow\/models\/rnn\/translate\/translate.py#L162\" rel=\"nofollow\">here<\/a>):<\/p>\n\n<pre><code>summary_writer = tf.train.SummaryWriter(\"path\/to\/logs\", sess.graph_def)\n<\/code><\/pre><\/li>\n<li><p>You'll need to create summary events for the perplexity in each bucket. These values are computed in Python, so you can't use the usual <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#scalar_summary\" rel=\"nofollow\"><code>tf.scalar_summary()<\/code><\/a> op. Instead, you'll create a <code>tf.Summary<\/code> directly by modifying <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3a955eb32d919bde8580ff53fb4e753a9579b8f5\/tensorflow\/models\/rnn\/translate\/translate.py#L194\" rel=\"nofollow\">this loop<\/a>:<\/p>\n\n<pre><code>perplexity_summary = tf.Summary()\n\n# Run evals on development set and print their perplexity.\nfor bucket_id in xrange(len(_buckets)):\n  encoder_inputs, decoder_inputs, target_weights = model.get_batch(\n      dev_set, bucket_id)\n  _, eval_loss, _ = model.step(sess, encoder_inputs, decoder_inputs,\n                               target_weights, bucket_id, True)\n  eval_ppx = math.exp(eval_loss) if eval_loss &lt; 300 else float('inf')\n  print(\"  eval: bucket %d perplexity %.2f\" % (bucket_id, eval_ppx))\n\n  bucket_value = perplexity_summary.value.add()\n  bucket_value.tag = \"peplexity_bucket)%d\" % bucket_id\n  bucket_value.simple_value = eval_ppx\n\nsummary_writer.add_summary(perplexity_summary, model.global_step.eval())\n<\/code><\/pre><\/li>\n<li><p>You can add other metrics by constructing <code>tf.Summary<\/code> values yourself and calling <code>summary_writer.add_summary()<\/code>.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":35958139,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-12T13:19:15.687","AcceptedAnswerId":35960657.0,"OwnerUserId":4829166.0,"Title":"TensorFlow: how can I sum a list of tf.Variables?","Body":"<p>I've got a 3D array that is of  type.  only works on individual tensors. I've tried doing:<\/p>\n\n<pre><\/pre>\n\n<p>...but  expects a tensor and not a list. Can I convert it to a tensor somehow, or is there another, better way to do it? I haven't found anything in the API.<\/p>\n","answers":[{"AnswerId":"35960657","CreationDate":"2016-03-12T17:04:35.387","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#reduce_sum\" rel=\"noreferrer\"><code>tf.reduce_sum()<\/code><\/a> op works on 3-D tensors and variables (and in general any rank or tensor or variable). However, if you have a <strong>list<\/strong> of 2-D tensors (or variables), you should use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#add_n\" rel=\"noreferrer\"><code>tf.add_n()<\/code><\/a> op to add together the values in the list:<\/p>\n\n<pre><code>var_3Dlist = ...  # List of 3-D variables.\nsum_list = [tf.reduce_sum(mat) for mat in var_3Dlist]\nsum = tf.add_n(sum_list)\n<\/code><\/pre>\n"}]}
{"QuestionId":35961216,"AnswerCount":3,"Tags":"<time-series><tensorflow><prediction><lstm>","CreationDate":"2016-03-12T17:49:07.590","AcceptedAnswerId":49816576.0,"OwnerUserId":2721272.0,"Title":"How to train a RNN with LSTM cells for time series prediction","Body":"<p>I'm currently trying to build a simple model for predicting time series. The goal would be to train the model with a sequence so that the model is able to predict future values.<\/p>\n\n<p>I'm using tensorflow and lstm cells to do so. The model is trained with truncated backpropagation through time. My question is how to structure the data for training.<\/p>\n\n<p>For example let's assume we want to learn the given sequence:<\/p>\n\n<pre><\/pre>\n\n<p>And we unroll the network for .<\/p>\n\n<p><strong>Option 1<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Option 2<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Option 3<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Option 4<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"49816576","CreationDate":"2018-04-13T12:03:14.333","ParentId":null,"OwnerUserId":"1389680","Title":null,"Body":"<p>After reading several LSTM introduction blogs e.g. <a href=\"http:\/\/www.jakob-aungiers.com\/articles\/a\/LSTM-Neural-Network-for-Time-Series-Prediction\" rel=\"nofollow noreferrer\">Jakob Aungiers'<\/a>, option 3 seems to be the right one for stateless LSTM.<\/p>\n\n<p>If your LSTMs need to remember data longer ago than your <code>num_steps<\/code>, your can train in a stateful way - for a Keras example see <a href=\"https:\/\/philipperemy.github.io\/keras-stateful-lstm\/\" rel=\"nofollow noreferrer\">Philippe Remy's blog post \"Stateful LSTM in Keras\"<\/a>. Philippe does not show an example for batch size greater than one, however. I guess that in your case a batch size of four with stateful LSTM could be used with the following data (written as <code>input -&gt; label<\/code>):<\/p>\n\n<pre><code>batch #0:\n1,2,3,4 -&gt; 5\n2,3,4,5 -&gt; 6\n3,4,5,6 -&gt; 7\n4,5,6,7 -&gt; 8\n\nbatch #1:\n5,6,7,8 -&gt; 9\n6,7,8,9 -&gt; 10\n7,8,9,10 -&gt; 11\n8,9,10,11 -&gt; 12\n\nbatch #2:\n9,10,11,12 -&gt; 13\n...\n<\/code><\/pre>\n\n<p>By this, the state of e.g. the 2nd sample in batch #0 is correctly reused to continue training with the 2nd sample of batch #1.<\/p>\n\n<p>This is somehow similar to your option 4, however you are not using all available labels there.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>In extension to my suggestion where <code>batch_size<\/code> equals the <code>num_steps<\/code>, Alexis Huet <a href=\"https:\/\/stackoverflow.com\/a\/48521460\/1389680\">gives an answer<\/a> for the case of <code>batch_size<\/code> being a divisor of <code>num_steps<\/code>, which can be used for larger <code>num_steps<\/code>. He <a href=\"https:\/\/ahstat.github.io\/RNN-Keras-time-series\/\" rel=\"nofollow noreferrer\">describes it nicely<\/a> on his blog.<\/p>\n"},{"AnswerId":"36017615","CreationDate":"2016-03-15T17:02:11.347","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>I believe Option 1 is closest to the reference implementation in \/tensorflow\/models\/rnn\/ptb\/reader.py<\/p>\n\n<pre><code>def ptb_iterator(raw_data, batch_size, num_steps):\n  \"\"\"Iterate on the raw PTB data.\n\n  This generates batch_size pointers into the raw PTB data, and allows\n  minibatch iteration along these pointers.\n\n  Args:\n    raw_data: one of the raw data outputs from ptb_raw_data.\n    batch_size: int, the batch size.\n    num_steps: int, the number of unrolls.\n\n  Yields:\n    Pairs of the batched data, each a matrix of shape [batch_size, num_steps].\n    The second element of the tuple is the same data time-shifted to the\n    right by one.\n\n  Raises:\n    ValueError: if batch_size or num_steps are too high.\n  \"\"\"\n  raw_data = np.array(raw_data, dtype=np.int32)\n\n  data_len = len(raw_data)\n  batch_len = data_len \/\/ batch_size\n  data = np.zeros([batch_size, batch_len], dtype=np.int32)\n  for i in range(batch_size):\n    data[i] = raw_data[batch_len * i:batch_len * (i + 1)]\n\n  epoch_size = (batch_len - 1) \/\/ num_steps\n\n  if epoch_size == 0:\n    raise ValueError(\"epoch_size == 0, decrease batch_size or num_steps\")\n\n  for i in range(epoch_size):\n    x = data[:, i*num_steps:(i+1)*num_steps]\n    y = data[:, i*num_steps+1:(i+1)*num_steps+1]\n    yield (x, y)\n<\/code><\/pre>\n\n<p>However, another Option is to select a pointer into your data array randomly for each training sequence.<\/p>\n"},{"AnswerId":"37528208","CreationDate":"2016-05-30T14:32:15.827","ParentId":null,"OwnerUserId":"3053918","Title":null,"Body":"<p>I'm just about to learn LSTMs in TensorFlow and try to implement an example which (luckily) tries to predict some time-series \/ number-series genereated by a simple math-fuction.<\/p>\n\n<p>But I'm using a different way to structure the data for training, motivated by <a href=\"http:\/\/arxiv.org\/pdf\/1502.04681v3.pdf\" rel=\"noreferrer\">Unsupervised Learning of Video Representations using LSTMs<\/a>:<\/p>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/1Ngkk.png\" rel=\"noreferrer\">LSTM Future Predictor Model<\/a><\/p>\n\n<p><strong>Option 5:<\/strong><\/p>\n\n<pre><code>input data               label     \n1,2,3,4                  5,6,7,8\n2,3,4,5                  6,7,8,9\n3,4,5,6                  7,8,9,10\n...\n<\/code><\/pre>\n\n<p>Beside this paper, I (tried) to take inspiration by the given TensorFlow RNN examples. My current complete solution looks like this:<\/p>\n\n<pre><code>import math\nimport random\nimport numpy as np\nimport tensorflow as tf\n\nLSTM_SIZE = 64\nLSTM_LAYERS = 2\nBATCH_SIZE = 16\nNUM_T_STEPS = 4\nMAX_STEPS = 1000\nLAMBDA_REG = 5e-4\n\n\ndef ground_truth_func(i, j, t):\n    return i * math.pow(t, 2) + j\n\n\ndef get_batch(batch_size):\n    seq = np.zeros([batch_size, NUM_T_STEPS, 1], dtype=np.float32)\n    tgt = np.zeros([batch_size, NUM_T_STEPS], dtype=np.float32)\n\n    for b in xrange(batch_size):\n        i = float(random.randint(-25, 25))\n        j = float(random.randint(-100, 100))\n        for t in xrange(NUM_T_STEPS):\n            value = ground_truth_func(i, j, t)\n            seq[b, t, 0] = value\n\n        for t in xrange(NUM_T_STEPS):\n            tgt[b, t] = ground_truth_func(i, j, t + NUM_T_STEPS)\n    return seq, tgt\n\n\n# Placeholder for the inputs in a given iteration\nsequence = tf.placeholder(tf.float32, [BATCH_SIZE, NUM_T_STEPS, 1])\ntarget = tf.placeholder(tf.float32, [BATCH_SIZE, NUM_T_STEPS])\n\nfc1_weight = tf.get_variable('w1', [LSTM_SIZE, 1], initializer=tf.random_normal_initializer(mean=0.0, stddev=1.0))\nfc1_bias = tf.get_variable('b1', [1], initializer=tf.constant_initializer(0.1))\n\n# ENCODER\nwith tf.variable_scope('ENC_LSTM'):\n    lstm = tf.nn.rnn_cell.LSTMCell(LSTM_SIZE)\n    multi_lstm = tf.nn.rnn_cell.MultiRNNCell([lstm] * LSTM_LAYERS)\n    initial_state = multi_lstm.zero_state(BATCH_SIZE, tf.float32)\n    state = initial_state\n    for t_step in xrange(NUM_T_STEPS):\n        if t_step &gt; 0:\n            tf.get_variable_scope().reuse_variables()\n\n        # state value is updated after processing each batch of sequences\n        output, state = multi_lstm(sequence[:, t_step, :], state)\n\nlearned_representation = state\n\n# DECODER\nwith tf.variable_scope('DEC_LSTM'):\n    lstm = tf.nn.rnn_cell.LSTMCell(LSTM_SIZE)\n    multi_lstm = tf.nn.rnn_cell.MultiRNNCell([lstm] * LSTM_LAYERS)\n    state = learned_representation\n    logits_stacked = None\n    loss = 0.0\n    for t_step in xrange(NUM_T_STEPS):\n        if t_step &gt; 0:\n            tf.get_variable_scope().reuse_variables()\n\n        # state value is updated after processing each batch of sequences\n        output, state = multi_lstm(sequence[:, t_step, :], state)\n        # output can be used to make next number prediction\n        logits = tf.matmul(output, fc1_weight) + fc1_bias\n\n        if logits_stacked is None:\n            logits_stacked = logits\n        else:\n            logits_stacked = tf.concat(1, [logits_stacked, logits])\n\n        loss += tf.reduce_sum(tf.square(logits - target[:, t_step])) \/ BATCH_SIZE\n\nreg_loss = loss + LAMBDA_REG * (tf.nn.l2_loss(fc1_weight) + tf.nn.l2_loss(fc1_bias))\n\ntrain = tf.train.AdamOptimizer().minimize(reg_loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n\n    total_loss = 0.0\n    for step in xrange(MAX_STEPS):\n        seq_batch, target_batch = get_batch(BATCH_SIZE)\n\n        feed = {sequence: seq_batch, target: target_batch}\n        _, current_loss = sess.run([train, reg_loss], feed)\n        if step % 10 == 0:\n            print(\"@{}: {}\".format(step, current_loss))\n        total_loss += current_loss\n\n    print('Total loss:', total_loss)\n\n    print('### SIMPLE EVAL: ###')\n    seq_batch, target_batch = get_batch(BATCH_SIZE)\n    feed = {sequence: seq_batch, target: target_batch}\n    prediction = sess.run([logits_stacked], feed)\n    for b in xrange(BATCH_SIZE):\n        print(\"{} -&gt; {})\".format(str(seq_batch[b, :, 0]), target_batch[b, :]))\n        print(\" `-&gt; Prediction: {}\".format(prediction[0][b]))\n<\/code><\/pre>\n\n<p>Sample output of this looks like this:<\/p>\n\n<pre><code>### SIMPLE EVAL: ###\n# [input seq] -&gt; [target prediction]\n#  `-&gt; Prediction: [model prediction]  \n[  33.   53.  113.  213.] -&gt; [  353.   533.   753.  1013.])\n `-&gt; Prediction: [ 19.74548721  28.3149128   33.11489105  35.06603241]\n[ -17.  -32.  -77. -152.] -&gt; [-257. -392. -557. -752.])\n `-&gt; Prediction: [-16.38951683 -24.3657589  -29.49801064 -31.58583832]\n[ -7.  -4.   5.  20.] -&gt; [  41.   68.  101.  140.])\n `-&gt; Prediction: [ 14.14126873  22.74848557  31.29668617  36.73633194]\n...\n<\/code><\/pre>\n\n<p>The model is a <em>LSTM-autoencoder<\/em> having 2 layers each.<\/p>\n\n<p>Unfortunately, as you can see in the results, this model does not learn the sequence properly. I might be the case that I'm just doing a bad mistake somewhere, or that 1000-10000 training steps is just way to few for a LSTM. As I said, I'm also just starting to understand\/use LSTMs properly.\nBut hopefully this can give you some inspiration regarding the implementation.<\/p>\n"}]}
{"QuestionId":35961348,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-12T18:01:23.910","AcceptedAnswerId":null,"OwnerUserId":1930168.0,"Title":"TensorBoard not working","Body":"<p>I'm able to use TensorFlow just fine. But I can't yet use TensorBoard at all. I'm following the instructions on tensorflow.org's <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"noreferrer\">Visualizing Learning page<\/a>.<\/p>\n\n<p>When I run<\/p>\n\n<pre><\/pre>\n\n<p>I get the following<\/p>\n\n<pre><\/pre>\n\n<p>If I nagivate to , I see a blank white page.<\/p>\n\n<p>[Stack: OSX, Google Chrome, Python 3, TensorFlow installed via pip]<\/p>\n","answers":[{"AnswerId":"48706036","CreationDate":"2018-02-09T12:38:06.283","ParentId":null,"OwnerUserId":"2420305","Title":null,"Body":"<p>I had a similar issue with TensorBoard 1.4 (no errors on start but black page in browser). In my case it helped to set the host explicitly to 127.0.0.1 with --host<\/p>\n"},{"AnswerId":"37884049","CreationDate":"2016-06-17T14:24:26.350","ParentId":null,"OwnerUserId":"1712419","Title":null,"Body":"<p>Try navigating to <a href=\"http:\/\/localhost:6006\/\" rel=\"noreferrer\">localhost:6006<\/a> instead (equivalent to <a href=\"http:\/\/127.0.0.1:6006\/\" rel=\"noreferrer\">http:\/\/127.0.0.1:6006\/<\/a>).<\/p>\n\n<p>It works fine for me while <a href=\"http:\/\/0.0.0.0:6006\" rel=\"noreferrer\">http:\/\/0.0.0.0:6006<\/a> yields no result.<\/p>\n"}]}
{"QuestionId":35961434,"AnswerCount":1,"Tags":"<python><ocr><tensorflow>","CreationDate":"2016-03-12T18:09:20.183","AcceptedAnswerId":null,"OwnerUserId":5599861.0,"Title":"TensorFlow: parameters do not update when training","Body":"<p>I'm implementing a classification model using TensorFlow<\/p>\n\n<p>The problem that I'm facing is that my weights and error are not being updated when I run the training step. As a result, my network keeps returning the same results.<\/p>\n\n<p>I've developed my model based on the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/pros\/index.html\" rel=\"nofollow\">MNIST example<\/a> from the TensorFlow website.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35962343","CreationDate":"2016-03-12T19:32:52.383","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue probably arises from how you initialize the weight matrix, <code>W<\/code>. If it is initialized to all zeroes, all of the neurons will follow the same gradient in each step, which leads to the network not training. Replacing the line<\/p>\n\n<pre><code>W = tf.Variable(tf.zeros([1200,26]))\n<\/code><\/pre>\n\n<p>...with something like<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([1200,26], stddev=0.001))\n<\/code><\/pre>\n\n<p>...should cause it to start training.<\/p>\n\n<p><a href=\"https:\/\/stats.stackexchange.com\/questions\/27112\/danger-of-setting-all-initial-weights-to-zero-in-backpropagation\">This question<\/a> on the CrossValidated site has a good explanation of why you should not initialize all of your weights to zero.<\/p>\n"}]}
{"QuestionId":35963661,"AnswerCount":1,"Tags":"<lua><neural-network><torch><conv-neural-network>","CreationDate":"2016-03-12T21:42:55.290","AcceptedAnswerId":null,"OwnerUserId":700023.0,"Title":"exception using the `apply` method on a `CudaTensor`","Body":"<p>This problem emerged only after trying to use torch's CUDA capabilities instead of relying to the CPU.<\/p>\n\n<p>I'm trying to initialize a  of weights of a convolutional neural network. The function is the following:<\/p>\n\n<pre><\/pre>\n\n<p>and it's invoked in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>with  being a convolutional module,  being a  and  is a floating point value.<\/p>\n\n<p>The exception that I get is the following:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on what might cause it?<\/p>\n\n<p>I also tried to convert the value returned by  into\na 1-element , but it didn't help.<\/p>\n","answers":[{"AnswerId":"35963871","CreationDate":"2016-03-12T22:03:55.550","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>You can directly call :normal on the tensor.<\/p>\n\n<pre><code>function fill_0normal(t,sigma) do\n  t:normal(0, sigma)\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":35964410,"AnswerCount":1,"Tags":"<python><dataset><tensorflow>","CreationDate":"2016-03-12T23:04:55.573","AcceptedAnswerId":null,"OwnerUserId":5599861.0,"Title":"Feeding external datataset in TensorFlow","Body":"<p>i'have successfully tested the simple model of mnist classifier available here: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html#mnist-for-ml-beginners<\/a><\/p>\n\n<p>Now, i'm trying to run a similar model with external data, that i've used to train classifiers in python.<\/p>\n\n<p>The dataset i'm using consist of the matrix X(11527 by 1200) containing the pixels values for 11527 images and Y(1200 by 26) containing the one hot encoded labels.<\/p>\n\n<p>The problem i'm facing is that given the new data, the weights aren't changing with the iterations, and i'm stuck on a single accuracy value of 10,945 no matter how much iterations the system perform.<\/p>\n\n<p>original code, from tensor flows website(this one works fine):<\/p>\n\n<pre><\/pre>\n\n<p>code using external dataset (this one does not work)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42315519","CreationDate":"2017-02-18T13:04:51.577","ParentId":null,"OwnerUserId":"6969060","Title":null,"Body":"<p>The problem is that your weights and bias are initialized as zeros (check pesos and bias).\nYou should initialize bias to 0.1 and weights to a normal distribution.<\/p>\n"}]}
{"QuestionId":35964994,"AnswerCount":5,"Tags":"<python><macos><scikit-learn><pycharm><tensorflow>","CreationDate":"2016-03-13T00:16:17.840","AcceptedAnswerId":40260973.0,"OwnerUserId":5758094.0,"Title":"pycharm error while importing, even though it works in the terminal","Body":"<p>I have installed the packages TensorFlow and scikit_learn and all its dependencies. When I try importing them using python 2.7.6 or 2.7.10 (I have tried both) in the terminal, it works fine. However, when I do it using pycharm it gives an error.<\/p>\n\n<p>In the case of <strong>scikit_learn<\/strong> with the <strong>launcher 2.7.6<\/strong> says:<\/p>\n\n<blockquote>\n  <p>ImportError: dynamic module does not define init function (init_check_build)<\/p>\n<\/blockquote>\n\n<p>In the case of <strong>scikit_learn<\/strong> with the <strong>launcher 2.7.10<\/strong> says:<\/p>\n\n<blockquote>\n  <p>ValueError: numpy.dtype has the wrong size, try recompiling<\/p>\n<\/blockquote>\n\n<p>In the case of <strong>TensorFlow<\/strong> with the <strong>launcher 2.7.6<\/strong> says:<\/p>\n\n<blockquote>\n  <p>ImportError: dlopen(\/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so, 2): no suitable image found.  Did find:\n          \/Library\/Python\/2.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow.so: mach-o, but wrong architecture<\/p>\n<\/blockquote>\n\n<p>In the case of <strong>TensorFlow<\/strong> with the <strong>launcher 2.7.10<\/strong> says:<\/p>\n\n<blockquote>\n  <p>ImportError: No module named copyreg<\/p>\n  \n  <p>Error importing tensorflow.  Unless you are using bazel,\n      you should not try to import tensorflow from its source directory;\n      please exit the tensorflow source tree, and relaunch your python interpreter\n      from there.<\/p>\n<\/blockquote>\n\n<p>I have tried to search in the net but the solutions did not work for me. I have tried to uninstall them and install them again with pip, conda and directly the source code and it gives always the same errors. I have even tried reinstalling pycharm with no better luck.<\/p>\n\n<p>Other libraries, such as scilab or numpy, work fine in pycharm.<\/p>\n\n<p>Any ideas? It is just driving me mental. By the way, I am using a Mac OS 10.10.5.<\/p>\n","answers":[{"AnswerId":"38749298","CreationDate":"2016-08-03T16:41:05.303","ParentId":null,"OwnerUserId":"6564941","Title":null,"Body":"<p>you should start PyCharm from terminal<\/p>\n\n<p><code>cd \/usr\/lib\/pycharm-community\/bin \n.\/pycharm.sh\n<\/code><\/p>\n"},{"AnswerId":"40121055","CreationDate":"2016-10-19T02:24:46.423","ParentId":null,"OwnerUserId":"4323093","Title":null,"Body":"<p>Add this 'DYLD_LIBRARY_PATH=\/usr\/local\/cuda\/lib' to Python environment variable. Run-> Edit Configurations -> Environment variables. Hope it works. <\/p>\n"},{"AnswerId":"40260973","CreationDate":"2016-10-26T11:20:47.960","ParentId":null,"OwnerUserId":"5758094","Title":null,"Body":"<p>At the end, I ended up creating a virtual environment, reinstalling everything in there, and calling it through pycharm. I am not entirely sure what was the problem between conda and pycharm, I probably messed up somewhere. <\/p>\n\n<p>I am now using a different virtual environment depending on the project and I am happier than ever :).<\/p>\n"},{"AnswerId":"47130715","CreationDate":"2017-11-06T06:05:44.310","ParentId":null,"OwnerUserId":"3667626","Title":null,"Body":"<p>I had a similar problem. My code was was not working on PyCharm professional. I had PyCharm CE previously installed and it worked from there. I had configured PyCharm CE a while ago and I had forgotten what setup I used but if issues persist, make sure that the packages are installed under Preferences > Project > Project Interpreter  <\/p>\n"},{"AnswerId":"49679794","CreationDate":"2018-04-05T19:10:27.937","ParentId":null,"OwnerUserId":"9559884","Title":null,"Body":"<p>hey if there are people still with this problem I would like to tell you the solution. It works for me everytime.<\/p>\n\n<p>1.Open Pycharm<\/p>\n\n<p>2.Open PyCharmProject or the folder you are using<\/p>\n\n<p>3.Click File and there Settings<\/p>\n\n<p>4.Click Project: PyCharmProjects , there project interpreter<\/p>\n\n<p>5.Select the global virtual enviroment for this ( including photo thats showing what i did )<\/p>\n\n<p>6.Check if there are all packages you have installed<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IPa3c.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IPa3c.png\" alt=\"This is the picture\"><\/a><\/p>\n"}]}
{"QuestionId":35968973,"AnswerCount":3,"Tags":"<neural-network><theano><keras>","CreationDate":"2016-03-13T10:10:21.217","AcceptedAnswerId":36798335.0,"OwnerUserId":5766416.0,"Title":"Keras IndexError: indices are out-of-bounds","Body":"<p>I'm new to Keras and im trying to do Binary MLP on a dataset, and keep getting indices out of bounds with no idea why.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Does anyone have any idea why this is happening? Im able to run other models just fine<\/p>\n","answers":[{"AnswerId":"42009631","CreationDate":"2017-02-02T18:25:11.527","ParentId":null,"OwnerUserId":"7284903","Title":null,"Body":"<p>From the <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/#training\" rel=\"nofollow noreferrer\">official Keras Page:<\/a> <\/p>\n\n<blockquote>\n  <p>Keras models are trained on Numpy arrays of input data and labels. For training a model, you will typically use the  fit function. <\/p>\n<\/blockquote>\n\n<p>To convert a pandas dataframe to numpy array you can use <code>np.array(dataframe)<\/code>. For example: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_train = np.array(x_train)\n<\/code><\/pre>\n"},{"AnswerId":"36798335","CreationDate":"2016-04-22T15:53:27.610","ParentId":null,"OwnerUserId":"1001658","Title":null,"Body":"<p>Answer from the comment - trainx and trainy should be numpy arrays. You can convert the data frame to numpy array using <code>as_matrix()<\/code> method. I also faced this issue. It's weird that Keras does not give meaningful error message.<\/p>\n"},{"AnswerId":"39837685","CreationDate":"2016-10-03T18:01:07.703","ParentId":null,"OwnerUserId":"4634344","Title":null,"Body":"<p>I came here looking for the same issue resolution for the auto-sklearn and pandas dataframe. The solution is to pass the X dataframe as X.values. I.e. fit(X.values,y)<\/p>\n"}]}
{"QuestionId":35969833,"AnswerCount":1,"Tags":"<cmake><torch><luarocks>","CreationDate":"2016-03-13T11:46:50.917","AcceptedAnswerId":null,"OwnerUserId":5576325.0,"Title":"How can I fix this \"luarocks make\" error when compiling a local torch7 package?","Body":"<p>I am trying to building a package written in torch7(C,cuda,lua) from github <a href=\"https:\/\/github.com\/mbhenaff\/spectral-lib\" rel=\"nofollow\">spectral-lib<\/a>(Ubuntu 14.04, cuda 7.0, torch7 installed). I use this command to build it:\n.<\/p>\n\n<p>However, strange errors occurred as below:<\/p>\n\n<pre><\/pre>\n\n<p>I can find that \"\" is in .\nMaybe this error is related to the CMakeLists.txt, but I'm not familiar with it. Can anyone give me a hand?<\/p>\n","answers":[{"AnswerId":"35971079","CreationDate":"2016-03-13T13:52:34.410","ParentId":null,"OwnerUserId":"5576325","Title":null,"Body":"<p>Solved by modifying the CMakeLists.txt. I copy <code>libcufft.so<\/code> to .\/cuda and add:<\/p>\n\n<pre><code>SET(PROJECT_LINK_LIBS libcufft.so )\nLINK_DIRECTORIES(\/net\/wanggu\/spectral-lib\/cuda)\n<\/code><\/pre>\n\n<p>The author's CMakeLists.txt may have some problem with the directions of cuda dynamic libraries. There may exist some better solutions, but this is a temporary solution for now.<\/p>\n"}]}
{"QuestionId":35972165,"AnswerCount":1,"Tags":"<macos><python-2.7><tensorflow><tensorboard>","CreationDate":"2016-03-13T15:34:54.480","AcceptedAnswerId":null,"OwnerUserId":6057339.0,"Title":"Error with tensorboard on Mac","Body":"<p>I have recently installed tensorflow, and it seems working fine until I was trying to display computing graph using tensorboard. I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestion about how to solve the problem?<\/p>\n","answers":[{"AnswerId":"36095137","CreationDate":"2016-03-18T22:16:52.520","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>This looks like a TensorBoard bug rather than a usage issue. Could you file a bug against it on our <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/new\" rel=\"nofollow\">Github<\/a>?<\/p>\n\n<p>It will be way easier for us to track it down if you can give us a clean repro, including:\n- How you installed TensorFlow\n- The events file that triggers this bug<\/p>\n\n<p>The issue you're encountering is related to the latest code from <code>master<\/code>; most people install from pip, but I've been seeing a few issues where people's pip installations are somehow sucking in new code from master. Are you intending to use cutting-edge unreleased source from our master branch, or did you install e.g. version 0.7.1 and are surprised that you have unstable code?<\/p>\n"}]}
{"QuestionId":35972194,"AnswerCount":1,"Tags":"<python><loops><for-loop><dynamic><tensorflow>","CreationDate":"2016-03-13T15:37:29.167","AcceptedAnswerId":36075770.0,"OwnerUserId":2776945.0,"Title":"variable length for loop - tensorflow","Body":"<p>I want to do for loop's with dynamic inputs in tensorflow.<\/p>\n\n<p>I know for the code below  is sufficient, but I want to be able to formulate code in the manner described below. So given that I have an array with different sizes, I want to traverse that array. <\/p>\n\n<p>How can I traverse dynamic length arrays?<\/p>\n\n<p>Problem: <\/p>\n\n<blockquote>\n  <p>TypeError: range() integer end argument expected, got Tensor<\/p>\n<\/blockquote>\n\n<p>How can I evaluate get_sum that is outside the session? <\/p>\n\n<p> shouldn't be in the  loop inside session<\/p>\n\n<p>For example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36075770","CreationDate":"2016-03-18T03:48:06.930","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>Dynamic loops are still young in TensorFlow, but they exist. Take a look for example at the fold operators and their implementation.\n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#foldl\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#foldl<\/a><\/p>\n"}]}
{"QuestionId":35974714,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-13T19:26:26.147","AcceptedAnswerId":null,"OwnerUserId":2267819.0,"Title":"Tensorboard __init__() got an unexpected keyword argument 'purge_orphaned_data'","Body":"<p>Running tensorboard for first time in a new VM and I end up with this error at tensorboard init(). Thanks for the help.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36086749","CreationDate":"2016-03-18T14:19:38.513","ParentId":null,"OwnerUserId":"5089833","Title":null,"Body":"<p>I have the same problem. I deleted \"purge_orphaned_data=FLAGS.purge_orphaned_data\" and it worked.<\/p>\n"},{"AnswerId":"36095024","CreationDate":"2016-03-18T22:07:50.940","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>TensorBoard developer here. I'm quite confused how this happened, because while the purge_orphaned_data flag exists in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tensorboard\/tensorboard.py\" rel=\"nofollow\">master<\/a>, it doesn't exist in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v0.7.1\/tensorflow\/tensorboard\/backend\/tensorboard.py\" rel=\"nofollow\">r0.7.1<\/a>. Actually, the file path tensorflow\/tensorboard\/tensorboard.py doesn't exist in 0.7.1 (it is in tensorboard\/backend\/tensorboard.py in 0.7.1).<\/p>\n\n<p>So it looks like somehow you got an installation that mixes r0.7.1 but also files from master. Could you please file a Github issue on TensorFlow with a repro on how you got into this state?<\/p>\n"}]}
{"QuestionId":35974826,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-03-13T19:36:31.390","AcceptedAnswerId":35975060.0,"OwnerUserId":1191087.0,"Title":"TensorFlow - Setting the shape of an image while reading and writing a TFRecords file?","Body":"<p>I'm running into trouble setting the shape of my image data when trying to work with TFRecords format. I've been going over the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html\" rel=\"nofollow\">how-to for reading data<\/a> and have taken the code from the MNIST example for <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow\">converting the image data to a TFRecords<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.7\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow\">reading the data from the TFRecords<\/a>. However, this example code originally expects the image to be used in a format where all the pixel data is in one long vector.<\/p>\n\n<p>I've been trying to alter this code to work with NumPy arrays that still are in the original image shape. So in my code below,  is a NumPy array with shape . I'm not sure if my problem is on the side of how I'm writing the data to the TFRecords or how I'm reading it back out. However, when I try to set the shape of the decoded image, I get the error  (note: 464 x 624 x 3 are the image dimensions). Any suggestions as to what I may be doing wrong?<\/p>\n\n<p><strong>Relavent code (slightly altered from example code)<\/strong>:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"35975060","CreationDate":"2016-03-13T19:56:49.407","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Note that <code>set_shape<\/code> doesn't change the shape of underlying buffer, it merely sets a graph-level annotation of possible set of shapes that will be seen at this tensor.<\/p>\n\n<p>To change the actual shape you need to use <code>tf.reshape<\/code><\/p>\n"}]}
{"QuestionId":35975845,"AnswerCount":1,"Tags":"<python><theano><lasagne>","CreationDate":"2016-03-13T21:03:07.690","AcceptedAnswerId":null,"OwnerUserId":4451856.0,"Title":"How to get the output target values for the test data in the following MLP of Lasagne","Body":"<p><a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/user\/tutorial.html#id2\" rel=\"nofollow\">http:\/\/lasagne.readthedocs.org\/en\/latest\/user\/tutorial.html#id2<\/a><\/p>\n\n<p>I have tried the following<\/p>\n\n<pre><\/pre>\n\n<p>however I get NAN for all the samples in Y_hat here.<\/p>\n\n<p>EDIT: \nI was able to solve the NAN issue. However now my prediction returns only one class (1)<\/p>\n","answers":[{"AnswerId":"36154476","CreationDate":"2016-03-22T12:31:55.210","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>network_output is a matrix of shape (N,K) where N is the number of datapoints and K is the number of classes; it is a raw scores matrix. In your code network_output[:,-1] will be all the raw score values for the K-th or last label. In other words, you are returning a column vector, not the entire raw scores matrix. To output all the scores simply modify your theano.function to this:<\/p>\n\n<pre><code>f = theano.function([input_var], network_output)\n<\/code><\/pre>\n\n<p>and you should be returning all the raw class scores. <\/p>\n"}]}
{"QuestionId":35976511,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-13T22:10:53.217","AcceptedAnswerId":null,"OwnerUserId":5948325.0,"Title":"Issue with TensorFlow Summary","Body":"<p>I tried to modify \"\" for analyzing image data from my research.<\/p>\n\n<p>When I tried to use TensorBoard for visualization, following error appeared ():<\/p>\n\n<pre><\/pre>\n\n<p>I was wondering why ? what does this \"Out of range\" mean ? <\/p>\n\n<p>Here is part of my modified code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Please help me!<\/p>\n","answers":[]}
{"QuestionId":35976954,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-03-13T22:56:51.797","AcceptedAnswerId":null,"OwnerUserId":2569930.0,"Title":"calculating cross entropy manually vs using softmax_cross_entropy_with_logits in tensorflow","Body":"<p>I'm running into an issue where I'm trying to create a deep ReLU network using tensorflow for the MNIST dataset. It's working fine when I use my loss as the built in <strong><em>tf.nn.softmax_cross_entropy_with_logits()<\/em><\/strong>, but calculating the entropy term manually doesn't seem to work. <\/p>\n\n<p>Here is what the network looks like:<\/p>\n\n<pre><\/pre>\n\n<p>In other words this works fine:<\/p>\n\n<pre><\/pre>\n\n<p>but not this:<\/p>\n\n<pre><\/pre>\n\n<p>The latter runs fine, but the accuracy gets stuck after an initial step. The former using the softmax_cross_entropy_with_logits function actually does learn something. I've seen the latter's setup being used for the deep MNIST example, which is why I'm wondering what it is about my setup here that is causing the optimization procedure to stall.<\/p>\n","answers":[{"AnswerId":"43418648","CreationDate":"2017-04-14T20:17:17.587","ParentId":null,"OwnerUserId":"6398153","Title":null,"Body":"<p><strong>Update:<\/strong><\/p>\n\n<p>Finally, I can solve this problem with implementing the inside of <code>softmax_cross_entropy_with_logits()<\/code> function by myself, you can find the code <a href=\"https:\/\/github.com\/AliAbbasi\/Numerically-Stable-Cross-Entropy-Loss-Function-Tensorflow\" rel=\"nofollow noreferrer\">here<\/a> on my GitHub. It is in two version for normal and multi-label problems.<\/p>\n\n<hr>\n\n<p><strong>Previous Answer:<\/strong><\/p>\n\n<p>Originally from tensorflow API:<\/p>\n\n<p>\"(Note that in the source code, we don't use this formulation,<\/p>\n\n<p><code>cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))<\/code><\/p>\n\n<p>because it is numerically unstable. Instead, we apply tf.nn.softmax_cross_entropy_with_logits on the unnormalized logits (e.g., we call softmax_cross_entropy_with_logits on tf.matmul(x, W) + b), because this more numerically stable function internally computes the softmax activation. In your code, consider using tf.nn.(sparse_)softmax_cross_entropy_with_logits instead)\"<\/p>\n\n<p>Source: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/mnist\/beginners\/<\/a> <\/p>\n"},{"AnswerId":"42989168","CreationDate":"2017-03-24T00:06:59.390","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>I think you're missing a few steps in order to accurately compute your cost. Consider looking at the source code in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d489dc57cfa991eb9471945ac3e1a21ccaf02aea\/tensorflow\/python\/ops\/nn_ops.py#L1585\" rel=\"nofollow noreferrer\"><code>nn_ops.py<\/code><\/a> to see what else <code>softmax_cross_entropy_with_logits<\/code> is doing.<\/p>\n"}]}
{"QuestionId":35978035,"AnswerCount":0,"Tags":"<caffe><activation><conv-neural-network><imagenet>","CreationDate":"2016-03-14T01:15:47.787","AcceptedAnswerId":null,"OwnerUserId":3698878.0,"Title":"Caffe: Why in Imagenet model activation dimension of fc6 is 4096?","Body":"<p>I was looking at pool5 layer in the Imagenet trained model (ILSVRC challenge) whose output size is 256*6*6 (~9000), after which there is a fc6 layer whose num_outputs is 4096. Can anyone please explain how is 4096 chosen?<\/p>\n","answers":[]}
{"QuestionId":35980044,"AnswerCount":3,"Tags":"<neural-network><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2016-03-14T05:20:18.517","AcceptedAnswerId":null,"OwnerUserId":2886263.0,"Title":"Getting the output shape of deconvolution layer using tf.nn.conv2d_transpose in tensorflow","Body":"<p>According to this <a href=\"http:\/\/www.matthewzeiler.com\/pubs\/cvpr2010\/cvpr2010.pdf\" rel=\"nofollow\">paper<\/a>, the output shape is ,  is input height or width,  is kernel height or width. This is obvious inverse process of convolution. This <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow\">tutorial<\/a> gives a formula to calculate the output shape of convolution which is ,  - input size,  - filter size,  - padding size,  - stride. But in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/nn_test.py\" rel=\"nofollow\">Tensorflow<\/a>, there are test cases like:         <\/p>\n\n<pre><\/pre>\n\n<p>So we use ,  and , according to formula  to calculate padding size . From , we get , which is not an integer. How does deconvolution works in Tensorflow?<\/p>\n","answers":[{"AnswerId":"49347788","CreationDate":"2018-03-18T12:18:37.227","ParentId":null,"OwnerUserId":"1587156","Title":null,"Body":"<p>for deconvolution,<\/p>\n\n<pre><code>output_size = strides * (input_size-1) + kernel_size - 2*padding\n<\/code><\/pre>\n\n<p>strides, input_size, kernel_size, padding are integer\npadding is zero for 'valid'<\/p>\n"},{"AnswerId":"41416361","CreationDate":"2017-01-01T14:54:48.740","ParentId":null,"OwnerUserId":"6734219","Title":null,"Body":"<p>This discussion is really helpful. Just add some additional information.\n<code>padding='SAME'<\/code> can also let the bottom and right side get the one additional padded pixel. According to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution\" rel=\"nofollow noreferrer\">TensorFlow document<\/a>, and the test case below<\/p>\n\n<pre><code>strides = [1, 2, 2, 1]\n# Input, output: [batch, height, width, depth]\nx_shape = [2, 6, 4, 3]\ny_shape = [2, 12, 8, 2]\n\n# Filter: [kernel_height, kernel_width, output_depth, input_depth]\nf_shape = [3, 3, 2, 3]\n<\/code><\/pre>\n\n<p>is using padding='SAME'. We can interpret padding='SAME' as:<\/p>\n\n<pre><code>(W\u2212F+pad_along_height)\/S+1 = out_height,\n(W\u2212F+pad_along_width)\/S+1 = out_width.\n<\/code><\/pre>\n\n<p>So <code>(12 - 3 + pad_along_height) \/ 2 + 1 = 6<\/code>, and we get <code>pad_along_height=1<\/code>. And <code>pad_top=pad_along_height\/2 = 1\/2 = 0<\/code>(integer division), <code>pad_bottom=pad_along_height-pad_top=1<\/code>. <\/p>\n\n<p>As for padding='VALID', as the name suggested, we use padding when it is proper time to use it. At first, we assume that the padded pixel = 0, if this doesn't work well, then we add 0 padding where any value outside the original input image region. For example, the test case below,<\/p>\n\n<pre><code>strides = [1, 2, 2, 1]\n\n# Input, output: [batch, height, width, depth]\nx_shape = [2, 6, 4, 3]\ny_shape = [2, 13, 9, 2]\n\n# Filter: [kernel_height, kernel_width, output_depth, input_depth]\nf_shape = [3, 3, 2, 3]\n<\/code><\/pre>\n\n<p>The output shape of <code>conv2d<\/code> is <\/p>\n\n<pre><code>out_height = ceil(float(in_height - filter_height + 1) \/ float(strides[1]))\n           = ceil(float(13 - 3 + 1) \/ float(3)) = ceil(11\/3) = 6\n           = (W\u2212F)\/S + 1.\n<\/code><\/pre>\n\n<p>Cause <code>(W\u2212F)\/S+1 = (13-3)\/2+1 = 6<\/code>, the result is an integer, we don't need to add 0 pixels around the border of the image, and <code>pad_top=1\/2<\/code>, <code>pad_left=1\/2<\/code> in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution\" rel=\"nofollow noreferrer\">TensorFlow document<\/a> padding='VALID' section are all 0.<\/p>\n"},{"AnswerId":"36019063","CreationDate":"2016-03-15T18:16:18.093","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>The formula for the output size from the tutorial assumes that the padding <code>P<\/code> is the same before and after the image (left &amp; right or top &amp; bottom).\nThen, the number of places in which you put the kernel is:\n<code>W (size of the image) - F (size of the kernel) + P (additional padding before) + P (additional padding after)<\/code>.<\/p>\n\n<p>But tensorflow also handles the situation where you need to pad more pixels to one of the sides than to the other, so that the kernels would fit correctly. You can read more about the strategies to choose the padding (<code>\"SAME\"<\/code> and <code>\"VALID\"<\/code>) in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#convolution\" rel=\"nofollow\">docs<\/a>. The test you're talking about uses method <code>\"VALID\"<\/code>.<\/p>\n"}]}
{"QuestionId":35983047,"AnswerCount":0,"Tags":"<ubuntu><caffe><feature-extraction>","CreationDate":"2016-03-14T08:52:16.940","AcceptedAnswerId":null,"OwnerUserId":6032253.0,"Title":"Segmentation fault (core dump) when extract features in caffe?","Body":"<p>When I follow the tutorial about feature extraction using a pre-trained model with caffe, I met some problem.\n As the tutorial says, I have made the temp.txt to generate a list of the files to process. To add expects labels after each filenames I have also made the file_list.txt. Then I used imagenet_mean.binaryproto and imagenet_val.prototxt to define the feature extraction network architecture. After this, I input the command in the terminal like this \".\/build\/tools\/extract_features.bin models\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel examples\/_temp\/imagenet_val.prototxt fc7 examples\/_temp\/features 10 leveldb\". \nThe expected result is just like the following image shown<a href=\"http:\/\/i.stack.imgur.com\/eqWT2.png\" rel=\"nofollow\">enter image description here<\/a>\nWhile I met the problem just like this<a href=\"http:\/\/i.stack.imgur.com\/Vot60.png\" rel=\"nofollow\">enter image description here<\/a>\nI have found all the necessary files mentioned in the tutorial. I really can not figure out the problem hoping someone will help me. I will be so appreciate! English is not my mother tongue, I hope readers can be more patient.<\/p>\n","answers":[]}
{"QuestionId":35989069,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-03-14T13:37:28.030","AcceptedAnswerId":null,"OwnerUserId":3919542.0,"Title":"How to test the model created using mnist dataset","Body":"<p>I am new to caffe and thus was trying to play around with the MNIST dataset . I ran the following commands <\/p>\n\n<pre><\/pre>\n\n<p>I haven't changed any of the files in caffe and has executed just the above 3 lines.<\/p>\n\n<p>My example\/mnist folder structure is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I want to test the model created by passing an image that is of size 256*256 but don't know how to do it, and want to do it using python . Any help would be appreciated .<\/p>\n","answers":[{"AnswerId":"37862919","CreationDate":"2016-06-16T15:05:10.817","ParentId":null,"OwnerUserId":"1335334","Title":null,"Body":"<ol>\n<li><p>You will need to downscale your image to the desired size 28x28.<\/p><\/li>\n<li><p>You have already created a dataset in LMDB format:<\/p>\n\n<ul>\n<li>.\/data\/mnist\/get_mnist.sh<\/li>\n<li>.\/examples\/mnist\/create_mnist.sh<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>And have already trained the model:<\/p>\n\n<ul>\n<li>.\/examples\/mnist\/lenet_train_test.prototxt<\/li>\n<\/ul>\n\n<p>whose weights are stored as:<\/p>\n\n<ul>\n<li>.\/examples\/mnist\/lenet_iter_10000.caffemodel<\/li>\n<\/ul>\n\n<p>You now want to deploy your classifier. lenet.prototxt is the same model at the core, except it's input and outputs have changed. When you deploy, you want to have as input a regular image in practice.<\/p>\n\n<ul>\n<li>.\/examples\/mnist\/lenet.prototxt : deployment model<\/li>\n<li>.\/examples\/mnist\/lenet_iter_10000.caffemodel : learned weights<\/li>\n<\/ul>\n\n<p>.<\/p>\n\n<p>I am not aware if you can simply use the caffe tool and run the test command on an image.<\/p>\n\n<p>The easiest ways to classify seem to be using matlab or python.<\/p>\n\n<p>Here is a matlab example:<\/p>\n\n<pre><code>model = 'lenet.prototxt';\nweights = 'lenet_iter_10000.caffemodel';\n%caffe.set_mode_gpu();\n%caffe.set_device(0);\nnet = caffe.Net(model, weights, 'test');\nimage = imread('example.png');\nres = net.forward({image});\nprob = res{1}\n<\/code><\/pre>\n\n<p>Here is a classification tutorial using python (<a href=\"https:\/\/nbviewer.ipython.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">.\/examples\/00-classification.ipynb<\/a>). However, it's designed for ilsvrc12 type datasets. If you simply ignore subtracting the mean image, it might just work.<\/p>\n\n<p>There is also a CPP example (.\/examples\/cpp_classification), but again I'm not sure about the mean image.<\/p>\n"}]}
{"QuestionId":35990446,"AnswerCount":1,"Tags":"<ubuntu><tensorflow>","CreationDate":"2016-03-14T14:36:30.903","AcceptedAnswerId":null,"OwnerUserId":6040027.0,"Title":"TensorFlow Android example: Cannot find WORKSPACE file","Body":"<p>I installed TensorFlow on Ubuntu. I also installed Android NDK and SDK, in order to build the example application. According to the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/README.md\" rel=\"nofollow\">instructions<\/a>, I now need to edit the  file to set the NDK and SDK installation paths, but I cannot find the  file.<\/p>\n\n<p>Could you please advise me how I can locate the WORKSPACE file?<\/p>\n","answers":[{"AnswerId":"35991020","CreationDate":"2016-03-14T15:02:09.693","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/WORKSPACE\" rel=\"nofollow\"><code>WORKSPACE<\/code> file<\/a> is in the root directory of the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow\">TensorFlow git repository<\/a>. It sounds like you may have installed a binary version of TensorFlow&mdash;at present you must build the Android example <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">from source<\/a>.<\/p>\n"}]}
{"QuestionId":35990447,"AnswerCount":0,"Tags":"<multithreading><caffe><openblas>","CreationDate":"2016-03-14T14:36:30.963","AcceptedAnswerId":null,"OwnerUserId":6061106.0,"Title":"Can't use multi CPU cores to train NNs using Caffe and OpenBLAS","Body":"<p>I've read this discussion: <a href=\"https:\/\/stackoverflow.com\/questions\/30195837\/how-to-use-multi-cpu-cores-to-train-nns-using-caffe-and-openblas\">How to use multi CPU cores to train NNs using caffe and OpenBLAS<\/a><\/p>\n\n<p>My machine has 4 CPUs.<\/p>\n\n<p>I've set the following variables:<\/p>\n\n<pre><\/pre>\n\n<p>however it doesn't help:<\/p>\n\n<pre><\/pre>\n\n<p>For the comparison here is the result for the case when<\/p>\n\n<pre><\/pre>\n\n<p><strong>one iteration still takes about 80 seconds.<\/strong><\/p>\n\n<p>As I see it, the variables mentioned don't make any effect (the difference in timing can be explained by different load of the machine).\nHow can I ensure that  really does not exploit the 4 CPUs and why?<\/p>\n","answers":[]}
{"QuestionId":35993175,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2016-03-14T16:39:19.020","AcceptedAnswerId":36005826.0,"OwnerUserId":6061106.0,"Title":"How can I get two output values (for each of the two classes) for a binary classifier in Caffe?","Body":"<p>I'm experimenting with LeNet network as a binary classifier (yes, no).\nThe first and several last layers in the configuration file for testing  is the following:<\/p>\n\n<pre><\/pre>\n\n<p>For testing I've set the batch_size=1, thus I ran testing with the following command:<\/p>\n\n<pre><\/pre>\n\n<p>My intent is to be able to analyze result for each test image separately.\nCurrently I get the following info for each iteration:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>However since I have two outputs in my network, on testing I want to see two separate values for each of the outputs: one for class \"0\" (\"no\") and one for class \"1\" (\"yes\"). It should be something like that:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p><strong>How should I modify the testing configuration file to make it happen?<\/strong><\/p>\n","answers":[{"AnswerId":"36005826","CreationDate":"2016-03-15T08:20:46.660","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You want to see the <code>\"Softmax\"<\/code> probability output (not just the loss).<br>\nFor this end you might try to use <code>\"SoftmaxWithLoss\"<\/code> with two <code>\"top\"<\/code>s (I'm not 100% sure this option is fully functional\/supported):<\/p>\n\n<pre><code>layer {\n  name: \"loss\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"ip2\"\n  bottom: \"label\"\n  top: \"loss\"\n  top: \"prob\" # add class probability output\n}\n<\/code><\/pre>\n\n<hr>\n\n<p>Alternatively, if the former solution does not work, explicitly add a <code>\"Softmax\"<\/code> layer:<\/p>\n\n<pre><code>layer {\n  name: \"prob\"\n  type: \"Softmax\"\n  bottom: \"ip2\"\n  top: \"prob\"\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":35997148,"AnswerCount":0,"Tags":"<machine-learning><neural-network><theano><starcluster>","CreationDate":"2016-03-14T20:11:57.223","AcceptedAnswerId":null,"OwnerUserId":6062847.0,"Title":"Theano lock issue Starcluster","Body":"<p>I have a problem with theano when I run it on Starcluster,  I create a cluster with 8 nodes one master and 7 nodes, when I run 108 neural network on the cluster they try to access the theano compile directory and because that directory is locked  it gives me the following  error:<\/p>\n\n<p>INFO (theano.gof.compilelock): To manually release the lock, delete \/data\/users\/yifeic\/.theano\/compiledir_Linux-*****\/lock_dir<\/p>\n\n<p>it tries to access to this directory several times and then it just kicks out the job.  So I run 108 neural network and I ended up having about 50 of them running. <\/p>\n\n<p>I have tried several things like changing the config.compiledir so that each nodes uses their own directory for compliledir <\/p>\n\n<p>and 2)  To force the theano to release the lock using theano.gof.compilelock.release_lock() <\/p>\n\n<p>but still  starcluster kicks out the jobs. <\/p>\n\n<p>Does anyone have had experience with theano and starcluster or any experience with this lock issue?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
{"QuestionId":35997851,"AnswerCount":0,"Tags":"<deep-learning><caffe><conv-neural-network><nvidia-digits>","CreationDate":"2016-03-14T20:52:29.277","AcceptedAnswerId":null,"OwnerUserId":6063255.0,"Title":"Using RCNN to autocrop an image","Body":"<p>I am new to machine learning. I've been messing around with NVIDIA Digits to train a new dataset. My dataset however is too inaccurate and I think it is because there is too much background in the image that it is getting confused as to what the actual object is. My question:<\/p>\n\n<p>Is there a way (possibly using RCNN) to crop out the background and then proceed to train using the cropped image? The object is consistent (ex only one object like a singular person but there may be people in the background) and always by itself.<\/p>\n","answers":[]}
{"QuestionId":35998515,"AnswerCount":2,"Tags":"<python><gpu><theano>","CreationDate":"2016-03-14T21:32:52.620","AcceptedAnswerId":35999446.0,"OwnerUserId":395857.0,"Title":"Why does the floatX's flag impact whether GPU is used in Theano?","Body":"<p>I am testing Theano with GPU using the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#testing-theano-with-gpu\" rel=\"nofollow\">script provided in the tutorial for that purpose<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>If I specify , it runs on GPU:<\/p>\n\n<pre><\/pre>\n\n<p>If I do not specify , it runs on CPU:<\/p>\n\n<pre><\/pre>\n\n<p>If I specify , it runs on CPU:<\/p>\n\n<pre><\/pre>\n\n<p>Why does the  flag impact whether GPU is used in Theano?<\/p>\n\n<p>I use:<\/p>\n\n<ul>\n<li>Theano 0.7.0 (according to ), <\/li>\n<li>Python 2.7.6 64 bits (according to ), <\/li>\n<li>Nvidia-smi 361.28 (according to ), <\/li>\n<li>CUDA 7.5.17 (according to ),<\/li>\n<li>GeForce GTX Titan X (according to ), <\/li>\n<li>Ubuntu 14.04.4 LTS x64 (according to  and ).<\/li>\n<\/ul>\n\n<p>I read the documentation on <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.floatX\" rel=\"nofollow\"><\/a> but it didn't help. It simply says:<\/p>\n\n<blockquote>\n  <p><br> String value: either \u2018float64\u2019 or \u2018float32\u2019<br>\n  Default: \u2018float64\u2019<br><\/p>\n  \n  <p>This sets the default dtype returned by tensor.matrix(),\n  tensor.vector(), and similar functions. It also sets the default\n  theano bit width for arguments passed as Python floating-point\n  numbers.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"35999446","CreationDate":"2016-03-14T22:38:30.433","ParentId":null,"OwnerUserId":"99989","Title":null,"Body":"<p>As far as I know, it's because they haven't yet implemented float64 for GPUs.<\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html<\/a> :<\/p>\n\n<blockquote>\n  <p>Only computations with float32 data-type can be accelerated. Better support for float64 is expected in upcoming hardware but float64 computations are still relatively slow (Jan 2010).<\/p>\n<\/blockquote>\n"},{"AnswerId":"36766571","CreationDate":"2016-04-21T10:16:16.023","ParentId":null,"OwnerUserId":"3583290","Title":null,"Body":"<p>From <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#gpuarray-backend\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html#gpuarray-backend<\/a> I read that it is possible to perform float64 calculations on GPU, but you have to install the <code>libgpuarray<\/code> from source.<\/p>\n\n<p>I managed to install it, see <a href=\"https:\/\/gist.github.com\/gaebor\/61323e72a7b415e6487e9590ce181857\" rel=\"nofollow\">this script<\/a>, I used <a href=\"https:\/\/pypi.python.org\/pypi\/virtualenv\" rel=\"nofollow\">virtualenv<\/a>, you don't even have to have <code>sudo<\/code>.<\/p>\n\n<p>After installation you can use the old backend with <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">config flag<\/a> <code>device=gpu<\/code> and the new backend with <code>device=cuda<\/code>.<\/p>\n\n<p>The new backend can perform 64 bit calculations, but it works differently for me. Some operations stopped working. <code>ABSOLUTELY NO WARRANTY, to the extent permitted by applicable law<\/code> :)<\/p>\n"}]}
{"QuestionId":36002325,"AnswerCount":0,"Tags":"<python><theano><convolution>","CreationDate":"2016-03-15T03:46:17.663","AcceptedAnswerId":null,"OwnerUserId":5303293.0,"Title":"Lenet convolution neural network implementation in python using theano giving same error value in simultaneous training","Body":"<p>I am a newbie to convolution neural network and theano. So I tried to run the code from the tutorial <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html<\/a> . The code was working good but the output seems to be wrong. For the initialization of weights, we are using the \"numpy.random.RandomState(23455)\" values. I ran two instances of the code and in each instance, the error value was \"cost_ij=2.302585\". How is that possible ?? We are using different random numbers on both neural networks, So how could the output be the same??  for getting fast result, I changed the number of epochs as 1, patience=10 and patience increase as 2.<\/p>\n\n<p>I am getting a warning:: \"\/convolutional_mlp.py:97: UserWarning: The  keyword argument to  is deprecated, it has been renamed to .\" Is this the reason for the network to behave like the same??? <\/p>\n\n<p>If we are using different weights, don't we need to get different outputs ??<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance for the answers...<\/p>\n","answers":[]}
{"QuestionId":36002387,"AnswerCount":1,"Tags":"<python-2.7><caffe><pycaffe><mnist>","CreationDate":"2016-03-15T03:52:25.350","AcceptedAnswerId":36075282.0,"OwnerUserId":3919542.0,"Title":"Channel swap needs to have the same number of dimensions as the input channels error in caffe","Body":"<p>I am new to caffe thus was trying to build a model using the MNIST dataset , I have executed the following instructions :<\/p>\n\n<pre><\/pre>\n\n<p>At the end of the last instruction, a series of outputs were printed on the console , the last few lines of it is :<\/p>\n\n<pre><\/pre>\n\n<p>I tried to now use the model created as binay proto file in examples\/mnist\/lenet_iter_10000.caffemodel<\/p>\n\n<p>i ran the following command giving my image as input for testing the model:<\/p>\n\n<pre><\/pre>\n\n<p><strong>My image is a gray scale 256*256 image (I have tried even with a 256*256 RGB image)<\/strong><\/p>\n\n<p>However an error occurs saying :<\/p>\n\n<pre><\/pre>\n\n<p>I haven't changed any of the files , except classify.py where i have added these line of code :<\/p>\n\n<pre><\/pre>\n\n<p>My file structure for the .\/examples\/mnist :<\/p>\n\n<pre><\/pre>\n\n<p>File structure for .\/data\/mnist :<\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"36075282","CreationDate":"2016-03-18T02:49:28.833","ParentId":null,"OwnerUserId":"3919542","Title":null,"Body":"<p>I figured out what was the problem , Actually classify.py in caffe is written for the imagenet model by default and takes a 3 dimensional channel as input . However <strong><em>The MNIST model requires one dimensional channel as input<\/em><\/strong>. So when ever i was passing the image , It was showing me this dimension mismatch error , Thus necessary code had to be written to make this MNIST caffemodel to work .<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2359\/commits\/dd3a5f9268ca3bdf19a17760bd6f568e21c1b521\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/pull\/2359\/commits\/dd3a5f9268ca3bdf19a17760bd6f568e21c1b521<\/a><\/p>\n\n<p>All necessary that need to be written has been given in this link . Hope it helps .<\/p>\n"}]}
{"QuestionId":36003229,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-15T05:21:23.610","AcceptedAnswerId":null,"OwnerUserId":6047866.0,"Title":"TensorFlow setup on window","Body":"<p>I am starting to do TensorFlow for my project but sadly I use Window 32bit. On the website they showed us how to install on Linux or iOS could anyone know how to setup on Windows.<\/p>\n\n<p>Thanks very much.<\/p>\n","answers":[{"AnswerId":"36050591","CreationDate":"2016-03-17T02:27:20.700","ParentId":null,"OwnerUserId":"6047866","Title":null,"Body":"<p>I set up my project on clound9, It run on Unbuntu. It is not great as destop but could be used anyway. \nThanks<\/p>\n"},{"AnswerId":"36007074","CreationDate":"2016-03-15T09:21:39.513","ParentId":null,"OwnerUserId":"4327543","Title":null,"Body":"<p>The best way or the only way for you is to using Pre-built Docker container with TensorFlow. To get started with TensorFlow quickly and work on your project, follow the instructions in following links:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md<\/a><\/p>\n"}]}
{"QuestionId":36007883,"AnswerCount":6,"Tags":"<python><machine-learning><linear-regression><tensorflow>","CreationDate":"2016-03-15T09:55:38.467","AcceptedAnswerId":36016117.0,"OwnerUserId":3940080.0,"Title":"TensorFlow: \"Attempting to use uninitialized value\" in variable initialization","Body":"<p>I am trying to implement multivariate linear regression in Python using TensorFlow, but have run into some logical and implementation issues. My code throws the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Ideally the  output should be <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43118291","CreationDate":"2017-03-30T12:45:40.763","ParentId":null,"OwnerUserId":"7792003","Title":null,"Body":"<p>I want to give my resolution, it work when i replace the line <code>[session = tf.Session()]<\/code> with <code>[sess = tf.InteractiveSession()]<\/code>. Hope this will be useful to others.<\/p>\n"},{"AnswerId":"55797959","CreationDate":"2019-04-22T16:52:35.387","ParentId":null,"OwnerUserId":"8041407","Title":null,"Body":"<p>run both: <\/p>\n\n<p>sess.run(tf.global_variables_initializer())<\/p>\n\n<p>sess.run(tf.local_variables_initializer())<\/p>\n"},{"AnswerId":"45840835","CreationDate":"2017-08-23T13:16:04.177","ParentId":null,"OwnerUserId":"1125621","Title":null,"Body":"<p>There is another the error happening which related to the order when calling initializing global variables. I've had the sample of code has similar error <strong>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value W<\/strong><\/p>\n\n<pre><code>def linear(X, n_input, n_output, activation = None):\n    W = tf.Variable(tf.random_normal([n_input, n_output], stddev=0.1), name='W')\n    b = tf.Variable(tf.constant(0, dtype=tf.float32, shape=[n_output]), name='b')\n    if activation != None:\n        h = tf.nn.tanh(tf.add(tf.matmul(X, W),b), name='h')\n    else:\n        h = tf.add(tf.matmul(X, W),b, name='h')\n    return h\n\nfrom tensorflow.python.framework import ops\nops.reset_default_graph()\ng = tf.get_default_graph()\nprint([op.name for op in g.get_operations()])\nwith tf.Session() as sess:\n    # RUN INIT\n    sess.run(tf.global_variables_initializer())\n    # But W hasn't in the graph yet so not know to initialize \n    # EVAL then error\n    print(linear(np.array([[1.0,2.0,3.0]]).astype(np.float32), 3, 3).eval())\n<\/code><\/pre>\n\n<p>You should change to following<\/p>\n\n<pre><code>from tensorflow.python.framework import ops\nops.reset_default_graph()\ng = tf.get_default_graph()\nprint([op.name for op in g.get_operations()])\nwith tf.Session() as \n    # NOT RUNNING BUT ASSIGN\n    l = linear(np.array([[1.0,2.0,3.0]]).astype(np.float32), 3, 3)\n    # RUN INIT\n    sess.run(tf.global_variables_initializer())\n    print([op.name for op in g.get_operations()])\n    # ONLY EVAL AFTER INIT\n    print(l.eval(session=sess))\n<\/code><\/pre>\n"},{"AnswerId":"36016117","CreationDate":"2016-03-15T15:52:28.727","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It's not 100% clear from the code example, but if the list <code>initial_parameters_of_hypothesis_function<\/code> is a list of <code>tf.Variable<\/code> objects, then the line <code>session.run(init)<\/code> will fail because TensorFlow isn't (yet) smart enough to figure out the dependencies in variable initialization. To work around this, you should change the loop that creates <code>parameters<\/code> to use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable.initialized_value\" rel=\"noreferrer\"><code>initial_parameters_of_hypothesis_function[i].initialized_value()<\/code><\/a>, which adds the necessary dependency:<\/p>\n\n<pre><code>parameters = []\nfor i in range(0, number_of_attributes, 1):\n    parameters.append(tf.Variable(\n        initial_parameters_of_hypothesis_function[i].initialized_value()))\n<\/code><\/pre>\n"},{"AnswerId":"40700037","CreationDate":"2016-11-20T02:24:03.843","ParentId":null,"OwnerUserId":"5629831","Title":null,"Body":"<p>Run this:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\nsess.run(init)\n<\/code><\/pre>\n\n<p>Or (depending on the version of TF that you have):<\/p>\n\n<pre><code>init = tf.initialize_all_variables()\nsess.run(init)\n<\/code><\/pre>\n"},{"AnswerId":"47523324","CreationDate":"2017-11-28T03:06:59.977","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>Normally there are two ways of initializing variables, 1) using the <code>sess.run(tf.global_variables_initializer())<\/code> as the previous answers noted; 2) the load the graph from checkpoint. <\/p>\n\n<p>You can do like this: <\/p>\n\n<pre><code>sess = tf.Session(config=config)\nsaver = tf.train.Saver(max_to_keep=3)\ntry:\n    saver.restore(sess, tf.train.latest_checkpoint(FLAGS.model_dir))\n    # start from the latest checkpoint, the sess will be initialized \n    # by the variables in the latest checkpoint\nexcept ValueError:\n    # train from scratch\n    init = tf.global_variables_initializer()\n    sess.run(init)\n<\/code><\/pre>\n\n<p>And the third method is to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.3\/api_docs\/python\/tf\/train\/Supervisor\" rel=\"nofollow noreferrer\">tf.train.Supervisor<\/a>. The session will be <\/p>\n\n<blockquote>\n  <p>Create a session on 'master', recovering or initializing the model as needed, or wait for a session to be ready.<\/p>\n<\/blockquote>\n\n<pre><code>sv = tf.train.Supervisor([parameters])\nsess = sv.prepare_or_wait_for_session()\n<\/code><\/pre>\n"}]}
{"QuestionId":36008183,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-15T10:08:35.427","AcceptedAnswerId":36008622.0,"OwnerUserId":2255757.0,"Title":"TensorFlow: removing\/adding Variable to computational graph","Body":"<p>I am learning TensorFlow, but can't find any conclusive documentation on this. Is it possible to dynamically remove variables from an initialized session? Or add variables similarly? Would this require recompiling\/initialising the computational graph every time?<\/p>\n","answers":[{"AnswerId":"36008622","CreationDate":"2016-03-15T10:28:52.040","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>You can add individual variables to the session using<\/p>\n\n<pre><code>init_new_vars_op = tf.initialize_variables([v_6, v_7, v_8])\nsess.run(init_new_vars_op)\n<\/code><\/pre>\n\n<p>Doesn't require recompiling the graph.<\/p>\n\n<p>Don't know about removing variables, I haven't found anything either, anyone else have any ideas?<\/p>\n"}]}
{"QuestionId":36015170,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-03-15T15:13:44.927","AcceptedAnswerId":null,"OwnerUserId":5703903.0,"Title":"How can I add labels to TensorBoard Images?","Body":"<p>TensorBoard is a great tool, but can it be more robust? The image below shows the visualization in TensorBoard.<\/p>\n\n<p>It's called by the following code:<\/p>\n\n<pre><\/pre>\n\n<p>As the API suggests, the last digit is the \"image number\", from 0 to 99 in this case since I specified max_images = 100. I would like to ask, if I can append the label of this image into the text? This would be a great functionality to have as it allows users to see in real time the images and their respective labels during training. In the event whereby some images are totally mislabelled, a fix can be implemented. In other words, I would like the corresponding text in the image below to be:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/piMpi.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/piMpi.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"36018271","CreationDate":"2016-03-15T17:34:46.197","ParentId":null,"OwnerUserId":"2125302","Title":null,"Body":"<p>I haven't been able to find a way to do this using only tensorflow, so instead I do the following:<\/p>\n\n<ol>\n<li>Create a placeholder for the summary images (e.g. like a (10, 224, 224, 3) for ten summary images).<\/li>\n<li>Create the image summary based on that placeholder.<\/li>\n<li>During validation (or training, if you like), pull the images and labels for your summary into python using something like <code>session.run([sample_images, sample_labels])<\/code>.<\/li>\n<li>Iterate through the batch and use OpenCV to write the label onto the image using <code>cv2.putText<\/code>.<\/li>\n<li>Run the summary op providing the labeled images for the placeholder.<\/li>\n<\/ol>\n"},{"AnswerId":"52858712","CreationDate":"2018-10-17T15:39:17.847","ParentId":null,"OwnerUserId":"7217513","Title":null,"Body":"<p>Here is a small improvement for approach proposed by Vince Gatto. We can use <code>tf.py_func<\/code> to avoid creating extra placeholders and doing extra <code>session.run<\/code>.<\/p>\n\n<p>First, we define these function (you will need <code>opencv-python<\/code> installed):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import cv2\nimport tensorflow as tf\n\ndef put_text(imgs, texts):\n    result = np.empty_like(imgs)\n    for i in range(imgs.shape[0]):\n        text = texts[i]\n        if isinstance(text, bytes):\n            text = text.decode()\n        # You may need to adjust text size and position and size.\n        # If your images are in [0, 255] range replace (0, 0, 1) with (0, 0, 255)\n        result[i, :, :, :] = cv2.putText(imgs[i, :, :, :], str(text), (0, 30), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 1), 2)\n    return result\n\ndef tf_put_text(imgs, texts):\n    return tf.py_func(put_text, [imgs, texts], Tout=imgs.dtype)\n<\/code><\/pre>\n\n<p>Now we can use <code>tf_put_text<\/code> to print labels on top images just before feeding them to image summary:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>annotated_images = tf_put_text(images, labels)\ntf.summary.image('annotated_images', annotated_images, 4)\n<\/code><\/pre>\n"}]}
{"QuestionId":36019125,"AnswerCount":1,"Tags":"<python><numpy><theano><theano-cuda>","CreationDate":"2016-03-15T18:19:34.410","AcceptedAnswerId":null,"OwnerUserId":5340003.0,"Title":"What is the right way to manage memory in Theano for training sets that cannot fit in RAM?","Body":"<p>TL;DR:\nHow do I give more data to a Theano function without taking more memory?<\/p>\n\n<p>The problem I'm having is that training my ML algorithm on the GPU with Theano causes the GPU to eventually run out of memory. I took a slight departure from the tutorial because my dataset is too big to read entirely into memory (this must be an issue for video algorithms too, right?), so rather than using an index input and update scheme, I just pass the Theano function the ndarrays directly.<\/p>\n\n<p>Let me give an example of what I mean. In the Logistic Regression tutorial in Theano it says to do something along the lines of:<\/p>\n\n<pre><\/pre>\n\n<p>This requires  and  to be loaded into memory, and the tutorial uses a  to store the <em>complete<\/em> dataset.<\/p>\n\n<p>Ok for me, the dataset is <em>huge<\/em> (many many gigabytes), which means it cannot all be loaded into memory at once, so I modified mine to take the data directly, thusly:<\/p>\n\n<pre><\/pre>\n\n<p>and then I do something that looks vaguely like this:<\/p>\n\n<pre><\/pre>\n\n<p>I think I may be fundamentally misunderstanding how to properly hand data to the GPU without some nasty python garbage-collection dirtiness. It seems like this is just occupying more and more memory in the model internally, because after training this after a (large) number of batches, I get an error like this:<\/p>\n\n<pre><\/pre>\n\n<p><strong>How do I give more data to a Theano function, without taking up more memory?<\/strong><\/p>\n","answers":[{"AnswerId":"40622006","CreationDate":"2016-11-16T00:15:54.957","ParentId":null,"OwnerUserId":"2444909","Title":null,"Body":"<p>If the dataset does not fit in memory, the idea is to <strong>take a portion of it and load it each time you need<\/strong>.<\/p>\n\n<p>If your data does not fit in the gpu memory, as seen in the classic <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow noreferrer\">lasagne<\/a> tutorial, you can <strong>iterate over portion of the dataset<\/strong>, called minibatches<\/p>\n\n<p>Then, if your data does not fit in your RAM, you need to load the minibatch each time you need it. Best way to do that is to <strong>make a separate process load the next minibatch<\/strong> (cpu working) as you are analysing the current one (gpu working)<\/p>\n\n<p>You can inspire yourself from <a href=\"https:\/\/github.com\/uoguelph-mlrg\/theano_alexnet\" rel=\"nofollow noreferrer\">AlexNet<\/a> : <\/p>\n"}]}
{"QuestionId":36020162,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network><tensorboard>","CreationDate":"2016-03-15T19:15:32.993","AcceptedAnswerId":null,"OwnerUserId":5551505.0,"Title":"How to interpret Tensorboard Histograms","Body":"<p>I am new to Tensorflow. I was wondering if someone could guide me as to how to interpret the Tensorflow histograms. The following are the histograms for the network prediction and filters at the first layer. W_conv1 is of the shape[5 5 3 10]. How can I view each filter individually? <a href=\"https:\/\/i.stack.imgur.com\/q7BqC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/q7BqC.png\" alt=\"This is the histograms for graph&#39;s prediction\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aXlqA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aXlqA.png\" alt=\"This is the histogram for one of the weights of one of the constitutional layers\"><\/a><\/p>\n","answers":[]}
{"QuestionId":36020471,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-03-15T19:32:30.160","AcceptedAnswerId":36095211.0,"OwnerUserId":1191087.0,"Title":"TensorBoard - Viewing image summaries side-by-side?","Body":"<p>The network I'm working on in TensorFlow produces a pixel-wise labeling result. TensorBoard is very helpful for viewing the data during training, but if I simply add 3 s I get an layout of the images something like:<\/p>\n\n<pre><\/pre>\n\n<p>(where  is the original image,  is the ground truth labels, and  is the predicted labels). What I would prefer to see is a layout like:<\/p>\n\n<pre><\/pre>\n\n<p>As this would make comparing the results much easier. Does TensorBoard provide any convenient way of achieving this? Or would taking all the images, mapping non-RGB ones RGB, and simply concatenating them to make one long image be the best solution? Thank you much!<\/p>\n","answers":[{"AnswerId":"36095211","CreationDate":"2016-03-18T22:22:32.567","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>TensorBoard doesn't have solution for this at the moment; concatenating the images, like you proposed, would be a reasonable workaround.<\/p>\n\n<p>I suggest filing a feature request on our Github :) the image dashboard is in need of some love, lots of people are using it and want improved layout and behavior.<\/p>\n"}]}
{"QuestionId":36020518,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-03-15T19:34:51.080","AcceptedAnswerId":36021397.0,"OwnerUserId":5407700.0,"Title":"Meta table error","Body":"<p>I am trying to set up a toy example for threading in torch but I am getting an error from running the code below. I think it might just be the way I set up the table but I am not sure. <\/p>\n\n<pre><\/pre>\n\n<p>The error occurs in  in the .new function. However calling the same  later on in the  function works. Am I not allowed to call the  functions before the metatable is set? Here is the error, which I think means  is not returning anything;<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36021397","CreationDate":"2016-03-15T20:23:24.417","ParentId":null,"OwnerUserId":"258523","Title":null,"Body":"<p>When the code in your <code>new<\/code> function runs the metatable hasn't been set up yet so when you use <code>self._getFromThreads<\/code> and <code>self._pushResult<\/code> there's nothing there and you get <code>nil<\/code> back and that's not valid.<\/p>\n\n<p>It works <em>after<\/em> the <code>new<\/code> function returns because at that point the metatable has been added so the lookups use the <code>__index<\/code> metamethod to look the entries up.<\/p>\n\n<p>You can do<\/p>\n\n<pre><code>local self = setmetatable({}, DataThreads)\n<\/code><\/pre>\n\n<p>to set the metatable up immediately and then just <\/p>\n\n<pre><code>return self\n<\/code><\/pre>\n\n<p>at the end.<\/p>\n"}]}
{"QuestionId":36021615,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-03-15T20:37:57.287","AcceptedAnswerId":null,"OwnerUserId":4697714.0,"Title":"Why the output of Theano shared variable is an array?","Body":"<p>I am using my own dataset to train a SdA using Theano with this code:<\/p>\n\n<pre><\/pre>\n\n<p>Then I am printing train_set_x and train_set_y with this code:<\/p>\n\n<pre><\/pre>\n\n<p>And these are my result:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see the output is a numpy array. However when I am printing the  MNIST dataset in SdK.py provided by theano tutorial, by these codes:<\/p>\n\n<pre><\/pre>\n\n<p>I see these results:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see this is not a numpy array. Do you have any idea how can I fix my codes and dataset in a way that my ouyput looks like the Theano tutorial output?<\/p>\n","answers":[{"AnswerId":"36027072","CreationDate":"2016-03-16T04:54:04.887","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>It looks like the Theano tutorial is written in Python3 and you're using Python 2.x.<\/p>\n\n<p>To get the same output format for Python 2,x, you can just remove the parenthesis after <code>print<\/code>.<\/p>\n\n<pre><code>a = numpy.asarray([1,2,3])\nprint (\"x=\", a)  # this will output ('x=', array([1, 2, 3]))\nprint \"x=\", a  # this will output x= [1 2 3]\n<\/code><\/pre>\n"}]}
{"QuestionId":36022239,"AnswerCount":1,"Tags":"<lua><machine-learning><neural-network><deep-learning><torch>","CreationDate":"2016-03-15T21:11:46.257","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"In Torch \/ Lua, what is the differece between loading a saved model and using Xavier weight initialization method?","Body":"<p>I've been working with Torch and Lua on an artificial neural network project.<\/p>\n\n<p>I sometimes used the commands  and  to save a trained model to a file and to load it into my scripts.<\/p>\n\n<p>Recently, I've been studying the so called Xavier weight initialization method <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf\" rel=\"nofollow\">originally developed by Xavier Glorot<\/a> in 2010.<\/p>\n\n<p>I found an implementation of this weight initialization method in the <a href=\"https:\/\/github.com\/e-lab\/torch-toolbox\/tree\/master\/Weight-init\" rel=\"nofollow\">torch-toolbox package<\/a>. The example contains this code:<\/p>\n\n<pre><\/pre>\n\n<p>I ran this example in the Torch console and tried to analyze the content of the objects, such as model and model_new. I cannot see immediate difference, so I'm wondering: <strong>what is the difference between using this Xavier  model weight initialization and the classical torch.load() command?<\/strong><\/p>\n","answers":[{"AnswerId":"36039929","CreationDate":"2016-03-16T15:21:10.907","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>Citing the mentioned docs:<\/p>\n\n<pre><code>The weights and biases in the network model will be reset based on the method provided.\n<\/code><\/pre>\n\n<p>So, <code>weight-init<\/code> <strong>resets<\/strong> all the weights in the network to some value (in case of Xavier method to <code>math.sqrt(2\/(fan_in + fan_out)<\/code> (see <a href=\"https:\/\/github.com\/e-lab\/torch-toolbox\/blob\/master\/Weight-init\/weight-init.lua#L19\" rel=\"nofollow\">here<\/a>)).<\/p>\n\n<p>On the other hand <code>torch.load()<\/code> reads the weights saved from a pre-trained file (that was saved with <code>torch.save()<\/code>). So, unless your weights didn't change at all during the training, if you do:<\/p>\n\n<p><code>weight-init() -&gt; train -&gt; torch.save() -&gt; torch.load()\n<\/code>\nyou should see other results than with just weight-init().<\/p>\n"}]}
{"QuestionId":36023667,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2016-03-15T22:44:22.287","AcceptedAnswerId":null,"OwnerUserId":4697714.0,"Title":"How can we map training data to 0 to 1 for Theano?","Body":"<p>As you might know for training task in Theano we have to map our training data to 0 to 1. My training data has negative values also. Currently I am using this formula:<\/p>\n\n<p>x'=(x - min(x))\/(max(x) - min(x))<\/p>\n\n<p>which is implemented by this code:<\/p>\n\n<pre><\/pre>\n\n<p>Is this formula correct? Do you have a better idea regarding feature rescaling?<\/p>\n","answers":[{"AnswerId":"36060122","CreationDate":"2016-03-17T12:07:41.463","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>so this is my way to do normalization for variable <code>t<\/code> use your formula:<\/p>\n\n<pre><code>import numpy as np\nt = np.array([[-1,-2,3,2],[2,1,3,-1]],dtype='float32')\nb = (t-t.min(1).reshape(2,1))\/t.ptp(1).reshape(2,1)\nprint b\n<\/code><\/pre>\n\n<p>it gives the correct output:<\/p>\n\n<pre><code>[[0.2, 0,   1, 0.8]\n[0.75, 0.5, 1,   0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":36026865,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-16T04:35:04.610","AcceptedAnswerId":null,"OwnerUserId":6069611.0,"Title":"Python tensorflow error, sys has no attribute getdlflags","Body":"<p>I Just attempted to install tensorflow for python and when I went to the console to see if the <strong>init<\/strong>.py was working it returned this error. I installed it manualy without pip or any other package manager.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36040880","CreationDate":"2016-03-16T16:01:07.263","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow is not supported on Windows yet. Please follow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow\">this<\/a> github issue which tracks TensorFlow Windows support.<\/p>\n"}]}
{"QuestionId":36026892,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-16T04:37:37.140","AcceptedAnswerId":36026969.0,"OwnerUserId":5703903.0,"Title":"How can I convert TFRecords into numpy arrays?","Body":"<p>The main idea is to convert TFRecords into numpy arrays. Assume that the TFRecord stores images. Specifically:<\/p>\n\n<ol>\n<li>Read a TFRecord File and convert each image into a numpy array.<\/li>\n<li>Write the image into 1.jpg, 2.jpg, etc.<\/li>\n<li>At the same time, write the file name and label to the text file like this:\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>I currently use the following code:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I'm just reading trying to get at least one image for starters. The code just gets stuck when I run this.<\/p>\n","answers":[{"AnswerId":"36026969","CreationDate":"2016-03-16T04:44:42.083","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>Oops, it was a silly mistake on my part. I used a string_input_producer but forgot to run the queue_runners.<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>with tf.Session() as sess:\n  filename_queue = tf.train.string_input_producer([\"..\/data\/svhn\/svhn_train.tfrecords\"])\n  image, label, height, width, depth = read_and_decode(filename_queue)\n  image = tf.reshape(image, tf.pack([height, width, 3]))\n  image.set_shape([32,32,3])\n  init_op = tf.initialize_all_variables()\n  sess.run(init_op)\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(coord=coord)\n  for i in range(1000):\n    example, l = sess.run([image, label])\n    print (example,l)\n  coord.request_stop()\n  coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":36030331,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-03-16T08:31:14.840","AcceptedAnswerId":36030638.0,"OwnerUserId":5545794.0,"Title":"Error in prototxt of caffe, caffe.SolverParameter has no field named \"name\"","Body":"<p>I wrote a caffe net about multi-label classification using HDF5, here is the prototxt file named 'auto_train.prototxt'   <\/p>\n\n<pre><\/pre>\n\n<p>and this is the train.sh<\/p>\n\n<pre><\/pre>\n\n<p>But when I run this script, it got something wrong<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I don't know what happened, look for help<\/p>\n","answers":[{"AnswerId":"36030638","CreationDate":"2016-03-16T08:48:30.073","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are confusion the net structure definition prototxt (a.k.a <code>train_val.prototxt<\/code>) with the solver definition prototxt (a.k.a <code>solver.prototxt<\/code>).<\/p>\n\n<p>See, e.g., <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet\" rel=\"noreferrer\">AlexNet example<\/a> for these two <em>different<\/em> prototxt files.<\/p>\n\n<p>The net structure definition, <code>train_val.prototxt<\/code> defines the net structure and looks like the <code>auto_train.prototxt<\/code> file you wrote.<\/p>\n\n<p>However, you are missing the <a href=\"https:\/\/stackoverflow.com\/q\/33780779\/1714410\">solver definition prototxt<\/a>, <code>solver.prototxt<\/code> that defines meta parameters for the optimization process.<br>\nIn your case a <code>solver.prototxt<\/code> would look something like:<\/p>\n\n<pre><code>net: \"examples\/corel5k\/auto_train.prototxt\" # here is where you put the net structure file\ntest_iter: 1000  \ntest_interval: 1000\nbase_lr: 0.01\nlr_policy: \"step\"\ngamma: 0.1\nstepsize: 100000\ndisplay: 20\nmax_iter: 450000\nmomentum: 0.9\nweight_decay: 0.0005\nsnapshot: 10000\nsnapshot_prefix: \"examples\/corel5k\/my_auto_snapshots\"\nsolver_mode: GPU\n<\/code><\/pre>\n\n<p>You can find information on how to set the meta parameters in the <code>solver.protoxt<\/code> <a href=\"https:\/\/stackoverflow.com\/q\/33780779\/1714410\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/32177764\/what-is-weight-decay-meta-parameter-in-caffe\">here<\/a>.<\/p>\n\n<p>Once you have a proper <code>solver.prototxt<\/code> you can run caffe:<\/p>\n\n<pre><code>.\/build\/tools\/caffe train \\\n  -solver examples\/corel5k\/solver.prototxt \\\n  -weights examples\/corel5k\/bvlc_reference_caffenet.caffemodel\n<\/code><\/pre>\n"}]}
{"QuestionId":36031324,"AnswerCount":1,"Tags":"<python><machine-learning><linear-regression><tensorflow>","CreationDate":"2016-03-16T09:20:29.163","AcceptedAnswerId":36045817.0,"OwnerUserId":3940080.0,"Title":"How to implement multivariate linear stochastic gradient descent algorithm in tensorflow?","Body":"<p>I started with simple implementation of single variable linear gradient descent but don't know to extend it to multivariate stochastic gradient descent algorithm ?<\/p>\n\n<p>Single variable linear regression <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36045817","CreationDate":"2016-03-16T20:01:25.970","ParentId":null,"OwnerUserId":"2642845","Title":null,"Body":"<p>You have two part in your question:<\/p>\n\n<ul>\n<li>How to change this problem to a higher dimension space.<\/li>\n<li>How to change from the batch gradient descent to a stochastic gradient descent.<\/li>\n<\/ul>\n\n<p>To get a higher dimensional setting, you can define your linear problem <code>y = &lt;x, w&gt;<\/code>. Then, you just need to change the dimension of your Variable <code>W<\/code> to match the one of <code>w<\/code> and replace the multiplication <code>W*x_data<\/code> by a scalar product <code>tf.matmul(x_data, W)<\/code> and your code should run just fine.<\/p>\n\n<p>To change the learning method to a stochastic gradient descent, you need to abstract the input of your cost function by using <code>tf.placeholder<\/code>.<br>\nOnce you have defined <code>X<\/code> and <code>y_<\/code> to hold your input at each step, you can construct the same cost function. Then, you need to call your step by feeding the proper mini-batch of your data.<\/p>\n\n<p>Here is an example of how you could implement such behavior and it should show that <code>W<\/code> quickly converges to <code>w<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Define dimensions\nd = 10     # Size of the parameter space\nN = 1000   # Number of data sample\n\n# create random data\nw = .5*np.ones(d)\nx_data = np.random.random((N, d)).astype(np.float32)\ny_data = x_data.dot(w).reshape((-1, 1))\n\n# Define placeholders to feed mini_batches\nX = tf.placeholder(tf.float32, shape=[None, d], name='X')\ny_ = tf.placeholder(tf.float32, shape=[None, 1], name='y')\n\n# Find values for W that compute y_data = &lt;x, W&gt;\nW = tf.Variable(tf.random_uniform([d, 1], -1.0, 1.0))\ny = tf.matmul(X, W, name='y_pred')\n\n# Minimize the mean squared errors.\nloss = tf.reduce_mean(tf.square(y_ - y))\noptimizer = tf.train.GradientDescentOptimizer(0.01)\ntrain = optimizer.minimize(loss)\n\n# Before starting, initialize the variables\ninit = tf.initialize_all_variables()\n\n# Launch the graph.\nsess = tf.Session()\nsess.run(init)\n\n# Fit the line.\nmini_batch_size = 100\nn_batch = N \/\/ mini_batch_size + (N % mini_batch_size != 0)\nfor step in range(2001):\n    i_batch = (step % n_batch)*mini_batch_size\n    batch = x_data[i_batch:i_batch+mini_batch_size], y_data[i_batch:i_batch+mini_batch_size]\n    sess.run(train, feed_dict={X: batch[0], y_: batch[1]})\n    if step % 200 == 0:\n        print(step, sess.run(W))\n<\/code><\/pre>\n\n<p>Two side notes: <\/p>\n\n<ul>\n<li><p>The implementation below is called a mini-batch gradient descent as at each step, the gradient is computed using a subset of our data of size <code>mini_batch_size<\/code>. This is a variant from the stochastic gradient descent that is usually used to stabilize the estimation of the gradient at each step. The stochastic gradient descent can be obtained by setting <code>mini_batch_size = 1<\/code>.<\/p><\/li>\n<li><p>The dataset can be shuffle at every epoch to get an implementation closer to the theoretical consideration. Some recent work also consider only using one pass through your dataset as it prevent over-fitting. For a more mathematical and detailed explanation, you can see <a href=\"http:\/\/research.microsoft.com\/pubs\/192769\/tricks-2012.pdf\" rel=\"nofollow noreferrer\">Bottou12<\/a>. This can be easily change according to your problem setup and the statistic property your are looking for.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":36032661,"AnswerCount":2,"Tags":"<torch>","CreationDate":"2016-03-16T10:15:22.013","AcceptedAnswerId":36058125.0,"OwnerUserId":133374.0,"Title":"3D tensor * 2D tensor dot in Torch","Body":"<p>In Theano, when I have a 3D tensor  with shape [A,B,C] and a 2D tensor  with shape [C,D], then  returns a 3D tensor with shape [A,B,D].<\/p>\n\n<p>What would be the equivalent operation in Torch?  doesn't seem to do that, and  and  complain that they want a 2D tensor for both arguments, and  wants 3D tensors for both arguments.<\/p>\n","answers":[{"AnswerId":"36038672","CreationDate":"2016-03-16T14:31:54.793","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>you need to expand y to be of size [A,C,D] and then use torch.bmm. Look at the documentation of torch.expand or torch.repeatTensor.<\/p>\n"},{"AnswerId":"36058125","CreationDate":"2016-03-17T10:41:06.390","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>As @smhx proposed, the possible solution is to repeat the second tensor (there's a way to do it without memory allocating) and then perform a batch matrix matrix product:<\/p>\n\n<pre><code>function repeatNoCopy(tensor, k)\n    local tens_size = tensor:size():totable()\n    return torch.expand(tensor:view(1, unpack(tens_size)), k, unpack(tens_size))\nend\n\nA = torch.rand(3, 2, 5)\nB = torch.rand(5, 4)\nB_rep = repeatNoCopy(B, 3)\n\nresult = torch.bmm(A, B_rep)\n\nprint(result)\n&gt; [torch.DoubleTensor of size 3x2x4]\n<\/code><\/pre>\n"}]}
{"QuestionId":36033850,"AnswerCount":0,"Tags":"<c++><deep-learning><caffe><backpropagation><conv-neural-network>","CreationDate":"2016-03-16T11:05:42.490","AcceptedAnswerId":null,"OwnerUserId":3550694.0,"Title":"Adding Mask Entry in Protobuf, and multiplying the same with gradient","Body":"<p>I have been working on pruning the neural network by discarding update to weight matrix by shielding weights below certain threshold.<br>\nTo achieve this I have to add a field \"mask\" in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\" rel=\"nofollow\">protobuf file<\/a>, and later multiply (dot product) the same mask with diff (gradient) during back propagation by modifying the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/blob.cpp\" rel=\"nofollow\"><\/a>.<\/p>\n\n<p>Any suggestions on how to accomplish this will be appreciated.<\/p>\n","answers":[]}
{"QuestionId":36035092,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-16T11:59:38.620","AcceptedAnswerId":39357384.0,"OwnerUserId":1611416.0,"Title":"Multi-dimensional gather in Tensorflow","Body":"<p>The general solution to this question is being worked on in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow\">this<\/a> github issue, but I was wondering if there are workarounds using  (or something else) to achieve array indexing using a multi-index. One solution I came up with was to broadcast multiply each index in the multi-idx with the cumulative product of the tensor shape, which produces indices suitable for indexing the flattened tensor:<\/p>\n\n<pre><\/pre>\n\n<p>However, this only works on tensors of rank 3 due to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1519\" rel=\"nofollow\">this<\/a> (current) limitation limiting broadcasts to tensors of rank 3.<\/p>\n\n<p>At the moment I can only think of doing a chained gather, transpose, gather, transpose, gather, but this is unlikely to be efficient. e.g.<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"39357384","CreationDate":"2016-09-06T20:31:57.123","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>It sounds like you want <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/array_ops.html#gather_nd\" rel=\"nofollow\">gather_nd<\/a>.<\/p>\n"}]}
{"QuestionId":36036163,"AnswerCount":1,"Tags":"<ffi><torch><luajit>","CreationDate":"2016-03-16T12:48:29.853","AcceptedAnswerId":36038698.0,"OwnerUserId":2086557.0,"Title":"How to use Lua FFI with C++ functions","Body":"<p>I have followed <a href=\"https:\/\/github.com\/Atcold\/Torch-Developer-Guide\/blob\/master\/FFI-so\/README.md\" rel=\"nofollow\">this<\/a> tutorial to create an FFI interface to some C code to torch<\/p>\n\n<p>However, it doesn't work for C++ functions. Is it even possible, if so how?<\/p>\n\n<p>I created a shared library from the C++ file (simple.cpp) like so:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36038698","CreationDate":"2016-03-16T14:33:09.030","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>The C++ functions need to have C linkage to be used via FFI. Wrap the function signatures with extern \"C\" calls. More details here: <a href=\"https:\/\/stackoverflow.com\/questions\/1041866\/in-c-source-what-is-the-effect-of-extern-c\">In C++ source, what is the effect of extern &quot;C&quot;?<\/a><\/p>\n"}]}
{"QuestionId":36036302,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow>","CreationDate":"2016-03-16T12:54:15.150","AcceptedAnswerId":36040301.0,"OwnerUserId":1534992.0,"Title":"Sharing variables between threads in Tensorflow","Body":"<p>I am trying to implement asynchronous gradient descent with TensorFlow using Python threads. In the main code, I define the graph, including a training operation, which gets a variable to keep count of the :<\/p>\n\n<pre><\/pre>\n\n<p>If I print the name of  I get:<\/p>\n\n<blockquote>\n  <p>scope_global_step\/global_step:0<\/p>\n<\/blockquote>\n\n<p>The main code also launches several threads to execute a  method:<\/p>\n\n<pre><\/pre>\n\n<p>I would like each thread to stop execution if the value of  is greater than or equal to . For that, I build the  method as it follows:<\/p>\n\n<pre><\/pre>\n\n<p>This fails with the message:<\/p>\n\n<blockquote>\n  <p>ValueError: Under-sharing: Variable scope_global_step\/global_step does not exist, disallowed. Did you mean to set reuse=None in VarScope?<\/p>\n<\/blockquote>\n\n<p>I can see that  is added to the name of the variable when it is first created, and when I try to retrieve it, that suffix is not used. Why is this? \nIf I manually add the suffix to the name of the variable when I try to retrieve it, it still claims the variable does not exist. Why does TensorFlow not find the variable? Shouldn't the variables automatically be shared among threads? I mean, all threads are running inside the same session, right? <\/p>\n\n<p>And another question related to my  method: will  execute the graph again, or will it just fetch the value assigned to  after the execution of the  and  operations? In general, what is the recommended way of fetching values from variables to be used inside Python code?<\/p>\n","answers":[{"AnswerId":"36040301","CreationDate":"2016-03-16T15:36:10.753","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> Pass the <code>global_step<\/code> <code>tf.Variable<\/code> object that you created in the first code fragment as one of the training thread arguments, and call <code>sess.run(global_step)<\/code> on the passed-in variable.<\/p>\n\n<p>As a general rule, your training loop (and especially a training loop in a separate thread) should not modify the graph. The <code>tf.variable_scope()<\/code> context manager and <code>tf.get_variable()<\/code> <em>can<\/em> modify the graph (even though they don't always), so you shouldn't use them in your training loop. The safest thing to do is to pass the <code>global_step<\/code> object (that you created in the first place) as one of the <code>args<\/code> tuple when you create the training threads. Then you can simply rewrite your training function as:<\/p>\n\n<pre><code>def training(sess, train_op, loss, global_step):\n    while (True):\n         _, loss_value = sess.run([train_op, loss])\n         current_step = sess.run(global_step)\n         if current_step &gt;= FLAGS.max_steps: break\n<\/code><\/pre>\n\n<p>To answer your other question, running <code>global_step.eval(session=sess)<\/code> or <code>sess.run(global_step)<\/code> only fetches the current value of the <code>global_step<\/code> variable, and doesn't re-execute the rest of your graph. This is the recommended way to fetch <code>tf.Variable<\/code> values for use in Python code.<\/p>\n"}]}
{"QuestionId":36036957,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><keras>","CreationDate":"2016-03-16T13:22:22.327","AcceptedAnswerId":null,"OwnerUserId":3038478.0,"Title":"Keras extremely high loss, not decreasing with each epoch","Body":"<p>I'm using Keras to build and train a recurrent neural network.<\/p>\n\n<pre><\/pre>\n\n<p> is an an array of sequences of [latitude, longitude, temperature], all padded to the same length with values of [0,0,0]. When I train this network, the first epoch gives me a loss of about 63, and <em>increases<\/em> after more epochs.\nThis is causing a  call later in the code to give values that are completely off of the training values. For example, most of the training values in each sequence is around , but the RNN outputs values consistently around , which causes me to think something is wrong with the masking layer. <\/p>\n\n<p>The training X () data looks something like this (only much larger): <\/p>\n\n<pre><\/pre>\n\n<p>and the training Y () data looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38406047","CreationDate":"2016-07-15T22:54:54.163","ParentId":null,"OwnerUserId":"4727781","Title":null,"Body":"<p>I am somewhat certain that you're misusing the Masking layer from Keras. Check the documentation <a href=\"http:\/\/keras.io\/layers\/core\/#masking\" rel=\"nofollow\">here<\/a> for more details. <\/p>\n\n<p>Try using a masking layer like: <\/p>\n\n<p><code>model.add(Masking(0, input_shape=(max_sequence_length, in_dimension)))<\/code><\/p>\n\n<p>because I believe it just needs the masking value in the timestep dimension, not the entire time-dimension and value (ie [0,0,0]). <\/p>\n\n<p>Best of luck.<\/p>\n"}]}
{"QuestionId":36039071,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-03-16T14:48:15.877","AcceptedAnswerId":36039480.0,"OwnerUserId":133374.0,"Title":"lua_pcall error message is missing","Body":"<p>I'm using  to call some function and I want to catch the error.\nIn some cases, the error seem to get lost. How can that happen?\nThis is for both the case where I use an error handler and when I don't use not. In both cases, the top of the stack is not a string.<\/p>\n\n<p>The C code:<\/p>\n\n<pre><\/pre>\n\n<p>The Lua function which is called looks like this (for testing):<\/p>\n\n<pre><\/pre>\n\n<p>It somehow gets an error when it calls .\nBut I don't exactly know why. And I don't get any meaningful error.\nThis is what my question is about.<\/p>\n\n<p>The output:<\/p>\n\n<pre><\/pre>\n\n<p>Or maybe my code is correct and it really should return the error string here? So maybe there is some memory corruption when calling , i.e. something gets messed up?<\/p>\n","answers":[{"AnswerId":"36039480","CreationDate":"2016-03-16T15:04:03.077","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>It seems I need to call <code>torch.updateerrorhandlers()<\/code>. Then I get some meaningful output:<\/p>\n\n<pre><code>hi from Lua func\nLua error code 2: inconsistent tensor size at \/tmp\/luarocks_torch-scm-1-1092\/torch7\/lib\/TH\/generic\/THTensorMath.c:384\nstack traceback:\n        [C]: at 0x7f63cd831360\n        [C]: in function 'dot'\n        [string \"return ...\"]:9: in function &lt;[string \"return ...\"]:2&gt;\n<\/code><\/pre>\n\n<p>But only if I have <code>torch.updateerrorhandlers()<\/code> inside the Lua function.<\/p>\n\n<p>I tried with this C code and that doesn't work:<\/p>\n\n<pre><code>    lua_getglobal(L, \"torch\");\n    lua_getfield(L, -1, \"updateerrorhandlers\");\n    lua_replace(L, -2);\n    assert(lua_pcall(L, 0, 0, 0) == 0);\n<\/code><\/pre>\n\n<p>I figured out that if I do another <code>torch.updateerrorhandlers()<\/code> call right before my actual <code>my_func_index<\/code> <code>lua_pcall<\/code>, it works.\nThis is unexpected but maybe this is because this might be another thread\n(which I would not have expected).\nActually, I found in the Torch code the function <code>torch.updatethreadlocals()<\/code> which is exactly for this purpose and I'm calling this one now, right before my other <code>lua_pcall<\/code>:<\/p>\n\n<pre><code>    lua_getglobal(L, \"torch\");\n    lua_getfield(L, -1, \"updatethreadlocals\");\n    lua_replace(L, -2);\n    assert(lua_pcall(L, 0, 0, 0) == 0);\n<\/code><\/pre>\n\n<p>This works now.<\/p>\n"}]}
{"QuestionId":36039263,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2016-03-16T14:56:27.007","AcceptedAnswerId":36046061.0,"OwnerUserId":3006737.0,"Title":"Using Keras\/ Neural Networks for categorical data","Body":"<p>In the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"nofollow noreferrer\">CNN example for the minst dataset for Keras<\/a> they tell you how to make a good CNN network to recognise hand written digits. The issue is that it doesn't tell you how to predict new digits.<\/p>\n\n<p>For example give an image, if I do this:<\/p>\n\n<pre><\/pre>\n\n<p>instead of telling me what digits it think it is, it instead gives me a list of 10 numbers (presumably probabilities)<\/p>\n","answers":[{"AnswerId":"36046061","CreationDate":"2016-03-16T20:15:46.077","ParentId":null,"OwnerUserId":"2602240","Title":null,"Body":"<p>You can use numpy's argmax to find out class that has maximum probability<\/p>\n\n<pre><code>import numpy as np\nprobabilities = model.predict(image)\nclasses = np.argmax(probabilities, axis=-1)\n<\/code><\/pre>\n"}]}
{"QuestionId":36039429,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-16T15:02:33.037","AcceptedAnswerId":null,"OwnerUserId":6072264.0,"Title":"Tensorflow extract_glimpse offset","Body":"<p>I am trying to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/image.html#extract_glimpse\" rel=\"nofollow\"><em>extract_glimpse<\/em><\/a> function of tensorflow but I encounter some difficulties with the <em>offset<\/em> parameter.<\/p>\n\n<p>Let's assume that I have a batch of one single channel 5x5 matrix called M and that I want to extract a 3x3 matrix of it.<\/p>\n\n<p>When I call <em>extract_glimpse([M], [3,3], [[1,1]], centered=False, normalized=False)<\/em>, it returns the result I am expecting: the 3x3 matrix centered at the position (1,1) in M.<\/p>\n\n<p>But when I call <em>extract_glimpse([M], [3,3], [[2,1]], centered=False, normalized=False)<\/em>, it doesn't return the 3x3 matrix centered at the position (2,1) in M but it returns the same as in the first call.<\/p>\n\n<p>What is the point that I don't get?<\/p>\n","answers":[{"AnswerId":"40785802","CreationDate":"2016-11-24T11:56:45.193","ParentId":null,"OwnerUserId":"4807429","Title":null,"Body":"<p>The pixel coordinates actually have a range of 2 times the size (not documented - so it is a bug indeed). This is at least true in the case of <code>centered=True<\/code> and <code>normalized=False<\/code>. With those settings, the offsets range from minus the size to plus the size of the tensor. I therefore wrote a wrapper that is more intuitive to numpy users, using pixel coordinates starting at (0,0). This wrapper and more details about the problem are available on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2134#issuecomment-262525617\" rel=\"nofollow noreferrer\">the tensorflow GitHub page<\/a>.<\/p>\n\n<p>For your particular case, I would try something like:<\/p>\n\n<pre><code>offsets1 = [-5 + 3,\n            -5 + 3]\nextract_glimpse([M], [3,3], [offsets1], centered=True, normalized=False)\noffsets2 = [-5 + 3 + 2,\n            -5 + 3]\nextract_glimpse([M], [3,3], [offsets2], centered=True, normalized=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":36041112,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><text-classification>","CreationDate":"2016-03-16T16:11:09.187","AcceptedAnswerId":null,"OwnerUserId":4550032.0,"Title":"Sparse matrices in python tensorflow text classification","Body":"<p>I've been trying to implement a text classification routine using the tensorflow package in python. I already had a successful perceptron version working in the scikit-learn enviroment but scikit-learn does not have multilayer neural networks (except for some mythical version 0.18 that I can't seem to find\/install anywhere).<\/p>\n\n<p>I thought it was best to try something simpler in tensorflow first, to learn how the package works and what it can and cannot do, so I went with nearest neighbors. So far so good, except I just can't find a way to feed a sparse version of the vocabulary matrix (bag-of-words vectorizations of the texts) to a placeholder in tensorflow (in scikit-learn this is no problem at all). Converting the vocabulary matrix to a dense matrix solves the problem but severely slows down the algorithm and clogs up RAM.<\/p>\n\n<p>Is there any way around this? From what I found on the web it seems tensorflow has very limited support for sparse objects (only certain operations will accept them as input), but I hope I'm wrong. <\/p>\n\n<p>P.S. Yes, I read <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/342\" rel=\"nofollow\">this<\/a> thread and it did not solve my problem. And yes, I know I could stick to the perceptron of the scikit-learn or choose another package, but that's a decision I'll make based on the answers I get here.<\/p>\n","answers":[{"AnswerId":"42807169","CreationDate":"2017-03-15T10:32:59.793","ParentId":null,"OwnerUserId":"1677276","Title":null,"Body":"<p>With TensorFlow 1.0.1, I can do this:<\/p>\n\n<pre><code>a = sparse.csr_matrix([[0, 1, 2], [5, 0, 0], [0, 0, 5],\n                       [10, 1, 0], [0, 0, 4]])\n# w = np.arange(6, dtype=np.float32).reshape([3, 2])\na = a.tocoo()\n\na_ = tf.sparse_placeholder('float32')\nw_ = tf.Variable(tf.random_normal([3, 2], stddev=1.0))\n\na_sum = tf.sparse_reduce_sum(a_, 1)\na_mul = tf.sparse_tensor_dense_matmul(a_, w_)\n\n# Initializing the variables\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    indices = np.array(zip(a.row, a.col), dtype=np.int32)\n    values = np.array(a.data, dtype=np.float32)\n    shape = np.array(a.shape, dtype=np.int32)\n\n    print sess.run(a_mul, feed_dict={a_: tf.SparseTensorValue(indices, values, shape)})\n    w = sess.run(w_)\n    print np.dot(a.todense(), w)\n<\/code><\/pre>\n\n<p>You can find the code from API page: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_placeholder\" rel=\"nofollow noreferrer\" title=\"sparse_placeholder\">sparse placeholder<\/a>. After the first layer, other layers of neural network will be dense matrix. <\/p>\n"}]}
{"QuestionId":36041171,"AnswerCount":1,"Tags":"<concat><tensorflow>","CreationDate":"2016-03-16T16:13:45.287","AcceptedAnswerId":36042389.0,"OwnerUserId":1044011.0,"Title":"TensorFlow concat a variable-sized placeholder with a vector","Body":"<p>Let's say I have a placeholder<\/p>\n\n<p><\/p>\n\n<p>and a vector<\/p>\n\n<p><\/p>\n\n<p>In this example  is filled with zeros for simplicity but in the actual case it will be changed by other variables and will have different values.<\/p>\n\n<p>I want to efficiently do a  on  and  on dimension  and get a new tensor with shape . Unfortunately  needs all input tensors to have the same shape except the , which my example does not meet.<\/p>\n\n<p>I was considering expand  to the same shape as the data that feed to  but am not exactly sure how to do that with the placeholder itself. If I get the shape from the input data directly then I guess it is not necessary to use the placeholder.<\/p>\n","answers":[{"AnswerId":"36042389","CreationDate":"2016-03-16T17:07:24.480","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The most general solution is to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#shape\" rel=\"noreferrer\"><code>tf.shape()<\/code><\/a> op to get the run-time size of the placeholder, and the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#tile\" rel=\"noreferrer\"><code>tf.tile()<\/code><\/a> op to expand <code>h<\/code> to the appropriate size:<\/p>\n\n<pre><code>ph_input = tf.placeholder(dtype=tf.int32, shape=[None, 1])\nh = tf.zeros([1, 2], dtype=tf.int32)  # ...or some other tensor of shape [1, 2]\n\n# Get the number of rows in the fed value at run-time.\nph_num_rows = tf.shape(ph_input)[0]\n\n# Makes a `ph_num_rows x 2` matrix, by tiling `h` along the row dimension.\nh_tiled = tf.tile(h, tf.pack([ph_num_rows, 1]))\n\nresult = tf.concat(1, [ph_input, h_tiled])\n<\/code><\/pre>\n"}]}
{"QuestionId":36043717,"AnswerCount":3,"Tags":"<python><numpy><theano>","CreationDate":"2016-03-16T18:11:41.697","AcceptedAnswerId":36423504.0,"OwnerUserId":2014591.0,"Title":"Theano: Operate on nonzero elements of sparse matrix","Body":"<p>I'm trying to take the  of nonzero elements in a sparse theano variable. I have the current code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to do something that's equivalent to the following numpy syntax:<\/p>\n\n<pre><\/pre>\n\n<p>but Theano doesn't support  masks yet. (And  doesn't seem to work either.)<\/p>\n\n<p>How can I achieve this?<\/p>\n","answers":[{"AnswerId":"36430283","CreationDate":"2016-04-05T15:11:07.823","ParentId":null,"OwnerUserId":"2014591","Title":null,"Body":"<p><code>T.where<\/code> works.<\/p>\n\n<p><code>A_sparse = T.where(A_sparse == 0, 0, T.exp(A_sparse))<\/code><\/p>\n\n<p>@Seppo Envari's answer seems faster though. So I'll accept his answer.<\/p>\n"},{"AnswerId":"36046399","CreationDate":"2016-03-16T20:34:24.040","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>Here's a way of doing this with the <code>scipy.sparse<\/code> module.  I don't know how <code>theano<\/code> implements its sparse.  It's likely to be based on similar ideas (since it uses name like <code>csc<\/code>)<\/p>\n\n<pre><code>In [224]: A=sparse.csc_matrix([[1.,0,0,2,0],[0,0,3,0,0],[0,1,1,2,0]])\n\nIn [225]: A.A\nOut[225]: \narray([[ 1.,  0.,  0.,  2.,  0.],\n       [ 0.,  0.,  3.,  0.,  0.],\n       [ 0.,  1.,  1.,  2.,  0.]])\n\nIn [226]: A.data\nOut[226]: array([ 1.,  1.,  3.,  1.,  2.,  2.])\n\nIn [227]: A.data[:]=np.exp(A.data)\n\nIn [228]: A.A\nOut[228]: \narray([[  2.71828183,   0.        ,   0.        ,   7.3890561 ,   0.        ],\n       [  0.        ,   0.        ,  20.08553692,   0.        ,   0.        ],\n       [  0.        ,   2.71828183,   2.71828183,   7.3890561 ,   0.        ]])\n<\/code><\/pre>\n\n<p>The main attributes of the <code>csc<\/code> format at <code>data<\/code>, <code>indices<\/code>, <code>indptr<\/code>.  It's possible that <code>data<\/code> has some 0 values if you fiddle with them after creation, but a freshly created matrix shouldn't.<\/p>\n\n<p>The matrix also has a <code>nonzero<\/code> method modeled on the <code>numpy<\/code> one.  In practice it converts the matrix to <code>coo<\/code> format, filters out any zero values, and returns the <code>row<\/code> and <code>col<\/code> attributes:<\/p>\n\n<pre><code>In [229]: A.nonzero()\nOut[229]: (array([0, 0, 1, 2, 2, 2]), array([0, 3, 2, 1, 2, 3]))\n<\/code><\/pre>\n\n<p>And the <code>csc<\/code> format allows indexing just as a dense numpy array:<\/p>\n\n<pre><code>In [230]: A[A.nonzero()]\nOut[230]: \nmatrix([[  2.71828183,   7.3890561 ,  20.08553692,   2.71828183,\n           2.71828183,   7.3890561 ]])\n<\/code><\/pre>\n"},{"AnswerId":"36423504","CreationDate":"2016-04-05T10:16:18.000","ParentId":null,"OwnerUserId":"651246","Title":null,"Body":"<p>The support for sparse matrices in Theano is incomplete, so some things are tricky to achieve. You can use <code>theano.sparse.structured_exp(A_sparse)<\/code> in that particular case, but I try to answer your question more generally below.<\/p>\n\n<p><strong>Comparison<\/strong><\/p>\n\n<p>In Theano one would normally use the comparison operators described here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html<\/a><\/p>\n\n<p>For example, instead of <code>A != 0<\/code>, one would write <code>T.neq(A, 0)<\/code>. With sparse matrices one has to use the comparison operators in <code>theano.sparse<\/code>. Both operators have to be sparse matrices, and the result is also a sparse matrix:<\/p>\n\n<pre><code>mask = theano.sparse.neq(A_sparse, theano.sparse.sp_zeros_like(A_sparse))\n<\/code><\/pre>\n\n<p><strong>Modifying a Subtensor<\/strong><\/p>\n\n<p>In order to modify part of a matrix, one can use <code>theano.tensor.set_subtensor<\/code>. With dense matrices this would work:<\/p>\n\n<pre><code>indices = mask.nonzero()\nA = T.set_subtensor(A[indices], T.exp(A[indices]))\n<\/code><\/pre>\n\n<p>Notice that Theano doesn't have a separated boolean type\u2014the mask is zeros and ones\u2014so <code>nonzero()<\/code> has to be called first to take the indices of the nonzero elements. Furthermore, this is not implemented for sparse matrices. <\/p>\n\n<p><strong>Operating on Nonzero Sparse Elements<\/strong><\/p>\n\n<p>Theano provides sparse operations that are said to be structured and operate only on the nonzero elements. See:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/sparse.html#structured-operation\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/sparse.html#structured-operation<\/a><\/p>\n\n<p>More precisely, they operate on the <code>data<\/code> attribute of a sparse matrix, independent of the indices of the elements. Such operations are straightforward to implement. Note that the structured operations will operate on all the values in the <code>data<\/code> array, also those that are explicitly set to zero.<\/p>\n"}]}
{"QuestionId":36044197,"AnswerCount":4,"Tags":"<c++><opencv><tensorflow>","CreationDate":"2016-03-16T18:34:38.107","AcceptedAnswerId":36093347.0,"OwnerUserId":6073278.0,"Title":"How do I pass an OpenCV Mat into a C++ Tensorflow graph?","Body":"<p>In Tensorflow C++ I can load an image file into the graph using<\/p>\n\n<pre><\/pre>\n\n<p>For an embedded application I would like to instead pass an OpenCV Mat into this graph.  <\/p>\n\n<p>How would I convert the Mat to a tensor that could be used as input to tensorflow::ops::Cast or tensorflow::ops::ExpandDims?<\/p>\n","answers":[{"AnswerId":"59024051","CreationDate":"2019-11-25T01:14:01.980","ParentId":null,"OwnerUserId":"470835","Title":null,"Body":"<pre><code>Tensor convertMatToTensor(Mat &amp;input)\n{\n    int height = input.rows;\n    int width = input.cols;\n    int depth = input.channels();\n\n    Tensor imgTensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({height, width, depth}));\n\n    float* p = imgTensor.flat&lt;float&gt;().data();\n    Mat outputImg(height, width, CV_32FC3, p);\n    input.convertTo(outputImg, CV_32FC3);\n\n    return imgTensor;\n}\n<\/code><\/pre>\n"},{"AnswerId":"36718671","CreationDate":"2016-04-19T12:39:44.610","ParentId":null,"OwnerUserId":"4985869","Title":null,"Body":"<p>I had tried to run inception model on the opencv <code>Mat<\/code> file and following code worked for me <a href=\"https:\/\/gist.github.com\/kyrs\/9adf86366e9e4f04addb\" rel=\"nofollow\">https:\/\/gist.github.com\/kyrs\/9adf86366e9e4f04addb<\/a>. Although there are some issue with integration of opencv and  tensorflow. Code worked without any issue for <code>.png<\/code> files but failed to load  <code>.jpg<\/code> and <code>.jpeg<\/code>. You can follow this for more info <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1924\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1924<\/a> <\/p>\n"},{"AnswerId":"52856207","CreationDate":"2018-10-17T13:33:50.027","ParentId":null,"OwnerUserId":"2953509","Title":null,"Body":"<p>Here is complete example to read and feed:<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>Mat image;\nimage = imread(\"flowers.jpg\", CV_LOAD_IMAGE_COLOR);\ncv::resize(image, image, cv::Size(input_height, input_width), 0, 0, CV_INTER_CUBIC);\n\nint depth = 3;\ntensorflow::Tensor input_tensor(tensorflow::DT_FLOAT,\n                                tensorflow::TensorShape({1, image.rows, image.cols, depth}));\n\nfor (int y = 0; y &lt; image.rows; y++) {\n    for (int x = 0; x &lt; image.cols; x++) {\n        Vec3b pixel = image.at&lt;Vec3b&gt;(y, x);\n\n        input_tensor_mapped(0, y, x, 0) = pixel.val[2]; \/\/R\n        input_tensor_mapped(0, y, x, 1) = pixel.val[1]; \/\/G\n        input_tensor_mapped(0, y, x, 2) = pixel.val[0]; \/\/B\n    }\n}\n\nauto result = Sub(root.WithOpName(\"subtract_mean\"), input_tensor, {input_mean});\nClientSession session(root);\nTF_CHECK_OK(session.Run({result}, out_tensors));\n<\/code><\/pre>\n"},{"AnswerId":"36093347","CreationDate":"2016-03-18T20:07:55.863","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>It's not directly from a CvMat, but you can see an example of how to initialize a Tensor from an in-memory array in the TensorFlow Android example:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0.6.0\/tensorflow\/examples\/android\/jni\/tensorflow_jni.cc#L173\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0.6.0\/tensorflow\/examples\/android\/jni\/tensorflow_jni.cc#L173<\/a><\/p>\n\n<p>You would start off by creating a new tensorflow::Tensor object, with something like this (all code untested):<\/p>\n\n<p><code>tensorflow::Tensor input_tensor(tensorflow::DT_FLOAT,\n      tensorflow::TensorShape({1, height, width, depth}));\n<\/code><\/p>\n\n<p>This creates a Tensor object with float values, with a batch size of 1, and a size of <code>width<\/code>x<code>height<\/code>, and with <code>depth<\/code> channels. For example a 128 wide by 64 high image with 3 channels would pass in a shape of <code>{1, 64, 128, 3}<\/code>. The batch size is just used when you need to pass in multiple images in a single call, and for simple uses you can leave it as 1.<\/p>\n\n<p>Then you would get the underlying array behind the tensor using a line like this:<\/p>\n\n<p><code>auto input_tensor_mapped = input_tensor.tensor&lt;float, 4&gt;();<\/code><\/p>\n\n<p>The <code>input_tensor_mapped<\/code> object is an interface to the data in your newly-created tensor, and you can then copy your own data into it. Here I'm assuming you've set <code>source_data<\/code> as a pointer to your source data, for example:<\/p>\n\n<p><code>const float* source_data = some_structure.imageData;<\/code><\/p>\n\n<p>You can then loop through your data and copy it over:<\/p>\n\n<pre><code>for (int y = 0; y &lt; height; ++y) {\n    const float* source_row = source_data + (y * width * depth);\n    for (int x = 0; x &lt; width; ++x) {\n        const float* source_pixel = source_row + (x * depth);\n        for (int c = 0; c &lt; depth; ++c) {\n           const float* source_value = source_pixel + c;\n           input_tensor_mapped(0, y, x, c) = *source_value;\n        }\n    }\n}\n<\/code><\/pre>\n\n<p>There are obvious opportunities to optimize this naive approach, and I don't have sample code on hand to show how to deal with the OpenCV side of getting the source data, but hopefully this is helpful to get you started.<\/p>\n"}]}
{"QuestionId":36046125,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-16T20:18:43.827","AcceptedAnswerId":null,"OwnerUserId":1968073.0,"Title":"Can I use TensorBoard also with jupyter notebooks","Body":"<p>I am experimenting (learning) TensorBoard and use the following code I got from the internet (simple regression function)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The code runs fine when I create a python file and run the file with<\/p>\n\n<pre><\/pre>\n\n<p>also it runs fine in the jupyter notebook<\/p>\n\n<p>However, while Tensorboard gets the information from running the python file (that is to say, it creates the xyz....home file), the interactive version does not create any info usable for Tensorboard.<\/p>\n\n<p>Can somebody explain to me why, please!<\/p>\n\n<p>Thanks\nPeter<\/p>\n","answers":[{"AnswerId":"41921643","CreationDate":"2017-01-29T13:52:47.623","ParentId":null,"OwnerUserId":"7483498","Title":null,"Body":"<p>Be sure that you use the full path when starting tensorboard.<\/p>\n\n<pre><code>tensorboard --logdir='.\/somedirectory\/mnist_logs'\n<\/code><\/pre>\n"}]}
{"QuestionId":36046448,"AnswerCount":1,"Tags":"<docker><ipython><tensorflow>","CreationDate":"2016-03-16T20:37:38.230","AcceptedAnswerId":null,"OwnerUserId":6073688.0,"Title":"Windows 7 jupyter notebook - ImportError: No module named 'tensorflow'","Body":"<p>I am novice to tensorflow and unix\/linux usage. So my questions might be too simple.<\/p>\n\n<p>I have been running .ipynb files from Anaconda\/Jupyter notebook successfully. However after I installed tensorflow using pip, i am getting an error when I try to import tensorflow.<\/p>\n\n<pre><\/pre>\n\n<p>Alternatively I have also downloaded tensorflow image using docker and I can run it from docker container. However I have no clue how to execute .ipynb files (&amp; also how to add that to my docker container with tensorflow)<\/p>\n","answers":[{"AnswerId":"37280604","CreationDate":"2016-05-17T15:35:09.317","ParentId":null,"OwnerUserId":"6346566","Title":null,"Body":"<p>I solved the problem using <a href=\"https:\/\/stackoverflow.com\/questions\/36046448\/windows-7-jupyter-notebook-executing-tensorflow\/37280604#comment59749704_36046448\">Yaroslav Bulatov's solution<\/a> from the comments.<\/p>\n\n<p>Use <\/p>\n\n<p><code>conda install -c http:\/\/conda.anaconda.org\/jjhelmus tensorflow<\/code> <\/p>\n\n<p>instead of <\/p>\n\n<p><code>conda install -c conda.anaconda.org\/jjhelmus tensorflow<\/code>.<\/p>\n"}]}
{"QuestionId":36046840,"AnswerCount":1,"Tags":"<machine-learning><theano>","CreationDate":"2016-03-16T21:01:49.037","AcceptedAnswerId":36047956.0,"OwnerUserId":2979687.0,"Title":"Theano: expected 2, got 1 with shape (128,).')","Body":"<p>Im trying to make a neural network. I have followed the video from <\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=S75EdAcXHKk\" rel=\"nofollow\">https:\/\/www.youtube.com\/watch?v=S75EdAcXHKk<\/a><\/p>\n\n<p>I have loaded the adult.data training set. <\/p>\n\n<p>I am now on my way of training and i have these lines where the code fails.<\/p>\n\n<pre><\/pre>\n\n<p>I am strugling with an error message which is: <\/p>\n\n<pre><\/pre>\n\n<p>File \"C:\\Users\\Bjornars\\PycharmProjects\\cogs-118a\\Project\\NN\\Network.py\", line 101, in training\n    self.cost = train(tr.reshape(128,106), trY[start:end])\n  File \"C:\\Anaconda3\\lib\\site-packages\\theano\\compile\\function_module.py\", line 513, in <strong>call<\/strong>\n    allow_downcast=s.allow_downcast)\n  File \"C:\\Anaconda3\\lib\\site-packages\\theano\\tensor\\type.py\", line 169, in filter\n    data.shape))\nTypeError: ('Bad input argument to theano function with name \"C:\\Users\\Bjornars\\PycharmProjects\\cogs-118a\\Project\\NN\\Network.py:84\"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 1 with shape (128,).')<\/p>\n\n<p>The shape of the array im sending in is (5000,106)<\/p>\n\n<p>---Solved----<\/p>\n","answers":[{"AnswerId":"36047956","CreationDate":"2016-03-16T22:13:46.207","ParentId":null,"OwnerUserId":"2979687","Title":null,"Body":"<p>Used this, it expected array not number in trY<\/p>\n\n<pre><code>    def preprocess(self,trDmatrix,labels):\n          for i in range(len(trDmatrix)):\n              numbers = [0.0]*2\n              numbers[int(labels[i])]= 1.0\n              labels[i] = numbers\n          return trDmatrix, labels\n<\/code><\/pre>\n"}]}
{"QuestionId":36047937,"AnswerCount":1,"Tags":"<c++><performance><gpu><tensorflow><mxnet>","CreationDate":"2016-03-16T22:12:17.880","AcceptedAnswerId":null,"OwnerUserId":258834.0,"Title":"Maximize tensorflow multi gpu performance","Body":"<p>I was wondering if anybody could advise on how to get peak performance out of tensorflow in a 4 GPU setting.<\/p>\n\n<p>As a test I created two of the same network (18 ish layer residual network with small filter banks (ranging from 16-128) on 32x32 inputs. Batch size 512, 128 per GPU.). One in MXNet and one I have modelled off of <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\">the inception example<\/a>.<\/p>\n\n<p>My MXNet network can train at around 7k examples a second where tensorflow is only capable of 4.2k with dummy data and 3.7 with real data.<\/p>\n\n<p>(when running on 1 GPU the numbers are 1.2k examples a second vs 2.1k)<\/p>\n\n<p>In my experiment I have a few questions in hopes to speed things up.<\/p>\n\n<ol>\n<li><p>GPU utilization seems quite low when training. I noticed that in the tensorflow white paper there is support for running multiple streams on the same GPU. Is this possible in the public release?<\/p><\/li>\n<li><p>Is there anyway to perform multiple train operations in one execution of ? Or have async execution? This would allow for weight updates to be done at the same time as the next batches forward pass? I have tried using 2 threads (both system and with 's), but this only resulted in a slowdown. MXNet is able to increase speeds by running weight updates on the CPU so that the gpu's can be used for the next batch.<\/p><\/li>\n<li><p>Will the new distributed run time get around some of these issues by letting me run more than one worker on a single machine?<\/p><\/li>\n<li><p>Is there something else that can be done?<\/p><\/li>\n<\/ol>\n\n<p>I know there are a number of similar questions here on stack overflow, but though my searching I couldn't find a solution to my problems that I have not already tried.<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>I did a little bit of CUDA profiling to see what the expensive kernels were. According to my run, 21.4% of the time is spent inside:<\/p>\n\n<pre><\/pre>\n\n<p>and 20.0% of the time were spent in <\/p>\n\n<pre><\/pre>\n\n<p>Off of the Signature I am not exactly sure what these are doing. Do these make sense?<\/p>\n\n<p>In addition to this, the analysis reports low kernel concurrency, 0%, as expected.\nAnd Low compute utilization 34.9% (granted this includes start-up time and a little bit of python in train loop. Around 32 seconds total out of 91. This comes out to around 50% utilization inside tensorflow.)<\/p>\n\n<p><strong>Edit 2:<\/strong><\/p>\n\n<p>I have attached a copy of the trimmed down <a href=\"https:\/\/gist.github.com\/lukemetz\/2072e9f7b3f2b9325e25\">source code<\/a>. In general though I am more concerned about question 1-3 and don't want to take too much of ever bodies time.<\/p>\n\n<p>In addition I am running on tensorflow built from: f07234db2f7b316b08f7df25417245274b63342a<\/p>\n\n<p><strong>Edit 3:<\/strong><\/p>\n\n<p>Updated to the most recent tensorflow (63409bd23facad471973b110df998782c0e19c06) same code, default data format (NHWC) and that seemed to speed this up a lot.\nOn fake data 6.7k-6.8k (thermal dependence I think?) examples a second 4gpu. 1gpu -- 2.0k examples a second.\nReal data performance is around 4.9k examples a second for 4gpu. 1gpu -- 1.7k examples a second.<\/p>\n\n<p><strong>Edit 4:<\/strong><\/p>\n\n<p>In addition I tried out switching data formats to BCHW. I made the conversion modelled off of <a href=\"https:\/\/github.com\/soumith\/convnet-benchmarks\/blob\/master\/tensorflow\/benchmark_googlenet.py\">Soumith's benchmarks<\/a>. The convolution parts were indeed faster, but batch norm appears to be messing everything up. With a <a href=\"https:\/\/gist.github.com\/lukemetz\/2072e9f7b3f2b9325e25#file-batchnorm_function-py\">naive implementation<\/a> (fixing axis, and making weights [1,C,1,1] instead of [C,]) I am only able to get 1.2k examples a second on 4 gpu (fake data). Where as with a transpose before and after the batch norm op I am able to get 6.2k examples a second (fake data). Still slower than the NHWC data_format.<\/p>\n","answers":[{"AnswerId":"36077298","CreationDate":"2016-03-18T06:08:14.610","ParentId":null,"OwnerUserId":"5545462","Title":null,"Body":"<p>It's a bit hard to diagnose your program's performance problem without seeing the code. Is it possible for us to read your test code somehow?<\/p>\n\n<p>TensorPadding showing on the top is a bit strange. I'd expect cudnn calls should be on the top of the profile. Anyway, showing us the test code will be helpful.<\/p>\n"}]}
{"QuestionId":36050686,"AnswerCount":4,"Tags":"<python><neural-network><tensorflow><deep-learning><mnist>","CreationDate":"2016-03-17T02:37:17.167","AcceptedAnswerId":null,"OwnerUserId":4552561.0,"Title":"How to save weights of training data from MNIST testing on tensorflow for future use?","Body":"<p>I am facing problems with saving the training weight (W) for MNIST tensorflow example as described here. <a href=\"https:\/\/www.tensorflow.org\/versions\/0.6.0\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">MNIST tensorflow<\/a>. If my understanding is correct, we need the training weight in future for other testing cases (not the MNIST testing case). I tried print W.eval() to get the weight; but it happens to provide me with a zero matrix of 784x10. How can I get the weights in an array form or in .csv format (not .cpkt)? <\/p>\n","answers":[{"AnswerId":"36057536","CreationDate":"2016-03-17T10:17:00.913","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>I've done something similar like so:<\/p>\n\n<pre><code>weight_values = session.run(W)\nnumpy.savetxt(\"myfilename.csv\", weight_values, delimiter=\",\")\n<\/code><\/pre>\n"},{"AnswerId":"44757790","CreationDate":"2017-06-26T10:34:25.153","ParentId":null,"OwnerUserId":"86643","Title":null,"Body":"<p>Has anyone used <a href=\"https:\/\/github.com\/gradientzoo\/python-gradientzoo\" rel=\"nofollow noreferrer\">gradientzoo<\/a> ?  <\/p>\n\n<blockquote>\n  <p>save, version, and share your neural net model weights<\/p>\n<\/blockquote>\n\n<p>(haven't used it myself)<\/p>\n"},{"AnswerId":"41754307","CreationDate":"2017-01-20T00:49:57.477","ParentId":null,"OwnerUserId":"7433039","Title":null,"Body":"<p>You can use <code>tf.train.Saver().save<\/code> to save your model after it has been created. to use it simply run:<\/p>\n\n<pre><code>   tf.train.Saver().save( SESSION-NAME-HERE , \"SAVE-LOCATION-HERE\")\n<\/code><\/pre>\n\n<p>At some point after both <code>tf.Session() as SESSION-NAME-HERE:<\/code> and <code>SESSION-NAME-HERE.run(tf.global_variables_initializer())<\/code> have been used (But before you close the session).<\/p>\n\n<p>To Restore you can use <code>tf.train.Saver().restore()<\/code> with similar syntax to before:<\/p>\n\n<pre><code>   tf.train.Saver().restore(SESSION-NAME-HERE , \"SAVE-LOCATION-HERE\")\n<\/code><\/pre>\n\n<p>Run this After <code>tf.Session() as SESSION-NAME-HERE:<\/code>. This code restores previous variables used.<\/p>\n\n<p>For futher information see: <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/variables\/<\/a><\/p>\n"},{"AnswerId":"36483784","CreationDate":"2016-04-07T17:55:58.943","ParentId":null,"OwnerUserId":"2332621","Title":null,"Body":"<p>I ran into a similar problem. Recording the answer here in case it's your issue as well \/ for posterity. I was printing <code>W<\/code> to the console and it looked like it was all zeros:<\/p>\n\n<pre><code>[[ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n ..., \n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]\n [ 0.  0.  0. ...,  0.  0.  0.]]\n<\/code><\/pre>\n\n<p>But actually, the weights displayed in each column correspond to the first 3 and last 3 pixels in each image (upper left and lower right, I think) ... and there is no color in any MNIST image in those locations. So the weights stay zero.<\/p>\n\n<p>If you follow Daniel Slater's advice you can see the non-zero weights in the output.<\/p>\n"}]}
{"QuestionId":36055090,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-03-17T08:26:44.693","AcceptedAnswerId":36055812.0,"OwnerUserId":3919542.0,"Title":"zlib.error: Error -3 while decompressing: invalid distance code","Body":"<p>I was trying to create a model using Tensorflow using the mnist dataset .\nI have installed Tensorflow Properly and then when i try to create the model using the command .<\/p>\n\n<pre><\/pre>\n\n<p>I am getting an error message on the console :<\/p>\n\n<pre><\/pre>\n\n<p>i am using Python2.7 installed in a 14.04 Ubuntu operating system .<\/p>\n","answers":[{"AnswerId":"36055812","CreationDate":"2016-03-17T09:02:17.790","ParentId":null,"OwnerUserId":"5088142","Title":null,"Body":"<p>It seems like the source of the problem is invalid data file: \"data\/train-images-idx3-ubyte.gz\". \nIs there any way to get a fresh copy of the file?\nPlease validate the new copy of the file using the following shell command:<\/p>\n\n<pre><code>gunzip -t  data\/train-images-idx3-ubyte.gz\n<\/code><\/pre>\n\n<p>Once the above test will return answer that the gz file is valid, execute your script:<\/p>\n\n<pre><code>python convolutional.py\n<\/code><\/pre>\n"}]}
{"QuestionId":36055536,"AnswerCount":1,"Tags":"<python><variables><if-statement><scope><tensorflow>","CreationDate":"2016-03-17T08:49:42.083","AcceptedAnswerId":null,"OwnerUserId":2886263.0,"Title":"Why variable defined in a if else blocks can be used outside blocks?","Body":"<pre><\/pre>\n\n<p>Here y is tensorflow variable and is defined inside the if else block, why does it can be used outside the block?<\/p>\n","answers":[{"AnswerId":"36056377","CreationDate":"2016-03-17T09:27:19.653","ParentId":null,"OwnerUserId":"2411320","Title":null,"Body":"<p>This is more general than <a href=\"\/questions\/tagged\/tensorflow\" class=\"post-tag\" title=\"show questions tagged &#39;tensorflow&#39;\" rel=\"tag\">tensorflow<\/a>, it has to do with <a href=\"\/questions\/tagged\/python\" class=\"post-tag\" title=\"show questions tagged &#39;python&#39;\" rel=\"tag\">python<\/a>'s scope of variables. Remember this:<\/p>\n\n<blockquote>\n  <p>Python <strong>does not<\/strong> have a block scope!<sup>*<\/sup><\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Consider this trivial example:<\/p>\n\n<pre><code>x = 2\na = 5\nif a == 5:\n    y = 2 * x\nelse:\n    y = 3 * x\n\nz = 4 * y\nprint z     # prints 16\n<\/code><\/pre>\n\n<p>What I am trying to say is that <code>y<\/code> is not a local variable defined in the scope of the body of the <em>if statement<\/em>, thus it's OK to use it after the if statement.<\/p>\n\n<hr>\n\n<p>For more: <a href=\"http:\/\/python-textbok.readthedocs.org\/en\/1.0\/Variables_and_Scope.html\" rel=\"nofollow noreferrer\">Variables_and_Scope<\/a>.<\/p>\n\n<p><sub>*<a href=\"https:\/\/stackoverflow.com\/questions\/6167923\/block-scope-in-python\">Block scope in Python<\/a><\/sub><\/p>\n"}]}
{"QuestionId":36057715,"AnswerCount":1,"Tags":"<nlp><tensorflow>","CreationDate":"2016-03-17T10:24:33.243","AcceptedAnswerId":36066107.0,"OwnerUserId":2662629.0,"Title":"TensorFlow with a NER-Tagger","Body":"<p>I was wondering if there is any possibility to use Named-Entity-Recognition with a self trained model in tensorflow.<\/p>\n\n<p>There is a word2vec implementation, but I could not find the 'classic' POS or NER tagger.<\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"36066107","CreationDate":"2016-03-17T16:16:05.410","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>You can adapt the Sequence-to-Sequence model for NER tagging. Your training text is the source vocabulary\/sequences to the encoder:<\/p>\n\n<pre><code>Yesterday afternoon , Mike Smith drove to New York .\n<\/code><\/pre>\n\n<p>your BIO \/ BILOU NER tags are your target vocabulary\/sequences to the decoder for NER tagging:<\/p>\n\n<pre><code>O O O B_PER I_PER O O B_LOC I_LOC O\n<\/code><\/pre>\n\n<p>or instead use POS tags to the decoder for POS tagging:<\/p>\n\n<pre><code>NN NN ,  NNP NNP VBD TO NNP NNP . \n<\/code><\/pre>\n\n<p>[IMHO using a deep learning approach usually eliminates the need POS tagging as an intermediate step, unless you specifically need those features as an output for something.]<\/p>\n\n<p>You would probably want to switch off the word embeddings for the decoder.<\/p>\n\n<p>This well-known paper applies sequence-to-sequence models to syntactic parsing which has some similarities to the POS and\/or NER tasks: <a href=\"http:\/\/arxiv.org\/abs\/1412.7449\">Grammar as a Foreign Language<\/a><\/p>\n"}]}
{"QuestionId":36058099,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-17T10:40:14.200","AcceptedAnswerId":null,"OwnerUserId":2902280.0,"Title":"Stride with different height and width in Tensorflow","Body":"<p>I'm using Tensorflow to process 1D signals (shape being ) and would like to perform stride operations. However, when I try , I get a ValueError:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a known workaround to the limits of the current implementation, which would allow me to have an horizontal stride only?<\/p>\n","answers":[{"AnswerId":"36075886","CreationDate":"2016-03-18T04:00:03.300","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>We're working on supporting strides with different lengths.<\/p>\n"}]}
{"QuestionId":36064720,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2016-03-17T15:17:06.407","AcceptedAnswerId":null,"OwnerUserId":4397140.0,"Title":"Caffe: How to output probabilities and class labels per image to file with lmdb?","Body":"<p>I have a network I am trying to test using caffe- however I have lots of test images and would like to use lmdb blobs to speed up the testing\/experimentation process (as opposed to doing it image by image in MatCaffe\/PyCaffe). At the end I would like to have a file that outputs something like: <\/p>\n\n<pre><\/pre>\n\n<p>for each of the 1-k top guesses, where each line of this file is the image being tested. <\/p>\n\n<p>I know how to modify the prototxt files to get the probabilities, argmax, and do logging, but I can't seem to figure out how to write it to a file in this specific way. <\/p>\n\n<p>Anyone have any ideas?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Added this to the bottom of my prototxt:<\/p>\n\n<pre><\/pre>\n\n<p>and I generate the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>It seems to be relatively similar to this <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1726\" rel=\"nofollow\">post<\/a>, however the link to that solution is a different version of caffe (eg, in my version, there is no file data_layers.hpp - I only have data_layer.hpp )\/ <\/p>\n","answers":[]}
{"QuestionId":36065634,"AnswerCount":2,"Tags":"<for-loop><lua><iteration><torch>","CreationDate":"2016-03-17T15:54:14.687","AcceptedAnswerId":36068126.0,"OwnerUserId":399397.0,"Title":"parallel iteration in lua","Body":"<p>I would like to for-loop through multiple tables in parallel in Lua. I could just do:<\/p>\n\n<pre><\/pre>\n\n<p>But I'd rather something like python's :<\/p>\n\n<pre><\/pre>\n\n<p>Is there such a thing in standard Lua (or at least in whatever comes packaged with torch?).<\/p>\n","answers":[{"AnswerId":"36096338","CreationDate":"2016-03-19T00:18:11.053","ParentId":null,"OwnerUserId":"3735873","Title":null,"Body":"<pre><code>--------------------------------------------------------------------------------\n-- Python-like zip() iterator\n--------------------------------------------------------------------------------\n\nfunction zip(...)\n  local arrays, ans = {...}, {}\n  local index = 0\n  return\n    function()\n      index = index + 1\n      for i,t in ipairs(arrays) do\n        if type(t) == 'function' then ans[i] = t() else ans[i] = t[index] end\n        if ans[i] == nil then return end\n      end\n      return unpack(ans)\n    end\nend\n\n--------------------------------------------------------------------------------\n-- Example use:\n--------------------------------------------------------------------------------\n\na = {'a','b','c','d'}\nb = {3,2,1}\nc = {7,8,9,10,11}\n\nfor a,b,c,line in zip(a,b,c,io.lines(arg[0])) do\n  print(a,b,c,line)\nend\n\nprint '\\n--- Done! ---'\n<\/code><\/pre>\n"},{"AnswerId":"36068126","CreationDate":"2016-03-17T17:50:27.633","ParentId":null,"OwnerUserId":"3677376","Title":null,"Body":"<p>If you want something in Lua that's similar to some Python function, you should look at <a href=\"https:\/\/github.com\/stevedonovan\/Penlight\" rel=\"noreferrer\">Penlight<\/a> first. For this specific case there is the <a href=\"http:\/\/stevedonovan.github.io\/Penlight\/api\/libraries\/pl.seq.html#zip\" rel=\"noreferrer\"><code>seq.zip<\/code><\/a> function. It <a href=\"https:\/\/github.com\/torch\/distro\/tree\/master\/extra\" rel=\"noreferrer\">seems<\/a> that Penlight is installed together with Torch, but you can also get it via LuaRocks (which again is bundled with at least one Torch distribution).<\/p>\n\n<p>Anyway, the <code>seq.zip<\/code> function in Penlight only supports zipping two sequences. Here is a version that should behave more like Python's <code>zip<\/code>, i.e. allowing more (or less) than two sequences:<\/p>\n\n<pre><code>local zip\ndo\n  local unpack = table.unpack or unpack\n  local function zip_select( i, var1, ... )\n    if var1 then\n      return var1, select( i, var1, ... )\n    end\n  end\n\n  function zip( ... )\n    local iterators = { n=select( '#', ... ), ... }\n    for i = 1, iterators.n do\n      assert( type( iterators[i] ) == \"table\",\n              \"you have to wrap the iterators in a table\" )\n      if type( iterators[i][1] ) ~= \"number\" then\n        table.insert( iterators[i], 1, -1 )\n      end\n    end\n    return function()\n      local results = {}\n      for i = 1, iterators.n do\n        local it = iterators[i]\n        it[4], results[i] = zip_select( it[1], it[2]( it[3], it[4] ) )\n        if it[4] == nil then return nil end\n      end\n      return unpack( results, 1, iterators.n )\n    end\n  end\nend\n\n-- example code (assumes that this file is called \"zip.lua\"):\nlocal t1 = { 2, 4, 6, 8, 10, 12, 14 }\nlocal t2 = { \"a\", \"b\", \"c\", \"d\", \"e\", \"f\" }\nfor a, b, c in zip( {ipairs( t1 )}, {ipairs( t2 )}, {io.lines\"zip.lua\"} ) do\n  print( a, b, c )\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":36066617,"AnswerCount":1,"Tags":"<theano><lstm><recurrent-neural-network>","CreationDate":"2016-03-17T16:36:41.183","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"How do I use Theanets LSTM RNN's on my time series data?","Body":"<p>I have a simple dataframe consisting of one column. In that column are 10320 observations (numerical). I'm simulating Time-Series data by inserting the data into a plot with a window of 200 observations each. Here is the code for plotting.<\/p>\n\n<pre><\/pre>\n\n<p>This simulates a flow of real-time data and visualizes it. What I want is to apply theanets RNN LSTM to the data to detect anomalies unsupervised. Because I am doing it unsupervised I don't think that I need to split my data into training and test sets. I haven't found much of anything that makes sense to me so far and have been googling for about 2 hours. Just hoping that you guys may be able to help. I want to put the prediction output of the RNN on the graph as well and define a threshold that, if the error is too large, the values will be identified as anomalous. If you need more information please comment and let me know. Thank you!<\/p>\n","answers":[{"AnswerId":"42591909","CreationDate":"2017-03-04T03:57:25.243","ParentId":null,"OwnerUserId":"5288679","Title":null,"Body":"<p><strong>READING<\/strong><\/p>\n\n<ol>\n<li>Like neurons, LSTM networks are build of interconnected <a href=\"https:\/\/en.wikipedia.org\/wiki\/Long_short-term_memory#\/media\/File:Long_Short_Term_Memory.png\" rel=\"nofollow noreferrer\">LSTM Blocks<\/a> whose training is done via <a href=\"https:\/\/en.wikipedia.org\/wiki\/Backpropagation_through_time\" rel=\"nofollow noreferrer\">BackPropogation Through Time<\/a>. <\/li>\n<li>Classical anomaly detection using time series required prediction of time series output in future (<em>at one or more points<\/em>) and finding error on these points with true values. <strong>Prediction Error<\/strong> above a threshold will reflect and amomly<\/li>\n<\/ol>\n\n<p><strong>SOLUTION<\/strong><\/p>\n\n<p>Having said this<\/p>\n\n<ol>\n<li>You've to train network so you need <em>training sets<\/em> and <em>test sets<\/em> both<\/li>\n<li>Use N inputs to predict M outputs (<em>decide upon N and M with experimentation - values for which training error is low<\/em>)<\/li>\n<li>Scroll a window of (N+M) elements in input data and use this data array of (N+M) items also termed as frame to train or test network.<\/li>\n<li>Typically we use 90% of starting series for training and 10% for testing.<\/li>\n<\/ol>\n\n<p><strong>This scheme will fail as if training is not proper there will be false prediction errors which are not-anomaly<\/strong>. So make sure to provide enough training, and most important <em>shuffle training frames<\/em> and <em>consider all variations<\/em>.<\/p>\n"}]}
{"QuestionId":36068292,"AnswerCount":2,"Tags":"<machine-learning><artificial-intelligence><deep-learning><prediction><keras>","CreationDate":"2016-03-17T17:59:15.177","AcceptedAnswerId":null,"OwnerUserId":4497662.0,"Title":"Incorporating user feedback in a ML model","Body":"<p>I have developed a ML model for a classification (0\/1) NLP task and deployed it in production environment. The prediction of the model is displayed to users, and the users have the option to give a feedback (if the prediction was right\/wrong). <\/p>\n\n<p>How can I continuously incorporate this feedback in my model ? From a UX stand point you dont want a user to correct\/teach the system more than twice\/thrice for a specific input, system shld learn fast i.e. so the feedback shld be incorporated \"fast\". (Google priority inbox does this in a seamless way)<\/p>\n\n<p>How does one build this \"feedback loop\" using which my system can improve ? I have searched a lot on net but could not find relevant material. any pointers will be of great help. <\/p>\n\n<p>Pls dont say retrain the model from scratch by including new data points. Thats surely not how google and facebook build their smart systems<\/p>\n\n<p>To further explain my question - think of google's spam detector or their priority inbox or their recent feature of \"smart replies\". Its a well known fact that they have the ability to learn \/ incorporate (fast) user feed. <\/p>\n\n<p><strong>All the while when it incorporates the user feedback fast (i.e. user has to teach the system correct output atmost 2-3 times per data point and the system start to give correct output for that data point) AND it also ensure it maintains old learnings and does not start to give wrong outputs on older data points (where it was giving right output earlier) while incorporating the learning from new data point.<\/strong> <\/p>\n\n<p><em>I have not found any blog\/literature\/discussion w.r.t how to build such systems - An intelligent system that explains in detaieedback loop\" in ML systems<\/em><\/p>\n\n<p>Hope my question is little more clear now.<\/p>\n\n<p>Update: Some related questions I found are:             <\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/23056460\/does-the-svm-in-sklearn-support-incremental-online-learning\">Does the SVM in sklearn support incremental (online) learning?<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/datascience.stackexchange.com\/questions\/1073\/libraries-for-online-machine-learning\">https:\/\/datascience.stackexchange.com\/questions\/1073\/libraries-for-online-machine-learning<\/a><\/p><\/li>\n<li><p><a href=\"http:\/\/mlwave.com\/predicting-click-through-rates-with-online-machine-learning\/\" rel=\"noreferrer\">http:\/\/mlwave.com\/predicting-click-through-rates-with-online-machine-learning\/<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Concept_drift\" rel=\"noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Concept_drift<\/a><\/p><\/li>\n<\/ul>\n\n<p>Update: I still dont have a concrete answer but such a recipe does exists. Read the section \"Learning from the feedback\" in the following blog  <a href=\"https:\/\/www.ibm.com\/developerworks\/community\/blogs\/jfp\/entry\/Machine_Learning_Learning_Machine\" rel=\"noreferrer\">Machine Learning != Learning Machine<\/a>. In this Jean talks about \"adding a feedback ingestion loop to machine\". Same in <a href=\"https:\/\/stackoverflow.com\/questions\/22854131\/use-feedback-or-reinforcement-in-machine-learning\">here<\/a>, <a href=\"http:\/\/siliconangle.com\/blog\/2015\/12\/11\/machine-learning-can-machines-ever-be-taught-to-think-like-us\/\" rel=\"noreferrer\">here<\/a>, here<a href=\"http:\/\/static.googleusercontent.com\/media\/research.google.com\/en\/\/pubs\/archive\/43146.pdf\" rel=\"noreferrer\">4<\/a>.<\/p>\n","answers":[{"AnswerId":"44259054","CreationDate":"2017-05-30T09:57:31.197","ParentId":null,"OwnerUserId":"5242337","Title":null,"Body":"<p>There could be couple of ways to do this:<\/p>\n\n<p>1) You can incorporate the feedback that you get from the user to only train the last layer of your model, keeping the weights of all other layers intact. Intuitively, for example, in case of CNN this means you are extracting the features using your model but slightly adjusting the classifier to account for the peculiarities of your specific user.<\/p>\n\n<p>2) Another way could be to have a global model ( which was trained on your large training set) and a simple logistic regression which is user specific. For final predictions, you can combine the results of the two predictions. See <a href=\"http:\/\/static.googleusercontent.com\/media\/research.google.com\/en\/\/pubs\/archive\/36955.pdf\" rel=\"nofollow noreferrer\">this paper<\/a> by google on how they do it for their priority inbox.<\/p>\n"},{"AnswerId":"39830595","CreationDate":"2016-10-03T11:32:12.770","ParentId":null,"OwnerUserId":"4497662","Title":null,"Body":"<p>Build a simple, light model(s) that can be updated per feedback. Online Machine learning gives a number of candidates for this<\/p>\n\n<p>Most good online classifiers are linear. In which case we can have a couple of them and achieve non-linearity by combining them via a small shallow neural net<\/p>\n\n<p><a href=\"https:\/\/stats.stackexchange.com\/questions\/126546\/nonlinear-dynamic-online-classification-looking-for-an-algorithm\">https:\/\/stats.stackexchange.com\/questions\/126546\/nonlinear-dynamic-online-classification-looking-for-an-algorithm<\/a><\/p>\n"}]}
{"QuestionId":36069260,"AnswerCount":4,"Tags":"<python><time-series><deep-learning><keras><lstm>","CreationDate":"2016-03-17T18:50:14.123","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"Why do I get a Keras LSTM RNN input_shape error?","Body":"<p>I keep getting an input_shape error from the following code.<\/p>\n\n<pre><\/pre>\n\n<p>All of the above runs smoothly. This is where it goes wrong.<\/p>\n\n<pre><\/pre>\n\n<p>It returns this error.<\/p>\n\n<pre><\/pre>\n\n<p>When I check <a href=\"http:\/\/keras.io\/layers\/core\/\" rel=\"noreferrer\">the website<\/a> it says to specify a tuple \"(e.g. (100,) for 100-dimensional inputs).\"<\/p>\n\n<p>That being said, my data set consists of one column with a length of 10320. I assume that that means that I should be putting  in as the input_shape, but I get the error anyways. Does anyone have a solution?<\/p>\n","answers":[{"AnswerId":"38084402","CreationDate":"2016-06-28T19:09:10.420","ParentId":null,"OwnerUserId":"681402","Title":null,"Body":"<p>My understanding is that both your input and your output are one dimensional vectors. The trick is to reshape them per Keras requirements:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers.core import Dense, Activation, Dropout\nfrom keras.layers.recurrent import LSTM\nimport numpy as np\n\nX= np.random.rand(1000)\ny = 2*X\n\npoi = int(len(X) * .8)\nX_train = X[:poi]\ny_train = y[:poi]\n\nX_test = X[poi:]\ny_test = y[poi:]\n\n# you have to change your input shape (nb_samples, timesteps, input_dim)\nX_train = X_train.reshape(len(X_train), 1, 1)\n# and also the output shape (note that the output *shape* is 2 dimensional)\ny_train = y_train.reshape(len(y_train), 1)\n\n\n#in_out_neurons = 2 \nin_out_neurons = 1\n\nhidden_neurons = 300\nmodel = Sequential()\n#model.add(Masking(mask_value=0, input_shape=(input_dim,)))\nmodel.add(LSTM(hidden_neurons, return_sequences=False, batch_input_shape=X_train.shape))\n# only specify the output dimension\nmodel.add(Dense(in_out_neurons))\nmodel.add(Activation(\"linear\"))\nmodel.compile(loss=\"mean_squared_error\", optimizer=\"rmsprop\")\nmodel.fit(X_train, y_train, nb_epoch=10, validation_split=0.05)\n\n# calculate test set MSE\npreds = model.predict(X_test).reshape(len(y_test))\nMSE = np.mean((preds-y_test)**2)\n<\/code><\/pre>\n\n<p>Here are the key points:<\/p>\n\n<ul>\n<li>when you add your first layer, you are required to specify the number of hidden nodes, and your input shape. Consequent layers don't require the input shape as they can infer it from the hidden nodes of the previous layer<\/li>\n<li>Similarly, for your output layer you <em>only<\/em> specify the number of output nodes<\/li>\n<\/ul>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"36273104","CreationDate":"2016-03-28T23:48:33.313","ParentId":null,"OwnerUserId":"2877875","Title":null,"Body":"<p>Try to use the LSTM layer without specifying the input shape. Let Keras do the work for you. I think you commented the masking as well because your getting similar issue. I faced it before and it turns out the input_shape = (time_steps, input_dim). I think this happens due to the new automatic shape inference in Keras.<\/p>\n"},{"AnswerId":"43089677","CreationDate":"2017-03-29T09:33:27.560","ParentId":null,"OwnerUserId":"2323035","Title":null,"Body":"<p>Some more information: when using RNN (like LSTM) with sequences of variable length you have to take of the format of your data.<\/p>\n\n<p>When you group sequences in order to pass it to the fit method, keras will try to build a matrix of samples, which implies that all input sequences must have the same size, otherwise you won't have a matrix of the correct dimension.<\/p>\n\n<p>There several possible solutions:<\/p>\n\n<ol>\n<li>train your network using samples one by one (using fit_generator for example)<\/li>\n<li>pad all your sequences so they have the same size<\/li>\n<li>group sequences by size (eventually padding them) and train your network group by group (again using generator based fit)<\/li>\n<\/ol>\n\n<p>The third solution corresponds to the most common strategy with variable size. And if you pad sequences (second or third solution) you may want to add a masking layer as input.<\/p>\n\n<p>If you're not sure, try to print the shape of your data (using the shape attribute of the numpy array.)<\/p>\n\n<p>You may need to look at: <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/sequence\/<\/a> (pad_sequences) and <a href=\"https:\/\/keras.io\/layers\/core\/#masking\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/core\/#masking<\/a> <\/p>\n"},{"AnswerId":"43434751","CreationDate":"2017-04-16T07:15:25.710","ParentId":null,"OwnerUserId":"825707","Title":null,"Body":"<p>Following is the working version with <strong>Keras 2.0.0<\/strong>, Modified radix's code<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers.core import Dense, Activation, Dropout\nfrom keras.layers.recurrent import LSTM\nimport numpy as np\n\nX= np.random.rand(1000)\ny = 2 * X\n\npoi = int(len(X) * .8)\nX_train = X[:poi]\ny_train = y[:poi]\n\nX_test = X[poi:]\ny_test = y[poi:]\n\n# you have to change your input shape (nb_samples, timesteps, input_dim)\nX_train = X_train.reshape(len(X_train), 1, 1)\n# and also the output shape (note that the output *shape* is 2 dimensional)\ny_train = y_train.reshape(len(y_train), 1)\n\n# Change test data's dimension also.\nX_test = X_test.reshape(len(X_test),1,1)\ny_test = y_test.reshape(len(y_test),1)\n\n\n#in_out_neurons = 2\nin_out_neurons = 1\n\nhidden_neurons = 300\nmodel = Sequential()\n# model.add(Masking(mask_value=0, input_shape=(input_dim,)))\n# Remove batch_input_shape and add input_shape = (1,1) - Imp change for Keras 2.0.0\nmodel.add(LSTM(hidden_neurons, return_sequences=False, input_shape=(X_train.shape[1],X_train.shape[2])))\n# only specify the output dimension\nmodel.add(Dense(in_out_neurons))\nmodel.add(Activation(\"linear\"))\nmodel.compile(loss=\"mean_squared_error\", optimizer=\"rmsprop\")\nmodel.summary()\nmodel.fit(X_train, y_train, epochs=10, validation_split=0.05)\n\n# calculate test set MSE\npreds = model.predict(X_test).reshape(len(y_test))\nprint(preds)\nMSE = np.mean((preds-y_test)**2)\nprint('MSE ', MSE)\n<\/code><\/pre>\n"}]}
{"QuestionId":36070587,"AnswerCount":1,"Tags":"<theano><deep-learning><lasagne>","CreationDate":"2016-03-17T20:01:51.877","AcceptedAnswerId":null,"OwnerUserId":6079200.0,"Title":"ReLu with a threshold in Lasagne","Body":"<p>I am building a simple convolutional network using the Lasagne package and wanted to add a ReLu layer with a simple threshold [max(0,x-threshold)] but could only find rectifiers without a trainable parameter (lasagne.layers.NonlinearityLayer) or that has a parameter that is being multiplied (lasagne.layers.ParametricRectifierLayer).  Does this layer exist or am I missing something obvious?<\/p>\n\n<p>Thank you for any help! Terry <\/p>\n","answers":[{"AnswerId":"36080227","CreationDate":"2016-03-18T09:06:34.777","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>I don't think that exists. With the reason being that you usually have a trainable layer before the relu (e.g. convolutional or fully connected), which has a bias included. Moving the data by some bias is equivalent to having a threshold at the relu.\nIf you don't have a trainable layer before the relu, you can also explicitly add a lasagne.layers.BiasLayer (<a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/modules\/layers\/special.html\" rel=\"nofollow\">http:\/\/lasagne.readthedocs.org\/en\/latest\/modules\/layers\/special.html<\/a>)<\/p>\n\n<p>Hope this helps<\/p>\n\n<p>Michael<\/p>\n"}]}
{"QuestionId":36070588,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-17T20:01:53.203","AcceptedAnswerId":36072131.0,"OwnerUserId":1056563.0,"Title":"Distributed tensorflow fails with \"BUILD file not found on package\"","Body":"<p>When attempting to build in the core\/distributed_runtime module using:<\/p>\n\n<pre><\/pre>\n\n<p>We get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Are there additional steps required (and not mentioned in the README.md) ?<\/p>\n","answers":[{"AnswerId":"36072131","CreationDate":"2016-03-17T21:35:51.383","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This sounds like a git submodule issue&mdash;and it would affect building <em>any<\/em> part of TensorFlow from source. To recover, run the following command in your git repository:<\/p>\n\n<pre><code>$ git submodule update --init --recursive\n<\/code><\/pre>\n\n<p>(There are many other ways to do the same thing: see <a href=\"https:\/\/stackoverflow.com\/questions\/1030169\/easy-way-pull-latest-of-all-submodules\">this question<\/a> for some suggestions.)<\/p>\n"}]}
{"QuestionId":36072672,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-17T22:15:32.230","AcceptedAnswerId":36075984.0,"OwnerUserId":4524141.0,"Title":"How to import word2vec into TensorFlow Seq2Seq model?","Body":"<p>I am playing with Tensorflow sequence to sequence translation model. I was wondering if I could import my own word2vec into this model? Rather than using its original 'dense representation' mentioned in the tutorial. <\/p>\n\n<p>From my point of view, it looks TensorFlow is using One-Hot representation for seq2seq model. Firstly,for function  the encoder's input is a tokenized symbol, e.g. 'a' would be '4' and 'dog' would be '15715' etc. and  requires num_encoder_symbols. So I think it makes me provide the position of the word and the total number of words, then the function could represent the word in One-Hot representation. I am still learning the source code, but it hard to understand.<\/p>\n\n<p>Could anyone give me an idea on above problem?<\/p>\n","answers":[{"AnswerId":"36199996","CreationDate":"2016-03-24T12:15:40.510","ParentId":null,"OwnerUserId":"5264769","Title":null,"Body":"<p>I guess with the scope style, which Matthew mentioned, you can get variable:<\/p>\n\n<pre><code> with tf.variable_scope(\"embedding_attention_seq2seq\"):\n        with tf.variable_scope(\"RNN\"):\n            with tf.variable_scope(\"EmbeddingWrapper\", reuse=True):\n                  embedding = vs.get_variable(\"embedding\", [shape], [trainable=])\n<\/code><\/pre>\n\n<p>Also, I would imagine you would want to inject embeddings into the decoder as well, the key (or scope) for it would be somthing like:<\/p>\n\n<blockquote>\n  <p>\"embedding_attention_seq2seq\/embedding_attention_decoder\/embedding\"<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Thanks for your answer, Lukasz! <\/p>\n\n<p>I was wondering, what exactly in the code snippet <code>&lt;b&gt;model.vocab[word]&lt;\/b&gt;<\/code> stands for? Just the position of the word in the vocabulary? <\/p>\n\n<p>In this case wouldn't that be faster to iterate through the vocabulary and inject w2v vectors for the words that exist in w2v model.<\/p>\n"},{"AnswerId":"36075984","CreationDate":"2016-03-18T04:10:23.963","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>The seq2seq embedding_* functions indeed create embedding matrices very similar to those from word2vec. They are a variable named sth like this:<\/p>\n\n<blockquote>\n  <p>EMBEDDING_KEY = \"embedding_attention_seq2seq\/RNN\/EmbeddingWrapper\/embedding\"<\/p>\n<\/blockquote>\n\n<p>Knowing this, you can just modify this variable. I mean -- get your word2vec vectors in some format, say a text file. Assuming you have your vocabulary in model.vocab you can then assign the read vectors in a way illustrated by the snippet below (it's just a snippet, you'll have to change it to make it work, but I hope it shows the idea).<\/p>\n\n<pre><code>   vectors_variable = [v for v in tf.trainable_variables()\n                        if EMBEDDING_KEY in v.name]\n    if len(vectors_variable) != 1:\n      print(\"Word vector variable not found or too many.\")\n      sys.exit(1)\n    vectors_variable = vectors_variable[0]\n    vectors = vectors_variable.eval()\n    print(\"Setting word vectors from %s\" % FLAGS.word_vector_file)\n    with gfile.GFile(FLAGS.word_vector_file, mode=\"r\") as f:\n      # Lines have format: dog 0.045123 -0.61323 0.413667 ...\n      for line in f:\n        line_parts = line.split()\n        # The first part is the word.\n        word = line_parts[0]\n        if word in model.vocab:\n          # Remaining parts are components of the vector.\n          word_vector = np.array(map(float, line_parts[1:]))\n          if len(word_vector) != vec_size:\n            print(\"Warn: Word '%s', Expecting vector size %d, found %d\"\n                     % (word, vec_size, len(word_vector)))\n          else:\n            vectors[model.vocab[word]] = word_vector\n    # Assign the modified vectors to the vectors_variable in the graph.\n    session.run([vectors_variable.initializer],\n                {vectors_variable.initializer.inputs[1]: vectors})\n<\/code><\/pre>\n"}]}
{"QuestionId":36073195,"AnswerCount":0,"Tags":"<lua><torch><lstm><cudnn>","CreationDate":"2016-03-17T22:58:46.070","AcceptedAnswerId":null,"OwnerUserId":6079717.0,"Title":"Saving lstm language model in Torch","Body":"<p>I am using the lstm language model in <a href=\"https:\/\/github.com\/wojzaremba\/lstm\/blob\/master\/main.lua\" rel=\"nofollow\">https:\/\/github.com\/wojzaremba\/lstm\/blob\/master\/main.lua<\/a><\/p>\n\n<p>I want to save the model at the end of training for later use. I added the following line at the end of training<\/p>\n\n<pre><\/pre>\n\n<p>Which seems to successfully save the model. However, when I try to load that model and test it, I get a very large perplexity. Just for testing, I ran a small training instance, which resulted in a test set perplexity of 134, then saved the model. I then loaded the saved model and applied exactly the same testing method (function run_test) on the same test set, but I got a huge perplexity of 71675.134 (even using random weights gives much lower perplexity than that!). I tried saving and loading only the weights, converting them to float() before saving, or saving them as cudaTensors, and all gave me the same result.<\/p>\n\n<p>Here is the code for loading and testing after saving the whole model; I only modified the main method from the original main.lua: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36075319,"AnswerCount":1,"Tags":"<python><image-processing><machine-learning><jpeg><tensorflow>","CreationDate":"2016-03-18T02:53:58.383","AcceptedAnswerId":36091574.0,"OwnerUserId":4333440.0,"Title":"How do you load, label, and feed jpeg data into Tensorflow?","Body":"<p>I have been trying to feed 1750 * 1750 images into Tensorflow, but I do not know how to label and feed the data after I convert the images into a Tensor using the tf.image.decode_jpeg() function.<\/p>\n\n<p>Currently, my code is:<\/p>\n\n<pre><\/pre>\n\n<p>Now, as I said earlier, I need to feed and label the data. I have seen the CIFAR-10 tutorial files, but they stored the labels in a file and I plan on not doing that way.<\/p>\n\n<p>I am quite new to Tensorflow so please keep the response as detailed as possible.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36091574","CreationDate":"2016-03-18T18:16:18.803","ParentId":null,"OwnerUserId":"5869947","Title":null,"Body":"<p>Depending on what you are trying to do, there are several directions to consider.<\/p>\n\n<ol>\n<li><p>If you just wish to run inference on an arbitrary JPEG file (i.e. labels are not required), then you can follow the example of classify_image.py which feeds in a JPEG image into a pre-trained Inception network:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/imagenet\/classify_image.py\" rel=\"nofollow noreferrer\">github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/imagenet\/classify_image.py<\/a><\/p><\/li>\n<li><p>If you do wish to train (or fine-tune) a model on a <strong>small<\/strong> custom data set of JPEG images, then take a look at this example for how to train a model off a small set of JPEG images.<\/p>\n\n<p><a href=\"http:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py<\/a><\/p><\/li>\n<li><p>If you do wish to train (or fine-tune) a model on a <strong>large<\/strong> custom data set of JPEG images, then reading many individual JPEG files will be inefficient and slow down training tremendously.<\/p><\/li>\n<\/ol>\n\n<p>I would suggest following the procedure of described in the inception\/ model library that converts a directory of JPEG images into sharded RecordIO containing serialized JPEG images.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/data\/build_image_data.py\" rel=\"nofollow noreferrer\">github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/data\/build_image_data.py<\/a><\/p>\n\n<p>Instructions for running the conversion script are available here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/README.md#how-to-construct-a-new-dataset-for-retraining\" rel=\"nofollow noreferrer\">github.com\/tensorflow\/models\/blob\/master\/research\/inception\/README.md#how-to-construct-a-new-dataset-for-retraining<\/a><\/p>\n\n<p>After running the conversion, you may then employ\/copy the image preprocessing pipeline used by the inception\/ model.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/image_processing.py\" rel=\"nofollow noreferrer\">github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/image_processing.py<\/a><\/p>\n"}]}
{"QuestionId":36078411,"AnswerCount":1,"Tags":"<tensorflow><cross-entropy>","CreationDate":"2016-03-18T07:21:58.460","AcceptedAnswerId":36086477.0,"OwnerUserId":3583384.0,"Title":"TensorFlow: Are my logits in the right format for cross entropy function?","Body":"<p>Alright, so I'm getting ready to run the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><\/a> function in Tensorflow.<\/p>\n\n<p>It's my understanding that the 'logits' should be a Tensor of probabilities, each one corresponding to a certain pixel's probability that it is part of an image that will ultimately be a \"dog\" or a \"truck\" or whatever... a finite number of things.<\/p>\n\n<p>These logits will get plugged into this cross entropy equation:\n<a href=\"https:\/\/i.stack.imgur.com\/e6RZx.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/e6RZx.png\" alt=\"Cross entropy formula from Wikipedia\"><\/a><\/p>\n\n<p>As I understand it, the logits are plugged into the right side of the equation.  That is, they are the q of every x (image).  If they were probabilities from 0 to 1... that would make sense to me.  But when I'm running my code and ending up with a tensor of logits, I'm not getting probabilities. Instead I'm getting floats that are both positive and negative:<\/p>\n\n<pre><\/pre>\n\n<p>So my question is... is that right?  Do I have to somehow calculate all my logits and turn them into probabilities from 0 to 1?  <\/p>\n","answers":[{"AnswerId":"36086477","CreationDate":"2016-03-18T14:05:12.543","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The crucial thing to note is that <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits(logits, labels)<\/code><\/a> performs an internal softmax on each row of <code>logits<\/code> so that they <em>are<\/em> interpretable as probabilities before they are fed to the cross entropy equation.<\/p>\n\n<p>Therefore, the \"<code>logits<\/code>\" need not be probabilities (or even true log probabilities, as the name would suggest), because of the internal normalization that happens within that op.<\/p>\n\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 the softmax may compute much larger values) and (ii) less efficient (since some redundant computation would happen in the backprop). For real uses, we recommend that you use <code>tf.nn.softmax_cross_entropy_with_logits()<\/code>.<\/p>\n"}]}
{"QuestionId":36080445,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-18T09:18:02.673","AcceptedAnswerId":36088396.0,"OwnerUserId":5650267.0,"Title":"TensorFlow in_top_k evaluation input argumants","Body":"<p>I am following the tutorial in <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html\" rel=\"noreferrer\">this link<\/a> and trying to change the evaluation method for the model (at the bottom). I would like to get a top-5 evaluation and I'm trying to use to following code:<\/p>\n\n<pre><\/pre>\n\n<p>However, this yields the following error:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I can tell, the problem is that  only works for  or  data, but my data is in  format. Is there any workaround for this?<\/p>\n","answers":[{"AnswerId":"36088396","CreationDate":"2016-03-18T15:33:13.407","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>targets<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/nn.html#in_top_k\" rel=\"noreferrer\"><code>tf.nn.in_top_k(predictions, targets, k)<\/code><\/a> must be a vector of class IDs (i.e. indices of columns in the <code>predictions<\/code> matrix). This means that it only works for single-class classification problems.<\/p>\n\n<p>If your problem is a single-class problem, then I assume that your <code>y_<\/code> tensor is a one-hot encoding of the true labels for your examples (for example because you also pass them to an op like <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits()<\/code><\/a>. In that case, you have two options:<\/p>\n\n<ul>\n<li>If the labels were originally stored as integer labels, pass them directly to <code>tf.nn.in_top_k()<\/code> without converting them to one-hot. (Also, consider using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/nn.html#sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a> as your loss function, because it may be more efficient.)<\/li>\n<li><p>If the labels were originally stored in the one-hot format, you can convert them to integers using <code>tf.argmax()<\/code>:<\/p>\n\n<pre><code>labels = tf.argmax(y_, 1)\ntopFiver = tf.nn.in_top_k(y, labels, 5)\n<\/code><\/pre><\/li>\n<\/ul>\n"}]}
{"QuestionId":36087816,"AnswerCount":1,"Tags":"<neural-network><classification><conv-neural-network><keras>","CreationDate":"2016-03-18T15:06:39.313","AcceptedAnswerId":null,"OwnerUserId":6077694.0,"Title":"Binary classification of images in Keras","Body":"<p>I am trying to classify images (29 x 29) using only 1 channel in Keras. If the middle pixel is in a certain range, then the output is 1, if not then it is 0.<\/p>\n\n<p>For the training set I have 10000 images with output 1 and 30000 images with output 0 and I am using the following architecture:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>For compilation, I am using the binary_crossentropy loss <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The accuracy fluctuates, but it reaches a final value of 0.75 after each epoch. \nIs there anything I am missing? Is there anything I should change or add in order for this to work? <\/p>\n","answers":[{"AnswerId":"36114466","CreationDate":"2016-03-20T13:40:40.227","ParentId":null,"OwnerUserId":"6089596","Title":null,"Body":"<p>Trying to interpret your approach, you want your CNN to only look at 1 of 29*29 pixels. I'm not an expert but this does not sound like a good problem for a CNN. However, usually if your network achieves bad results on the training data, your network is too small. So you could try to make it bigger. Another things you can think about: What do your settings (kernel size, layers, pooling) actually achieve? Could your settings be bad for this kind of task? Here is what I would do:<\/p>\n\n<ul>\n<li>try this without CNN, i.e. use 1 or 2 fully connected layer and see what happens<\/li>\n<li>make the resolution 'finer', i.e. use (3,3) filter size<\/li>\n<li>remove maxpooling (maxpooling throws away information to make features more invariant, I don't see that you would need this)<\/li>\n<li>add more conv layers<\/li>\n<li>might not be relevant, but your data set is very small so you don't need the speed benefit of sgd, go for another optimizer instead<\/li>\n<\/ul>\n\n<p>How exactly does your training and testing look like? Is this a toy project?<\/p>\n"}]}
{"QuestionId":36088277,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-18T15:27:58.560","AcceptedAnswerId":36088778.0,"OwnerUserId":234167.0,"Title":"How to select rows from a 3-D Tensor in TensorFlow?","Body":"<p>I have a tensor  with the dimensions  (i.e. each logit in the batch is a matrix). In my case batch size is 2, there's 4 rows and 4 coordinates.<\/p>\n\n<pre><\/pre>\n\n<p>I want to select the first and second row of the first batch and the second and fourth row of the second batch.<\/p>\n\n<pre><\/pre>\n\n<p>So the desired output would be<\/p>\n\n<pre><\/pre>\n\n<p>How do I do this using TensorFlow? I tried using  but it did not return what I expected. Thanks!<\/p>\n","answers":[{"AnswerId":"42691442","CreationDate":"2017-03-09T09:25:51.990","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>mrry's answer is great, but I think with the function <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a> the problem can be solved with much fewer lines of code (probably this function was not yet available at the time of mrry's writing):<\/p>\n\n<pre><code>logits = tf.constant([[[10.0, 10.0, 20.0, 20.0],\n                      [11.0, 10.0, 10.0, 30.0],\n                      [12.0, 10.0, 10.0, 20.0],\n                      [13.0, 10.0, 10.0, 20.0]],\n                     [[14.0, 11.0, 21.0, 31.0],\n                      [15.0, 11.0, 11.0, 21.0],\n                      [16.0, 11.0, 11.0, 21.0],\n                      [17.0, 11.0, 11.0, 21.0]]])\n\nindices = tf.constant([[[0, 0], [0, 1]], [[1, 1], [1, 3]]])\n\nresult = tf.gather_nd(logits, indices)\nwith tf.Session() as sess:\n    print(sess.run(result))\n<\/code><\/pre>\n\n<p>This will print<\/p>\n\n<pre><code>[[[ 10.  10.  20.  20.]\n  [ 11.  10.  10.  30.]]\n\n [[ 15.  11.  11.  21.]\n  [ 17.  11.  11.  21.]]]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a> should be available as of v0.10. Check out <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">this github issue<\/a> for more discussions on this.<\/p>\n"},{"AnswerId":"36088778","CreationDate":"2016-03-18T15:50:41.560","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is possible in TensorFlow, but slightly inconvenient, because <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather()<\/code><\/a> currently only works with one-dimensional indices, and only selects slices from the 0th dimension of a tensor. However, it is still possible to solve your problem efficiently, by transforming the arguments so that they can be passed to <code>tf.gather()<\/code>:<\/p>\n\n<pre><code>logits = ... # [2 x 4 x 4] tensor\nindices = tf.constant([[0, 1], [1, 3]])\n\n# Use tf.shape() to make this work with dynamic shapes.\nbatch_size = tf.shape(logits)[0]\nrows_per_batch = tf.shape(logits)[1]\nindices_per_batch = tf.shape(indices)[1]\n\n# Offset to add to each row in indices. We use `tf.expand_dims()` to make \n# this broadcast appropriately.\noffset = tf.expand_dims(tf.range(0, batch_size) * rows_per_batch, 1)\n\n# Convert indices and logits into appropriate form for `tf.gather()`. \nflattened_indices = tf.reshape(indices + offset, [-1])\nflattened_logits = tf.reshape(logits, tf.concat(0, [[-1], tf.shape(logits)[2:]]))\n\nselected_rows = tf.gather(flattened_logits, flattened_indices)\n\nresult = tf.reshape(selected_rows,\n                    tf.concat(0, [tf.pack([batch_size, indices_per_batch]),\n                                  tf.shape(logits)[2:]]))\n<\/code><\/pre>\n\n<p>Note that, since this uses <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\"><code>tf.reshape()<\/code><\/a> and not <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\"><code>tf.transpose()<\/code><\/a>, it doesn't need to modify the (potentially large) data in the <code>logits<\/code> tensor, so it should be fairly efficient.<\/p>\n"}]}
{"QuestionId":36088609,"AnswerCount":1,"Tags":"<python-2.7><theano><lasagne>","CreationDate":"2016-03-18T15:42:52.530","AcceptedAnswerId":36096764.0,"OwnerUserId":1995263.0,"Title":"python lasagne ImportError cannot import BatchNormLayer","Body":"<p>When I try to play with the code <a href=\"https:\/\/github.com\/larsmaaloee\/auxiliary-deep-generative-models\" rel=\"noreferrer\">here<\/a>, I met a very strange error. \nEvery other modules can be imported properly, except for one. <\/p>\n\n<p>Specifically, the error is:   <\/p>\n\n<pre><\/pre>\n\n<p>from the file <a href=\"https:\/\/github.com\/larsmaaloee\/auxiliary-deep-generative-models\/blob\/master\/models\/sdgmssl.py\" rel=\"noreferrer\">here<\/a>. And the lasagne_extensions.layers is as following:    <\/p>\n\n<pre><\/pre>\n\n<p>So, I believe the problem should be that I didn't install  or  properly. <\/p>\n\n<p>I have tried to upgrade the relevant modules including , ,  and  to newest version with  respectively. There is no  support for , so I downloaded it and installed it again.\nHowever, the error remains. <\/p>\n\n<p>Can anyone give me some advice about what I should look into? <\/p>\n","answers":[{"AnswerId":"36096764","CreationDate":"2016-03-19T01:20:58.960","ParentId":null,"OwnerUserId":"1995263","Title":null,"Body":"<p>The solution is to install the cutting-edge version, and <code>pip install --upgrade<\/code> does not do that. <\/p>\n\n<p>But the problem can be solved by:<\/p>\n\n<pre><code>pip install --upgrade https:\/\/github.com\/Theano\/Theano\/archive\/master.zip\npip install --upgrade https:\/\/github.com\/Lasagne\/Lasagne\/archive\/master.zip\n<\/code><\/pre>\n\n<p>Details are <a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/user\/installation.html#bleeding-edge-version\" rel=\"noreferrer\">here<\/a>. <\/p>\n"}]}
{"QuestionId":36088671,"AnswerCount":1,"Tags":"<python><pandas><regression><tensorflow><keras>","CreationDate":"2016-03-18T15:46:19.600","AcceptedAnswerId":36088938.0,"OwnerUserId":2390577.0,"Title":"How to prepare input data for a LSTM regression in keras?","Body":"<p>I have a dataset with 5K rows (-1K for validation) and 17 columns, including the last one (the target integer binary label).<\/p>\n\n<p>My model is simply this 2-layer LSTM:<\/p>\n\n<pre><\/pre>\n\n<p>After loading my dataset with pandas<\/p>\n\n<pre><\/pre>\n\n<p>and trying to run my model, I get this error: <\/p>\n\n<blockquote>\n  <p><strong>Exception<\/strong>: When using TensorFlow, you should define explicitly the number of timesteps of your sequences. - If your first layer is an Embedding, make sure to pass it an \"input_length\" argument. Otherwise, make sure the first layer has an \"input_shape\" or \"batch_input_shape\" argument, including the time axis.<\/p>\n<\/blockquote>\n\n<p>What should I put in ? The total rowcount?<\/p>\n\n<p>Since my dataframe has a shape as train_X=(4000, 17) train_y=(4000,) how can I prepare it to feed this kind of model? I have to change my input data shape?<\/p>\n\n<p>Thanks for any help!! (=<\/p>\n","answers":[{"AnswerId":"36088938","CreationDate":"2016-03-18T15:58:19.157","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like Keras uses the static unrolling approach to build recurrent networks (such as LSTMs) on TensorFlow. The <code>input_length<\/code> should be the length of the longest sequence that you want to train: so if each row of your CSV file <code>train_file<\/code> is a comma-delimited sequence of symbols, it should be the number of symbols in the longest row.<\/p>\n"}]}
{"QuestionId":36089868,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-03-18T16:45:14.520","AcceptedAnswerId":36094131.0,"OwnerUserId":3449534.0,"Title":"Torch, how to select a subset of a tensor?","Body":"<p>For example, I have a tensor . And I want to select the tensor subset of at the rows of  and concatenate them as new tensor. Is there any elegant way? Thank you.<\/p>\n","answers":[{"AnswerId":"36094131","CreationDate":"2016-03-18T21:00:41.907","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>There's <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#tensor-indexdim-index\" rel=\"nofollow\"><code>:index<\/code><\/a> method:<\/p>\n\n<pre><code>tens = torch.rand(10,1)\nind = torch.LongTensor{1,3,7,9}\nsubset = tens:index(1, ind)\n<\/code><\/pre>\n"}]}
{"QuestionId":36091256,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-18T17:58:31.060","AcceptedAnswerId":null,"OwnerUserId":4328698.0,"Title":"Applying tf.abs() on multidimensional tensor (tensorflow)","Body":"<p>This should be super simple but I cannot find the right way to do it:<\/p>\n\n<p>I have a tensor with shape (1, 1, 1024, 66) and I want to apply the operation tf.abs() on all values.\nDoing tf.abs(tensor) gives the error:<\/p>\n\n<pre><\/pre>\n\n<p>tf.abs() does not have a parameter to specify the dimension.<\/p>\n\n<p>How can I do this?<\/p>\n","answers":[{"AnswerId":"36091401","CreationDate":"2016-03-18T18:06:35.247","ParentId":null,"OwnerUserId":"4328698","Title":null,"Body":"<p>The problem was actually a different one. My tensor was not shape (1,1,1024,66), but the tensor was shape (1,1024,66) inside of a python list with length 1.<\/p>\n\n<p>Doing <code>tf.abs(tensor[0])<\/code> did the job.<\/p>\n"}]}
{"QuestionId":36094601,"AnswerCount":1,"Tags":"<python><python-2.7><gpu><theano><keras>","CreationDate":"2016-03-18T21:32:31.677","AcceptedAnswerId":36100128.0,"OwnerUserId":4802797.0,"Title":"Keras: CuDNN not available?","Body":"<p>I am running the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/kaggle_otto_nn.py\" rel=\"nofollow\">example<\/a>  of  with backend of . <\/p>\n\n<p>In the following print out, line 5, something says: <\/p>\n\n<blockquote>\n  <p>CNMeM is enabled with initial size: 90.0% of memory, CuDNN not available<\/p>\n<\/blockquote>\n\n<p>I am wondering, does this  matter since the GPU device is detectable? Am I running my program on GPU correctly? Or it's actually not running on GPU.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36100128","CreationDate":"2016-03-19T09:38:22.710","ParentId":null,"OwnerUserId":"2370124","Title":null,"Body":"<p>cuDNN is a library by NVidia which increases the performance of neural networks on GPU. So your program still runs on GPU but much slower than it could have been if you had installed cuDNN.<\/p>\n"}]}
{"QuestionId":36096386,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-19T00:23:20.980","AcceptedAnswerId":36097678.0,"OwnerUserId":2063031.0,"Title":"Tensorflow `set_random_seed` not working","Body":"<h3>Calling  has no effect that I can tell...<\/h3>\n\n<p>For example, running the code below several times <em>inside an IPython notebook<\/em> produces different output each time:<\/p>\n\n<pre><\/pre>\n\n<p>If I set the seed explicitly: , the output is the same after each run. But why do I need to set the seed if I already call ?<\/p>\n\n<hr>\n\n<p>The equivalent code using numpy just works:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36097678","CreationDate":"2016-03-19T04:05:48.460","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>That only sets the graph-level random seed. If you execute this snippet several times in a row, the graph will change, and two shuffle statements will get different operation-level seeds. The details are described in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/random_seed.py#L65\" rel=\"noreferrer\">doc string<\/a> for <code>set_random_seed<\/code><\/p>\n\n<p>To get deterministic <code>a_shuf<\/code> you can either<\/p>\n\n<ol>\n<li>Call <code>tf.reset_default_graph()<\/code> between invocations or<\/li>\n<li>Set operation-level seed for shuffle: <code>a_shuf = tf.random_shuffle(a, seed=42)<\/code> <\/li>\n<\/ol>\n"}]}
{"QuestionId":36098278,"AnswerCount":1,"Tags":"<plot><torch>","CreationDate":"2016-03-19T05:37:13.057","AcceptedAnswerId":null,"OwnerUserId":1217329.0,"Title":"Changing axis endpoints in iTorch plot","Body":"<p>I'm trying to graph some data in an iTorch notebook. I can generate plots fine, but I want to change the endpoints of the axes. (My autogenerated y-axis is from 20-100, but I'd rather it be from 0-100 since it's a graph of percentages and I want the lower left corner to be the origin.)<\/p>\n\n<p>I looked in the documentation, and in the list of methods implemented, and in the source code, but didn't find anything that lets me do this. I can zoom the generated graph, but it preserves aspect ratio, so I can't zoom just one of the axes.<\/p>\n\n<p>Does anyone know how to do this? I'm half convinced this isn't implemented, but it seems like a very strange feature to leave out.<\/p>\n","answers":[{"AnswerId":"36118667","CreationDate":"2016-03-20T19:47:41.003","ParentId":null,"OwnerUserId":"2775876","Title":null,"Body":"<p>You can change the axis scale by mouse-scrolling over the axis ticks, and translate by drag-and-drop.<\/p>\n\n<p>I know it's not ideal, but I had the same problem and that's the best I could find so far.<\/p>\n"}]}
{"QuestionId":36098947,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-19T07:11:38.657","AcceptedAnswerId":36108627.0,"OwnerUserId":2260767.0,"Title":"How does Tensorflow support Cuda streams?","Body":"<p>Does Tensorflow utilize Cuda streams automatically for concurrent execution of the computation graph on a single GPU or should streams be assigned manually to ops\/tensors ?<\/p>\n","answers":[{"AnswerId":"36108627","CreationDate":"2016-03-19T23:43:54.960","ParentId":null,"OwnerUserId":"6088060","Title":null,"Body":"<p>For now, TensorFlow only uses one compute stream, and multiple copy streams. Some kernels may choose to use multiple streams for computation, while maintaining a single-stream semantics. <\/p>\n\n<p>Our experiment showed that enabling multi-stream automatically does not bring much performance gains, since most of our kernels are large enough to utilize all processors in GPU. But enabling multi-stream would disable our current design to recycle GPU memory aggressively. <\/p>\n\n<p>This is a decision we might revisit in the future. If that happens, it is likely for TensorFlow to automatically assign ops\/kernels to different Cuda streams, without exposing them to users.<\/p>\n"}]}
{"QuestionId":36099555,"AnswerCount":1,"Tags":"<amazon-web-services><amazon-ec2><machine-learning><amazon-ami><torch>","CreationDate":"2016-03-19T08:27:43.823","AcceptedAnswerId":null,"OwnerUserId":3863931.0,"Title":"Using Torch with neural-style arts in AWS linux AMI, error loading module 'libjpeg' and <image.loadJPG> libjpeg package not found","Body":"<p>As title, I use the AWS EC2 with AWS linux AMI to set the environment run the neural-style algorithm.<\/p>\n\n<p><a href=\"https:\/\/github.com\/jcjohnson\/neural-style\" rel=\"nofollow\">Nerual-style<\/a><\/p>\n\n<p>I have install torch, Lua, Loadcaffe, Google protocal buffer and neural-style package succesfully.<\/p>\n\n<p>When I try run the testing instruction :<\/p>\n\n<pre><\/pre>\n\n<p>I got this :<\/p>\n\n<pre><\/pre>\n\n<p>How can I correctly load the libjpeg package ???<\/p>\n","answers":[{"AnswerId":"36211221","CreationDate":"2016-03-24T22:43:28.987","ParentId":null,"OwnerUserId":"4917583","Title":null,"Body":"<p>You are missing libjpeg. SSH to your AWS instance and run \n<code>sudo apt-get install libjpeg-dev<\/code>.<\/p>\n"}]}
{"QuestionId":36099918,"AnswerCount":1,"Tags":"<python><python-2.7><memory><gpu><theano>","CreationDate":"2016-03-19T09:11:52.103","AcceptedAnswerId":36269522.0,"OwnerUserId":4802797.0,"Title":"Theano: Initialisation of device gpu failed! Reason=CNMEM_STATUS_OUT_OF_MEMORY","Body":"<p>I am running the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/kaggle_otto_nn.py\">example<\/a>  of  with backend of . When I set , the following error comes out: <\/p>\n\n<pre><\/pre>\n\n<p>It seems like I cannot set the  to a very large value (about > 0.9) since it may cause the GPU's memory overflow. And when I set , it's is working correctly.  According to <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.config.lib.cnmem\">this<\/a>, it <\/p>\n\n<blockquote>\n  <p>represents the start size (in MB or % of total GPU memory) of the memory pool.<\/p>\n<\/blockquote>\n\n<p>And<\/p>\n\n<blockquote>\n  <p>This could cause memory fragmentation. So if you have a memory error while using cnmem, try to allocate more memory at the start or disable it. If you try this, report your result on :ref.<\/p>\n<\/blockquote>\n\n<p>But if I got memory error, why should I allocate more memory at the start\uff1f And in my case, allocating more memory at the start seems like causing the error. <\/p>\n","answers":[{"AnswerId":"36269522","CreationDate":"2016-03-28T19:21:44.473","ParentId":null,"OwnerUserId":"4802797","Title":null,"Body":"<p>This is question is solved based on <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/4302\" rel=\"noreferrer\">this<\/a>. <\/p>\n\n<p>As indicated <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html#config.config.lib.cnmem\" rel=\"noreferrer\">here<\/a>, the <code>cnmem<\/code> is actually only allowed to be assigned as a <code>float<\/code>. <\/p>\n\n<blockquote>\n  <p>0: not enabled.<\/p>\n  \n  <p>0 &lt; N &lt;= 1: use this fraction of the total GPU memory (clipped to .95 for driver memory).<\/p>\n  \n  <p>> 1: use this number in megabytes (MB) of memory.<\/p>\n<\/blockquote>\n\n<p>So it will be working if <code>cnmem=1.0<\/code> instead of <code>cnmem=1<\/code>.<\/p>\n"}]}
{"QuestionId":36101191,"AnswerCount":0,"Tags":"<python><numpy><theano><matrix-inverse>","CreationDate":"2016-03-19T11:36:37.420","AcceptedAnswerId":null,"OwnerUserId":2451504.0,"Title":"Hessian matrix always singular in Theano","Body":"<p>I'm trying to implement training an MLP with Newton's method in Theano. I've managed to find out how to obtain the inverse hessian (it's pretty easy, just needed to remove some s in the  module). However, when I try to update my weight vectors with my  function, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Obviously, if my hessian is singular, it can't be inverted. But the weights are random, so how come it's always singular? I don't know much about matrix computation, but it seems to me that there must be a flaw in my process. Here's the relevant part of my code:<\/p>\n\n<pre><\/pre>\n\n<p>Would really appreciate any pointers as to what I'm doing wrong.<\/p>\n","answers":[]}
{"QuestionId":36101402,"AnswerCount":1,"Tags":"<python><theano><elementwise-operations>","CreationDate":"2016-03-19T11:56:36.900","AcceptedAnswerId":null,"OwnerUserId":3368154.0,"Title":"elementwise square of theano matrix","Body":"<p>I have a theano covariance matrix, and i am trying to calculate element wise square of it. I have written following code for same:<\/p>\n\n<pre><\/pre>\n\n<p>where cov is covariance matrix, calculated as:<\/p>\n\n<pre><\/pre>\n\n<p>But it is throwing following error:<\/p>\n\n<pre><\/pre>\n\n<p>I know it is trivial to do, but still could not find possible fix. Please help me in this regard.<\/p>\n","answers":[{"AnswerId":"36101912","CreationDate":"2016-03-19T12:45:41.427","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>Your input to the Theano function you compiled cannot be a symbolic expression, it has to be a NumPy array or a shared variable. For example:<\/p>\n\n<pre><code>A = T.matrix('input matrix')\nB = T.matrix('dummy matrix')\nC = np.random.rand(5,5).astype(theano.config.floatX)\nsquared = A**2 \nget_squared = theano.function([A], squared)\n<\/code><\/pre>\n\n<p>If I run the following command:<\/p>\n\n<pre><code>get_squared(B)\n<\/code><\/pre>\n\n<p>I will get the following error:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name \":1\"  at index 0(0-based)', 'Expected an array-like object, but found a Variable: maybe you are trying to call a function on a (possibly shared) variable instead of a numeric array?')<\/p>\n<\/blockquote>\n\n<p>However if I run:<\/p>\n\n<pre><code>get_squared(C)\n<\/code><\/pre>\n\n<p>I get the squared matrix back. <\/p>\n\n<p>I'm not sure about your codebase, how it's structured, but one very direct (possibly naive but it will work) solution is to create a symbolic expression for your squared covariance matrix and return that as part of your function. For example, if y1 and y2 are part of the graph to compute cov, you can create a theano function which returns the covariance squared:<\/p>\n\n<pre><code>cov = ... # (some expressions involving y1 and y2 as in your original post)\nget_cov_squared = theano.function([y1,y2], cov**2)\n<\/code><\/pre>\n\n<p>But again, your input to the function must be actual arrays or shared variables, not symbolic expressions. <\/p>\n"}]}
{"QuestionId":36103955,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2016-03-19T16:07:20.937","AcceptedAnswerId":36152358.0,"OwnerUserId":4343920.0,"Title":"How to reconstruct a caffemodel by some parameters?","Body":"<p>I have met a problem that I have a sparse-matrix-stored version neural network which means I have only some parameters and its index like which layer does it belong and which location does it belong? How can I use these information and a  to reconstruct the  by python.Thanks<\/p>\n","answers":[{"AnswerId":"36152358","CreationDate":"2016-03-22T10:49:07.760","ParentId":null,"OwnerUserId":"809993","Title":null,"Body":"<p>The caffemodel binary file is just a binary proto-buffer file, which you can create\/edit with the proto python interfrace module (caffe.proto.caffe_pb2). You should use the helper methods of caffe.io to convert numpy arrays to the array format acceptable by the protobuffer interface (or from the protobuffer to numpy). \nAlternatively, you could use the caffe.Net python class that wraps most these methods for easier access -- you can look at <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">this example from Caffe's homepage<\/a> for a more detailed tutorial.<\/p>\n\n<p>For simplicity, I would use the prototxt to train a dummy network to create a baseline caffemodel structure, and then just edit the weights of the layers using your own external weights using one of the above methods.<\/p>\n"}]}
{"QuestionId":36105763,"AnswerCount":1,"Tags":"<python><tensorflow><training-data><prefetch>","CreationDate":"2016-03-19T18:52:49.037","AcceptedAnswerId":36106559.0,"OwnerUserId":1104696.0,"Title":"How to handle different queue batch size and feed value batch size in tensorflow?","Body":"<p>My code used to work on tensorflow 0.6, but it no longer works on the lastest tensorflow.<\/p>\n\n<p>I would like to perform inference every few training iterations. My training data is pulled from a queue, my inference data is from feed_dict. The training batch size is 128 while the inference batch size is 1. What Should I do to make the network accept the two different batch sizes?<\/p>\n\n<pre><\/pre>\n\n<p>I got the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36106559","CreationDate":"2016-03-19T20:06:56.067","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>We added this check recently to prevent errors (and add a few optimization opportunities). You can make your program work again by changing the declaration of <code>x<\/code> to use the new <code>tf.placeholder_with_default()<\/code> op:<\/p>\n\n<pre><code>x = tf.placeholder_with_default(q.dequeue(), shape=[None, 100])\n<\/code><\/pre>\n"}]}
{"QuestionId":36108120,"AnswerCount":2,"Tags":"<caffe><lmdb>","CreationDate":"2016-03-19T22:45:04.950","AcceptedAnswerId":null,"OwnerUserId":4358739.0,"Title":"Shuffle data in LMDB file","Body":"<p>I already have an existing <a href=\"http:\/\/symas.com\/mdb\/\" rel=\"nofollow\">LMDB<\/a>(Symas Lightning Memory-Mapped Database) file which is created for <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow\">caffe<\/a>. Is there any possible way to shuffle the data in already existing LMDB to create a new LMDB with data shuffled. Any suggestions or ideas would be helpful.<\/p>\n","answers":[{"AnswerId":"42601748","CreationDate":"2017-03-04T21:42:08.870","ParentId":null,"OwnerUserId":"5867750","Title":null,"Body":"<p>To add to the answer given by @Manolo, while creating the LMDB dataset as suggested <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow noreferrer\">here<\/a>, I concatenated a random int to the beginning as follows:<\/p>\n\n<pre><code>random.seed(i)\nstr_id = '{:05}'.format(random.randint(1,70000))+'{:05}'.format(i)\n<\/code><\/pre>\n\n<p>I chose 70000, since my LMDB had about 72000 images. <\/p>\n"},{"AnswerId":"37021645","CreationDate":"2016-05-04T07:51:35.283","ParentId":null,"OwnerUserId":"159836","Title":null,"Body":"<p>LMDB traverses data according to the lexicological order of the key. You can prepend a random number to your current key and the data will be shuffled accordingly. I am also investigating an efficient way to rewrite the keys randomly between epochs as I would like to use Batch normalization in my data set.<\/p>\n"}]}
{"QuestionId":36110834,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-20T06:22:36.277","AcceptedAnswerId":36113230.0,"OwnerUserId":3251207.0,"Title":"What's difference between tf.sub and just minus operation in tensorflow?","Body":"<p>I am trying to use Tensorflow. Here is an very simple code.<\/p>\n\n<pre><\/pre>\n\n<p>Just ignore the optimization part (4th line). It will take a floating number and train W1 so as to increase squared difference.<\/p>\n\n<p>My question is simple. If I use just minus sign instead of \ntf.sub\" as below, what is different? Will it cause a wrong result? <\/p>\n\n<pre><\/pre>\n\n<p>When I replace it, the result looks the same. If they are the same, why do we need to use the \"tf.add\/tf.sub\" things?<\/p>\n\n<p>Built-in back propagation calculation can be done only by the \"tf.*\" things? <\/p>\n","answers":[{"AnswerId":"45089398","CreationDate":"2017-07-13T19:24:55.137","ParentId":null,"OwnerUserId":"2184122","Title":null,"Body":"<p>(tf.sub appears to have been replaced with tf.subtract)<\/p>\n\n<p>The only advantage I see is that you can specify a name of the operation as in:<\/p>\n\n<pre><code>tf.subtract(train, W1, name='foofoo')\n<\/code><\/pre>\n\n<p>This helps identify the operation causing an error as the name you provide is also shown:<\/p>\n\n<pre><code>ValueError: Dimensions must be equal, but are 28 and 40 for 'foofoo' (op: 'Sub') with input shapes\n<\/code><\/pre>\n\n<p>it may also help with TensorBoard understanding.  It might be overkill for most people as python also shows the line number which triggered the error.<\/p>\n"},{"AnswerId":"36113230","CreationDate":"2016-03-20T11:40:10.093","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>Yes, - and + resolve to tf.sub ad tf.add. If you look at the tensorflow code you will see that these operators on tf.Variable are overloaded with the tf.* methods.<\/p>\n\n<p>As to why both exists I assume the tf.* ones exist for consistency. So sub and say matmul operation can be used in the same way. While the operator overloading is for convenience. <\/p>\n"}]}
{"QuestionId":36112192,"AnswerCount":2,"Tags":"<image-processing><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-03-20T09:40:06.750","AcceptedAnswerId":41274924.0,"OwnerUserId":1326644.0,"Title":"Are 16 bit images supported by Caffe? If not, how to implement support?","Body":"<p><strong>Background information<\/strong>: I need to load some 16 bit grayscale PNGs.<\/p>\n\n<p>Does Caffe support loading 16 bit images through the ?<\/p>\n\n<p>After some googling, the answer seems it doesn't.\nThe <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/image_data_layer.cpp\" rel=\"nofollow\"><\/a> relies on <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/util\/io.cpp\" rel=\"nofollow\">this io routine<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Which uses opencv's  function. This function will read the input as 8bits unless the appropiate flag is set<\/p>\n\n<blockquote>\n  <p>CV_LOAD_IMAGE_ANYDEPTH - If set, return 16-bit\/32-bit image when the\n  input has the corresponding depth, otherwise convert it to 8-bit.<\/p>\n<\/blockquote>\n\n<p>Simply adding the appropriate flag will not work because later in the code [io.cpp] they check for 8bit depth:<\/p>\n\n<pre><\/pre>\n\n<p>I could just remove the check but I'm afraid it's there for a reason and unpredictable results might happen. \nCan anybody shine light on this issue?<\/p>\n","answers":[{"AnswerId":"36113085","CreationDate":"2016-03-20T11:24:20.940","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe works with <code>float32<\/code> variables, by default. An image is usually represented as a <code>C<\/code>-by-<code>H<\/code>-by-<code>W<\/code> blob, where <code>C=3<\/code> for three color channels. So, working with three channels of type <code>float32<\/code> allows you to deal with images in <code>uint16<\/code> provided you convert properly to <code>float32<\/code>.<\/p>\n\n<p>I do not have personal experience with <code>\"ImageData\"<\/code> layer, so I cannot comment on how you can or cannot load <code>uint16<\/code> image data using this layer.<\/p>\n\n<p><strong>However<\/strong>, you might find <code>\"HDF5Data\"<\/code> layer useful: you can externally read and convert your images to <code>hdf5<\/code> data format (that supports <code>float32<\/code>) and then feed the converted data to caffe via <code>\"HDF5Data\"<\/code> layer.<\/p>\n\n<p>You can find more information on <code>\"HDF5Data\"<\/code> layer <a href=\"https:\/\/stackoverflow.com\/a\/31634295\/1714410\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/a\/34261942\/1714410\">here<\/a>.<\/p>\n"},{"AnswerId":"41274924","CreationDate":"2016-12-22T02:27:41.637","ParentId":null,"OwnerUserId":"1923364","Title":null,"Body":"<p>You can patch ImageDataLayer to read 16bit images like this:<\/p>\n\n<ol>\n<li>Add appropriate flag as you mentioned (io.cpp):<\/li>\n<\/ol>\n\n<p>after<\/p>\n\n<pre><code>int cv_read_flag = (is_color ? CV_LOAD_IMAGE_COLOR :\n    CV_LOAD_IMAGE_GRAYSCALE);\n<\/code><\/pre>\n\n<p>add<\/p>\n\n<pre><code>cv_read_flag |= CV_LOAD_IMAGE_ANYDEPTH;\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Modify the check you mentioned (data_transformer.cpp):<\/li>\n<\/ol>\n\n<p>this<\/p>\n\n<pre><code>CHECK(cv_img.depth() == CV_8U) &lt;&lt; \"Image data type must be unsigned byte\";\n<\/code><\/pre>\n\n<p>becomes<\/p>\n\n<pre><code>CHECK(cv_img.depth() == CV_8U || cv_img.depth() == CV_16U) &lt;&lt; \"Image data type must be uint8 or uint16\";\nbool is16bit = cv_img.depth() == CV_16U;\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>Modify the way DataTransformer reads cv::Mat like this (same function below):<\/li>\n<\/ol>\n\n<p>add pointer of uint16_t type to:<\/p>\n\n<pre><code>const uchar* ptr = cv_cropped_img.ptr&lt;uchar&gt;(h);\n<\/code><\/pre>\n\n<p>like this<\/p>\n\n<pre><code>const uint16_t* ptr_16 = cv_cropped_img.ptr&lt;uint16_t&gt;(h);\n<\/code><\/pre>\n\n<p>Then read using appropriate pointer:<\/p>\n\n<pre><code>Dtype pixel = static_cast&lt;Dtype&gt;(ptr[img_index++]);\n<\/code><\/pre>\n\n<p>becomes<\/p>\n\n<pre><code>Dtype pixel;\nif(is16bit)\n    pixel = static_cast&lt;Dtype&gt;(ptr_16[img_index++]);\nelse\n    pixel = static_cast&lt;Dtype&gt;(ptr[img_index++]);\n<\/code><\/pre>\n"}]}
{"QuestionId":36113090,"AnswerCount":8,"Tags":"<tensorflow>","CreationDate":"2016-03-20T11:25:09.933","AcceptedAnswerId":36117853.0,"OwnerUserId":2245776.0,"Title":"How to get the global_step when restoring checkpoints in Tensorflow?","Body":"<p>I'm saving my session state like so:<\/p>\n\n<pre><\/pre>\n\n<p>When I later restore I want to get the value of the global_step for the checkpoint I restore from. This is in order to set some hyper parameters from it.<\/p>\n\n<p>The hacky way to do this would be to run through and parse the file names in the checkpoint directory. But surly there has to be a better, built in way to do this?<\/p>\n","answers":[{"AnswerId":"39337109","CreationDate":"2016-09-05T20:11:24.197","ParentId":null,"OwnerUserId":"3649117","Title":null,"Body":"<p>The current 0.10rc0 version seems to be different, there's no tf.saver() any more.  Now it's tf.train.Saver().  Also, the save command adds info onto save_path filename for the global_step, so we can't just call restore on the same save_path since that not the actual save file.<\/p>\n\n<p>The easiest way I see right now is to use the SessionManager along with a saver like this:<\/p>\n\n<pre><code>my_checkpoint_dir = \"\/tmp\/checkpoint_dir\"\n# make a saver to use with SessionManager for restoring\nsaver = tf.train.Saver()\n# Build an initialization operation to run below.\ninit = tf.initialize_all_variables()\n# use a SessionManager to help with automatic variable restoration\nsm = tf.train.SessionManager()\n# try to find the latest checkpoint in my_checkpoint_dir, then create a session with that restored\n# if no such checkpoint, then call the init_op after creating a new session\nsess = sm.prepare_session(\"\", init_op=init, saver=saver, checkpoint_dir=my_checkpoint_dir))\n<\/code><\/pre>\n\n<p>That's it.  Now you have a session that's either restored from the my_checkpoint_dir (make sure that directory exists before calling this), or if there's no checkpoint there then it creates a new session and calls the init_op to initialize your variables.<\/p>\n\n<p>When you want to save, you just save to any name you want in that directory and pass the global_step in.  Here's an example where I save the step variable in a loop as the global_step, so it comes back to that point if you kill the program and restart it so it restores the checkpoint:<\/p>\n\n<pre><code>checkpoint_path = os.path.join(my_checkpoint_dir, 'model.ckpt')\nsaver.save(sess, checkpoint_path, global_step=step)\n<\/code><\/pre>\n\n<p>This creates files in my_checkpoint_dir like \"model.ckpt-1000\" where 1000 is the global_step passed in.  If it keeps running, then you get more like \"model.ckpt-2000\".  The SessionManager above picks up the latest one of these when the program is restarted.  The checkpoint_path can be whatever file name you want, as long as it's in the checkpoint_dir.  The save() will create that file with the global_step appended (as shown above).  It also creates a \"checkpoint\" index file, which is how the SessionManager then finds the latest save checkpoint.<\/p>\n"},{"AnswerId":"36117853","CreationDate":"2016-03-20T18:35:34.463","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>General pattern is to have a <code>global_step<\/code> variable to keep track of steps<\/p>\n\n<pre><code>global_step = tf.Variable(0, name='global_step', trainable=False)\ntrain_op = optimizer.minimize(loss, global_step=global_step)\n<\/code><\/pre>\n\n<p>Then you can save with<\/p>\n\n<pre><code>saver.save(sess, save_path, global_step=global_step)\n<\/code><\/pre>\n\n<p>When you restore, the value of <code>global_step<\/code> is restored as well<\/p>\n"},{"AnswerId":"43690580","CreationDate":"2017-04-29T00:25:59.943","ParentId":null,"OwnerUserId":"5092283","Title":null,"Body":"<p>This is a bit of a hack, but the other answers did not work for me at all<\/p>\n\n<pre><code>ckpt = tf.train.get_checkpoint_state(checkpoint_dir) \n\n#Extract from checkpoint filename\nstep = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])\n<\/code><\/pre>\n\n<p><strong>Update 9\/2017<\/strong><\/p>\n\n<p>I'm not sure if this started working due to updates, but the following method seems to be effective in getting global_step to update and load properly:<\/p>\n\n<p>Create two ops. One to hold global_step and another to increment it:<\/p>\n\n<pre><code>    global_step = tf.Variable(0, trainable=False, name='global_step')\n    increment_global_step = tf.assign_add(global_step,1,\n                                            name = 'increment_global_step')\n<\/code><\/pre>\n\n<p>Now in your training loop run the increment op every time you run your training op.<\/p>\n\n<pre><code>sess.run([train_op,increment_global_step],feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>If you ever want to retrieve you global step value as an integer at any point, just use the following command after loading the model: <\/p>\n\n<pre><code>sess.run(global_step)\n<\/code><\/pre>\n\n<p>This can be useful for creating filenames or calculating what your current epoch is without having a second tensorflow Variable for holding that value. For instance, calculating the current epoch on loading would be something like:<\/p>\n\n<pre><code>loaded_epoch = sess.run(global_step)\/\/(batch_size*num_train_records)\n<\/code><\/pre>\n"},{"AnswerId":"44742186","CreationDate":"2017-06-25T00:42:04.557","ParentId":null,"OwnerUserId":"889617","Title":null,"Body":"<p>I had the same issue as Lawrence Du, I could not find a way to get the global_step by restoring the model. So I applied <a href=\"https:\/\/stackoverflow.com\/a\/43690580\/889617\">his hack<\/a> to <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/inception\/inception\/inception_train.py\" rel=\"nofollow noreferrer\">the inception v3 training code in the Tensorflow\/models github repo<\/a> I'm using. The code below also contains a fix related to the <code>pretrained_model_checkpoint_path<\/code>.<\/p>\n\n<p>If you have a better solution, or know what I'm missing please leave a comment!<\/p>\n\n<p>In any case, this code works for me:<\/p>\n\n<pre><code>...\n\n# When not restoring start at 0\nlast_step = 0\nif FLAGS.pretrained_model_checkpoint_path:\n    # A model consists of three files, use the base name of the model in\n    # the checkpoint path. E.g. my-model-path\/model.ckpt-291500\n    #\n    # Because we need to give the base name you can't assert (will always fail)\n    # assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)\n\n    variables_to_restore = tf.get_collection(\n        slim.variables.VARIABLES_TO_RESTORE)\n    restorer = tf.train.Saver(variables_to_restore)\n    restorer.restore(sess, FLAGS.pretrained_model_checkpoint_path)\n    print('%s: Pre-trained model restored from %s' %\n          (datetime.now(), FLAGS.pretrained_model_checkpoint_path))\n\n    # HACK : global step is not restored for some unknown reason\n    last_step = int(os.path.basename(FLAGS.pretrained_model_checkpoint_path).split('-')[1])\n\n    # assign to global step\n    sess.run(global_step.assign(last_step))\n\n...\n\nfor step in range(last_step + 1, FLAGS.max_steps):\n\n  ...\n<\/code><\/pre>\n"},{"AnswerId":"51519865","CreationDate":"2018-07-25T13:16:24.947","ParentId":null,"OwnerUserId":"9674919","Title":null,"Body":"<p>You can use the <code>global_step<\/code> variable to keep track of steps, but if in your code, you are initializing or assigning this value to another <code>step<\/code> variable, it may not be consistent.<\/p>\n\n<p>For instance, you define your <code>global_step<\/code> using:<\/p>\n\n<pre><code>global_step = tf.Variable(0, name='global_step', trainable=False)\n<\/code><\/pre>\n\n<p>Assign to your training operation:<\/p>\n\n<pre><code>train_op = optimizer.minimize(loss, global_step=global_step)\n<\/code><\/pre>\n\n<p>Save in your checkpoint:<\/p>\n\n<pre><code>saver.save(sess, checkpoint_path, global_step=global_step)\n<\/code><\/pre>\n\n<p>And restore from your checkpoint:<\/p>\n\n<pre><code>saver.restore(sess, checkpoint_path) \n<\/code><\/pre>\n\n<p>the value of <code>global_step<\/code> is restored as well but if you are assigning it to another variable, say <code>step<\/code>, then you must do something like:<\/p>\n\n<pre><code>step = global_step.eval(session=sess)\n<\/code><\/pre>\n\n<p>The variable <code>step<\/code>, contains the last saved <code>global_step<\/code> in the checkpoint.<\/p>\n\n<p>It will be nice to also define the global_step from graph than as zero variable (as earlier defined):<\/p>\n\n<pre><code>global_step = tf.train.get_or_create_global_step()\n<\/code><\/pre>\n\n<p>This will get your last <code>global_step<\/code> if exist or create one if not.<\/p>\n"},{"AnswerId":"53865486","CreationDate":"2018-12-20T09:09:55.360","ParentId":null,"OwnerUserId":"1210272","Title":null,"Body":"<p><\/p>\n\n<h1>TL;DR<\/h1>\n\n<p>As tensorflow variable (will be evaluated in the session)<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>global_step = tf.train.get_or_create_global_step()\n# use global_step variable to calculate your hyperparameter \n# this variable will be evaluated later in the session\nsaver = tf.train.Saver()\nwith tf.Session() as sess:\n    # restore all variables from checkpoint\n    saver.restore(sess, checkpoint_path)\n    # than init table and local variables and start training\/evaluation ...\n<\/code><\/pre>\n\n<p><strong>Or:<\/strong> As numpy integer (without any session):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>reader = tf.train.NewCheckpointReader(absolute_checkpoint_path)\nglobal_step = reader.get_tensor('global_step')\n<\/code><\/pre>\n\n<p><br\/><\/p>\n\n<h1>Long Answer<\/h1>\n\n<p>There are at least two ways retrieving the global from a checkpoint. As tensorflow variable or as numpy integer. Parsing the filename will not work, if the <code>global_step<\/code> was not provided as a parameter in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver#save\" rel=\"nofollow noreferrer\"><code>save<\/code><\/a> method of the <code>Saver<\/code>. For pretrained models see the remark at the end of the answer.<\/p>\n\n<h1>As Tensorflow variable<\/h1>\n\n<p>If you need the <code>global_step<\/code> variable to calculate some hyperparameters you can just use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/get_or_create_global_step\" rel=\"nofollow noreferrer\"><code>tf.train.get_or_create_global_step()<\/code><\/a>. This will return a tensorflow variable. Because the variable will be evaluated later in the session you can only use tensorflow operations to calculate your hyperparameters. So e.g.: <code>max(global_step, 100)<\/code> will not work. You have to use tensorflow equivalent <code>tf.maximum(global_step, 100)<\/code> that can be evaluated later in the session. <\/p>\n\n<p>Within the session you can initialize the global step variable with a checkpoint using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver#restore\" rel=\"nofollow noreferrer\"><code>saver.restore(sess, checkpoint_path)<\/code><\/a><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>global_step = tf.train.get_or_create_global_step()\n# use global_step variable to calculate your hyperparameter \n# this variable will be evaluated later in the session\nhyper_parameter = tf.maximum(global_step, 100) \nsaver = tf.train.Saver()\nwith tf.Session() as sess:\n    # restore all variables from checkpoint\n    saver.restore(sess, checkpoint_path)\n    # than init table and local variables and start training\/evaluation ...\n\n    # for verification you can print the global step and your hyper parameter\n    print(sess.run([global_step, hyper_parameter]))\n<\/code><\/pre>\n\n<h1>Or: As numpy integer (without session)<\/h1>\n\n<p>If you need the global step variable as scalar without starting a session you can also read this variable directly from your checkpoint file(s). You just need a <code>NewCheckpointReader<\/code>. Because of a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4921\" rel=\"nofollow noreferrer\">bug<\/a> in older tensorflow versions you should convert the path of the checkpoint file to an absolute path. With the reader you can get all the tensors of the model as numpy variables. \nThe name of the global step variable is a constant string <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GraphKeys#GLOBAL_STEP\" rel=\"nofollow noreferrer\"><code>tf.GraphKeys.GLOBAL_STEP<\/code><\/a> defined as <code>'global_step'<\/code>.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>absolute_checkpoint_path = os.path.abspath(checkpoint_path)\nreader = tf.train.NewCheckpointReader(absolute_checkpoint_path)\nglobal_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)\n<\/code><\/pre>\n\n<p><strong>Remark to pretrained models:<\/strong> In most pretrained models that are available online the global step is reset to zero. So, these models can be used to initialize the model parameters for finetuning without overwrite the global step.<\/p>\n"},{"AnswerId":"46642471","CreationDate":"2017-10-09T08:58:17.683","ParentId":null,"OwnerUserId":"6187135","Title":null,"Body":"<p>just note my solution on global step saving and restore.<\/p>\n\n<p>Save:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>global_step = tf.Variable(0, trainable=False, name='global_step')\nsaver.save(sess, model_path + model_name, global_step=_global_step)\n<\/code><\/pre>\n\n<p>Restore:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if os.path.exists(model_path):\n    saver.restore(sess, tf.train.latest_checkpoint(model_path))\n    print(\"Model restore finished, current globle step: %d\" % global_step.eval())\n<\/code><\/pre>\n"},{"AnswerId":"55912059","CreationDate":"2019-04-29T23:32:03.087","ParentId":null,"OwnerUserId":"11065060","Title":null,"Body":"<p>The reason that a variable is not restored as expected is most likely due to the fact that it was created after your <code>tf.Saver()<\/code> object was created.<\/p>\n\n<p>The place where you create the <code>tf.Saver()<\/code> object matters when you don't explicitly specify a <code>var_list<\/code>, or specify <code>None<\/code> for <code>var_list<\/code>. The expected behavior for many programmers is that all variables in the graph are saved when the <code>save()<\/code> method is called, but this is not the case, and it should perhaps be documented as such. A snapshot of all variables in the graph is saved at the time of object creation.<\/p>\n\n<p>Unless you're having any performance issues, it's safest to create the saver object right when you decide to save your progress. Otherwise, make sure to create the saver object after you create all your variables.<\/p>\n\n<p>Also, the <code>global_step<\/code> that is passed to <code>saver.save(sess, save_path, global_step=global_step)<\/code> is merely a counter used for creating the filename and has nothing to do with whether it will be restored as a <code>global_step<\/code> variable. This is a parameter misnomer IMO since if you're saving your progress at the end of each epoch, it's probably best to pass your epoch number for this parameter. <\/p>\n"}]}
{"QuestionId":36115026,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2016-03-20T14:33:40.943","AcceptedAnswerId":36115263.0,"OwnerUserId":6089550.0,"Title":"InfogainLoss layer blobs failure","Body":"<p>I am trying to build up caffe's infogain loss layer to work. I have seen the posts, solutions but for me it's still doesn't work<\/p>\n\n<p>My data lmdb dimensions are Nx1xHxW (grayscale images) and my target image lmdb dimensions are Nx3xH\/8xW\/8 (rgb images). My last convolutional layer's dimension is 1x3x20x80. The output_size is 3,\n so I have 3 classes as my label numbers are (0,1,2) in the target lmdb image dataset.<\/p>\n\n<p>I want to try out infogain loss layer, because I think I have class imbalance problem. Most of my images contains too much background.<\/p>\n\n<p>After My last convolutional layer (conv3) I have these:<\/p>\n\n<pre><\/pre>\n\n<p>My infogain matrix was generated by  <a href=\"https:\/\/stackoverflow.com\/questions\/27632440\/infogainloss-layer\">InfogainLoss layer<\/a> post (as Shai suggested) so my H matrix is 1x1x3x3 dimension (an identity matrix). So my  is 3 as I have 3 classes.\nWhen I run the prototxt file everything is fine (dimensions are ok), but after my last convolution layer (conv3 layer) I get the following error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I double checked, every lmdb dataset filename has been set correctly. I don't know what can be the problem. Any idea?<\/p>\n\n<p>Dear @Shai<\/p>\n\n<p>Thank you for your answer. I did the following as you mentioned:<\/p>\n\n<pre><\/pre>\n\n<p>But I still have error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36115263","CreationDate":"2016-03-20T14:54:37.817","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h3>What went wrong?<\/h3>\n\n<p>Your error comes from <code>\"loss\"<\/code> layer, not the <code>\"InfogainLoss\"<\/code> layer: You confused <code>\"Softmax\"<\/code> layer that outputs class probabilities, with <code>\"SoftmaxWithLoss\"<\/code> layer that outputs a (scalar) loss value.<\/p>\n\n<h3>What should you do?<\/h3>\n\n<ol>\n<li><p>Replace the <code>\"loss\"<\/code> layer with a <code>\"prob\"<\/code> layer of type <code>\"Softmax\"<\/code> layer:<\/p>\n\n<pre><code>layer {\n  name: \"prob\"\n  type: \"Softmax\" # NOT SoftmaxWithLoss\n  bottom: \"conv3\"\n  top: \"prob\"\n  softmax_param { axis: 1 } # compute prob along 2nd axis\n}\n<\/code><\/pre><\/li>\n<li><p>You need to compute the loss across the second dimension, currently it seems like <code>\"InfogainLoss\"<\/code> layer does not support this feature. You might need to tweak <code>\"InfogainLoss\"<\/code> layer to have functionality like <code>\"SoftmaxWithLoss\"<\/code> that allows loss to be computed along arbitrary axis.<br>\n<strong>Update:<\/strong> I created a <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3855\" rel=\"nofollow noreferrer\">pull request<\/a> on <code>BVLC\/caffe<\/code> that \"upgrades\" infogain loss layer. This upgraded version supports \"loss along axis\" like you are after. Moreover, it makes the \"Softmax\" layer redundant as it computes the probabilities internally (see <a href=\"https:\/\/stackoverflow.com\/a\/34917052\/1714410\">this thread<\/a>).<br>\nThe upgraded layer can be used like this:<\/p>\n\n<pre><code>layer {\n  bottom: \"conv3\" # prob is computed internally\n  bottom: \"label\"\n  top: \"infoGainLoss\"\n  name: \"infoGainLoss\"\n  type: \"InfogainLoss\"\n  infogain_loss_param {\n    source: \"infogainH.binaryproto\"\n    axis: 1  # compute loss and probability along axis\n  }\n}\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":36115579,"AnswerCount":1,"Tags":"<android><linux><torch><flashlight>","CreationDate":"2016-03-20T15:19:17.613","AcceptedAnswerId":null,"OwnerUserId":4171029.0,"Title":"Torch app trouble","Body":"<p>UPDATE: I have edited my question to include the working code. I have solved my issue. See my latest\/last post.<\/p>\n\n<p>UPDATE2: Code has been updated as of 2018-06-08. Here is the repo: <\/p>\n\n<p><a href=\"https:\/\/github.com\/amboxer21\/FlashLightApp\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amboxer21\/FlashLightApp<\/a><\/p>\n\n<p>NOTE: I compile from the Linux command line with ant.<\/p>\n\n<p>I have a torch app I am trying to get working but I am having no luck. Can any one point out what I am missing and explain? I would really appreciate it!<\/p>\n\n<p>I can compile with no issues. I can push with no issues. I can toggle the on and off button with no issues. I can see this with the toasts I have put in place. Basically all functions seem to get called properly but the LED light on the back camera just does not come on. I am using a Nexus 6 on 5.1.1 and developing on Linux. Also I figure its worth mentioning here but I am able to compile a 3rd party torch app on my Linux box and the torch woks.<\/p>\n\n<p>Main Activity:<\/p>\n\n<pre><\/pre>\n\n<p>main.xml:<\/p>\n\n<pre><\/pre>\n\n<p>Manifest:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36132744","CreationDate":"2016-03-21T13:47:10.790","ParentId":null,"OwnerUserId":"4171029","Title":null,"Body":"<p>Turns out I needed a surfaceHolder and surfaceView in order to call the startPreview method.<\/p>\n\n<p>Here is my working torch app via github ->  <a href=\"https:\/\/github.com\/amboxer21\/FlashLightApp\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amboxer21\/FlashLightApp<\/a><\/p>\n"}]}
{"QuestionId":36115768,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2016-03-20T15:36:15.117","AcceptedAnswerId":36261080.0,"OwnerUserId":6089914.0,"Title":"Using an optimizer from the backend in keras","Body":"<p>I can see that keras has reimplemented the optimizers. I would like to use an optimizer that is not available in keras, but is available in the backend, e.g. from tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>Could you please guide me with some bullet points what would I need to do, to hack the framework, to be able to use an optimizer from the backend?<\/p>\n\n<p>Thank you,\nJozef<\/p>\n","answers":[{"AnswerId":"36261080","CreationDate":"2016-03-28T11:08:03.927","ParentId":null,"OwnerUserId":"6089914","Title":null,"Body":"<p>OK figured it out, the hack was to add in a wrapper optimizer<\/p>\n"}]}
{"QuestionId":36117187,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><theano><convolution>","CreationDate":"2016-03-20T17:41:34.157","AcceptedAnswerId":37457823.0,"OwnerUserId":4566639.0,"Title":"Advice using the theano implementation of Conv3D","Body":"<p>I am trying to run a 3D convolutional neural network using theano, however I am not completely sure of the usage of the function .<\/p>\n\n<p>I am used to using lasagne, however due to not having access to a GPU at this time I am unable to use the function. <\/p>\n\n<p>Is anyone able to advise me in terms of inputs and outputs what I need to do to be able to use the theano function? I have data in the form N x 9 x 9 x 9 with 1 channel, and have initialised the theano tensors. I have also created an input layer with lasagne like so:<\/p>\n\n<pre><\/pre>\n\n<p>Any advice is very welcome! Thanks.<\/p>\n","answers":[{"AnswerId":"37457823","CreationDate":"2016-05-26T10:11:33.147","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>There is currently no 3D conv support in Lasagne other than cuDNN.\nHowever I have fork of lasagne that implements it:\n<a href=\"https:\/\/github.com\/gyglim\/Lasagne\" rel=\"nofollow\">https:\/\/github.com\/gyglim\/Lasagne<\/a> (Conv3DLayer)<\/p>\n\n<p>Its outputs are identical with the ones from Conv3DDNNLayer.<\/p>\n\n<p>There is also discussion about adding it to the core lasagne (see <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/issues\/27\" rel=\"nofollow\">https:\/\/github.com\/Lasagne\/Lasagne\/issues\/27<\/a>), but it hasn't happend yet.<\/p>\n\n<p>Hope this helps.<\/p>\n\n<p>Cheers,\nMichael<\/p>\n"}]}
{"QuestionId":36117501,"AnswerCount":1,"Tags":"<python><caffe><lmdb>","CreationDate":"2016-03-20T18:05:21.990","AcceptedAnswerId":null,"OwnerUserId":4358739.0,"Title":"Copying LMDB to another LMDB reduces file size","Body":"<p>To shuffle the data in already existing lmdb (<a href=\"https:\/\/stackoverflow.com\/questions\/36108120\/shuffle-data-in-lmdb-file\">Trying to solve this problem<\/a>). I retrieved the data, shuffled and wrote back to new lmdb. But the when I checked the lmdb file size, it is reduced. Old lmdb file size : 3792896 but the New lmdb file size : 2314240.<\/p>\n\n<p>Python code Inplemented : <\/p>\n\n<pre><\/pre>\n\n<p>Reference for the code is taken from <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow noreferrer\">here<\/a>.\nAny suggestions\/ideas would be helpful.<\/p>\n","answers":[{"AnswerId":"37646102","CreationDate":"2016-06-05T19:49:25.990","ParentId":null,"OwnerUserId":"5129319","Title":null,"Body":"<p>You can use <code>mdb_stat<\/code> to see the number of entries in the database. This should confirm if your copy worked correctly.<\/p>\n"}]}
{"QuestionId":36120002,"AnswerCount":0,"Tags":"<python><theano><conv-neural-network><cudnn>","CreationDate":"2016-03-20T21:47:24.660","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Compiling convolution layers with Theano. issue with CuDNN \" ('The following error happened while compiling the node', GpuDnnConvDesc ...\"","Body":"<p>Theano seems to have some issues when compiling convolution layers. When I try to run the <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">Convolutional Neural Networks (LeNet)<\/a> example, I get the error message:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to run the example using: <\/p>\n\n<pre><\/pre>\n\n<p>or <\/p>\n\n<pre><\/pre>\n\n<p>The example runs fine if I disable CuDNN with the  flag:<\/p>\n\n<pre><\/pre>\n\n<p>What could be the issue with CuDNN?<\/p>\n\n<p>I use:<\/p>\n\n<ul>\n<li>Theano 0.7.0 (according to ), <\/li>\n<li>Python 2.7.6 64 bits (according to ), <\/li>\n<li>Nvidia-smi 361.28 (according to ), <\/li>\n<li>CUDA 7.5.17 (according to ),<\/li>\n<li>GeForce GTX Titan X (according to ), <\/li>\n<li>Ubuntu 14.04.4 LTS x64 (according to  and ).<\/li>\n<\/ul>\n\n<hr>\n\n<p>Full error message:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36120302,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><logistic-regression><softmax>","CreationDate":"2016-03-20T22:15:06.843","AcceptedAnswerId":36574629.0,"OwnerUserId":2235817.0,"Title":"How to replace Softmax ouput Layer with Logistic Layer in Tensorflow?","Body":"<p>I need a bit help in my work. Right now, I am using Softmax layer as output layer for classification scores in neural network. But, I need to replace Softmax layer with logistic layer on the output layer. I have some inputs that belongs to multiple classes. Softmax is showing probability over all the classes and assigned the class to the highest probability and its hard to decide a threshold to predict more than one classes at a time. While in case of logistic function each neuron will display a number between (0-1) and I can decide a threshold in that case.\nHere is my code:<\/p>\n\n<p><strong>2 layer Network Initialization<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>How we can modify this network to have a probability in between (0-1) on each output Neuron?<\/p>\n","answers":[{"AnswerId":"36574629","CreationDate":"2016-04-12T13:21:28.960","ParentId":null,"OwnerUserId":"2235817","Title":null,"Body":"<p>just change line:<\/p>\n\n<pre><code># Construct model\npred = multilayer_perceptron(x, weights, biases)\n<\/code><\/pre>\n\n<p>To<\/p>\n\n<pre><code># Construct model\nmodel pred = tf.nn.sigmoid(multilayer_perceptron(x, weights, biases))\n<\/code><\/pre>\n"}]}
{"QuestionId":36120914,"AnswerCount":1,"Tags":"<argparse><tensorflow>","CreationDate":"2016-03-20T23:28:36.593","AcceptedAnswerId":36121207.0,"OwnerUserId":1930168.0,"Title":"Argparse error with TensorFlow's cifar10.py","Body":"<p>I get the following error when I run :<\/p>\n\n<p><\/p>\n\n<p>Here's the full output of the run including a complete trace:<\/p>\n\n<pre><\/pre>\n\n<p>This error seems to come from the following line in : <\/p>\n\n<p>It seems like the  library thinks that I've already defined the option string , but I haven't.<\/p>\n\n<p>[Stack: Amazon g2.2xlarge spot instance, Python 2.7.6]<\/p>\n","answers":[{"AnswerId":"36121207","CreationDate":"2016-03-21T00:05:01.560","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>In the <code>cifr10.py<\/code> file:<\/p>\n\n<pre><code>import tensorflow as tf\n\nfrom tensorflow.models.image.cifar10 import cifar10_input\n\nFLAGS = tf.app.flags.FLAGS\n\n# Basic model parameters.\ntf.app.flags.DEFINE_integer('batch_size', 128,\n                            \"\"\"Number of images to process in a batch.\"\"\")\n....\n<\/code><\/pre>\n\n<p>The error is produced by this last statement, which, in the <code>_flags.py<\/code> file, defines an <code>argparse<\/code> argument with that name.  Evidently at this point the <code>tf.app<\/code> already has such an argument define.<\/p>\n\n<p>So we need to look further back at <code>import tensorflow as tf<\/code> to see how <code>tf.app<\/code> was created?<\/p>\n\n<p>What's the <code>Amazon g2.2xlarge<\/code>?  Could that defining <code>batch_size<\/code> as well?<\/p>\n\n<p>Looks like <code>tf.app<\/code> comes from <\/p>\n\n<pre><code>tensorflow\/python\/platform\/app.py\n<\/code><\/pre>\n\n<p>which in turn gets it from something like<\/p>\n\n<pre><code>from tensorflow.python.platform.google._app import *\n<\/code><\/pre>\n\n<p>So if you are running this on some <code>google<\/code> or <code>amazon<\/code> platform that itself accepts <code>batch_size<\/code> parameter, it could produce this error.<\/p>\n\n<hr>\n\n<p>Another question about <code>cifr10<\/code> and the <code>batch_size<\/code> argument:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/35241827\/how-to-use-flags-command-line-switches-in-tensorflow\">How to use &quot;FLAGS&quot; (command line switches) in TensorFlow?<\/a><\/p>\n\n<p>Same error here: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/34734054\/tensorflow-argumenterror-running-cifar-10-example\">Tensorflow ArgumentError Running CIFAR-10 example<\/a>\nThe answer says to use <code>cifar10_train.py<\/code>,<code>cifar10_eval.py<\/code>, not <code>cifar10.py<\/code>.<\/p>\n"}]}
{"QuestionId":36122499,"AnswerCount":2,"Tags":"<computer-vision><caffe><vgg-net>","CreationDate":"2016-03-21T03:03:38.847","AcceptedAnswerId":36123435.0,"OwnerUserId":4182595.0,"Title":"Finetuning Cifar data with VGG net","Body":"<p>The original VGG network accepts input of size 256x256x3. In order to train Cifar-10 data, do I need to resize the data into 256x256x3? Or is there any other way?<\/p>\n\n<p>I am trying to fine tune VGG network using Cifar-10 data, which is of size 32x32x3.<\/p>\n","answers":[{"AnswerId":"36123435","CreationDate":"2016-03-21T04:58:32.870","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>In order to train the Cifar-10 data on VGG, you will have to give input data of size 256x256x3. Various input layers have transform parameters to take care of this image resizing.<\/p>\n\n<p>But upscaling a 32x32 image to 256x256 is not a good method as a major portion of the image data is created by an approximation from the available 32x32 image data. So I do not recommend doing this. But if this is just to try a demo VGG finetuning, you could do it by setting the transform parameter.<\/p>\n"},{"AnswerId":"43271921","CreationDate":"2017-04-07T07:15:43.917","ParentId":null,"OwnerUserId":"6729283","Title":null,"Body":"<p>You should use the reduced version of VGG for smaller images since VGG-16 is too complex for small images.<\/p>\n\n<p>I think this model works better for your problem and it's not hard to train it all with a gpu:<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Conv2D(32, (3, 3), padding='same',\n                 input_shape=x_train.shape[1:]))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes))\nmodel.add(Activation('softmax'))\n<\/code><\/pre>\n"}]}
{"QuestionId":36122601,"AnswerCount":1,"Tags":"<tensorflow><word2vec>","CreationDate":"2016-03-21T03:17:54.243","AcceptedAnswerId":null,"OwnerUserId":6091551.0,"Title":"Where is the result of word embedding after running word2vec.py in TensorFlow?","Body":"<p>I have run the 'word2vec.py' file under the folder of '\/tensorflow\/models\/embedding\/'. But I can't find the result of word embedding in the output folder, and there are only model files and some unknown files like this:<\/p>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/wbT2m.png\" rel=\"nofollow\">result files of word2vec<\/a> <\/p>\n","answers":[{"AnswerId":"44882684","CreationDate":"2017-07-03T09:58:18.620","ParentId":null,"OwnerUserId":"8162632","Title":null,"Body":"<p>Since, from the paper of Mikolov, it is clear that the ouput of word embeddings of all the words are not obtained in the final layer. Rather they chop off the final layer and take the weights of the hidden layer which is of the dimension: [vocab_size x embd_dim], i.e. Each row is a word vector (dimension: embd_dim) of the corresponding word.<\/p>\n\n<p>Refer this for better understanding: <a href=\"http:\/\/mccormickml.com\/2016\/04\/19\/word2vec-tutorial-the-skip-gram-model\/\" rel=\"nofollow noreferrer\">http:\/\/mccormickml.com\/2016\/04\/19\/word2vec-tutorial-the-skip-gram-model\/<\/a> <\/p>\n\n<p>So, in the tensorflow implementation, the <code>tf.embedding_lookup<\/code> has the word embeddings of all the words. \nIt's mentioned in <strong>line 26<\/strong> of the file <code>word2vec.py<\/code> in <code>\/tensorflow\/models\/embedding\/<\/code><\/p>\n\n<pre><code>24 The key ops used are:\n25 * placeholder for feeding in tensors for each example.\n26 * embedding_lookup for fetching rows from the embedding matrix.\n27 * sigmoid_cross_entropy_with_logits to calculate the loss.\n28 * GradientDescentOptimizer for optimizing the loss.\n29 * skipgram custom op that does input processing.\n<\/code><\/pre>\n"}]}
{"QuestionId":36123470,"AnswerCount":0,"Tags":"<lua><machine-learning><deep-learning><torch><conv-neural-network>","CreationDate":"2016-03-21T05:02:13.233","AcceptedAnswerId":null,"OwnerUserId":6091401.0,"Title":"Torch7 Convolutional Neural Network, multiple outputs with probabilities","Body":"<p>I'm using Torch7 for image recognition (Convolutionnal Neural Network), when I use the default model, I get one output even if I have 4 output classes. The generated output is the final decision, I want to get the probability for each one of the classes (4), something like:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the code I use for the definition of my model:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you for helping me. <\/p>\n","answers":[]}
{"QuestionId":36123740,"AnswerCount":4,"Tags":"<gpu><tensorflow>","CreationDate":"2016-03-21T05:27:50.163","AcceptedAnswerId":37931964.0,"OwnerUserId":349760.0,"Title":"Is there a way of determining how much GPU memory is in use by TensorFlow?","Body":"<p>Tensorflow tends to preallocate the entire available memory on it's GPUs. For debugging, is there a way of telling how much of that memory is actually in use?<\/p>\n","answers":[{"AnswerId":"46067272","CreationDate":"2017-09-06T05:16:53.957","ParentId":null,"OwnerUserId":"7583767","Title":null,"Body":"<p>The TensorFlow profiler has improved memory timeline that is based on real gpu memory allocator information\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/profiler#visualize-time-and-memory\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/profiler#visualize-time-and-memory<\/a><\/p>\n"},{"AnswerId":"50880817","CreationDate":"2018-06-15T18:17:13.987","ParentId":null,"OwnerUserId":"9947951","Title":null,"Body":"<p>There's some code in <code>tensorflow.contrib.memory_stats<\/code> that will help with this:<\/p>\n\n<pre><code>from tensorflow.contrib.memory_stats.python.ops.memory_stats_ops import BytesInUse\nwith tf.device('\/device:GPU:0'):  # Replace with device you are interested in\n  bytes_in_use = BytesInUse()\nwith tf.Session() as sess:\n  print(sess.run(bytes_in_use))\n<\/code><\/pre>\n"},{"AnswerId":"37931964","CreationDate":"2016-06-20T21:17:22.650","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>(1) There is some limited support with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/client\/timeline.py\" rel=\"noreferrer\">Timeline<\/a> for logging memory allocations. Here is an example for its usage:<\/p>\n\n<pre><code>    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\n    run_metadata = tf.RunMetadata()\n    summary, _ = sess.run([merged, train_step],\n                          feed_dict=feed_dict(True),\n                          options=run_options,\n                          run_metadata=run_metadata)\n    train_writer.add_run_metadata(run_metadata, 'step%03d' % i)\n    train_writer.add_summary(summary, i)\n    print('Adding run metadata for', i)\n    tl = timeline.Timeline(run_metadata.step_stats)\n    print(tl.generate_chrome_trace_format(show_memory=True))\n    trace_file = tf.gfile.Open(name='timeline', mode='w')\n    trace_file.write(tl.generate_chrome_trace_format(show_memory=True))\n<\/code><\/pre>\n\n<p>You can give this code a try with the MNIST example (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"noreferrer\">mnist with summaries<\/a>)<\/p>\n\n<p>This will generate a tracing file named timeline, which you can open with chrome:\/\/tracing. Note that this only gives an approximated GPU memory usage statistics. It basically simulated a GPU execution, but doesn't have access to the full graph metadata. It also can't know how many variables have been assigned to the GPU.<\/p>\n\n<p>(2) For a very coarse measure of GPU memory usage, nvidia-smi will show the total device memory usage at the time you run the command.<\/p>\n\n<p>nvprof can show the on-chip shared memory usage and register usage at the CUDA kernel level, but doesn't show the global\/device memory usage.<\/p>\n\n<p>Here is an example command: nvprof --print-gpu-trace matrixMul<\/p>\n\n<p>And more details here:\n<a href=\"http:\/\/docs.nvidia.com\/cuda\/profiler-users-guide\/#abstract\" rel=\"noreferrer\">http:\/\/docs.nvidia.com\/cuda\/profiler-users-guide\/#abstract<\/a><\/p>\n"},{"AnswerId":"52026799","CreationDate":"2018-08-26T13:38:09.737","ParentId":null,"OwnerUserId":"896111","Title":null,"Body":"<p>Here's a practical solution that worked well for me:  <\/p>\n\n<p>Disable GPU memory pre-allocation using TF session configuration:  <\/p>\n\n<pre><code>config = tf.ConfigProto()  \nconfig.gpu_options.allow_growth=True  \nsess = tf.Session(config=config)  \n<\/code><\/pre>\n\n<p>run <strong>nvidia-smi -l<\/strong> (or some other utility) to monitor GPU memory consumption.<\/p>\n\n<p>Step through your code with the debugger until you see the unexpected GPU memory consumption.<\/p>\n"}]}
{"QuestionId":36123814,"AnswerCount":1,"Tags":"<algorithm><machine-learning><computer-vision><deep-learning><caffe>","CreationDate":"2016-03-21T05:33:39.687","AcceptedAnswerId":36125298.0,"OwnerUserId":5572849.0,"Title":"Which is right tripletLoss backpropagation formula?","Body":"<p>I saw <a href=\"https:\/\/stackoverflow.com\/questions\/33330779\/whats-the-triplet-loss-back-propagation-gradient-formula\/33349475#33349475\">What's the triplet loss back propagation gradient formula?<\/a>\nFor gradient, anchor is , positive is  and negative is .<br>\nBut from 80 line to 92 line of <a href=\"https:\/\/github.com\/happynear\/caffe-windows\/blob\/master\/src\/caffe\/layers\/triplet_loss_layer.cpp\" rel=\"nofollow noreferrer\"><\/a> is different to this.<br>\nNamely, for gradient anchor is  and positive is .\nwhich is really right?<\/p>\n","answers":[{"AnswerId":"36125298","CreationDate":"2016-03-21T07:31:30.203","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Lines 80-92 in <a href=\"https:\/\/github.com\/happynear\/caffe-windows\/blob\/master\/src\/caffe\/layers\/triplet_loss_layer.cpp#L80-L92\" rel=\"nofollow noreferrer\"><code>triplet_loss_layer.cpp<\/code><\/a> are part of <code>forward_cpu<\/code> function - that is the actual loss computation and NOT the gradient computation.<\/p>\n\n<p>The gradient is computed in <a href=\"https:\/\/github.com\/happynear\/caffe-windows\/blob\/master\/src\/caffe\/layers\/triplet_loss_layer.cpp#L116-L171\" rel=\"nofollow noreferrer\"><code>backward_cpu<\/code><\/a>, where you can see that each <code>bottom<\/code> is assigned its <code>diff<\/code> according to the derivation presented <a href=\"https:\/\/stackoverflow.com\/a\/33349475\/1714410\">here<\/a>.<\/p>\n"}]}
{"QuestionId":36124790,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-03-21T06:53:36.207","AcceptedAnswerId":36133259.0,"OwnerUserId":349760.0,"Title":"How do you pin gradients to a certain GPU in TensorFlow?","Body":"<p>I created a model with an embedding layer on CPU, an LSTM on gpu:0 and a softmax layer on gpu:1.<\/p>\n\n<p>I then created a training op like so:<\/p>\n\n<pre><\/pre>\n\n<p>Training fails with an Out of Memory error. Enabling device placement logging reveals that the gradient for the softmax has been placed on gpu:0 rather than gpu:1. How can get TensorFlow to place it correctly?<\/p>\n","answers":[{"AnswerId":"36133259","CreationDate":"2016-03-21T14:09:23.437","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To place the gradients on the same device as the original (forward-pass) op, you should pass the optional argument <code>colocate_gradients_with_ops=True<\/code> to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.minimize\" rel=\"nofollow\"><code>optim.minimize()<\/code><\/a>. Your code will look like:<\/p>\n\n<pre><code>optim = tf.train.GradientDescentOptimizer(LEARNING_RATE)\ntrain_op = optim.minimize(final_loss, colocate_gradients_with_ops=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":36125985,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-03-21T08:20:34.973","AcceptedAnswerId":36137931.0,"OwnerUserId":6056683.0,"Title":"Why I can't use theano.tensor.argmax and theano.tensor.mean correctly","Body":"<p>I am learning Theano now but there are always some problems.my code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I thought it would print the index of '4',but the result is:<\/p>\n\n<pre><\/pre>\n\n<p>what's more,when I am using  T.neq().just as follows:<\/p>\n\n<pre><\/pre>\n\n<p>the result shows:<\/p>\n\n<pre><\/pre>\n\n<p>I really new to this and have no idea,did I miss anything?thank you in advance..<\/p>\n","answers":[{"AnswerId":"36137931","CreationDate":"2016-03-21T17:33:16.223","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>T.argmax() is expecting a Theano TensorVariable type. Some of the types of Variables used in Theano are listed <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#all-fully-typed-constructors\" rel=\"noreferrer\">here<\/a>. Don't let the name \"fully typed constructors\" scare you. Think about them more in terms of what type of data you want to use as your input. Are you using float matrices? Then the relevant TensorVariable type is probably \"fmatrix.\" Are you dealing with batches of RGB image data? Then the relevant TensorVariable type is probably \"tensor4.\" <\/p>\n\n<p>In your code, we are trying to input a list type into T.argmax(). So from the above point of view, that isn't going to work. Also, note that type(T.argmax(a)) is a theano.tensor.var.TensorVariable type. So it is expecting a TensorVariable as input, and it outputs a TensorVariable type as well. <em>So this isn't going to return the actual argmax<\/em>.<\/p>\n\n<p>Okay, so what does work? How can we do this computation in Theano? <\/p>\n\n<p>Let's first identify the type of data you want to deal with. This is going to be the starting point of our computational graph that we will be building. In this case, it looks like we want to deal with arrays or vectors. Theano has an ivector type, which is a vector of integers, or an fvector type which is an vector of float32 values. Let's stick with your data and do ivector since we have integer values:<\/p>\n\n<pre><code>x = T.ivector('input')\n<\/code><\/pre>\n\n<p>This line just created a TensorVariable x that represents our intended input type, an array of integers. <\/p>\n\n<p>Now let's define a TensorVariable for the argmax of the elements of x:<\/p>\n\n<pre><code>y = T.argmax(x)\n<\/code><\/pre>\n\n<p>So far we have built a computational graph, which is expecting an array of integers as input and will output the argmax of that array. However, in order to actually do this, we have to compile this into a function:<\/p>\n\n<pre><code>get_argmax = theano.function([x], y)\n<\/code><\/pre>\n\n<p>The theano.function syntax can be found <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/function.html\" rel=\"noreferrer\">here<\/a>. <\/p>\n\n<p>Think of this function as now actually performing the computation that we have defined using x and y. <\/p>\n\n<p>When I execute:<\/p>\n\n<pre><code>get_argmax([1,2,3,4,19,1])\n<\/code><\/pre>\n\n<p>It returns: <\/p>\n\n<pre><code>array(4)\n<\/code><\/pre>\n\n<p>So what did we really do? By defining Theano variables and using theano.tensor functions, we build a computational graph. We then used theano.function to compile a function that actually performs that computation on actual inputs that we specify. <\/p>\n\n<p>To end: how to do the not equals operation? <\/p>\n\n<pre><code>a = T.ivector('a')\nb = T.ivector('b')\nout = T.neq(a,b)\nget_out = theano.function([a,b], out)\nprint get_out([1,2,3,4], [7,8,9,10])\n<\/code><\/pre>\n\n<p>will return:<\/p>\n\n<pre><code>[1,1,1,1]\n<\/code><\/pre>\n\n<p>One of the key conceptual differences is that I treat the a,b as theano TensorVariables, rather than assigning them explicit variables. <\/p>\n\n<p>You'll get the hang out of it, just remember that you need to define your computation in terms of Theano TensorVariables, and then to actually \"use it\" you have to compile it using theano.function. <\/p>\n"}]}
{"QuestionId":36126147,"AnswerCount":2,"Tags":"<bash><shell><terminal><.bash-profile><torch>","CreationDate":"2016-03-21T08:31:32.930","AcceptedAnswerId":null,"OwnerUserId":5099765.0,"Title":"Torch installation error with the shell profile","Body":"<p>I am using mac osx el capitan. \nI did the first two parts without errors:<\/p>\n\n<pre><\/pre>\n\n<p>When I do the third part ,  I get the following error:<\/p>\n\n<p>There are two parts of the error I believe. One is regarding using iPython and second has something to do with the . Please help me out here.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36128314","CreationDate":"2016-03-21T10:18:53.077","ParentId":null,"OwnerUserId":"4571544","Title":null,"Body":"<p>It looks like <code>install-deps<\/code> doesn't install anything python-related for MAC OSX. Try installing ipython yourself via <code>easy_install<\/code> or <code>pip<\/code>prior to running <code>install.sh<\/code><\/p>\n"},{"AnswerId":"36128787","CreationDate":"2016-03-21T10:38:00.330","ParentId":null,"OwnerUserId":"874188","Title":null,"Body":"<p>The <code>permission denied<\/code> suggests that you may have incorrectly run some part of this (or another?) installation script as <code>root<\/code> or <code>sudo<\/code> previously, causing some directories in your home directory to be unwritable by yourself. Maybe you can <code>chown<\/code> them to yourself as <code>root<\/code>, then fix the permissions; or maybe just remove them and start over.<\/p>\n"}]}
{"QuestionId":36126570,"AnswerCount":1,"Tags":"<python><theano><mxnet>","CreationDate":"2016-03-21T08:56:39.803","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"MXNET vs Theano Experience","Body":"<p>Am looking to do some distributed calc. using GPU for machine learning ?\nJust wondering if anybody has experience with MXNET (perf. vs Theano)<\/p>\n\n<p>Reference\n<a href=\"http:\/\/www.cs.cmu.edu\/~muli\/file\/mxnet-learning-sys.pdf\" rel=\"nofollow\">http:\/\/www.cs.cmu.edu\/~muli\/file\/mxnet-learning-sys.pdf<\/a><\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"36295821","CreationDate":"2016-03-29T21:55:39.220","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>I had a lot of experience with both <code>mxnet<\/code> and <code>Theano<\/code> (via lasagne and keras)<\/p>\n\n<p>Benchmarking is always biased, so I will not comment on that, except to note that all the frameworks are very fast. Here are several things that should help you decide:<\/p>\n\n<ol>\n<li><code>Theano<\/code> compared to <code>mxnet<\/code> is like assembly compared to python. Theano has low-level primitives to build machine learning models, and on itself does not define any layers or optimizers, and you would usually use it with some deep-learning library, such as Lasagne or Keras, while <code>mxnet<\/code> is higher level. So fare comparison would be <code>mxnet<\/code> vs <code>Keras<\/code>, not <code>mxnet<\/code> vs <code>Theano<\/code>.<\/li>\n<li><code>mxnet<\/code> is a more recent library, and certain things in it are not as polished yet, and there's way fewer resources online than for <code>Theano<\/code>.<\/li>\n<li><code>Theano<\/code> (and therefore Lasagne and Keras) compile models when they run them for the first time into C++ and Cuda, which is very slow. For a very complex model, such as an unrolled LSTM, it can take good couple minutes to compile. It is usually very little compared to the time it will take for the model to train (hours to weeks), but is very annoying when you prototype.<\/li>\n<\/ol>\n\n<p>Overall, if you choose between these two frameworks, I would suggest Theano + Keras for everything except for the recurrent or very deep networks, otherwise the compilation in <code>Theano<\/code> will be killing you.<\/p>\n\n<p>Also look into <code>TensorFlow<\/code>. It is (subjectively) slower than <code>mxnet<\/code>, but is more mature and has more resources online.<\/p>\n"}]}
{"QuestionId":36127436,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-03-21T09:38:48.677","AcceptedAnswerId":36134261.0,"OwnerUserId":1165907.0,"Title":"Tensorflow predicts always the same result","Body":"<p>I'm trying to get the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">TensorFlow example<\/a> running with my own data, but somehow the classifier always picks the same class for every test example. The input data is always shuffled prior. I have about 4000 images as a training set and 500 images as a testing set.<\/p>\n\n<p>The result I get looks like:<\/p>\n\n<pre><\/pre>\n\n<p>The right side remains for all 500 images . The classification is binary so I just have two labels.<\/p>\n\n<p>Here is my source code:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing a point?<\/p>\n","answers":[{"AnswerId":"36134261","CreationDate":"2016-03-21T14:49:24.513","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are three potential issues in your code:<\/p>\n\n<ol>\n<li><p>The weights, <code>W<\/code>, are initialized to zero. <a href=\"https:\/\/stats.stackexchange.com\/q\/27112\/96788\">This question from stats.stackexchange.com<\/a> has a good discussion of why this can lead to poor training outcomes (such as getting stuck in a local minimum). Instead, you should initialize them randomly, for example as follows:<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([pixels, labels],\n                                    stddev=1.\/math.sqrt(pixels)))\n<\/code><\/pre><\/li>\n<li><p>The <code>cross_entropy<\/code> should be aggregated to a single, scalar value before minimizing it, using for example <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops.html#reduce_mean\" rel=\"noreferrer\"><code>tf.reduce_mean()<\/code><\/a>:<\/p>\n\n<pre><code>cross_entropy = tf.reduce_mean(\n    tf.nn.softmax_cross_entropy_with_logits(y_prime, y_))\n<\/code><\/pre><\/li>\n<li><p>You may get faster convergence if you train on mini-batches (or even single examples) rather than training on the entire dataset at once:<\/p>\n\n<pre><code>for i in range(10):\n        for j in range(4000):\n            res = train_step.run({x: train_images[j:j+1],\n                                  y_: train_labels[j:j+1]})\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"43772553","CreationDate":"2017-05-04T01:48:40.540","ParentId":null,"OwnerUserId":"2498105","Title":null,"Body":"<p>The other problem you may be having is a Class imbalance.  If you have one class that greatly outweighs the other your function may be converging to that value.  Try balancing the classes in your training sample as well as using smaller batches.  For example if your labels are binary then make sure there is an equal amount of zeros and one labels in your training sample. <\/p>\n"}]}
{"QuestionId":36128348,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-21T10:20:11.053","AcceptedAnswerId":36140651.0,"OwnerUserId":2902280.0,"Title":"All runs are not visible on TensorBoard","Body":"<p>I am using tensorboard to visualize three runs. I have a folder, , which contains three files:  <\/p>\n\n<pre><\/pre>\n\n<p>when I run  (from logs), only \n is visible:<a href=\"https:\/\/i.stack.imgur.com\/254nZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/254nZ.png\" alt=\"blu\"><\/a><\/p>\n\n<p>And if I delete  from logs and restart tensorboard, then only  is visible. Any idea of what's happening here?<\/p>\n\n<p>Edit: the log files are (to my surprise), quite big: here is the result of :<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong>\nThe above run structure can be obtained via logging and checkpointing into <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/how_tos\/summaries_and_tensorboard\/index.html#launching-tensorboard\" rel=\"nofollow noreferrer\">subdirectories of the tensorboard log-dir<\/a>:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36140651","CreationDate":"2016-03-21T20:11:45.310","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>Tensorboard takes a while to parse the log files. If you refresh the graph, you can see it showing more and more iterations. When it finished parsing one, the next run will pop up.<\/p>\n\n<p>It is a bit annoying that there is no visual indication for this, maybe this is worth a feature request at <a href=\"https:\/\/github.com\/tensorflow\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow<\/a>.<\/p>\n"}]}
{"QuestionId":36133088,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-21T14:01:33.357","AcceptedAnswerId":36133338.0,"OwnerUserId":6093883.0,"Title":"What is a nice way to obtain a slice partially indexed by another tensor in TensorFlow?","Body":"<p>Suppose we have a tensor  with unknown first dimension (for example ), and another tensor  that is actually a scalar. Is there a nice way to obtain the -th slice of  split by first dimension, for example, to get a tensor of dimension ? I'm new to TensorFlow and was only able to come up with this extremely clumsy solution. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36133338","CreationDate":"2016-03-21T14:12:47.407","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can take advantage of the fact that <code>-1<\/code> is a special argument to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#slice\" rel=\"nofollow\"><code>tf.slice()<\/code><\/a> <code>size<\/code> argument, meaning \"all remaining elements in that dimension\". Then, assuming <code>i<\/code> is a scalar (and not a length-1 vector as it seems to be in your code snippet), you can do:<\/p>\n\n<pre><code>result = tf.squeeze(tf.slice(x, tf.pack([index, 0, 0, 0]), [1, -1, -1, -1]), [0])\n<\/code><\/pre>\n\n<p>Alternatively, you can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#gather\" rel=\"nofollow\"><code>tf.gather()<\/code><\/a> to select one or more slices from a tensor on the zeroth dimension. In this case, <code>i<\/code> must be a vector:<\/p>\n\n<pre><code>i = tf.expand_dims(i, 0)  # Converts `i` to a vector if it is a scalar.\nresult = tf.squeeze(tf.gather(x, i), [0])\n<\/code><\/pre>\n\n<p>In both cases, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#squeeze\" rel=\"nofollow\"><code>tf.squeeze()<\/code><\/a> op removes the 0th dimension to give a three-dimensional result.<\/p>\n"}]}
{"QuestionId":36133662,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><recurrent-neural-network>","CreationDate":"2016-03-21T14:26:03.670","AcceptedAnswerId":36925269.0,"OwnerUserId":4802797.0,"Title":"epoch vs. iteration in recurrent neural networks","Body":"<p>I am looking at the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"nofollow noreferrer\">text generation example<\/a> of Keras using RNN and LSTM but still confused about the difference between the terms <em>epoch<\/em> and <em>iteration<\/em>. <\/p>\n\n<p>Even though, <a href=\"https:\/\/stackoverflow.com\/q\/4752626\/4802797\">here<\/a> is a previous question asking the same thing, I cannot understand the <a href=\"https:\/\/stackoverflow.com\/a\/31842945\/4802797\">answer<\/a>, or this answer is different from how I understand it and also different from how the following example handles it. Based on this <a href=\"https:\/\/stackoverflow.com\/a\/31842945\/4802797\">answer<\/a>, it is said that <\/p>\n\n<blockquote>\n  <p>one <strong>epoch<\/strong> = one forward pass and one backward pass of all the training examples<\/p>\n  \n  <p>number of <strong>iterations<\/strong> = number of passes, each pass using [batch size] number of examples. <\/p>\n  \n  <p>Example: if you have 1000 training examples, and your batch size is 500, then it will take 2 iterations to complete 1 epoch.<\/p>\n<\/blockquote>\n\n<p>Conclude this: .<\/p>\n\n<p><strong>However<\/strong>, the following <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"nofollow noreferrer\">example<\/a>, as I understand, is different from the previous conclusion. <\/p>\n\n<pre><\/pre>\n\n<p>Here, the <em>iteration<\/em> is <strong>60<\/strong> and number of <em>epoch<\/em> is set to <strong>1<\/strong>, which confused me a lot. It seems like, there are 60 <em>iterations<\/em>, as stated . And for each of the <em>iteration<\/em>, one <em>epoch<\/em> was done as stated  for each for-loop. And again, here, there is a . So what does the <em>iteration<\/em> exactly mean?<\/p>\n\n<p>Anyone can explain the difference between <em>iteration<\/em> and <em>epoch<\/em> <strong>based on this example<\/strong>?<\/p>\n","answers":[{"AnswerId":"42116354","CreationDate":"2017-02-08T14:52:26.450","ParentId":null,"OwnerUserId":"6397107","Title":null,"Body":"<p>In this case iteration using just for displaing intermediate result.\nWe can delete this code:<\/p>\n\n<pre><code>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\n<p>and instead:<\/p>\n\n<pre><code>for iteration in range(1, 60):\n  model.fit(X, y, batch_size=128, nb_epoch=1)\n<\/code><\/pre>\n\n<p>write:<\/p>\n\n<pre><code>model.fit(X, y, batch_size=128, nb_epoch=60)\n<\/code><\/pre>\n"},{"AnswerId":"36925269","CreationDate":"2016-04-28T20:54:17.343","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>I think that in this example the <strong>iteration<\/strong> means something different : you are iterating through the learning process and after every epoch you are doing something with partially learnt model. You are doing it iteratievly and that's why an <strong>iteration<\/strong> word is used.<\/p>\n"}]}
{"QuestionId":36135266,"AnswerCount":0,"Tags":"<docker><windows-7><ipython><tensorflow><jupyter-notebook>","CreationDate":"2016-03-21T15:31:23.063","AcceptedAnswerId":null,"OwnerUserId":6073688.0,"Title":"docker windows 7 execute ipython tensorflow","Body":"<p>I recently installed docker on my windows 7 machine. I understand that docker is the way to go if I want to run tensorflow on windows machine. So  I have also downloaded tensorflow image using docker and I can run it from docker container. <\/p>\n\n<p>Now I have bunch of ipython\/.ipynb files. I don't know how I can execute them in docker. The ipython files import tensorflow. <\/p>\n\n<p>I consider myself beginner. So the answers may be quite simple.<\/p>\n","answers":[]}
{"QuestionId":36136562,"AnswerCount":2,"Tags":"<python><keras><recurrent-neural-network>","CreationDate":"2016-03-21T16:27:42.973","AcceptedAnswerId":36305644.0,"OwnerUserId":5606352.0,"Title":"Wrong number of dimensions on model.fit","Body":"<p>I'm trying to run this SimpleRNN:<\/p>\n\n<pre><\/pre>\n\n<p>The error is on model.fit, as you can see below:<\/p>\n\n<pre><\/pre>\n\n<p>The error is telling me it's got the wrong number of dimensions, it should be 3 and it has only got 2. <strong>What are the dimensions it is referring to?<\/strong><\/p>\n","answers":[{"AnswerId":"36159529","CreationDate":"2016-03-22T16:08:43.360","ParentId":null,"OwnerUserId":"179372","Title":null,"Body":"<p>The error is probably because your <strong>input dimensions<\/strong> are not in the format of:<\/p>\n\n<pre><code>(nb_samples, timesteps, input_dim)\n<\/code><\/pre>\n\n<p>It is expecting 3 dimensions, and you're providing only 2 of them <code>(88,88)<\/code>.<\/p>\n"},{"AnswerId":"36305644","CreationDate":"2016-03-30T10:07:14.230","ParentId":null,"OwnerUserId":"6099650","Title":null,"Body":"<p>You are trying to run a RNN. This means that you want to <strong>include previous time steps<\/strong> in your calculation. In order to do so, you have to preprocess your data before giving it to the SimpleRNN layer.<\/p>\n\n<p>For simplicity, let us assume that instead of 88 samples with 88 features each you have 8 samples with 4 features each. Now, when using a RNN you will have to decide on a maximum for the backpropagation (i.e. number of previous time steps that are included in the calculation). In this case, you could choose to include a maximum of 2 previous time steps. Therefore, for the calculation of the weights of the RNN you will have to provide at each time step the input of the current time step (with its 4 features) and the input of the 2 previous time steps (with 4 features each). Just like in this visualization:<\/p>\n\n<pre><code>sequence    sample0  sample1  sample2  sample3  sample4  sample5  sample6 sample7       \n   0        |-----------------------|\n   1                 |-----------------------|\n   2                          |-----------------------|\n   3                                   |-----------------------|\n   4                                             |----------------------|\n   5                                                      |----------------------|\n<\/code><\/pre>\n\n<p>So instead of giving a (nb_samples, nb_features) matrix as an input to the SimpleRNN, you will have to give it a (nb_sequences, nb_timesteps, nb_features) shaped input. In this example, it means that instead of giving a (8x4) input you give it a (5x3x4) input.<\/p>\n\n<p>The keras <a href=\"http:\/\/keras.io\/layers\/embeddings\/\">Embedding<\/a> layer might do this job but in this case you can also write a short code for it:<\/p>\n\n<pre><code>input = np.random.rand(8,4)\nnb_timesteps = 3    # 2 (previous) + 1 (current)\nnb_sequences = input.shape[0] - nb_timesteps    #8-3=5\n\ninput_3D = np.array([input[i:i+nb_timesteps] for i in range(nb_sequences)])\n<\/code><\/pre>\n"}]}
{"QuestionId":36137636,"AnswerCount":3,"Tags":"<python><linux><ubuntu><anaconda><tensorflow>","CreationDate":"2016-03-21T17:17:11.277","AcceptedAnswerId":36138102.0,"OwnerUserId":3320135.0,"Title":"Installing packages with Anaconda","Body":"<p>On Ubuntu 14.04, I have installed Anaconda, which I use as my main Python interpreter. I now want to install the TensorFlow library and use this via Anaconda. So, I downloaded the relevant  file from the TensorFlow website, and then ran . After this, I ran , and it showed me  indicating that it was installed successfully.<\/p>\n\n<p>However, using the Anaconda interpreter, when I run a Python file which has , it tells me . Additionally, if I search my Anaconda directory, there is no reference to TensorFlow.<\/p>\n\n<p>Now at first, I thought this was because  was using  that comes with the native Ubuntu installation. However, I have the line  in my  file, and so this suggests it would use Anaconda's pip.<\/p>\n\n<p>Any idea what's going on? Thanks!<\/p>\n","answers":[{"AnswerId":"40056602","CreationDate":"2016-10-15T07:59:14.540","ParentId":null,"OwnerUserId":"6929661","Title":null,"Body":"<p>First uninstall all the dependencies of tensorflow using<\/p>\n\n<pre><code>pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>Then install tensorflow package with conda run:<\/p>\n\n<pre><code> conda install -c jjhelmus tensorflow=0.10.0rc0\n<\/code><\/pre>\n\n<p>If you want to install tensorflow package with pip run:<\/p>\n\n<pre><code>pip install -i https:\/\/pypi.anaconda.org\/jjhelmus\/simple tensorflow\n<\/code><\/pre>\n\n<p>Sources: <a href=\"https:\/\/anaconda.org\/jjhelmus\/tensorflow\" rel=\"nofollow\">https:\/\/anaconda.org\/jjhelmus\/tensorflow<\/a><\/p>\n"},{"AnswerId":"36240986","CreationDate":"2016-03-26T21:15:47.727","ParentId":null,"OwnerUserId":"3568373","Title":null,"Body":"<p>Try without <code>sudo<\/code>:<\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>instead of<\/p>\n\n<pre><code>sudo pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.7.1-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"},{"AnswerId":"36138102","CreationDate":"2016-03-21T17:42:30.420","ParentId":null,"OwnerUserId":"5573572","Title":null,"Body":"<p>You can try the similar answer here:\n<a href=\"https:\/\/stackoverflow.com\/a\/33698750\/5573572\">https:\/\/stackoverflow.com\/a\/33698750\/5573572<\/a><\/p>\n\n<p>Pretty much do these steps:<\/p>\n\n<pre><code>1. Uninstall TensorFlow from pip:\n\npip uninstall tensorflow\n<\/code><\/pre>\n\n<p>Do the above to avoid conflicts.<\/p>\n\n<pre><code>2. Install Python 3 in a virtual environment (version 0.7.1 as of this writing):\n\nconda create -n &lt;environment_name&gt; python==3.5.1\n\n3. Activate your virtual environment (do this every time you want to use TensorFlow):\n\nsource activate &lt;environment_name&gt;\n\n4. Install a Conda version of TensorFlow in that environment (version 0.7.1 as of this writing):\n\nconda install -c https:\/\/conda.anaconda.org\/jjhelmus tensorflow\n<\/code><\/pre>\n\n<p>Remember to change \"environment_name\" to whatever you want to name your environment. After these, you should hopefully be able to import tensorflow. If not, then anaconda might be having trouble installing TensorFlow's dependencies. <strike>I'll run this on my machine to check real quick :p.<\/strike> <strong>I have confirmed that this works.<\/strong><\/p>\n\n<p>A possible reason that your installation attempt was not working is because Ubuntu 14.04 has Python 2.7 installed, in which many system programs depend on for the time being. As an aside, the Ubuntu development team is working on porting all of those programs to use Python 3 instead:\n<a href=\"https:\/\/wiki.ubuntu.com\/Python\/Python35Transition\" rel=\"nofollow noreferrer\">https:\/\/wiki.ubuntu.com\/Python\/Python35Transition<\/a><\/p>\n\n<p><strong>Update<\/strong>: added instructions to include creating a virtual environment. The virtual environment helps because it allows you to use the Python commands within the environment instead of any system Python commands. So, commands like \"pip\" and \"python\" will use the ones in the environment, which also contains the TensorFlow libraries. To get out of the environment, do:<\/p>\n\n<pre><code>source deactivate\n<\/code><\/pre>\n"}]}
{"QuestionId":36139508,"AnswerCount":1,"Tags":"<networking><theano><convolution><keras>","CreationDate":"2016-03-21T19:03:50.467","AcceptedAnswerId":null,"OwnerUserId":6095208.0,"Title":"CNN - Model predicting all the same class","Body":"<p>I have read some of the other answers to a similar problem on stackoverflow, however I have found nothing to help me in this case. I have a set of 539 RGB images, dimensions 607 x 607 x 3, and each image is one of 6 classes. I have had success with the MNIST and CIFAR10 datasets, however when I create CNNs for this dataset, the test val_acc produced when training is remaining constant \/ produced by predicting all the same classes (which may vary). Below I have included my code and an example CNN, as well as the output on the GPU:<\/p>\n\n<pre><\/pre>\n\n<p>And the output for the first 9 epochs:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried changing the number of nodes in the hidden layer, creating more complicated models, changing activation functions, everything I can think of. If I were to run the CIFAR10 dataset through this (and change the last layer to Dense(10) rather than Dense(6)) I get successful results - Not sure if there is a problem with the data I am importing, however the np.shape structure for my data is exactly the same as the np.shape structure for the CIFAR10 dataset<\/p>\n","answers":[{"AnswerId":"36141876","CreationDate":"2016-03-21T21:23:44.127","ParentId":null,"OwnerUserId":"6089914","Title":null,"Body":"<p>val_acc is not constant, it's jumping back and forth. Try lowering the learning rate 10 times then 10 times again until it starts learning. You will have to create an RMSProp object and pass the object in instead of a string.<\/p>\n"}]}
{"QuestionId":36139889,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-21T19:27:36.860","AcceptedAnswerId":36192590.0,"OwnerUserId":349760.0,"Title":"RNN model running out of memory in TensorFlow","Body":"<p>I implemented a Sequence to Sequence model using the rnn.rnn helper in TensorFlow.<\/p>\n\n<pre><\/pre>\n\n<p>The model is running out of memory on a Titan X with 16 GB of memory while allocating gradients for the LSTM cells:<\/p>\n\n<pre><\/pre>\n\n<p>If I reduce the length of the input and output sequences to 4 or less the model runs without a problem.<\/p>\n\n<p>This indicates to me that TF is trying to allocate the gradients for all time steps at the same time. Is there a way of avoiding this?<\/p>\n","answers":[{"AnswerId":"36192590","CreationDate":"2016-03-24T03:05:47.813","ParentId":null,"OwnerUserId":"349760","Title":null,"Body":"<p>The function <code>tf.gradients<\/code> as well as the <code>minimize<\/code> method of the optimizers allow you to set parameter called <code>aggregation_method<\/code>. The default value is <code>ADD_N<\/code>. This method constructs the graph in such a way that all gradients need to be computed at the same time. <\/p>\n\n<p>There are two other undocumented methods called  <code>tf.AggregationMethod.EXPERIMENTAL_TREE<\/code> and  <code>tf.AggregationMethod.EXPERIMENTAL_ACCUMULATE_N<\/code>, which do not have this requirement.<\/p>\n"}]}
{"QuestionId":36139980,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2016-03-21T19:33:25.577","AcceptedAnswerId":null,"OwnerUserId":6095189.0,"Title":"Prevention of overfitting in convolutional layers of a CNN","Body":"<p>I'm using TensorFlow to train a Convolutional Neural Network (CNN) for a sign language application.  The CNN has to classify 27 different labels, so unsurprisingly, a major problem has been addressing overfitting.  I've taken several steps to accomplish this:<\/p>\n\n<ol>\n<li>I've collected a large amount of high-quality training data (over 5000 samples per label).<\/li>\n<li>I've built a reasonably sophisticated pre-processing stage to help maximize invariance to things like lighting conditions.<\/li>\n<li>I'm using dropout on the fully-connected layers.<\/li>\n<li>I'm applying L2 regularization to the fully-connected parameters.<\/li>\n<li>I've done extensive hyper-parameter optimization (to the extent possible given HW and time limitations) to identify the simplest model that can achieve close to 0% loss on training data.<\/li>\n<\/ol>\n\n<p>Unfortunately, even after all these steps, I'm finding that I can't achieve much better that about 3% test error.  (It's not terrible, but for the application to be viable, I'll need to improve that substantially.)<\/p>\n\n<p>I suspect that the source of the overfitting lies in the convolutional layers since I'm not taking any explicit steps there to regularize (besides keeping the layers as small as possible).  But based on examples provided with TensorFlow, it doesn't appear that regularization or dropout is typically applied to convolutional layers.<\/p>\n\n<p>The only approach I've found online that explicitly deals with prevention of overfitting in convolutional layers is a fairly new approach called <a href=\"http:\/\/www.matthewzeiler.com\/pubs\/iclr2013\/iclr2013.pdf\" rel=\"noreferrer\">Stochastic Pooling<\/a>.  Unfortunately, it appears that there is no implementation for this in TensorFlow, at least not yet.<\/p>\n\n<p>So in short, is there a recommended approach to prevent overfitting in convolutional layers that can be achieved in TensorFlow?  Or will it be necessary to create a custom pooling operator to support the Stochastic Pooling approach?<\/p>\n\n<p>Thanks for any guidance!<\/p>\n","answers":[{"AnswerId":"42532658","CreationDate":"2017-03-01T13:08:45.890","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<h2>How can I fight overfitting?<\/h2>\n\n<ul>\n<li>Get more data (or data augmentation)<\/li>\n<li>Dropout (see <a href=\"https:\/\/arxiv.org\/abs\/1207.0580\" rel=\"noreferrer\">paper<\/a>, <a href=\"https:\/\/www.cs.toronto.edu\/~hinton\/absps\/JMLRdropout.pdf\" rel=\"noreferrer\">explanation<\/a>, <a href=\"https:\/\/datascience.stackexchange.com\/q\/16045\/8820\">dropout for cnns<\/a>)<\/li>\n<li>DropConnect<\/li>\n<li>Regularization (see <a href=\"https:\/\/arxiv.org\/pdf\/1707.09725.pdf#page=134\" rel=\"noreferrer\">my masters thesis<\/a>, page 85 for examples)<\/li>\n<li>Feature scale clipping<\/li>\n<li>Global average pooling<\/li>\n<li>Make network smaller<\/li>\n<li>Early stopping<\/li>\n<\/ul>\n\n<h2>How can I improve my CNN?<\/h2>\n\n<blockquote>\n  <p>Thoma, Martin. \"<a href=\"https:\/\/arxiv.org\/pdf\/1707.09725.pdf\" rel=\"noreferrer\">Analysis and Optimization of Convolutional Neural Network Architectures<\/a>.\" arXiv preprint arXiv:1707.09725 (2017).<\/p>\n<\/blockquote>\n\n<p>See chapter 2.5 for analysis techniques. As written in the beginning of that chapter, you can usually do the following:<\/p>\n\n<ul>\n<li>(I1) Change the problem definition (e.g., the classes which are to be distinguished)<\/li>\n<li>(I2) Get more training data<\/li>\n<li>(I3) Clean the training data<\/li>\n<li>(I4) Change the preprocessing (see Appendix B.1)<\/li>\n<li>(I5) Augment the training data set (see Appendix B.2)<\/li>\n<li>(I6) Change the training setup (see Appendices B.3 to B.5)<\/li>\n<li>(I7) Change the model (see Appendices B.6 and B.7)<\/li>\n<\/ul>\n\n<h2>Misc<\/h2>\n\n<blockquote>\n  <p>The CNN has to classify 27 different labels, so unsurprisingly, a major problem has been addressing overfitting.<\/p>\n<\/blockquote>\n\n<p>I don't understand how this is connected. You can have hundreds of labels without a problem of overfitting.<\/p>\n"}]}
{"QuestionId":36145065,"AnswerCount":0,"Tags":"<neural-network><tensorflow><conv-neural-network><autoencoder>","CreationDate":"2016-03-22T02:14:07.067","AcceptedAnswerId":null,"OwnerUserId":6000679.0,"Title":"Convolutional autoencoder not learning meaningful filters","Body":"<p>I am playing with TensorFlow to understand convolutional autoencoders. I have implemented a simple single-layer autoencoder which does this:<\/p>\n\n<p>Input (Dimension: 95x95x1) ---> Encoding (convolution with 32 5x5 filters) ---> Latent representation (Dimension: 95x95x1x32) ---> Decoding (using tied weights) ---> Reconstructed input (Dimension: 95x95x1) <\/p>\n\n<p>The inputs are black-and-white edge images i.e. the results of edge detection on RGB images.<\/p>\n\n<p>I initialised the filters randomly and then trained the model to minimise loss, where loss is defined as the mean-squared-error of the input and the reconstructed input.<\/p>\n\n<pre><\/pre>\n\n<p>After training with 1000 steps, my loss converges and the network is able to reconstruct the images well. However, when I visualise the learned filters, they do not look very different from the randomly-initialised filters! But the values of the filters change from training step to training step.<\/p>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/YrVqG.png\" rel=\"nofollow\">Example of learned filters<\/a><\/p>\n\n<p>I would have expected at least horizontal and vertical edge filters. Or if my network was learning \"identity filters\" I would have expected the filters to all be white or something?<\/p>\n\n<p>Does anyone have any idea about this? Or are there any suggestions as to what I can do to analyse what is happening? Should I include pooling and depooling layers before decoding?<\/p>\n\n<p>Thank you!<\/p>\n\n<p>P\/S: I tried the same model on RGB images and again the filters look random (like random blotches of colours).<\/p>\n","answers":[]}
{"QuestionId":36145808,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-22T03:53:25.767","AcceptedAnswerId":36145962.0,"OwnerUserId":6093883.0,"Title":"In tensorflow, is variable value the only context information a session stores?","Body":"<p>Suppose we run , is it true that the only resident information is the tf.Variable values, and all other evaluations results are either returned from the call, or discarded? <\/p>\n\n<p>Suppose we have two sessions, do they share anything except for the default graph?<\/p>\n","answers":[{"AnswerId":"36145962","CreationDate":"2016-03-22T04:13:00.540","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are various forms of state retained between <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.run\" rel=\"noreferrer\"><code>run()<\/code><\/a> calls (\"steps\") in a TensorFlow <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\" rel=\"noreferrer\">session<\/a>:<\/p>\n\n<ul>\n<li>As you already pointed out, <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> objects store values between invocations, which may be read and written by any step.<\/li>\n<li>TensorFlow queues (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/threading_and_queues\/index.html\" rel=\"noreferrer\">tutorial<\/a>) allow you to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.enqueue\" rel=\"noreferrer\">enqueue one<\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.enqueue_many\" rel=\"noreferrer\">many<\/a> values into a bounded buffer in a step, and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.dequeue\" rel=\"noreferrer\">dequeue one<\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.dequeue_many\" rel=\"noreferrer\">many<\/a> values in a later step. Queues also support coordination between steps, such as back-pressure in a producer\/consumer relationship. They also allow you to switch between element-by-element and batched computation, using functions like <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#batch\" rel=\"noreferrer\"><code>tf.train.batch()<\/code><\/a>.<\/li>\n<li>TensorFlow readers (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html\" rel=\"noreferrer\">tutorial<\/a>) act like implicit file pointers that remember their current position in a file between steps. Subsequent executions of the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#ReaderBase.read\" rel=\"noreferrer\"><code>read()<\/code><\/a> op yield, for examples, different <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#TextLineReader\" rel=\"noreferrer\">lines of a text file<\/a>.<\/li>\n<\/ul>\n\n<p>In the single-process version of TensorFlow, sessions do not share any state. They may share the same graph (if they are both created with the same default graph), but the stateful components (such as the <code>tf.Variable<\/code> objects) will assume different values in the different sessions.<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/distributed_runtime\/README.md\" rel=\"noreferrer\">distributed runtime<\/a> adds support for \"resource containers\" that are shared between sessions. These cover variables, queues, and readers, and can be configured by passing the optional <code>container<\/code> argument to the constructor for these objects.<\/p>\n"}]}
{"QuestionId":36148900,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-03-22T07:57:09.683","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Why using drop_out layer with tensorflow cause memory allocation error?","Body":"<p>I have a simple network with a few layers (3 conv + 1 fully connected). The last convolution layer output a tensor of shape 32x32x192 and is followed by a fully connected layer of output dim 4096. Without dropout layer, this network have a normal memory footprint. No more than 2G on my PC. But when I had a dropout layer between the last convolution and the fully connected layer, there's a memory allocation error. The total memory of my PC is 16G.<\/p>\n","answers":[]}
{"QuestionId":36150834,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-22T09:40:16.363","AcceptedAnswerId":36156697.0,"OwnerUserId":6093883.0,"Title":"How does TensorFlow name tensors?","Body":"<p>I wonder if this is the correct understanding: <\/p>\n\n<p>All tensors are derived from some operation, and operations are either given a name in the constructor, or given the default name for a particular kind of operation. If the name is not unique, TensorFlow automatically handles this by appending , , etc. An operation with n tensor outputs name these tensors , , ..., .<\/p>\n\n<p>One problem seems to arise: if  is a , then  gives . This is confusing: to what does  refer?<\/p>\n","answers":[{"AnswerId":"36156697","CreationDate":"2016-03-22T14:09:59.393","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Your observations on <code>Tensor<\/code> naming are absolutely correct: the name of a <code>Tensor<\/code> is the concatenation of <\/p>\n\n<ol>\n<li>the name of the operation that produced it, <\/li>\n<li>a colon (<code>:<\/code>), and <\/li>\n<li>the index of that tensor in the outputs of the operation that produced it. <\/li>\n<\/ol>\n\n<p>Therefore the tensor named <code>\"foo:2\"<\/code> is the output of the op named <code>\"foo\"<\/code> at position 2 (with indices starting from zero).<\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable.name\" rel=\"noreferrer\">naming<\/a> of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> objects is slightly strange. Every <code>tf.Variable<\/code> contains a <em>mutable tensor<\/em> object that holds the state of the variable (and a few other tensors). A <code>\"Variable\"<\/code> op (which has the name <code>\"variable_name\"<\/code> in your example) \"produces\" this mutable tensor each time it is run as its 0th output, so the name of the mutable tensor is <code>\"variable_name:0\"<\/code>.<\/p>\n\n<p>Since a <code>tf.Variable<\/code> is mostly indistinguishable from a <code>tf.Tensor<\/code>&mdash;in that it can be used in the same places&mdash;we took the decision to make variable names resemble tensor names, so the <code>Variable.name<\/code> property returns the name of the mutable tensor. (This contrasts with <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase\" rel=\"noreferrer\"><code>tf.QueueBase<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#ReaderBase\" rel=\"noreferrer\"><code>tf.ReaderBase<\/code><\/a> objects, which are <em>not<\/em> usable directly as tensors (instead you have to call methods on them to create ops that operate on their state), so these do not have a tensor-like name.)<\/p>\n"}]}
{"QuestionId":36151690,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-03-22T10:19:46.730","AcceptedAnswerId":36151984.0,"OwnerUserId":3145671.0,"Title":"tensorflow GradientDescentOptimizer function trains the paramter layer by layer?","Body":"<p>Does tensorflow's GradientDescentOptimizer trains the internal network layer by layer or just use a regular method like Back propagation. How deep the network could be for GradientDescentOptimizer?<\/p>\n","answers":[{"AnswerId":"36151984","CreationDate":"2016-03-22T10:31:55.210","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>It just uses standard back prop, not layer by layer. <\/p>\n\n<p>The network could be theoretically as deep as you can hold in memory. But the deeper it gets the longer it will take to train and the less useful successive layers will be. At a certain point extra layers can even start to increase the error on training data.<\/p>\n"}]}
{"QuestionId":36152571,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-22T10:58:14.227","AcceptedAnswerId":36159461.0,"OwnerUserId":1611353.0,"Title":"Tensorflow reshape tensor gives None dimension","Body":"<p>I have used the model described <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">here<\/a> on the 0.6.0 branch. The code can be found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0.6.0\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow\">here<\/a>. I have done some minor changes to the linked code.<\/p>\n\n<p>In my code I create two models, one for training and one for validation, very similar as it is done in the Tensorflow Tutorial.<\/p>\n\n<pre><\/pre>\n\n<p>The first model, the one for training, seems to be created just fine, but the second, used for validation, does not. The output gets a None dimension! The row I'm refering to is on row 134 in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0.6.0\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow\">linked<\/a> code:<\/p>\n\n<p><\/p>\n\n<p>I've added these lines right after the reshape of the output:<\/p>\n\n<pre><\/pre>\n\n<p>and that gives me this:<\/p>\n\n<pre><\/pre>\n\n<p>This problem only happens with the 'validation model', not with the 'training model'. For the 'training model' I get expected output:<\/p>\n\n<pre><\/pre>\n\n<p>(Note that with the 'validation model' I use a  instead of  that I use for the training model)<\/p>\n\n<p>From what I understand, using  as input to the  function, will figure the output shape out automagically! But then why do I get None? Nothing in my config fed to the model has a None value.<\/p>\n\n<p>Thank you for all the help and tips!<\/p>\n","answers":[{"AnswerId":"36159461","CreationDate":"2016-03-22T16:05:35.810","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> A dimension being <code>None<\/code> simply means that shape inference could not determine an exact shape for the <code>output<\/code> tensor, <em>at graph-building time<\/em>. When you run the graph, the tensor will have the appropriate <em>run-time<\/em> shape.<\/p>\n\n<p>If you're not interested in how shape inference works, you can stop reading now.<\/p>\n\n<p>Shape inference applies local rules, based on a \"shape function\" that takes the shapes of the inputs to an operation and computes (possibly incomplete) shapes for the outputs of an operation. To figure out why <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#reshape\" rel=\"noreferrer\"><code>tf.reshape()<\/code><\/a> gives an incomplete shape, we have to look at its inputs, and work backwards:<\/p>\n\n<ul>\n<li>The <code>shape<\/code> argument to <code>tf.reshape()<\/code> includes a <code>[-1]<\/code>, which means \"figure the output shape automagically\" based on the shape of the <code>tensor<\/code> input.<\/li>\n<li>The <code>tensor<\/code> input is the output of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#concat\" rel=\"noreferrer\"><code>tf.concat()<\/code><\/a> on the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4be56f381cd000e91f79209aaf150636db6fb840\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L134\" rel=\"noreferrer\">same line<\/a>.<\/li>\n<li>The inputs to <code>tf.concat()<\/code> are computed by a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#mul\" rel=\"noreferrer\"><code>tf.mul()<\/code><\/a> in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/fd464caaa40cfa16c81712939e39bd14d88c6fd4\/tensorflow\/python\/ops\/rnn_cell.py#L209\" rel=\"noreferrer\"><code>BasicLSTMCell.__call__()<\/code><\/a>. The <code>tf.mul()<\/code> op multiplies the result of a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#tanh\" rel=\"noreferrer\"><code>tf.tanh()<\/code><\/a> and a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#sigmoid\" rel=\"noreferrer\"><code>tf.sigmoid()<\/code><\/a> op.<\/li>\n<li>The <code>tf.tanh()<\/code> op produces an output of size <code>[?, hidden_size]<\/code>, and the <code>tf.sigmoid()<\/code> op produces an output of size <code>[batch_size, hidden_size]<\/code>.<\/li>\n<\/ul>\n\n<p>The <code>tf.mul()<\/code> op performs <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/user\/basics.broadcasting.html\" rel=\"noreferrer\">NumPy-style broadcasting<\/a>. A dimension will only be broadcast if it has size 1. Consider three cases where we compute <code>tf.mul(x, y)<\/code>:<\/p>\n\n<ol>\n<li>If <code>x<\/code> has shape <code>[1, 10]<\/code>, and <code>y<\/code> has shape <code>[5, 10]<\/code>, then broadcasting will happen, and the output shape will be <code>[5, 10]<\/code>.<\/li>\n<li>If <code>x<\/code> has shape <code>[1, 10]<\/code>, and <code>y<\/code> has shape <code>[1, 10]<\/code>, then there will be no broadcasting, and the output shape will be <code>[1, 10]<\/code>.<\/li>\n<li>However, if <code>x<\/code> has shape <code>[1, 10]<\/code>, and <code>y<\/code> has shape <code>[?, 10]<\/code>, there is insufficient static information to tell whether broadcasting will happen (even though we happen to know that case 2 applies at runtime).<\/li>\n<\/ol>\n\n<p>Therefore, when <code>batch_size<\/code> is 1, the <code>tf.mul()<\/code> op produces an output with the shape <code>[?, hidden_size]<\/code>; but when <code>batch_size<\/code> is greater than 1, the output shape is <code>[batch_size, hidden_size]<\/code>.<\/p>\n\n<p>Where shape inference breaks down, it can be appropriate to use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Tensor.set_shape\" rel=\"noreferrer\"><code>Tensor.set_shape()<\/code><\/a> method to add information. This would potentially be useful in the <code>BasicLSTMCell<\/code> implementation, where we know more than it is possible to infer about the shapes of the outputs.<\/p>\n"}]}
{"QuestionId":36155397,"AnswerCount":0,"Tags":"<mex><caffe><solver><matcaffe>","CreationDate":"2016-03-22T13:16:21.567","AcceptedAnswerId":null,"OwnerUserId":4708607.0,"Title":"Matcaffe training net produces \"Data layer prefetch queue empty\"","Body":"<p>I'm trying to figure out why my MatCaffe implementation cannot pop from my train lmdb, which I've created using convert_imageset.bin.<\/p>\n\n<p><strong>What I do is basically just this:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>and when looking at the terminal, the output after the last statement is this:<\/p>\n\n<pre><\/pre>\n\n<p>I can reproduce this problem even when I delete the locks.mdb to make sure that no locks are left when I restart this procedure. After the message I only can do a hard Matlab shutdown.\nI've checked the lmdb with Matlab LMDB and the contents of both, my train and test lmdb seem to be ok. Parameters I've used to generate the lmdb: shuffle.<\/p>\n\n<p>Note (might be the source of problems here): currently I'm facing MEX problems with this constellation. On the first run of my implementation I get the error message<\/p>\n\n<p>\"<strong>Unexpected unknown exception from MEX file..<\/strong>\"<\/p>\n\n<p>for which the terminal output looks like this:<\/p>\n\n<pre><\/pre>\n\n<p><strong>What I've tried so far:<\/strong>\nI implemented a try-catch block so that the pointers, spaces etc. are freed (hopefully) by using 'caffe.reset_all();\" so that in ANY case this method is called.\nOn the second run, I get the above mentioned output. It seems that my first run blocks the lmdb access, what led me to delete locks.mdb manually between the first and the second run -> same effect unfortunately. A \"manual\" train by command line does work with the same lmdb's. Only the matcaffe run seems to raise these problems and questions. Note that I want to use Matcaffe for manual initialization of my weights for the layers - \"weight_filler\" in .prototxt isn't an option.\nMy MatCaffe implementation is from January 2016 and I've also recompiled the mex-file for caffe_ with the correct gcc version (before it gave me the warning that my gcc version should be \"x\" -> changed to \"x\" and recompiled).<\/p>\n\n<p>Do you have any other ideas, recommendations or inputs please?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
{"QuestionId":36158058,"AnswerCount":2,"Tags":"<csv><export-to-csv><torch>","CreationDate":"2016-03-22T15:07:12.953","AcceptedAnswerId":36195768.0,"OwnerUserId":3957707.0,"Title":"Torch : Save tensor to csv file","Body":"<p>I have been working with Torch. And my current program requires to export a Tensor containing reduced feature matrix. \nI tried doing the following:<\/p>\n\n<pre><\/pre>\n\n<p>and the output was: <\/p>\n\n<pre><\/pre>\n\n<p>Expected output: <\/p>\n\n<pre><\/pre>\n\n<p>I'm hoping somebody has an idea as to how I can do this?<\/p>\n","answers":[{"AnswerId":"36195768","CreationDate":"2016-03-24T08:19:20.983","ParentId":null,"OwnerUserId":"2011802","Title":null,"Body":"<p>When saving tensor, torch saves not only data but also -- as you can see -- several other useful information for later deserialisation.<\/p>\n\n<p>If you need csv serialisation, you are good to implement it yourself.<\/p>\n\n<p>Fortunately, this is very straightforward.<\/p>\n\n<p>Here is a quick example :<\/p>\n\n<pre><code>require 'torch'\n\nmatrix = torch.Tensor(5,3) -- a 5x3 matrix\n\nmatrix:random(1,10) -- matrix initialized with random numbers in [1,10]\n\nprint(matrix) -- let's see the matrix content\n\nsubtensor = matrix[{{1,3}, {2,3}}] -- let's create a view on the row 1 to 3, for which we take columns 2 to 3 (the view is a 3x2 matrix, note that values are bound to the original tensor)\n\nlocal out = assert(io.open(\".\/dump.csv\", \"w\")) -- open a file for serialization\n\nsplitter = \",\"\nfor i=1,subtensor:size(1) do\n    for j=1,subtensor:size(2) do\n        out:write(subtensor[i][j])\n        if j == subtensor:size(2) then\n            out:write(\"\\n\")\n        else\n            out:write(splitter)\n        end\n    end\nend\n\nout:close()\n<\/code><\/pre>\n\n<p>The output on my computer for the matrix is :<\/p>\n\n<pre><code> 10  10   6\n  4   8   3\n  3   8   5\n  5   5   5\n  1   6   8\n[torch.DoubleTensor of size 5x3]\n<\/code><\/pre>\n\n<p>and the file dumped content :<\/p>\n\n<pre><code>10,6\n8,3\n8,5\n<\/code><\/pre>\n\n<p>HTH<\/p>\n"},{"AnswerId":"41618021","CreationDate":"2017-01-12T16:15:27.743","ParentId":null,"OwnerUserId":"848503","Title":null,"Body":"<p>You can first convert the tensor to a Lua table using <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/ff11731\/doc\/utility.md#table-torchtotableobject\" rel=\"nofollow noreferrer\">torch.totable<\/a>. Then use the <a href=\"https:\/\/github.com\/clementfarabet\/lua---csv\" rel=\"nofollow noreferrer\">csvigo<\/a> library to save the table as a csv file. This may be a workaround but I haven't had any problem with it.<\/p>\n"}]}
{"QuestionId":36159194,"AnswerCount":5,"Tags":"<python><ubuntu><cuda><tensorflow>","CreationDate":"2016-03-22T15:54:18.623","AcceptedAnswerId":36161286.0,"OwnerUserId":3559686.0,"Title":"TensorFlow: libcudart.so.7.5: cannot open shared object file: No such file or directory","Body":"<p>I am running TensorFlow on Ubuntu 15.10. When I enter , I see that TF has been installed properly.<\/p>\n\n<p>However, when I write , I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>For what it's worth, I have followed the instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#optional-linux-enable-gpu-support\" rel=\"noreferrer\">here<\/a> and set my  and  environment variables.<\/p>\n\n<p>Any advice?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I have installed CUDA 7.5 and added these to my  file:<\/p>\n\n<pre><\/pre>\n\n<p>However, I continue to see the same error message.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I see the following output when I run :<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>If it is relevant, I use GeForce GT640.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I followed <a href=\"https:\/\/stackoverflow.com\/a\/36161286\/3559686\">@tommus' advice<\/a> and called  before running TensorFlow, and now it works like a charm.<\/p>\n\n<p>Thanks to everyone who tried to help me in the comments -- this is my first experience with any kind of Linux distribution, and I am really <strong>very appreciative<\/strong> of all the patient assistance I have received so far :-) You guys are absolutely awesome!<\/p>\n","answers":[{"AnswerId":"36161286","CreationDate":"2016-03-22T17:29:11.193","ParentId":null,"OwnerUserId":"2057291","Title":null,"Body":"<p>After editing <code>.profile<\/code> file you either need to log out and log back in or run the following command:<\/p>\n\n<p><code>source ~\/.profile<\/code><\/p>\n\n<p>The solution is persistent so there is no need to perform it ever again.<\/p>\n"},{"AnswerId":"45875834","CreationDate":"2017-08-25T06:48:37.677","ParentId":null,"OwnerUserId":"8515096","Title":null,"Body":"<p><code>sudo nano \/etc\/ld.so.conf.d\/nvidia.conf<\/code><\/p>\n\n<p>Add this:<\/p>\n\n<p><code>\/usr\/local\/cuda-8.0\/lib64<\/code><\/p>\n\n<p><code>\/usr\/local\/cuda-8.0\/lib<\/code><\/p>\n\n<p>run <\/p>\n\n<p><code>sudo ldconfig<\/code><\/p>\n\n<p>if it works well ,please tell me.<\/p>\n"},{"AnswerId":"48254800","CreationDate":"2018-01-14T22:28:40.343","ParentId":null,"OwnerUserId":"5719764","Title":null,"Body":"<p>If you are using pip but other answers did not work for you, try this.<\/p>\n\n<p>In my case this exact problem was solved by <strong>reinstalling tensorflow<\/strong> using:<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>Pip must have messed up its dependencies or something when installing other packages after Tensorflow installation (I've run <code>pip install -r requirements.txt<\/code> afterwards for my own purposes --> that broke Tensorflow).<\/p>\n"},{"AnswerId":"51359616","CreationDate":"2018-07-16T10:22:49.323","ParentId":null,"OwnerUserId":"5815687","Title":null,"Body":"<p>I had same error. I have to have <code>Cuda 8.0<\/code> and <code>CuDNN 5.1<\/code> for running some experiments!! <\/p>\n\n<p>SOLUTION: Run python with <code>Sudo Python<\/code><\/p>\n\n<pre><code>python\nPython 2.7.15 |Anaconda, Inc.| (default, May 1 2018, 23:32:55)\n[GCC 7.2.0] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n\nimport tensorflow\nTraceback (most recent call last):\nFile \"\", line 1, in \nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/init.py\", line 22, in \nfrom tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import\nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/init.py\", line 49, in \nfrom tensorflow.python import pywrap_tensorflow\nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 74, in \nraise ImportError(msg)\nImportError: Traceback (most recent call last):\nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in \nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in \n_pywrap_tensorflow_internal = swig_import_helper()\nFile \"\/home\/ujjval\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>Failed to load the native <code>TensorFlow<\/code> runtime.<\/p>\n"},{"AnswerId":"45989104","CreationDate":"2017-08-31T19:53:04.150","ParentId":null,"OwnerUserId":"2487309","Title":null,"Body":"<p>You maybe need to update your cuDNN version as Tensorflow 1.3 doesn't work with cuDNN 5.1 or earlier. Download cuDNN 6.0 or upper and install to fix the issue<\/p>\n"}]}
{"QuestionId":36162180,"AnswerCount":3,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-03-22T18:16:38.643","AcceptedAnswerId":44225502.0,"OwnerUserId":859453.0,"Title":"Gradient Descent vs Adagrad vs Momentum in TensorFlow","Body":"<p>I'm studying <em>TensorFlow<\/em> and how to use it, even if I'm not an expert of neural networks and deep learning (just the basics).<\/p>\n\n<p>Following tutorials, I don't understand the real and practical differences between the three optimizers for loss. I look at the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#optimizers\" rel=\"noreferrer\">API<\/a> and I understand the principles, but my questions are:<\/p>\n\n<p><strong>1. When is it preferable to use one instead of the others ?<\/strong><\/p>\n\n<p><strong>2. Are there important differences to know ?<\/strong><\/p>\n","answers":[{"AnswerId":"52474058","CreationDate":"2018-09-24T06:56:40.883","ParentId":null,"OwnerUserId":"5728089","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/44225502\/5728089\">Salvador Dali's answer<\/a> already explains about the differences between some popular methods (i.e. optimizers), but I would try to elaborate on them some more.<br\/>\n(Note that our answers disagree about some points, especially regarding ADAGRAD.)<\/p>\n\n<h3>Classical Momentum (CM) vs Nesterov's Accelerated Gradient (NAG)<\/h3>\n\n<p>(Mostly based on section 2 in the paper <a href=\"http:\/\/proceedings.mlr.press\/v28\/sutskever13.pdf\" rel=\"noreferrer\">On the importance of initialization and momentum in deep learning<\/a>.)<br\/><\/p>\n\n<p>Each step in both CM and NAG is actually composed of two sub-steps:<\/p>\n\n<ul>\n<li>A momentum sub-step - This is simply a fraction (typically in the range <code>[0.9,1)<\/code>) of the last step.<\/li>\n<li>A gradient dependent sub-step - This is like the usual step in <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent#Iterative_method\" rel=\"noreferrer\">SGD<\/a> - it is the product of the learning rate and the vector opposite to the gradient, while the gradient is computed where this sub-step starts from.<\/li>\n<\/ul>\n\n<p>CM takes the gradient sub-step first, while NAG takes the momentum sub-step first.<\/p>\n\n<p>Here is a demonstration from <a href=\"https:\/\/stats.stackexchange.com\/a\/368179\/215801\">an answer about intuition for CM and NAG<\/a>:<br\/><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pAwIf.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pAwIf.png\" alt=\"CM vs NAG example\"><\/a><\/p>\n\n<p>So NAG seems to be better (at least in the image), but why?<br\/><\/p>\n\n<p>The important thing to note is that it doesn't matter when the momentum sub-step comes - it would be the same either way. Therefore, we might as well behave is if the momentum sub-step has already been taken.<br\/><\/p>\n\n<p>Thus, the question actually is: Assuming that the gradient sub-step is taken after the momentum sub-step, should we calculate the gradient sub-step as if it started in the position before or after taking the momentum sub-step?<br\/><\/p>\n\n<p>\"After it\" seems like the right answer, as generally, the gradient at some point <code>\u03b8<\/code> roughly points you in the direction from <code>\u03b8<\/code> to a minimum (with the relatively right magnitude), while the gradient at some other point is less likely to point you in the direction from <code>\u03b8<\/code> to a minimum (with the relatively right magnitude).<br\/><\/p>\n\n<p>Here is a demonstration (from the gif below):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/T7Ori.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T7Ori.png\" alt=\"CM vs NAG in a specific moment in the awesome gif\"><\/a><\/p>\n\n<ul>\n<li>The minimum is where the star is, and the curves are <a href=\"https:\/\/en.wikipedia.org\/wiki\/Contour_line\" rel=\"noreferrer\">contour lines<\/a>. (For an explanation about contour lines and why they are perpendicular to the gradient, see videos <a href=\"https:\/\/www.youtube.com\/watch?v=WsZj5Rb6do8\" rel=\"noreferrer\">1<\/a> and <a href=\"https:\/\/www.youtube.com\/watch?v=ZTbTYEMvo10\" rel=\"noreferrer\">2<\/a> by the legendary <a href=\"https:\/\/www.youtube.com\/c\/3blue1brown\" rel=\"noreferrer\">3Blue1Brown<\/a>.)<\/li>\n<li>The (long) purple arrow is the momentum sub-step.<\/li>\n<li>The transparent red arrow is the gradient sub-step if it starts before the momentum sub-step.<\/li>\n<li>The black arrow is the gradient sub-step if it starts after the momentum sub-step.<\/li>\n<li>CM would end up in the target of the dark red arrow.<\/li>\n<li>NAG would end up in the target of the black arrow.<\/li>\n<\/ul>\n\n<p>Note that this argument for why NAG is better is independent of whether the algorithm is close to a minimum.<br\/>\nIn general, both NAG and CM often have the problem of accumulating more momentum than is good for them, so whenever they should change direction, they have an embarrassing \"response time\". The advantage of NAG over CM that we explained doesn't prevent the problem, but only makes the \"response time\" of NAG less embarrassing (but embarrassing still).<br\/><\/p>\n\n<p>This \"response time\" problem is beautifully demonstrated in the gif by <a href=\"https:\/\/twitter.com\/AlecRad\" rel=\"noreferrer\">Alec Radford<\/a> (which appeared in <a href=\"https:\/\/stackoverflow.com\/a\/44225502\/5728089\">Salvador Dali's answer<\/a>):<br\/>\n<a href=\"https:\/\/i.stack.imgur.com\/ihlS2.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ihlS2.gif\" alt=\"An example of the embarrassing response time of both momentum methods\"><\/a><\/p>\n\n<h3>ADAGRAD<\/h3>\n\n<p>(Mostly based on section 2.2.2 in <a href=\"https:\/\/arxiv.org\/abs\/1212.5701\" rel=\"noreferrer\">ADADELTA: An Adaptive Learning Rate Method<\/a> (the original ADADELTA paper), as I find it much more accessible than <a href=\"http:\/\/www.jmlr.org\/papers\/v12\/duchi11a.html\" rel=\"noreferrer\">Adaptive Subgradient Methods for Online Learning and Stochastic Optimization<\/a> (the original ADAGRAD paper).)<br\/><\/p>\n\n<p>In <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent#Iterative_method\" rel=\"noreferrer\">SGD<\/a>, the step is given by <code>- learning_rate * gradient<\/code>, while <code>learning_rate<\/code> is a hyperparameter.<br\/>\nADAGRAD also has a <code>learning_rate<\/code> hyperparameter, but the actual learning rate for each component of the gradient is calculated individually.<br\/>\nThe <code>i<\/code>-th component of the <code>t<\/code>-th step is given by:<br\/><\/p>\n\n<pre><code>              learning_rate \n- --------------------------------------- * gradient_i_t\n  norm((gradient_i_1, ..., gradient_i_t))\n<\/code><\/pre>\n\n<p>while:<\/p>\n\n<ul>\n<li><code>gradient_i_k<\/code> is the <code>i<\/code>-th component of the gradient in the <code>k<\/code>-th step<\/li>\n<li><code>(gradient_i_1, ..., gradient_i_t)<\/code> is a vector with <code>t<\/code> components. This isn't intuitive (at least to me) that constructing such a vector makes sense, but that's what the algorithm does (conceptually).<\/li>\n<li><code>norm(vector)<\/code> is the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Norm_(mathematics)#Euclidean_norm\" rel=\"noreferrer\">Eucldiean norm<\/a> (aka <code>l2<\/code> norm) of <code>vector<\/code>, which is our intuitive notion of length of <code>vector<\/code>.<\/li>\n<li>Confusingly, in ADAGRAD (as well as in some other methods) the expression that is multiplied by <code>gradient_i_t<\/code> (in this case, <code>learning_rate \/ norm(...)<\/code>) is often called \"the learning rate\" (in fact, I called it \"the actual learning rate\" in the previous paragraph). I guess this is because in <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent#Iterative_method\" rel=\"noreferrer\">SGD<\/a> the <code>learning_rate<\/code> hyperparameter and this expression are one and the same.<\/li>\n<li>In a real implementation, some constant would be added to the denominator, to prevent a division by zero.<\/li>\n<\/ul>\n\n<p>E.g. if:<\/p>\n\n<ul>\n<li>The <code>i<\/code>-th component of the gradient in the first step is <code>1.15<\/code><\/li>\n<li>The <code>i<\/code>-th component of the gradient in the second step is <code>1.35<\/code><\/li>\n<li>The <code>i<\/code>-th component of the gradient in the third step is <code>0.9<\/code><\/li>\n<\/ul>\n\n<p>Then the norm of <code>(1.15, 1.35, 0.9)<\/code> is the length of the yellow line, which is:<br\/>\n<code>sqrt(1.15^2 + 1.35^2 + 0.9^2) = 1.989<\/code>.<br\/>\nAnd so the <code>i<\/code>-th component of the third step is: <code>- learning_rate \/ 1.989 * 0.9<\/code><br\/><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pjyp7.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pjyp7.png\" alt=\"l2 norm example\"><\/a><\/p>\n\n<p>Note two things about the <code>i<\/code>-th component of the step:<\/p>\n\n<ol>\n<li>It is proportional to <code>learning_rate<\/code>.<\/li>\n<li>In the calculations of it, the norm is increasing, and so the learning rate is decreasing.<\/li>\n<\/ol>\n\n<p>This means that ADAGRAD is sensitive to the choice of the hyperparameter <code>learning_rate<\/code>.<br\/>\nIn addition, it might be that after some time the steps become so small, that ADAGRAD virtually gets stuck.<\/p>\n\n<h3>ADADELTA and RMSProp<\/h3>\n\n<p>From the <a href=\"https:\/\/arxiv.org\/abs\/1212.5701\" rel=\"noreferrer\">ADADELTA paper<\/a>:<\/p>\n\n<blockquote>\n  <p>The idea presented in this paper was derived from ADAGRAD in order to improve upon the two main drawbacks\n  of the method: 1) the continual decay of learning rates\n  throughout training, and 2) the need for a manually selected\n  global learning rate.<\/p>\n<\/blockquote>\n\n<p>The paper then explains an improvement that is meant to tackle the first drawback:<\/p>\n\n<blockquote>\n  <p>Instead of accumulating the sum of squared gradients over all\n  time, we restricted the window of past gradients that are accumulated\n  to be some fixed size <code>w<\/code> [...]. This ensures that learning continues\n  to make progress even after many iterations of updates have\n  been done.<br\/>\n  Since storing <code>w<\/code> previous squared gradients is inefficient,\n  our methods implements this accumulation as an exponentially\n  decaying average of the squared gradients.<\/p>\n<\/blockquote>\n\n<p>By \"exponentially decaying average of the squared gradients\" the paper means that for each <code>i<\/code> we compute a weighted average of all of the squared <code>i<\/code>-th components of all of the gradients that were calculated.<br\/>\nThe weight of each squared <code>i<\/code>-th component is bigger than the weight of the squared <code>i<\/code>-th component in the previous step.<\/p>\n\n<p>This is an approximation of a window of size <code>w<\/code> because the weights in earlier steps are very small.<\/p>\n\n<p>(When I think about an exponentially decaying average, I like to visualize a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Comet\" rel=\"noreferrer\">comet's<\/a> trail, which becomes dimmer and dimmer as it gets further from the comet:<br\/><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UIzkI.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UIzkI.png\" alt=\"a comet&#39;s trail as an intuition for a moving average\"><\/a>)<\/p>\n\n<p>If you make only this change to ADAGRAD, then you will get RMSProp, which is a method that was proposed by Geoff Hinton in <a href=\"http:\/\/www.cs.toronto.edu\/~tijmen\/csc321\/slides\/lecture_slides_lec6.pdf\" rel=\"noreferrer\">Lecture 6e of his Coursera Class<\/a>.<\/p>\n\n<p>So in RMSProp, the <code>i<\/code>-th component of the <code>t<\/code>-th step is given by:<br\/><\/p>\n\n<pre><code>                   learning_rate\n- ------------------------------------------------ * gradient_i_t\n  sqrt(exp_decay_avg_of_squared_grads_i + epsilon)\n<\/code><\/pre>\n\n<p>while:<\/p>\n\n<ul>\n<li><code>epsilon<\/code> is a hyperparameter that prevents a division by zero.<\/li>\n<li><code>exp_decay_avg_of_squared_grads_i<\/code> is an exponentially decaying average of the squared <code>i<\/code>-th components of all of the gradients calculated (including <code>gradient_i_t<\/code>).<\/li>\n<\/ul>\n\n<p>But as aforementioned, ADADELTA also aims to get rid of the <code>learning_rate<\/code> hyperparameter, so there must be more stuff going on in it.<br\/><\/p>\n\n<p>In ADADELTA, the <code>i<\/code>-th component of the <code>t<\/code>-th step is given by:<br\/><\/p>\n\n<pre><code>  sqrt(exp_decay_avg_of_squared_steps_i + epsilon)\n- ------------------------------------------------ * gradient_i_t\n  sqrt(exp_decay_avg_of_squared_grads_i + epsilon) \n<\/code><\/pre>\n\n<p>while <code>exp_decay_avg_of_squared_steps_i<\/code> is an exponentially decaying average of the squared <code>i<\/code>-th components of all of the steps calculated (until the <code>t-1<\/code>-th step).<br\/>\n<code>sqrt(exp_decay_avg_of_squared_steps_i + epsilon)<\/code> is somewhat similar to momentum, and according to <a href=\"https:\/\/arxiv.org\/abs\/1212.5701\" rel=\"noreferrer\">the paper<\/a>, it \"acts as an acceleration term\". (The paper also gives another reason for why it was added, but my answer is already too long, so if you are curious, check out section 3.2.)<\/p>\n\n<h3>Adam<\/h3>\n\n<p>(Mostly based on <a href=\"https:\/\/arxiv.org\/abs\/1412.6980\" rel=\"noreferrer\">Adam: A Method for Stochastic Optimization<\/a>, the original Adam paper.)<\/p>\n\n<p>Adam is short for Adaptive Moment Estimation (see <a href=\"https:\/\/stats.stackexchange.com\/a\/367987\/215801\">this answer<\/a> for an explanation about the name).<br\/>\nThe <code>i<\/code>-th component of the <code>t<\/code>-th step is given by:<br\/><\/p>\n\n<pre><code>                   learning_rate\n- ------------------------------------------------ * exp_decay_avg_of_grads_i\n  sqrt(exp_decay_avg_of_squared_grads_i) + epsilon\n<\/code><\/pre>\n\n<p>while:<\/p>\n\n<ul>\n<li><code>exp_decay_avg_of_grads_i<\/code> is an exponentially decaying average of the <code>i<\/code>-th components of all of the gradients calculated (including <code>gradient_i_t<\/code>).<\/li>\n<li>Actually, both <code>exp_decay_avg_of_grads_i<\/code> and <code>exp_decay_avg_of_squared_grads_i<\/code> are also corrected to account for a bias toward <code>0<\/code> (for more about that, see section 3 in <a href=\"https:\/\/arxiv.org\/abs\/1412.6980\" rel=\"noreferrer\">the paper<\/a>, and also <a href=\"https:\/\/stats.stackexchange.com\/a\/234686\/215801\">an answer in stats.stackexchange<\/a>).<\/li>\n<\/ul>\n\n<p>Note that Adam uses an exponentially decaying average of the <code>i<\/code>-th components of the gradients where most <a href=\"https:\/\/en.wikipedia.org\/wiki\/Stochastic_gradient_descent#Iterative_method\" rel=\"noreferrer\">SGD<\/a> methods use the <code>i<\/code>-th component of the current gradient. This causes Adam to behave like \"a heavy ball with friction\", as explained in the paper <a href=\"https:\/\/arxiv.org\/abs\/1706.08500\" rel=\"noreferrer\">GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium<\/a>.<br\/>\nSee <a href=\"https:\/\/datascience.stackexchange.com\/a\/39224\/56981\">this answer<\/a> for more about how Adam's momentum-like behavior is different from the usual momentum-like behavior.<\/p>\n"},{"AnswerId":"55657011","CreationDate":"2019-04-12T17:38:07.683","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>Let's boil it down to a couple of simple question:<\/p>\n\n<p><strong>Which optimizer would give me the best result\/accuracy?<\/strong><\/p>\n\n<p>There is no silver bullet. Some optimizers for your task would probably work better than the others. There is no way to tell beforehand, you have to try a few to find the best one. Good news is that the results of different optimizers would probably be close to each other. You have to find the best hyperparameters for any single optimizer you choose, though.<\/p>\n\n<p><strong>Which optimizer should I use right now?<\/strong><\/p>\n\n<p>Maybe, take AdamOptimizer and run it for learning_rate 0.001 and 0.0001. If you want better results, try running for other learning rates. Or try other optimizers and tune their hyperparameters.<\/p>\n\n<h3>Long story<\/h3>\n\n<p>There are a few aspects to consider when choosing your optimizer:<\/p>\n\n<ul>\n<li>Easy of use (i.e. how fast you can find parameters that work for you);<\/li>\n<li>Convergence speed (basic as SGD or faster as any other);<\/li>\n<li>Memory footprint (typically between 0 and x2 sizes of your model);<\/li>\n<li>Relation to other parts of the training process.<\/li>\n<\/ul>\n\n<p><strong>Plain SGD<\/strong> is the bare minimum that can be done:  it simply multiplies the gradients by the learning rate and adds the result to the weights. SGD has a number of beautiful qualities: it has only 1 hyperparameter; it does not need any additional memory; it has minimal effect on the other parts of training. It also has 2 drawbacks: it might be too sensitive to learning rate choice and training can take longer than with other methods.<\/p>\n\n<p>From these drawbacks of plain SGD we can see what are the more complicated update rules (optimizers) are for: we sacrifice a part of our memory to achieve faster training and, possibly, simplify the choice of hyperparameters.<\/p>\n\n<p><strong>Memory overhead<\/strong> is typically non-significant and can be ignored. Unless the model is extremely large, or you are training on GTX760, or fighting for ImageNet leadership. Simpler methods like momentum or Nesterov accelerated gradient need 1.0 or less of model size (size of the model hyperparameters). Second order methods (Adam,  might need twice as much memory and computation.<\/p>\n\n<p><strong>Convergence speed<\/strong>-wise pretty much anything is better than SGD and anything else is hard to compare. One note might be that AdamOptimizer is good at starting training almost immediately, without a warm-up.<\/p>\n\n<p>I consider <strong>easy-of-use<\/strong> to be the most important in the choice of an optimizer. Different optimizers have a different number of hyperparameters and have a different sensibility to them. I consider Adam the most simple of all readily-available ones. You typically need to check 2-4 learning_rates between <code>0.001<\/code> and <code>0.0001<\/code> to figure out if the model converges nicely. For comparison for SGD (and momentum) I typically try <code>[0.1, 0.01, ... 10e-5]<\/code>. Adam has 2 more hyperparameters that rarely have to be changed. <\/p>\n\n<p><strong>Relation between optimizer and other parts of training<\/strong>. Hyperparameter tuning typically involves selecting <code>{learning_rate, weight_decay, batch_size, droupout_rate}<\/code> simultaneously. All of them are interrelated and each can be viewed as a form of model regularization. One, for example, has to pay close attention if, exactly, weight_decay or L2-norm is used and possibly choose <code>AdamWOptimizer<\/code> instead of <code>AdamOptimizer<\/code>.<\/p>\n"},{"AnswerId":"44225502","CreationDate":"2017-05-28T09:01:02.457","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Here is a brief explanation based on my understanding:<\/p>\n\n<ul>\n<li><strong>momentum<\/strong> <a href=\"https:\/\/www.youtube.com\/watch?v=8yg2mRJx-z4\" rel=\"noreferrer\">helps<\/a> SGD to navigate along the relevant directions and softens the oscillations in the irrelevant. It simply adds a fraction of the direction of the previous step to a current step. This achieves amplification of speed in the correct direction and softens oscillation in wrong directions. This fraction is usually in the (0, 1) range. It also makes sense to use adaptive momentum. In the beginning of learning a big momentum will only hinder your progress, so it makes sense to use something like 0.01 and once all the high gradients disappeared you can use a bigger momentum. There is one problem with momentum: when we are very close to the goal, our momentum in most of the cases is very high and it does not know that it should slow down. This can cause it to miss or oscillate around the minima<\/li>\n<li><strong>nesterov accelerated gradient<\/strong> overcomes this problem by starting to slow down early. In momentum we first compute gradient and then make a jump in that direction amplified by whatever momentum we had previously. NAG does the same thing but in another order: at first we make a big jump based on our stored information, and then we calculate the gradient and make a small correction. This seemingly irrelevant change gives significant practical speedups.<\/li>\n<li><strong>AdaGrad<\/strong> or adaptive gradient allows the learning rate to adapt based on parameters. It performs larger updates for infrequent parameters and smaller updates for frequent one. Because of this it is well suited for sparse data (NLP or image recognition). Another advantage is that it basically eliminates the need to tune the learning rate. Each parameter has its own learning rate and due to the peculiarities of the algorithm the learning rate is monotonically decreasing. This causes the biggest problem: at some point of time the learning rate is so small that the system stops learning.<\/li>\n<li><strong>AdaDelta<\/strong> <a href=\"http:\/\/int8.io\/comparison-of-optimization-techniques-stochastic-gradient-descent-momentum-adagrad-and-adadelta\/#AdaGrad\" rel=\"noreferrer\">resolves<\/a> the problem of monotonically decreasing learning rate in AdaGrad. In AdaGrad the learning rate was calculated approximately as one divided by the sum of square roots. At each stage you add another square root to the sum, which causes denominator to constantly increase. In AdaDelta instead of summing all past square roots it uses sliding window which allows the sum to decrease. <strong>RMSprop<\/strong> is very similar to AdaDelta<\/li>\n<li><p><strong>Adam<\/strong> or adaptive momentum is an algorithm similar to AdaDelta. But in addition to storing learning rates for each of the parameters it also stores momentum changes for each of them separately.<\/p>\n\n<p>A <a href=\"http:\/\/ruder.io\/optimizing-gradient-descent\/index.html#visualizationofalgorithms\" rel=\"noreferrer\">few visualizations<\/a>:\n<img src=\"https:\/\/i.stack.imgur.com\/qAx2i.gif\" alt=\"enter image description here\">\n<img src=\"https:\/\/i.stack.imgur.com\/1obtV.gif\" alt=\"enter image description here\"><\/p><\/li>\n<\/ul>\n\n<p>I would say that SGD, Momentum and Nesterov are inferior than the last 3.<\/p>\n"}]}
{"QuestionId":36163613,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2016-03-22T19:37:47.260","AcceptedAnswerId":36184707.0,"OwnerUserId":636866.0,"Title":"how does the validation split work when calling fit?","Body":"<p>When I specify the validation_split parameter when calling the fit method of the model class, is the same validation data used at every epoch? Or does the validation data change for every epoch?<\/p>\n","answers":[{"AnswerId":"36184707","CreationDate":"2016-03-23T17:20:03.470","ParentId":null,"OwnerUserId":"4293020","Title":null,"Body":"<p>It uses the same validation data for every epoch.<\/p>\n\n<p>If it didn't then it wouldn't be validation data as the model would learn have seen elements in the validation set during previous epochs and and thus would have the possibility of overfitting to them.<\/p>\n"}]}
{"QuestionId":36164743,"AnswerCount":2,"Tags":"<python><c++><tensorflow>","CreationDate":"2016-03-22T20:39:09.223","AcceptedAnswerId":null,"OwnerUserId":640898.0,"Title":"Tensorflow C++ evaluation performance worse than Python one","Body":"<p>I built an LSTM classifier in Tensorflow (using Python) and now I am doing a series of benchmarks to measure the execution performance. The benchmark code loads the model that was saved during training and executes it for a number of inputs. I have both a Python benchmark and a C++ one. <\/p>\n\n<p>I get  per input when using Python (on CPU) and  when using C++. I was expecting C++ version to have at least similar performance.<\/p>\n\n<p>What can cause this difference ? I am not familiar with the internals of Tensorflow, but one of my guesses is that Python version uses BLAS (my numpy is linked with blas) and C++ not. If yes, does that mean that some graph operations are executed in Python ?<\/p>\n","answers":[{"AnswerId":"41343470","CreationDate":"2016-12-27T10:34:33.573","ParentId":null,"OwnerUserId":"1551094","Title":null,"Body":"<p>I'm not sure if it is too late to post an answer. I faced the same issue and got answer from tensorflow github page:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3471\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3471<\/a><\/p>\n\n<p>The solution is simple, just use the <code>-c opt<\/code> flag. <\/p>\n"},{"AnswerId":"36165555","CreationDate":"2016-03-22T21:26:55.063","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Can you try a simpler benchmark to narrow down the problem. The following takes 2-5 seconds on CPU<\/p>\n\n<pre><code>import tensorflow as tf\na=tf.Variable(tf.random_normal([8192, 8192], stddev=0.35))\nb=tf.Variable(tf.random_normal([8192, 8192], stddev=0.35))\nc=tf.matmul(a, b)\n\nsess.run([tf.initialize_all_variables()])\n\nimport time\nstart_time = time.time()\nsess.run(c)\nprint time.time()-start_time\n<\/code><\/pre>\n"}]}
{"QuestionId":36165164,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-22T21:03:09.953","AcceptedAnswerId":36165503.0,"OwnerUserId":234167.0,"Title":"How do I draw a resized image in TensorFlow?","Body":"<p>It seems like images in TensorFlow get transformed to a different kind of image coordinate system after any transformation (e.g. resize) is applied. Drawing the image results in this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cFnWi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cFnWi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>If I remove the resize operation it draws the regular image. How do I get matplotlib to draw the resized image correctly or tell Tensorflow to return it to RGB?<\/p>\n","answers":[{"AnswerId":"36165503","CreationDate":"2016-03-22T21:22:43.410","ParentId":null,"OwnerUserId":"234167","Title":null,"Body":"<p>Seems like there is no image transformation besides unsigned integer to float. Converting back to unsigned integer fixed the problem.<\/p>\n\n<pre><code>plt.imshow(image_val[0].astype(np.uint8), interpolation='nearest')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2nAfw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2nAfw.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":36166473,"AnswerCount":1,"Tags":"<python><numpy><theano><lasagne>","CreationDate":"2016-03-22T22:30:20.517","AcceptedAnswerId":36175381.0,"OwnerUserId":4566639.0,"Title":"Theano dimensionality error - target dimensions","Body":"<p>I am using lasagne's Conv3DDNNLayer, and have input dimensions of (N x 1 x 9 x 9 x 9), where each 9x9x9 cube is my sample to be classified. <\/p>\n\n<p>Therefore I have a target dimension of (N x 1), with each entry corresponding to a cube. This is raising the error:<\/p>\n\n<pre><\/pre>\n\n<p>Which dimensions should I have my targets in in this case?<\/p>\n\n<pre><\/pre>\n\n<p>edit - added code<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36175381","CreationDate":"2016-03-23T10:18:47.737","ParentId":null,"OwnerUserId":"4566639","Title":null,"Body":"<p>In case any one is interested, it was because the data was (N, 1) not (N, ). <\/p>\n\n<p>seemed to solve the problem! - on to the next..<\/p>\n"}]}
{"QuestionId":36166486,"AnswerCount":0,"Tags":"<image><neural-network><image-segmentation><caffe><crf>","CreationDate":"2016-03-22T22:31:24.870","AcceptedAnswerId":null,"OwnerUserId":1565344.0,"Title":"Image segmentation with a challenging background","Body":"<p>[cross-posted from stats and datascience]<\/p>\n\n<p>I'm working on an animal classification problem, with the data extracted from a video feed. The recording was made in a pen, so the problem is quite challenging with a dark background and many shadows: <a href=\"https:\/\/i.stack.imgur.com\/2B5BE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2B5BE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Initially I tried scikit-image, but then someone helped me with an advanced tool called crf-rnn (<a href=\"http:\/\/crfasrnn.torr.vision\/\" rel=\"nofollow noreferrer\">http:\/\/crfasrnn.torr.vision\/<\/a>) that does a great job segmenting and labelling objects in an image. I did the following: <\/p>\n\n<pre><\/pre>\n\n<p>The resulting image segmentation is rather poor (although two cows are recognized correctly):\n<a href=\"https:\/\/i.stack.imgur.com\/eBEr4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eBEr4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I use a trained crf-rnn (MODEL_FILE, PRETRAINED), which works well for other problems, but this one is harder. I would appreciate any suggestions on how to pre-process this sort of image to extract the shape of most cows. <\/p>\n","answers":[]}
{"QuestionId":36167895,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-03-23T00:48:25.997","AcceptedAnswerId":36177213.0,"OwnerUserId":3130081.0,"Title":"About binary classification in Caffe","Body":"<p>When solving a binary classification problem, I think there are two possible ways in caffe.<br>\nThe first one is using <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SigmoidCrossEntropyLossLayer.html\" rel=\"nofollow\"><\/a> with one output unit.<br>\nThe other one is using <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1SoftmaxWithLossLayer.html\" rel=\"nofollow\"><\/a> with two output units.\nMy question is what\u2019s the difference between these two approaches?<br>\nWhich one should I use?<br>\nThank you very much!<\/p>\n","answers":[{"AnswerId":"36177213","CreationDate":"2016-03-23T11:40:33.547","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you play a bit with the math, you can \"duplicate\" the predicted class probability of the <code>\"Sigmoid\"<\/code> layer to <code>0.5*x_i<\/code> for class 1 and <code>-0.5*x_i<\/code> for class 0, then the <code>\"SoftmaxWithLoss\"<\/code> layer amounts to <code>\"SigmoindWithCrossEntropy\"<\/code> on the single output predictions <code>x_i<\/code>.  <\/p>\n\n<p>So I believe it can be said that these two methods can be regarded as equivalent for predicting binary outputs.<\/p>\n"}]}
{"QuestionId":36172716,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><gradient-descent>","CreationDate":"2016-03-23T08:05:45.417","AcceptedAnswerId":null,"OwnerUserId":4239601.0,"Title":"Inception-v3 using RMSProp epsilon=1","Body":"<p>I just read the <a href=\"http:\/\/arxiv.org\/abs\/1512.00567\" rel=\"nofollow\">Inception-v3 paper<\/a>, and its training code released by the authors. And I found that when do RMSProp\noptimization, the authors used <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/inception_train.py#L81\" rel=\"nofollow\">epsilon=1<\/a>. However, to my knowledge, people usually used 1e-10 or some small values and Tensorflow set the default value to 1e-10.<\/p>\n\n<p>So my question is how will the epsilon influence the optimization? Why we need a epsilon=1 here?\nThank you!<\/p>\n","answers":[{"AnswerId":"43144725","CreationDate":"2017-03-31T15:44:25.650","ParentId":null,"OwnerUserId":"380038","Title":null,"Body":"<p>Through this <a href=\"https:\/\/github.com\/ibab\/tensorflow-wavenet\/issues\/149\" rel=\"nofollow noreferrer\">github issue<\/a> I found a post explaining the value: <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msg\/discuss\/1JDPbuHN39w\/AVXdYeaDCAAJ\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msg\/discuss\/1JDPbuHN39w\/AVXdYeaDCAAJ<\/a><\/p>\n\n<p>In this post one of the inception v3 authors explains that they needed such a high epsilon to train it with async SGD and that lower values might actually work (better) for sync SGD.<\/p>\n"}]}
{"QuestionId":36173205,"AnswerCount":2,"Tags":"<tensorflow><sequence>","CreationDate":"2016-03-23T08:34:07.003","AcceptedAnswerId":null,"OwnerUserId":6103081.0,"Title":"Weights of Seq2Seq Models","Body":"<p>I went through the code and I'm afraid I don't grasp an important point.<\/p>\n\n<p>I can't seem to find the weights matrix of the model for the encoder and decoder, neither where they are updated. I found the target_weights but it seems to be reinitialized at every  call so I don't really understand what they stand for either.<\/p>\n\n<p>My actual goal is to concatenate two hidden states of two source encoders for one decoder by applying a linear transformation with a weight matrix that I'll have to train along with the model (I'm building a manytoone model), but I have no idea where to start because of my problem mentionned above.<\/p>\n","answers":[{"AnswerId":"36183103","CreationDate":"2016-03-23T16:04:50.087","ParentId":null,"OwnerUserId":"5765409","Title":null,"Body":"<p>This might help you start. There are a couple of models implemented in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py\" rel=\"nofollow\">tensorflow.python.ops.seq2seq.py<\/a> (with\/without buckets, attention, etc.) but take a look at the definition for <code>embedding_attention_seq2seq<\/code> (which is the one called in their example model <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/seq2seq.py\" rel=\"nofollow\">seq2seq_model.py<\/a> that you seem to be referencing):<\/p>\n\n<pre><code>def embedding_attention_seq2seq(encoder_inputs, decoder_inputs, cell,\n                                num_encoder_symbols, num_decoder_symbols,\n                                num_heads=1, output_projection=None,\n                                feed_previous=False, dtype=dtypes.float32,\n                                scope=None, initial_state_attention=False):\n\n  with variable_scope.variable_scope(scope or \"embedding_attention_seq2seq\"):\n    # Encoder.\n    encoder_cell = rnn_cell.EmbeddingWrapper(cell, num_encoder_symbols)\n    encoder_outputs, encoder_state = rnn.rnn(\n        encoder_cell, encoder_inputs, dtype=dtype)\n\n    # First calculate a concatenation of encoder outputs to put attention on.\n    top_states = [array_ops.reshape(e, [-1, 1, cell.output_size])\n                  for e in encoder_outputs]\n    attention_states = array_ops.concat(1, top_states)\n    ....\n<\/code><\/pre>\n\n<p>You can see where it picks out the top layer of encoder outputs as <code>top_states<\/code> before handing them off to the decoder.<\/p>\n\n<p>So you could implement a similar function with two encoders and concatenate those states before handing off to the decoder.<\/p>\n"},{"AnswerId":"44801392","CreationDate":"2017-06-28T11:40:30.280","ParentId":null,"OwnerUserId":"5493885","Title":null,"Body":"<p>The value created in the get_batch function is only used for the first iteration. Even though the weights are passed every time into the function, their value gets updated as a global variable in the Seq2Seq model class in the init function. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    with tf.name_scope('Optimizer'):\n        # Gradients and SGD update operation for training the model.\n        params = tf.trainable_variables()\n        if not forward_only:\n            self.gradient_norms = []\n            self.updates = []\n            opt = tf.train.GradientDescentOptimizer(self.learning_rate)\n            for b in range(len(buckets)):\n                gradients = tf.gradients(self.losses[b], params)\n                clipped_gradients, norm = tf.clip_by_global_norm(gradients,\n                                                                 max_gradient_norm)\n                self.gradient_norms.append(norm)\n                self.updates.append(opt.apply_gradients(\n                    zip(clipped_gradients, params), global_step=self.global_step))\n\n    self.saver = tf.train.Saver(tf.global_variables())\n<\/code><\/pre>\n\n<p>The weights are fed seperately as a place-holder because they are normalized in the get_batch function to create zero weights for the PAD inputs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    # Batch decoder inputs are re-indexed decoder_inputs, we create weights.\n    for length_idx in range(decoder_size):\n        batch_decoder_inputs.append(\n            np.array([decoder_inputs[batch_idx][length_idx]\n                      for batch_idx in range(self.batch_size)], dtype=np.int32))\n\n        # Create target_weights to be 0 for targets that are padding.\n        batch_weight = np.ones(self.batch_size, dtype=np.float32)\n        for batch_idx in range(self.batch_size):\n            # We set weight to 0 if the corresponding target is a PAD symbol.\n            # The corresponding target is decoder_input shifted by 1 forward.\n            if length_idx &lt; decoder_size - 1:\n                target = decoder_inputs[batch_idx][length_idx + 1]\n            if length_idx == decoder_size - 1 or target == data_utils.PAD_ID:\n                batch_weight[batch_idx] = 0.0\n        batch_weights.append(batch_weight)\n<\/code><\/pre>\n"}]}
{"QuestionId":36173686,"AnswerCount":1,"Tags":"<machine-learning><segmentation-fault><neural-network><deep-learning><caffe>","CreationDate":"2016-03-23T09:00:09.003","AcceptedAnswerId":36174031.0,"OwnerUserId":6103162.0,"Title":"When I run my own images on caffe , it stop at Iteration 0, Testing net (#0)","Body":"<p>I ran caffe and got this output:<br>\n<img src=\"https:\/\/i.stack.imgur.com\/Xh1Pr.png\" alt=\"enter image description here\"><\/p>\n\n<p>who can tell me what is the problem?<br>\nI will really appreciate!!<\/p>\n","answers":[{"AnswerId":"36174031","CreationDate":"2016-03-23T09:18:09.923","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like one (or more) of your label values are invalid, see <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3043\" rel=\"nofollow noreferrer\">this PR<\/a> for information:<\/p>\n\n<blockquote>\n  <p>If you have an invalid ground truth label, <code>\"SoftmaxWithLoss\"<\/code> will silently access invalid memory [...] The old check only worked in DEBUG mode and also only worked for CPU.<\/p>\n<\/blockquote>\n\n<p>Make sure your prediction vector length matches the number of labels you try to predict.<\/p>\n\n<p>From your comments, it seems like you have labels in the range <a href=\"https:\/\/stackoverflow.com\/questions\/36173686\/when-i-run-my-own-images-on-caffe-it-stop-at-iteration-0-testing-net-0\/36174031?noredirect=1#comment59985696_36174031\"><code>0..10575<\/code><\/a>, but on the other hand, your classification layer, <a href=\"https:\/\/stackoverflow.com\/questions\/36173686\/when-i-run-my-own-images-on-caffe-it-stop-at-iteration-0-testing-net-0\/36174031?noredirect=1#comment59987277_36174031\"><code>\"fc7\"<\/code><\/a> only predicts probabilities for 1000 classes. Thus, <code>\"SoftmaxWithLoss\"<\/code> layer tries to compute the loss for predicting label <code>l<\/code>>1000, and access memory outside the probability array, resulting with a segmentation fault.<\/p>\n"}]}
{"QuestionId":36176890,"AnswerCount":2,"Tags":"<neural-network><gradient><theano>","CreationDate":"2016-03-23T11:26:25.870","AcceptedAnswerId":null,"OwnerUserId":6103829.0,"Title":"Theano -- Mean of squared gradients","Body":"<p>In theano, given a batch cost  with shape (batch_size,), it is easy to compute the gradient of the mean cost, as in  with p being a parameter used in the computation of . This is done efficiently by backpropagating the gradient through the computational graph. What I would now like to do is to compute the mean of the squared gradients over the batch. This can be done using the following piece of code:  <\/p>\n\n<pre><\/pre>\n\n<p>Where for convenience  is assumed to be a single theano tensor and not a list of tensors.\nThe computation could be performed efficiently by simply backpropagating the gradient until the last step, and squaring the components of the last operation (which should be a sum over the batch index). I might be wrong on this one, but the computation should be as easy, and nearly as fast as a simple backpropagation. However, theano seems unable to optimize the computation, and it keeps using a loop, making computations extremely slow.<\/p>\n\n<p>Would anyone know of a solution to make the computation efficient, either by forcing optimizations, expressing the computation in a different way, or even going through the backpropagation process ?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"40486159","CreationDate":"2016-11-08T11:48:01.990","ParentId":null,"OwnerUserId":"7113979","Title":null,"Body":"<p>After digging deeper in the Theano docs I found a solution that would <strong>work in the compute graph<\/strong>. Key idea is that you clone the graph of your network inside the scan function, thereby explicitly slicing the input tensor. I tried the following code and empirically it shows O(batch_size) as expected:<\/p>\n\n<pre><code> # x: input data batch\n # assuming cost = network(x,p)\n\n from theano.gof.graph import clone_get_equiv\n\n def g_square(cost,p):\n\n    g = T.zeros_like(p)\n\n    def scan_fn( i, g, cost, p):\n        # clone the graph computing cost, but slice it's input \n        cloned = clone_get_equiv([],[cost], \n                                 copy_inputs_and_orphans=False,\n                                 memo={x: x[i:i+1]})\n        cost_slice = cloned[cost].reshape([])\n        return  g+T.grad(cost_slice,p)**2 \n\n    result,updates = theano.reduce( scan_fn,\n                                    outputs_info=g,\n                                    sequences=[T.arange(cost.size)],\n                                    non_sequences=[cost.flatten(),p])\n\n    return result\n<\/code><\/pre>\n"},{"AnswerId":"40418875","CreationDate":"2016-11-04T09:13:09.847","ParentId":null,"OwnerUserId":"7113979","Title":null,"Body":"<p>Your function <code>g_square<\/code> happens to have complexity O(batch_size**2) instead of O(batch_size) as expected. This lets it appear incredibly slow for larger batch sizes.<\/p>\n\n<p>The reason is because in every iteration the forward and backward pass is computed over the whole batch, even though just <code>cost[i]<\/code> for one data point is needed. \nI assume the input to the <code>cost<\/code> computation graph, <code>x<\/code>, is a tensor with the first dimension of size <code>batch_size<\/code>. Theano has no means to automatically slice this tensor along this dimension. Therefore computation is always done over the whole batch.<\/p>\n\n<p>Unfortunately I see no better solution than slicing your input and doing the loop outside Theano:<\/p>\n\n<pre><code># x: input data batch\nbatch_size = x.shape[0]\ng_square_fun = theano.function( [p], T.grad(cost[0],p)**2) \n\ng_square_value = 0\nfor i in batch_size:\n    g_square_value += g_square_fun( x[i:i+1])\n<\/code><\/pre>\n\n<p>Perhaps when future versions of Theano come with better build in capabilities to compute Jacobians there will be more elegant solutions. <\/p>\n"}]}
{"QuestionId":36177788,"AnswerCount":1,"Tags":"<python><batch-processing><keras>","CreationDate":"2016-03-23T12:07:43.157","AcceptedAnswerId":null,"OwnerUserId":3921875.0,"Title":"Creating a generator from list of sequences for RNN","Body":"<p>I need to create a generator for my data to pass into my RNN training function. I have a list of patient samples, where each sample is a time series of length <em>n<sub>i<\/sub><\/em> (which varies, annoyingly) in three dimensions, and I want to create batches of data where each sample in a batch only belongs to a single patient but each batch may contain multiple patient samples. Doing it this way should maximise the number of samples I can train using with no consequences as my RNN is not <a href=\"http:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow\">stateful<\/a>. At first I had the following function<\/p>\n\n<pre><\/pre>\n\n<p>However this trims each of the patient samples in order to fit the batch size. So I want something that creates all possible batches, including the undersized one at the end. However my unfamiliarity with generators has left me pretty confused. So far I've worked out it's going to have to use modulo aritmetic, but exactly how I'm not sure, so I've only got to this point:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT<\/strong>\nHere's a short example to show how I would solve the problem without using generators<\/p>\n\n<pre><\/pre>\n\n<p>Which produces:<\/p>\n\n<pre><\/pre>\n\n<p>Which as you can see has two batches of size three, both of which contain two different 'patients' in them, but the time series for each 'patient' do not overlap.<\/p>\n","answers":[{"AnswerId":"36370027","CreationDate":"2016-04-02T06:33:07.850","ParentId":null,"OwnerUserId":"5828883","Title":null,"Body":"<p>It's not exactly clear what you are looking for.  A small sample of input and desired output would help.  Nevertheless, I'll take a stab at what I think you are asking:<\/p>\n\n<pre><code>def dataIterator(data, batchSize=batchSize):\n    for patient_data in data:\n        for n in range(0, len(patient_data), batchSize):\n            yield patient_data[n:n+batchSize]\n<\/code><\/pre>\n"}]}
{"QuestionId":36177958,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-23T12:17:12.923","AcceptedAnswerId":36184935.0,"OwnerUserId":234167.0,"Title":"TensorFlow: Rerun network with a different input tensor?","Body":"<p>Suppose I have a typical CNN model in TensorFlow.<\/p>\n\n<pre><\/pre>\n\n<p>A typical forward pass could be done like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now suppose my  function now returns a pair of arguments,  and  (stereo camera). I want to run  up to  and  up to . So something like this<\/p>\n\n<pre><\/pre>\n\n<p>This however fails with <\/p>\n\n<blockquote>\n  <p>TypeError: The value of a feed cannot be a tf.Tensor object.\n  Acceptable feed values include Python scalars, strings, lists, or\n  numpy ndarrays.<\/p>\n<\/blockquote>\n\n<p>I want to avoid having to evaluate  to then feed it back to TensorFlow. Calling  twice with different arguments will also not work because functions like  create variables.<\/p>\n\n<p>Is it possible to rerun the network with a different input tensor?<\/p>\n","answers":[{"AnswerId":"36184935","CreationDate":"2016-03-23T17:31:10.933","ParentId":null,"OwnerUserId":"2594660","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variable_scope\/index.html\" rel=\"nofollow\">Tensorflow shared Variables<\/a> is what you are looking for. Replace all calls of <code>tf.Variable<\/code> with <code>tf.get_variable()<\/code> in inference. Then you can run:<\/p>\n\n<pre><code>images_left, images_right = input()\nwith tf.variable_scope(\"logits\") as scope:\n    logits_left = inference(images_left)\n    scope.reuse_variables()\n    logits_right = inference(images_right)\noutput = sess.run([logits_left, logits_right])\n<\/code><\/pre>\n\n<p>Variables are not created again in the second call of inference. Left and right images are processed using the same weights. Also check out my <a href=\"https:\/\/github.com\/MarvinTeichmann\/tensorflow_toolkit\" rel=\"nofollow\">Tensorflow CNN training toolkit<\/a> (Look at <a href=\"https:\/\/github.com\/TensorVision\/TensorVision\/blob\/master\/bin\/tv-train\" rel=\"nofollow\">training<\/a> code). I utilize this technique to run validation and training forwards in the same TensorFlow graph. <\/p>\n"}]}
{"QuestionId":36182380,"AnswerCount":3,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-23T15:31:58.220","AcceptedAnswerId":36188682.0,"OwnerUserId":349760.0,"Title":"How do display different runs in TensorBoard?","Body":"<p>TensorBoard seems to have a feature to display multiple different runs and toggle them. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5E3eQ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5E3eQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>How can I make multiple runs show up here and how can assign a name to them to differentiate them?<\/p>\n","answers":[{"AnswerId":"57889604","CreationDate":"2019-09-11T12:50:29.390","ParentId":null,"OwnerUserId":"5992145","Title":null,"Body":"<p>It seems that just declaring it like this is ok:<\/p>\n\n<pre><code>writter = SummaryWritter(logdir='\/runs\/you_tag')\n<\/code><\/pre>\n\n<p>Then tensorboard will create this <code>you_tag<\/code> folder below <code>runs\/<\/code>, in the meantime, the web application will refresh and find <code>you_tag<\/code>.<\/p>\n"},{"AnswerId":"36182548","CreationDate":"2016-03-23T15:39:38.860","ParentId":null,"OwnerUserId":"349760","Title":null,"Body":"<p>I found the answer to my own question on github (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1548\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1548<\/a>).<\/p>\n\n<p>You need to put your logs in a subfolder e.g. \/logs\/run1\/ and then run tensorboard on the root folder e.g. \/logs\/.<\/p>\n"},{"AnswerId":"36188682","CreationDate":"2016-03-23T21:06:46.510","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>In addition to TensorBoard scanning subdirectories (so you can pass a directory containing the directories with your runs), you can also pass multiple directories to TensorBoard explicitly and give custom names (example taken from the --help output): <\/p>\n\n<pre><code>tensorboard --logdir=name1:\/path\/to\/logs\/1,name2:\/path\/to\/logs\/2\n<\/code><\/pre>\n\n<p>More information can be found at the <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/README.md#runs-comparing-different-executions-of-your-model\" rel=\"noreferrer\">TensorBoard documentation<\/a>.<\/p>\n"}]}
{"QuestionId":36185639,"AnswerCount":1,"Tags":"<python><machine-learning><caffe><lasagne>","CreationDate":"2016-03-23T18:09:37.423","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"Load a pretrained caffe model to lasagne?","Body":"<p>I am trying to reproduce the <a href=\"http:\/\/arxiv.org\/abs\/1411.4389\" rel=\"nofollow\">Long-term Recurrent Convolutional Networks<\/a> paper.<\/p>\n\n<p>I have a pretrained caffe model that I'd like to use in .\nI have the <a href=\"http:\/\/www.eecs.berkeley.edu\/~lisa_anne\/single_frame_all_layers_hyb_flow_iter_50000.caffemodel\" rel=\"nofollow\"><\/a> for this file, and the <a href=\"https:\/\/github.com\/LisaAnne\/lisa-caffe-public\/blob\/lstm_video_deploy\/examples\/LRCN_activity_recognition\/train_test_singleFrame_flow.prototxt\" rel=\"nofollow\"><\/a>.\nI have used the <a href=\"https:\/\/github.com\/Lasagne\/Recipes\/blob\/master\/examples\/Using%20a%20Caffe%20Pretrained%20Network%20-%20CIFAR10.ipynb\" rel=\"nofollow\">lasagne example<\/a> to load the caffe weights to the caffe model.\nThis is the <a href=\"http:\/\/pastebin.com\/xdD3wZCe\" rel=\"nofollow\">code I used<\/a>, but the data is not loaded to the lasagne model.\nI check it by using the   command, which throws this error.<\/p>\n\n<pre><\/pre>\n\n<p>TRIAL\/Test Code:-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36926099","CreationDate":"2016-04-28T21:49:38.957","ParentId":null,"OwnerUserId":"6268849","Title":null,"Body":"<p>try:<\/p>\n\n<blockquote>\n  <p>x=lasagne.layers.get_all_param_values(net['prob'])<\/p>\n<\/blockquote>\n\n<p>or make your net in this way:<\/p>\n\n<pre><code>def build_model():\n    net = {}\n    # Input layer\n    net = InputLayer((None, 3, 227, 227))\n    # First Conv Layer\n    net = ConvLayer(net, num_filters=96,filter_size=7, pad=0,     flip_filters=False,stride=2,nonlinearity=rectify)\n    net = PoolLayer(net, pool_size=3,stride=2,mode='max')\n    ....\n    net= NonlinearityLayer(net, softmax)\n    return net\n<\/code><\/pre>\n"}]}
{"QuestionId":36185914,"AnswerCount":2,"Tags":"<python><machine-learning><caffe>","CreationDate":"2016-03-23T18:23:26.093","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"Long-term Recurrent Convolutional Networks paper reproduction error in LSTM?","Body":"<p>I am trying to reproduce the <a href=\"http:\/\/arxiv.org\/abs\/1411.4389\" rel=\"nofollow\">Long-term Recurrent Convolutional Networks paper.<\/a><\/p>\n\n<p>I have used their <a href=\"https:\/\/github.com\/LisaAnne\/lisa-caffe-public\/tree\/lstm_video_deploy\" rel=\"nofollow\">given code<\/a>. And followed their instructions and generated the single frame model. But when trying to train the LSTM hybrid network it fails. I have already made necessary changes as mentioned in the <a href=\"http:\/\/www.eecs.berkeley.edu\/~lisa_anne\/LRCN_video\" rel=\"nofollow\">instructions.<\/a><\/p>\n\n<p>The command i run is the \nthe output I get is <\/p>\n\n<pre><\/pre>\n\n<p>This is my changed <a href=\"http:\/\/pastebin.com\/ZuRb3tQR\" rel=\"nofollow\">sequence_input_layer.py<\/a> and <a href=\"http:\/\/pastebin.com\/Ha4S1ZBU\" rel=\"nofollow\">prototext<\/a> files.\nMy input train and test txts to the network is of this <a href=\"http:\/\/pastebin.com\/4gNPFssb\" rel=\"nofollow\">format<\/a>.<\/p>\n\n<p>I think the main problem is the <\/p>\n\n<p>I was not able to solve this is confused me quite a bit.\nBut I might be wrong. <\/p>\n","answers":[{"AnswerId":"50686432","CreationDate":"2018-06-04T18:09:26.830","ParentId":null,"OwnerUserId":"6026051","Title":null,"Body":"<p>I know I am a bit late, hope my answer helps someone in future. I too faced the same error. The problem was with scikit-image version. I was using some version newer than scikit-image 0.9.3. Following line<\/p>\n\n<pre><code>processed_image = transformer.preprocess('data_in',data_in)\n<\/code><\/pre>\n\n<p>uses scikit-image library functions being imported by <code>lisa-caffe-public-lstm_video_deploy\/python\/caffe\/io.py<\/code>. With newer version, <code>transformer.preprocess()<\/code> was crashing, and therefore <code>processImageCrop()<\/code> function didnot return and thus <code>KeyError: 'data'<\/code>.<\/p>\n\n<p>In short, you should use scikit-image-0.9.3 to get rid of your error :)\nHope that helps :) <\/p>\n"},{"AnswerId":"36616728","CreationDate":"2016-04-14T07:43:59.877","ParentId":null,"OwnerUserId":"2932001","Title":null,"Body":"<p>I encounter the same question. I adapt several places in <code>sequence_input_layer.py<\/code> and it works well for me at now:<br>\n1. I change line 95 and remove <code>%i<\/code> at the end since I don't know what it is used for. i.e., change<br>\n<code>\n frames.append(self.video_dict[key]['frames'] %i)\n<\/code>\nto<br>\n<code>\n frames.append(self.video_dict[key]['frames'])\n<\/code>\n2. I change line around 157 to make sure the <code>frames<\/code> variable is right for my data since the data directory of author is different of mine.<\/p>\n\n<p>Hope this will be helpful.<\/p>\n"}]}
{"QuestionId":36186069,"AnswerCount":1,"Tags":"<ubuntu><tensorflow><bazel>","CreationDate":"2016-03-23T18:30:45.387","AcceptedAnswerId":36186634.0,"OwnerUserId":3559686.0,"Title":"TensorFlow: Error while parsing .d file","Body":"<p>I use Ubuntu 15.10. Currently I am building TensorFlow from source, following the instructions <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">here<\/a>.<\/p>\n\n<p>My installation is successful, save for the last step: \"Build your target with GPU support\".<\/p>\n\n<p>When I run this command from the root of my source tree:<\/p>\n\n<p><\/p>\n\n<p>I see the following output:<\/p>\n\n<pre><\/pre>\n\n<p>There are several things I do not understand:<\/p>\n\n<ol>\n<li>Why is the  file missing?<\/li>\n<li>What does this mean? <\/li>\n<\/ol>\n\n<blockquote>\n  <p>nvcc warning : option '--relaxed-constexpr' has been deprecated and replaced by option '--expt-relaxed-constexpr'.<\/p>\n<\/blockquote>\n\n<ol start=\"3\">\n<li>My GCC version is 5.2.1. Does this mean there is no way I can build TensorFlow from source?<\/li>\n<\/ol>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"36186634","CreationDate":"2016-03-23T19:04:25.183","ParentId":null,"OwnerUserId":"2057291","Title":null,"Body":"<p>Consider installing and temporary selecting an older version of the <code>gcc<\/code>:<\/p>\n\n<pre><code>apt-get install gcc-4.8\n<\/code><\/pre>\n\n<p>Then update your alternatives:<\/p>\n\n<pre><code>update-alternatives --remove-all gcc\nupdate-alternatives --install \/usr\/bin\/gcc gcc \/usr\/bin\/gcc-4.8 10\n<\/code><\/pre>\n\n<p>After compiling TensorFlow you can use previous version of <code>gcc<\/code> calling above command with proper path to <code>\/usr\/bin\/gcc-5.x<\/code>.<\/p>\n\n<p><a href=\"https:\/\/askubuntu.com\/a\/26518\/519613\">Here<\/a> you can find more detailed answer.<\/p>\n"}]}
{"QuestionId":36187825,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-03-23T20:11:43.443","AcceptedAnswerId":36244956.0,"OwnerUserId":3427580.0,"Title":"Where is layer module defined in PyCaffe","Body":"<p>I am modifying a <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/pascal-multilabel-with-datalayer.ipynb\" rel=\"noreferrer\">Caffe tutorial<\/a> to implement a neural network, but I am struggling to identify where some of the pycaffe modules are location in order to see certain function definitions.<\/p>\n\n<p>For example, the tutorial mentions:<\/p>\n\n<pre><\/pre>\n\n<p>Where can I find these function definitions and where can I see what other types of layers are pre-defined? I see that  and  are <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/net_spec.py\" rel=\"noreferrer\">defined here<\/a> but there's no mention of the types (e.g. , etc).<\/p>\n\n<p>The reason I am trying to figure this out is because there are other prototxt parameters left out of the pycaffe tutorials that I would like to be able to define from Python when generating the prototxts. These include,  and .<\/p>\n","answers":[{"AnswerId":"36244956","CreationDate":"2016-03-27T07:35:04.773","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can add the <code>blob_lr<\/code> and <code>phase<\/code> like this:<\/p>\n\n<pre><code>import caffe\nfrom caffe import layers a L, params as P\n\nns = caffe.NetSpec()\nns.conv = L.Convolution(bottom, convolution_param={'kernel_size':ks,\n                                                   'stride':stride,\n                                                   'num_output':nout, \n                                                   'pad':pad, \n                                                   'group':group},\n                                param=[{'lr_mult':1, 'decay_mult':1},\n                                       {'lr_mult':2, 'decay_mult':0}],\n                                include={'phase': caffe.TRAIN})\n<\/code><\/pre>\n\n<p>You can see some more examples in <a href=\"https:\/\/stackoverflow.com\/a\/35967589\/1714410\">this answer<\/a>.<\/p>\n"}]}
{"QuestionId":36188958,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2016-03-23T21:23:54.510","AcceptedAnswerId":36189235.0,"OwnerUserId":1104696.0,"Title":"How to use variable batch size for bidirectional RNN in tensorflow","Body":"<p>It seems tensorflow does not support variable batch size for bidirectional RNN. In this example the  is tied to , which is a Python integer:<\/p>\n\n<pre><\/pre>\n\n<p>How can I use different batch sizes for training and testing?<\/p>\n","answers":[{"AnswerId":"36189235","CreationDate":"2016-03-23T21:42:21.713","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The bidirectional code works with variable batch sizes. For example, take a look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/95e06c5fea9e730155196f42dea94ece2f56eb62\/tensorflow\/python\/kernel_tests\/rnn_test.py#L779\" rel=\"noreferrer\">this test code<\/a>, which creates a <code>tf.placeholder(..., shape=(None, input_size))<\/code> (where <code>None<\/code> means that the batch size can be variable).<\/p>\n\n<p>You can convert your code snippet to work with variable batch sizes with a small modification:<\/p>\n\n<pre><code># Compute the batch size based on the shape of the (presumably fed-in) `input`\n# tensor. (Assumes that `input = tf.placeholder(..., shape=[None, input_size])`.)\nbatch_size = tf.shape(input)[0]\n\n_seq_len = tf.fill(tf.expand_dims(batch_size, 0),\n                   tf.constant(n_steps, dtype=tf.int64))\noutputs, state1, state2 = rnn.bidirectional_rnn(rnn_fw_cell, rnn_bw_cell, input,\n                                                dtype=tf.float32,\n                                                sequence_length=_seq_len)\n<\/code><\/pre>\n"}]}
{"QuestionId":36189453,"AnswerCount":1,"Tags":"<python><gcc><theano>","CreationDate":"2016-03-23T21:55:42.567","AcceptedAnswerId":null,"OwnerUserId":3573182.0,"Title":"How to change Python default compiler to GCC?","Body":"<p>I have Windows 10 and Python 2.7 installed. When I run IDLE I find this:\nPython 2.7.10 (default, Oct 14 2015, 16:09:02) \n[MSC v.1500 32 bit (Intel)]<\/p>\n\n<p>I want the default compiler here to be MinGW's GCC (I already installed MinGW) becaue I cannot import Theano with the MSC compiler<\/p>\n\n<p>I tried all the tutorials out there and every time I successfully install Theano but when I try to import it I get the error \"Problem occurred during compilation with the command line below:\" and I get a huge list of errors. Btw, I don't have VS installed on my system<\/p>\n","answers":[{"AnswerId":"36189588","CreationDate":"2016-03-23T22:04:42.200","ParentId":null,"OwnerUserId":"6085157","Title":null,"Body":"<p>Edit Distutils config file C:\\Python2.7\\Lib\\distutils\\distutils.cfg (Create the file if it already does not exist).<\/p>\n\n<p>Add the following to the file:<\/p>\n\n<blockquote>\n  <p>[build]<\/p>\n  \n  <p>compiler = mingw32<\/p>\n<\/blockquote>\n\n<p>This should work.<\/p>\n"}]}
{"QuestionId":36190928,"AnswerCount":0,"Tags":"<ubuntu><pycharm><theano><theano-cuda><cudnn>","CreationDate":"2016-03-23T23:53:18.237","AcceptedAnswerId":null,"OwnerUserId":3025898.0,"Title":"theano running on GPU using terminal not in pycharm","Body":"<p>I have installed Theano + CUDA + CUDNN as per instructions on official webpage.<\/p>\n\n<p>When I run on terminal:<\/p>\n\n<pre><\/pre>\n\n<p>I get:<\/p>\n\n<pre><\/pre>\n\n<p>When I do the same in Pycharm I get:<\/p>\n\n<pre><\/pre>\n\n<p>I have added the CUDA folders to the interpreter path without any success. as you can see in the following pic:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ojQ55.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ojQ55.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any alternative recommendations on how to fix this?<\/p>\n","answers":[]}
{"QuestionId":36193553,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-03-24T04:58:58.457","AcceptedAnswerId":36193677.0,"OwnerUserId":3106889.0,"Title":"Get the value of some weights in a model trained by TensorFlow","Body":"<p>I have trained a ConvNet model with TensorFlow, and I want to get a particular weight in layer. For example in torch7 I would simply access . to get the weights of layer 2. How would I do the same thing in TensorFlow?<\/p>\n","answers":[{"AnswerId":"53741369","CreationDate":"2018-12-12T10:50:49.963","ParentId":null,"OwnerUserId":"9928929","Title":null,"Body":"<p>So if you proceed this code step by step, you will first get a list of used\/trainable variables. Then you could sort them in a list where you sort weight matrices\/lists to variable names, just for example how you were possible to deal with that information.  <\/p>\n\n<pre><code>vars = tf.trainable_variables()\nprint(vars) #some infos about variables...\nvars_vals = sess.run(vars)\nfor var, val in zip(vars, vars_vals):\n    print(\"var: {}, value: {}\".format(var.name, val)) #...or sort it in a list....\n<\/code><\/pre>\n"},{"AnswerId":"36193677","CreationDate":"2016-03-24T05:13:23.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow, trained weights are represented by <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\"><code>tf.Variable<\/code><\/a> objects. If you created a <code>tf.Variable<\/code>&mdash;e.g. called <code>v<\/code>&mdash;yourself, you can get its value as a NumPy array by calling <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session.run\"><code>sess.run(v)<\/code><\/a> (where <code>sess<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\"><code>tf.Session<\/code><\/a>).<\/p>\n\n<p>If you do not currently have a pointer to the <code>tf.Variable<\/code>, you can get a list of the trainable variables in the current graph by calling <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#trainable_variables\"><code>tf.trainable_variables()<\/code><\/a>. This function returns a list of all trainable <code>tf.Variable<\/code> objects in the current graph, and you can select the one that you want by matching the <code>v.name<\/code> property. For example:<\/p>\n\n<pre><code># Desired variable is called \"tower_2\/filter:0\".\nvar = [v for v in tf.trainable_variables() if v.name == \"tower_2\/filter:0\"][0]\n<\/code><\/pre>\n"}]}
{"QuestionId":36194228,"AnswerCount":0,"Tags":"<python><ipython><anaconda><theano>","CreationDate":"2016-03-24T06:11:08.610","AcceptedAnswerId":null,"OwnerUserId":3353185.0,"Title":"why does ipython hang with theano","Body":"<p>I've been trying to run various configurations of neural networks over the past few days using Keras, Theano and iPython (through Spyder). For a few days the code ran well but then iPython started hanging. This was on exactly the same code that ran without trouble in the past. <\/p>\n\n<p>I tried all sorts of remedies such as installing Anaconda updates and re-updating Theano to the 'bleeding edge' version. But to no avail. <\/p>\n\n<p>Then I just tried to run my code on the Python console in Spyder instead of iPython and it worked like a charm (no pun intended :-)). Therefore my guess is that there's something wrong with my iPython installation.<\/p>\n\n<p>So I have two questions: 1) is my diagnosis accurate and 2) whether it is or not, what could I try to remedy the problem? <\/p>\n\n<p>My config is:<\/p>\n\n<ul>\n<li>2015 15\" MacBook Pro with El Capitan<\/li>\n<li>Python 3.4.4<\/li>\n<li>Anaconda 2.4.1 (x86_64)<\/li>\n<li>IPython 4.1.2<\/li>\n<\/ul>\n\n<p>Thanks in advance,<\/p>\n\n<p>Edit: Response to Ogaday's questions: It's difficult to provide a mimimal working example as it depends on a fair amount of pre-processing and also needs inputs from other neural networks which I've trained separately. \nAnd on your suggestion I did try to run the code on iPython from the command line and it works! So maybe something's wrong with Spyder?<br>\nFinally 'which ipython' returns \/\/anaconda\/bin\/ipython. <\/p>\n","answers":[]}
{"QuestionId":36194394,"AnswerCount":1,"Tags":"<python><gpu><tensorflow>","CreationDate":"2016-03-24T06:28:53.597","AcceptedAnswerId":36204078.0,"OwnerUserId":349760.0,"Title":"How I reduce memory consumption in a loop in TensorFlow?","Body":"<p>I have a loop in TensorFlow that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting an OOM error when allocating the gradients for this layer, since each matrix multiplication is a different operation in the graph taking memory. Is there a way of preventing TensorFlow from allocating all these operations at the same time?<\/p>\n","answers":[{"AnswerId":"36204078","CreationDate":"2016-03-24T15:33:55.167","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is a challenging graph for TensorFlow to optimize, since the activations from each layer must be kept to aggregate a single gradient for <code>W<\/code>. One possibility is to pass the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0249729a26b2cd3bdbbc5880f8d634a56860b0fd\/tensorflow\/python\/ops\/gradients.py#L580\">experimental<\/a> <code>aggregation_method<\/code> argument when calling <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.minimize\"><code>optimizer.optimize()<\/code><\/a>.<\/p>\n\n<p>For example, you could try the following:<\/p>\n\n<pre><code>optimizer = tf.train.AdagradOptimizer(...)  # Or another optimization algorithm.\ntrain_op = optimizer.minimize(\n    total_loss,\n    aggregation_method=tf.AggregationMethod.EXPERIMENTAL_ACCUMULATE_N)\n<\/code><\/pre>\n\n<p>This option eagerly aggregates the gradients for recurrently-used variables in place, rather than keeping them all in memory until all of the gradients have been computed. If this doesn't work, the <code>tf.AggregationMethod.EXPERIMENTAL_TREE<\/code> may work better.<\/p>\n"}]}
{"QuestionId":36195454,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-24T07:56:00.037","AcceptedAnswerId":36203288.0,"OwnerUserId":234167.0,"Title":"What is the TensorFlow checkpoint meta file?","Body":"<p>When saving a checkpoint, TensorFlow often saves a meta file: . What is in that file, can we still restore a model even if we delete it and what kind of info did we lose if we restore a model without the meta file?<\/p>\n","answers":[{"AnswerId":"36203288","CreationDate":"2016-03-24T14:59:22.703","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This file contains a serialized <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/meta_graph.proto\"><code>MetaGraphDef<\/code> protocol buffer<\/a>. The <code>MetaGraphDef<\/code> is designed as a serialization format that includes all of the information required to restore a training or inference process (including the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/graph.proto\"><code>GraphDef<\/code><\/a> that describes the dataflow, and additional annotations that describe the variables, input pipelines, and other relevant information). For example, the <code>MetaGraphDef<\/code> is used by <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic.html\">TensorFlow Serving<\/a> to start an inference service based on your trained model. We are investigating other tools that could use the <code>MetaGraphDef<\/code> for training.<\/p>\n\n<p>Assuming that you still have the Python code for your model, you do not need the <code>MetaGraphDef<\/code> to restore the model, because you can reconstruct all of the information in the <code>MetaGraphDef<\/code> by re-executing the Python code that builds the model. To restore from a checkpoint, you only need the checkpoint files that contain the trained weights, which are written periodically to the same directory.<\/p>\n"}]}
{"QuestionId":36196719,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-03-24T09:21:07.253","AcceptedAnswerId":null,"OwnerUserId":3919542.0,"Title":"How to train a simple model using Tensorflow","Body":"<p>I am completely new to Tensorflow and machine learning and was thus going through the Tensorflow tutorials provided by google in udacity , The data is ready , However i don't have any idea as how can i apply a classifier to my data . It is the first assignment given in the tutotial .<\/p>\n\n<p>Link :<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/1_notmnist.ipynb\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/1_notmnist.ipynb<\/a><\/p>\n\n<p>I am facing problems with Problem 6 as i don't have much of a clue as what should i write there to create my own model and test it by giving my own images , Thus any help would help me understand the field of machine learning and get started with it . <\/p>\n","answers":[{"AnswerId":"36203924","CreationDate":"2016-03-24T15:26:41.443","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The tutorial suggests using <a href=\"http:\/\/scikit-learn.org\/\" rel=\"nofollow\">scikit-learn<\/a> to train an \"off-the-shelf\" classifier using the data. For example, the scikit-learn documentation has a section on <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/linear_model.html#logistic-regression\" rel=\"nofollow\">Logistic Regression<\/a>, which is a much simpler model than a deep neural network. There are <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.linear_model.LogisticRegression.html#examples-using-sklearn-linear-model-logisticregression\" rel=\"nofollow\">several examples<\/a> of how to use scikit-learn to perform classification.<\/p>\n"}]}
{"QuestionId":36197759,"AnswerCount":1,"Tags":"<gradient><theano><shared><lasagne>","CreationDate":"2016-03-24T10:16:35.877","AcceptedAnswerId":null,"OwnerUserId":5939610.0,"Title":"Calculate gradient for only part of a shared variable array","Body":"<p>I want to do the following:<\/p>\n\n<pre><\/pre>\n\n<p>Here, w_sub is for example w[1] but I do not want to explicitly write out b in function of w_sub. Despite going through <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/faq_tutorial.html\" rel=\"nofollow noreferrer\">this<\/a> and other related issues I can't solve it.<\/p>\n\n<p>This is just to show you my problem. Actually, what I really want to do is a sparse convolution with Lasagne. The zero entries in the weight matrix do not need to be updated and therefore there is no need to calculate the gradient for these entries of .<\/p>\n\n<p>This is now the complete error message:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"37466441","CreationDate":"2016-05-26T16:30:45.730","ParentId":null,"OwnerUserId":"6220857","Title":null,"Body":"<p>When theano compiles the graph, it only <em>sees<\/em> the variables as explicitly defined in the graph. In your example, <code>w_sub<\/code> is not explicitly used in the computation of <code>b<\/code> and therefore is not part of the computation graph.<\/p>\n\n<p>Using theano printing library with the following code, you can see on this \n<a href=\"http:\/\/i.stack.imgur.com\/ULg6S.png\" rel=\"nofollow\">graph vizualization<\/a> that indeed w_sub is not part of the graph of b.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import theano\nimport theano.tensor as T\nimport numpy\nimport theano.d3viz as d3v\n\na = T.fvector('a')\nw = theano.shared(numpy.array([1, 2, 3, 4], dtype=theano.config.floatX))\nw_sub = w[1]\nb = T.sum(a * w)\n\no = b, w_sub\n\nd3v.d3viz(o, 'b.html')\n<\/code><\/pre>\n\n<p>To fix the problem, you need to explicitly use <code>w_sub<\/code> in the computation of <code>b<\/code>.<\/p>\n\n<p>Then you will be able to compute the gradients of <code>b<\/code> wrt <code>w_sub<\/code> and update the values of the shared variable as in the following example : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import theano\nimport theano.tensor as T\nimport numpy\n\n\na = T.fvector('a')\nw = theano.shared(numpy.array([1, 2, 3, 4], dtype=theano.config.floatX))\nw_sub = w[1]\nb = T.sum(a * w_sub)\ngrad = T.grad(b, w_sub)\nupdates = [(w, T.inc_subtensor(w_sub, -0.1*grad))]\n\nf = theano.function([a], b, updates=updates, allow_input_downcast=True)\n\nf(numpy.arange(10))\n<\/code><\/pre>\n"}]}
{"QuestionId":36198648,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><backpropagation>","CreationDate":"2016-03-24T11:02:04.060","AcceptedAnswerId":null,"OwnerUserId":6108941.0,"Title":"What bookkeeping is caffe doing?","Body":"<p>Caffe <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/net_layer_blob.html\" rel=\"nofollow\">tutorial<\/a> states:<\/p>\n\n<blockquote>\n  <p>The net is a set of layers connected in a computation graph \u2013 a directed acyclic graph (DAG) to be exact. <em>Caffe does all the bookkeeping<\/em> for any DAG of layers to ensure correctness of the forward and backward passes. <\/p>\n<\/blockquote>\n\n<p>What is the meaning by \"all the bookkeeping\"?  I don't understand it.<br>\nHow to do all the bookkeeping?<\/p>\n","answers":[{"AnswerId":"36247990","CreationDate":"2016-03-27T13:40:43.727","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe, like many other deep-learning frameworks, trains its models using stochastic gradient decent (SGD), implemented as gradient back propagation. That is, for a mini-batch of training examples, caffe feed the batch through the net (\"forward pass\") to compute the loss w.r.t the net's parameters. Then, it propagates the loss gradient back (\"backward pass\") to update all the parameters according to the estimated gradient.<\/p>\n\n<p>By \"bookkeeping\" the tutorial means, you do not need to worry about estimating the gradients and updating the parameters. Once you are using existing layers (e.g., <code>\"Convolution\"<\/code>, <code>\"ReLU\"<\/code>, <code>\"Sigmoid\"<\/code> etc.) you only need to define the graph structure (the net's architecture) and supply the training data, and caffe will take care of the rest of the training process: It will forward\/backward each mini batch, compute the loss, estimate the gradients and update the parameters - all for you.<br>\nPretty awesome, don't you think? ;)<\/p>\n"}]}
{"QuestionId":36200043,"AnswerCount":2,"Tags":"<tensorflow><bazel>","CreationDate":"2016-03-24T12:18:03.557","AcceptedAnswerId":36201251.0,"OwnerUserId":3559686.0,"Title":"TensorFlow: Extension file not found: 'google\/protobuf\/protobuf.bzl'","Body":"<p>I am following <a href=\"http:\/\/www.aymeric.net\/journal\/2016\/01\/install-tensorflow-with-cuda-3-0-compute-compatibility-devices\/\" rel=\"nofollow\">this tutorial<\/a> to install GPU-enabled TensorFlow that is compatible with CUDA Compute Capability 3.0.<\/p>\n\n<p>I installed Java-JDK8, Bazel 0.1.0, TensorFlow 0.6.0, and changed the configurations to run on CUDA Compute Capability 3.0. Everything is good so far.<\/p>\n\n<p>But when I enter this command:<\/p>\n\n<blockquote>\n  <p>$HOME\/bin\/bazel build -c opt --config=cuda\n  \/\/tensorflow\/cc:tutorials_example_trainer<\/p>\n<\/blockquote>\n\n<p>I see this output:<\/p>\n\n<pre><\/pre>\n\n<p>Any advice?<\/p>\n","answers":[{"AnswerId":"36263735","CreationDate":"2016-03-28T13:49:53.753","ParentId":null,"OwnerUserId":"2259635","Title":null,"Body":"<p>I've had issues with the command above, -recurse-submodules does not exist<\/p>\n\n<p>Try this:<\/p>\n\n<p>$ git clone --recursive git@github.com:tensorflow\/tensorflow.git<\/p>\n"},{"AnswerId":"36201251","CreationDate":"2016-03-24T13:21:54.117","ParentId":null,"OwnerUserId":"3559686","Title":null,"Body":"<p>The problem is fixed by running this command:<\/p>\n\n<blockquote>\n  <p>$ git clone -b 0.6.0 \u2013recurse-submodules <a href=\"https:\/\/github.com\/tensorflow\/tensorflow.git\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow.git<\/a><\/p>\n<\/blockquote>\n\n<p>The error message I received is documented <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/605\" rel=\"nofollow\">here<\/a>. Pulling all submodules fixed the problem.<\/p>\n"}]}
{"QuestionId":36200052,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-dataflow>","CreationDate":"2016-03-24T12:18:23.240","AcceptedAnswerId":null,"OwnerUserId":1499038.0,"Title":"Running tensorflow model training from dataflow","Body":"<p>I am playing around with tensorflow and today I have noticed that google also open-sourced Python SDK for their dataflow.\nCurrently when I need to train and evaluate several networks in parallel I usually use either  and run one model training after another or I use  and I am performing each model training within the map step.\nWhole this data processing is just a part of the pipeline.<\/p>\n\n<p>I am wondering if there is or if there is planned something like perform tensorflow model training step inside of the dataflow pipeline? <\/p>\n\n<p>Is there currently some best practice around this? <\/p>\n\n<p>Or do I have to run each model setting within the map step?<\/p>\n\n<p>I went through the documentation and for now it seems to be really vague, so I'm asking here if someone has some experience with this.<\/p>\n","answers":[{"AnswerId":"36207559","CreationDate":"2016-03-24T18:32:29.820","ParentId":null,"OwnerUserId":"4539304","Title":null,"Body":"<p>There is nothing planned at this time.<\/p>\n\n<p>If you can run the Tensorflow training on a single machine (it sounds like this is what you were doing with Spark) then it should be possible to do the training within a <code>DoFn<\/code> of a Dataflow pipeline.<\/p>\n"}]}
{"QuestionId":36201062,"AnswerCount":1,"Tags":"<python><machine-learning><linear-regression><theano>","CreationDate":"2016-03-24T13:12:42.707","AcceptedAnswerId":36201791.0,"OwnerUserId":3906918.0,"Title":"Linear Regression with Theano - Dimension Mis-match","Body":"<p>I am familiarizing myself with Theano and machine learning. To that end, I'd like to calculate a linear regression. My code is inspired by the <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg\" rel=\"nofollow\">logistic regression example<\/a> from the introduction to Theano.    <\/p>\n\n<p>I wrote the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, Python returns the following error message when I run the code: <\/p>\n\n<pre><\/pre>\n\n<p>I suspect the error is related to the sample data being of dimension (50,1) and the regressor only of dimension of (1,1). Nevertheless, I failed to correct the error in my code since some time. Can somebody provide a hint for how to correct the mistake? I'm grateful for any help!<\/p>\n","answers":[{"AnswerId":"36201791","CreationDate":"2016-03-24T13:51:11.350","ParentId":null,"OwnerUserId":"262432","Title":null,"Body":"<p>You need to broadcast <code>b<\/code>:<\/p>\n\n<pre><code>self.y_pred = T.dot(input, self.W) + self.b[:, None]\n<\/code><\/pre>\n\n<p>I would expect Theano to do this automatically, but it seems this is not the case.<\/p>\n\n<hr>\n\n<p>To locate the problem follow the error message suggestion and run Theano with high exception verbosity<\/p>\n\n<pre><code>$ THEANO_FLAGS='exception_verbosity=high' python path\/to\/script.py\n<\/code><\/pre>\n\n<p>This produces quite a lot of output including the problematic node along with its operands<\/p>\n\n<pre><code>Debugprint of the apply node:\nElemwise{Composite{((-i0) + i1)}}[(0, 1)] [@A] &lt;TensorType(float64, vector)&gt; ''\n |b [@B] &lt;TensorType(float64, vector)&gt;\n |CGemv{no_inplace} [@C] &lt;TensorType(float64, vector)&gt; ''\n   |&lt;TensorType(float64, vector)&gt; [@D] &lt;TensorType(float64, vector)&gt;\n   |TensorConstant{-1.0} [@E] &lt;TensorType(float64, scalar)&gt;\n   |&lt;TensorType(float64, matrix)&gt; [@F] &lt;TensorType(float64, matrix)&gt;\n   |W [@G] &lt;TensorType(float64, vector)&gt;\n   |TensorConstant{1.0} [@H] &lt;TensorType(float64, scalar)&gt;\n<\/code><\/pre>\n\n<p>The node corresponds to a subtraction of <code>b<\/code> from a temporary node <code>CGemv{no_inplace}<\/code>. The only line of code involving <code>b<\/code> is <\/p>\n\n<pre><code>self.y_pred = T.dot(input, self.W) + self.b\n<\/code><\/pre>\n"}]}
{"QuestionId":36203585,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-03-24T15:12:30.900","AcceptedAnswerId":null,"OwnerUserId":5628492.0,"Title":"How to do matrix-scalar multiplication in TensorFlow?","Body":"<p>What would be the best way to multiply a matrix by a scalar in TensorFlow?\nI simply want to scale up the matrix by some scalar value. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52654589","CreationDate":"2018-10-04T20:21:32.297","ParentId":null,"OwnerUserId":"4360557","Title":null,"Body":"<p>very simple:<\/p>\n\n<p><strong>scalar * matrix<\/strong><\/p>\n\n<p>TensorFlow converts that to tf.multiply and broadcasts everything.<\/p>\n"},{"AnswerId":"56357491","CreationDate":"2019-05-29T09:38:36.577","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Equivalent variants:<\/p>\n\n<pre><code>x = tf.constant([[1.0, 0.0], [0.0, 1.0]]\ny1 = tf.scalar_mul(-1.0, x)\ny2 = tf.multiply(-1.0, x)\ny3 = -1.0 * x\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>sess.run(y1)\narray([[-1., -0.],\n       [-0., -1.]], dtype=float32)\n\nsess.run(y2)\narray([[-1., -0.],\n       [-0., -1.]], dtype=float32)\n\nsess.run(y3)\narray([[-1., -0.],\n       [-0., -1.]], dtype=float32)\n<\/code><\/pre>\n"},{"AnswerId":"46929968","CreationDate":"2017-10-25T10:27:32.313","ParentId":null,"OwnerUserId":"1587156","Title":null,"Body":"<pre><code>scalar_mul(scalar, x)\n<\/code><\/pre>\n\n<p>Multiplies a scalar times a <code>Tensor<\/code> or <code>IndexedSlices<\/code> object.<\/p>\n\n<p>Intended for use in gradient code which might deal with <code>IndexedSlices<\/code>\n  objects, which are easy to multiply by a scalar but more expensive to\n  multiply with arbitrary tensors.<\/p>\n\n<p>Args:\n    scalar: A 0-D scalar <code>Tensor<\/code>. Must have known shape.\n    x: A <code>Tensor<\/code> or <code>IndexedSlices<\/code> to be scaled.<\/p>\n\n<p>Returns:\n    <code>scalar * x<\/code> of the same type (<code>Tensor<\/code> or <code>IndexedSlices<\/code>) as <code>x<\/code>.<\/p>\n"},{"AnswerId":"36203807","CreationDate":"2016-03-24T15:21:49.937","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can multiply a matrix (or any other tensor) by a scalar using the <em>element-wise<\/em> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/multiply\" rel=\"nofollow noreferrer\"><code>tf.multiply()<\/code><\/a> operation, which implicitly broadcasts its arguments to match sizes:<\/p>\n\n<pre><code>x = tf.constant([[1.0, 0.0], [0.0, 1.0]])\ny = tf.multiply(x, 2.0)\n\nsess = tf.Session()\nprint sess.run(y)\n# ==&gt; [[2.0, 0.0], [0.0, 2.0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":36204225,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-24T15:39:59.617","AcceptedAnswerId":null,"OwnerUserId":1426205.0,"Title":"tensorflow custom op gradient","Body":"<p>We want to create a custom layer in tensorflow. Therefore, we decided to simply start with a toy example: a copy-Layer. After some try and error, we got to the point where it seems like the gradient would pass the right values through. However, in the second iteration the features get NAN's.\nIt may be a simple mistake, but currently I can't see it.<\/p>\n\n<p>In general, I have two questions:<\/p>\n\n<ol>\n<li>Can someone spot the problem here and how to solve it?<\/li>\n<li>What is a good method to debug a tensorflow session?<\/li>\n<\/ol>\n\n<hr>\n\n<p><strong>copy_op.cc<\/strong><\/p>\n\n<pre class=\"lang-cc prettyprint-override\"><\/pre>\n\n<p>We used the simple MNIST example as the basis:<\/p>\n\n<p><strong>layer_test.py<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><strong>compile<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>output:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Thanks a lot in advance!<\/p>\n","answers":[{"AnswerId":"47367518","CreationDate":"2017-11-18T15:08:56.917","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>from mrry in a comment:  There are known stability issues with using -<code>tf.reduce_sum(y_ * tf.log(y))<\/code> to compute cross-entropy (use <code>tf.nn.softmax_cross_entropy_with_logits(y, y_)<\/code> instead), and initializing your <code>W<\/code> variable to zeros often leads to a worse outcome than initializing it randomly.  <a href=\"https:\/\/stackoverflow.com\/questions\/36127436\/tensorflow-predicts-always-the-same-result\/36134261#36134261\">This answer<\/a> has more detail about the weight initialization issue.<\/p>\n"}]}
{"QuestionId":36204404,"AnswerCount":2,"Tags":"<python><theano><factorial>","CreationDate":"2016-03-24T15:47:32.857","AcceptedAnswerId":36204685.0,"OwnerUserId":6110242.0,"Title":"How to calculate factorial in Theano","Body":"<p>Here is a simple example. <\/p>\n\n<pre><\/pre>\n\n<p>There is few useful information on Internet. Something I found like \n<a href=\"http:\/\/matthewrocklin.com\/blog\/\/work\/2013\/08\/14\/SymPy-Theano-part-4\" rel=\"nofollow\">http:\/\/matthewrocklin.com\/blog\/\/work\/2013\/08\/14\/SymPy-Theano-part-4<\/a>\nwhich told me it belonged to \"the future work\", which means there is no solution for this problem.\nHow could that possible? I mean is there no body need calculate factorial in your theano project?\nI had this problem when I designed a RBM with Poisson conditional possibility which needs calculate factorial of visible units in the energy function. <\/p>\n","answers":[{"AnswerId":"36205568","CreationDate":"2016-03-24T16:45:00.763","ParentId":null,"OwnerUserId":"6016122","Title":null,"Body":"<pre><code>def factorial(x):\n    y = x-1\n    temp = x*y\n    it = y-1\n    for i in range(it):\n        y -= 1\n        temp = temp*y\n    print temp\n<\/code><\/pre>\n\n<p>Here is the edited solution without the Boolean operators. I know you found another answer but this may be helpful to you as well as others in the future.<\/p>\n"},{"AnswerId":"36204685","CreationDate":"2016-03-24T16:02:09.033","ParentId":null,"OwnerUserId":"6110242","Title":null,"Body":"<p>I have found the answer that we use gamma function to solve the factorial problem.\n:( Sorry for bothering everybody!\nHere is the detail:\n<a href=\"https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/ytqep8AickA\/cE7QeJZxXzUJ\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!msg\/theano-users\/ytqep8AickA\/cE7QeJZxXzUJ<\/a><\/p>\n\n<p>So here should be:<\/p>\n\n<pre><code>import theano \nv = theano.tensor.scalar(\"variable\") \nfactorial = theano.tensor.gamma(v)\n# Calculate factorial of v , like numpy.math.factorial(v), but it didn't work, \n# Actually it should be a integer as a parameter of numpy.math.factorial function \nresult = theano.function([v], factorial)\n<\/code><\/pre>\n"}]}
{"QuestionId":36205137,"AnswerCount":1,"Tags":"<machine-learning><theano><deep-learning><training-data><stochastic>","CreationDate":"2016-03-24T16:24:00.540","AcceptedAnswerId":36220369.0,"OwnerUserId":2776945.0,"Title":"theano - how to have many of the same function","Body":"<p>The input is a variable sized array. I can only process a given example one sample at a time in train_model. I want to accumulate the sum of objectives for the elements in the batch then apply regularization and gradient descent. <\/p>\n\n<p>Currently, this is the training stage where updates are done for each element xi.<\/p>\n\n<pre><\/pre>\n\n<p>How can I get results from train_model(xi) for the number of elements in the batch and then do the updates?<\/p>\n","answers":[{"AnswerId":"36220369","CreationDate":"2016-03-25T12:55:46.433","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>just use all the elements in <code>dataset.get_next_xi(batch_size)<\/code> as input and create a theano function to calculate the average cost (instead of only one cost) and do the updates using the average cost. You can see the example code from <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow\">here<\/a><\/p>\n\n<p>they use theano function from train model like this:<\/p>\n\n<pre><code>train_model = theano.function(\n    inputs=[index],\n    outputs=cost,\n    updates=updates,\n    givens={\n        x: train_set_x[index * batch_size: (index + 1) * batch_size],\n        y: train_set_y[index * batch_size: (index + 1) * batch_size]\n    }\n)\n<\/code><\/pre>\n\n<p>with <code>cost<\/code> is average cost of datasets batch<\/p>\n"}]}
{"QuestionId":36207551,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><keras>","CreationDate":"2016-03-24T18:31:45.017","AcceptedAnswerId":36925362.0,"OwnerUserId":4380945.0,"Title":"changes in keras syntax?","Body":"<p>I have a script that worked just fine in a previous version of keras. Running the script on distribution 0.3.2, I get all sort of error messages. Below you will find some lines of the script:<\/p>\n\n<pre><\/pre>\n\n<p>Some of the errors that I get are:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any 'from-to' type of document to go through and be in an easier position to update older keras scripts?<\/p>\n","answers":[{"AnswerId":"36925362","CreationDate":"2016-04-28T20:59:29.317","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>You need to set an input dimension because it's the only dimension which cannot be inferred from previous layers.<\/p>\n"}]}
{"QuestionId":36207963,"AnswerCount":0,"Tags":"<gpu><tensorflow>","CreationDate":"2016-03-24T18:55:30.840","AcceptedAnswerId":null,"OwnerUserId":6111004.0,"Title":"Tensorflow: GPU memory exhuastion and max pool layers","Body":"<p>I have a pretty simple CNN setup (3 convolution layers into a fully connected layer) , using a Geforce GTX 960 GPU with 4GB of memory. I was hitting a GPU resource exhausted issue, and while playing around I discovered if I put a pooling layer (which reduces the overall size of the layer by a factor of 4) it actually made the problem worse. In other words, I was hitting the resource exhausted, and then after removing the pooling layer, I was able to train. I intuitively thought that since a pooling layer would reduce the number of operations downstream during the optimization piece of the training that it would help the memory constraints. Does any know how a max pooling layer could potentially add operations and increase memory like this?  <\/p>\n","answers":[]}
{"QuestionId":36209335,"AnswerCount":2,"Tags":"<anaconda><tensorflow><skflow>","CreationDate":"2016-03-24T20:21:21.053","AcceptedAnswerId":36299462.0,"OwnerUserId":1810559.0,"Title":"Install SKFLOW to work with Anaconda","Body":"<p>Now that SKFLOW has been moved within TensorFlow I am not sure how to use it. I am running Anaconda on Ubuntu server and I installed Tensorflow as suggested in this question: <a href=\"https:\/\/stackoverflow.com\/questions\/33646541\/tensorflow-and-anaconda-on-ubuntu\">Tensorflow and Anaconda on Ubuntu?<\/a><\/p>\n\n<p>My scripts are seeing Tensorflow but not SKFLOW.<\/p>\n","answers":[{"AnswerId":"36212090","CreationDate":"2016-03-25T00:08:12.537","ParentId":null,"OwnerUserId":"1193874","Title":null,"Body":"<p>I don't think that skflow is in the latest, 0.7.1, tensorflow release. And therefore you won't find it if you have just conda installed tensorflow via the <a href=\"https:\/\/stackoverflow.com\/questions\/33646541\/tensorflow-and-anaconda-on-ubuntu\">post you mentioned<\/a>. <\/p>\n\n<p>For example, checking the <a href=\"https:\/\/anaconda.org\/jjhelmus\/tensorflow\/files\" rel=\"nofollow noreferrer\">anaconda channel for tensorflow<\/a> it looks like 0.7.1 was uploaded on March 3rd while the first check-in of skflow in the tensorflow code base was on March 9th. I also don't see any notice of skflow in the recent release mentions of tensorflow.<\/p>\n\n<p>That said, when it is released with skflow as part of it, I expect you'll be able to import it like this: <\/p>\n\n<p><code>import tensorflow.contrib.skflow<\/code><\/p>\n"},{"AnswerId":"36299462","CreationDate":"2016-03-30T04:03:28.987","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>I have shared this in our Gitter channel but it might useful to repost it here. <\/p>\n\n<p>You can use <code>pip<\/code> to install TensorFlow's <strong><em>nightly binaries<\/em><\/strong> as mentioned in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#installation\" rel=\"nofollow\">README<\/a> file if you don't want to wait until next official release. <\/p>\n\n<p>Then at least for the date I posted this, you can import skflow directly from TensorFlow via <code>from tensorflow.contrib.skflow.python import skflow<\/code> and then everything else works the same as usual. <\/p>\n\n<p>Hope this helps. <\/p>\n\n<p><strong>Edit<\/strong>: Once <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/1712\" rel=\"nofollow\">this PR<\/a> is merged, you'll be able to import as done in examples folder in TensorFlow, e.g. <code>from tensorflow.contrib import skflow<\/code>.<\/p>\n"}]}
{"QuestionId":36210162,"AnswerCount":1,"Tags":"<python><multithreading><queue><tensorflow>","CreationDate":"2016-03-24T21:19:01.477","AcceptedAnswerId":36211686.0,"OwnerUserId":216880.0,"Title":"Tensorflow: Stopping threads via Coordinator seems not to work?","Body":"<p>Following a <a href=\"https:\/\/stackoverflow.com\/questions\/34594198\/how-to-prefetch-data-using-a-custom-python-function-in-tensorflow\">standard prefetching queue<\/a>, I extended the aformentioned example by some additional validation code, see attached code. That is, every ith training step, the learned model is evaluated on a validation set (several, in my case). The validation set can't be fed through the queue, so one possible idea is to build an additional inference graph using the shared variables.<\/p>\n\n<p>This somehow works, but after training is finished, the program hangs (at ) and eventually throws an exception:  and then the asynchronous loading thread also throws an exception. The  exception can be tackled via a  clause (see code below), but the async thread still throws an exception (which does not hamper the main program, though, but should not happen in my opinon---it has the  loop that should tell it to stop). <\/p>\n\n<p>Interestingly, if training is done without any evaluation code running (that is, the block after  commented out), then  does not hang at all.<\/p>\n\n<p>My question: What am I doing wrong here? It seems as if  is not doing what I hope it should?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36211686","CreationDate":"2016-03-24T23:26:57.183","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is a race condition in your code. The thread running <code>async_load()<\/code> will block forever if the following events happen:<\/p>\n\n<ol>\n<li><code>async_load()<\/code> calls <code>coord.should_stop()<\/code> which returns <code>False<\/code>.<\/li>\n<li><code>async_load()<\/code> calls <code>session.run(queue, ...)<\/code> but the queue is full so the call blocks indefinitely.<\/li>\n<li>Main thread calls <code>coord.request_stop()<\/code>.<\/li>\n<li>Main thread calls <code>coord.join([t])<\/code>, and this blocks forever because of (2).<\/li>\n<\/ol>\n\n<p>One way to avoid this is to create a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.close\" rel=\"noreferrer\"><code>queue.close(cancel_pending_enqueues=True)<\/code><\/a> op, and run it in the main thread before calling <code>coord.request_stop()<\/code>. This will unblock the <code>async_load()<\/code> thread, and enable <code>coord.join([t])<\/code> to return.<\/p>\n"}]}
{"QuestionId":36210887,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><tensorflow>","CreationDate":"2016-03-24T22:12:23.330","AcceptedAnswerId":36211137.0,"OwnerUserId":2235817.0,"Title":"How to fix MatMul Op has type float64 that does not match type float32 TypeError?","Body":"<p>I am trying to save Nueral Network weights into a file and then restoring those weights by initializing the network instead of random initialization. My code works fine with random initialization. But, when i initialize weights from file it is showing me an error  I don't know how do i solve this issue.Here is my code:<\/p>\n\n<p><strong>Model Initialization<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Model parameters<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>ERROR DETAILS<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/2ZNj9.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2ZNj9.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"36211137","CreationDate":"2016-03-24T22:35:33.180","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/math_ops.html#matmul\" rel=\"noreferrer\"><code>tf.matmul()<\/code><\/a> op does not perform automatic type conversions, so both of its inputs must have the same element type. The error message you are seeing indicates that you have a call to <code>tf.matmul()<\/code> where the first argument has type <code>tf.float32<\/code>, and the second argument has type <code>tf.float64<\/code>. You must convert one of the inputs to match the other, for example using <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#cast\" rel=\"noreferrer\"><code>tf.cast(x, tf.float32)<\/code><\/a>.<\/p>\n\n<p>Looking at your code, I don't see anywhere that a <code>tf.float64<\/code> tensor is explicitly created (the default <code>dtype<\/code> for floating-point values in the TensorFlow Python API&mdash;e.g. for <code>tf.constant(37.0)<\/code>&mdash;is <code>tf.float32<\/code>). I would guess that the errors are caused by the <code>np.loadtxt(filename)<\/code> calls, which might be loading an <code>np.float64<\/code> array. You can explicitly change them to load <code>np.float32<\/code> arrays (which are converted to <code>tf.float32<\/code> tensors) as follows:<\/p>\n\n<pre><code>initial = np.loadtxt(filename).astype(np.float32)\n<\/code><\/pre>\n"},{"AnswerId":"47604528","CreationDate":"2017-12-02T04:17:17.000","ParentId":null,"OwnerUserId":"8163412","Title":null,"Body":"<p>Although It's an old question but I would like you include that I came across the same problem. I resolved it using <code>dtype=tf.float64<\/code> for parameter initialization and for creating X and Y placeholders as well.<\/p>\n\n<p>Here is the snap of my code.<\/p>\n\n<pre><code>X = tf.placeholder(shape=[n_x, None],dtype=tf.float64)\nY = tf.placeholder(shape=[n_y, None],dtype=tf.float64)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>parameters['W' + str(l)] = tf.get_variable('W' + str(l), [layers_dims[l],layers_dims[l-1]],dtype=tf.float64, initializer = tf.contrib.layers.xavier_initializer(seed = 1))\nparameters['b' + str(l)] = tf.get_variable('b' + str(l), [layers_dims[l],1],dtype=tf.float64, initializer = tf.zeros_initializer())\n<\/code><\/pre>\n\n<p>Declaring all placholders and parameters with float64 datatype will resolve this issue.<\/p>\n"}]}
{"QuestionId":36211594,"AnswerCount":0,"Tags":"<python><neural-network><sparse-matrix><theano><backpropagation>","CreationDate":"2016-03-24T23:18:21.060","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"convolution neural network with backpropagation and sparsity in python","Body":"<p>I am trying to modify the code provided by neural-networks-and-deep-learning on github for <a href=\"https:\/\/github.com\/mnielsen\/neural-networks-and-deep-learning\/blob\/master\/src\/network3.py\" rel=\"nofollow\">network3.py<\/a>. This code basically constructs a convolution neural network and trains the MNIST data set.<\/p>\n\n<p>What I am trying to do is add the concept of back propagation and sparsity to this code. The part of code which I added is outlined between the two lines of #. I get an <\/p>\n\n<p>I understand that the size should be of 4D (1,1,28,28) but I am not sure where and how to do this modification. <\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know how to fix this?<\/p>\n\n<p>The main code I run to call the above script is<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36212718,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-03-25T01:29:31.350","AcceptedAnswerId":null,"OwnerUserId":1322301.0,"Title":"TensorFlow: Labels as relations between examples","Body":"<p>I have a somewhat unusual learning task: the \"labels\" for my examples (i.e., images) are actually relations between specific images. For instance, suppose each batch consists of  images, where the pair  has label  while the pair  has label . Suppose most (i.e., the <em>vast<\/em> majority) pairs possible from the set of images lack any label, and as such random sampling will only rarely yield even a single relation to train on. Therefore, in this example, each batch consists of a set  images,  pairs of which are known to have some relation.<\/p>\n\n<p>TensorFlow is so flexible it's easy to design a net that can compute the appropriate loss function. However, creating the input is problematic. I don't remember C++ well enough to implement a custom input op, so I either have to read the data into python in the appropriate manner and feed it to TensorFlow (easy, but slow) or glue some ops together to get it to work if possible. <\/p>\n\n<p>Accomplishing this appears very unwieldy, requiring many queue objects (why do readers only accept queues in TensorFlow?!). Assuming that I have some set of files which are serialized examples, each a list of the form , is there a simple way to serve them to the net as a tensor of shape  (with image i and image i + 1 belonging to a single relation, where i is  i.e., )? My current creation, individual components of which work but works as a whole only in theory, employs 5 queue objects of various kinds and does not shuffle perfectly.<\/p>\n","answers":[]}
{"QuestionId":36213642,"AnswerCount":2,"Tags":"<macos><python-3.x><clang><theano>","CreationDate":"2016-03-25T03:41:06.803","AcceptedAnswerId":36222154.0,"OwnerUserId":3335387.0,"Title":"MAC OSX10.11.4 python3 import theano error","Body":"<p>I upgraded my Mac to a OSX 10.11.4, and sadly I found my theano cannot be imported anymore. \nHere is information about my machine:<\/p>\n\n<pre><\/pre>\n\n<p>The logs are list below and I hope somebody can help me out. The output is very long and I just listed part of it here. The full version can be found <a href=\"http:\/\/liujr.com\/2016\/03\/25\/theano-errors\/\">here<\/a>.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36222154","CreationDate":"2016-03-25T14:45:39.263","ParentId":null,"OwnerUserId":"3688908","Title":null,"Body":"<p>The problem comes from an interface change in Xcode 7.3.  It is fixed in the current master of Theano and there will be release 0.8.1 next Tuesday (2016\/03\/29) that will be 0.8.0 + the fix.<\/p>\n\n<p>In the meantime you can either run the master: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions\">http:\/\/deeplearning.net\/software\/theano\/install.html#bleeding-edge-install-instructions<\/a> or install version 7.2 of the command line tools and select those with xcode-select to make Theano work.<\/p>\n"},{"AnswerId":"36274688","CreationDate":"2016-03-29T03:09:04.920","ParentId":null,"OwnerUserId":"2178800","Title":null,"Body":"<p>A new Numpy release (1.11.0) came out yesterday, 3\/27\/16. Installed it and finally everything on my MacBook OS-X Canopy Python started working correctly. Installed scikit-neuralnetwork, which installs Theano, Numpy, and others.  Whew...<\/p>\n"}]}
{"QuestionId":36215346,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-25T06:47:34.383","AcceptedAnswerId":null,"OwnerUserId":315046.0,"Title":"Tensorflow: 'module' object has no attribute 'FixedLenFeature'","Body":"<p>I was running through Google's Tensorflow's fully_connected_reader.py example (I have executed convert_to_record.py, and source code is here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/how_tos\/reading_data\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/how_tos\/reading_data<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>Anyone got any idea?   (it is running from the latest github version).<\/p>\n","answers":[{"AnswerId":"36217548","CreationDate":"2016-03-25T09:45:30.347","ParentId":null,"OwnerUserId":"3318517","Title":null,"Body":"<p><code>FixedLenFeature<\/code> was <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/492e55ff2297dc13f5ed540056ad40603f2b1dd0\" rel=\"nofollow\">added<\/a> in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/tag\/v0.7.0\" rel=\"nofollow\">TensorFlow 0.7.0<\/a>. You have an earlier version installed. You can check your version with <code>print tf.__version__<\/code>.<\/p>\n"}]}
{"QuestionId":36215503,"AnswerCount":0,"Tags":"<macos><boost><caffe>","CreationDate":"2016-03-25T07:03:59.610","AcceptedAnswerId":null,"OwnerUserId":1910545.0,"Title":"Installing Boost C++ on Mac OS X Capitan","Body":"<p>I am trying to install the boost C++ library as a dependency for Caffe on Mac OS X EL Capitan. I have tried to build from source using <\/p>\n\n<pre><\/pre>\n\n<p>but the build process never passes the below stage for almost a day. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HICdf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HICdf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So, I have tried installing the pre-packaged version using just<\/p>\n\n<pre><\/pre>\n\n<p>this time it stops at the below stage for more than 2 hours.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9uI79.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9uI79.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any help or work around for this problem is really appreciated. \nThank you in advance.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>I figured out that for me there is a problem installing boost 1.57.0 either from source or the pre-packaged version. <\/p>\n\n<p>So, I did  then\n  and that solved it. However, installing with --build-from-souce option still gets stuck.<\/p>\n","answers":[]}
{"QuestionId":36217978,"AnswerCount":1,"Tags":"<python><performance><theano>","CreationDate":"2016-03-25T10:14:08.230","AcceptedAnswerId":null,"OwnerUserId":6007803.0,"Title":"there is no speed-up using multi-gpus","Body":"<p>I try to test the performance of multi-gpus using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I got results as the attached picture showed, there is no improvment when I used two different gpu device.(THEANO_FLAGS='contexts=dev0->cuda0;dev1->cuda1')<\/p>\n\n<p>The log in the terminal is as follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41625088","CreationDate":"2017-01-12T23:46:54.807","ParentId":null,"OwnerUserId":"6442723","Title":null,"Body":"<p>A few changes will reveal the improvement:<\/p>\n\n<ol>\n<li>Tell the function outputs to stay on the GPU using e.g. <code>theano.tensor.dot(v_00, v_01).transfer('dev0')<\/code>.<\/li>\n<li>Call the function once before timing (it will compile), then many times in a loop while timing, and drop the sizes from 10,240 back down to 1,024.<\/li>\n<li>Only take the final answer off the GPU e.g. <code>np.asarray(res0)<\/code> or maybe even <code>print(res_0)<\/code> after the loop is complete, and time when this is done.  This ensures all the computations are actually done.<\/li>\n<\/ol>\n\n<p>What you have to avoid to observe the speedup is transferring the output to the CPU every time the math is executed.  The time elapsed from the start of the loop to the end of the <code>np.assarray<\/code> call will be the total time to compute and to transfer the last answer.  The time elapsed during the loop will only be the time to queue up the operations in the GPU, but not to complete them.  I'm not sure how to time the GPU execution time in isolation; you have to infer it.<\/p>\n"}]}
{"QuestionId":36218823,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><recurrent-neural-network>","CreationDate":"2016-03-25T11:12:13.150","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"Char-RNN in Tensorflow","Body":"<p>I'm trying to get a simple RNN to work in tensorflow, but I'm having several problems with it.<\/p>\n\n<p>What I'm trying to do for now, is to simply run a forward pass of the RNN with an LSTM as its cell type.<\/p>\n\n<p>I have scraped some news articles and want to feed them into the RNN. I have split the string, which consists of the concatenation of all articles, into chars and mapped the chars to integers. Then I have one-hot encoded those integers.<\/p>\n\n<pre><\/pre>\n\n<p>Now for the tensorflow code. I want to run through all the characters in the data and use 25 characters for each forward pass.\nMy first question is with regards to the batch size, if I want to do this the way I just mentioned, my batch size is 1, right? So each vector that corresponds to one char in the input has shape [1,vocab_size] and there are 25 of those vectors in my input. So I have used the following tensors:<\/p>\n\n<pre><\/pre>\n\n<p>I had to create the last tensor, because thats the format that the rnn function expects.<\/p>\n\n<p>Then I encountered a problem with the variable scope, I get the error below:<\/p>\n\n<pre><\/pre>\n\n<p>And I'm not sure why I get this error, because I actually haven't specified any variables in my code, the variables are only created inside the rnn and rnn_cell functions, can someone tell me how to fix this error?<\/p>\n\n<p>Another error I'm currently getting is a type error, since my input is of type tf.int32, but the hidden layer that is created inside the LSTM is of type tf.float32 and the Linear function inside the rnn_cell.py code, concatenates those two tensor and multiplies them by a weight matrix. Why is this not possible, I'd assume that it's relatively common for input to be one-hot encoded and thus having type int32?<\/p>\n\n<p>In general, is this approach of having a batch size of 1 standard when training char-rnns? I've only seen code by Andrej Karpathy, where he trains a char-rnn in basic numpy and he uses the same procedure, where he simply goes through the whole text in sequences of length 25. Here is the code: <a href=\"https:\/\/gist.github.com\/karpathy\/d4dee566867f8291f086\" rel=\"nofollow\">https:\/\/gist.github.com\/karpathy\/d4dee566867f8291f086<\/a><\/p>\n","answers":[]}
{"QuestionId":36221588,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-03-25T14:13:15.303","AcceptedAnswerId":null,"OwnerUserId":1104696.0,"Title":"How to resume training from *.meta in tensorflow?","Body":"<p>In the latest version of tensorflow, when I save the model I find two files are produced: model_xxx and model_xxx.meta.<\/p>\n\n<p>Does model_xxx.meta specify the network? Can I resume training using model_xxx and model_xxx.meta without specify the network in the code? What about training queue structure, are they stored in model_xxx.meta?<\/p>\n","answers":[{"AnswerId":"38489976","CreationDate":"2016-07-20T20:11:11.670","ParentId":null,"OwnerUserId":"6471140","Title":null,"Body":"<p>Not sure if this will work for you, but at least for <code>DNNCLassifiers<\/code> you can specify the <code>model_dir<\/code> parameter when creating it and that will construct the model from the files and then you can continue the training.<\/p>\n\n<p>For the <code>DNNClassifiers<\/code> you specify <code>model_dir<\/code> when first creating the object and the training will store checkpoints and other files on this directory. You can come then after that, and create another <code>DNNClassifier<\/code> specifying the same <code>model_dir<\/code> and that will restore your pre-trained model.<\/p>\n"}]}
{"QuestionId":36222076,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-03-25T14:41:03.040","AcceptedAnswerId":null,"OwnerUserId":6114330.0,"Title":"Theano indexing details","Body":"<p>I'm following Theano tutorial (particularly <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/logreg.html<\/a>) and I'm trying to understand this \"self evident\" code:<\/p>\n\n<pre><\/pre>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>what is the [T.arange(y.shape[0]), y] after the function call? <\/li>\n<li>How does it return [LP[0,y[0]], LP[1,y[1]], LP[2,y[2]], ..., LP[n-1,y[n-1]]]<\/li>\n<li>How does T.mean know how to compute the mean over something like that (whatever it is)?<\/li>\n<li>Could you please generalize concepts I don't understand?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"36222968","CreationDate":"2016-03-25T15:34:41.853","ParentId":null,"OwnerUserId":"6114330","Title":null,"Body":"<p>Found the answer (NLL stands for negative log likelihood): <\/p>\n\n<p>NLL is a symbolic variable ; to get the actual value of NLL, this symbolic\nexpression has to be compiled into a Theano function (see the Theano\ntutorial for more details)\n<code>NLL = -T.sum(T.log(p_y_given_x)[T.arange(y.shape[0]), y])<\/code>\nnote on syntax: <code>T.arange(y.shape[0])<\/code> is a vector of integers <code>[0,1,2,...,len(y)]<\/code>.\nIndexing a matrix M by the two vectors <code>[0,1,...,K], [a,b,...,k]<\/code> returns the\nelements <code>M[0,a], M[1,b], ..., M[K,k]<\/code> as a vector.  Here, we use this\nsyntax to retrieve the log-probability of the correct labels, y.<\/p>\n"}]}
{"QuestionId":36223157,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-03-25T15:45:42.937","AcceptedAnswerId":36223553.0,"OwnerUserId":6114511.0,"Title":"Set weight and bias tensors of tensorflow conv2d operation","Body":"<p>I have been given a trained neural network in torch and I need to rebuild it exactly in tensorflow. I believe I have correctly defined the network's architecture in tensorflow but I am having trouble transferring the weight and bias tensors. Using a third party package, I converted all the weight and bias tensors from the torch network to numpy arrays then wrote them to disk. I can load them back into my python program but I cannot figure out a way to assign them to the corresponding layers in my tensorflow network. <\/p>\n\n<p>For instance, I have a convolution layer defined in tensorflow as<\/p>\n\n<pre><\/pre>\n\n<p>According to the tensorflow documentation, the tf.nn.conv2d operation uses the shape defined in the kernel_1 variable to construct the weight tensor. However, I cannot figure out how to access that weight tensor to set it to the weight array I have loaded from file. <\/p>\n\n<p><strong>Is it possible to explicitly set the weight tensor? And if so, how?<\/strong> <\/p>\n\n<p>(The same question applies to bias tensor.)<\/p>\n","answers":[{"AnswerId":"36223553","CreationDate":"2016-03-25T16:06:43.547","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you have the weights and biases in a NumPy array, it should be easy to connect them into your TensorFlow network:<\/p>\n\n<pre><code>weights_1_array = ...  # ndarray of weights for layer 1\nbiases_1_array = ...  # ndarray of biases for layer 1\n\nconv_kernel_1 = tf.nn.conv2d(input, weights_1_array, [1, 4, 4, 1], padding='SAME')\nbias_layer_1 = tf.nn.bias_add(conv_kernel_1, biases_1_array)\n<\/code><\/pre>\n\n<p>Note that you must ensure that <code>weights_1_array<\/code> and <code>biases_1_array<\/code> are in the correct data format. See the documentation for <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn.html#conv2d\" rel=\"noreferrer\"><code>tf.nn.conv2d()<\/code><\/a> for an explanation of the required filter shape.<\/p>\n"}]}
{"QuestionId":36225498,"AnswerCount":1,"Tags":"<python><theano><deep-learning><keras>","CreationDate":"2016-03-25T18:12:42.483","AcceptedAnswerId":36242055.0,"OwnerUserId":4380945.0,"Title":"theano error from keras","Body":"<p>I am running a keras script (no direct call to theano in my script) and I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have seen the error from folks running theano directly, but not through keras. Not sure what I should do, since I am not dealing with tensors directly.<\/p>\n","answers":[{"AnswerId":"36242055","CreationDate":"2016-03-26T23:21:05.083","ParentId":null,"OwnerUserId":"4380945","Title":null,"Body":"<p>the problem was that there is a change in keras version (I am currently using keras 0.3.2 with theano 0.8.0) and what used to be fine does not work well with he new keras version.<\/p>\n\n<p>The following was the original code, and see the fix below.<\/p>\n\n<pre><code>from keras.models import Sequential\nimport keras.optimizers\nfrom keras.layers.core import Dense, Dropout\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.layers.advanced_activations import PReLU\nfrom keras.layers.core import Activation\nfrom keras.optimizers import SGD, Adam\n\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.base import BaseEstimator, RegressorMixin\n\nclass NnRegression(BaseEstimator, RegressorMixin):\n        def __init__(self, apply_standart_scaling=True,\n             dropx=[0.2, 0.5, 0.5], nb_neuronx=[50, 30], nb_epoch=105, validation_split=0.,\n             verbose=1):\n        self.apply_standart_scaling = apply_standart_scaling\n        self.dropx = dropx\n        self.nb_neuronx = nb_neuronx\n        self.nb_epoch = nb_epoch\n        self.validation_split = validation_split\n        self.verbose = verbose\n\n    def fit(self, X, y):\n\n        nb_features = X.shape[1]\n        self.standart_scaling = StandardScaler() if self.apply_standart_scaling else None\n\n        if self.standart_scaling:\n            X = self.standart_scaling.fit_transform(X)\n\n        model = Sequential()\n        model.add(Dropout(input_shape = (nb_features,),p= self.dropx[0]))\n        model.add(Dense(output_dim = self.nb_neuronx[0], init='glorot_uniform'))\n        model.add(PReLU())\n        model.add(BatchNormalization(self.nb_neuronx[0],)))\n        model.add(Dropout(self.dropx[1]))\n\n        model.add(Dense(self.nb_neuronx[1], init='glorot_uniform'))\n        model.add(PReLU())\n        model.add(BatchNormalization(self.nb_neuronx[0],)))\n        model.add(Dropout(self.dropx[2]))\n\n        model.add(Dense(1, init='glorot_uniform'))\n\n        nn_verbose = 1 if self.verbose&gt;0 else 0\n        optz = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)\n        model.compile(optimizer=Adam(),loss='mse')\n        model.fit(X, y, batch_size=16,\n              nb_epoch=self.nb_epoch, validation_split=self.validation_split, verbose=nn_verbose)\n\n        self.model = model\n\n    def predict(self, X):\n        if self.standart_scaling:\n            X = self.standart_scaling.transform(X)\n        return self.model.predict_proba(X, verbose=0)\n<\/code><\/pre>\n\n<p>well, it turns out that the problem is this single line of code:<\/p>\n\n<pre><code>model.add(BatchNormalization(self.nb_neuronx[0],)))\n<\/code><\/pre>\n\n<p>It should actually be:<\/p>\n\n<pre><code>model.add(BatchNormalization())\n<\/code><\/pre>\n\n<p>because the number of neurons has no business within the normalization layer (however this did not bother in a previous keras version).<\/p>\n\n<p>This apparently causes theano to generate new weights that are not float32 but float64, and that triggers the message above.<\/p>\n"}]}
{"QuestionId":36225584,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-03-25T18:18:17.240","AcceptedAnswerId":null,"OwnerUserId":1079110.0,"Title":"How to use the last state of a variable as the next state in Tensorflow?","Body":"<p>For learning purposes, I want to build my own LSTM model in Tensorflow. The problem is, how to train is in a way that the states at a certain timestep get initialized using the states from the previous timestep. Is there a mechanism for this in Tensorflow?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36240035","CreationDate":"2016-03-26T19:42:45.983","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>@danijar - you may want to look at the 'Variables' section of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/basic_usage.html\" rel=\"nofollow\">this page<\/a> for a simple example of how to maintain state across calls to a subgraph.<\/p>\n"}]}
{"QuestionId":36225929,"AnswerCount":1,"Tags":"<python><tensorflow><convolution>","CreationDate":"2016-03-25T18:41:06.367","AcceptedAnswerId":null,"OwnerUserId":6115219.0,"Title":"What's the best way to implement 3D convolution in tensorflow?","Body":"<p>It seems the current version of Tensorflow doesn't support any command to directly perform 3D convolution. Theano does so, but I don't want to switch to Theano. I was just wondering is there any simple way to perform 3D convolution by perhaps using the available commands like conv2d? If anybody has the code to share, I would appreciate it.<\/p>\n","answers":[{"AnswerId":"37604294","CreationDate":"2016-06-03T00:46:36.907","ParentId":null,"OwnerUserId":"464273","Title":null,"Body":"<p>Yes, 3d convolution was added at the beginning of March.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/2213#issuecomment-217081951\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/2213#issuecomment-217081951<\/a><\/p>\n\n<p>See commit: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/6a187ccddaebb741ea77fc3201c6e36625f0aadb\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/commit\/6a187ccddaebb741ea77fc3201c6e36625f0aadb<\/a><\/p>\n"}]}
{"QuestionId":36226872,"AnswerCount":2,"Tags":"<windows><powershell><docker><machine-learning><tensorflow>","CreationDate":"2016-03-25T19:42:42.890","AcceptedAnswerId":36227750.0,"OwnerUserId":6115382.0,"Title":"Having difficulty accessing the Jupyter notebook when trying to run TensorFlow through Docker","Body":"<p><strong>My problem:<\/strong><\/p>\n\n<p>Newbie to the ML field here. I am currently attempting to set up tensorflow through docker but am having some troubles getting to the jupyter notebook once tensorflow appears to be running. For future reference, I am using windows 10, powershell, python 2.7, and oracle virtualbox (v.5.0.16).<\/p>\n\n<p>I am pretty confident that docker is working properly because I went through their setup tutorial (\"Using docker from Powershell\") without any issues--I was able to run the hello world example there just fine. I was also able to successfully run an ubuntu terminal with the command <\/p>\n\n<pre><\/pre>\n\n<p>I then followed the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#test-the-tensorflow-installation\" rel=\"nofollow noreferrer\">tensorflow docker installation instructions<\/a> and everything appeared to be normal--I entered the docker run command for tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>and got this output:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I attempt to get to the jupyter notebook at localhost:8888, chrome says the site cannot be reached within a few seconds and that the server refused to connect.<\/p>\n\n<p><strong>Solutions I have tried:<\/strong><\/p>\n\n<ul>\n<li><p>Followed the advice in <a href=\"https:\/\/stackoverflow.com\/questions\/33636925\/how-do-i-start-tensorflow-docker-jupyter-notebook\">this stackoverflow post<\/a>. When I run the command they suggest to start a new tensorflow container, I get the error: <\/p>\n\n<pre><\/pre>\n\n<p>I also changed the port settings using virtualbox like they suggested, and this also did not seem to make a difference. <\/p><\/li>\n<li><p>Instead of going to localhost, I tried the IP addresses listed for the driver when I use the command <em>docker-machine ls<\/em>. No luck here either.<\/p><\/li>\n<li><p>My first attempt was using the quickstart terminal that comes with docker, and I got stuck in the same place. So now I have a \"default\", from when I set up docker using the quickstart terminal, and a \"my-default\", from when I set up docker using powershell. Stackexchange comments have said that the quickstart terminal doesn't work as well as just using powershell, so I have started using powershell instead. <\/p><\/li>\n<li><p>Tried changing which driver was active using <em>docker-machine env<\/em> and then running the tensorflow run command for that driver's IP address, and still had no luck. <\/p><\/li>\n<li><p>Tried skipping the jupyter notebook and just running everything from the command line through python as suggested in the next step of the tensorflow install (\"Run tensorflow from the command line\") and I could not even import tensorflow from there:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know if this is possible without running jupyter first but I thought it was worth a shot. This problem might be unrelated to the one this post is about, but if anyone has an idea as to why this is, that would be helpful too. <\/p><\/li>\n<\/ul>\n\n<p>I appreciate any and all help that is offered. Please let me know if anything I stated is vague, or if I can offer any more details on what I have tried. Thank you very much in advance. <\/p>\n","answers":[{"AnswerId":"36227750","CreationDate":"2016-03-25T20:45:27.643","ParentId":null,"OwnerUserId":"823732","Title":null,"Body":"<p>Try adding -p 8888:888 to the docker run command.<\/p>\n\n<pre><code>docker run -p 8888:8888 -it b.gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p>Got the solution <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/42\" rel=\"nofollow\">here<\/a>.<\/p>\n"},{"AnswerId":"37371340","CreationDate":"2016-05-22T06:40:33.100","ParentId":null,"OwnerUserId":"5597937","Title":null,"Body":"<p>For ubuntu 14.04 I had to use the below command to specify '\/bin\/bash' so that it would enter the interactive console.\n<p><code>docker run -it gcr.io\/tensorflow\/tensorflow \/bin\/bash<\/code><\/p>\n"}]}
{"QuestionId":36227354,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-03-25T20:15:56.663","AcceptedAnswerId":36227535.0,"OwnerUserId":3761534.0,"Title":"Saving tensorflow model after training is finished","Body":"<p>I have finished running a big model in tensorflow python. But I have not saved it inside the session. Now that the training is over, I want to save the variables. I am doing the following: <\/p>\n\n<pre><\/pre>\n\n<p>This returns : ValueError: No variables to save. According to their website what is missing is initialize_all_variables(). The documentation says little about what exactly that does. The word \"initialize\" scares me, I do not want to reset all my trained values. Any way to save my model without re-running it? <\/p>\n","answers":[{"AnswerId":"38518996","CreationDate":"2016-07-22T05:43:07.320","ParentId":null,"OwnerUserId":"6471140","Title":null,"Body":"<p>How about using skflow ? With skflow(now skflow is integrated in tensorflow) you can specify the parameter model_dir on your constructor and that automatically will  save your model while training(it will save checkpoints so if something goes wrong during training, you can restart from last checkpooint).<\/p>\n"},{"AnswerId":"38532375","CreationDate":"2016-07-22T17:38:19.450","ParentId":null,"OwnerUserId":"6536794","Title":null,"Body":"<p>I believe its because you are not initializing all of your variables in the saver. This should work <\/p>\n\n<pre><code>with tf.Session() as sess:\n      tf.initialize_all_variables().run()\n      saver = tf.train.Saver(tf.all_variables())\n     -------everything your session does -------------\n      checkpoint_path = os.path.join(save_dir, 'model.ckpt')\n      saver.save(sess, checkpoint_path, global_step = your_global_step)\n<\/code><\/pre>\n"},{"AnswerId":"36227535","CreationDate":"2016-03-25T20:28:54.883","ParentId":null,"OwnerUserId":"2773607","Title":null,"Body":"<p>It seems like from the tensorflow documentation, the \"session\" is the thing that holds the information from the trained model. So presumably somewhere you called <code>sess.run()<\/code> to train your model - what you want to do is call <code>sess.save()<\/code> using THAT session, not a new one you create with this saver object.<\/p>\n"}]}
{"QuestionId":36228723,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-03-25T21:59:22.363","AcceptedAnswerId":36246038.0,"OwnerUserId":468565.0,"Title":"Implementing seq2seq with sampled decoder outputs","Body":"<p>I'm trying to implement seq2seq with a feature to generate sampled outputs from the decoder i.e. at every step, rather than taking the argmax of the output logits from previous state,  it should sample from them according to the logit distribution and use that as input for the next step. <\/p>\n\n<p>After poking around I found the loop_function in seq2seq.py as a promising place to start.  It looks like I have to write a loop function that looks like this (modified from the one in the file that extracts argmax+embedding):<\/p>\n\n<pre><\/pre>\n\n<p>Then i use this loop function generator in the seq2seq_embedding_with_attention model. However, the op I need that samples from a tensor of floats does not exist in tensorflow, so do I need to write my own? How do I do that?<\/p>\n\n<ol start=\"2\">\n<li><p>In searching for guidance, I found that in tensorflow\/tensorflow\/python\/ops\/candidate_sampling_ops there is a reference to :<\/p>\n\n<pre><\/pre>\n\n<p>but I can't find this file. I'm guessing it's auto-generated from somewhere. where?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"43829077","CreationDate":"2017-05-07T07:49:47.910","ParentId":null,"OwnerUserId":"6885326","Title":null,"Body":"<p>I meet the same problem today, and my solution is:<br>\nReplace line\n<code>\nprev_symbol = math_ops.sample(prev)\n<\/code>\nwith \n<code>\nprev_symbol = squeeze(multinomial(prev, 1), axis=1)\n<\/code><\/p>\n\n<p>The function tf.multinomial() draws samples from a multinomial distribution. It took a 2-D Tensor \"logits\" with shape [batch_size, num_classes], and a 0-D scalar \"num_samples\" as input. And output a drawn samples of shape [batch_size, num_samples].<\/p>\n\n<p>Meanwhile, math_ops.sample() outputs samples of shape [batch_size], therefore we need tf.squeeze() to reduce dimension.<\/p>\n\n<p>This implementation is simpler.<\/p>\n"},{"AnswerId":"36246038","CreationDate":"2016-03-27T09:56:12.323","ParentId":null,"OwnerUserId":"1134366","Title":null,"Body":"<p>Currently, you can also do it as follows, with <a href=\"https:\/\/hips.seas.harvard.edu\/blog\/2013\/04\/06\/the-gumbel-max-trick-for-discrete-distributions\/\" rel=\"nofollow\">the gumbel max trick for directe distributions<\/a>: <\/p>\n\n<pre><code>def batch_gumbel_max_sample(a, max_gumbel_noise = 1.0):\n    matrix_U = -1.0*tf.log(-1.0*tf.log(tf.random_uniform(tf.shape(a),\n                            minval = 0.0, maxval = max_gumbel_noise)))\n    return tf.argmax(tf.sub(a, matrix_U), dimension = 1)\n<\/code><\/pre>\n\n<p>There is also a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/456\" rel=\"nofollow\">discussion<\/a> on Tensorflows issue tracker about this at the moment. I guess sooner or later a multinomial sample function will be added to Tensorflow. LeavesBreathe also posted a work around on that Github page, which isn't entirely correct in my opinion:<\/p>\n\n<pre><code>def batch_sample_with_temperature(a, temperature=1.0):\n'''this function is like sample_with_temperature except it can handle batch input a of [batch_size x logits] \n    this function takes logits input, and produces a specific number from the array. This is all done on the gpu\n    because this function uses tensorflow\n    As you increase the temperature, you will get more diversified output but with more errors (usually gramatical if you're \n        doing text)\nargs: \n    Logits -- this must be a 2d array [batch_size x logits]\n    Temperature -- how much variance you want in output\nreturns:\n    Selected number from distribution\n'''\n\n'''\nEquation can be found here: https:\/\/en.wikipedia.org\/wiki\/Softmax_function (under reinforcement learning)\n    Karpathy did it here as well: https:\/\/github.com\/karpathy\/char-rnn\/blob\/4297a9bf69726823d944ad971555e91204f12ca8\/sample.lua'''\n'''a is [batch_size x logits]'''\nwith tf.op_scope([a,temperature], \"batch_sample_with_temperature\"):\n\n    exponent_raised = tf.exp(tf.div(a, temperature)) #start by reduction of temperature, and get rid of negative numbers with exponent\n    matrix_X = tf.div(exponent_raised, tf.reduce_sum(exponent_raised, reduction_indices = 1)) #this will yield probabilities!\n    matrix_U = tf.random_uniform(tf.shape(a), minval = 0, maxval = 1)\n    final_number = tf.argmax(tf.sub(matrix_X, matrix_U), dimension = 1) #you want dimension = 1 because you are argmaxing across rows.\n\nreturn final_number\n<\/code><\/pre>\n"}]}
{"QuestionId":36230559,"AnswerCount":1,"Tags":"<python><tensorflow><optimization>","CreationDate":"2016-03-26T01:21:00.600","AcceptedAnswerId":36233277.0,"OwnerUserId":6024751.0,"Title":"Processing time gets longer and longer after each iteration (TensorFlow)","Body":"<p>I am training a CNN with TensorFlow for medical images application.<\/p>\n\n<p>As I don't have a lot of data, I am trying to apply random modifications to my training batch during the training loop to artificially increase my training dataset. I made the following function in a different script and call it on my training batch:<\/p>\n\n<pre><\/pre>\n\n<p>The code works well for applying modifications to my images. <\/p>\n\n<p><strong>The problem is :<\/strong><\/p>\n\n<p>After each iteration of my training loop (feedfoward + backpropagation), applying this same function to my next training batch steadily takes 5 seconds longer than the last time.<\/p>\n\n<p>It takes around 1 second to process and reaches over a minute of processing after a bit more than 10 iterations.<\/p>\n\n<p>What causes this slowing?\nHow can I prevent it?<\/p>\n\n<p>(I suspect something with  but I'm not quite sure. Am opening a new session each time? TensorFlow isn't supposed to close automatically the session as I use in a \"with tf.Session()\" block?)<\/p>\n","answers":[{"AnswerId":"36233277","CreationDate":"2016-03-26T08:38:47.033","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>You call that code in each iteration, so each iteration you add these operations to the graph. You don't want to do that. You want to build the graph at the start and in the training loop only execute it. Also, why do you need to convert to ndimage again afterwards, instead of putting things into your TF graph once and just use tensors all the way through? <\/p>\n"}]}
{"QuestionId":36230853,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><tensorflow>","CreationDate":"2016-03-26T02:11:16.823","AcceptedAnswerId":36240769.0,"OwnerUserId":3570004.0,"Title":"Tensorflow: Setting an array element with a sequence","Body":"<p>I'm trying to train a CNN using my own image dataset, but when passing the batch data and label to the  I get the error  from what I read <a href=\"https:\/\/stackoverflow.com\/questions\/34156639\/tensorflow-python-valueerror-setting-an-array-element-with-a-sequence-in-t\">here<\/a>, this is a dimension issue, and probably coming from my  Tensor, but I couldn't figure out how to make it a one-hot Tensor (what my graph expects).<\/p>\n\n<p>I uploaded the full code as a gist here: <a href=\"https:\/\/gist.github.com\/guivn\/f7f753547f77a3b12992\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/guivn\/f7f753547f77a3b12992<\/a><\/p>\n","answers":[{"AnswerId":"36240769","CreationDate":"2016-03-26T20:55:17.000","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> You can't feed a <code>tf.Tensor<\/code> object (viz. <code>batch_data<\/code> and <code>batch_labels<\/code> in your <a href=\"https:\/\/gist.github.com\/guivn\/f7f753547f77a3b12992\" rel=\"nofollow\">gist<\/a>) as the value for another tensor. (I believe the error message should be clearer about this in more recent versions of TensorFlow.)<\/p>\n\n<p>Unfortunately you can't currently use the feed\/<code>tf.placeholder()<\/code> mechanism to pass the result of one TensorFlow graph to another. We are investigating ways to make this easier, since it is a common confusion and feature request. For your exact program, it should be easy to solve this however. Simply move the <a href=\"https:\/\/gist.github.com\/guivn\/f7f753547f77a3b12992#file-cnn-py-L112\" rel=\"nofollow\">lines<\/a> that create the input and replace the placeholders with them. Your program will then look something like:<\/p>\n\n<pre><code>with graph.as_default():\n\n  # Input data.\n  filename_and_label_tensor = tf.train.string_input_producer(['train.txt'], shuffle=True)\n  data, label = parse_csv(filename_and_label_tensor)\n  tf_train_dataset, tf_train_labels = tf.train.batch([data, label], batch_size, num_threads=4)\n\n  # Rest of the model construction goes here....\n<\/code><\/pre>\n\n<p>Typically, if you want to pass another dataset through the same model&mdash;e.g. for evaluation&mdash;it's easiest to make another copy of the graph (perhaps sharing the same <code>tf.Variable<\/code> objects).<\/p>\n"}]}
{"QuestionId":36232819,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2016-03-26T07:33:15.610","AcceptedAnswerId":null,"OwnerUserId":375267.0,"Title":"Difficulty adapting Keras example to customer data","Body":"<p>I'm trying to run the following network with Keras, powered by a TensorFlow backend. It's an adaptation of the \"VGG-like\" convnet from <a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow\">http:\/\/keras.io\/examples\/<\/a>: <\/p>\n\n<pre><\/pre>\n\n<p>This produces the error: <\/p>\n\n<pre><\/pre>\n\n<p>I've checked out the following posts but haven't been able to solve this one yet. Any help would be appreciated, along with an explanation of what's going wrong.<\/p>\n\n<ul>\nhttps:\/\/groups.google.com\/forum\/#!topic\/keras-users\/Vb7MhSqne0Y\nhttp:\/\/stackoverflow.com\/questions\/33974231\/tensorflow-error-using-my-own-data\n<\/ul>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Posted this question to the Keras issue board (<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2092\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/issues\/2092<\/a>). There are <a href=\"https:\/\/gist.github.com\/aaronpolhamus\/39c7a71151b8560d02dd\" rel=\"nofollow\">Gist<\/a> and <a href=\"https:\/\/github.com\/fchollet\/keras\/files\/190676\/keras_ex_data.zip\" rel=\"nofollow\">sample data<\/a> links that will allow you to re-create the problem. <\/p>\n","answers":[{"AnswerId":"36242933","CreationDate":"2016-03-27T01:35:10.553","ParentId":null,"OwnerUserId":"375267","Title":null,"Body":"<p>Solved it. In the script, the following line is needed to \"reshape\" the input array:<\/p>\n\n<p><code>X_train = X_train.reshape(X_train.shape[0], 1, 128, 256)<\/code><\/p>\n\n<p>Really, all we're doing here is adding a somewhat redundant channel dimension to make the shape of the array <code>(8144, 1, 128, 256)<\/code> instead of <code>(8144, 128, 256)<\/code>. If we were using an RGB array this wouldn't be redundant at all, since it would be <code>(8144, 3, 128, 256)<\/code>. Bottom line: my input array was missing the channel dimension, which I thought I could get away with omitting for greyscales. Turns out you still need to explicitly define the shape. <\/p>\n\n<p>Great package. Once fixed the code should execute as-is.<\/p>\n"}]}
{"QuestionId":36232824,"AnswerCount":1,"Tags":"<shape><theano><mismatch><lasagne>","CreationDate":"2016-03-26T07:34:10.220","AcceptedAnswerId":null,"OwnerUserId":3141047.0,"Title":"Bad argument to theano function, Wrong number of dimensions","Body":"<p>I am predicting 20 numbers for a regression tasks from a VGG, using lasagne and theano. For an example script I wrote, number of images are 100. I think I am doing something stupid, but am stuck. <\/p>\n\n<p>Looking online, for people who use nolearn, it can be fixed by specifying , but I am just using lasagne<\/p>\n\n<p>So:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the exact error message<\/p>\n\n<pre><\/pre>\n\n<p>Here's the model<\/p>\n\n<pre><\/pre>\n\n<p>Generators:   <\/p>\n\n<pre><\/pre>\n\n<p>Relevant training snippet<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36263356","CreationDate":"2016-03-28T13:27:04.637","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>Your prediction output has shape (batchsize, 20) but your y_sym variable is an ivector type, so it is a vector of length batchsize, presumably. I'm not sure if this is what is causing the error, but I don't think you can compute a squared error term for these two quantities. One is a matrix, the other is a vector, their shapes do not seem to align? <\/p>\n\n<p>What are your regression targets? If you are predicting 20 numbers for each datapoint, your y_sym should probably be a matrix, then you can compute the squared error term. <\/p>\n\n<p>Another possibility is to change your last layer to have nonlinearity be sigmoid. This way you can interpret your convolutional neural network as producing multi-label probabilities. You can then synthetically produce multi-label probability target variables for your regression. One example is, say a datapoint x has multi-labels 0, 1, 10. You could create a 20 length vector where 0, 1, 10 are like 1-a for some small a, and the other entries are small, positive numbers. <\/p>\n\n<p>You could also switch your objective function to be binary cross entropy. In this case we aren't performing regression. However, we still have that our network outputs a class score <em>matrix<\/em>, not a <em>vector<\/em>. This is typically the loss function used when detecting the presence of K different objects in an image (think cat, dog, human, car, bike, etc), e.g. multi-label classification. If you want to try out this route, we will change our last layer as follows: <\/p>\n\n<pre><code>net['fc8'] = DenseLayer(net['drop7'], num_units=20, nonlinearity=sigmoid)\n<\/code><\/pre>\n\n<p>Now I will interpret my network as outputting probabilities; each probability will represent a confidence score as to whether or not our network believes that object class is in the image. <\/p>\n\n<p>We will have the following loss function now: <\/p>\n\n<pre><code>X_sym = T.tensor4()\ny_sym = T.imatrix() \noutput_layer = net['fc8']\nprobabilities = lasagne.layers.get_output(output_layer, X_sym)\nloss = lasagne.objectives.binary_crossentropy(probabilities, y_sym)\nloss = loss.mean()\n... the rest of the update stuff...\ntrain_fn = theano.function([X_sym, y_sym], loss, updates=updates)\n<\/code><\/pre>\n\n<p>The key difference here is that our target variable y_sym is now an imatrix. This must be a {0,1} valued (batchsize,K) sized matrix, where 0 represents the object not being present in the image and 1 represents the object being present in the image. <\/p>\n\n<p>And to compute the accuracy for multi-label prediction one typically uses the F1-score. <a href=\"https:\/\/stackoverflow.com\/questions\/32572212\/how-to-calculate-f1-micro-score-using-lasagne\">Here<\/a> is a reference to how to compute the F1-score using scikit-learn. Our validation function would be different in this case, it may look something like this:<\/p>\n\n<pre><code> from sklearn.metrics import f1_score\n probabilities_fn = theano.function([X_sym], probabilities)\n confidence_threshold = .5 \n predictions = np.where(probabilities_fn(X_sym) &gt; confidence_threshold, 1, 0)  \n f1_score(y_true, predictions, average='micro')\n<\/code><\/pre>\n\n<p>The above code snippet with return the probabilities\/confidence scores from our network, and any probability over our confidence_threshold parameter, here I choose .5, we will interpret that as the class label being present. <\/p>\n"}]}
{"QuestionId":36234086,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2016-03-26T10:13:31.250","AcceptedAnswerId":null,"OwnerUserId":1079110.0,"Title":"How to use Tensorflow's LSTM cell as part of a larger graph?","Body":"<p>The built-in LSTM cells give you a callable. However, I'd like to add static layers before and after the LSTM layers and train everything through backprop. Is that possible?<\/p>\n","answers":[{"AnswerId":"36241501","CreationDate":"2016-03-26T22:12:16.137","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The cell <code>callable<\/code> is a function that takes an input <code>tf.Tensor<\/code> and the current state as a <code>tf.Tensor<\/code>, and returns an output <code>tf.Tensor<\/code> and the new state as a <code>tf.Tensor<\/code>. There is nothing special about the input\/output and state tensors: these can be computed from and\/or used as the inputs to other TensorFlow operations.<\/p>\n\n<p>For an example, look at the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9dde1657e95f64be2c045aba140178b5220043de\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L120\" rel=\"nofollow\"><code>ptb_word_lm.py<\/code><\/a> model:<\/p>\n\n<pre><code>for time_step in range(num_steps):\n  # ...\n  (cell_output, state) = cell(inputs[:, time_step, :], state)\n  outputs.append(cell_output)\n<\/code><\/pre>\n\n<p>Here, <code>inputs<\/code> is a 3-D tensor from which a slice is taken to get the input for one cell of the LSTM, and <code>state<\/code> is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9dde1657e95f64be2c045aba140178b5220043de\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L98\" rel=\"nofollow\">initially<\/a> computed using <code>cell.zero_state()<\/code>, then each subsequent iteration uses the state from the previous iteration.<\/p>\n\n<p>The <code>inputs<\/code> tensor is the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9dde1657e95f64be2c045aba140178b5220043de\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L102\" rel=\"nofollow\">result<\/a> of a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#embedding_lookup\" rel=\"nofollow\"><code>tf.nn.embedding_lookup()<\/code><\/a> operation; the <code>outputs<\/code> list is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9dde1657e95f64be2c045aba140178b5220043de\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L125\" rel=\"nofollow\">later concatenated<\/a> and used as the input to a loss calculation. TensorFlow backprops from the loss through the RNN and the embedding lookup back to the model variables.<\/p>\n"}]}
{"QuestionId":36234946,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2016-03-26T11:52:08.060","AcceptedAnswerId":null,"OwnerUserId":6117372.0,"Title":"Modifying Caffe to accept 16 bit data inside lmdb","Body":"<p>I'm trying to make some modifications on Caffe in order to accept my 16 bit data.<br>\nI succeded into creating a lmdb dataset filled with 16 bit unsigned, unencoded 256x256 image instead of usual caffe 8 bit unsigned data, saved as \"string\" as would be the usual 8 bit lmdb that can be created with  or DIGITS utilities.<br>\nI've changed the  functions  and  in order to create this lmdb at 16bit \"string\" data.\nNow, if I use this lmdb on caffe (just 4 classes), the networks runs but doesn't converge. I strongly suspect that is not reading my data properly.\nProblem is, the  functions  and  doesn't seems to make any distinction between internal data contents, and I cannot find the code where I should modify for dealing with 16bits.<br>\nCould anyone give me an hint to where to work on? <\/p>\n\n<p>Edit: \nMessing around in the code, I think that one of the possibility should be to create a new data layer or a new image data layer if I would like to work directly on the png without going throught lmdb. But trying to modify that C++ code is not a trivial task for me, esp. I cannot easily follow the data flow inside the code. I see that new layer can be written in python. Do you  think that a new input data layer could work nicely or would slow down the cnn performance? <\/p>\n","answers":[{"AnswerId":"36244710","CreationDate":"2016-03-27T06:57:02.473","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I don't know much about converting and adapting caffe\/lmdb interface, but it seems like a very risky path to take if you are not 100% certain with what you are doing.<br>\nFor instance, you changed the io functions in the python interface, but I don't think caffe is using this interface when running from command line (e.g., <code>$CAFFE_ROOT\/build\/tools\/caffe train ...<\/code>). Have you looked into the cpp io functions in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/util\/io.cpp\" rel=\"nofollow noreferrer\"><code>io.cpp<\/code><\/a> file?<\/p>\n\n<p>I would strongly suggest an alternative path: use hdf5 binary inputs instead of lmdb.<br>\nYou can convert the 16bit images you have to float32 and store them in hdf5 files, then input them to caffe via <code>\"HDF5Data\"<\/code> layer.<\/p>\n\n<p>Read more about hdf5 and caffe <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">in this thread<\/a>.<\/p>\n"}]}
{"QuestionId":36237427,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-26T15:58:43.160","AcceptedAnswerId":null,"OwnerUserId":1259390.0,"Title":"Understanding variable_scope and name_scope in tensorflow and variable sharing","Body":"<p>I'd like to share variables between two subgraphs. More precisely, I'd like to do the fowolling operation : given 4 tensors , , ,  and a weight variable , compute , , ,  but in different subgraph. The code, I have is as follow :<\/p>\n\n<pre><\/pre>\n\n<p>This example seems to run but I'm not sure whether the variable  is reused especially inside the  When I add  I got an error saying that there's no variable to share.\nWhen I visualize the graph in tensorboard, I do have several  inside the  subgraph.<\/p>\n","answers":[{"AnswerId":"39426901","CreationDate":"2016-09-10T14:12:26.030","ParentId":null,"OwnerUserId":"38626","Title":null,"Body":"<p>The following code does what you want:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef forward(inputs):\n    init = tf.random_normal_initializer()\n    w = tf.get_variable(\"weights\", shape=(3,2), initializer=init)\n    return tf.matmul(w, inputs)\n\nwith tf.name_scope(\"group_1\"):\n    a = tf.placeholder(tf.float32, shape=(2, 3), name=\"a\")\n    b = tf.placeholder(tf.float32, shape=(2, 3), name=\"b\")\n    c = tf.placeholder(tf.float32, shape=(2, 3), name=\"c\")\n    with tf.variable_scope(\"foo\", reuse=False):\n        aa = forward(a)\n    with tf.variable_scope(\"foo\", reuse=True):\n        bb = forward(b)\n        cc = forward(c)\n\nwith tf.name_scope(\"group_2\"):\n    d = tf.placeholder(tf.float32, shape=(2, 3), name=\"d\")\n    with tf.variable_scope(\"foo\", reuse=True):\n        dd = forward(d)\n\ninit = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    print(bb.eval(feed_dict={b: np.array([[1,2,3],[4,5,6]])}))\n    for var in tf.all_variables():\n        print(var.name)\n        print(var.eval())\n<\/code><\/pre>\n\n<p>A few important things to understand:<\/p>\n\n<ul>\n<li><code>name_scope()<\/code>affects all ops <strong>except variables created with <code>get_variable()<\/code><\/strong>.<\/li>\n<li>To place a variable in a scope, you need to use a <code>variable_scope()<\/code>. For example, placeholders <code>a<\/code>, <code>b<\/code>, and <code>c<\/code> are actually named <code>\"group_1\/a\"<\/code>, <code>\"group_1\/b\"<\/code>, <code>\"group_1\/c\"<\/code> and <code>\"group_1\/d\"<\/code>, but the <code>weights<\/code> variable is named <code>\"foo\/weights\"<\/code>.\nSo <code>get_variable(\"weights\")<\/code> in the name scope <code>\"group_1\"<\/code> and variable scope <code>\"foo\"<\/code> actually looks for <code>\"foo\/weights\"<\/code>.<\/li>\n<\/ul>\n\n<p>The <code>all_variables()<\/code> function is useful if you are not sure what variables exist and how they are named.<\/p>\n"}]}
{"QuestionId":36237531,"AnswerCount":1,"Tags":"<machine-learning><neural-network><theano><deep-learning>","CreationDate":"2016-03-26T16:07:09.270","AcceptedAnswerId":36237582.0,"OwnerUserId":5257450.0,"Title":"How to understand the batch size in CNN?","Body":"<p>When using Theano with conv3d2d.conv3d, I found the input condition of signal is<\/p>\n\n<pre><\/pre>\n\n<p>I think I do understand what does the batch size mean, which is the update frequency, but why would the batch size here being a dimension? <\/p>\n","answers":[{"AnswerId":"36237582","CreationDate":"2016-03-26T16:10:45.913","ParentId":null,"OwnerUserId":"326402","Title":null,"Body":"<p>The <code>batch_size<\/code> is the number of examples you are going to use for this minibatch. For example, if your <code>batch_size<\/code> is 50, that means that you are training\/testing 50 examples at a time. Consider if you had a 2d matrix to contain your data. It would look like an excel spreadsheet where each row is a separate example and each column is a <code>feature<\/code> of that example. The same thing is happening here, just in a higher dimension.<\/p>\n"}]}
{"QuestionId":36239838,"AnswerCount":2,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2016-03-26T19:24:28.133","AcceptedAnswerId":null,"OwnerUserId":455684.0,"Title":"Simple Lasange Neural Network not working","Body":"<p>I'm using the Lasagne package to build a simple 3 layer neural network, and I'm testing it on a very simple dataset (just 4 examples).<\/p>\n\n<pre><\/pre>\n\n<p>However it fails to learn this, and results in the prediction:<\/p>\n\n<pre><\/pre>\n\n<p>Full code:<\/p>\n\n<pre><\/pre>\n\n<p>Training:<\/p>\n\n<pre><\/pre>\n\n<p>I'm guessing there might be an issue with the nonlinear functions used in the hidden layers, or with the learning method. <\/p>\n","answers":[{"AnswerId":"36244500","CreationDate":"2016-03-27T06:22:18.720","ParentId":null,"OwnerUserId":"6088739","Title":null,"Body":"<p>I feel like you can't really judge whether the model is correctly learning based on the above. <\/p>\n\n<ol>\n<li><p><strong>Number of training instances<\/strong>\nYou have 4 training instances. The neural network you constructed contains 3*4 + 4*4 + 4*2 = 36 weights which it has to learn. Not to mention you have 4 different types of outputs. The network is definitely underfitting, which may explain unexpected results.<\/p><\/li>\n<li><p><strong>How to test if a model is working<\/strong>\nIf I wanted to test whether a neural network is correctly learning, I would test on a working dataset (like MNIST) and ensure my model is learning with high probability. \nYou could also try comparing with another neural network library you've already written or with literature.\nIf I really wanted to go micro, I would use boosting with a linearly separable dataset. <\/p><\/li>\n<\/ol>\n\n<p>If your model still doesn't learn properly, I would be concerned then.<\/p>\n"},{"AnswerId":"36244137","CreationDate":"2016-03-27T05:25:00.583","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>this is my guess for the problem,<\/p>\n\n<p>First, I don't know why is there output like <code>[0,0]<\/code>? is that means that sample not categorize in all classes?<\/p>\n\n<p>Second, You are using Softmax in the last layer, that usually use for classification, are you build this network for classification? if you confuse about the output, the output is actually probability of each class So I think the output is correct:<\/p>\n\n<ul>\n<li><p>second sample prediction is <code>[0.98 0.02]<\/code> so it means the second sample is belong to first class, like your target <code>[1 0]<\/code><\/p><\/li>\n<li><p>third sample prediction is <code>[0.25 0.75]<\/code> so it means the third sample is belong to second class, like your target <code>[1 1]<\/code> (regardless your first class value, it is classification, so it'll be count as correct classification by system)<\/p><\/li>\n<li><p>fourth sample prediction is <code>[0.25 0.75]<\/code> so it means the fourth sample is belong to second class, like your target <code>[0 1]<\/code><\/p><\/li>\n<li><p>first sample prediction is <code>[0.5 0.5]<\/code> this one seems a bit confusing to me, so I guess Lasagne will predict the first sample which have the same probability in each class as not a member of any classes<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":36241731,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2016-03-26T22:40:36.293","AcceptedAnswerId":36287841.0,"OwnerUserId":5606352.0,"Title":"Wrong number of dimensions: expected 3, got 2 with shape (119, 80)","Body":"<p>I'm new to Keras and having some trouble with shapes, specially when it comes to RNNs and LSTMs.<\/p>\n\n<p><strong>I'm running this code:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The variable predictor_train is a numpy array with 119 inner arrays, each one having 80 different items.<\/p>\n\n<p><strong>I'm having this error:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>So far what I found out is that an RNN receives 3D tensor with shape of (batch_size, timesteps, dimension) and when you set input_shape the batch_size is usually omitted, and you should just provide a tuple of (timesteps, dimension). But in what part of the code should that be changed (if possible, add your code changing suggestions)? <\/p>\n\n<hr>\n\n<h2>Extra info<\/h2>\n\n<blockquote>\n  <p><strong>About pred_frame<\/strong><\/p>\n<\/blockquote>\n\n<p>type: class 'pandas.core.frame.DataFrame'<\/p>\n\n<p>shape: (206,80)<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p><strong>About target_train<\/strong><\/p>\n<\/blockquote>\n\n<p>type: class 'pandas.core.series.Series'<\/p>\n\n<p>shape: (119,)<\/p>\n\n<p>dtype: float64<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p><strong>About predictor_train<\/strong><\/p>\n<\/blockquote>\n\n<p>type: 'numpy.ndarray'<\/p>\n\n<p>shape: (119,80)<\/p>\n\n<p>dtype: float64<\/p>\n\n<pre><\/pre>\n\n<h2>Edit<\/h2>\n\n<p>Thanks to @y300 apparently the 3d problem is surpassed. My shape now is (119,1,80).<\/p>\n\n<pre><\/pre>\n\n<p>However, I'm still getting a shaping problem in the model.fit line, as you can see below:<\/p>\n\n<pre><\/pre>\n\n<p>Why is it happening and how could I fix it?<\/p>\n","answers":[{"AnswerId":"36287841","CreationDate":"2016-03-29T14:54:33.363","ParentId":null,"OwnerUserId":"3427404","Title":null,"Body":"<p>You can check what your model looks like by doing<\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n\n<p>In this case, your should look something like this (actual values may differ):<\/p>\n\n<pre><code>--------------------------------------------------------------------------------\nInitial input shape: (None, None, 100)\n--------------------------------------------------------------------------------\nLayer (name)                  Output Shape                  Param #             \n--------------------------------------------------------------------------------\nSimpleRNN (simplernn)         (None, 1)                     102                 \n  --------------------------------------------------------------------------------\nTotal params: 102\n--------------------------------------------------------------------------------\n<\/code><\/pre>\n\n<p>As you can see, the input is a 3D tensor, not a 2D one. So you need to reshape your arrays to fit what keras is expecting. In particular, the input X_train should have dimensions (num_samples,1,input_dim). Here's a working example with some randomly generated x\/y data:<\/p>\n\n<pre><code>model.add(keras.layers.SimpleRNN(init='uniform',output_dim=1,input_dim=100))\nmodel.compile(loss=\"mse\", optimizer=\"sgd\")\nX_train = np.random.rand(300,1,100)\ny_train = np.random.rand(300)\nmodel.fit(X=X_train, y=y_train, batch_size=32,show_accuracy=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":36242286,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-26T23:58:51.717","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"bazel build tensorflow_serving\/... no such package 'inception'","Body":"<p>I'm trying to build tensorfolw_serving on OSX El Capitan and keep getting an error. I installed Bazel through homebrew.<\/p>\n\n<p>Here is the error:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried giving it the package path on my system and got a different error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43858818","CreationDate":"2017-05-08T23:39:43.077","ParentId":null,"OwnerUserId":"4156490","Title":null,"Body":"<p>use cd.. and run this script out of tensorflow folder<\/p>\n"},{"AnswerId":"36242487","CreationDate":"2016-03-27T00:26:26.553","ParentId":null,"OwnerUserId":"563762","Title":null,"Body":"<p>I installed Bazel through the download off their website and worked properly.<\/p>\n\n<p>Here's a link: <a href=\"http:\/\/bazel.io\/docs\/install.html\" rel=\"nofollow\">http:\/\/bazel.io\/docs\/install.html<\/a><\/p>\n\n<p>Seems that installing from homebrew causes some errors.<\/p>\n"}]}
{"QuestionId":36242678,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2016-03-27T00:53:27.593","AcceptedAnswerId":36243933.0,"OwnerUserId":4749396.0,"Title":"How to enable parallel processing on Theano ? No GPU","Body":"<p>I have been trying to find ways to enable parallel processing in theano while training a neural network , but I can't seem to find it. Right now when I train a network theano is only using a single core. <\/p>\n\n<p>Also I do not have access to a GPU , so if I could make theano use all the cores on the machine, then it will hopefull speed things up.<\/p>\n\n<p>Any tips on speeding up theano is very welcome !<\/p>\n","answers":[{"AnswerId":"36243933","CreationDate":"2016-03-27T04:46:08.547","ParentId":null,"OwnerUserId":"4749396","Title":null,"Body":"<p>This is what I have been able to figure out. <\/p>\n\n<p>Follow the instructions on this page \n<a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html<\/a>\nIt seems that I did not install BLAS properly. So I reinstalled everything according to the instructions on the website. <\/p>\n\n<p>Theano has config flags that have to be set.<\/p>\n\n<p>And follow the discussion here <a href=\"https:\/\/stackoverflow.com\/questions\/15639779\/why-does-multiprocessing-use-only-a-single-core-after-i-import-numpy\">Why does multiprocessing use only a single core after I import numpy?<\/a><\/p>\n\n<p>Using all this when I run the script <\/p>\n\n<pre><code>THEANO_FLAGS='openmp=True' OMP_NUM_THREADS=N OPENBLAS_MAIN_FREE=1 python &lt;script&gt;.py \n\/\/Where N is the number of cores \n<\/code><\/pre>\n\n<p>Theano uses all the cores on my machine.<\/p>\n"}]}
{"QuestionId":36242881,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2016-03-27T01:26:49.867","AcceptedAnswerId":null,"OwnerUserId":6119302.0,"Title":"Is it possible to train tensorflow on android?","Body":"<p>It seems no C++ API to train tensorflow graph and save to pb.So,any way to do that on Android platform? can I build tensorflow workspace with python API on Android devices?<\/p>\n","answers":[{"AnswerId":"45839117","CreationDate":"2017-08-23T11:59:03.410","ParentId":null,"OwnerUserId":"4620643","Title":null,"Body":"<p>Since you would need to install tensorflow on your android device and then run a python script on your android device, I highly doubt that this is possible.<br>\nAlso, since training is always relatively resource intensive, it does not really make sense to do this on mobile devices. For most problems, you'll even want to have decent GPU power for training which obviously can't be provided by an android device.<br>\nThe way to get a tensorflow network onto your android device would be to train it locally and then ship the trained network to your device, which then only has to perform the evaluation and no training. Go to <a href=\"https:\/\/www.tensorflow.org\/mobile\/\" rel=\"nofollow noreferrer\">Tensorflow Mobile<\/a> for more insights on that.<\/p>\n"}]}
{"QuestionId":36243479,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2016-03-27T03:21:51.953","AcceptedAnswerId":null,"OwnerUserId":1462620.0,"Title":"2DPooling in Keras doesn't pool last column","Body":"<p>When performing 2DPooling in keras over an input with odd dimension, say 8x24x128, the output is appropriately 4x12x128 if 2x2 pooling is used. When the input has an odd dimension, say 8x25x128, the output is 4x12x128. The pooling does NOT operate on the last column (25) of the input. I would like to zero pad the input to 8x26x128 with an extraneous zero column. Is this possible?<\/p>\n\n<p>In general terms: what is the proper etiquette for pooling over odd dimensional inputs?<\/p>\n","answers":[]}
{"QuestionId":36243536,"AnswerCount":3,"Tags":"<machine-learning><neural-network><theano><convolution>","CreationDate":"2016-03-27T03:31:21.823","AcceptedAnswerId":36243662.0,"OwnerUserId":5257450.0,"Title":"What is the number of filter in CNN?","Body":"<p>I am currently seeing the API of theano,<\/p>\n\n<pre><\/pre>\n\n<p>where the  is a tuple of , I am confusing about this because isn't that the number of filter decided by the stride while sliding the filter window on the image? How can I specify on filter number just like this? It would be reasonable to me if it is calculated by the parameter stride (if there is any).<\/p>\n\n<p>Also, I am confused with the term feature map as well, is it the neurons at each layer? How about the batch size? How are they correlated?<\/p>\n","answers":[{"AnswerId":"51372228","CreationDate":"2018-07-17T02:03:03.807","ParentId":null,"OwnerUserId":"10085419","Title":null,"Body":"<p>The number of filters is a hyper-parameter that can be tuned. The number of neurons in a convolutional layer equals to the size of the output of the layer. In the case of images, it's the size of the feature map.<\/p>\n"},{"AnswerId":"56163537","CreationDate":"2019-05-16T07:53:22.143","ParentId":null,"OwnerUserId":"8557216","Title":null,"Body":"<p>There is no correct answer as to what the best number of filters is. This strongly depends on the type and complexity of your (image) data. A suitable number of features is learnd from experience after working with similar types of datasets repeatedly over time. In general, the more features you want to capture (and are potentially available) in an image the higher the number of filters required in a CNN. <\/p>\n"},{"AnswerId":"36243662","CreationDate":"2016-03-27T03:56:22.430","ParentId":null,"OwnerUserId":"834518","Title":null,"Body":"<p>The number of filters is the number of neurons, since each neuron performs a different convolution on the input to the layer (more precisely, the neurons' input weights form convolution kernels).<\/p>\n\n<p>A feature map is the result of applying a filter (thus, you have as many feature maps as filters), and its size is a result of window\/kernel size of your filter and stride.<\/p>\n\n<p>The following image was the best I could find to explain the concept at high level:\n<a href=\"https:\/\/i.imgur.com\/TqTJCrz.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.imgur.com\/TqTJCrz.png\" alt=\"enter image description here\"><\/a>\nNote that 2 different convolutional filters are applied to the input image, resulting in 2 different feature maps (the output of the filters). Each pixel of each feature map is an output of the convolutional layer.<\/p>\n\n<p>For instance, if you have 28x28 input images and a convolutional layer with 20 7x7 filters and stride 1, you will get 20 22x22 feature maps at the output of this layer. Note that this is presented to the next layer as a volume with width = height = 22 and depth = num_channels = 20. You could use the same representation to train your CNN on RGB images such as the ones from the CIFAR10 dataset, which would be 32x32x3 volumes (convolution is applied only to the 2 spatial dimensions).<\/p>\n"}]}
{"QuestionId":36244973,"AnswerCount":1,"Tags":"<python><neural-network><theano><logistic-regression>","CreationDate":"2016-03-27T07:37:08.617","AcceptedAnswerId":36249365.0,"OwnerUserId":2147347.0,"Title":"Simple Binary Classification Using Theano Error","Body":"<p>I got an error when trying to create a simple binary classification for XOR case using Theano. It said dimension mismatch, but I can't find out what variable cause that.<\/p>\n\n<p>and the strange part, my program is works when I change the number of neuron in the last layer. When I change to use 2 neuron in the last layer, and change that layer to softmax layer, and also use the negative log likelihood (multiclass classification style), this program is works fine.<\/p>\n\n<p>This is my full code:<\/p>\n\n<pre><\/pre>\n\n<p>and the error:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you in advance for any help.<\/p>\n","answers":[{"AnswerId":"36249365","CreationDate":"2016-03-27T15:53:16.160","ParentId":null,"OwnerUserId":"6103829","Title":null,"Body":"<p>Your problem is with your y and inpy variables: what you are trying to do is to have y be the expected output of the network. Your network is given a dataset with 4 elements, each having 2 features, you thus have 4 rows in your input matrix, and 2 columns. You are thus expected to have 4 elements in your predicted output, that is 4 rows in your y or inpy matrix, but you are using a vector, which in theano is a row vector and thus has only one row. You need either to transpose your y vector when computing the cost, or to define your y variable as a matrix, and thus to have inpy as a (4,1) matrix instead of a (4,) vector (once again, vectors are row vectors in theano).<\/p>\n\n<p>Hope this helps,<\/p>\n\n<p>Best<\/p>\n"}]}
{"QuestionId":36245065,"AnswerCount":2,"Tags":"<computer-vision><tensorflow><image-recognition>","CreationDate":"2016-03-27T07:48:01.590","AcceptedAnswerId":null,"OwnerUserId":2454753.0,"Title":"Building a simple image search using TensorFlow","Body":"<p>I need to implement a simple image search in my app using TensorFlow.\nThe requirements are these:<\/p>\n\n<ol>\n<li>The dataset contains around a million images, all of the same size, each containing one unique object and <strong>only<\/strong> that object.<\/li>\n<li>The search parameter is an image taken with a phone camera of some object that is potentially in the dataset.<\/li>\n<\/ol>\n\n<p>I've managed to extract the image from the camera picture and straighten it to rectangular form and as a result, a reverse-search image indexer like TinEye was able to find a match.<\/p>\n\n<p>Now I want to reproduce that indexer by using TensorFlow to create a model based on my data-set (make each image's file name a unique index).<\/p>\n\n<p>Could anyone point me to tutorials\/code that would explain how to achieve such thing without diving too much into computer vision terminology?<\/p>\n\n<p>Much appreciated!<\/p>\n","answers":[{"AnswerId":"39863339","CreationDate":"2016-10-04T23:38:59.020","ParentId":null,"OwnerUserId":"5463996","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/wuzhenyusjtu\/VisualSearchServer\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wuzhenyusjtu\/VisualSearchServer<\/a>\nIt is a simple implementation of similar image searching using TensorFlow and InceptionV3 model. The code implements two methods, a server that handles image search, and a simple indexer that do Nearest Neighbor matching based on the pool3 features extracted. <a href=\"https:\/\/i.stack.imgur.com\/v1Rwz.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/v1Rwz.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"36409382","CreationDate":"2016-04-04T17:43:54.730","ParentId":null,"OwnerUserId":"1930168","Title":null,"Body":"<p>The <a href=\"https:\/\/en.wikipedia.org\/wiki\/TinEye\" rel=\"noreferrer\">Wikipedia article on TinEye<\/a> says that <a href=\"https:\/\/en.wikipedia.org\/wiki\/Perceptual_hashing\" rel=\"noreferrer\">Perceptual Hashing<\/a> will yield results similar to TinEye's. They reference <a href=\"http:\/\/www.hackerfactor.com\/blog\/index.php?\/archives\/432-Looks-Like-It.html\" rel=\"noreferrer\">this detailed description of the algorithm<\/a>. But <a href=\"http:\/\/blog.tineye.com\/tineye-is-magic\/\" rel=\"noreferrer\">TinEye refuses to comment.<\/a><\/p>\n\n<hr>\n\n<p>The biggest issue with the Perceptual Hashing approach is that while it's efficient for identifying the same image (subject to skews, contrast changes, etc.), it's not great at identifying a completely different image of the same object (e.g. the front of a car vs. the side of a car).<\/p>\n\n<p>TensorFlow has great support for deep neural nets which might give you better results. Here's a high level description of how you might use a deep neural net in TensorFlow to solve this problem:<\/p>\n\n<p>Start with a pre-trained NN (such as GoogLeNet) or train one yourself on a dataset like <a href=\"http:\/\/www.image-net.org\/\" rel=\"noreferrer\">ImageNet<\/a>. Now we're given a new picture we're trying to identify. Feed that into the NN. Look at the activations of a fairly deep layer in the NN. This vector of activations is like a 'fingerprint' for the image. Find the picture in your database with the closest fingerprint. If it's sufficiently close, it's probably the same object.<\/p>\n\n<p>The intuition behind this approach is that unlike Perceptual Hashing, the NN is building up a high-level representation of the image including identifying edges, shapes, and important colors. For example, the fingerprint of an apple might include information about its circular shape, red color, and even its small stem.<\/p>\n\n<hr>\n\n<p>You could also try something like <a href=\"http:\/\/link.springer.com\/article\/10.1007\/s00530-012-0265-1\" rel=\"noreferrer\">this 2012 paper on image retrieval<\/a> which uses a slew of hand-picked features such as SIFT, regional color moments and object contour fragments. This is probably a lot more work and it's not what TensorFlow is best at.<\/p>\n\n<hr>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>OP has provided an example pair of images from his application:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hZK6q.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hZK6q.jpg\" alt=\"Image in the database\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/K5paJ.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/K5paJ.jpg\" alt=\"Image from the user. Should match something in the database.\"><\/a><\/p>\n\n<p>Here are the results of using <a href=\"http:\/\/www.phash.org\/demo\/\" rel=\"noreferrer\">the demo on the pHash.org website<\/a> on that pair of similar images as well as on a pair of completely dissimilar images.<\/p>\n\n<p><strong>Comparing the two images provided by the OP:<\/strong><\/p>\n\n<p>RADISH (radial hash): <code>pHash determined your images are not similar with PCC = 0.518013<\/code><\/p>\n\n<p>DCT hash: <code>pHash determined your images are not similar with hamming distance = 32.000000.<\/code><\/p>\n\n<p>Marr\/Mexican hat wavelet: <code>pHash determined your images are not similar with normalized hamming distance = 0.480903.<\/code><\/p>\n\n<p><strong>Comparing one of his images with a random image from my machine:<\/strong><\/p>\n\n<p>RADISH (radial hash): <code>pHash determined your images are not similar with PCC = 0.690619.<\/code><\/p>\n\n<p>DCT hash: <code>pHash determined your images are not similar with hamming distance = 27.000000.<\/code><\/p>\n\n<p>Marr\/Mexican hat wavelet: <code>pHash determined your images are not similar with normalized hamming distance = 0.519097.<\/code><\/p>\n\n<p><strong>Conclusion<\/strong><\/p>\n\n<p>We'll have to test more images to really know. But so far pHash does not seem to be doing very well. With the default thresholds it doesn't consider the similar images to be similar. And for one algorithm, it actually considers a completely random image to be more similar.<\/p>\n"}]}
{"QuestionId":36245481,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-27T08:44:20.400","AcceptedAnswerId":36250569.0,"OwnerUserId":6120011.0,"Title":"TensorFlow: slow performance when getting gradients at inputs","Body":"<p>I'm building a simple multilayer perceptron with TensorFlow, and I also need to obtain the gradients (or error signal) of the loss at the neural network's inputs.<\/p>\n\n<p>Here's my code, which works:<\/p>\n\n<pre><\/pre>\n\n<p>(edited to include training loop)<\/p>\n\n<p>Without the last line (), this runs really fast on a CUDA machine. However,  reduces performance greatly by tenfold or more.<\/p>\n\n<p>I recall that the error signals at the nodes are computed as intermediate values in the backpropagation algorithm, and I have successfully done this using the Java library DeepLearning4j. I was also under the impression that this would be a slight modification to the computation graph already built by .<\/p>\n\n<p>How can this be made faster, or is there any other way to compute the gradients of the loss w.r.t. the inputs?<\/p>\n","answers":[{"AnswerId":"36250569","CreationDate":"2016-03-27T17:47:33.617","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.gradients()<\/code> function builds a new backpropagation graph each time it is called, so the reason for the slowdown is that TensorFlow has to parse a new graph on each iteration of the loop. (This can be surprisingly expensive: the current version of TensorFlow is optimized for executing the <em>same<\/em> graph a large number of times.)<\/p>\n\n<p>Fortunately the solution is easy: just compute the gradients once, outside the loop. You can restructure your code as follows:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(self.network, self.y))\noptimizer = tf.train.AdagradOptimizer(learning_rate=nn_learning_rate).minimize(cost)\ngrads_wrt_input_tensor = tf.gradients(cost, self.x)[0]\n# ...\nfor i in range(epochs):\n    # ...\n    for batch in batches:\n        # ...\n        _, grads_wrt_input = sess.run([optimizer, grads_wrt_input_tensor],\n                                      feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>Note that, for performance, I also combined the two <code>sess.run()<\/code> calls. This ensures that the forward propagation, and much of the backpropagation, will be reused.<\/p>\n\n<hr>\n\n<p>As an aside, one tip to find performance bugs like this is to call <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Graph.finalize\" rel=\"noreferrer\"><code>tf.get_default_graph().finalize()<\/code><\/a> before starting your training loop. This will raise an exception if you inadvertantly add any nodes to the graph, which makes it easier to trace the cause of these bugs.<\/p>\n"}]}
{"QuestionId":36247621,"AnswerCount":0,"Tags":"<tensorflow><bazel><tensorflow-serving>","CreationDate":"2016-03-27T13:00:44.167","AcceptedAnswerId":null,"OwnerUserId":3733814.0,"Title":"How to use tensorflow serving in external directory using bazel","Body":"<p>I have a repository, in the following hierarchy :<\/p>\n\n<pre><\/pre>\n\n<p>Then how to build target 'model_train.py' in 'face_ver', without changing any file in 'serving' directory, using bazel (bazel build)<\/p>\n\n<p>Please, post the code of sample BUILD file and WORKSPACE file(if required) to be placed in face_ver.<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":36248056,"AnswerCount":2,"Tags":"<python><theano><cudnn>","CreationDate":"2016-03-27T13:48:04.630","AcceptedAnswerId":36464973.0,"OwnerUserId":5743949.0,"Title":"how to setup cuDnn with theano on Windows 7 64 bit","Body":"<p>I have installed  framework and enabled CUDA on my machine, however when I \"import theano\" in my python console, I got the following message:<\/p>\n\n<pre><\/pre>\n\n<p>Now that \"CuDNN not available\",  I downloaded  from Nvidia website. I also updated 'path' in environment, and added 'optimizer_including=cudnn' in '.theanorc.txt' config file. <\/p>\n\n<p>Then, I tried again, but failed, with:<\/p>\n\n<pre><\/pre>\n\n<p>anyone can help me? Thanks.<\/p>\n","answers":[{"AnswerId":"39513996","CreationDate":"2016-09-15T14:42:00.797","ParentId":null,"OwnerUserId":"5041140","Title":null,"Body":"<p>In addition to all the stuffs you did I updated following content of .theanorc.txt in my home folder and it worked after that. <\/p>\n\n<pre><code>[lib] \n#cnmem=1.0 \ncudnn=1.0\n<\/code><\/pre>\n"},{"AnswerId":"36464973","CreationDate":"2016-04-07T01:10:01.137","ParentId":null,"OwnerUserId":"420774","Title":null,"Body":"<p>There should be a way to do it by setting only the Path environment variable but I could never get that to work. The only thing that worked for me was to manually copy the CuDNN files into the appropriate folders in your CUDA installation. <\/p>\n\n<p>For example, if your CUDA installation is in C:\\CUDA\\v7.0 and you extracted CuDNN to C:\\CuDNN you would copy as follows:<\/p>\n\n<ul>\n<li>The contents of C:\\CuDNN\\lib\\x64\\ would be copied to C:\\CUDA\\v7.0\\lib\\x64\\    <\/li>\n<li>The contents of C:\\CuDNN\\include\\ would be copied to C:\\CUDA\\v7.0\\include\\<\/li>\n<li>The contents of C:\\CuDNN\\bin\\ would be copied to C:\\CUDA\\v7.0\\bin\\<\/li>\n<\/ul>\n\n<p>After that it should work.<\/p>\n"}]}
{"QuestionId":36251165,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-27T18:43:28.947","AcceptedAnswerId":null,"OwnerUserId":278191.0,"Title":"Tensorboard not listing any event","Body":"<p>Running Tensorflow and Tensorboard on docker here.<\/p>\n\n<p>I was trying to write the simplest code to just demonstrate how tensorboard may work:<\/p>\n\n<pre><\/pre>\n\n<p>I then ran  but no event was listed there. Is there anything I should have written differently in the code maybe?<\/p>\n\n<p>Note that  does contain files like .<\/p>\n","answers":[{"AnswerId":"37381662","CreationDate":"2016-05-23T02:05:43.543","ParentId":null,"OwnerUserId":"4794308","Title":null,"Body":"<p>I am not sure whether it is your case.<\/p>\n\n<p><code>SummaryWriter<\/code> by default will store summaries in its buffer, it will flush every period of time(I guess 120 seconds? Not sure).<\/p>\n\n<p>So maybe you just did not wait until your the flush happens. Try to manually flush <code>SummaryWriter<\/code> or just <code>close()<\/code> it at the end of your program.<\/p>\n"}]}
{"QuestionId":36252540,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-27T20:54:03.880","AcceptedAnswerId":36267185.0,"OwnerUserId":2245776.0,"Title":"Tensorflow apply op to each element of a 2d tensor","Body":"<p>What I'm after is the ability to apply a tensorflow op to each element of a 2d tensor e.g.<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"36267185","CreationDate":"2016-03-28T17:12:45.320","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The next version of TensorFlow (0.8, which is currently available if you build from source or download a nightly build) includes <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#higher-order-operators\">higher-order operators<\/a> including <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#map_fn\"><code>tf.map_fn()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops.html#scan\"><code>tf.scan()<\/code><\/a> that allow you to apply a <em>function<\/em> made up of TensorFlow ops to subtensors of a larger tensor.<\/p>\n\n<p>The <code>tf.map_fn(fn, elems, ...)<\/code> function unpacks the <code>N<\/code>-dimensional input <code>elems<\/code> along the first dimension into multiple <code>N-1<\/code>-dimensional subtensors and applies <code>fn<\/code> to each subtensor. This seems to fit your use case perfectly:<\/p>\n\n<pre><code>input = tf.Variable([[1.0, 2.0], [3.0, 4.0]])\nfunction_to_map = lambda x: f(x)  # Where `f` instantiates myCustomOp.\nfinal_result = tf.map_fn(function_to_map, input)\n<\/code><\/pre>\n"}]}
{"QuestionId":36253557,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-27T22:40:54.507","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"Bazel Build Mnist Serving Interface Fail on OSX El Capitan","Body":"<p>I'm trying to build the TensorFlow serving basic tutorial on OSX El Capitan but run into this error. Any ideas?<\/p>\n\n<p>Here's the link to the tutorial: <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow\">https:\/\/tensorflow.github.io\/serving\/serving_basic<\/a><\/p>\n\n<pre><\/pre>\n\n<p>This change was also made to my WORKSPACE file (based on <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/6\" rel=\"nofollow\">this<\/a> discussion):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36254293","CreationDate":"2016-03-28T00:33:05.400","ParentId":null,"OwnerUserId":"563762","Title":null,"Body":"<p>Found a temporary fix to use docker. Seems like there are known issues with OSX right now.<\/p>\n\n<p>Docker install instructions can be found here: <a href=\"http:\/\/tensorflow.github.io\/serving\/docker\" rel=\"nofollow\">http:\/\/tensorflow.github.io\/serving\/docker<\/a><\/p>\n"}]}
{"QuestionId":36253964,"AnswerCount":2,"Tags":"<c++><deep-learning><caffe>","CreationDate":"2016-03-27T23:40:23.367","AcceptedAnswerId":null,"OwnerUserId":5457326.0,"Title":"Caffe C++ save network caffemodel file","Body":"<p>I have successfully built and trained an audioCaffe demo, but the demo doesn't save the network.<\/p>\n\n<p>I have found documentation for saving the network in Python in MatLab, but I can't find any documentation on C++.<\/p>\n\n<p>I would think there would be a similar function like  but I tried that and it didn't work.<\/p>\n\n<p>In the train function in  I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>But I got a <\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"36421526","CreationDate":"2016-04-05T08:53:51.947","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>You should look at <code>Snapshot()<\/code> and <code>SnapshotToBinaryProto()<\/code> - src\/caffe\/solver.cpp.<br>\nCaller code is in Solver::Step:<\/p>\n\n<pre><code>\/\/ Save a snapshot if needed.\n    if ((param_.snapshot()\n         &amp;&amp; iter_ % param_.snapshot() == 0\n         &amp;&amp; Caffe::root_solver()) ||\n         (request == SolverAction::SNAPSHOT)) {\n      Snapshot();\n    }\n<\/code><\/pre>\n"},{"AnswerId":"41567592","CreationDate":"2017-01-10T11:21:13.607","ParentId":null,"OwnerUserId":"3246146","Title":null,"Body":"<p>Please try this...<\/p>\n\n<pre><code>caffe::NetParameter net_param;\nnet_-&gt;ToProto(&amp;net_param);\ncaffe::WriteProtoToBinaryFile(net_param, caffe_model_path);\n<\/code><\/pre>\n"}]}
{"QuestionId":36254114,"AnswerCount":1,"Tags":"<tensorflow><eigen><convolution>","CreationDate":"2016-03-28T00:05:35.180","AcceptedAnswerId":null,"OwnerUserId":6122099.0,"Title":"Custom Spatial Convolution in TensorFlow","Body":"<p>I'd like to extend TensorFlow's  function by having it threshold the input (with a sigmoid or ReLU function) in each receptive field before multiplying by the filter. The actual convolution operation is implemented in Eigen, and Eigen's convolution functionality is in a function called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/third_party\/eigen3\/unsupported\/Eigen\/CXX11\/src\/NeuralNetworks\/SpatialConvolutions.h#L768\" rel=\"nofollow\">choose<\/a>. However, I have not been able to find the implementation for the  function.<\/p>\n\n<p>Where is the  function located in the TensorFlow github repository, and is there any simple way to extend Eigen's convolution function to do input processing?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36292541","CreationDate":"2016-03-29T18:41:30.783","ParentId":null,"OwnerUserId":"3309044","Title":null,"Body":"<p>Eigen supports both row-major and col-major memory layouts. The 'choose' template selects the appropriate convolution code path depending on the layout of the input data. The template itself is implemented in <a href=\"http:\/\/eigen.tuxfamily.org\/dox-devel\/unsupported\/TensorMeta_8h_source.html\" rel=\"nofollow\">http:\/\/eigen.tuxfamily.org\/dox-devel\/unsupported\/TensorMeta_8h_source.html<\/a><\/p>\n\n<p>The actual convolution is peformed as a matrix multiplication (called a contraction) between the convolution weights (aka the kernels) and the 2d patches extracted from the input image.<\/p>\n"}]}
{"QuestionId":36255918,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-28T04:19:45.237","AcceptedAnswerId":null,"OwnerUserId":6122578.0,"Title":"SummaryWriter not writing summaries to file","Body":"<p>I'm trying to use tensorflow's SummaryWriter, however it does not seem to write events, images, or histograms to file. However it does write the graph to file, (which I can then see in tensorboard), indicating at least, tensorboard and SummaryWriter know where my logdir is. <\/p>\n\n<p>Here is my (simplified) code, broken up by omitted code blocks:<\/p>\n\n<pre><\/pre>\n\n<p>.... More layers<\/p>\n\n<pre><\/pre>\n\n<p>.... Reccurrent Part<\/p>\n\n<pre><\/pre>\n\n<p>.... and then the training code:<\/p>\n\n<pre><\/pre>\n\n<p>Then when I load tensorboard, and look at the events tab, I see the following error:<\/p>\n\n<blockquote>\n  <p>No scalar summary tags were found.<\/p>\n  \n  <p>Maybe data hasn't loaded yet, or maybe you need to add some >tf.scalar_summary ops to your graph, and serialize them using the >tf.training.summary_io.SummaryWriter.<\/p>\n<\/blockquote>\n\n<p>I added the writer.flush() statement, because after searching on both stack exchange on github, that's a common suggestion. Issue not fixed. <\/p>\n\n<p>In my logs file, only the graph_def is written, no other files are written during training. <\/p>\n\n<p>I'm using tensorflow '0.7.1' on mac 0SX el-capitan. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"45742634","CreationDate":"2017-08-17T18:31:52.163","ParentId":null,"OwnerUserId":"222489","Title":null,"Body":"<p>I know this is an old post, but I was experiencing the same thing in a virtual environment running TensorFlow 1.1.0.  Running version 1.2.1 I don't seem to have this problem.  You can execute the following at the command line to determine which version of TensorFlow you're running:<\/p>\n\n<pre><code>python -c \"import tensorflow as tf; print(tf.__version__)\"\n<\/code><\/pre>\n\n<p>Hope it helps someone out there!<\/p>\n\n<p>Cheers,<\/p>\n\n<p>-Maashu<\/p>\n"}]}
{"QuestionId":36256520,"AnswerCount":1,"Tags":"<python><tensorflow><conda>","CreationDate":"2016-03-28T05:29:17.230","AcceptedAnswerId":null,"OwnerUserId":2646845.0,"Title":"Can't install tensorflow with conda package manager. Pip3 not found and virtualenv not compatible","Body":"<p>Trying to install tensorflow on OSX with python installed with anaconda : <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#virtualenv-installation\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#virtualenv-installation<\/a><\/p>\n\n<p>Get stuck on : <\/p>\n\n<p>sudo pip3 install --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp35-none-any.whl\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp35-none-any.whl<\/a><\/p>\n\n<p>and get error : <\/p>\n\n<p>tensorflow-0.7.1-cp35-none-any.whl is not a supported wheel on this platform.<\/p>\n\n<p>When I try virtualenv installation <\/p>\n\n<p>New python executable in \/Users\/xxx\/tensorflow\/bin\/python\nERROR: The executable \/Users\/xxx\/tensorflow\/bin\/python is not functioning\nERROR: It thinks sys.prefix is '\/Users\/xxx\/tmp' (should be '\/Users\/xxx\/tensorflow')\nERROR: virtualenv is not compatible with this system or executable<\/p>\n\n<p>What's the best way to install tensorflow with python installed with anaconda? <\/p>\n","answers":[{"AnswerId":"36607989","CreationDate":"2016-04-13T19:37:01.840","ParentId":null,"OwnerUserId":"722263","Title":null,"Body":"<p>I had the same issue trying to install with <code>python 3.4.3<\/code> and <code>pip 8.1<\/code>. <\/p>\n\n<pre><code> $ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\n  tensorflow-0.7.1-cp35-none-any.whl is not a supported wheel on this platform.\n<\/code><\/pre>\n\n<p>To fix I simply upgraded to <code>python 3.5.1<\/code> (I use pyenv to allow installation local of multiple versions of python).  <\/p>\n\n<p>Running with <code>python 3.5.1<\/code> I can now successfully execute: <\/p>\n\n<pre><code>$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp35-none-any.whl\n ...happy install message fun follows....\n<\/code><\/pre>\n\n<p>I have no clue as to why python 3.4.3 fails, but this version of tensorflow <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.8.0rc0-py3-none-any.whl\" rel=\"nofollow\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.8.0rc0-py3-none-any.whl<\/a> seems to work with python 3.4.3 fine.<\/p>\n"}]}
{"QuestionId":36259396,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-28T09:22:58.767","AcceptedAnswerId":36267491.0,"OwnerUserId":2260767.0,"Title":"Graph usage pattern for RNNs","Body":"<p>I am confused about the lifetime of tensors\/ops\/variables inside the default graph. <\/p>\n\n<p>In RNNs, the graph structure will change on each iteration due to the variable length of input data. Should separate graphs thus be created\/destroyed on each iteration of training to accommodate the changing structure of the graph, or is there a way to stick with the default graph?<\/p>\n\n<p>Also, if new graphs are constructed, how can they share variables?<\/p>\n","answers":[{"AnswerId":"36267491","CreationDate":"2016-03-28T17:30:34.473","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>For RNNs specifically, there are two options: <code>tf.nn.rnn<\/code>, and <code>tf.nn.dynamic_rnn<\/code>.  Neither one creates or destroys graphs temporarily.  <\/p>\n\n<p>The first function creates <code>T<\/code> subgraphs, where <code>T<\/code> is the length of the python list of inputs you provide (that is, inputs is a len <code>T<\/code> python list of shape <code>[batch, depth]<\/code> tensors).  <code>tf.nn.rnn<\/code> always expects a fixed number of time steps.  Note, you can control <em>which<\/em> subgraphs are run for a given step by passing the <code>sequence_length<\/code> parameter; the function then uses conditional evaluation (<code>tf.cond<\/code>) to determine what operations get run.<\/p>\n\n<p>In contrast, <code>dynamic_rnn<\/code> uses a special TensorFlow while loop and other trickery that introduces a limited type of loop in the graph structure.  In this case, there is exactly one subgraph for the \"time step\", and it gets run over and over until your input has been processed.  In this case your input is a <code>3D<\/code> tensor with dimensions <code>[batch, time, depth]<\/code> (it can be <code>[time, batch, depth]<\/code> if you set <code>time_major=True<\/code>); and the first two dimensions can vary from step to step.<\/p>\n"}]}
{"QuestionId":36260753,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-03-28T10:48:03.650","AcceptedAnswerId":36278337.0,"OwnerUserId":492372.0,"Title":"How does the syntax of this torch statement work?","Body":"<p>I am trying to understand how this particular torch statement works.<\/p>\n\n<p>I am referring to line number 115 at     <\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/1_data.lua#L115\" rel=\"nofollow\">https:\/\/github.com\/torch\/tutorials\/blob\/master\/2_supervised\/1_data.lua#L115<\/a><\/p>\n\n<p>The line reads <\/p>\n\n<pre><\/pre>\n\n<p>I was reading through slicing of data in torch and I understood that something like <\/p>\n\n<pre><\/pre>\n\n<p>means you are referring to \"all the rows and 1st column\" of tensor t4.<\/p>\n\n<p>However, in the statement I printed above, we have a {} inside another { } outside. What does that mean?<\/p>\n\n<p>I do understand that <\/p>\n\n<pre><\/pre>\n\n<p>refers to the batch size of the trainData which probably is the number of images.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"36278337","CreationDate":"2016-03-29T07:43:45.357","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>This acts as a narrow. Please refer to these <a href=\"https:\/\/github.com\/torch\/torch7\/issues\/501#issuecomment-171290546\" rel=\"nofollow\">detailed explanations<\/a>:<\/p>\n\n<blockquote>\n  <p>When you have double curly-braces, it returns a <code>narrow<\/code> of the tensor, and a <code>narrow<\/code>ed tensor is always a tensor (even if it only has one element). With double curly-braces, you can specify a range in which the tensor will be narrowed, which is not possible with single curly-braces. For example, you can do <code>ten[{{1,2},1}]<\/code>, which will be a 1D tensor of dimension <code>2<\/code>, and if your do <code>ten[{{1,2},{2}}]<\/code> it will return a 2D tensor of size <code>2x1<\/code>.<\/p>\n<\/blockquote>\n\n<p>e.g.:<\/p>\n\n<pre><code>th&gt; trsize = 10\nth&gt; trdata = torch.Tensor(trsize, 3, 32, 32)\nth&gt; subdata = trdata[{ {1, 5} }]\nth&gt; subdata:size()\n  5\n  3\n 32\n 32\n[torch.LongStorage of size 4]\n<\/code><\/pre>\n"}]}
{"QuestionId":36261468,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-03-28T11:33:05.287","AcceptedAnswerId":36278491.0,"OwnerUserId":4716167.0,"Title":"In lua\/torch, write tensor out without printing type","Body":"<p>I'm new to lua and just cannot find the answer to what seems a terribly simple question.<\/p>\n\n<p>I want to print out some tensors which correspond to word embeddings in a Word2Vec style. Each line should start with a word, followed by the tensor elements. I have the following code:<\/p>\n\n<pre><\/pre>\n\n<p>This is all fine and good, but I keep getting the tensor type and size printed out too:<\/p>\n\n<pre><\/pre>\n\n<p>How can I tell lua not to return the type? Like this:<\/p>\n\n<pre><\/pre>\n\n<p>Sorry if the answer is out there and I haven't searched for the right keywords. I'm still new to lua's concepts. <\/p>\n","answers":[{"AnswerId":"36278491","CreationDate":"2016-03-29T07:52:21.150","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>You can write your own dump function, e.g.:<\/p>\n\n<pre><code>local dump = function(vec)\n    vec = vec:view(vec:nElement())\n    local t = {}\n    for i=1,vec:nElement() do\n        t[#t+1] = string.format('%.4f', vec[i])\n    end\n    return table.concat(t, '  ')\nend\n<\/code><\/pre>\n\n<p>And use it instead of <code>tostring<\/code>.<\/p>\n"}]}
{"QuestionId":36264339,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-28T14:24:52.417","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"Tensorflow Serving Compile Error Using Docker on OSX","Body":"<p>I'm trying to install TensorFlow serving on OSX El Capitan using Docker but keep running into an error. Here is the tutorial I'm following:<\/p>\n\n<p><a href=\"https:\/\/tensorflow.github.io\/serving\/docker.html\" rel=\"nofollow\">https:\/\/tensorflow.github.io\/serving\/docker.html<\/a><\/p>\n\n<p>Here is the command causing the error:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the error I'm getting:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36268413","CreationDate":"2016-03-28T18:19:40.203","ParentId":null,"OwnerUserId":"563762","Title":null,"Body":"<p>Solved! Looks like the issues was with running out of memory in the VM.<\/p>\n\n<p>Here's how I fixed it:<\/p>\n\n<p>1) When creating the machine, make sure it has more memory (mine was only 1GB). Here is how you create a docker machine with 4GB:<\/p>\n\n<pre><code>docker-machine create -d virtualbox --virtualbox-memory 4096 default\n<\/code><\/pre>\n\n<p>2) When running the bazel command pass in a parameter limiting the amount of memory to use. Here I'm running the command using only 2GB:<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --verbose_failures --local_resources 2048,2.0,1.0 -j 1 \/\/tensorflow_serving\/example:mnist_export \n<\/code><\/pre>\n\n<p>Where the original command was:<\/p>\n\n<pre><code>bazel build \/\/tensorflow_serving\/example:mnist_export\n<\/code><\/pre>\n"}]}
{"QuestionId":36264514,"AnswerCount":1,"Tags":"<nvidia><tensorflow><theano><deep-learning>","CreationDate":"2016-03-28T14:34:37.063","AcceptedAnswerId":null,"OwnerUserId":3654968.0,"Title":"Can Tensorflow be installed alongside Theano?","Body":"<p>I'm trying to install tensorflow alongside Theano on a Nvidia Tesla K80. I'm working with Cuda 7.5 and following the instructions given <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md\" rel=\"nofollow\">here<\/a>\nTheano by itself works well, but as soon as I install tensorflow from source code following the instructions OR using pip install, nvidia-smi as well as Theano stops working.<\/p>\n\n<p>More specifically, nvidia-smi hangs indefinitely whereas Theano just refuses to run in GPU mode.<\/p>\n\n<p>I'm also using the latest version of cudnn v4. \nDoes Tensorflow have known issues with respect to causing nvidia-smi to hang and being non-compatible with Theano?<\/p>\n","answers":[{"AnswerId":"36720413","CreationDate":"2016-04-19T13:49:34.483","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>TensorFlow pick all the avaible GPU. So if you start it before Theano, Theano won't have any GPUs available by default. If you start Theano first, TensorFlow will segfault when it can't get the GPU Theano use.<\/p>\n\n<p>To work around that, make the NVIDIA driver only show to TensorFlow the device that you want him to see with this environment variable:<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES=0 python ...\n<\/code><\/pre>\n"}]}
{"QuestionId":36265507,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-03-28T15:31:00.677","AcceptedAnswerId":null,"OwnerUserId":1013125.0,"Title":"How to retrieve a layer in Caffe","Body":"<p>Given a  object, what's the best way to access a particular layer?<br>\nRight now I only figured how to iterate over them, which is not very useful:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36265993","CreationDate":"2016-03-28T15:59:15.367","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can get all the layers' names by<\/p>\n\n<pre><code>all_names = [n for n in net._layer_names]\n<\/code><\/pre>\n\n<p>of course if you want to inspect the values of the learned parameters, you can see how it's done in this <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"noreferrer\" title=\"net surgery\">net surgery<\/a> example.<\/p>\n\n<p>For instance, if you want to inspect the filters of <code>conv1<\/code> layer (assuming you have a layer with that name in your model) you can access<\/p>\n\n<pre><code>In [1]: net.params['conv1'][0].data.shape\nOut[1]: (64, 3, 3, 3)\n<\/code><\/pre>\n\n<p>And the bias term of this layer<\/p>\n\n<pre><code>In [2]: net.params['conv1'][1].data.shape\nOut[2]: (64,)\n<\/code><\/pre>\n\n<p>As you can see, this is the first layer of an image processing net, it has 64 filters acting on 3x3 patches of BGR (3 channels) input.<\/p>\n\n<p>If you already fed data through the net (using <code>net.forward<\/code>, or <code>net.backward<\/code>) you can inspect the responses of the different layers to the specific inputs fed through the net:<\/p>\n\n<pre><code>In [3]: net.blobs['conv1'].data.shape\nOut[3]: (1, 64, 198, 198)\n<\/code><\/pre>\n\n<p>The output shape of <code>conv1<\/code> layer is 198x198 pixels with 64 channels (there are 64 filters in this layer) and the batch size is 1.<br>\nIf you performed a backward pass as well, you can also inspect the gradients computed at this layer:<\/p>\n\n<pre><code>In [4]: net.blobs['conv1'].diff.shape\nOut[4]: (1, 64, 198, 198)\n<\/code><\/pre>\n"}]}
{"QuestionId":36267028,"AnswerCount":1,"Tags":"<lua><neural-network><deep-learning><torch>","CreationDate":"2016-03-28T17:03:29.830","AcceptedAnswerId":null,"OwnerUserId":3480490.0,"Title":"Torch: the implementation of embedding layer","Body":"<p>I'm a beginner of torch and I'm reading the code of LSTM language models. I'm confused on the implementation of embedding layer (<a href=\"https:\/\/github.com\/oxford-cs-ml-2015\/practical6\/blob\/master\/Embedding.lua\" rel=\"nofollow\">https:\/\/github.com\/oxford-cs-ml-2015\/practical6\/blob\/master\/Embedding.lua<\/a>).<\/p>\n\n<p>This is the  function in embedding layer:<\/p>\n\n<pre><\/pre>\n\n<p>My first question is why there is a  statement in this function rather than just like below:<\/p>\n\n<pre><\/pre>\n\n<p>I think  will always be true because  was initialized in  of .<\/p>\n\n<p>My second question is why  directly instead of calculates  using  and ? I mean .<\/p>\n\n<p>My last question is what the effect of this code in <a href=\"https:\/\/github.com\/oxford-cs-ml-2015\/practical6\/blob\/master\/Embedding.lua\" rel=\"nofollow\">Embedding.lua<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>What does it mean of \"we do not need to accumulate parameters when sharing\"?<\/p>\n","answers":[{"AnswerId":"36267884","CreationDate":"2016-03-28T17:51:57.717","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>(1) is so that if someone explicitly nils gradInput, then dont compute it\n(2) is because embedding layers don't actually have a gradInput, the inputs are not differentiable<\/p>\n"}]}
{"QuestionId":36271934,"AnswerCount":0,"Tags":"<build><dependencies><install><redhat><caffe>","CreationDate":"2016-03-28T21:58:25.693","AcceptedAnswerId":null,"OwnerUserId":4561745.0,"Title":"Installing Caffe dependencies locally","Body":"<p>I'm trying to build Caffe on a Red-Hat machine on which I'm logged in remotely and don't have root access. Is there any way to build its dependencies locally without asking the super-user to get them installed?<\/p>\n\n<p>The Caffe dependencies and installation instructions for RHEL are listed <a href=\"http:\/\/caffe.berkeleyvision.org\/install_yum.html\" rel=\"nofollow\">here<\/a>.<\/p>\n","answers":[]}
{"QuestionId":36272502,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-03-28T22:47:30.753","AcceptedAnswerId":null,"OwnerUserId":6080641.0,"Title":"What is the best way to top k pool elements instead of only the max one in Tensorflow?","Body":"<p>Now the max pooling function in tensorflow is<\/p>\n\n<pre>\ntf.nn.max_pool(value, ksize, strides, padding, name=None)\nReturns:\nA Tensor with type tf.float32. The max pooled output tensor.\n<\/pre>\n\n<p>I would like to have an extend version of max_pool, like<\/p>\n\n<pre>\ntf.nn.top_k_pool(value, ksize, strides, padding, k=1, name=None)\n\nPerforms the top k pooling on the input.\n\nArgs:\n\nvalue: A 4-D Tensor with shape [batch, height, width, channels] and type tf.float32.\nksize: A list of ints that has length >= 4. The size of the window for each dimension of the input tensor.\nstrides: A list of ints that has length >= 4. The stride of the sliding window for each dimension of the input tensor.\npadding: A string, either 'VALID' or 'SAME'. The padding algorithm.\nk: 0-D int32 Tensor. Number of top elements to look in each pool.\nname: Optional name for the operation.\nReturns:\n\nA Tensor with type tf.float32. The max pooled output tensor. There will be an additional dimension saving the top k values.\n<\/pre>\n\n<p>I know that I can expend the tensorflow operation following <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/adding_an_op\/index.html\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/adding_an_op\/index.html<\/a><\/p>\n\n<p>I would like to know if there is an easier way to achieve that.<\/p>\n","answers":[{"AnswerId":"36273376","CreationDate":"2016-03-29T00:18:14.290","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>Your best bet is probably the TopK op: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#top_k\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#top_k<\/a><\/p>\n"},{"AnswerId":"41137918","CreationDate":"2016-12-14T08:26:09.140","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>Here is a function to using <code>top_k<\/code>to take the max k activations of the channels. You can modify it to fit your purpose:<\/p>\n\n<pre><code>def make_sparse_layer(inp_x,k, batch_size=None):\n\n    in_shape = tf.shape(inp_x)\n    d = inp_x.get_shape().as_list()[-1]\n    matrix_in = tf.reshape(inp_x, [-1,d])\n    values, indices = tf.nn.top_k(matrix_in, k=k, sorted=False)\n    out = []\n    vals = tf.unpack(values, axis=0, num=batch_size)\n    inds = tf.unpack(indices, axis=0, num=batch_size)\n    for i, idx in enumerate(inds):\n        out.append(tf.sparse_tensor_to_dense(tf.SparseTensor(tf.reshape(tf.cast(idx,tf.int64),[-1,1]),vals[i], [d]), validate_indices=False ))\n    shaped_out = tf.reshape(tf.pack(out), in_shape)\n    return shaped_out \n<\/code><\/pre>\n"},{"AnswerId":"43131535","CreationDate":"2017-03-31T02:46:55.490","ParentId":null,"OwnerUserId":"1999041","Title":null,"Body":"<p>usd tf.reshape(), tf.matrix_transpose(), tf.nn.top_k(sorted=False), and the 'data_format' argument in tf.nn.conv2d(), see <a href=\"http:\/\/www.infocool.net\/kb\/OtherCloud\/201703\/318346.html\" rel=\"nofollow noreferrer\">http:\/\/www.infocool.net\/kb\/OtherCloud\/201703\/318346.html<\/a> for details<\/p>\n"}]}
{"QuestionId":36274235,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-03-29T02:14:56.677","AcceptedAnswerId":null,"OwnerUserId":1911639.0,"Title":"TensorFlow --use_gpu failed","Body":"<h3>Environment info<\/h3>\n\n<p>Operating System: Ubuntu 16.04 Beta 2 \nNVidia Geforce 980M, driver 361.28 (used for display at the same time) Cuda-7.5 <\/p>\n\n<p>If installed from sources, provide the commit hash: <\/p>\n\n<h3>Steps to reproduce<\/h3>\n\n<ol>\n<li><p><strong>bazel build -c opt --config=cuda --verbose_failures --genrule_strategy=standalone --spawn_strategy=standalone \/\/tensorflow\/cc:tutorials_example_trainer<\/strong> succeeded <\/p><\/li>\n<li><p><strong>bazel-bin\/tensorflow\/cc\/tutorials_example_trainer<\/strong> is OK, but <strong>bazel-bin\/tensorflow\/cc\/tutorials_example_trainer --use_gpu<\/strong> gives the following result: <\/p><\/li>\n<\/ol>\n\n<blockquote>\n  <p>$ bazel-bin\/tensorflow\/cc\/tutorials_example_trainer --use_gpu F\n  tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) F tensorflow\/cc\/tutorials\/example_trainer.cc:116] Check failed:\n  ::tensorflow::Status::OK() == (session->Run({{\"x\", x}}, {\"y:0\",\n  \"y_normalized:0\"}, {}, &amp;outputs)) (OK vs. Invalid argument: Cannot\n  assign a device to node 'Const\/_2': Could not satisfy explicit device\n  specification '\/gpu:0' because no devices matching that specification\n  are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0 [[Node: Const\/_2 =\n  Constdtype=DT_INT32, value=Tensor,\n  _device=\"\/gpu:0\"]]) Aborted (core dumped) ><\/p>\n<\/blockquote>\n\n<ol start=\"3\">\n<li>BTW, build succeed doesn't mean tensorflow has been installed. How to install then? <\/li>\n<\/ol>\n\n<h3>What have you tried?<\/h3>\n\n<ol>\n<li>Looks like GPU conflict? Does that mean, if I'm using the NVidia card for display, I'm NOT allowed to use it for GPU computation? \nCheers\nPei<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":36275146,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-29T03:57:11.223","AcceptedAnswerId":null,"OwnerUserId":3919542.0,"Title":"How to pass an image to a model for classification after the model has been created in Tensorflow","Body":"<p>I have created the model using Tensorflow <\/p>\n\n<pre><\/pre>\n\n<p>This is the accuracy that i am getting after all the iterations are done . I want to save this model so that i can pass a random image and display what my model outputs(classify) given the image . <\/p>\n\n<p>In caffe we used to create .caffemodel file and then used to pass an image in the command line and used to display the first 5 probable objects that the images is most likely to be , I want to perform a similar thing in Tensorflow but don't know as how to do so.<\/p>\n\n<p>Any help would be appreciated .<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>After the suggestion from mrry i tried looking for the tf.train.saver() obejct as how it work.<\/p>\n\n<p>This is what is written in the website for storing variables:<\/p>\n\n<pre><\/pre>\n\n<p><strong>For restoring :<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>How can i pass my image to the model now ? , It's written in the comment that do some work with the model but exactly how should i pass my image to the model .  <\/p>\n","answers":[{"AnswerId":"36275218","CreationDate":"2016-03-29T04:04:14.340","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can reuse your trained network for evaluation, using TensorFlow's <strong>feed<\/strong> mechanism. Typically you will have a tensor called <code>input<\/code> that contains a batch of images and serves as the input to the first layer of your model. Let's also assume that you have a tensor called <code>logits<\/code> that is the output of the final layer of your network (before it is passed to the loss function).<\/p>\n\n<p>You can get the top-five most probable classes by passing the <code>logits<\/code> to a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#softmax\" rel=\"nofollow\"><code>tf.nn.softmax()<\/code><\/a> op followed by a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#top_k\" rel=\"nofollow\"><code>tf.nn.top_k()<\/code><\/a> op. The resulting program will look like the following:<\/p>\n\n<pre><code>input = ...  # batch_size x height x width x channels tensor\n# [...rest of network defined as a function of `input`...]\nlogits = ...\n\npredictions = tf.nn.top_k(tf.nn.softmax(logits), k=5)\n\neval_image = ...  # 1 x height x width x channels NumPy array.\n                  # NOTE: you may need to reshape this to match `input`.\n\npredicted_classes = sess.run(predictions, feed_dict={input: eval_image})\nprint predicted_classes\n<\/code><\/pre>\n"}]}
{"QuestionId":36275828,"AnswerCount":1,"Tags":"<theano><deep-learning><recurrent-neural-network>","CreationDate":"2016-03-29T05:03:04.530","AcceptedAnswerId":36282279.0,"OwnerUserId":1303432.0,"Title":"the parameter of RNN in Theano tutorial","Body":"<pre><\/pre>\n\n<p>I am following this Theano tutorial about RNN.(<a href=\"http:\/\/deeplearning.net\/tutorial\/rnnslu.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/rnnslu.html<\/a>) But I have two questions about it.\nFirst. In this tutorial, recurrence function like this:<\/p>\n\n<p> <\/p>\n\n<p>I wounder why do not plus h0 in h_t ? (i.e. ) <\/p>\n\n<p>Second, why ? I know outputs_info is the Initialization result. So I think <\/p>\n","answers":[{"AnswerId":"36282279","CreationDate":"2016-03-29T10:53:40.693","ParentId":null,"OwnerUserId":"6128491","Title":null,"Body":"<pre><code>def recurrence(x_t, h_tm1):\n        h_t = T.nnet.sigmoid(T.dot(x_t, self.wx)\n                             + T.dot(h_tm1, self.wh) + self.bh)\n        s_t = T.nnet.softmax(T.dot(h_t, self.w) + self.b)\n        return [h_t, s_t]\n<\/code><\/pre>\n\n<p>So, first you ask why we don't use h0 in the recurrence function. Let's breakdown this part,<\/p>\n\n<pre><code>   h_t = T.nnet.sigmoid(T.dot(x_t, self.wx)+ T.dot(h_tm1, self.wh) + self.bh)\n<\/code><\/pre>\n\n<p>What we expect is 3 terms. <\/p>\n\n<ol>\n<li><p>The first term is the input layer multiplied by the weighting matrix <code>T.dot(x_t, self.wx)<\/code>.<\/p><\/li>\n<li><p>The second term is the hidden layer muliplied by another weighting matrix (this is what makes it recurrent) <code>T.dot(h_tm1, self.wh)<\/code>. Note that you must have a weighting matrix, you proposed to add <code>self.h0<\/code> as a bias basically.<\/p><\/li>\n<li><p>The third term is the bias of the hidden layer, <code>self.bh<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>Now, after every iteration we want to keep track of the hidden layer activations, contained in <code>self.h0<\/code>. However, <code>self.h0<\/code> is meant to contain the CURRENT activations and what we need is the previous activations.<\/p>\n\n<pre><code>[h, s], _ = theano.scan(fn=recurrence,\n                            sequences=x,\n                            outputs_info=[self.h0, None],\n                            n_steps=x.shape[0])\n<\/code><\/pre>\n\n<p>So, look at the scan function again. You are right that <code>outputs_info=[self.h0, None]<\/code> initializes the values, but the values are also linked to the outputs. There are two outputs from <code>recurrence()<\/code>, namely <code>[h_t, s_t]<\/code>.<\/p>\n\n<p>So what the outputs_info does as well is that after every iteration, the value of <code>self.h0<\/code> is overwritten with the value of <code>h_t<\/code> (the first returned value). The second element of outputs_info is <code>None<\/code>, because we do not save or initialize the value for <code>s_t<\/code> anywhere (the second argument of outputs_info is linked to the returned values of the recurrence function this way.)<\/p>\n\n<p>In the next iteration, the first argument of <code>outputs_info<\/code> is used again as input, such that <code>h_tm1<\/code> is the same value as <code>self.h0<\/code>. But, since we must have an argument for <code>h_tm<\/code> we must initialize this value. Since we don't need to initialize a second argument in <code>outputs_info<\/code>, we leave the second term as <code>None<\/code>.<\/p>\n\n<p>Granted, the <code>theano.scan()<\/code> function is very confusing at times and I'm new at it too. But, this is what I understood from doing this same tutorial.<\/p>\n"}]}
{"QuestionId":36277060,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><gradient-descent><softmax>","CreationDate":"2016-03-29T06:34:23.857","AcceptedAnswerId":36283863.0,"OwnerUserId":2362514.0,"Title":"Gradient calculation for softmax version of triplet loss","Body":"<p>I have been trying to implement the softmax version of the triplet loss in Caffe described in<br>\n<a href=\"http:\/\/arxiv.org\/pdf\/1412.6622v3.pdf\" rel=\"nofollow\">Hoffer and Ailon, <em>Deep Metric Learning Using Triplet Network<\/em>, ICLR 2015<\/a>.<\/p>\n\n<p>I have tried this but I am finding it hard to calculate the gradient as the L2 in exponent is not squared. <\/p>\n\n<p>Can someone please help me here? <\/p>\n","answers":[{"AnswerId":"36281942","CreationDate":"2016-03-29T10:38:56.567","ParentId":null,"OwnerUserId":"1648011","Title":null,"Body":"<p>This is a math question, but here it goes.   The first equation is what you're used to, and the second is what you do when it's not squared.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4WrLN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4WrLN.png\" alt=\"Derivation of norm\"><\/a><\/p>\n"},{"AnswerId":"36283863","CreationDate":"2016-03-29T12:08:44.830","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Implementing the L2 norm using existing layers of caffe can save you all the hustle.<\/p>\n\n<p>Here's one way to compute <code>||x1-x2||_2<\/code> in caffe for \"bottom\"s <code>x1<\/code> and <code>x2<\/code> (assuming <code>x1<\/code> and <code>x2<\/code> are <code>B<\/code>-by-<code>C<\/code> blobs, computing <code>B<\/code> norms for <code>C<\/code> dimensional diffs)<\/p>\n\n<pre><code>layer {\n  name: \"x1-x2\"\n  type: \"Eltwise\"\n  bottom: \"x1\"\n  bottom: \"x1\"\n  top: \"x1-x2\"\n  eltwise_param { \n    operation: SUM\n    coeff: 1 coeff: -1\n  }\n}\nlayer {\n  name: \"sqr_norm\"\n  type: \"Reduction\"\n  bottom: \"x1-x2\"\n  top: \"sqr_norm\"\n  reduction_param { operation: SUMSQ axis: 1 }\n}\nlayer {\n  name: \"sqrt\"\n  type: \"Power\"\n  bottom: \"sqr_norm\"\n  top: \"sqrt\"\n  power_param { power: 0.5 }\n}\n<\/code><\/pre>\n\n<p>For the triplet loss defined in the paper, you need to compute L2 norm for <code>x-x+<\/code> and for <code>x-x-<\/code>, concat these two blobs and feed the concat blob to a <code>\"Softmax\"<\/code> layer.<br>\nNo need for dirty gradient computations.<\/p>\n"}]}
{"QuestionId":36279144,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-29T08:29:01.297","AcceptedAnswerId":36288748.0,"OwnerUserId":1461746.0,"Title":"Tensorflow saved models get bigger when running cross-validation","Body":"<p>What's the correct way to run cross-validation on Tensorflow?\nBelow are my code snippets:<br><\/p>\n\n<pre><\/pre>\n\n<p>The saved model size for fold 0 is around 2M. But for fold 1 around 4M, fold 2 around 6M, and so on.<\/p>\n","answers":[{"AnswerId":"36288748","CreationDate":"2016-03-29T15:33:46.930","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>My guess is that the <code>TextCNN<\/code> constructor and <code>train()<\/code> method are adding nodes to the default graph (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#get_default_graph\" rel=\"nofollow\"><code>tf.get_default_graph()<\/code><\/a>) and the saved model includes all previous graphs, so it is \"accidentally quadratic\" and growing with each iteration of the <code>__main__<\/code> loop.<\/p>\n\n<p>The solution is easy, fortunately. Just rewrite your main loop as follows:<\/p>\n\n<pre><code>if __name__ == \"__main__\":\n  for i in range(CV_SIZE):\n    with tf.Graph().as_default():  # Performs training in a new, empty graph.\n      cnn = TextCNN(i)\n      cnn.train()\n<\/code><\/pre>\n\n<p>This will create a new, empty graph for each iteration of the loop. Therefore the saved model will not contain the nodes (and variables) from the previous iteration, and the model size should remain constant.<\/p>\n\n<p>Note that, if possible, you should try to reuse the same graph for all iterations. However, I realize that this might not be possible if the structure of the graph changes from one iteration to the next.<\/p>\n"}]}
{"QuestionId":36281129,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-29T10:02:08.447","AcceptedAnswerId":36286136.0,"OwnerUserId":3919542.0,"Title":"No variable to save error in Tensorflow","Body":"<p>I am trying to save the model and then reuse it for classifying my images but unfortunately i am getting errors in restoring the model that i have saved.<\/p>\n\n<p><strong>The code in which model has been created<\/strong> :<\/p>\n\n<pre><\/pre>\n\n<p>Everything works fine and the model is stored in the respective folder .<\/p>\n\n<p><strong>I have created one more python file where i have tried restoring the model but getting an error there<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>error that i am getting is :<\/p>\n\n<p><strong><em>No variables to save<\/em><\/strong> <\/p>\n\n<p>Any help would be appreciated<\/p>\n","answers":[{"AnswerId":"36282423","CreationDate":"2016-03-29T11:00:13.577","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You are creating a new session in <code>In[17]<\/code> which wipes your variables. Also, you don't need to use <code>with<\/code> blocks if you only have one default graph and one default session, you can instead do something like this<\/p>\n\n<pre><code>sess = tf.InteractiveSession()\nlayer1_weights = tf.Variable(tf.truncated_normal(\n  [patch_size, patch_size, num_channels, depth], stddev=0.1),name=\"layer1_weights\")\ntf.train.Saver().restore(sess, \"\/tmp\/model.ckpt\")\n<\/code><\/pre>\n"},{"AnswerId":"36286136","CreationDate":"2016-03-29T13:42:51.860","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The error here is quite subtle. In <code>In[8]<\/code> you create a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Graph\" rel=\"noreferrer\"><code>tf.Graph<\/code><\/a> called <code>graph<\/code> and set it as default for the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#Graph.as_default\" rel=\"noreferrer\"><code>with graph.as_default():<\/code><\/a> block. This means that all of the variables are created in <code>graph<\/code>, and if you print <code>graph.all_variables()<\/code> you should see a list of your variables.<\/p>\n\n<p><strong>However<\/strong>, you exit the <code>with<\/code> block before creating (i) the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\" rel=\"noreferrer\"><code>tf.Session<\/code><\/a>, and (ii) the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Saver\" rel=\"noreferrer\"><code>tf.train.Saver<\/code><\/a>. This means that the session and saver are created in a <strong>different graph<\/strong> (the global default <code>tf.Graph<\/code> that is used when you don't explicitly create one and set it as default), which doesn't contain any variables&mdash;or any nodes at all.<\/p>\n\n<p>There are at least two solutions:<\/p>\n\n<ol>\n<li><p>As <a href=\"https:\/\/stackoverflow.com\/a\/36282423\/3574081\">Yaroslav suggests<\/a>, you can write your program without using the <code>with graph.as_default():<\/code> block, which avoids the confusion with multiple graphs. However, this can lead to name collisions between different cells in your IPython notebook, which is awkward when using the <code>tf.train.Saver<\/code>, since it uses the <code>name<\/code> property of a <code>tf.Variable<\/code> as the key in the checkpoint file.<\/p><\/li>\n<li><p>You can create the saver <em>inside<\/em> the <code>with graph.as_default():<\/code> block, and create the <code>tf.Session<\/code> with an explicit graph, as follows:<\/p>\n\n<pre><code>with graph.as_default():\n    # [Variable and model creation goes here.]\n\n    saver = tf.train.Saver()  # Gets all variables in `graph`.\n\nwith tf.Session(graph=graph) as sess:\n    saver.restore(sess)\n    # Do some work with the model....\n<\/code><\/pre>\n\n<p>Alternatively, you can create the <code>tf.Session<\/code> <em>inside<\/em> the <code>with graph.as_default():<\/code> block, in which case it will use <code>graph<\/code> for all of its operations.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":36281861,"AnswerCount":2,"Tags":"<gpu><cpu><theano><theano-cuda>","CreationDate":"2016-03-29T10:35:24.330","AcceptedAnswerId":null,"OwnerUserId":6108696.0,"Title":"theano has different results on CPU and GPU","Body":"<p>I wrote a Theano machine learning program. But I got two absolutely different results between on CPU and on GPU.<\/p>\n\n<p>Below is the log.(only a tiny part of log)<\/p>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/U1JVx.jpg\" rel=\"nofollow\">result on GPU<\/a><\/p>\n\n<p><a href=\"http:\/\/i.stack.imgur.com\/t6mXI.jpg\" rel=\"nofollow\">result on CPU<\/a><\/p>\n\n<p>The loss function will quickly decrease and then converg to 0.2 on CPU.<\/p>\n\n<p>However, the loss function will increase and finally become NaN on GPU.<\/p>\n\n<p>What mistakes may be in my program? Or what should I take into attention? Thank you!<\/p>\n","answers":[{"AnswerId":"38171246","CreationDate":"2016-07-03T15:03:59.167","ParentId":null,"OwnerUserId":"6290016","Title":null,"Body":"<p>Check out this thread... \n<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/511\" rel=\"nofollow\">https:\/\/github.com\/fchollet\/keras\/issues\/511<\/a>\nBasically adding;<\/p>\n\n<pre><code>optimizer_excluding=cudnn \n<\/code><\/pre>\n\n<p>in the .theanorc solved it for me, but solution runs a slower.<\/p>\n"},{"AnswerId":"36282169","CreationDate":"2016-03-29T10:48:57.037","ParentId":null,"OwnerUserId":"380038","Title":null,"Body":"<p>Could it be that CPU is using <code>float64<\/code> (double precision) and the GPU is using <code>float32<\/code> (single precision)? Here you can look up the configuration flags: <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/config.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/library\/config.html<\/a><\/p>\n"}]}
{"QuestionId":36286594,"AnswerCount":2,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-03-29T14:00:24.153","AcceptedAnswerId":36289926.0,"OwnerUserId":4034105.0,"Title":"Predicting the next word using the LSTM ptb model tensorflow example","Body":"<p>I am trying to use the tensorflow <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html#lstm\" rel=\"nofollow noreferrer\">LSTM model<\/a> to make next word predictions.<\/p>\n\n<p>As described in this <a href=\"https:\/\/stackoverflow.com\/questions\/33773661\/predicting-next-word-using-the-language-model-tensorflow-example\">related question<\/a> (which has no accepted answer) the example contains pseudocode to extract next word probabilities:<\/p>\n\n<pre><\/pre>\n\n<p>I am confused about how to interpret the probabilities vector. I modified the  function of the  in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow noreferrer\">ptb_word_lm.py<\/a> to store the probabilities and logits:<\/p>\n\n<pre><\/pre>\n\n<p>Then printed some info about them in the  function:<\/p>\n\n<pre><\/pre>\n\n<p>This produces output like this:<\/p>\n\n<pre><\/pre>\n\n<p>I was expecting the  vector to be an array of probabilities, with one for each word in the vocabulary (eg with shape ), meaning that I could get the predicted word using  as suggested in the other question. <\/p>\n\n<p>However, the first dimension of the vector is actually equal to the number of steps in the unrolled LSTM (20 if the small config settings are used), which I'm not sure what to do with. To access to the predicted word, do I just need to use the last value (because it's the output of the final step)? Or is there something else that I'm missing?<\/p>\n\n<p>I tried to understand how the predictions are made and evaluated by looking at the implementation of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py\" rel=\"nofollow noreferrer\">seq2seq.sequence_loss_by_example<\/a>, which must perform this evaluation, but this ends up calling , which doesn't seem to be included in the github repo, so I'm not sure where else to look.<\/p>\n\n<p>I'm quite new to both tensorflow and LSTMs, so any help is appreciated!  <\/p>\n","answers":[{"AnswerId":"36289926","CreationDate":"2016-03-29T16:27:36.570","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>output<\/code> tensor contains the concatentation of the LSTM cell outputs for each timestep (see its definition <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/5d62000743498770191bbb76ad0141989d600163\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py#L125\" rel=\"noreferrer\">here<\/a>). Therefore you can find the prediction for the next word by taking <code>chosen_word[-1]<\/code> (or <code>chosen_word[sequence_length - 1]<\/code> if the sequence has been padded to match the unrolled LSTM).<\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#sparse_softmax_cross_entropy_with_logits\" rel=\"noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a> op is documented in the public API under a different name. For technical reasons, it calls a generated wrapper function that does not appear in the GitHub repository. The implementation of the op is in C++, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/sparse_xent_op.h\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"36360943","CreationDate":"2016-04-01T16:07:35.277","ParentId":null,"OwnerUserId":"6145309","Title":null,"Body":"<p>I am implementing seq2seq model too.<\/p>\n\n<p>So lets me try to explain with my understanding:<\/p>\n\n<p>The <strong>outputs<\/strong> of your LSTM model is a list (with length <strong>num_steps<\/strong>) of 2D tensor of size [<strong>batch_size<\/strong>, <strong>size<\/strong>].  <\/p>\n\n<p>The code line:  <\/p>\n\n<p><code>output = tf.reshape(tf.concat(1, outputs), [-1, size])<\/code><\/p>\n\n<p>will produce a new <strong>output<\/strong> which is a 2D tensor of size [<strong>batch_size<\/strong> x <strong>num_steps<\/strong>, <strong>size<\/strong>].  <\/p>\n\n<p>For your case, batch_size = 1 and num_steps = 20 --> output shape is [<strong>20<\/strong>, <strong>size<\/strong>].  <\/p>\n\n<p>Code line:  <\/p>\n\n<p><code>logits = tf.nn.xw_plus_b(output, tf.get_variable(\"softmax_w\", [size, vocab_size]), tf.get_variable(\"softmax_b\", [vocab_size]))<\/code><\/p>\n\n<p>&lt;=> <strong>output<\/strong>[batch_size x num_steps, size] x <strong>softmax_w<\/strong>[size, vocab_size] will output <strong>logits<\/strong> of size [<strong>batch_size<\/strong> x <strong>num_steps<\/strong>, <strong>vocab_size<\/strong>].<br>\nFor your case, <strong>logits<\/strong> of size [<strong>20<\/strong>, <strong>vocab_size<\/strong>]\n--> <strong>probs<\/strong> tensor has same size as <strong>logits<\/strong> by [<strong>20<\/strong>, <strong>vocab_size<\/strong>].  <\/p>\n\n<p>Code line: <\/p>\n\n<p><code>chosen_word = np.argmax(probs, 1)<\/code><\/p>\n\n<p>will output <strong>chosen_word<\/strong> tensor of size [<strong>20<\/strong>, <strong>1<\/strong>] with each value is the next prediction word index of current word.  <\/p>\n\n<p>Code line: <\/p>\n\n<p><code>loss = seq2seq.sequence_loss_by_example([logits], [tf.reshape(self._targets, [-1])], [tf.ones([batch_size * num_steps])])<\/code><\/p>\n\n<p>is to compute the softmax cross entropy loss for <em>batch_size<\/em> of sequences.<\/p>\n"}]}
{"QuestionId":36286701,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><tensorflow>","CreationDate":"2016-03-29T14:05:08.017","AcceptedAnswerId":36287541.0,"OwnerUserId":2527680.0,"Title":"TensorFlow: Dimension not compatible error in convolutional network","Body":"<pre><\/pre>\n\n<p>I would like to run the code above and feed images of size  to this network. However, when I try to build the network I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I presume that the weights variable  has the wrong shape, but I'm not sure what the correct one is.<\/p>\n","answers":[{"AnswerId":"36287541","CreationDate":"2016-03-29T14:40:56.137","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The issue is that your input image (in <code>_X<\/code>) has 3 channels (presumably Red, Green, and Blue), whereas the convolutional filter for layer <code>conv1<\/code> (in <code>_weights['wc1']<\/code>) expects 1 input channel.<\/p>\n\n<p>There are at least two possibilities for how to solve this problem:<\/p>\n\n<ol>\n<li><p>Redefine <code>_weights['wc1']<\/code> to accept 3 input channels:<\/p>\n\n<pre><code>weights = {\n    'wc1': tf.Variable(tf.random_normal([7, 7, 3, 96])), # ...\n    # ...\n}\n<\/code><\/pre><\/li>\n<li><p>Convert your input image <code>_X<\/code> to have 1 input channel using the <code>tf.image.rgb_to_grayscale()<\/code> op:<\/p>\n\n<pre><code>_X = tf.image.rgb_to_grayscale(_X)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":36288235,"AnswerCount":3,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-03-29T15:11:05.430","AcceptedAnswerId":null,"OwnerUserId":6129871.0,"Title":"How to get stable results with TensorFlow, setting random seed","Body":"<p>I'm trying to run a neural network multiple times with different parameters in order to calibrate the networks parameters (dropout probabilities, learning rate e.d.). However I am having the problem that running the network while keeping the parameters the same still gives me a different solution when I run the network in a loop as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I am using the following code at the start of my train_network function before setting up the layers and error function of my network:<\/p>\n\n<pre><\/pre>\n\n<p>I have also tried adding this code before the TensorFlow graph creation, but I keep getting different solutions in my results output. <\/p>\n\n<p>I am using an AdamOptimizer and am initializing network weights using . Additionally I am using  to shuffle the incoming images for each epoch.<\/p>\n","answers":[{"AnswerId":"58293360","CreationDate":"2019-10-08T20:05:42.947","ParentId":null,"OwnerUserId":"8279925","Title":null,"Body":"<p>Please add all random seed functions before your code:<\/p>\n\n<p><strong>tf.reset_default_graph()<\/strong><\/p>\n\n<p><strong>tf.set_random_seed(0)<\/strong><\/p>\n\n<p><strong>random.seed(0)<\/strong><\/p>\n\n<p><strong>np.random.seed(0)<\/strong><\/p>\n\n<p>I think, some models in TensorFlow are using numpy or the python random function.<\/p>\n"},{"AnswerId":"38423188","CreationDate":"2016-07-17T16:01:15.577","ParentId":null,"OwnerUserId":"353091","Title":null,"Body":"<p>Deterministic behaviour can be obtained either by supplying a graph-level or an operation-level seed. Both worked for me. A graph-level seed can be placed with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/set_random_seed\" rel=\"noreferrer\">tf.set_random_seed<\/a>. An operation-level seed can be placed e.g, in a variable intializer as in:<\/p>\n\n<pre><code>myvar = tf.Variable(tf.truncated_normal(((10,10)), stddev=0.1, seed=0))\n<\/code><\/pre>\n"},{"AnswerId":"36289575","CreationDate":"2016-03-29T16:10:26.807","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Setting the current TensorFlow random seed affects the current default graph only. Since you are creating a new graph for your training and setting it as default (<code>with g.as_default():<\/code>), you must set the random seed within the scope of that <code>with<\/code> block.<\/p>\n\n<p>For example, your loop should look like the following:<\/p>\n\n<pre><code>for i in range(3):\n  g = tf.Graph()\n  with g.as_default():\n    tf.set_random_seed(1)\n    accuracy_result, average_error = network.train_network(\n        parameters, inputHeight, inputWidth, inputChannels, outputClasses)\n<\/code><\/pre>\n\n<p>Note that this will use the same random seed for each iteration of the outer <code>for<\/code> loop. If you want to use a different&mdash;but still deterministic&mdash;seed in each iteration, you can use <code>tf.set_random_seed(i + 1)<\/code>.<\/p>\n"}]}
{"QuestionId":36289196,"AnswerCount":1,"Tags":"<snapshot><caffe><pre-trained-model><resuming-training>","CreationDate":"2016-03-29T15:53:40.493","AcceptedAnswerId":42948550.0,"OwnerUserId":2736559.0,"Title":"Caffe: resuming from a trained snapshot fails with an error","Body":"<p>I trained my network with some configuration , and then saved a snapshot of it.<br>\nNow I am trying to resume from the last snapshot and it fails with this error message :  <\/p>\n\n<pre><\/pre>\n\n<p>This is how I'm trying to resume it :<\/p>\n\n<pre><\/pre>\n\n<p>I then gave up, and I tried to use  instead of , but I get this error :    <\/p>\n\n<pre><\/pre>\n\n<p>when I tried different times with different models , the history part changes ( like 58 vs 28, 32 vs 28 for example and like this, the over all error is the same, the number differs though!)<\/p>\n\n<p>What should I do? this is making me crazy!<\/p>\n","answers":[{"AnswerId":"42948550","CreationDate":"2017-03-22T10:12:27.847","ParentId":null,"OwnerUserId":"673242","Title":null,"Body":"<p>As the value for the --snapshot argument, you have to pass the .solverstate.h5 file, not the .caffemodel.h5 file.<\/p>\n"}]}
{"QuestionId":36289743,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-03-29T16:18:31.817","AcceptedAnswerId":36291952.0,"OwnerUserId":3932212.0,"Title":"Torch7: Slice Tensor using ByteTensor mask","Body":"<p>I have two tensors:<\/p>\n\n<ol>\n<li>labels is a 1D Tensor (5000)<\/li>\n<li>dataset is  4D Tensor (5000,1,32,32)<\/li>\n<\/ol>\n\n<p>I would like to efficiently slice the labels and dataset corresponding to label of value 1. I succeed in slicing the labels but not the dataset.<\/p>\n\n<p>Slicing the labels:<\/p>\n\n<pre><\/pre>\n\n<p>I tried doing the following to slice the dataset and failed:<\/p>\n\n<pre><\/pre>\n\n<p>Is there an <strong>elegant<\/strong> approach to perform this in Torch7?<\/p>\n","answers":[{"AnswerId":"36291952","CreationDate":"2016-03-29T18:10:34.303","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<pre><code>sliced_dataset = dataset:index(1, positive_mask:nonzero():squeeze())\n<\/code><\/pre>\n"}]}
{"QuestionId":36290181,"AnswerCount":1,"Tags":"<python><theano><deep-learning><keras>","CreationDate":"2016-03-29T16:40:08.367","AcceptedAnswerId":37212814.0,"OwnerUserId":2515942.0,"Title":"AssertionError while training mnist dataset using theano and keras","Body":"<p>I was doing a tutorial on deep learning and there was mnist dataset classification example:<\/p>\n\n<pre><\/pre>\n\n<p>However, on the line:<\/p>\n\n<pre><\/pre>\n\n<p>I get an exception, \"AssertionError was unhandled by user code\". I get no other message in the exception.<\/p>\n\n<p>Any help would be welcome. Thank you.<\/p>\n","answers":[{"AnswerId":"37212814","CreationDate":"2016-05-13T14:36:08.973","ParentId":null,"OwnerUserId":"2515942","Title":null,"Body":"<p>The correct code should be:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\n\nnp.random.seed(123)\n\nfrom keras.datasets import mnist\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Convolution2D, MaxPooling2D\nfrom keras.utils import np_utils\n\nbatch_size = 128\nnb_classes = 10 # number of classes\n\n# input image dimensions\nimg_rows, img_cols = 28, 28\n\n(X_train, y_train) , (X_test, y_test) = mnist.load_data()\n\n# reshape data\nX_train =  X_train.reshape(X_train.shape[0], 1, img_rows, img_cols) # 1 is 1 channnel i.e. grayscale\nX_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)\n\nX_train = X_train.astype(\"float32\")\nX_test = X_test.astype(\"float32\")\n\n# normalization\nX_train \/= 255;\nX_test \/= 255;\n\nprint(\"X_train shape: \", X_train.shape)\nprint(X_train.shape[0], \" train samples\")\nprint(X_test.shape[0], \" test samples\")\n\n# convert class vectors to binary class matrices\n\ny_train = np_utils.to_categorical(y_train, nb_classes)\ny_test = np_utils.to_categorical(y_test, nb_classes)\nprint(\"One hot encoding: {}\".format(y_train[0, :]))\n\nmodel = Sequential();\n\nmodel.add(Convolution2D(6, 5, 5, input_shape = (1, img_rows, img_cols), border_mode = \"same\"))\n\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Convolution2D(16, 5, 5, border_mode= \"same\"))\n\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Convolution2D(120, 5, 5, border_mode= \"same\"))\n\nmodel.add(Activation(\"relu\"))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(84))\nmodel.add(Activation(\"relu\"))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10)) # because 10 output classes\nmodel.add(Activation(\"softmax\"))\n\n# TRAIN\nmodel.compile(loss = \"categorical_crossentropy\", optimizer=\"adadelta\")\nnb_epoch = 2\n\nmodel.fit(X_train, y_train, batch_size = batch_size, nb_epoch = nb_epoch, show_accuracy= True, verbose = 1, validation_data = (X_test, y_test))\nscore = model.evaluate(X_test, y_test, show_accuracy = True, verbose = 0)\nprint(\"Test score: \", score[0])\nprint(\"Test accuracy: \", score[1])\n<\/code><\/pre>\n"}]}
{"QuestionId":36290883,"AnswerCount":1,"Tags":"<neural-network><tensorflow>","CreationDate":"2016-03-29T17:17:55.130","AcceptedAnswerId":36294703.0,"OwnerUserId":5195629.0,"Title":"Implementing Adversarial Training in TensorFlow","Body":"<p>I would like to implement the following cost function for my neural network:\n<img src=\"https:\/\/latex.codecogs.com\/gif.latex?%5Cwidetilde&space;J(%5Ctheta,x,t)&space;=&space;J(%5Ctheta,x,t)&space;&plus;&space;J(%5Ctheta,x&plus;%5Cepsilon&space;%5Cnabla&space;J(%5Ctheta,&space;x,&space;t))\" title=\"\\widetilde J(\\theta,x,t) = J(\\theta,x,t) + J(\\theta,x+\\epsilon \\nabla J(\\theta, x, t))\" \/><\/p>\n\n<p>This makes use of adversarial inputs for the neural network to improve generalization <a href=\"http:\/\/arxiv.org\/pdf\/1412.6572.pdf\" rel=\"noreferrer\">[ref]<\/a>.<\/p>\n\n<p>Specifically, I am having trouble with the\n<img src=\"https:\/\/latex.codecogs.com\/gif.latex?x&plus;%5Cepsilon%5Cnabla&space;J(%5Ctheta,&space;x,&space;t)\" title=\"\\nabla J(\\theta, x, t)\" \/> part.\nIn my TensorFlow graph, I have defined <img src=\"https:\/\/latex.codecogs.com\/gif.latex?J(%5Ctheta,&space;x,&space;t)\" title=\"J(\\theta, x, t)\" \/> as an operation. How can I feed <img src=\"https:\/\/latex.codecogs.com\/gif.latex?J\" title=\"J\" \/> with an argument other than <img src=\"https:\/\/latex.codecogs.com\/gif.latex?x\" title=\"x\" \/>?<\/p>\n\n<p>The only way I have found to do this so far is to define a parallel network <img src=\"https:\/\/latex.codecogs.com\/gif.latex?J%27(%5Ctheta,&space;x&space;&plus;&space;%5CDelta,&space;t)\" title=\"J'(\\theta, x + \\Delta, t)\" \/> that shares the weights with my original network and passing it <img src=\"https:\/\/latex.codecogs.com\/gif.latex?x&space;&plus;&space;%5CDelta\" title=\"x + \\Delta\" \/> in its  argument. If possible, I would like to avoid having to redefine my entire network. How can I do this?<\/p>\n\n<hr>\n\n<p>My TensorFlow model is written as:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Possibly relevant:<\/p>\n\n<blockquote>\n  <h2>tf.stop_gradient(input, name=None)<\/h2>\n  \n  <p>Stops gradient computation.<\/p>\n  \n  <p><em>...lots more stuff...<\/em><\/p>\n  \n  <ul>\n  <li>Adversarial training, where no backprop should happen through the adversarial example generation process.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#stop_gradient\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#stop_gradient<\/a><\/p>\n","answers":[{"AnswerId":"36294703","CreationDate":"2016-03-29T20:42:52.253","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>You need to write your model in such a way that it supports calls like<\/p>\n\n<pre><code>output = model.fprop(input_tensor)\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>output = model.fprop(input_tensor, params)\n<\/code><\/pre>\n\n<p>The fprop method builds the same forward propagation expression twice, but with a different input tensor on each call:<\/p>\n\n<pre><code>raw_output = model.fprop(clean_examples)\nadv_examples = ...\nadv_output = model.fprop(adv_examples)\n<\/code><\/pre>\n\n<p>If you want to apply this to one of our open source models and it doesn't support the interface to do this, file an issue on github.<\/p>\n"}]}
{"QuestionId":36293208,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2016-03-29T19:16:26.243","AcceptedAnswerId":36310065.0,"OwnerUserId":null,"Title":"Keras: 2D input -> 2D output?","Body":"<p>I want to build a neural network to learn a set of standard feature vectors. The set is thus of shape (N,100), where N is the number of samples. However, set of labels is of shape (Nx18) (e.g. each \"label\" is another array of 18 elements). I'm quite new to keras and neural nets, and I only know how to deal with the label when it is one dimensional (e.g. 0 or 1 in binary classification). How do I can deal with multi-dimensional output?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36310065","CreationDate":"2016-03-30T13:18:26.420","ParentId":null,"OwnerUserId":"6099650","Title":null,"Body":"<p>Maybe I don't completely understand the question but the simplest way would be to have an output layer with 18 neuron. Each neuron output one value, i.e. the output will be a vector of 18 values. <\/p>\n\n<p>One possible way of doing this would be a feed-forward neural network with on hidden layer, e.g. containing 100 neurons. You will need the <a href=\"http:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow\">Dense layer<\/a> in Keras for this.<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>nb_hidden = 100\n\nmodel = Sequential()\nmodel.add(Dense(input_dim = 100, output_dim = nb_hidden)\nmodel.add(Dense(output_dim = 18, activation = 'softmax')\nmodel.compile(loss='categorical_crossentropy', optimizer='adadelta')\n<\/code><\/pre>\n\n<p>Consider varying the number of hidden layers, the general network topology (e.g. include a Dropout layer) and the activation functions until you get a good result.<\/p>\n"}]}
{"QuestionId":36295004,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-03-29T21:00:49.300","AcceptedAnswerId":36299982.0,"OwnerUserId":280069.0,"Title":"How to fix dimensional error in Theano v0.8 tutorial code","Body":"<p>I'm following the <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/examples.html\" rel=\"nofollow\">tutorial<\/a> in the v0.8 release of Theano. <\/p>\n\n<p>It includes an example of creating &amp; training a logistical regression as shown below:<\/p>\n\n<pre><\/pre>\n\n<p>Theano is throwing the following error:<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name\n  \".\/tut.py:206\"  at index 1(0-based)', 'Wrong number of dimensions:\n  expected 2, got 1 with shape (400,).')<\/p>\n<\/blockquote>\n\n<p>I'm reasonably certain the fix is simple (due to my newb status at Theano), and perhaps involves a reshape step. The tutorial doesn't have a good hint. Suggestions?<\/p>\n","answers":[{"AnswerId":"36299982","CreationDate":"2016-03-30T04:53:58.300","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>try to reshape <code>D[1]<\/code> before you use it in the theano function, maybe something like this (I haven't try it, tell me if it doesn't work):<\/p>\n\n<pre><code>pred, err = train(D[0], np.reshape(D[1],(400,1))\n<\/code><\/pre>\n\n<p>the error happen because you initialize <code>D[1]<\/code> in one dimensional array using <code>rng.randint(size=N, low=0, high=2)<\/code> but it's pass to matrix (two dimensional) variable <code>y<\/code><\/p>\n\n<p>or<\/p>\n\n<p>the other simple solution is by using vector instead of matrix for variable <code>y<\/code>:<\/p>\n\n<pre><code>y = T.vector(\"y\")\n<\/code><\/pre>\n"}]}
{"QuestionId":36295516,"AnswerCount":1,"Tags":"<tensorflow><bazel><raspberry-pi3>","CreationDate":"2016-03-29T21:34:13.713","AcceptedAnswerId":null,"OwnerUserId":2628128.0,"Title":"Distributed compilation for Tensorflow","Body":"<p>I am trying to use  for speeding-up the Tensorflow compilation. My  installations (both on host and client) are working correctly with traditional source packages that use Makefiles for compilation (). But apparently  does not use  for compiling the Tensorflow sources.<\/p>\n\n<p>Note that the target platform is a Raspberry Pi 3 running Raspbian Jessy (Ubuntu), thus it is important to offload the compilation on a more powerful machine.<\/p>\n\n<p>Do you have any experience or suggestions?<\/p>\n","answers":[{"AnswerId":"46627144","CreationDate":"2017-10-08T03:04:58.043","ParentId":null,"OwnerUserId":"1131467","Title":null,"Body":"<p>Yes.  We are using bazel with distcc.  You have to write a CROSSTOOL file and then create a wrapper script for gcc that prepends distcc and then forwards the arguments to real gcc.<\/p>\n"}]}
{"QuestionId":36295624,"AnswerCount":3,"Tags":"<python><numpy><machine-learning><tensorflow>","CreationDate":"2016-03-29T21:41:40.353","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"Tensorflow loss not decreasing and acuraccy stuck at 0.00%?","Body":"<h1>Code<\/h1>\n\n<p>I\"m trying to train CNN with UCF101 single frame data.\nAs far as I understand the problem is either with the weight Initialization or \nthe loss with either the or the cost and optimizer function.<\/p>\n\n<p>Also is there any way to use xavier initialization ??<\/p>\n\n<pre><\/pre>\n\n<h1>Output<\/h1>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"40648525","CreationDate":"2016-11-17T06:57:22.897","ParentId":null,"OwnerUserId":"5367279","Title":null,"Body":"<p>I faced the same problem. drop the relu activation from last dense layer. worked for me.<\/p>\n\n<pre><code> dense2 = tf.nn.relu(tf.matmul(dense1, _weights['wd2']) + _biases['bd2'], name='fc7') # Relu activation\n<\/code><\/pre>\n\n<p>after this again<\/p>\n\n<pre><code>out = tf.nn.softmax(tf.matmul(dense2, _weights['out']) + _biases['out'])\n<\/code><\/pre>\n\n<p>Thus the first \"relu\" seems to be redundant to me. I did a similar thing. After dropping that line it solved the problem.<\/p>\n"},{"AnswerId":"36317143","CreationDate":"2016-03-30T18:41:51.460","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>You can use a different initialization scheme by creating numpy arrays for the initial values of your variables.<\/p>\n\n<p>Your loss isn't changing at all, so one way to debug is to confirm that the variables are actually changing as you do updates.<\/p>\n"},{"AnswerId":"36715531","CreationDate":"2016-04-19T10:21:26.643","ParentId":null,"OwnerUserId":"6224293","Title":null,"Body":"<p>From Tensorflow <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/api_docs\/python\/nn.html#softmax_cross_entropy_with_logits\" rel=\"nofollow\">documentation<\/a>!<\/p>\n\n<p><em>WARNING: This op expects unscaled logits, since it performs a softmax on logits internally for efficiency. Do not call this op with the output of softmax, as it will produce incorrect results.<\/em><\/p>\n\n<p>With this in mind you should take out <\/p>\n\n<pre><code>out = tf.nn.softmax(tf.matmul(dense2, _weights['out']) + _biases['out'])\n<\/code><\/pre>\n\n<p>from your network definition and replace with <\/p>\n\n<pre><code>out = tf.matmul(dense2, _weights['out']) + _biases['out']\n<\/code><\/pre>\n\n<p>Since you are not fine-tuning i.e not porting weights trained on a similar problem the training would tend to be slow. Keeping in mind this could be one of many problems with the training. Hope it helps. <\/p>\n"}]}
{"QuestionId":36296760,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-29T23:10:46.653","AcceptedAnswerId":36296897.0,"OwnerUserId":3919542.0,"Title":"How to pass images to the model for classification in Tensorflow","Body":"<p>I have created a model using the following code below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have saved the model and wrote another python program where i am trying to restore the model and use it for classification of my images , but i am not being able to create a 4D tensor of the image , that i have to pass as input to the model.<\/p>\n\n<p>The code of the python file is as follows :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>As you can see in ln[19] i have restored my model and want to pass an image to the model by creating a 4d Tensor , I am reading the image and then trying to convert it to a 4d tensor but the ysntax for creating it is wrong in my code , thus need help in correcting it .<\/p>\n","answers":[{"AnswerId":"36296897","CreationDate":"2016-03-29T23:24:18.663","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Assuming that <code>image_data<\/code> is a <em>grayscale<\/em> image, it should be a 2-D NumPy array. You can convert it to a 4-D array with the following:<\/p>\n\n<pre><code>data = image_data[np.newaxis, ..., np.newaxis]\n<\/code><\/pre>\n\n<p>The <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis\" rel=\"noreferrer\"><code>np.newaxis<\/code><\/a> adds a new dimension of size 1 in the first (batch size) and last (channels) dimensions. It is equivalent to the following, using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy-1.10.1\/reference\/generated\/numpy.expand_dims.html\" rel=\"noreferrer\"><code>np.expand_dims()<\/code><\/a>:<\/p>\n\n<pre><code>data = np.expand_dims(np.expand_dims(image_data, 0), -1)\n<\/code><\/pre>\n\n<p>On the other hand, if you are working with RGB data, you will need to convert it to fit the model. You could for example define a placeholder for the image input:<\/p>\n\n<pre><code>input_placeholder = tf.placeholder(tf.float32, shape=[None, image_size, image_size, 3])\ninput_grayscale = tf.image.rgb_to_grayscale(input_placeholder)\n\nprediction = tf.nn.softmax(model(input_grayscale))\n\nimage_data = ...  # Load from file\ndata = image_data[np.newaxis, ...]  # Only add a batch dimension.\n\nprediction_val = sess.run(prediction, feed_dict={input_placeholder: data})\n<\/code><\/pre>\n"}]}
{"QuestionId":36297662,"AnswerCount":0,"Tags":"<python-2.7><deep-learning><caffe>","CreationDate":"2016-03-30T00:42:45.483","AcceptedAnswerId":null,"OwnerUserId":1556826.0,"Title":"Layers not activated in Caffe","Body":"<p>I am using the pretrained AlexNet to classify a set of images. However, I find that many of the layers are not activated at all. For example, the output of the fc7 layer is all zero. I'm stuck at a dead end trying to figure out what is going wrong?<\/p>\n\n<p>I have my training and testing data stored in separate .txt files in the following format<\/p>\n\n<pre><\/pre>\n\n<p>The code that I have written is<\/p>\n\n<pre><\/pre>\n\n<p>The fc7 layer contains only zeros. <\/p>\n","answers":[]}
{"QuestionId":36302137,"AnswerCount":1,"Tags":"<machine-learning><neural-network><theano><deep-learning>","CreationDate":"2016-03-30T07:19:58.123","AcceptedAnswerId":36312289.0,"OwnerUserId":5257450.0,"Title":"Can I have the shape information of the shared variable in theano?","Body":"<p>It seems that variable.shape would notify me that<\/p>\n\n<pre><\/pre>\n\n<p>while theano.tensor.shape(variable) will return me a <\/p>\n\n<p>I am really confused why can't I get a shape information on that? The same problem occurs when I want to get the shape information of a symbolic variable. It is just so weird.<\/p>\n\n<pre><\/pre>\n\n<p>In the example above, the x (symbolic variable) has been reshaped to some shape, if would not make sense to me if I can't retrieve its shape information while could still assigning it new shape.<\/p>\n","answers":[{"AnswerId":"36312289","CreationDate":"2016-03-30T14:45:05.447","ParentId":null,"OwnerUserId":"3489247","Title":null,"Body":"<p>The first error is probably due to the fact that you tried to evaluate the <code>shape<\/code> property on the data type <code>SharedVariable<\/code>, not on an actual shared variable.<\/p>\n\n<p>Otherwise, obtaining <code>shape.0<\/code> is completely normal: This is a symbolic expression representing the shape, which is a priori unknown. As soon as you evaluate with data, you will see the shape:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport numpy as np\ns = theano.shared(np.arange(2 * 3 * 5).reshape(2, 3, 5))\nprint(s.shape)  # gives you shape.0\nprint(s.shape.eval())  # gives you an array containing 2, 3, 5\n\na = T.tensor3()\nprint(a.shape)  # gives you shape.0\nprint(a.shape.eval({a: np.arange(2 * 3 * 5).reshape(2, 3, 5).astype(theano.config.floatX)}))  # gives 2, 3, 5\n<\/code><\/pre>\n"}]}
{"QuestionId":36303340,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-03-30T08:21:30.640","AcceptedAnswerId":36311677.0,"OwnerUserId":1994648.0,"Title":"Specify dynamic shape as an input argument","Body":"<p>For tensorflow functions that require a shape argument, how can we feed a dynamic shape as one of the input argument(s)?<\/p>\n\n<p>For instance, let's say I have a placeholder<\/p>\n\n<pre><\/pre>\n\n<p>And now I want to create a 1D random tensor with the same shape as the first dimension of .<\/p>\n\n<pre><\/pre>\n\n<p>This will not work since  will return a static shape \"None\". Is there a a way to assign a dynamic shape to  based on the dynamic shape of ?<\/p>\n","answers":[{"AnswerId":"36303885","CreationDate":"2016-03-30T08:47:40.977","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<pre><code>x = tf.placeholder(tf.float32, [None, 3])\ns = tf.shape(x)\ns2 = tf.slice(s, [0], [1])\nr = tf.random_uniform(s2)\nsess = create_session()\nprint r.eval(feed_dict={x:np.ones((4,3))})\n<\/code><\/pre>\n\n<blockquote>\n  <p>[ 0.69890845  0.64149153  0.16378665  0.89732885]<\/p>\n<\/blockquote>\n"},{"AnswerId":"36311677","CreationDate":"2016-03-30T14:20:50.847","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can get the dynamic shape of a tensor using the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#shape\" rel=\"nofollow\"><code>tf.shape()<\/code><\/a> op, and use that as the argument to <code>tf.random_uniform()<\/code>:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, 3])\nr = tf.random_uniform(tf.shape(x)[0:1])\n<\/code><\/pre>\n\n<p>(Note that the slice of <code>tf.shape(x)<\/code>, i.e. <code>[0:1]<\/code> must specify a range rather than a single element, because <code>tf.random_uniform()<\/code> expects a vector of dimensions.)<\/p>\n"}]}
{"QuestionId":36303610,"AnswerCount":1,"Tags":"<python><machine-learning><sentiment-analysis><keras><lstm>","CreationDate":"2016-03-30T08:34:16.653","AcceptedAnswerId":null,"OwnerUserId":4583366.0,"Title":"Whats the minimum training set size needed for sentiment classification task on short texts","Body":"<p>I'm trying to train an LSTM model for the task of sentiment classification on short texts such as products reviews and tweets. <\/p>\n\n<p>I'm looking for a training set that labels positive\/negative\/neutral, is there such thing (free for research) out there that is really based on human tags and not on starts or emoticons? Iv'e found only small training sets which led me to poor results. Iv'e tried to increase the size of my network and stacked layers but no improvement. <\/p>\n\n<p>Whats the minimum size for such a training set in order to start getting reasonable results (F1 > 0.8).<\/p>\n","answers":[{"AnswerId":"36303806","CreationDate":"2016-03-30T08:43:38.550","ParentId":null,"OwnerUserId":"375828","Title":null,"Body":"<p>You really only would need to make a list of 100 or so, each, of negative and positive adjectives.<\/p>\n\n<p>See:<br\/>\n<a href=\"http:\/\/na2english.wikispaces.com\/file\/view\/ADJECTIVES%20TO%20DESCRIBE%20FILMS.pdf\/400672720\/ADJECTIVES%20TO%20DESCRIBE%20FILMS.pdf\" rel=\"nofollow\">http:\/\/na2english.wikispaces.com\/file\/view\/ADJECTIVES%20TO%20DESCRIBE%20FILMS.pdf\/400672720\/ADJECTIVES%20TO%20DESCRIBE%20FILMS.pdf<\/a><br\/><br\/>\n<a href=\"http:\/\/arxiv.org\/ftp\/arxiv\/papers\/1011\/1011.4623.pdf\" rel=\"nofollow\">http:\/\/arxiv.org\/ftp\/arxiv\/papers\/1011\/1011.4623.pdf<\/a><br\/><br\/><\/p>\n\n<p>Obviously cite them if you use them, but language is free, so you can use those for your work.<\/p>\n\n<p>Probably more important than the size of the database you construct will be picking words that target your specific application for increased efficacy.<\/p>\n\n<p>Are you aiming this project at a specific commercial use or as a more generalized research effort?<\/p>\n"}]}
{"QuestionId":36305495,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-30T10:01:31.250","AcceptedAnswerId":36311407.0,"OwnerUserId":3919542.0,"Title":"No classification done after passing an image to the model in Tensorflow","Body":"<p>I am trying to pass an image to the model that i have created by following the 2_fullyconnected.ipynb udacity assignment. <\/p>\n\n<p>The code in which i have created the model is shown below . <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The model is saved in \/tmp\/important_model\/.<\/p>\n\n<p>Tree structure for that folder is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Now i am creating a new file in which i am trying to restore my model and then pass an image to the model for classification .<\/p>\n\n<p>I have created the graph in the new python file as well , which is necessary for restoring the model (I think, I could be wrong. please correct me if i am wrong).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When i am executing ln[34] in this ipython notebookthe output that is coming is :<\/p>\n\n<pre><\/pre>\n\n<p>I want to tell the 5 probable labels which the given image may belong to but don't know how to do it , The above program doesn't shows any error but neither shows the desired output . I thought i will get the probabilities of the image being in all classes as i have passed my image in tf.nn.softmax function but unfortunately not getting anything .<\/p>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"36311407","CreationDate":"2016-03-30T14:10:50.030","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The following line in your code computes a probability distribution across the possible output labels for each image in your data set (in this case a single image):<\/p>\n\n<pre><code>session.run(valid_prediction,feed_dict={tf_valid_dataset:resized_data})\n<\/code><\/pre>\n\n<p>The result of this method is a NumPy array of shape <code>(1, 10)<\/code>. To see the probabilities, you can simply print the array:<\/p>\n\n<pre><code>result = session.run(valid_prediction,feed_dict={tf_valid_dataset:resized_data})\nprint(result)\n<\/code><\/pre>\n\n<p>There are many ways that you can get the top <em>k<\/em> predictions for your image. One of the easiest is to use TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/nn.html#top_k\" rel=\"nofollow\"><code>tf.nn.top_k()<\/code><\/a> operator when defining your graph:<\/p>\n\n<pre><code>valid_prediction = tf.nn.softmax(tf.matmul(tf_valid_dataset, weights) + biases)\ntop_5_labels = tf.nn.top_k(valid_prediction, k=5)\n\n# ...\nresult = session.run(top_5_labels, feed_dict={tf_valid_dataset: resized_data})\nprint(result)\n<\/code><\/pre>\n"}]}
{"QuestionId":36309439,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-03-30T12:51:52.007","AcceptedAnswerId":null,"OwnerUserId":5587428.0,"Title":"mnist_with_summaries.py error: 'Graph' object has no attribute 'SerializeToString'","Body":"<p>When I try to run mnist_with_summaries.py I get the following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"\/home\/rob\/tf_from_source\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\", line 110, in \n      tf.app.run()\n    File \"\/home\/rob\/.virtualenvs\/tensorflow\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/platform\/default\/_app.py\", line 30, in run\n      sys.exit(main(sys.argv))\n    File \"\/home\/rob\/tf_from_source\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\", line 85, in main\n      writer = tf.train.SummaryWriter(FLAGS.summaries_dir, sess.graph)\n    File \"\/home\/rob\/.virtualenvs\/tensorflow\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/summary_io.py\", line 104, in <strong>init<\/strong>\n      self.add_graph(graph_def)\n    File \"\/home\/rob\/.virtualenvs\/tensorflow\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/summary_io.py\", line 168, in add_graph\n      graph_bytes = graph_def.SerializeToString()\n  AttributeError: 'Graph' object has no attribute 'SerializeToString'<\/p>\n<\/blockquote>\n\n<p>I'm also seeing this error on some of my own code when I try to generate a tensorboard graph. Any ideas about the problem and solution would be appreciated.<\/p>\n","answers":[{"AnswerId":"36310123","CreationDate":"2016-03-30T13:21:07.457","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>Looks like the answer was provided <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1690\" rel=\"nofollow\">here<\/a>. I rolled back to the r0.7 branch and the problem was resolved.<\/p>\n"}]}
{"QuestionId":36309718,"AnswerCount":4,"Tags":"<deep-learning><caffe>","CreationDate":"2016-03-30T13:03:27.830","AcceptedAnswerId":null,"OwnerUserId":1525479.0,"Title":"Caffe Autoencoder","Body":"<p>I wanna compare the performance of CNN and autoencoder in caffe. I'm completely familiar with cnn in caffe but I wanna is the autoencoder also has deploy.prototxt file ? is there any differences in using this two models rather than the architecture? <\/p>\n","answers":[{"AnswerId":"43566823","CreationDate":"2017-04-23T02:36:53.173","ParentId":null,"OwnerUserId":"5768578","Title":null,"Body":"<p>You need to implement an auto-encoder example using python or matlab. The example in Caffe is not true auto-encoder because it doesn't set layer-wise training stage and during training stage, it doesn't fix W{L->L+1} = W{L+1->L+2}^T. It is easily to find a 1D auto-encoder in github, but 2D auto-encoder may be hard to find.<\/p>\n"},{"AnswerId":"36326119","CreationDate":"2016-03-31T06:51:34.487","ParentId":null,"OwnerUserId":"4095689","Title":null,"Body":"<p>Yes it also has a deploy.prototxt.<\/p>\n\n<p>both <code>train_val.prototxt<\/code> and 'deploy.prototxt' are cnn architecture description files. The sole difference between them is, <code>train_val.prototxt<\/code> takes training data and loss as input\/output, but 'deploy.prototxt' takes testing image as input, and predicted value as out put.<\/p>\n"},{"AnswerId":"36357536","CreationDate":"2016-04-01T13:23:33.367","ParentId":null,"OwnerUserId":"5299473","Title":null,"Body":"<p>Here is an example of a cnn and autoencoder for MINST: <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/mnist\" rel=\"nofollow\">Caffe Examples<\/a>. (I have not tried the examples.) Using the models is generally the same. Learning rates etc. depend on the model.<\/p>\n"},{"AnswerId":"44716056","CreationDate":"2017-06-23T07:57:12.097","ParentId":null,"OwnerUserId":"3246146","Title":null,"Body":"<p>The main difference between the Auto encoders and conventional network is <\/p>\n\n<ol>\n<li>In Auto encoder your input is your label image for training.<\/li>\n<li>Auto encoder tries to approximate the output similar as input.<\/li>\n<li>Auto encoders does not have softmax layer while training.<\/li>\n<li>It can be used as a pre-trained model for your network which converge faster comparing to other pre-trained models. It is because your network has already extracted the features for your data.<\/li>\n<li>The Conventional training and testing you can perform on pre trained auto encoder network for faster convergence and accuracy.<\/li>\n<\/ol>\n"}]}
{"QuestionId":36311451,"AnswerCount":0,"Tags":"<opencv><caffe><kurento>","CreationDate":"2016-03-30T14:12:28.187","AcceptedAnswerId":null,"OwnerUserId":5176361.0,"Title":"How to integrate caffe in KurentoMediaServer","Body":"<p><strong>What I do<\/strong>:<\/p>\n\n<p>1.I installed the opencv-plugin-sample in the kurenoMediaPlayer(from the link <a href=\"https:\/\/github.com\/Kurento\/kms-opencv-plugin-sample\" rel=\"nofollow\">https:\/\/github.com\/Kurento\/kms-opencv-plugin-sample<\/a>) and run a sample opencv face detection.<\/p>\n\n<p><strong>What i have to do:<\/strong><\/p>\n\n<p>1.Now i need to integrate <strong>caffe<\/strong>(<a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/install_apt.html<\/a>) into KMS ,so I can able to run more complex opencv algorithms in KMS.<\/p>\n\n<p><strong>What I need to know:<\/strong><\/p>\n\n<p>1.Is there any specific way to integrate caffe in KurentoMediaServer?<\/p>\n","answers":[]}
{"QuestionId":36312018,"AnswerCount":2,"Tags":"<lua><ubuntu-14.04><torch><luajit>","CreationDate":"2016-03-30T14:34:01.280","AcceptedAnswerId":36489444.0,"OwnerUserId":2838606.0,"Title":"Unable to import (require) 'cutorch' in Torch","Body":"<p>I have CUDA 7.5 installed on my machine and installed Torch by following the instructions <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">here<\/a>. Before installing Torch I upgraded CMake to version 3.2.x.x. Then I installed  and  using the following commands:<\/p>\n\n<pre><\/pre>\n\n<p>While installing cutorch I saw some messages say8ing that CUDA 7.5 was found on my machine. However, I am unable to import  using  and get the following error when I try to run it using the torch (th) interpreter in Terminal:<\/p>\n\n<pre><\/pre>\n\n<p>Then I followed the instructions <a href=\"http:\/\/www.lighting-torch.com\/2015\/07\/27\/configuring-eclipse-with-torch\/\" rel=\"nofollow\">here<\/a> to be able to use Eclipse to write Lua\/Torch code. There, I get the following error when trying to execute :<\/p>\n\n<pre><\/pre>\n\n<p>This is however the file  is in the mentioned directory. <a href=\"https:\/\/github.com\/torch\/cutorch\/issues\/210\" rel=\"nofollow\">This<\/a> also did not help me resolve the issue. Does anyone know should I go about resolving the issue?<\/p>\n\n<p>When I enter  I get the followings:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36489444","CreationDate":"2016-04-08T00:20:12.800","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Okay, here's how you can resolve the problem (thanks to alban desmaison on Torch7 google group):<\/p>\n\n<ul>\n<li>Make sure you follow all the steps I mentioned in the question in order to make sure CUDA is installed correctly.<\/li>\n<li>If you have Theano installed, execute <code>sudo python<\/code><\/li>\n<li>Then do <code>import theano<\/code><\/li>\n<li>Then go to CUDA samples folder and <code>make<\/code> one of the samples (does not matter which one)<\/li>\n<li>Then do <code>.\/[name_of_the_sample]<\/code><\/li>\n<li>Then Torch will easily be able use CUDA backend<\/li>\n<\/ul>\n\n<p><strong>If you do not have<\/strong> Theano installed, try running the CUDA samples using <code>sudo<\/code>. The <strong>important step<\/strong> is to be able to run one of the CUDA samples.<\/p>\n\n<p>Now you're all set!<\/p>\n"},{"AnswerId":"42378060","CreationDate":"2017-02-21T21:11:20.237","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<p>Try <code>sudo bash \/torch\/update.sh<\/code><\/p>\n"}]}
{"QuestionId":36312253,"AnswerCount":1,"Tags":"<python><numpy><theano><deep-learning>","CreationDate":"2016-03-30T14:43:17.277","AcceptedAnswerId":null,"OwnerUserId":5257450.0,"Title":"Why can't I evaluate the reshaped tensorvariable in theano?","Body":"<p>Why can't I evaluate the reshaped tensorvariable through the code I wrote below?<\/p>\n\n<pre><\/pre>\n\n<p>Since I have reshape the tensorvariable x, and pass a numpy array of same dimension to it, it simply reports,<\/p>\n\n<blockquote>\n  <p>TypeError: ('Bad input argument to theano function with name\n  \":17\"  at index 0(0-based)', 'Wrong\n  number of dimensions: expected 2, got 5 with shape (2592, 1, 51, 61,\n  23).')<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"36314540","CreationDate":"2016-03-30T16:26:31.847","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>I think the problem is because you are using <code>matrix<\/code> (two dimensional) as data type of <code>x<\/code> that receive a five dimensional input <code>xTrain<\/code>. As said <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#custom-tensor-types\" rel=\"nofollow\">here<\/a>, for five dimensional input, you should create a custom data type.<\/p>\n\n<p>sample code:<\/p>\n\n<pre><code>from theano import tensor as T\nimport numpy as np\nxTrain = np.random.rand(1,1,2,3,3).astype('float32')\n\ndtensor5 = T.TensorType('float32', (False,)*5)\nx = dtensor5('x')\n\nlayer1_input = x\nprint layer1_input.eval({x:xTrain})\n<\/code><\/pre>\n\n<p>and about<\/p>\n\n<blockquote>\n  <p>Since I have reshape the tensorvariable x, and pass a numpy array of\n  same dimension to it<\/p>\n<\/blockquote>\n\n<p>I think what actually happen is variable <code>x<\/code> recieve the input first (raise an error) and then you reshape it for <code>layer1_input<\/code><\/p>\n"}]}
{"QuestionId":36312410,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><conv-neural-network>","CreationDate":"2016-03-30T14:50:27.223","AcceptedAnswerId":null,"OwnerUserId":1079110.0,"Title":"How to implement convolutional connections without tied weights?","Body":"<p>Given two layers of a neural network that have a 2D representation, i.e. fields of activation. I'd like to connect each neuron of the lower layer to the near neurons of the upper layer, say within a certain radius. Is this possible with TensorFlow?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/ZSZJT.png\" alt=\"Connection pattern\"><\/p>\n\n<p>This is similar to a convolution, but the weight kernels should not be tied. I'm trying to avoid connecting both layers fully first and masking out most of the parameters, in order to keep the number of parameters low.<\/p>\n","answers":[{"AnswerId":"36317484","CreationDate":"2016-03-30T19:01:20.313","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>I don't see a simple way to do this with existing TensorFlow ops efficiently, but there might be some tricks with sparse things. However, ops for efficient locally connected, non-convolutional neural net layers would be very useful, so you might want to file a feature request as a GitHub issue.<\/p>\n"}]}
{"QuestionId":36312967,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-03-30T15:14:11.353","AcceptedAnswerId":36317544.0,"OwnerUserId":2527680.0,"Title":"Tensorflow CNN model not training? Constant loss and accuracy","Body":"<p>I have built a model using <a href=\"https:\/\/github.com\/soumith\/convnet-benchmarks\/blob\/master\/tensorflow\/benchmark_alexnet.py\" rel=\"nofollow\">this<\/a> as a base. <\/p>\n\n<p>And the train portion from <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/alexnet.py\" rel=\"nofollow\">this code<\/a>.<\/p>\n\n<p>This model does not train and always gives the cost\/loss output for every iteration.<\/p>\n\n<p>I dont think it learns anything.<\/p>\n\n<p>I have checked for the usual stuff like shuffled inputs.\nMaking sure that each batch is new. <\/p>\n\n<p>Any idea why ?<\/p>\n\n<p>This is my <a href=\"http:\/\/pastebin.com\/Z4gA5AEG\" rel=\"nofollow\">code.<\/a><\/p>\n\n<h1>Output<\/h1>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36317544","CreationDate":"2016-03-30T19:05:27.990","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>The code you started from is just a benchmark of the forward and backward pass and isn't designed to do training. You should start from an example that actually trains a model instead and ignore the benchmark code.<\/p>\n\n<p>You might have an easier time starting from a completely working training example program instead of trying to combine two pieces.<\/p>\n"}]}
{"QuestionId":36313934,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow>","CreationDate":"2016-03-30T15:55:10.133","AcceptedAnswerId":36320000.0,"OwnerUserId":1104696.0,"Title":"Is it possible to split a network across multiple GPUs in tensorflow?","Body":"<p>I plan to run a very large recurrent network (e.g. 2048x5), is it possible to define one layer at one GPU in tensorflow? How should I implement the model to achieve the best efficiency. I understand there is overhead for inter-GPU or GPU-CPU-GPU communication.<\/p>\n","answers":[{"AnswerId":"36320000","CreationDate":"2016-03-30T21:21:56.337","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Splitting a large model across multiple GPUs is certainly possible in TensorFlow, but doing it optimally is a hard research problem. In general, you will need to do the following:<\/p>\n\n<ol>\n<li><p>Wrap large contiguous regions of your code in a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#device\"><code>with tf.device(...):<\/code><\/a> block, naming the different GPUs:<\/p>\n\n<pre><code>with tf.device(\"\/gpu:0\"):\n  # Define first layer.\n\nwith tf.device(\"\/gpu:1\"):\n  # Define second layer.\n\n# Define other layers, etc.\n<\/code><\/pre><\/li>\n<li><p>When building your optimizer, pass the optional argument <code>colocate_gradients_with_ops=True<\/code> to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#Optimizer.minimize\"><code>optimizer.minimize()<\/code><\/a> method:<\/p>\n\n<pre><code>loss = ...\noptimizer = tf.train.AdaGradOptimizer(0.01)\ntrain_op = optimizer.minimize(loss, colocate_gradients_with_ops=True)\n<\/code><\/pre><\/li>\n<li><p>(Optionally.) You may need to enable \"soft placement\" in the <code>tf.ConfigProto<\/code> when you create your <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/client.html#Session\"><code>tf.Session<\/code><\/a>, if any of the operations in your model cannot run on GPU:<\/p>\n\n<pre><code>config = tf.ConfigProto(allow_soft_placement=True)\nsess = tf.Session(config=config)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":36315851,"AnswerCount":1,"Tags":"<lua><machine-learning><torch>","CreationDate":"2016-03-30T17:34:46.177","AcceptedAnswerId":36342467.0,"OwnerUserId":4341948.0,"Title":"Testing a single image with torch7 model","Body":"<p>I trained my model based on the link <a href=\"http:\/\/torch.ch\/blog\/2015\/07\/30\/cifar.html\" rel=\"nofollow\">here<\/a>. It achieved almost  accuracy when i trained it. I am using the  model that you will find on the link. But the problem is, i don't know how to test it for a single image.<\/p>\n\n<p>I know how to test a model. By forward passing the image through the network.<\/p>\n\n<p>So, testing the model will require . Where the  is the name of the model that i trained and  is used to forward the model and 'image' is the image location that i want to forward. Now, I cannot figure what the dimension of the single image will be from this network.<\/p>\n\n<p>So,\nwhat i want to do is, take an image. Say the image has dimension [3x32x32]. Pass it through the network and get the result. Is it possible with this network? <\/p>\n\n<p>There was no documentation there how to test it for a single image.<\/p>\n\n<p>What i tried so far was,<\/p>\n\n<p>1) \nDeclare a tensor of size (3x32x32). Lets call it image. `image = torch.Tensor(3x32x32).\nforward pass this.<\/p>\n\n<p><\/p>\n\n<p>It produces the error <\/p>\n\n<p>2) I reshaped the image to (1,3,32,32) <\/p>\n\n<p>\nforward pass this<\/p>\n\n<p><\/p>\n\n<p>It produces the error\n<\/p>\n\n<p>So I tried out approaches. But could not figure out how to pass a single image to that network. Can you help me out?<\/p>\n\n<p>The model definition is<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36342467","CreationDate":"2016-03-31T19:50:41.477","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p>Well, the error is clear: <code>nn.BatchNormalization<\/code> expects a 2D tensor as an input (a batch), but receives a 1D tensor. You added batch dimension to your input (<code>image:reshape(1,3,32,32)<\/code>), but passing through your network, the dimension was lost. <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/simple.md#view\" rel=\"nofollow\"><code>nn.View<\/code><\/a> module is guilty of this. <\/p>\n\n<p>Suppose the module was instantiated with the following parameter:<\/p>\n\n<pre><code>output_size = channels*height*width    -- (512 in your case)\nview = nn.View(output_size)\n<\/code><\/pre>\n\n<p>and it is given an input tensor of shape <code>batch_size x channels x height x width<\/code> (1x512x1x1). \nThe module now has to decide whether it is expected to return a batch or a single non-batch output. <\/p>\n\n<ol>\n<li>If <code>batch_size<\/code> > 1, the answer is obvious: <code>batch_size*channels*height*width<\/code> is a multiple of <code>output_size<\/code> => the input is a batch => the output must be a batch.<\/li>\n<li>If <code>batch_size<\/code> == 1, what then? <code>1*channels*height*width<\/code> == <code>output_size<\/code>, is the input a batch or not? <code>nn.View<\/code> assumes it's not and produces a single output (without batch dimension).<\/li>\n<\/ol>\n\n<p>To fix the misunderstanding, one can specify the number <code>NB<\/code> of non-batch dimensions (if input has NB+1 dimensions, it's a batch):<\/p>\n\n<pre><code>view:setNumInputDims(NB)\n<\/code><\/pre>\n\n<p>In light of the above, this will solve your problem:<\/p>\n\n<p>vgg:add(nn.View(512):<strong>setNumInputDims(3)<\/strong>)<\/p>\n"}]}
{"QuestionId":36317664,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2016-03-30T19:12:03.557","AcceptedAnswerId":36318514.0,"OwnerUserId":6136342.0,"Title":"Dot product dimension issue with theano","Body":"<p>This is my first question on stackoverflow. Hopes it fits the expected format.<\/p>\n\n<p>I'm simply trying to make a function that computes a dot product between a matrix and a vector:<\/p>\n\n<pre><\/pre>\n\n<p>When I ask the product to numpy, I have no issue.  returns my expected <\/p>\n\n<p>But when I ask  to Theano, I've got this error: <\/p>\n\n<p>My understanding was that Theano typing expected my  declared as vector to necessarily be in 1D.\nI've therefore tried with  instead. Maybe theano will understand that it has to be a column to be able to compute ? \nNow the result I get is , which is even more of a mystery to me.<\/p>\n\n<p>I've find similar topics around but none with an explanation I could grasp. I realize it must appear obvious for many of you, but I'd appreciate a hand !<\/p>\n","answers":[{"AnswerId":"36318514","CreationDate":"2016-03-30T19:56:58.090","ParentId":null,"OwnerUserId":"3929051","Title":null,"Body":"<p>When you type the following <\/p>\n\n<pre><code>b = numpy.array([1,0])\n<\/code><\/pre>\n\n<p>it internally creates is a column matrix of shape Rx1. That's how vectors are represented generally. <\/p>\n\n<p>Now when you type in, <\/p>\n\n<pre><code>b = numpy.array([[1],[0]])\n<\/code><\/pre>\n\n<p>And then pass it to function f, it throws an error because it is expecting a Vector and you are passing a Matrix as the second parameter. Change B to type matrix, and it works fine. <\/p>\n\n<p>So, either do this :-<\/p>\n\n<pre><code>A = T.matrix() \nB = T.matrix()\nC = T.dot(A,B)\n\na = numpy.array([[1,2],[3,4]])\nb = numpy.array([[1], [0]])\n\nf = function([A,B],C)\n\nprint f(a,b)\n<\/code><\/pre>\n\n<p>Or you can treat B as a vector itself, and pass a vector which would be <\/p>\n\n<pre><code>A = T.matrix()\nB = T.vector()\nC = T.dot(A,B)\n\na = numpy.array([[1,2],[3,4]])\nb = numpy.array([1,0])\n\nf = function([A,B],C)\n\nprint f(a,b)\n<\/code><\/pre>\n\n<p>Hope this helped!<\/p>\n"}]}
{"QuestionId":36318362,"AnswerCount":1,"Tags":"<python><performance><neural-network><theano><lasagne>","CreationDate":"2016-03-30T19:48:40.883","AcceptedAnswerId":36429758.0,"OwnerUserId":5170255.0,"Title":"Simple lasagne network output is very slow","Body":"<p>I'm trying to train an extremely simple neural network with Lasagne: one dense layer with one output, without nonlinearity (so it's simply a linear regression). Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not training a network here yet, just doing 1K evals (with initial random weights) to see how much time it will take to get 1K actual predictions of my network. Comparing to 1K dot products it's a terrible slowdown!<\/p>\n\n<pre><\/pre>\n\n<p>So my question is: why it takes so much time to evaluate such simple network?<\/p>\n\n<p>Also, I'm confused about the predicted value. If the dot product is less \nthan zero, the network outputs 0, otherwise these two values are the same:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing something about how DenseLayer works?<\/p>\n","answers":[{"AnswerId":"36429758","CreationDate":"2016-04-05T14:48:48.247","ParentId":null,"OwnerUserId":"5170255","Title":null,"Body":"<p>Thanks to Jan Schlueter on <a href=\"https:\/\/groups.google.com\/forum\/#!forum\/lasagne-users\" rel=\"nofollow\">https:\/\/groups.google.com\/forum\/#!forum\/lasagne-users<\/a> there is an answer to this.<\/p>\n\n<p>Here I have not only done 1K passes through the network, but compiled 1K different functions and call each of them once. Instead of using eval() on 1K different expressions (they are different because they each include a different numpy array as a constant), I should have compiled a single prediction function (similar to train_fn, but returning the prediction instead of returning the loss and performing updates) and call that one 1K times in a loop.<\/p>\n\n<p>Question about DenseLayer also solved:<\/p>\n\n<p>The DenseLayer includes a nonlinearity, which defaults to the rectifier. The rectifier sets all outputs smaller than zero to zero.<\/p>\n\n<p>It seems that all the Lasagne questions are much more likely to be answered on googlegroups rather than on StackOverflow. According to Jan, they're more focused on the mailing list.<\/p>\n"}]}
{"QuestionId":36319947,"AnswerCount":1,"Tags":"<python><multithreading><queue><tensorflow>","CreationDate":"2016-03-30T21:18:58.953","AcceptedAnswerId":36320123.0,"OwnerUserId":1322301.0,"Title":"Does Tensorflow's QueueBase.enqueue_many preserve order across threads?","Body":"<p>Let's say two threads are simultaneously trying to enqueue N tensors each, into an instance of a FIFOQueue. I.e., they're calling<\/p>\n\n<p> where  is a list of tensors of length . <\/p>\n\n<p>Let's label each tensor from the first thread as  to , and  to  for the second thread. When all is said and done (both calls to  have completed), will the order be preserved? I.e., will the queue contain either  or ? Or could the tensors be enqueued in an interleaved manner, i.e., ? Or, I suppose, a third option is that there is no rhyme or reason whatsoever: the tensors are enqueued in an arbitrary order.<\/p>\n","answers":[{"AnswerId":"36320123","CreationDate":"2016-03-30T21:31:31.147","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The order within the batches enqueued by each thread is preserved. Another way of putting this is that <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/io_ops.html#QueueBase.enqueue_many\" rel=\"nofollow\"><code>FIFOQueue.enqueue_many()<\/code><\/a> executes atomically* with respect to other enqueue operations on the same queue.<\/p>\n\n<hr>\n\n<p>*&nbsp;However, if the remaining capacity of the queue is less than the size of the batch being enqueued, it is possible to <em>dequeue<\/em> elements from the batch before the enqueue completes&mdash;i.e. you are able to enqueue batches that are <em>larger<\/em> than the capacity of the queue, and the enqueue op will block until enough elements have been dequeued.<\/p>\n"}]}
{"QuestionId":36321395,"AnswerCount":1,"Tags":"<neural-network><classification><deep-learning><caffe><pycaffe>","CreationDate":"2016-03-30T23:14:07.800","AcceptedAnswerId":36325285.0,"OwnerUserId":2377388.0,"Title":"How to use more generalized lables when using Caffe for classification","Body":"<p>I am using a pre-trained AlexNet in Caffe for classification. However, the output from classification is around 1000 classes from the  file. Is there a way to use more generic classes?<\/p>\n\n<p>For example instead of having as output the type of dog in the picture, i just want the output to be \"dog\".<\/p>\n\n<p>I know that there's a hierarchy in the classes, but is there a way to retrieve that hierarchy and use it to get more generic labels?<\/p>\n\n<p>thanks :)<\/p>\n","answers":[{"AnswerId":"36325285","CreationDate":"2016-03-31T05:58:53.527","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Quoting from <a href=\"http:\/\/image-net.org\/download-API\" rel=\"nofollow\">image-net API documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>WordNet Hierarchy<\/strong><br>\n  The <a href=\"http:\/\/www.image-net.org\/archive\/wordnet.is_a.txt\" rel=\"nofollow\">following file<\/a> provides all \"is-a\" relations provided by WordNet 3.0. Each line is a parent-child pair.<\/p>\n<\/blockquote>\n\n<p>You need to download the <code>'wordnet.ia_a.txt'<\/code> file, parse it and build the wordnet hierarchy.<\/p>\n"}]}
{"QuestionId":36321400,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning><conv-neural-network>","CreationDate":"2016-03-30T23:15:12.733","AcceptedAnswerId":36323861.0,"OwnerUserId":5257450.0,"Title":"Strange TypeError with Theano","Body":"<pre><\/pre>\n\n<p>And here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>It is really weird as the error was not thrown in Jupyter (but it takes long long time to run and finally the kernel is down I really don't know why), but as I move it to the shell and run  the error was thrown.<\/p>\n","answers":[{"AnswerId":"36323861","CreationDate":"2016-03-31T03:56:45.553","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>The problem lies in <code>loadmat<\/code> from <code>scipy<\/code>. The typeerror you are getting is thrown by this code in Theano:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if not data.flags.aligned:\n    ...\n    raise TypeError(...)\n<\/code><\/pre>\n\n<p>Now, when you create a new array in numpy from raw data, it would usually be aligned:<\/p>\n\n<pre><code>&gt;&gt;&gt; a = np.array(2)\n&gt;&gt;&gt; a.flags.aligned\nTrue\n<\/code><\/pre>\n\n<p>But if you <code>savemat<\/code> \/ <code>loadmat<\/code> it, the value of the flag gets lost:<\/p>\n\n<pre><code>&gt;&gt;&gt; savemat('test', {'a':a})\n&gt;&gt;&gt; a2 = loadmat('test')['a']\n&gt;&gt;&gt; a2.flags.aligned\nFalse\n<\/code><\/pre>\n\n<p>(seems like this particular issue is discussed <a href=\"https:\/\/mail.scipy.org\/pipermail\/scipy-user\/2009-December\/023646.html\" rel=\"nofollow\">here<\/a>)<\/p>\n\n<p>One quick and dirty way to address it is to create a new numpy array from the array you loaded:<\/p>\n\n<pre><code>&gt;&gt;&gt; a2 = loadmat('test')['a']\n&gt;&gt;&gt; a3 = np.array(a2)\n&gt;&gt;&gt; a3.flags.aligned\nTrue\n<\/code><\/pre>\n\n<p>So, for your code:<\/p>\n\n<pre><code>dataReadyForCNN = np.array(sio.loadmat(\"DataReadyForCNN.mat\"))\n<\/code><\/pre>\n"}]}
{"QuestionId":36321511,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2016-03-30T23:24:32.787","AcceptedAnswerId":null,"OwnerUserId":1305150.0,"Title":"In torch trepl on mac, the tab completion is not working. It works fine on the linux machine I have.","Body":"<p>In torch trepl on mac, the tab completion is not working. It works fine on the linux machine I have. Any idea what could went wrong?<\/p>\n","answers":[]}
{"QuestionId":36321955,"AnswerCount":2,"Tags":"<pycharm><tensorflow><skflow>","CreationDate":"2016-03-31T00:06:12.030","AcceptedAnswerId":null,"OwnerUserId":3147590.0,"Title":"Why Pycharm cannot recognize functions of Skflow?","Body":"<p>I install Skflow and ran digits.py example on Pycharm and saw that it returns an error \"AttributeError: 'module' object has no attribute 'TensorFlowDNNRegressor\". I proceed and ran the same program on Ipython and everything's fine. What should be the problem?<\/p>\n\n<pre><\/pre>\n\n<p>Furthermore, I learned that I have problem with any functions of Skflow on Pycharm when it works on Ipython very well. Any speculation on this?<\/p>\n","answers":[{"AnswerId":"36341205","CreationDate":"2016-03-31T18:41:53.423","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>You might check to ensure that Pycharm is using the the same python interpreter and environment as used by ipython. You can set the interpreter in pycharm in settings|Project|Project Interpreter<\/p>\n"},{"AnswerId":"36380758","CreationDate":"2016-04-03T01:47:15.240","ParentId":null,"OwnerUserId":"5927529","Title":null,"Body":"<p>Is TensorFlow working well in Pycharm? If so, you can now install a <strong>nightly built TensorFlow<\/strong> with skflow in its contrib module. <\/p>\n\n<p>To install a nightly built TensorFlow, see the README file on TensorFlow's Github page for instructions. <\/p>\n\n<p>and then you can import skflow via <code>from tensorflow.contrib.skflow.python import skflow<\/code>.<\/p>\n\n<p>Hope this helps. <\/p>\n"}]}
{"QuestionId":36322579,"AnswerCount":1,"Tags":"<ubuntu><pycharm><anaconda><theano><lasagne>","CreationDate":"2016-03-31T01:21:47.820","AcceptedAnswerId":36322706.0,"OwnerUserId":5269749.0,"Title":"add all my python path to pycharm","Body":"<p>I'm using PyCharm in Ubuntu 14.04.\nSo when I use iPython I can see the libraries that I am using by:<\/p>\n\n<pre><\/pre>\n\n<p>results:<\/p>\n\n<pre><\/pre>\n\n<p>I can import the libraries in iPython but they do not work in PyCharm.<\/p>\n\n<p>How can I insert all of these paths and libraries into PyCharm.\nRight now my PyCharm does not recognize most of these libraries.<\/p>\n\n<p>If you are familiar with deep learning libraries, I am using Anaconda, Theano, Lasagne, but none of them are added to my PyCharm. however they all work fine in iPython.<\/p>\n\n<p>In other words, how I can add all the paths that I have in my iPython to my PyCharm path.<\/p>\n","answers":[{"AnswerId":"36322706","CreationDate":"2016-03-31T01:40:35.560","ParentId":null,"OwnerUserId":"4891502","Title":null,"Body":"<p>PyCharm includes a package management interface which can be found in <strong>File > Settings > Project Interpreter<\/strong> if you use a packaging tool like <a href=\"https:\/\/pypi.python.org\/pypi\" rel=\"nofollow noreferrer\">PyPi<\/a>. If it is the case that you do not have a package manager installed, PyCharm will prompt you to install one at this screen:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1xa9M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1xa9M.png\" alt=\"Install package\"><\/a><\/p>\n\n<p>Then you can install the packages you want locally:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5cpfr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5cpfr.png\" alt=\"Choose to install\"><\/a><\/p>\n\n<p>Search for the correct packages, click \"Install to user's site packages\", and select the version you would like to use before clicking \"Install Package\":<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/boCxF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/boCxF.png\" alt=\"Install Package\"><\/a><\/p>\n\n<p><strong>Edit<\/strong>: I have added a more detailed description of PyCharm's built-in package management. I find it useful to do all this work inside PyCharm, but it will be a little different than how you currently set up your project if you are not putting all the libraries into the preconfigured python paths.<\/p>\n"}]}
{"QuestionId":36324146,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-31T04:25:53.140","AcceptedAnswerId":36329872.0,"OwnerUserId":1001930.0,"Title":"Tensorflow: Placeholder Size Issue in Tutorial","Body":"<p>I'm following <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/pros\/index.html#deep-mnist-for-experts\" rel=\"nofollow\">Tensorflow's Deep MNIST tutorial<\/a>. When I ran the following line<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>It seem unhappy with how I'm feeding . I set  to be the following placeholder:<\/p>\n\n<pre><\/pre>\n\n<p>I checked that  is a float32 numpy array of size 50x748. What's the proper way of feeding an array to ?<\/p>\n","answers":[{"AnswerId":"36329872","CreationDate":"2016-03-31T09:49:30.483","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>This is just a typo in your placeholder, notice<\/p>\n\n<pre><code>Cannot feed value of shape (50, 784) for Tensor 'Placeholder:0', which has shape '(?, 748)'\n<\/code><\/pre>\n\n<p>It's shape 784 in the feed, while the placeholder is 748, so you just need to change the placeholder to<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=[None, 784]) \n<\/code><\/pre>\n\n<p>Also don't worry this happens to everyone :)<\/p>\n"}]}
{"QuestionId":36330670,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-03-31T10:22:30.360","AcceptedAnswerId":null,"OwnerUserId":6138987.0,"Title":"Feeding data through an embedding wrapper in TensorFlow","Body":"<p>I'm working on a text summarization network and need to implement an\nencoder to use with . As part of that I need to encode varying batches of sequences into representing vectors but the innermost encoding doesn't go through. <\/p>\n\n<p>Here's a simplified snippet giving the same error:<\/p>\n\n<pre><\/pre>\n\n<p>This fails with the following stack trace:<\/p>\n\n<pre><\/pre>\n\n<p>When debugging the input size of the  is  and the elements in  all have dimension  which is infact .  <\/p>\n\n<p>Switching to  in the call to  makes the snippet works but give me the following stack trace in my code:<\/p>\n\n<pre><\/pre>\n\n<p>What am I missing?<\/p>\n","answers":[]}
{"QuestionId":36331419,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-03-31T10:56:52.423","AcceptedAnswerId":43189934.0,"OwnerUserId":234167.0,"Title":"TensorFlow: How to measure how much GPU memory each tensor takes?","Body":"<p>I'm currently implementing <a href=\"http:\/\/pjreddie.com\/media\/files\/papers\/yolo.pdf\" rel=\"noreferrer\">YOLO<\/a> in TensorFlow and I'm a little surprised on how much memory that is taking. On my GPU I can train YOLO using their <a href=\"http:\/\/pjreddie.com\/darknet\/\" rel=\"noreferrer\">Darknet<\/a> framework with batch size 64. On TensorFlow I can only do it with batch size 6, with 8 I already run out of memory. For the test phase I can run with batch size 64 without running out of memory.<\/p>\n\n<ol>\n<li><p>I am wondering how I can calculate how much memory is being consumed by each tensor? Are all tensors by default saved in the GPU? Can I simply calculate the total memory consumption as the shape * 32 bits?<\/p><\/li>\n<li><p>I noticed that since I'm using momentum, all my tensors also have a  tensor. Could that also be using a lot of memory?<\/p><\/li>\n<li><p>I am augmenting my dataset with a method , very similar to the one defined in the <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/image\/cifar10\/cifar10.py\" rel=\"noreferrer\">CIFAR-10 tutorial<\/a>. Could it be that this part is occupying a huge chunk of memory? I believe Darknet does the modifications in the CPU.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"43189934","CreationDate":"2017-04-03T16:40:05.143","ParentId":null,"OwnerUserId":"599139","Title":null,"Body":"<p>Now that <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1258\" rel=\"noreferrer\">1258<\/a> has been closed, you can enable memory logging in Python by setting an environment variable before importing TensorFlow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nos.environ['TF_CPP_MIN_VLOG_LEVEL']='3'\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>There will be a <em>lot<\/em> of logging as a result of this. You'll want to grep the results to find the appropriate lines. For example:<\/p>\n\n<pre><code>grep MemoryLogTensorAllocation train.log\n<\/code><\/pre>\n"},{"AnswerId":"36335064","CreationDate":"2016-03-31T13:36:12.617","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>See the description in this (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/ec1403e7dc2b919531e527d36d28659f60621c9e\" rel=\"nofollow\">commit<\/a>).\nThe memory allocation is raw info is there although it needs a script to collect the information in an easy to read form.<\/p>\n"},{"AnswerId":"36505898","CreationDate":"2016-04-08T17:32:56.050","ParentId":null,"OwnerUserId":"5823391","Title":null,"Body":"<p>Sorry for the slow reply. Unfortunately right now the only way to set the log level is to edit tensorflow\/core\/platform\/logging.h and recompile with e.g.<\/p>\n\n<pre><code>#define VLOG_IS_ON(lvl) ((lvl) &lt;= 1)\n<\/code><\/pre>\n\n<p>There is a bug open <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1258\" rel=\"noreferrer\" title=\"1258\">1258<\/a> to control logging more elegantly.<\/p>\n\n<p><code>MemoryLogTensorOutput<\/code> entries are logged at the end of each Op execution, and indicate the tensors that hold the outputs of the Op. It's useful to know these tensors since the memory is not released until the downstream Op consumes the tensors, which may be much later on in a large graph.<\/p>\n"}]}
{"QuestionId":36334371,"AnswerCount":3,"Tags":"<python><multithreading><queue><tensorflow>","CreationDate":"2016-03-31T13:05:35.167","AcceptedAnswerId":36334919.0,"OwnerUserId":612911.0,"Title":"Tensorflow: Batching input queues then changing the queue source","Body":"<p>I have a model that runs over a set of images and calculates some stats on them using - for simplicity say it just outputs the average image from the set (it does more than that in practice). I have a number of directories that contain images and I want to get the output from each directory. Each directory has a variable number of images in it.<\/p>\n\n<p>I have constructed the graph, output variable and loss function once for my script. The inputs are batched using slightly adapted <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_input.py\" rel=\"nofollow noreferrer\">code from here<\/a>. I adapted it to take an array of paths which I feed in using a placeholder of variable size. I got the inspiration for that <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/reading_data\/index.html#preloaded-data\" rel=\"nofollow noreferrer\">from here<\/a>.<\/p>\n\n<p>Then I loop over the directories and run the following:<\/p>\n\n<ol>\n<li>Initialize the variables (this resets the previous output variable from the results calculated on the previous directory)<\/li>\n<li>Set the image paths variable to the current files array from the new directory: <\/li>\n<li>Start the queues running: <\/li>\n<li>Run for a number of epochs to get results<\/li>\n<li>Close threads <\/li>\n<li>Return results, save results, move onto the next directory...<\/li>\n<\/ol>\n\n<p>The problem is that when it comes to the second directory the queue runner threads refuse to start (I can see this by debugging the  variable). This gives errors like:<\/p>\n\n<pre><\/pre>\n\n<p>If I don't close the threads (and don't start them a second time) then they don't produce files from the new directory - they ignore the variable assignment op in (2). Is it just not possible to restart the queues like this?<\/p>\n\n<p>I have tried setting up the queues in their own separate session and pulling the batches from them, but that gives me various CUDA \/ memory errors. If I do this and add debug stops I can get it to run quite far before it hits this - but I don't know if it is possible to add control dependencies between disjoint sessions\/graphs? <\/p>\n\n<p>It would be possible to start from scratch for each new directory but that adds a lot of overhead to the process that I am trying to avoid. I have done similar things to this (ie, resetting variables and rerunning with different inputs) without queues and it saves a lot of time, so I know that bit works.<\/p>\n\n<p>Can any of you wonderful SO folk think of a way out of this?<\/p>\n","answers":[{"AnswerId":"36334919","CreationDate":"2016-03-31T13:29:42.673","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><code>string_input_producer<\/code> is a <code>FIFOQueue<\/code> + <code>QueueRunner<\/code>. You get more control if you use a <code>FIFOQueue<\/code> and enqueue things manually. Something like this<\/p>\n\n<pre><code>filename_queue = tf.FIFOQueue(100, tf.string)\nenqueue_placeholder = tf.placeholder(dtype=tf.string)\nenqueue_op = filename_queue.enqueue(enqueue_placeholder)\n\nconfig = tf.ConfigProto()\nconfig.operation_timeout_in_ms=2000  # for debugging queue hangs\nsess = tf.InteractiveSession(config=config)\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(coord=coord)\n\nsess.run([enqueue_op], feed_dict={enqueue_placeholder:\"\/temp\/dir1\/0\"})\nsess.run([enqueue_op], feed_dict={enqueue_placeholder:\"\/temp\/dir1\/1\"})\n\n# do stats for \/temp\/dir1\n\nsess.run([enqueue_op], feed_dict={enqueue_placeholder:\"\/temp\/dir2\/0\"})\nsess.run([enqueue_op], feed_dict={enqueue_placeholder:\"\/temp\/dir2\/1\"})\n\n# do stats for \/temp\/dir2\n\ncoord.request_stop()\ncoord.join(threads)\n<\/code><\/pre>\n"},{"AnswerId":"36355267","CreationDate":"2016-04-01T11:32:18.777","ParentId":null,"OwnerUserId":"612911","Title":null,"Body":"<p>Big thanks to @yaroslav-bulatov for pointing me in the right direction on this. <\/p>\n\n<p>It seems that my biggest problem was with the queue runners. When I replaced the filename queue with a <code>FIFOQueue<\/code> and manually enqueued filenames to it that worked fine, but because I was also using a <code>shuffle_batch<\/code> queue this got upset when I tried to empty the filenames queue for the next directory. I could not empty this queue either as it causes lockups or breaks the queue, so the best I could manage was to let it fill up with the new images while keeping the leftovers from the previous directory - obviously no good! In the end I replaced that with a <code>RandomShuffleQueue<\/code> and again enqueue items manually in the same way as with filenames. I think this is giving good-enough mixing of images and isn't overkill for the problem. No threading, but as soon as I did away with that things got a lot simpler.<\/p>\n\n<p>I've included my final solution as below. Any suggestions welcome!<\/p>\n\n<pre><code>import os\nimport tensorflow as tf\nimport numpy as np\nfrom itertools import cycle\n\noutput_dir = '\/my\/output\/dir'\n\nmy_dirs = [\n    [\n        '\/path\/to\/datasets\/blacksquares\/black_square_100x100.png',\n        '\/path\/to\/datasets\/blacksquares\/black_square_200x200.png',\n        '\/path\/to\/datasets\/blacksquares\/black_square_300x300.png'\n    ],\n    [\n        '\/path\/to\/datasets\/whitesquares\/white_square_100x100.png',\n        '\/path\/to\/datasets\/whitesquares\/white_square_200x200.png',\n        '\/path\/to\/datasets\/whitesquares\/white_square_300x300.png',\n        '\/path\/to\/datasets\/whitesquares\/white_square_400x400.png'\n    ],\n    [\n        '\/path\/to\/datasets\/mixedsquares\/black_square_200x200.png',\n        '\/path\/to\/datasets\/mixedsquares\/white_square_200x200.png'\n    ]\n]\n\n# set vars\npatch_size = (100, 100, 1)\nbatch_size = 20\nqueue_capacity = 1000\n\n# setup filename queue\nfilename_queue = tf.FIFOQueue(\n    capacity=queue_capacity,\n    dtypes=tf.string,\n    shapes=[[]]\n)\nfilenames_placeholder = tf.placeholder(dtype='string', shape=(None))\nfilenames_enqueue_op = filename_queue.enqueue_many(filenames_placeholder)\n\n# read file and preprocess\nimage_reader = tf.WholeFileReader()\nkey, file = image_reader.read(filename_queue)\nuint8image = tf.image.decode_png(file)\ncropped_image = tf.random_crop(uint8image, patch_size) # take a random 100x100 crop\nfloat_image = tf.div(tf.cast(cropped_image, tf.float32), 255) # put pixels in the [0,1] range\n\n# setup shuffle batch queue for training images\nimages_queue = tf.RandomShuffleQueue(\n    capacity=queue_capacity,\n    min_after_dequeue=0,  # allow queue to become completely empty (as we need to empty it)\n    dtypes=tf.float32,\n    shapes=patch_size\n)\nimages_enqueue_op = images_queue.enqueue(float_image)\n\n# setup simple computation - calculate an average image patch\ninput = tf.placeholder(shape=(None,) + patch_size, dtype=tf.float32)\navg_image = tf.Variable(np.random.normal(loc=0.5, scale=0.5, size=patch_size).astype(np.float32))\nloss = tf.nn.l2_loss(tf.sub(avg_image, input))\ntrain_op = tf.train.AdamOptimizer(2.).minimize(loss)\n\n# start session and initialize variables\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\n\n# note - no need to start any queue runners as I've done away with them\n\nfor dir_index, image_paths in enumerate(my_dirs):\n    image_paths_cycle = cycle(image_paths)\n\n    # reset the optimisation and training vars\n    sess.run(tf.initialize_all_variables())\n\n    num_epochs = 1000\n    for i in range(num_epochs):\n        # keep the filename queue at capacity\n        size = sess.run(filename_queue.size())\n        image_paths = []\n        while size &lt; queue_capacity:\n            image_paths.append(next(image_paths_cycle))\n            size += 1\n        sess.run(filenames_enqueue_op, feed_dict={filenames_placeholder: image_paths})\n\n        # keep the shuffle batch queue at capacity\n        size = sess.run(images_queue.size())\n        while size &lt; queue_capacity:\n            sess.run([images_enqueue_op])\n            size += 1\n\n        # get next (random) batch of training images\n        batch = images_queue.dequeue_many(batch_size).eval()\n\n        # run train op\n        _, result, loss_i = sess.run([train_op, avg_image, loss], feed_dict={input: batch})\n        print('Iteration {:d}. Loss: {:.2f}'.format(i, loss_i))\n\n        # early stopping :)\n        if loss_i &lt; 0.05:\n            break\n\n    # empty filename queue and verify empty\n    size = sess.run(filename_queue.size())\n    sess.run(filename_queue.dequeue_many(size))\n    size = sess.run(filename_queue.size())\n    assert size == 0\n\n    # empty batch queue and verify empty\n    size = sess.run(images_queue.size())\n    sess.run(images_queue.dequeue_many(size))\n    size = sess.run(filename_queue.size())\n    assert size == 0\n\n    # save the average image output\n    result_image = np.clip(result * 255, 0, 255).astype(np.uint8)\n    with open(os.path.join(output_dir, 'result_' + str(dir_index)), 'wb') as result_file:\n        result_file.write(tf.image.encode_png(result_image).eval())\n\nprint('Happy days!')\nexit(0)\n<\/code><\/pre>\n\n<p>This ouputs <code>result_0<\/code> - a black square, <code>result_1<\/code> - a white square and <code>result_2<\/code> - a (mostly) grey square. <\/p>\n"},{"AnswerId":"44869035","CreationDate":"2017-07-02T08:52:36.357","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Looking at your own answer, I would suggest you to take a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/match_filenames_once\" rel=\"nofollow noreferrer\"><code>tf.train.match_filenames_once<\/code><\/a>. It allows you to specify the pattern for your files and pass it to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/string_input_producer\" rel=\"nofollow noreferrer\">string_input_producer<\/a>.<\/p>\n\n<pre><code>q = tf.train.string_input_producer(\n  tf.train.match_filenames_once('\/path\/to\/datasets\/*.png')\n)\n<\/code><\/pre>\n\n<p>Here is a more <a href=\"https:\/\/gist.github.com\/eerwitt\/518b0c9564e500b4b50f\" rel=\"nofollow noreferrer\">comprehensive example<\/a>. By doing this not only you will fix your problem, but also will not need to iterate over all your directories.<\/p>\n"}]}
{"QuestionId":36337736,"AnswerCount":0,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-03-31T15:32:36.333","AcceptedAnswerId":null,"OwnerUserId":4229825.0,"Title":"Extremely low accuracy on own data in caffe","Body":"<p>I'm trying to train a network on my own data. Whole dataset consists of 256x256 jpeg images. There is 236 objects for classification. Training and validation sets have ~247K and ~61K images, respectively. I've made LMDBs from them using  utility. <\/p>\n\n<p>Just for starting I'm using caffenet's topology for my model. During training I come across the weird message \"Data layer prefetch queue empty\" that I never seen before.<\/p>\n\n<p>Moreover, initially, network has an abnormal accuracy (~0.00378378) and during next 1000 iterations, it reaches max ~0.01 and further does not increase (just fluctuates).<\/p>\n\n<p>What I'm doing wrong and how can I improve the accuracy?<\/p>\n\n<p>Runtime log:\n<a href=\"http:\/\/paste.ubuntu.com\/15568421\/\" rel=\"nofollow\">http:\/\/paste.ubuntu.com\/15568421\/<\/a><\/p>\n\n<p>Model:\n<a href=\"http:\/\/paste.ubuntu.com\/15568426\/\" rel=\"nofollow\">http:\/\/paste.ubuntu.com\/15568426\/<\/a><\/p>\n\n<p>Solver:\n<a href=\"http:\/\/paste.ubuntu.com\/15568430\/\" rel=\"nofollow\">http:\/\/paste.ubuntu.com\/15568430\/<\/a><\/p>\n\n<p>P.S. I'm using the latest version of Caffe, Ubuntu Server 14.04 LTS and g2.2xlarge instance on AWS.<\/p>\n","answers":[]}
{"QuestionId":36339059,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-03-31T16:37:07.800","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"Exporter classification_signature","Body":"<p>I'm trying to modify the <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow\">serving tutorial<\/a> to work with my model, which is basically the CIFAR example modified to work with a CSV file and JPEGs. I can't seem to find the documentation for the Exporter class, but here is what I have so far. It's in the train() function in the cifar10_train.py file:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the code I use to train the model:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas how I can set up Exporter correctly?<\/p>\n","answers":[{"AnswerId":"37036108","CreationDate":"2016-05-04T19:12:22.247","ParentId":null,"OwnerUserId":"6291805","Title":null,"Body":"<p>Please take a look at the <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/example\/mnist_export.py\" rel=\"nofollow\">MNIST export example<\/a>.<\/p>\n\n<p>That shows how x and y are generated then placed in the signature.<\/p>\n\n<p>Also, the <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/example\/inception_export.py\" rel=\"nofollow\">Inception example<\/a> shows how to extend an existing model to create exports and serving.  In particular the <code>cifar10.inference<\/code> call looks similar to <code>inception_model.inference<\/code>.<\/p>\n"}]}
{"QuestionId":36341473,"AnswerCount":2,"Tags":"<gpu><tensorflow>","CreationDate":"2016-03-31T18:57:26.497","AcceptedAnswerId":36346815.0,"OwnerUserId":3444294.0,"Title":"Recommended GPUs for Tensorflow","Body":"<p>I understand that Tensorflow requires (for GPU computation) a GPU with Nvidia Compute Capability >= 3.0. There are many such GPUs to choose from. The gaming oriented GPUs, e.g. GeForce models, are much less expensive than the compute-oriented models, e.g. Tesla. My limited undertanding is that the compute-oriented models may lack video output (not needed for computation) and that the gaming models may be doing 32-bit math instead of 64. Assuming that Tensorflow uses (or prefers) 64-bit, does this mean that the gaming models will not work or will produce deficient results if used with Tensorflow? What attributes should one look for in choosing a GPU to use with Tensorflow?<\/p>\n","answers":[{"AnswerId":"43777343","CreationDate":"2017-05-04T08:12:44.530","ParentId":null,"OwnerUserId":"7941008","Title":null,"Body":"<p>The GPU-enabled version of TensorFlow has the following requirements:<\/p>\n\n<ul>\n<li>64-bit Linux<\/li>\n<li>Python 2.7<\/li>\n<li>NVIDIA CUDA\u00ae 7.5 (CUDA 8.0 required for Pascal GPUs)<\/li>\n<li>NVIDIA cuDNN v4.0 (minimum) or v5.1 (recommended)<\/li>\n<\/ul>\n\n<p>TensorFlow GPU support requires having a GPU card with NVidia Compute Capability >= 3.0. Supported cards include but are not limited to:<\/p>\n\n<ul>\n<li>NVidia Titan<\/li>\n<li>NVidia Titan X<\/li>\n<li>NVidia K20<\/li>\n<li>NVidia K40<\/li>\n<\/ul>\n\n<p>You can see their official docs <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/get_started\/os_setup\" rel=\"noreferrer\">Tensorflow GPU support<\/a><\/p>\n"},{"AnswerId":"36346815","CreationDate":"2016-04-01T02:00:56.270","ParentId":null,"OwnerUserId":"5732171","Title":null,"Body":"<p>Gaming GPUs can work quite well. You want a very recent GPU with lots of memory and CUDA cores. Most people training neural nets these days on GPU use 32 bit floats.<\/p>\n"}]}
{"QuestionId":36343199,"AnswerCount":2,"Tags":"<indexing><torch>","CreationDate":"2016-03-31T20:32:08.880","AcceptedAnswerId":null,"OwnerUserId":4395366.0,"Title":"torch logical indexing of tensor","Body":"<p>I looking for an elegant way to select a subset of a torch tensor which satisfies some constrains.\nFor example, say I have:<\/p>\n\n<pre><\/pre>\n\n<p>and  is a 10x1 tensor,<\/p>\n\n<pre><\/pre>\n\n<p>I would like to be able to do logical indexing, as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But that doesn't work. \nSo there's the  function which accepts a  but I could not find a simple way to convert  to a , except the following:<\/p>\n\n<pre><\/pre>\n\n<p>which returns a K x 2 tensor (K being the number of values of S >= 5). So then I have to convert it to a 1 dimensional array, which finally allows me to index A:<\/p>\n\n<pre><\/pre>\n\n<p>This is very cumbersome; in e.g. Matlab all I'd have to do is <\/p>\n\n<pre><\/pre>\n\n<p>Can anyone suggest a better way? <\/p>\n","answers":[{"AnswerId":"43884411","CreationDate":"2017-05-10T05:45:51.717","ParentId":null,"OwnerUserId":"1866656","Title":null,"Body":"<p>There are two simpler alternatives:<\/p>\n\n<ol>\n<li><p>Use <code>maskedSelect<\/code>:<\/p>\n\n<p><code>result=A:maskedSelect(your_byte_tensor)<\/code><\/p><\/li>\n<li><p>Use a simple element-wise multiplication, for example<\/p>\n\n<p><code>result=torch.cmul(A,S:gt(0))<\/code><\/p><\/li>\n<\/ol>\n\n<p>The second one is very useful if you need to keep the shape of the original matrix (<em>i.e<\/em> A), for example to select neurons in a layer at backprop. However, since it puts zeros in the resulting matrix whenever the condition dictated by the ByteTensor doesn't apply, you can't use it to compute the product (or median, etc.). The first one only returns the elements that satisfy the condittion, so this is what I'd use to compute products or medians or any other thing where I don't want zeros.  <\/p>\n"},{"AnswerId":"36350806","CreationDate":"2016-04-01T07:47:19.390","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>One possible alternative is:<\/p>\n\n<pre><code>sel = S:ge(5):expandAs(A)   -- now you can use this mask with the [] operator\nA1 = A[sel]:unfold(1, 2, 2) -- unfold to get back a 2D tensor\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre><code>&gt; A = torch.rand(3,2)-1\n-0.0047 -0.7976\n-0.2653 -0.4582\n-0.9713 -0.9660\n[torch.DoubleTensor of size 3x2]\n\n&gt; S = torch.Tensor{{6}, {1}, {5}}\n 6\n 1\n 5\n[torch.DoubleTensor of size 3x1]\n\n&gt; sel = S:ge(5):expandAs(A)\n1  1\n0  0\n1  1\n[torch.ByteTensor of size 3x2]\n\n&gt; A[sel]\n-0.0047\n-0.7976\n-0.9713\n-0.9660\n[torch.DoubleTensor of size 4]\n\n&gt; A[sel]:unfold(1, 2, 2)\n-0.0047 -0.7976\n-0.9713 -0.9660\n[torch.DoubleTensor of size 2x2]\n<\/code><\/pre>\n"}]}
{"QuestionId":36343542,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-03-31T20:51:41.957","AcceptedAnswerId":36345391.0,"OwnerUserId":6142034.0,"Title":"TensorFlow shape error in feed_dict","Body":"<p>I'm trying to adapt existing logreg example to my data and are getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Source code can be found here: <a href=\"https:\/\/github.com\/ilautar\/tensorflow-test\/blob\/master\/tflin.py\" rel=\"nofollow\">https:\/\/github.com\/ilautar\/tensorflow-test\/blob\/master\/tflin.py<\/a><\/p>\n\n<p>I'm sure it is obvious, any pointers?<\/p>\n\n<p>Thank you,\nIgor<\/p>\n","answers":[{"AnswerId":"36345391","CreationDate":"2016-03-31T23:12:37.617","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The error occurs because you are trying to feed a <code>60000 x 6<\/code> matrix into a <code>tf.placeholder()<\/code> that is defined to be a vector of length 6. This happens when you <a href=\"https:\/\/github.com\/ilautar\/tensorflow-test\/blob\/master\/tflin.py#L80\" rel=\"nofollow\">try to feed the whole <code>train_X<\/code> matrix<\/a> (as opposed to <a href=\"https:\/\/github.com\/ilautar\/tensorflow-test\/blob\/master\/tflin.py#L76\" rel=\"nofollow\">feeding a single row<\/a>, which succeeds).<\/p>\n\n<p>The best way to make this work is to do the following:<\/p>\n\n<ol>\n<li><p>Define your placeholders (and model) in terms of batch inputs, which can have varying shape:<\/p>\n\n<pre><code># tf Graph Input\nX = tf.placeholder(tf.float32, [None, n_input])\nY = tf.placeholder(tf.float32, [None])\n<\/code><\/pre><\/li>\n<li><p>When feeding in a single example, extend it to be a <code>1 x 6<\/code> matrix using <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis\" rel=\"nofollow\"><code>numpy.newaxis<\/code><\/a>:<\/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[numpy.newaxis, ...],\n                                       Y: y[numpy.newaxis, ...]})\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":36344842,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-03-31T22:19:14.150","AcceptedAnswerId":36362192.0,"OwnerUserId":4882455.0,"Title":"Caffe - num_output in prototxt gives strange behaviour","Body":"<p>I am doing some experiments where I split the Cifar-10 dataset into two halves, such that each half contains five random classes. I trained on one half with the  architecture. Therefore, I changed  to  and made some other small tweaks to the network. When I inspect the logfile, I find that the loss increases to around <strong>80<\/strong>, with a test accuracy of <strong>0<\/strong>.<\/p>\n\n<p>However, when I change   to , training seems to go normal i.e. loss steadily decreases, and results in a test accuracy of about <strong>70%<\/strong>.<\/p>\n\n<p>How can this be explained?<\/p>\n\n<p><strong>train_val.prototxt<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>This split contains the classes 0, 4, 5, 6 and 8. I used the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/6eae122a8eb84f8371dde815986cd7524fc4cbaa\/examples\/imagenet\/create_imagenet.sh\" rel=\"nofollow\">create_imagenet.sh<\/a> script to create the lmdb files.<\/p>\n\n<p><strong>Sample of the train.txt<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Sample of the val.txt<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36362192","CreationDate":"2016-04-01T17:17:06.330","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>As was pointed out in the comments, Caffe expects that the labels are integers between 0 and <code>num_classes - 1<\/code>. In your case, as you set number of labels to 5, Caffe will create five output neurons in the last layer. When you ask it to predict class 6 or 8, you are asking it to maximize the output of a non-existent neuron, which Caffe obviously cannot do.<\/p>\n\n<p>Now, when you relabel your data, and set <code>num_classes<\/code> to 5, you do the correct thing, and it, therefore, works. When you set <code>num_classes<\/code> to 10, the network will still work, because now it has 10 output neurons, which is enough to be able to classify five classes. It will learn that classes from 5 to 9 are never present, and therefore should never be predicted, and it will just adjust the weights in a way that will always result in very small values returned by those output neurons. It is important to note, however, that neural networks are naturally random, so it might still occasionally return a class that was never presented to it, so I would expect a NN with <code>num_classes<\/code> larger than the actual number of classes to perform worse than the one with the correct <code>num_classes<\/code>.<\/p>\n"}]}
{"QuestionId":36345633,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2016-03-31T23:37:14.653","AcceptedAnswerId":36346159.0,"OwnerUserId":2241766.0,"Title":"GrpcServer class not found when using distributed TensorFlow","Body":"<p>I'm trying to test the distributed computing features of TensorFlow with the following simple commands:<\/p>\n\n<pre><\/pre>\n\n<p>However, as you can see above, it throws an error of indicating that  can't be found. I've installed TensorFlow into a \"Virtualenv\" and the MNIST example works fine. How can I test the distributed capability of TensorFlow?<\/p>\n","answers":[{"AnswerId":"36346159","CreationDate":"2016-04-01T00:39:04.747","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The class <code>tf.GrpcServer<\/code> is only available if you install a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#installation\" rel=\"nofollow\">nightly build<\/a> of TensorFlow, or build from source. It will be included in the next release (version 0.8).<\/p>\n"}]}
{"QuestionId":36347908,"AnswerCount":0,"Tags":"<python><arrays><audio><mp3><tensorflow>","CreationDate":"2016-04-01T04:12:33.217","AcceptedAnswerId":null,"OwnerUserId":5449456.0,"Title":"Python store mp3 file as data in array","Body":"<p>I'm looking for a module that will read in an mp3 file (which is a binary file from what I understand) and store it's data values in an array. I'm new to the audio world, so I'm not sure what the values represent. Frequency maybe? <\/p>\n\n<p>I'm working in a Tensorflow project that learns from an audio Byte stream. So I need some sort of way to represent an audio Byte stream. <\/p>\n\n<p>A standard  might work, but each audio file needs to be represented as an entire array. So maybe each element represents a frame of the audio stream.<\/p>\n\n<p>I found this <a href=\"http:\/\/zulko.github.io\/blog\/2013\/10\/04\/read-and-write-audio-files-in-python-using-ffmpeg\/\" rel=\"nofollow\">link<\/a>, but I can't get it to work. <\/p>\n\n<p>I have:<\/p>\n\n<pre><\/pre>\n\n<p>But it's throwing an error: <\/p>\n\n<p>Note that  is a path to the file --> <\/p>\n\n<p>The entire snippet looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36349049,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-04-01T05:56:48.200","AcceptedAnswerId":36358913.0,"OwnerUserId":1762211.0,"Title":"overcome Graphdef cannot be larger than 2GB in tensorflow","Body":"<p>I am using tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/models\/image\/imagenet\" rel=\"noreferrer\">imageNet trained model<\/a> to extract the last pooling layer's features as representation vectors for a new dataset of images. <\/p>\n\n<p>The model as is predicts on a new image as follows: <\/p>\n\n<pre><\/pre>\n\n<p>I edited the main function so that I can take a folder of images and return the prediction on all images at once and write the feature vectors in a csv file. Here is how I did that: <\/p>\n\n<pre><\/pre>\n\n<p>It worked just fine for around 21 images but then crashed with the following error: <\/p>\n\n<pre><\/pre>\n\n<p>I thought by calling the method  the previous image data would be overwritten to only consider the new image data, which doesn't seem to be the case. How to resolve this issue? <\/p>\n","answers":[{"AnswerId":"43052381","CreationDate":"2017-03-27T17:01:21.363","ParentId":null,"OwnerUserId":"4225496","Title":null,"Body":"<p>The simplest way is put <code>create_graph()<\/code> at the first of main function.\nThen, it just create graph only <\/p>\n"},{"AnswerId":"51989984","CreationDate":"2018-08-23T16:03:26.170","ParentId":null,"OwnerUserId":"7319232","Title":null,"Body":"<p>A good explanation of why such errors is mentioned <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5383\" rel=\"nofollow noreferrer\">here<\/a>, I encountered the same error while using tf dataset api and came to the understanding that data when iterated over in the session, its getting appended on the existing graph. so what I did is used <code>tf.reset_default_graph()<\/code> before the dataset iterator to make sure that previous graph is cleared away. <\/p>\n\n<p>Hope this helps for such a scenario.<\/p>\n"},{"AnswerId":"41291482","CreationDate":"2016-12-22T20:51:46.753","ParentId":null,"OwnerUserId":"309653","Title":null,"Body":"<p>You can move the <code>create_graph()<\/code> to somewhere <em>before<\/em> this loop <code>for image in list_of_images:<\/code> (which loops over files).<\/p>\n\n<p>What it does is performing inference multiple times on the same graph.<\/p>\n"},{"AnswerId":"36358913","CreationDate":"2016-04-01T14:28:16.270","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem here is that each call to <code>run_inference_on_image()<\/code> <em>adds<\/em> nodes to the same graph, which eventually exceeds the maximum size. There are at least two ways to fix this:<\/p>\n\n<ol>\n<li><p>The <strong>easy but slow<\/strong> way is to use a different default graph for each call to <code>run_inference_on_image()<\/code>:<\/p>\n\n<pre><code>for image in list_of_images:\n  # ...\n  with tf.Graph().as_default():\n    current_features = run_inference_on_image(images_folder+\"\/\"+image)\n  # ...\n<\/code><\/pre><\/li>\n<li><p>The <strong>more involved but more efficient<\/strong> way is to modify <code>run_inference_on_image()<\/code> to run on multiple images. Relocate your <code>for<\/code> loop to surround <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3c8780451007c27b69f10925d43d1f3501d94106\/tensorflow\/models\/image\/imagenet\/classify_image.py#L171\">this <code>sess.run()<\/code> call<\/a>, and you will no longer have to reconstruct the entire model on each call, which should make processing each image much faster.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":36354101,"AnswerCount":1,"Tags":"<python><performance><regression><tensorflow>","CreationDate":"2016-04-01T10:33:48.003","AcceptedAnswerId":null,"OwnerUserId":3468216.0,"Title":"Linear regression with tensorflow is very slow","Body":"<p>I am trying to implement a simple linear regression in tensorflow (with the goal of eventually extending it to more advanced models). My current code looks as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The code works, and finds the correct coefficients. However, it is extremely slow. On my MacBook Pro, it takes several minutes just to run a few training epochs for a data set with 1000 data points and 10 features. Since I'm running OSX I don't have GPU acceleration, which could explain some of the slowness, but I would think that it could be faster than this. I have experimented with different optimizers, but the performance is very similar.<\/p>\n\n<p>Is there some obvious way to speed up this code? Otherwise, it feels like tensorflow is pretty much useless for these types of problems.<\/p>\n","answers":[{"AnswerId":"36354705","CreationDate":"2016-04-01T11:03:59.770","ParentId":null,"OwnerUserId":"2370124","Title":null,"Body":"<p>It is so slow, since you train the network point by point which requires <code>NUM_STEPS * num_datapoints<\/code> iterations (which leads to <strong>5 hundred thousands<\/strong> cycles).<\/p>\n\n<p>All you actually need to train your network is<\/p>\n\n<pre><code>for i in range(NUM_STEPS):\n    sess.run(train_step, feed_dict={x: data, y_:labels})\n<\/code><\/pre>\n\n<p>This would take just a couple of seconds.<\/p>\n"}]}
{"QuestionId":36355290,"AnswerCount":2,"Tags":"<caffe><matcaffe>","CreationDate":"2016-04-01T11:33:14.577","AcceptedAnswerId":null,"OwnerUserId":570593.0,"Title":"Caffe matlab examples for training and testing CNN","Body":"<p>How can I train and test a CNN using Matlab? \nCan anybody give me an example code?<\/p>\n","answers":[{"AnswerId":"36387149","CreationDate":"2016-04-03T15:10:17.940","ParentId":null,"OwnerUserId":"732973","Title":null,"Body":"<p>you can find Matlab examples in the Caffe repository: <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/matlab\" rel=\"nofollow\">https:\/\/github.com\/BVLC\/caffe\/tree\/master\/matlab<\/a><\/p>\n"},{"AnswerId":"39294438","CreationDate":"2016-09-02T14:09:41.740","ParentId":null,"OwnerUserId":"6205009","Title":null,"Body":"<p>Try this simple tutorial. it contains sample code\n<a href=\"https:\/\/ahmedibrahimvt.wordpress.com\/2016\/04\/18\/caffe-in-matlab-matcaffe-simple-tutorial\/\" rel=\"nofollow\">https:\/\/ahmedibrahimvt.wordpress.com\/2016\/04\/18\/caffe-in-matlab-matcaffe-simple-tutorial\/<\/a><\/p>\n"}]}
{"QuestionId":36356004,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2016-04-01T12:10:11.117","AcceptedAnswerId":36359184.0,"OwnerUserId":1013394.0,"Title":"Continue Training from a Specific Epoch","Body":"<p>A resource manager I'm using to fit a Keras model limits the access to a server to 1 day at a time. After this day, I need to start a new job. Is it possible with Keras to save the current model at epoch K, and then load that model to continue training epoch K+1 (i.e., with a new job)?<\/p>\n","answers":[{"AnswerId":"36359184","CreationDate":"2016-04-01T14:40:29.343","ParentId":null,"OwnerUserId":"3427404","Title":null,"Body":"<p>You can save weights after every epoch by specifying a callback:<\/p>\n\n<pre><code>weight_save_callback = ModelCheckpoint('\/path\/to\/weights.{epoch:02d}-{val_loss:.2f}.hdf5', monitor='val_loss', verbose=0, save_best_only=False, mode='auto')\nmodel.fit(X_train,y_train,batch_size=batch_size,nb_epoch=nb_epoch,callbacks=[weight_save_callback])\n<\/code><\/pre>\n\n<p>This will save the weights after every epoch. You can then load them with:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(...)\nmodel.load('path\/to\/weights.hf5')\n<\/code><\/pre>\n\n<p>Of course your model needs to be the same in both cases.<\/p>\n"},{"AnswerId":"56048311","CreationDate":"2019-05-08T20:08:06.620","ParentId":null,"OwnerUserId":"9770403","Title":null,"Body":"<p>You can add the <code>initial_epoch<\/code> argument. This will allow you to continue training from a specific epoch.<\/p>\n"}]}
{"QuestionId":36356277,"AnswerCount":2,"Tags":"<python><tensorflow><adam>","CreationDate":"2016-04-01T12:24:07.707","AcceptedAnswerId":null,"OwnerUserId":6024628.0,"Title":"Tensorflow Adam Multigpu Gradient","Body":"<p>I am trying to implement a network multiple gpu on tensorflow using ADAM Optimization.<\/p>\n\n<p>I am coping the code from the Cifar10_multigpu, but it looks that when the gradient calls the second tower it calls the gradient of the first and generated error on the average of the two towers.\nthe code of the two tower is this<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>and this generates each tower:<\/p>\n\n<pre><\/pre>\n\n<p>looking at the gradients the first tower generate this:<\/p>\n\n<pre><\/pre>\n\n<p>and the second generates this;<\/p>\n\n<pre><\/pre>\n\n<p>I am wondering how to remove from the second the first None, but no targeting indexes so i can make for more towers.<\/p>\n","answers":[{"AnswerId":"36371738","CreationDate":"2016-04-02T09:51:31.177","ParentId":null,"OwnerUserId":"6024628","Title":null,"Body":"<p>I already find the error. I was using a trainable variable for the learning rate(I wanted to trak the lr but it does not look possible), and also add the list of variables to compute by the op at adam. I am not sure if this is a correct way but it looks like it works.<\/p>\n\n<pre><code>with tf.Graph().as_default(), tf.device('\/cpu:0'):\n        devs = ['\/job:prs\/task:0\/gpu:0','\/job:worker\/task:0\/gpu:0'] #\n        global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False)\n        num_batches_per_epoch = dt_fdr.FLS_PER_ANGLE\/ FLAGS.batch_size\n        #lr = tf.Variable(tf.constant(FLAGS.learning_rate, dtype=tf.float32))\n        opt = tf.train.AdamOptimizer(FLAGS.learning_rate)\n        tower_grads = []\n        for i in xrange(FLAGS.num_gpus):\n            with tf.device(devs[i]):\n                with tf.name_scope('%s_%d' % (tf_model.TOWER_NAME, i)) as scope:\n                    loss = tower_loss(scope)\n                    tf.get_variable_scope().reuse_variables()\n                    summaries = tf.get_collection(tf.GraphKeys.SUMMARIES, scope)\n                    #\"print('\\n'.join('{}: {}'.format(*k) for k in enumerate(summaries)))\n                    grads = opt.compute_gradients(loss, tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope))\n                    #print('\\n'.join('{}: {}'.format(*k) for k in enumerate(grads)))\n                    tower_grads.append(grads)\n        grads = average_gradients(tower_grads)\n        #summaries.append(tf.scalar_summary('learning_rate', lr))\n        for grad, var in grads:\n            if grad:\n                summaries.append(\n                    tf.histogram_summary(var.op.name + '\/gradients', grad))\n        apply_gradient_op = opt.apply_gradients(grads, global_step=global_step) \n        for var in tf.trainable_variables():\n            summaries.append(tf.histogram_summary(var.op.name, var))\n\n        train_op = apply_gradient_op\n\n        saver = tf.train.Saver(tf.all_variables())\n\n        summary_op = tf.merge_summary(summaries)\n\n        init = tf.initialize_all_variables()\n\n        sess = tf.Session(\"grpc:\/\/nelson-lab:2500\",config=tf.ConfigProto(\n            allow_soft_placement=True,\n            log_device_placement=FLAGS.log_device_placement))\n        sess.run(init)\n<\/code><\/pre>\n\n<p>I wonder if someone also tried to do some double gpu training using adam.<\/p>\n\n<p>Regards<\/p>\n"},{"AnswerId":"45730326","CreationDate":"2017-08-17T08:39:27.223","ParentId":null,"OwnerUserId":"8476923","Title":null,"Body":"<p><strong>Additional<\/strong> <\/p>\n\n<p>If you have a plan to update learning rate in training phase, declare like below.<\/p>\n\n<pre><code>lr = tf.Variable(FLAGS.learning_rate, trainable=False)\nopt = tf.train.AdamOptimizer(lr)\n<\/code><\/pre>\n\n<p>After that <\/p>\n\n<pre><code>sess.run(tf.assign(lr, new_lr))\n<\/code><\/pre>\n"}]}
{"QuestionId":36357193,"AnswerCount":1,"Tags":"<python><python-2.7><numpy><theano><lasagne>","CreationDate":"2016-04-01T13:07:24.687","AcceptedAnswerId":38747015.0,"OwnerUserId":3846213.0,"Title":"Data type incompatibility issue in Theano","Body":"<p>I've got a dense dropout ANN with a softmax output layer. Here is the training method: <\/p>\n\n<pre><\/pre>\n\n<p>The  method is defined in the base class: <\/p>\n\n<pre><\/pre>\n\n<p>In the beginning of   is initialised as a  with the same number of dimensions and type as the numpy array used to feed it. For a reason beyond my comprehension, if the datatype is not  or  I get this error: <\/p>\n\n<pre><\/pre>\n\n<p>where  represents the type of  inferred from the numpy array (I tested that with , , , , , ). What is the reason it only takes  and ?<\/p>\n","answers":[{"AnswerId":"38747015","CreationDate":"2016-08-03T14:51:59.223","ParentId":null,"OwnerUserId":"3846213","Title":null,"Body":"<p>Seems like the reason for that originates from Theano trying to be flexible by assuring your code is 100% compatible with both CPU and GPU backends at the same time. Since most versions of CUDA only support 32 and 64 bit arithmetic types, you can't use any other. <\/p>\n"}]}
{"QuestionId":36362190,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2016-04-01T17:17:02.177","AcceptedAnswerId":null,"OwnerUserId":6146107.0,"Title":"Slice error when using MultiRNNCell","Body":"<p>I am using MultiRNNCell from . This is how declare my MultiRNNCell<\/p>\n\n<p><strong>Code:<\/strong><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>Later on I use it from inside seq2seq.embedding_rnn_decoder as follows<\/p>\n\n<p><\/p>\n\n<p>On doing this I get the following error<\/p>\n\n<p><strong>Error:<\/strong><\/p>\n\n<p><\/p>\n\n<p>Has anyone seen a similar error? Any pointers?<\/p>\n","answers":[{"AnswerId":"38411187","CreationDate":"2016-07-16T12:27:10.703","ParentId":null,"OwnerUserId":"5769621","Title":null,"Body":"<p>This problem is occurring because you have doubled your GRU cell but your initial vector is not doubled. \nIf your initial_vector size is [batch_size,50].<\/p>\n\n<p>Then initial_vector = tf.concat(1,[initial_vector,initial_vector])<\/p>\n\n<p>Now input this to decoder as initial vector.<\/p>\n"},{"AnswerId":"36371050","CreationDate":"2016-04-02T08:34:04.017","ParentId":null,"OwnerUserId":"5264769","Title":null,"Body":"<p>Seem like an invalid argument into embedding_rnn_decoder. \nMaybe try to change enc_state:<\/p>\n\n<p><code>ouputs, mem_states = seq2seq.embedding_rnn_decoder(decoder_inputs, enc_state[-1], e_cell, vocab_size, output_projection=(W, b), feed_previous=False)<\/code><\/p>\n"},{"AnswerId":"36794050","CreationDate":"2016-04-22T12:37:25.483","ParentId":null,"OwnerUserId":"6240347","Title":null,"Body":"<p>I've got a problem similar to yours.<code>{tensorflow.python.framework.errors.InvalidArgumentError: Expected size[1] in [0, 0], but got 40}<\/code><br>\nI also use <code>rnn_cell.GRUCell(self.rnn_size)<\/code>\nI want to share my experience ,maybe it's helpful.\nHere is how I fixed it.\nI want to use gru cell and basic rnn cell ,so I adapt the programme from others which is coded for lstm cell.\nThe different between lstm and GRU\/BasicRnn is the state_size.\nHere is lstm cell<code>def state_size(self):return 2 * self._num_units<\/code>\nHere is GRU\/BasicRnn cell <code>def state_size(self):return self._num_units<\/code>\nTherefore the shape of matrix is different,and the tensor is not suit for the op.I advice you to check your code contains <code>tf.slice<\/code><\/p>\n"},{"AnswerId":"38138674","CreationDate":"2016-07-01T06:40:11.670","ParentId":null,"OwnerUserId":"6536794","Title":null,"Body":"<p>I had a similar issue with the embedding_rnn_seq2seq (Tensorflow v0.9.0). I switched my inputs around to run with the embedding_rnn_decoder and was able to run a session with these inputs.<\/p>\n\n<pre><code>enc_inp = [tf.placeholder(tf.int32, shape=(batch_size,),\n           name=\"inp%i\" % t) for t in range(seq_length)]\n\nlabels = [tf.placeholder(tf.int32, shape=(batch_size,),\n           name=\"labels%i\" % t) for t in range(seq_length)]\n\nweights = [tf.ones_like(labels_t, dtype=tf.float32) for labels_t in labels]\n\ndec_inp = ([tf.zeros_like(enc_inp[0], dtype=np.int32, name=\"GO\")] + enc_inp[:-1])\n\nprev_mem = tf.zeros((batch_size, memory_dim))\n\ncell = rnn_cell.GRUCell(memory_dim)\ncell = rnn_cell.MultiRNNCell([cell]*2)\n\ndec_outputs, mem_states = seq2seq.embedding_rnn_decoder(dec_inp, prev_mem, \ncell, n_x, n_y,feed_previous=True)\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":36363222,"AnswerCount":1,"Tags":"<python><numpy><theano><keras>","CreationDate":"2016-04-01T18:18:14.337","AcceptedAnswerId":36400844.0,"OwnerUserId":5606352.0,"Title":"Python\/Keras\/Theano - Index out of bounds","Body":"<p>I'm new to Keras and having some trouble with shapes, specially when it comes to RNNs and LSTMs.<\/p>\n\n<p><strong>I'm running this code:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>I'm having this error:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I don't understand where is that \"index 143\" coming from and how do I fix it.<\/p>\n\n<p>Anyone available to enlight my journey?<\/p>\n\n<p>Extra info below.<\/p>\n\n<p>-- EDIT --\nThis \"index 143\" actually varies everytime I run the code. The numbers do not follow any apparent logic, the only thing I could notice is that, coincidentally or not, the smallest number that appeared was 80 (I ran the code more than 20 times)<\/p>\n\n<p><br>\n<br><\/p>\n\n<hr>\n\n<p><strong>EXTRA INFO<\/strong><\/p>\n\n<p><br><\/p>\n\n<blockquote>\n  <p><strong>About predictor_train (X)<\/strong><\/p>\n<\/blockquote>\n\n<p>type: 'numpy.ndarray'<\/p>\n\n<p>shape: (119,80)<\/p>\n\n<p>dtype: float64<\/p>\n\n<blockquote>\n  <p><strong>About target_train (Y)<\/strong><\/p>\n<\/blockquote>\n\n<p>type: class 'pandas.core.series.Series'<\/p>\n\n<p>shape: (119,)<\/p>\n\n<p>dtype: float64<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p><strong>model.summary()<\/strong><\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p><strong>FULL TRACEBACK<\/strong><\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p><br><\/p>\n","answers":[{"AnswerId":"36400844","CreationDate":"2016-04-04T11:02:39.420","ParentId":null,"OwnerUserId":"118630","Title":null,"Body":"<p>Your <code>X<\/code> variable probably contains the value 143. The <code>Embedding<\/code> layer will be of dimensions 80x70. <\/p>\n\n<p>I'm assuming this is in the NLP field. This means you have a vocabulary size of 80 words each represented by vectors of length 70. Your <code>X<\/code> variable represents 119 sentences of length 80 (or 80 sentences of length 119) with its contents representing indices to your vocabulary. If it contains a word index greater than 80 this error will pop-up.<\/p>\n\n<p>A more common value for your <code>col<\/code> variable is above 10.000. Of course, it depends on what you are doing.<\/p>\n"}]}
{"QuestionId":36363678,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-04-01T18:45:40.757","AcceptedAnswerId":null,"OwnerUserId":6146467.0,"Title":"Tensor flow: nn.tanh is not converging (C++ version works very well)","Body":"<p>I ported a very simple working C++ neural network to Tensorflow, however, tensor flow version does not converge in training. It's a combinator of many Netflix predictors and operates on Netflix prize prediction data set. <\/p>\n\n<h2>Here is the tensorflow program:<\/h2>\n\n<pre><\/pre>\n\n<hr>\n\n<p>Training runs, with training errors GROWING over time (first iteration has about similar training error to C++ version):<\/p>\n\n<pre><\/pre>\n\n<p>Anything wrong? The C++ version will train 1000 iterations and training errors consistently go down from 1.0x level to 0.87. Tried a few changes, still not converging.<\/p>\n","answers":[]}
{"QuestionId":36364190,"AnswerCount":1,"Tags":"<machine-learning><neural-network><torch><conv-neural-network>","CreationDate":"2016-04-01T19:15:23.217","AcceptedAnswerId":36365633.0,"OwnerUserId":2838606.0,"Title":"Obtaining intermediate layers' activation values given input example(s)","Body":"<p>Suppose I have defined my sequential model as follow:<\/p>\n\n<pre><\/pre>\n\n<p>And here's the model printed:<\/p>\n\n<pre><\/pre>\n\n<p>Simply using  returns the last layer's output after LogSoftMax has been applied to which is what I do not want. Instead, I would like to get the activations of some of the intermediate layers (e.g. module 6). <\/p>\n\n<p>So, How can I get the activations of each of the intermediate layers when feeding an input? i.e. I feed an input example to the network and want to extract the activation results of the 6th layer (a convolutional layer), not just the final layer.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"36365633","CreationDate":"2016-04-01T20:50:07.353","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Via <code>net:get(6).output<\/code> (see <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/containers.md#getindex\" rel=\"noreferrer\">get<\/a> and <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/module.md#output\" rel=\"noreferrer\">output<\/a>).<\/p>\n"}]}
{"QuestionId":36365013,"AnswerCount":1,"Tags":"<neural-network><theano><reinforcement-learning>","CreationDate":"2016-04-01T20:09:02.677","AcceptedAnswerId":null,"OwnerUserId":781608.0,"Title":"How to calculate gradients for a neural network with theano when using Q-Learning","Body":"<p>I am trying to use a standard fully-connected neural net as the basis for action values in Q-Learning.  I am using <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html#mlp\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/mlp.html#mlp<\/a> as a reference specifically this line:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to calculate the error for my output unit associated with the last action using the Q-Learning policy control method (as described in <a href=\"http:\/\/webdocs.cs.ualberta.ca\/~sutton\/book\/ebook\/node65.html\" rel=\"nofollow\">http:\/\/webdocs.cs.ualberta.ca\/~sutton\/book\/ebook\/node65.html<\/a>) and set the other output errors to zero.<\/p>\n\n<p>How can I use Theano's grad function to back-propagate the errors to the rest of the neural network?<\/p>\n","answers":[{"AnswerId":"36517362","CreationDate":"2016-04-09T13:24:17.747","ParentId":null,"OwnerUserId":"6113498","Title":null,"Body":"<p>Neural nets are just one possible way of parametrizing your Q-function. The way you perform gradient descent in this case is explained in this <a href=\"https:\/\/webdocs.cs.ualberta.ca\/~sutton\/book\/ebook\/node87.html\" rel=\"nofollow\">section<\/a> of Sutton and Barto's book. Just treat the weights of your neural net as a vector of parameters.<\/p>\n"}]}
{"QuestionId":36365118,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-01T20:15:36.460","AcceptedAnswerId":null,"OwnerUserId":5948619.0,"Title":"How do you write and retrieve TFRecord features that are lists?","Body":"<p>I have a CNN model that takes N classification labels per training example and I am trying to create TFRecords from my data set that have a label feature that is a list of int64s.<\/p>\n\n<p>On the shard creation side I am using something like the following. I have put the label data explicitly in the code but obviously it would be different for each sample:<\/p>\n\n<pre><\/pre>\n\n<p>On the reading side, I am doing something like the following. I assume a fixed number of labels (4):<\/p>\n\n<pre><\/pre>\n\n<p>When I try this Tensorflow reports:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly, I'm not understanding something fairly fundamental<\/p>\n","answers":[{"AnswerId":"36377005","CreationDate":"2016-04-02T18:08:08.253","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>Try setting default value= [-1]*4<\/p>\n"}]}
{"QuestionId":36365425,"AnswerCount":1,"Tags":"<python><machine-learning><theano><conv-neural-network><lasagne>","CreationDate":"2016-04-01T20:35:21.730","AcceptedAnswerId":null,"OwnerUserId":4846917.0,"Title":"Train convolutional neural network with theano\/lasagne","Body":"<p>I'm trying to implement a CNN using theano\/lasagne. \nI've made a neural network but can't figure out how to train it with the current state.<\/p>\n\n<p>This is how I'm trying to get the output of the network with the  as input. <\/p>\n\n<pre><\/pre>\n\n<p>However I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Why is current_states not used? <\/p>\n\n<p>I want to get the output of the model on the current_states. How do I do this?<\/p>\n\n<p>(the CNN build code: <a href=\"http:\/\/pastebin.com\/Gd35RncU\" rel=\"nofollow\">http:\/\/pastebin.com\/Gd35RncU<\/a>)<\/p>\n","answers":[{"AnswerId":"36374027","CreationDate":"2016-04-02T13:42:25.303","ParentId":null,"OwnerUserId":"5630712","Title":null,"Body":"<p>The following code snippet works for me:<\/p>\n\n<pre><code> import lasagne, theano\n import theano.tensor as T\n import numpy as np\n input_var = theano.tensor.tensor4('inputs')\n l_out = build_cnn(input_var)\n train = theano.function([input_var], lasagne.layers.get_output(l_out))\n x = np.random.randn(10, 4, 80, 80).astype(theano.config.floatX)\n train(x)\n<\/code><\/pre>\n\n<p>You didn't post your entire code, but you can check to see if in your script you are passing in the input_var variable to your build_cnn function. If you do not, then input_var will not be part of your computational graph, which is why Theano is raising the error. <\/p>\n"}]}
{"QuestionId":36367641,"AnswerCount":2,"Tags":"<python><neural-network><theano><keras>","CreationDate":"2016-04-01T23:59:51.940","AcceptedAnswerId":null,"OwnerUserId":778720.0,"Title":"Keras ValueError: I\/O operation on closed file","Body":"<p>I am trying to write a single layer network. When it starts to train through , at some random epoch it will throw the following error:<\/p>\n\n<p><\/p>\n\n<p>Here is how I am using <\/p>\n\n<p><\/p>\n\n<p>Please let me know if you have any thoughts or is encountering the same problem.<\/p>\n\n<p>Thanks<\/p>\n\n<p>Here is the full output of the error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36876338","CreationDate":"2016-04-26T21:45:15.687","ParentId":null,"OwnerUserId":"3801886","Title":null,"Body":"<p>As mentioned in the question comments (didn't see until just now), this is actually due to a bug in IPython\/Jupyter IO and how it handles the verbose output from Keras. You can disable reporting by setting <code>verbose=False<\/code> on the <code>train<\/code> and <code>predict<\/code> or <code>predict_proba<\/code> methods invoked on the model as a workaround in the mean time, or just run the model outside of the notebook.<\/p>\n\n<p>There's an <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2110\">issue on the Keras Github<\/a> that summarizes the problem.<\/p>\n"},{"AnswerId":"38579937","CreationDate":"2016-07-26T01:54:33.847","ParentId":null,"OwnerUserId":"6637508","Title":null,"Body":"<p>I got the same problem and had it fixed now. By setting verbose = 2 in model.fit can help to solve this problem. In Keras document for model.fit, you can find this: \"verbose: 0 for no logging to stdout, 1 for progress bar logging, 2 for one log line per epoch.\". Setting verbose = 2 will update the progress after each epoch, and help to reduce the amount of logging information output to the screen, which i guess caused the IO closed problem. Some people suggest to set verbose = 0 to disable the output but in that case we can't track the progress. I hope this can help to solve your problem as well.<\/p>\n"}]}
{"QuestionId":36369679,"AnswerCount":1,"Tags":"<computer-vision><normalization><deep-learning><caffe>","CreationDate":"2016-04-02T05:47:18.727","AcceptedAnswerId":null,"OwnerUserId":2362514.0,"Title":"L2 normalization in Caffe using already existing layers","Body":"<p>I am trying to perform L2 normalization in Caffe for a layer. The idea is sort of to use these L2 normalized fc7 features in contrastive loss like <a href=\"http:\/\/www.cs.cornell.edu\/~kb\/publications\/SIG15ProductNet.pdf\" rel=\"noreferrer\">http:\/\/www.cs.cornell.edu\/~kb\/publications\/SIG15ProductNet.pdf<\/a> . <\/p>\n\n<p>I could find some links where people posted there code for L2 normalization layer. However I was wondering if it's possible to do using Local Response Normalization layer of Caffe or possibly any other.<\/p>\n\n<p>I have a final fc vector of 1x2048 (2048 channels of size 1x1). Can someone please guide me about this?<\/p>\n","answers":[{"AnswerId":"55570804","CreationDate":"2019-04-08T09:59:10.000","ParentId":null,"OwnerUserId":"4136920","Title":null,"Body":"<p>You can perform L2 normalization in Caffe using a combination of simple layers:<\/p>\n\n<pre><code>layer {\n  name: \"denom\"\n  type: \"Reduction\"\n  bottom: \"loss\"\n  top: \"denom\"\n  reduction_param {\n    operation: SUMSQ\n    axis: 1\n  }\n}\nlayer {\n  name: \"power\"\n  type: \"Power\"\n  bottom: \"denom\"\n  top: \"power\"\n  power_param {\n    power: -0.5\n    shift: 9.99999996004e-13\n  }\n}\nlayer {\n  name: \"reshape\"\n  type: \"Reshape\"\n  bottom: \"power\"\n  top: \"reshape\"\n  reshape_param {\n    shape {\n      dim: 1\n    }\n    axis: -1\n    num_axes: 0\n  }\n}\nlayer {\n  name: \"tile\"\n  type: \"Tile\"\n  bottom: \"reshape\"\n  top: \"tile\"\n  tile_param {\n    axis: 1\n    tiles: 300\n  }\n}\nlayer {\n  name: \"elwise\"\n  type: \"Eltwise\"\n  bottom: \"loss\"\n  bottom: \"tile\"\n  top: \"elwise\"\n  eltwise_param {\n    operation: PROD\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":36370129,"AnswerCount":3,"Tags":"<math><tensorflow>","CreationDate":"2016-04-02T06:46:33.307","AcceptedAnswerId":null,"OwnerUserId":6027909.0,"Title":"Does tensorflow use automatic or symbolic gradients?","Body":"<p>I haven't been able to find a clear statement of whether tensorflow uses automatic or symbolic differentiation.<\/p>\n\n<p>I skimmed the tensorflow <a href=\"http:\/\/download.tensorflow.org\/paper\/whitepaper2015.pdf\" rel=\"noreferrer\">paper<\/a> and they mention automatic gradients, but it is unclear if they just mean symbolic gradients, as they also mention that it has that capability.<\/p>\n","answers":[{"AnswerId":"36373220","CreationDate":"2016-04-02T12:25:57.380","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>By \"automatic differentiation\" you may be thinking of \"differentiation by finite differences\" where you approximate derivative of <code>f(x)<\/code> as <code>[f(x+e)-f(x-e)]\/(2e)<\/code>. However, automatic differentiation is different and the finite difference method is an example of \"numerical differentiation\".<\/p>\n\n<p>TensorFlow uses reverse mode automatic differentiation for it's <code>gradients<\/code> operation and finite difference method for tests that check validity of gradient operation like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/kernel_tests\/gradient_checker.py\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Finite difference method is not practical for high dimensional problems, whereas reverse mode automatic differentiation gives you derivative of a \"many->1\" function at roughly the same cost as computing original function<\/p>\n\n<p>You can see the core of the implementation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/gradients.py#L307\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Implementation of differentiation method is symbolic in a sense that <code>gradients<\/code> operation takes computational graph and produces computational graph that can be fed back into \"gradients\" operation to get higher level derivatives. <\/p>\n\n<p>Here's an example<\/p>\n\n<pre><code>tf.reset_default_graph()\nx = tf.Variable(0.)\ny = tf.square(x)\nz = tf.gradients([y], [x])\n<\/code><\/pre>\n\n<p>Here's the graph you get<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tuWGf.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tuWGf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>There are some extra operators since same graph code will work for higher dimensional <code>x<\/code>, but the point is that you see both <code>x^2<\/code> and <code>2*x<\/code> in the same graph.<\/p>\n"},{"AnswerId":"44963454","CreationDate":"2017-07-07T05:55:08.287","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Afaik symbolic differentiation means working with a mathematical, symbolic equation (i.e., symbolic math equation in, derivative of the equation out). Automatic differentiation computes derivatives based on computational functions (which in turn are broken down into basic operations such as addition\/subtraction and multipliation\/division). <\/p>\n\n<p>Since TensorFlow does differentiation based on a computation graph of operations, I'd intuitively say that it's automatic differentiation (I don't know of any other technique that would be appropriate here; I think the possibility that TensorFlow is converting the computation graph into a mathematical equation that is then parsed to compute the derivative of that equation is prob. out of question). The authors say \"symbolic differentiation\" in the TensorFlow <a href=\"http:\/\/download.tensorflow.org\/paper\/whitepaper2015.pdf\" rel=\"nofollow noreferrer\">whitepaper<\/a> though -- however, I think this may be a misnomer similar to \"Tensor\" instead of \"(multi-dimensional) data array\" if you'd ask a mathematician.<\/p>\n"},{"AnswerId":"45133884","CreationDate":"2017-07-16T22:36:58.753","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>TF uses automatic differentiation and more specifically reverse-mode auto differentiation. <\/p>\n\n<hr>\n\n<p>There are 3 popular methods to calculate the derivative:<\/p>\n\n<ol>\n<li>Numerical differentiation<\/li>\n<li>Symbolic differentiation<\/li>\n<li>Automatic differentiation<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Numerical_differentiation\" rel=\"noreferrer\"><strong>Numerical differentiation<\/strong><\/a> relies on the definition of the derivative: <a href=\"https:\/\/i.stack.imgur.com\/ZUwpC.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZUwpC.png\" alt=\"enter image description here\"><\/a>, where you put a very small <code>h<\/code> and evaluate function in two places. This is the most basic formula and on practice people use other formulas which give smaller estimation error. This way of calculating a derivative is suitable mostly if you do not know your function and can only sample it. Also it requires a lot of computation for a high-dim function.<\/p>\n\n<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Symbolic_computation\" rel=\"noreferrer\"><strong>Symbolic differentiation<\/strong><\/a> manipulates mathematical expressions. If you ever used matlab or mathematica, then you <a href=\"https:\/\/www.wolframalpha.com\/input\/?i=derivative%20%20x%5E2*cos(x-7)%2F(sin(x))\" rel=\"noreferrer\">saw something like this<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/XZgZL.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XZgZL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here for every math expression they know the derivative and use various rules (product rule, chain rule) to calculate the resulting derivative. Then they simplify the end expression to obtain the resulting expression.<\/p>\n\n<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation\" rel=\"noreferrer\"><strong>Automatic differentiation<\/strong><\/a> manipulates blocks of computer programs. A differentiator has the rules for taking the derivative of each element of a program (when you define any op in core TF, you need to <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op#implement_the_gradient_in_python\" rel=\"noreferrer\">register a gradient<\/a> for this op). It also uses chain rule to break complex expressions into simpler ones. Here is a <a href=\"https:\/\/stackoverflow.com\/q\/44342432\/1090562\">good example how it works in real TF programs with some explanation<\/a>.<\/p>\n\n<hr>\n\n<p>You might think that Automatic differentiation is the same as Symbolic differentiation (in one place they operate on math expression, in another on computer programs). And yes, they are sometimes very similar. But for control flow statements (`if, while, loops) the results can be very <a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation\" rel=\"noreferrer\">different<\/a>:<\/p>\n\n<blockquote>\n  <p>symbolic differentiation leads to inefficient code (unless carefully\n  done) and faces the difficulty of converting a computer program into a\n  single expression<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":36370247,"AnswerCount":2,"Tags":"<python><numpy><pandas><keras>","CreationDate":"2016-04-02T07:00:23.983","AcceptedAnswerId":36370605.0,"OwnerUserId":5185388.0,"Title":"How to convert an excel file data into numpy array using pandas?","Body":"<p>I am really new in keras library and also Python. I am trying to import an excel file using pandas and convert it to a  using  function of pandas. But it seams to read my file wrong. Like I have a 90x1049 data set in Excel file. But when i am trying to convert it into numpy array it reads my data as 89x1049. I am using the following code, which is not working:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36370605","CreationDate":"2016-04-02T07:42:15.670","ParentId":null,"OwnerUserId":"2681632","Title":null,"Body":"<p>Probably what happens is that your Excel file has no header row and so <code>pandas.read_excel<\/code> consumes your first data row as such.<\/p>\n\n<p>I tried creating an xlsx containing<\/p>\n\n<pre><code>1   2   3\n2   3   4\n3   4   5\n4   5   6\n5   6   7\n6   7   8\n7   8   9\n8   9   10\n9   10  11\n10  11  12\n<\/code><\/pre>\n\n<p>Reading that resulted in<\/p>\n\n<pre><code>In [3]: df = pandas.read_excel('test.xlsx')\n\nIn [4]: df\nOut[4]: \n    1   2   3\n0   2   3   4\n1   3   4   5\n2   4   5   6\n3   5   6   7\n4   6   7   8\n5   7   8   9\n6   8   9  10\n7   9  10  11\n8  10  11  12\n<\/code><\/pre>\n\n<p>As can be seen, the first data row has been used as labels for columns.<\/p>\n\n<p>To avoid consuming the first data row as headers, pass <code>headers=None<\/code> to <code>read_excel<\/code>. Interestingly the <a href=\"http:\/\/pandas.pydata.org\/pandas-docs\/stable\/generated\/pandas.read_excel.html\" rel=\"nofollow noreferrer\">documentation<\/a> did not mention this usage before, but has been fixed since:<\/p>\n\n<blockquote>\n  <p><strong>header<\/strong> : <em>int, list of ints, default 0<\/em><\/p>\n  \n  <p>Row (0-indexed) to use for the column labels of the parsed DataFrame. If a list of integers is passed those row positions will be combined into a <code>MultiIndex<\/code>. Use None if there is no header.<\/p>\n<\/blockquote>\n"},{"AnswerId":"36370645","CreationDate":"2016-04-02T07:47:14.430","ParentId":null,"OwnerUserId":"4531270","Title":null,"Body":"<p>If you have no header, try the following:<\/p>\n\n<pre><code>training_data = pd.read_excel(\"\/home\/workstation\/ANN\/new_input.xlsx\", header=None)\n\nX_train = training_data_x.as_matrix()\n<\/code><\/pre>\n\n<p>See also answers from a <a href=\"https:\/\/stackoverflow.com\/questions\/29539742\/pandas-read-csv-without-header-which-might-be-there\">previous question<\/a>.<\/p>\n"}]}
{"QuestionId":36370677,"AnswerCount":1,"Tags":"<python><tensorflow><word2vec>","CreationDate":"2016-04-02T07:51:40.457","AcceptedAnswerId":null,"OwnerUserId":5264769.0,"Title":"Injecting pre-trained word2vec vectors into TensorFlow seq2seq","Body":"<p>I was trying to inject pretrained word2vec vectors into existing tensorflow seq2seq model.<\/p>\n\n<p>Following <a href=\"https:\/\/stackoverflow.com\/questions\/36072672\/how-to-import-word2vec-into-tensorflow-seq2seq-model\/36075984#36075984\">this answer<\/a>, I produced the following code. But it doesn't seem to improve performance as it should, although the values in the variable are updated.<\/p>\n\n<p>In my understanding the error might be due to the fact that EmbeddingWrapper or embedding_attention_decoder create embeddings independently of the vocabulary order?<\/p>\n\n<p>What would be the best way to load pretrained vectors into tensorflow model?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36414580","CreationDate":"2016-04-04T23:23:36.657","ParentId":null,"OwnerUserId":"640898","Title":null,"Body":"<p>I am not familiar with the seq2seq example, but in general you can use the following code snippet to inject your embeddings:<\/p>\n\n<p>Where you build you graph:<\/p>\n\n<pre><code>with tf.device(\"\/cpu:0\"):\n  embedding = tf.get_variable(\"embedding\", [vocabulary_size, embedding_size])      \n  inputs = tf.nn.embedding_lookup(embedding, input_data)\n<\/code><\/pre>\n\n<p>When you execute (after building your graph and before stating the training), just assign your saved embeddings to the embedding variable:<\/p>\n\n<pre><code>session.run(tf.assign(embedding, embeddings_that_you_want_to_use))\n<\/code><\/pre>\n\n<p>The idea is that the <code>embedding_lookup<\/code> will replace <code>input_data<\/code> values with those present in the <code>embedding<\/code> variable. <\/p>\n"}]}
{"QuestionId":36371137,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-04-02T08:44:41.577","AcceptedAnswerId":null,"OwnerUserId":1968073.0,"Title":"google.protobuf.message.DecodeError: Error parsing message","Body":"<p>I am doing the tutorial on <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/image_recognition\/index.html<\/a><\/p>\n\n<p>and working on the Python API section, I launched\n    <\/p>\n\n<p>Here I am getting the error message out like so:<\/p>\n\n<p><\/p>\n\n<p>So, I tried to access the file coded_stream.h, but I cannot get to it as the directory tree is getting me to google\/protobuf and no further. There is no io directory.\nAny hints on what I should do now?<\/p>\n\n<p>thanks\nPeter<\/p>\n","answers":[{"AnswerId":"36377767","CreationDate":"2016-04-02T19:23:59.783","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You need to install the custom protobuf pip package as detailed here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#protobuf-library-related-issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#protobuf-library-related-issues<\/a><\/p>\n\n<p><strong>EDIT :<\/strong> (updating link) <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/install\/install_linux#protobuf_pip_package_31\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.0\/install\/install_linux#protobuf_pip_package_31<\/a><\/p>\n"}]}
{"QuestionId":36372938,"AnswerCount":1,"Tags":"<lua><machine-learning><ubuntu-14.04><deep-learning><torch>","CreationDate":"2016-04-02T11:57:51.000","AcceptedAnswerId":36378137.0,"OwnerUserId":4001193.0,"Title":"require('image') not working in Torch7","Body":"<p>I have installed 'image' package using luarocks<\/p>\n\n<pre><\/pre>\n\n<p>and everything gets installed without any error. Whenever I try to use it the following error is returned:<\/p>\n\n<pre><\/pre>\n\n<p>Can anybody please help me in this? I am a newbie in Torch7 &amp; LUA.<\/p>\n\n<p>PS: I am running Ubuntu 14.04 and have installed Torch7 according to the official instructions and everything else is working fine.<\/p>\n","answers":[{"AnswerId":"36378137","CreationDate":"2016-04-02T20:04:24.913","ParentId":null,"OwnerUserId":"431895","Title":null,"Body":"<p>Hit the same problem. Doing<\/p>\n\n<pre><code>luarocks install sys\n<\/code><\/pre>\n\n<p>as well worked for me. Not sure what happened to sys with the image install.<\/p>\n"}]}
{"QuestionId":36376066,"AnswerCount":0,"Tags":"<python><archlinux><caffe><undefined-symbol>","CreationDate":"2016-04-02T16:50:40.483","AcceptedAnswerId":null,"OwnerUserId":5572659.0,"Title":"Get \"undefined symbol: cairo_surface_get_device_scale\" when use caffe in python","Body":"<p>After successfully compiled and installed  and , I got  when trying to import  in Python.<\/p>\n\n<pre><\/pre>\n\n<p>I have googled that upgrade the  might help. But it did not help. I use Arch Linux and have every thing upgraded and reinstalled , which is the latest version, but it still did not work.<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for your help!<\/p>\n","answers":[]}
{"QuestionId":36381488,"AnswerCount":1,"Tags":"<python><theano><deep-learning>","CreationDate":"2016-04-03T04:03:04.203","AcceptedAnswerId":36406057.0,"OwnerUserId":3654968.0,"Title":"Why does Theano throw NaNs when I use dropouts?","Body":"<p>I am training a simple feed-forward model with 3 or 4 hidden layers and dropouts between each (hidden layer + non linearity) combination.\nSometimes after a few epochs (about 10-11) the model starts outputting Infs and NaNs as the error of the NLL and the accuracy falls to 0.0%.  This problem does not happen when I do not use dropouts. Is this a known issue with dropouts in Theano? The way I implement dropouts is:<\/p>\n\n<pre><\/pre>\n\n<p>where input is the feature-vector on which we want to apply dropouts.\nI have also observed that the occurance of NaNs happens earlier if the dropout probability (self.p) is higher. p = 0.5 would cause NaNs to occur around epoch 1 or 2 but p = 0.7 would cause NaNs to occur around epoch 10 or 11.\nAlso the occurrence of NaNs happens only when hidden layer sizes are large. For example (800,700,700) gives NaNs whereas (500,500,500) does not.<\/p>\n","answers":[{"AnswerId":"36406057","CreationDate":"2016-04-04T14:54:07.443","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>in my experience, NaNs, when training a network usually happen because of two problems:<\/p>\n\n<ul>\n<li>first, mathematical error, e.g. log of negative value. It could happen when you are using log() in your loss function.<\/li>\n<li>Second, there is a value that becomes too big so python can't handle.<\/li>\n<\/ul>\n\n<p>In your case, from your good observation, I think it's a second case. Your loss value may become too big to handled by python. Try to initialize smaller weight when you try to expand your network. Or just use different approach to initialize weight like explained by <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf\" rel=\"nofollow\">Glorot (2010)<\/a> or <a href=\"http:\/\/arxiv.org\/abs\/1502.01852\" rel=\"nofollow\">He (2015)<\/a>. Hope it helps.<\/p>\n"}]}
{"QuestionId":36382076,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-04-03T05:51:29.000","AcceptedAnswerId":36382246.0,"OwnerUserId":5576325.0,"Title":"What is the difference between `tensorflow.nn.np` and `numpy`?","Body":"<p>Does  simply call  or it is implemented in 's c++ back-end?<\/p>\n","answers":[{"AnswerId":"36382246","CreationDate":"2016-04-03T06:21:29.147","ParentId":null,"OwnerUserId":"2272172","Title":null,"Body":"<p>When in doubt, print out the module.<\/p>\n\n<pre><code>import tensorflow\nprint(tensorflow.nn.np)\n<\/code><\/pre>\n\n<p>For me it prints:<\/p>\n\n<pre><code>&lt;module 'numpy' from '\/Users\/ch\/miniconda\/envs\/sb35\/lib\/python3.5\/site-packages\/numpy\/__init__.py'&gt;\n<\/code><\/pre>\n\n<p>So <code>tensorflow.nn.np<\/code> is just an alias for numpy.<\/p>\n"}]}
{"QuestionId":36383466,"AnswerCount":1,"Tags":"<machine-learning><torch>","CreationDate":"2016-04-03T08:58:57.537","AcceptedAnswerId":36391774.0,"OwnerUserId":4341948.0,"Title":"Unusual behavior of image saving and loading in torch7","Body":"<p>I noticed an unusual behavior with torch7. I know a little about torch7. So I don't know how this behavior can be explained or corrected.<\/p>\n\n<p>So, I am using CIFAR-10 dataset. Simply I fetched data for an image from CIFAR-10 and then saved it in my directory. When I loaded that saved image, it was different.<\/p>\n\n<p>Here is my code - <\/p>\n\n<pre><\/pre>\n\n<p>Is this behavior expected? I thought  was supposed to be lossless.<\/p>\n\n<p>If so how this can be corrected?<\/p>\n\n<p>Code for loading CIFAR-10 dataset-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36391774","CreationDate":"2016-04-03T22:02:08.103","ParentId":null,"OwnerUserId":"4850610","Title":null,"Body":"<p><code>==<\/code> operator compares pointers to two tensors, not contents:<\/p>\n\n<pre><code>a = torch.Tensor(3, 5):fill(1)\nb = torch.Tensor(3, 5):fill(1)\nprint(a == b)\n&gt; false\nprint(a:eq(b):all())\n&gt; true\n<\/code><\/pre>\n"}]}
{"QuestionId":36384243,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><matcaffe>","CreationDate":"2016-04-03T10:22:36.767","AcceptedAnswerId":36386448.0,"OwnerUserId":6151595.0,"Title":"How can I convert data which form is matlab file to LMDB as my caffe input?","Body":"<p>I have a data set which form is matlab file.<br>\nThe data set contains 600,000 samples and every sample is a matrix of 7-by-256.<br>\nMy  data  is not image but signal.<br>\nI want to use CNN of caffe to train the data.<br>\nSo how can I convert it to LMDB as my input of CNN.<\/p>\n\n<p>I'm badly need the solution!<\/p>\n","answers":[{"AnswerId":"36386448","CreationDate":"2016-04-03T14:10:57.420","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Converting data in matlab directly to lmdb might be a little tricky.<\/p>\n\n<p>Why don't you try exporting your data to hdf5 binary files (supported both by matlab and caffe)?<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/34261942\/1714410\">Here<\/a> is an answer describing how this can be done.<\/p>\n"}]}
{"QuestionId":36384618,"AnswerCount":3,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-04-03T10:59:00.207","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"Tensorflow getting data into it (SVHN)","Body":"<p>i sucessfully installed tensorflow and followed the easy tutorial on the MNIST data.\nnow i wanna build models to train on the SVHN data. Unfortunately i cannot find anywhere how i can get the data into the model.\nBasically the first step of every model.\nthe data is saved in a dict, with the keys \"X\" for the parameters and 'y' for the labels.\nthe shapes are as follows:<\/p>\n\n<p>print traindata['X'].shape<\/p>\n\n<p>(32, 32, 3, 73257)<\/p>\n\n<p>print traindata['y'].shape<\/p>\n\n<p>(73257, 1)<\/p>\n\n<p>Can anyone give me a tip or a link how to get this successfully into tensorflow?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"41733113","CreationDate":"2017-01-19T03:16:42.470","ParentId":null,"OwnerUserId":"5046896","Title":null,"Body":"<p>The SVNH's <code>digitStruct.mat<\/code> is in matlab's file format, so you should convert it.<\/p>\n\n<p>Here is a code to covert the <code>digitStruct.mat<\/code> to json, or you can use <code>scipy.io.loadmat<\/code><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code># coding: utf-8\n# SVHN extracts data from the digitStruct.mat full numbers files.  The data can be downloaded\n# the Street View House Number (SVHN)  web site: http:\/\/ufldl.stanford.edu\/housenumbers.\n#\n# This is an A2iA tweak (YG -9 Jan 2014) of the script found here :\n# http:\/\/blog.grimwisdom.com\/python\/street-view-house-numbers-svhn-and-octave\n#\n# The digitStruct.mat files in the full numbers tars (train.tar.gz, test.tar.gz, and extra.tar.gz)\n# are only compatible with matlab.  This Python program can be run at the command line and will generate\n# a json version of the dataset.\n#\n# Command line usage:\n#       SVHN_dataextract.py [-f input] [-o output_without_extension]\n#    &gt;  python SVHN_dataextract.py -f digitStruct.mat -o digitStruct\n#\n# Issues:\n#    The alibility to split in several files has been removed from the original\n#    script.\n#\n\nimport tqdm\nimport h5py\nimport optparse\nfrom json import JSONEncoder\n\nparser = optparse.OptionParser()\nparser.add_option(\"-f\", dest=\"fin\", help=\"Matlab full number SVHN input file\", default=\"digitStruct.mat\")\nparser.add_option(\"-o\", dest=\"filePrefix\", help=\"name for the json output file\", default=\"digitStruct\")\noptions, args = parser.parse_args()\n\nfin = options.fin\n\n\n# The DigitStructFile is just a wrapper around the h5py data.  It basically references\n#    inf:              The input h5 matlab file\n#    digitStructName   The h5 ref to all the file names\n#    digitStructBbox   The h5 ref to all struc data\nclass DigitStructFile:\n    def __init__(self, inf):\n        self.inf = h5py.File(inf, 'r')\n        self.digitStructName = self.inf['digitStruct']['name']\n        self.digitStructBbox = self.inf['digitStruct']['bbox']\n\n    # getName returns the 'name' string for for the n(th) digitStruct.\n    def getName(self, n):\n        return ''.join([chr(c[0]) for c in self.inf[self.digitStructName[n][0]].value])\n\n    # bboxHelper handles the coding difference when there is exactly one bbox or an array of bbox.\n    def bboxHelper(self, attr):\n        if len(attr) &gt; 1:\n            attr = [self.inf[attr.value[j].item()].value[0][0] for j in range(len(attr))]\n        else:\n            attr = [attr.value[0][0]]\n        return attr\n\n    # getBbox returns a dict of data for the n(th) bbox.\n    def getBbox(self, n):\n        bbox = {}\n        bb = self.digitStructBbox[n].item()\n        bbox['height'] = self.bboxHelper(self.inf[bb][\"height\"])\n        bbox['label'] = self.bboxHelper(self.inf[bb][\"label\"])\n        bbox['left'] = self.bboxHelper(self.inf[bb][\"left\"])\n        bbox['top'] = self.bboxHelper(self.inf[bb][\"top\"])\n        bbox['width'] = self.bboxHelper(self.inf[bb][\"width\"])\n        return bbox\n\n    def getDigitStructure(self, n):\n        s = self.getBbox(n)\n        s['name'] = self.getName(n)\n        return s\n\n    # getAllDigitStructure returns all the digitStruct from the input file.\n    def getAllDigitStructure(self):\n        print('Starting get all digit structure')\n        return [self.getDigitStructure(i) for i in tqdm.tqdm(range(len(self.digitStructName)))]\n\n    # Return a restructured version of the dataset (one structure by boxed digit).\n    #\n    #   Return a list of such dicts :\n    #      'filename' : filename of the samples\n    #      'boxes' : list of such dicts (one by digit) :\n    #          'label' : 1 to 9 corresponding digits. 10 for digit '0' in image.\n    #          'left', 'top' : position of bounding box\n    #          'width', 'height' : dimension of bounding box\n    #\n    # Note: We may turn this to a generator, if memory issues arise.\n    def getAllDigitStructure_ByDigit(self):\n        pictDat = self.getAllDigitStructure()\n        result = []\n        structCnt = 1\n        print('Starting pack josn dict')\n        for i in tqdm.tqdm(range(len(pictDat))):\n            item = {'filename': pictDat[i][\"name\"] }\n            figures = []\n            for j in range(len(pictDat[i]['height'])):\n                figure = dict()\n                figure['height'] = pictDat[i]['height'][j]\n                figure['label']  = pictDat[i]['label'][j]\n                figure['left']   = pictDat[i]['left'][j]\n                figure['top']    = pictDat[i]['top'][j]\n                figure['width']  = pictDat[i]['width'][j]\n                figures.append(figure)\n            structCnt += 1\n            item['boxes'] = figures\n            result.append(item)\n        return result\n\n\ndsf = DigitStructFile(fin)\ndataset = dsf.getAllDigitStructure_ByDigit()\nfout = open(options.filePrefix + \".json\", 'w')\nfout.write(JSONEncoder(indent=True).encode(dataset))\nfout.close()\n<\/code><\/pre>\n\n<p>After that, you should write code to load the data into numpy.<\/p>\n\n<p>And in my opinion, your task is not load the data into TensorFlow, but load the all the images into numpy. So, you should also use <code>PIL<\/code> library to read the images into numpy format.<\/p>\n"},{"AnswerId":"36393069","CreationDate":"2016-04-04T00:48:23.417","ParentId":null,"OwnerUserId":"6104317","Title":null,"Body":"<p>TensorFlow uses this concept where: first, you define a graph; next, you train a graph; lastly, you employ the graph.<\/p>\n\n<p>At the moment you define the graph, you make placeholders. That are like input nodes for your graph. However, at this time, these variables are \"empty\" in the sense that they do not relate to your input data. <\/p>\n\n<p>Both at train-time and test-time, you \"feed\" your data into the graph by referring to this predefined input nodes.<\/p>\n\n<p>This concept can be new to you and I recommend you take some tutorials on it. TensorFlow itself has a nice page called \"Tensorflow Mechanics 101\" and \"basic usage\". \nIf you are more visual, I can recommend the YouTube channel \"Dan Does Data\" he explores the TensorFlow concepts in a humorous way.<\/p>\n\n<p>If you are more the person who likes example-code, you can consider <a href=\"http:\/\/robromijnders.github.io\/tensorflow_basic\/\" rel=\"nofollow\">this<\/a> example, where I make a small CNN for MNIST. Have a look at the placeholders \"x\" and \"y_\", they are the variables of your interest.<\/p>\n"},{"AnswerId":"36389264","CreationDate":"2016-04-03T18:18:02.830","ParentId":null,"OwnerUserId":"5774288","Title":null,"Body":"<p>I think a good idea is to reshape your data into something like:<\/p>\n\n<p>print traindata['X'].shape <br \/>\n(73257, 32*32*3) = (73257, 3072)<\/p>\n\n<p>I'm new to tensorflow and python too but i think u can do this by using numpy with np.reshape<\/p>\n\n<p>look at the following documentation :\n<a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow\">http:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html<\/a><\/p>\n\n<p>Tell me if it worked for you :)<\/p>\n"}]}
{"QuestionId":36384664,"AnswerCount":0,"Tags":"<caffe><matcaffe>","CreationDate":"2016-04-03T11:04:52.010","AcceptedAnswerId":null,"OwnerUserId":570593.0,"Title":"Caffe matlab wrapper random outputs","Body":"<p>I trained a CNN.\nI used matcaffe to test it. My code is<\/p>\n\n<pre><\/pre>\n\n<p>the results I am getting is (2 class classification),<\/p>\n\n<pre><\/pre>\n\n<p>Why I am getting different results for the same input?<\/p>\n","answers":[]}
{"QuestionId":36385159,"AnswerCount":1,"Tags":"<matlab><lua><machine-learning><torch>","CreationDate":"2016-04-03T12:03:38.197","AcceptedAnswerId":36385419.0,"OwnerUserId":4001193.0,"Title":"How to convert LUA Table to matlab MAT File?","Body":"<p>I want to save a LUA table as a MAT file. Structure of Table is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Can anybody guide me how to do that ?<\/p>\n\n<p>I tried matio.save(string of path, table to save) but it returned following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36385419","CreationDate":"2016-04-03T12:29:06.743","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Install and use <a href=\"https:\/\/github.com\/soumith\/matio-ffi.torch\" rel=\"nofollow\">https:\/\/github.com\/soumith\/matio-ffi.torch<\/a>, e.g.:<\/p>\n\n<pre><code>local matio = require \"matio\"\nmatio.save(\"archive.mat\",{foo=\"bar\", scm=\"git\"})\n<\/code><\/pre>\n"}]}
{"QuestionId":36388431,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-03T17:06:07.410","AcceptedAnswerId":null,"OwnerUserId":6057163.0,"Title":"Tensorflow multi-dimension argmax","Body":"<p>Say I have a tensor of size BxWxHxD. I want to process the tensor such that I have a new BxWxHxD tensor where only the maximum element in each WxH slice is kept, and all other values are zero.<\/p>\n\n<p>In other words, I think the best way to achieve this is to somehow take a 2D argmax across the WxH slices, resulting in BxD index tensors for the rows and colums that can then be converted to a one-hot BxWxHxD tensor to be used as a mask. How do I make this work?<\/p>\n","answers":[{"AnswerId":"47097095","CreationDate":"2017-11-03T13:36:49.233","ParentId":null,"OwnerUserId":"3502012","Title":null,"Body":"<p>You can use the following function as a starting point. It calculates the indexes of the maximum element for each batch and for each channel. The resulting array is in the format (batch size, 2, number of channels).<\/p>\n\n<pre><code>def argmax_2d(tensor):\n\n  # input format: BxHxWxD\n  assert rank(tensor) == 4\n\n  # flatten the Tensor along the height and width axes\n  flat_tensor = tf.reshape(tensor, (tf.shape(tensor)[0], -1, tf.shape(tensor)[3]))\n\n  # argmax of the flat tensor\n  argmax = tf.cast(tf.argmax(flat_tensor, axis=1), tf.int32)\n\n  # convert indexes into 2D coordinates\n  argmax_x = argmax \/\/ tf.shape(tensor)[2]\n  argmax_y = argmax % tf.shape(tensor)[2]\n\n  # stack and return 2D coordinates\n  return tf.stack((argmax_x, argmax_y), axis=1)\n\ndef rank(tensor):\n\n  # return the rank of a Tensor\n  return len(tensor.get_shape())\n<\/code><\/pre>\n"}]}
{"QuestionId":36388985,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-03T17:54:51.387","AcceptedAnswerId":36393015.0,"OwnerUserId":5774288.0,"Title":"How to print some debug?","Body":"<p>I'm doing my first script with tensorflow.\nI wanted to try a simple Logistic regression to start, i'm working on kaggle titanic dataset.<\/p>\n\n<p>My problem is that i'm not able to print some tensor to debug things that i'm doing wrong.<\/p>\n\n<p>i read this post (<a href=\"https:\/\/stackoverflow.com\/questions\/33633370\/how-to-print-the-value-of-a-tensor-object-in-tensorflow\">How to print the value of a Tensor object in TensorFlow?<\/a>) but i did not understand how i can print tensors. :(<\/p>\n\n<p>I'm pretty sure to be close but can't figure it out<\/p>\n\n<p>Let me show you what i am doing;<\/p>\n\n<pre><\/pre>\n\n<p>want to print here <\/p>\n\n<pre><\/pre>\n\n<p>everything i'm doing is basics i guess. If somebody can help me i would be so great ! i'm kinda stuck and can't imporve my model. Don't hesitate to tell me good practice if you feel like :)<\/p>\n\n<p>Thanks for reading this !<\/p>\n","answers":[{"AnswerId":"36393015","CreationDate":"2016-04-04T00:41:01.110","ParentId":null,"OwnerUserId":"6104317","Title":null,"Body":"<p>@Ghoso here.<\/p>\n\n<p>Good you are starting on TensorFlow. I like your idea to start with a simple classifier to get familiar.<\/p>\n\n<p>Tensorflow has this concept where first you define a computational graph. Next you train it. After and during training, you can feed and fetch your variables of interest.<\/p>\n\n<p>In this code, you seem to define a Tensor \"accuracy\" during your session. In my opinion, \"accuracy\" is a part of your graph. That is, \"accuracy\" depends on other nodes in the graph and can be considered itself a node.<\/p>\n\n<p>Next, you can expand <code>sess.run()<\/code> with a list of fetches, rather than one fetch. Like this:\n<code>output_list = sess.run([train_step, accuracy], feed_dict = feed_dict)<\/code><\/p>\n\n<p>I have a small example online where I also define \"accuracy\" as a node and fetch it in a session. You might want to look at <a href=\"http:\/\/robromijnders.github.io\/tensorflow_basic\/\" rel=\"nofollow\">this<\/a> example. The variables of your interest are named 'accuracy' and 'correct_prediction'. <\/p>\n\n<p>I hope this helps!<\/p>\n"}]}
{"QuestionId":36389356,"AnswerCount":0,"Tags":"<numpy><neural-network><torch><pybrain><lstm>","CreationDate":"2016-04-03T18:25:11.167","AcceptedAnswerId":null,"OwnerUserId":3340234.0,"Title":"Using Torch for Time Series prediction using LSTMs","Body":"<p>My main problem is how should I pre-process my dataset that is basically a 60 minutely sequenced numbers inputs that will result in a 1 hourly output. Knowing that each input vector every minute is producing some output, but unfortunately this output can't be observed until 1 hour is passed.<\/p>\n\n<p>I thought about considering putting 60 inputs as one big input vector which corresponds to 1 hourly output on a normal ML classfier, hence having 1 sample at a time. But I don't think it would be time series anymore.<\/p>\n\n<p>How can I represent that to be doable in an LSTM environment?<\/p>\n","answers":[]}
{"QuestionId":36390767,"AnswerCount":3,"Tags":"<python><memory><gpu><tensorflow>","CreationDate":"2016-04-03T20:28:26.350","AcceptedAnswerId":null,"OwnerUserId":3320135.0,"Title":"Tensor Flow: Ran out of memory trying to allocate","Body":"<p>I am running Tensor Flow version 0.7.1, 64-bit GPU-enabled, installed with pip, and on a PC with Ubuntu 14.04. My issue is that Tensor Flow is running out of memory when building my network, even though based on my calculations, there should be sufficient room on my GPU.<\/p>\n\n<p>Below is a minimal example of my code, which is based on the Tensor Flow MNIST tutorial. The network is a two-layer fully-connected network, and the number of nodes in the hidden layer is defined by the variable . The size of the training minibatch is 1. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Now, if , then the network runs fine. However, if , I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>However, according to my calculations, there should not be a problem with the memory. The number of parameters in the network is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>So with , and using  data, the total memory required for the network should therefore be 73.1 MB.<\/p>\n\n<p>Now, my graphics card is the NVIDIA GeForce GTX 780 Ti, which has 3072 MB of memory. After finding my graphics card, Tensor Flow prints out the following:<\/p>\n\n<pre><\/pre>\n\n<p>So, there should be around 2.32 GB memory available, which is far greater than the 73.1 MB calculated above. The minibatch size is 1, so this has minimal effect. Why am I getting this error?<\/p>\n\n<hr>\n\n<p>I have also now tried this on my laptop, which has an NVida GeForce GTX 880M GPU. Here, Tensor Flow reads out . Running the same code as above, it gives me a memory error at around , which is equivalent to 2.2 GB. This makes a bit more sense, and is significantly higher than the point at which my PC code breaks. However, it is still puzzling to me why it does not break closer to the 7.6 GB mark.<\/p>\n\n<hr>\n\n<p>The full output from Tensor Flow whilst running the above code on my PC, with , is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43125887","CreationDate":"2017-03-30T18:37:41.230","ParentId":null,"OwnerUserId":"4175613","Title":null,"Body":"<p>Just for your information. I had same error on my Macbook pro.\nAfter I closed few of other Applications, that issue is gone.\nBut I still got other errors like:<\/p>\n\n<blockquote>\n  <p>Blockquote\n  W tensorflow\/core\/common_runtime\/bfc_allocator.cc:217] Ran out of memory trying to allocate 214.51MiB. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory is available.<\/p>\n<\/blockquote>\n\n<p>So, that's a real out of memory issue.<\/p>\n"},{"AnswerId":"43695575","CreationDate":"2017-04-29T12:13:27.973","ParentId":null,"OwnerUserId":"3422219","Title":null,"Body":"<p>Encounter the same error, just restart the program from jupyter notebook, it runs properly. Still not find the reason. Even run the single <code>session = tf.InteractiveSession()<\/code> the same error appears. Hope it helps.<\/p>\n"},{"AnswerId":"37910983","CreationDate":"2016-06-19T19:24:04.280","ParentId":null,"OwnerUserId":"6478364","Title":null,"Body":"<p>Judging from the error, tensorflow OOM'ed trying to allocate a <code>[10000, 23000]<\/code>-sized tensor. Given that 10,000 happens to be the number of examples usually in the MNIST test set, I'm going to assume that you have some evaluation code that attempts to evaluate the whole test set at once. For just the activations you'd need <code>10000 * (784 + n + 10) ~= 1GB<\/code>, which by itself shouldn't be enough to OOM. But there's also 1.7GB tensor allocated for some reason which is hard to explain.<\/p>\n\n<p>For the case on the laptop, you're missing some variables in your calculation. Adam <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/adam.py#L117\" rel=\"noreferrer\">tracks the first and second moments for each variable<\/a> so the 2.2GB triples to become 6.6GB. Add some overhead for the gradients that will be in memory and that explains that OOM.<\/p>\n\n<p>I'm sorry this doesn't completely answer your question, I would have added this as a comment but I don't have the reputation for that yet.<\/p>\n"}]}
{"QuestionId":36390776,"AnswerCount":1,"Tags":"<memory-management><lua><machine-learning><gpgpu><torch>","CreationDate":"2016-04-03T20:29:12.147","AcceptedAnswerId":36489412.0,"OwnerUserId":2838606.0,"Title":"How to deal with GPU memory leakage issues in Torch?","Body":"<p>My machine's GPU has 2 GBs of memory. When I run the following code for the first time, I get no errors. However, the second time I run the code I get memory error. As a short-time remedy, the only thing I can do is to cast the data into float32 using . However, the problem still persists and the occupied memory is not released after the process is done, or the process is terminated while running. This is the case for the machine RAM as well. How should one prevent memory leakage in Torch or release the memory?<\/p>\n\n<pre><\/pre>\n\n<p>Side question: Why Torch initializes the network parameters as double although GPUs are pretty slow at computing double-precision operations and there is actually no need of having 64-bit parameter values for almost all neural network applications? How can I initialize a model with float (32-bit) parameter vectors?<\/p>\n\n<p>I found the answer to the side question. You can easily make torch's default data type as float using the following at the beginning of your code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36489412","CreationDate":"2016-04-08T00:15:44.933","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>I could resolve the issue (almost) by upgrading from CUDA 6.5 to CUDA 7.5 on the machine which I was doing the above experiments. Now, for most of the time when the program crashes while running the GPU memory gets released. However, still sometimes it does not happen and I have to restart the machine.<\/p>\n\n<p>Also, I would do the followings in order to make sure the program clears the GPU memory when the program successfully runs:<\/p>\n\n<pre><code>net = nil\ntrainset = nil\ntestset = nil\npred = nil\ninputGrad = nil\ncriterion = nil\n\ncollectgarbage()\n<\/code><\/pre>\n"}]}
{"QuestionId":36392114,"AnswerCount":0,"Tags":"<python><macos><boost><caffe><pycaffe>","CreationDate":"2016-04-03T22:40:14.933","AcceptedAnswerId":null,"OwnerUserId":5179625.0,"Title":"Building boost via homebrew hangs","Body":"<p>I am trying to get Caffe running on OS X 10.11 but have been having trouble building build boost from source via homebrew, as <a href=\"http:\/\/caffe.berkeleyvision.org\/install_osx.html\" rel=\"nofollow\">the guide<\/a> recommends. I have a fresh install of OS X with command-line tools, homebrew (python and python3 installed via homebrew) and all dependencies listed in the guide installed (built protobuf successfully). I am using MKL and CUDA 7.5.\nAfter running the command<\/p>\n\n<pre><\/pre>\n\n<p>the install script runs, shows configuration, and says\n\"...updating 13743 targets...\"\nthen copies a bunch of files and says \"...updated 13743 targets...\"<\/p>\n\n<p>It then hangs. It's been about two hours now, with Activity Monitor showing a ruby process that comes and goes taking about 85% of the CPU, with about 10% CPU in use by User. No gcc process. I've uploaded the full log <a href=\"https:\/\/drive.google.com\/file\/d\/0B5CvB-unqP4tYU9MVmtVbmlvbEk\/view?usp=sharing\" rel=\"nofollow\">here<\/a>. Am I just being impatient? Fresh OS X install gives the same result.<\/p>\n","answers":[]}
{"QuestionId":36392952,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><tensorboard>","CreationDate":"2016-04-04T00:29:49.330","AcceptedAnswerId":null,"OwnerUserId":6104317.0,"Title":"TensorBoard: How to plot histogram for gradients?","Body":"<p>TensorBoard had the function to plot histograms of Tensors at session-time. I want a histogram for the gradients during training.<\/p>\n\n<p> returns a list a gradients.<\/p>\n\n<p>However,  accepts only Tensors, not lists of Tensors. <\/p>\n\n<p>For the time being, I made a work-around. I flatten all Tensors to a column vector and concatenate them:<\/p>\n\n<p><\/p>\n\n<p>What would be a better way to plot the histogram for the gradient?<\/p>\n\n<p>It seems a common thing to do, so I hope TensorFlow would have a dedicated function for this.<\/p>\n","answers":[{"AnswerId":"49180546","CreationDate":"2018-03-08T18:59:06.383","ParentId":null,"OwnerUserId":"2379009","Title":null,"Body":"<p>Another solution (based on <a href=\"https:\/\/www.quora.com\/Are-there-any-examples-of-tensorflow-that-shows-how-to-monitor-the-jacobian-and-or-the-hessian#uzPgw\" rel=\"noreferrer\">this quora answer<\/a>) is to access the gradients directly from the optimizer you are already using. <\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(..)\ngrads = optimizer.compute_gradients(loss)\ngrad_summ_op = tf.summary.merge([tf.summary.histogram(\"%s-grad\" % g[1].name, g[0]) for g in grads])\ngrad_vals = sess.run(fetches=grad_summ_op, feed_dict = feed_dict)\nwriter['train'].add_summary(grad_vals)\n<\/code><\/pre>\n"},{"AnswerId":"43948624","CreationDate":"2017-05-13T02:24:28.897","ParentId":null,"OwnerUserId":"7447161","Title":null,"Body":"<p>Following the suggestion from <a href=\"https:\/\/stackoverflow.com\/users\/728291\/user728291\">@user728291<\/a>, I was able to view gradients in tensorboard by using the the <code>optimize_loss<\/code> function as follows. \nThe function calling syntax for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/optimize_loss\" rel=\"noreferrer\">optimize_loss<\/a> is<\/p>\n\n<pre><code>optimize_loss(\nloss,\nglobal_step,\nlearning_rate,\noptimizer,\ngradient_noise_scale=None,\ngradient_multipliers=None,\nclip_gradients=None,\nlearning_rate_decay_fn=None,\nupdate_ops=None,\nvariables=None,\nname=None,\nsummaries=None,\ncolocate_gradients_with_ops=False,\nincrement_global_step=True\n)\n<\/code><\/pre>\n\n<p>The function requires <code>global_step<\/code> and is dependent on some other imports as shown next.<\/p>\n\n<pre><code>from tensorflow.python.ops import variable_scope\nfrom tensorflow.python.framework import dtypes\nfrom tensorflow.python.ops import init_ops\nglobal_step = variable_scope.get_variable(  # this needs to be defined for tf.contrib.layers.optimize_loss()\n      \"global_step\", [],\n      trainable=False,\n      dtype=dtypes.int64,\n      initializer=init_ops.constant_initializer(0, dtype=dtypes.int64))\n<\/code><\/pre>\n\n<p>Then replace your typical training operation<\/p>\n\n<pre><code>training_operation = optimizer.minimize(loss_operation)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>training_operation = tf.contrib.layers.optimize_loss(\n      loss_operation, global_step, learning_rate=rate, optimizer='Adam',\n      summaries=[\"gradients\"])\n<\/code><\/pre>\n\n<p>Then have a merge statement for your summaries<\/p>\n\n<pre><code>summary = tf.summary.merge_all()\n<\/code><\/pre>\n\n<p>Then in your tensorflow session at the end of each run\/epoch:<\/p>\n\n<pre><code>summary_writer = tf.summary.FileWriter(logdir_run_x, sess.graph) \nsummary_str = sess.run(summary, feed_dict=feed_dict)\nsummary_writer.add_summary(summary_str, i)\nsummary_writer.flush()  # evidently this is needed sometimes or scalars will not show up on tensorboard.\n<\/code><\/pre>\n\n<p>Where <code>logdir_run_x<\/code> is a different directory for each run. That way when TensorBoard runs, you can look at each run separately. The gradients will be under the histogram tab and will have the label <code>OptimizeLoss<\/code>. It will show all the weights, all the biases, and the <code>beta<\/code> parameter as histograms. <\/p>\n\n<p>UPDATE: Using tf slim, there is another way that also works and is perhaps cleaner. <\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(learning_rate = rate)\ntraining_operation = slim.learning.create_train_op(loss_operation, optimizer,summarize_gradients=True)\n<\/code><\/pre>\n\n<p>By setting <code>summarize_gradients=True<\/code>, which is not the default, you will then get gradient summaries for all weights. These will be viewable in Tensorboard under <code>summarize_grads<\/code><\/p>\n"}]}
{"QuestionId":36392966,"AnswerCount":2,"Tags":"<python><rgb><theano><conv-neural-network><keras>","CreationDate":"2016-04-04T00:31:59.130","AcceptedAnswerId":36405818.0,"OwnerUserId":5390189.0,"Title":"Keras CNN images and kernel size mismatch even after image transformation to fit","Body":"<p>I'm trying to run a CNN similar to the one in the <a href=\"http:\/\/keras.io\/examples\/\" rel=\"nofollow\">Keras documantation \"VGG-like convnet\"<\/a> but for a custom set of images and binary classification instead of a 10-class output.<\/p>\n\n<p>When I try to fit the CNN, i get this longwinded error that I assume is telling me my input image size is not the right size for the CNN input.<\/p>\n\n<pre><\/pre>\n\n<p>The thing is I thought I reshaped all my images to fit. My input is a stack of 4000 232x300 px RBG images and the output is an array of 4000 boolean values.<\/p>\n\n<p><strong>Input<\/strong>: <\/p>\n\n<p><strong>Output<\/strong>: <\/p>\n\n<p>This is the function to build the CNN<\/p>\n\n<pre><\/pre>\n\n<p>I've been banging my head against the wall with this long enough that I thought maybe someone else has had this problem. Any thoughts? Thanks in advance.<\/p>\n","answers":[{"AnswerId":"36405818","CreationDate":"2016-04-04T14:43:40.407","ParentId":null,"OwnerUserId":"3427404","Title":null,"Body":"<p>You specified input as <code>(depth,width,height)<\/code>. So the array you must must have dimensions <code>(N,depth,width,height)<\/code>, where N is the number of training examples. <\/p>\n\n<p>The input you are actually passing, <code>(4000, 232, 300, 3)<\/code>, doesn't match. It should be reshaped to be <code>(4000, depth, width, height)<\/code>. This means you'll have to resize each image, and reorder the axes.<\/p>\n"},{"AnswerId":"36418447","CreationDate":"2016-04-05T06:12:53.483","ParentId":null,"OwnerUserId":"5390189","Title":null,"Body":"<p>The above answer is correct: for posterity, my problem was solved by a simple:<\/p>\n\n<pre><code>im_list = im_list.transpose((0,3,1,2))\n<\/code><\/pre>\n"}]}
{"QuestionId":36393410,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-04T01:38:49.940","AcceptedAnswerId":36398195.0,"OwnerUserId":5774288.0,"Title":"Where do these nan come from?","Body":"<p>EDIT : <br \/>\nI had some nan in my data but the anwser is correct, you have to initialize your weight with some noise !\nThanks !<\/p>\n\n<p>I'm doing my first script with tensorflow. I had some problem printing value but now i got it.\nI wanted to try a simple Logistic regression to start, i'm working on kaggle titanic dataset.<\/p>\n\n<p>My problem is that i don't know why but i got some nan in my Weight and Bias so in my y (prediction) vector too ...<\/p>\n\n<p>EDIT :\nMy weight was initialized at 0 so i was a null gradient i guess.\nAccording to the answer provided i add <\/p>\n\n<pre><\/pre>\n\n<p>instead of <\/p>\n\n<pre><\/pre>\n\n<p>But i still got some issues. My b variable a y variable are still nan and when i tried the same thing for b i got the following error :\nValueError: No variables to optimize<br \/>\nI tried several way to assign my biais of a tensor [1, 1] but it looks like i'm missing something<br \/>\nlooks like y is nan because cross entropy is nan because b is nan... :(\nEND - EDIT<\/p>\n\n<p>i read this post (<a href=\"https:\/\/stackoverflow.com\/questions\/33922937\/why-does-tensorflow-return-nan-nan-instead-of-probabilities-from-a-csv-file\">Why does TensorFlow return [[nan nan]] instead of probabilities from a CSV file?<\/a>) who give me an hint, during my cross entropy calcul 0*log(0) return nan so i applied the solution given, that is to add 1e-50 like :<\/p>\n\n<p>cross_entropy = -tf.reduce_sum(y_*tf.log(y  + 1e-50))<\/p>\n\n<p>unfortunatly it wasn't the problem i guess, i still got nan everywhere :(<\/p>\n\n<p>this is the interisting (i guess) part of my very simple model :<\/p>\n\n<pre><\/pre>\n\n<p>It returns me a 0.0 accuracy of course and after 2 array of nan\nI tried to print values everywhere but nan everywhere :'(<\/p>\n\n<p>Do somebody has an idea ? I may forgot something or doing it wrong<\/p>\n\n<p>The thing is that i tried a similar script with mnist (google tutorial) with included data and it's works (no nan). I get my data with panda reading the csv file.<\/p>\n\n<p>Thanks for reading !<\/p>\n","answers":[{"AnswerId":"36398195","CreationDate":"2016-04-04T08:51:21.570","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You are getting a division by zero in <code>tf.nn.softmax<\/code> since your weight matrices are zero. Use a different normalization method, like <code>truncated_normal<\/code> from the MNIST example<\/p>\n"}]}
{"QuestionId":36393590,"AnswerCount":1,"Tags":"<numpy><gpu><theano><pycuda>","CreationDate":"2016-04-04T02:08:36.257","AcceptedAnswerId":null,"OwnerUserId":1045053.0,"Title":"sorting a numpy matrix on gpu","Body":"<p>I have a large matrix -  which I want to sort based on the 1st column. Since, it's a numpy matrix, I can use argsort to get the result<\/p>\n\n<p><\/p>\n\n<p>It takes about  to complete this step which seems a little to high to me.<\/p>\n\n<p>My question is if there is a way to speed up the computation using the GPU? I was looking at some libraries like pycuda, theano but I'm finding it hard to understand the documentation. Specially, the part around passing numpy arrays and matrices to the gpu. Any pointers will be extremely helpful.<\/p>\n","answers":[{"AnswerId":"36674486","CreationDate":"2016-04-17T08:49:41.537","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>There is no native sorting functionality in PyCUDA or scikit-cuda that I am aware of. However, efficient parallel sorting on the GPU is a mostly solved problem, and you can find state-of-the-art sorting algorithms in the Thrust template library, which ships with every version of the toolkit since about CUDA 4.0.<\/p>\n\n<p>The PyCUDA wiki includes a <a href=\"https:\/\/andreask.cs.illinois.edu\/PyCuda\/Examples\/ThrustInterop\" rel=\"nofollow\">complete example<\/a> of calling the Thrust sort algorithm from PyCUDA.<\/p>\n"}]}
{"QuestionId":36394130,"AnswerCount":0,"Tags":"<python><gpu><tensorflow><memcpy>","CreationDate":"2016-04-04T03:28:44.000","AcceptedAnswerId":null,"OwnerUserId":1930168.0,"Title":"TensorFlow Error: \"CopyCPUTensorToGPU: GPU Memcpy failed\"","Body":"<p>I'm running the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/deepdream\/deepdream.ipynb\" rel=\"nofollow\"><\/a> example that came with TensorFlow. The line  gives me .<\/p>\n\n<p>One possibility is that it's a mismatch between the notebook which is from TensorFlow HEAD and my TensorFlow installation which is from a recent pip release.<\/p>\n\n<p>But if that were the case I would expect many more folks to have the same issue. It seems like this is a very rare bug because a <a href=\"https:\/\/www.google.com\/search?q=CopyCPUTensorToGPU%3A%20GPU%20Memcpy%20failed&amp;oq=CopyCPUTensorToGPU%3A%20GPU%20Memcpy%20failed&amp;aqs=chrome..69i57j69i58j69i60.198j0j1&amp;sourceid=chrome&amp;ie=UTF-8\" rel=\"nofollow\">Google search<\/a> for  yields a grand total of 5 results.<\/p>\n\n<p>Second possibility is that this is an issue with the specific machine that I'm running on, EC2's g2.2xlarge.<\/p>\n\n<p>Stack: EC2's g2.2xlarge machine, Python 2, Jupyter<\/p>\n\n<p>Complete trace:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36398041,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2016-04-04T08:43:15.827","AcceptedAnswerId":null,"OwnerUserId":6145309.0,"Title":"How to generate a sample sentence with LSTM model in Tensorflow?","Body":"<p>I'm working with the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html#recurrent-neural-networks\" rel=\"nofollow\">LSTM model in Tensorflow<\/a>.<br>\nI already trained and saved the LSTM model. Now I'm coming up to the last task to generate the sentences. \nHere is my pseudo code:<\/p>\n\n<pre><\/pre>\n\n<p><strong>My question is:<\/strong><br>\nWhen training, validating, or testing model I have to feed both of the <em>inputs<\/em> and their <em>labels<\/em> (by shifted inputs one) into model via <em>feed_dict<\/em> dictionary. But in the generating task, I have only one input which is the generating sentence <em>new_sentence<\/em>.  <\/p>\n\n<p><strong>How can I build the right <em>get_logits<\/em> function or full <em>generate<\/em> function also?<\/strong><\/p>\n","answers":[{"AnswerId":"42194129","CreationDate":"2017-02-12T22:38:57.077","ParentId":null,"OwnerUserId":"7523983","Title":null,"Body":"<p>when you train you have an output of the neural network, based on that output you calculate the error, based on error you create the optimizer to minimize the error. <\/p>\n\n<p>In order to generate a new sentence you need to get just the output of the neural network(rnn). <\/p>\n\n<p>Edited: <\/p>\n\n<pre><code>\"\"\"\nPlaceholders\n\"\"\"\n\nx = tf.placeholder(tf.int32, [batch_size, num_steps], name='input_placeholder')\ny = tf.placeholder(tf.int32, [batch_size, num_steps], name='labels_placeholder')\ninit_state = tf.zeros([batch_size, state_size])\n\n\"\"\"\nRNN Inputs\n\"\"\"\n\n# Turn our x placeholder into a list of one-hot tensors:\n# rnn_inputs is a list of num_steps tensors with shape [batch_size, num_classes]\nx_one_hot = tf.one_hot(x, num_classes)\nrnn_inputs = tf.unpack(x_one_hot, axis=1)\n\n\"\"\"\nDefinition of rnn_cell\n\nThis is very similar to the __call__ method on Tensorflow's BasicRNNCell. See:\nhttps:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn_cell.py\n\"\"\"\nwith tf.variable_scope('rnn_cell'):\n    W = tf.get_variable('W', [num_classes + state_size, state_size])\n    b = tf.get_variable('b', [state_size], initializer=tf.constant_initializer(0.0))\n\ndef rnn_cell(rnn_input, state):\n    with tf.variable_scope('rnn_cell', reuse=True):\n        W = tf.get_variable('W', [num_classes + state_size, state_size])\n        b = tf.get_variable('b', [state_size], initializer=tf.constant_initializer(0.0))\n    return tf.tanh(tf.matmul(tf.concat(1, [rnn_input, state]), W) + b)\n\nstate = init_state\nrnn_outputs = []\nfor rnn_input in rnn_inputs:\n    state = rnn_cell(rnn_input, state)\n    rnn_outputs.append(state)\nfinal_state = rnn_outputs[-1]\n\n#logits and predictions\nwith tf.variable_scope('softmax'):\n    W = tf.get_variable('W', [state_size, num_classes])\n    b = tf.get_variable('b', [num_classes], initializer=tf.constant_initializer(0.0))\nlogits = [tf.matmul(rnn_output, W) + b for rnn_output in rnn_outputs]\npredictions = [tf.nn.softmax(logit) for logit in logits]\n\n# Turn our y placeholder into a list labels\ny_as_list = [tf.squeeze(i, squeeze_dims=[1]) for i in tf.split(1, num_steps, y)]\n\n#losses and train_step\nlosses = [tf.nn.sparse_softmax_cross_entropy_with_logits(logit,label) for \\\n          logit, label in zip(logits, y_as_list)]\ntotal_loss = tf.reduce_mean(losses)\ntrain_step = tf.train.AdagradOptimizer(learning_rate).minimize(total_loss)\n def train():\n  with tf.Session() as sess:\n    #load the model\n    training_losses = []\n    for idx, epoch in enumerate(gen_epochs(num_epochs, num_steps)):\n        training_loss = 0\n        training_state = np.zeros((batch_size, state_size))\n        if verbose:\n            print(\"\\nEPOCH\", idx)\n        for step, (X, Y) in enumerate(epoch):\n            tr_losses, training_loss_, training_state, _ = \\\n                sess.run([losses,\n                          total_loss,\n                          final_state,\n                          train_step],\n                              feed_dict={x:X, y:Y, init_state:training_state})\n            training_loss += training_loss_\n            if step % 100 == 0 and step &gt; 0:\n                if verbose:\n                    print(\"Average loss at step\", step,\n                          \"for last 250 steps:\", training_loss\/100)\n                training_losses.append(training_loss\/100)\n                training_loss = 0\n     #save the model\n\ndef generate_seq():\n  with tf.Session() as sess:\n    #load the model\n    # load saved model for generating task\n    new_sentence = [START_TOKEN]\n    # Here I want to generate a sentence until END_TOKEN is generated.\n    while new_sentence[-1] != END_TOKEN:\n      logits = sess.run(final_state,{x:np.asarray([new_sentence])})\n      # get argmax(logits) or sample(logits)\n      next_word = argmax(logits[0])\n      new_sentence.append(next_word)\n  print(new_sentence)\n<\/code><\/pre>\n"}]}
{"QuestionId":36399337,"AnswerCount":1,"Tags":"<debugging><tensorflow>","CreationDate":"2016-04-04T09:48:53.887","AcceptedAnswerId":36405902.0,"OwnerUserId":2492058.0,"Title":"debugging Tensorflow's C++ code behind the SWIG interface","Body":"<p>I'm not sure how to debug (presumably with GDB) the Python code behind a SWIG interface.<\/p>\n\n<p>I can use  to watch the execution of Tensorflow's Python code all the way to the SWIG wrapper (e.g.  in ), but I would like to debug the C++ code behind the SWIG interface.<\/p>\n\n<p>Presumably I build Tensorflow with , but how do I attach  to the resulting code when called from the Python interface?<\/p>\n","answers":[{"AnswerId":"36405902","CreationDate":"2016-04-04T14:48:07.977","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow's C++ code executes in the same process as the Python code that calls it (or, if you are using the distributed version, in the same process as one of the Python programs that created a <code>tf.GrpcServer<\/code>).<\/p>\n\n<p>The simplest interface between Python and C++ is the pure-C API in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/public\/tensor_c_api.h\" rel=\"noreferrer\"><code>tensor_c_api.h<\/code><\/a>. To intercept one of these calls, you can attach <code>gdb<\/code> to the process ID of the Python interpreter that is running TensorFlow, and create a breakpoint on one of these functions.<\/p>\n\n<p>For example, using an interactive Python session, in the first terminal enter:<\/p>\n\n<pre><code>$ python\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; import os\n&gt;&gt;&gt; os.getpid()\n14680\n<\/code><\/pre>\n\n<p>Then, in another terminal, start <code>gdb<\/code>:<\/p>\n\n<pre><code>$ gdb -p 14680\n[...]\n(gdb) break TF_NewSession\nBreakpoint 1 at 0x7f15f450a4d0\n(gdb) continue\nContinuing.\n<\/code><\/pre>\n\n<p>Back in the Python interpreter, create a new session:<\/p>\n\n<pre><code>&gt;&gt;&gt; sess = tf.Session()\n<\/code><\/pre>\n\n<p>The interpreter will pause, and your debugger will print something like the following:<\/p>\n\n<pre><code>Breakpoint 1, 0x00007f15f450a4d0 in TF_NewSession () from [...]\/tensorflow\/python\/_pywrap_tensorflow.so\n(gdb) backtrace\n#0  0x00007f15f450a4d0 in TF_NewSession () from [...]\/tensorflow\/python\/_pywrap_tensorflow.so\n#1  0x00007f15f3ac5cdb in _wrap_TF_NewSession () from [...]\/tensorflow\/python\/_pywrap_tensorflow.so\n#2  0x000000000049968d in PyEval_EvalFrameEx ()\n#3  0x00000000004a090c in PyEval_EvalCodeEx ()\n#4  0x0000000000499a52 in PyEval_EvalFrameEx ()\n[...]\n<\/code><\/pre>\n\n<p>You can now use the full power of <code>gdb<\/code> to debug TensorFlow.<\/p>\n"}]}
{"QuestionId":36400150,"AnswerCount":3,"Tags":"<python><tensorflow><image-segmentation><image-masking>","CreationDate":"2016-04-04T10:27:38.343","AcceptedAnswerId":null,"OwnerUserId":1007618.0,"Title":"Image Segmentation with TensorFlow","Body":"<p>I am trying to see the feasibility of using TensorFlow to identify features in my image data. I have 50x50px grayscale images of nuclei that I would like to have segmented- the desired output would be either a 0 or 1 for each pixel.  0 for the background, 1 as the nucleus.<\/p>\n\n<p>Example input: <a href=\"http:\/\/i.stack.imgur.com\/OvoWa.jpg\" rel=\"nofollow\">raw input data<\/a><\/p>\n\n<p>Example label (what the \"label\"\/real answer would be): <a href=\"http:\/\/i.stack.imgur.com\/CR2aL.jpg\" rel=\"nofollow\">output data (label)<\/a><\/p>\n\n<p>Is it even possible to use TensorFlow to perform this type of machine learning on my dataset?  I could potentially have thousands of images for the training set.   <\/p>\n\n<p>A lot of the examples have a label correspond to a single category, for example, a 10 number array [0,0,0,0,0,0,0,0,0,0,0] for the handwritten digit data set, but I haven't seen many examples that would output a larger array.  I would assume I the label would be a 50x50 array? <\/p>\n\n<p>Also, any ideas on the processing CPU time for this time of analysis?  <\/p>\n","answers":[{"AnswerId":"39646179","CreationDate":"2016-09-22T18:17:47.623","ParentId":null,"OwnerUserId":"1533097","Title":null,"Body":"<p>Try using a convolutional filters for the model. A stacking of convolution and downsampling layers. The input should be the normalized pixel image and output should be the mask. The last layer should be a softmaxWithLoss. HTH.<\/p>\n"},{"AnswerId":"37196403","CreationDate":"2016-05-12T20:01:49.217","ParentId":null,"OwnerUserId":"1913898","Title":null,"Body":"<p>SoftmaxWithLoss() works for your image segmentation problem, if you reshape the predicted label and true label map from [batch, height, width, channel] to [N, channel]. <\/p>\n\n<p>In your case, your final predicted map will be channel = 2, and after reshaping, N = batch<em>height<\/em>width, then you can use SoftmaxWithLoss() or similar loss function in tensorflow to run the optimization. <\/p>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/questions\/36850531\/per-pixel-softmax-for-fully-convolutional-network\/36850726#36850726\">this question<\/a> that may help. <\/p>\n"},{"AnswerId":"36409074","CreationDate":"2016-04-04T17:27:31.287","ParentId":null,"OwnerUserId":"1930168","Title":null,"Body":"<p><strong>Yes<\/strong>, this is possible with TensorFlow. In fact, there are many ways to approach it. Here's a very simple one:<\/p>\n\n<p>Consider this to be a binary classification task. Each pixel needs to be classified as foreground or background. Choose a set of features by which each pixel will be classified. These features could be local features (such as a patch around the pixel in question) or global features (such as the pixel's location in the image). Or a combination of the two.<\/p>\n\n<p>Then train a model of your choosing (such as a NN) on this dataset. Of course your results will be highly dependant upon your choice of features.<\/p>\n\n<hr>\n\n<p>You could also take a <a href=\"http:\/\/link.springer.com\/article\/10.1007\/s11263-006-7934-5#\/page-1\" rel=\"nofollow\">graph-cut approach<\/a> if you can represent that computation as a computational graph using the primitives that TensorFlow provides. You could then either not make use of TensorFlow's optimization functions such as backprop or if there are some differentiable variables in your computation you could use TF's optimization functions to optimize those variables.<\/p>\n"}]}
{"QuestionId":36402855,"AnswerCount":0,"Tags":"<python><math><scipy><theano><minimize>","CreationDate":"2016-04-04T12:36:16.507","AcceptedAnswerId":null,"OwnerUserId":5427064.0,"Title":"Is there any Theano function to minimize\/optimize an expression?","Body":"<p>Any equivalent function to this one to minimize an expression in 1 line: \n<a href=\"http:\/\/docs.scipy.org\/doc\/scipy-0.15.1\/reference\/generated\/scipy.optimize.minimize.html\" rel=\"nofollow\">http:\/\/docs.scipy.org\/doc\/scipy-0.15.1\/reference\/generated\/scipy.optimize.minimize.html<\/a><\/p>\n\n<pre><\/pre>\n\n<p>also tensorflow functions are welcome.<\/p>\n","answers":[]}
{"QuestionId":36403865,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-04T13:21:00.443","AcceptedAnswerId":null,"OwnerUserId":3320135.0,"Title":"TensorFlow: defining variables in global scope","Body":"<p>I'm building an image classifier in TensorFlow, and there is a class imbalance in my training data. Therefore, when computing the loss, I need to weight the loss for each class by the inverse frequency of that class in the training data.<\/p>\n\n<p>So, here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>My question is: Can I store  as just a  in global scope? Or do I need to pass it as a parameter when computing ?<\/p>\n\n<p>The reason I am wondering is that  is different for every batch. Therefore, I am worried that if it is just defined in global scope, then when the Tensor Flow graph is constructed, it just takes the initial values in , and then never updates them. Whereas if I were to pass  using the  when computing , then I am explicitly telling Tensor Flow to use the recent, updated values in .<\/p>\n\n<p>Any help would be appreciated. Thanks! <\/p>\n","answers":[{"AnswerId":"36405441","CreationDate":"2016-04-04T14:28:00.073","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>I think having <code>class_weights<\/code> a tf.constant is fine. Class weighting should be done for the whole dataset, not per mini-batch.<\/p>\n\n<p>Also another approach to this you might want to consider is sampling so each batch has equal numbers of each class?<\/p>\n"}]}
{"QuestionId":36405677,"AnswerCount":1,"Tags":"<python><numpy><optimization><caffe><pycaffe>","CreationDate":"2016-04-04T14:38:11.263","AcceptedAnswerId":36422577.0,"OwnerUserId":2527680.0,"Title":"Caffe feature extraction is too slow? caffe.Classifier or caffe.Net","Body":"<p>I have trained a model with images.\nAnd now would like to extract the  features to  files.\nI'm using to run the   and extract the features.<\/p>\n\n<p>Instead of extracting and saving the feature per frame.\nI save all the features of a folder to a temp variable and the result of the complete video to a npy file(decreasing the number of write operations to disk).<\/p>\n\n<p>I have also heard that I could use the Caffe.Net and then pass a batch of images. But I'm not sure of what preprocessing has to be done and if this is faster ?<\/p>\n\n<pre><\/pre>\n\n<h1>Output<\/h1>\n\n<pre><\/pre>\n\n<h1>Time needed per image Around 0.9 Seconds now-<\/h1>\n","answers":[{"AnswerId":"36422577","CreationDate":"2016-04-05T09:39:31.727","ParentId":null,"OwnerUserId":"2527680","Title":null,"Body":"<p>I found the best answer here in <a href=\"http:\/\/sites.duke.edu\/rachelmemo\/2015\/05\/21\/setup-a-network-in-caffes-python-wrapper\/\" rel=\"nofollow\">this post.<\/a><\/p>\n\n<p>Till now I had used a <\/p>\n\n<pre><code>net = caffe.Classifier(model_prototxt, model_trained,\n                           mean=np.array([128, 128, 128]),\n                           channel_swap=(2,1,0),\n                           raw_scale=255,\n                           image_dims=(255, 255))\n<\/code><\/pre>\n\n<p>to initialize a model and get the output per image.\nBut this method is really slow and requires around .9 seconds per image.<\/p>\n\n<p>The best Idea is to pass a batch of images(maybe 100,200,250) changing. Depending on how much memory you have on your GPU.<\/p>\n\n<p>for this I set <code>caffe.set_mode_gpu()<\/code> as I have one and It's faster when you send large batches.\nInitialize the model with ur trained model.<\/p>\n\n<pre><code>net=caffe.Net(model_prototxt,model_trained,caffe.TEST)\n<\/code><\/pre>\n\n<p>Create a Transformer and make sure to set mean and other values depending on how u trained your model.<\/p>\n\n<pre><code>transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})\ntransformer.set_transpose('data', (2,0,1)) # height*width*channel -&gt; channel*height*width\nmean_file = np.array([128, 128, 128])\ntransformer.set_mean('data', mean_file) #### subtract mean ####\ntransformer.set_raw_scale('data', 255) # pixel value range\ntransformer.set_channel_swap('data', (2,1,0)) # RGB -&gt; BGR\ndata_blob_shape = net.blobs['data'].data.shape\ndata_blob_shape = list(data_blob_shape)\n<\/code><\/pre>\n\n<p>Read a group of images and convert to the network input.<\/p>\n\n<pre><code>net.blobs['data'].reshape(len(all_images), data_blob_shape[1], data_blob_shape[2], data_blob_shape[3])\nimages = [temp_pres_dir+str(x) for x in all_images]\nnet.blobs['data'].data[...] = map(lambda x: \ntransformer.preprocess('data',caffe.io.load_image(x)), images)\n<\/code><\/pre>\n\n<p>Pass the batch of images through network.<\/p>\n\n<pre><code>out = net.forward()\n<\/code><\/pre>\n\n<p>You can use this output as you wish.<\/p>\n\n<h1>Speed for each image is now 20 msec<\/h1>\n"}]}
{"QuestionId":36405726,"AnswerCount":0,"Tags":"<c++><linker><g++><caffe><torcs>","CreationDate":"2016-04-04T14:39:46.113","AcceptedAnswerId":null,"OwnerUserId":6151543.0,"Title":"Linking caffe library and dependencies g++ make","Body":"<p>I'm trying to add caffe to an existing project. I get 99 undefined reference errors, suggesting that it has something to do with linking libraries. The errors mention caffe, boost and google (relates to gflags or glog).<\/p>\n\n<p>Linker command: (libcaffe is in \/home\/torcs\/lib)<\/p>\n\n<pre><\/pre>\n\n<p>Linker error (first 2 and final 2):<\/p>\n\n<pre><\/pre>\n\n<p>I added  and tried  and  separately but the number of errors is not decreasing as I add libraries. If I misspel  it says cannot find , which it doesn't do otherwise, so the lib is included correctly I guess.<\/p>\n\n<p>I read up on linker order (windows background) and found that the symbols only get added if they are on the required symbols list. More specific libs should be last, dependent libs should be first. I figured that if I only add  at the end, the unrecognized symbol list should be populated, and at least the caffe symbols would be recognized, and maybe replaced by other dependencies. But that doesn't happen.<\/p>\n\n<p>I tried to find whether  supplies the references. For the final undefined reference (),  finds:<\/p>\n\n<pre><\/pre>\n\n<p>Which suggest the reference is in the lib. But the undefined reference error doesn't go away.<\/p>\n\n<p>I also tried adding boost libs and glog, didn't help.<\/p>\n\n<p>Edit. There is some additional weird behaviour going on. I think it is unrelated, but am not sure. The first time I build after a  I get an error about :<\/p>\n\n<pre><\/pre>\n\n<p>This prevents the  files and  from getting built. But if I  again, the c++11 error disappears, this lib does get build, and I see the 99 undefined reference errors.<\/p>\n","answers":[]}
{"QuestionId":36406723,"AnswerCount":1,"Tags":"<neural-network><deep-learning><torch><recurrent-neural-network>","CreationDate":"2016-04-04T15:25:32.793","AcceptedAnswerId":36407076.0,"OwnerUserId":5676466.0,"Title":"General rules for training the RNN model when loss stops decreasing","Body":"<p>I have an RNN model. After about 10K iterations, the loss stops decreasing, but the loss is not very small yet. Does it always mean the optimization is trapped in a local minimum? <\/p>\n\n<p>In general, what would be the actions should I take to address this issue? Add more training data? Change a different optimization scheme (SGD now)? Or Other options?<\/p>\n\n<p>Many thanks!<\/p>\n\n<p>JC<\/p>\n","answers":[{"AnswerId":"36407076","CreationDate":"2016-04-04T15:39:54.903","ParentId":null,"OwnerUserId":"316799","Title":null,"Body":"<p>If you are training you neural network using a gradient vector based algorithm such as <code>Back Propagation<\/code> or <code>Resilient Propagation<\/code> it can stop improving when it finds a local minimum and it is normal because of the nature of this type fo algorithm. In this case, the propagation algorithms is used to search what a (gradient) vector is pointing. <\/p>\n\n<p>As a suggestion you could add a different strategy during the training to explore the space of search instead only searching. For sample, a <code>Genetic Algorithm<\/code> or the <code>Simulated Annealing<\/code> algorithm. These approaches will provide a exploration of possibilities and it can find a global minimum. You could implement 10 itegrations for each 200 iterations of the propagation algorithm, creating a hybrid strategy. For sample (it's just a pseudo-code):<\/p>\n\n<pre><code>int epochs = 0;\ndo\n{\n   train();\n\n   if (epochs % 200 == 0)\n      traingExplorativeApproach();\n\n   epochs++;\n} while (epochs &lt; 10000);\n<\/code><\/pre>\n\n<p>I've developed a strategy like this using <code>Multi-Layer Perceptrons<\/code> and <code>Elman recurrent neural network<\/code> in <code>classification<\/code> and <code>regressions<\/code> problems and both cases a hybrid strategy has provided better results then a single propagation training. <\/p>\n"}]}
{"QuestionId":36407124,"AnswerCount":1,"Tags":"<lua><logical-operators><torch>","CreationDate":"2016-04-04T15:42:22.677","AcceptedAnswerId":36407177.0,"OwnerUserId":987185.0,"Title":"Logical negation of ByteTensor in torch","Body":"<p>Is there a way of taking a logical negation (i.e. performing a NOT operation) on a  in Torch:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36407177","CreationDate":"2016-04-04T15:44:54.347","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Via <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torchnea-b\" rel=\"nofollow\">torch.ne<\/a> (not-equal):<\/p>\n\n<pre><code> a[b:ne(1)]\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong>:<\/p>\n\n<p>Or equivalently via <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/maths.md#torcheqa-b\" rel=\"nofollow\">torch.eq<\/a>:<\/p>\n\n<pre><code> a[b:eq(0)]\n<\/code><\/pre>\n"}]}
{"QuestionId":36409874,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-04-04T18:12:44.420","AcceptedAnswerId":null,"OwnerUserId":3320135.0,"Title":"Defining constants in TensorFlow","Body":"<p>I am using Python 2.7 with the TensorFlow library. In this library, there is a data type called , which means a constant float of value 3. The value needs to be assigned at initialisation and cannot be changed, similar to a  in C++.<\/p>\n\n<p>In my code however, I do not know the value of this constant at the start. Instead, I need to call a function, which does some processing and then creates the constant. For example, if the constant is called :<\/p>\n\n<pre><\/pre>\n\n<p>The problem with this, is that  is now not available outside the scope of the function .<\/p>\n\n<p>So, I want to do something like:<\/p>\n\n<pre><\/pre>\n\n<p>But this is not possible, because the constant can only be defined once.<\/p>\n\n<p>How can I solve this?<\/p>\n","answers":[{"AnswerId":"36409969","CreationDate":"2016-04-04T18:18:38.227","ParentId":null,"OwnerUserId":"4463948","Title":null,"Body":"<p>I'm not sure how your program looks like, but how about making a function that returns a constant?<\/p>\n\n<pre><code>def initialize_x():\n    # Do something\n    y = ...\n    return tf.constant(y)\n\nif __name__ == \"__main__\":\n    ...\n    x = initialize_x()\n    ...\n<\/code><\/pre>\n"}]}
{"QuestionId":36410926,"AnswerCount":1,"Tags":"<python><machine-learning><theano><conv-neural-network><lasagne>","CreationDate":"2016-04-04T19:10:24.427","AcceptedAnswerId":36443955.0,"OwnerUserId":4846917.0,"Title":"Convolutional neural network: how to train it? (unsupervised)","Body":"<p>I'm trying to implement a CNN to play a game. \nI'm using python with theano\/lasagne. I've build the network and am now figuring out how to train it. <\/p>\n\n<p>So now I have a batch of 32 states and for each state in that batch the <strong>action<\/strong> and the <strong>expected rewards<\/strong> for that action.<\/p>\n\n<p>Now how can I train the network so that it learn that these actions in these states lead to these rewards?<\/p>\n\n<p>EDIT: Clarifying my problem.<\/p>\n\n<p>Here is my full code: <a href=\"http:\/\/pastebin.com\/zY8w98Ng\" rel=\"nofollow\">http:\/\/pastebin.com\/zY8w98Ng<\/a>\nThe snake import: <a href=\"http:\/\/pastebin.com\/fgGCabzR\" rel=\"nofollow\">http:\/\/pastebin.com\/fgGCabzR<\/a><\/p>\n\n<p>I'm having trouble with this bit:<\/p>\n\n<pre><\/pre>\n\n<p>I want to update the model using previous_states, actions and agent_expected_rewards so that it learn that those actions lead to those rewards.<\/p>\n\n<p>I expect it might look something like this:<\/p>\n\n<pre><\/pre>\n\n<p>I just don't get how the givens would effect the model because when building the network I don't specify them. I can't find it in the theano and lasagne documentation either. <\/p>\n\n<p>So how can I update the model\/network so that it 'learns'.<\/p>\n\n<p>If its still not clear, comment on what information is still needed. I've been trying to figure this out for a few days now.<\/p>\n","answers":[{"AnswerId":"36443955","CreationDate":"2016-04-06T07:12:08.513","ParentId":null,"OwnerUserId":"4846917","Title":null,"Body":"<p>After going through the documentation I've finally found the answer. I was looking in the wrong places before.<\/p>\n\n<pre><code>    network = self._output_layer\n    prediction = lasagne.layers.get_output(network)\n    loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)\n    loss = loss.mean()\n\n    params = lasagne.layers.get_all_params(network, trainable=True)\n    updates = lasagne.updates.sgd(loss, params, self.LEARN_RATE)\n    givens = {\n        states: current_states,\n        expected: agents_expected_reward,\n        real_rewards: rewards\n    }\n    train_fn = theano.function([input_var, target_var], loss,\n                                    updates=updates, on_unused_input='warn',\n                                    givens=givens,\n                                    allow_input_downcast='True')\n    train_fn(current_states, agents_expected_reward)\n<\/code><\/pre>\n"}]}
{"QuestionId":36412065,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow><conv-neural-network>","CreationDate":"2016-04-04T20:14:11.087","AcceptedAnswerId":36412802.0,"OwnerUserId":2052105.0,"Title":"Get weights from tensorflow model","Body":"<p>Hello I would like to finetune VGG model from tensorflow. I have two questions.<\/p>\n\n<p>How to get the weights from network? The trainable_variables returns empty list for me.<\/p>\n\n<p>I used existing model from here: <a href=\"https:\/\/github.com\/ry\/tensorflow-vgg16\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ry\/tensorflow-vgg16<\/a> .\nI find the post about getting weights however this doesn't work for me because of import_graph_def. <a href=\"https:\/\/stackoverflow.com\/questions\/36193553\/get-the-value-of-some-weights-in-a-model-trained-by-tensorflow\">Get the value of some weights in a model trained by TensorFlow<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36412802","CreationDate":"2016-04-04T21:02:29.530","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This <a href=\"https:\/\/github.com\/ry\/tensorflow-vgg16\" rel=\"noreferrer\">pretrained VGG-16 model<\/a> encodes all of the model parameters as <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/constant_op.html#constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> ops. (See, for example, the calls to <code>tf.constant()<\/code> <a href=\"https:\/\/github.com\/ry\/tensorflow-vgg16\/blob\/97a9e1ff663ac7238f986d8b8247f4ab81f83432\/caffe_to_tensorflow.py#L88\" rel=\"noreferrer\">here<\/a>.) As a result, the model parameters would not appear in <code>tf.trainable_variables()<\/code>, and the model is not mutable without substantial surgery: you would need to replace the constant nodes with <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> objects that start with the same value in order to continue training.<\/p>\n\n<p>In general, when importing a graph for retraining, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#import_meta_graph\" rel=\"noreferrer\"><code>tf.train.import_meta_graph()<\/code><\/a> function should be used, as this function loads additional metadata (including the collections of variables). The <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/framework.html#import_graph_def\" rel=\"noreferrer\"><code>tf.import_graph_def()<\/code><\/a> function is lower level, and does not populate these collections.<\/p>\n"}]}
{"QuestionId":36412098,"AnswerCount":5,"Tags":"<c++><protocol-buffers><tensorflow><keras>","CreationDate":"2016-04-04T20:16:07.040","AcceptedAnswerId":null,"OwnerUserId":915501.0,"Title":"Convert Keras model to TensorFlow protobuf","Body":"<p>We're currently training various neural networks using Keras, which is ideal because it has a nice interface and is relatively easy to use, but we'd like to be able to apply them in our production environment. <\/p>\n\n<p>Unfortunately the production environment is C++, so our plan is to:<\/p>\n\n<ul>\n<li>Use the TensorFlow backend to save the model to a protobuf<\/li>\n<li>Link our production code to TensorFlow, and then load in the protobuf<\/li>\n<\/ul>\n\n<p>Unfortunately I don't know how to access the TensorFlow saving utilities from Keras, which normally saves to HDF5 and JSON. How do I save to protobuf?<\/p>\n","answers":[{"AnswerId":"36785508","CreationDate":"2016-04-22T05:26:01.517","ParentId":null,"OwnerUserId":"6133502","Title":null,"Body":"<p>You can access TensorFlow backend by:<\/p>\n\n<pre><code>import keras.backend.tensorflow_backend as K\n<\/code><\/pre>\n\n<p>Then you can call any TensorFlow utility or function like:<\/p>\n\n<pre><code>K.tf.ConfigProto\n<\/code><\/pre>\n"},{"AnswerId":"43884259","CreationDate":"2017-05-10T05:33:07.453","ParentId":null,"OwnerUserId":"5465000","Title":null,"Body":"<p>Save your keras model as an HDF5 file.<\/p>\n\n<p>You can then do the conversion with the following code:<\/p>\n\n<pre><code>from keras import backend as K\nfrom tensorflow.python.framework import graph_util\nfrom tensorflow.python.framework import graph_io\n\nweight_file_path = 'path to your keras model'\nnet_model = load_model(weight_file_path)\nsess = K.get_session()\n\nconstant_graph = graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), 'name of the output tensor')\ngraph_io.write_graph(constant_graph, 'output_folder_path', 'output.pb', as_text=False)\nprint('saved the constant graph (ready for inference) at: ', osp.join('output_folder_path', 'output.pb'))\n<\/code><\/pre>\n\n<p>Here is my sample code which handles multiple input and multiple output cases:\n<a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amir-abdi\/keras_to_tensorflow<\/a><\/p>\n"},{"AnswerId":"48082593","CreationDate":"2018-01-03T17:50:58.703","ParentId":null,"OwnerUserId":"1866775","Title":null,"Body":"<p>In case you don't need to utilize a GPU in the environment you are deploying to, you could also use my library, called frugally-deep. It is available on GitHub and published under the MIT License: <a href=\"https:\/\/github.com\/Dobiasd\/frugally-deep\" rel=\"noreferrer\">https:\/\/github.com\/Dobiasd\/frugally-deep<\/a><\/p>\n\n<p>frugally-deep allows running forward passes on already-trained Keras models directly in C++ without the need to link against TensorFlow or any other backend.<\/p>\n"},{"AnswerId":"37140568","CreationDate":"2016-05-10T13:52:16.253","ParentId":null,"OwnerUserId":"4255993","Title":null,"Body":"<p>This seems to be answered in <a href=\"http:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"nofollow noreferrer\">\"Keras as a simplified interface to TensorFlow: tutorial\"<\/a>, posted on <em>The Keras Blog<\/em> by Francois Chollet.<\/p>\n\n<p>In particular, <a href=\"http:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html#using-keras-models-with-tensorflow\" rel=\"nofollow noreferrer\">section II, \"Using Keras models with TensorFlow\"<\/a>.<\/p>\n"},{"AnswerId":"55453044","CreationDate":"2019-04-01T10:32:00.757","ParentId":null,"OwnerUserId":"4311535","Title":null,"Body":"<p>Make sure you change the <strong>learning phase of keras backend<\/strong> to store proper values of the layers (like dropout or batch normalization). Here is a <a href=\"https:\/\/ai-pool.com\/d\/problem_after_converting_keras_model_into_tensorflow_pb\" rel=\"nofollow noreferrer\">discussion<\/a> about it.<\/p>\n"}]}
{"QuestionId":36413949,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-04-04T22:22:58.853","AcceptedAnswerId":36421613.0,"OwnerUserId":2827214.0,"Title":"How can I see or save one tensor at runtime in Tensorflow?","Body":"<p>I have a model:<\/p>\n\n<pre><\/pre>\n\n<p>The model is called by:<\/p>\n\n<pre><\/pre>\n\n<p>How can I print\/save\/see what the \"states\" are, returned by rnn.rnn()?    <\/p>\n\n<p>I have tried tf.Print(states[-1], [states[-1]])  which gives me the shape of the tensor.<\/p>\n\n<pre><\/pre>\n\n<p>I have tried states[-1].eval() which provides a series of errors such as:   <\/p>\n\n<pre><\/pre>\n\n<p>I have also tried adding the var to the model to return it, but that did not work:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36421613","CreationDate":"2016-04-05T08:57:30.297","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>In order to see the value of a tensor in the eval method you can't have a dependency on any placeholders in the graph. In this case the error message tells you that <code>states[-1]<\/code> is dependent on <code>'encoder1'<\/code>. <\/p>\n\n<p>You can call seesion.run and feed in the value of the placeholder like so:<\/p>\n\n<pre><code>session.run(states[-1], feed_dict={encoder1:[#values for encoder1 here\n                                            ]})\n<\/code><\/pre>\n\n<p>Where <code>encoder1<\/code> is the placeholder object. This should return you the value of <code>states[-1]<\/code> which you can then serialize to save.<\/p>\n\n<p>In your specific case <code>encoder1<\/code> may be an internal placeholder within the <code>rnn<\/code> so you might want to run something like:<\/p>\n\n<pre><code>_, state_values = session.run([output_feed, states[-1]], input_feed)\n<\/code><\/pre>\n\n<p>To get the value of the variable in the context that your running it.<\/p>\n"}]}
{"QuestionId":36415325,"AnswerCount":0,"Tags":"<text><tensorflow><skflow><word-embedding>","CreationDate":"2016-04-05T00:52:00.237","AcceptedAnswerId":null,"OwnerUserId":5863915.0,"Title":"skflow load pre-train word embeddings into text_classification_cnn","Body":"<p><pre><\/p>\n\n<\/pre>\n\n<p><\/p>\n\n<p>this is a skflow example on convolutional text classification<\/p>\n\n<p>when I debug this patch of code, i can't explain how it works.\nHow to use pre-trained word embeddings instead of it?<\/p>\n","answers":[]}
{"QuestionId":36415784,"AnswerCount":1,"Tags":"<machine-learning><cuda><nvidia><deep-learning><caffe>","CreationDate":"2016-04-05T01:50:43.273","AcceptedAnswerId":36419298.0,"OwnerUserId":5196164.0,"Title":"Using the built-in graphics cards without a NVIDIA graphics card, Can I use the CUDA and Caffe library?","Body":"<p>Using the built-in graphics cards without a NVIDIA graphics card, Can I use the CUDA and Caffe library ?<\/p>\n\n<p>My O\/S is ubuntu 15\n   CPU is (Intel)i5-4670(3.40GHz 4 core)\n   RAM is 12.0GB<\/p>\n\n<p>I want to start to study deep learning.<\/p>\n","answers":[{"AnswerId":"36419298","CreationDate":"2016-04-05T07:01:26.320","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>CUDA is for NVidia graphics cards. Therefore you cannot make use of the CUDA built of Caffe in your PC.<\/p>\n\n<p>Still, you could make use of the CPU_ONLY flag during built and get Caffe working on your system. But you can expect the training to be too slow compared to what you should get when using a graphics card.<\/p>\n\n<p>There are various fork projects of Caffe which are made to make it work on other graphics card and are available on Git. But most of these projects are seriously outdated when compared with the master branch of original Caffe.<\/p>\n\n<p>An OpenCL implementation of Caffe can be found <a href=\"https:\/\/github.com\/amd\/OpenCL-caffe\" rel=\"nofollow\">here<\/a>.<\/p>\n"}]}
{"QuestionId":36416941,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-04-05T04:06:56.713","AcceptedAnswerId":null,"OwnerUserId":4941215.0,"Title":"Is there anyway to define placeholders for types rather than tf.dtypes in Tensorflow?","Body":"<p>The definition of placeholder is <\/p>\n\n<pre><\/pre>\n\n<p>but I was wondering is there anyway to put other types of data in the placeholder, instead of dtype data?<\/p>\n\n<p>In my case, I want to put a list of python strings in this placeholder, but I can only put tf.string, which is difficult to use for later processes. In addition, I can't find a way to transform tf.string to python string.<\/p>\n\n<p>Anyone can help? Thx!!<\/p>\n","answers":[]}
{"QuestionId":36420068,"AnswerCount":1,"Tags":"<android><torch>","CreationDate":"2016-04-05T07:42:34.883","AcceptedAnswerId":null,"OwnerUserId":6159772.0,"Title":"How to build torch-android","Body":"<p>I want to build torch-android on server(Ubuntu) in the light of the procedure showed on <a href=\"https:\/\/github.com\/soumith\/torch-android\" rel=\"nofollow\">https:\/\/github.com\/soumith\/torch-android<\/a>. But when I run the build.sh, the script can not find the GPU(the server has Tesla K80):<\/p>\n\n<pre><\/pre>\n\n<p>Open build.sh, we can find<\/p>\n\n<pre><\/pre>\n\n<p>I think I should modify the build.sh, but I have no idea of it. Does anyone know how to solve this problem? <\/p>\n","answers":[{"AnswerId":"37783224","CreationDate":"2016-06-13T06:39:28.510","ParentId":null,"OwnerUserId":"6458301","Title":null,"Body":"<p>Have you run these command in the ternimal?<\/p>\n\n<ol>\n<li><code>cd ~\/torch-android;<\/code><\/li>\n<li><code>git submodule update --init --recursive;<\/code><\/li>\n<li><code>.\/build;<\/code><\/li>\n<\/ol>\n"}]}
{"QuestionId":36425107,"AnswerCount":0,"Tags":"<computer-vision><neural-network><deep-learning><caffe><object-detection>","CreationDate":"2016-04-05T11:28:37.157","AcceptedAnswerId":null,"OwnerUserId":6080605.0,"Title":"Faster RCNN training get stuck in iteration 660","Body":"<p>I met one problem where I trained py-faster-rcnn on my own dataset, but it get stuck in iteration 660 for a long time and no response. Has anyone met this problem before? How to solve it? Thanks a lot!<\/p>\n\n<p>FYR:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n","answers":[]}
{"QuestionId":36426949,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-04-05T12:54:01.990","AcceptedAnswerId":null,"OwnerUserId":3280780.0,"Title":"How to prevent Tensorflow unpack() method casting to float64","Body":"<p>I'm trying to set up a sequential RNN in Tensorflow with seq2seq.rnn_decoder().  The input that rnn_decoder() wants is a list of tensors, so to generate this I've passed in a rank-3 tensor and used tf.unpack() to make it into a list.  The problem arises when the float32 array that I pass in in turned into a float64 tensor by tf.unpack(), making it incompatible with the rest of the model.   Here's the code I put together to convince me that the culprit is tf.unpack():<\/p>\n\n<pre><\/pre>\n\n<p>output (BATCH_SIZE is five, sequence length is ten):<\/p>\n\n<pre><\/pre>\n\n<p>Then I can load this data in a Tensorflow session:<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p>What's going on here?   Using a FOR loop to iterate through each of the sequential entries and re-cast it seems like the wrong way to do this, and I can't find a method inside Tensorflow to cast every member of a list.<\/p>\n","answers":[{"AnswerId":"43119046","CreationDate":"2017-03-30T13:17:21.180","ParentId":null,"OwnerUserId":"1082776","Title":null,"Body":"<p>You don't need a for-loop: you can use tf.cast().<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>input_tensor = tf.unpack(inputDat) # result is 64-bit\ninput_tensor = tf.cast(input_tensor, tf.float32) # now it's 32-bit\n<\/code><\/pre>\n"}]}
{"QuestionId":36427219,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-04-05T13:05:13.777","AcceptedAnswerId":36430319.0,"OwnerUserId":5576325.0,"Title":"Can I write a custom layer in tensorflow in python with existing ops such as conv2d and tensor operations?","Body":"<p>As stated in the title, I want to implement a custom layer in tensorflow with the existing ops and tensor operations. I wonder if I can do it in python just like what can be done in theano. In this layer, the input may be some matrices, the batch input and some weights and biases need to be learned. After some tensor operations, the output will be fed to the following layer. The computation in this layer may be complex, so I wonder if tensorflow can do auto-diff for me.<\/p>\n","answers":[{"AnswerId":"36430319","CreationDate":"2016-04-05T15:12:49.550","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>If your layer is a composition of existing ops, it will just work, sure. That's how for example TF-Slim works.<\/p>\n\n<pre><code># Skeleton code, just to demonstrate the concept\ndef conv(input, ...):\n    kernel = tf.Variable(...)\n    tmp = tf.nn.conv2d(input, kernel, ...)\n    bias = tf.Variable(...)\n    tmp = tf.nn.bias_add(tmp, bias, ...)\n    return tf.nn.relu(tmp, ...)\n<\/code><\/pre>\n\n<p>defines a function that gives you an \"integrated\" layer which performs the usual steps of a basic convolution layer, which you then can use as<\/p>\n\n<pre><code>layer_1 = conv(input, ...)\nlayer_2 = conv(layer_1, ...)\n<\/code><\/pre>\n\n<p>and so on. As long as you're just composing ops that have gradients, the automatic differentiation will work.<\/p>\n"}]}
{"QuestionId":36428157,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><lstm>","CreationDate":"2016-04-05T13:43:36.877","AcceptedAnswerId":36561191.0,"OwnerUserId":2527680.0,"Title":"Understanding Tensorflow LSTM models input?","Body":"<p>I have some trouble understanding LSTM models in TensorFlow.<\/p>\n\n<p>I use the <a href=\"http:\/\/tflearn.org\/\" rel=\"nofollow noreferrer\">tflearn<\/a> as a wrapper, as it does all the initialization and other higher level stuff automatically. For simplicity, let's consider <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/nlp\/lstm.py\" rel=\"nofollow noreferrer\">this example program<\/a>. Until <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/nlp\/lstm.py#L40\" rel=\"nofollow noreferrer\">line 42<\/a>, , it's pretty clear what happens. You load a dataset into variables and make it of a standard length (in this case, 200). Both the input variables and also the 2 classes are, in this case, converted to one-hot vectors.<\/p>\n\n<p><em>How does the LSTM take the input<\/em>? <em>Across how many samples does it predict the output?<\/em><\/p>\n\n<p><em>What does  represent?<\/em><\/p>\n\n<p>My goal is to <a href=\"http:\/\/jeffdonahue.com\/lrcn\/\" rel=\"nofollow noreferrer\">replicate<\/a> the activity recognition dataset in the <a href=\"http:\/\/arxiv.org\/abs\/1411.4389\" rel=\"nofollow noreferrer\">paper<\/a>. For example, I would like to input a 4096 vector as input to the LSTM, and the idea is to take 16 of such vectors, and then produce the classification result. I think the code would look like this, but I don't know how the input to the LSTM should be given.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36561191","CreationDate":"2016-04-11T23:44:01.017","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Basically, lstm takes the size of your vector for once cell:<\/p>\n\n<pre><code>lstm = rnn_cell.BasicLSTMCell(lstm_size, forget_bias=1.0)\n<\/code><\/pre>\n\n<p>Then, how many time series do you want to feed? It's up to your fed vector. The number of arrays in the <code>X_split<\/code> decides the number of time steps:<\/p>\n\n<pre><code>X_split = tf.split(0, time_step_size, X)\noutputs, states = rnn.rnn(lstm, X_split, initial_state=init_state)\n<\/code><\/pre>\n\n<p>In your example, I guess the <code>lstm_size<\/code> is 256, since it's the vector size of one word. The <code>time_step_size<\/code> would be the max word count in your training\/test sentences.<\/p>\n\n<p>Please see this example: <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/07_lstm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/07_lstm.py<\/a><\/p>\n"}]}
{"QuestionId":36428323,"AnswerCount":5,"Tags":"<machine-learning><neural-network><deep-learning><keras><recurrent-neural-network>","CreationDate":"2016-04-05T13:50:53.883","AcceptedAnswerId":38015983.0,"OwnerUserId":1013394.0,"Title":"LSTM Followed by Mean Pooling","Body":"<p>I'm using Keras 1.0. My problem is identical to this one (<a href=\"https:\/\/stackoverflow.com\/questions\/35355528\">How to implement a Mean Pooling layer in Keras<\/a>), but the answer there does not seem to be sufficient for me.<\/p>\n\n<p>I want to implement this network:\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>The following code does not work:<\/p>\n\n<pre><\/pre>\n\n<p>If I don't set , I get this error when I call : <\/p>\n\n<pre><\/pre>\n\n<p>Otherwise, I get this error when I call :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38015983","CreationDate":"2016-06-24T14:37:07.043","ParentId":null,"OwnerUserId":"1013394","Title":null,"Body":"<p>Adding <code>TimeDistributed(Dense(1))<\/code> helped:<\/p>\n\n<pre><code>sequence = Input(shape=(max_sent_len,), dtype='int32')\nembedded = Embedding(vocab_size, word_embedding_size)(sequence)\nlstm = LSTM(hidden_state_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True)(embedded)\ndistributed = TimeDistributed(Dense(1))(lstm)\npool = AveragePooling1D()(distributed)\noutput = Dense(1, activation='sigmoid')(pool)\n<\/code><\/pre>\n"},{"AnswerId":"40854921","CreationDate":"2016-11-28T22:50:06.423","ParentId":null,"OwnerUserId":"5351387","Title":null,"Body":"<p>I think the accepted answer is basically wrong. A solution was found at:\n<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2151\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/2151<\/a>\nHowever, it only works for theano backend. I have modified the code so that it supports both theano and tensorflow.<\/p>\n\n<pre><code>from keras.engine.topology import Layer, InputSpec\nfrom keras import backend as T\n\nclass TemporalMeanPooling(Layer):\n    \"\"\"\nThis is a custom Keras layer. This pooling layer accepts the temporal\nsequence output by a recurrent layer and performs temporal pooling,\nlooking at only the non-masked portion of the sequence. The pooling\nlayer converts the entire variable-length hidden vector sequence\ninto a single hidden vector, and then feeds its output to the Dense\nlayer.\n\ninput shape: (nb_samples, nb_timesteps, nb_features)\noutput shape: (nb_samples, nb_features)\n\"\"\"\ndef __init__(self, **kwargs):\n    super(TemporalMeanPooling, self).__init__(**kwargs)\n    self.supports_masking = True\n    self.input_spec = [InputSpec(ndim=3)]\n\ndef get_output_shape_for(self, input_shape):\n    return (input_shape[0], input_shape[2])\n\ndef call(self, x, mask=None): #mask: (nb_samples, nb_timesteps)\n    if mask is None:\n        mask = T.mean(T.ones_like(x), axis=-1)\n    ssum = T.sum(x,axis=-2) #(nb_samples, np_features)\n    mask = T.cast(mask,T.floatx())\n    rcnt = T.sum(mask,axis=-1,keepdims=True) #(nb_samples)\n    return ssum\/rcnt\n    #return rcnt\n\ndef compute_mask(self, input, mask):\n    return None\n<\/code><\/pre>\n"},{"AnswerId":"36524166","CreationDate":"2016-04-09T23:55:45.920","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Please try this (I hope this will solve your problem :) ):<\/p>\n\n<p><a href=\"http:\/\/keras.io\/layers\/core\/#timedistributedmerge\" rel=\"nofollow\">http:\/\/keras.io\/layers\/core\/#timedistributedmerge<\/a> <\/p>\n"},{"AnswerId":"43531172","CreationDate":"2017-04-20T22:51:51.687","ParentId":null,"OwnerUserId":"6454119","Title":null,"Body":"<p>I just attempted to implement the same model as the original poster, and I'm using <code>Keras 2.0.3<\/code>. The mean pooling after LSTM worked when I used <code>GlobalAveragePooling1D<\/code>, just make sure <code>return_sequences=True<\/code> in the LSTM layer. Give it a try!<\/p>\n"},{"AnswerId":"43014228","CreationDate":"2017-03-25T08:31:08.103","ParentId":null,"OwnerUserId":"1105231","Title":null,"Body":"<p>Thanks, I also meet the question, but I think TimeDistributed layer not working as you want, you can try Luke Guye's TemporalMeanPooling layer, it works for me. Here is the example:<\/p>\n\n<pre><code>sequence = Input(shape=(max_sent_len,), dtype='int32')\nembedded = Embedding(vocab_size, word_embedding_size)(sequence)\nlstm = LSTM(hidden_state_size, return_sequences=True)(embedded)\npool = TemporalMeanPooling()(lstm)\noutput = Dense(1, activation='sigmoid')(pool)\n<\/code><\/pre>\n"}]}
{"QuestionId":36428665,"AnswerCount":1,"Tags":"<neural-network><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-04-05T14:04:15.747","AcceptedAnswerId":36430715.0,"OwnerUserId":1349516.0,"Title":"Tensorflow translate training - when will it stop?","Body":"<p>I am using example from Google's Tensorflow. I am able to run it well, but the training doesn't seem to stop as it has a  loop.\nreference:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/rnn\/translate\/translate.py\" rel=\"nofollow\" title=\"translate.py\">translate.py<\/a><\/p>\n\n<p>Coe in :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36430715","CreationDate":"2016-04-05T15:27:56.910","ParentId":null,"OwnerUserId":"2662629","Title":null,"Body":"<p>This is true, at one point you can abort the training process. The moment of it depends on the model parameters you have choosen. According to the tensorflow <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/seq2seq\/index.html#sequence-to-sequence-models\" rel=\"nofollow\">documentation<\/a> it takes about 340K steps with choosen batch-size of 64.\nIf have selected the option to create checkpoints, you can use any of the checkpoints to check your models performance with the decoder. <\/p>\n"}]}
{"QuestionId":36429205,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><theano><deep-learning><caffe>","CreationDate":"2016-04-05T14:25:08.327","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"How to make sense of tensorflow tensorboard Histograms?","Body":"<p>I would like to know how to make sense of the tensor flow graphs\/Histograms generated.\nThe code for this can be found <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/images\/convnet_mnist.py\" rel=\"noreferrer\">here.<\/a>\nThis graph is easy to understand Accuracy and loss are straight forward to understand.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LVAWR.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LVAWR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Basically what the graph signifies and how i could use it to understand my network and if possible what changes i can make to improve it.<\/p>\n\n<h1>How Do I interpret the histograms?<\/h1>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/P5cMY.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/P5cMY.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"45996950","CreationDate":"2017-09-01T09:20:01.153","ParentId":null,"OwnerUserId":"4393970","Title":null,"Body":"<p><strong>tf.summary.histogram<\/strong> takes an arbitrarily sized and shaped Tensor, and compresses it into a histogram data structure consisting of many bins with widths and counts. For example, let's say we want to organize the numbers [0.5, 1.1, 1.3, 2.2, 2.9, 2.99] into bins. We could make three bins: a bin containing everything from 0 to 1 (it would contain one element, 0.5), a bin containing everything from 1-2 (it would contain two elements, 1.1 and 1.3), * a bin containing everything from 2-3 (it would contain three elements: 2.2, 2.9 and 2.99).<\/p>\n\n<p>Please follow below links for more details: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42315202\/understanding-tensorboard-weight-histograms\">sunside answer<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/get_started\/tensorboard_histograms\" rel=\"nofollow noreferrer\">Tensorflow documentation<\/a><\/p>\n"}]}
{"QuestionId":36430601,"AnswerCount":3,"Tags":"<python><tensorflow><recurrent-neural-network><gated-recurrent-unit>","CreationDate":"2016-04-05T15:23:48.557","AcceptedAnswerId":null,"OwnerUserId":4442323.0,"Title":"Mixing feed forward layers and recurrent layers in Tensorflow?","Body":"<p>Has anyone been able to mix feedforward layers and recurrent layers in Tensorflow?<\/p>\n\n<p>For example:\ninput->conv->GRU->linear->output<\/p>\n\n<p>I can imagine one can define his own cell with feedforward layers and no state which can then be stacked using the MultiRNNCell function, something like:<\/p>\n\n<p>cell = tf.nn.rnn_cell.MultiRNNCell([conv_cell,GRU_cell,linear_cell])<\/p>\n\n<p>This would make life a whole lot easier...<\/p>\n","answers":[{"AnswerId":"41360678","CreationDate":"2016-12-28T11:03:06.803","ParentId":null,"OwnerUserId":"2641587","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/alrojo\/tensorflow-tutorial\/blob\/master\/lab4_Kaggle\/lab4_Kaggle.ipynb\" rel=\"nofollow noreferrer\">This tutoria<\/a>l gives an example of how to use convolutional layers together with recurrent ones. For example, having last convolution layers like this:<\/p>\n\n<pre><code>...\nl_conv4_a = conv_pre(l_pool3, 16, (5, 5), scope=\"l_conv4_a\")\nl_pool4 = pool(l_conv3_a, scope=\"l_pool4\")\nl_flatten = flatten(l_pool4, scope=\"flatten\")\n<\/code><\/pre>\n\n<p>and having defined RNN cell:<\/p>\n\n<pre><code>_, shape_state = tf.nn.dynamic_rnn(cell=shape_cell,\n    inputs=tf.expand_dims(batch_norm(x_shape_pl), 2), dtype=tf.float32, scope=\"shape_rnn\")\n<\/code><\/pre>\n\n<p>You can concatenate both outputs and use it as the input to the next layer:<\/p>\n\n<pre><code>features = tf.concat(concat_dim=1, values=[x_margin_pl, shape_state, x_texture_pl, l_flatten], name=\"features\")\n<\/code><\/pre>\n\n<p>Or you can just use the output of CNN layer as the input to the RNN cell:<\/p>\n\n<pre><code>_, shape_state = tf.nn.dynamic_rnn(cell=shape_cell,\n    inputs=l_flatten, dtype=tf.float32, scope=\"shape_rnn\")\n<\/code><\/pre>\n"},{"AnswerId":"38767552","CreationDate":"2016-08-04T12:38:18.220","ParentId":null,"OwnerUserId":"6678014","Title":null,"Body":"<p>can't you just do the following:<\/p>\n\n<pre><code>rnnouts, _ = rnn(grucell, inputs)\nlinearout = [tf.matmul(rnnout, weights) + bias for rnnout in rnnouts]\n<\/code><\/pre>\n\n<p>etc.<\/p>\n"},{"AnswerId":"51439704","CreationDate":"2018-07-20T09:42:50.600","ParentId":null,"OwnerUserId":"4726173","Title":null,"Body":"<p>This is what I have so far; improvements welcome:<\/p>\n\n<pre><code>class LayerCell(rnn_cell_impl.RNNCell):\n\n    def __init__(self, tf_layer, **kwargs):\n        ''' :param tf_layer: a tensorflow layer, e.g. tf.layers.Conv2D or \n            tf.keras.layers.Conv2D. NOT tf.layers.conv2d !\n            Can pass all other layer params as well, just need to give the \n            parameter name: paramname=param'''\n        self.layer_fn = tf_layer(**kwargs)\n\n    def __call__(self, inputs, state, scope=None):\n        ''' Every `RNNCell` must implement `call` with\n          the signature `(output, next_state) = call(input, state)`.  The optional\n          third input argument, `scope`, is allowed for backwards compatibility\n          purposes; but should be left off for new subclasses.'''\n        return (self.layer_fn(inputs), state)\n\n    def __str__(self):\n            return \"Cell wrapper of \" + str(self.layer_fn)\n\n    def __getattr__(self, attr):\n        '''credits to https:\/\/stackoverflow.com\/questions\/1382871\/dynamically-attaching-a-method-to-an-existing-python-object-generated-with-swig\/1383646#1383646'''\n        return getattr(self.layer_fn, attr)\n\n    @property\n    def state_size(self):\n        \"\"\"size(s) of state(s) used by this cell.\n\n        It can be represented by an Integer, a TensorShape or a tuple of Integers\n        or TensorShapes.\n        \"\"\"\n        return  (0,) \n\n    @property\n    def output_size(self):\n        \"\"\"Integer or TensorShape: size of outputs produced by this cell.\"\"\"\n        # use with caution; could be uninitialized\n        return self.layer_fn.output_shape\n<\/code><\/pre>\n\n<hr>\n\n<p>(Naturally, don't use with recurrent layers because state-keeping will be destroyed.)<\/p>\n\n<p>Seems to work with: tf.layers.Conv2D, tf.keras.layers.Conv2D, tf.keras.layers.Activation, tf.layers.BatchNormalization<\/p>\n\n<p><em>Does NOT work with: tf.keras.layers.BatchNormalization.<\/em>\nAt least it failed for me when using it in a tf.while loop;  complaining about combining variables from different frames, similar to <a href=\"https:\/\/stackoverflow.com\/questions\/42564698\/invalidargumenterror-the-node-has-inputs-from-different-frames\">here<\/a>. Maybe keras uses <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8604\" rel=\"nofollow noreferrer\">tf.Variable() instead of tf.get_variable()<\/a> ...?<\/p>\n\n<hr>\n\n<p>Usage:<\/p>\n\n<pre><code>cell0 = tf.contrib.rnn.ConvLSTMCell(conv_ndims=2, input_shape=[40, 40, 3], output_channels=16, kernel_shape=[5, 5])\ncell1 = LayerCell(tf.keras.layers.Conv2D, filters=8, kernel_size=[5, 5], strides=(1, 1), padding='same')\ncell2 = LayerCell(tf.layers.BatchNormalization, axis=-1)\n\ninputs =  np.random.rand(10, 40, 40, 3).astype(np.float32)\nmulticell = tf.contrib.rnn.MultiRNNCell([cell0, cell1, cell2])\nstate = multicell.zero_state(batch_size=10, dtype=tf.float32)\n\noutput = multicell(inputs, state)\n<\/code><\/pre>\n"}]}
{"QuestionId":36434030,"AnswerCount":0,"Tags":"<python><caffe>","CreationDate":"2016-04-05T18:17:04.893","AcceptedAnswerId":null,"OwnerUserId":4780721.0,"Title":"caffe python example ERROR caffe.NetParameter","Body":"<p>I\u2019m trying this <a href=\"http:\/\/www.marekrei.com\/blog\/transforming-images-to-feature-vectors\/#comment-24909\" rel=\"nofollow\">example<\/a> of caffe and I'm getting this error. I checked all the paths and they are fine.<\/p>\n\n<pre><\/pre>\n\n<p>Do you have some idea what could be.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":36434878,"AnswerCount":1,"Tags":"<python><c++><python-2.7><theano><lasagne>","CreationDate":"2016-04-05T19:00:49.140","AcceptedAnswerId":36435226.0,"OwnerUserId":5988583.0,"Title":"Python 2.7 compilation failure (theano), windows 10","Body":"<p>I tried looking at what others have done with problem similar to mine, but I still get the same error. I have no idea what's causing it. I've tried following: \n<a href=\"https:\/\/stackoverflow.com\/questions\/33687103\/how-to-install-theano-on-anaconda-python-2-7-x64-on-windows\">How to install Theano on Anaconda Python 2.7 x64 on Windows?<\/a> <\/p>\n\n<p>but that didn't help.<\/p>\n\n<p>I'm trying to run this example, but with my own dataset: <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py<\/a><\/p>\n\n<p>Error message is as follows, I didn't include the 787 other lines inbetween: <\/p>\n\n<pre><\/pre>\n\n<p><strong>Solved<\/strong>: Followed every part of the following instructions: <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html#install-windows<\/a><\/p>\n\n<p>did this as well: Why can't nvcc find my Visual C++ installation?<\/p>\n\n<p>Theano is working now and is using my GPU.<\/p>\n","answers":[{"AnswerId":"36435226","CreationDate":"2016-04-05T19:20:56.650","ParentId":null,"OwnerUserId":"6037118","Title":null,"Body":"<p>Use Microsoft Visual Studio 2010 with Anaconda as you have to use the same compiler they used for packages you compile to be compatible.  They do have mingw bindings but I'm not sure how compatible they are.<\/p>\n"}]}
{"QuestionId":36435190,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras>","CreationDate":"2016-04-05T19:18:57.623","AcceptedAnswerId":36436256.0,"OwnerUserId":4701050.0,"Title":"Keras\/Machine Learning: Any pros and cons of flattening input data vs having a higher dimensional input?","Body":"<p>I am training a model that takes two 300d word vectors as an input. I was originally planning on supplying a 600d vector as the input data.<\/p>\n\n<p>However, I noticed on the Keras datasets page: <a href=\"http:\/\/keras.io\/datasets\/\" rel=\"nofollow\">http:\/\/keras.io\/datasets\/<\/a><\/p>\n\n<p>They use higher dimensional input data instead of flattening the input. For example, they represent a 32x32 RBG image with dimensions:  (3, 32, 32)<\/p>\n\n<p>Should I do a (2,300) input or a (1,600) input? Does it matter? <\/p>\n","answers":[{"AnswerId":"36436256","CreationDate":"2016-04-05T20:19:34.087","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>It matters if and only if your model is a spatial model. Typical examples involve convolutional neural networks. Otherwise - if no spatial based module is used - data will be internally flattened either way.<\/p>\n"}]}
{"QuestionId":36437989,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><theano><deep-learning>","CreationDate":"2016-04-05T22:11:07.667","AcceptedAnswerId":null,"OwnerUserId":1452257.0,"Title":"Implementing sparse connections in neural network (Theano)","Body":"<p>Some use cases for neural networks requires that not all neurons are connected between two consecutive layers. For my neural network architecture, I need to have a layer, where each neuron only has connections to some prespecified neurons in the previous layer (at somewhat arbitrary places, not with a pattern such as a convolution layer). This is needed in order to model data on a specific graph. I need to implement this \"Sparse\" layer in Theano, but I'm not used to the Theano way of programming. <\/p>\n\n<p>It seems that the most efficient way of programming sparse connections in Theano would be to use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/blocksparse.html\">theano.tensor.nnet.blocksparse.SparseBlockGemv<\/a>. An alternative would be to do matrix multiplication, where many weights are set to 0 (= no connection), but that would be very inefficient compared to  as each neuron is only connected to 2-6 neurons in the previous layer out of ~100000 neurons. Moreover, a weight matrix of 100000x100000 would not fit on my RAM\/GPU. Could someone therefore provide an example of how to implement sparse connections using the  method or another computationally-efficient method?<\/p>\n\n<p>A perfect example would be to extend the <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html#mlp\">MLP Theano Tutorial<\/a> with an extra layer after the hidden layer (and before softmax), where each neuron only has connections to a subset of neurons in the previous layer. However, other examples are also very welcome!<\/p>\n\n<p><strong>Edit:<\/strong> Note that the layer must be implemented in Theano as it is just a small part of a larger architecture.<\/p>\n","answers":[{"AnswerId":"36641729","CreationDate":"2016-04-15T08:13:21.063","ParentId":null,"OwnerUserId":"118630","Title":null,"Body":"<p>The output of a fully-connected layer is given by the dot product of the input and the weights of that layer. In theano or numpy you can use the <code>dot<\/code> method.<\/p>\n\n<pre><code>y = x.dot(w)\n<\/code><\/pre>\n\n<p>If you only have connections to some neurons in the previous layer and those connections are predefined you could do something like this:<\/p>\n\n<pre><code>y = [x[edges[i]].dot(w[i])) for i in neurons]\n<\/code><\/pre>\n\n<p>Where <code>edges[i]<\/code> contains the indices for neurons connected to neuron <code>i<\/code> and <code>w[i]<\/code> the weights of this connection.<\/p>\n\n<p>Please note, that theano doesn't know about layers or other high-level details.<\/p>\n"},{"AnswerId":"49284013","CreationDate":"2018-03-14T17:19:11.880","ParentId":null,"OwnerUserId":"669960","Title":null,"Body":"<p>Apologies for resurrecting an old thread, but this was the simplest guidance I found that was useful in extending the guidance at <a href=\"https:\/\/iamtrask.github.io\/2015\/07\/12\/basic-python-network\/\" rel=\"nofollow noreferrer\">https:\/\/iamtrask.github.io\/2015\/07\/12\/basic-python-network\/<\/a> for partially-connected inputs. However, it took me a while to make sense of basaundi's answer and I think I can improve upon it.<\/p>\n\n<p>There were a couple of things that I needed to change to make it work. In my case, I am trying to map from N inputs to M neurons in my first hidden layer. My inputs are in a NxF array, where F is the number of features for my inputs, and my synapse values (weights) between inputs and the first layer are in a FxM array. Therefore, the output of Inputs &lt;dot&gt; Weights is a NxM array. My edge matrix is an MxF array that specifies for each neuron in layer 1 (rows), which of the features of the input data are relevant (columns).<\/p>\n\n<p>In this setup, at least, it required me to slice my arrays differently than specified above. Also, the list comprehension returns a list of matrices, which  must be summed to get the correct NxM (otherwise you get an MxNxM array). <\/p>\n\n<p>So I have used the following (util.sigmoid is a helper function of my own):<\/p>\n\n<pre><code>y = [numpy.dot(x[:, edges[i]], w[edges[i]]) \n     for i in range(M)]\ny = util.sigmoid(numpy.sum(y, 0))\n<\/code><\/pre>\n\n<p>This seems to work for me.<\/p>\n"}]}
{"QuestionId":36438800,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-05T23:24:09.897","AcceptedAnswerId":null,"OwnerUserId":2666216.0,"Title":"Tensorflow: transferring variables across graphs","Body":"<p>I am trying to figure out how to import a learned saved model as a part of a larger model.<\/p>\n\n<p>To be concrete, I have an RNN and some embedding matrices (model A) that have been trained on a large dataset stored in a  file. I also have another model (model B), that uses this exact same RNN and embedding model as a submodel, with additional ops to manipulate the output of this RNN and use it for classification.<\/p>\n\n<p>The graph definitions for these two models are different, but I'd like to initialize model B to the saved version of model A.  How can this be done?<\/p>\n\n<p>My attempt so far is to try and load model A as another model under a different session and graph and then assign the relevant matrices in model B to those from A, but this isn't working.<\/p>\n\n<p>Here's the relevant portion of code:<\/p>\n\n<pre><\/pre>\n\n<p>Note that both models have explicit variables assigned to the matrices from  for the RNN using  when instantiating the class.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"36605802","CreationDate":"2016-04-13T17:42:46.460","ParentId":null,"OwnerUserId":"2666216","Title":null,"Body":"<p>The answer turned out to be so easy.  This was just creating the assign ops... To actually do the assignment, run the ops in the appropriate session.<\/p>\n"}]}
{"QuestionId":36438991,"AnswerCount":1,"Tags":"<python><neural-network><theano><deep-learning>","CreationDate":"2016-04-05T23:43:32.570","AcceptedAnswerId":null,"OwnerUserId":6108129.0,"Title":"Addition of unequal sized Tensors in theano","Body":"<p>I am new to theano and was looking at this tutorial : <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html<\/a> . I have pasted the line in question. The pooled_out(output after convolution and maxout) is a 4D Tensor of size (a,b,c,d) whereas <strong><em>b<\/em><\/strong>(bias) is a vector. Even after dimshuffle dimension for <strong><em>b<\/em><\/strong> is (1,b,1,1). How can we add two unequal sized Tensors. Or does theano automatically copies the same value for <strong><em>b<\/em><\/strong> at all other places  for same value of second dimension of tensor <strong><em>b<\/em><\/strong> ? The comment on top of it says \"Each bias will thus be broadcasted across mini-batches and feature map width &amp; height\" is that what it means ? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36506653","CreationDate":"2016-04-08T18:21:02.637","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>\"Or does theano automatically copies the same value for b at all other places for same value of second dimension of tensor b ?\"<\/p>\n\n<p>This is exactly what will happen. <code>'x'<\/code> in <code>dimshuffle<\/code> means to match the required by the operation dimension of your tensor, 'copying' the value if needed.<\/p>\n\n<p>That process is called <strong>broadcasting<\/strong>.<\/p>\n"}]}
{"QuestionId":36439483,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-06T00:37:08.073","AcceptedAnswerId":null,"OwnerUserId":456105.0,"Title":"how to get the time consumed to execute each node in tensorflow graph","Body":"<p>Currently I can run through my tensorflow graph correctly, but the running time is longer than my expectation, so I'd like to know how to profile execution time for each node in the graph.<\/p>\n","answers":[{"AnswerId":"36439551","CreationDate":"2016-04-06T00:48:19.563","ParentId":null,"OwnerUserId":"2216822","Title":null,"Body":"<p>You could probably use fields recorded in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/step_stats.proto\" rel=\"nofollow\">step_stats<\/a>. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/3c8780451007c27b69f10925d43d1f3501d94106\/tensorflow\/python\/client\/timeline_test.py\" rel=\"nofollow\">TimelineTest<\/a> shows an example of how to get stats about execution.<\/p>\n"}]}
{"QuestionId":36440468,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2016-04-06T02:33:57.137","AcceptedAnswerId":36442940.0,"OwnerUserId":5545794.0,"Title":"How can I define a constant vector layer in Caffe?","Body":"<p>I have a data with 10-d label vector, and I want to use a caffe model to make regression against these data with 10-d output. But now, I only want to check loss of some outputs (for example, 1, 3, 4, 5, 6-d of 10-d vector), so I define a layer with 5-d output at the bottom of the last output layer, But I'v no idea how to get corresponding 5-d label vector groundtruth, I think  may be I can define a constant layer to indicate which entries I want get. Please help me if you have any ideas.<\/p>\n\n<p><strong>update: example<\/strong><\/p>\n\n<p>This is my original InnerProduct and Loss layer<\/p>\n\n<pre><\/pre>\n\n<p>I  care more about $n_1$ (like 1,3,4,5,6) entries of the 10-dimension output and their loss, so I want to fetch the loss of these entries, like<\/p>\n\n<pre><\/pre>\n\n<p>How can I get score1 from score directly?<\/p>\n","answers":[{"AnswerId":"36442940","CreationDate":"2016-04-06T06:14:00.170","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<p>From what I interpret from your question, I think you would like to calculate the loss of the output layer w.r.t the labels for a regression model. But you would like to not bring some of the labels to the equation. <\/p>\n\n<p>If my interpretation is true, being a regression model, I am expecting your new layer to be something similar to that of EuclidieanLayer. If it is so, the <code>caffe_sub<\/code> function in the layer could be replaced by the following code segment.<\/p>\n\n<pre><code>  int arrayPos[5] = {1,3,4,5,6};\n  int count = 5;\n  Dtype *newBottom0=(Dtype*)malloc(sizeof(Dtype)*count);\n  Dtype *newBottom1=(Dtype*)malloc(sizeof(Dtype)*count);\n  for(int varI=0; varI&lt;count; varI++)\n  {\n    newBottom0[varI] = (Dtype) bottom[0]-&gt;cpu_data()[arrayPos[varI]];\n    newBottom1[varI] = (Dtype) bottom[1]-&gt;cpu_data()[arrayPos[varI]];\n  }\n  caffe_sub( count, newBottom0, newBottom1, diff_.mutable_cpu_data());\n  free(newBottom0);\n  free(newBottom1);\n<\/code><\/pre>\n"}]}
{"QuestionId":36440826,"AnswerCount":1,"Tags":"<neural-network><deep-learning><backpropagation><torch><gradient-descent>","CreationDate":"2016-04-06T03:12:14.697","AcceptedAnswerId":36445175.0,"OwnerUserId":5576325.0,"Title":"How to write the updateGradInput and accGradParameters in torch?","Body":"<p>I know the two functions are for torch's backward propagation and the interface is as follows\n\nI'm confused about what the  and  really mean in a layer.\nAssume the network's cost is  and a layer . Do  and  of layer  mean  and ?<\/p>\n\n<p>If so, how to compute  accorading to ? <\/p>\n\n<p>Moreover, does  mean to accumulate  and ? If so, how to compute these values?<\/p>\n","answers":[{"AnswerId":"36445175","CreationDate":"2016-04-06T08:11:38.737","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<blockquote>\n  <p>Do <code>gradInput<\/code> and <code>gradOutput<\/code> of layer L mean <code>d_C\/d_input_L<\/code> and <code>d_C\/d_output_L<\/code><\/p>\n<\/blockquote>\n\n<p>Yes:<\/p>\n\n<ul>\n<li><code>gradInput<\/code> = derivative of the cost w.r.t layer's input,<\/li>\n<li><code>gradOutput<\/code> = derivative of the cost w.r.t layer's output.<\/li>\n<\/ul>\n\n<blockquote>\n  <p>how to compute <code>gradInput<\/code> according to <code>gradOutput<\/code><\/p>\n<\/blockquote>\n\n<p>Adapting the schema from <a href=\"http:\/\/deepdish.io\/2015\/11\/21\/building-blocks-of-deep-learning\/\" rel=\"nofollow noreferrer\">The building blocks of Deep Learning<\/a> (<strong>warning<\/strong>: in this schema the cost is denoted <code>L<\/code> = <code>Loss<\/code>, and the layer <code>f<\/code>) we have:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MfD46.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MfD46.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>For a concrete, step-by-step example of such a computation on a LogSoftMax layer you can refer to <a href=\"https:\/\/stackoverflow.com\/a\/35328031\/1688185\">this answer<\/a>.<\/p>\n\n<blockquote>\n  <p>does <code>accGradParameters<\/code> mean to accumulate <code>d_C\/d_Weight_L<\/code> and <code>d_C\/d_bias_L<\/code><\/p>\n<\/blockquote>\n\n<p>Yes. Named <code>gradWeight<\/code> and <code>gradBias<\/code> in torch\/nn.<\/p>\n\n<blockquote>\n  <p>how to compute these values?<\/p>\n<\/blockquote>\n\n<p>Similarly as above. Still using a formula from the above blog post:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/r1v2p.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/r1v2p.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Except the jacobian has not the same dimensionality (see the blog post for more details). As an example, for a Linear layer this translates into:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SmEiJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SmEiJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is the outer product between the layer's input and <code>gradOutput<\/code>. In Torch <a href=\"https:\/\/github.com\/torch\/nn\/blob\/872f4b9407368e1190d38d01775f13994c517f28\/Linear.lua#L90\" rel=\"nofollow noreferrer\">we have<\/a>:<\/p>\n\n<pre><code>self.gradWeight:addr(scale, gradOutput, input)\n<\/code><\/pre>\n\n<p>And:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pp5CT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pp5CT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Which is <code>gradOutput<\/code>. In Torch <a href=\"https:\/\/github.com\/torch\/nn\/blob\/872f4b9407368e1190d38d01775f13994c517f28\/Linear.lua#L91\" rel=\"nofollow noreferrer\">we have<\/a>:<\/p>\n\n<pre><code>self.gradBias:add(scale, gradOutput)\n<\/code><\/pre>\n\n<p>In both cases <code>scale<\/code> is a scale factor used in practice as the learning rate.<\/p>\n"}]}
{"QuestionId":36441224,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras><autoencoder>","CreationDate":"2016-04-06T03:58:17.997","AcceptedAnswerId":36476794.0,"OwnerUserId":6164459.0,"Title":"Keras: Wrong Number of Training Epochs","Body":"<p>I'm trying to build a class to quickly initialize and train an autoencoder for rapid prototyping.  One thing I'd like to be able to do is quickly adjust the number of epochs I train for.  However, it seems like no matter what I do, the model trains each layer for 100 epochs!  I'm using the tensorflow backend.  <\/p>\n\n<p>Here is the code from the two offending methods.<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions on how to fix the problem would be appreciated.  My original suspicion was that it was something to do with tensorflow, so I tried running with the theano backend but encountered the same problem. <\/p>\n\n<p><a href=\"https:\/\/github.com\/efrubin\/SpringJP-Autoencoders\/blob\/master\/tools\/JPtools.py\" rel=\"nofollow\">Here<\/a> is a link to the full program. <\/p>\n","answers":[{"AnswerId":"36476794","CreationDate":"2016-04-07T12:53:04.003","ParentId":null,"OwnerUserId":"5179470","Title":null,"Body":"<p>Following the <a href=\"http:\/\/keras.io\/models\/#sequential\" rel=\"nofollow\">Keras doc<\/a>, the <code>fit<\/code> method uses a default of 100 training epochs (<code>nb_epoch=100<\/code>):<\/p>\n\n<pre><code>fit(X, y, batch_size=128, nb_epoch=100, verbose=1, callbacks=[], validation_split=0.0, validation_data=None, shuffle=True, show_accuracy=False, class_weight=None, sample_weight=None)\n<\/code><\/pre>\n\n<p>I'm sure how you are running these methods, but following the \"Typical usage\" from the <a href=\"https:\/\/github.com\/efrubin\/SpringJP-Autoencoders\/blob\/master\/tools\/JPtools.py\" rel=\"nofollow\">original code<\/a>, you should be able to run something like (adjusting the variable <code>num_epoch<\/code> as required):<\/p>\n\n<pre><code>#Typical usage:\nnum_epoch = 10\nae = JPAutoEncoder(dims)\nae.pretrain(X_train, nb_epoch = num_epoch)\nae.train(X_train, nb_epoch = num_epoch)\nae.predict(X_val)\n<\/code><\/pre>\n"}]}
{"QuestionId":36443521,"AnswerCount":3,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-04-06T06:48:39.400","AcceptedAnswerId":37315453.0,"OwnerUserId":4229825.0,"Title":"Effective reading of own images in tensorflow","Body":"<p>I've skimmed over all tensorflow tutorials in which all data sets were loaded in RAM due to their small size. However, my own data (~30 Gb of images) can not be loaded in memory, therefore I'm looking for effective ways of reading images for further processing. Could anyone provide me examples of how can I do that?<\/p>\n\n<p>P.S. I have two files  and  that contain:<\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"37315453","CreationDate":"2016-05-19T06:30:56.380","ParentId":null,"OwnerUserId":"3824287","Title":null,"Body":"<p>This is what you're looking for: <a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\">Tensorflow read images with labels<\/a><\/p>\n\n<p>The exact code snippet is like this:<\/p>\n\n<pre><code>def read_labeled_image_list(image_list_file):\n    \"\"\"Reads a .txt file containing pathes and labeles\n    Args:\n       image_list_file: a .txt file with one \/path\/to\/image per line\n       label: optionally, if set label will be pasted after each line\n    Returns:\n       List with all filenames in file image_list_file\n    \"\"\"\n    f = open(image_list_file, 'r')\n    filenames = []\n    labels = []\n    for line in f:\n        filename, label = line[:-1].split(' ')\n        filenames.append(filename)\n        labels.append(int(label))\n    return filenames, labels\n\ndef read_images_from_disk(input_queue):\n    \"\"\"Consumes a single filename and label as a ' '-delimited string.\n    Args:\n      filename_and_label_tensor: A scalar string tensor.\n    Returns:\n      Two tensors: the decoded image, and the string label.\n    \"\"\"\n    label = input_queue[1]\n    file_contents = tf.read_file(input_queue[0])\n    example = tf.image.decode_png(file_contents, channels=3)\n    return example, label\n\n# Reads pfathes of images together with their labels\nimage_list, label_list = read_labeled_image_list(filename)\n\nimages = ops.convert_to_tensor(image_list, dtype=dtypes.string)\nlabels = ops.convert_to_tensor(label_list, dtype=dtypes.int32)\n\n# Makes an input queue\ninput_queue = tf.train.slice_input_producer([images, labels],\n                                            num_epochs=num_epochs,\n                                            shuffle=True)\n\nimage, label = read_images_from_disk(input_queue, num_labels=num_labels)\n\n# Optional Preprocessing or Data Augmentation\n# tf.image implements most of the standard image augmentation\nimage = preprocess_image(image)\nlabel = preprocess_label(label)\n\n# Optional Image and Label Batching\nimage_batch, label_batch = tf.train.batch([image, label],\n                                          batch_size=batch_size)\n<\/code><\/pre>\n"},{"AnswerId":"36451264","CreationDate":"2016-04-06T12:31:02.457","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>The recommended way is to put it into sharded protobuf files, where encoded jpeg and label(s) are features of a <code>tf.Example<\/code>. <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/405348050cff4d2d329669850da2a96a84946b08\/inception\/inception\/data\/build_image_data.py\" rel=\"nofollow\">build_image_data.py<\/a> in the <a href=\"https:\/\/github.com\/tensorflow\/models\/\" rel=\"nofollow\">tensorflow\/models<\/a> repository shows how to create such a database of image\/label pairs from a directory structure, you'll need to adapt it a bit to your case (it's straightforward). Then for training time you can look at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/405348050cff4d2d329669850da2a96a84946b08\/inception\/inception\/image_processing.py#L333\" rel=\"nofollow\">image_processing.py<\/a> where it shows how to go from the <code>tf.Example<\/code> proto to image\/label tensors (extract decoded jpg and label from the Example record, decode jpg, resize, apply augmentations as needed, then enqueue).<\/p>\n"},{"AnswerId":"36450272","CreationDate":"2016-04-06T11:48:37.157","ParentId":null,"OwnerUserId":"1068885","Title":null,"Body":"<p>Tutorial on udacity has stochastic method explained in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/4_convolutions.ipynb\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/4_convolutions.ipynb<\/a>, you can use the same with one change, instead of saving all images in single pickle file, save them in chunks of batch_size that you are using. That way at a time, you can load only as much data as used in the one batch.<\/p>\n"}]}
{"QuestionId":36445776,"AnswerCount":1,"Tags":"<machine-learning><nvidia><deep-learning><caffe><conv-neural-network>","CreationDate":"2016-04-06T08:38:23.130","AcceptedAnswerId":36448161.0,"OwnerUserId":3060545.0,"Title":"Is it possible to use Caffe Only for classification without any training?","Body":"<p>Some users might see this as opinion-based-question but if you look closely, I am trying to explore use of Caffe as a purely testing platform as opposed to currently popular use as training platform.<\/p>\n\n<p><strong>Background:<\/strong><\/p>\n\n<ol>\n<li>I have installed all dependencies using Jetpack 2.0 on Nvidia TK1.<\/li>\n<li>I have installed caffe and its dependencies successfully.<\/li>\n<li>The MNIST example is working fine.<\/li>\n<\/ol>\n\n<p><strong>Task:<\/strong><\/p>\n\n<ol>\n<li>I have been given a convnet with all standard layers. (Not an opensource model)<\/li>\n<li>The network weights and bias values etc are available <em>after training<\/em>. The training has not been done via caffe. (Pretrained Network)<\/li>\n<li>The weights and bias are all in the form of MATLAB matrices. (Actually in a .txt file but I can easily write code to get them to be matrices)<\/li>\n<li>I CANNOT do training of this network with caffe and must used the given weights and bias values ONLY for classification.<\/li>\n<li>I have my own dataset in the form of 32x32 pixel images.<\/li>\n<\/ol>\n\n<p><strong>Issue:<\/strong>\nIn all tutorials, details are given on how to deploy and train a network, and then use the generated .proto and .caffemodel files to validate and classify. Is it possible to implement this network on caffe and directly use my weights\/bias and training set to classify images? What are the available options here? I am a caffe-virgin so be kind. Thank you for the help!<\/p>\n","answers":[{"AnswerId":"36448161","CreationDate":"2016-04-06T10:19:25.177","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The only issue here is:<br>\n<em>How to initialize caffe net from text file weights<\/em>?<\/p>\n\n<p>I assume you have a <code>'deploy.prototxt'<\/code> describing the net's architecture (layer types, connectivity, filter sizes etc.). The only issue remaining is how to set the internal weights of <code>caffe.Net<\/code> to pre-defined values saved as text files.<\/p>\n\n<p>You can get access to <code>caffe.Net<\/code> internals, see <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\" title=\"net surgery\">net surgery<\/a> tutorial on how this can be done in python.<\/p>\n\n<p>Once you are able to set the weights according to your text file, you can <code>net.save(...)<\/code> the new weights into a binary <code>caffemodel<\/code> file to be used from now on. You do not have to train the net if you already have trained weights, and you can use it for generating predictions (\"test\").<\/p>\n"}]}
{"QuestionId":36446899,"AnswerCount":1,"Tags":"<python-2.7><lua><deep-learning><torch>","CreationDate":"2016-04-06T09:25:52.847","AcceptedAnswerId":null,"OwnerUserId":648896.0,"Title":"How can I load and use torch deep learning models from python?","Body":"<p>I use python for my real application and like to harness a torch model in python. Is there any available interface or a tenuous way to do so ?<\/p>\n","answers":[{"AnswerId":"36895941","CreationDate":"2016-04-27T16:44:54.277","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>There are several libraries enables you to interact with torch in python, but I would recommend <strong>lutorpy<\/strong>, here are some details:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow noreferrer\"><strong>lutorpy<\/strong><\/a>: the most promising one for deep learning with torch, it's based on lupa, easier to use(more pythonic). It keeps a lua engine in python, all lua variables and functions are automatically converted into python objects. So, you can import any torch module by using \"require\" function. \nYou can easily convert almost any lua\/torch code into python with very minimal change. It also provide conversion functions for numpy array and torch tensor,n the conversion is instant because they share the underlying memory. <\/p><\/li>\n<li><p>lunatic-python:\na minimal library for using lua in python, but the features are limited<\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/scoder\/lupa\" rel=\"nofollow noreferrer\">lupa<\/a>:\na library for using lua in python with a lot of useful features, but it's for lua in general, not torch, for example you can't easily convert torch tensor into numpy array.<\/p><\/li>\n<li><p>pytorch:\na wrapper for torch, the code are taken from torch and compiled into python module, so you can use torch pythonicly, but the module you can use is very limited (only nn are supported so far).<\/p><\/li>\n<\/ul>\n\n<p>Here is one example: <a href=\"https:\/\/stackoverflow.com\/a\/36896347\/6262499\">https:\/\/stackoverflow.com\/a\/36896347\/6262499<\/a><\/p>\n"}]}
{"QuestionId":36447505,"AnswerCount":1,"Tags":"<python><numpy><deep-learning><caffe><lmdb>","CreationDate":"2016-04-06T09:50:06.240","AcceptedAnswerId":null,"OwnerUserId":6092721.0,"Title":"Creating large LMDBs for Caffe with numpy arrays","Body":"<p>I have two 60 x 80921 matrices, one filled with data, one with reference.<br>\nI would like to store the values as key\/value pairs in two different LMDBs, one for training (say I'll slice around the 60000 column mark) and one for testing. Here is my idea; does it work?<\/p>\n\n<pre><\/pre>\n\n<p>I'm really not sure of the code. And what does the last line  refer to?<\/p>\n","answers":[{"AnswerId":"36450753","CreationDate":"2016-04-06T12:11:05.417","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It's not 100% clear what you are trying to do: are you treating each entry as a separate data sample, or are you trying to train on 60K 1D vectors of dim=60...<\/p>\n\n<p>Assuming you have 60K training samples of dim 60, you can write the training lmdbs like this:<\/p>\n\n<pre><code>env_x = lmdb.open('sensormatrix_train_x_lmdb', map_size=map_size) # you can put map_size a little bigger \nenv_y = lmdb.open('sensormatrix_train_y_lmdb', map_size=map_size)\nwith env_x.begin(write=True) as txn_x, env_y.begin(write=True) as txn_y:\n    for i in xrange(X_train.shape[1]):\n        x = X_train[:,i]\n        y = Y_train[:,i] \n\n        datum_x = caffe.io.array_to_datum(arr=x.reshape((60,1,1)),label=i)\n        datum_y = caffe.io.array_to_datum(arr=y.reshape((60,1,1)),label=i)\n        keystr = '{:0&gt;10d}'.format(i) # format an lmdb key for this entry\n        txn_x.put( keystr, datum_x.SerializeToString() ) # actual write to lmdb\n        txn_y.put( keystr, datum_y.SerializeToString() )\n<\/code><\/pre>\n\n<p>Now you have two lmdb for training, in your <code>'prototxt'<\/code> you should have two corresponding <code>\"Data\"<\/code> layers:<\/p>\n\n<pre><code>layer {\n  name: \"input_x\"\n  top: \"x\"\n  top: \"idx_x\"\n  type: \"Data\"\n  data_param { source: \"sensormatrix_train_x_lmdb\" batch_size: 32 }\n  include { phase: TRAIN }\n}\nlayer {\n  name: \"input_y\"\n  top: \"y\"\n  top: \"idx_y\"\n  type: \"Data\"\n  data_param { source: \"sensormatrix_train_y_lmdb\" batch_size: 32 }\n  include { phase: TRAIN }\n}\n<\/code><\/pre>\n\n<p>To make sure you read corresponding <code>x<\/code> <code>y<\/code> pairs, you can add a sanity check<\/p>\n\n<pre><code>layer {\n  name: \"sanity\"\n  type: \"EuclideanLoss\"\n  bottom: \"idx_x\"\n  bottom: \"idx_y\"\n  top: \"sanity\"\n  loss_weight: 0 \n  propagate_down: false\n  propagate_down: false\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":36448023,"AnswerCount":1,"Tags":"<deep-learning><torch>","CreationDate":"2016-04-06T10:12:16.223","AcceptedAnswerId":36451411.0,"OwnerUserId":648896.0,"Title":"What is the best way to read large datasets with less latency for torch model?","Body":"<p>I am familiar to use some data specific layers or tools by the habits of my in use libraries like mxnet (recordIO), caffe(LMDB). They also handle data augmentation tricks as well. I decided to give a shot for Torch to see what is good or bad. From the initial gaze, there is no such data loading standard for Torch or I am missing it. Could you point some good ways to load large scale datasets for my model training? <\/p>\n","answers":[{"AnswerId":"36451411","CreationDate":"2016-04-06T12:38:06.477","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>Some pointers:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/twitter\/torch-dataset\" rel=\"nofollow\">twitter\/torch-dataset<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/facebook\/fb.resnet.torch\/blob\/8f433818e9466a049d8844ae4462569ae38e5a61\/dataloader.lua\" rel=\"nofollow\">fb.resnet.torch\/dataloader<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/Element-Research\/dataload\" rel=\"nofollow\">Element-Research\/dataload<\/a><\/li>\n<\/ul>\n\n<p>More resources can also be found in the <a href=\"https:\/\/github.com\/torch\/torch7\/wiki\/Cheatsheet#loading-popular-datasets-\" rel=\"nofollow\">torch7\/wiki<\/a>.<\/p>\n"}]}
{"QuestionId":36448200,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><theano><lasagne>","CreationDate":"2016-04-06T10:20:46.827","AcceptedAnswerId":null,"OwnerUserId":4846917.0,"Title":"Theano error: ValueError: y_i value out of bounds.","Body":"<p>I'm trying to get my own data to work with a convolutional neural network in theano\/lasagne. <\/p>\n\n<p>A state consists of 4 80x80 images. There are 32 states in a batch and the batch is the input to the neural network. The output of the network has 5 units. (Relating to the 5 possible actions in the game) <strong>also see edit 2<\/strong><\/p>\n\n<p>Before training the model, the program observes states. When testing the training function I set the observations to only 500 and everything was fine. But when I set the observations to 50 000, I suddenly get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Debugprint of the apply node:<\/p>\n\n<pre><\/pre>\n\n<p>The storage map footprint:<\/p>\n\n<pre><\/pre>\n\n<p>My training function looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand why it happens at 50 000 observations but not at 500.\nThe only thing that changes is the amount of observations, why is it suddenly 'out of bounds'. Any ideas on why this is happening? Every answer is hugely appreciated. Thank you.<\/p>\n\n<p>The full code is here:<\/p>\n\n<ul>\n<li>main file: <a href=\"http:\/\/pastebin.com\/y53VAauT\" rel=\"nofollow\">http:\/\/pastebin.com\/y53VAauT<\/a><\/li>\n<li>Snake game file: <a href=\"http:\/\/pastebin.com\/Xx4aPRu7\" rel=\"nofollow\">http:\/\/pastebin.com\/Xx4aPRu7<\/a><\/li>\n<li>pygame player file: <a href=\"http:\/\/pastebin.com\/EKAYt5N8\" rel=\"nofollow\">http:\/\/pastebin.com\/EKAYt5N8<\/a> (to get the game score and send actions)<\/li>\n<\/ul>\n\n<p>Edit: It also happens at 1000 and up. Still have no clue why.<\/p>\n\n<p><strong>Edit 2:<\/strong> I've found that the problem relates to the reward. Because the model is not trained yet, it mostly makes random moves. So that's why the problem occurs at the higher number of observations. (Higher chance of the reward being higher than 5) And it gives an error when the reward (the length of the snake) is higher then 5 which is odd because 5 is the number of possible actions (output of the neural network). Getting closer to the solution!                     <\/p>\n","answers":[]}
{"QuestionId":36449222,"AnswerCount":1,"Tags":"<android><python-2.7><protocol-buffers><tensorflow>","CreationDate":"2016-04-06T11:03:21.750","AcceptedAnswerId":null,"OwnerUserId":1340801.0,"Title":"Tensor Flow Could not load custom protobuf files in","Body":"<p>I have just created a protobuf file ( file) for my own custom images using a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/image_retraining\/index.html\" rel=\"nofollow\">TensorFlow tutorial<\/a>.<\/p>\n\n<p>But when I replaced the same file into the assets folder in  and try to build and generate an APK, the APK gets generated, but when I run the APK in an Android device, the APK crashes.<\/p>\n\n<p>If I run the  from Python, it gives me proper results.<\/p>\n\n<p>Appreciate any help.<\/p>\n","answers":[{"AnswerId":"39032216","CreationDate":"2016-08-19T06:13:14.430","ParentId":null,"OwnerUserId":"4393031","Title":null,"Body":"<p>Since DecodeJpeg isn't supported as part of the core, you'll need to strip it out of the graph first.<\/p>\n\n<pre><code>bazel build tensorflow\/python\/tools:strip_unused &amp;&amp; \\\nbazel-bin\/tensorflow\/python\/tools\/strip_unused \\\n--input_graph=your_retrained_graph.pb \\\n--output_graph=stripped_graph.pb \\\n--input_node_names=Mul \\\n--output_node_names=final_result \\\n--input_binary=true\n<\/code><\/pre>\n\n<p>Change few parameters in this file <\/p>\n\n<p>\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/TensorFlowImageListener.java<\/p>\n\n<p>The input sizes need to be 299, not 224. You'll also need to change the mean and std values both to 128. <\/p>\n\n<p>INPUT_NAME to \"Mul:0\" ,\nOUTPUT_NAME to \"final_result:0\"<\/p>\n\n<p>after which you will be able to compile the apk. <\/p>\n\n<p>Good Luck  <\/p>\n"}]}
{"QuestionId":36456436,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-04-06T16:03:48.233","AcceptedAnswerId":36480182.0,"OwnerUserId":393885.0,"Title":"How Can I Define Only the Gradient for a Tensorflow Subgraph?","Body":"<p>First: I am only a few days in with Tensorflow, so please bear with me.<\/p>\n\n<p>I started out from the cifar10 tutorial code and I am now using a combination of convolutions and eigenvalue decompositions that break the symbolic differentiation. I.e. the graph gets built, then upon calling  the script halts with \"No gradient defined for operation [...] (op type: SelfAdjointEig)\". No surprise there.<\/p>\n\n<p>The inputs to the subgraph in question are still only the input feature maps and the filters being used, and I have the formulas for the gradients at hand and they should be straight-forward to implement given the inputs to the subgraph and the gradient with respect to its output.<\/p>\n\n<p>From what I can see in the docs, I can register a gradient method for custom Ops with <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#RegisterGradient\"><\/a> or override them with the experimental .\nBoth of those should give me access to exactly the things I need. For example, <a href=\"https:\/\/github.com\/search?p=3&amp;q=RegisterGradient&amp;ref=searchresults&amp;type=Code&amp;utf8=%E2%9C%93\">searching on Github<\/a> I find a lot of examples that access the op's inputs as  or such.<\/p>\n\n<p>The problem I have is that I want to essentially \"shortcut\" a whole subgraph, not a single op, so I have no single op to decorate.\nSince this is happening in one of the convolutional layers of the cifar example I tried using the scope object for that layer. \nConceptually, what enters and exits that scope's graph is exactly what I want so if I could somehow override the whole scope's gradients that would \"already\" do it.<\/p>\n\n<p>I saw <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Graph.create_op\"><\/a> which (I think) I could use to register a new type of operation and I could then override that Operation type's gradient computation with aforementioned methods. But I don't see a way of defining that op's <em>forward<\/em> pass without writing it in C++...<\/p>\n\n<p>Maybe I am approaching this the wrong way entirely?\nSince all of my forward or backward operations can be implemented with the python interface I obviously want to avoid implementing anything in C++.<\/p>\n","answers":[{"AnswerId":"36480182","CreationDate":"2016-04-07T15:06:16.630","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's a trick from Sergey Ioffe:<\/p>\n\n<p>Suppose you want group of ops that behave as f(x) in forward mode, but as g(x) in the backward mode. You implement it as<\/p>\n\n<pre><code>t = g(x)\ny = t + tf.stop_gradient(f(x) - t)\n<\/code><\/pre>\n\n<p>So in your case your g(x) could be an identity op, with a custom gradient using <code>gradient_override_map<\/code><\/p>\n"},{"AnswerId":"58170593","CreationDate":"2019-09-30T15:05:38.017","ParentId":null,"OwnerUserId":"1133157","Title":null,"Body":"<p>Here is the approach which works for TensorFlow 2.0. Note that in 2.0 we are happy to have 2 different autodiff algorithms: <code>GradientTape<\/code> for eager mode and <code>tf.gradient<\/code> for the non-eager mode (here called \"lazy\"). We demonstrate that <code>tf.custom_gradient<\/code> works both ways.<\/p>\n\n<pre><code>import tensorflow as tf\nassert tf.version.VERSION.startswith('2.')\nimport numpy as np\nfrom tensorflow.python.framework.ops import disable_eager_execution, enable_eager_execution\nfrom tensorflow.python.client.session import Session\n\n@tf.custom_gradient\ndef mysquare(x):\n  res = x * x\n  def _grad(dy):\n    return dy * (2*x)\n  return res, _grad\n\ndef run_eager():\n  enable_eager_execution()\n\n  x = tf.constant(np.array([[1,2,3],[4,5,6]]).astype('float32'))\n  with tf.GradientTape() as tape:\n    tape.watch(x)\n    y = tf.reduce_sum(mysquare(x))\n\n    dy_dx = tape.gradient(y,x)\n    print('Eager mode')\n    print('x:\\n',x.numpy())\n    print('y:\\n',y.numpy())\n    print('dy_dx:\\n',dy_dx.numpy())\n\n\ndef run_lazy():\n  disable_eager_execution()\n\n  x = tf.constant(np.array([[1,2,3],[4,5,6]]).astype('float32'))\n  y = tf.reduce_sum(mysquare(x))\n  dy_dx = tf.gradients(y,x)\n\n  with Session() as s:\n    print('Lazy mode')\n    print('x:\\n',x.eval(session=s))\n    print('y:\\n',y.eval(session=s))\n    assert len(dy_dx)==1\n    print('dy_dx:\\n',dy_dx[0].eval(session=s))\n\nif __name__ == '__main__':\n  run_eager()\n  run_lazy()\n<\/code><\/pre>\n"},{"AnswerId":"54015702","CreationDate":"2019-01-03T02:48:45.740","ParentId":null,"OwnerUserId":"3623290","Title":null,"Body":"<p>From TensorFlow 1.7 onward, <code>tf.custom_gradient<\/code> is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/custom_gradient\" rel=\"nofollow noreferrer\">the way to go<\/a>.<\/p>\n"},{"AnswerId":"43724419","CreationDate":"2017-05-01T18:34:25.010","ParentId":null,"OwnerUserId":"5185653","Title":null,"Body":"<p>How about multiply and divide, instead of adding and subtracting t?<\/p>\n\n<pre><code>t = g(x)\ny = tf.stop_gradient(f(x) \/ t) * t\n<\/code><\/pre>\n"}]}
{"QuestionId":36456833,"AnswerCount":1,"Tags":"<cuda><gpu><tensorflow><theano><theano-cuda>","CreationDate":"2016-04-06T16:21:34.180","AcceptedAnswerId":null,"OwnerUserId":1986270.0,"Title":"GPU Compatibility for Theano [NVidia GeForce 8800GT]","Body":"<p>I am currently working on an ML project on my personal computer that has an AMD graphics card. I have an old NVidia 8800GT card that I could plug in for CUDA accelerated convolution, but I haven't found if it is compatible with Theano. Googling has surprisingly been unsuccessful.<\/p>\n\n<p>I know the 8800GT supports CUDA and I've done some CUDA work with it in the past, but is compatible with Theano? (or TensorFlow?)<\/p>\n\n<p>Best,\nJoe<\/p>\n","answers":[{"AnswerId":"37058946","CreationDate":"2016-05-05T19:39:10.340","ParentId":null,"OwnerUserId":"3688908","Title":null,"Body":"<p>Theano has no specific requirements for cards other than \"it works with cuda\".<\/p>\n\n<p>If you want to use the cuDNN layers or other specilized things, then you might need a more recent card and the requirements for those is specified in the documentation for those libraries.<\/p>\n"}]}
{"QuestionId":36457364,"AnswerCount":0,"Tags":"<python-3.x><numpy><tensorflow><theano><keras>","CreationDate":"2016-04-06T16:45:53.323","AcceptedAnswerId":null,"OwnerUserId":853462.0,"Title":"Does keras support custom numpy strides?","Body":"<p>Consider a 1D numpy array. One can simulate sliding windows upon this array by changing the shape and strides of the array. For example consider the a half parabola on which we apply two sliding windows:<\/p>\n\n<pre><\/pre>\n\n<p>Here each row in  has a matching row in  which precede each other in . That is:<\/p>\n\n<pre><\/pre>\n\n<p>The output are the row tuples of inputs and outputs, to which we'll fit a neural network shortly.<\/p>\n\n<pre><\/pre>\n\n<p>Notice that the 2 's shown above refer to the same memory. That is, we've created a sliding window <em>view<\/em> instead of a <em>copy<\/em>. The same goes for all 's, all 's, etc.<\/p>\n\n<p>I would like to know whether it is safe to use these for training a Keras neural network. For example:<\/p>\n\n<pre><\/pre>\n\n<p>The code seems to work, as the weights have normal values between -2 and 2 and the resulting RMSE is around 0.65.<\/p>\n\n<p>However, I might be pushing my luck. Does Keras consider that<\/p>\n\n<ol>\n<li>strides may take unusual values, in particular smaller than what  creates<\/li>\n<li>different elements in the data may refer to the same memory<\/li>\n<li>elements in inputs may refer to the same memory als elements in the output.<\/li>\n<\/ol>\n\n<p>Please do not answer with the alternative to simply materialize all the windows separately, nor with alternative libraries.<\/p>\n\n<p>I would like to know whether it is safe to use Keras like this, and whether the results are the same as when the data was materialized completely. If the answer depends on the backend, please answer it for both backend (Theano \/ Tensorflow).<\/p>\n","answers":[]}
{"QuestionId":36458573,"AnswerCount":1,"Tags":"<python><matlab><numpy><lua><torch>","CreationDate":"2016-04-06T17:46:11.823","AcceptedAnswerId":null,"OwnerUserId":2838606.0,"Title":"Torch\/Lua equivalent function to MATLAB or Numpy 'Unique'","Body":"<p>In Python one can do the following to get the unique values in a vector\/matrix\/tensor:<\/p>\n\n<pre><\/pre>\n\n<p>There is a similar function in MATLAB as well:<\/p>\n\n<pre><\/pre>\n\n<p>Is there an equivalent function in Torch\/Lua as well?<\/p>\n","answers":[{"AnswerId":"36459293","CreationDate":"2016-04-06T18:21:00.723","ParentId":null,"OwnerUserId":"1150918","Title":null,"Body":"<p>Nope, there is no such a standard function in stock Lua and\/or Torch.<\/p>\n\n<p>Considering using some implementation of <code>set<\/code> data structure, rolling Your own implementation of <code>unique()<\/code> or redesigning Your application not to require this kind of functionality.<\/p>\n\n<p>Example 11-liner:<\/p>\n\n<pre><code>function vector_unique(input_table)\n    local unique_elements = {} --tracking down all unique elements\n    local output_table = {} --result table\/vector\n\n    for _, value in ipairs(input_table) do\n        unique_elements[value] = true\n    end\n\n    for key, _ in pairs(unique_elements) do\n        table.insert(output_table, key)\n    end\n\n    return output_table\nend\n<\/code><\/pre>\n\n<p>Related questions:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/6618843\/lua-smartest-way-to-add-to-table-only-if-not-already-in-table-or-remove-duplic\">Lua: Smartest way to add to table only if not already in table, or remove duplicates<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/20066835\/lua-remove-duplicate-elements\">Lua : remove duplicate elements<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":36458820,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-04-06T17:57:16.837","AcceptedAnswerId":null,"OwnerUserId":1789784.0,"Title":"Exporting graphdef protobuf during training of inception model","Body":"<p>Has anyone incorporated the ability to save a graphdef protobuf along with each checkpoint in the inception-v3 model? <\/p>\n\n<p>Lively discussions at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/616\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/616<\/a> provided some solutions to export graph protobufs, but I cannot get any of them working using the inception training model (inception_train.py).<\/p>\n\n<p>I am trying to implement graph.util.convery_variables_to_constants, but I am failing to capture trainable variables, as I always get assertion errors: [variable] is not in graph. I am using output names from tensorboard, which may be incorrect procedure. <\/p>\n\n<p>Any solutions to this issue would definitely be of interest to the general community, not only for learning but deployment.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":36459233,"AnswerCount":1,"Tags":"<machine-learning><neural-network><theano><deep-learning>","CreationDate":"2016-04-06T18:17:58.133","AcceptedAnswerId":null,"OwnerUserId":5257450.0,"Title":"Weird erro in theano when running 3D CNN","Body":"<p>I am modifying the tutorial here, <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/tutorial\/lenet.html<\/a>, to a 3D convolutional neural network. <\/p>\n\n<p>However, I met a problem when I run, as,<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code, it seems the problem occurs when I change the dimension order with ? This is too weird, and I completely can't figure out why.<\/p>\n\n<p>Here is my code.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"39263713","CreationDate":"2016-09-01T06:00:33.917","ParentId":null,"OwnerUserId":"6779006","Title":null,"Body":"<p>Because I did not have your data. So I changed the getting data part of your code and then tested it. I did not get any bug.<\/p>\n\n<p>Make sure your input data is in the right form(with the shape of (50, 1, 51, 61, 23) in your case).<\/p>\n"}]}
{"QuestionId":36459266,"AnswerCount":1,"Tags":"<python><caffe><pycaffe>","CreationDate":"2016-04-06T18:19:35.210","AcceptedAnswerId":null,"OwnerUserId":2697940.0,"Title":"caffe python manual sgd","Body":"<p>I am trying to implement the SGD functionality to update weights in python manually in caffe python instead of using  function. The goal is to match the weight updates after doing  and that by manually updating the weights.<\/p>\n\n<p>The setup is as follows:\nUse MNIST data. Set the random seed in solver.prototxt as: . Make sure  and, , . Above is done so that, I can simply implement the SGD update equation (with out momentum, regularization etc.). The equation is simply: \nW_t+1 = W_t - mu * W_t_diff <\/p>\n\n<p>Following are the two tests:<\/p>\n\n<p><strong>Test1:<\/strong>\nUsing caffe's forward() and backward() to calculate the forward propagation and backward propagation. \nFor each layer that contain weights I do:<\/p>\n\n<pre><\/pre>\n\n<p>Next, update the weight\/biases as:<\/p>\n\n<pre><\/pre>\n\n<p>I run this for 5 iterations. <\/p>\n\n<p><strong>Test2<\/strong>: Run caffe's .<\/p>\n\n<p>Now, what I expect is the two tests should yield exactly same weights after the two iterations.<\/p>\n\n<p>I save the weights values after each of the above tests and calculate the norm difference between the weight vectors by the two tests, and I see that they are not bit-exact. Can some one spot something that I might be missing?<\/p>\n\n<p>Following is the entire code for reference:<\/p>\n\n<pre><\/pre>\n\n<p>The last line that compares the weights with the two tests produces:<\/p>\n\n<p>\nwhere as I expect this difference to be 0.0<\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"40708905","CreationDate":"2016-11-20T20:17:08.007","ParentId":null,"OwnerUserId":"7043417","Title":null,"Body":"<p>You got it almost right, you just need to set the diffs to zero after each update.\nCaffe wont do this automatically to give you the opportunity to implement batch accumulation (acummulate the gradients over multiple batches for one weight update, this can be helpful if your memory is not large enough for your desired batch size).<\/p>\n\n<p>Another possible problem could be the use of cudnn, its implementation of the convolution is non-deterministic (or how it is set to be used in caffe to be precise).\nIn general, this should be no problem, but in your case it causes slightly different results each time and therefore different weights.\nIf you compiled caffe with cudnn you can simply set the mode to cpu to prevent that from happening while testing.<\/p>\n\n<p>Also, the solver parameters have an impact on the weight updates. As you noted, you should be aware of:<\/p>\n\n<ul>\n<li>lr_policy: \"fixed\"<\/li>\n<li>momentum: 0<\/li>\n<li>weight_decay: 0<\/li>\n<li>random_seed: 52 # or any other constant<\/li>\n<\/ul>\n\n<p>In the nets, be sure not to use learn rate multipliers, often the biases are learned twice as fast as the weights, but this is not the behaviour you implemented. So you need to be sure to set them to one in the layer definitions:<\/p>\n\n<pre><code>param {\n    lr_mult: 1 # weight lr multiplier\n  }\nparam {\n    lr_mult: 1 # bias lr multiplier\n  }\n<\/code><\/pre>\n\n<p>Last but not least, here an example how your code would look like with momentum, weight decay and lr_mult. In CPU mode, this produces the expected output (no differences):<\/p>\n\n<pre><code>import caffe\ncaffe.set_device(0)\ncaffe.set_mode_cpu()\nimport numpy as np\n\nniter = 5\nsolver = None\nsolver = caffe.SGDSolver('solver.prototxt')\n\n# Automatic SGD: TEST2\nsolver.step(niter)\n# save the weights to compare later\nw_solver_step = solver.net.layers[1].blobs[0].data.copy()\nb_solver_step = solver.net.layers[1].blobs[1].data.copy()\n\n# Manual SGD: TEST1\nsolver = None\nsolver = caffe.SGDSolver('solver.prototxt')\nbase_lr = 0.01\nmomentum = 0.9\nweight_decay = 0.0005\nlr_w_mult = 1\nlr_b_mult = 2\n\nmomentum_hist = {}\nfor layer in solver.net.params:\n    m_w = np.zeros_like(solver.net.params[layer][0].data)\n    m_b = np.zeros_like(solver.net.params[layer][1].data)\n    momentum_hist[layer] = [m_w, m_b]\n\nfor i in range(niter):\n    solver.net.forward()\n    solver.net.backward()\n    for layer in solver.net.params:\n        momentum_hist[layer][0] = momentum_hist[layer][0] * momentum + (solver.net.params[layer][0].diff + weight_decay *\n                                                       solver.net.params[layer][0].data) * base_lr * lr_w_mult\n        momentum_hist[layer][1] = momentum_hist[layer][1] * momentum + (solver.net.params[layer][1].diff + weight_decay *\n                                                       solver.net.params[layer][1].data) * base_lr * lr_b_mult\n        solver.net.params[layer][0].data[...] -= momentum_hist[layer][0]\n        solver.net.params[layer][1].data[...] -= momentum_hist[layer][1]\n        solver.net.params[layer][0].diff[...] *= 0\n        solver.net.params[layer][1].diff[...] *= 0\n\n# save the weights to compare later\nw_fwdbwd_update = solver.net.layers[1].blobs[0].data.copy()\nb_fwdbwd_update = solver.net.layers[1].blobs[1].data.copy()\n\n# Compare\nprint \"after iter\", niter, \": weight diff: \", np.linalg.norm(w_solver_step - w_fwdbwd_update), \"and bias diff:\", np.linalg.norm(b_solver_step - b_fwdbwd_update)\n<\/code><\/pre>\n"}]}
{"QuestionId":36460144,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-04-06T19:03:05.387","AcceptedAnswerId":null,"OwnerUserId":1322301.0,"Title":"TensorFlow: Labeling summary images with filenames","Body":"<p>I'm training a CNN to learn an unusual task where the labels for each image depend on the other images in the minibatch. Further, the data undergo a number of transformations and through a bunch of different threads and queues before the net can actually axes them. <\/p>\n\n<p>I'd like to be able to validate that the label and image mapping is correct. However, it doesn't seem like it's possible to get TensorFlow to surface the filename for each image summary as part of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/train.html#image_summary\" rel=\"nofollow\"><\/a>. Does anyone know if it's possible to do this? Because the labeling regime is so unusual, you cannot gather immediately from the image if the label itself is correct. I need to be able to get the filename to recover the label properly. I can provide the filenames along with the images themselves and their labels in the queue without any problem.<\/p>\n\n<p>Edit: To clarify, I'm using TensorBoard to display the image summaries.<\/p>\n","answers":[]}
{"QuestionId":36461329,"AnswerCount":1,"Tags":"<lua><machine-learning><neural-network><backpropagation><torch>","CreationDate":"2016-04-06T20:08:07.380","AcceptedAnswerId":36488938.0,"OwnerUserId":2838606.0,"Title":"How to compute the gradient of loss with repect to an arbitrary layer\/weight in Torch?","Body":"<p>I'm transiting from Theano to Torch. So please bear with me. In Theano, it was kind of straight-forward to compute the gradients of loss function w.r.t even a specific weight. I wonder, how can one do this in Torch?<\/p>\n\n<p>Assume we have the following code which generates some data\/labels and defines a model :<\/p>\n\n<pre><\/pre>\n\n<p>How can I compute the gradient w.r.t weights of convolutinal layer?<\/p>\n","answers":[{"AnswerId":"36488938","CreationDate":"2016-04-07T23:25:54.347","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>Okay, I found the answer (thanks to alban desmaison on Torch7 Google group).\nThe code in the question has a bug and does not work. So I re-write the code. Here's how you can get the gradients with respect to each node\/parameter:<\/p>\n\n<pre><code>t = require 'torch'\nrequire 'cunn'\nrequire 'nn'\nrequire 'cutorch'\n\n\n\n-- A function to generate some random labels\nfunction randLabels(nExamples, nClasses)\n    -- nClasses: number of classes\n    -- nExamples: number of examples\n    label = {}\n    for i=1, nExamples do\n        label[i] = t.random(1, nClasses)\n    end\n    return t.FloatTensor(label)\nend\n\n-- Declare some variables\nnClass = 10\nkernelSize = 5\nstride = 2\npoolKernelSize = 2\nnData = 100\nnChannel = 3\nimageSize = 32\n\n-- Generate some [random] data\ndata = t.rand(nData, nChannel, imageSize, imageSize) -- 100 Random images with 3 channels\ndata = data:cuda() -- Transfer to the GPU (remove this line if you're not using GPU)\nlabel = randLabels(data:size()[1], nClass)\nlabel = label:cuda() -- Transfer to the GPU (remove this line if you're not using GPU)\n\n-- Define model\nnet = nn.Sequential()\nnet:add(nn.SpatialConvolution(3, 6, 5, 5))\nnet:add(nn.ReLU())\nnet:add(nn.SpatialMaxPooling(poolKernelSize, poolKernelSize, stride, stride))\nnet:add(nn.View(6*14*14))\nnet:add(nn.Linear(6*14*14, 350))\nnet:add(nn.ReLU())\nnet:add(nn.Linear(350, 10))\nnet = net:cuda() -- Transfer to the GPU (remove this line if you're not using GPU)\n\ncriterion = nn.CrossEntropyCriterion()\ncriterion = criterion:cuda() -- Transfer to the GPU (remove this line if you're not using GPU)\n\n-- Do forward pass and get the gradient for each node\/parameter:\n\nnet:forward(data) -- Do the forward propagation\ncriterion:forward(net.output, label) -- Computer the overall negative log-likelihood error\ncriterion:backward(net.output, label); -- Don't forget to put ';'. Otherwise you'll get everything printed on the screen\nnet:backward(data, criterion.gradInput); -- Don't forget to put ';'. Otherwise you'll get everything printed on the screen\n\n-- Now you can access the gradient values\n\nlayer1InputGrad = net:get(1).gradInput\nlayer1WeightGrads = net:get(1).gradWeight\n\nnet = nil\ndata = nil\nlabel = nil\ncriterion = nil\n<\/code><\/pre>\n\n<p>Copy and paste the code and it works like charm :)<\/p>\n"}]}
{"QuestionId":36462229,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-04-06T20:59:10.153","AcceptedAnswerId":36462756.0,"OwnerUserId":3251207.0,"Title":"TensorFlow: How do I release a model without source code?","Body":"<p>I am using Tensorflow + Python.<\/p>\n\n<p>I am curious if I can release a saved Tensorflow model (architecture + trained variables) without detailed source code. I'm aware of , but it looks to save only variables, and in order to restore\/run them, a user needs to \"define\" the same architecture. <\/p>\n\n<p>For the testing\/running purpose only, is there a way to release a saved {architecture+trained variables} without source code, so that a user can just cast a query and get a result?<\/p>\n","answers":[{"AnswerId":"38921538","CreationDate":"2016-08-12T15:29:03.630","ParentId":null,"OwnerUserId":"1121123","Title":null,"Body":"<p>You can build a <code>Saver<\/code> from the MetaGraphDef (saved with checkpoints by default: those .meta files). and then use that Saver to restore your model. So users don't have to re-define your graph in their code. But then they still need to figure out the model signature (input, output variables). I solve this using <code>tf.Collection<\/code> (but i am interested to find better ways to do it as well). <\/p>\n\n<p>You can take a look at my example implementation (the <code>eval.py<\/code> evaluate a model without re-defining a model):<\/p>\n\n<ul>\n<li>reconstruct saver from meta graph <a href=\"https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/eval.py#L18\" rel=\"nofollow\">https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/eval.py#L18<\/a><\/li>\n<li>get input variables from collections <a href=\"https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/eval.py#L58\" rel=\"nofollow\">https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/eval.py#L58<\/a><\/li>\n<li>how to define your model <a href=\"https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/models\/cp2f3d.py\" rel=\"nofollow\">https:\/\/github.com\/falcondai\/cifar10\/blob\/master\/models\/cp2f3d.py<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"36462756","CreationDate":"2016-04-06T21:32:28.693","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/tensorflow.github.io\/serving\/\" rel=\"nofollow\">TensorFlow Serving<\/a> project is intended to make this use case straightforward (assuming that the end user is only using the model for inference, not training). TensorFlow Serving includes an <code>Exporter<\/code> class that takes your <code>tf.train.Saver<\/code>, the <code>tf.GraphDef<\/code> that defines your overall model, and a \"signature\" that describes the inputs to and output from your model.<\/p>\n\n<p>The <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow\">basics tutorial<\/a> has a good introduction to exporting your model.<\/p>\n"}]}
{"QuestionId":36462853,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-06T21:39:58.400","AcceptedAnswerId":36463678.0,"OwnerUserId":563762.0,"Title":"Getting Probabilities of Outputs for CIFAR Example (or similar)","Body":"<p>Hey I'm trying to figure out how to get the probabilities for each of the classes in the CIFAR example. Currently I can input the predicted class in the input tensor and get back a true or false, but it won't tell me the class it thinks it is (out of 10) and the probability it places on that fact. Any ideas how to get this information after sess.run()?<\/p>\n","answers":[{"AnswerId":"36463678","CreationDate":"2016-04-06T22:50:37.247","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The CIFAR-10 example code for evaluation (in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py\" rel=\"noreferrer\"><code>cifar10_eval.py<\/code><\/a>) produces (as an intermediate output) a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/641cb259fb2622a1269e2a3332250cd0808539e1\/tensorflow\/models\/image\/cifar10\/cifar10_eval.py#L125\" rel=\"noreferrer\"><code>logits<\/code><\/a> tensor that contains log-probabilities that a given image is in each of the classes. Passing this tensor to <code>sess.run()<\/code> will give you the necessary information to tell you what the predicted class for an image is. If instead you compute <code>tf.nn.softmax(logits)<\/code> and pass that to <code>sess.run()<\/code>, you will get a probability (ranging from 0 to 1) for each class.<\/p>\n"}]}
{"QuestionId":36462962,"AnswerCount":4,"Tags":"<neural-network><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2016-04-06T21:47:13.300","AcceptedAnswerId":null,"OwnerUserId":5813490.0,"Title":"Loss clipping in tensor flow (on DeepMind's DQN)","Body":"<p>I am trying my own implementation of the DQN paper by Deepmind in tensor flow and am running into difficulty with clipping of the loss function. <\/p>\n\n<p>Here is an excerpt from the nature paper describing the loss clipping:<\/p>\n\n<blockquote>\n  <p>We also found it helpful to clip the error term from the update to be between \u22121 and 1. Because the absolute value loss function |x| has a derivative of \u22121 for all negative values of x and a derivative of 1 for all positive values of x, clipping the squared error to be between \u22121 and 1 corresponds to using an absolute value loss function for errors outside of the (\u22121,1) interval. This form of error clipping further improved the stability of the algorithm.<\/p>\n<\/blockquote>\n\n<p>(link to full paper: <a href=\"http:\/\/www.nature.com\/nature\/journal\/v518\/n7540\/full\/nature14236.html\" rel=\"noreferrer\">http:\/\/www.nature.com\/nature\/journal\/v518\/n7540\/full\/nature14236.html<\/a>)<\/p>\n\n<p>What I have tried so far is using <\/p>\n\n<pre><\/pre>\n\n<p>to clip the loss I calculate between -1 and +1. The agent is not learning the proper policy in this case. I printed out the gradients of the network and realized that if the loss falls below -1, the gradients all suddenly turn to 0!<\/p>\n\n<p>My reasoning for this happening is that the clipped loss is a constant function in (-inf,-1) U (1,inf), which means it has zero gradient in those regions. This in turn ensures that the gradients throughout the network are zero (think of it as, whatever input image I provide the network, the loss stays at -1 in the local neighborhood because it has been clipped).<\/p>\n\n<p>So, my question is two parts: <\/p>\n\n<ol>\n<li><p>What exactly did Deepmind mean in the excerpt? Did they mean that the loss below -1 is clipped to -1 and above +1 is clipped to +1. If so, how did they deal with the gradients (i.e. what is all that part about absolute value functions?) <\/p><\/li>\n<li><p>How should I implement loss clipping in tensor flow such that the gradients do not go to zero outside the clipped range (but maybe stay at +1 and -1)?\nThanks!<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"42985363","CreationDate":"2017-03-23T19:31:33.427","ParentId":null,"OwnerUserId":"5728213","Title":null,"Body":"<ol>\n<li>In the Deep Mind paper you reference, they limit the <em>gradient<\/em> of the loss. This prevents giant gradients and so improves robustness. They do this by using a quadratic loss function for errors inside a small range, and using an absolute value loss for larger errors.<\/li>\n<li><p>I suggest implementing the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Huber_loss\" rel=\"nofollow noreferrer\">Huber loss function<\/a>. Below is a python tensorflow implementation.<\/p>\n\n<pre><code>def huber_loss(y_true, y_pred, max_grad=1.):\n    \"\"\"Calculates the huber loss.\n\n    Parameters\n    ----------\n    y_true: np.array, tf.Tensor\n      Target value.\n    y_pred: np.array, tf.Tensor\n      Predicted value.\n    max_grad: float, optional\n      Positive floating point value. Represents the maximum possible\n      gradient magnitude.\n\n    Returns\n    -------\n    tf.Tensor\n      The huber loss.\n    \"\"\"\n    err = tf.abs(y_true - y_pred, name='abs')\n    mg = tf.constant(max_grad, name='max_grad')\n\n    lin = mg*(err-.5*mg)\n    quad=.5*err*err\n\n    return tf.where(err &lt; mg, quad, lin)\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"43720168","CreationDate":"2017-05-01T13:47:33.327","ParentId":null,"OwnerUserId":"252152","Title":null,"Body":"<p>First of all, the code for the paper is <a href=\"https:\/\/sites.google.com\/a\/deepmind.com\/dqn\/\" rel=\"nofollow noreferrer\">available online<\/a>, which constitutes an invaluable reference.<\/p>\n\n<h2>Part 1<\/h2>\n\n<p>If you take a look at the code you will see that, in <code>nql:getQUpdate<\/code> (<code>NeuralQLearner.lua<\/code>, line 180), they clip the error term of the Q-learning function:<\/p>\n\n<pre class=\"lang-lua prettyprint-override\"><code>-- delta = r + (1-terminal) * gamma * max_a Q(s2, a) - Q(s, a)\nif self.clip_delta then\n    delta[delta:ge(self.clip_delta)] = self.clip_delta\n    delta[delta:le(-self.clip_delta)] = -self.clip_delta\nend\n<\/code><\/pre>\n\n<h2>Part 2<\/h2>\n\n<p>In TensorFlow, assuming the last layer of your neural network is called <code>self.output<\/code>, <code>self.actions<\/code> is a one-hot encoding of all actions, <code>self.q_targets_<\/code> is a placeholder with the targets, and <code>self.q<\/code> is your computed Q:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code># The loss function\none = tf.Variable(1.0)\ndelta = self.q - self.q_targets_\nabsolute_delta = tf.abs(delta)\ndelta = tf.where(\n    absolute_delta &lt; one,\n    tf.square(delta),\n    tf.ones_like(delta) # squared error: (-1)^2 = 1\n)\n<\/code><\/pre>\n\n<p>Or, using <code>tf.clip_by_value<\/code> (and having an implementation closer to the original):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>delta = tf.clip_by_value(\n    self.q - self.q_targets_,               \n    -1.0,                \n    +1.0                 \n)                                                 \n<\/code><\/pre>\n"},{"AnswerId":"40235773","CreationDate":"2016-10-25T09:05:44.600","ParentId":null,"OwnerUserId":"5734313","Title":null,"Body":"<ol>\n<li><p>No. They talk about error clipping actually, not about loss clipping which is however as far as I know referring to the same thing but leads to confusion. They DO NOT mean that the loss below -1 is clipped to -1 and the loss above +1 is clipped to +1 because that leads to zero gradients outside the error range [-1;1] as you realized. Instead, they suggest to use a linear loss instead of a quadratic loss for error values &lt; -1 and error values > 1.<\/p><\/li>\n<li><p>Compute the error value (r + \\gamma \\max_{a'} Q(s',a'; \\theta_i^-) - Q(s,a; \\theta_i)). If this error value is within the range [-1;1], square it, if the error value is &lt; -1 multiply by -1, if the error value is > 1 leave it as it is. If you use this as loss function the gradients outside the interval [-1;1] won't vanish.<\/p><\/li>\n<\/ol>\n\n<p>In order to have a \"smooth-looking\" compound loss function you could also replace the squared loss outside the error range [-1;1] with a first-order Taylor approximation at the border values -1 and 1. In this case, if e was your error value, you would square it in case e \\in [-1;1], in case e &lt; -1, replace it by -2e-1, in case e > 1, replace it by 2e-1.<\/p>\n"},{"AnswerId":"38363141","CreationDate":"2016-07-13T23:09:01.117","ParentId":null,"OwnerUserId":"781723","Title":null,"Body":"<p>I suspect they mean that you should clip the <em>gradient<\/em> to [-1,1], not clip the <em>loss function<\/em>.  Thus, you compute the gradient as usual, but then clip each component of the gradient to be in the range [-1,1] (so if it is larger than +1, you replace it with +1; if it is smaller than -1, you replace it with -1); and then you use the result in the gradient descent update step instead of using the unmodified gradient.<\/p>\n\n<p><strong>Equivalently:<\/strong> Define a function <code>f<\/code> as follows:<\/p>\n\n<pre><code>f(x) = x^2          if x in [-0.5,0.5]\nf(x) = |x| - 0.25   if x &lt; -0.5 or x &gt; 0.5\n<\/code><\/pre>\n\n<p>Instead of using something of the form <code>s^2<\/code> as the loss function (where <code>s<\/code> is some complicated expression), they suggest to use <code>f(s)<\/code> as the loss function.  This is some kind of hybrid between squared-loss and absolute-value-loss: will behave like <code>s^2<\/code> when <code>s<\/code> is small, but when <code>s<\/code> gets larger, it will behave like the absolute value (<code>|s|<\/code>).<\/p>\n\n<p>Notice that the derivative of <code>f<\/code> has the nice property that its derivative will always be in the range [-1,1]:<\/p>\n\n<pre><code>f'(x) = 2x    if x in [-0.5,0.5]\nf'(x) = +1    if x &gt; +1\nf'(x) = -1    if x &lt; -1\n<\/code><\/pre>\n\n<p>Thus, when you take the gradient of this <code>f<\/code>-based loss function, the result will be the same as computing the gradient of a squared-loss and then clipping it.<\/p>\n\n<p>Thus, what they're doing is effectively replacing a squared-loss with a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Huber_loss\" rel=\"noreferrer\">Huber  loss<\/a>.  The function <code>f<\/code> is just two times the Huber loss for delta = 0.5.<\/p>\n\n<p>Now the point is that the following two alternatives are equivalent:<\/p>\n\n<ul>\n<li><p>Use a squared loss function.  Compute the gradient of this loss function, but the gradient to [-1,1] before doing the update step of the gradient descent.<\/p><\/li>\n<li><p>Use a Huber loss function instead of a squared loss function.  Compute the gradient of this loss function directly (unchanged) in the gradient descent.<\/p><\/li>\n<\/ul>\n\n<p>The former is easy to implement.  The latter has nice properties (improves stability; it's better than absolute-value-loss because it avoids oscillating around the minimum).  Because the two are equivalent, this means we get an easy-to-implement scheme that has the simplicity of squared-loss with the stability and robustness of the Huber loss.<\/p>\n"}]}
{"QuestionId":36464163,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2016-04-06T23:37:16.547","AcceptedAnswerId":null,"OwnerUserId":4332326.0,"Title":"installing hdf5, mattorch\/matio for lua\/torch","Body":"<p>I am configuring the packages for torch.I can't install these packages using .<\/p>\n\n<p>for the matio, I followed the instruction on github, the matio is installed. But I cannot require it. \nfor the hdf5,see following.<\/p>\n\n<p>Does anyone know what is the problem? I am new to Lua. Thanks very much!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38769814","CreationDate":"2016-08-04T14:12:13.917","ParentId":null,"OwnerUserId":"2687955","Title":null,"Body":"<p>I encountered the same problem on ubuntu >= 13.04, this command solves my problem:<\/p>\n\n<pre><code>sudo apt-get install libhdf5-serial-dev hdf5-tools\n<\/code><\/pre>\n\n<p>Then you can use luarocks to install hdf5<\/p>\n\n<pre><code>luarocks install hdf5\n<\/code><\/pre>\n\n<p>For other systems, you can reference <a href=\"https:\/\/github.com\/deepmind\/torch-hdf5\/blob\/master\/doc\/usage.md\" rel=\"nofollow\">this link<\/a> for dependencies.<\/p>\n"}]}
{"QuestionId":36464923,"AnswerCount":2,"Tags":"<python><deep-learning><caffe><pycaffe><vgg-net>","CreationDate":"2016-04-07T01:03:01.287","AcceptedAnswerId":36509155.0,"OwnerUserId":3427580.0,"Title":"Caffe shape mismatch error using pretrained VGG-16 model","Body":"<p>I am using PyCaffe to implement a neural network inspired by the VGG 16 layer network. I want to use the pre-trained model available from their <a href=\"https:\/\/gist.github.com\/ksimonyan\/211839e770f7b538e2d8#file-readme-md\" rel=\"nofollow\">GitHub page<\/a>. Generally this works by matching layer names.<\/p>\n\n<p>For my  layer I have the following definition in my train.prototxt file:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"http:\/\/cs.stanford.edu\/people\/karpathy\/vgg_train_val.prototxt\" rel=\"nofollow\">Here<\/a> is the prototxt file for the VGG-16 deploy architecture. Note that the in their prototxt is identical to mine (except for the learning rate, but that's irrelevant). It's also worth noting that the inputs are all the same size in my model too: 3-channel 224x224px images.<\/p>\n\n<p>I have been following <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/pascal-multilabel-with-datalayer.ipynb\" rel=\"nofollow\">this tutorial<\/a> pretty closely, and the block of code that's giving me an issue is the following:<\/p>\n\n<pre><\/pre>\n\n<p>The first line loads my solver prototxt and then the second line copies the weights from the pre-trained model (). When the solver runs, I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>The gist of it is that their model expects the layer to be of size 1x1x4096 while mine is just 4096. But I don't get how I can change this?<\/p>\n\n<p>I found <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/MJ9sIMn3krY\/OUY-TjW8DgAJ\" rel=\"nofollow\">this answer<\/a> in the Users Google group instructing me to do net surgery to reshape the pre-trained model before copying, but in order to do that I need the  files from the original architecture's data layers, which I don't have (it throws an error when I try to run the net surgery script).<\/p>\n","answers":[{"AnswerId":"44594228","CreationDate":"2017-06-16T16:45:21.700","ParentId":null,"OwnerUserId":"2737713","Title":null,"Body":"<p>This error comes up if you're cropping the images to 224, rather than the 227 that was done with the original dataset. Adjust that and you should be good to go. <\/p>\n"},{"AnswerId":"36509155","CreationDate":"2016-04-08T20:51:22.503","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>The problem is not with 4096 but rather with 25088. You need to calculate the output feature maps for each layer of your network based on the input feature maps. Note that the <code>fc<\/code> layer takes an input of fixed size so the output of the previous <code>conv<\/code> layer must match the input size required by the <code>fc<\/code> layer. Calculate your fc6 input feature map size (this is the output feature map of the previous <code>conv<\/code> layer) using the input feature map size of the previous <code>conv<\/code> layer. Here's the formula:<\/p>\n\n<pre><code>H_out = ( H_in + 2 x Padding_Height - Kernel_Height ) \/ Stride_Height + 1\nW_out = (W_in + 2 x Padding_Width - Kernel_Width) \/ Stride_Width + 1\n<\/code><\/pre>\n"}]}
{"QuestionId":36465665,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-07T02:28:21.363","AcceptedAnswerId":36466655.0,"OwnerUserId":2002387.0,"Title":"Performance of RNN with long sequences in Tensorflow","Body":"<p>I have build an RNN which has a sequence length of about 10,000. It took very long time to create the optimizer\n and took very long time to train. I am wondering how to expedite the computation process.<\/p>\n","answers":[{"AnswerId":"36466655","CreationDate":"2016-04-07T04:21:59.660","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The latest version of TensorFlow (current nightly build\/upcoming version 0.8) includes an experimental feature called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/51f5f6be3ede7242b01c4bdf136de938fd00839f\/tensorflow\/python\/ops\/rnn.py#L373\" rel=\"nofollow\"><code>tf.nn.dynamic_rnn()<\/code><\/a>. This function lets you build a RNN using a loop in the TensorFlow graph, which dramatically cuts the number of nodes used in the graph (and hence the size of the graph that must be built for gradient descent).<\/p>\n"}]}
{"QuestionId":36468701,"AnswerCount":1,"Tags":"<python><tensorflow><gradient-descent>","CreationDate":"2016-04-07T06:50:20.893","AcceptedAnswerId":36481418.0,"OwnerUserId":6170554.0,"Title":"AdamOptimizer and GradientDescentOptimizer from tensorflow not able to fit simple data","Body":"<p>Similar question: <a href=\"https:\/\/stackoverflow.com\/questions\/33720645\/why-is-this-tensorflow-implementation-vastly-less-successful-than-matlabs-nn\">Here<\/a><\/p>\n\n<p>I am trying out TensorFlow. I generated simple data which is linearly separable and tried to fit a linear equation to it. Here is the code.<\/p>\n\n<pre><\/pre>\n\n<p>I get the following output:<\/p>\n\n<blockquote>\n  <p>0.536667, 0.46, 0.46, 0.46, 0.46, 0.46, 0.46, 0.46, 0.46, 0.46, 0.46<\/p>\n<\/blockquote>\n\n<p>Right after the first iteration it gets struck at .<\/p>\n\n<p>And following is the plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3Cxf2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3Cxf2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Then I changed the code to use gradient descent:<\/p>\n\n<pre><\/pre>\n\n<p>Now i got the following: 0.54, 0.54, 0.63, 0.70, 0.75, 0.8, 0.84, 0.89, 0.92, 0.94, 0.94<\/p>\n\n<p>And following is the plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FKeZG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FKeZG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My questions:<\/p>\n\n<p>1) Why is the AdamOptimizer failing?<\/p>\n\n<p>2) If the issue is with learning rate, or other parameters which I need to tune, how do I generally debug them?<\/p>\n\n<p>3) I ran gradient descent for 50 iterations (I ran for 10 above) and printed the accuracy every 5 iterations and this is the output:<\/p>\n\n<blockquote>\n  <p>0.54, 0.8, 0.95, 0.96, 0.92, 0.89, 0.87, 0.84, 0.81, 0.79, 0.77.<\/p>\n<\/blockquote>\n\n<p>Clearly it started to diverge, looks like the issue is with fixed learning rate (it is overshooting after a point). Am I right?<\/p>\n\n<p>4) In this toy example what can be done to get a better fit. Ideally it should have 1.0 accuracy as the data is linearly separable.<\/p>\n\n<p>[EDIT]<\/p>\n\n<p>As requested by @Yaroslav, here is the code used for plots<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36481418","CreationDate":"2016-04-07T15:57:30.153","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TLDR; your loss is wrong. Loss goes to zero without decreasing accuracy.<\/p>\n\n<p>The problem is that your probabilities are not normalized. If you look at your loss, it's going down, but probabilities for both <code>y[:0]<\/code> and <code>y[:1]<\/code> are going to 1, so argmax is meaningless.<\/p>\n\n<p>Traditional solution is to use only 1 degree of freedom instead of 2, so your probability for first class is <code>sigmoid(y)<\/code>, and for second class it is <code>1-sigmoid(y)<\/code> so cross entropy is something like <code>-y[0]log(sigmoid(y0)) - y[1]log(1-sigmoid(y0))<\/code><\/p>\n\n<p>Alternatively you could change your code is to use <code>tf.nn.softmax<\/code> instead of <code>tf.sigmoid<\/code>. This divides by the sum of the probabilities so the optimizer can't decrease loss by driving both probabilities to 1 simultaneously. <\/p>\n\n<p>The following gets to <code>0.99666673<\/code> accuracy.<\/p>\n\n<pre><code>tf.reset_default_graph()\nnp.random.seed(2010)\nn = 300\nx_data = np.random.random([n, 2]).tolist()\ny_data = [[1., 0.] if v[0]&gt; 0.5 else [0., 1.] for v in x_data]\n\nx = tf.placeholder(tf.float32, [None, 2]) \nW = tf.Variable(tf.zeros([2, 2]))\nb = tf.Variable(tf.zeros([2]))\ny = tf.nn.softmax(tf.matmul(x , W) + b)\n\ny_ = tf.placeholder(tf.float32, [None, 2]) \ncross_entropy = -tf.reduce_sum(y_ * tf.log(y))\nregularizer = tf.reduce_sum(tf.square(y))\ntrain_step = tf.train.AdamOptimizer(1.0).minimize(cross_entropy+regularizer)\n\ncorrect_predict = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1)) \naccuracy = tf.reduce_mean(tf.cast(correct_predict, tf.float32))\n\ns = tf.Session()\ns.run(tf.initialize_all_variables())\n\nfor i in range(30):\n        s.run(train_step, feed_dict = {x: x_data, y_: y_data})\n        cost1,cost2=s.run([cross_entropy,accuracy], feed_dict = {x: x_data, y_: y_data})\n        print(cost1, cost2)\n<\/code><\/pre>\n\n<p>PS: can you share the code you used for making the plots above?<\/p>\n"}]}
{"QuestionId":36471758,"AnswerCount":2,"Tags":"<regression><keras>","CreationDate":"2016-04-07T09:19:02.313","AcceptedAnswerId":null,"OwnerUserId":3439054.0,"Title":"Simple regression with Keras seems not working properly","Body":"<p>I am trying, just for practising with <strong>Keras<\/strong>, to train a network to learn a very easy function.\nThe input of the network is <strong>2Dimensional<\/strong> . The output is <strong>one dimensional<\/strong>.\nThe function can indeed represented with an image, and the same is for the approximate function. \nAt the moment I'm not looking for any good generalization, I just want that the network is at least good in representing the training set.\nHere I place my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>As you can see, the two function are completely different, and I can't understand why.\nI think that maybe <em>model.predict<\/em> doesn't work as I axpect.<\/p>\n","answers":[{"AnswerId":"36843894","CreationDate":"2016-04-25T14:45:13.743","ParentId":null,"OwnerUserId":"2637126","Title":null,"Body":"<p>Your understanding is correct; it's just a question of hyperparameter tuning.<\/p>\n\n<p>I just tried your code, and it looks like you're not giving your training enough time:<\/p>\n\n<p>Look at the loss, under 100 epochs, it's stuck at around 0.23. But try using the 'adam' otimizer instead of SGD, and increase the number of epochs up to 10,000: the loss now decreases down to 0.09 and your picture looks much better.<\/p>\n\n<p>If it's still not precise enough for you, you may also want to try increasing the number of parameters: just add a few layers; this will make overfitting much easier ! :-)<\/p>\n"},{"AnswerId":"42832104","CreationDate":"2017-03-16T10:57:38.350","ParentId":null,"OwnerUserId":"5129506","Title":null,"Body":"<p>I have changed just your network structure and added a training dataset. The loss decreases down to 0.01.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/H5QQa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/H5QQa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code># -*- coding: utf-8 -*-\n\"\"\"\nCreated on Thu Mar 16 15:26:52 2017\n\n@author: Administrator\n\"\"\"\n\nimport matplotlib.pyplot as plt\nimport numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation\nfrom keras.optimizers import SGD\nimport random as rnd\nimport math\nfrom keras.optimizers import Adam,SGD\nm = [\n[1,1,1,1,0,0,0,0,1,1],\n[1,1,0,0,0,0,0,0,1,1],\n[1,0,0,0,1,1,0,1,0,0],\n[1,0,0,1,0,0,0,0,0,0],\n[0,0,0,0,1,1,0,0,0,0],\n[0,0,0,0,1,1,0,0,0,0],\n[0,0,0,0,0,0,1,0,0,1],\n[0,0,1,0,1,1,0,0,0,1],\n[1,1,0,0,0,0,0,0,1,1],\n[1,1,0,0,0,0,1,1,1,1]]                      #A representation of the function that I would like to approximize\n\nmatrix = np.matrix(m)\n\nevaluation = np.zeros((1000,1000))\n\nx_train = np.zeros((1000000,2))\ny_train = np.zeros((1000000,1))\n\nfor x in range(0,1000):\n    for y in range(0,1000):\n         x_train[x+1000*y,0] = x\/1000.                #I normilize the input of the function, between [0,1)\n         x_train[x+1000*y,1] = y\/1000.\n         y_train[x+1000*y,0] = matrix[int(x\/100),int(y\/100)] +0.0\n\n\n#Here I show graphically what I would like to have\nplt.matshow(matrix, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))\n\n#Here I built the model\nmodel = Sequential()\nmodel.add(Dense(50, input_dim=2, init='uniform'))## init\u662f\u5173\u952e\u5b57\uff0c\u2019uniform\u2019\u8868\u793a\u7528\u5747\u5300\u5206\u5e03\u53bb\u521d\u59cb\u5316\u6743\u91cd\nmodel.add(Activation('tanh'))\nmodel.add(Dense(20, init='uniform'))\nmodel.add(Activation('tanh'))\nmodel.add(Dense(1, init='uniform'))\nmodel.add(Activation('sigmoid'))\n\n#Here I train it\n#sgd = SGD(lr=0.01)\nadam = Adam(lr = 0.01)\nmodel.compile(loss='mean_squared_error', optimizer=adam)\n\nmodel.fit(x_train, y_train,\n          nb_epoch=100,\n          batch_size=100,\n          show_accuracy=True)\n\n#Here (I'm not sure), I'm using the network over the given example\nx = model.predict(x_train,batch_size=1)\n\n#Here I show the approximated function\nprint (x)\nprint (x_train)\nfor i in range(0, 1000000):\n    evaluation[int(x_train[i,0]*1000),int(x_train[i,1]*1000)] = x[i]\n\nplt.matshow(evaluation, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))\nplt.colorbar()\nplt.show()\n<\/code><\/pre>\n"}]}
{"QuestionId":36475189,"AnswerCount":1,"Tags":"<java><linux><terminal><torch>","CreationDate":"2016-04-07T11:43:39.397","AcceptedAnswerId":null,"OwnerUserId":4299527.0,"Title":"How to execute multiple commands interactively in linux terminal from Java?","Body":"<p>I want to run multiple commands in the  and that will be one like as follows:<\/p>\n\n<p><strong>1.<\/strong> I will run suppose  and i wrote  command and it opened the torch promt.\n<br>\n<strong>2.<\/strong> Now if i execute next command from java then it will run in the torch promt of the linux terminal.<br><br>\n <strong><em>you can take another example like:<\/em><\/strong><br><br>\n<strong>1.<\/strong> At fist i will run  in the linux terminal from java.<br>\n<strong>2.<\/strong> Then run  or anything in the python interpreter from java.<br>\nSo here my second command is dependent on the first command. I want to run the commands in the terminal in a sequential way.<\/p>\n\n<p><strong>Edit<\/strong>: Another approach will do.Suppose i have a python interpreter running in linux terminal and now i want to execute a command from java that will run in the python interpreter of that particular opened linux terminal. <strong>Can i do that??<\/strong><\/p>\n\n<p>I have tried to to run a command using this:<\/p>\n\n<pre><\/pre>\n\n<p>But now after this command how can i run a command from java in the promt that means i want it interactively ??<\/p>\n\n<p><strong>So how can i save the state of the previous command and use it for the next command??<\/strong><\/p>\n","answers":[{"AnswerId":"36477282","CreationDate":"2016-04-07T13:12:03.107","ParentId":null,"OwnerUserId":"65464","Title":null,"Body":"<p>Use <code>p.getOutputStream()<\/code> to get a stream into which you can write the desired input for the program you run. (I know, the naming is a bit confusing).<\/p>\n\n<p>Just the relevant part:<\/p>\n\n<pre><code>    Process p = Runtime.getRuntime().exec(execstr);\n    BufferedReader input = new BufferedReader(new InputStreamReader(p.getInputStream()));\n    OutputStream ops = p.getOutputStream();\n    ops.write(\"echo hello world\".getBytes());\n    ops.close();            \n\n    while ((line = input.readLine()) != null) \n    {\n       System.out.println(line);\n    }\n    input.close();\n<\/code><\/pre>\n"}]}
{"QuestionId":36477056,"AnswerCount":2,"Tags":"<python><speech-recognition><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2016-04-07T13:03:05.570","AcceptedAnswerId":36885929.0,"OwnerUserId":6171980.0,"Title":"Tensor flow LSTM for speech recognition slows down when training each subsequent word","Body":"<p>so I am trying to use tensorflows LSTM to recognise spoken words. However, after each training word is passed through the LSTM the next word takes longer to process (specifically the calculation of gradient by back propogation and the application of these to the network). I am currently working on an iMac without CUDA support so I am having to use a CPU not a GPU (I will be switching to GPU once I am able to).<\/p>\n\n<p>I am programming using Python-2.7<\/p>\n\n<p>I am using a very small vocabulary, 8 word classes with 10 training examples of each and each word is isolated (is not part of a sentence, just a word on its own).<\/p>\n\n<p>Each word is preprocessed into Mel Frequency Cepstral coefficients and then these are clustered using Kmeans with K = 100. So the input to the LSTM is a list of float32s fed in one item at a time.<\/p>\n\n<p>The slowdown is definitely occurring in the LSTM as the time taken to fetch each item from the list and pass it to the LSTM is staying roughly the same for each item. The size of each item passed to the LSTM is also the same each time (longer words simply have longer lists of items); however, even shorter words (with less items in their list) are still taking progressively longer as the training continues.<\/p>\n\n<p>I am using gradient decent and backpropogation to train the network and have tried clipping gradients to 10 time steps or not clipping at all, which makes no difference.<\/p>\n\n<p>The LSTM is instantiated with: <\/p>\n\n<pre><\/pre>\n\n<p>After the LSTM I have a softmax layer; the output of which is compared to a one hot vector which represents the correct output, using cross enthalpy loss.<\/p>\n\n<p>Pipeline sudo code:<\/p>\n\n<pre><\/pre>\n\n<p>Finally, in case I haven't been clear about what my problem is, here is the timings from my network:<\/p>\n\n<pre><\/pre>\n\n<p>If anyone has any ideas as to why it's taking longer and longer <\/p>\n","answers":[{"AnswerId":"36885929","CreationDate":"2016-04-27T09:37:47.747","ParentId":null,"OwnerUserId":"6171980","Title":null,"Body":"<p>The issue was caused by a using tf.assign each time a word segment was fed in. This causes the graph to grow as a new variable is created each time. In my case, each word was broken into around 250 pieces hence the rapid and large increase in time taken.<\/p>\n\n<p>The issue was solved by removing the tf.assign method and replacing the offending variable with a placeholder which is set using the feed_dict not tf.assign.<\/p>\n\n<p>Thanks to @mrry for helping me!<\/p>\n"},{"AnswerId":"43685456","CreationDate":"2017-04-28T17:04:33.753","ParentId":null,"OwnerUserId":"6784405","Title":null,"Body":"<p>A nice trick to help with this kind of bug is <code>Graph.finalize()<\/code>... once you have added all the graph ops you intended to, call finalize and any attempts to add additional ops to the graph will throw an error. Was very helpful to me when I started and hadn't fully internalized the \"I build the graph ONCE and get pointers to tensors as I do, then I exclusively run ops\/call eval on the tensors to do stuff\" mantra<\/p>\n"}]}
{"QuestionId":36479335,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><deep-learning><lstm>","CreationDate":"2016-04-07T14:32:12.067","AcceptedAnswerId":null,"OwnerUserId":6172644.0,"Title":"Dynamic LSTM in Tensorflow - Optimizer Issues","Body":"<p>I have implemented the following LSTM class in Tensorflow, where the unroll operation is inspired by the dynamic_rnn() implementation within Tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>I define a graph using the LSTM with some cost function. The graph compiles properly, and I'm able to forward propagate using 'in_batch' which is of size [sequence_length, batch_size, input_dim]. 'sequence_length' can vary for different batches. However, when I use an optimizer(Adam) with the cost function, I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>Here 1000 is 'dim_x' and 512 is 4*'dim_h', so the error is for 'W_x_h'. I have tried using '.set_shape()' for 'x_t', 'h_t', 'c_t' and 'output' in 'lstmUnroll()', still fails.<\/p>\n\n<p>Any ideas to make it work with the optimizer?<\/p>\n","answers":[]}
{"QuestionId":36480456,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-07T15:17:19.823","AcceptedAnswerId":null,"OwnerUserId":6172975.0,"Title":"dynamic_rnn() and array_ops.reverse_sequence() problems","Body":"<p>I am trying to reverse my inputs with  before sending it to , the inference graph can be build with no problem, but when building the training graph, I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what went wrong?<\/p>\n","answers":[{"AnswerId":"36594699","CreationDate":"2016-04-13T09:50:58.760","ParentId":null,"OwnerUserId":"6172975","Title":null,"Body":"<p>This has been fixed at the mater in TensorFlow<\/p>\n"}]}
{"QuestionId":36481372,"AnswerCount":1,"Tags":"<python><while-loop><theano>","CreationDate":"2016-04-07T15:55:25.973","AcceptedAnswerId":36561778.0,"OwnerUserId":2014591.0,"Title":"Theano: Using theano.scan with theano.scan_module.until","Body":"<p>I'm new to using  and . From the docs <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#theano.scan\" rel=\"nofollow\">here<\/a>, I'm not sure how to set variables in my  loop, and I'm uncertain how to adapt this post <a href=\"http:\/\/theano-users.narkive.com\/cCNzuH3M\/is-scan-module-until-more-like-while-or-do-while#post2\" rel=\"nofollow\">here<\/a> to use .<\/p>\n\n<p>This is the code I'd like to translate to equivalent theano. Someone wanna take a shot in translating this? (And perhaps explaining the translated code.)<\/p>\n\n<pre><\/pre>\n\n<p>I see three assignment operations here, and one if-statement. But I don't know how to translate this to theano.<\/p>\n\n<p>And if you were curious, you could paste this code above to set the variables<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36561778","CreationDate":"2016-04-12T00:53:34.687","ParentId":null,"OwnerUserId":"1828879","Title":null,"Body":"<p>Given:<\/p>\n\n<pre><code>N = 3\ntol = 1.0e-6\n<\/code><\/pre>\n\n<p>You can define your symbolic function like this:<\/p>\n\n<pre><code>r_init = T.vector()\nW = T.matrix()\n\ndef step(r):\n    r_prime = T.dot(r, W)\n    delta = (r_prime - r).norm(1)\n    condition = T.lt(delta, N * tol)\n    return r_prime, theano.scan_module.until(condition)\n\noutputs, updates = theano.scan(step, outputs_info=[r_init], n_steps=1024)\nr_final = outputs[-1]\n\nsolve = theano.function(inputs=[r_init, W], outputs=r_final)\n<\/code><\/pre>\n\n<p>And then use it like this:<\/p>\n\n<pre><code>M = np.random.rand(N, N)\nM \/= M.sum(axis=1).reshape((-1, 1))\n\nr = np.ones(N, dtype=np.float) \/ N\n\nprint solve(r, M)\n<\/code><\/pre>\n\n<p>By the way, you are not performing a \"random walk.\"  You are solving for r such that rW = r, typically called the stationary distribution of the Markov chain.<\/p>\n"}]}
{"QuestionId":36484283,"AnswerCount":1,"Tags":"<python-2.7><numpy><theano>","CreationDate":"2016-04-07T18:23:25.207","AcceptedAnswerId":36505300.0,"OwnerUserId":4565943.0,"Title":"How to do element-wise conditional indexing comparison in Theano?","Body":"<p>The operation consists of two arrays  and  of equal length where the values of  can vary between 0 to (k-1) with the value of k given.<\/p>\n\n<p>This is the general Python code to illustrate this.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Sample output:<\/p>\n\n<pre><\/pre>\n\n<p>Note that there is actually a reason for me to represent  as a matrix and not as a vector. It was initialised to  as part of its computation, where  the size of . <\/p>\n\n<p>I tried implement this in Theano like so<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>But I received this error at line where  is defined:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a simple way to implement this in Theano?<\/p>\n","answers":[{"AnswerId":"36505300","CreationDate":"2016-04-08T16:59:51.713","ParentId":null,"OwnerUserId":"4565943","Title":null,"Body":"<p>Use <code>nonzero()<\/code> and correct the dimensions of <code>idx<\/code>.<\/p>\n\n<p>This code solved the problem<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import theano\nimport theano.tensor as T\n\nX = T.vector('X')\nidx = T.vector('idx')\nk = T.scalar()\nc, updates = theano.scan(lambda i: X[T.eq(idx,i).nonzero()], sequences=T.arange(k)) \nf = function([X,idx,k],c)\n<\/code><\/pre>\n\n<p>For the same example, through the use of Theano:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import numpy as np\n\nX = np.arange(6) \nk = 3\nidx = np.array([[0, 1, 2, 2, 0, 1]]).T\n\nf(X, idx.T[0], k).astype(int)\n<\/code><\/pre>\n\n<p>This gives the output as<\/p>\n\n<pre><code>array([[0, 4],\n       [1, 5],\n       [2, 3]])\n<\/code><\/pre>\n\n<p>If idx is defined as <code>np.array([0, 1, 2, 2, 0, 1])<\/code>, then <code>f(X, idx, k)<\/code> can be used instead.<\/p>\n"}]}
{"QuestionId":36484408,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-07T18:30:06.490","AcceptedAnswerId":36498589.0,"OwnerUserId":3320135.0,"Title":"Implementing a Siamese Network in Tensor Flow","Body":"<p>\nI want to implement a Siamese Convolutional Neural Network, where two images share weights in the convolutional layers, and are then concatenated before being passed through the fully-connected layers. I have tried an implementation, but it seems rather a \"hacked\" solution. In particular, I have defined an operation on tensors as simply a Python function, and I'm not sure whether this is allowed.<\/p>\n\n<p>Here is the code I have tried:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Whilst this compiles and seems to run fine, the results are not really as expected. So, I'm wondering if there is a better way to implement a Siamese network using built-in operations in TensorFlow?<\/p>\n","answers":[{"AnswerId":"36498589","CreationDate":"2016-04-08T11:35:55.480","ParentId":null,"OwnerUserId":"4524845","Title":null,"Body":"<p>\nYou can make use of <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#pack\" rel=\"nofollow noreferrer\">tf.pack<\/a> and <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/array_ops.html#unpack\" rel=\"nofollow noreferrer\">tf.unpack<\/a>, somewhat like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>pairs = tf.pack(tf.split(0, 64, pool3_flat))\nleft, right = tf.unpack(tf.transpose(pairs, perm=[1,0,2]))\nmerged_pairs = tf.concat(1, [left, right])\n<\/code><\/pre>\n\n<p>A cleaner way to do this is to keep your pairs separate from the beginning, so that you can define two networks and use the same trainable variables in each network.<\/p>\n\n<p>You would have something like (skipping the convolutional layers):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>image_left = tf.placeholder(tf.float32, shape=[None, 64, 64, 1])\nimage_right = tf.placeholder(tf.float32, shape=[None, 64, 64, 1])\n\npool_left = tf.nn.max_pool(image_left, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\npool_right = tf.nn.max_pool(image_left, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n\npool_flat_left = tf.reshape(pool_left, [-1, 32*32])\npool_flat_right = tf.reshape(pool_right, [-1, 32*32])\n<\/code><\/pre>\n\n<p>Then simply concat left and right in dimension 1.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>concat_layer = tf.concat(1, [pool_flat_left, pool_flat_right])\n<\/code><\/pre>\n\n<p>This way you can also vary the batch size later.\nMake sure to use the same weights and biases on each size (left and right).<\/p>\n"}]}
{"QuestionId":36484695,"AnswerCount":1,"Tags":"<macos><installation><gpu><theano><gpu-programming>","CreationDate":"2016-04-07T18:46:04.683","AcceptedAnswerId":null,"OwnerUserId":4565943.0,"Title":"How to make Theano use the integrated GPU in MacBook Air?","Body":"<p>I tried running the test program for GPU usage:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>It only shows this (even after installing ):<\/p>\n\n<pre><\/pre>\n\n<p>I would like to know how to utilise the integrated GPU of MacBook Air (Early 2014). <\/p>\n\n<p>My processor has Intel HD Graphics 5000 -- not NVIDIA and hence not CUDA compatible Many <a href=\"https:\/\/developer.apple.com\/library\/mac\/documentation\/Performance\/Conceptual\/OpenCL_MacProgGuide\/Introduction\/Introduction.html\" rel=\"nofollow\">links<\/a> suggest usage of OpenCL. This was also supposed to be pre-installed with OS-X. But I can't make head or tail out of the links in the web.<\/p>\n\n<p>I could not find much help of how to set up Theano in the <a href=\"http:\/\/deeplearning.net\/software\/theano\/\" rel=\"nofollow\">docs<\/a> either.<\/p>\n\n<p>I just need to make Theano use my Mac's integrated GPU. Is this possible? If so, how? What are its prerequisites?<\/p>\n","answers":[{"AnswerId":"44287274","CreationDate":"2017-05-31T14:19:12.533","ParentId":null,"OwnerUserId":"8092557","Title":null,"Body":"<pre><code>THEANO_FLAGS=device=opencl0:1 python ~\/check_GPU.py\n<\/code><\/pre>\n"}]}
{"QuestionId":36488429,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-04-07T22:37:03.350","AcceptedAnswerId":null,"OwnerUserId":2827214.0,"Title":"How can I take the 2nd max from a Tensorflow tensor?","Body":"<p>Right now, my function uses argmax:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried using the following, but the dimn are incompatible: <\/p>\n\n<pre><\/pre>\n\n<p>My TF version might be 0.5, which is why top_k only has 2 args.<\/p>\n","answers":[{"AnswerId":"36494377","CreationDate":"2016-04-08T07:54:25.827","ParentId":null,"OwnerUserId":"2267819","Title":null,"Body":"<p>Check the documentation for tf.nn.top_k(). The function returns values and indices. So something like below should work.<\/p>\n\n<pre><code>values, indices = tf.nn.top_k(prev,2)\np = tf.stop_gradient(indices[1])\n<\/code><\/pre>\n"}]}
{"QuestionId":36489214,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-04-07T23:53:23.023","AcceptedAnswerId":null,"OwnerUserId":5324588.0,"Title":"Torch: tensor vs table: disk space","Body":"<p>I have ~ 15000 3-D tensors (each one with size 2x300x20) that i need to save to disk. I've checked 2 options: \n1. One big Tensor of size 15000x2x300x20\n2. table of the 15000 tensors.<\/p>\n\n<p>I prefer to save it in one big tensor (via torch.save()) but unfortunately the new file is much much bigger in the first option. why is that ?  is there an efficient way to save tensor to disk ?\n(for example, for 160 tensors (instead of 15000) the file is 1.3 GB relative to 10 MG in the second option).<\/p>\n","answers":[{"AnswerId":"36494910","CreationDate":"2016-04-08T08:25:27.863","ParentId":null,"OwnerUserId":"1688185","Title":null,"Body":"<p>The 2nd option (table of tensors) has an overhead since it includes N (= 15,000) times the header of each 2x300x20 tensor (see below). But this overhead is here negligible w.r.t. the total amount of data. So both options should be <em>roughly<\/em> equivalent in terms of space.<\/p>\n\n<p>That being said please note that the underlying storage is part of the archive. That means that if for some reason, the storage is larger than the current tensor size, the archive will be <em>large<\/em>, e.g.:<\/p>\n\n<pre><code>x = torch.Tensor(100000)\nx[1] = 1234\nx:resize(1)\ntorch.save(\"x.t7\", x)\n\ny = torch.Tensor(1)\ny[1] = 1234\ntorch.save(\"y.t7\", y)\n<\/code><\/pre>\n\n<p>Here <code>x.t7<\/code> is around 782KB vs. 119B for <code>y.t7<\/code>, because it refers to an underlying storage of 100,000 elements.<\/p>\n\n<p>In your first option, you should double check that you are not in this case.<\/p>\n\n<p>--<\/p>\n\n<p>e.g. serializing a dummy Torch tensor in ASCII mode:<\/p>\n\n<pre><code>$ th -e \"torch.save('foo.t7', torch.Tensor{1234}, 'ascii')\"\n$ xxd -g1 foo.t7\n00000000: 34 0a 31 0a 33 0a 56 20 31 0a 31 38 0a 74 6f 72  4.1.3.V 1.18.tor\n00000010: 63 68 2e 44 6f 75 62 6c 65 54 65 6e 73 6f 72 0a  ch.DoubleTensor.\n00000020: 31 0a 31 0a 31 0a 31 0a 34 0a 32 0a 33 0a 56 20  1.1.1.1.4.2.3.V \n00000030: 31 0a 31 39 0a 74 6f 72 63 68 2e 44 6f 75 62 6c  1.19.torch.Doubl\n00000040: 65 53 74 6f 72 61 67 65 0a 31 0a 31 32 33 34 0a  eStorage.1.1234.\n<\/code><\/pre>\n\n<p>As you can see the archive includes: a first integer (4 here) that denotes the <a href=\"https:\/\/github.com\/torch\/torch7\/blob\/68d2854283a6aafcbb1ca2db5270005602dc2dab\/File.lua#L15-L23\" rel=\"nofollow\">type of the object<\/a>, and for a Torch class other metadata like its version (<code>V 1<\/code> here), etc, and then the final value (<code>1234<\/code> here).<\/p>\n"}]}
{"QuestionId":36489345,"AnswerCount":0,"Tags":"<python><amazon-web-services><apache-spark><docker><caffe>","CreationDate":"2016-04-08T00:08:08.727","AcceptedAnswerId":null,"OwnerUserId":3785114.0,"Title":"Best way to to implement Spark + AWS + Caffe\/CUDA?","Body":"<p>I am looking to deploy an application that already has a trained caffemodel file and I need to deploy it to a Spark cluster on AWS for processing due to GPU computation power needed (20K patches per image). From my research it seems that the best way to do it is to use Spark to create an AWS cluster which then runs a Docker image or Amazon AMI to install project dependencies automatically. Once everything is installed, the job can run in the cluster through Spark. What I am wondering is how to do this from start to finish. I have seen several guides, and have taken some online courses on Spark (BerkeleyX, Udemy) and Docker (Udemy); however almost all the information I have seen are examples of how to implement the simplest application that has little to no heavy software dependencies (CUDA drivers, CuDNN, Caffe, DIGITS). I have deployed Spark clusters on AWS and ran simple examples that had no dependencies, but have found little to no information on running an application that would require even a small dependency such as numpy. I would like to leverage the group to see if anyone has experience in such an implementation and can point me in the right direction or offer some help\/suggestions?<\/p>\n\n<p>Here are some things I have looked into:<br>\nDocker+NVIDIA: <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\" rel=\"nofollow\">https:\/\/github.com\/NVIDIA\/nvidia-docker<\/a><br>\nbitfusion AMI: <a href=\"https:\/\/aws.amazon.com\/marketplace\/pp\/B01DJ93C7Q\/ref=sp_mpg_product_title?ie=UTF8&amp;sr=0-13\" rel=\"nofollow\">https:\/\/aws.amazon.com\/marketplace\/pp\/B01DJ93C7Q\/ref=sp_mpg_product_title?ie=UTF8&amp;sr=0-13<\/a><\/p>\n\n<p>My question is in regards to how to implement a small sample application from start to end with the Spark cluster getting created automatically side-by-side while installing the dependencies needed through either Docker or an AMI from above?<\/p>\n\n<p>Notes:<br>\nPlatform: Ubuntu 14.04<br>\nLanguage: Python<br>\nDependencies: CUDA 7.5, caffe\u00adnv, libcudnn4, NVIDIA Graphics Driver (346-352)<\/p>\n","answers":[]}
{"QuestionId":36490609,"AnswerCount":1,"Tags":"<python><python-import><caffe>","CreationDate":"2016-04-08T02:40:35.483","AcceptedAnswerId":null,"OwnerUserId":1270789.0,"Title":"Understanding imports and scopes and caffe","Body":"<p>I have the two files:<\/p>\n\n<h1>main.py<\/h1>\n\n<pre><\/pre>\n\n<h1>mylib.py<\/h1>\n\n<pre><\/pre>\n\n<p>However, the code runs on the CPU despite the almost identical code in an alternative version of  (my two modules currently communicate via , so I want to instead call directly) working as expected when invoked directly via .<\/p>\n\n<p>So, is there something I don't know about importing modules and contexts? Do I need some extra scope information to make sure I'm setting GPU mode in the correct context?<\/p>\n\n<p><strong>UPDATE<\/strong>: I tried pasting the  into , and the code still fails to use the GPU, so perhaps there is some obscure clash between  and other imports, like  or various ROS modules.<\/p>\n","answers":[{"AnswerId":"36492002","CreationDate":"2016-04-08T05:12:06.153","ParentId":null,"OwnerUserId":"1270789","Title":null,"Body":"<p>That was bizarre - the solution for my problem was:<\/p>\n\n<h1>mylib.py<\/h1>\n\n<pre><code>import caffe\n\nclass MyObj:\n    def __init__(self, gpuid):\n        caffe.set_device(gpuid)\n        self.nd = caffe.Net(...)\n\n    def docaffe():\n        caffe.set_mode_gpu()\n        self.nd.forward(...)\n<\/code><\/pre>\n\n<p>It didn't work moving <code>caffe.set_mode_gpu()<\/code> to either the start or end of the <code>__init__<\/code> method, it was only before calling the guts of <code>caffe<\/code> that it seemed to stick. I believe the problem was related to using threads and creating <code>MyObj<\/code> in one thread, but then using it in another.<\/p>\n"}]}
{"QuestionId":36490781,"AnswerCount":1,"Tags":"<matlab><deep-learning><caffe>","CreationDate":"2016-04-08T02:59:10.437","AcceptedAnswerId":36491943.0,"OwnerUserId":6151595.0,"Title":"There is a data blob of hdf5 [600000,1,7,256]. How to do mean subtraction using Caffe in matlab?","Body":"<p>There is a data blob of hdf5 .The function 'compute_image_mean.cpp' in Caffe is not fit for HDF5. So how to do mean subtraction using Caffe in matlab?<\/p>\n","answers":[{"AnswerId":"36491943","CreationDate":"2016-04-08T05:06:44.667","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Suppose you have a matrix in MATLAB of 60K vectors, how would you subtract the mean in MATLAB?<\/p>\n\n<pre><code>X=bsxfun(@minus, X, mean(X, 1));\n<\/code><\/pre>\n\n<p>Now that you subtracted the mean you can save X to hdf5 and feed it to caffe.<\/p>\n"}]}
{"QuestionId":36490849,"AnswerCount":1,"Tags":"<python><install><pip><tensorflow><prettytensor>","CreationDate":"2016-04-08T03:08:42.910","AcceptedAnswerId":36502872.0,"OwnerUserId":678572.0,"Title":"Installing prettytensor for TensorFlow not working on OSX","Body":"<p>I can't install . I've tried doing it through  and also by installing the sourcecode and doing  in the directory. <\/p>\n\n<p><strong>I updated my  . Does anyone have any idea on how I can install  on my machine?<\/strong> <\/p>\n\n<pre><\/pre>\n\n<p>when i try to import <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36502872","CreationDate":"2016-04-08T14:57:26.500","ParentId":null,"OwnerUserId":"2220465","Title":null,"Body":"<p>Please check a version of tensorflow.<\/p>\n\n<p>If your tensorflow is not 0.7.1, you can upgrade like below.<\/p>\n\n<p>(see also <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html<\/a>)<\/p>\n\n<pre><code># Mac OSX, CPU only:\n$ pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.7.1-cp27-none-any.whl\n<\/code><\/pre>\n\n<p>Then install prettytensor again.<\/p>\n\n<pre><code>$ pip uninstall prettytensor\n$ pip install prettytensor\n<\/code><\/pre>\n\n<p>For your reference.\nHere is my results.(OSX 10.11.3)<\/p>\n\n<pre><code>In [1]: import tensorflow as tf\n\nIn [2]: tf.__version__\nOut[2]: '0.7.1'\n\nIn [3]: import prettytensor as pt\n\nIn [4]: pt.PrettyTensor\nOut[4]: prettytensor.pretty_tensor_class.PrettyTensor\n<\/code><\/pre>\n"}]}
{"QuestionId":36491298,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-04-08T04:01:20.397","AcceptedAnswerId":36547541.0,"OwnerUserId":6175310.0,"Title":"How to extract pairwise diagonal in tensor in theano without for loop?","Body":"<p>I have a 4-dim tensor \"a\" with dimensions m x n x n x k.\nI want to obtain the following tensor \"b\": using python I have:<\/p>\n\n<pre><\/pre>\n\n<p>how can I convert this to theano code without using a for loop? n is around 500 so it crashes if I just do the for loop in theano.<\/p>\n","answers":[{"AnswerId":"36547541","CreationDate":"2016-04-11T11:33:18.673","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>you can try this:<\/p>\n\n<pre><code>i = theano.shared(numpy.asarray(range(n)))\nb = a[:,i,i,3:-3]\n<\/code><\/pre>\n\n<p>Theano basically inherited <a href=\"http:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#integer-array-indexing\" rel=\"nofollow\">Numpy's array indexing<\/a>.<\/p>\n"}]}
{"QuestionId":36493297,"AnswerCount":3,"Tags":"<python><rest><machine-learning><tensorflow><tensorflow-serving>","CreationDate":"2016-04-08T06:50:12.560","AcceptedAnswerId":null,"OwnerUserId":1515940.0,"Title":"Serve trained Tensorflow model with REST API using Flask?","Body":"<p>I have got a trained Tensorflow model and I want to serve the prediction method with REST API. What I can think of is to use Flask to build a simple REST API that receive JSON as input and then call the predict method in Tensorflow and then return the predicted result to the client side.<\/p>\n\n<p>I would like to know is there any concern to do it this way especially in production environment?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"36510952","CreationDate":"2016-04-08T23:35:21.697","ParentId":null,"OwnerUserId":"1238847","Title":null,"Body":"<p>The first concern which comes into my mind is the performance.<\/p>\n\n<p>TensorFlow team seems to have worked out server\/client usage.\nYou may want to look into <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow\">tensorflow serving<\/a>.\nAs a default, it uses gRPC for communication protocol.<\/p>\n"},{"AnswerId":"45495284","CreationDate":"2017-08-03T22:06:16.557","ParentId":null,"OwnerUserId":"1424575","Title":null,"Body":"<p>We use Flask + TensorFlow serving at work. Our setup might not be the most optimal way to serve models, but it gets the job done and it works fine for us so far.<\/p>\n\n<p>The setup is the following:<\/p>\n\n<ol>\n<li>Because tfserving takes forever to build, we built a docker image (not GPU support or anything, but it works for just serving a model and it's faster and better than serving it directly from within a huge Python\/Flask monolite). The model server image can be found here: <a href=\"https:\/\/hub.docker.com\/r\/epigramai\/model-server\/\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/epigramai\/model-server\/<\/a><\/li>\n<li>Then Flask is used to setup an API. In order to send requests to the model server we need a grcp prediction client, so we built one in Python that we can import directly into the flask API, <a href=\"https:\/\/github.com\/epigramai\/tfserving_predict_client\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/epigramai\/tfserving_predict_client\/<\/a>.<\/li>\n<\/ol>\n\n<p>The good thing here is that the model is not served by the Flask API application. The docker image model server can easily be replaced with a model server running on a GPU compiled for the machines hardware instead of the docker container. <\/p>\n"},{"AnswerId":"51959354","CreationDate":"2018-08-22T03:09:09.243","ParentId":null,"OwnerUserId":"7953862","Title":null,"Body":"<p>I think that one of your main concerns might be batching the requests. For example, let's say that your model is a trained CNN like VGG, Inception or similar. If you implement a regular web service with Flask, for each prediction request you receive (assuming you're running on GPU) you will do the prediction of a single image in the GPU, which can be suboptimal since you could batch similar requests, for example.<\/p>\n\n<p>That's one of the things that TensorFlow Serving aims to offer, being able to combine requests for the same model\/signature into a single batch before sending to GPU, being more efficient in the use of resources and (potentially) in throughput. You can find more information here: <a href=\"https:\/\/github.com\/tensorflow\/serving\/tree\/master\/tensorflow_serving\/batching\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/tree\/master\/tensorflow_serving\/batching<\/a><\/p>\n\n<p>That said, it depends on the scenario very much. But batching of the predictions is something important to keep in mind.<\/p>\n"}]}
{"QuestionId":36495253,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><softmax>","CreationDate":"2016-04-08T08:44:27.590","AcceptedAnswerId":null,"OwnerUserId":6176094.0,"Title":"tensorflow cifar-10 evaluation example softmax outputs","Body":"<p>I want to view the probabilities of each test image, so I modified the code(cifar10_eval.py) like this<\/p>\n\n<pre><\/pre>\n\n<p>and I run the python code like this.<\/p>\n\n<pre><\/pre>\n\n<p>The screen results are like this<\/p>\n\n<pre><\/pre>\n\n<p>The Label values seem to be good, but the logits outputs seem to be same values!\nWhy?\nAnyone can tell me the reason?<\/p>\n\n<p>This is new cifar10_eval.py source code.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36503671","CreationDate":"2016-04-08T15:34:45.103","ParentId":null,"OwnerUserId":"6176094","Title":null,"Body":"<p>I have trained with 1k steps(accuracy 10% or so).\nBut after I trained with 100k steps(accuracy 86% or so), the result is very good<\/p>\n\n<pre><code>.................\nStep: 9991 result: [[  1.30259633   0.71064955  -2.6035285   -1.30183697  -4.1291523   -3.00246906   0.30873945  -4.02916574  13.05054665  -0.42556083]] Label: 8\nStep: 9992 result: [[-1.05670786 -1.86572766  0.28350741  1.78929067  0.03841069  1.23079467    2.97172165 -1.18722486 -1.17184007 -1.02505279]] Label: 6\nStep: 9993 result: [[  1.50454926   2.34122658  -3.45632267  -0.55308843  -4.35214806    -2.28931832  -1.74908364  -4.71527719  11.44062901   1.72015083]] Label: 8\nStep: 9994 result: [[ 1.96891284 -2.57139373  0.29864013  1.30923986  1.72708285  0.95571399   -0.49331608  0.49454236 -2.26134181 -1.39561605]] Label: 0\nStep: 9995 result: [[-0.65523863  1.58577776  0.13226865  1.43122363 -2.34669352  0.18927786   -2.51019335 -1.70729315 -0.21297894  4.06098557]] Label: 9\nStep: 9996 result: [[-2.17944765 -3.22895575  2.29571438  2.63287306  0.46685112  4.42715979   -0.76104468  2.39603662 -3.21783161 -2.8433671 ]] Label: 2\nStep: 9997 result: [[  4.26957560e+00   1.95574760e-03   1.91038296e-01  -8.00723195e-01    -2.36319876e+00  -2.12906289e+00  -3.35138845e+00   7.97132492e-01   6.60009801e-01   2.73786736e+00]] Label: 0\nStep: 9998 result: [[ 0.42694128 -2.07150149  0.47749567  2.62247086  1.11608386  3.05186462   -0.19805858  0.03386561 -2.87092948 -2.59781456]] Label: 5\nStep: 9999 result: [[ 0.23629765 -3.21540785  1.01075113  0.46802399  3.44423246  0.25743011    4.71304989 -1.12128389 -3.07727337 -2.7076664 ]] Label: 6\n2016-04-09 00:32:49.861650 Total:10000 True:8631: precision @ 1 = 0.863\n<\/code><\/pre>\n"}]}
{"QuestionId":36497671,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-08T10:44:32.637","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"Variable length dimension in tensor","Body":"<p>I'm trying to implement the paper \"End-to-End memory networks\" (<a href=\"http:\/\/arxiv.org\/abs\/1503.08895\" rel=\"nofollow\">http:\/\/arxiv.org\/abs\/1503.08895<\/a>)<\/p>\n\n<p>Each training example consists of a number of phrases, a question and then the answer. The number of sentences is variable, as is the number of words in each sentence and the question. Each word is encoded as an integer. So my input would have the form [batch size, # of sentences, # words in sentence].<\/p>\n\n<p>Now my problem is that the second and third dimension are unknown for each mini-batch. Can I still somehow represent this input as a single tensor or do I have to use lists of tensors, so that I have a list of length batch_size, and then a sublist of length number of sentences and then for each sentence a tensor, whose size is also not known in advance, corresponding to the words encoded as integers.<\/p>\n\n<p>Can I use this second approach or will tensorflow then not be able to backpropagate, e.g. I have an operation where I have to calculate the following sum: \\sum_i tf.scalar_mul(p_i, c_i), where p_i is a scalar and c_i is an embedding vector that was previously calculated. The tensors for the p and c values are then stored in a list, so I would have to sum over the elements in the two lists in a loop. I'm assuming that tensorflow would not be able to incoorporate this loop in the computation graph, correct? I'm sceptical since theano has a special scan function that allows one to loop over input, so I'm assuming that a regular loop would cause problems in the computation graph. How does tensorflow handle this?<\/p>\n","answers":[{"AnswerId":"47367502","CreationDate":"2017-11-18T15:07:06.190","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Moving Yaroslav's comment to an answer:<\/p>\n\n<p>TensorFlow has <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scan\" rel=\"nofollow noreferrer\"><code>tf.scan<\/code><\/a>. Dimensions may also be dynamic as in Theano.<\/p>\n"}]}
{"QuestionId":36498127,"AnswerCount":6,"Tags":"<python><machine-learning><tensorflow><deep-learning><lstm>","CreationDate":"2016-04-08T11:09:55.200","AcceptedAnswerId":36501922.0,"OwnerUserId":2527680.0,"Title":"How to apply gradient clipping in TensorFlow?","Body":"<p>Considering the <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3%20-%20Neural%20Networks\/recurrent_network.py\">example code<\/a>.<\/p>\n\n<p>I would like to know How to apply gradient clipping on this network on the RNN where there is a possibility of exploding gradients.<\/p>\n\n<pre><\/pre>\n\n<p>This is an example that could be used but where do I introduce this ?\nIn the def of RNN <\/p>\n\n<pre><\/pre>\n\n<p>But this doesn't make sense as the tensor _X is the input and not the grad what is to be clipped? <\/p>\n\n<p>Do I have to define my own Optimizer for this or is there a simpler option?<\/p>\n","answers":[{"AnswerId":"51138713","CreationDate":"2018-07-02T14:42:17.547","ParentId":null,"OwnerUserId":"243158","Title":null,"Body":"<p>IMO the best solution is wrapping your optimizer with TF's estimator decorator <code>tf.contrib.estimator.clip_gradients_by_norm<\/code>:<\/p>\n\n<pre><code>original_optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\noptimizer = tf.contrib.estimator.clip_gradients_by_norm(original_optimizer, clip_norm=5.0)\ntrain_op = optimizer.minimize(loss)\n<\/code><\/pre>\n\n<p>This way you only have to define this once, and not run it after every gradients calculation.<\/p>\n\n<p>Documentation:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/estimator\/clip_gradients_by_norm\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/estimator\/clip_gradients_by_norm<\/a><\/p>\n"},{"AnswerId":"43959981","CreationDate":"2017-05-14T03:04:16.413","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This is actually properly <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Optimizer#processing_gradients_before_applying_them\" rel=\"nofollow noreferrer\">explained in the documentation.<\/a>:<\/p>\n\n<blockquote>\n  <p>Calling minimize() takes care of both computing the gradients and\n  applying them to the variables. If you want to process the gradients\n  before applying them you can instead use the optimizer in three steps:<\/p>\n  \n  <ul>\n  <li>Compute the gradients with compute_gradients().<\/li>\n  <li>Process the gradients as you wish.<\/li>\n  <li>Apply the processed gradients with apply_gradients().<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>And in the example they provide they use these 3 steps:<\/p>\n\n<pre><code># Create an optimizer.\nopt = GradientDescentOptimizer(learning_rate=0.1)\n\n# Compute the gradients for a list of variables.\ngrads_and_vars = opt.compute_gradients(loss, &lt;list of variables&gt;)\n\n# grads_and_vars is a list of tuples (gradient, variable).  Do whatever you\n# need to the 'gradient' part, for example cap them, etc.\ncapped_grads_and_vars = [(MyCapper(gv[0]), gv[1]) for gv in grads_and_vars]\n\n# Ask the optimizer to apply the capped gradients.\nopt.apply_gradients(capped_grads_and_vars)\n<\/code><\/pre>\n\n<p>Here <code>MyCapper<\/code> is any function that caps your gradient. The list of useful functions (other than <code>tf.clip_by_value()<\/code>) is <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/train\/gradient_clipping\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"49207469","CreationDate":"2018-03-10T09:26:45.653","ParentId":null,"OwnerUserId":"5611409","Title":null,"Body":"<p>Gradient Clipping basically helps in case of exploding or vanishing gradients.Say your loss is too high which will result in exponential gradients to flow through the network which may result in Nan values . To overcome this we clip gradients within a specific range (-1 to 1 or any range as per condition) .<\/p>\n\n<p><code>clipped_value=tf.clip_by_value(grad, -range, +range), var) for grad, var in grads_and_vars\n<\/code><\/p>\n\n<p>where grads _and_vars are the pairs of gradients (which you calculate via tf.compute_gradients) and their variables they will be applied to.<\/p>\n\n<p>After clipping we simply apply its value using an optimizer.\n<code>optimizer.apply_gradients(clipped_value)<\/code><\/p>\n"},{"AnswerId":"43486487","CreationDate":"2017-04-19T04:37:20.113","ParentId":null,"OwnerUserId":"1079110","Title":null,"Body":"<p>Despite what seems to be popular, you probably want to clip the whole gradient by its global norm:<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(1e-3)\ngradients, variables = zip(*optimizer.compute_gradients(loss))\ngradients, _ = tf.clip_by_global_norm(gradients, 5.0)\noptimize = optimizer.apply_gradients(zip(gradients, variables))\n<\/code><\/pre>\n\n<p>Clipping each gradient matrix individually changes their relative scale but is also possible:<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(1e-3)\ngradients, variables = zip(*optimizer.compute_gradients(loss))\ngradients = [\n    None if gradient is None else tf.clip_by_norm(gradient, 5.0)\n    for gradient in gradients]\noptimize = optimizer.apply_gradients(zip(gradients, variables))\n<\/code><\/pre>\n"},{"AnswerId":"47725302","CreationDate":"2017-12-09T04:24:42.673","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>For those who would like to understand the idea of gradient clipping (by norm):<\/p>\n\n<p>Whenever the gradient norm is greater than a particular threshold, we clip the gradient norm so that it stays within the threshold. This threshold is sometimes set to <code>5<\/code>.<\/p>\n\n<p>Let the gradient be <strong>g<\/strong> and the max_norm_threshold be <em>j<\/em>. <\/p>\n\n<p>Now, if ||<strong>g<\/strong>|| > <em>j<\/em> , we do:<\/p>\n\n<p><strong>g<\/strong> = (  <em>j<\/em> * <strong>g<\/strong> ) \/  ||<strong>g<\/strong>||<\/p>\n\n<p>This is the implementation done in <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.3\/api_docs\/python\/tf\/clip_by_norm\" rel=\"noreferrer\"><code>tf.clip_by_norm<\/code><\/a><\/p>\n"},{"AnswerId":"36501922","CreationDate":"2016-04-08T14:15:39.890","ParentId":null,"OwnerUserId":"118173","Title":null,"Body":"<p>Gradient clipping needs to happen after computing the gradients, but before applying them to update the model's parameters. In your example, both of those things are handled by the <code>AdamOptimizer.minimize()<\/code> method.<\/p>\n\n<p>In order to clip your gradients you'll need to explicitly compute, clip, and apply them as described in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/train\/Optimizer#processing_gradients_before_applying_them\" rel=\"noreferrer\">this section in TensorFlow's API documentation<\/a>. Specifically you'll need to substitute the call to the <code>minimize()<\/code> method with something like the following:<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\ngvs = optimizer.compute_gradients(cost)\ncapped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs]\ntrain_op = optimizer.apply_gradients(capped_gvs)\n<\/code><\/pre>\n"}]}
{"QuestionId":36498416,"AnswerCount":1,"Tags":"<theano><deep-learning><dot-product>","CreationDate":"2016-04-08T11:26:05.653","AcceptedAnswerId":36502534.0,"OwnerUserId":3654968.0,"Title":"How to multiply Tensors in Theano","Body":"<p>I have two tensors and a weight matrix in Theano. \nTensor A has dimension (k, 5, 40). Tensor B has dimension (k, 5, 40). Weight matrix W has dimension (40,40). I would like to compute\nA<em>W<\/em>B. What is the correct sequence of Theano tensor operations to achieve this? Note that k can vary during run time but other dimensions are fixed.\nThe semantics we want from A<em>W<\/em>B is the following:<\/p>\n\n<p>Think of A as a collection of k (5,40) matrices. Call them A_1, ..., A_k\nThink of B as a collection of k (5,40) matrices. Call them B_1, ..., B_k.\nWe want to find A_{i} * W * B_{i}^{T} for all i from 1 to k. I tried using theano.tensor.dot but it seems quite confusing.<\/p>\n\n<p>Note that a non-efficient way of doing this is to use the scan function. But this would make an inherently parallel code sequential.<\/p>\n","answers":[{"AnswerId":"36502534","CreationDate":"2016-04-08T14:43:01.433","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>I'm sorry, but I don't know what did you mean \"confusing\"?<\/p>\n\n<p>I've try using small case, I hope it can represent your case. Dot-product using <code>theano.tensor.dot<\/code> between three dimensional tensor and two dimensional matrix:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\na = T.tensor3('a', dtype='int64')\nc = T.matrix('c',dtype='int64')\nd = T.dot(a,c)\n\ng = theano.function([a,c],d)\n\nx = np.array([[[1,2],[1,3]],[[2,2],[1,1]]], dtype=int)\ny = np.array([[1,2],[1,3]], dtype=int)\nprint g(x,y)\n<\/code><\/pre>\n\n<p>the output:<\/p>\n\n<pre><code>[[[ 3  8]\n  [ 4 11]]\n\n [[ 4 10]\n  [ 2  5]]]\n<\/code><\/pre>\n\n<p>It works like your logic, matrix <code>c<\/code> do dot-product only in the second and third dimension.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>that first code above, you can use for the first operation in your case (A*W). Sorry I'm not calculate carefully, of course after that operation the output become three dimensional tensor. So, to perform (AW)*B you must use different approach. To perform multiplication between two three dimensional tensors I usually using scan:<\/p>\n\n<pre><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\na = T.tensor3('a', dtype='int64')\nc = T.tensor3('c',dtype='int64')\nd, b = theano.scan(lambda i: T.dot(a[i,:], c[i,:]),sequences=T.arange(2))\ng = theano.function([a,c],d)\n\nx = np.array([[[1,2],[1,3]],[[2,2],[1,1]]], dtype=int)\ny = np.array([[[1,2],[1,3]],[[2,2],[1,1]]], dtype=int)\nprint g(x,y)\n<\/code><\/pre>\n\n<p><strong>but<\/strong> I know there are another approach using <code>theano.tensor.batched_dot<\/code> (<code>theano.tensor.dot<\/code> I think only  for 2D and 1D array). In your case it's simple to code like this:<\/p>\n\n<pre><code>e = T.batched_dot(a,c)\ng = theano.function([a,c],e)\n<\/code><\/pre>\n\n<p>the code above give the same results. Hope it helps.<\/p>\n"}]}
{"QuestionId":36500777,"AnswerCount":1,"Tags":"<python><syntax><subset><theano><matrix-indexing>","CreationDate":"2016-04-08T13:24:42.237","AcceptedAnswerId":36501588.0,"OwnerUserId":4565943.0,"Title":"How to to 0\/1 subsetting in Theano?","Body":"<p>The objective is to obtain the subset of an array of elements through values provided in another array.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>This prints <\/p>\n\n<pre><\/pre>\n\n<p>instead of<\/p>\n\n<pre><\/pre>\n\n<p>How do I get the expected result?<\/p>\n","answers":[{"AnswerId":"36501588","CreationDate":"2016-04-08T14:00:42.150","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>try to use <code>nonzero()<\/code> function.<\/p>\n\n<p>Example in your case:<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\n\na = T.vector('X', dtype='int64')\nb = T.vector('Y', dtype='int64')\nc = a[b.nonzero()]\ng = function([a,b],c)\n\nx = np.array([5,3,2,3,4,6], dtype=int)\ny = np.array([0,0,1,0,0,1], dtype=int)\nprint g(x,y)\n<\/code><\/pre>\n\n<p>hope it helps<\/p>\n"}]}
{"QuestionId":36501853,"AnswerCount":1,"Tags":"<python><neural-network><theano><lasagne>","CreationDate":"2016-04-08T14:12:15.240","AcceptedAnswerId":36542841.0,"OwnerUserId":5985670.0,"Title":"How to keep the weight value to zero in a particular location using theano or lasagne?","Body":"<p>I'm a theano and lasagne user.<\/p>\n\n<p>I have a problem dealing with the variable length of the input matrix.<\/p>\n\n<p>i.e) <\/p>\n\n<pre><\/pre>\n\n<p>So, I try to use the padding.<\/p>\n\n<pre><\/pre>\n\n<p>But, after processing, theano updates the parameters matrix_padding_embedding, so, matrix_padding_embedding[-1]  no longer a 0.<\/p>\n\n<p>How to keep the weight value to zero in matrix_padding_embedding[-1]?<\/p>\n\n<p>Or, whether there are other ways of dealing with variable length?<\/p>\n","answers":[{"AnswerId":"36542841","CreationDate":"2016-04-11T07:52:28.240","ParentId":null,"OwnerUserId":"3041068","Title":null,"Body":"<p>you can create the padded matrix by concatenating two matrices, like,<\/p>\n\n<pre><code>import theano as the\nimport theano.tensor as ten\nimport numpy as np    \nmatrix_embedding = the.shared(np.asarray([[0.1, 0.2, 0.3],\n                                          [0.4, 0.5, 0.6],\n                                          [0.2, 0.3, 0.5],\n                                          [0.5, 0.6, 0.7]]))\nmatrix_padding_embedding = ten.concatenate((matrix_embedding, ten.zeros((1, 3))))\n\nx = ten.lvector()\ny = ten.sum(matrix_padding_embedding[x])\ngrad = the.grad(y, matrix_embedding)\nfn = the.function([x], [matrix_padding_embedding, grad])\n\nx2 = [1, 2, -1]\np, g = fn(x2)\nprint p\nprint g\n<\/code><\/pre>\n\n<p>results are <\/p>\n\n<pre><code># [[ 0.1  0.2  0.3]\n#  [ 0.4  0.5  0.6]\n#  [ 0.2  0.3  0.5]\n#  [ 0.5  0.6  0.7]\n#  [ 0.   0.   0. ]]\n# \n# [[ 0.  0.  0.]\n#  [ 1.  1.  1.]\n#  [ 1.  1.  1.]\n#  [ 0.  0.  0.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":36504628,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2016-04-08T16:25:11.713","AcceptedAnswerId":36518644.0,"OwnerUserId":3427580.0,"Title":"Caffe snapshots: .solverstate vs .caffemodel","Body":"<p>When training a network, the snapshots taken every N iterations come in two forms together. One is the .solverstate file, which I presume is exactly what it sounds like, storing the state of the loss functions and gradients, etc. The other is the .caffemodel file which I know stores the trained parameters.<\/p>\n\n<p>The .caffemodel is the file you need if you want a pre-trained model, so I imagine it's also the file you want if you are going to test your network.<\/p>\n\n<p>WWhat is the .solverstate good for? In <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html#resume-training\" rel=\"noreferrer\">this tutorial<\/a> it looks like you can restart training from it, but how does that differ than using the .caffemodel? Does .solverstate also include the same info as .caffemodel? Put another way, is .caffemodel just a subset of .solverstate?<\/p>\n","answers":[{"AnswerId":"36518644","CreationDate":"2016-04-09T15:11:12.367","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>The solverstate file, as its name conveys, stores the state of the solver and not any information related to classification results. The model is saved as caffemodel file, which you can use to obtain classification results for your data. If you want to fine-tune your network you may use a pre-trained caffemodel file. This will save time as your network does not need to learn from scratch. But, in case your present training needs to be halted, due to a power cut or an unexpected reboot, you may resume your training form the previous snapshot of the solverstate. The difference between using the solverstate and the caffemodel files is that the former allows you to complete your training in the pre-determined manner while the latter may require changes in certain training parameters such as the maximum number of iterations.<\/p>\n"}]}
{"QuestionId":36508322,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-08T19:58:59.007","AcceptedAnswerId":null,"OwnerUserId":563762.0,"Title":"per_image_whitening in Python","Body":"<p>I'm trying to set up TensorFlow to accept one image at a time but I believe I'm getting incorrect results because I pass a regular array without first performing tf.image.per_image_whitening() beforehand. Is there an easy way to do this in Python to an individual image without using the image queue?<\/p>\n\n<p>Here's my code so far:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41620739","CreationDate":"2017-01-12T18:48:10.137","ParentId":null,"OwnerUserId":"2799796","Title":null,"Body":"<p>If you want to avoid the image queue and do the predictions one by one, I think <\/p>\n\n<pre><code>image_array = (image_array - mean) \/ adjusted_stddev\n<\/code><\/pre>\n\n<p>should be able to do the trick. <\/p>\n\n<p>If you want to do the prediction by batches, it's a little bit complicated as  per_image_whitening (now per_image_standardization) only works with single images. So you need to do it before you form the batch like the way above or setup a preprocess procedure.<\/p>\n"}]}
{"QuestionId":36508489,"AnswerCount":1,"Tags":"<lua><parameter-passing><torch>","CreationDate":"2016-04-08T20:09:37.203","AcceptedAnswerId":null,"OwnerUserId":2838606.0,"Title":"How call a Torch .lua file given new\/different arguments?","Body":"<p>Suppose I have the followings in  file:<\/p>\n\n<pre><\/pre>\n\n<p>Now assume I want to call  given some <strong>different<\/strong> arguments through another lua file :<\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting an error which indicates that the file 'test.lua -visualize 0' cannot be found when trying to call the function through . <\/p>\n\n<p>So, how can I correctly run another lua file which contains torch code through another .lua file?<\/p>\n","answers":[{"AnswerId":"42617603","CreationDate":"2017-03-06T04:34:53.843","ParentId":null,"OwnerUserId":"4121334","Title":null,"Body":"<p>If you do not need to use the variables defined inside your 'test.lua', you can use os.execute:<\/p>\n\n<pre><code>os.execute(\"th test.lua -visiualize 0\")\n<\/code><\/pre>\n"}]}
{"QuestionId":36508860,"AnswerCount":1,"Tags":"<neural-network><tensorflow><deep-learning><lstm><recurrent-neural-network>","CreationDate":"2016-04-08T20:34:39.427","AcceptedAnswerId":null,"OwnerUserId":4185992.0,"Title":"How do I actually execute a saved TensorFlow model?","Body":"<p>Tensorflow newbie here. I'm trying to build an RNN. My input data is a set of vector instances of size  representing the (x,y) positions of a set of particles at each time step. (Since the instances already have semantic content, they do not require an embedding.) The goal is to learn to predict the positions of the particles at the next step. <\/p>\n\n<p>Following the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/recurrent\/index.html\" rel=\"nofollow\">RNN tutorial<\/a> and slightly adapting the included RNN code, I create a model more or less like this (omitting some details):<\/p>\n\n<pre><\/pre>\n\n<p>Then I create a , iterate over the data to train it using the given  method, and write out the parameters with . So far, so good.<\/p>\n\n<p>But how do I actually <em>use<\/em> the trained model? The tutorial stops at this point. From <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variables\/index.html#restoring-variables\" rel=\"nofollow\">the docs on <\/a>, in order to read back in the variables, I need to either set up exactly the same graph I was running when I saved the variables out, or selectively restore particular variables. Either way, that means my new model will require inputs of size . However, all I want now is to do a single forward pass through the model on an input of size  and read out a single -sized result (the prediction for the next time step); in other words, I want to create a model that accepts a different-size tensor than the one I trained on. I can kludge it by passing the existing model my intended data  times, but that doesn't seem like a best practice. What's the best way to do this? <\/p>\n","answers":[{"AnswerId":"36510993","CreationDate":"2016-04-08T23:40:20.673","ParentId":null,"OwnerUserId":"640898","Title":null,"Body":"<p>You have to create a new graph that has the same structure but with the <code>batch_size = 1<\/code> and import the saved variables with <code>tf.train.Saver.restore()<\/code>. You can take a look at how they define multiple models with variable batch size in ptb_word_lm.py: <a href=\"https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow\">https:\/\/tensorflow.googlesource.com\/tensorflow\/+\/master\/tensorflow\/models\/rnn\/ptb\/ptb_word_lm.py<\/a><\/p>\n\n<p>So you can have a separate file for instance, where you instantiate the graph with the batch_size that you want, then restore the saved variables. Then you can execute your graph.<\/p>\n"}]}
{"QuestionId":36509340,"AnswerCount":1,"Tags":"<machine-learning><dataset><computer-vision><caffe><vision>","CreationDate":"2016-04-08T21:05:15.553","AcceptedAnswerId":null,"OwnerUserId":5705743.0,"Title":"Create dataset for LeNet?","Body":"<p>I'm working on a project for which I'd like to create a dataset of drawn faces (similar in concept to the CUFS dataset). Hand-drawing the faces aside, how would I go from \"I have uploaded these image files to my computer and have ensured that they all have identical dimensions\" to having a ready-to-use dataset? (I'd like to train\/test LeNet with this dataset.) I've never created my own dataset before so am pretty unsure as to how to start. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"36532054","CreationDate":"2016-04-10T16:02:16.073","ParentId":null,"OwnerUserId":"5722565","Title":null,"Body":"<p>You should convert the images into levelDB or LMDB. You can follow the exsample of <code>convert_mnist_data.cpp<\/code>.<\/p>\n"}]}
{"QuestionId":36509857,"AnswerCount":3,"Tags":"<python><matlab><lua><eval><torch>","CreationDate":"2016-04-08T21:47:38.397","AcceptedAnswerId":null,"OwnerUserId":2838606.0,"Title":"How invoke a function given its name in string in Lua\/Torch?","Body":"<p>I want to evaluate a string name as a function in Lua\/Torch. Below is an example of what I want to do:<\/p>\n\n<pre><\/pre>\n\n<p>My goal is to be able to use the string in  as name of a function and evaluate that function like follow:<\/p>\n\n<pre><\/pre>\n\n<p>In MATLAB and Python there is a `eval()' function which can execute arbitrary strings. Is there such a function in Lua\/Torch?<\/p>\n\n<p>How can I do that in Torch\/Lua?<\/p>\n","answers":[{"AnswerId":"36510379","CreationDate":"2016-04-08T22:33:33.390","ParentId":null,"OwnerUserId":"3979429","Title":null,"Body":"<p>You can use <code>loadstring<\/code> but that's generally not recommended because it has to compile code at runtime. What is this for?<\/p>\n"},{"AnswerId":"37010376","CreationDate":"2016-05-03T17:15:31.567","ParentId":null,"OwnerUserId":"6262499","Title":null,"Body":"<p>You could also try <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a>, you will have a lua engine in python, thus, you can load any lua\/torch library, and you can run lua code with \"lua.eval('torch.DoubleTensor(3,4)')\". Check <a href=\"https:\/\/github.com\/imodpasteur\/lutorpy\" rel=\"nofollow\">lutorpy<\/a> for more details.<\/p>\n"},{"AnswerId":"36512076","CreationDate":"2016-04-09T02:33:45.507","ParentId":null,"OwnerUserId":"107090","Title":null,"Body":"<p>First extract the name of the field from the string:<\/p>\n\n<pre><code>k=tensorType:match(\"%.(.+)$\")\n<\/code><\/pre>\n\n<p>Then use the name to call the function:<\/p>\n\n<pre><code>torch[k](some arguments)\n<\/code><\/pre>\n"}]}
{"QuestionId":36512183,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><tensorboard>","CreationDate":"2016-04-09T02:52:29.580","AcceptedAnswerId":null,"OwnerUserId":1031775.0,"Title":"What does the \"Images\" tab show on Tensorboard?","Body":"<p>After launching the tensorboard, I see 3 rows of images under the \"IMAGES\" tab, along with a relative path url to the left. Can someone tell me what do these images and relative paths represent? What do the images in 1st and 2nd column represent?<\/p>\n\n<p>I have attached an image for reference.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HTdSJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HTdSJ.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"36512220","CreationDate":"2016-04-09T02:57:59.597","ParentId":null,"OwnerUserId":"6104317","Title":null,"Body":"<p>The relative path URL relates to TensorFlow's way of dealing with variables. <\/p>\n\n<p>You might want to browse the Documentation on \"name_scope\", \"variable_scope\" and \"Sharing Variables\". A good start would be <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/variable_scope\/index.html\" rel=\"nofollow\">this<\/a><\/p>\n\n<p>As you are doing images with Tensorflow, you might be interested in visualizing your filter kernels. In <a href=\"http:\/\/robromijnders.github.io\/tensorflow_basic\/\" rel=\"nofollow\">this<\/a> projects, I have working code to visualize the kernels of the first layer within TensorBoard<\/p>\n"},{"AnswerId":"40589301","CreationDate":"2016-11-14T12:54:54.620","ParentId":null,"OwnerUserId":"5975188","Title":null,"Body":"<p>From the tensorboard <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.10\/tensorflow\/tensorboard\/README.md\" rel=\"nofollow noreferrer\">readme<\/a> file:<\/p>\n\n<blockquote>\n  <p>The dashboard is set up so that each row corresponds to a different\n  tag, and each column corresponds to a run.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":36512311,"AnswerCount":0,"Tags":"<matcaffe>","CreationDate":"2016-04-09T03:12:59.837","AcceptedAnswerId":null,"OwnerUserId":6151595.0,"Title":"How can I train my data set in Matlab using Caffe?","Body":"<p>If I already have ,,.How can I code in matlab to call function of Caffe to train my data set.I know   there is a demo code but it just for classsification not train.Who have a complete code about how to train one's own data set?<\/p>\n","answers":[]}
{"QuestionId":36515648,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-04-09T10:37:32.900","AcceptedAnswerId":null,"OwnerUserId":3808401.0,"Title":"TensorFlow bidirectional LSTM encoding of word embeddings","Body":"<p>I have a word embedding matrix containing a vector for each word. I am trying to use TensorFlow to get the bidirectional LSTM encoding of each word given the embedding vectors. Unfortunately, I get the following error message:<\/p>\n\n<blockquote>\n  <p>ValueError: Shapes (1, 125) and () must have the same rank\n  Exception TypeError: TypeError(\"'NoneType' object is not callable\",) in  ignored<\/p>\n<\/blockquote>\n\n<p>Here is the code I used:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36525018","CreationDate":"2016-04-10T02:18:13.807","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>the sequence length passed to rnn must be a vector of length batch size.<\/p>\n"}]}
{"QuestionId":36516675,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2016-04-09T12:19:29.633","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"the confuse of 'nn' , 'dp' and 'dpnn' module in Torch","Body":"<p>When i read some torch codes, i was confused with the \"nn\" and \"dp\" model. I see in many codes which requires only  dp (no nn), but why it can use nn.methodA? why it can use nn without require it ? besides, the method is defined in the dpnn!  Should it be used as dpnn.methodA?  the codes do like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36517056","CreationDate":"2016-04-09T12:56:00.583","ParentId":null,"OwnerUserId":"359582","Title":null,"Body":"<p>'dp' has 'nn' as a dependency, and requires 'nn'. More specifically when calling <code>require 'dp'<\/code> the init script of 'dp' is run. This calls <code>require 'dpnn'<\/code> (<a href=\"https:\/\/github.com\/nicholas-leonard\/dp\/blob\/master\/init.lua\" rel=\"nofollow\">see<\/a>). Subsequently the init script of 'dpnn' is run and calls <code>require 'nn'<\/code> (<a href=\"https:\/\/github.com\/Element-Research\/dpnn\/blob\/master\/init.lua\" rel=\"nofollow\">see<\/a>).<\/p>\n\n<p>Additionally, modules share the same global namespace. This means modules can set abc or nn or whatever. Many modules extend the namespace of 'nn' to include extra functionality.<\/p>\n"}]}
{"QuestionId":36519138,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-04-09T15:51:55.620","AcceptedAnswerId":36547129.0,"OwnerUserId":2290331.0,"Title":"tensorflow\/tflearn input shape","Body":"<p>I'm trying to create a lstm-rnn to generate sequences of music. The training data is a sequence of vectors of size 4, representing various features (including MIDI note) of each note in some songs to train on.<\/p>\n\n<p>From my reading, it looks like what I'm trying to do is have for each input sample, the output sample is the next size 4 vector (i.e. it should be trying to predict the next note given the current one, and because of the LSTMs incorporating knowledge of samples that have come before).<\/p>\n\n<p>I'm using tflearn as I'm still very new to RNNs. I have the following code<\/p>\n\n<pre><\/pre>\n\n<p>Before this code I have split the trainX and trainY into sequences of length 20 (arbitrarily, but I read somewhere that training on sequences like this is a good way to do this).<\/p>\n\n<p>This seems to be fine but I get the error ValueError: Cannot feed         value of shape (128, 16, 4) for Tensor u'TargetsData\/Y:0', which has shape     '(?, 4)'<\/p>\n\n<p>SO: my assumptions so far is that the input shape [None, seqLength, 4] is saying to TF [batchLength (which gets fed by tflearn sequentially), sequence length, feature length of sample]. What I don't understand is why it's saying the output is the wrong shape? Am I assuming wrongly with the data sequence split? When I just try to feed in all my data without splitting into sequences, so the input shape is [None, 4], TF tells me the LSTM layer expects an input shape with at least 3 dimensions.<\/p>\n\n<p>I can't get my head round what the shapes of the inputs and outputs should be. It feels like this should be a simple thing -- I have a set of input sequences of vectors and I want the network to try and predict the next one in the sequence. There's very little online that doesn't assume a fairly advanced level of knowledge, so I've hit a brick wall. Really appreciate any insight anyone can give!<\/p>\n","answers":[{"AnswerId":"36547129","CreationDate":"2016-04-11T11:14:29.957","ParentId":null,"OwnerUserId":"2290331","Title":null,"Body":"<p>I solved this, so am writing the answer here for anyone having the same problem. It was based on a mis-understanding of how these networks work, but this is assumed knowledge in most tutorials I've read so may not be clear to other beginners.<\/p>\n\n<p>The LSTM networks are useful for these situations because they can take into account input history. The way the history is given to the LSTM is through the sequencing, but each sequence still leads to a single output data point. So the input must be of 3D shape, while the output is just 2D. <\/p>\n\n<p>Given an entire sequence and a desired historyLength, I split the input into sequences of historyLength and a single output vector. This solved my shape problem. <\/p>\n"}]}
{"QuestionId":36519724,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-09T16:38:30.920","AcceptedAnswerId":36520509.0,"OwnerUserId":3559686.0,"Title":"How to use a linear activation function in TensorFlow?","Body":"<p>In CUDA ConvNet, we can specify the neuron activation function to be linear by writing , such that .<\/p>\n\n<p>How can I achieve the same result in TensorFlow?<\/p>\n","answers":[{"AnswerId":"36520509","CreationDate":"2016-04-09T17:49:29.253","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The most basic way to write a linear activation in TensorFlow is using <code>tf.matmul()<\/code> and <code>tf.add()<\/code> (or the <code>+<\/code> operator). Assuming you have a matrix of outputs from the previous layer (let's call it <code>prev_layer<\/code>) with size <code>batch_size<\/code> x <code>prev_units<\/code>, and the size of the linear layer is <code>linear_units<\/code>:<\/p>\n\n<pre><code>prev_layer = \u2026\n\nlinear_W = tf.Variable(tf.truncated_normal([prev_units, linear_units], \u2026))\nlinear_b = tf.Variable(tf.zeros([linear_units]))\n\nlinear_layer = tf.matmul(prev_layer, linear_W) + linear_b\n<\/code><\/pre>\n"}]}
{"QuestionId":36520050,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-04-09T17:04:47.870","AcceptedAnswerId":36521460.0,"OwnerUserId":3834415.0,"Title":"How to install tensor flow on Linux","Body":"<p>On the main TensorFlow website, you must specifically enable the gpu: <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>On the anaconda website, you only have to install tensor flow, but there is no indication whether this is the GPU or CPU version: <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>After running the Anaconda command, I decided that I needed to make sure that it was the GPU version, so I tried to run the pip command inside the anaconda folder. <\/p>\n\n<p>When I ran the pip command, I got the following error: <\/p>\n\n<blockquote>\n  <p>tensorflow-0.7.1-cp34-none-linux_x86_64.whl is not a supported wheel on this platform<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>I have no idea what a wheel is...but I have a newish laptop with a multicore i7 (July last year) and a laptop nvidia card with a compute rating of 5.2 and CUDA installed (Ubuntu 15.10).<\/p>\n\n<p>There must be a version that is compatible with my specs.  <\/p>\n\n<p>How do I upgrade my tensor flow install to the GPU version?<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>When I run: , the supported wheels include <\/p>\n\n<blockquote>\n  <p>cp35<\/p>\n<\/blockquote>\n\n<p>But the wheel that is available for TensorFlow is cp34...so does that mean TensorFlow on the GPU does not work for my computer?<\/p>\n","answers":[{"AnswerId":"43797547","CreationDate":"2017-05-05T05:59:44.373","ParentId":null,"OwnerUserId":"3460262","Title":null,"Body":"<p>Everything changed in Tensorflow V1.0 and This guide explains how to install TensorFlow on Ubuntu of Course with GPU supported.I have only tested these instructions on Ubuntu 14.04.<\/p>\n\n<ul>\n<li><h2>nVidia Check<\/h2><\/li>\n<\/ul>\n\n<p>first of all you must check that nVidia installed for check this enter the command below:<\/p>\n\n<pre><code>nvidia-smi\n<\/code><\/pre>\n\n<p>and you should see somthing like this<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fOI2P.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fOI2P.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>if you see this your nvidia card is correctly installed.<\/p>\n\n<ul>\n<li><h2>Install CUDA,Cudnn<\/h2><\/li>\n<\/ul>\n\n<p>go to <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-downloads<\/a> and download cuda toolkit for ubutu 14.04 (download .deb file) and install it with these commands<\/p>\n\n<pre><code>sudo dpkg -i cuda-repo-ubuntu1404-8-0-local-ga2_8.0.61-1_amd64.deb\nsudo apt-get update\nsudo apt-get install cuda\n<\/code><\/pre>\n\n<p><strong>Be Careful<\/strong> tensorflow V1.0 is needed cuda-8.0 and is not worked with cuda-7.5 and if you have installed cuda-7.5 and lower , you got this error when import tensorflow in python<\/p>\n\n<p>ImportError: libcudart.so.7.5: cannot open shared object file: No such file or directory.<\/p>\n\n<p>after install cuda-8.0 go to <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cudnn<\/a> and download cuDNN v5.1 (Jan 20, 2017), for CUDA 8.0 and install with commands below<\/p>\n\n<pre><code>tar -xzvf cudnn-8.0-linux-x64-v5.1.tgz\nsudo cp -P include\/cudnn.h \/usr\/local\/include\nsudo cp -P lib64\/libcudnn* \/usr\/local\/lib64\nsudo chmod a+r \/usr\/local\/lib64\/libcudnn*\n<\/code><\/pre>\n\n<p>at the end add these lines to your path file maybe add these lines to ~\/.bashrc file for doing this use these commands below:<\/p>\n\n<p>open ~\/.bashrc<\/p>\n\n<pre><code>sudo nano ~\/.bashrc\n<\/code><\/pre>\n\n<p>add these lines<\/p>\n\n<pre><code>export CUDA_HOME=\/usr\/local\/cuda-8.0\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\n<\/code><\/pre>\n\n<ul>\n<li><h2>Tensorflow<\/h2><\/li>\n<\/ul>\n\n<p>for installing tensorflow with gpu supported you can install it with one command below<\/p>\n\n<pre><code>pip install tensorflow-gpu\n<\/code><\/pre>\n\n<blockquote>\n  <p>I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally\n  I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally\n  I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally\n  I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally\n  I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally<\/p>\n<\/blockquote>\n\n<p>after all if do everythings right when you import tensorflow in python console you can see these and this means your tensorflow with gpu supported <\/p>\n"},{"AnswerId":"36521460","CreationDate":"2016-04-09T19:12:07.870","ParentId":null,"OwnerUserId":"233293","Title":null,"Body":"<p>As dumb as it seems, just renaming the wheel from 3.4 to 3.5 is sufficient:<\/p>\n\n<pre><code>$ wget https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.7.1-cp34-none-linux_x86_64.whl\n...\n$ pip install tensorflow-0.7.1-cp34-none-linux_x86_64.whl\ntensorflow-0.7.1-cp34-none-linux_x86_64.whl is not a supported wheel on this platform.\n$ mv tensorflow-0.7.1-cp3{4,5}-none-linux_x86_64.whl\n$ pip install tensorflow-0.7.1-cp35-none-linux_x86_64.whl\nProcessing .\/tensorflow-0.7.1-cp35-none-linux_x86_64.whl\n...\nInstalling collected packages: tensorflow\nSuccessfully installed tensorflow-0.7.1\n<\/code><\/pre>\n"}]}
{"QuestionId":36521908,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-04-09T19:52:32.187","AcceptedAnswerId":null,"OwnerUserId":1483516.0,"Title":"How to Iteratively Create Tensorflow Graphs On The Fly Without Accumulating Memory?","Body":"<p>What is the idiomatic way to construct a new neural network during each iteration of the training loop? It is an unusual thing to do, but I am working on an unusual project.<\/p>\n\n<p>The tensorflow API documentation says \"A session may own resources, such as variables, queues, and readers. It is important to release these resources when they are no longer required. To do this, either invoke the close() method on the session, or use the session as a context manager.\", so my first attempt involved declaring all my tensorflow variables within a session context manager, so that they can be freed at the next iteration of the training loop:<\/p>\n\n<pre><\/pre>\n\n<p>The first iteration of the training loop executes correctly, but on the second iteration the call to  leads to a memory error \"ValueError: GraphDef cannot be larger than 2GB.\" Does exiting the session context not in fact free memory in the graph? Each iteration of the loop involves the same amount of data and the same number of parameters in the neural network, so I don't think it's my construction that leads to the memory error.<\/p>\n\n<p>The documentation is not very detailed on what exactly  does, but I've tried freeing memory by calling  at the end of the session, leading to the error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I iteratively create tensorflow graphs without accumulating memory?<\/p>\n","answers":[{"AnswerId":"47903235","CreationDate":"2017-12-20T09:56:45.020","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>In your code, as it was already said in the comments, you are adding to the same graph in each iteration. You should use a clean graph every time. <code>tf.reset_default_graph()<\/code> is one option, but maybe a more clean way is to make it explicit, like this:<\/p>\n\n<pre><code>while step * train_batchsize &lt; training_iters:\n  with tf.Graph().as_default() as graph:\n    with tf.device('\/gpu:1'):\n      with tf.Session(graph=graph, config=tf.ConfigProto(allow_soft_placement=True,log_device_placement=False)) as sess:\n        # ...\n<\/code><\/pre>\n\n<p>If you then end up with GPU memory problems, this is covered by <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1727\" rel=\"nofollow noreferrer\">this issue<\/a> (as indicated by the comments), which says that it is actually a bug in the Nvidia drivers and an update to the Nvidia drivers should fix it.<\/p>\n"}]}
{"QuestionId":36523768,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-04-09T23:03:04.430","AcceptedAnswerId":null,"OwnerUserId":6182502.0,"Title":"Installing Theano on windows machine","Body":"<p>I have read through <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/33687103\/how-to-install-theano-on-anaconda-python-2-7-x64-on-windows\/36523637#36523637\">How to install Theano on Anaconda Python 2.7 x64 on Windows?<\/a><\/p>\n\n<p>I installed Anaconda 2 and followed the steps and I can import theano, however once in a while I get a windows crash message window on python crash similar to this:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/31050976\/python-exe-crashes-when-importing-theano\">python.exe crashes when importing `theano`<\/a><\/p>\n\n<p>I have done \"conda install mingw libpython\" step.<\/p>\n\n<p>Also when I run theano.test() I get this warning:<\/p>\n\n<p>pycuda import failed in theano.misc.pycuda_init. <\/p>\n\n<p>Overall I think there were some more steps in the windows installation page regarding visual studio, and other modules. I have visual studio 2012 installed by microsoft.<\/p>\n","answers":[{"AnswerId":"39461258","CreationDate":"2016-09-13T02:12:01.807","ParentId":null,"OwnerUserId":"2280645","Title":null,"Body":"<p>Did you follow the theano install instructions fully?<\/p>\n\n<p>On windows 7 x64, I would advise to either use Canopy from Enthought, which, if you are a student , you can get the complete package for free.<\/p>\n\n<p>Or, i also used WinPython before, it worked fine.\nAlthough you have to follow all steps, including :<\/p>\n\n<p><strong>Installing Windows Software Development Kit version 7.1 or you can install the newest Visual Studio which comes with it and it comes with an option to install Microsoft Visual C++ Compiler for Python 2.7<\/strong><\/p>\n\n<p><strong>Installing TDM GCC<\/strong> <\/p>\n\n<p><strong><em>But be sure to check<\/em><\/strong> <\/p>\n\n<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow\">http:\/\/deeplearning.net\/software\/theano\/install_windows.html<\/a><\/p>\n\n<p><em>You might wanna check if you missed any steps.<\/em><\/p>\n\n<p>I had problems with Anaconda and Theano before. So i gave up on using it.\nHowever a friend had success with it and Theano as well.<\/p>\n\n<p>Just be sure you have your environment and python setup installed properly prior to theano install.<\/p>\n\n<p>Doing <\/p>\n\n<pre><code>where gcc\nwhere gendef\nwhere cl\nwhere nvcc\n<\/code><\/pre>\n\n<p><strong><em>on the Command Prompt Might reveal something...<\/em><\/strong><\/p>\n\n<p>Be sure to have installed Windows SDK and last visual studio.<\/p>\n"}]}
{"QuestionId":36524257,"AnswerCount":1,"Tags":"<python><neural-network><conv-neural-network><keras><recurrent-neural-network>","CreationDate":"2016-04-10T00:11:02.763","AcceptedAnswerId":36555574.0,"OwnerUserId":4013185.0,"Title":"Unable to add Embedding layer to a siamese network using keras","Body":"<p>I am trying to alter the example code for a siamese network and add an embedding layer like so:<\/p>\n\n<pre><\/pre>\n\n<p>Which follows the <a href=\"http:\/\/keras.io\/examples\/#single-shared-lstm-over-two-parallel-sequences-for-classification\" rel=\"nofollow\">last example<\/a> on their documentation very closely.<\/p>\n\n<p>Unfortunately I keep getting an error:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone help?<\/p>\n","answers":[{"AnswerId":"36555574","CreationDate":"2016-04-11T17:41:45.477","ParentId":null,"OwnerUserId":"2259979","Title":null,"Body":"<pre><code>model.add_input(name='input_a', input_shape=(timesteps,), dtype='int')\nmodel.add_input(name='input_b', input_shape=(timesteps,), dtype='int')\n<\/code><\/pre>\n"}]}
{"QuestionId":36525379,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-04-10T03:20:48.223","AcceptedAnswerId":null,"OwnerUserId":6177624.0,"Title":"How to handle the type of the theano","Body":"<p>I have a matrix 'y' and I want to add one of number in every row of y.Then I set a matrix 'x' which the first column set the one I want to add. I use the following code:<\/p>\n\n<pre><\/pre>\n\n<p>But this cann't work. There will be a grammar error. why dosen't this code work? <\/p>\n","answers":[{"AnswerId":"36553913","CreationDate":"2016-04-11T16:10:45.753","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>there are some problems in your code. first, define your data type carefully and then becareful when using <code>[]<\/code> bracket (see how you just make another array when you call your function by <code>f([x,y])<\/code>)<\/p>\n\n<p>so, after I modify your code become:<\/p>\n\n<pre><code>import numpy\nimport theano\nimport theano.tensor as T\n\nx = T.imatrix('x') #define integer matrix\ny = T.imatrix('y')\nz = T.sum(y[T.arange(5),x[:,0]])\n\nf = theano.function(inputs=[x,y],outputs=z)\n\nx = numpy.asarray([[1],[2],[3],[4],[0]])\ny = numpy.ones((5,5),dtype=\"int32\") #define integer array\nprint(f(x,y)) #remove square bracket\n<\/code><\/pre>\n\n<p>Hope it helps<\/p>\n"}]}
{"QuestionId":36525613,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2016-04-10T03:59:19.160","AcceptedAnswerId":36526739.0,"OwnerUserId":5895928.0,"Title":"Theano Matrix Inverse","Body":"<pre><\/pre>\n\n<p>The print(invA) statement prints out \"MatrixInverse.0\". The matrix does have an inverse. Could someone please help me print out the matrix inverse?\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"36526739","CreationDate":"2016-04-10T06:42:52.703","ParentId":null,"OwnerUserId":"5486261","Title":null,"Body":"<p>You are missing a Theano function to evaluate the symbolic matrix inverse<\/p>\n\n<pre><code>A = T.dmatrix('A')\ninvA = matrix_inverse(A)\nf = theano.function([A], invA)   \nprint(f(np.matrix('1 0 0; 1 2 0; 0 0 3', dtype=config.floatX)))\n#[[ 1.          0.          0.        ]\n#[-0.5         0.5         0.        ]\n#[ 0.          0.          0.33333333]]\n<\/code><\/pre>\n"}]}
{"QuestionId":36526959,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2016-04-10T07:11:14.310","AcceptedAnswerId":36532014.0,"OwnerUserId":1714410.0,"Title":"Caffe: What can I do if only a small batch fits into memory?","Body":"<p>I am trying to train a very large model. Therefore, I can only fit a very small batch size into GPU memory. Working with small batch sizes results with very <a href=\"https:\/\/stackoverflow.com\/a\/33717093\/1714410\">noisy gradient estimations<\/a>.<br>\nWhat can I do to avoid this problem?<\/p>\n","answers":[{"AnswerId":"36527804","CreationDate":"2016-04-10T08:58:47.023","ParentId":null,"OwnerUserId":"6113498","Title":null,"Body":"<p>As stated in <a href=\"https:\/\/stats.stackexchange.com\/questions\/68323\/choosing-an-appropriate-minibatch-size-for-stochastic-gradient-descent-sgd?rq=1\">this post<\/a>, the batch size is not a problem in theory (the efficiency of stochastic gradient descent has been proven with a batch of size 1). Make sure you implement your batch correctly (the samples should be randomly picked over your data).<\/p>\n"},{"AnswerId":"36532014","CreationDate":"2016-04-10T15:58:26.963","ParentId":null,"OwnerUserId":"5722565","Title":null,"Body":"<p>You can change the <code>iter_size<\/code> in the solver parameters.\nCaffe accumulates gradients over <code>iter_size<\/code> x <code>batch_size<\/code> instances in each stochastic gradient descent step.\nSo increasing <code>iter_size<\/code> can also get more stable gradient when you cannot use large batch_size due to the limited memory.<\/p>\n"}]}
{"QuestionId":36530699,"AnswerCount":1,"Tags":"<bash><lua><torch>","CreationDate":"2016-04-10T13:58:55.907","AcceptedAnswerId":36531823.0,"OwnerUserId":4611375.0,"Title":"Error importing packages into Torch","Body":"<p>I just installed Torch7 <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow\">as per the instructions<\/a>. However, I am unable to port any packages via the command-line interpreter:<\/p>\n\n<pre><\/pre>\n\n<p>Outside of the interpreter, (just from the command line)  I ran the command , but the same error described above persisted.<\/p>\n\n<p>Would love to get Torch7 up and running!<\/p>\n\n<hr>\n\n<p>Hi, @deltheil ...<\/p>\n\n<pre><\/pre>\n\n<p>Running the following after ...<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36531823","CreationDate":"2016-04-10T15:41:47.813","ParentId":null,"OwnerUserId":"4611375","Title":null,"Body":"<p>Thanks be to @deltheil for the assistance!<\/p>\n\n<p><a href=\"https:\/\/github.com\/torch\/trepl\/issues\/37\">As per the instructions on this page<\/a>, what worked for me (on Mac OSX) was:<\/p>\n\n<ol>\n<li>Locating and removing the file <code>.\/pkg\/torch\/lib\/TH\/cmake\/FindMKL.cmake<\/code> ...<\/li>\n<li>Doing a clean install by running <code>.\/clean.sh; .\/install.sh<\/code> ...<\/li>\n<\/ol>\n\n<p>After that Torch is running as intended :-). I am machine learning as we speak!<\/p>\n"}]}
{"QuestionId":36530944,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-04-10T14:22:35.997","AcceptedAnswerId":36544881.0,"OwnerUserId":6176753.0,"Title":"How to get the count of an element in a tensor","Body":"<p>I want to get the count of an element in a tensor, for example, t=[1,2,0,0,0,0] (t is a tensor), I can get the amount, 4, of '0' by calling t.count(0) in python, but in tensorflow, I can't find any functions to do this. How can i get the count of '0'? Could please someone help me? <\/p>\n","answers":[{"AnswerId":"36544881","CreationDate":"2016-04-11T09:32:29.577","ParentId":null,"OwnerUserId":"2245776","Title":null,"Body":"<p>There isn't a built in count method in tensor flow right now. But you could do it using the existing tools in a method like so: <\/p>\n\n<pre><code>def tf_count(t, val):\n    elements_equal_to_value = tf.equal(t, val)\n    as_ints = tf.cast(elements_equal_to_value, tf.int32)\n    count = tf.reduce_sum(as_ints)\n    return count\n<\/code><\/pre>\n"},{"AnswerId":"41498251","CreationDate":"2017-01-06T03:01:37.907","ParentId":null,"OwnerUserId":"3597868","Title":null,"Body":"<p>An addition to Slater's answer above. If you want to get the count of <em>all<\/em> the elements, you can use <code>one_hot<\/code> and <code>reduce_sum<\/code> to avoid any looping within python. For example, the code-snippet below returns a vocab, ordered by occurrences within a word_tensor.<\/p>\n\n<pre><code>def build_vocab(word_tensor, vocab_size): \n  unique, idx = tf.unique(word_tensor)\n  counts_one_hot = tf.one_hot(\n      idx, \n      tf.shape(unique)[0],\n      dtype=tf.int32\n  )\n  counts = tf.reduce_sum(counts_one_hot, 0)\n  _, indices = tf.nn.top_k(counts, k=vocab_size)\n  return tf.gather(unique, indices)\n<\/code><\/pre>\n\n<p>EDIT: After a little experimentation, I discovered it's pretty easy for the <code>one_hot<\/code> tensor to blow up beyond TF's maximum tensor size. It's likely more efficient (if a little less elegant) to replace the <code>counts<\/code> call with something like this:<\/p>\n\n<pre><code>counts = tf.foldl(\n  lambda counts, item: counts + tf.one_hot(\n      item, tf.shape(unique)[0], dtype=tf.int32),\n  idx,\n  initializer=tf.zeros_like(unique, dtype=tf.int32),\n  back_prop=False\n)\n<\/code><\/pre>\n"},{"AnswerId":"45207218","CreationDate":"2017-07-20T06:48:36.183","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>To count just a specific element you can create a boolean mask, convert it to <code>int<\/code> and sum it up:<\/p>\n\n<pre><code>import tensorflow as tf\n\nX = tf.constant([6, 3, 3, 3, 0, 1, 3, 6, 7])\nres = tf.reduce_sum(tf.cast(tf.equal(X, 3), tf.int32))\nwith tf.Session() as sess:\n    print sess.run(res)\n<\/code><\/pre>\n\n<p>Also you can count every element in the list\/tensor using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/unique_with_counts\" rel=\"noreferrer\">tf.unique_with_counts<\/a>;<\/p>\n\n<pre><code>import tensorflow as tf\n\nX = tf.constant([6, 3, 3, 3, 0, 1, 3, 6, 7])\ny, idx, cnts = tf.unique_with_counts(X)\nwith tf.Session() as sess:\n    a, _, b = sess.run([y, idx, cnts])\n    print a\n    print b\n<\/code><\/pre>\n"}]}
{"QuestionId":36532125,"AnswerCount":1,"Tags":"<android><python><c++><tensorflow><tensorflow-serving>","CreationDate":"2016-04-10T16:08:26.887","AcceptedAnswerId":null,"OwnerUserId":5452997.0,"Title":"Use Tensorflow model in c++ app","Body":"<p>Apologies in advance for this relatively newbie question. Using Tensorflow, I've trained a neural net in python and I'd like to use it to classify images in a c++ application, which I want to later integrate into an android app.<\/p>\n\n<p>Today I spent all day working through the Tensorflow Serving tutorial: <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow\">https:\/\/tensorflow.github.io\/serving\/serving_basic<\/a>.<\/p>\n\n<p>I have everything installed and working, but it's still not obvious to me what to do next. Is TF Serving the right thing to do? Can it be used to integrate a trained TF model into a an application? Or is it just something to build models that can be run from the Terminal?<\/p>\n\n<p>If TF Serving is not what I need, then where should I look in stead? Is there a simpler way to use trained models cross-platform?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"37041984","CreationDate":"2016-05-05T03:47:24.647","ParentId":null,"OwnerUserId":"6291805","Title":null,"Body":"<p>Please take a look at this <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic#load-exported-tensorflow-model\" rel=\"nofollow\">section<\/a> in the tutorial for how to load a <code>SessionBundle<\/code> in c++.<\/p>\n\n<p>Once you have a SessionBundle you can use utilities in <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/session_bundle\/signature.h\" rel=\"nofollow\">signature.h<\/a> to get a signature (e.g., <code>GetClassificationSignature<\/code>) then run an inference (e.g., <code>RunClassification<\/code>).<\/p>\n"}]}
{"QuestionId":36533622,"AnswerCount":1,"Tags":"<tensorflow><skflow>","CreationDate":"2016-04-10T18:16:11.867","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"Tensorflow skflow, Data seems compatible, Valuerror, Shape error","Body":"<p>Trying to run a very simple nn classifier with skflow.<\/p>\n\n<pre><\/pre>\n\n<p>output is :\n(73257, 3072)\n(73257, 10)<\/p>\n\n<p>and the error is :<\/p>\n\n<pre><\/pre>\n\n<p>I do not really understand what the problem is here :( <\/p>\n","answers":[{"AnswerId":"36543201","CreationDate":"2016-04-11T08:12:39.747","ParentId":null,"OwnerUserId":"2220465","Title":null,"Body":"<p>Perhaps the labels of your dataset are one-hot vectors. (In this case, I use mnist dataset. See also <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/mnist\/beginners\/index.html<\/a>)<\/p>\n\n<pre><code>In [1]: from tensorflow.examples.tutorials.mnist import input_data\n\nIn [2]: mnist = input_data.read_data_sets(\"MNIST_DATA\/\", one_hot=True)\nExtracting MNIST_DATA\/train-images-idx3-ubyte.gz\nExtracting MNIST_DATA\/train-labels-idx1-ubyte.gz\nExtracting MNIST_DATA\/t10k-images-idx3-ubyte.gz\nExtracting MNIST_DATA\/t10k-labels-idx1-ubyte.gz\n\nIn [3]: mnist.train.labels\nOut[3]: \narray([[ 0.,  0.,  0., ...,  1.,  0.,  0.],\n       [ 0.,  0.,  0., ...,  0.,  0.,  0.],\n       [ 0.,  0.,  0., ...,  0.,  0.,  0.],\n       ..., \n       [ 0.,  0.,  0., ...,  0.,  0.,  0.],\n       [ 0.,  0.,  0., ...,  0.,  0.,  0.],\n       [ 0.,  0.,  0., ...,  0.,  1.,  0.]])\n\nIn [4]: mnist.train.labels.shape\nOut[4]: (55000, 10)\n\nIn [5]: import skflow\n\nIn [6]: classifier = skflow.TensorFlowDNNClassifier(hidden_units=[10, 10, 10], n_classes=10, batch_size=100, learning_rate=0.05)\n\nIn [7]: classifier.fit(mnist.train.images, mnist.train.labels)\n<\/code><\/pre>\n\n<p>Then I got same error.<\/p>\n\n<pre><code>ValueError: Shapes (?, 10) and (?, 10, 10) must have the same rank\n<\/code><\/pre>\n\n<p>But skflow assumes that lables are numbers between 0 and 9. (one_hot=False)<\/p>\n\n<pre><code>In [5]: mnist = input_data.read_data_sets(\"MNIST_DATA\/\", one_hot=False)\nExtracting MNIST_DATA\/train-images-idx3-ubyte.gz\nExtracting MNIST_DATA\/train-labels-idx1-ubyte.gz\nExtracting MNIST_DATA\/t10k-images-idx3-ubyte.gz\nExtracting MNIST_DATA\/t10k-labels-idx1-ubyte.gz\n\nIn [6]: mnist.train.labels\nOut[6]: array([7, 3, 4, ..., 5, 6, 8], dtype=uint8)\n\nIn [7]: classifier.fit(mnist.train.images, mnist.train.labels)\nStep #99, avg. train loss: 2.31658\nStep #199, avg. train loss: 1.63361\nOut[7]: \nTensorFlowDNNClassifier(batch_size=100, class_weight=None,\n            config_addon=&lt;skflow.addons.config_addon.ConfigAddon object at 0x11cf7eb90&gt;,\n            continue_training=False, hidden_units=[10, 10, 10],\n            keep_checkpoint_every_n_hours=10000, learning_rate=0.05,\n            max_to_keep=5, n_classes=10, optimizer='SGD', steps=200,\n            tf_master='', tf_random_seed=42, verbose=1)\n<\/code><\/pre>\n\n<p>Please give it a try.<\/p>\n"}]}
{"QuestionId":36533709,"AnswerCount":0,"Tags":"<python><deep-learning><caffe><multilabel-classification>","CreationDate":"2016-04-10T18:23:06.033","AcceptedAnswerId":null,"OwnerUserId":6185028.0,"Title":"LMDB data for multi-label classification and layers definition in caffe","Body":"<p>I am a beginner in caffe framework and I would like to implement the training of a CNN model with multi-label data.\nI know how to create a lmdb data in python like this <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow\">example<\/a> but I couldn't find any complete tutorial of how to define the labels when an image has 2 or more labels.<\/p>\n\n<p>Do you know if there is a complete and clear example of how to create such data and how to define the accuracy and loss layers in the prototxt files.<\/p>\n","answers":[]}
{"QuestionId":36533723,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-04-10T18:24:01.887","AcceptedAnswerId":36536063.0,"OwnerUserId":1595417.0,"Title":"Tensorflow get all variables in scope","Body":"<p>I have some variables created within a certain scope like this:<\/p>\n\n<pre><\/pre>\n\n<p>I then want to get the list of all the variables in \"my_scope\" so I can pass it to an optimizer. What is the right way to do this?<\/p>\n","answers":[{"AnswerId":"43861652","CreationDate":"2017-05-09T05:14:28.753","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>User correctly pointed out that you need <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_collection\" rel=\"noreferrer\"><code>tf.get_collection()<\/code><\/a>. I will just give a simple example how to do this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nwith tf.name_scope('some_scope1'):\n    a = tf.Variable(1, 'a')\n    b = tf.Variable(2, 'b')\n    c = tf.Variable(3, 'c')\n\nwith tf.name_scope('some_scope2'):\n    d = tf.Variable(4, 'd')\n    e = tf.Variable(5, 'e')\n    f = tf.Variable(6, 'f')\n\nh = tf.Variable(8, 'h')\n\nfor i in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='some_scope'):\n    print i   # i.name if you want just a name\n<\/code><\/pre>\n\n<p>Notice that you can provide any of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GraphKeys\" rel=\"noreferrer\">graphKeys<\/a> and scope is a regular expression:<\/p>\n\n<blockquote>\n  <p>scope: (Optional.) If supplied, the resulting list is filtered to\n  include only items whose name attribute matches using re.match. Items\n  without a name attribute are never returned if a scope is supplied and\n  the choice or re.match means that a scope without special tokens\n  filters by prefix.<\/p>\n<\/blockquote>\n\n<p>So if you will pass 'some_scope' you will get 6 variables.<\/p>\n"},{"AnswerId":"36536063","CreationDate":"2016-04-10T22:03:38.030","ParentId":null,"OwnerUserId":"728291","Title":null,"Body":"<p>I think you want <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_collection\" rel=\"noreferrer\">tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='my_scope')<\/a>.  This will get all variables in a scope.<\/p>\n\n<p>To pass to an optimizer you do not want <em>all<\/em> variables you would just want the trainable variables.  Those are also kept in a default collection, which is <code>tf.GraphKeys.TRAINABLE_VARIABLES<\/code>.<\/p>\n"}]}
{"QuestionId":36534154,"AnswerCount":1,"Tags":"<matrix><lua><torch>","CreationDate":"2016-04-10T18:58:32.130","AcceptedAnswerId":null,"OwnerUserId":265521.0,"Title":"How to duplicate matrix rows in Torch","Body":"<p>With Torch, how would I turn this matrix:<\/p>\n\n<pre><\/pre>\n\n<p>Into this:<\/p>\n\n<pre><\/pre>\n\n<p>?<\/p>\n\n<p>In Matlab I would do it something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a similar method in Torch?<\/p>\n","answers":[{"AnswerId":"36534297","CreationDate":"2016-04-10T19:09:45.043","ParentId":null,"OwnerUserId":"265521","Title":null,"Body":"<p>This works ok:<\/p>\n\n<pre><code>a = a:repeatTensor(1, 3):reshape(9, 3)\n<\/code><\/pre>\n"}]}
{"QuestionId":36538327,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-04-11T01:11:40.077","AcceptedAnswerId":36540230.0,"OwnerUserId":3427580.0,"Title":"Caffe sigmoid cross entropy loss","Body":"<p>I am using the sigmoid cross entropy loss function for a multilabel classification problem as laid out by <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/pascal-multilabel-with-datalayer.ipynb\" rel=\"nofollow\">this tutorial<\/a>. However, in both their results on the tutorial and my results, the output predictions are in the range , while the range of a sigmoid is . Is the sigmoid only processed in the backprop? That is, shouldn't a forward pass squash the output?<\/p>\n","answers":[{"AnswerId":"36540230","CreationDate":"2016-04-11T05:07:09.347","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In this example the input to the <code>\"SigmoidCrossEntropyLoss\"<\/code> layer is the output of a fully-connect layer. Indeed there are no constraints on the values of the outputs of an <code>\"InnerProduct\"<\/code> layer and they can be in range <code>[-inf, inf]<\/code>.<br>\nHowever, if you examine carefully the <code>\"SigmoidCrossEntropyLoss\"<\/code> you'll notice that it includes a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/layers\/sigmoid_cross_entropy_loss_layer.hpp#L99\" rel=\"noreferrer\"><code>\"Sigmoid\"<\/code> layer inside<\/a> -- to ensure stable gradient estimation.<br>\nTherefore, at test time, you should replace the <code>\"SigmoidCrossEntropyLoss\"<\/code> with a simple <code>\"Sigmoid\"<\/code> layer to output per-class predictions.<\/p>\n"}]}
{"QuestionId":36539560,"AnswerCount":1,"Tags":"<python><ipython><jupyter><torch>","CreationDate":"2016-04-11T03:52:39.323","AcceptedAnswerId":36543189.0,"OwnerUserId":4013781.0,"Title":"Installing Torch has changed Jupyter Notebook configuration","Body":"<p>I was running Jupyter Notebook with Caffe. After installing Torch, Jupyter Notebook has stopped working. It is producing an error: <\/p>\n\n<pre><\/pre>\n\n<p>But the only option available is 'iTorch'. How can I reconfigure it please? Thanks. <\/p>\n\n<hr>\n\n<p>Typing jupyter kernelspec list produces: <\/p>\n\n<pre><\/pre>\n\n<p>How can I reinstall the python2 kernel? ipython kernelspec install --user did not work.<\/p>\n","answers":[{"AnswerId":"36543189","CreationDate":"2016-04-11T08:12:03.593","ParentId":null,"OwnerUserId":"4013781","Title":null,"Body":"<p>Installing Torch adversely affected my ipython installation. After removing ipython and reinstalling, everything is working as expected. This is what I did: <\/p>\n\n<pre><code>sudo apt-get --purge remove ipython\nsudo pip uninstall ipython\ngit clone https:\/\/github.com\/ipython\/ipython.git\ncd ipython\nsudo pip install -e . \n<\/code><\/pre>\n\n<p>If anyone knows what Torch did, please post. Thanks. <\/p>\n"}]}
{"QuestionId":36540563,"AnswerCount":0,"Tags":"<module><machine-learning><installation><torch>","CreationDate":"2016-04-11T05:38:56.217","AcceptedAnswerId":null,"OwnerUserId":4013781.0,"Title":"Installing Torch Extension fb-caffe-exts from github","Body":"<p>I am trying to install the fb-caffe-exts extension for Torch from Github <a href=\"https:\/\/github.com\/facebook\/fb-caffe-exts\" rel=\"nofollow\">fb-caffe-exts<\/a>, specifically I am trying to use torch2caffe. <\/p>\n\n<p>I am trying to run the test.lua demo for torch2caffe, but I am getting the following error, which refers to a line in the test.lua file:<\/p>\n\n<pre><\/pre>\n\n<p>All the tests fail, with the stack trace referring to this error. I have installed the dependencies (fb.lualib etc) and added the lua packages to my path. Fb.python and the other dependencies import without producing errors. <\/p>\n\n<p>Any help appreciated. <\/p>\n","answers":[]}
{"QuestionId":36542377,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-04-11T07:27:49.537","AcceptedAnswerId":null,"OwnerUserId":6186894.0,"Title":"diag_part shape issue in TensorFlow","Body":"<p>Running from master the following code gives an error.<\/p>\n\n<pre><\/pre>\n\n<p>The relevant part of the error is <\/p>\n\n<blockquote>\n  <p>\/tensorflow\/python\/ops\/array_ops.py\", line 960, in _DiagPartShape\n      \" do not match \")<\/p>\n  \n  <p><strong>ValueError<\/strong>: Invalid shape, shape[:mid] (?,) and shape[mid:] (?,) do not match <\/p>\n<\/blockquote>\n\n<p>Am I doing something wrong or is this a bug?<\/p>\n\n<p>Thanks,\nAlex<\/p>\n","answers":[]}
{"QuestionId":36547440,"AnswerCount":0,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-04-11T11:28:33.000","AcceptedAnswerId":null,"OwnerUserId":6061106.0,"Title":"How can I force Caffe to print the label near each probability or fix the order of their printing?","Body":"<p>I'm trying to analyze probabilities of a binary classifier according to the solution #2 suggested here: <a href=\"https:\/\/stackoverflow.com\/questions\/35993175\/how-can-i-get-two-output-values-for-each-of-the-two-classes-for-a-binary-class?answertab=votes#tab-top\">How can I get two output values (for each of the two classes) for a binary classifier in Caffe?<\/a>\nI've got the following output of my testing:<\/p>\n\n<pre><\/pre>\n\n<p>There is a problem that the two probabilities for each class are not labeled here. So in this particular case for iterations 0-1999 the probabilities are printed in the following order: first the probability for class 0 (\"no\") is printed and then for class 1 (\"yes\"). For iterations 2000-2999 it is printed vice versa: (first probability for class 1 (\"yes\") and then probability for class 0 (\"no\").\nThis is not convenient since I further make an automated parsing of this log.\n<strong>How can I force Caffe to print the label near each probability or fix the order of their printing?<\/strong><\/p>\n","answers":[]}
{"QuestionId":36547615,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2016-04-11T11:36:45.070","AcceptedAnswerId":null,"OwnerUserId":4919612.0,"Title":"Upsampling feature maps in TensorFlow","Body":"<p>I want to implement a convolution-deconvolution network for a image segmentation project. In the deconvolution part, I am planning to upsample the feature map by 2. e.g. The original feature map is of dimension 64*64*4 and I want to upsample it into 128*128*4. Does anyone know a tensor operation that does this? Thanks!<\/p>\n","answers":[{"AnswerId":"38622226","CreationDate":"2016-07-27T20:04:56.323","ParentId":null,"OwnerUserId":"3075902","Title":null,"Body":"<p>You could use <code>tf.image.resize_images()<\/code>. It takes batches of images or single images and supports the most common methods such as <code>bilinear<\/code> and <code>nearest_neighbor<\/code>. <\/p>\n\n<p>Here's the link to the TensorFlow API reference: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.9\/api_docs\/python\/image.html#resizing\" rel=\"noreferrer\">resizing<\/a><\/p>\n\n<p>You can also take a look at how the upsampling operation is implemented in a higher-level API such as <a href=\"http:\/\/tflearn.org\/\" rel=\"noreferrer\">tflearn<\/a>. You can find <code>upsample_2d<\/code> and <code>upscore_layer<\/code> in their Github repo:  <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/tflearn\/layers\/conv.py\" rel=\"noreferrer\">conv.py<\/a><\/p>\n\n<p><strong>Note:<\/strong> the output might be cast to <code>tf.float32<\/code> in older TF versions <\/p>\n"}]}
{"QuestionId":36547907,"AnswerCount":1,"Tags":"<python><gpu><tensorflow><deep-learning><keras>","CreationDate":"2016-04-11T11:51:43.653","AcceptedAnswerId":null,"OwnerUserId":1045285.0,"Title":"How to control memory while using Keras with tensorflow backend?","Body":"<p>I have created a wrapper class which initializes a  model and has a couple of methods for starting the training process and monitoring the progress. I instantiate this class in my  file and perform the training process. Fairly mundane stuff.<\/p>\n\n<p>My <strong>question<\/strong> is:<\/p>\n\n<p>How to free all the GPU memory allocated by . I tried the following with no luck:<\/p>\n\n<pre><\/pre>\n\n<p>Even after the session has been closed and reset to ,  does not reflect any reduction in memory usage. Any ideas?<\/p>\n\n<p><strong>Idea<\/strong><\/p>\n\n<p>Would it be meaningful to add a  method to my  class and instantiate it as:<\/p>\n\n<pre><\/pre>\n\n<p>How should I free up the resources of the keras model in this method?<\/p>\n\n<p><strong>Test environment<\/strong><\/p>\n\n<p>I'm using iPython Notebook on an Ubuntu 14.04 with 3 GTX 960 GPUs.<\/p>\n\n<p>Reference:<\/p>\n\n<ol>\n<li><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2102\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/2102<\/a><\/li>\n<li><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/MFUEY9P1sc8\" rel=\"noreferrer\">https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/MFUEY9P1sc8<\/a><\/li>\n<\/ol>\n","answers":[{"AnswerId":"43083767","CreationDate":"2017-03-29T03:28:14.113","ParentId":null,"OwnerUserId":"2436995","Title":null,"Body":"<p>The following works for me to reinitialize the state of Keras layers in my Jupyter notebook for every run:<\/p>\n\n<pre><code>from keras import backend as K\nK.clear_session()\nsess = tf.Session()\nK.set_session(sess)\n<\/code><\/pre>\n\n<p>Also, the graph is named and reset every time the notebook runs using:<\/p>\n\n<pre><code>graphr = K.get_session().graph\nwith graphr.as_default():\n    #...graph building statements...\n<\/code><\/pre>\n\n<p>Note : I am still trying to wrap my head around the concepts of Keras and tensorflow ( I believe they are described poorly in documentation and examples ) but the above works.<\/p>\n"}]}
{"QuestionId":36548096,"AnswerCount":1,"Tags":"<python><neural-network><tensorflow><multi-layer>","CreationDate":"2016-04-11T11:59:34.663","AcceptedAnswerId":null,"OwnerUserId":5967574.0,"Title":"MLP on TensorFlow is giving the same prediction for all observations after the training","Body":"<p>I am trying to train a sparse data with an MLP to predict a forecast. However, the forecast on the test data is giving the same value for all observations. Once I omit the activation function from each layer, the outcome starts being different. \nmy code is below:<\/p>\n\n<pre><\/pre>\n\n<p>PS: I am generating randomly my data just to reproduce the error. My data is sparse in fact, pretty similar to the one generated randomly. The problem I want to solve is: MLP is giving the same prediction for all observations in the test data.<\/p>\n","answers":[{"AnswerId":"36558505","CreationDate":"2016-04-11T20:21:14.203","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>That's a sign that your training failed. With GoogeLeNet Imagenet training I've seen it label everything as \"nematode\" when started with a bad choice of hyper-parameters. Things to check -- does your training loss decrease? If it doesn't decrease, try different learning rates\/architectures. If it decreases to zero maybe your loss is wrong like was case <a href=\"https:\/\/stackoverflow.com\/questions\/36468701\/adamoptimizer-and-gradientdescentoptimizer-from-tensorflow-not-able-to-fit-simpl\/36481418#36481418\">here<\/a><\/p>\n"}]}
{"QuestionId":36548375,"AnswerCount":0,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-04-11T12:12:49.660","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"Shape of gradients returned by tf.gradients()","Body":"<p>I have the following graph:<\/p>\n\n<pre><\/pre>\n\n<p>and then run the following session:<\/p>\n\n<pre><\/pre>\n\n<p>If I e.g. look at the gradient of cost with respect to the embedding matrix A, I get a gradient of shape (batch_size * memory_size * max_sent_length, d), so the first dimension of the gradient corresponds to the product of the dimensions of the input tensor. I don't understand why tensorflow returns a matrix of that shape for the gradient of A,  shouldn't you get a gradient with the shape of A?<\/p>\n","answers":[]}
{"QuestionId":36548736,"AnswerCount":4,"Tags":"<tensorflow><conv-neural-network><deconvolution>","CreationDate":"2016-04-11T12:29:21.293","AcceptedAnswerId":36557995.0,"OwnerUserId":5724595.0,"Title":"TensorFlow: Unpooling","Body":"<p>Is there TensorFlow native function that does unpooling for Deconvolutional Networks ? <\/p>\n\n<p>I have written this in normal python, but it is getting complicated when want to translate it to TensorFlow as it's objects does not even support item assignment at the moment, and I think this is a great inconvenience with TF.<\/p>\n","answers":[{"AnswerId":"58687655","CreationDate":"2019-11-04T05:30:22.250","ParentId":null,"OwnerUserId":"4030917","Title":null,"Body":"<p>I checked <a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/tensorpack\/models\/pool.py#L66\" rel=\"nofollow noreferrer\">this<\/a> which shagas mentioned <a href=\"https:\/\/stackoverflow.com\/a\/36557995\/4030917\">here<\/a> and it is working.<\/p>\n\n<pre><code>x = [[[[1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3]],\n  [[1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3]],\n[[1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3],\n  [1, 1, 2,2, 3, 3]]]]\n\nx = np.array(x)\n\ninp = tf.convert_to_tensor(x)\n\nout = UnPooling2x2ZeroFilled(inp)\n\nout\nOut[19]: \n&lt;tf.Tensor: id=36, shape=(1, 6, 12, 6), dtype=int64, numpy=\narray([[[[1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0]],\n\n        [[0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0]],\n\n        [[1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0]],\n\n        [[0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0]],\n\n        [[1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0],\n         [1, 1, 2, 2, 3, 3],\n         [0, 0, 0, 0, 0, 0]],\n\n        [[0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0]]]])&gt;\n\n\nout1 = tf.keras.layers.MaxPool2D()(out)\n\nout1\nOut[37]: \n&lt;tf.Tensor: id=118, shape=(1, 3, 6, 6), dtype=int64, numpy=\narray([[[[1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3]],\n\n        [[1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3]],\n\n        [[1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3],\n         [1, 1, 2, 2, 3, 3]]]])&gt;\n<\/code><\/pre>\n\n<p>If you need max unpooling then you can use (though I didn't check it) <a href=\"https:\/\/github.com\/fregu856\/segmentation\/blob\/f051c6a918e74bf54b8d44ad1cc2c8103ee0b9d6\/utilities.py#L34\" rel=\"nofollow noreferrer\">this one<\/a><\/p>\n"},{"AnswerId":"36557995","CreationDate":"2016-04-11T19:52:06.597","ParentId":null,"OwnerUserId":"3860525","Title":null,"Body":"<p>I don't think there is an official unpooling layer yet which is frustrating because you have to use image resize (bilinear interpolation or nearest neighbor) which is like an average unpooling operation and it's reaaaly slow. Look at the tf api in the section 'image' and you will find it. <\/p>\n\n<p>Tensorflow has a maxpooling_with_argmax thing where you get you maxpooled output as well as the activation map which is nice as you could use it in an unpooling layer to preserve the 'lost' spacial information but it seems as there isn't such an unpooling operation that does it. I guess that they are planning to add it ... soon.<\/p>\n\n<p>Edit: I found some guy on google discuss a week ago who seems to have implemented something like this but I personally haven't tried it yet.\n<a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/tensorpack\/models\/pool.py#L66\">https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/tensorpack\/models\/pool.py#L66<\/a><\/p>\n"},{"AnswerId":"39167075","CreationDate":"2016-08-26T13:06:37.230","ParentId":null,"OwnerUserId":"2057762","Title":null,"Body":"<p>I was searching for a <strong>max<\/strong>unpooling operation and tried implementing it. I came up with some kind of <a href=\"https:\/\/github.com\/jon-sch\/tensorflow\/blob\/2eec55bcd2ccad6fd4d5bb13ab638227d9ad0556\/tensorflow\/python\/ops\/nn_grad.py#L374-L390\" rel=\"nofollow noreferrer\">hacky implementation for the gradient<\/a>, as I was struggling with CUDA.<\/p>\n\n<p>The code is <a href=\"https:\/\/github.com\/jon-sch\/tensorflow\/pull\/3\/files\" rel=\"nofollow noreferrer\">here<\/a>, you will need to build it from source with GPU support.\nBelow is a demo application. No warranties, though!<\/p>\n\n<p>There also exists an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2169\" rel=\"nofollow noreferrer\">open issue<\/a> for this operation.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\ndef max_pool(inp, k=2):\n    return tf.nn.max_pool_with_argmax_and_mask(inp, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding=\"SAME\")\n\ndef max_unpool(inp, argmax, argmax_mask, k=2):\n    return tf.nn.max_unpool(inp, argmax, argmax_mask, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding=\"SAME\")\n\ndef conv2d(inp, name):\n    w = weights[name]\n    b = biases[name]\n    var = tf.nn.conv2d(inp, w, [1, 1, 1, 1], padding='SAME')\n    var = tf.nn.bias_add(var, b)\n    var = tf.nn.relu(var)\n    return var\n\ndef conv2d_transpose(inp, name, dropout_prob):\n    w = weights[name]\n    b = biases[name]\n\n    dims = inp.get_shape().dims[:3]\n    dims.append(w.get_shape()[-2]) # adpot channels from weights (weight definition for deconv has switched input and output channel!)\n    out_shape = tf.TensorShape(dims)\n\n    var = tf.nn.conv2d_transpose(inp, w, out_shape, strides=[1, 1, 1, 1], padding=\"SAME\")\n    var = tf.nn.bias_add(var, b)\n    if not dropout_prob is None:\n        var = tf.nn.relu(var)\n        var = tf.nn.dropout(var, dropout_prob)\n    return var\n\n\nweights = {\n    \"conv1\":    tf.Variable(tf.random_normal([3, 3,  3, 16])),\n    \"conv2\":    tf.Variable(tf.random_normal([3, 3, 16, 32])),\n    \"conv3\":    tf.Variable(tf.random_normal([3, 3, 32, 32])),\n    \"deconv2\":  tf.Variable(tf.random_normal([3, 3, 16, 32])),\n    \"deconv1\":  tf.Variable(tf.random_normal([3, 3,  1, 16])) }\n\nbiases = {\n    \"conv1\":    tf.Variable(tf.random_normal([16])),\n    \"conv2\":    tf.Variable(tf.random_normal([32])),\n    \"conv3\":    tf.Variable(tf.random_normal([32])),\n    \"deconv2\":  tf.Variable(tf.random_normal([16])),\n    \"deconv1\":  tf.Variable(tf.random_normal([ 1])) }\n\n\n## Build Miniature CEDN\nx = tf.placeholder(tf.float32, [12, 20, 20, 3])\ny = tf.placeholder(tf.float32, [12, 20, 20, 1])\np = tf.placeholder(tf.float32)\n\nconv1                                   = conv2d(x, \"conv1\")\nmaxp1, maxp1_argmax, maxp1_argmax_mask  = max_pool(conv1)\n\nconv2                                   = conv2d(maxp1, \"conv2\")\nmaxp2, maxp2_argmax, maxp2_argmax_mask  = max_pool(conv2)\n\nconv3                                   = conv2d(maxp2, \"conv3\")\n\nmaxup2                                  = max_unpool(conv3, maxp2_argmax, maxp2_argmax_mask)\ndeconv2                                 = conv2d_transpose(maxup2, \"deconv2\", p)\n\nmaxup1                                  = max_unpool(deconv2, maxp1_argmax, maxp1_argmax_mask)\ndeconv1                                 = conv2d_transpose(maxup1, \"deconv1\", None)\n\n\n## Optimizing Stuff\nloss        = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(deconv1, y))\noptimizer   = tf.train.AdamOptimizer(learning_rate=1).minimize(loss)\n\n\n## Test Data\nnp.random.seed(123)\nbatch_x = np.where(np.random.rand(12, 20, 20, 3) &gt; 0.5, 1.0, -1.0)\nbatch_y = np.where(np.random.rand(12, 20, 20, 1) &gt; 0.5, 1.0,  0.0)\nprob    = 0.5\n\n\nwith tf.Session() as session:\n    tf.set_random_seed(123)\n    session.run(tf.initialize_all_variables())\n\n    print \"\\n\\n\"\n    for i in range(10):\n        session.run(optimizer, feed_dict={x: batch_x, y: batch_y, p: prob})\n        print \"step\", i + 1\n        print \"loss\",  session.run(loss, feed_dict={x: batch_x, y: batch_y, p: 1.0}), \"\\n\\n\"\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Edit 29.11.17<\/strong><\/p>\n\n<p>Some time back, I reimplemented it in a clean fashion against TensorFlow 1.0, the forward operations are also available as CPU-version. You can find it <a href=\"https:\/\/github.com\/jon-sch\/tensorflow\/commits\/feature\/unpooling_r1.0\" rel=\"nofollow noreferrer\">in this branch<\/a>, I recommend you looking up the last few commits if you want to use it.<\/p>\n"},{"AnswerId":"42549265","CreationDate":"2017-03-02T07:31:13.320","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>There is a couple of tensorflow implementations here <a href=\"https:\/\/github.com\/yselivonchyk\/Tensorflow_WhatWhereAutoencoder\/blob\/master\/pooling.py\" rel=\"nofollow noreferrer\" title=\"pooling.py\">pooling.py<\/a><\/p>\n\n<p>Namely:<\/p>\n\n<p>1) unpool operation (<a href=\"https:\/\/arxiv.org\/abs\/1506.02351\" rel=\"nofollow noreferrer\">source<\/a>) that utilizes output of <code>tf.nn.max_pool_with_argmax<\/code>. Although please notice, that as of tensorflow 1.0 <code>tf.nn.max_pool_with_argmax<\/code> is GPU-only<\/p>\n\n<p>2) upsample operation that mimics inverse of max-pooling by filling positions of unpooled region with either zeros or copies of max element.\nComparing to <a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/tensorpack\/models\/pool.py#L66\" rel=\"nofollow noreferrer\">tensorpack<\/a> it allows copies of elements instead of zeros and supports strides other than <code>[2, 2]<\/code>. <\/p>\n\n<p>No recompile, back-prop friendly.<\/p>\n\n<p>Illustration:\n<a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/tensorpack\/models\/pool.py#L66\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7nKkL.png\" alt=\"Upsampling\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7nKkL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oHaap.png\" alt=\"Unpooling\"><\/a><\/p>\n"}]}
{"QuestionId":36550131,"AnswerCount":2,"Tags":"<machine-learning><computer-vision><deep-learning><caffe><lmdb>","CreationDate":"2016-04-11T13:28:14.220","AcceptedAnswerId":36554793.0,"OwnerUserId":5833345.0,"Title":"Check failed: mdb_status == 0 (2 vs. 0) No such file or directory","Body":"<p>I received the following error while I was training the data. I have tried all the solutions given on the internet and nothing seems to work for me. I have checked paths and size of the lmdb files are non-zero. But the problem still exists. I have no idea how to solve this issue.<\/p>\n\n<pre><\/pre>\n\n<p>Below is my file settings:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36554793","CreationDate":"2016-04-11T16:56:33.767","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>You have not set your paths to the LMDB directories correctly. Go to the directory where you have created your LMDBs and get the absolute paths using this command:<\/p>\n\n<pre><code>$ readlink -f &lt;LMDB_directory_name&gt;\n<\/code><\/pre>\n\n<p>Use this path, it should solve your problem.<\/p>\n"},{"AnswerId":"37330305","CreationDate":"2016-05-19T17:32:28.847","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>To expand on Harsh's answer:<\/p>\n\n<p>Make sure you carefully read the set-up steps on the <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/imagenet.html\" rel=\"nofollow\">Caffe Imagenet page<\/a>.  Some of the steps you have to carry out are embedded within the text; not all of them are in code boxes.<\/p>\n\n<p>Specific to this case, you have to edit file examples\/imagenet\/create_imagenet.sh, replacing the <strong>path\/to<\/strong> references with the correct path in your environment: this is wherever the imagenet files live.  Lines 9&amp;10 need your attention:<\/p>\n\n<pre><code>TRAIN_DATA_ROOT=\/path\/to\/imagenet\/train\/\nVAL_DATA_ROOT=\/path\/to\/imagenet\/val\/\n<\/code><\/pre>\n\n<p>Also, at line 5, make sure that your EXAMPLE variable is set to a location with enough space for the compressed images: <strong>train<\/strong> requires 41Gb, but the pre-processing high-water mark is at least 55Gb.  <strong>test<\/strong> occupies only 1.7Gb.<\/p>\n"}]}
{"QuestionId":36551186,"AnswerCount":0,"Tags":"<python><docker><pycharm><tensorflow>","CreationDate":"2016-04-11T14:12:28.497","AcceptedAnswerId":null,"OwnerUserId":5847441.0,"Title":"Settings up remote interpretor in pycharm for docker","Body":"<p>I'm using <strong>PyCharm community editor<\/strong> 5.0.3.I. I want to run my <strong>Tensorflow python<\/strong> project on <strong>docker<\/strong> running on <strong>windows<\/strong>. As per comment community edition doesn't support remote interpretor. What's the possible way to excepting using Enterprise edition of PyCharm? <\/p>\n","answers":[]}
{"QuestionId":36551284,"AnswerCount":3,"Tags":"<image-processing><computer-vision><neural-network><deep-learning><caffe>","CreationDate":"2016-04-11T14:16:24.310","AcceptedAnswerId":36551710.0,"OwnerUserId":570593.0,"Title":"How to train Caffe with only G and B channels","Body":"<p>Is there anyway to use only  and  channels for training Caffe using  input layer?<\/p>\n","answers":[{"AnswerId":"36551710","CreationDate":"2016-04-11T14:34:20.617","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can add a convolution layer on top of your input that will select G and B:<\/p>\n\n<pre><code>layer {\n  name: \"select_B_G\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"select_B_G\"\n  convolution_param { kernel_size: 1 num_output: 2 bias_term: false }\n  param { lr_mult: 0 } # do not learn parameters for this layer\n}\n<\/code><\/pre>\n\n<p>You'll need to do some <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow\">net surgery<\/a> prior to training to set the weights for this layer to be<\/p>\n\n<pre><code>net.params['select_B_G'][0].data[...] = np.array( [[1,0,0],[0,1,0]], dtype='f4')\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> sometimes images loaded to caffe are going through channel-swap transformation, i.e., RGB -> BGR, therefore you need to be careful what channels you pick.<\/p>\n"},{"AnswerId":"36564930","CreationDate":"2016-04-12T06:14:40.160","ParentId":null,"OwnerUserId":"3130081","Title":null,"Body":"<p>I wrote a simple python layer to do this, by the way, I did't test this code.<\/p>\n\n<pre><code>import caffe\n\nclass ExtractGBChannelLayer(caffe.Layer):\n  def setup(self,bottom,top):\n    pass\n  def reshape(self,bottom,top):\n    bottom_shape=bottom[0].data.shape\n    top_shape=[bottom_shape[0],2,bottom_shape[2],bottom_shape[3]] #because we only want G and B channels.\n    top[0].reshape(*top_shape)\n  def forward(self,bottom,top):\n    #copy G and B channel to top, note caffe BGR order!\n    top[0].data[:,0,...]=bottom[0].data[:,1,...]\n    top[0].data[:, 1, ...] = bottom[0].data[:, 0, ...]\n  def backward(self,top,propagate_down,bottom):\n    pass\n<\/code><\/pre>\n\n<p>You can save this file as <code>MyPythonLayer.py<\/code><\/p>\n\n<p>In you prototxt you can insert a layer after <code>ImageDataLayer<\/code> like this<\/p>\n\n<pre><code>layer {\n  name: \"GB\"\n  type: \"Python\"\n  bottom: \"data\"\n  top: \"GB\"\n  python_param {\n    module: \"MyPythonLayer\"\n    layer: \"ExtractGBChannelLayer\"\n  }\n}\n<\/code><\/pre>\n\n<p>Hope this works fine.<\/p>\n"},{"AnswerId":"36570718","CreationDate":"2016-04-12T10:34:57.730","ParentId":null,"OwnerUserId":"570593","Title":null,"Body":"<p>This is the Matlab code I used and it works.<\/p>\n\n<pre><code> caffe.reset_all(); % reset caffe\n caffe.set_mode_gpu();  \n gpu_id = 0;  % we will use the first gpu in this demo\n caffe.set_device(gpu_id);\n net_model = ['net_images.prototxt'];\n net = caffe.Net(net_model, 'train')\n a = zeros(1,1,3,2);\n a(1,1,:,:) = [[1,0,0];[0,1,0]]'; % caffe uses BGR color channel order\n net.layers('select_B_G').params(1).set_data(a);\n solver = caffe.Solver(solverFN);\n solver.solve();\n net.save(fullfile(model_dir, 'my_net.caffemodel'));\n<\/code><\/pre>\n"}]}
{"QuestionId":36551936,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-04-11T14:44:17.703","AcceptedAnswerId":null,"OwnerUserId":6176753.0,"Title":"ValueError: Tensor A must be from the same graph as Tensor B","Body":"<p>I'm doing text matching using tensorflow, before i call , I have to combine some words from 2 sentence(get m words from sentence S1 and also get m words from sentence S2, then combine them together as \"combine_result\"), but when the code gose to  it gives me the error: <\/p>\n\n<blockquote>\n  <p>ValueError: Tensor(\"Reshape_7:0\", shape=(1, 6), dtype=int32) must be\n  from the same graph as Tensor(\"word_embedding_matrix:0\", shape=(26320,\n  50), dtype=float32_ref).<\/p>\n<\/blockquote>\n\n<p>the code is as bellow:<\/p>\n\n<pre><\/pre>\n\n<p>could please someone help me? Thank you very much!<\/p>\n","answers":[{"AnswerId":"47367486","CreationDate":"2017-11-18T15:05:49.113","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Yaroslav answered in a comment above - moving to an answer:<\/p>\n\n<p>This error happens when you create new default graph. Try to do <code>tf.reset_default_graph()<\/code> before the computation and not create any more graphs (i.e., calls to tf.Graph) <\/p>\n"}]}
{"QuestionId":36552101,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-11T14:51:15.273","AcceptedAnswerId":36553123.0,"OwnerUserId":3688217.0,"Title":"prediction in tensor flow for multiple labels","Body":"<p>How can I get a prediction vector in the tensorflow graph apart from \n ? (since argmax only works for softmax classifiers)<\/p>\n\n<p>I have a multilabel classification problem so therefore I need something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36553123","CreationDate":"2016-04-11T15:36:35.080","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>Hope this helps :<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nsess = tf.InteractiveSession()\nprob = tf.constant(np.random.rand(10))\npredictions = tf.select(prob &gt; 0.5, tf.ones_like(prob), tf.zeros_like(prob))\nprint(predictions.eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":36552254,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2016-04-11T14:57:56.017","AcceptedAnswerId":36552737.0,"OwnerUserId":2886575.0,"Title":"torch testJacobian works with DoubleTensor, not with CudaTensor","Body":"<p>I'm working on a neural network using this module:  <a href=\"https:\/\/github.com\/qassemoquab\/stnbhwd\/blob\/master\/AffineGridGeneratorBHWD.lua\" rel=\"nofollow\">https:\/\/github.com\/qassemoquab\/stnbhwd\/blob\/master\/AffineGridGeneratorBHWD.lua<\/a><\/p>\n\n<p>nn.Jacobian.testJacobian is large when I run the module as :cuda() with a CudaTensor input, but not when I run it as :double() with the same DoubleTensor input<\/p>\n\n<p>The :forward for a :double() run and a :cuda() run are very close.<\/p>\n\n<p>The :backward for a :double() run and a :cuda() run are wildly different, so the problem is somewhere in the updateGradInput method, I think:<\/p>\n\n<pre><\/pre>\n\n<p>The bmm method was the only one I wasn't super familiar with, so I tested it out, and it gives comparable results with :double() and :cuda().<\/p>\n\n<p>Anybody have any experience with similar problems?  I'll continue trying to track down the problem in the meantime...<\/p>\n\n<p><strong>EDIT<\/strong>\nTo illustrate the extent of the discrepancy:<\/p>\n\n<pre><\/pre>\n\n<p>It is possible that I'm mistaken about the problem being in the updateGradInput method.... Still poking about.<\/p>\n","answers":[{"AnswerId":"36552737","CreationDate":"2016-04-11T15:17:56.687","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>This is expected (and not wrong or a  bug). Jacobian tests need a precision that's high enough and Float preicison (CudaTensor) doesn't give you that. <\/p>\n"}]}
{"QuestionId":36552860,"AnswerCount":1,"Tags":"<installation><cmake><makefile><libraries><caffe>","CreationDate":"2016-04-11T15:23:24.843","AcceptedAnswerId":null,"OwnerUserId":5295893.0,"Title":"Installing caffe brings up some questions, depending libraries and versions","Body":"<p>I wanted to install caffe on openSuse.\nJust for the record - it worked out for me, I just don't know what's the \"exact\" way to do this. The things I did maybe aren't really for someone who's new to this, and also it was kind of a \"bad installation\". My way was the following:<\/p>\n\n<p>First, I did<\/p>\n\n<pre><\/pre>\n\n<p>This worked, until it complained that some libraries weren't found (libclbas etc.). So I used<\/p>\n\n<pre><\/pre>\n\n<p>to change the paths to the libraries manually. I needed to manually type the paths to the snappy, boost_python, blas, cblas and lapack libs. After doing that I did <\/p>\n\n<pre><\/pre>\n\n<p>and then<\/p>\n\n<pre><\/pre>\n\n<p>and everything worked. My problem now is - why doesn't make find the libs, and is there a way to fix this? I think the problem was that I didn't have \/usr\/lib\/libcblas.so but \/usr\/lib\/libcblas.so.3, and similar \"problems\" with the other libraries.<\/p>\n\n<p>Another thing - when I tried using ccmake\/cmake right from the beginning (without the make part first), there weren't any files in my build directory (like $CAFFE_ROOT\/build\/examples or $CAFFE_ROOT\/build\/tools were empty), so the mnist tutorial for example wasn't working. That's why I first called <\/p>\n\n<pre><\/pre>\n\n<p>, what may seem strange to you.<\/p>\n\n<p>Of course I know how to fix this stuff, but I would like to know how the correct way for a \"clean and simple installation\" is. Did is miss anything when using make\/cmake, is this some kind of inconsistency in caffe or something else? And, what <em>is<\/em> the clean way to do this?<\/p>\n","answers":[{"AnswerId":"36556584","CreationDate":"2016-04-11T18:37:12.197","ParentId":null,"OwnerUserId":"387099","Title":null,"Body":"<p>Maybe look at the Ubuntu installation guide? <a href=\"http:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow\">http:\/\/caffe.berkeleyvision.org\/install_apt.html<\/a><\/p>\n\n<p>It mentions all the different packages you might need. I couldn't find openSuse installation instructions - but you should be able to translate the <code>apt-get<\/code> commands for your platform.<\/p>\n"}]}
{"QuestionId":36552913,"AnswerCount":1,"Tags":"<deep-learning><caffe><matcaffe>","CreationDate":"2016-04-11T15:26:12.377","AcceptedAnswerId":null,"OwnerUserId":570593.0,"Title":"caffe how to set a filter using matlab?","Body":"<p>I want to update a convolutional filter using matlab in Caffe.\nHow can I do it?\nso far my code is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36570919","CreationDate":"2016-04-12T10:43:38.913","ParentId":null,"OwnerUserId":"570593","Title":null,"Body":"<p>The following works fine.<\/p>\n\n<pre><code>caffe.reset_all(); % reset caffe\ncaffe.set_mode_gpu();  \ngpu_id = 0;  % we will use the first gpu in this demo    \ncaffe.set_device(gpu_id);\nnet_model = ['net_images.prototxt'];\nnet = caffe.Net(net_model, 'train')\na = zeros(1,1,3,2);\na(1,1,:,:) = [[1,0,0];[0,1,0]]'; % caffe uses BGR color channel order\n         net.layers('select_B_G').params(1).set_data(a);\nsolver = caffe.Solver(solverFN);\nsolver.solve();\nnet.save(fullfile(model_dir, 'my_net.caffemodel'));\n<\/code><\/pre>\n"}]}
{"QuestionId":36553860,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2016-04-11T16:08:11.160","AcceptedAnswerId":null,"OwnerUserId":2527680.0,"Title":"How to debug Tensorflow Network weights\/outputs\/cost per input sample?","Body":"<p>I have made a tensor flow model.\nBut always get a NAN loss for some reason.\nI would like to know how I could debug and view each value in each tensor.<\/p>\n\n<p>For eg:-<\/p>\n\n<pre><\/pre>\n\n<p>During Run time I'd like to View the values in this tensor and see where things go wrong.\nI have found something similar in this <a href=\"https:\/\/stackoverflow.com\/questions\/33633370\/how-to-print-the-value-of-a-tensor-object-in-tensorflow\">post<\/a><\/p>\n\n<p>Where they do something like this <\/p>\n\n<pre><\/pre>\n\n<p>But this gives and out put like this <\/p>\n\n<pre><\/pre>\n\n<p>Which is not really that Useful as I cant view all the values.\nI'd like to know if there is a step by step debugging option ?<\/p>\n","answers":[{"AnswerId":"41192507","CreationDate":"2016-12-16T21:14:39.030","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>TensorFlow now has a builtin debugger called <em>tfdbg<\/em>. It exposes the intermediate tensors in the graph, along with the graph structure, which should make it easier for you to debug this type of problems. Compared to print ops, it requires less code change and provides better coverage of the graph.<\/p>\n\n<p>Please take a look at its documentation \/ tutorial at master HEAD:\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"},{"AnswerId":"39017898","CreationDate":"2016-08-18T12:11:28.033","ParentId":null,"OwnerUserId":"507062","Title":null,"Body":"<p><code>tf.Print()<\/code> will display more results if you pass the <code>summarize<\/code> optional parameter. For example <code>summarize=100<\/code> would display the first 100 elements of the tensor.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/control_flow_ops.html#Print\" rel=\"nofollow\">API Reference<\/a><\/p>\n\n<p>(Not sure if the feature was there when the question was asked)<\/p>\n"}]}
{"QuestionId":36554827,"AnswerCount":1,"Tags":"<keras><deep-learning><theano><face-recognition><vgg-net>","CreationDate":"2016-04-11T16:58:44.360","AcceptedAnswerId":36559623.0,"OwnerUserId":5276460.0,"Title":"How to use the vgg-net when I load vgg16_weights.h5?","Body":"<p>I use the VGG-16 Net by keras. This is the <a href=\"https:\/\/gist.github.com\/baraldilorenzo\/07d7802847aaad0a35d3\" rel=\"nofollow noreferrer\">detail<\/a><\/p>\n\n<p>my problem is how to use this net to fine-tuning, and must I use the image size which is 224*224 for this net? And I must use 1000 classes when I use this net?\nif I don't use 1000 classes, it cause the error<\/p>\n\n<blockquote>\n  <p>Exception: Layer shape (4096L, 10L) not compatible with weight shape (4096, 1000).<\/p>\n<\/blockquote>\n\n<p>Asking for help, thank you!<\/p>\n","answers":[{"AnswerId":"36559623","CreationDate":"2016-04-11T21:31:10.940","ParentId":null,"OwnerUserId":"2930373","Title":null,"Body":"<p>I posted a detailed answer <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1728#issuecomment-184879068\" rel=\"noreferrer\">in this issue<\/a> if you want to take a look. The following snippet will help you with the dimension of your last layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import Sequential, Graph\nfrom keras.layers import Convolution2D, ZeroPadding2D, MaxPooling2D\nimport keras.backend as K\n\nimg_width, img_height = 128, 128\n\n# build the VGG16 network with our input_img as input\nfirst_layer = ZeroPadding2D((1, 1), input_shape=(3, img_width, img_height))\n\nmodel = Sequential()\nmodel.add(first_layer)\nmodel.add(Convolution2D(64, 3, 3, activation='relu', name='conv1_1'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(64, 3, 3, activation='relu', name='conv1_2'))\nmodel.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(128, 3, 3, activation='relu', name='conv2_1'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(128, 3, 3, activation='relu', name='conv2_2'))\nmodel.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(256, 3, 3, activation='relu', name='conv3_1'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(256, 3, 3, activation='relu', name='conv3_2'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(256, 3, 3, activation='relu', name='conv3_3'))\nmodel.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv4_1'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv4_2'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv4_3'))\nmodel.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv5_1'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv5_2'))\nmodel.add(ZeroPadding2D((1, 1)))\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv5_3'))\nmodel.add(MaxPooling2D((2, 2), strides=(2, 2)))\n\n# get the symbolic outputs of each \"key\" layer (we gave them unique names).\nlayer_dict = dict([(layer.name, layer) for layer in model.layers])\n\n# load the weights\n\nimport h5py\n\nweights_path = 'vgg16_weights.h5'\n\nf = h5py.File(weights_path)\nfor k in range(f.attrs['nb_layers']):\n    if k &gt;= len(model.layers):\n        # we don't look at the last (fully-connected) layers in the savefile\n        break\n    g = f['layer_{}'.format(k)]\n    weights = [g['param_{}'.format(p)] for p in range(g.attrs['nb_params'])]\n    model.layers[k].set_weights(weights)\nf.close()\nprint('Model loaded.')\n\n# Here is what you want:\n\ngraph_m = Graph()\ngraph_m.add_input('my_inp', input_shape=(3, img_width, img_height))\ngraph_m.add_node(model, name='your_model', input='my_inp')\ngraph_m.add_node(Flatten(), name='Flatten', input='your_model')\ngraph_m.add_node(Dense(4096, activation='relu'), name='Dense1',      input='Flatten')\ngraph_m.add_node(Dropout(0.5), name='Dropout1', input='Dense1')\ngraph_m.add_node(Dense(4096, activation='relu'), name='Dense2',  input='Dropout1')\ngraph_m.add_node(Dropout(0.5), name='Dropout2', input='Dense2')\ngraph_m.add_node(Dense(10, activation='softmax'), name='Final', input='Dropout2')\ngraph_m.add_output(name='out1', input='Final')\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\ngraph_m.compile(optimizer=sgd, loss={'out1': 'categorical_crossentropy'})\n<\/code><\/pre>\n\n<p>Note that you could freeze the training of the feature extraction layers and only fine tune the last fully connected layers.\nFrom the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/layers\/core.py#L25-L27\" rel=\"noreferrer\">doc<\/a>, you just have to add <code>trainable = False<\/code> to freeze the training of a layer.\nEx freezed:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>...\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv5_1', trainable=False))\n...\n<\/code><\/pre>\n\n<p>Ex trainable:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>...\nmodel.add(Convolution2D(512, 3, 3, activation='relu', name='conv5_1',     trainable=True))\n...\n<\/code><\/pre>\n\n<p><code>trainable<\/code> is <code>True<\/code> by default so that something happens if you don't know about the feature...<\/p>\n"}]}
{"QuestionId":36554893,"AnswerCount":1,"Tags":"<c++><graph><header><tensorflow>","CreationDate":"2016-04-11T17:02:14.743","AcceptedAnswerId":36555205.0,"OwnerUserId":5452997.0,"Title":"Which headers to include in Tensorflow load graph c++ example?","Body":"<p>The c++ API for Tensorflow looks simple enough, but I cannot figure out which header files are needed to run the shown example on their page:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/cc\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/cc\/index.html<\/a><\/p>\n\n<p>And if I get it working, will this API allow me to do classification with a neural net I built in Python?<\/p>\n","answers":[{"AnswerId":"36555205","CreationDate":"2016-04-11T17:20:50.540","ParentId":null,"OwnerUserId":"2748675","Title":null,"Body":"<p>For the headers I would assume some subset of these:<\/p>\n\n<pre><code>#include \"tensorflow\/cc\/ops\/standard_ops.h\"\n#include \"tensorflow\/core\/framework\/graph.pb.h\"\n#include \"tensorflow\/core\/framework\/tensor.h\"\n#include \"tensorflow\/core\/graph\/default_device.h\"\n#include \"tensorflow\/core\/graph\/graph_def_builder.h\"\n#include \"tensorflow\/core\/lib\/core\/threadpool.h\"\n#include \"tensorflow\/core\/lib\/strings\/stringprintf.h\"\n#include \"tensorflow\/core\/platform\/init_main.h\"\n#include \"tensorflow\/core\/platform\/logging.h\"\n#include \"tensorflow\/core\/platform\/types.h\"\n#include \"tensorflow\/core\/public\/session.h\"\n<\/code><\/pre>\n\n<p>From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/cc\/tutorials\/example_trainer.cc\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/cc\/tutorials\/example_trainer.cc<\/a><\/p>\n"}]}
{"QuestionId":36556190,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-04-11T18:15:39.493","AcceptedAnswerId":null,"OwnerUserId":1595417.0,"Title":"Tensorflow: Error on optimizer initialization with custom var_list","Body":"<p>I'm trying to optimize only a specific subset of my variables like so:<\/p>\n\n<pre><\/pre>\n\n<p>This gives me an error message<\/p>\n\n<pre><\/pre>\n\n<p>There is no error if I leave out the var_list, but this obviously isn't what I want, and from the error message it isn't clear to me what the problem is.<\/p>\n","answers":[]}
{"QuestionId":36556704,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><lasagne>","CreationDate":"2016-04-11T18:42:10.327","AcceptedAnswerId":36655213.0,"OwnerUserId":5988583.0,"Title":"Trying to use own dataset on Lasagne basic example (slightly modified) to make neural network for face recognition","Body":"<p>I'm doing the basic lasagne example:\n<a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py\" rel=\"nofollow\">https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/examples\/mnist.py<\/a><\/p>\n\n<p>which I've modified slightly by combining it with another similar example.<\/p>\n\n<p>I'm trying to run the CNN model, where I've added some extra input to the CNN def, but it shouldn't be any different. Also changed the default value of 28 for the input layer in the example to 60 (for height and width), classes is being used later on in the code, but the code \"hangs\" on at that last network line, meaning the code is still running, but nothing is happening.\n<a href=\"http:\/\/i.stack.imgur.com\/dJkBC.png\" rel=\"nofollow\">Output when running the code.<\/a> The input_var is defined as such in the mainloop:<\/p>\n\n<pre><\/pre>\n\n<p>rest of the code:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong>\nOkay, based on what I've tried so far it seems to be my own dataset that is the problem. I've reshaped my dataset to match the MNIST dataset where for ex. X_train has the shape [images, channel, height, width]. Where channel = 1 and height, width = 60. The code for retrieving these is:<\/p>\n\n<pre><\/pre>\n\n<p>where read_images1 is:<\/p>\n\n<pre><\/pre>\n\n<p>The code in main which is ran:<\/p>\n\n<pre><\/pre>\n\n<p>if mnist = 1 (in main) the code runs fine, if I try with my own dataset it gets stuck at in build_mlp (similar to original problem with cnn):<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT 2:<\/strong>\nAfters struggling with this for a while I found out that the the image resizing done in read_images1() caused the problem:<\/p>\n\n<pre><\/pre>\n\n<p>If I didn't pass any resizing and used the default image sizes that were in the folders, the neural network was able to compile. Does anyone have any idea why? I updated the read_images1() to this:<\/p>\n\n<pre><\/pre>\n\n<p>If I run the program with sz = None and na = True, then it works. If any size is given to the sz parameter the code gets stuck trying to compile the neural network again.<\/p>\n","answers":[{"AnswerId":"36655213","CreationDate":"2016-04-15T19:15:34.807","ParentId":null,"OwnerUserId":"1828289","Title":null,"Body":"<p>Okay, I think I can see a few issues here, not sure which one you're running into...<\/p>\n\n<ol>\n<li><p>In <code>read_images1()<\/code>, X is a python list of numpy arrays.  Where does it get converted to a single numpy array?  Try adding <code>X = numpy.asarray(X)<\/code>.  You also need to reshape it to (n_images, n_channels, width, height) where I assume n_channels=1 for grayscale.  The network expects 4D input not 3D.<\/p><\/li>\n<li><p>This code <code>list_of_labels = list(xrange(max(label) + 1)); classes = len(list_of_labels)<\/code> assumes the labels are sequential numbers from 0 to N.  Are they?<\/p><\/li>\n<li><p>Your <code>build_mlp(classes, height, width, input_var=None)<\/code> is pretty different from the original example <code>build_mlp(input_var=None)<\/code>.  The original example clearly works, so whatever is wrong has to do with the differences.  One of the big differences is that you keep assigning to the same variable, like this <code>network = lasagne.layers.DenseLayer(network, ...)<\/code> where the original has different variables for each layer <code>l_hid1 = lasagne.layers.DenseLayer(l_in_drop, ...)<\/code><\/p><\/li>\n<li><p>Also, if it hangs during <code>build_mlp()<\/code> then clearly the problem is not in how you read images.  Try using the original version of <code>build_mlp()<\/code> with your images.  Try running it by itself. Skip the image reading, just call <code>build_mlp()<\/code> with constant parameters.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":36556868,"AnswerCount":2,"Tags":"<python><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2016-04-11T18:50:46.280","AcceptedAnswerId":36557055.0,"OwnerUserId":4451856.0,"Title":"TypeError: 'numpy.float64' object is not iterable Keras","Body":"<p>The following line of code gives me the above error in Keras. model is a Graph model in Keras.<\/p>\n\n<pre><\/pre>\n\n<p>but when I change it to the following, it runs fine.<\/p>\n\n<pre><\/pre>\n\n<p>The Y_test here is a  of . A one-hot encoded vector. <\/p>\n\n<p>Sample :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36640091","CreationDate":"2016-04-15T06:50:37.637","ParentId":null,"OwnerUserId":"2219766","Title":null,"Body":"<p>Using keras 1.0 I was able to get the score print out but I would receive this error when I tried to get the score and accuracy.<\/p>\n\n<p>I downgraded my keras to 0.3.3, and ran the exact same code. I was able to get score and accuracy doing this.<\/p>\n"},{"AnswerId":"36557055","CreationDate":"2016-04-11T19:00:40.160","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>As you can see here :<\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py<\/a><\/p>\n\n<p>The evaluate method returns only test loss (or losses). So assigning result of this method to a pair results in error.<\/p>\n"}]}
{"QuestionId":36557327,"AnswerCount":0,"Tags":"<torch><lstm><cudnn>","CreationDate":"2016-04-11T19:14:57.627","AcceptedAnswerId":null,"OwnerUserId":6079717.0,"Title":"gradients in LSTM neural network (Torch)","Body":"<p>I am using the LSTM language model implemented in <a href=\"https:\/\/github.com\/wojzaremba\/lstm\" rel=\"nofollow\">https:\/\/github.com\/wojzaremba\/lstm<\/a><\/p>\n\n<p>It uses the following lstm function<\/p>\n\n<pre><\/pre>\n\n<p>Which is used in the following network (I removed the softmax and criterion layers and added them separately somewhere else in the code)<\/p>\n\n<pre><\/pre>\n\n<p>The above network returns the output of the network and the state of the lstm layers to be used in the next iteration. The states are saved in a table in the following order { cell_1, output_1, cell_2, output_2 } for a 2 layer lstm network. The network output and output_2 are the same.<\/p>\n\n<p>I have two questions:\n(1) When I apply forward and backward propagation on this network, how are the gradients of the states arranged? Will they have the same order as the above table or are they going to be reversed like this: {grad_cell_2, grad_output_2, grad_cell_1, grad_output_1}<\/p>\n\n<p>I initially thought they would be in the same order as the output table, but I have a reason to suspect that the order is reversed (based on some tests where I manually set the gradients for each iteration). I don't know for sure though and I don't know how to debug this code to know exactly what is going on. <\/p>\n\n<p>(2) In the backward step, if I know the gradients for the output only (which is the same as the last entry in the state table) should I pass the gradients for the output (res) or the state table (next_s) or both? I would think that passing the gradient to the output only or the last entry of the state table only would give me exactly the same result, since the output is simply the last entry in the table. However, I get different results when I try it both ways. <\/p>\n","answers":[]}
{"QuestionId":36559402,"AnswerCount":1,"Tags":"<c++><eclipse><git><tensorflow>","CreationDate":"2016-04-11T21:17:03.543","AcceptedAnswerId":null,"OwnerUserId":5452997.0,"Title":"Trouble compiling Tensorflow in c++\/Eclipse","Body":"<p>I have trouble getting Tensorflow to work in my c++ environment: Eclipse.\nI git cloned tensorflow, built with bazel, and installed the resulting pip-package exactly like described here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/get_started\/os_setup.html#installing-from-sources<\/a><\/p>\n\n<p>I created a new c++ project in Eclipse and added the \/home\/tensorflow to the include paths. However, when I try to build a simple \"hello world\" project with a few includes from TF I run into this error:<\/p>\n\n<pre><\/pre>\n\n<p>It seems to look for a file, which doesn't exist in the Tensorflow directory. Why is it looking for something that doesn't exist? I checked it out and the file really isn't there, even though I followed all steps.<\/p>\n\n<p>My program was very simple:<\/p>\n\n<pre><\/pre>\n\n<p>Where could this have gone wrong?<\/p>\n","answers":[{"AnswerId":"38069216","CreationDate":"2016-06-28T07:01:46.080","ParentId":null,"OwnerUserId":"1444499","Title":null,"Body":"<p>I think you are supposed to use <a href=\"http:\/\/bazel.io\" rel=\"nofollow\">bazel<\/a> instead of cmake and make to build you project.\nFrom what I know graph.pb.h is created by bazel during the build process.<\/p>\n\n<p>Someone else had a similar problem: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1890\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1890<\/a><\/p>\n"}]}
{"QuestionId":36559420,"AnswerCount":1,"Tags":"<python><opencl><theano><lasagne>","CreationDate":"2016-04-11T21:18:25.003","AcceptedAnswerId":36559592.0,"OwnerUserId":6190310.0,"Title":"Using convolutional layers in lasagne with OpenCL","Body":"<p>I have an AMD graphic card, so I have to use OpenCL. After a long time installing I almost get it working, and the only thing I am unable to do is to use convolutional layers. I get an error:<\/p>\n\n<blockquote>\n  <p>AssertionError: AbstractConv2d Theano optimization failed: there is no implementation available supporting the requested options. Did you exclude both \"conv_dnn\" and \"conv_gemm\" from the optimizer? If on GPU, is cuDNN available and does the GPU support it? If on CPU, do you have a BLAS library installed Theano can link against?<\/p>\n<\/blockquote>\n\n<p>So, is there a way to use convolutional layers in lasagne on GPU using OpenCL? <\/p>\n","answers":[{"AnswerId":"36559592","CreationDate":"2016-04-11T21:29:34.490","ParentId":null,"OwnerUserId":"1357527","Title":null,"Body":"<p><a href=\"http:\/\/lasagne.readthedocs.org\/en\/latest\/modules\/layers\/conv.html\" rel=\"nofollow\">The lasagne docs<\/a> note that on compilation on GPU, it will use a <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow\">cuDDN<\/a> implementation, and if this fails it will fall back to a CPU based implementation. Unfortunately, there seems to be no way to use lasagne with a card only supporting OpenCL.<\/p>\n"}]}
{"QuestionId":36559951,"AnswerCount":2,"Tags":"<neural-network><tensorflow><conv-neural-network><tensorboard>","CreationDate":"2016-04-11T21:51:32.807","AcceptedAnswerId":null,"OwnerUserId":2241766.0,"Title":"Improve accuracy on convolutional neural network","Body":"<p>I'm trying to train a convolutional neural network to classify sentences. I take the code from <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>The code works fine in one of my datasets. However, on another dataset, its performance is very poor. Two datasets are comparable in terms of volumes and sentence length. For the dataset with the poor performance, I notice the loss function decreases after steps. Here the blue is for training set and the read is for test set.\n<a href=\"https:\/\/i.stack.imgur.com\/w4PLu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/w4PLu.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And here is the accuracy:\n<a href=\"https:\/\/i.stack.imgur.com\/JCeET.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JCeET.png\" alt=\"enter image description here\"><\/a>\nAs you can see, the loss values on the test set decrease, but the accuracy values are almost the same.\nCan you suggest me how to fix it?\nThank you very much.<\/p>\n","answers":[{"AnswerId":"56332145","CreationDate":"2019-05-27T20:30:15.957","ParentId":null,"OwnerUserId":"11564211","Title":null,"Body":"<p>Try using an exponential decaying learning rate--starts big at first (say 0.1) and then slows down over iterations.<\/p>\n"},{"AnswerId":"36560831","CreationDate":"2016-04-11T23:07:17.550","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Can you try with a smaller learning rate? <\/p>\n"}]}
{"QuestionId":36560664,"AnswerCount":1,"Tags":"<c++><lua><torch>","CreationDate":"2016-04-11T22:50:09.300","AcceptedAnswerId":null,"OwnerUserId":5353679.0,"Title":"Call lua function (with require 'nn') in c++","Body":"<p>I have a Lua function:<\/p>\n\n<pre><\/pre>\n\n<p>I want to call this function in C++, but I am getting an error:<\/p>\n\n<pre><\/pre>\n\n<p>because of the require '...'<\/p>\n\n<p>What should I do, so that C++ can recognize all the packages or other Lua scripts (like data_augmentation)?<\/p>\n","answers":[{"AnswerId":"37670082","CreationDate":"2016-06-07T03:40:42.567","ParentId":null,"OwnerUserId":"3187749","Title":null,"Body":"<p>I was having this same issue on my mac and solved it with <\/p>\n\n<pre><code>source ~\/.profile\n<\/code><\/pre>\n\n<p>then running lua script.<\/p>\n\n<p>Full instructions found here: <a href=\"http:\/\/torch.ch\/docs\/getting-started.html#_\" rel=\"nofollow\">torch.ch<\/a><\/p>\n\n<pre><code># On Linux with bash\nsource ~\/.bashrc\n# On Linux with zsh\nsource ~\/.zshrc\n# On OSX or in Linux with none of the above.\nsource ~\/.profile\n<\/code><\/pre>\n"}]}
{"QuestionId":36561231,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2016-04-11T23:47:48.263","AcceptedAnswerId":null,"OwnerUserId":1322301.0,"Title":"de-Bazel-ing TensorFlow Serving","Body":"<p>While I admire, and am somewhat baffled by, the documentation's commitment to mediating everything related to TensorFlow Serving through Bazel, my understanding of it is tenuous at best. I'd like to minimize my interaction with it.<\/p>\n\n<p>I'm implementing my own TF Serving server by adapting code from the Inception + TF Serving tutorial. I find the BUILD files intimidating enough as it is, and rather than slogging through a lengthy debugging process, I decided to simply edit BUILD to refer to the  file, in lieu of also building the python stuff which (as I understand it?) isn't strictly necessary.<\/p>\n\n<p>However, my functional installation of TF Serving can't be imported into python. With normal TensorFlow you build a  file and install it that way; is there something similar you can do with TF Serving? That way I could keep the construction and exporting of models in the realm of the friendly python interactive shell rather than editing it, crossing all available fingers, building in bazel, and then . <\/p>\n\n<p>Simply adding the directory to my PYTHONPATH has so far been unsuccessful. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"37579943","CreationDate":"2016-06-01T22:38:55.993","ParentId":null,"OwnerUserId":"1576818","Title":null,"Body":"<p>You are close, you need to update the environment as they do in this script<\/p>\n\n<p>...\/serving\/bazel-bin\/tensorflow_serving\/example\/mnist_export<\/p>\n\n<p>I printed out the environment update, did it manually \nexport PYTHONPATH=...<\/p>\n\n<p>then I was able to import tensorflow_serving<\/p>\n"}]}
{"QuestionId":36561417,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2016-04-12T00:08:51.007","AcceptedAnswerId":null,"OwnerUserId":5543312.0,"Title":"tensorflow copy variable but not trainable to pretrain next layers","Body":"<p>I want to implement the autoencoder (to be exact stacked convolutional autoencoder)<\/p>\n\n<p>here I'd like to pretrain each layer first and then fine-tuning<\/p>\n\n<p>So I created variables for weight of each layer<\/p>\n\n<p>ex. W_1 = tf.Variable(initial_value, name,trainable=True etc) for first layer<\/p>\n\n<p>and I pretrained W_1 of first layer<\/p>\n\n<p>Then I want to pretrain weight of second layer (W_2)<\/p>\n\n<p>Here I should use W_1 for calculating input of second layer.<\/p>\n\n<p>However W_1 is trainable so if I use W_1 directly then tensorflow may train W_1 together.<\/p>\n\n<p>So I should create W_1_out that keep value of W_1 but not trainable<\/p>\n\n<p>To be honest I tried to modify code of this site<\/p>\n\n<p><a href=\"https:\/\/github.com\/cmgreen210\/TensorFlowDeepAutoencoder\/blob\/master\/code\/ae\/autoencoder.py\" rel=\"nofollow\">https:\/\/github.com\/cmgreen210\/TensorFlowDeepAutoencoder\/blob\/master\/code\/ae\/autoencoder.py<\/a><\/p>\n\n<p>At line 102 it creates variable by following code<\/p>\n\n<pre><\/pre>\n\n<p>However it calls error cause it use uninitialized value<\/p>\n\n<p>How should I do to copy variable but make it not trainable to pretrain next layers??<\/p>\n","answers":[{"AnswerId":"42069693","CreationDate":"2017-02-06T14:04:31.770","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>Not sure if still relevant, but I'll try anyway.<\/p>\n\n<p>Generally, what I do in a situation like that is the following:<\/p>\n\n<ul>\n<li><p>Populate the (default) graph according to the model you are building, e.g. for the first training step just create the first convolutional layer <code>W1<\/code> you mention. When you train the first layer you can store the saved model once training is finished, then reload it and add the ops required for the second layer <code>W2<\/code>. Or you can just build the whole graph for <code>W1<\/code> from scratch again directly in the code and then add the ops for <code>W2<\/code>. <\/p><\/li>\n<li><p>If you are using the restore mechanism provided by Tensorflow, you will have the advantage that the weights for <code>W1<\/code> are already the pre-trained ones. If you don't use the restore mechanism, you will have to set the <code>W1<\/code> weights manually, e.g. by doing something shown in the snippet further below.<\/p><\/li>\n<li>Then when you set up the training op, you can pass a list of variables as <code>var_list<\/code> to the optimizer which explicitly tells the optimizer which parameters are updated in order to minimize the loss. If this is set to <code>None<\/code> (the default), it just uses what it can find in <code>tf.trainable_variables()<\/code> which in turn is a collection of all <code>tf.Variables<\/code> that are trainable. May be check this <a href=\"https:\/\/stackoverflow.com\/questions\/35180888\/how-to-use-pre-trained-model-as-non-trainable-sub-network-in-tensorflow\">answer<\/a>, too, which basically says the same thing.<\/li>\n<li>When using the <code>var_list<\/code> argument, graph collections come in handy. E.g. you could create a separate graph collection for every layer you want to train. The collection would contain the trainable variables for each layer and then you could very easily just retrieve the required collection and pass it as the <code>var_list<\/code> argument (see example below and\/or the remark in the above linked documentation).<\/li>\n<\/ul>\n\n<p><strong>How to override the value of a variable:<\/strong> <code>name<\/code> is the name of the variable to be overriden, <code>value<\/code> is an array of the appropriate size and type and <code>sess<\/code> is the session: <\/p>\n\n<pre><code>variable = tf.get_default_graph().get_tensor_by_name(name)\nsess.run(tf.assign(variable, value))\n<\/code><\/pre>\n\n<p>Note that the <code>name<\/code> needs an additional <code>:0<\/code> in the end, so e.g. if the weights of your layer are called <code>'weights1'<\/code> the <code>name<\/code> in the example should be <code>'weights1:0'<\/code>.<\/p>\n\n<p><strong>To add a tensor to a custom collection<\/strong>: Use something along the following lines:<\/p>\n\n<pre><code>tf.add_to_collection('layer1_tensors', weights1)\ntf.add_to_collection('layer1_tensors', some_other_trainable_variable)\n<\/code><\/pre>\n\n<p>Note that the first line creates the collection because it does not yet exist and the second line adds the given tensor to the existing collection.<\/p>\n\n<p><strong>How to use the custom collection:<\/strong> Now you can do something like this:<\/p>\n\n<pre><code># loss = some tensorflow op computing the loss\nvar_list = tf.get_collection_ref('layer1_tensors')\noptim = tf.train.AdamOptimizer().minimize(loss=loss, var_list=var_list)\n<\/code><\/pre>\n\n<p>You could also use <code>tf.get_collection('layer_tensors')<\/code> which would return you a copy of the collection.<\/p>\n\n<p>Of course, if you don't wanna do any of this, you could just use <code>trainable=False<\/code> when creating the graph for all variables you don't want to be trainable as you hinted towards in your question. However, I don't like that option too much, because it requires you to pass in booleans into the functions that populate your graph, which is very easily overlooked and thus error-prone. Also, even if you decide to it like that, you would still have to restore the non-trainable variables manually.<\/p>\n"}]}
{"QuestionId":36561423,"AnswerCount":1,"Tags":"<python><optimization><keras><hyperparameters><spearmint>","CreationDate":"2016-04-12T00:09:40.650","AcceptedAnswerId":null,"OwnerUserId":4451856.0,"Title":"Issues integrating Keras with Spearmint","Body":"<p>When i  am using spearmint to optimize the hyperparamaters for Keras models, for the first time it runs fine. But the second job onwards it always throws the following error. <\/p>\n\n<pre><\/pre>\n\n<p>I am using the following code to load the pre-created numpy array of the train data and test data. The following params are passed by the optimization python script. But the set of parameters work fine if run without spearmint.  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36614343","CreationDate":"2016-04-14T05:13:47.343","ParentId":null,"OwnerUserId":"4451856","Title":null,"Body":"<p>The variables that I was setting with the spearmint, they had to be explicitly converted to basic python datatypes, using float(),int(). This helped in solving this issue. <\/p>\n"}]}
{"QuestionId":36565430,"AnswerCount":2,"Tags":"<python><neural-network><tensorflow><deep-learning>","CreationDate":"2016-04-12T06:42:40.600","AcceptedAnswerId":36577144.0,"OwnerUserId":158652.0,"Title":"Adding multiple layers to TensorFlow causes loss function to become Nan","Body":"<p>I'm writing a neural-network classifier in TensorFlow\/Python for the <a href=\"http:\/\/yaroslavvb.blogspot.kr\/2011\/09\/notmnist-dataset.html\">notMNIST<\/a> dataset.  I've implemented l2 regularization and dropout on the hidden layers.  It works fine as long as there is only one hidden layer, but when I added more layers (to improve accuracy), the loss function increases rapidly at each step, becoming NaN by step 5.  I tried temporarily disabling Dropout and L2 regularization, but I get the same behavior as long as there are 2+ layers.  I even rewrote my code from scratch (doing some refactoring to make it more flexible), but with the same results.  The number and size of layers is controlled by .  What am I missing?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"38331221","CreationDate":"2016-07-12T14:06:58.767","ParentId":null,"OwnerUserId":"3192082","Title":null,"Body":"<p>I had the same problem and reducing the batch size and learning rate worked for me.<\/p>\n"},{"AnswerId":"36577144","CreationDate":"2016-04-12T15:03:51.480","ParentId":null,"OwnerUserId":"158652","Title":null,"Body":"<p>Turns out this was not so much a coding issue as a Deep Learning Issue.  The extra layer made the gradients too unstable, and that lead to the loss function quickly devolving to NaN.  The best way to fix this is to use <a href=\"https:\/\/stackoverflow.com\/questions\/33640581\/how-to-do-xavier-initialization-on-tensorflow\">Xavier initialization<\/a>.  Otherwise, the variance of the initial values will tend to be too high, causing instability.  Also, decreasing the learning rate may help.<\/p>\n"}]}
{"QuestionId":36566333,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2016-04-12T07:24:28.797","AcceptedAnswerId":36567652.0,"OwnerUserId":5155533.0,"Title":"How to extract feature for an image after neural network training?","Body":"<p>Is there a way to learn unsupervised features from set of images. Similar to  or , where neural network is learnt and given new document we get its features.<\/p>\n\n<p>Expecting similar to this <a href=\"https:\/\/dato.com\/learn\/gallery\/notebooks\/food_retrieval-public.html\" rel=\"nofollow noreferrer\">example<\/a> shows that it can load learnt nn-model and predict features for new images.<\/p>\n\n<p>Is there any simple example how to implement cnn over images and get their features back will help !!<\/p>\n\n<p>Suppose in this <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"nofollow noreferrer\">example<\/a>\nIf I want to get cnn features for all  and  ... is there any way?\nAlso, if we can get weights per layer per image, we can stack them and use as features. In that case is there a way to get the same.\nUsing those features for unsupervised task would be easier, if we consider them as vectors.<\/p>\n","answers":[{"AnswerId":"36567652","CreationDate":"2016-04-12T08:29:43.787","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>If I correctly understood your question, this task is quite common in a deep learning field. In case of images what I consider the best is a convolutional autoencoder. You may read about this architecture e.g. here <\/p>\n\n<p><a href=\"http:\/\/people.idsia.ch\/~ciresan\/data\/icann2011.pdf\" rel=\"nofollow\">http:\/\/people.idsia.ch\/~ciresan\/data\/icann2011.pdf<\/a><\/p>\n\n<p>Previous version of Keras supported this architecture as one of core layers, though from version 1.0 I noticed that it disappeared from documentation. But - it's still quite easy to build it from a scratch :)<\/p>\n\n<p>In noimage cases there are also another approaches like e.g. Restricted Boltzmann Machines.<\/p>\n\n<p>UPDATE :<\/p>\n\n<p>When it comes to what sort of activations are the best for obtaining new features from neural network activations - from my personal experience - it depends on the size of the net which you use. If you use a network which last layer is wide (has a lot of nodes) it might be useful to get only last layer (due to number of parameters if you want to consider also previous layers - it may harm the performance of learning). But - if (like in case of some MNIST networks) your last layer is not sufficient for this task - you may try using also previous layers activation or even all net activity. To be honest - I'm not expecting much of improvement in this case - but you may try. I think that you should use both approaches - starting from taking only last layer activations - and then trying to check the behaviour of your new classifier when you add activations from previous layers.<\/p>\n\n<p>What I will strongly advise to you is also getting some insights from what sort of features network is learning - using T-SNE embeddings of it activations. In many cases I found it useful - e.g. checking if the size of a layer is sufficient. Using T-SNE you may check if the features obtained from last layer are good discriminators of your classes. It may also give you good insights about your data and what neural networks are really learning (alongside with amazing visualizations :) )<\/p>\n"}]}
{"QuestionId":36568520,"AnswerCount":0,"Tags":"<matlab><caffe>","CreationDate":"2016-04-12T09:06:17.023","AcceptedAnswerId":null,"OwnerUserId":6151595.0,"Title":"Invalid MEX-file '\/home\/hit\/caffe-master\/matlab\/+caffe\/private\/caffe_.mexa64':","Body":"<p>When I run classification_demo.m which is in \/caffe-master\/matlab\/demo,I got the error message,how to fix it? <a href=\"https:\/\/i.stack.imgur.com\/7fVXq.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7fVXq.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
{"QuestionId":36568568,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-04-12T09:07:51.693","AcceptedAnswerId":null,"OwnerUserId":4829166.0,"Title":"Tensorflow doesn't allow for training data to be used in (features,targets) style?","Body":"<p>I have  that is a ndarray containing features and targets with different dimensions respectively. This seems to give tensorflow problems.<\/p>\n\n<p>If I have a function:<\/p>\n\n<pre><\/pre>\n\n<p>This results in .<\/p>\n\n<p>It seems to be because  is not recognizing my inputs as numpy arrays. I think this is because features and targets have different dimensions and  has problems with that; if I have  with 100 pairs then its shape is . If I then slice it:  and , so the length of the feature\/target vectors is not recognized even after slicing.<\/p>\n\n<p>I've got around the problem by splitting  up beforehand into  and , who's shapes are returned correctly.<\/p>\n\n<p>My question is, is this normal - is this supposed to be like this? Or am I just doing something wrong? It would be nicer to just work with  instead of passing two variables around all the time.<\/p>\n\n<p>edit: <\/p>\n\n<pre><\/pre>\n\n<p>The dimensions of  should be self-explanatory.<\/p>\n","answers":[]}
{"QuestionId":36568951,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2016-04-12T09:22:55.153","AcceptedAnswerId":null,"OwnerUserId":5707306.0,"Title":"Does TensorBoard (TensorFlow) have the features to add labels for axes and legends on plots? If so, how?","Body":"<p>I need to know how to add legend and label axes on TensorBoard (TensorFlow) graphs. Please help, thanks!<\/p>\n","answers":[{"AnswerId":"44190531","CreationDate":"2017-05-25T22:06:29.067","ParentId":null,"OwnerUserId":"4212158","Title":null,"Body":"<p>As of today (5\/25\/17), no. Someone submitted this as a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7631\" rel=\"noreferrer\">feature request<\/a>, and Dandelion Man\u00e9 (from the Tensorflow dev team) labeled the issue as \"contributions welcome\".<\/p>\n"}]}
{"QuestionId":36569615,"AnswerCount":1,"Tags":"<gpu><tensorflow>","CreationDate":"2016-04-12T09:48:37.847","AcceptedAnswerId":36657555.0,"OwnerUserId":3489538.0,"Title":"tensorflow transfer learning problems to run on GPU","Body":"<p>I am trying to run the retrain.py script (available here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py<\/a>). I have noticed that the part starting with the line 747 is executed on CPU when the default should be GPU. So, I have added the following line to force it to work on GPU: <\/p>\n\n<pre><\/pre>\n\n<p>It causes the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1140\" rel=\"nofollow\">here<\/a> that it might be a problem that mean is not implemented on GPU but on the other hand there is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/reduction_ops_mean.cc\" rel=\"nofollow\">commit<\/a> on github which fixes counting mean on GPU.<\/p>\n\n<p>Previous part, e.g. generating bottlenecks (line 744) runs perfectly on GPU, without even forcing it.<\/p>\n\n<p>I would be grateful for any help!!<\/p>\n\n<p>Justyna<\/p>\n","answers":[{"AnswerId":"36657555","CreationDate":"2016-04-15T22:00:42.880","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This has now been fixed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/b874e2ccfda77c964bfb0a41a8dc339e320b185c\" rel=\"nofollow\">b874e2c<\/a>, nice catch<\/p>\n"}]}
{"QuestionId":36570729,"AnswerCount":0,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-04-12T10:35:30.797","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"tf.IndexedSlicesValue when returned from tf.gradients()","Body":"<p>I'm having the following problem, I have four embedding matrices and want to get the gradients of my loss function with respect to those matrices.<\/p>\n\n<p>When I run the session to return the values for the gradients, two of those returned objects are of type tensorflow.python.framework.ops.IndexedSlicesValue, the other two are numpy arrays. Now for the numpy arrays, their shape corresponds to the shape of their corresponding embedding matrix, but I'm having problems with the IndexedSlicesValue objects. <\/p>\n\n<p>If I call .values on one of those objects, I get an array whose shape does not match that of the gradient, the shape of the embedding matrix is [22,30], but calling .values on the IndexedSlicesValue object I get an array with shape [4200,30] ( The shape of my input tensor had dimensions of [30,20,7], the product of those dimensions equals 4200, not sure if this is relevant).\nThe IndexedSlicesValue object has an attribute called dense_shape, which is an array that holds the dimensions the gradient should have, i.e. array([22,30]) is value returned by .dense_shape.<\/p>\n\n<p>I don't really understand the docs here: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#IndexedSlices\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/api_docs\/python\/state_ops.html#IndexedSlices<\/a><\/p>\n\n<p>It says:<\/p>\n\n<blockquote>\n  <p>An IndexedSlices is typically used to represent a subset of a\n  larger tensor dense of shape [LARGE0, D1, .. , DN] where LARGE0 >> D0.\n  The values in indices are the indices in the first dimension of the\n  slices that have been extracted from the larger tensor.<\/p>\n<\/blockquote>\n\n<p>So this array of shape (4200,30) is extracted from an array corresponding to an even larger, dense tensor?<\/p>\n\n<p>What exactly is the gradient in this IndexedSlicesValue object and why does tensorflow automatically use this type for some gradients returned by tf.gradients()?<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":36570904,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-04-12T10:42:46.410","AcceptedAnswerId":36573850.0,"OwnerUserId":3559686.0,"Title":"How to define weight decay for individual layers in TensorFlow?","Body":"<p>In CUDA ConvNet, we can write something like this (<a href=\"https:\/\/code.google.com\/p\/cuda-convnet\/wiki\/LayerParams\" rel=\"noreferrer\">source<\/a>) for each layer:<\/p>\n\n<pre><\/pre>\n\n<p>where  refers to the L2 weight decay.<\/p>\n\n<p>How can the same be achieved in TensorFlow?<\/p>\n","answers":[{"AnswerId":"36573850","CreationDate":"2016-04-12T12:50:53.883","ParentId":null,"OwnerUserId":"234167","Title":null,"Body":"<p>You can add all the variables you want to add weight decay to, to a collection name 'variables' and then you calculate the L2 norm weight decay for the whole collection.<\/p>\n\n<pre><code>  # Create your variables\n  weights = tf.get_variable('weights', collections=['variables'])\n\n  with tf.variable_scope('weights_norm') as scope:\n    weights_norm = tf.reduce_sum(\n      input_tensor = WEIGHT_DECAY_FACTOR*tf.pack(\n          [tf.nn.l2_loss(i) for i in tf.get_collection('weights')]\n      ),\n      name='weights_norm'\n  )\n\n  # Add the weight decay loss to another collection called losses\n  tf.add_to_collection('losses', weights_norm)\n\n  # Add the other loss components to the collection losses     \n  # ...\n\n  # To calculate your total loss\n  tf.add_n(tf.get_collection('losses'), name='total_loss')\n<\/code><\/pre>\n"},{"AnswerId":"50780869","CreationDate":"2018-06-10T05:37:44.840","ParentId":null,"OwnerUserId":"2366315","Title":null,"Body":"<p>Both current answers are wrong in that they do not give you \"weight decay as in cuda-convnet\" but instead L2-regularization, which is different.<\/p>\n\n<p>When using pure SGD (without momentum) as an optimizer, weight decay is the same thing as adding a L2-regularization term to the loss. <strong>When using any other optimizer, this is not true.<\/strong><\/p>\n\n<p>Weight decay (don't know how to TeX here, so excuse my pseudo-notation):<\/p>\n\n<pre><code>w[t+1] = w[t] - learning_rate * dw - weight_decay * w\n<\/code><\/pre>\n\n<p>L2-regularization:<\/p>\n\n<pre><code>loss = actual_loss + lambda * 1\/2 sum(||w||_2 for w in network_params)\n<\/code><\/pre>\n\n<p>Computing the gradient of the extra term in L2-regularization gives <code>lambda * w<\/code> and thus inserting it into the SGD update equation<\/p>\n\n<pre><code>dloss_dw = dactual_loss_dw + lambda * w\nw[t+1] = w[t] - learning_rate * dw\n<\/code><\/pre>\n\n<p>gives the same as weight decay, but mixes <code>lambda<\/code> with the <code>learning_rate<\/code>. Any other optimizer, even SGD with momentum, gives a different update rule for weight decay as for L2-regularization! See the paper <a href=\"\/https:\/\/arxiv.org\/abs\/1711.05101\">Fixing weight decay in Adam<\/a> for more details. (Edit: AFAIK, <a href=\"http:\/\/www.cs.toronto.edu\/~hinton\/absps\/parle.pdf\" rel=\"nofollow noreferrer\">this 1987 Hinton paper<\/a> introduced \"weight decay\", literally as \"each time the weights are updated, their magnitude is also decremented by 0.4%\" at page 10)<\/p>\n\n<p>That being said, there doesn't seem to be support for \"proper\" weight decay in TensorFlow yet. There are a few issues discussing it, specifically because of above paper.<\/p>\n\n<p>One possible way to implement it is by writing an op that does the decay step manually after every optimizer step. A different way, which is what I'm currently doing, is using an additional SGD optimizer just for the weight decay, and \"attaching\" it to your <code>train_op<\/code>. Both of these are just crude work-arounds, though. My current code:<\/p>\n\n<pre><code># In the network definition:\nwith arg_scope([layers.conv2d, layers.dense],\n               weights_regularizer=layers.l2_regularizer(weight_decay)):\n    # define the network.\n\nloss = # compute the actual loss of your problem.\ntrain_op = optimizer.minimize(loss, global_step=global_step)\nif args.weight_decay not in (None, 0):\n    with tf.control_dependencies([train_op]):\n        sgd = tf.train.GradientDescentOptimizer(learning_rate=1.0)\n        train_op = sgd.minimize(tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)))\n<\/code><\/pre>\n\n<p>This somewhat makes use of TensorFlow's provided bookkeeping. Note that the <code>arg_scope<\/code> takes care of appending an L2-regularization term for every layer to the <code>REGULARIZATION_LOSSES<\/code> graph-key, which I then all sum up and optimize using SGD which, as shown above, corresponds to actual weight-decay.<\/p>\n\n<p>Hope that helps, and if anyone gets a nicer code snippet for this, or TensorFlow implements it better (i.e. in the optimizers), please share.<\/p>\n\n<p><strong>Edit:<\/strong> see also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/17438\" rel=\"nofollow noreferrer\">this PR<\/a> which just got merged into TF.<\/p>\n"},{"AnswerId":"44886139","CreationDate":"2017-07-03T12:53:54.143","ParentId":null,"OwnerUserId":"8246532","Title":null,"Body":"<pre><code>get_variable(\nname,\nshape=None,\ndtype=None,\ninitializer=None,\nregularizer=None,\ntrainable=True,\ncollections=None,\ncaching_device=None,\npartitioner=None,\nvalidate_shape=True,\nuse_resource=None,\ncustom_getter=None)\n<\/code><\/pre>\n\n<p>This is the usage of tensorflow function <code>get_variable<\/code>. You can easily specify the regularizer to do weight decay.<\/p>\n\n<p>Following is an example:<\/p>\n\n<pre><code>weight_decay = tf.constant(0.0005, dtype=tf.float32) # your weight decay rate, must be a scalar tensor.\nW = tf.get_variable(name='weight', shape=[4, 4, 256, 512], regularizer=tf.contrib.layers.l2_regularizer(weight_decay))\n<\/code><\/pre>\n"}]}
{"QuestionId":36571225,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-12T10:56:50.930","AcceptedAnswerId":36574034.0,"OwnerUserId":3559686.0,"Title":"Printing the value of loss that is minimized by an optimizer in TensorFlow","Body":"<p>I want to print the value of loss that has been minimized by an optimizer. Here is an example:<\/p>\n\n<pre><\/pre>\n\n<p>Now that I have minimized , how do I print its minimized value?<\/p>\n","answers":[{"AnswerId":"36574034","CreationDate":"2016-04-12T12:58:08.477","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>The easiest way to visualize a loss is to create a scalar summary of it:<\/p>\n\n<pre><code>mean_squared_error = tf.reduce_mean(tf.square(tf.sub(predictions, training_outputs)))\nloss_summ = tf.scalar_summary(\"loss\", mean_squared_error)\n<\/code><\/pre>\n\n<p>You then create a writer in the TensorFlow session, and add the summary <code>loss_summ<\/code> to the <code>sess.run()<\/code> call. You then get the value back in <code>mse_val<\/code> and can print it.<\/p>\n\n<pre><code>with tf.Session() as session:\n    writer = tf.train.SummaryWriter(\"log\", session.graph_def)\n    session.run(initialize)\n\n    print(\"Begin training...\")\n    _, mse_val, summ = session.run([train_step, mean_squared_error, loss_summ], feed_dict={training_inputs: features, training_outputs: labels})\n    writer.add_summary(summ)\n    print(\"Finished training! The mean squared error is: %f\" % mse_val)\n<\/code><\/pre>\n\n<p>As a bonus, you can even visualize the evolution of the loss in TensorBoard, by running <code>tensorboard --logdir log<\/code> (read <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">this tutorial<\/a> for more details).<\/p>\n\n<p>P.S: your code only runs 1 iteration of the training, you may want to add a loop.<\/p>\n"}]}
{"QuestionId":36574531,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><data-loss>","CreationDate":"2016-04-12T13:17:40.880","AcceptedAnswerId":null,"OwnerUserId":2235817.0,"Title":"Is Kullback Liebler Divergence already implented in TensorFlow?","Body":"<p>I am working with tensorflow and using Nueral Networks to solve multi-label classification problem. I was using Softmax cross entropy as my loss function:<\/p>\n\n<pre><\/pre>\n\n<p>Now, i thought that i should use KL divergence loss function. But, i didn't find it in tensorflow can any body help me to use KL Divergence loss function instead of Softmax loss?<\/p>\n","answers":[{"AnswerId":"40176919","CreationDate":"2016-10-21T12:39:36.577","ParentId":null,"OwnerUserId":"5687149","Title":null,"Body":"<p>Here you go:<\/p>\n\n<pre><code>tf.contrib.distributions.kl(distribution_1, distribution_2)\n<\/code><\/pre>\n"}]}
{"QuestionId":36576229,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2016-04-12T14:26:25.027","AcceptedAnswerId":36612666.0,"OwnerUserId":4308982.0,"Title":"caffe.io.Transformer to use for LeNet classification","Body":"<p>I am trying to classify digits (0-9), using LeNet.<\/p>\n\n<p>I read images of dimension 28x28, transpose them, and send one color channel into the net.<\/p>\n\n<pre><\/pre>\n\n<p>And get predictions that make no sense whatsoever.\nI see that PyCaffe has a Transformer that gets this job done but I don't know how I would use it for LeNet.<\/p>\n\n<hr>\n\n<p>I am using the net from caffe\/examples\/mnist (hopefully that answers what sort of data the net expects).<\/p>\n\n<p>The only thing that's different is that the input is now:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"36612666","CreationDate":"2016-04-14T02:12:53.523","ParentId":null,"OwnerUserId":"6998621","Title":null,"Body":"<p>Since you are directly using the deploy prototxt, you should take a look at the \"data\" layer in lenet_train_test.prototxt (or a file with a similar name) which will give you an indication of how data is preprocessed.<\/p>\n\n<p>By looking at the files I have, it seems that the LeNet example performs scaling of all input data. This is indicated by the following lines in \"data\" layer <\/p>\n\n<pre><code>transform_param {\n    scale: 0.00390625\n}\n<\/code><\/pre>\n\n<p>So that's one thing you might be missing. Have a look at <a href=\"http:\/\/nbviewer.jupyter.org\/github\/BVLC\/caffe\/blob\/master\/examples\/00-classification.ipynb\" rel=\"nofollow\">imagenet tutorial<\/a> for pointers on preprocessing (esp Transformer). Second, you are reading in the image as <\/p>\n\n<pre><code>bild = np.transpose(bild,(2,0,1))\n<\/code><\/pre>\n\n<p>This looks like an error based on the channel swaps you are performing, so make sure that the data sent to the network has the dimensions - 1 (batch size) x 1 (channels) x 28(height) x 28(width).<\/p>\n"}]}
{"QuestionId":36576571,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><tensorboard>","CreationDate":"2016-04-12T14:40:15.977","AcceptedAnswerId":36576914.0,"OwnerUserId":6193863.0,"Title":"Tensorboad - add_summary makes my code crashes","Body":"<p>I am very new to tensorflow, and I try to display my first tensorboard. <\/p>\n\n<p>I downloaded and executed ok the board for the example given here\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/how_tos\/summaries_and_tensorboard\/index.html<\/a><\/p>\n\n<p>Following the method, I have in my code:<\/p>\n\n<pre><\/pre>\n\n<p>and when I run the session<\/p>\n\n<pre><\/pre>\n\n<p>The process crashes with the following error<\/p>\n\n<pre><\/pre>\n\n<p>What am I missing ? \nAny help\/comment would be very welcome<\/p>\n\n<p>Thank you very much for the help<\/p>\n\n<p>K. <\/p>\n","answers":[{"AnswerId":"36576914","CreationDate":"2016-04-12T14:54:09.600","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>You should write: <\/p>\n\n<pre><code>_, summary_str, l, predictions = session.run(\n  [optimizer, merged, loss, train_prediction], feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>I added a 4th argument <code>merged<\/code> which corresponds to the summary you are trying to get (you were only getting the result of the optimization step).<\/p>\n"}]}
{"QuestionId":36583603,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-04-12T20:38:29.680","AcceptedAnswerId":null,"OwnerUserId":378469.0,"Title":"Is it possible to access the shape of tensorflow tensors without running them?","Body":"<p>I think there is a shape inference mechanism, but I'm not sure how I can access the inferred shapes.<\/p>\n","answers":[{"AnswerId":"36583728","CreationDate":"2016-04-12T20:45:40.047","ParentId":null,"OwnerUserId":"378469","Title":null,"Body":"<p>Rather than using the shape op on the tensor, just call tensor.get_shape().<\/p>\n"}]}
{"QuestionId":36584770,"AnswerCount":0,"Tags":"<tensorflow><quantization>","CreationDate":"2016-04-12T21:54:48.850","AcceptedAnswerId":null,"OwnerUserId":5587428.0,"Title":"Quantize Embeddings","Body":"<p>I would like to quantize embeddings to a single signed byte in each dimension. If I try to do this by scaling the values to [-127, 128], then casting to tf.int8, recasting to tf.float32 and rescaling to the original [-1, 1] range, I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: No gradients provided for any variable<\/p>\n<\/blockquote>\n\n<p>The same training script works fine without the quantization step.<\/p>\n\n<p>According to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/15\" rel=\"nofollow\">this thread<\/a>, quantized ops will be a future feature of tensorflow. In the mean time is there a good work-around for this simple quantization scenario?<\/p>\n","answers":[]}
{"QuestionId":36584907,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-04-12T22:06:22.107","AcceptedAnswerId":56436392.0,"OwnerUserId":3834415.0,"Title":"Tensor flow toggle between CPU\/GPU","Body":"<p>Having installed tensorflow GPU (running on a measly NVIDIA GeForce 950), I would like to compare performance with the CPU. <\/p>\n\n<p>I am running the tensorFlow MNIST tutorial code, and have noticed a dramatic increase in speed--estimated anyways (I ran the CPU version 2 days ago on a laptop i7 with a batch size of 100, and this on a desktop GPU, batch size of 10)--between the CPU and the GPU when I switched...but I only noticed the speed increase when I lowered the batch size on the GPU to 10 from 100...<\/p>\n\n<p>Now I lack an objective measure for what I am gaining.  <\/p>\n\n<p>Is there a way to toggle between the CPU and GPU tensor flows?<\/p>\n","answers":[{"AnswerId":"56436392","CreationDate":"2019-06-04T01:48:25.680","ParentId":null,"OwnerUserId":"3503565","Title":null,"Body":"<pre><code># Check if the server\/ instance is having GPU\/ CPU from python code\nimport sys\nimport tensorflow as tf\n# device_lib.list_local_devices()     ## this command list all the processing device GPU and CPU\n\n\ndevice_name = [x.name for x in device_lib.list_local_devices() if x.device_type == 'GPU']\nif device_name[0] == \"\/device:GPU:0\":\n    device_name = \"\/gpu:0\"\n    #print('GPU')\nelse:\n    #print('CPU')\n    device_name = \"\/cpu:0\"\n\nwith tf.device(device_name):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n    c = tf.matmul(a, b)\nwith tf.Session() as sess:\n    print (sess.run(c))    \n<\/code><\/pre>\n"},{"AnswerId":"41329803","CreationDate":"2016-12-26T10:26:12.157","ParentId":null,"OwnerUserId":"1468002","Title":null,"Body":"<p>Another option is to install the cpu version and the gpu version of tensorflow in two virtual environments, detailed instructions on how to install tensorflow in virtual environments are listed here <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup<\/a>; in this way, you can have the same code running in two terminals windows one uses CPU and the other uses GPU.<\/p>\n"},{"AnswerId":"39126114","CreationDate":"2016-08-24T14:28:54.553","ParentId":null,"OwnerUserId":"4282481","Title":null,"Body":"<p>try setting tf.device to cpu:0    <\/p>\n\n<pre><code>with tf.Session() as sess:\n     with tf.device(\"\/cpu:0\"):\n<\/code><\/pre>\n"},{"AnswerId":"36585271","CreationDate":"2016-04-12T22:33:06.257","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>To make GPU invisible<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=\"\"\n<\/code><\/pre>\n\n<p>To return to normal<\/p>\n\n<pre><code>unset CUDA_VISIBLE_DEVICES\n<\/code><\/pre>\n"}]}
{"QuestionId":36585017,"AnswerCount":0,"Tags":"<performance><tensorflow>","CreationDate":"2016-04-12T22:14:49.633","AcceptedAnswerId":null,"OwnerUserId":1238847.0,"Title":"For Tensorflow Inception v3 examples, Why C++ classifier runs much slower than python one?","Body":"<p>I'm running example \/ how-to given in Tensorflow documentation that are given at <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/image_recognition\/index.html\" rel=\"nofollow\">https:\/\/www.tensorflow.org\/versions\/r0.7\/tutorials\/image_recognition\/index.html<\/a><\/p>\n\n<p>For some reason, python classifier (.\/tensorflow\/models\/image\/imagenet\/classify_image.py) runs much faster than C++ classifier (.\/tensorflow\/examples\/label_image\/main.cc).  I expected otherwise.  Can anybody tell me why it is the case?<\/p>\n\n<ul>\n<li>environment: windows8.1 pro (core i7) + HyperV + Docker environment (without GPU).<\/li>\n<li>label_image (C++ example) takes 30sec.<\/li>\n<li>classify_image (python example) takes 3sec.<\/li>\n<li>I run once before measurement to make sure data is downloaded.<\/li>\n<li>I see a lot of Warning at label_image runtime.\n\n<blockquote>\n  <p>W tensorflow\/core\/kernels\/batch_norm_op.cc:36] Op is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().<\/p>\n<\/blockquote><\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":36585071,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-04-12T22:18:13.677","AcceptedAnswerId":36585161.0,"OwnerUserId":3471004.0,"Title":"Cannot feed value of shape (500,) for Tensor 'x_17:0', which has shape '(?, 500)'","Body":"<p>I'm just learning TensorFlow, so sorry if this is obvious. I've checked the documentation and experimented quite a bit and I just can't seem to get this to work.<\/p>\n\n<pre><\/pre>\n\n<p>yield_financials() outputs an numpy array of 500 numbers and the number that I want it to guess. I've tried shuffling OUT_DIMS and FIN_SIZE around, I tried accumulating them into batches to more closely match what the tutorial looked like, I tried setting OUT_DIMS to 0, removing it entirely, and I tried replacing None with other numbers, but have not made any progress.<\/p>\n","answers":[{"AnswerId":"40858994","CreationDate":"2016-11-29T06:18:06.327","ParentId":null,"OwnerUserId":"7223806","Title":null,"Body":"<p>I had the same problem and I solved this problem.I hope that it's helpful for u.<\/p>\n\n<p>firstly,I transformed load data into :<\/p>\n\n<pre><code>train_data = np.genfromtxt(train_data1, delimiter=',')\ntrain_label = np.transpose(train_label1, delimiter=',')\ntest_data = np.genfromtxt(test_data1, delimiter=',')\ntest_label = np.transpose(test_label1, delimiter=',')\n<\/code><\/pre>\n\n<p>Then,transformed trX, trY, teX, teY data into:<\/p>\n\n<pre><code># convert the data\ntrX, trY, teX, teY = train_data,train_label, test_data, test_label\ntemp = trY.shape\ntrY = trY.reshape(temp[0], 1)\ntrY = np.concatenate((1-trY, trY), axis=1)\ntemp = teY.shape\nteY = teY.reshape(temp[0], 1)\nteY = np.concatenate((1-teY, teY), axis=1)\n<\/code><\/pre>\n\n<p>Finally,I transformed launching the graph in a session into:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    # you need to initialize all variables\n    tf.initialize_all_variables().run()\n\n    for i in range(100):\n            sess.run(train_op, feed_dict={X:  trX, Y: trY})        \n            print(i, np.mean(np.argmax(teY, axis=1) == sess.run(predict_op, feed_dict={X: teX})))\n<\/code><\/pre>\n\n<p>That's all.<\/p>\n"},{"AnswerId":"36585161","CreationDate":"2016-04-12T22:24:18.570","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Try <\/p>\n\n<pre><code>    this_x = np.reshape(this_x,(1, FIN_SIZE))\n    sess.run(train_step, feed_dict={x:    this_x,\n                                    yhat: this_y})\n<\/code><\/pre>\n"}]}